diff --git a/.clang-format b/.clang-format index aff93435f58c522f5ed1090aef2005f76e91cf31..8b5830627348c6bff12260b7d9adbd357f074718 100644 --- a/.clang-format +++ b/.clang-format @@ -19,7 +19,7 @@ BasedOnStyle: Google IndentWidth: 2 TabWidth: 2 ContinuationIndentWidth: 4 -AccessModifierOffset: -2 # The private/protected/public has no indent in class +AccessModifierOffset: -1 # The private/protected/public has no indent in class Standard: Cpp11 AllowAllParametersOfDeclarationOnNextLine: true BinPackParameters: false diff --git a/.copyright.hook b/.copyright.hook index 09afff2072df3384a429d01d06188218ae6e85d1..86b16ebdc46047c7cb3d7731a71cbf9647a1f2fe 100644 --- a/.copyright.hook +++ b/.copyright.hook @@ -9,7 +9,7 @@ import subprocess import platform COPYRIGHT = ''' - Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. diff --git a/.pre-commit-config.yaml b/.pre-commit-config.yaml index 6140340890c0e5025eb08209e8ea78df918b4dc0..eeda759ff18ccb86ce6a585fe41cb972ea3ae295 100644 --- a/.pre-commit-config.yaml +++ b/.pre-commit-config.yaml @@ -34,6 +34,14 @@ repos: entry: bash ./tools/codestyle/cpplint_pre_commit.hook language: system files: \.(c|cc|cxx|cpp|cu|h|hpp|hxx)$ +- repo: local + hooks: + - id: pylint-doc-string + name: pylint + description: Check python docstring style using docstring_checker. + entry: bash ./tools/codestyle/pylint_pre_commit.hook + language: system + files: \.(py)$ - repo: https://github.com/PaddlePaddle/pre-commit-golang sha: 8337620115c25ff8333f1b1a493bd031049bd7c0 hooks: diff --git a/.travis.yml b/.travis.yml index 929c847bd36d64e79a199b2634ebf68c3225429b..8c772030925dcad3909f142b08e4d8057a3f89b7 100644 --- a/.travis.yml +++ b/.travis.yml @@ -12,45 +12,24 @@ services: os: - linux env: - - JOB=build_doc + - JOB=doc - JOB=check_style - JOB=build_android addons: - apt: - packages: - - gcc-4.8 - - g++-4.8 - - git - - build-essential - - python - - python-pip - - python2.7-dev - - python-wheel - - libboost-dev - - curl - - swig - - graphviz - - clang-format-3.8 - - automake - - libtool - - ccache ssh_known_hosts: 13.229.163.131 before_install: - - if [[ "$JOB" == "check_style" ]]; then sudo ln -s /usr/bin/clang-format-3.8 /usr/bin/clang-format; fi - # Paddle is using protobuf 3.1 currently. Protobuf 3.2 breaks the compatibility. So we specify the python - # protobuf version. - - sudo pip install -r $TRAVIS_BUILD_DIR/python/requirements.txt - - sudo pip install wheel sphinx==1.5.6 recommonmark sphinx-rtd-theme==0.1.9 virtualenv pre-commit LinkChecker + # For pylint dockstring checker + - sudo pip install pylint pytest astroid isort - | function timeout() { perl -e 'alarm shift; exec @ARGV' "$@"; } script: - | # 43min timeout - if [[ "$JOB" == "build_android" ]]; then timeout 2580 docker run -it --rm -v "$TRAVIS_BUILD_DIR:/paddle" paddlepaddle/paddle:latest-dev-android; - else timeout 2580 paddle/scripts/travis/${JOB}.sh; fi; - RESULT=$?; if [ $RESULT -eq 0 ] || [ $RESULT -eq 142 ]; then true; else exit 1; fi; + paddle/scripts/paddle_docker_build.sh ${JOB} + if [ $? -eq 0 ] || [ $? -eq 142 ]; then true; else exit 1; fi; - | - if [[ "$JOB" != "build_doc" ]]; then exit 0; fi; + if [[ "$JOB" != "doc" ]]; then exit 0; fi; + # For document only if [[ "$TRAVIS_PULL_REQUEST" != "false" ]]; then exit 0; fi; if [[ "$TRAVIS_BRANCH" != "develop" && ! "$TRAVIS_BRANCH" =~ ^v[[:digit:]]+\.[[:digit:]]+(\.[[:digit:]]+)?(-\S*)?$ ]]; then exit 0; fi; export DEPLOY_DOCS_SH=https://raw.githubusercontent.com/PaddlePaddle/PaddlePaddle.org/master/scripts/deploy/deploy_docs.sh diff --git a/AUTHORS.md b/AUTHORS.md index 9c6821d9f8681c5907c2fc9938fdb62ba64b9a92..4ee05420982d13f686cf13e8957ce41dfcdd2cb8 100644 --- a/AUTHORS.md +++ b/AUTHORS.md @@ -2,12 +2,14 @@ |---|---| | abhinavarora | Abhinav Arora | | backyes | Yan-Fei Wang | +| baiyfbupt | Yi-Fan Bai | | beckett1124 | Bin Qi | -| JiayiFeng | Jia-Yi Feng | | chengxiaohua1105 | Xiao-Hua Cheng | | cxwangyi, yiwangbaidu, wangkuiyi | Yi Wang | | cxysteven | Xing-Yi Cheng | | dzhwinter | Zhi-Hong Dong | +| dragonwarrior | Long Wang | +| dyning | Yuning Du | | emailweixu | Wei Xu | | gangliao | Gang Liao | | gongweibao | Wei-Bao Gong | @@ -16,6 +18,9 @@ | hedaoyuan | Dao-Yuan He | | helinwang | He-Lin Wang | | jacquesqiao | Long-Fei Qiao | +| jczaja | Jacek Czaja | +| JiayiFeng | Jia-Yi Feng | +| kbinias | Krzysztof Binias | | kuke | Yi-Bing Liu | | lcy-seso | Ying Cao | | lipeng-unisound | Peng Li | @@ -24,15 +29,20 @@ | llxxxll | Yong-Feng Liu | | luotao01 | Tao Luo | | lzhao4ever | Liang Zhao | +| mozga-intel | Mateusz Ozga | | NHZlX | Zhao-Long Xing | +| Noplz | Yuan Gao | | pakchoi | Chuan-Jiang Song | +| panyx0718 | Xin Pan | | pengli09 | Peng Li | | pkuyym | Ya-Ming Yang | +| pzelazko-intel | Pawel Zelazko | | QiJune | Jun Qi | | qingqing01 | Qing-Qing Dang | | reyoung | Yang Yu | | Superjom | Chun-Wei Yan | | tianbingsz | Tian-Bing Xu | +| tpatejko | Tomasz Patejko | | typhoonzero | Yi Wu | | wanghaoshuang | Hao-Shuang Wang | | wangyang59 | Yang Wang | diff --git a/CMakeLists.txt b/CMakeLists.txt index 23bbe829ac16180088bfa37df66e23f19b021ea3..cfaab206e1f321a55119d4a8d65c4a99d3819fff 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -25,7 +25,6 @@ message(STATUS "CXX compiler: ${CMAKE_CXX_COMPILER}, version: " message(STATUS "C compiler: ${CMAKE_C_COMPILER}, version: " "${CMAKE_C_COMPILER_ID} ${CMAKE_C_COMPILER_VERSION}") -find_package(Sphinx) if(NOT CMAKE_CROSSCOMPILING) find_package(CUDA QUIET) endif(NOT CMAKE_CROSSCOMPILING) @@ -42,7 +41,6 @@ option(WITH_MKL "Compile PaddlePaddle with MKL support." ${AVX_FO option(WITH_DSO "Compile PaddlePaddle with dynamic linked CUDA" ON) option(WITH_TESTING "Compile PaddlePaddle with unit testing" OFF) option(WITH_SWIG_PY "Compile PaddlePaddle with inference api" ON) -option(WITH_STYLE_CHECK "Compile PaddlePaddle with style check" ON) option(WITH_PYTHON "Compile PaddlePaddle with python interpreter" ON) option(WITH_DOUBLE "Compile PaddlePaddle with double precision" OFF) option(WITH_RDMA "Compile PaddlePaddle with RDMA support" OFF) @@ -59,8 +57,10 @@ option(GLIDE_INSTALL "Download and install go dependencies " ON) option(USE_NNPACK "Compile PaddlePaddle with NNPACK library" OFF) option(WITH_DISTRIBUTE "Compile with grpc distributed support" OFF) option(USE_EIGEN_FOR_BLAS "Use matrix multiplication in Eigen" OFF) +option(EIGEN_USE_THREADS "Compile with multi-threaded Eigen" OFF) option(WITH_ARM_FP16 "Use half precision support on armv8.2-a cpu" OFF) option(WITH_FAST_BUNDLE_TEST "Bundle tests that can be run in a single process together to reduce launch overhead" OFF) +option(WITH_CONTRIB "Compile the third-party contributation" OFF) # CMAKE_BUILD_TYPE if(NOT CMAKE_BUILD_TYPE) @@ -101,6 +101,9 @@ endif() set(THIRD_PARTY_PATH "${CMAKE_BINARY_DIR}/third_party" CACHE STRING "A path setting third party libraries download & build directories.") +set(FLUID_INSTALL_DIR "${CMAKE_BINARY_DIR}/fluid_install_dir" CACHE STRING + "A path setting fluid shared and static libraries") + if (WITH_C_API AND WITH_PYTHON) message(WARNING "It is suggest not embedded a python interpreter in Paddle " "when using C-API. It will give an unpredictable behavior when using a " @@ -118,13 +121,14 @@ else() endif() set(WITH_MKLML ${WITH_MKL}) -if (WITH_MKL AND AVX2_FOUND) - set(WITH_MKLDNN ON) -else() - message(STATUS "Do not have AVX2 intrinsics and disabled MKL-DNN") - set(WITH_MKLDNN OFF) +if (NOT DEFINED WITH_MKLDNN) + if (WITH_MKL AND AVX2_FOUND) + set(WITH_MKLDNN ON) + else() + message(STATUS "Do not have AVX2 intrinsics and disabled MKL-DNN") + set(WITH_MKLDNN OFF) + endif() endif() - ######################################################################################## include(external/mklml) # download mklml package @@ -153,7 +157,6 @@ include(cupti) include(configure) # add paddle env configuration include(generic) # simplify cmake module include(package) # set paddle packages -include(cpplint) # set paddle c++ style include(ccache) # set ccache for compilation include(util) # set unittest and link libs include(rdma) # set rdma libraries @@ -202,7 +205,7 @@ endif(USE_NNPACK) add_subdirectory(proto) -if(NOT MOBILE_INFERENCE) +if(NOT MOBILE_INFERENCE AND NOT WITH_FLUID_ONLY) # "add_subdirectory(go)" should be placed after the following loine, # because it depends on paddle/optimizer. add_subdirectory(paddle/optimizer) @@ -226,5 +229,11 @@ if(WITH_PYTHON) endif() if(WITH_DOC) + find_package(Sphinx REQUIRED) + find_python_module(recommonmark REQUIRED) add_subdirectory(doc) endif() + +if (WITH_CONTRIB) + add_subdirectory(paddle/contrib) +endif() diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md index 3c36cffcb4eeaaf7f8cff5167777628dd2697e7d..b1b02bcc2f4fd14297715bcf5bfd1617e3d5f0c9 100644 --- a/CONTRIBUTING.md +++ b/CONTRIBUTING.md @@ -58,6 +58,8 @@ PaddlePaddle uses this [Git branching model](http://nvie.com/posts/a-successful- create mode 100644 233 ``` + NOTE: The `yapf` installed by `pip install pre-commit` and `conda install -c conda-forge pre-commit` is slightly different. Paddle developers use `pip install pre-commit`. + 1. Build and test Users can build PaddlePaddle natively on Linux and Mac OS X. But to unify the building environment and to make it easy for debugging, the recommended way is [using Docker](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/dev/build_en.md). diff --git a/Dockerfile b/Dockerfile index c257dbfc2987323f8ed2a24dfffa8b3c15e09399..4d6165b79a1d94b8f27d7f3ee1b6e2cee5992d31 100644 --- a/Dockerfile +++ b/Dockerfile @@ -1,7 +1,6 @@ # A image for building paddle binaries # Use cuda devel base image for both cpu and gpu environment - -# When you modify it, please be aware of cudnn-runtime version +# When you modify it, please be aware of cudnn-runtime version # and libcudnn.so.x in paddle/scripts/docker/build.sh FROM nvidia/cuda:8.0-cudnn7-devel-ubuntu16.04 MAINTAINER PaddlePaddle Authors @@ -24,16 +23,16 @@ ENV HOME /root COPY ./paddle/scripts/docker/root/ /root/ RUN apt-get update && \ - apt-get install -y \ + apt-get install -y --allow-downgrades \ git python-pip python-dev openssh-server bison \ libnccl2=2.1.2-1+cuda8.0 libnccl-dev=2.1.2-1+cuda8.0 \ wget unzip unrar tar xz-utils bzip2 gzip coreutils ntp \ curl sed grep graphviz libjpeg-dev zlib1g-dev \ python-matplotlib gcc-4.8 g++-4.8 \ - automake locales clang-format swig doxygen cmake \ + automake locales clang-format swig cmake \ liblapack-dev liblapacke-dev \ clang-3.8 llvm-3.8 libclang-3.8-dev \ - net-tools libtool && \ + net-tools libtool ccache && \ apt-get clean -y # Install Go and glide @@ -71,7 +70,7 @@ RUN localedef -i en_US -f UTF-8 en_US.UTF-8 # specify sphinx version as 1.5.6 and remove -U option for [pip install -U # sphinx-rtd-theme] since -U option will cause sphinx being updated to newest # version(1.7.1 for now), which causes building documentation failed. -RUN pip install --upgrade pip==9.0.3 && \ +RUN easy_install -U pip && \ pip install -U wheel && \ pip install -U docopt PyYAML sphinx==1.5.6 && \ pip install sphinx-rtd-theme==0.1.9 recommonmark @@ -80,6 +79,9 @@ RUN pip install pre-commit 'ipython==5.3.0' && \ pip install 'ipykernel==4.6.0' 'jupyter==1.0.0' && \ pip install opencv-python +#For docstring checker +RUN pip install pylint pytest astroid isort + COPY ./python/requirements.txt /root/ RUN pip install -r /root/requirements.txt @@ -102,6 +104,3 @@ RUN echo 'root:root' | chpasswd RUN sed -ri 's/^PermitRootLogin\s+.*/PermitRootLogin yes/' /etc/ssh/sshd_config RUN sed -ri 's/UsePAM yes/#UsePAM yes/g' /etc/ssh/sshd_config EXPOSE 22 - -# development image default do build work -CMD ["bash", "/paddle/paddle/scripts/docker/build.sh"] diff --git a/Dockerfile.android b/Dockerfile.android index 848a7eba6f1421432addae8acff407b611adb4ae..48db2efea21a648657e3f490c95429b9a29ede52 100644 --- a/Dockerfile.android +++ b/Dockerfile.android @@ -40,5 +40,3 @@ RUN mkdir -p ${ANDROID_TOOLCHAINS_DIR} && \ unzip -q android-ndk-r14b-linux-x86_64.zip && \ mv android-ndk-r14b ${ANDROID_NDK_HOME} && \ rm -rf /opt/android-ndk-tmp - -CMD ["bash", "/paddle/paddle/scripts/docker/build_android.sh"] diff --git a/README.md b/README.md index d06375a444dd65675bdd75baccf8445c1638a87c..8d89c6b1ec9e4aefbd64328dedb4e8c7cc50c21b 100644 --- a/README.md +++ b/README.md @@ -62,9 +62,9 @@ Please refer to our [release announcement](https://github.com/PaddlePaddle/Paddl ## Installation It is recommended to check out the -[Docker installation guide](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/build_and_install/docker_install_en.html) +[Docker installation guide](http://www.paddlepaddle.org/docs/develop/documentation/fluid/en/build_and_install/docker_install_en.html) before looking into the -[build from source guide](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/build_and_install/build_from_source_en.html). +[build from source guide](http://www.paddlepaddle.org/docs/develop/documentation/fluid/en/build_and_install/build_from_source_en.html). ## Documentation @@ -75,19 +75,19 @@ We provide [English](http://www.paddlepaddle.org/docs/develop/documentation/en/g You might want to start from this online interactive book that can run in a Jupyter Notebook. -- [Distributed Training](http://www.paddlepaddle.org/docs/develop/documentation/en/howto/usage/cluster/cluster_train_en.html) +- [Distributed Training](http://www.paddlepaddle.org/docs/develop/documentation/en/howto/cluster/index_en.html) You can run distributed training jobs on MPI clusters. -- [Distributed Training on Kubernetes](http://www.paddlepaddle.org/docs/develop/documentation/en/howto/usage/cluster/k8s_en.html) +- [Distributed Training on Kubernetes](http://www.paddlepaddle.org/docs/develop/documentation/en/howto/cluster/multi_cluster/k8s_en.html) You can also run distributed training jobs on Kubernetes clusters. -- [Python API](http://www.paddlepaddle.org/docs/develop/documentation/en/api/index_en.html) +- [Python API](http://www.paddlepaddle.org/docs/develop/api/en/overview.html) Our new API enables much shorter programs. -- [How to Contribute](http://www.paddlepaddle.org/docs/develop/documentation/en/howto/dev/contribute_to_paddle_en.html) +- [How to Contribute](http://www.paddlepaddle.org/docs/develop/documentation/fluid/en/dev/contribute_to_paddle_en.html) We appreciate your contributions! diff --git a/benchmark/cluster/README.md b/benchmark/cluster/README.md deleted file mode 100644 index 64816098a524f064ec12474a736cd4c721227a70..0000000000000000000000000000000000000000 --- a/benchmark/cluster/README.md +++ /dev/null @@ -1,196 +0,0 @@ -# Cluster Training Benchmark - -## Setup - -- Platform - - Kubernetes: v1.6.2 - - Linux Kernel: v3.10.0 - -- Resource - - CPU: 10 Cores per Pod - - Memory: 5GB per Pod - -- Docker Image - - We use different base Docker Image to run the benchmark on Kubernetes: - - PaddlePaddle v2: paddlepaddle/paddle:0.11.0 - - PaddlePaddle Fluid: paddlepaddle/paddle:[commit-id] - - TensorFlow: tensorflow/tensorflow:1.5.0-rc0 - -- Model - vgg16 is used in this benchmark. - -## Cases - -- Variable - - Batch Size of training data. - - PServer count of the training job. - - The number of trainers. - -- Invariant - - The resource of trainer/pserver Pod. - -### Measure the Performance for Different Batch Size - -- PServer Count: 40 -- Trainer Count: 100 -- Metrics: mini-batch / sec - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
Batch Size 3264128 256
PaddlePaddle Fluid-- - -
PaddlePaddle v2 - - - -
TensorFlow - - - -
- -### Measure the Performance for Different PServer Count - -- Trainer Count: 100 -- Batch Size: 64 -- Metrics: mini-batch / sec - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
PServer Count 102040 60
PaddlePaddle Fluid-- - -
PaddlePaddle v2 - - - -
TensorFlow - - - -
- -### Measure Parallel Efficiency By Increasing Trainer Count - -- PServer Count: 20 -- Batch Size: 64 -- Metrics: - -$S = \div(T1, TN)$ - -which S is the ratio of T1 over TN, training time of 1 and N trainers. -The parallel efficiency is: - -$E = \div(S, N)$ - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
Trainer Counter 11020 30405060 708090100
PaddlePaddle Fluid-- - - -- - - -- -
PaddlePaddle v2 - - - - -- - - -- -
TensorFlow - - - - -- - - -- -
- - -## Reproduce the benchmark - -TODO diff --git a/benchmark/cluster/vgg16/Dockerfile b/benchmark/cluster/vgg16/Dockerfile deleted file mode 100644 index 13ad8e1b6237e6f41a076c4fb54311728832ae33..0000000000000000000000000000000000000000 --- a/benchmark/cluster/vgg16/Dockerfile +++ /dev/null @@ -1,35 +0,0 @@ -FROM nvidia/cuda:8.0-cudnn5-runtime-ubuntu16.04 - -# you can get mirror list here: -# https://launchpad.net/ubuntu/+archivemirrors -ARG UBUNTU_MIRROR -RUN /bin/bash -c 'if [[ -n ${UBUNTU_MIRROR} ]]; then sed -i 's#http://archive.ubuntu.com/ubuntu#${UBUNTU_MIRROR}#g' /etc/apt/sources.list; fi' - -RUN apt-get update && apt-get install -y python python-dev python-pip iputils-ping libgtk2.0-dev -RUN pip install -U kubernetes opencv-python - -RUN pip install paddlepaddle -# if network is slowly, you may need to add proxy here. -# ENV https_proxy= -RUN sh -c 'echo "import paddle.v2 as paddle\npaddle.dataset.cifar.train10()" | python' -RUN pip uninstall -y paddlepaddle -# unset proxy if it is setted. -# ENV https_proxy="" - -# NOTE: By default CI built wheel packages turn WITH_DISTRIBUTE=OFF, -# so we must build one with distribute support to install in this image. -ADD *.whl / -RUN pip install /*.whl && rm -f /*.whl -ENV LD_LIBRARY_PATH=/usr/local/lib - -# tf k8s -RUN pip install tensorflow==1.4.0 -ADD tf_k8s /usr/bin -RUN chmod +x /usr/bin/tf_k8s -ADD vgg16_tf.py /workspace/ - -# below lines may change a lot for debugging -ADD https://raw.githubusercontent.com/PaddlePaddle/cloud/develop/docker/paddle_k8s /usr/bin -ADD https://raw.githubusercontent.com/PaddlePaddle/cloud/develop/docker/k8s_tools.py /root -RUN chmod +x /usr/bin/paddle_k8s -ADD vgg16_fluid.py vgg16_v2.py /workspace/ diff --git a/benchmark/cluster/vgg16/README.md b/benchmark/cluster/vgg16/README.md deleted file mode 100644 index d56a912b9b03986e32693363f82df05a34b779e9..0000000000000000000000000000000000000000 --- a/benchmark/cluster/vgg16/README.md +++ /dev/null @@ -1,195 +0,0 @@ -# Performance for Distributed vgg16 - -## Test Result - -### Hardware Infomation - -- CPU: Intel(R) Xeon(R) CPU E5-2620 v4 @ 2.10GHz -- cpu MHz : 2101.000 -- cache size : 20480 KB - -### Blas settings - -Setting environment variable: `MKL_NUM_THREADS=1`. - -### Single Node Single Thread - -- Metrics: samples / sec - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
Batch Size 3264128 256
PaddlePaddle Fluid 15.44 16.32 16.74 16.79
PaddlePaddle v2 15.97 17.04 17.60 17.83
TensorFlow 9.09 9.10 9.24 8.66
- - -### Different Batch Size - -- PServer Count: 10 -- Trainer Count: 20 -- Metrics: samples / sec - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
Batch Size 3264128 256
PaddlePaddle Fluid 190.20 222.15 247.40 258.18
PaddlePaddle v2 170.96 233.71 256.14 329.23
TensorFlow - - - -
- -### Accelerate Rate - -- Pserver Count: 20 -- Batch Size: 128 -- Metrics: samples / sec - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
Trainer Count 204080100
PaddlePaddle Fluid 263.29 (78.64%) 518.80 (77.47%) 836.26 (62.44%) 1019.29 (60.89%)
PaddlePaddle v2 (need more tests) 326.85 (92.85%) 534.58 (75.93%) 853.30 (60.60%) 1041.99 (59.20%)
TensorFlow - - - -
- - -### Different Pserver Count - -- Trainer Count: 60 -- Batch Size: 128 -- Metrics: samples/ sec - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
PServer Count 361020
PaddlePaddle Fluid(should fix in next PR) 589.1 592.6 656.4 655.8
PaddlePaddle v2 (need more tests) 593.4 791.3 729.7 821.7
TensorFlow - - - -
- - -*The performance gap between Fuild and v2 comes from the network interference.* - - -## Steps to Run the Performance Test - -1. You must re-compile PaddlePaddle and enable `-DWITH_DISTRIBUTE` to build PaddlePaddle with distributed support. -1. When the build finishes, copy the output `whl` package located under `build/python/dist` to current directory. -1. Run `docker build -t [image:tag] .` to build the docker image and run `docker push [image:tag]` to push the image to reponsitory so kubernetes can find it. -1. Run `kubectl create -f pserver.yaml && kubectl create -f trainer.yaml` to start the job on your kubernetes cluster (you must configure the `kubectl` client before this step). -1. Run `kubectl get po` to get running pods, and run `kubectl logs [podID]` to fetch the pod log of pservers and trainers. - -Check the logs for the distributed training progress and analyze the performance. - -## Enable Verbos Logs - -Edit `pserver.yaml` and `trainer.yaml` and add an environment variable `GLOG_v=3` and `GLOG_logtostderr=1` to see what happend in detail. diff --git a/benchmark/cluster/vgg16/fluid_pserver.yaml b/benchmark/cluster/vgg16/fluid_pserver.yaml deleted file mode 100644 index ee8b0763b62fc011f40f6197e929a68b48a93e47..0000000000000000000000000000000000000000 --- a/benchmark/cluster/vgg16/fluid_pserver.yaml +++ /dev/null @@ -1,72 +0,0 @@ -apiVersion: extensions/v1beta1 -kind: ReplicaSet -metadata: - name: vgg16job-pserver -spec: - replicas: 10 - template: - metadata: - labels: - paddle-job-pserver: vgg16job - spec: - hostNetwork: true - imagePullSecrets: - - name: job-registry-secret - containers: - - name: pserver - image: "registry.baidu.com/paddlepaddle/fluid_benchmark:vgg16" - imagePullPolicy: Always - ports: - - name: jobport-30236 - containerPort: 30236 - env: - - name: PADDLE_JOB_NAME - value: vgg16job - - name: MKL_NUM_THREADS - value: "1" - - name: TRAINING_ROLE - value: "PSERVER" - - name: TRAINERS - value: "20" - - name: PSERVERS - value: "10" - - name: TOPOLOGY - value: "" - - name: ENTRY - value: "MKL_NUM_THREADS=1 python /workspace/vgg16_fluid.py --local 0" - - name: TRAINER_PACKAGE - value: "/workspace" - - name: PADDLE_INIT_PORT - value: "30236" - - name: PADDLE_INIT_NICS - value: "xgbe0" - - name: PADDLE_INIT_TRAINER_COUNT - value: "1" - - name: PADDLE_INIT_PORTS_NUM - value: "1" - - name: PADDLE_INIT_PORTS_NUM_FOR_SPARSE - value: "1" - - name: PADDLE_INIT_NUM_GRADIENT_SERVERS - value: "20" - - name: PADDLE_INIT_NUM_PASSES - value: "1" - - name: PADDLE_INIT_USE_GPU - value: "0" - - name: LD_LIBRARY_PATH - value: "/usr/local/lib:/usr/local/nvidia/lib64" - - name: NAMESPACE - valueFrom: - fieldRef: - fieldPath: "metadata.namespace" - - name: POD_IP - valueFrom: - fieldRef: - fieldPath: "status.podIP" - command: ["paddle_k8s", "start_fluid"] - resources: - requests: - memory: 10Gi - cpu: 4 - limits: - memory: 10Gi - cpu: 4 diff --git a/benchmark/cluster/vgg16/fluid_trainer.yaml b/benchmark/cluster/vgg16/fluid_trainer.yaml deleted file mode 100644 index 3d56caac009464d1073423bb63abff1f8b0cf28f..0000000000000000000000000000000000000000 --- a/benchmark/cluster/vgg16/fluid_trainer.yaml +++ /dev/null @@ -1,69 +0,0 @@ -apiVersion: batch/v1 -kind: Job -metadata: - name: vgg16job-trainer -spec: - parallelism: 20 - completions: 20 - template: - metadata: - labels: - paddle-job: vgg16job - spec: - imagePullSecrets: - - name: job-registry-secret - hostNetwork: true - containers: - - name: trainer - image: "registry.baidu.com/paddlepaddle/fluid_benchmark:vgg16" - imagePullPolicy: Always - command: ["paddle_k8s", "start_fluid"] - env: - - name: PADDLE_JOB_NAME - value: vgg16job - - name: TRAINING_ROLE - value: "TRAINER" - - name: TRAINERS - value: "20" - - name: PSERVERS - value: "10" - - name: TOPOLOGY - value: "" - - name: ENTRY - value: "MKL_NUM_THREADS=1 python /workspace/vgg16_fluid.py --local 0 --batch_size 128" - - name: TRAINER_PACKAGE - value: "/workspace" - - name: PADDLE_INIT_PORT - value: "30236" - - name: PADDLE_INIT_NICS - value: "xgbe0" - - name: PADDLE_INIT_TRAINER_COUNT - value: "1" - - name: PADDLE_INIT_PORTS_NUM - value: "1" - - name: PADDLE_INIT_PORTS_NUM_FOR_SPARSE - value: "1" - - name: PADDLE_INIT_NUM_GRADIENT_SERVERS - value: "20" - - name: PADDLE_INIT_NUM_PASSES - value: "1" - - name: PADDLE_INIT_USE_GPU - value: "0" - - name: LD_LIBRARY_PATH - value: "/usr/local/lib:/usr/local/nvidia/lib64" - - name: NAMESPACE - valueFrom: - fieldRef: - fieldPath: "metadata.namespace" - - name: POD_IP - valueFrom: - fieldRef: - fieldPath: "status.podIP" - resources: - requests: - memory: 40Gi - cpu: 2 - limits: - memory: 40Gi - cpu: 2 - restartPolicy: Never diff --git a/benchmark/cluster/vgg16/tf_k8s b/benchmark/cluster/vgg16/tf_k8s deleted file mode 100644 index 4fc263d5f681aeabfa71f1758714d269d987b272..0000000000000000000000000000000000000000 --- a/benchmark/cluster/vgg16/tf_k8s +++ /dev/null @@ -1,82 +0,0 @@ -#!/bin/bash -check_trainer_ret() { - ret=$1 - stdbuf -oL echo "job returned $ret...setting pod return message..." - stdbuf -oL echo "===============================" - - if [ $ret -eq 136 ] ; then - echo "Error Arithmetic Operation(Floating Point Exception)" > /dev/termination-log - elif [ $ret -eq 139 ] ; then - echo "Segmentation Fault" > /dev/termination-log - elif [ $ret -eq 1 ] ; then - echo "General Error" > /dev/termination-log - elif [ $ret -eq 134 ] ; then - echo "Program Abort" > /dev/termination-log - fi - stdbuf -oL echo "termination log wroted..." - exit $ret -} - -g_pservers="" -g_trainers="" - -wait_running_pods(){ - pserver_label="tf-job-pserver=${JOB_NAME}" - trainer_label="tf-job-trainer=${JOB_NAME}" - - stdbuf -oL python /root/k8s_tools.py wait_pods_running ${pserver_label} ${PSERVERS_NUM} - stdbuf -oL python /root/k8s_tools.py wait_pods_running ${trainer_label} ${TRAINERS_NUM} - - g_pservers=$(python /root/k8s_tools.py fetch_endpoints ${pserver_label} ${PORT}) - g_trainers=$(python /root/k8s_tools.py fetch_endpoints ${trainer_label} ${PORT}) -} - -start_tf_pserver(){ - wait_running_pods - - label="tf-job-pserver=${JOB_NAME}" - pserver_id=$(python /root/k8s_tools.py fetch_id ${label}) - - cmd="${ENTRY} --ps_hosts=${g_pservers} --worker_hosts=${g_trainers} \ - --job_name=${TF_JOB_NAME} --task_index=${pserver_id}" - - stdbuf -oL sh -c "cd ${TRAINER_PACKAGE} && ${cmd}" -} - -start_tf_trainer(){ - wait_running_pods - - label="tf-job-trainer=${JOB_NAME}" - trainer_id=$(python /root/k8s_tools.py fetch_id ${label}) - - cmd="${ENTRY} --ps_hosts=${g_pservers} --worker_hosts=${g_trainers} \ - --job_name=${TF_JOB_NAME} --task_index=${trainer_id} --batch_size=${BATCH_SIZE}" - - stdbuf -oL sh -c "cd ${TRAINER_PACKAGE} && ${cmd}" - check_trainer_ret $? -} - -start_tf(){ - if [[ "${TF_JOB_NAME}" == "worker" ]]; then - start_tf_trainer - else - start_tf_pserver - fi -} - -usage() { - echo "usage: tf_k8s []:" - echo " start_tf Start tensorflow jobs" -} - -case "$1" in - start_tf) - start_tf - ;; - --help) - usage - ;; - *) - usage - ;; -esac diff --git a/benchmark/cluster/vgg16/tf_pserver.yaml b/benchmark/cluster/vgg16/tf_pserver.yaml deleted file mode 100644 index 5e37c700819119c8af05c40fe4b8d13911efc3e1..0000000000000000000000000000000000000000 --- a/benchmark/cluster/vgg16/tf_pserver.yaml +++ /dev/null @@ -1,56 +0,0 @@ -apiVersion: extensions/v1beta1 -kind: ReplicaSet -metadata: - name: vgg16job-tf-pserver -spec: - replicas: 10 - template: - metadata: - labels: - tf-job-pserver: vgg16job-tf - spec: - hostNetwork: true - imagePullSecrets: - - name: job-registry-secret - containers: - - name: pserver - image: "registry.baidu.com/paddlepaddle/fluid_benchmark_tf:vgg16" - imagePullPolicy: Always - command: ["tf_k8s", "start_tf"] - ports: - - name: jobport-30236 - containerPort: 30236 - env: - - name: PORT - value: "32036" - - name: ENTRY - value: "python vgg16_tf.py" - - name: JOB_NAME - value: vgg16job-tf - - name: PSERVERS_NUM - value: "10" - - name: TF_JOB_NAME - value: "ps" - - name: TRAINERS_NUM - value: "20" - - name: BATCH_SIZE - value: "128" - - name: TRAINER_PACKAGE - value: "/workspace" - - name: NUM_PASSES - value: "1" - - name: NAMESPACE - valueFrom: - fieldRef: - fieldPath: "metadata.namespace" - - name: POD_IP - valueFrom: - fieldRef: - fieldPath: "status.podIP" - resources: - requests: - memory: 10Gi - cpu: 4 - limits: - memory: 10Gi - cpu: 4 diff --git a/benchmark/cluster/vgg16/tf_trainer.yaml b/benchmark/cluster/vgg16/tf_trainer.yaml deleted file mode 100644 index 08795df3addfa7b618db24a65e57be190e268f06..0000000000000000000000000000000000000000 --- a/benchmark/cluster/vgg16/tf_trainer.yaml +++ /dev/null @@ -1,58 +0,0 @@ -apiVersion: batch/v1 -kind: Job -metadata: - name: vgg16job-tf-trainer -spec: - parallelism: 20 - completions: 20 - template: - metadata: - labels: - tf-job-trainer: vgg16job-tf - spec: - imagePullSecrets: - - name: job-registry-secret - hostNetwork: true - containers: - - name: trainer - image: "registry.baidu.com/paddlepaddle/fluid_benchmark_tf:vgg16" - imagePullPolicy: Always - command: ["tf_k8s", "start_tf"] - ports: - - name: jobport-30236 - containerPort: 30236 - env: - - name: PORT - value: "32036" - - name: JOB_NAME - value: vgg16job-tf - - name: TF_JOB_NAME - value: "worker" - - name: ENTRY - value: "python vgg16_tf.py" - - name: PSERVERS_NUM - value: "10" - - name: BATCH_SIZE - value: "128" - - name: TRAINERS_NUM - value: "20" - - name: TRAINER_PACKAGE - value: "/workspace" - - name: NUM_PASSES - value: "1" - - name: NAMESPACE - valueFrom: - fieldRef: - fieldPath: "metadata.namespace" - - name: POD_IP - valueFrom: - fieldRef: - fieldPath: "status.podIP" - resources: - requests: - memory: 40Gi - cpu: 2 - limits: - memory: 40Gi - cpu: 2 - restartPolicy: Never diff --git a/benchmark/cluster/vgg16/v2_pserver.yaml b/benchmark/cluster/vgg16/v2_pserver.yaml deleted file mode 100644 index dd1271e0cf399184134c06b3200ee1202c65cef0..0000000000000000000000000000000000000000 --- a/benchmark/cluster/vgg16/v2_pserver.yaml +++ /dev/null @@ -1,64 +0,0 @@ -apiVersion: extensions/v1beta1 -kind: ReplicaSet -metadata: - name: vgg16v2job-pserver -spec: - replicas: 10 - template: - metadata: - labels: - paddle-job-pserver: vgg16v2job - spec: - hostNetwork: true - imagePullSecrets: - - name: job-registry-secret - containers: - - name: pserver - image: "registry.baidu.com/paddlepaddle/fluid_benchmark:vgg16" - imagePullPolicy: Always - ports: - - name: jobport-30236 - containerPort: 30236 - env: - - name: PADDLE_JOB_NAME - value: vgg16v2job - - name: TRAINERS - value: "20" - - name: PSERVERS - value: "10" - - name: TOPOLOGY - value: "" - - name: ENTRY - value: "python train.py" - - name: TRAINER_PACKAGE - value: "/workspace" - - name: PADDLE_INIT_PORT - value: "30236" - - name: PADDLE_INIT_NICS - value: "xgbe0" - - name: PADDLE_INIT_TRAINER_COUNT - value: "1" - - name: PADDLE_INIT_PORTS_NUM - value: "1" - - name: PADDLE_INIT_PORTS_NUM_FOR_SPARSE - value: "1" - - name: PADDLE_INIT_NUM_GRADIENT_SERVERS - value: "20" - - name: PADDLE_INIT_NUM_PASSES - value: "1" - - name: PADDLE_INIT_USE_GPU - value: "0" - - name: LD_LIBRARY_PATH - value: "/usr/local/lib:/usr/local/nvidia/lib64" - - name: NAMESPACE - valueFrom: - fieldRef: - fieldPath: "metadata.namespace" - command: ["paddle_k8s", "start_pserver"] - resources: - requests: - memory: 10Gi - cpu: 4 - limits: - memory: 10Gi - cpu: 4 diff --git a/benchmark/cluster/vgg16/v2_trainer.yaml b/benchmark/cluster/vgg16/v2_trainer.yaml deleted file mode 100644 index 12c8964066cbcfe8d2a44de2f51a3d12ea422fe2..0000000000000000000000000000000000000000 --- a/benchmark/cluster/vgg16/v2_trainer.yaml +++ /dev/null @@ -1,65 +0,0 @@ -apiVersion: batch/v1 -kind: Job -metadata: - name: vgg16v2job-trainer -spec: - parallelism: 20 - completions: 20 - template: - metadata: - labels: - paddle-job: vgg16v2job - spec: - imagePullSecrets: - - name: job-registry-secret - hostNetwork: true - containers: - - name: trainer - image: "registry.baidu.com/paddlepaddle/fluid_benchmark:vgg16" - imagePullPolicy: Always - command: ["paddle_k8s", "start_trainer", "v2"] - env: - - name: PADDLE_JOB_NAME - value: vgg16v2job - - name: BATCH_SIZE - value: "256" - - name: TRAINERS - value: "20" - - name: PSERVERS - value: "10" - - name: TOPOLOGY - value: "" - - name: ENTRY - value: "cd /workspace && MKL_NUM_THREADS=1 python /workspace/vgg16_v2.py" - - name: TRAINER_PACKAGE - value: "/workspace" - - name: PADDLE_INIT_PORT - value: "30236" - - name: PADDLE_INIT_NICS - value: "xgbe0" - - name: PADDLE_INIT_TRAINER_COUNT - value: "1" - - name: PADDLE_INIT_PORTS_NUM - value: "1" - - name: PADDLE_INIT_PORTS_NUM_FOR_SPARSE - value: "1" - - name: PADDLE_INIT_NUM_GRADIENT_SERVERS - value: "20" - - name: PADDLE_INIT_NUM_PASSES - value: "2" - - name: PADDLE_INIT_USE_GPU - value: "0" - - name: LD_LIBRARY_PATH - value: "/usr/local/lib:/usr/local/nvidia/lib64" - - name: NAMESPACE - valueFrom: - fieldRef: - fieldPath: "metadata.namespace" - resources: - requests: - memory: 40Gi - cpu: 2 - limits: - memory: 40Gi - cpu: 2 - restartPolicy: Never diff --git a/benchmark/cluster/vgg16/vgg16_fluid.py b/benchmark/cluster/vgg16/vgg16_fluid.py deleted file mode 100644 index 8b29227cfab2a36d5b9f6d17b837b33da8d2a92e..0000000000000000000000000000000000000000 --- a/benchmark/cluster/vgg16/vgg16_fluid.py +++ /dev/null @@ -1,293 +0,0 @@ -# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -"""VGG16 benchmark in Fluid""" -from __future__ import print_function - -import sys -import time -import numpy as np -import paddle.v2 as paddle -import paddle.fluid as fluid -import paddle.fluid.core as core -import paddle.fluid.profiler as profiler -import argparse -import functools -import os -from paddle.fluid import debuger - - -def str2bool(v): - if v.lower() in ('yes', 'true', 't', 'y', '1'): - return True - elif v.lower() in ('no', 'false', 'f', 'n', '0'): - return False - else: - raise argparse.ArgumentTypeError('Boolean value expected.') - - -parser = argparse.ArgumentParser(description=__doc__) -parser.add_argument( - '--batch_size', type=int, default=128, help="Batch size for training.") -parser.add_argument( - '--learning_rate', - type=float, - default=1e-3, - help="Learning rate for training.") -parser.add_argument('--num_passes', type=int, default=50, help="No. of passes.") -parser.add_argument( - '--device', - type=str, - default='CPU', - choices=['CPU', 'GPU'], - help="The device type.") -parser.add_argument('--device_id', type=int, default=0, help="The device id.") -parser.add_argument( - '--data_format', - type=str, - default='NCHW', - choices=['NCHW', 'NHWC'], - help='The data order, now only support NCHW.') -parser.add_argument( - '--data_set', - type=str, - default='cifar10', - choices=['cifar10', 'flowers'], - help='Optional dataset for benchmark.') -parser.add_argument( - '--local', - type=str2bool, - default=True, - help='Whether to run as local mode.') - -parser.add_argument( - "--ps_hosts", - type=str, - default="", - help="Comma-separated list of hostname:port pairs") -parser.add_argument( - "--trainer_hosts", - type=str, - default="", - help="Comma-separated list of hostname:port pairs") - -# Flags for defining the tf.train.Server -parser.add_argument( - "--task_index", type=int, default=0, help="Index of task within the job") -args = parser.parse_args() - - -def vgg16_bn_drop(input): - def conv_block(input, num_filter, groups, dropouts): - return fluid.nets.img_conv_group( - input=input, - pool_size=2, - pool_stride=2, - conv_num_filter=[num_filter] * groups, - conv_filter_size=3, - conv_act='relu', - conv_with_batchnorm=True, - conv_batchnorm_drop_rate=dropouts, - pool_type='max') - - conv1 = conv_block(input, 64, 2, [0.3, 0]) - conv2 = conv_block(conv1, 128, 2, [0.4, 0]) - conv3 = conv_block(conv2, 256, 3, [0.4, 0.4, 0]) - conv4 = conv_block(conv3, 512, 3, [0.4, 0.4, 0]) - conv5 = conv_block(conv4, 512, 3, [0.4, 0.4, 0]) - - drop = fluid.layers.dropout(x=conv5, dropout_prob=0.5) - fc1 = fluid.layers.fc(input=drop, size=4096, act=None) - bn = fluid.layers.batch_norm(input=fc1, act='relu') - drop2 = fluid.layers.dropout(x=bn, dropout_prob=0.5) - fc2 = fluid.layers.fc(input=drop2, size=4096, act=None) - return fc2 - - -def main(): - if args.data_set == "cifar10": - classdim = 10 - if args.data_format == 'NCHW': - data_shape = [3, 32, 32] - else: - data_shape = [32, 32, 3] - else: - classdim = 102 - if args.data_format == 'NCHW': - data_shape = [3, 224, 224] - else: - data_shape = [224, 224, 3] - - # Input data - images = fluid.layers.data(name='pixel', shape=data_shape, dtype='float32') - label = fluid.layers.data(name='label', shape=[1], dtype='int64') - - # Train program - net = vgg16_bn_drop(images) - predict = fluid.layers.fc(input=net, size=classdim, act='softmax') - cost = fluid.layers.cross_entropy(input=predict, label=label) - avg_cost = fluid.layers.mean(x=cost) - - # Evaluator - batch_size = fluid.layers.create_tensor(dtype='int64') - batch_acc = fluid.layers.accuracy( - input=predict, label=label, total=batch_size) - - # inference program - inference_program = fluid.default_main_program().clone() - with fluid.program_guard(inference_program): - inference_program = fluid.io.get_inference_program(batch_acc) - - # Optimization - optimizer = fluid.optimizer.Adam(learning_rate=args.learning_rate) - optimize_ops, params_grads = optimizer.minimize(avg_cost) - - # Initialize executor - place = core.CPUPlace() if args.device == 'CPU' else core.CUDAPlace( - args.device_id) - exe = fluid.Executor(place) - - # test - def test(exe): - test_pass_acc = fluid.average.WeightedAverage() - for batch_id, data in enumerate(test_reader()): - img_data = np.array(map(lambda x: x[0].reshape(data_shape), - data)).astype("float32") - y_data = np.array(map(lambda x: x[1], data)).astype("int64") - y_data = y_data.reshape([-1, 1]) - - outs = exe.run(inference_program, - feed={"pixel": img_data, - "label": y_data}, - fetch_list=[batch_acc, batch_size]) - test_pass_acc.add(value=np.array(outs[0]), weight=np.array(outs[1])) - - return test_pass_acc.eval() - - def train_loop(exe, trainer_prog): - iters = 0 - ts = time.time() - train_pass_acc = fluid.average.WeightedAverage() - for pass_id in range(args.num_passes): - # train - start_time = time.time() - num_samples = 0 - train_pass_acc.reset() - for batch_id, data in enumerate(train_reader()): - ts = time.time() - img_data = np.array( - map(lambda x: x[0].reshape(data_shape), data)).astype( - "float32") - y_data = np.array(map(lambda x: x[1], data)).astype("int64") - y_data = y_data.reshape([-1, 1]) - - loss, acc, b_size = exe.run( - trainer_prog, - feed={"pixel": img_data, - "label": y_data}, - fetch_list=[avg_cost, batch_acc, batch_size]) - iters += 1 - num_samples += len(data) - train_pass_acc.add(value=acc, weight=b_size) - print( - "Pass = %d, Iters = %d, Loss = %f, Accuracy = %f, Speed = %.2f img/s" - % (pass_id, iters, loss, acc, - len(data) / (time.time() - ts)) - ) # The accuracy is the accumulation of batches, but not the current batch. - - pass_elapsed = time.time() - start_time - pass_train_acc = train_pass_acc.eval() - pass_test_acc = test(exe) - print( - "Pass = %d, Training performance = %f imgs/s, Train accuracy = %f, Test accuracy = %f\n" - % (pass_id, num_samples / pass_elapsed, pass_train_acc, - pass_test_acc)) - - if args.local: - # Parameter initialization - exe.run(fluid.default_startup_program()) - - # data reader - train_reader = paddle.batch( - paddle.reader.shuffle( - paddle.dataset.cifar.train10() if args.data_set == 'cifar10' - else paddle.dataset.flowers.train(), - buf_size=5120), - batch_size=args.batch_size) - test_reader = paddle.batch( - paddle.dataset.cifar.test10() - if args.data_set == 'cifar10' else paddle.dataset.flowers.test(), - batch_size=args.batch_size) - train_loop(exe, fluid.default_main_program()) - else: - trainers = int(os.getenv("TRAINERS")) # total trainer count - print("trainers total: ", trainers) - - training_role = os.getenv( - "TRAINING_ROLE", - "TRAINER") # get the training role: trainer/pserver - - t = fluid.DistributeTranspiler() - t.transpile( - optimize_ops, - params_grads, - trainer_id=args.task_index, - pservers=args.ps_hosts, - trainers=trainers) - - if training_role == "PSERVER": - current_endpoint = os.getenv("POD_IP") + ":" + os.getenv( - "PADDLE_INIT_PORT") - if not current_endpoint: - print("need env SERVER_ENDPOINT") - exit(1) - pserver_prog = t.get_pserver_program(current_endpoint) - pserver_startup = t.get_startup_program(current_endpoint, - pserver_prog) - exe.run(pserver_startup) - exe.run(pserver_prog) - elif training_role == "TRAINER": - # Parameter initialization - exe.run(fluid.default_startup_program()) - - # data reader - train_reader = paddle.batch( - paddle.reader.shuffle( - paddle.dataset.cifar.train10() if args.data_set == 'cifar10' - else paddle.dataset.flowers.train(), - buf_size=5120), - batch_size=args.batch_size) - test_reader = paddle.batch( - paddle.dataset.cifar.test10() if args.data_set == 'cifar10' else - paddle.dataset.flowers.test(), - batch_size=args.batch_size) - - trainer_prog = t.get_trainer_program() - feeder = fluid.DataFeeder(feed_list=[images, label], place=place) - # TODO(typhoonzero): change trainer startup program to fetch parameters from pserver - exe.run(fluid.default_startup_program()) - train_loop(exe, trainer_prog) - else: - print("environment var TRAINER_ROLE should be TRAINER os PSERVER") - - -def print_arguments(): - print('----------- Configuration Arguments -----------') - for arg, value in sorted(vars(args).iteritems()): - print('%s: %s' % (arg, value)) - print('------------------------------------------------') - - -if __name__ == "__main__": - print_arguments() - main() diff --git a/benchmark/cluster/vgg16/vgg16_tf.py b/benchmark/cluster/vgg16/vgg16_tf.py deleted file mode 100644 index 2d220478acae46566760209dbc012cff316946aa..0000000000000000000000000000000000000000 --- a/benchmark/cluster/vgg16/vgg16_tf.py +++ /dev/null @@ -1,366 +0,0 @@ -# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -"""VGG16 benchmark in TensorFlow -You can get distribution example template structure here: -https://medium.com/clusterone/how-to-write-distributed-tensorflow-code-with-an-example-on-tensorport-70bf3306adcb -https://www.tensorflow.org/deploy/distributed -""" - -import tensorflow as tf -import paddle.v2 as paddle -import numpy as np -import argparse -import time - -parser = argparse.ArgumentParser(description=__doc__) -parser.add_argument( - '--batch_size', type=int, default=128, help="Batch size for training.") -parser.add_argument( - '--learning_rate', - type=float, - default=1e-3, - help="Learning rate for training.") -parser.add_argument('--num_passes', type=int, default=50, help="No. of passes.") -parser.add_argument( - '--device', - type=str, - default='CPU', - choices=['CPU', 'GPU'], - help="The device type.") -parser.add_argument( - '--data_format', - type=str, - default='NHWC', - choices=['NCHW', 'NHWC'], - help='The data order, NCHW=[batch, channels, height, width].' - 'Only support NHWC right now.') -parser.add_argument( - '--data_set', - type=str, - default='cifar10', - choices=['cifar10', 'flowers'], - help='Optional dataset for benchmark.') - -parser.add_argument( - "--ps_hosts", - type=str, - default="", - help="Comma-separated list of hostname:port pairs") -parser.add_argument( - "--worker_hosts", - type=str, - default="", - help="Comma-separated list of hostname:port pairs") -parser.add_argument( - "--job_name", type=str, default="", help="One of 'worker', 'ps'") -# Flags for defining the tf.train.Server -parser.add_argument( - "--task_index", type=int, default=0, help="Index of task within the job") - -args = parser.parse_args() - - -class VGG16Model(object): - def __init__(self): - self.parameters = [] - - def batch_norm_relu(self, inputs, is_training): - """Performs a batch normalization followed by a ReLU.""" - # We set fused=True for a significant speed boost. See - # https://www.tensorflow.org/speed/speed_guide#common_fused_ops - inputs = tf.layers.batch_normalization( - inputs=inputs, - axis=1 if args.data_format == 'NCHW' else -1, - momentum=0.9, - epsilon=1e-05, - center=True, - scale=True, - training=is_training, - fused=True) - inputs = tf.nn.relu(inputs) - return inputs - - def conv_bn_layer(self, - name, - images, - kernel_shape, - is_training, - drop_rate=0.0): - with tf.name_scope(name) as scope: - kernel = tf.Variable( - tf.truncated_normal( - kernel_shape, dtype=tf.float32, stddev=1e-1), - name='weights') - conv = tf.nn.conv2d( - images, - kernel, [1, 1, 1, 1], - data_format=args.data_format, - padding='SAME') - biases = tf.Variable( - tf.constant( - 0.0, shape=[kernel_shape[-1]], dtype=tf.float32), - trainable=True, - name='biases') - out = tf.nn.bias_add(conv, biases) - out = self.batch_norm_relu(out, is_training) - out = tf.layers.dropout(out, rate=drop_rate, training=is_training) - return out - - def fc_layer(self, name, inputs, shape): - with tf.name_scope(name) as scope: - fc_w = tf.Variable( - tf.truncated_normal( - shape, dtype=tf.float32, stddev=1e-1), - name='weights') - fc_b = tf.Variable( - tf.constant( - 0.0, shape=[shape[-1]], dtype=tf.float32), - trainable=True, - name='biases') - out = tf.nn.bias_add(tf.matmul(inputs, fc_w), fc_b) - return out - - def network(self, images, class_dim, is_training): - """ VGG16 model structure. - - TODO(kuke): enable this network to support the 'NCHW' data format - """ - - # conv1 - conv1_1 = self.conv_bn_layer( - 'conv1_1', images, [3, 3, 3, 64], is_training, drop_rate=0.3) - conv1_2 = self.conv_bn_layer( - 'conv1_2', conv1_1, [3, 3, 64, 64], is_training, drop_rate=0.0) - # pool1 - pool1 = tf.nn.max_pool( - conv1_2, - ksize=[1, 2, 2, 1], - strides=[1, 2, 2, 1], - padding='SAME', - name='pool1') - # conv2 - conv2_1 = self.conv_bn_layer( - 'conv2_1', pool1, [3, 3, 64, 128], is_training, drop_rate=0.4) - conv2_2 = self.conv_bn_layer( - 'conv2_2', conv2_1, [3, 3, 128, 128], is_training, drop_rate=0.0) - # pool2 - pool2 = tf.nn.max_pool( - conv2_2, - ksize=[1, 2, 2, 1], - strides=[1, 2, 2, 1], - padding='SAME', - name='pool2') - # conv3 - conv3_1 = self.conv_bn_layer( - 'conv3_1', pool2, [3, 3, 128, 256], is_training, drop_rate=0.4) - conv3_2 = self.conv_bn_layer( - 'conv3_2', conv3_1, [3, 3, 256, 256], is_training, drop_rate=0.4) - conv3_3 = self.conv_bn_layer( - 'conv3_3', conv3_2, [3, 3, 256, 256], is_training, drop_rate=0.0) - # pool3 - pool3 = tf.nn.max_pool( - conv3_3, - ksize=[1, 2, 2, 1], - strides=[1, 2, 2, 1], - padding='SAME', - name='pool3') - # conv4 - conv4_1 = self.conv_bn_layer( - 'conv4_1', pool3, [3, 3, 256, 512], is_training, drop_rate=0.4) - conv4_2 = self.conv_bn_layer( - 'conv4_2', conv4_1, [3, 3, 512, 512], is_training, drop_rate=0.4) - conv4_3 = self.conv_bn_layer( - 'conv4_3', conv4_2, [3, 3, 512, 512], is_training, drop_rate=0.0) - # pool4 - pool4 = tf.nn.max_pool( - conv4_3, - ksize=[1, 2, 2, 1], - strides=[1, 2, 2, 1], - padding='SAME', - name='pool4') - # conv5 - conv5_1 = self.conv_bn_layer( - 'conv5_1', pool4, [3, 3, 512, 512], is_training, drop_rate=0.4) - conv5_2 = self.conv_bn_layer( - 'conv5_2', conv5_1, [3, 3, 512, 512], is_training, drop_rate=0.4) - conv5_3 = self.conv_bn_layer( - 'conv5_3', conv5_2, [3, 3, 512, 512], is_training, drop_rate=0.0) - # pool5 - pool5 = tf.nn.max_pool( - conv5_3, - ksize=[1, 2, 2, 1], - strides=[1, 2, 2, 1], - padding='SAME', - name='pool4') - # flatten - shape = int(np.prod(pool5.get_shape()[1:])) - pool5_flat = tf.reshape(pool5, [-1, shape]) - # fc1 - drop = tf.layers.dropout(pool5_flat, rate=0.5, training=is_training) - fc1 = self.fc_layer('fc1', drop, [shape, 512]) - # fc2 - bn = self.batch_norm_relu(fc1, is_training) - drop = tf.layers.dropout(bn, rate=0.5, training=is_training) - fc2 = self.fc_layer('fc2', drop, [512, 512]) - - fc3 = self.fc_layer('fc3', fc2, [512, class_dim]) - - return fc3 - - -def run_benchmark(cluster_spec, server): - """Run benchmark on cifar10 or flowers.""" - - if args.data_set == "cifar10": - class_dim = 10 - raw_shape = (3, 32, 32) - dat_shape = (None, 32, 32, 3) if args.data_format == 'NHWC' else ( - None, 3, 32, 32) - else: - class_dim = 102 - raw_shape = (3, 224, 224) - dat_shape = (None, 224, 224, 3) if args.data_format == 'NHWC' else ( - None, 3, 224, 224) - - device = tf.train.replica_device_setter( - worker_device="/job:worker/task:{}".format(args.task_index), - cluster=cluster_spec) - - with tf.device(device): - images = tf.placeholder(tf.float32, shape=dat_shape) - labels = tf.placeholder(tf.int64, shape=(None, )) - is_training = tf.placeholder('bool') - onehot_labels = tf.one_hot(labels, depth=class_dim) - - vgg16 = VGG16Model() - logits = vgg16.network(images, class_dim, is_training) - loss = tf.losses.softmax_cross_entropy( - onehot_labels=onehot_labels, logits=logits) - avg_loss = tf.reduce_mean(loss) - - correct = tf.equal(tf.argmax(logits, 1), labels) - accuracy = tf.reduce_mean(tf.cast(correct, tf.float32)) - - optimizer = tf.train.AdamOptimizer(learning_rate=args.learning_rate) - update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS) - global_step = tf.Variable(0, name='global_step', trainable=False) - with tf.control_dependencies(update_ops): - train_op = optimizer.minimize(avg_loss, global_step=global_step) - - summary_op = tf.summary.merge_all() - init_op = tf.global_variables_initializer() - - # data reader - train_reader = paddle.batch( - paddle.reader.shuffle( - paddle.dataset.cifar.train10() - if args.data_set == 'cifar10' else paddle.dataset.flowers.train(), - buf_size=5120), - batch_size=args.batch_size) - test_reader = paddle.batch( - paddle.reader.shuffle( - paddle.dataset.cifar.test10() - if args.data_set == 'cifar10' else paddle.dataset.flowers.test(), - buf_size=5120), - batch_size=args.batch_size) - - # test - def test(): - test_accs = [] - for batch_id, data in enumerate(test_reader()): - test_images = np.array( - map(lambda x: np.transpose(x[0].reshape(raw_shape), - axes=[1, 2, 0]) if args.data_format == 'NHWC' else x[0], data)).astype("float32") - test_labels = np.array(map(lambda x: x[1], data)).astype('int64') - test_accs.append( - accuracy.eval(feed_dict={ - images: test_images, - labels: test_labels, - is_training: False - })) - return np.mean(test_accs) - - config = tf.ConfigProto( - intra_op_parallelism_threads=1, - inter_op_parallelism_threads=1, - log_device_placement=True) - config.gpu_options.allow_growth = True - - hooks = [tf.train.StopAtStepHook(last_step=1000000)] - - with tf.train.MonitoredTrainingSession( - master=server.target, - is_chief=(args.task_index == 0), - hooks=hooks, - config=config) as sess: - iters, num_samples, start_time = 0, 0, 0.0 - for pass_id in range(args.num_passes): - # train - num_samples = 0 - start_time = time.time() - for batch_id, data in enumerate(train_reader()): - train_images = np.array( - map(lambda x: np.transpose(x[0].reshape(raw_shape), - axes=[1, 2, 0]) if args.data_format == 'NHWC' else x[0], data)).astype("float32") - train_labels = np.array(map(lambda x: x[1], data)).astype( - 'int64') - iter_begin_time = time.time() - _, loss, acc = sess.run([train_op, avg_loss, accuracy], - feed_dict={ - images: train_images, - labels: train_labels, - is_training: True - }) - iters += 1 - print( - "Pass = %d, Iters = %d, Loss = %f, Accuracy = %f, Speed=%.2f imgs/sec" - % (pass_id, iters, loss, acc, - len(data) / (time.time() - iter_begin_time))) - num_samples += len(data) - train_elapsed = time.time() - start_time - # test - pass_test_acc = test() - print("Pass = %d, Train speed = %f imgs/s, Test accuracy = %f\n" % - (pass_id, num_samples / train_elapsed, pass_test_acc)) - - -def print_arguments(): - print('----------- Configuration Arguments -----------') - for arg, value in sorted(vars(args).iteritems()): - print('%s: %s' % (arg, value)) - print('------------------------------------------------') - - -if __name__ == '__main__': - print_arguments() - - ps_hosts = args.ps_hosts.split(",") - worker_hosts = args.worker_hosts.split(",") - - # Create a cluster from the parameter server and worker hosts. - cluster_spec = tf.train.ClusterSpec({ - "ps": ps_hosts, - "worker": worker_hosts - }) - - # Create and start a server for the local task. - server = tf.train.Server( - cluster_spec, job_name=args.job_name, task_index=args.task_index) - - if args.job_name == "ps": - print("start pserver") - server.join() - elif args.job_name == "worker": - print("start worker") - run_benchmark(cluster_spec, server) diff --git a/benchmark/cluster/vgg16/vgg16_v2.py b/benchmark/cluster/vgg16/vgg16_v2.py deleted file mode 100644 index 1a66af32d7131997c63bd3c3042875f33a467084..0000000000000000000000000000000000000000 --- a/benchmark/cluster/vgg16/vgg16_v2.py +++ /dev/null @@ -1,154 +0,0 @@ -# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -# -#Licensed under the Apache License, Version 2.0 (the "License"); -#you may not use this file except in compliance with the License. -#You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -#Unless required by applicable law or agreed to in writing, software -#distributed under the License is distributed on an "AS IS" BASIS, -#WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -#See the License for the specific language governing permissions and -#limitations under the License. - -import gzip - -import paddle.v2.dataset.cifar as cifar -import paddle.v2 as paddle -import time -import os - -DATA_DIM = 3 * 32 * 32 -CLASS_DIM = 10 -BATCH_SIZE = os.getenv("BATCH_SIZE") -if BATCH_SIZE: - BATCH_SIZE = int(BATCH_SIZE) -else: - BATCH_SIZE = 128 -print "batch_size", BATCH_SIZE -NODE_COUNT = int(os.getenv("TRAINERS")) -ts = 0 - - -def vgg(input, nums, class_dim): - def conv_block(input, num_filter, groups, num_channels=None): - return paddle.networks.img_conv_group( - input=input, - num_channels=num_channels, - pool_size=2, - pool_stride=2, - conv_num_filter=[num_filter] * groups, - conv_filter_size=3, - conv_act=paddle.activation.Relu(), - pool_type=paddle.pooling.Max()) - - assert len(nums) == 5 - # the channel of input feature is 3 - conv1 = conv_block(input, 64, nums[0], 3) - conv2 = conv_block(conv1, 128, nums[1]) - conv3 = conv_block(conv2, 256, nums[2]) - conv4 = conv_block(conv3, 512, nums[3]) - conv5 = conv_block(conv4, 512, nums[4]) - - fc_dim = 512 - fc1 = paddle.layer.fc(input=conv5, - size=fc_dim, - act=paddle.activation.Relu(), - layer_attr=paddle.attr.Extra(drop_rate=0.5)) - fc2 = paddle.layer.fc(input=fc1, - size=fc_dim, - act=paddle.activation.Relu(), - layer_attr=paddle.attr.Extra(drop_rate=0.5)) - out = paddle.layer.fc(input=fc2, - size=class_dim, - act=paddle.activation.Softmax()) - return out - - -def vgg13(input, class_dim): - nums = [2, 2, 2, 2, 2] - return vgg(input, nums, class_dim) - - -def vgg16(input, class_dim): - nums = [2, 2, 3, 3, 3] - return vgg(input, nums, class_dim) - - -def vgg19(input, class_dim): - nums = [2, 2, 4, 4, 4] - return vgg(input, nums, class_dim) - - -def main(): - global ts - paddle.init(use_gpu=False) - image = paddle.layer.data( - name="image", type=paddle.data_type.dense_vector(DATA_DIM)) - lbl = paddle.layer.data( - name="label", type=paddle.data_type.integer_value(CLASS_DIM)) - - extra_layers = None - # NOTE: for v2 distributed training need averaging updates. - learning_rate = 1e-3 / NODE_COUNT - out = vgg16(image, class_dim=CLASS_DIM) - cost = paddle.layer.classification_cost(input=out, label=lbl) - - # Create parameters - parameters = paddle.parameters.create(cost) - - # Create optimizer - optimizer = paddle.optimizer.Momentum( - momentum=0.9, - regularization=paddle.optimizer.L2Regularization(rate=0.0005 * - BATCH_SIZE), - learning_rate=learning_rate / BATCH_SIZE, - learning_rate_decay_a=0.1, - learning_rate_decay_b=128000 * 35, - learning_rate_schedule="discexp", ) - - train_reader = paddle.batch( - paddle.reader.shuffle( - cifar.train10(), - # To use other data, replace the above line with: - # reader.train_reader('train.list'), - buf_size=1000), - batch_size=BATCH_SIZE) - test_reader = paddle.batch( - cifar.test10(), - # To use other data, replace the above line with: - # reader.test_reader('val.list'), - batch_size=BATCH_SIZE) - - # Create trainer - trainer = paddle.trainer.SGD(cost=cost, - parameters=parameters, - update_equation=optimizer, - extra_layers=extra_layers, - is_local=False) - - # End batch and end pass event handler - def event_handler(event): - global ts, ts_pass - if isinstance(event, paddle.event.BeginPass): - ts_pass = time.time() - if isinstance(event, paddle.event.BeginIteration): - ts = time.time() - if isinstance(event, paddle.event.EndIteration): - if event.batch_id % 1 == 0: - print "\nPass %d, Batch %d, Cost %f, %s, spent: %f" % ( - event.pass_id, event.batch_id, event.cost, event.metrics, - time.time() - ts) - if isinstance(event, paddle.event.EndPass): - print "Pass %d end, spent: %f" % (event.pass_id, - time.time() - ts_pass) - result = trainer.test(reader=test_reader) - print "\nTest with Pass %d, %s" % (event.pass_id, result.metrics) - - trainer.train( - reader=train_reader, num_passes=200, event_handler=event_handler) - - -if __name__ == '__main__': - main() diff --git a/benchmark/fluid/README.md b/benchmark/fluid/README.md new file mode 100644 index 0000000000000000000000000000000000000000..7071e9fdcd394a5a4db4d0d599610a72d98c0a3c --- /dev/null +++ b/benchmark/fluid/README.md @@ -0,0 +1,71 @@ +# Fluid Benchmark + +This directory contains several models configurations and tools that used to run +Fluid benchmarks for local and distributed training. + + +## Run the Benchmark + +To start, run the following command to get the full help message: + +```bash +python fluid_benchmark.py --help +``` + +Currently supported `--model` argument include: + +* mnist +* resnet + * you can chose to use different dataset using `--data_set cifar10` or + `--data_set flowers`. +* vgg +* stacked_dynamic_lstm +* machine_translation + +* Run the following command to start a benchmark job locally: + ```bash + python fluid_benchmark.py --model mnist --device GPU + ``` + You can choose to use GPU/CPU training. With GPU training, you can specify + `--gpus ` to run multi GPU training. +* Run distributed training with parameter servers: + * start parameter servers: + ```bash + PADDLE_TRAINING_ROLE=PSERVER PADDLE_PSERVER_PORT=7164 PADDLE_PSERVER_IPS=127.0.0.1 PADDLE_TRAINERS=1 PADDLE_CURRENT_IP=127.0.0.1 PADDLE_TRAINER_ID=0 python fluid_benchmark.py --model mnist --device GPU --update_method pserver + ``` + * start trainers: + ```bash + PADDLE_TRAINING_ROLE=TRAINER PADDLE_PSERVER_PORT=7164 PADDLE_PSERVER_IPS=127.0.0.1 PADDLE_TRAINERS=1 PADDLE_CURRENT_IP=127.0.0.1 PADDLE_TRAINER_ID=0 python fluid_benchmark.py --model mnist --device GPU --update_method pserver + ``` +* Run distributed training using NCCL2 + ```bash + PADDLE_PSERVER_PORT=7164 PADDLE_TRAINER_IPS=192.168.0.2,192.168.0.3 PADDLE_CURRENT_IP=127.0.0.1 PADDLE_TRAINER_ID=0 python fluid_benchmark.py --model mnist --device GPU --update_method nccl2 + ``` + +## Run Distributed Benchmark on Kubernetes Cluster + +We provide a script `kube_gen_job.py` to generate Kubernetes yaml files to submit +distributed benchmark jobs to your cluster. To generate a job yaml, just run: + +```bash +python kube_gen_job.py --jobname myjob --pscpu 4 --cpu 8 --gpu 8 --psmemory 20 --memory 40 --pservers 4 --trainers 4 --entry "python fluid_benchmark.py --model mnist --parallel 1 --device GPU --update_method pserver " --disttype pserver +``` + +Then the yaml files are generated under directory `myjob`, you can run: + +```bash +kubectl create -f myjob/ +``` + +The job shall start. + + +## Notes for Run Fluid Distributed with NCCL2 and RDMA + +Before running NCCL2 distributed jobs, please check that whether your node has multiple network +interfaces, try to add the environment variable `export NCCL_SOCKET_IFNAME=eth0` to use your actual +network device. + +To run high-performance distributed training, you must prepare your hardware environment to be +able to run RDMA enabled network communication, please check out [this](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/howto/cluster/nccl2_rdma_training.md) +note for details. diff --git a/benchmark/fluid/fluid_benchmark.py b/benchmark/fluid/fluid_benchmark.py new file mode 100644 index 0000000000000000000000000000000000000000..c1d458970a58bfac2a3369e8964eb100568b28f2 --- /dev/null +++ b/benchmark/fluid/fluid_benchmark.py @@ -0,0 +1,391 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import argparse +import cProfile +import time +import os + +import numpy as np + +import paddle.fluid as fluid +import paddle.fluid.core as core +import paddle.fluid.profiler as profiler +import paddle.fluid.transpiler.distribute_transpiler as distribute_transpiler + +BENCHMARK_MODELS = [ + "machine_translation", "resnet", "vgg", "mnist", "stacked_dynamic_lstm" +] + + +def parse_args(): + parser = argparse.ArgumentParser('Fluid model benchmarks.') + parser.add_argument( + '--model', + type=str, + choices=BENCHMARK_MODELS, + default='resnet', + help='The model to run benchmark with.') + parser.add_argument( + '--batch_size', type=int, default=32, help='The minibatch size.') + parser.add_argument( + '--learning_rate', + type=float, + default=0.001, + help='The minibatch size.') + # TODO(wuyi): add "--use_fake_data" option back. + parser.add_argument( + '--skip_batch_num', + type=int, + default=5, + help='The first num of minibatch num to skip, for better performance test' + ) + parser.add_argument( + '--iterations', type=int, default=80, help='The number of minibatches.') + parser.add_argument( + '--pass_num', type=int, default=100, help='The number of passes.') + parser.add_argument( + '--data_format', + type=str, + default='NCHW', + choices=['NCHW', 'NHWC'], + help='The data data_format, now only support NCHW.') + parser.add_argument( + '--device', + type=str, + default='GPU', + choices=['CPU', 'GPU'], + help='The device type.') + parser.add_argument( + '--gpus', + type=int, + default=1, + help='If gpus > 1, will use ParallelExecutor to run, else use Executor.') + parser.add_argument( + '--data_set', + type=str, + default='flowers', + choices=['cifar10', 'flowers'], + help='Optional dataset for benchmark.') + parser.add_argument( + '--infer_only', action='store_true', help='If set, run forward only.') + parser.add_argument( + '--use_cprof', action='store_true', help='If set, use cProfile.') + parser.add_argument( + '--use_nvprof', + action='store_true', + help='If set, use nvprof for CUDA.') + parser.add_argument( + '--no_test', + action='store_false', + help='If set, test the testset during training.') + parser.add_argument( + '--memory_optimize', + action='store_true', + help='If set, optimize runtime memory before start.') + parser.add_argument( + '--use_fake_data', + action='store_true', + help='If set ommit the actual read data operators.') + parser.add_argument( + '--profile', action='store_true', help='If set, profile a few steps.') + parser.add_argument( + '--update_method', + type=str, + default='local', + choices=['local', 'pserver', 'nccl2'], + help='Choose parameter update method, can be local, pserver, nccl2.') + args = parser.parse_args() + return args + + +def append_nccl2_prepare(trainer_id): + if trainer_id >= 0: + # append gen_nccl_id at the end of startup program + trainer_id = int(os.getenv("PADDLE_TRAINER_ID")) + port = os.getenv("PADDLE_PSERVER_PORT") + worker_ips = os.getenv("PADDLE_TRAINER_IPS") + worker_endpoints = [] + for ip in worker_ips.split(","): + worker_endpoints.append(':'.join([ip, port])) + num_trainers = len(worker_endpoints) + current_endpoint = os.getenv("PADDLE_CURRENT_IP") + ":" + port + worker_endpoints.remove(current_endpoint) + + nccl_id_var = fluid.default_startup_program().global_block().create_var( + name="NCCLID", + persistable=True, + type=fluid.core.VarDesc.VarType.RAW) + fluid.default_startup_program().global_block().append_op( + type="gen_nccl_id", + inputs={}, + outputs={"NCCLID": nccl_id_var}, + attrs={ + "endpoint": current_endpoint, + "endpoint_list": worker_endpoints, + "trainer_id": trainer_id + }) + return nccl_id_var, num_trainers, trainer_id + else: + raise Exception("must set positive PADDLE_TRAINER_ID env variables for " + "nccl-based dist train.") + + +def dist_transpile(trainer_id): + if trainer_id < 0: + return None, None + + # the port of all pservers, needed by both trainer and pserver + port = os.getenv("PADDLE_PSERVER_PORT", "6174") + # comma separated ips of all pservers, needed by trainer and + # pserver + pserver_ips = os.getenv("PADDLE_PSERVER_IPS", "") + eplist = [] + for ip in pserver_ips.split(","): + eplist.append(':'.join([ip, port])) + pserver_endpoints = ",".join(eplist) + # total number of workers/trainers in the job, needed by + # trainer and pserver + trainers = int(os.getenv("PADDLE_TRAINERS")) + # the IP of the local machine, needed by pserver only + current_endpoint = os.getenv("PADDLE_CURRENT_IP", "") + ":" + port + # the role, should be either PSERVER or TRAINER + training_role = os.getenv("PADDLE_TRAINING_ROLE") + + t = distribute_transpiler.DistributeTranspiler() + t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers) + if training_role == "PSERVER": + pserver_program = t.get_pserver_program(current_endpoint) + pserver_startup_program = t.get_startup_program(current_endpoint, + pserver_program) + return pserver_program, pserver_startup_program + elif training_role == "TRAINER": + train_program = t.get_trainer_program() + return train_program, fluid.default_startup_program() + else: + raise ValueError( + 'TRAINING_ROLE environment variable must be either TRAINER or PSERVER' + ) + + +def test(exe, inference_program, test_reader, feeder, batch_acc): + accuracy_evaluator = fluid.metrics.Accuracy() + for batch_id, data in enumerate(test_reader()): + acc = exe.run(inference_program, + feed=feeder.feed(data), + fetch_list=[batch_acc]) + accuracy_evaluator.update(value=np.array(acc), weight=len(data)) + + return accuracy_evaluator.eval() + + +# TODO(wuyi): replace train, train_parallel, test functions with new trainer +# API once it is ready. +def train(avg_loss, infer_prog, optimizer, train_reader, test_reader, batch_acc, + args, train_prog, startup_prog): + if os.getenv("PADDLE_TRAINING_ROLE") == "PSERVER": + place = core.CPUPlace() + exe = fluid.Executor(place) + exe.run(startup_prog) + exe.run(train_prog) + return + + if args.use_fake_data: + raise Exception( + "fake data is not supported in single GPU test for now.") + + place = core.CPUPlace() if args.device == 'CPU' else core.CUDAPlace(0) + exe = fluid.Executor(place) + exe.run(startup_prog) + feed_var_list = [ + var for var in train_prog.global_block().vars.itervalues() + if var.is_data + ] + feeder = fluid.DataFeeder(feed_var_list, place) + + iters, num_samples, start_time = 0, 0, time.time() + for pass_id in range(args.pass_num): + train_losses = [] + for batch_id, data in enumerate(train_reader()): + if iters == args.skip_batch_num: + start_time = time.time() + num_samples = 0 + if iters == args.iterations: + break + loss = exe.run(train_prog, + feed=feeder.feed(data), + fetch_list=[avg_loss]) + iters += 1 + num_samples += len(data) + train_losses.append(loss) + print("Pass: %d, Iter: %d, Loss: %f\n" % + (pass_id, iters, np.mean(train_losses))) + train_elapsed = time.time() - start_time + examples_per_sec = num_samples / train_elapsed + print('\nTotal examples: %d, total time: %.5f, %.5f examples/sec\n' % + (num_samples, train_elapsed, examples_per_sec)) + print("Pass: %d, Loss: %f" % (pass_id, np.mean(train_losses))) + # evaluation + if not args.no_test and batch_acc != None: + pass_test_acc = test(exe, infer_prog, test_reader, feeder, + batch_acc) + print(", Test Accuracy: %f" % pass_test_acc) + print("\n") + # TODO(wuyi): add warmup passes to get better perf data. + exit(0) + + +# TODO(wuyi): replace train, train_parallel, test functions with new trainer +# API once it is ready. +def train_parallel(avg_loss, infer_prog, optimizer, train_reader, test_reader, + batch_acc, args, train_prog, startup_prog, nccl_id_var, + num_trainers, trainer_id): + feed_var_list = [ + var for var in train_prog.global_block().vars.itervalues() + if var.is_data + ] + # generate fake: + if args.use_fake_data: + for var in feed_var_list: + v = startup_prog.global_block().clone_variable(var) + var.persistable = True + v.persistable = True + + real_shape = list(var.shape) + real_shape[0] = args.batch_size / args.gpus + startup_prog.global_block().append_op( + outputs={"Out": v}, + type="fill_constant", + attrs={"shape": real_shape, + "value": 1.0, + "dtype": var.dtype}) + + place = core.CPUPlace() if args.device == 'CPU' else core.CUDAPlace(0) + if nccl_id_var and trainer_id == 0: + #FIXME(wuyi): wait other trainer to start listening + time.sleep(30) + + startup_exe = fluid.Executor(place) + startup_exe.run(startup_prog) + strategy = fluid.ExecutionStrategy() + strategy.num_threads = 1 + strategy.allow_op_delay = False + exe = fluid.ParallelExecutor( + True, + avg_loss.name, + exec_strategy=strategy, + num_trainers=num_trainers, + trainer_id=trainer_id) + + feeder = fluid.DataFeeder(feed_var_list, place) + for pass_id in range(args.pass_num): + num_samples = 0 + iters = 0 + start_time = time.time() + for batch_id, data in enumerate(train_reader()): + if args.profile and pass_id == 0 and batch_id == 5: + profiler.start_profiler("All") + elif args.profile and pass_id == 0 and batch_id == 10: + profiler.stop_profiler("total", "/tmp/profile_%d" % trainer_id) + + if iters == args.skip_batch_num: + start_time = time.time() + num_samples = 0 + if iters == args.iterations: + break + if args.use_fake_data: + loss, = exe.run([avg_loss.name]) + else: + loss, = exe.run([avg_loss.name], feed=feeder.feed(data)) + if args.update_method == "pserver": + exe.bcast_params() + num_samples += len(data) + iters += 1 + if batch_id % 1 == 0: + print("Pass %d, batch %d, loss %s" % + (pass_id, batch_id, np.array(loss))) + train_elapsed = time.time() - start_time + examples_per_sec = num_samples / train_elapsed + print('\nTotal examples: %d, total time: %.5f, %.5f examples/sed\n' % + (num_samples, train_elapsed, examples_per_sec)) + if not args.no_test and batch_acc != None: + test_acc = test(startup_exe, infer_prog, test_reader, feeder, + batch_acc) + print("Pass: %d, Test Accuracy: %f\n" % (pass_id, test_acc)) + exit(0) + + +def print_arguments(args): + vars(args)['use_nvprof'] = (vars(args)['use_nvprof'] and + vars(args)['device'] == 'GPU') + print('----------- resnet Configuration Arguments -----------') + for arg, value in sorted(vars(args).iteritems()): + print('%s: %s' % (arg, value)) + print('------------------------------------------------') + + +def main(): + args = parse_args() + print_arguments(args) + + # the unique trainer id, starting from 0, needed by trainer + # only + nccl_id_var, num_trainers, trainer_id = ( + None, 1, int(os.getenv("PADDLE_TRAINER_ID", "-1"))) + + if args.use_cprof: + pr = cProfile.Profile() + pr.enable() + model_def = __import__("models.%s" % args.model, fromlist=["models"]) + train_args = list(model_def.get_model(args)) + train_args.append(args) + # Run optimizer.minimize(avg_loss) + train_args[2].minimize(train_args[0]) + if args.memory_optimize: + fluid.memory_optimize(fluid.default_main_program()) + + if args.update_method == "pserver": + train_prog, startup_prog = dist_transpile(trainer_id) + if not train_prog: + raise Exception( + "Must configure correct environments to run dist train.") + train_args.extend([train_prog, startup_prog]) + if args.gpus > 1 and os.getenv("PADDLE_TRAINING_ROLE") == "TRAINER": + train_args.extend([nccl_id_var, num_trainers, trainer_id]) + train_parallel(*train_args) + train(*train_args) + exit(0) + + # for other update methods, use default programs + train_args.append(fluid.default_main_program()) + train_args.append(fluid.default_startup_program()) + + if args.update_method == "nccl2": + nccl_id_var, num_trainers, trainer_id = append_nccl2_prepare(trainer_id) + if args.gpus == 1: + # NOTE: parallel executor use profiler interanlly + if args.use_nvprof and args.device == 'GPU': + with profiler.cuda_profiler("cuda_profiler.txt", 'csv') as nvprof: + train(*train_args) + else: + train(*train_args) + else: + if args.device == "CPU": + raise Exception("Only support GPU perf with parallel exe") + train_args.extend([nccl_id_var, num_trainers, trainer_id]) + train_parallel(*train_args) + + +if __name__ == "__main__": + main() diff --git a/benchmark/fluid/kube_gen_job.py b/benchmark/fluid/kube_gen_job.py new file mode 100644 index 0000000000000000000000000000000000000000..39ba207fd96f71563504017e77dc0e87c249b3f8 --- /dev/null +++ b/benchmark/fluid/kube_gen_job.py @@ -0,0 +1,191 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import yaml +import copy +import argparse +import random +import os +from kube_templates import pserver, trainer, envs + + +def parse_args(): + parser = argparse.ArgumentParser(description='Generate dist job yamls.') + + parser.add_argument( + '--jobname', default="paddlejob", help='unique job name') + parser.add_argument( + '--cpu', default=1, type=int, help='CPU cores per trainer node') + parser.add_argument( + '--pscpu', default=1, type=int, help='CPU cores per pserver node') + parser.add_argument( + '--gpu', default=0, type=int, help='num of GPUs per node') + parser.add_argument( + '--image', + default="bootstrapper:5000/fluid_benchmark:gpu", + help='num of GPUs per node') + parser.add_argument( + '--pservers', default=1, type=int, help='num of pservers') + parser.add_argument( + '--trainers', default=1, type=int, help='num of trainers') + parser.add_argument('--memory', default=1, type=int, help='trainer memory') + parser.add_argument( + '--psmemory', default=1, type=int, help='pserver memory') + parser.add_argument( + '--port', default=30236, type=int, help='num of trainers') + parser.add_argument( + '--entry', default="python train.py", help='command to run') + parser.add_argument( + '--fluid', default=1, type=int, help='whether is fluid job') + parser.add_argument( + '--rdma', action='store_ture', help='whether mount rdma libs') + parser.add_argument( + '--disttype', + default="pserver", + type=str, + choices=['pserver', 'nccl2', 'local'], + help='pserver or nccl2 or local') + + args = parser.parse_args() + return args + + +def gen_job(): + ps = pserver + tn = trainer + args = parse_args() + + ps_container = ps["spec"]["template"]["spec"]["containers"][0] + tn_container = tn["spec"]["template"]["spec"]["containers"][0] + + if args.fluid == 1: + ps_container["command"] = \ + ["paddle_k8s", "start_fluid"] + tn_container["command"] = \ + ["paddle_k8s", "start_fluid"] + ps["metadata"]["name"] = args.jobname + "-pserver" + ps["spec"]["template"]["metadata"]["labels"][ + "paddle-job-pserver"] = args.jobname + tn["metadata"]["name"] = args.jobname + "-trainer" + tn["spec"]["template"]["metadata"]["labels"]["paddle-job"] = args.jobname + + ps_container["image"] = args.image + tn_container["image"] = args.image + + ps_container["resources"]["requests"]["cpu"] = str(args.pscpu) + ps_container["resources"]["requests"]["memory"] = str(args.psmemory) + "Gi" + ps_container["resources"]["limits"]["cpu"] = str(args.pscpu) + ps_container["resources"]["limits"]["memory"] = str(args.psmemory) + "Gi" + + tn_container["resources"]["requests"]["cpu"] = str(args.cpu) + tn_container["resources"]["requests"]["memory"] = str(args.memory) + "Gi" + tn_container["resources"]["limits"]["cpu"] = str(args.cpu) + tn_container["resources"]["limits"]["memory"] = str(args.memory) + "Gi" + if args.gpu > 0: + tn_container["resources"]["requests"][ + "alpha.kubernetes.io/nvidia-gpu"] = str(args.gpu) + tn_container["resources"]["limits"][ + "alpha.kubernetes.io/nvidia-gpu"] = str(args.gpu) + + ps["spec"]["replicas"] = int(args.pservers) + tn["spec"]["parallelism"] = int(args.trainers) + tn["spec"]["completions"] = int(args.trainers) + ps_container["ports"][0]["name"] = "jobport-" + str(args.port) + ps_container["ports"][0]["containerPort"] = args.port + spreadport = random.randint(40000, 60000) + tn_container["ports"][0]["name"] = "spr-" + str(spreadport) + tn_container["ports"][0]["containerPort"] = spreadport + + envs.append({"name": "PADDLE_JOB_NAME", "value": args.jobname}) + envs.append({"name": "TRAINERS", "value": str(args.trainers)}) + envs.append({"name": "PSERVERS", "value": str(args.pservers)}) + envs.append({"name": "ENTRY", "value": args.entry}) + envs.append({"name": "PADDLE_INIT_PORT", "value": str(args.port)}) + envs.append({"name": "PADDLE_PSERVER_PORT", "value": str(args.port)}) + # NOTE: these directories below are cluster specific, please modify + # this settings before you run on your own cluster. + envs.append({ + "name": "LD_LIBRARY_PATH", + "value": + "/usr/local/lib:/usr/local/nvidia/lib64:/usr/local/rdma/lib64:/usr/lib64/mlnx_ofed/valgrind" + }) + + volumes = [{ + "name": "nvidia-driver", + "hostPath": { + "path": "/usr/local/nvidia/lib64" + } + }] + volumeMounts = [{ + "mountPath": "/usr/local/nvidia/lib64", + "name": "nvidia-driver" + }] + + if args.rdma: + volumes.extend([{ + "name": "ibetc", + "hostPath": { + "path": "/etc/libibverbs.d" + } + }, { + "name": "iblibs", + "hostPath": { + "path": "/usr/local/rdma" + } + }, { + "name": "valgrind", + "hostPath": { + "path": "/usr/lib64/mlnx_ofed/valgrind" + } + }]) + volumeMounts.extend([{ + "mountPath": "/etc/libibverbs.d", + "name": "ibetc" + }, { + "mountPath": "/usr/local/rdma", + "name": "iblibs" + }, { + "mountPath": "/usr/lib64/mlnx_ofed/valgrind", + "name": "valgrind" + }]) + # append shm for NCCL2 + volumes.append({"name": "dshm", "emptyDir": {"medium": "Memory"}}) + volumeMounts.append({"mountPath": "/dev/shm", "name": "dshm"}) + + tn["spec"]["template"]["spec"]["volumes"] = volumes + tn_container["volumeMounts"] = volumeMounts + + ps_container["env"] = envs + ps_container["env"].append({"name": "TRAINING_ROLE", "value": "PSERVER"}) + tn_container["env"] = envs + if args.disttype == "pserver": + tn_container["env"].append({ + "name": "TRAINING_ROLE", + "value": "TRAINER" + }) + elif args.disttype == "nccl2" or args.disttype == "local": + # NCCL2 have no training role, set to plain WORKER + tn_container["env"].append({"name": "TRAINING_ROLE", "value": "WORKER"}) + + os.mkdir(args.jobname) + if args.disttype == "pserver": + with open("%s/pserver.yaml" % args.jobname, "w") as fn: + yaml.dump(ps, fn) + + with open("%s/trainer.yaml" % args.jobname, "w") as fn: + yaml.dump(tn, fn) + + +if __name__ == "__main__": + gen_job() diff --git a/benchmark/fluid/kube_templates/__init__.py b/benchmark/fluid/kube_templates/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..2d09d940a5ee638e4b55405d05924e2d76006cfc --- /dev/null +++ b/benchmark/fluid/kube_templates/__init__.py @@ -0,0 +1,66 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from pserver import pserver +from trainer import trainer + +__all__ = ["pserver", "trainer", "envs"] + +envs = [ + # envs that don't need to change + { + "name": "GLOG_v", + "value": "0" + }, + { + "name": "GLOG_logtostderr", + "value": "1" + }, + { + "name": "TOPOLOGY", + "value": "" + }, + { + "name": "TRAINER_PACKAGE", + "value": "/workspace" + }, + { + "name": "PADDLE_INIT_NICS", + "value": "eth2" + }, + { + "name": "NAMESPACE", + "valueFrom": { + "fieldRef": { + "fieldPath": "metadata.namespace" + } + } + }, + { + "name": "POD_IP", + "valueFrom": { + "fieldRef": { + "fieldPath": "status.podIP" + } + } + }, + { + "name": "PADDLE_CURRENT_IP", + "valueFrom": { + "fieldRef": { + "fieldPath": "status.podIP" + } + } + } +] diff --git a/benchmark/fluid/kube_templates/pserver.py b/benchmark/fluid/kube_templates/pserver.py new file mode 100644 index 0000000000000000000000000000000000000000..b54982c806ad4229fbd4bd7edf82a4e7eb4c5ad1 --- /dev/null +++ b/benchmark/fluid/kube_templates/pserver.py @@ -0,0 +1,58 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +pserver = { + "apiVersion": "extensions/v1beta1", + "kind": "ReplicaSet", + "metadata": { + "name": "jobname-pserver" + }, + "spec": { + "replicas": 1, + "template": { + "metadata": { + "labels": { + "paddle-job-pserver": "jobname" + } + }, + "spec": { + "hostNetwork": True, + "imagePullSecrets": [{ + "name": "job-registry-secret" + }], + "containers": [{ + "name": "pserver", + "image": "", + "imagePullPolicy": "Always", + "ports": [{ + "name": "jobport-1", + "containerPort": 1 + }], + "env": [], + "command": ["paddle_k8s", "start_pserver"], + "resources": { + "requests": { + "memory": "10Gi", + "cpu": "4" + }, + "limits": { + "memory": "10Gi", + "cpu": "4" + } + } + }] + } + } + } +} diff --git a/benchmark/fluid/kube_templates/trainer.py b/benchmark/fluid/kube_templates/trainer.py new file mode 100644 index 0000000000000000000000000000000000000000..b915d31e371d9d787ff64d705e32baf301e16abe --- /dev/null +++ b/benchmark/fluid/kube_templates/trainer.py @@ -0,0 +1,70 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +trainer = { + "apiVersion": "batch/v1", + "kind": "Job", + "metadata": { + "name": "jobname-pserver" + }, + "spec": { + "parallelism": 4, + "completions": 4, + "template": { + "metadata": { + "labels": { + "paddle-job": "jobname" + } + }, + "spec": { + "hostNetwork": True, + "imagePullSecrets": [{ + "name": "job-registry-secret" + }], + "restartPolicy": "Never", + "containers": [{ + "name": "trainer", + "image": "", + "imagePullPolicy": "Always", + # to let container set rlimit + "securityContext": { + "privileged": True + # TODO(wuyi): use below specific cap instead of privileged, + # using privileged will cause all GPU device are visible + # in the container. + # "capabilities": { + # "add": ["SYS_RESOURCE"] + # } + }, + "ports": [{ + "name": "jobport-1", + "containerPort": 1 + }], + "env": [], + "command": ["paddle_k8s", "start_trainer", "v2"], + "resources": { + "requests": { + "memory": "10Gi", + "cpu": "4", + }, + "limits": { + "memory": "10Gi", + "cpu": "4", + } + } + }] + } + } + } +} diff --git a/benchmark/fluid/mnist.py b/benchmark/fluid/mnist.py deleted file mode 100644 index 1e2185dfac1072d1f1046f4616a9d53a8fc76061..0000000000000000000000000000000000000000 --- a/benchmark/fluid/mnist.py +++ /dev/null @@ -1,224 +0,0 @@ -# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -from __future__ import absolute_import -from __future__ import division -from __future__ import print_function - -import numpy as np -import argparse -import time - -import paddle -import paddle.fluid as fluid -import paddle.fluid.profiler as profiler - -SEED = 1 -DTYPE = "float32" - -# random seed must set before configuring the network. -# fluid.default_startup_program().random_seed = SEED - - -def parse_args(): - parser = argparse.ArgumentParser("mnist model benchmark.") - parser.add_argument( - '--batch_size', type=int, default=128, help='The minibatch size.') - parser.add_argument( - '--skip_batch_num', - type=int, - default=5, - help='The first num of minibatch num to skip, for better performance test' - ) - parser.add_argument( - '--iterations', type=int, default=35, help='The number of minibatches.') - parser.add_argument( - '--pass_num', type=int, default=5, help='The number of passes.') - parser.add_argument( - '--device', - type=str, - default='GPU', - choices=['CPU', 'GPU'], - help='The device type.') - parser.add_argument( - '--infer_only', action='store_true', help='If set, run forward only.') - parser.add_argument( - '--use_cprof', action='store_true', help='If set, use cProfile.') - parser.add_argument( - '--use_nvprof', - action='store_true', - help='If set, use nvprof for CUDA.') - parser.add_argument( - '--with_test', - action='store_true', - help='If set, test the testset during training.') - args = parser.parse_args() - return args - - -def cnn_model(data): - conv_pool_1 = fluid.nets.simple_img_conv_pool( - input=data, - filter_size=5, - num_filters=20, - pool_size=2, - pool_stride=2, - act="relu") - conv_pool_2 = fluid.nets.simple_img_conv_pool( - input=conv_pool_1, - filter_size=5, - num_filters=50, - pool_size=2, - pool_stride=2, - act="relu") - - # TODO(dzhwinter) : refine the initializer and random seed settting - SIZE = 10 - input_shape = conv_pool_2.shape - param_shape = [reduce(lambda a, b: a * b, input_shape[1:], 1)] + [SIZE] - scale = (2.0 / (param_shape[0]**2 * SIZE))**0.5 - - predict = fluid.layers.fc( - input=conv_pool_2, - size=SIZE, - act="softmax", - param_attr=fluid.param_attr.ParamAttr( - initializer=fluid.initializer.NormalInitializer( - loc=0.0, scale=scale))) - return predict - - -def eval_test(exe, batch_acc, batch_size_tensor, inference_program): - test_reader = paddle.batch( - paddle.dataset.mnist.test(), batch_size=args.batch_size) - test_pass_acc = fluid.average.WeightedAverage() - for batch_id, data in enumerate(test_reader()): - img_data = np.array(map(lambda x: x[0].reshape([1, 28, 28]), - data)).astype(DTYPE) - y_data = np.array(map(lambda x: x[1], data)).astype("int64") - y_data = y_data.reshape([len(y_data), 1]) - - acc, weight = exe.run(inference_program, - feed={"pixel": img_data, - "label": y_data}, - fetch_list=[batch_acc, batch_size_tensor]) - test_pass_acc.add(value=acc, weight=weight) - pass_acc = test_pass_acc.eval() - return pass_acc - - -def run_benchmark(model, args): - if args.use_cprof: - pr = cProfile.Profile() - pr.enable() - start_time = time.time() - # Input data - images = fluid.layers.data(name='pixel', shape=[1, 28, 28], dtype=DTYPE) - label = fluid.layers.data(name='label', shape=[1], dtype='int64') - - # Train program - predict = model(images) - cost = fluid.layers.cross_entropy(input=predict, label=label) - avg_cost = fluid.layers.mean(x=cost) - - # Evaluator - batch_size_tensor = fluid.layers.create_tensor(dtype='int64') - batch_acc = fluid.layers.accuracy( - input=predict, label=label, total=batch_size_tensor) - - # inference program - inference_program = fluid.default_main_program().clone() - - # Optimization - opt = fluid.optimizer.AdamOptimizer( - learning_rate=0.001, beta1=0.9, beta2=0.999) - opt.minimize(avg_cost) - - fluid.memory_optimize(fluid.default_main_program()) - - # Initialize executor - place = fluid.CPUPlace() if args.device == 'CPU' else fluid.CUDAPlace(0) - exe = fluid.Executor(place) - - # Parameter initialization - exe.run(fluid.default_startup_program()) - - # Reader - train_reader = paddle.batch( - paddle.dataset.mnist.train(), batch_size=args.batch_size) - - accuracy = fluid.metrics.Accuracy() - iters, num_samples, start_time = 0, 0, time.time() - for pass_id in range(args.pass_num): - accuracy.reset() - train_accs = [] - train_losses = [] - for batch_id, data in enumerate(train_reader()): - if iters == args.skip_batch_num: - start_time = time.time() - num_samples = 0 - if iters == args.iterations: - break - img_data = np.array( - map(lambda x: x[0].reshape([1, 28, 28]), data)).astype(DTYPE) - y_data = np.array(map(lambda x: x[1], data)).astype("int64") - y_data = y_data.reshape([len(y_data), 1]) - - outs = exe.run( - fluid.default_main_program(), - feed={"pixel": img_data, - "label": y_data}, - fetch_list=[avg_cost, batch_acc, batch_size_tensor] - ) # The accuracy is the accumulation of batches, but not the current batch. - accuracy.update(value=outs[1], weight=outs[2]) - iters += 1 - num_samples += len(y_data) - loss = np.array(outs[0]) - acc = np.array(outs[1]) - train_losses.append(loss) - train_accs.append(acc) - print("Pass: %d, Iter: %d, Loss: %f, Accuracy: %f" % - (pass_id, iters, loss, acc)) - - print("Pass: %d, Loss: %f, Train Accuray: %f\n" % - (pass_id, np.mean(train_losses), np.mean(train_accs))) - train_elapsed = time.time() - start_time - examples_per_sec = num_samples / train_elapsed - - print('\nTotal examples: %d, total time: %.5f, %.5f examples/sed\n' % - (num_samples, train_elapsed, examples_per_sec)) - # evaluation - if args.with_test: - test_avg_acc = eval_test(exe, batch_acc, batch_size_tensor, - inference_program) - exit(0) - - -def print_arguments(args): - vars(args)['use_nvprof'] = (vars(args)['use_nvprof'] and - vars(args)['device'] == 'GPU') - print('----------- mnist Configuration Arguments -----------') - for arg, value in sorted(vars(args).iteritems()): - print('%s: %s' % (arg, value)) - print('------------------------------------------------') - - -if __name__ == '__main__': - args = parse_args() - print_arguments(args) - if args.use_nvprof and args.device == 'GPU': - with profiler.cuda_profiler("cuda_profiler.txt", 'csv') as nvprof: - run_benchmark(cnn_model, args) - else: - run_benchmark(cnn_model, args) diff --git a/benchmark/fluid/models/__init__.py b/benchmark/fluid/models/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..1c3fcac8dd4a1ba0496ef013bd4eb468a0075125 --- /dev/null +++ b/benchmark/fluid/models/__init__.py @@ -0,0 +1,17 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +__all__ = [ + "machine_translation", "resnet", "vgg", "mnist", "stacked_dynamic_lstm" +] diff --git a/benchmark/fluid/machine_translation.py b/benchmark/fluid/models/machine_translation.py similarity index 60% rename from benchmark/fluid/machine_translation.py rename to benchmark/fluid/models/machine_translation.py index adde5f21acd4e77d58a453d6868abeccfca4bb5a..635b3373dd27b21f83afae10b1d24833b81d57eb 100644 --- a/benchmark/fluid/machine_translation.py +++ b/benchmark/fluid/models/machine_translation.py @@ -27,74 +27,6 @@ import paddle.fluid.core as core import paddle.fluid.framework as framework from paddle.fluid.executor import Executor -parser = argparse.ArgumentParser(description=__doc__) -parser.add_argument( - "--embedding_dim", - type=int, - default=512, - help="The dimension of embedding table. (default: %(default)d)") -parser.add_argument( - "--encoder_size", - type=int, - default=512, - help="The size of encoder bi-rnn unit. (default: %(default)d)") -parser.add_argument( - "--decoder_size", - type=int, - default=512, - help="The size of decoder rnn unit. (default: %(default)d)") -parser.add_argument( - "--batch_size", - type=int, - default=16, - help="The sequence number of a mini-batch data. (default: %(default)d)") -parser.add_argument( - '--skip_batch_num', - type=int, - default=5, - help='The first num of minibatch num to skip, for better performance test') -parser.add_argument( - '--iterations', type=int, default=80, help='The number of minibatches.') -parser.add_argument( - "--dict_size", - type=int, - default=30000, - help="The dictionary capacity. Dictionaries of source sequence and " - "target dictionary have same capacity. (default: %(default)d)") -parser.add_argument( - "--pass_num", - type=int, - default=2, - help="The pass number to train. (default: %(default)d)") -parser.add_argument( - "--learning_rate", - type=float, - default=0.0002, - help="Learning rate used to train the model. (default: %(default)f)") -parser.add_argument( - "--infer_only", action='store_true', help="If set, run forward only.") -parser.add_argument( - "--beam_size", - type=int, - default=3, - help="The width for beam searching. (default: %(default)d)") -parser.add_argument( - '--device', - type=str, - default='GPU', - choices=['CPU', 'GPU'], - help="The device type.") -parser.add_argument( - "--max_length", - type=int, - default=250, - help="The maximum length of sequence when doing generation. " - "(default: %(default)d)") -parser.add_argument( - '--with_test', - action='store_true', - help='If set, test the testset during training.') - def lstm_step(x_t, hidden_t_prev, cell_t_prev, size): def linear(inputs): @@ -264,116 +196,37 @@ def lodtensor_to_ndarray(lod_tensor): return ndarray -def train(): +def get_model(args): + embedding_dim = 512 + encoder_size = 512 + decoder_size = 512 + dict_size = 30000 + beam_size = 3 + max_length = 250 avg_cost, feeding_list = seq_to_seq_net( - args.embedding_dim, - args.encoder_size, - args.decoder_size, - args.dict_size, - args.dict_size, + embedding_dim, + encoder_size, + decoder_size, + dict_size, + dict_size, False, - beam_size=args.beam_size, - max_length=args.max_length) + beam_size=beam_size, + max_length=max_length) # clone from default main program inference_program = fluid.default_main_program().clone() optimizer = fluid.optimizer.Adam(learning_rate=args.learning_rate) - optimizer.minimize(avg_cost) - - fluid.memory_optimize(fluid.default_main_program()) train_batch_generator = paddle.batch( paddle.reader.shuffle( - paddle.dataset.wmt14.train(args.dict_size), buf_size=1000), + paddle.dataset.wmt14.train(dict_size), buf_size=1000), batch_size=args.batch_size) test_batch_generator = paddle.batch( paddle.reader.shuffle( - paddle.dataset.wmt14.test(args.dict_size), buf_size=1000), + paddle.dataset.wmt14.test(dict_size), buf_size=1000), batch_size=args.batch_size) - place = core.CPUPlace() if args.device == 'CPU' else core.CUDAPlace(0) - exe = Executor(place) - exe.run(framework.default_startup_program()) - - def do_validation(): - total_loss = 0.0 - count = 0 - for batch_id, data in enumerate(test_batch_generator()): - src_seq = to_lodtensor(map(lambda x: x[0], data), place)[0] - trg_seq = to_lodtensor(map(lambda x: x[1], data), place)[0] - lbl_seq = to_lodtensor(map(lambda x: x[2], data), place)[0] - - fetch_outs = exe.run(inference_program, - feed={ - feeding_list[0]: src_seq, - feeding_list[1]: trg_seq, - feeding_list[2]: lbl_seq - }, - fetch_list=[avg_cost], - return_numpy=False) - - total_loss += lodtensor_to_ndarray(fetch_outs[0])[0] - count += 1 - - return total_loss / count - - iters, num_samples, start_time = 0, 0, time.time() - for pass_id in xrange(args.pass_num): - train_accs = [] - train_losses = [] - for batch_id, data in enumerate(train_batch_generator()): - if iters == args.skip_batch_num: - start_time = time.time() - num_samples = 0 - if iters == args.iterations: - break - src_seq, word_num = to_lodtensor(map(lambda x: x[0], data), place) - num_samples += word_num - trg_seq, word_num = to_lodtensor(map(lambda x: x[1], data), place) - num_samples += word_num - lbl_seq, _ = to_lodtensor(map(lambda x: x[2], data), place) - - fetch_outs = exe.run(framework.default_main_program(), - feed={ - feeding_list[0]: src_seq, - feeding_list[1]: trg_seq, - feeding_list[2]: lbl_seq - }, - fetch_list=[avg_cost]) - - iters += 1 - loss = np.array(fetch_outs[0]) - print( - "Pass = %d, Iter = %d, Loss = %f" % (pass_id, iters, loss) - ) # The accuracy is the accumulation of batches, but not the current batch. - - train_elapsed = time.time() - start_time - examples_per_sec = num_samples / train_elapsed - print('\nTotal examples: %d, total time: %.5f, %.5f examples/sed\n' % - (num_samples, train_elapsed, examples_per_sec)) - # evaluation - if args.with_test: - test_loss = do_validation() - exit(0) - - -def infer(): - pass - - -def print_arguments(args): - print('----------- seq2seq Configuration Arguments -----------') - for arg, value in sorted(vars(args).iteritems()): - print('%s: %s' % (arg, value)) - print('------------------------------------------------') - - -if __name__ == '__main__': - args = parser.parse_args() - print_arguments(args) - if args.infer_only: - infer() - else: - train() + return avg_cost, inference_program, optimizer, train_batch_generator, \ + test_batch_generator, None diff --git a/benchmark/fluid/models/mnist.py b/benchmark/fluid/models/mnist.py new file mode 100644 index 0000000000000000000000000000000000000000..d264bfc12bdb159c06dae81db4949b9ee17268e2 --- /dev/null +++ b/benchmark/fluid/models/mnist.py @@ -0,0 +1,94 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import numpy as np +import argparse +import time +import cProfile + +import paddle +import paddle.fluid as fluid +import paddle.fluid.profiler as profiler + +SEED = 1 +DTYPE = "float32" + +# random seed must set before configuring the network. +# fluid.default_startup_program().random_seed = SEED + + +def cnn_model(data): + conv_pool_1 = fluid.nets.simple_img_conv_pool( + input=data, + filter_size=5, + num_filters=20, + pool_size=2, + pool_stride=2, + act="relu") + conv_pool_2 = fluid.nets.simple_img_conv_pool( + input=conv_pool_1, + filter_size=5, + num_filters=50, + pool_size=2, + pool_stride=2, + act="relu") + + # TODO(dzhwinter) : refine the initializer and random seed settting + SIZE = 10 + input_shape = conv_pool_2.shape + param_shape = [reduce(lambda a, b: a * b, input_shape[1:], 1)] + [SIZE] + scale = (2.0 / (param_shape[0]**2 * SIZE))**0.5 + + predict = fluid.layers.fc( + input=conv_pool_2, + size=SIZE, + act="softmax", + param_attr=fluid.param_attr.ParamAttr( + initializer=fluid.initializer.NormalInitializer( + loc=0.0, scale=scale))) + return predict + + +def get_model(args): + # Input data + images = fluid.layers.data(name='pixel', shape=[1, 28, 28], dtype=DTYPE) + label = fluid.layers.data(name='label', shape=[1], dtype='int64') + + # Train program + predict = cnn_model(images) + cost = fluid.layers.cross_entropy(input=predict, label=label) + avg_cost = fluid.layers.mean(x=cost) + + # Evaluator + batch_size_tensor = fluid.layers.create_tensor(dtype='int64') + batch_acc = fluid.layers.accuracy( + input=predict, label=label, total=batch_size_tensor) + + # inference program + inference_program = fluid.default_main_program().clone() + + # Optimization + opt = fluid.optimizer.AdamOptimizer( + learning_rate=0.001, beta1=0.9, beta2=0.999) + + # Reader + train_reader = paddle.batch( + paddle.dataset.mnist.train(), batch_size=args.batch_size) + test_reader = paddle.batch( + paddle.dataset.mnist.test(), batch_size=args.batch_size) + return avg_cost, inference_program, opt, train_reader, test_reader, batch_acc diff --git a/benchmark/fluid/models/resnet.py b/benchmark/fluid/models/resnet.py new file mode 100644 index 0000000000000000000000000000000000000000..9dec8911ed64e09285fb461c4a12adb601535316 --- /dev/null +++ b/benchmark/fluid/models/resnet.py @@ -0,0 +1,161 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import functools +import numpy as np +import time + +import cProfile, pstats, StringIO + +import paddle +import paddle.fluid as fluid +import paddle.fluid.core as core +import paddle.fluid.profiler as profiler + + +def conv_bn_layer(input, ch_out, filter_size, stride, padding, act='relu'): + conv1 = fluid.layers.conv2d( + input=input, + filter_size=filter_size, + num_filters=ch_out, + stride=stride, + padding=padding, + act=None, + bias_attr=False) + return fluid.layers.batch_norm(input=conv1, act=act) + + +def shortcut(input, ch_out, stride): + ch_in = input.shape[1] # if args.data_format == 'NCHW' else input.shape[-1] + if ch_in != ch_out: + return conv_bn_layer(input, ch_out, 1, stride, 0, None) + else: + return input + + +def basicblock(input, ch_out, stride): + short = shortcut(input, ch_out, stride) + conv1 = conv_bn_layer(input, ch_out, 3, stride, 1) + conv2 = conv_bn_layer(conv1, ch_out, 3, 1, 1, act=None) + return fluid.layers.elementwise_add(x=short, y=conv2, act='relu') + + +def bottleneck(input, ch_out, stride): + short = shortcut(input, ch_out * 4, stride) + conv1 = conv_bn_layer(input, ch_out, 1, stride, 0) + conv2 = conv_bn_layer(conv1, ch_out, 3, 1, 1) + conv3 = conv_bn_layer(conv2, ch_out * 4, 1, 1, 0, act=None) + return fluid.layers.elementwise_add(x=short, y=conv3, act='relu') + + +def layer_warp(block_func, input, ch_out, count, stride): + res_out = block_func(input, ch_out, stride) + for i in range(1, count): + res_out = block_func(res_out, ch_out, 1) + return res_out + + +def resnet_imagenet(input, class_dim, depth=50, data_format='NCHW'): + + cfg = { + 18: ([2, 2, 2, 1], basicblock), + 34: ([3, 4, 6, 3], basicblock), + 50: ([3, 4, 6, 3], bottleneck), + 101: ([3, 4, 23, 3], bottleneck), + 152: ([3, 8, 36, 3], bottleneck) + } + stages, block_func = cfg[depth] + conv1 = conv_bn_layer(input, ch_out=64, filter_size=7, stride=2, padding=3) + pool1 = fluid.layers.pool2d( + input=conv1, pool_type='avg', pool_size=3, pool_stride=2) + res1 = layer_warp(block_func, pool1, 64, stages[0], 1) + res2 = layer_warp(block_func, res1, 128, stages[1], 2) + res3 = layer_warp(block_func, res2, 256, stages[2], 2) + res4 = layer_warp(block_func, res3, 512, stages[3], 2) + pool2 = fluid.layers.pool2d( + input=res4, + pool_size=7, + pool_type='avg', + pool_stride=1, + global_pooling=True) + out = fluid.layers.fc(input=pool2, size=class_dim, act='softmax') + return out + + +def resnet_cifar10(input, class_dim, depth=32, data_format='NCHW'): + assert (depth - 2) % 6 == 0 + + n = (depth - 2) // 6 + + conv1 = conv_bn_layer( + input=input, ch_out=16, filter_size=3, stride=1, padding=1) + res1 = layer_warp(basicblock, conv1, 16, n, 1) + res2 = layer_warp(basicblock, res1, 32, n, 2) + res3 = layer_warp(basicblock, res2, 64, n, 2) + pool = fluid.layers.pool2d( + input=res3, pool_size=8, pool_type='avg', pool_stride=1) + out = fluid.layers.fc(input=pool, size=class_dim, act='softmax') + return out + + +def get_model(args): + model = resnet_cifar10 + if args.data_set == "cifar10": + class_dim = 10 + if args.data_format == 'NCHW': + dshape = [3, 32, 32] + else: + dshape = [32, 32, 3] + model = resnet_cifar10 + else: + class_dim = 102 + if args.data_format == 'NCHW': + dshape = [3, 224, 224] + else: + dshape = [224, 224, 3] + model = resnet_imagenet + + input = fluid.layers.data(name='data', shape=dshape, dtype='float32') + label = fluid.layers.data(name='label', shape=[1], dtype='int64') + predict = model(input, class_dim) + cost = fluid.layers.cross_entropy(input=predict, label=label) + avg_cost = fluid.layers.mean(x=cost) + + batch_size_tensor = fluid.layers.create_tensor(dtype='int64') + batch_acc = fluid.layers.accuracy( + input=predict, label=label, total=batch_size_tensor) + + inference_program = fluid.default_main_program().clone() + with fluid.program_guard(inference_program): + inference_program = fluid.io.get_inference_program( + target_vars=[batch_acc, batch_size_tensor]) + + optimizer = fluid.optimizer.Momentum(learning_rate=0.01, momentum=0.9) + + train_reader = paddle.batch( + paddle.reader.shuffle( + paddle.dataset.cifar.train10() + if args.data_set == 'cifar10' else paddle.dataset.flowers.train(), + buf_size=5120), + batch_size=args.batch_size) + test_reader = paddle.batch( + paddle.dataset.cifar.test10() + if args.data_set == 'cifar10' else paddle.dataset.flowers.test(), + batch_size=args.batch_size) + + return avg_cost, inference_program, optimizer, train_reader, test_reader, batch_acc diff --git a/benchmark/fluid/stacked_dynamic_lstm.py b/benchmark/fluid/models/stacked_dynamic_lstm.py similarity index 52% rename from benchmark/fluid/stacked_dynamic_lstm.py rename to benchmark/fluid/models/stacked_dynamic_lstm.py index 73bcc47b4d404af2c01d61ca3dfb11971bbcfe9c..81a28b5f3aed0c325398b909d700c23df545824a 100644 --- a/benchmark/fluid/stacked_dynamic_lstm.py +++ b/benchmark/fluid/models/stacked_dynamic_lstm.py @@ -29,57 +29,6 @@ import paddle.fluid as fluid import paddle.batch as batch import paddle.fluid.profiler as profiler - -def parse_args(): - parser = argparse.ArgumentParser("Understand Sentiment by Dynamic RNN.") - parser.add_argument( - '--batch_size', - type=int, - default=32, - help='The sequence number of a batch data. (default: %(default)d)') - parser.add_argument( - '--skip_batch_num', - type=int, - default=5, - help='The first num of minibatch num to skip, for better performance test' - ) - parser.add_argument( - '--iterations', type=int, default=80, help='The number of minibatches.') - parser.add_argument( - '--emb_dim', - type=int, - default=512, - help='Dimension of embedding table. (default: %(default)d)') - parser.add_argument( - '--hidden_dim', - type=int, - default=512, - help='Hidden size of lstm unit. (default: %(default)d)') - parser.add_argument( - '--pass_num', - type=int, - default=100, - help='Epoch number to train. (default: %(default)d)') - parser.add_argument( - '--device', - type=str, - default='CPU', - choices=['CPU', 'GPU'], - help='The device type.') - parser.add_argument( - '--crop_size', - type=int, - default=int(os.environ.get('CROP_SIZE', '1500')), - help='The max sentence length of input. Since this model use plain RNN,' - ' Gradient could be explored if sentence is too long') - parser.add_argument( - '--with_test', - action='store_true', - help='If set, test the testset during training.') - args = parser.parse_args() - return args - - word_dict = imdb.word_dict() @@ -94,14 +43,15 @@ def crop_sentence(reader, crop_size): return __impl__ -def main(): - args = parse_args() - lstm_size = args.hidden_dim +def get_model(args): + lstm_size = 512 + emb_dim = 512 + crop_size = 1500 data = fluid.layers.data( name="words", shape=[1], lod_level=1, dtype='int64') sentence = fluid.layers.embedding( - input=data, size=[len(word_dict), args.emb_dim]) + input=data, size=[len(word_dict), emb_dim]) sentence = fluid.layers.fc(input=sentence, size=lstm_size, act='tanh') @@ -161,51 +111,17 @@ def main(): target_vars=[batch_acc, batch_size_tensor]) adam = fluid.optimizer.Adam() - adam.minimize(loss) - - fluid.memory_optimize(fluid.default_main_program()) - - place = fluid.CPUPlace() if args.device == 'CPU' else fluid.CUDAPlace(0) - exe = fluid.Executor(place) - exe.run(fluid.default_startup_program()) train_reader = batch( paddle.reader.shuffle( - crop_sentence(imdb.train(word_dict), args.crop_size), - buf_size=25000), + crop_sentence(imdb.train(word_dict), crop_size), buf_size=25000), + batch_size=args.batch_size) + test_reader = batch( + paddle.reader.shuffle( + crop_sentence(imdb.test(word_dict), crop_size), buf_size=25000), batch_size=args.batch_size) - iters, num_samples, start_time = 0, 0, time.time() - for pass_id in range(args.pass_num): - train_accs = [] - train_losses = [] - for batch_id, data in enumerate(train_reader()): - if iters == args.skip_batch_num: - start_time = time.time() - num_samples = 0 - if iters == args.iterations: - break - tensor_words = to_lodtensor([x[0] for x in data], place) - label = numpy.array([x[1] for x in data]).astype("int64") - label = label.reshape((-1, 1)) - loss_np, acc, weight = exe.run( - fluid.default_main_program(), - feed={"words": tensor_words, - "label": label}, - fetch_list=[loss, batch_acc, batch_size_tensor]) - iters += 1 - for x in data: - num_samples += len(x[0]) - print( - "Pass = %d, Iter = %d, Loss = %f, Accuracy = %f" % - (pass_id, iters, loss_np, acc) - ) # The accuracy is the accumulation of batches, but not the current batch. - - train_elapsed = time.time() - start_time - examples_per_sec = num_samples / train_elapsed - print('\nTotal examples: %d, total time: %.5f, %.5f examples/sed\n' % - (num_samples, train_elapsed, examples_per_sec)) - exit(0) + return loss, inference_program, adam, train_reader, test_reader, batch_acc def to_lodtensor(data, place): @@ -221,16 +137,3 @@ def to_lodtensor(data, place): res.set(flattened_data, place) res.set_lod([lod]) return res - - -def print_arguments(args): - print('----------- lstm Configuration Arguments -----------') - for arg, value in sorted(vars(args).iteritems()): - print('%s: %s' % (arg, value)) - print('------------------------------------------------') - - -if __name__ == '__main__': - args = parse_args() - print_arguments(args) - main() diff --git a/benchmark/fluid/models/vgg.py b/benchmark/fluid/models/vgg.py new file mode 100644 index 0000000000000000000000000000000000000000..53856c5f7acd3a4e1476ec57154a880bb6f984c9 --- /dev/null +++ b/benchmark/fluid/models/vgg.py @@ -0,0 +1,104 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +"""VGG16 benchmark in Fluid""" +from __future__ import print_function + +import sys +import time +import numpy as np +import paddle +import paddle.fluid as fluid +import paddle.fluid.core as core +import argparse +import functools + + +def vgg16_bn_drop(input): + def conv_block(input, num_filter, groups, dropouts): + return fluid.nets.img_conv_group( + input=input, + pool_size=2, + pool_stride=2, + conv_num_filter=[num_filter] * groups, + conv_filter_size=3, + conv_act='relu', + conv_with_batchnorm=True, + conv_batchnorm_drop_rate=dropouts, + pool_type='max') + + conv1 = conv_block(input, 64, 2, [0.3, 0]) + conv2 = conv_block(conv1, 128, 2, [0.4, 0]) + conv3 = conv_block(conv2, 256, 3, [0.4, 0.4, 0]) + conv4 = conv_block(conv3, 512, 3, [0.4, 0.4, 0]) + conv5 = conv_block(conv4, 512, 3, [0.4, 0.4, 0]) + + drop = fluid.layers.dropout(x=conv5, dropout_prob=0.5) + fc1 = fluid.layers.fc(input=drop, size=512, act=None) + bn = fluid.layers.batch_norm(input=fc1, act='relu') + drop2 = fluid.layers.dropout(x=bn, dropout_prob=0.5) + fc2 = fluid.layers.fc(input=drop2, size=512, act=None) + return fc2 + + +def get_model(args): + if args.data_set == "cifar10": + classdim = 10 + if args.data_format == 'NCHW': + data_shape = [3, 32, 32] + else: + data_shape = [32, 32, 3] + else: + classdim = 102 + if args.data_format == 'NCHW': + data_shape = [3, 224, 224] + else: + data_shape = [224, 224, 3] + + # Input data + images = fluid.layers.data(name='pixel', shape=data_shape, dtype='float32') + label = fluid.layers.data(name='label', shape=[1], dtype='int64') + + # Train program + net = vgg16_bn_drop(images) + predict = fluid.layers.fc(input=net, size=classdim, act='softmax') + cost = fluid.layers.cross_entropy(input=predict, label=label) + avg_cost = fluid.layers.mean(x=cost) + + # Evaluator + batch_size_tensor = fluid.layers.create_tensor(dtype='int64') + batch_acc = fluid.layers.accuracy( + input=predict, label=label, total=batch_size_tensor) + + # inference program + inference_program = fluid.default_main_program().clone() + with fluid.program_guard(inference_program): + inference_program = fluid.io.get_inference_program( + target_vars=[batch_acc, batch_size_tensor]) + + # Optimization + optimizer = fluid.optimizer.Adam(learning_rate=args.learning_rate) + + # data reader + train_reader = paddle.batch( + paddle.reader.shuffle( + paddle.dataset.cifar.train10() + if args.data_set == 'cifar10' else paddle.dataset.flowers.train(), + buf_size=5120), + batch_size=args.batch_size) + test_reader = paddle.batch( + paddle.dataset.cifar.test10() + if args.data_set == 'cifar10' else paddle.dataset.flowers.test(), + batch_size=args.batch_size) + + return avg_cost, inference_program, optimizer, train_reader, test_reader, batch_acc diff --git a/benchmark/fluid/resnet.py b/benchmark/fluid/resnet.py deleted file mode 100644 index 831fa2c019fc2868cd85b1ca7b2c8c76a2f1628c..0000000000000000000000000000000000000000 --- a/benchmark/fluid/resnet.py +++ /dev/null @@ -1,313 +0,0 @@ -# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -from __future__ import absolute_import -from __future__ import division -from __future__ import print_function - -import argparse -import functools -import numpy as np -import time - -import cProfile, pstats, StringIO - -import paddle -import paddle.fluid as fluid -import paddle.fluid.core as core -import paddle.fluid.profiler as profiler - - -def parse_args(): - parser = argparse.ArgumentParser('Convolution model benchmark.') - parser.add_argument( - '--model', - type=str, - choices=['resnet_imagenet', 'resnet_cifar10'], - default='resnet_imagenet', - help='The model architecture.') - parser.add_argument( - '--batch_size', type=int, default=32, help='The minibatch size.') - parser.add_argument( - '--use_fake_data', - action='store_true', - help='use real data or fake data') - parser.add_argument( - '--skip_batch_num', - type=int, - default=5, - help='The first num of minibatch num to skip, for better performance test' - ) - parser.add_argument( - '--iterations', type=int, default=80, help='The number of minibatches.') - parser.add_argument( - '--pass_num', type=int, default=100, help='The number of passes.') - parser.add_argument( - '--data_format', - type=str, - default='NCHW', - choices=['NCHW', 'NHWC'], - help='The data data_format, now only support NCHW.') - parser.add_argument( - '--device', - type=str, - default='GPU', - choices=['CPU', 'GPU'], - help='The device type.') - parser.add_argument( - '--data_set', - type=str, - default='flowers', - choices=['cifar10', 'flowers'], - help='Optional dataset for benchmark.') - parser.add_argument( - '--infer_only', action='store_true', help='If set, run forward only.') - parser.add_argument( - '--use_cprof', action='store_true', help='If set, use cProfile.') - parser.add_argument( - '--use_nvprof', - action='store_true', - help='If set, use nvprof for CUDA.') - parser.add_argument( - '--with_test', - action='store_true', - help='If set, test the testset during training.') - args = parser.parse_args() - return args - - -def conv_bn_layer(input, ch_out, filter_size, stride, padding, act='relu'): - conv1 = fluid.layers.conv2d( - input=input, - filter_size=filter_size, - num_filters=ch_out, - stride=stride, - padding=padding, - act=None, - bias_attr=False) - return fluid.layers.batch_norm(input=conv1, act=act) - - -def shortcut(input, ch_out, stride): - ch_in = input.shape[1] if args.data_format == 'NCHW' else input.shape[-1] - if ch_in != ch_out: - return conv_bn_layer(input, ch_out, 1, stride, 0, None) - else: - return input - - -def basicblock(input, ch_out, stride): - short = shortcut(input, ch_out, stride) - conv1 = conv_bn_layer(input, ch_out, 3, stride, 1) - conv2 = conv_bn_layer(conv1, ch_out, 3, 1, 1, act=None) - return fluid.layers.elementwise_add(x=short, y=conv2, act='relu') - - -def bottleneck(input, ch_out, stride): - short = shortcut(input, ch_out * 4, stride) - conv1 = conv_bn_layer(input, ch_out, 1, stride, 0) - conv2 = conv_bn_layer(conv1, ch_out, 3, 1, 1) - conv3 = conv_bn_layer(conv2, ch_out * 4, 1, 1, 0, act=None) - return fluid.layers.elementwise_add(x=short, y=conv3, act='relu') - - -def layer_warp(block_func, input, ch_out, count, stride): - res_out = block_func(input, ch_out, stride) - for i in range(1, count): - res_out = block_func(res_out, ch_out, 1) - return res_out - - -def resnet_imagenet(input, class_dim, depth=50, data_format='NCHW'): - - cfg = { - 18: ([2, 2, 2, 1], basicblock), - 34: ([3, 4, 6, 3], basicblock), - 50: ([3, 4, 6, 3], bottleneck), - 101: ([3, 4, 23, 3], bottleneck), - 152: ([3, 8, 36, 3], bottleneck) - } - stages, block_func = cfg[depth] - conv1 = conv_bn_layer(input, ch_out=64, filter_size=7, stride=2, padding=3) - pool1 = fluid.layers.pool2d( - input=conv1, pool_type='avg', pool_size=3, pool_stride=2) - res1 = layer_warp(block_func, pool1, 64, stages[0], 1) - res2 = layer_warp(block_func, res1, 128, stages[1], 2) - res3 = layer_warp(block_func, res2, 256, stages[2], 2) - res4 = layer_warp(block_func, res3, 512, stages[3], 2) - pool2 = fluid.layers.pool2d( - input=res4, - pool_size=7, - pool_type='avg', - pool_stride=1, - global_pooling=True) - out = fluid.layers.fc(input=pool2, size=class_dim, act='softmax') - return out - - -def resnet_cifar10(input, class_dim, depth=32, data_format='NCHW'): - assert (depth - 2) % 6 == 0 - - n = (depth - 2) // 6 - - conv1 = conv_bn_layer( - input=input, ch_out=16, filter_size=3, stride=1, padding=1) - res1 = layer_warp(basicblock, conv1, 16, n, 1) - res2 = layer_warp(basicblock, res1, 32, n, 2) - res3 = layer_warp(basicblock, res2, 64, n, 2) - pool = fluid.layers.pool2d( - input=res3, pool_size=8, pool_type='avg', pool_stride=1) - out = fluid.layers.fc(input=pool, size=class_dim, act='softmax') - return out - - -def run_benchmark(model, args): - if args.use_cprof: - pr = cProfile.Profile() - pr.enable() - - if args.data_set == "cifar10": - class_dim = 10 - if args.data_format == 'NCHW': - dshape = [3, 32, 32] - else: - dshape = [32, 32, 3] - else: - class_dim = 102 - if args.data_format == 'NCHW': - dshape = [3, 224, 224] - else: - dshape = [224, 224, 3] - - input = fluid.layers.data(name='data', shape=dshape, dtype='float32') - label = fluid.layers.data(name='label', shape=[1], dtype='int64') - predict = model(input, class_dim) - cost = fluid.layers.cross_entropy(input=predict, label=label) - avg_cost = fluid.layers.mean(x=cost) - - batch_size_tensor = fluid.layers.create_tensor(dtype='int64') - batch_acc = fluid.layers.accuracy( - input=predict, label=label, total=batch_size_tensor) - - inference_program = fluid.default_main_program().clone() - with fluid.program_guard(inference_program): - inference_program = fluid.io.get_inference_program( - target_vars=[batch_acc, batch_size_tensor]) - - optimizer = fluid.optimizer.Momentum(learning_rate=0.01, momentum=0.9) - opts = optimizer.minimize(avg_cost) - - fluid.memory_optimize(fluid.default_main_program()) - - train_reader = paddle.batch( - paddle.reader.shuffle( - paddle.dataset.cifar.train10() - if args.data_set == 'cifar10' else paddle.dataset.flowers.train(), - buf_size=5120), - batch_size=args.batch_size) - test_reader = paddle.batch( - paddle.dataset.cifar.test10() - if args.data_set == 'cifar10' else paddle.dataset.flowers.test(), - batch_size=args.batch_size) - - def test(exe): - test_accuracy = fluid.average.WeightedAverage() - for batch_id, data in enumerate(test_reader()): - img_data = np.array(map(lambda x: x[0].reshape(dshape), - data)).astype("float32") - y_data = np.array(map(lambda x: x[1], data)).astype("int64") - y_data = y_data.reshape([-1, 1]) - - acc, weight = exe.run(inference_program, - feed={"data": img_data, - "label": y_data}, - fetch_list=[batch_acc, batch_size_tensor]) - test_accuracy.add(value=acc, weight=weight) - - return test_accuracy.eval() - - place = core.CPUPlace() if args.device == 'CPU' else core.CUDAPlace(0) - exe = fluid.Executor(place) - exe.run(fluid.default_startup_program()) - accuracy = fluid.average.WeightedAverage() - if args.use_fake_data: - data = train_reader().next() - image = np.array(map(lambda x: x[0].reshape(dshape), data)).astype( - 'float32') - label = np.array(map(lambda x: x[1], data)).astype('int64') - label = label.reshape([-1, 1]) - - iters, num_samples, start_time = 0, 0, time.time() - for pass_id in range(args.pass_num): - accuracy.reset() - train_accs = [] - train_losses = [] - for batch_id, data in enumerate(train_reader()): - if iters == args.skip_batch_num: - start_time = time.time() - num_samples = 0 - if iters == args.iterations: - break - if not args.use_fake_data: - image = np.array(map(lambda x: x[0].reshape(dshape), - data)).astype('float32') - label = np.array(map(lambda x: x[1], data)).astype('int64') - label = label.reshape([-1, 1]) - loss, acc, weight = exe.run( - fluid.default_main_program(), - feed={'data': image, - 'label': label}, - fetch_list=[avg_cost, batch_acc, batch_size_tensor]) - iters += 1 - num_samples += len(label) - accuracy.add(value=acc, weight=weight) - train_losses.append(loss) - train_accs.append(acc) - print("Pass: %d, Iter: %d, Loss: %f, Accuracy: %f" % - (pass_id, iters, loss, acc)) - print("Pass: %d, Loss: %f, Train Accuray: %f\n" % - (pass_id, np.mean(train_losses), np.mean(train_accs))) - train_elapsed = time.time() - start_time - examples_per_sec = num_samples / train_elapsed - print('\nTotal examples: %d, total time: %.5f, %.5f examples/sed\n' % - (num_samples, train_elapsed, examples_per_sec)) - # evaluation - if args.with_test: - pass_test_acc = test(exe) - exit(0) - - -def print_arguments(args): - vars(args)['use_nvprof'] = (vars(args)['use_nvprof'] and - vars(args)['device'] == 'GPU') - print('----------- resnet Configuration Arguments -----------') - for arg, value in sorted(vars(args).iteritems()): - print('%s: %s' % (arg, value)) - print('------------------------------------------------') - - -if __name__ == '__main__': - model_map = { - 'resnet_imagenet': resnet_imagenet, - 'resnet_cifar10': resnet_cifar10 - } - args = parse_args() - print_arguments(args) - if args.data_format == 'NHWC': - raise ValueError('Only support NCHW data_format now.') - if args.use_nvprof and args.device == 'GPU': - with profiler.cuda_profiler("cuda_profiler.txt", 'csv') as nvprof: - run_benchmark(model_map[args.model], args) - else: - run_benchmark(model_map[args.model], args) diff --git a/benchmark/fluid/vgg.py b/benchmark/fluid/vgg.py deleted file mode 100644 index 53e34e0cbd15914791c305db6797f826ebfae34e..0000000000000000000000000000000000000000 --- a/benchmark/fluid/vgg.py +++ /dev/null @@ -1,224 +0,0 @@ -# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -"""VGG16 benchmark in Fluid""" -from __future__ import print_function - -import sys -import time -import numpy as np -import paddle -import paddle.fluid as fluid -import paddle.fluid.core as core -import argparse -import functools - -parser = argparse.ArgumentParser(description=__doc__) -parser.add_argument( - '--batch_size', type=int, default=128, help="Batch size for training.") -parser.add_argument( - '--skip_batch_num', - type=int, - default=5, - help='The first num of minibatch num to skip, for better performance test') -parser.add_argument( - '--iterations', type=int, default=80, help='The number of minibatches.') -parser.add_argument( - '--learning_rate', - type=float, - default=1e-3, - help="Learning rate for training.") -parser.add_argument('--pass_num', type=int, default=50, help="No. of passes.") -parser.add_argument( - '--device', - type=str, - default='GPU', - choices=['CPU', 'GPU'], - help="The device type.") -parser.add_argument( - '--data_format', - type=str, - default='NCHW', - choices=['NCHW', 'NHWC'], - help='The data order, now only support NCHW.') -parser.add_argument( - '--data_set', - type=str, - default='cifar10', - choices=['cifar10', 'flowers'], - help='Optional dataset for benchmark.') -parser.add_argument( - '--with_test', - action='store_true', - help='If set, test the testset during training.') -args = parser.parse_args() - - -def vgg16_bn_drop(input): - def conv_block(input, num_filter, groups, dropouts): - return fluid.nets.img_conv_group( - input=input, - pool_size=2, - pool_stride=2, - conv_num_filter=[num_filter] * groups, - conv_filter_size=3, - conv_act='relu', - conv_with_batchnorm=True, - conv_batchnorm_drop_rate=dropouts, - pool_type='max') - - conv1 = conv_block(input, 64, 2, [0.3, 0]) - conv2 = conv_block(conv1, 128, 2, [0.4, 0]) - conv3 = conv_block(conv2, 256, 3, [0.4, 0.4, 0]) - conv4 = conv_block(conv3, 512, 3, [0.4, 0.4, 0]) - conv5 = conv_block(conv4, 512, 3, [0.4, 0.4, 0]) - - drop = fluid.layers.dropout(x=conv5, dropout_prob=0.5) - fc1 = fluid.layers.fc(input=drop, size=512, act=None) - bn = fluid.layers.batch_norm(input=fc1, act='relu') - drop2 = fluid.layers.dropout(x=bn, dropout_prob=0.5) - fc2 = fluid.layers.fc(input=drop2, size=512, act=None) - return fc2 - - -def main(): - if args.data_set == "cifar10": - classdim = 10 - if args.data_format == 'NCHW': - data_shape = [3, 32, 32] - else: - data_shape = [32, 32, 3] - else: - classdim = 102 - if args.data_format == 'NCHW': - data_shape = [3, 224, 224] - else: - data_shape = [224, 224, 3] - - # Input data - images = fluid.layers.data(name='pixel', shape=data_shape, dtype='float32') - label = fluid.layers.data(name='label', shape=[1], dtype='int64') - - # Train program - net = vgg16_bn_drop(images) - predict = fluid.layers.fc(input=net, size=classdim, act='softmax') - cost = fluid.layers.cross_entropy(input=predict, label=label) - avg_cost = fluid.layers.mean(x=cost) - - # Evaluator - batch_size_tensor = fluid.layers.create_tensor(dtype='int64') - batch_acc = fluid.layers.accuracy( - input=predict, label=label, total=batch_size_tensor) - - # inference program - inference_program = fluid.default_main_program().clone() - with fluid.program_guard(inference_program): - inference_program = fluid.io.get_inference_program( - target_vars=[batch_acc, batch_size_tensor]) - - # Optimization - optimizer = fluid.optimizer.Adam(learning_rate=args.learning_rate) - opts = optimizer.minimize(avg_cost) - - fluid.memory_optimize(fluid.default_main_program()) - - # Initialize executor - place = core.CPUPlace() if args.device == 'CPU' else core.CUDAPlace(0) - exe = fluid.Executor(place) - - # Parameter initialization - exe.run(fluid.default_startup_program()) - - # data reader - train_reader = paddle.batch( - paddle.reader.shuffle( - paddle.dataset.cifar.train10() - if args.data_set == 'cifar10' else paddle.dataset.flowers.train(), - buf_size=5120), - batch_size=args.batch_size) - test_reader = paddle.batch( - paddle.dataset.cifar.test10() - if args.data_set == 'cifar10' else paddle.dataset.flowers.test(), - batch_size=args.batch_size) - - # test - def test(exe): - test_accuracy = fluid.average.WeightedAverage() - for batch_id, data in enumerate(test_reader()): - img_data = np.array(map(lambda x: x[0].reshape(data_shape), - data)).astype("float32") - y_data = np.array(map(lambda x: x[1], data)).astype("int64") - y_data = y_data.reshape([-1, 1]) - - acc, weight = exe.run(inference_program, - feed={"pixel": img_data, - "label": y_data}, - fetch_list=[batch_acc, batch_size_tensor]) - test_accuracy.add(value=acc, weight=weight) - return test_accuracy.eval() - - iters, num_samples, start_time = 0, 0, time.time() - accuracy = fluid.average.WeightedAverage() - for pass_id in range(args.pass_num): - accuracy.reset() - train_accs = [] - train_losses = [] - for batch_id, data in enumerate(train_reader()): - if iters == args.skip_batch_num: - start_time = time.time() - num_samples = 0 - if iters == args.iterations: - break - img_data = np.array(map(lambda x: x[0].reshape(data_shape), - data)).astype("float32") - y_data = np.array(map(lambda x: x[1], data)).astype("int64") - y_data = y_data.reshape([-1, 1]) - - loss, acc, weight = exe.run( - fluid.default_main_program(), - feed={"pixel": img_data, - "label": y_data}, - fetch_list=[avg_cost, batch_acc, batch_size_tensor]) - accuracy.add(value=acc, weight=weight) - iters += 1 - num_samples += len(y_data) - print( - "Pass = %d, Iter = %d, Loss = %f, Accuracy = %f" % - (pass_id, iters, loss, acc) - ) # The accuracy is the accumulation of batches, but not the current batch. - - # pass_train_acc = accuracy.eval() - train_losses.append(loss) - train_accs.append(acc) - print("Pass: %d, Loss: %f, Train Accuray: %f\n" % - (pass_id, np.mean(train_losses), np.mean(train_accs))) - train_elapsed = time.time() - start_time - examples_per_sec = num_samples / train_elapsed - print('\nTotal examples: %d, total time: %.5f, %.5f examples/sed\n' % - (num_samples, train_elapsed, examples_per_sec)) - # evaluation - if args.with_test: - pass_test_acc = test(exe) - exit(0) - - -def print_arguments(): - print('----------- vgg Configuration Arguments -----------') - for arg, value in sorted(vars(args).iteritems()): - print('%s: %s' % (arg, value)) - print('------------------------------------------------') - - -if __name__ == "__main__": - print_arguments() - main() diff --git a/cmake/configure.cmake b/cmake/configure.cmake index e490397cc0624c310949a4b571bd00cac6e8953b..682614742cf1bd3130c638020a2545e16226d4d6 100644 --- a/cmake/configure.cmake +++ b/cmake/configure.cmake @@ -41,6 +41,10 @@ if(USE_EIGEN_FOR_BLAS) add_definitions(-DPADDLE_USE_EIGEN_FOR_BLAS) endif(USE_EIGEN_FOR_BLAS) +if(EIGEN_USE_THREADS) + add_definitions(-DEIGEN_USE_THREADS) +endif(EIGEN_USE_THREADS) + if(NOT WITH_PROFILER) add_definitions(-DPADDLE_DISABLE_PROFILER) endif(NOT WITH_PROFILER) diff --git a/cmake/cpplint.cmake b/cmake/cpplint.cmake deleted file mode 100644 index 4823dc3e91390002aefac70f7931b4197db05789..0000000000000000000000000000000000000000 --- a/cmake/cpplint.cmake +++ /dev/null @@ -1,62 +0,0 @@ -# util to check C++ file style -# * it basically use google cpplint.py. -# * It provide "add_style_check_target" for cmake. -# Usage see add_style_check_target's document -# -# TODO(yuyang18): Add python style check. - -set(STYLE_FILTER) - -# diable unwanted filters - -# paddle do not indent public/potected/private in class -set(STYLE_FILTER "${STYLE_FILTER}-whitespace/indent,") -# paddle use mutable reference. BUT IT IS NOT RECOMMANDED -set(STYLE_FILTER "${STYLE_FILTER}-runtime/references,") -# paddle use relative path for include. -set(STYLE_FILTER "${STYLE_FILTER}-build/include,") -# paddle use , , etc. -set(STYLE_FILTER "${STYLE_FILTER}-build/c++11,") -# paddle use c style casting. BUT IT IS NOT RECOMMANDED -set(STYLE_FILTER "${STYLE_FILTER}-readability/casting") - - -# IGNORE SOME FILES -set(IGNORE_PATTERN - .*ImportanceSampler.* - .*cblas\\.h.* - .*\\.pb\\.txt - .*MultiDataProvider.* - .*pb.* - .*pybind.h) - -# add_style_check_target -# -# attach check code style step for target. -# -# first argument: target name to attach -# rest arguments: source list to check code style. -# -# NOTE: If WITH_STYLE_CHECK is OFF, then this macro just do nothing. -macro(add_style_check_target TARGET_NAME) - if(WITH_STYLE_CHECK) - set(SOURCES_LIST ${ARGN}) - list(REMOVE_DUPLICATES SOURCES_LIST) - foreach(filename ${SOURCES_LIST}) - foreach(pattern ${IGNORE_PATTERN}) - if(filename MATCHES ${pattern}) - list(REMOVE_ITEM SOURCES_LIST ${filename}) - endif() - endforeach() - endforeach() - - if(SOURCES_LIST) - add_custom_command(TARGET ${TARGET_NAME} POST_BUILD - COMMAND "${PYTHON_EXECUTABLE}" "${PADDLE_SOURCE_DIR}/paddle/scripts/cpplint.py" - "--filter=${STYLE_FILTER}" - ${SOURCES_LIST} - COMMENT "cpplint: Checking source code style" - WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}) - endif() - endif() -endmacro() diff --git a/cmake/cuda.cmake b/cmake/cuda.cmake index 7edc8637727e300539a46bc3941ace87c87903b8..b520c03a836a9e3f263ba050f151877ffe0d071d 100644 --- a/cmake/cuda.cmake +++ b/cmake/cuda.cmake @@ -172,6 +172,8 @@ set(CUDA_PROPAGATE_HOST_FLAGS OFF) list(APPEND CUDA_NVCC_FLAGS "-std=c++11") list(APPEND CUDA_NVCC_FLAGS "--use_fast_math") list(APPEND CUDA_NVCC_FLAGS "-Xcompiler -fPIC") +# in cuda9, suppress cuda warning on eigen +list(APPEND CUDA_NVCC_FLAGS "-w") # Set :expt-relaxed-constexpr to suppress Eigen warnings list(APPEND CUDA_NVCC_FLAGS "--expt-relaxed-constexpr") diff --git a/cmake/external/boost.cmake b/cmake/external/boost.cmake index 10662fc96704685f030a5d76c6857d4bc20a63d9..73713d93d5a52738651dda498fac5ea66e3589d2 100644 --- a/cmake/external/boost.cmake +++ b/cmake/external/boost.cmake @@ -23,8 +23,12 @@ set(BOOST_PROJECT "extern_boost") # checked that the devtools package of CentOS 6 installs boost 1.41.0. # So we use 1.41.0 here. set(BOOST_VER "1.41.0") -set(BOOST_TAR "boost_1_41_0") -set(BOOST_URL "http://paddlepaddledeps.bj.bcebos.com/${BOOST_TAR}.tar.gz") +if((NOT DEFINED BOOST_TAR) OR (NOT DEFINED BOOST_URL)) + message(STATUS "use pre defined download url") + set(BOOST_TAR "boost_1_41_0" CACHE STRING "" FORCE) + set(BOOST_URL "http://paddlepaddledeps.cdn.bcebos.com/${BOOST_TAR}.tar.gz" CACHE STRING "" FORCE) +endif() +MESSAGE(STATUS "BOOST_TAR: ${BOOST_TAR}, BOOST_URL: ${BOOST_URL}") set(BOOST_SOURCES_DIR ${THIRD_PARTY_PATH}/boost) set(BOOST_DOWNLOAD_DIR "${BOOST_SOURCES_DIR}/src/${BOOST_PROJECT}") set(BOOST_INCLUDE_DIR "${BOOST_DOWNLOAD_DIR}/${BOOST_TAR}" CACHE PATH "boost include directory." FORCE) diff --git a/cmake/external/grpc.cmake b/cmake/external/grpc.cmake index e90948782bb5e333bbdb47ef9d61c1e37e3cf9e4..4b6840578fd155027c895b6ed5d1f9133868f312 100644 --- a/cmake/external/grpc.cmake +++ b/cmake/external/grpc.cmake @@ -23,17 +23,28 @@ SET(GRPC_SOURCES_DIR ${THIRD_PARTY_PATH}/grpc) SET(GRPC_INSTALL_DIR ${THIRD_PARTY_PATH}/install/grpc) SET(GRPC_INCLUDE_DIR "${GRPC_INSTALL_DIR}/include/" CACHE PATH "grpc include directory." FORCE) SET(GRPC_CPP_PLUGIN "${GRPC_INSTALL_DIR}/bin/grpc_cpp_plugin" CACHE FILEPATH "GRPC_CPP_PLUGIN" FORCE) + +include(ProcessorCount) +ProcessorCount(NUM_OF_PROCESSOR) + IF(APPLE) - SET(BUILD_CMD make -n HAS_SYSTEM_PROTOBUF=false -s -j static grpc_cpp_plugin | sed "s/-Werror//g" | sh) + SET(BUILD_CMD make -n HAS_SYSTEM_PROTOBUF=false -s -j ${NUM_OF_PROCESSOR} static grpc_cpp_plugin | sed "s/-Werror//g" | sh) ELSE() - SET(BUILD_CMD make HAS_SYSTEM_PROTOBUF=false -s -j static grpc_cpp_plugin) + SET(BUILD_CMD make HAS_SYSTEM_PROTOBUF=false -s -j ${NUM_OF_PROCESSOR} static grpc_cpp_plugin) ENDIF() +# FIXME(wuyi): do not build zlib cares protobuf twice, find a way to build grpc with them ExternalProject_Add( extern_grpc DEPENDS protobuf zlib - GIT_REPOSITORY "https://github.com/grpc/grpc.git" - GIT_TAG "v1.10.x" + # NOTE(wuyi): + # this package is generated by following steps: + # 1. git clone -b v1.8.x https://github.com/grpc/grpc.git + # 2. submodule update --init + # 3. keep only zlib, cares, protobuf, boringssl under "third_party", + # checkout and clean other dirs under third_party + # 4. remove .git, and package the directory. + URL "http://paddlepaddledeps.bj.bcebos.com/grpc-v1.8.x.tar.gz" PREFIX ${GRPC_SOURCES_DIR} UPDATE_COMMAND "" CONFIGURE_COMMAND "" @@ -46,7 +57,6 @@ ExternalProject_Add( INSTALL_COMMAND make prefix=${GRPC_INSTALL_DIR} install ) -# FIXME(typhoonzero): hack to get static lib path, try a better way like merge them. ADD_LIBRARY(grpc++_unsecure STATIC IMPORTED GLOBAL) SET_PROPERTY(TARGET grpc++_unsecure PROPERTY IMPORTED_LOCATION "${GRPC_INSTALL_DIR}/lib/libgrpc++_unsecure.a") diff --git a/cmake/external/mkldnn.cmake b/cmake/external/mkldnn.cmake index 5759e5c489724332793bf103b7aacf7ffb068611..25c07850dda7b2f69c2207c37b9d2368632104ec 100644 --- a/cmake/external/mkldnn.cmake +++ b/cmake/external/mkldnn.cmake @@ -45,15 +45,15 @@ IF(${CBLAS_PROVIDER} STREQUAL "MKLML") ELSE() MESSAGE(FATAL_ERROR "Should enable MKLML when build MKLDNN") ENDIF() - -SET(MKLDNN_CFLAG "${CMAKE_C_FLAGS} -Wno-error=strict-overflow") -SET(MKLDNN_CXXFLAG "${CMAKE_CXX_FLAGS} -Wno-error=strict-overflow") +SET(MKLDNN_FLAG "-Wno-error=strict-overflow -Wno-error=unused-result -Wno-unused-result") +SET(MKLDNN_CFLAG "${CMAKE_C_FLAGS} ${MKLDNN_FLAG}") +SET(MKLDNN_CXXFLAG "${CMAKE_CXX_FLAGS} ${MKLDNN_FLAG}") ExternalProject_Add( ${MKLDNN_PROJECT} ${EXTERNAL_PROJECT_LOG_ARGS} DEPENDS ${MKLDNN_DEPENDS} GIT_REPOSITORY "https://github.com/01org/mkl-dnn.git" - GIT_TAG "v0.11" + GIT_TAG "db3424ad44901513c03a1ea31ccaacdf633fbe9f" PREFIX ${MKLDNN_SOURCES_DIR} UPDATE_COMMAND "" CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${MKLDNN_INSTALL_DIR} @@ -61,6 +61,7 @@ ExternalProject_Add( CMAKE_ARGS -DMKLROOT=${MKLML_ROOT} CMAKE_ARGS -DCMAKE_C_FLAGS=${MKLDNN_CFLAG} CMAKE_ARGS -DCMAKE_CXX_FLAGS=${MKLDNN_CXXFLAG} + CMAKE_ARGS -DWITH_TEST=OFF -DWITH_EXAMPLE=OFF CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${MKLDNN_INSTALL_DIR} -DMKLROOT:PATH=${MKLML_ROOT} ) diff --git a/cmake/external/mklml.cmake b/cmake/external/mklml.cmake index 796bcf28a1dfb308ccb7a2f839742c5c2fcf2002..82c424fb79d5596c31891bc395699bf9ff4e7e7e 100644 --- a/cmake/external/mklml.cmake +++ b/cmake/external/mklml.cmake @@ -27,8 +27,12 @@ ENDIF() INCLUDE(ExternalProject) SET(MKLML_PROJECT "extern_mklml") -SET(MKLML_VER "mklml_lnx_2018.0.1.20171007") -SET(MKLML_URL "http://paddlepaddledeps.bj.bcebos.com/${MKLML_VER}.tgz") +IF((NOT DEFINED MKLML_VER) OR (NOT DEFINED MKLML_URL)) + MESSAGE(STATUS "use pre defined download url") + SET(MKLML_VER "mklml_lnx_2018.0.3.20180406" CACHE STRING "" FORCE) + SET(MKLML_URL "http://paddlepaddledeps.cdn.bcebos.com/${MKLML_VER}.tgz" CACHE STRING "" FORCE) +ENDIF() +MESSAGE(STATUS "MKLML_VER: ${MKLML_VER}, MKLML_URL: ${MKLML_URL}") SET(MKLML_SOURCE_DIR "${THIRD_PARTY_PATH}/mklml") SET(MKLML_DOWNLOAD_DIR "${MKLML_SOURCE_DIR}/src/${MKLML_PROJECT}") SET(MKLML_DST_DIR "mklml") diff --git a/cmake/external/protobuf.cmake b/cmake/external/protobuf.cmake index 0fde4373a4be58e71ff1a305bd4991cc554d7a34..2665996432b1f6681927320a85d6835094abe4cd 100644 --- a/cmake/external/protobuf.cmake +++ b/cmake/external/protobuf.cmake @@ -212,6 +212,7 @@ FUNCTION(build_protobuf TARGET_NAME BUILD_FOR_HOST) ${CMAKE_COMMAND} ${PROTOBUF_SOURCES_DIR}/src/${TARGET_NAME}/cmake ${OPTIONAL_ARGS} -Dprotobuf_BUILD_TESTS=OFF + -DCMAKE_SKIP_RPATH=ON -DCMAKE_POSITION_INDEPENDENT_CODE=ON -DCMAKE_BUILD_TYPE=${THIRD_PARTY_BUILD_TYPE} -DCMAKE_INSTALL_PREFIX=${PROTOBUF_INSTALL_DIR} diff --git a/cmake/external/snappy.cmake b/cmake/external/snappy.cmake index 80282329c6ac65fbd1493a6838efca4bd9cadaad..af09ed4d5d6e21cc50aba5198a7e9ea56f49451a 100644 --- a/cmake/external/snappy.cmake +++ b/cmake/external/snappy.cmake @@ -47,8 +47,6 @@ ExternalProject_Add( -DCMAKE_INSTALL_LIBDIR:PATH=${SNAPPY_INSTALL_DIR}/lib -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON -DCMAKE_BUILD_TYPE:STRING=${THIRD_PARTY_BUILD_TYPE} - BUILD_COMMAND make -j8 - INSTALL_COMMAND make install ) add_library(snappy STATIC IMPORTED GLOBAL) diff --git a/cmake/external/snappystream.cmake b/cmake/external/snappystream.cmake index 20a96430823d07a07d4bb4602e7fc0cfe55c3bf2..6df636d7fa8757ade73892bda03a80ba9767472b 100644 --- a/cmake/external/snappystream.cmake +++ b/cmake/external/snappystream.cmake @@ -46,8 +46,6 @@ ExternalProject_Add( -DCMAKE_INSTALL_PREFIX:PATH=${SNAPPYSTREAM_INSTALL_DIR} -DCMAKE_INSTALL_LIBDIR:PATH=${SNAPPYSTREAM_INSTALL_DIR}/lib -DCMAKE_BUILD_TYPE:STRING=${THIRD_PARTY_BUILD_TYPE} - BUILD_COMMAND make -j8 - INSTALL_COMMAND make install DEPENDS snappy ) diff --git a/cmake/external/warpctc.cmake b/cmake/external/warpctc.cmake index a631ad14b18310598f7eea3a51839d61a9e456ff..07e1137e16afc1e4e9ab9640e1ccaea8008a0cd2 100644 --- a/cmake/external/warpctc.cmake +++ b/cmake/external/warpctc.cmake @@ -38,8 +38,7 @@ ENDIF() ExternalProject_Add( extern_warpctc ${EXTERNAL_PROJECT_LOG_ARGS} - GIT_REPOSITORY "https://github.com/gangliao/warp-ctc.git" - GIT_TAG b63a0644654a3e0ed624c85a1767bc8193aead09 + GIT_REPOSITORY "https://github.com/dzhwinter/warp-ctc.git" PREFIX ${WARPCTC_SOURCES_DIR} UPDATE_COMMAND "" CMAKE_ARGS -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER} diff --git a/cmake/generic.cmake b/cmake/generic.cmake index 1d3e2ade6d393c6e4c37eea0dc1064cdb18808a5..9ddd05b3d9404df29ca1bf634105314b7e6a5b70 100644 --- a/cmake/generic.cmake +++ b/cmake/generic.cmake @@ -206,8 +206,6 @@ function(cc_library TARGET_NAME) list(APPEND cc_library_HEADERS ${CMAKE_CURRENT_SOURCE_DIR}/${source}.h) endif() endforeach() - add_style_check_target(${TARGET_NAME} ${cc_library_SRCS} ${cc_library_HEADERS}) - else(cc_library_SRCS) if(cc_library_DEPS) merge_static_libs(${TARGET_NAME} ${cc_library_DEPS}) @@ -231,7 +229,7 @@ endfunction(cc_binary) function(cc_test TARGET_NAME) if(WITH_TESTING) - set(options "") + set(options SERIAL) set(oneValueArgs "") set(multiValueArgs SRCS DEPS ARGS) cmake_parse_arguments(cc_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) @@ -241,6 +239,9 @@ function(cc_test TARGET_NAME) add_test(NAME ${TARGET_NAME} COMMAND ${TARGET_NAME} ${cc_test_ARGS} WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}) + if (${cc_test_SERIAL}) + set_property(TEST ${TARGET_NAME} PROPERTY SERIAL 1) + endif() endif() endfunction(cc_test) @@ -268,7 +269,6 @@ function(nv_library TARGET_NAME) list(APPEND nv_library_HEADERS ${CMAKE_CURRENT_SOURCE_DIR}/${source}.h) endif() endforeach() - add_style_check_target(${TARGET_NAME} ${nv_library_SRCS} ${nv_library_HEADERS}) else(nv_library_SRCS) if (nv_library_DEPS) merge_static_libs(${TARGET_NAME} ${nv_library_DEPS}) @@ -295,7 +295,7 @@ endfunction(nv_binary) function(nv_test TARGET_NAME) if (WITH_GPU AND WITH_TESTING) - set(options "") + set(options SERIAL) set(oneValueArgs "") set(multiValueArgs SRCS DEPS) cmake_parse_arguments(nv_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) @@ -303,6 +303,9 @@ function(nv_test TARGET_NAME) target_link_libraries(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main memory gtest gflags glog) add_dependencies(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main memory gtest gflags glog) add_test(${TARGET_NAME} ${TARGET_NAME}) + if (nv_test_SERIAL) + set_property(TEST ${TARGET_NAME} PROPERTY SERIAL 1) + endif() endif() endfunction(nv_test) @@ -338,7 +341,6 @@ function(hip_library TARGET_NAME) list(APPEND hip_library_HEADERS ${CMAKE_CURRENT_SOURCE_DIR}/${source}.h) endif() endforeach() - add_style_check_target(${TARGET_NAME} ${hip_library_SRCS} ${hip_library_HEADERS}) else(hip_library_SRCS) if (hip_library_DEPS) merge_static_libs(${TARGET_NAME} ${hip_library_DEPS}) diff --git a/cmake/inference_lib.cmake b/cmake/inference_lib.cmake index cc758019827b9a5416a801e4da43d754d4492a73..236a55d332a91c88d1c5515e7aca4142930a079f 100644 --- a/cmake/inference_lib.cmake +++ b/cmake/inference_lib.cmake @@ -52,66 +52,91 @@ function(copy TARGET) endfunction() # third party -set(dst_dir "${CMAKE_INSTALL_PREFIX}/third_party/eigen3") +set(dst_dir "${FLUID_INSTALL_DIR}/third_party/eigen3") copy(eigen3_lib SRCS ${EIGEN_INCLUDE_DIR}/Eigen/Core ${EIGEN_INCLUDE_DIR}/Eigen/src ${EIGEN_INCLUDE_DIR}/unsupported/Eigen DSTS ${dst_dir}/Eigen ${dst_dir}/Eigen ${dst_dir}/unsupported + DEPS eigen3 ) -set(dst_dir "${CMAKE_INSTALL_PREFIX}/third_party/install/gflags") +set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/gflags") copy(gflags_lib SRCS ${GFLAGS_INCLUDE_DIR} ${GFLAGS_LIBRARIES} DSTS ${dst_dir} ${dst_dir}/lib + DEPS gflags ) -set(dst_dir "${CMAKE_INSTALL_PREFIX}/third_party/install/glog") +set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/glog") copy(glog_lib SRCS ${GLOG_INCLUDE_DIR} ${GLOG_LIBRARIES} DSTS ${dst_dir} ${dst_dir}/lib + DEPS glog +) + +set(dst_dir "${FLUID_INSTALL_DIR}/third_party/boost/") +copy(boost_lib + SRCS ${BOOST_INCLUDE_DIR}/boost + DSTS ${dst_dir} + DEPS boost ) if(NOT PROTOBUF_FOUND) - set(dst_dir "${CMAKE_INSTALL_PREFIX}/third_party/install/protobuf") + set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/protobuf") copy(protobuf_lib SRCS ${PROTOBUF_INCLUDE_DIR} ${PROTOBUF_LIBRARY} DSTS ${dst_dir} ${dst_dir}/lib + DEPS extern_protobuf ) endif() if(NOT CBLAS_FOUND) - set(dst_dir "${CMAKE_INSTALL_PREFIX}/third_party/install/openblas") + set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/openblas") copy(openblas_lib SRCS ${CBLAS_INSTALL_DIR}/lib ${CBLAS_INSTALL_DIR}/include DSTS ${dst_dir} ${dst_dir} + DEPS extern_openblas ) elseif (WITH_MKLML) - set(dst_dir "${CMAKE_INSTALL_PREFIX}/third_party/install/mklml") + set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/mklml") copy(mklml_lib SRCS ${MKLML_LIB} ${MKLML_IOMP_LIB} ${MKLML_INC_DIR} DSTS ${dst_dir}/lib ${dst_dir}/lib ${dst_dir} + DEPS mklml ) endif() +if(WITH_MKLDNN) + set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/mkldnn") + copy(mkldnn_lib + SRCS ${MKLDNN_INC_DIR} ${MKLDNN_SHARED_LIB} + DSTS ${dst_dir} ${dst_dir}/lib + DEPS mkldnn + ) +endif() + if(NOT MOBILE_INFERENCE AND NOT RPI) - set(dst_dir "${CMAKE_INSTALL_PREFIX}/third_party/install/snappy") + set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/snappy") copy(snappy_lib SRCS ${SNAPPY_INCLUDE_DIR} ${SNAPPY_LIBRARIES} - DSTS ${dst_dir} ${dst_dir}/lib) + DSTS ${dst_dir} ${dst_dir}/lib + DEPS snappy) - set(dst_dir "${CMAKE_INSTALL_PREFIX}/third_party/install/snappystream") + set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/snappystream") copy(snappystream_lib SRCS ${SNAPPYSTREAM_INCLUDE_DIR} ${SNAPPYSTREAM_LIBRARIES} - DSTS ${dst_dir} ${dst_dir}/lib) + DSTS ${dst_dir} ${dst_dir}/lib + DEPS snappystream) - set(dst_dir "${CMAKE_INSTALL_PREFIX}/third_party/install/zlib") + set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/zlib") copy(zlib_lib SRCS ${ZLIB_INCLUDE_DIR} ${ZLIB_LIBRARIES} - DSTS ${dst_dir} ${dst_dir}/lib) + DSTS ${dst_dir} ${dst_dir}/lib + DEPS zlib) endif() # paddle fluid module set(src_dir "${PADDLE_SOURCE_DIR}/paddle/fluid") -set(dst_dir "${CMAKE_INSTALL_PREFIX}/paddle/fluid") +set(dst_dir "${FLUID_INSTALL_DIR}/paddle/fluid") set(module "framework") copy(framework_lib DEPS framework_py_proto SRCS ${src_dir}/${module}/*.h ${src_dir}/${module}/details/*.h ${PADDLE_BINARY_DIR}/paddle/fluid/framework/framework.pb.h @@ -142,4 +167,31 @@ copy(string_lib DSTS ${dst_dir}/${module} ${dst_dir}/${module}/tinyformat ) +set(module "pybind") +copy(pybind_lib + SRCS ${CMAKE_CURRENT_BINARY_DIR}/paddle/fluid/${module}/pybind.h + DSTS ${dst_dir}/${module} +) + +# CMakeCache Info +copy(cmake_cache + SRCS ${CMAKE_CURRENT_BINARY_DIR}/CMakeCache.txt + DSTS ${FLUID_INSTALL_DIR}) + add_custom_target(inference_lib_dist DEPENDS ${inference_lib_dist_dep}) + +# paddle fluid version +execute_process( + COMMAND ${GIT_EXECUTABLE} log --pretty=format:%H -1 + WORKING_DIRECTORY ${PADDLE_SOURCE_DIR} + OUTPUT_VARIABLE PADDLE_GIT_COMMIT) +set(version_file ${FLUID_INSTALL_DIR}/version.txt) +file(WRITE ${version_file} + "GIT COMMIT ID: ${PADDLE_GIT_COMMIT}\n" + "WITH_MKL: ${WITH_MKL}\n" + "WITH_GPU: ${WITH_GPU}\n") +if(WITH_GPU) + file(APPEND ${version_file} + "CUDA version: ${CUDA_VERSION}\n" + "CUDNN version: v${CUDNN_MAJOR_VERSION}\n") +endif() diff --git a/doc/fluid/CMakeLists.txt b/doc/fluid/CMakeLists.txt index 8086507bb4b7e870ad6d6091945ed07a00b5100b..be92af3902769a65c77953c9f3cb1f3aa3738d79 100644 --- a/doc/fluid/CMakeLists.txt +++ b/doc/fluid/CMakeLists.txt @@ -15,6 +15,9 @@ set(SPHINX_CACHE_DIR_EN "${CMAKE_CURRENT_BINARY_DIR}/en/_doctrees") # HTML output director set(SPHINX_HTML_DIR_EN "${CMAKE_CURRENT_BINARY_DIR}/en/html") +set(IMPORT_PADDLE_STRING "") +set(IMPORT_PADDLEV2_STRING "") + configure_file( "${CMAKE_CURRENT_SOURCE_DIR}/../templates/conf.py.en.in" "${BINARY_BUILD_DIR_EN}/conf.py" @@ -27,8 +30,6 @@ sphinx_add_target(paddle_fluid_docs ${CMAKE_CURRENT_SOURCE_DIR} ${SPHINX_HTML_DIR_EN}) -add_dependencies(paddle_fluid_docs gen_proto_py paddle_python) - # configured documentation tools and intermediate build results set(BINARY_BUILD_DIR_CN "${CMAKE_CURRENT_BINARY_DIR}/cn/_build") @@ -50,6 +51,4 @@ sphinx_add_target(paddle_fluid_docs_cn ${CMAKE_CURRENT_SOURCE_DIR} ${SPHINX_HTML_DIR_CN}) -add_dependencies(paddle_fluid_docs_cn gen_proto_py paddle_python) - add_subdirectory(api) diff --git a/doc/fluid/api/CMakeLists.txt b/doc/fluid/api/CMakeLists.txt index 48b396f0786adad1ba6cd41f72497f853e54bc38..435d6e10fb02e9b2a8147f37da33e8848cc9b98a 100644 --- a/doc/fluid/api/CMakeLists.txt +++ b/doc/fluid/api/CMakeLists.txt @@ -7,6 +7,9 @@ set(SPHINX_CACHE_DIR_EN "${CMAKE_CURRENT_BINARY_DIR}/en/_doctrees") # HTML output director set(SPHINX_HTML_DIR_EN "${CMAKE_CURRENT_BINARY_DIR}/en/html") +set(IMPORT_PADDLE_STRING "import paddle") +set(IMPORT_PADDLEV2_STRING "import paddle.v2") + configure_file( "${CMAKE_CURRENT_SOURCE_DIR}/../../templates/conf.py.en.in" "${BINARY_BUILD_DIR_EN}/conf.py" diff --git a/doc/fluid/api/clip.rst b/doc/fluid/api/clip.rst new file mode 100644 index 0000000000000000000000000000000000000000..3ba096388fc87dda3096a9030fe5749e61112c06 --- /dev/null +++ b/doc/fluid/api/clip.rst @@ -0,0 +1,47 @@ +.. THIS FILE IS GENERATED BY `gen_doc.{py|sh}` + !DO NOT EDIT THIS FILE MANUALLY! + +==== +clip +==== + +ErrorClipByValue +---------------- + +.. autoclass:: paddle.fluid.clip.ErrorClipByValue + :members: + :noindex: + +GradientClipByValue +------------------- + +.. autoclass:: paddle.fluid.clip.GradientClipByValue + :members: + :noindex: + +GradientClipByNorm +------------------ + +.. autoclass:: paddle.fluid.clip.GradientClipByNorm + :members: + :noindex: + +GradientClipByGlobalNorm +------------------------ + +.. autoclass:: paddle.fluid.clip.GradientClipByGlobalNorm + :members: + :noindex: + +append_gradient_clip_ops +------------------------ + +.. autofunction:: paddle.fluid.clip.append_gradient_clip_ops + :noindex: + +error_clip_callback +------------------- + +.. autofunction:: paddle.fluid.clip.error_clip_callback + :noindex: + diff --git a/doc/fluid/api/data/data_reader.rst b/doc/fluid/api/data/data_reader.rst index d7c896a6270b488ca4449e5211d0d0879eda6ac5..1a35d0bbc8f9d751f49c7e1fc26feb1bcb3ae7f0 100644 --- a/doc/fluid/api/data/data_reader.rst +++ b/doc/fluid/api/data/data_reader.rst @@ -56,11 +56,11 @@ DataFeeder Reader ====== -.. automodule:: paddle.v2.reader +.. automodule:: paddle.reader :members: :noindex: -.. automodule:: paddle.v2.reader.creator +.. automodule:: paddle.reader.creator :members: :noindex: diff --git a/doc/fluid/api/evaluator.rst b/doc/fluid/api/evaluator.rst index f80b87c7d2704a144c02028c4925530a67d11289..c0dc9a0d1d9f2f70948dc3c905dca25d7dd43742 100644 --- a/doc/fluid/api/evaluator.rst +++ b/doc/fluid/api/evaluator.rst @@ -5,24 +5,3 @@ evaluator ========= -ChunkEvaluator --------------- - -.. autoclass:: paddle.fluid.evaluator.ChunkEvaluator - :members: - :noindex: - -EditDistance --------------- - -.. autoclass:: paddle.fluid.evaluator.EditDistance - :members: - :noindex: - -DetectionMAP --------------- - -.. autoclass:: paddle.fluid.evaluator.DetectionMAP - :members: - :noindex: - diff --git a/doc/fluid/api/executor.rst b/doc/fluid/api/executor.rst index a9cdf264e49691afc4b9425b7bfe54f8157ae6c2..f67a14c49f372e67d18ec8e6f87da01109376d22 100644 --- a/doc/fluid/api/executor.rst +++ b/doc/fluid/api/executor.rst @@ -30,3 +30,9 @@ switch_scope .. autofunction:: paddle.fluid.executor.switch_scope :noindex: +fetch_var +--------- + +.. autofunction:: paddle.fluid.executor.fetch_var + :noindex: + diff --git a/doc/fluid/api/gen_doc.sh b/doc/fluid/api/gen_doc.sh index ba7b7ba8e51399deb852b0a7c8ddd3128f521e85..0f0539355559446fd91f659d61b636db214b5a40 100755 --- a/doc/fluid/api/gen_doc.sh +++ b/doc/fluid/api/gen_doc.sh @@ -1,7 +1,7 @@ #!/bin/bash python gen_doc.py layers --submodules control_flow device io nn ops tensor > layers.rst -for module in io data_feeder evaluator executor initializer io nets optimizer param_attr profiler regularizer +for module in data_feeder clip metrics executor initializer io nets optimizer param_attr profiler regularizer do python gen_doc.py ${module} > ${module}.rst done diff --git a/doc/fluid/api/index_en.rst b/doc/fluid/api/index_en.rst index 06c686d9508635abd41571983e00be174e94743e..29cea9c68221b921939e8e09072d87f9f604e21b 100644 --- a/doc/fluid/api/index_en.rst +++ b/doc/fluid/api/index_en.rst @@ -9,8 +9,9 @@ Fluid data_feeder.rst executor.rst initializer.rst - evaluator.rst + metrics.rst nets.rst + clip.rst optimizer.rst param_attr.rst profiler.rst diff --git a/doc/fluid/api/initializer.rst b/doc/fluid/api/initializer.rst index 2f02c5de097945a45a3e053427104bd17bea1279..c49a98c744cdf907630ea8c74791ff2021d996e8 100644 --- a/doc/fluid/api/initializer.rst +++ b/doc/fluid/api/initializer.rst @@ -33,11 +33,16 @@ Xavier :members: :noindex: -MSRA ------- +force_init_on_cpu +----------------- -.. autoclass:: paddle.fluid.initializer.MSRA - :members: +.. autofunction:: paddle.fluid.initializer.force_init_on_cpu + :noindex: + +init_on_cpu +----------- + +.. autofunction:: paddle.fluid.initializer.init_on_cpu :noindex: ConstantInitializer @@ -68,9 +73,3 @@ XavierInitializer :members: :noindex: - -MSRAInitializer ------------------ -.. autoclass:: paddle.fluid.initializer.MSRAInitializer - :members: - :noindex: diff --git a/doc/fluid/api/layers.rst b/doc/fluid/api/layers.rst index 3790f09c84563fe541bd8d0bc08e23b19d4287ca..f53da4d194f8d2428b4121fa1bb31f3fc95a9f64 100644 --- a/doc/fluid/api/layers.rst +++ b/doc/fluid/api/layers.rst @@ -55,6 +55,13 @@ While :members: :noindex: +Switch +------ + +.. autoclass:: paddle.fluid.layers.Switch + :members: + :noindex: + lod_rank_table -------------- @@ -67,12 +74,6 @@ max_sequence_len .. autofunction:: paddle.fluid.layers.max_sequence_len :noindex: -topk ----- - -.. autofunction:: paddle.fluid.layers.topk - :noindex: - lod_tensor_to_array ------------------- @@ -109,6 +110,12 @@ less_than .. autofunction:: paddle.fluid.layers.less_than :noindex: +equal +----- + +.. autofunction:: paddle.fluid.layers.equal + :noindex: + array_read ---------- @@ -212,6 +219,42 @@ Send .. autofunction:: paddle.fluid.layers.Send :noindex: +open_recordio_file +------------------ + +.. autofunction:: paddle.fluid.layers.open_recordio_file + :noindex: + +open_files +---------- + +.. autofunction:: paddle.fluid.layers.open_files + :noindex: + +read_file +--------- + +.. autofunction:: paddle.fluid.layers.read_file + :noindex: + +shuffle +------- + +.. autofunction:: paddle.fluid.layers.shuffle + :noindex: + +batch +----- + +.. autofunction:: paddle.fluid.layers.batch + :noindex: + +double_buffer +------------- + +.. autofunction:: paddle.fluid.layers.double_buffer + :noindex: + nn == @@ -281,12 +324,6 @@ square_error_cost .. autofunction:: paddle.fluid.layers.square_error_cost :noindex: -accuracy --------- - -.. autofunction:: paddle.fluid.layers.accuracy - :noindex: - chunk_eval ---------- @@ -311,6 +348,18 @@ sequence_pool .. autofunction:: paddle.fluid.layers.sequence_pool :noindex: +sequence_softmax +---------------- + +.. autofunction:: paddle.fluid.layers.sequence_softmax + :noindex: + +softmax +------- + +.. autofunction:: paddle.fluid.layers.softmax + :noindex: + pool2d ------ @@ -323,12 +372,6 @@ batch_norm .. autofunction:: paddle.fluid.layers.batch_norm :noindex: -layer_norm ----------- - -.. autofunction:: paddle.fluid.layers.layer_norm - :noindex: - beam_search_decode ------------------ @@ -377,6 +420,12 @@ reduce_min .. autofunction:: paddle.fluid.layers.reduce_min :noindex: +reduce_prod +----------- + +.. autofunction:: paddle.fluid.layers.reduce_prod + :noindex: + sequence_first_step ------------------- @@ -425,6 +474,12 @@ matmul .. autofunction:: paddle.fluid.layers.matmul :noindex: +topk +---- + +.. autofunction:: paddle.fluid.layers.topk + :noindex: + warpctc ------- @@ -473,25 +528,34 @@ multiplex .. autofunction:: paddle.fluid.layers.multiplex :noindex: -label_smooth ------------- +layer_norm +---------- -.. autofunction:: paddle.fluid.layers.label_smooth +.. autofunction:: paddle.fluid.layers.layer_norm :noindex: -ops -=== +softmax_with_cross_entropy +-------------------------- -mean ----- +.. autofunction:: paddle.fluid.layers.softmax_with_cross_entropy + :noindex: -.. autofunction:: paddle.fluid.layers.mean +smooth_l1 +--------- + +.. autofunction:: paddle.fluid.layers.smooth_l1 :noindex: -mul ---- +one_hot +------- -.. autofunction:: paddle.fluid.layers.mul +.. autofunction:: paddle.fluid.layers.one_hot + :noindex: + +autoincreased_step_counter +-------------------------- + +.. autofunction:: paddle.fluid.layers.autoincreased_step_counter :noindex: reshape @@ -500,12 +564,52 @@ reshape .. autofunction:: paddle.fluid.layers.reshape :noindex: +lod_reset +--------- + +.. autofunction:: paddle.fluid.layers.lod_reset + :noindex: + +lrn +--- + +.. autofunction:: paddle.fluid.layers.lrn + :noindex: + pad --- .. autofunction:: paddle.fluid.layers.pad :noindex: +label_smooth +------------ + +.. autofunction:: paddle.fluid.layers.label_smooth + :noindex: + +roi_pool +-------- + +.. autofunction:: paddle.fluid.layers.roi_pool + :noindex: + + +ops +=== + +mean +---- + +.. autofunction:: paddle.fluid.layers.mean + :noindex: + +mul +--- + +.. autofunction:: paddle.fluid.layers.mul + :noindex: + scale ----- @@ -572,10 +676,70 @@ clip_by_norm .. autofunction:: paddle.fluid.layers.clip_by_norm :noindex: -sequence_softmax ----------------- +logical_and +----------- -.. autofunction:: paddle.fluid.layers.sequence_softmax +.. autofunction:: paddle.fluid.layers.logical_and + :noindex: + +logical_or +---------- + +.. autofunction:: paddle.fluid.layers.logical_or + :noindex: + +logical_xor +----------- + +.. autofunction:: paddle.fluid.layers.logical_xor + :noindex: + +logical_not +----------- + +.. autofunction:: paddle.fluid.layers.logical_not + :noindex: + +uniform_random +-------------- + +.. autofunction:: paddle.fluid.layers.uniform_random + :noindex: + +uniform_random_batch_size_like +------------------------------ + +.. autofunction:: paddle.fluid.layers.uniform_random_batch_size_like + :noindex: + +gaussian_random +--------------- + +.. autofunction:: paddle.fluid.layers.gaussian_random + :noindex: + +gaussian_random_batch_size_like +------------------------------- + +.. autofunction:: paddle.fluid.layers.gaussian_random_batch_size_like + :noindex: + +cumsum +------ + +.. autofunction:: paddle.fluid.layers.cumsum + :noindex: + +scatter +------- + +.. autofunction:: paddle.fluid.layers.scatter + :noindex: + +sum +--- + +.. autofunction:: paddle.fluid.layers.sum :noindex: sigmoid @@ -644,6 +808,18 @@ floor .. autofunction:: paddle.fluid.layers.floor :noindex: +cos +--- + +.. autofunction:: paddle.fluid.layers.cos + :noindex: + +sin +--- + +.. autofunction:: paddle.fluid.layers.sin + :noindex: + round ----- @@ -820,3 +996,16 @@ topk .. autofunction:: paddle.fluid.layers.topk :noindex: + +dice_loss +---- + +.. autofunction:: paddle.fluid.layers.dice_loss + :noindex: + +upsampling_bilinear2d +____ + +.. autofunction:: paddle.fluid.layers.upsampling_bilinear2d + :noindex: + diff --git a/doc/fluid/api/metrics.rst b/doc/fluid/api/metrics.rst new file mode 100644 index 0000000000000000000000000000000000000000..ddf07775d7ea293acd421b8549d03b277ff0611d --- /dev/null +++ b/doc/fluid/api/metrics.rst @@ -0,0 +1,56 @@ +.. THIS FILE IS GENERATED BY `gen_doc.{py|sh}` + !DO NOT EDIT THIS FILE MANUALLY! + +======= +metrics +======= + +MetricBase +---------- + +.. autoclass:: paddle.fluid.metrics.MetricBase + :members: + :noindex: + +CompositeMetric +--------------- + +.. autoclass:: paddle.fluid.metrics.CompositeMetric + :members: + :noindex: + +Accuracy +-------- + +.. autoclass:: paddle.fluid.metrics.Accuracy + :members: + :noindex: + +ChunkEvaluator +-------------- + +.. autoclass:: paddle.fluid.metrics.ChunkEvaluator + :members: + :noindex: + +EditDistance +------------ + +.. autoclass:: paddle.fluid.metrics.EditDistance + :members: + :noindex: + +DetectionMAP +------------ + +.. autoclass:: paddle.fluid.metrics.DetectionMAP + :members: + :noindex: + +Auc +--- + +.. autoclass:: paddle.fluid.metrics.Auc + :members: + :noindex: + diff --git a/doc/fluid/api/optimizer.rst b/doc/fluid/api/optimizer.rst index 7a92caf9b7139cf091eff834dbed3586b23ac3af..df2bd2eace52e78805433bea320f5de95d45bfc7 100644 --- a/doc/fluid/api/optimizer.rst +++ b/doc/fluid/api/optimizer.rst @@ -47,6 +47,28 @@ DecayedAdagrad :members: :noindex: +Adadelta +----------------- + +.. autoclass:: paddle.fluid.optimizer.Adadelta + :members: + :noindex: + +RMSProp +----------------- + +.. autoclass:: paddle.fluid.optimizer.RMSProp + :members: + :noindex: + +ModelAverage +----------------- + +.. autoclass:: paddle.fluid.optimizer.ModelAverage + :members: + :noindex: + + SGDOptimizer ------------ @@ -89,9 +111,25 @@ DecayedAdagradOptimizer :members: :noindex: -Adadelta --------------- + +AdadeltaOptimizer +----------------- .. autoclass:: paddle.fluid.optimizer.AdadeltaOptimizer :members: :noindex: + + +RMSPropOptimizer +----------------- + +.. autoclass:: paddle.fluid.optimizer.RMSPropOptimizer + :members: + :noindex: + +Optimizer +--------- + +.. autoclass:: paddle.fluid.optimizer.Optimizer + :members: + :noindex: diff --git a/doc/fluid/api/regularizer.rst b/doc/fluid/api/regularizer.rst index 837c67111c6e98e6a3859be802addc20a1c64f2b..756bc53baa0625aef48dad0c35e7ae57421a70d0 100644 --- a/doc/fluid/api/regularizer.rst +++ b/doc/fluid/api/regularizer.rst @@ -11,6 +11,13 @@ append_regularization_ops .. autofunction:: paddle.fluid.regularizer.append_regularization_ops :noindex: +WeightDecayRegularizer +---------------------- + +.. autoclass:: paddle.fluid.regularizer.WeightDecayRegularizer + :members: + :noindex: + L1Decay ------- @@ -26,15 +33,16 @@ L2Decay :noindex: L1DecayRegularizer ---------------------- +------------------ .. autoclass:: paddle.fluid.regularizer.L1DecayRegularizer :members: :noindex: L2DecayRegularizer ---------------------- +------------------ .. autoclass:: paddle.fluid.regularizer.L2DecayRegularizer :members: :noindex: + diff --git a/doc/fluid/build_and_install/paddleci.png b/doc/fluid/build_and_install/paddleci.png new file mode 120000 index 0000000000000000000000000000000000000000..c3eb1457acc77cab9360e654240d1e8f548035b4 --- /dev/null +++ b/doc/fluid/build_and_install/paddleci.png @@ -0,0 +1 @@ +../../v2/build_and_install/paddleci.png \ No newline at end of file diff --git a/doc/fluid/design/algorithm/parameter_average.md b/doc/fluid/design/algorithm/parameter_average.md index 340bc302d57429a9bf10a9d23ed9b0cdc7a2a568..28ad6495d97515442eb8af2050158829814acd33 100644 --- a/doc/fluid/design/algorithm/parameter_average.md +++ b/doc/fluid/design/algorithm/parameter_average.md @@ -1,7 +1,7 @@ # Averaging Parameter in PaddlePaddle ## Why Averaging -In a large scale machine learning setup where the size of the training data is huge, it could take us a large number of iterations over the training data before we can achieve the optimal values of parameters of our model. Looking at the problem setup, it is desirable if we can obtain the optimal values of parameters by going through the data in as few passes as we can. +In a large scale machine learning setup where the size of the training data is huge, it could take us a large number of iterations over the training data before we can achieve the optimal values of parameters of our model. Looking at the problem setup, it is desirable to obtain the optimal values of parameters by going through the data in as few passes as possible. Polyak and Juditsky (1992) showed that the test performance of simple average of parameters obtained by Stochastic Gradient Descent (SGD) is as good as that of parameter values that are obtained by training the model over and over again, over the training dataset. @@ -16,16 +16,16 @@ We propose averaging for any optimizer similar to how ASGD performs it, as menti ### How to perform Parameter Averaging in PaddlePaddle Parameter Averaging in PaddlePaddle works in the following way during training : -1. It will take in an instance of a normal optimizer as an input, e.g. RMSPropOptimizer +1. It will take in an instance of an optimizer as an input, e.g. RMSPropOptimizer 2. The optimizer itself is responsible for updating the parameters. 3. The ParameterAverageOptimizer maintains a separate copy of the parameters for itself: - 1. In concept, the values of this copy are the average of the values of the parameters in the most recent N batches. - 2. However, saving all the N instances of the parameters in memory is not feasible. + 1. In theory, the values of this copy are the average of the values of the parameters in the most recent N batches. + 2. However, saving all N instances of the parameters in memory is not feasible. 3. Therefore, an approximation algorithm is used. Hence, overall we have have two copies of the parameters: one for the optimizer itself, and one for the ParameterAverageOptimizer. The former should be used in back propagation, while the latter should be used during testing and should be saved. -During the testing/ saving the model phase, we perform the following steps: +During the testing/saving the model phase, we perform the following steps: 1. Perform the delayed operations. 2. Save current values of the parameters to a temporary variable. 3. Replace the values of the parameters with the averaged values. diff --git a/doc/fluid/design/concepts/functions_operators_layers.md b/doc/fluid/design/concepts/functions_operators_layers.md index 30bc488a18a28d349645d9d2502aae6691a69931..1f86b99e5197c3e0b85fd76fe704520ef21b06d3 100644 --- a/doc/fluid/design/concepts/functions_operators_layers.md +++ b/doc/fluid/design/concepts/functions_operators_layers.md @@ -40,7 +40,7 @@ template class FCOp : public OperatorBase { public: void Run(...) { - add(mul(Input("X"), Input("W")), Input("b"); + add(mul(Input("X"), Input("W")), Input("b")); } }; REGISTER_OP(FCOp, "fc"); diff --git a/doc/fluid/design/concepts/lod_tensor.md b/doc/fluid/design/concepts/lod_tensor.md index a88292e7888d0ebc64ee89ca315dfea38a12c71d..d606d7a790b4b0dc18553f2220d39cec8aa619ec 100644 --- a/doc/fluid/design/concepts/lod_tensor.md +++ b/doc/fluid/design/concepts/lod_tensor.md @@ -155,7 +155,7 @@ into offsets 3 2+3 4+5 1+9 2+10 3+12 ``` -so we know that the first sentence is from word 0 to word 3, and the second sentence from work 3 to word 5. +so we know that the first sentence is from word 0 to word 3, and the second sentence from word 3 to word 5. Similarly, the lengths in the top level LoD diff --git a/doc/fluid/design/data_type/float16.md b/doc/fluid/design/data_type/float16.md index 1ea95ed6b5d6792171569b6ff76d09be92fcb13e..844d2aafcf257b85057e1ac200ed3d5cf0be2ff0 100644 --- a/doc/fluid/design/data_type/float16.md +++ b/doc/fluid/design/data_type/float16.md @@ -3,7 +3,7 @@ ## Why float16 Half precision (float16) is a binary floating-point format that occupies 16 bits in memory. float16 is half the size of traditional 32-bit single precision format (float) and has lower precision and smaller range. -When high precision computation is not required, using float16 data type could potentially +When high precision computation is not required (which is usually the case at least in the deep learning inference stage), using float16 data type could potentially - reduce storage space, memory bandwidth, and power usages; - increase the chance of data fitting into a smaller cache of lower latency; @@ -12,7 +12,7 @@ When high precision computation is not required, using float16 data type could p ## Survey of current float16 support A brief survey of float16 support on different compilers, hardwares, and libraries can be found below. Interested readers can refer to [link1](https://github.com/PaddlePaddle/Paddle/issues/4853) and [link2](https://github.com/Xreki/Xreki.github.io/blob/master/multi_data_types_in_dl_framework/ppt/float16_and_quantized_type.md) for more info. -The goal of float16 is to serve as a key for the executor to find and run the correct version of compute method specialized for float16 in operator kernel. It should be compatible with various natively supported float16 implementations including `__half` for cuda, `float16_t` for ARM, and `Eigen::half` for Eigen to make writing customized float16 kernels easier. +The goal of float16 is to serve as a key for the executor to find and run the correct version of compute method specialized for float16 in operator kernels. It should be compatible with various natively supported float16 implementations including `__half` for cuda, `float16_t` for ARM, and `Eigen::half` for Eigen to make writing customized float16 kernels easier. ### Compiler - nvcc supports `__half` data type after CUDA 7.5. @@ -95,11 +95,89 @@ float half_to_float(float16 h); ``` which provides one-to-one conversion between float32 and float16. These twos functions will do different conversion routines based on the current hardware. CUDA/ARM instrinsics will be used when the corresonding hardware is available. If the hardware or compiler level does not support float32 to float16 conversion, software emulation will be performed to do the conversion. -## To do -After float16 class is available, some of the future items are below: +## float16 inference +In Fluid, a neural network is represented as a protobuf message called [ProgramDesc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/program.md), whose Python wrapper is a [Program](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#program). The basic structure of a program is some nested [blocks](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#block), where each block consists of some [variable](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#variable) definitions and a sequence of [operators](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#operator). An [executor](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/executor.md) will run a given program desc by executing the sequence of operators in the entrance block of the program one by one. -- Update pybind/tensor_py.h to bind c++ float16 with numpy float16. +### Operator level requirement +Each operator has many kernels for different data types, devices, and library types. The operator will select the appropriate kernel to run based on, among other things, the data type of the input variables. By default, every Fluid operator has a float data type kernel that takes float variables as input and generates float output. -- Modify `GetKernelType()` method in `framework/operator.h` to make it compatible with float16. +This means that if we provide float input to the first operator in a program, then each opeartor will use float kernel to compute float output and send it as input to the next operator to trigger the float kernel. Overall, the program will run in float mode and give us a final output of float data type. -- Create a type-casting operator that can convert the data type in tensor between float16 and other types. +The same principle applies if we want a program to run in float16 mode. We provide input variable of float16 data type to the first operator, and then one by one, each operator in the program will run the float16 kernel (provided that each operator in this program has float16 kernels registered) until we finally obtain a float16 output variable. + +So the preliminary requirement for float16 inference is to add float16 kernel to operators that are needed in a specific kind of program. For example, float16 inference on an image classification neural network like Vgg or Resnet, typically requires the following operators to have float16 kernels: convolution, pooling, multiplication, addition, batch norm, dropout, relu, and softmax. Please refer to [new_op_en](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/dev/new_op_en.md) for details of how to add new kernels to an operator. + +### Variable level requirement +Operators including convolution and multiplication (used in fully-connected layers) takes as input not only the variables generated by the preceding operators but also [parameter](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#parameter) variables, which contains the trained weights to apply to the input data. These weights are obtained in the Fluid training process and are by default of float data type. + +When these operators are running in float16 mode, the float16 kernel requires those parameter variables to contain weights of Fluid float16 data type. Thus, we need a convenient way to convert the original float weights to float16 weights. + +In Fluid, we use tensor to hold actual data for a variable on the c++ end. [Pybind](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/pybind/tensor_py.h) is used to bind c++ tensors of certain data type with numpy array of the correponding numpy data type on the Python end. Each common c++ built-in data type has a corresponding numpy data type of the same name. However, since there is no built-in float16 type in c++, we cannot directly bind numpy float16 data type with the Fluid float16 class. Since both Fluid float16 and numpy float16 use uint16 as the internal data storage type, we use c++ built-in type `uint16_t` and the corresponding numpy uint16 data type to bridge the gap via [Pybind](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/pybind/tensor_py.h). + +The following code demonstrates how to do the tensor conversion. +```Python +# var is the variable of float weights +# tensor is a numpy array of data copied from the tensor data in var +# fp16_var is the variable that will contain float16 weights converted from var +tensor = numpy.array(var.get_tensor()) +fp16_tensor = fp16_var.get_tensor() + +# After the original tensor data is converted to numpy float16 data type, +# view(numpy.uint16) is used so that the internal memory of the numpy array +# will be reinterpreted to be of uint16 data type, which is binded to +# Fluid float16 class via pybind with the help of uint16_t built-in c++ type +fp16_tensor.set(tensor.astype(numpy.float16).view(numpy.uint16), GPUPlace) +``` + +### Consistent API requirement +The basic inference in float16 mode requires users to feed input and obtain output both of float16 data type. However, in this way, the inference APIs are not consistent between float16 mode and float mode, and users may find it confusing and diffcult to use float16 inference since they need to do extra steps to provide float16 input data and convert float16 output data back to float. To have consistent API for different inference modes, we need to transpile the program desc in some way so that we can run float16 inference by feeding and fetching variables of float data type. + +This problem can be solved by introducing a type-casting operator which takes an input variable of certain data type, cast it to another specified data type, and put the casted data into the output variable. Insert cast operator where needed can make a program internally run in float16 mode. + +### float16 transpiler +Put all the above requirements in mind, we designed a float16 inference transpiler that can tranpile a float32 mode inference program desc to a float16 mode one. + +Given a float inference program and the corresponding variables of float32 weights in the [scope](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/scope.md), +this transpiler mainly does the following modifications: + +1. Insert cast operators at the beginning of the program so that the input float data will be converted to float16 data type before feeding to subsequent operators to invoke the float16 kernel. + +2. Insert cast operators at the end of the program so that the output float16 data will be converted back to float data type before users obtain the result. + +3. For each parameter variable of float weights, create in the scope a corresponding variable of float16 weights which are converted from the corresponding float weights and add this new float16 variable to the program. + +4. Update the operator information in the program so that each relevant operator use the newly created float16 variable instead of its float counterpart. + +Below is an example of usage: +```Python +# Get the float inference program +[float_inference_program, feed_target_names, + fetch_targets] = fluid.io.load_inference_model(save_dirname, exe) + +# Prepare the float input data +tensor_img = numpy.random.rand(1, 3, 32, 32).astype(numpy.float32) + +# Running inference_program in float mode +float_results = exe.run(float_inference_program, + feed={feed_target_names[0]: tensor_img}, + fetch_list=fetch_targets) + +# Use float16 transpiler to speedup +float16_inference_program = float_inference_program.clone() +t = fluid.InferenceTranspiler() +t.float16_transpile(float16_inference_program, GPUPlace) + +# Running +float16_results = exe.run(float16_inference_program, + feed={feed_target_names[0]: tensor_img}, + fetch_list=fetch_targets) +``` + +As we can see from the example above, users can simply use the `float16_transpile` method provided by the infernece transpiler class on an existing float inference program to run inference in float16 mode. + +### Speedup on GPU +Currently, Fluid inference in float16 mode is only supported on Nvidia GPU device. There is no motivation to support float16 inference on non-ARM CPUs because float16 is not natively supported there and float16 calculation will only be slower than its float counterpart. + +Nvidia started to support its native float16 data type (which has the same internal memory representation as Fluid float16 class) on CUDA 7.5. Moreover, float16 speedups on common computational intensive tasks including GEMM (general matrix-matrix multiplication) and convolution are supported since cublas 7.5 and cuDNN 5.0. + +Recently, the introduction of [tensor core](https://devblogs.nvidia.com/programming-tensor-cores-cuda-9/) in volta architecture GPUs and the support of tensor core calculation in CUDA 9.0 and cuDNN 7.0 make float16 truly superior to float in certain deep learning applications. Please refer to this [benchmark report](https://github.com/kexinzhao/Paddle_benchmark/blob/master/float16_benchmark.md) for more details. diff --git a/doc/fluid/design/dist_train/async_update.md b/doc/fluid/design/dist_train/async_update.md index 6a0835b761b69030ba30697e6e8863928efbf57f..248d2ec18dafdecac9184527638754b6ba4d85b8 100644 --- a/doc/fluid/design/dist_train/async_update.md +++ b/doc/fluid/design/dist_train/async_update.md @@ -4,34 +4,37 @@ For the typical synchronous distributed training, some significant steps are as follows: -1. A Trainer will compute the gradients and SEND them to the Parameter Server(PServer) nodes. -1. After the PServer node received gradients came from all the Trainers, It will aggregate the +1. A trainer process will compute the gradients and **send** them to the parameter server (PS) nodes. +1. After the PS node received gradients came from all the Trainers, It will aggregate the gradient variables for the same parameter into one gradient variable and then apply the aggregated gradient to the respective parameter, finally using an optimize algorithms(SGD, Monument...) to update the parameters. -1. The Trainer would wait for the PServers finished the optimize stage, and GET the parameters from PServer, +1. The Trainer would wait for the PS finished the optimize stage, and GET the parameters from PS, so all the Trainers would get the same parameters. -In the synchronously distributed training, there should be a `Barrier` to synchronise the -parameters after the optimizing stage. The performance of a distributed training job would -depend on the slowest node if there were hundreds or thousands of training nodes in a -Job, the performance of synchronously distributed training might be very poor because of -the slow node. So this design doc would introduce an approach to implement -*asynchronously* distributed training in PaddlePaddle Fluid. +In Synchronous Distributed Training, there is a **barrier** on each PS to wait until all trainers processes +have completed running current mini-batch. After that, all trainers can continue to run the next +mini-batch. So, we can find that the overall performance of Synchronous Distributed Training depends +on the slowest node. + +In Asynchronous Distributed Training, we don't need to wait for a global mini-bach, the optimizer on +the PS will run immediately when the gradient is uploaded to the PS from one trainer. This mode would +train such models that achieve scaling, better throughput. In this design doc, we will introduce how to +implement the Asynchronous Distributed Training base on PaddlePaddle Fluid. ## Design -As the figure above, we describe a global view of asynchronously update process and use +As the figure above, we describe a global view of the asynchronous update process and use the parameter `w1` as an example to introduce the steps: 1. For each gradient variables, they may distribute on different GPU card and aggregate them while they are all calculated. -1. Split the gradient variable into multiple blocks according to the number of PServer +1. Split the gradient variable into multiple blocks according to the number of PS instances and then send them. -1. PServer would run an `Optimize Block` using a specified optimize algorithm to update +1. PS would run an `Optimize Block` using a specified optimize algorithm to update the specified parameter. -1. The trainer will fetch latest parameter from PServer before running forward Op which depends +1. The trainer will fetch the latest parameter from PS before running forward Op which depends on the specified parameter. 1. Broadcast the received variable into multiple GPU cards and continue to run the next mini-batch. @@ -40,8 +43,8 @@ mini-batch. - For the multiple devices distributed training, we need to aggregate the gradient variables which placed on different devices firstly and then schedule a `SendVars` Operator to -send the gradient variables to the multiple PServer instances. -- Schedule `FetchVars` operator to fetch the latest parameter from PServer before running +send the gradient variables to the multiple PS instances. +- Schedule `FetchVars` operator to fetch the latest parameter from PS before running the forward ops. - There could be a large number of gradient variables to be sent, so we need to use another thread pool(IO Threadpool) whose a number of the schedulable threads is larger than the diff --git a/doc/fluid/design/dist_train/distributed_traing_review.md b/doc/fluid/design/dist_train/distributed_traing_review.md new file mode 100644 index 0000000000000000000000000000000000000000..c09b7c99159ace9b3df989f803ede20bc3585d92 --- /dev/null +++ b/doc/fluid/design/dist_train/distributed_traing_review.md @@ -0,0 +1,44 @@ +# Parallelism, Asynchronous, Synchronous, Codistillation + + +For valuable models, it’s worth using more hardware resources to reduce the training time and improve the final model quality. This doc discuss various solutions, their empirical results and some latest researches. + +# Model Parallelism +In some situations, larger and more complex models can improve the model quality. Sometimes, such models cannot fit in one device. Sometimes, parts of the model can be executed in parallel to improve speed. Model Parallelism address the issues by partitioning a single model and place the shards on several devices for execution. + +A common way of model parallelism is partition the logic of “gradient application” to parameter servers, while leaving the forward and backward computation at training servers. + +More flexible model parallelism is challenging. For example, multi-level-single-direction LSTM can be partitioned by layers, while such solution is not helpful for bi-directional LSTM. Different models can have quite different ways of partitioning and the benefits also depend on the underlying hardware. Framework needs to provide flexible APIs for user to define the customized partition scheme. For example, in TensorFlow, user can use tf.device() to specify the device placement. In MxNet, mx.AttrScope(ctx_group='dev1') does similar things. Recent research proposes to automatically find the optimal partition scheme with Reinforcement Learning, which is essentially solution space search algorithm that could cost a lot of extra hardware sources. + +# Data Parallelism +Data Parallelism runs the same model on multiple devices, each taking in a partition of the input batch. It’s more commonly used for a few reasons. It generally applies to common SGD mini-batch training. Compared with model parallelism, which requires users to carefully partition their model and tune for good performance, data parallelism usually involves no more than calling an extra API and speed up is more predictable. + +# Asynchronous Training +In asynchronous training, it usually involves a set of trainers and a set of parameter servers. The parameter servers collectively hold a single copy of shared parameters. While the trainers each holds a unique copy of model and trains the model independently. Each trainer pulls parameters from parameter servers and sends gradients to the parameter servers independently. Similarly the parameter servers applies the gradients to parameters as soon as the gradients are received and sends parameters whenever they are requested. + +In theory, asynchronous training is not safe and unstable. Each trainer is very likely using stale copy of parameters and parameters are also likely to apply stale gradients. However, in practice, especially for large-scale nonconvex optimization, it is effective [1]. Compared with synchronous solution, which will be discussed later, asynchronous distributed training is easier to implement and scales to a few dozen workers without losing much performance due to network communication or other overhead. Besides, asynchronous training can make progress even in case of random trainer failure in the cluster. + +Many production models, such as [3], are trained with distributed asynchronous solutions due to its scalability and effectiveness in practice. However, asynchronous training has its limitations. Usually, it’s not as stable as synchronous training. A warm-up phase is sometimes needed. Learning rate is usually smaller compared with synchronous training and decay is also often needed. Normally, asynchronous training doesn’t scale beyond 100 trainers. In other words, when putting more trainers beyond that, the model cannot converge faster. + +# Synchronous Training +Unlike asynchronous training, synchronous training requires step barriers. Parameter servers needs to wait for gradients from all trainers before they are applied to parameters and trainers will always pull the latest parameters. + +An obvious advantage of synchronous training is that the behavior is more clearly defined. Usually, it's more stable than asynchronous training. Learning rate can be set larger and for some vision tasks, the final accuracy can be slightly higher. (In my practical experience, for some models, it can actually be worse). + +Synchronous training usually faces scalability and performance issues, if not carefully implemented or deployed. In [2], native synchronous training can be 20%~40% slower than asynchronous training. A common trick to avoid slowness, discussed in [1] and [2], is to have backups. N+M replicas are scheduled while only the first N is needed for the training step the proceed. + +Similar to asynchronous training, the benefit of synchronous training diminishes quickly. Depending on the models, increasing the number of trainers (effectively batch size) beyond a point won’t delivers faster converge time or better final model quality. + +# Codistillation +Codistillation is a technique that tries to scale the training further. A few training instance (each training instance can be distributed) are performed during the same period. Each training instance has extra losses that comes from the prediction of other training instances. (likey teacher and student) The training process converges faster and usually converge to a better model quality. [4] + + +# Reference + +[1] Jeffrey Dean, Greg Corrado, Rajat Monga, Kai Chen, Matthieu Devin, Mark Mao, Andrew Senior, Paul Tucker, Ke Yang, Quoc V Le, et al. Large scale distributed deep networks. + +[2] Jianmin Chen, Rajat Monga, Samy Bengio, and Rafal Jozefowicz. Revisiting distributed synchronous SGD. + +[3] Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V Le, Mohammad Norouzi, Wolfgang Macherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, et al. Google’s neural machine translation system: Bridging the gap between human and machine translation. + +[4] LARGE SCALE DISTRIBUTED NEURAL NETWORK TRAINING THROUGH ONLINE DISTILLATION diff --git a/doc/fluid/design/motivation/api.md b/doc/fluid/design/motivation/api.md index e6a4638d9100d9b07c3ee6b92b530a17eae1c162..bc222564e3ec28e306ca0572b6a23104f6e9cbc5 100644 --- a/doc/fluid/design/motivation/api.md +++ b/doc/fluid/design/motivation/api.md @@ -77,8 +77,7 @@ print "The sematic-vector of testA: ", paddle.infer(fA, parameters, testA) ### Example 2. Sharing Parameters between "Models" -We use [GAN](https://github.com/PaddlePaddle/book/tree/develop/gan) in -this example. In the following example program, `d0` and `d1` +We use GAN in this example. In the following example program, `d0` and `d1` correspond to the two networks in the following figure: diff --git a/doc/fluid/design/motivation/refactorization.md b/doc/fluid/design/motivation/refactorization.md index 4e1d660cef6369f04db8e1e83360f6af25259f96..ad9d0f6d3f3ad9884f108826e8410871fffd51bf 100644 --- a/doc/fluid/design/motivation/refactorization.md +++ b/doc/fluid/design/motivation/refactorization.md @@ -125,12 +125,12 @@ Compile Time -> IR -> Runtime ## Operator/OpWithKernel/OpKernel -![class_diagram](http://api.paddlepaddle.org/graphviz?dot=https://gist.githubusercontent.com/reyoung/53df507f6749762675dff3e7ce53372f/raw/49caf1fb70820fb4a6c217634317c9306f361f36/op_op_with_kern_class_diagram.dot) +![class_diagram](https://raw.githubusercontent.com/PaddlePaddle/Paddle/develop/doc/fluid/images/op_op_with_kern_class_diagram.dot) --- ## Operator -![class_diagram](http://api.paddlepaddle.org/graphviz?dot=https://gist.githubusercontent.com/reyoung/53df507f6749762675dff3e7ce53372f/raw/dd598e8f1976f5759f58af5e5ef94738a6b2e661/op.dot) +![class_diagram](https://raw.githubusercontent.com/PaddlePaddle/Paddle/develop/doc/fluid/images/op.dot) * `Operator` is the fundamental building block of the user interface. * Operator stores input/output variable names and attributes. @@ -141,7 +141,7 @@ Compile Time -> IR -> Runtime ## OpWithKernel/Kernel -![class_diagram](http://api.paddlepaddle.org/graphviz?dot=https://gist.githubusercontent.com/reyoung/53df507f6749762675dff3e7ce53372f/raw/9d7f4eba185cf41c8e2fbfb40ae21890dbddcd39/op_with_kernel.dot) +![class_diagram](https://raw.githubusercontent.com/PaddlePaddle/Paddle/develop/doc/fluid/images/op_with_kernel.dot) * `OpWithKernel` inherits `Operator`. * `OpWithKernel` contains a Kernel map. diff --git a/doc/fluid/design/multi_devices/operator_kernel_type.md b/doc/fluid/design/multi_devices/operator_kernel_type.md index 8c1bc8f76a337006497e5ab5e5a710f9f49261b8..5e391bd62b4f4e123a9a6f35b7adf5726f205635 100644 --- a/doc/fluid/design/multi_devices/operator_kernel_type.md +++ b/doc/fluid/design/multi_devices/operator_kernel_type.md @@ -75,7 +75,7 @@ Different layout leads to different implementation of the operator kernel. There - The inference of Layout is at run-time, not at compile-time. -- Every operator has to implement different kernels for different layouts. Let's take MKLDNN as an example. If we want to implement an MKLDNN convolution operator, we have to implement all the kernels for different layouts, which are listed [here](http://01org.github.io/mkl-dnn/structmkldnn_1_1memory.html). And we will have a special macro to register kernels for MKLDNN operators. +- Every operator has to implement different kernels for different layouts. Let's take MKLDNN as an example. If we want to implement an MKLDNN convolution operator, we have to implement all the kernels for different layouts, which are listed [here](http://intel.github.io/mkl-dnn/structmkldnn_1_1memory.html). And we will have a special macro to register kernels for MKLDNN operators. `Layout` is also defined as a enum variable: diff --git a/doc/fluid/getstarted/Developer's_Guide_to_Paddle_Fluid.md b/doc/fluid/getstarted/Developer's_Guide_to_Paddle_Fluid.md new file mode 100644 index 0000000000000000000000000000000000000000..79df6c59578e2acf495a3453ab61f069c3f09a49 --- /dev/null +++ b/doc/fluid/getstarted/Developer's_Guide_to_Paddle_Fluid.md @@ -0,0 +1,1819 @@ + +# Paddle Fluid 开发者指南 + +--- + +### ==1==. 为什么需要 PaddlePaddle Fluid？ + +--- + +### 两个基础问题 + + + +1. 如何描述机器学习模型和优化过程？ + - 完备自洽，表达能力足以支持潜在出现的各种计算需求 +1. 如何充分利用资源高效计算？ + - 支持异步设备、多卡、分布式计算 + - 降低计算/计算优化的开发成本 + - …… + + + +--- + +### 如何描述模型和优化过程？ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
一组连续执行的layersvariable和operator构成的计算图 不再有模型的概念
2013 Caffe，Theano, Torch, PaddlePaddle
2015 TensorFlow, MxNet, Caffe2, ONNX, n-graph
2016 PyTorch, TensorFlow Eager Execution, **==PaddlePaddle Fluid==**
+ +--- + + +###

目标

+ + + +- 提高对各类机器学习任务的描述能力：能够描述潜在出现的任意机器学习模型。 +- 代码结构逻辑清晰，各模块充分解耦：内外部贡献者能够专注于自己所需的功能模块，基于框架进行再次开发。 +- 从设计上，留下技术优化的空间和潜力。 +- 代码解耦后降低多设备支持、计算优化等的开发成本。 +- 在统一的设计理念下，实现自动可伸缩，自动容错的分布式计算。 + + + +--- + +## ==2.== Design Overview + +--- + +# Fluid: 系统形态 + +- [编译器式的执行流程，区分编译时和运行时](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/motivation/fluid_compiler.md) +
+ +

+ +

+ +--- + +#### 让我们在Fluid程序实例中，区分编译时和运行时 + +--- +### Fluid 编译时 + + + +- ==**定义前向计算**== + + ```python + x = fluid.layers.data(name='x',shape=[13], dtype='float32') + y_predict = fluid.layers.fc(input=x, size=1, act=None) + y = fluid.layers.data(name='y', shape=[1], dtype='float32') + cost = fluid.layers.square_error_cost(input=y_predict, label=y) + avg_cost = fluid.layers.mean(x=cost) + ``` + +- ==**添加反向、正则、优化**== + ```python + learning_rate = 0.01 + sgd_optimizer = fluid.optimizer.SGD(learning_rate) + sgd_optimizer.minimize(avg_cost) + ``` + + +--- + +### `Program` vs. 计算图 + + + +- 在科学计算领域，计算图是一种描述计算的经典方式。下图展示了从前向计算图（蓝色）开始，通过添加反向（红色）和优化算法相关（绿色）操作，构建出整个计算图的过程： +- +

+ +

+ + +- Fluid ==使用`Program`而不是计算图==来描述模型和优化过程。`Program`由`Block`、`Operator`和`Variable`构成，相关概念会在后文详细展开。 +- 编译时 Fluid 接受前向计算（这里可以先简单的理解为是一段有序的计算流）`Program`，为这段前向计算按照：前向 -> 反向 -> 梯度 clip -> 正则 -> 优化 的顺序，添加相关 `Operator`和`Variable`到`Program`到完整的计算。 + + + +--- + +### Fluid 运行时 + + + +- ==**读入数据**== + + ```python + train_reader = paddle.batch( + paddle.reader.shuffle(paddle.dataset.uci_housing.train(), buf_size=500), + batch_size=20) + feeder = fluid.DataFeeder(place=place, feed_list=[x, y]) + ``` +- ==**定义执行程序的设备**== + ```python + place = fluid.CPUPlace() + feeder = fluid.DataFeeder(place=place,feed_list=[x, y]) + ``` + +- ==创建执行器（Executor），执行初始化 `Program`和训练`Program`== + + ```python + exe = fluid.Executor(place) + exe.run(fluid.default_startup_program()) + PASS_NUM = 100 + for pass_id in range(PASS_NUM): + for data in train_reader(): + avg_loss_value, = exe.run(fluid.default_main_program(), + feed=feeder.feed(data), + fetch_list=[avg_cost]) + print(avg_loss_value) + ``` + + +--- + +### 总结：框架做什么？用户做什么？ +
+ + + + + + + + + + + + + + + + +
构建训练执行训练
+用户：描述前向运算
框架：添加反向运算
框架：添加优化运算
框架：添加内存优化
框架：添加并行/多设备/分布式相关的计算单元 +		 +框架：创建Operator（计算）+ Variable（数据）
框架：创建`Block`
框架：内存管理/设备管理
框架：执行计算 +
+ + +--- + +###

总结：编译时

+ + +**用户编写一段Python程序，描述模型的前向计算** +1. 创建变量描述 `VarDesc` +1. 创建operators的描述 `OpDesc` +1. 创建operators的属性 +1. 推断变量的类型和形状，进行静态检查：`inferShape` +1. 规划变量的内存复用 +1. 创建反向计算 +1. 添加优化相关的Operators +1. （可选）添加多卡/多机相关的Operator，生成在多卡/多机上运行的程序 + + + +--- + +###

总结：运行时

+ + +**执行规划好的计算** +1. 创建`Executor` +1. 为将要执行的一段计算，在层级式的`Scope`空间中创建`Scope` +1. 创建`Block`，依次执行`Block` + +

+
+ Figure. 编译时运行时概览 +

+ + + +--- + +## ==3==. 用户如何描述计算？ +--- + +### Fluid：==像写程序一样==定义计算 + + +- 顺序执行 + ```python + x = fluid.layers.data(name='x',shape=[13], dtype='float32') + y_predict = fluid.layers.fc(input=x, size=1, act=None) + y = fluid.layers.data(name='y', shape=[1], dtype='float32') + cost = fluid.layers.square_error_cost(input=y_predict, label=y) + ``` + +- 条件分支: [swith](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/execution/switch.md)、[ifelse](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/execution/if_else_op.md) + + ```python + a = fluid.Var(10) + b = fluid.Var(0) + + switch = fluid.switch() + with switch.block(): + with switch.case(fluid.less_equal(a, 10)): + fluid.print("Case 1") + with switch.case(fluid.larger(a, 0)): + fluid.print("Case 2") + with switch.default(): + fluid.print("Case 3") + ``` + +>[A Lisp cond form may be compared to a continued if-then-else as found in many algebraic programming languages](https://www.cs.cmu.edu/Groups/AI/html/cltl/clm/node84.html). + + + +--- + +### Fluid: ==像写程序一样==定义计算 + + + +- 循环：[while](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/tests/book/test_machine_translation.py#L105) + + ```python + d0 = layers.data("d0", shape=[10], dtype='float32') + data_array = layers.array_write(x=d0, i=i) + array_len = layers.fill_constant(shape=[1],dtype='int64', value=3) + + cond = layers.less_than(x=i, y=array_len) + while_op = layers.While(cond=cond) + with while_op.block(): + d = layers.array_read(array=data_array, i=i) + i = layers.increment(x=i, in_place=True) + layers.array_write(result, i=i, array=d) + layers.less_than(x=i, y=array_len, cond=cond) + ``` + +- 完整实例请点查看 [->](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/tests/unittests/test_while_op.py#L36-L44) +- beam search [->]( https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/tests/book/test_machine_translation.py#L105) + + + +--- + +####

总结

+ + + +1. 用户层提供的描述语法具有完备性、自洽性，有能力支持对复杂计算过程描述 +1. 使用方式和核心概念可以类比编程语言，认知能够直接迁移 +1. 能够支持：定义问题，逐步求解 + + + +--- + +## ==3.== 核心概念 + +--- +### 编译时概念 ：==变量和计算的描述== + + + +- `VarDesc` + `TensorDesc` + `OpDesc` -> `BlockDesc` -> `ProgramDesc` + - https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/framework/framework.proto + +- 什么是 Fluid Program + + - 在Fluid中，一个神经网络任务（训练/预测）被描述为一段`Program` + - `Program`包含对`Variable`（数据）和 `Operator`（对数据的操作）的描述 + - `Variable` 和 `Operator` 被组织为多个可以嵌套的`Block`，构成一段完整的`Fluid Program` + + +>编译阶段最终，经过 Transpiler 的执行规划，变换处理，生成使用`protobuf`序列化后的`ProgramDesc`。可以发送给多卡或者网络中的其它计算节点执行 + + + +--- + +### 编译时概念 ：==**[Transpiler](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/motivation/fluid_compiler.md)**== + + +1. 接受一段`ProgramDesc`作为输入，生成一段新的`ProgramDesc` + + - *Memory optimization transpiler*：向原始`ProgramDesc` 中插入 `FreeMemoryOps`，在一次迭代优化结束前提前释放内存，使得能够维持较小的 memory footprint + + - *Distributed training transpiler*：将原始的`ProgramDesc`中转化为对应的分布式版本，生成两段新的`ProgramDesc`: + 1. trainer进程执行的`ProgramDesc` + 1. parameter server执行的`ProgramDesc` + +1. ==**WIP**==: 接受一段`ProgramDesc`，生成可直接被`gcc`, `nvcc`, `icc`等编译的代码，编译后得到可执行文件 + + + +--- +### Transplier + +

+ +

+ +--- + +### 打印 `ProgramDesc` + +

+ +

+ + + +- `default_startup_program`：创建可学习参数，对参数进行初始化 +- `default_main_program`：由用户定义的模型，包括了前向、反向、优化及所有必要的计算 + +- 打印可读的 `Program` + ```python + from paddle.v2.fluid import debuger + print debuger.pprint_program_codes(framework.default_main_program().desc) + ``` + + +--- +### 输出效果 + + + + + + + + + + + + + + +
variable in block 0variable in block 0
+ + +--- + +### 运行时概念 + + + +- 数据相关 + - `Tensor` / `LoDTensor` / `Variable` + - `Scope` + +- 计算相关 + - `Block` + - `Kernel`、`OpWithKernel`、`OpWithoutKernel` + + + + + + + + + + + + + + + + + + + + + + + + + + + +
protobuf messagesC++ class objects
Data[VarDesc](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/framework/framework.proto#L107) +[Variable](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/framework/variable.h#L24) +
Operation[OpDesc](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/framework/framework.proto#L35) +[Operator](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/framework/operator.h#L64) +
BlockBlockDesc +Block +
+ +- 执行相关 ：`Executor` + + + +--- +#### Tensor 和 LoD(Level-of-Detail) Tensor + + +- Tensor 是$n$-dimensional arry的推广，LoDTensor是在Tensor基础上附加了序列信息 +- Fluid中输入、输出，网络中的可学习参数全部统一使用LoDTensor（n-dimension array）表示 +- 一个mini-batch输入数据是一个LoDTensor + - 在Fluid中，RNN 处理变长序列无需padding，得益于 `LoDTensor`表示 + - 可以简单将 LoD 理解为：`std::vector>` + - 对非序列数据，LoD 信息为空 + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
TensorFlowPaddlePaddle
RNNSupport +Support +
recursive RNNSupport +Support +
padding zerosMust +No need +
blob data typeTensor +LODTensor +
+ + + +--- +#### LoD 信息实例 + + + +

+ +

+ +- 图(a)的LoD 信息 + ```cpp + [0, 5, 8, 10, 14] + ``` +- 图(b)的 LoD 信息 + ```cpp + [[0, 5, 8, 10, 14] /*level=1*/, [0, 2, 3, 5, 7, 8, 10, 13, 14] /*level=2*/] + ``` + + +--- +#### Tensor, Variable, Scope 之间的关系 + +

+ +

+ + +1. `Block` 是一个实现层的概念，不在应用层暴露给用户。目前用户无法自行创建并利用`Block`，用户能够感知的只有`Program`这个概念。 +1. 逻辑上，可以将 `Block` 类比为编程语言中的大括号：定义了一段作用域，其中运行一段代码 +1. `Executor`会为每一个`Block`创建一个`Scope`，`Block`是可嵌套的，因此`Scope`也是可嵌套的 + + + +--- +### Executor + + + + + + + + + + + + + + +
接口说明

+ +

输入
1. `ProgramDesc`
2. `Scope`
3.`block_id`

解释执行步骤
1. 创建所有 Variables
2. 逐一创建 Operator 并运行 +
+ +--- +### Operator/OpWithKernel/Kernel + + +

+ +

+ +- operator 无状态，Operator的核心是==Run==方法 +- 一个operator可以注册多个kernel +- operator 可以无 kernel：while_op 、ifelse op + + + +--- +#### Fluid Operator vs. PaddlePaddle layers + + + + + + + + + + + + + + + + + + +
LayerOperator

+ +

+ +

1. 内部维护状态
2. 包含forward和backward方法		1. 内部无状态
2. 只有Run方法
+ + + +--- + +### ==4.== 内存管理 + +--- +### 目标 + +- 为异构设备提供统一的内存分配、回收接口 +- 最小化管理内存所需的时间，最小化管理开销 +- 减少内存碎片 +- 将内存管理与计算（Operators/Kernels）完全剥离 +- 统一内存管理是内存优化的基础 + +--- + + + +### Memory 接口 + +- 内存管理模块向上层应用逻辑提供三个基础接口： + ```cpp + template + void* Alloc(Place place, size_t size); + + template + void Free(Place place, void* ptr); + + template + size_t Used(Place place); + + struct Usage : public boost::static_visitor { + size_t operator()(const platform::CPUPlace& cpu) const; + size_t operator()(const platform::CUDAPlace& gpu) const; + }; + ``` +- 模板参数 `Place` 指示内存分配发生的设备 +- 实现时，需特化支持的 `Place`， 提供以上三个接口的实现 + + + +--- +### 代码结构 + + + +内存管理模块可以理解为由以下两部分构成： + +1. SystemAllocator：实际从物理设备上分配、释放的内存的接口 +1. BuddyAllocator：内存管理算法 + + + +--- +### System Allocator + + + +- SystemAllocator 是实现物理内存分配、回收的基类 + - 不同设备上的内存分配和回收终将转化为标准接口调用 + - 为不同设备实现MemoryAllocator，继承自SystemAllocator + + ```cpp + class SystemAllocator { + public: + virtual ~SystemAllocator() {} + virtual void* Alloc(size_t& index, size_t size) = 0; + virtual void Free(void* p, size_t size, size_t index) = 0; + virtual bool UseGpu() const = 0; + }; + ``` + + +--- + +### CPU/GPU Allocator + + + +```cpp +class CPUAllocator : public SystemAllocator { + public: + virtual void* Alloc(size_t& index, size_t size); + virtual void Free(void* p, size_t size, size_t index); + virtual bool UseGpu() const; +}; + +#ifdef PADDLE_WITH_CUDA +class GPUAllocator : public SystemAllocator { + public: + virtual void* Alloc(size_t& index, size_t size); + virtual void Free(void* p, size_t size, size_t index); + virtual bool UseGpu() const; + private: + size_t gpu_alloc_size_ = 0; + size_t fallback_alloc_size_ = 0; +}; +#endif +``` +- CPUAllocator和GPUAllocator分别继承自SystemAllocator，分别调用相应的标准库函数实现物理内存的分配和释放。 +- 一旦大块、连续的物理内存分配之后，将通过内存管理算法实现内存的按块分配、回收、重用等。 + + + +--- +### CPU Allocator + + + +- CPU 内存的分配提供两种选项： + 1. non-pinned memory：可分页内存 + 2. pinned memory：页锁定内存 + - 分配过大的页锁定内存有可能因为系统可使用的分页内存减少，影响系统性能，默认CPU下分配的是可分页内存 + +- 通过gflags进行设置一次性分配内存的大小以及是否使用页锁定内存。 + + ```cpp + DEFINE_bool(use_pinned_memory, true, "If set, allocate cpu pinned memory."); + DEFINE_double(fraction_of_cpu_memory_to_use, 1, + "Default use 100% of CPU memory for PaddlePaddle," + "reserve the rest for page tables, etc"); + ``` + + + +--- +### GPU Allocator + + + +- 通过 cudaMalloc 分配GPU显存 +- GPUAllocator::Alloc 首先会计算指定GPU device上的可用显存 + - 如果可用显存小于请求分配大小，调用cudaMalloc进行分配 + - 如果可用显存不足，目前会报错退出。 +- 通过gflags控制GPU下一次性分配显存的大小： + + ```cpp + DEFINE_double(fraction_of_gpu_memory_to_use, 0.92, + "Default use 92% of GPU memory for PaddlePaddle," + "reserve the rest for page tables, etc"); + ``` + + + +--- +#### 内存管理算法: [Buddy Memory Allocation](https://en.wikipedia.org/wiki/Buddy_memory_allocation) + + + +- Memory Arena：一次性分配大块连续内存，之后会基于这块内存进行内存管理：动态分配、释放、重用内存块。 +- 伙伴内存分配： + - 将内存划分为 2 的幂次方个分区，使用 best-fit 方法来分配内存请求。 + - 当释放内存时，检查 buddy 块，查看相邻的内存块是否也已被释放。如果是，将内存块合并，以最小化内存碎片。 + - 分配的内存在物理内存的自然边界对齐，提高内存访问效率。 + - 算法的时间效率高，单使用 best-fit 方法的缘故，会产生一定的内存浪费 + + + +--- + +### Buddy Allocator + + + +- BuddyAllocator 是一个单例，每个设备（如： GPU/CPU(0)/GPU(1)） 拥有一个BuddyAllocator +- BuddyAllocator 内部拥有一个私有成员变量 SystemAllocator +- 当请求的内存超过BuddyAllocator管理的空余内存时，将会调用SystemAllocator去指定的设备上分配物理内存 + + + +--- +### 实例：CPU 下内存管理接口的实现 + + + +- 对上层应用，统一通过BuddyAllocator来实现内存的分配、释放以及用量查询 + ```cpp + template <> + void* Alloc(platform::CPUPlace place, size_t size) { + VLOG(10) << "Allocate " << size << " bytes on " << platform::Place(place); + void* p = GetCPUBuddyAllocator()->Alloc(size); + VLOG(10) << " pointer=" << p; + return p; + } + + template <> + void Free(platform::CPUPlace place, void* p) { + VLOG(10) << "Free pointer=" << p << " on " << platform::Place(place); + GetCPUBuddyAllocator()->Free(p); + } + + template <> + size_t Used(platform::CPUPlace place) { + return GetCPUBuddyAllocator()->Used(); + } + ``` + + +--- +### ==5.== 多设备支持 + +--- +### 多设备支持（一） + + + +- step 1：添加Place类型，由用户实现添加到框架 + - 可以将Place类型理解为一个整数加上一个枚举型，包括：设备号 + 设备类型 + +

+ +

+- DeviceContext + - 不同的Place会对应一个相应的DeviceContext，用于组织管理与设备相关的信息 + - 例如，GpuDeviceContext中会管理Cuda stream + - 目前实现中一些特殊的库也会对应有自己的DeviceContext：例如： + ```cpp + class MKLDNNDeviceContext : public CPUDeviceContext {……} + ``` + - 每种设备对应的DeviceContext需要管理的内容不尽相同，视具体需求来实现 + + + +--- + +### 多设备支持（二） + + + +- step 2: 增加KernelType，为相应的KernelType注册Kernel对象，由用户实现注册给框架 可以按照： + 1. Place 执行设备 + 1. DataType 执行数据类型 FP32/FP64/INT32/INT64 + 1. Memory layout： 运行时 Tensor 在内存中的排布格式 NCHW、 NHWC + 1. 使用的库 + + 来区分Kernel，为同一个operator注册多个 Kernel。 + + ```cpp + struct OpKernelType { + proto::DataType data_type_; + DataLayout data_layout_; + platform::Place place_; + LibraryType library_type_; + } + ``` + + + +--- + +### 多设备支持（三） + + + +step 3: 运行时的 KernelType 推断和Kernel切换，按需要修改Kernel推断和Kernel切换规则 +- Expected Kernel：期待调用的Kernel：由（1）`Place`和计算精度决定；或（2）用户在配置中显示指定使用的计算库，如`cudnn`、`mkldnn`等。 +- Actual Kernel：运行时从`Operator`的输入（`Variable`）可以推断出实际需要的`KernelType` +- 当Expected Kernel和Actual Kernel不一致的时候，框架会插入`data_transformer`或者`data_layerout_transform`等，保证Expected Kernel可以执行，包括： + - CPUPlace -> GPUPlace ：跨设备内存复制 + - NCHW -> nChw8c ：Layout转换 + - FP32 -> FP16 ：精度转换 _**尚未支持**_ + - …… +- 以上过程实现在OperatorWithKernel类的Run方法中 [->](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/framework/operator.cc#L497) + + + +--- +## ==6.== while_op + +--- +### while_op + + + +- 循环执行一段`Program`，直到条件operator判断循环条件不满足时终止循环 +- while_op 的特殊之处： + 1. while_op 没有 kernel + 1. while_op 拥有自己的`Block`，会形成一段嵌套的`Block` + 1. ==while_op 内部创建了一个 Executor，来循环执行`Block`== + +- while_op 输入输出 ： LoDTensorArray + ```cpp + namespace paddle { + namespace framework { + using LoDTensorArray = std::vector; + } + } + ``` + - 每一次循环，从原始输入中“切出”一个片段 + - LoDTensorArray 在Python端暴露，是Fluid支持的基础数据结构之一，用户可以直接创建并使用 + + + +--- +### while_op [Run](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/operators/while_op.cc#L42) 方法概览 + + + +```cpp + +void Run(const framework::Scope &scope, + const platform::Place &dev_place) const override { + PADDLE_ENFORCE_NOT_NULL(scope.FindVar(Input(kCondition))); + auto &cond = scope.FindVar(Input(kCondition))->Get(); + PADDLE_ENFORCE_EQ(cond.dims(), paddle::framework::make_ddim({1})); + + framework::Executor executor(dev_place); + auto *block = Attr(kStepBlock); + + auto *program = block->Program(); + auto step_scopes = + scope.FindVar(Output(kStepScopes))->GetMutable(); + + while (cond.data()[0]) { + auto ¤t_scope = scope.NewScope(); + step_scopes->push_back(¤t_scope); + executor.Run(*program, ¤t_scope, block->ID(), + false /*create_local_scope*/); + } +} + +``` + + + +--- +### while_op 的重要应用：Dynamic RNN + +--- + +### 什么是 `dynamicRNN` ? + + +
+ +1. 用户可以自定义在一个时间步之内的计算, 框架接受序列输入数据，在其上循环调用用户定义的单步计算 +1. 可学习参数在多个时间步之间共享 +1. `dynamicRNN` 由 `while_op` 实现 +1. 如果`dynamicRNN`中定义了`memory`，将会构成一个循环神经网络，否则其行为就等于在输入序列上循环调用预定义的单步计算 + + + +--- + +#### `dynamic RNN` 用户接口 + + +

+ +

+ +- `dynamicRNN` 中的重要元素 + 1. **step input**: `dynamicRNN` 每个时间步的输入 + 1. **step function**: 用户定义的单步计算 + 1. **memory**: 用于形成循环连接 + 1. **external/static memory**：单步计算的每一步都可以全部读取到的外部输入 + + + +--- + +#### dynamicRNN 中的 Memory + + + +`dynamicRNN`中`memory`的行为非常类似于 C++ 中的引用变量 + - `memory` “指向” 一个operator的输出变量，记作： A + - `memory` 可以被 LoDTensor 初始化（当LoD信息为空时，为非序列，否则为序列）,默认`memory`被初始化为零 + - `memory` 在 operator A 前向计算之后，进行前向计算 + - 当 `memory` 的前向计算会 "指向" A 的输出 LoDTensor + - `memory` 的输出可以是另一个 operator 的输入，于是形成了“循环”连接 + + + +--- + +### DynamicRNN 实现细节 + + + +- `while_op` 无法独立构成dynamicRNN，必须和一组相关的 operator 及数据结构配合 + - 依赖的 operators (这里仅列出最重要的，并非全部): + - `lod_rank_table` operator + - `lod_tensor_to_array` operator + - `array_to_lod_tensor` operator + - `shrink_memory` operator + - 依赖的数据结构 + - `TensorArray` + - `LoDRankTable` + +- 在Fluid中，RNN接受变长序列输入，无需填充，以上数据结构和相关的operator配合工作，实现了对变长输入以batch计算 + + + +--- + +### `dynamicRNN` 如何实现 batch 计算 ? + + + +- 问题： + - RNN 可以看作是一个展开的前向网络，前向网络的深度是最长序列的长度 + - 如果不对变长序列进行填充，将它们填充到一样长度，每个mini-batch输入将会不等长，每个样本展开长度不一致，导致前向和反向计算实现困难 + + + +---- +##### 实例 ：RNN encoder-decoder with attention + + + +- 以机器翻译的RNN encoder-decoder 模型（涉及了`dynamicRNN`的所有设计要素）为例，下图是 RNN encoder-decoder 的原始输入： +

+
Figure. RNN encoder-decoder 原始batch 输入数据 +

+ +- source word sequences 是encoder RNN的输出，是一个LoDTensor +- target word sequences 是look_uptable的输入，是一个LoDTensor +- 上图中一个矩形方块是CPU/GPU内存中一片连续的内存空间，表示一个dense vector + + + +--- + +### `dynamicRNN` 如何实现 batch 计算 ? + + + +1. 对一个mini batch中不等长样本进行排序，最长样本变成batch中的第一个，最短样本是batch中最后一个 + - `LoDTensor` -> `LoDRankTable` :heavy_plus_sign: `lod_rank_table operaator` + - 可以将`LoDRankTable`理解为对LoDTensor中的多个序列按照长度排序LoDRankTable 存储了排序之后的index + +2. 构建每个时间步的batch输入：随着时间步增加，每个时间步的batch输入可能会逐渐缩小 + - `TensorArray` :heavy_plus_sign: `lod_tensor_to_array` -> `LoDTensor` (without LoD) +3. 每个时间步输出写入一个输出 `LoDTensorArray` +3. `dynamicRNN`循环结束后, 按照`LoDRankTable`中记录的信息对输出`LoDTensorArray`重排序，还原会原始输入顺序 + - `TensorArray` :heavy_plus_sign: `array_to_lod_tensor` -> `LoDTensor` + + + +--- + +### 运行实例 + +

+ +

+ +--- +### 运行实例 + +

+ +

+ + + +- 执行到第5~7个batch时，batch size将会缩小 + + + +--- +### 运行实例 + +

+ +

+ + + +- 第5 ~ 7个batch时RNN的`memory`会发生什么？ + - `memory` 指向某个operator的输出Tensor，在该operator前向计算之后，“取回”其计算结果 + - 5 ~ 7时，遇到了序列的结束，==下一个时间步计算不再需要在已经结束的序列上展开== + - 在`dynamicRNN`中`shrink_memory` operator 用来缩小`memory`的batch输入 + + + +--- +### 运行实例：batch 1 ~ 2 + +

+
Figure. 第1、2个batch输入dynamicRNN的batch输入 +

+ +--- +### 运行实例：batch 3 ~ 4 + +

+
Figure. 第3、4个batch输入dynamicRNN的batch输入 +

+ +--- + +### 运行实例：batch 5 ~ 7 + +

+
Figure. 第5、6、7个batch输入dynamicRNN的batch输入 +

+ +--- +### ==7.== Fluid 代码结构 + +--- +### Fluid 代码结构 + + + + + + + + + + + + + + + +
代码结构模块结构
+

+ +

+		 +

+ +

+
+ +--- + +### ==8.== 文档总结 + +--- + + +- 设计概览 + - 重构概览 [->](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/refactorization.md) + - fluid [->](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/fluid.md) + - fluid_compiler [->](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/motivation/fluid_compiler.md) +- 核心概念 + - variable 描述 [->](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/var_desc.md) + - Tensor [->](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/tensor.md) + - LoDTensor [->](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/lod_tensor.md) + - TensorArray [->](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/tensor_array.md) + - Program [->](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/program.md) + - Block [->](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/block.md) + - Scope [->](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/scope.md) + +--- + +- 重要功能模块 + - backward [->](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/backward.md) + - 内存优化 [->](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/memory_optimization.md) + - evaluator [->](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/executor.md) + - python API [->](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/python_api.md) + - regularization [->](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/regularization.md) + +- 开发指南 + - 支持新设硬件设备库 [->](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/support_new_device.md) + - 添加新的Operator [->](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/dev/new_op_cn.md) + - 添加新的Kernel [->]( +https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/dev/new_op_kernel_en.md) + + + +--- + +### ==9.== 开发指南 + +--- + +#### 建议开发环境：使用 Docker 编译和测试 + + + +Docker编译PaddlePaddle源码: [->](http://www.paddlepaddle.org/docs/develop/documentation/fluid/zh/build_and_install/docker_install_cn.html) + +PaddlePaddle 在 Dockerhub 地址：[->]( + https://hub.docker.com/r/paddlepaddle/paddle/tags/) + +1. 获取PaddlePaddle的Docker镜像 + ```bash + docker pull paddlepaddle/paddle:latest-dev + ``` + +1. 启动 docker container + + ```bash + docker run -it -v $PWD/Paddle:/paddle paddlepaddle/paddle:latest-dev /bin/bash + ``` + +1. 进入docker container后，从源码编译，请参考文档 [->]( http://www.paddlepaddle.org/docs/develop/documentation/fluid/zh/build_and_install/build_from_source_cn.html) + + + +--- + +### 一些说明 + + + +1. PaddlePaddle的Docker镜像为了减小体积，默认没有安装vim，可以在容器中执行`apt-get install -y vim`来安装vim。 +1. 开发推荐使用tag为`latest-dev`的镜像，其中打包了所有编译依赖。`latest`及`lastest-gpu`是production镜像，主要用于运行PaddlePaddle程序。 +2. 在Docker中运行GPU程序，推荐使用nvidia-docker，[否则需要将CUDA库和设备挂载到Docker容器内](http://www.paddlepaddle.org/docs/develop/documentation/fluid/zh/build_and_install/docker_install_cn.html)。 + + + ```bash + nvidia-docker run -it -v $PWD/Paddle:/paddle paddlepaddle/paddle:latest-dev /bin/bash + ``` + + + + + +--- + +### [如何贡献](http://www.paddlepaddle.org/docs/develop/documentation/fluid/zh/dev/contribute_to_paddle_cn.html) + + + +- ==提交PullRequest前请务必阅读==： [->](http://www.paddlepaddle.org/docs/develop/documentation/fluid/zh/dev/contribute_to_paddle_cn.html) +- 代码要求 + 1. 代码注释遵守 Doxygen 的样式 + 1. 确保编译器选项 WITH_STYLE_CHECK 已打开，并且编译能通过代码样式检查 + 1. 所有代码必须具有单元测试，且能够通过所有单元测试 +- 使用 `pre-commit` 钩子提交Pull Request + 1. 帮助格式化源代码（C++，Python） + 1. 在提交前自动检查一些基本事宜：如每个文件只有一个 EOL，Git 中不要添加大文件等 + 1. 安装pre-commit，并在PaddlePaddle根目录运行： + ```bash + ➜ pip install pre-commit + ➜ pre-commit install + ``` + + +--- + +### 如何贡献 + + + +1. 开始开发之前请先建立issue。 + - 让其它同学知道某项工作已经有人在进行，以避免多人开发同一功能的情况。 +1. 提交PR必须关联相关的issue。做法请参考：[->](https://help.github.com/articles/closing-issues-using-keywords/) + - 目的：为了在提交的版本中留有记录描述这个PR是为了开发什么样的功能，为了解决什么样的问题。 + - 当PR被merge后，关联的issue会被自动关闭。 +1. PR review 中，reviewer的每条comment都必须回复。 + - 如修改完可直接回复：Done。 + - 目的：review comment 中可能会有（1）询问类型的问题；（2）可以在下一个PR修改的问题；（3）comment意见不合理等。需要明确回复，以便reviewer和其他人有历史可查，便于区分是否已经进行修改，或者准备下一个PR修改，或者意见不合理可以不用进行修改。 + + + +--- + +### ==10.== 添加新的 Operator + +--- + +### 概念简介 + + + +添加一个新的operator，会涉及实现以下C++类的派生类： + +1. `framework::OperatorBase`: Operator(简写，Op)基类。 +1. `framework::OpKernel`: Op计算函数的基类，称作Kernel。 +1. `framework::OperatorWithKernel`：继承自OperatorBase，Op有计算函数，称作有Kernel。 +1. `class OpProtoAndCheckerMaker`：描述该Op的输入、输出、属性、注释,主要用于Python API接口生成 + +依据是否包含kernel，可以将Op分为两种： +1. 包含Kernel的Op：继承自OperatorWithKernel，==绝大多数operator都属于这一类== +1. 不包含kernel的Op，继承自OperatorBase，只有少量Op属于这一类，例如while_op，ifelse_op + +这里主要介绍带Kernel的Op如何编写。 + + + +--- + +#### 添加新的Operator需要修改/添加哪些文件？ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
内容定义位置
+OpProtoMake定义 + +`.cc`文件，Backward Op不需要OpProtoMaker +
+Op定义 + +`.cc`文件 +
+Kernel实现 + +CPU、CUDA共享Kernel实现在`.h`文件中，否则，CPU 实现在`.cc`文件中，CUDA 实现在`.cu`文件中。 +
+注册Op + +Op注册实现在`.cc`文件；Kernel注册CPU实现在`.cc`文件中，CUDA实现在`.cu`文件中 +
+ +- 添加 Operator 之前请阅读：[Operator 命名规范](https://github.com/PaddlePaddle/Paddle/blob/63cca04cfd488a4dab6d6273fd04a8017ef45932/doc/fluid/dev/name_convention.md)及[Operator Markdown注释规范](https://github.com/PaddlePaddle/Paddle/blob/63cca04cfd488a4dab6d6273fd04a8017ef45932/doc/fluid/dev/op_markdown_format.md)。 +- 实现新的op都添加至目录[paddle/operators](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid/operators)下，文件命名以`*_op.h`（如有） 、 `*_op.cc` 、`*_op.cu`（如有）结尾。 +- 根据文件名自动构建op和Python端绑定，请务必遵守以上命名，否则需要进一步修改PyBind相关文件及CMakeLists.txt。 + + +--- + +###### 实现带Kernel的Operator step1: 定义ProtoMaker类 + + + +下面均以[clip_op](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/operators/clip_op.h)为例进行介绍 + +- clip_op计算公式：$Out = \min(\max(X, min), max)$ +- 首先定义`ProtoMaker`来描述该Op的输入、输出，并添加注释（*下面代码段的中注释进行了简化，实现时需按照规范添加注释*）： + + ```cpp + template + class ClipOpMaker : public framework::OpProtoAndCheckerMaker { + public: + ClipOpMaker(OpProto* proto, OpAttrChecker* op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X","(Tensor)The input of clip op."); + AddOutput("Out", "(Tensor),The output of clip op."); + AddAttr( + "min", "(float),Minimum value."); + AddAttr( + "max", "(float),Maximum value."); + AddComment(R"DOC( + …… + )DOC"); + } + }; + ``` + + + +--- + +###### 实现带Kernel的Operator step2: 定义Operator类 + + + +下面的代码段实现了`clip_op`的定义： + +```cpp +class ClipOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of ClipOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of ClipOp should not be null."); + auto x_dims = ctx->GetInputDim("X"); + auto max = ctx->Attrs().Get("max"); + auto min = ctx->Attrs().Get("min"); + PADDLE_ENFORCE_LT(min, max, "max should be greater than min."); + ctx->SetOutputDim("Out", x_dims); + ctx->ShareLoD("X", /*->*/ "Out"); + } +}; +``` + + +--- + +### Operator 类中需要完成的工作 + + + +1. clip_op 继承自`OperatorWithKernel`， + + ```cpp + using framework::OperatorWithKernel::OperatorWithKernel; + ``` + 表示使用基类`OperatorWithKernel`的构造函数。 + +1. 重写`InferShape`接口。 + - `InferShape` 为const函数，不能修改Op的成员变 + - `InferShape` 的参数为 `const framework::InferShapeContext &ctx`，从中可获取到输入输出以及属性 + - `InferShape` 会被调用两次，一次是编译时（创建op），一次是运行时（调用op的`Run`方法时），需要完成以下功能： + 1. 做检查， 尽早报错：检查输入数据维度、类型等是否合法 + 2. 设置输出Tensor的形状 + +通常`OpProtoMaker`和`Op`类的定义写在`.cc`文件中。 + + + +--- + +### 补充说明 + + + +1. `InferShape`目前支持两种实现方式，二者最后都会生成一个functor注册给OpInfo结构体。 + 1. 继承framework::InferShapeBase，实现为一个functor（参考 [mul_op](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/operators/mul_op.cc#L22)） + 2. override InferShape函数（参考 [clip_op](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/operators/clip_op.cc#L24)） + +1. 什么是`functor` ? + + - 类或结构体仅重载了`()`，一般是可被多个kernel复用的计算函数。 + + + + ```cpp + template + class CrossEntropyFunctor { + public: + void operator()(const platform::CPUDeviceContext& ctx, + framework::Tensor* out, + const framework::Tensor* prob, + const framework::Tensor* labels, const bool softLabel) { + …… + } + }; + ``` + + + - 在 clip_op 内也会看到将一段计算函数抽象为functor的使用法： [->](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/operators/clip_op.h#L27)。 + + + +--- + +###### 实现带Kernel的Operator step3: 定义OpKernel类 + + + +- `ClipKernel`继承自`framework::OpKernel`，带有下面两个模板参数: + 1. `typename DeviceContext`: 表示设备类型，不同设备共享同一个Kernel时，需添加该模板参数。不共享时，需要提供针对不同设备的特化实现。 + 1. `typename T` : 表示支持的数据类型，如`float`, `double`等 + +- 在`ClipKernel`类中重写`Compute`方法 + 1. `Compute`接受输入参数：`const framework::ExecutionContext& context` + - `ExecutionContext` 是从 `Scope`中将运行时Op的输入、输出`Variable`组织在一起，使得Op在调用`Compute`方法时，能够简单地通过名字拿到需要的输入输出`Variable` + - 与`InferShapeContext`相比，`ExecutionContext` 中增加了设备类型 + 1. 在`Compute`函数里实现`OpKernel`的具体计算逻辑 + + + +--- +#### ClipKernel 代码概览 + + + +```cpp +template +class ClipKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto max = context.Attr("max"); + auto min = context.Attr("min"); + auto* x = context.Input("X"); + auto* out = context.Output("Out"); + T* out_data = out->mutable_data(context.GetPlace()); + const T* x_data = x->data(); + int64_t numel = x->numel(); + Transform trans; + trans(context.template device_context(), x_data, + x_data + numel, out_data, ClipFunctor(min, max)); + } +}; +``` + +- 为了使`OpKernel`的计算过程书写更加简单，并且CPU、CUDA的代码可以复用， Fluid 使用 Eigen 作为基础的矩阵运算库 +- Fluid对Eigen unsupported Tensor提供了一些基本的封装，可以在`Compute`接口中直接调用 + - 关于在PaddlePaddle中如何使用Eigen库，请参考[使用文档](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/dev/use_eigen_cn.md)。 + + + +--- +###### 实现带Kernel的Operator step4: 实现反向Op + + + +- ==**反向Op没有`ProtoMaker`**==，除此之外定义与实现方式前向Op完全一致，不再赘述 +- 这里仅对反向Op的输入输出进行说明： + 1. 反向Op的输入 + - 前向Op的输出 + - 反向传播过程中传递给当前Op的梯度 + - 需要注意，Fluid中，不区分Cost Op和中间层Op，所有Op都必须正确处理接收到的梯度 + 2. 反向Op的输出 + - 对可学习参数的求导结果 + - 对所有输入的求导结果 + + + + +--- + +###### 实现带Kernel的Operator step5: 注册Op及Kernel + + + +至此Op和Op kernel都已经实现完毕，接下来，需要在`.cc`和`cu`文件中注册op和kernel + +1. 在`.cc`文件中注册前向、反向Op类，注册CPU Kernel。 + + + + ```cpp + namespace ops = paddle::operators; + REGISTER_OP(clip, ops::ClipOp, ops::ClipOpMaker, clip_grad, + ops::ClipOpGrad); + REGISTER_OP_CPU_KERNEL( + clip, ops::ClipKernel); + REGISTER_OP_CPU_KERNEL( + clip_grad, ops::ClipGradKernel); + ``` + + - 在上面的代码片段中： + + 1. `REGISTER_OP` ： 注册`ops::ClipOp`类，类型名为`clip`，该类的`ProtoMaker`为`ops::ClipOpMaker`，注册`ops::ClipOpGrad`，类型名为`clip_grad` + 1. `REGISTER_OP_WITHOUT_GRADIENT` ： 用于注册没有反向的Op，例如：优化算法相关的Op + 1. `REGISTER_OP_CPU_KERNEL` ：注册`ops::ClipKernel`类，并特化模板参数为`paddle::platform::CPUPlace`和`float`类型，同理，注册`ops::ClipGradKernel`类 + + +1. 按照同样方法，在`.cu`文件中注册GPU Kernel + - 如果CUDA Kernel的实现基于Eigen，需在 `.cu`的开始加上宏定义 `#define EIGEN_USE_GPU` + + + +--- + +##### 编译和Python端绑定 + + + +- 运行下面命令可以仅编译新添加的Op： + + ``` + make mul_op + ``` + - 需注意，运行单元测试需要编译整个工程 + +- 如果遵循前文的文件命名规则，构建过程中，会自动为新增的op添加Python端绑定，并链接到生成的lib库中 + + + +--- + +###### 实现带Kernel的Operator step6: 添加前向单测及梯度检测 + + + +- 新增Op的单元测试统一添加至：[python/paddle/v2/fluid/tests/unittests](https://github.com/PaddlePaddle/Paddle/tree/develop/python/paddle/fluid/tests/unittests)目录 +- 前向Operator单测 + + 1. Op单元测试继承自`OpTest`，各项具体的单元测试在`TestClipOp`里完成，所有单测case都以`TestXX`命名 + 1. 单元测试Operator，需要： + 1. 在`setUp`函数定义输入、输出，以及相关的属性参数 + 1. 生成随机的输入数据 + 1. 在Python脚本中实现与前向operator相同的计算逻辑，得到输出值，与operator前向计算的输出进行对比 + 1. 反向梯度检测流程测试框架已经实现，直接调用相应接口`check_grad`即可 + +- `clip_op` 单测代码请参考 [->](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/tests/unittests/test_clip_op.py)，这里不再展开 + + + +--- +#### 编译执行单测 + + + +- `python/paddle/v2/framework/tests` 目录下新增的 `test_*.py` 单元测试会被自动加入工程进行编译 + + - 运行单元测试测时需要编译整个工程，并且编译时需要打开`WITH_TESTING`, 即`cmake paddle_dir -DWITH_TESTING=ON` +- 编译成功后，执行下面的命令来运行单元测试： + + ```bash + make test ARGS="-R test_mul_op -V" + ``` + + 或者: + + ``` + ctest -R test_mul_op + ``` + + +--- + +### 添加Op的一些注意事项 + + + +- 为每个Op创建单独的`*_op.h`（如有）、`*_op.cc`和`*_op.cu`（如有）。不允许一个文件中包含多个Op，将会导致编译出错。 +- 注册Op时的类型名，需要和该Op的名字一样。不允许在`A_op.cc`里面，注册`REGISTER_OP(B, ...)`，会导致单元测试出错。 +- 如果Op没有实现CUDA Kernel，不要创建空的`*_op.cu`，会导致单元测试出错。 +- 如果多个Op依赖一些共用的函数，可以创建非`*_op.*`格式的文件来存放，如`gather.h`文件。 + + + +--- + +### ==10.== 使用相关问题 + +--- + +### 定义前向计算 + + + +- 当在python端执行时： + ```python + import paddle.v2.fluid as fluid + ``` + [`framework.py`](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/framework.py#L1040)定义了两个全局`Program`: + ```python + # program is a global instance. + _main_program_ = Program() + _startup_program_ = Program() + ``` + +- 前向定义的过程就是不断往`mian_program`中添加Op和Variable +- 如果需要执行一个新的`mian_program`时，可以调用调用： + ```python + def switch_main_program(program): + """ + Switch the main program to a new program. + This funtion returns the previous main program. + """ + …… + ``` + + +--- + +### 自定义参数的初始化 + + + +- 调用`fluid.ParamAttr(……)`接口，自定义参数的初始化 + + ```python + w_param_attrs = ParamAttr(name=None, + initializer=UniformInitializer(low=-1.0, high=1.0, seed=0), + learning_rate=1.0, + regularizer=L1Decay(1.0), + trainable=True, + clip=GradientClipByValue(-1.0, 1.0), + ) + y_predict = fluid.layers.fc(input=x, size=1, param_attr=w_param_attrs) + ``` + +- 补充问题：如何创建 `Variable` + ```python + cur_program = Program() + cur_block = cur_program.current_block() + new_var = cur_block.create_var(name="X", shape=[-1, 16, 16], dtype="float32") + ``` + + + +--- + +### 添加反向Op + + + +- 调用`fluid.backward.append_backward(X)`（`X`是一个Variable），来为一段前向`ProgramDesc`添加反Op + + ```python + data = fluid.layers.data(name="data", shape=(2,3,4)) + out = fluid.layers.fc(input=data,size=128,act=None) + loss = fluid.layers.reduce_sum(out) + fluid.backward.append_backward(loss=loss) + ``` + +- 添加优化相关的Op + ```python + sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.001) + sgd_optimizer.minimize(loss) + ``` + +- 可以随时调用`print(fluid.default_main_program())`来输出当前的`main_program` + +- 当构建完成整个`Program`后，调用下面的接口执行内存优化： + ```python + fluid.memory_optimize(fluid.default_main_program()) + ``` + - _注：内存优化目前仍在持续开发中，有可能不够稳定。_ + + + +--- + +### 总结：编译时执行流程 + + + +- 用户定义前向计算 +- 添加反向Op到`default_main_program` +- 添加 gradient clipping Op 到 +- 添加 regularization Op 到`default_main_program` +- 为指定的优化算法，添加相关的状态 variable of optimizer 到`default_startup_program` + - 状态相关 variable是指如学习率, 历史 momentum, 二阶momentum等 +- 添加初始化 variable 的Op 到 `default_startup_program` +- 为整个网络最后一个op，添加设置其接受到的梯度的Op到`default_main_program` +- 进行内存优化规划 + + + +--- + +### Feed 数据 (一)：通过 feed 字典 + + + +- 执行executor的run方法时，指定feed字典，feed op 会将指定的数据放到`x`和`y`两个Variable中 + ```python + y_data = np.random.randint(0, 8, [1]).astype("int32") + y_tensor = core.Tensor() + y_tensor.set(y_data, place) + + x_data = np.random.uniform(0.1, 1, [11, 8]).astype("float32") + x_tensor = core.Tensor() + x_tensor.set(x_data, place) + …… + cost = exe.run( + fluid.default_main_program(), + feed={'x': x_tensor, + 'y': y_tensor}, + fetchlist=[avg_cost]) + ``` + +- 这种方法较为底层，一般用于单测中 + + + +--- + +### Feed 数据 (二)：使用 DataFeeder接口 + + + +- 编写一个data_reader函数，data_reader是一个Python generator + + ```python + def demo_reader(): + def random_generator(): + yield np.random.uniform(0.1, 1, [4]), np.random.randint(0, 1, [1]) + return random_generator + ``` +- 在训练任务中使用 DataFeeder 接口 + ```python + cost = exe.run( + fluid.default_main_program(), + feed={'x': x_tensor, + 'y': y_tensor}, + fetchlist=[avg_cost]) + + train_reader = paddle.batch( + paddle.reader.shuffle(demo_reader(), buf_size=500), batch_size=4) + feeder = fluid.DataFeeder(place=place, feed_list=[x, y]) + for data in train_reader(): + cost = exe.run( + fluid.default_main_program(), + feed=feeder.feed(data), + fetch_list=[cost]) + ``` + + + +--- + +### 常见问题 + + + +- 如何使用 evaluator ? [->](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/tests/book/test_label_semantic_roles.py#L168) + + ```python + accuracy = fluid.evaluator.Accuracy(input=predict, label=label) + for pass_id in range(PASS_NUM): + accuracy.reset() + for data in train_reader(): + loss, acc = exe.run(fluid.default_main_program(), + feed=feeder.feed(data), + fetch_list=[avg_cost] + accuracy.metrics) + pass_acc = accuracy.eval(exe) + # acc 当前一个batch 的 accuracy + # pass_acc 当前batch 的 accuracy + pass_total_acc = accuracy.eval(exe) # 整个pass的accuracy + ``` + +- 如何在训练中测试？[->](https://github.com/dzhwinter/benchmark/blob/master/fluid/vgg16.py#L144) +- 如何保存训练好的模型？[->](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/tests/book/test_recognize_digits.py#L143) +- 如何加载训练好的模型进行预测？[->](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/tests/book/test_recognize_digits.py#L154) +- 如何在同一个训练任务中定义多个Program，并交替运行？ [->](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/tests/demo/fc_gan.py) +- 如何profile？Fluid 实现了profile 工具，可以直接调用。请参考示例 [->](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/tests/unittests/test_profiler.py) + + + + +--- diff --git a/doc/fluid/getstarted/index_cn.rst b/doc/fluid/getstarted/index_cn.rst index 75af7354be93a6eeabfa9ccf86903505402a7ca6..3daea71d0933a2774227ff2b5e744392ca6b1765 100644 --- a/doc/fluid/getstarted/index_cn.rst +++ b/doc/fluid/getstarted/index_cn.rst @@ -17,3 +17,4 @@ :maxdepth: 1 concepts/use_concepts_cn.rst + developer's_guide_to_paddle_fluid.md diff --git a/doc/fluid/getstarted/index_en.rst b/doc/fluid/getstarted/index_en.rst index 75a43f4af87c34830ec940068196e6ca72640501..fb20bb4f245281c3acf67c417979dc63c144fef3 100644 --- a/doc/fluid/getstarted/index_en.rst +++ b/doc/fluid/getstarted/index_en.rst @@ -16,3 +16,4 @@ Here is an example of linear regression. It introduces workflow of PaddlePaddle, :maxdepth: 1 concepts/index_en.rst + developer's_guide_to_paddle_fluid.md diff --git a/doc/fluid/getstarted/quickstart_cn.rst b/doc/fluid/getstarted/quickstart_cn.rst index 135beb75d0330f39d062753aa2aa83a077f36bb1..6a964d4f8561f30aa10936d2399698c51583442c 100644 --- a/doc/fluid/getstarted/quickstart_cn.rst +++ b/doc/fluid/getstarted/quickstart_cn.rst @@ -11,7 +11,7 @@ PaddlePaddle支持使用pip快速安装，目前支持CentOS 6以上, Ubuntu 14. pip install paddlepaddle -如果需要安装支持GPU的版本（cuda7.5_cudnn5_avx_openblas），需要执行： +如果需要安装支持GPU的版本（cuda8.0_cudnn5_avx_openblas），需要执行： .. code-block:: bash @@ -28,18 +28,18 @@ PaddlePaddle支持使用pip快速安装，目前支持CentOS 6以上, Ubuntu 14. import paddle.dataset.uci_housing as uci_housing import paddle.fluid as fluid - + with fluid.scope_guard(fluid.core.Scope()): # initialize executor with cpu exe = fluid.Executor(place=fluid.CPUPlace()) - # load inference model + # load inference model [inference_program, feed_target_names,fetch_targets] = \ fluid.io.load_inference_model(uci_housing.fluid_model(), exe) # run inference - result = exe.run(inference_program, - feed={feed_target_names[0]: uci_housing.predict_reader()}, + result = exe.run(inference_program, + feed={feed_target_names[0]: uci_housing.predict_reader()}, fetch_list=fetch_targets) - # print predicted price is $12,273.97 + # print predicted price is $12,273.97 print 'Predicted price: ${:,.2f}'.format(result[0][0][0] * 1000) 执行 :code:`python housing.py` 瞧！ 它应该打印出预测住房数据的清单。 diff --git a/doc/fluid/getstarted/quickstart_en.rst b/doc/fluid/getstarted/quickstart_en.rst index df6619cfd039fc1fdca8cde57db9cc6aebf8f029..680122f25893a5a48fac103266bda4788f891f6d 100644 --- a/doc/fluid/getstarted/quickstart_en.rst +++ b/doc/fluid/getstarted/quickstart_en.rst @@ -12,7 +12,7 @@ Simply run the following command to install, the version is cpu_avx_openblas: pip install paddlepaddle -If you need to install GPU version (cuda7.5_cudnn5_avx_openblas), run: +If you need to install GPU version (cuda8.0_cudnn5_avx_openblas), run: .. code-block:: bash @@ -31,18 +31,18 @@ code: import paddle.dataset.uci_housing as uci_housing import paddle.fluid as fluid - + with fluid.scope_guard(fluid.core.Scope()): # initialize executor with cpu exe = fluid.Executor(place=fluid.CPUPlace()) - # load inference model + # load inference model [inference_program, feed_target_names,fetch_targets] = \ fluid.io.load_inference_model(uci_housing.fluid_model(), exe) # run inference - result = exe.run(inference_program, - feed={feed_target_names[0]: uci_housing.predict_reader()}, + result = exe.run(inference_program, + feed={feed_target_names[0]: uci_housing.predict_reader()}, fetch_list=fetch_targets) - # print predicted price is $12,273.97 + # print predicted price is $12,273.97 print 'Predicted price: ${:,.2f}'.format(result[0][0][0] * 1000) Run :code:`python housing.py` and voila! It should print out a list of predictions diff --git a/doc/fluid/howto/cluster/nccl2_rdma_training.md b/doc/fluid/howto/cluster/nccl2_rdma_training.md new file mode 100644 index 0000000000000000000000000000000000000000..cecd5c3a7a7339e3be6772543a534728ec132105 --- /dev/null +++ b/doc/fluid/howto/cluster/nccl2_rdma_training.md @@ -0,0 +1,110 @@ +# Distributed Training with NCCL2 and RDMA + +When doing distributed multi-GPU training, network bandwith often becomes the +bottle neck. We introduce a way to use NCCL2 to do such training job to +achieve best performace. + +## Prepare Hardwares with RDMA and Multiple GPUs + +I'm using two Linux servers each of them is installed with 8 GPUs and +one 100Gb RDMA card. +Base environment is: + +* OS: CentOS 7.4 +* RDMA device: "Mellanox Technologies MT27700 Family [ConnectX-4]" +* Kernel version: `4.4.88-1.el7.elrepo.x86_64` +* Docker version: `1.12.6` +* Docker storage driver: `overlay2` +* IP addresses: 192.168.16.30,192.168.16.34 + +In general, the steps including: + +1. Install GPU drivers +1. Install RDMA drivers +1. Install "InfiniBand Support" +1. Use docker to run tests and make sure GPUs and RDMA can work inside + the container. + +I'll ommit section "Install GPU drivers" because we can find it easily +somewhere else. + +### Install RDMA drivers + +For my case, I've got two machines with device +"Mellanox Technologies MT27700 Family [ConnectX-4]" installed. The OS was +"CentOS 7.4" and I updated the kernel to version 4.4 so that docker can +work with latest overlay2 filesystem. + +***NOTE: before you start, make sure you have a way to get a console +of the server other than ssh because we may need to re-configure the +network device.*** + +1. Go to http://www.mellanox.com/page/products_dyn?product_family=26, + download `MLNX_OFED` software in the bottom of the page, and upload it + onto the server. +1. Run `./mlnxofedinstall --add-kernel-support` in the software package. +1. Run `/etc/init.d/openibd restart` to make everything work, note that + this operation may cause the network goes down if you are using this + RDMA device as default network device and use ssh to login the server. +1. Re-configure the network interface, for example: + `ifconfig eth2 192.168.16.30/20 up`, then add routes if needed: + `ip route add default via 192.168.16.1 dev eth2`. +1. Do the same thing on the other node. +1. Use `ping` to test if the two nodes have typical ICMP connection. +1. Use either `udaddy` or `ib_write_bw` to test the network connection is + ready and have the desired bandwith. + +### Prepare Docker Image to Run RDMA Programs + +1. Build a docker image using cuda base image like: `nvidia/cuda:8.0-cudnn5-devel-ubuntu16.04` and install paddlepaddle whl + package in it. +1. Start a docker container and mount GPU driver libs into it (you can + skip this step if you are using nvidia-docker). +1. Mount RDMA dirvers and libs into the docker image (see below section), + also `udaddy` and `ib_write_bw` if needed. +1. Mount GPU devices and RDMA devices into the container using `--device` + or just use privileged mode `--privileged`. +1. Start the container using host network mode: `--net=host` + +### RDMA Library Files Needed + +Usually, `MLNX_OFED` install latest supported libs under +`/usr/lib64/mlnx_ofed/valgrind`. Other libs also needed to run RDMA programs +is listed below. These libs must be mounted into the docker container. + +* Libs under `/usr/lib64/mlnx_ofed/valgrind` + * libibcm.so + * libibverbs.so + * libmlx4.so + * libmlx5.so + * libmlx5-rdmav2.so + * librdmacm.so +* Other libs: + * libnl-3.so.200 + * libnl-route-3.so.200 + * libnuma.so.1 + +## Start to Run the Training Job + +Setting NCCL environment variables to turn NCCL switches on and off: + + +| Env Name | Description | +| --- | --- | +| NCCL_SOCKET_IFNAME | The RDMA device, e.g. eth2 | +| NCCL_P2P_DISABLE | Set to 1 to disable P2P transfer between GPUs | +| NCCL_IB_DISABLE | Set to 1 to disable using RDMA | +| NCCL_IB_CUDA_SUPPORT | Set to 1 to enable GPU Direct if supported | +| NCCL_DEBUG | Set debug level: VERSION, WARN, INFO | + +My two servers are: `192.168.16.30,192.168.16.34`, On node 1, Run : + +```bash +PADDLE_TRAINER_ID=0 PADDLE_PORT=48372 PADDLE_WORKERS=192.168.16.30,192.168.16.34 POD_IP=192.168.16.30 stdbuf -oL python vgg16.py +``` + +On node 2, Run: + +```bash +PADDLE_TRAINER_ID=1 PADDLE_PORT=48372 PADDLE_WORKERS=192.168.16.30,192.168.16.34 POD_IP=192.168.16.34 stdbuf -oL python vgg16.py +``` diff --git a/doc/fluid/howto/index_cn.rst b/doc/fluid/howto/index_cn.rst index 97aeaf167d329529f2b120b5a3d4085e0510fe16..b7c620179724ebe97a0a47b75a57b376b21ccf90 100644 --- a/doc/fluid/howto/index_cn.rst +++ b/doc/fluid/howto/index_cn.rst @@ -3,5 +3,6 @@ .. toctree:: :maxdepth: 1 - + optimization/index_cn.rst + inference/inference_support_in_fluid.md diff --git a/doc/fluid/howto/index_en.rst b/doc/fluid/howto/index_en.rst index fd21e167ce3a46da167db1e9d7013804f730e047..f3ca41cdbf1d40ec8afaf045233a38755d8a777a 100644 --- a/doc/fluid/howto/index_en.rst +++ b/doc/fluid/howto/index_en.rst @@ -5,3 +5,4 @@ HOW TO :maxdepth: 1 optimization/index_en.rst + inference/inference_support_in_fluid.md diff --git a/doc/fluid/howto/inference/inference_support_in_fluid.md b/doc/fluid/howto/inference/inference_support_in_fluid.md new file mode 100644 index 0000000000000000000000000000000000000000..d272cd3e3bdac49b9ed1a21531de1b0be03d881e --- /dev/null +++ b/doc/fluid/howto/inference/inference_support_in_fluid.md @@ -0,0 +1,361 @@ +# Fluid Inference使用指南 + +## 目录： + +- Python Inference API +- 编译Fluid Inference库 +- Inference C++ API +- Inference实例 +- Inference计算优化 + +## Python Inference API **[改进中]** +- 保存Inference模型 ([链接](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/io.py#L295)) + + ```python + def save_inference_model(dirname, + feeded_var_names, + target_vars, + executor, + main_program=None, + model_filename=None, + params_filename=None): + ``` + Inference模型和参数将会保存到`dirname`目录下： + - 序列化的模型 + - `model_filename`为`None`，保存到`dirname/__model__` + - `model_filename`非`None`，保存到`dirname/model_filename` + - 参数 + - `params_filename`为`None`，单独保存到各个独立的文件，各文件以参数变量的名字命名 + - `params_filename`非`None`，保存到`dirname/params_filename` + +- 两种存储格式 + - 参数保存到各个独立的文件 + - 如，设置`model_filename`为`None`、`params_filename`为`None` + + ```bash + $ cd recognize_digits_conv.inference.model + $ ls + $ __model__ batch_norm_1.w_0 batch_norm_1.w_2 conv2d_2.w_0 conv2d_3.w_0 fc_1.w_0 batch_norm_1.b_0 batch_norm_1.w_1 conv2d_2.b_0 conv2d_3.b_0 fc_1.b_0 + ``` + - 参数保存到同一个文件 + - 如，设置`model_filename`为`None`、`params_filename`为`__params__` + + ```bash + $ cd recognize_digits_conv.inference.model + $ ls + $ __model__ __params__ + ``` +- 加载Inference模型([链接](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/io.py#L380)) + ```python + def load_inference_model(dirname, + executor, + model_filename=None, + params_filename=None): + ... + return [program, feed_target_names, fetch_targets] + ``` + + +## 编译Fluid Inference库 + + - **不需要额外的CMake选项** + - 1、 配置CMake命令，更多配置请参考[源码编译PaddlePaddle](http://www.paddlepaddle.org/docs/develop/documentation/zh/build_and_install/build_from_source_cn.html) + ```bash + $ git clone https://github.com/PaddlePaddle/Paddle.git + $ cd Paddle + $ mkdir build + $ cd build + $ cmake -DCMAKE_INSTALL_PREFIX=your/path/to/paddle_inference_lib \ + -DCMAKE_BUILD_TYPE=Release \ + -DWITH_PYTHON=ON \ + -DWITH_MKL=OFF \ + -DWITH_GPU=OFF \ + .. + ``` + + - 2、 编译PaddlePaddle + ```bash + $ make + ``` + + - 3、 部署。执行如下命令将PaddlePaddle Fluid Inference库部署到`your/path/to/paddle_inference_lib`目录。 + ```bash + $ make inference_lib_dist + ``` + +- 目录结构 + + ```bash + $ cd your/path/to/paddle_inference_lib + $ tree + . + |-- paddle + | `-- fluid + | |-- framework + | |-- inference + | | |-- io.h + | | `-- libpaddle_fluid.so + | |-- memory + | |-- platform + | `-- string + |-- third_party + | |-- eigen3 + | `-- install + | |-- gflags + | |-- glog + | `-- protobuf + `-- ... + ``` + + 假设`PADDLE_ROOT=your/path/to/paddle_inference_lib`。 + + + +## 链接Fluid Inference库 +- 示例项目([链接](https://github.com/luotao1/fluid_inference_example.git)) + + - GCC配置 + ```bash + $ g++ -o a.out -std=c++11 main.cc \ + -I${PADDLE_ROOT}/ \ + -I${PADDLE_ROOT}/third_party/install/gflags/include \ + -I${PADDLE_ROOT}/third_party/install/glog/include \ + -I${PADDLE_ROOT}/third_party/install/protobuf/include \ + -I${PADDLE_ROOT}/third_party/eigen3 \ + -L${PADDLE_ROOT}/paddle/fluid/inference -lpaddle_fluid \ + -lrt -ldl -lpthread + ``` + + - CMake配置 + ```cmake + include_directories(${PADDLE_ROOT}/) + include_directories(${PADDLE_ROOT}/third_party/install/gflags/include) + include_directories(${PADDLE_ROOT}/third_party/install/glog/include) + include_directories(${PADDLE_ROOT}/third_party/install/protobuf/include) + include_directories(${PADDLE_ROOT}/third_party/eigen3) + target_link_libraries(${TARGET_NAME} + ${PADDLE_ROOT}/paddle/fluid/inference/libpaddle_fluid.so + -lrt -ldl -lpthread) + ``` + + - 设置环境变量： + `export LD_LIBRARY_PATH=${PADDLE_ROOT}/paddle/fluid/inference:$LD_LIBRARY_PATH` + + + +## C++ Inference API + +- 推断流程([链接](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/inference/tests/test_helper.h#L91)) + + - 1、 初始化设备 + ```cpp + #include "paddle/fluid/framework/init.h" + paddle::framework::InitDevices(false); + ``` + + - 2、 定义place，executor，scope + ```cpp + auto place = paddle::platform::CPUPlace(); + auto executor = paddle::framework::Executor(place); + auto* scope = new paddle::framework::Scope(); + ``` + + - 3、 加载模型 + ```cpp + #include "paddle/fluid/inference/io.h" + auto inference_program = paddle::inference::Load(executor, *scope, dirname); + // or + auto inference_program = paddle::inference::Load(executor, + *scope, + dirname + "/" + model_filename, + dirname + "/" + params_filename); + ``` + + - 4、 获取`feed_target_names`和`fetch_target_names` + ```cpp + const std::vector& feed_target_names = inference_program->GetFeedTargetNames(); + const std::vector& fetch_target_names = inference_program->GetFetchTargetNames(); + ``` + + - 5、 准备`feed`数据 + ```cpp + #include "paddle/fluid/framework/lod_tensor.h" + std::vector cpu_feeds; + ... + std::map feed_targets; + for (size_t i = 0; i < feed_target_names.size(); ++i) { + // Please make sure that cpu_feeds[i] is right for feed_target_names[i] + feed_targets[feed_target_names[i]] = cpu_feeds[i]; + } + ``` + + - 6、 定义`Tensor`来`fetch`结果 + ```cpp + std::vector cpu_fetchs; + std::map fetch_targets; + for (size_t i = 0; i < fetch_target_names.size(); ++i) { + fetch_targets[fetch_target_names[i]] = cpu_fetchs[i]; + } + ``` + + - 7、 执行`inference_program` + ```cpp + executor.Run(*inference_program, scope, feed_targets, fetch_targets); + ``` + + - 8、 使用`fetch`数据 + ```cpp + for (size_t i = 0; i < cpu_fetchs.size(); ++i) { + std::cout << "lod_i: " << cpu_fetchs[i]->lod(); + std::cout << "dims_i: " << cpu_fetchs[i]->dims(); + std::cout << "result:"; + float* output_ptr = cpu_fetchs[i]->data(); + for (int j = 0; j < cpu_fetchs[i]->numel(); ++j) { + std::cout << " " << output_ptr[j]; + } + std::cout << std::endl; + } + ``` + 针对不同的数据，4. - 8.可执行多次。 + + - 9、 释放内存 + ```cpp + delete scope; + ``` + + +- 接口说明 + + ```cpp + void Run(const ProgramDesc& program, Scope* scope, + std::map& feed_targets, + std::map& fetch_targets, + bool create_vars = true, + const std::string& feed_holder_name = "feed", + const std::string& fetch_holder_name = "fetch"); + ``` + - 使用Python API `save_inference_model`保存的`program`里面包含了`feed_op`和`fetch_op`，用户提供的`feed_targets`、`fetch_targets`必须和`inference_program`中的`feed_op`、`fetch_op`保持一致。 + - 用户提供的`feed_holder_name`和`fetch_holder_name`也必须和`inference_program`中`feed_op`、`fetch_op`保持一致，可使用`SetFeedHolderName`和`SetFetchHolderName`接口重新设置`inferece_program` + - 默认情况下，除了`persistable`属性设置为`True`的`Variable`之外，每次执行`executor.Run`会创建一个局部`Scope`，并且在这个局部`Scope`中创建和销毁所有的`Variable`，以最小化空闲时的内存占用。 + - `persistable`属性为`True`的`Variable`有： + - Operators的参数`w`、`b`等 + - `feed_op`的输入变量 + - `fetch_op`的输出变量 + + +- **不在每次执行时创建和销毁变量 + ([PR](https://github.com/PaddlePaddle/Paddle/pull/9301))** + - 执行`inference_program` + ```cpp + // Call once + executor.CreateVariables(*inference_program, scope, 0); + // Call as many times as you like + executor.Run( + *inference_program, scope, feed_targets, fetch_targets, false); + ``` + - **优点** + - 节省了频繁创建、销毁变量的时间（约占每次`Run`总时间的1% ~ 12%） + - 执行结束后可获取所有Operators的计算结果 + - **缺点** + - 空闲时也会占用大量的内存 + - 在同一个`Scope`中，相同的变量名是公用同一块内存的，容易引起意想不到的错误 + + +- **不在每次执行时创建Op([PR](https://github.com/PaddlePaddle/Paddle/pull/9630))** + - 执行`inference_program` + ```cpp + // Call once + auto ctx = executor.Prepare(*inference_program, 0); + // Call as many times as you like if you have no need to change the inference_program + executor.RunPreparedContext(ctx.get(), scope, feed_targets, fetch_targets); + ``` + - **优点** + - 节省了频繁创建、销毁Op的时间 + - **缺点** + - 一旦修改了`inference_program`，则需要重新创建`ctx` + + +- **多线程共享Parameters([链接](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/inference/tests/test_multi_thread_helper.h))** + - 主线程 + - 1、 初始化设备 + - 2、 定义`place`，`executor`，`scope` + - 3、 加载模型，得到`inference_program` + - 从线程 + - **复制`inference_program`得到`copy_program`，修改`copy_program`的`feed_holder_name`和`fetch_holder_name`** + ```cpp + auto copy_program = std::unique_ptr( + new paddle::framework::ProgramDesc(*inference_program)); + std::string feed_holder_name = "feed_" + paddle::string::to_string(thread_id); + std::string fetch_holder_name = "fetch_" + paddle::string::to_string(thread_id); + copy_program->SetFeedHolderName(feed_holder_name); + copy_program->SetFetchHolderName(fetch_holder_name); + ``` + - 4、 获取`copy_program`的`feed_target_names`和`fetch_target_names` + - 5、 准备feed数据，定义Tensor来fetch结果 + - 6、 执行`copy_program` + ```cpp + executor->Run(*copy_program, scope, feed_targets, fetch_targets, true, feed_holder_name, fetch_holder_name); + ``` + - 7、 使用fetch数据 + - 主线程 + - 8、 释放资源 + + +- 基本概念 + - 数据相关： + - [Tensor](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/tensor.md)，一个N维数组，数据可以是任意类型（int，float，double等） + - [LoDTensor](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/lod_tensor.md)，带LoD(Level-of-Detail)即序列信息的Tensor + - [Scope](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/scope.md)，记录了变量Variable + - 执行相关： + - [Executor](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/executor.md)，无状态执行器，只跟设备相关 + - Place + - CPUPlace，CPU设备 + - CUDAPlace，CUDA GPU设备 + - 神经网络表示： + - [Program](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/program.md). + + 详细介绍请参考[**Paddle Fluid开发者指南**](https://github.com/lcy-seso/learning_notes/blob/master/Fluid/developer's_guid_for_Fluid/Developer's_Guide_to_Paddle_Fluid.md) + + + +## Inference实例 + + 1. fit a line: [Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/tests/book/test_fit_a_line.py), [C++](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/inference/tests/book/test_inference_fit_a_line.cc) + 1. image classification: [Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/tests/book/test_image_classification.py), [C++](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/inference/tests/book/test_inference_image_classification.cc) + 1. label semantic roles: [Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/tests/book/test_label_semantic_roles.py), [C++](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/inference/tests/book/test_inference_label_semantic_roles.cc) + 1. recognize digits: [Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/tests/book/test_recognize_digits.py), [C++](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/inference/tests/book/test_inference_recognize_digits.cc) + 1. recommender system: [Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/tests/book/test_recommender_system.py), [C++](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/inference/tests/book/test_inference_recommender_system.cc) + 1. understand sentiment: [Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/tests/book/test_understand_sentiment.py), [C++](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/inference/tests/book/test_inference_understand_sentiment.cc) + 1. word2vec: [Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/tests/book/test_word2vec.py), [C++](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/inference/tests/book/test_inference_word2vec.cc) + + +## Inference计算优化 +- 使用Python推理优化工具([inference_transpiler](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/inference_transpiler.py)) + ```python + class InferenceTranspiler: + def transpile(self, program, place, scope=None): + ... + if scope is None: + scope = global_scope() + ... + ``` + - 使用`InferenceTranspiler`将会直接修改`program`。 + - 使用`InferenceTranspiler`会修改参数的值，请确保`program`的参数在`scope`内。 +- 支持的优化 + - 融合batch_norm op的计算 +- 使用示例([链接](https://github.com/Xreki/Xreki.github.io/blob/master/fluid/inference/inference_transpiler.py)) + ```python + import paddle.fluid as fluid + # NOTE: Applying the inference transpiler will change the inference_program. + t = fluid.InferenceTranspiler() + t.transpile(inference_program, place, inference_scope) + ``` + + + + +## 内存使用优化 +- 使用Python内存优化工具([memory_optimization_transipiler](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/memory_optimization_transpiler.py)) + ```python + fluid.memory_optimize(inference_program) + ``` diff --git a/doc/fluid/howto/optimization/cpu_profiling_cn.md b/doc/fluid/howto/optimization/cpu_profiling_cn.md index 8266dec3c6125a09b90ac0ccd4aa5464f5c7db31..198a05a79e19227e90eaafe116217a164cd51a7d 100644 --- a/doc/fluid/howto/optimization/cpu_profiling_cn.md +++ b/doc/fluid/howto/optimization/cpu_profiling_cn.md @@ -1,3 +1,5 @@ +# CPU性能调优 + 此教程会介绍如何使用Python的cProfile包、Python库yep、Google perftools来进行性能分析 (profiling) 与调优（performance tuning）。 Profling 指发现性能瓶颈。系统中的瓶颈可能和程序员开发过程中想象的瓶颈相去甚远。Tuning 指消除瓶颈。性能优化的过程通常是不断重复地 profiling 和 tuning。 @@ -8,7 +10,7 @@ PaddlePaddle 用户一般通过调用 Python API 编写深度学习程序。大 * Python 与 C++ 混合代码的性能分析 -# Python代码的性能分析 +## Python代码的性能分析 ### 生成性能分析文件 diff --git a/doc/fluid/howto/optimization/cpu_profiling_en.md b/doc/fluid/howto/optimization/cpu_profiling_en.md index e95556dd608b7ff0a3eb18873df0015a2da94e7c..216694965b3c878a8a5f3ccd2a0cba8d21d9ce05 100644 --- a/doc/fluid/howto/optimization/cpu_profiling_en.md +++ b/doc/fluid/howto/optimization/cpu_profiling_en.md @@ -1,3 +1,5 @@ +# Tune CPU performance + This tutorial introduces techniques we use to profile and tune the CPU performance of PaddlePaddle. We will use Python packages `cProfile` and `yep`, and Google's `perftools`. @@ -14,7 +16,7 @@ the profiling and tuning of 1. the Python code and 1. the mixture of Python and C++ code. -# Profiling the Python Code +## Profiling the Python Code ### Generate the Performance Profiling File diff --git a/doc/fluid/images/1.png b/doc/fluid/images/1.png new file mode 100644 index 0000000000000000000000000000000000000000..67daf566f91aab570e60971c4ea8e2be876e214d Binary files /dev/null and b/doc/fluid/images/1.png differ diff --git a/doc/fluid/images/2.png b/doc/fluid/images/2.png new file mode 100644 index 0000000000000000000000000000000000000000..43367777f41449a666e7a3b571f09ac5d5dfb1ae Binary files /dev/null and b/doc/fluid/images/2.png differ diff --git a/doc/fluid/images/3.png b/doc/fluid/images/3.png new file mode 100644 index 0000000000000000000000000000000000000000..481021ef306e2596818aab7fe17a570754f63635 Binary files /dev/null and b/doc/fluid/images/3.png differ diff --git a/doc/fluid/images/4.png b/doc/fluid/images/4.png new file mode 100644 index 0000000000000000000000000000000000000000..4279f41e06de459f18b9a622539511d555e9a0af Binary files /dev/null and b/doc/fluid/images/4.png differ diff --git a/doc/fluid/images/LoDTensor.png b/doc/fluid/images/LoDTensor.png new file mode 100644 index 0000000000000000000000000000000000000000..75369f5378309e0f304b83f6bb69bdb195eac079 Binary files /dev/null and b/doc/fluid/images/LoDTensor.png differ diff --git a/doc/fluid/images/compile_run_time.png b/doc/fluid/images/compile_run_time.png new file mode 100644 index 0000000000000000000000000000000000000000..0bc9b2fd0e81b4851e6d96171ccb9a05d0f42a48 Binary files /dev/null and b/doc/fluid/images/compile_run_time.png differ diff --git a/doc/fluid/images/executor.png b/doc/fluid/images/executor.png new file mode 100644 index 0000000000000000000000000000000000000000..b29c0d779e3d46b779b5baeabe3176adaeb00a6d Binary files /dev/null and b/doc/fluid/images/executor.png differ diff --git a/doc/fluid/images/fluid_examples.png b/doc/fluid/images/fluid_examples.png new file mode 100644 index 0000000000000000000000000000000000000000..aa99472c0f914cde128fd7b3bd8dc29ac24f94b6 Binary files /dev/null and b/doc/fluid/images/fluid_examples.png differ diff --git a/doc/fluid/images/fluid_module_1.png b/doc/fluid/images/fluid_module_1.png new file mode 100644 index 0000000000000000000000000000000000000000..554782ba54e43efc3d6babbb94e3cac3530ac649 Binary files /dev/null and b/doc/fluid/images/fluid_module_1.png differ diff --git a/doc/fluid/images/fluid_module_2.png b/doc/fluid/images/fluid_module_2.png new file mode 100644 index 0000000000000000000000000000000000000000..4219efccbb1e87839adf6b5720fe46808b7d2fcf Binary files /dev/null and b/doc/fluid/images/fluid_module_2.png differ diff --git a/doc/fluid/images/layer.png b/doc/fluid/images/layer.png new file mode 100644 index 0000000000000000000000000000000000000000..e46db4c9c6f5b65ff274b498b716b11de343a8b0 Binary files /dev/null and b/doc/fluid/images/layer.png differ diff --git a/doc/fluid/images/op.dot b/doc/fluid/images/op.dot new file mode 100644 index 0000000000000000000000000000000000000000..c8ad839cb88788e9b5906402257cc7bbc3ddcb54 --- /dev/null +++ b/doc/fluid/images/op.dot @@ -0,0 +1,4 @@ +digraph sample { + graph [rankdir=TD]; node [shape=record]; + op [label="{Operator| InferShape()=0\lRun()=0\l | map<string, string[]> inputs_\lmap<string, string[]> outputs_ \l AttributeMap attrs_\l}"]; +} \ No newline at end of file diff --git a/doc/fluid/images/op_op_with_kern_class_diagram.dot b/doc/fluid/images/op_op_with_kern_class_diagram.dot new file mode 100644 index 0000000000000000000000000000000000000000..8f24e9ea83acf879c7008f2d97113c0a4cc111c3 --- /dev/null +++ b/doc/fluid/images/op_op_with_kern_class_diagram.dot @@ -0,0 +1,38 @@ +digraph sample { + graph [rankdir=TD]; node [shape=record]; + op [label="{Operator| InferShape()=0\lRun()=0\l | map<string, string[]> inputs_\lmap<string, string[]> outputs_ \l AttributeMap attrs_\l}"]; + op_with_kern [label="{OpWithKernel | InferShape()=0\lRun()\l | map<OpKernelKey,OpKernel>kernels_ }"] + op_kernel [label="{OpKernel | Compute()=0}"] + op_kernel_key [label="{OpKernelKey| Place place\n...}"] + + op -> op_with_kern [dir=back, arrowtail=onormal] + op_with_kern -> op_kernel [arrowhead=vee, label="contains many"] + + { + rank=same; + op_with_kern + op_kernel + } + + op_kernel -> op_kernel_key [style=invis] + + { + rank=same; + op_kernel + op_kernel_key + } + + op_with_kern -> op_kernel_key [arrowhead=vee, label ="\nas map key"] + + mul_op [label="MulOp"] + op_with_kern -> mul_op [dir=back, arrowtail=onormal] + mul_kernel [label="template <typename Place>\lclass MulOpKernel\l"] + op_kernel -> mul_kernel [dir=back, arrowtail=onormal] + mul_op -> mul_kernel [arrowhead=vee, label="register many"] + + { + rank=same; + mul_op; + mul_kernel; + } +} \ No newline at end of file diff --git a/doc/fluid/images/op_with_kernel.dot b/doc/fluid/images/op_with_kernel.dot new file mode 100644 index 0000000000000000000000000000000000000000..4f5af4f7b5f5a69693a058c99eb658900136077a --- /dev/null +++ b/doc/fluid/images/op_with_kernel.dot @@ -0,0 +1,26 @@ +digraph sample { + graph [rankdir=TD]; node [shape=record]; + op [label="{Operator}"]; + op_with_kern [label="{OpWithKernel | InferShape()=0\lRun()\l | map<OpKernelKey,OpKernel>kernels_ }"] + op_kernel [label="{OpKernel | Compute()=0}"] + op_kernel_key [label="{OpKernelKey| Place place\n...}"] + + op -> op_with_kern [dir=back, arrowtail=onormal] + op_with_kern -> op_kernel [arrowhead=vee, label="contains many"] + + { + rank=same; + op_with_kern + op_kernel + } + + op_kernel -> op_kernel_key [style=invis] + + { + rank=same; + op_kernel + op_kernel_key + } + + op_with_kern -> op_kernel_key [arrowhead=vee, label ="\nas map key"] +} \ No newline at end of file diff --git a/doc/fluid/images/operator1.png b/doc/fluid/images/operator1.png new file mode 100644 index 0000000000000000000000000000000000000000..3975b06f615b7a88dfc11e71b6451fdf4ce42d60 Binary files /dev/null and b/doc/fluid/images/operator1.png differ diff --git a/doc/fluid/images/operator2.png b/doc/fluid/images/operator2.png new file mode 100644 index 0000000000000000000000000000000000000000..b7bb1fae2050d3a70797517bc20dbbdef3dfcb7c Binary files /dev/null and b/doc/fluid/images/operator2.png differ diff --git a/doc/fluid/images/place.png b/doc/fluid/images/place.png new file mode 100644 index 0000000000000000000000000000000000000000..14e77511d639af155e5a3725cde05323e0cc94f2 Binary files /dev/null and b/doc/fluid/images/place.png differ diff --git a/doc/fluid/images/print_fluid_program.png b/doc/fluid/images/print_fluid_program.png new file mode 100644 index 0000000000000000000000000000000000000000..e8e459e1b3d5c8706b3caa05dc371db8d46df4a5 Binary files /dev/null and b/doc/fluid/images/print_fluid_program.png differ diff --git a/doc/fluid/images/program_desc1.png b/doc/fluid/images/program_desc1.png new file mode 100644 index 0000000000000000000000000000000000000000..0656336914ece957f2e5bb4d70ad337a63e31d88 Binary files /dev/null and b/doc/fluid/images/program_desc1.png differ diff --git a/doc/fluid/images/program_desc2.png b/doc/fluid/images/program_desc2.png new file mode 100644 index 0000000000000000000000000000000000000000..db5bfa1231345add8661b4f8ef0fc9d861f40d24 Binary files /dev/null and b/doc/fluid/images/program_desc2.png differ diff --git a/doc/fluid/images/raw_input.png b/doc/fluid/images/raw_input.png new file mode 100644 index 0000000000000000000000000000000000000000..0725f92d2b169c2b59ec7c68b402859c2a2dd1d8 Binary files /dev/null and b/doc/fluid/images/raw_input.png differ diff --git a/doc/fluid/images/scope_variable_tensor.png b/doc/fluid/images/scope_variable_tensor.png new file mode 100644 index 0000000000000000000000000000000000000000..59b0de6fb36f9f6b469227c05760a7612bb30b4d Binary files /dev/null and b/doc/fluid/images/scope_variable_tensor.png differ diff --git a/doc/fluid/images/sorted_input.png b/doc/fluid/images/sorted_input.png new file mode 100644 index 0000000000000000000000000000000000000000..ff601128368ee179e3fd33e5e295a9ddd3dcbaeb Binary files /dev/null and b/doc/fluid/images/sorted_input.png differ diff --git a/doc/fluid/images/transpiler.png b/doc/fluid/images/transpiler.png new file mode 100644 index 0000000000000000000000000000000000000000..422973c0dc7aa2b544d2fc86a97ace706388cb9e Binary files /dev/null and b/doc/fluid/images/transpiler.png differ diff --git a/doc/fluid/images/user_interface.png b/doc/fluid/images/user_interface.png new file mode 100644 index 0000000000000000000000000000000000000000..ffc94e3d8945ec6291460afd90e8fcc600828390 Binary files /dev/null and b/doc/fluid/images/user_interface.png differ diff --git a/doc/mobile/CMakeLists.txt b/doc/mobile/CMakeLists.txt index b104a6318d474d6531670b8ac3569448774850c7..7b34ba8d0768427802b11614c6962f3c3f6ef4e3 100644 --- a/doc/mobile/CMakeLists.txt +++ b/doc/mobile/CMakeLists.txt @@ -15,6 +15,9 @@ set(SPHINX_CACHE_DIR_EN "${CMAKE_CURRENT_BINARY_DIR}/en/_doctrees") # HTML output director set(SPHINX_HTML_DIR_EN "${CMAKE_CURRENT_BINARY_DIR}/en/html") +set(IMPORT_PADDLE_STRING "") +set(IMPORT_PADDLEV2_STRING "") + configure_file( "${CMAKE_CURRENT_SOURCE_DIR}/../templates/conf.py.en.in" "${BINARY_BUILD_DIR_EN}/conf.py" @@ -27,8 +30,6 @@ sphinx_add_target(paddle_mobile_docs ${CMAKE_CURRENT_SOURCE_DIR} ${SPHINX_HTML_DIR_EN}) -add_dependencies(paddle_mobile_docs gen_proto_py paddle_python) - # configured documentation tools and intermediate build results set(BINARY_BUILD_DIR_CN "${CMAKE_CURRENT_BINARY_DIR}/cn/_build") @@ -49,5 +50,3 @@ sphinx_add_target(paddle_mobile_docs_cn ${SPHINX_CACHE_DIR_CN} ${CMAKE_CURRENT_SOURCE_DIR} ${SPHINX_HTML_DIR_CN}) - -add_dependencies(paddle_mobile_docs_cn gen_proto_py paddle_python) diff --git a/doc/mobile/index_cn.rst b/doc/mobile/index_cn.rst index 8297316e8fbb2b8f41954030293feadbcd81295e..56d1515005f6e40b084c6b2184c6a0b3e3a00496 100644 --- a/doc/mobile/index_cn.rst +++ b/doc/mobile/index_cn.rst @@ -1,9 +1,9 @@ 移动端 -===== +====== .. toctree:: :maxdepth: 1 cross_compiling_for_android_cn.md cross_compiling_for_ios_cn.md - cross_compiling_for_raspberry_cn.md \ No newline at end of file + cross_compiling_for_raspberry_cn.md diff --git a/doc/templates/conf.py.cn.in b/doc/templates/conf.py.cn.in index 76b82fd97f1ed642696c4414676b694ebda9ad81..890f70615538af23cd05b9ffd685e870a5644cdb 100644 --- a/doc/templates/conf.py.cn.in +++ b/doc/templates/conf.py.cn.in @@ -16,8 +16,8 @@ import os, subprocess sys.path.insert(0, os.path.abspath('@PADDLE_BINARY_DIR@/python')) import shlex from recommonmark import parser, transform -import paddle -import paddle.v2 +@IMPORT_PADDLE_STRING@ +@IMPORT_PADDLEV2_STRING@ MarkdownParser = parser.CommonMarkParser AutoStructify = transform.AutoStructify diff --git a/doc/templates/conf.py.en.in b/doc/templates/conf.py.en.in index 5aa5c1381fa3fad4ebc181c7868da03ae0138016..5b09464cb991f96127edec40f7dbbc97a8d82582 100644 --- a/doc/templates/conf.py.en.in +++ b/doc/templates/conf.py.en.in @@ -16,8 +16,8 @@ import os, subprocess sys.path.insert(0, os.path.abspath('@PADDLE_BINARY_DIR@/python')) import shlex from recommonmark import parser, transform -import paddle -import paddle.v2 +@IMPORT_PADDLE_STRING@ +@IMPORT_PADDLEV2_STRING@ MarkdownParser = parser.CommonMarkParser diff --git a/doc/v2/CMakeLists.txt b/doc/v2/CMakeLists.txt index be957d37b14c618e9346251b3bd3dbaf1541773f..d230a1b9217eea6740419822f350096e361a4435 100644 --- a/doc/v2/CMakeLists.txt +++ b/doc/v2/CMakeLists.txt @@ -15,6 +15,9 @@ set(SPHINX_CACHE_DIR_EN "${CMAKE_CURRENT_BINARY_DIR}/en/_doctrees") # HTML output director set(SPHINX_HTML_DIR_EN "${CMAKE_CURRENT_BINARY_DIR}/en/html") +set(IMPORT_PADDLE_STRING "") +set(IMPORT_PADDLEV2_STRING "") + configure_file( "${CMAKE_CURRENT_SOURCE_DIR}/../templates/conf.py.en.in" "${BINARY_BUILD_DIR_EN}/conf.py" @@ -27,8 +30,6 @@ sphinx_add_target(paddle_v2_docs ${CMAKE_CURRENT_SOURCE_DIR} ${SPHINX_HTML_DIR_EN}) -add_dependencies(paddle_v2_docs gen_proto_py paddle_python) - # configured documentation tools and intermediate build results set(BINARY_BUILD_DIR_CN "${CMAKE_CURRENT_BINARY_DIR}/cn/_build") @@ -50,6 +51,4 @@ sphinx_add_target(paddle_v2_docs_cn ${CMAKE_CURRENT_SOURCE_DIR} ${SPHINX_HTML_DIR_CN}) -add_dependencies(paddle_v2_docs_cn gen_proto_py paddle_python) - add_subdirectory(api) diff --git a/doc/v2/api/CMakeLists.txt b/doc/v2/api/CMakeLists.txt index 2670a21a227546ffcee4f10f395feef3c58df9b4..0c74522cb089b17c8419e9058f76631b0fe0df93 100644 --- a/doc/v2/api/CMakeLists.txt +++ b/doc/v2/api/CMakeLists.txt @@ -7,6 +7,9 @@ set(SPHINX_CACHE_DIR_EN "${CMAKE_CURRENT_BINARY_DIR}/en/_doctrees") # HTML output director set(SPHINX_HTML_DIR_EN "${CMAKE_CURRENT_BINARY_DIR}/en/html") +set(IMPORT_PADDLE_STRING "import paddle") +set(IMPORT_PADDLEV2_STRING "import paddle.v2") + configure_file( "${CMAKE_CURRENT_SOURCE_DIR}/../../templates/conf.py.en.in" "${BINARY_BUILD_DIR_EN}/conf.py" diff --git a/doc/v2/api/config/layer.rst b/doc/v2/api/config/layer.rst index 29388f5005bf779a1bfa63c0d46d35996c0c792d..1a6496968cae1fef88142ba9ca3f9e63a81b196d 100644 --- a/doc/v2/api/config/layer.rst +++ b/doc/v2/api/config/layer.rst @@ -142,7 +142,7 @@ gated_unit ----------- .. autoclass:: paddle.v2.layer.gated_unit :noindex: - + Recurrent Layer Group ===================== @@ -354,7 +354,7 @@ dropout -------- .. autoclass:: paddle.v2.layer.dropout :noindex: - + dot_prod --------- .. autoclass:: paddle.v2.layer.dot_prod @@ -460,6 +460,11 @@ multi_binary_label_cross_entropy_cost .. autoclass:: paddle.v2.layer.multi_binary_label_cross_entropy_cost :noindex: +classification_cost +------------------- +.. autoclass:: paddle.v2.layer.classification_cost + :noindex: + huber_regression_cost ------------------------- .. autoclass:: paddle.v2.layer.huber_regression_cost @@ -534,7 +539,7 @@ detection_output ---------------- .. autoclass:: paddle.v2.layer.detection_output :noindex: - + Check Layer ============ diff --git a/doc/v2/api/data/data_reader.rst b/doc/v2/api/data/data_reader.rst index d7c896a6270b488ca4449e5211d0d0879eda6ac5..1a35d0bbc8f9d751f49c7e1fc26feb1bcb3ae7f0 100644 --- a/doc/v2/api/data/data_reader.rst +++ b/doc/v2/api/data/data_reader.rst @@ -56,11 +56,11 @@ DataFeeder Reader ====== -.. automodule:: paddle.v2.reader +.. automodule:: paddle.reader :members: :noindex: -.. automodule:: paddle.v2.reader.creator +.. automodule:: paddle.reader.creator :members: :noindex: diff --git a/doc/v2/build_and_install/build_from_source_cn.rst b/doc/v2/build_and_install/build_from_source_cn.rst index 115b92a33888abf1e1be400e1abbb58b632a2976..741c01ce5428c0046daa5a784da70d4bb492438c 100644 --- a/doc/v2/build_and_install/build_from_source_cn.rst +++ b/doc/v2/build_and_install/build_from_source_cn.rst @@ -19,8 +19,9 @@ ---------------- PaddlePaddle需要使用Docker环境完成编译，这样可以免去单独安装编译依赖的步骤，可选的不同编译环境Docker镜像 -可以在 `这里 `_ 找到。或者 -参考下述可选步骤，从源码中构建用于编译PaddlePaddle的Docker镜像。 +可以在 `这里 `__ 找到，您也可以 +在 `这里 `__ 找到 paddle_manylinux_devel +镜像的编译以及使用方法。或者参考下述可选步骤，从源码中构建用于编译PaddlePaddle的Docker镜像。 如果您选择不使用Docker镜像，则需要在本机安装下面章节列出的 `编译依赖`_ 之后才能开始编译的步骤。 @@ -34,13 +35,11 @@ PaddlePaddle需要使用Docker环境完成编译，这样可以免去单独安 # 2. 可选步骤：源码中构建用于编译PaddlePaddle的Docker镜像 docker build -t paddle:dev . # 3. 执行下面的命令编译CPU-Only的二进制 - docker run -it -v $PWD:/paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=OFF" paddlepaddle/paddle_manylinux_devel:cuda8.0_cudnn5 bash -x /paddle/paddle/scripts/docker/build.sh + docker run -it -v $PWD:/paddle -w /paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=OFF" paddlepaddle/paddle_manylinux_devel:cuda8.0_cudnn5 ./paddle/scripts/paddle_build.sh build # 4. 或者也可以使用为上述可选步骤构建的镜像（必须先执行第2步） - docker run -it -v $PWD:/paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=OFF" paddle:dev + docker run -it -v $PWD:/paddle -w /paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=OFF" paddle:dev ./paddle/scripts/paddle_build.sh build -注：上述命令把当前目录（源码树根目录）映射为 container 里的 :code:`/paddle` 目录。如果使用自行 -构建的镜像（上述第4步）会执行 :code:`Dockerfile` 描述的默认入口程序 :code:`build.sh` 可以省略步骤3中 -最后的执行脚本的命令。 +注：上述命令把当前目录（源码树根目录）映射为 container 里的 :code:`/paddle` 目录。 编译完成后会在build/python/dist目录下生成输出的whl包，可以选在在当前机器安装也可以拷贝到目标机器安装： @@ -71,15 +70,15 @@ PaddlePaddle需要使用Docker环境完成编译，这样可以免去单独安 .. code-block:: bash - docker run -it -v $PWD:/paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=ON" -e "RUN_TEST=ON" paddlepaddle/paddle_manylinux_devel:cuda8.0_cudnn5 bash -x /paddle/paddle/scripts/docker/build.sh + docker run -it -v $PWD:/paddle -w /paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=ON" -e "RUN_TEST=ON" paddlepaddle/paddle_manylinux_devel:cuda8.0_cudnn5 ./paddle/scripts/paddle_build.sh test 如果期望执行其中一个单元测试，（比如 :code:`test_sum_op` ）： .. code-block:: bash - docker run -it -v $PWD:/paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=ON" -e "RUN_TEST=OFF" paddlepaddle/paddle_manylinux_devel:cuda8.0_cudnn5 /bin/bash - bash /paddle/paddle/scripts/docker/build.sh - cd /paddle/build + docker run -it -v $PWD:/paddle -w /paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=ON" -e "RUN_TEST=OFF" paddlepaddle/paddle_manylinux_devel:cuda8.0_cudnn5 /bin/bash + ./paddle/scripts/paddle_build.sh build + cd build ctest -R test_sum_op -V .. _faq_docker: @@ -115,11 +114,10 @@ PaddlePaddle需要使用Docker环境完成编译，这样可以免去单独安 很多 PaddlePaddle 开发者使用 Emacs。他们在自己的 `~/.emacs` 配置文件里加两行 - ```emacs - (global-set-key "\C-cc" 'compile) - (setq compile-command - "docker run --rm -it -v $(git rev-parse --show-toplevel):/paddle paddle:dev") - ``` + .. code-block:: emacs + + (global-set-key "\C-cc" 'compile) + (setq compile-command "docker run --rm -it -v $(git rev-parse --show-toplevel):/paddle paddle:dev") 就可以按 `Ctrl-C` 和 `c` 键来启动编译了。 diff --git a/doc/v2/build_and_install/build_from_source_en.rst b/doc/v2/build_and_install/build_from_source_en.rst index 8fef9e7347e8d924026999bfda985381750c6b51..b06c43e19dcfc52ad0f074a85517a16744895a3a 100644 --- a/doc/v2/build_and_install/build_from_source_en.rst +++ b/doc/v2/build_and_install/build_from_source_en.rst @@ -22,6 +22,8 @@ How To Build You need to use Docker to build PaddlePaddle to avoid installing dependencies by yourself. We have several pre-built Docker images `here `_ , +you can also find how to build and use paddle_manylinux_devel Docker image from +`here `__ Or you can build your own image from source as the optional step below: .. code-block:: bash @@ -32,14 +34,12 @@ Or you can build your own image from source as the optional step below: # 2. Optional: build development docker image from source docker build -t paddle:dev . # 3. Run the following command to build a CPU-Only binaries - docker run -it -v $PWD:/paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=OFF" paddlepaddle/paddle_manylinux_devel:cuda8.0_cudnn5 bash -x /paddle/paddle/scripts/docker/build.sh + docker run -it -v $PWD:/paddle -w /paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=OFF" paddlepaddle/paddle_manylinux_devel:cuda8.0_cudnn5 ./paddle/scripts/paddle_build.sh build # 4. Or, use your built Docker image to build PaddlePaddle (must run step 2) - docker run -it -v $PWD:/paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=OFF" paddle:dev + docker run -it -v $PWD:/paddle -w /paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=OFF" paddle:dev ./paddle/scripts/paddle_build.sh build NOTE: The above command try to mount the current working directory (root directory of source code) -into :code:`/paddle` directory inside docker container. If you are using your own image -(Step 4) it will run default entry-point :code:`build.sh` , so you could omit the last -command in step 3. +into :code:`/paddle` directory inside docker container. When the compile finishes, you can get the output whl package under build/python/dist, then you can choose to install the whl on local @@ -72,21 +72,21 @@ Set :code:`WITH_GPU=ON` Can also run tests on GPU. .. code-block:: bash - docker run -it -v $PWD:/paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=ON" -e "RUN_TEST=ON" paddlepaddle/paddle_manylinux_devel:cuda8.0_cudnn5 bash -x paddle/paddle/scripts/docker/build.sh + docker run -it -v $PWD:/paddle -w /paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=ON" -e "RUN_TEST=ON" paddlepaddle/paddle_manylinux_devel:cuda8.0_cudnn5 ./paddle/scripts/paddle_build.sh test If you wish to run only one unit test, like :code:`test_sum_op`: .. code-block:: bash - docker run -it -v $PWD:/paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=ON" -e "RUN_TEST=OFF" paddlepaddle/paddle_manylinux_devel:cuda8.0_cudnn5 /bin/bash - bash /paddle/paddle/scripts/docker/build.sh - cd /paddle/build + docker run -it -v $PWD:/paddle -w /paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=ON" -e "RUN_TEST=OFF" paddlepaddle/paddle_manylinux_devel:cuda8.0_cudnn5 /bin/bash + ./paddle/scripts/paddle_build.sh build + cd build ctest -R test_sum_op -V .. _faq_docker: Frequently Asked Questions ----------------- +--------------------------- - What is Docker? @@ -116,11 +116,10 @@ Frequently Asked Questions Many PaddlePaddle developers are using Emacs. They add the following few lines into their `~/.emacs` configure file: - ```emacs - (global-set-key "\C-cc" 'compile) - (setq compile-command - "docker run --rm -it -v $(git rev-parse --show-toplevel):/paddle paddle:dev") - ``` + .. code-block:: emacs + + (global-set-key "\C-cc" 'compile) + (setq compile-command "docker run --rm -it -v $(git rev-parse --show-toplevel):/paddle paddle:dev") so they could type `Ctrl-C` and `c` to build PaddlePaddle from source. @@ -143,7 +142,7 @@ Frequently Asked Questions .. _compile_deps: Appendix: Compile Dependencies ----------------- +------------------------------- PaddlePaddle need the following dependencies when compiling, other dependencies will be downloaded automatically. @@ -164,11 +163,11 @@ will be downloaded automatically. .. _build_options: Appendix: Build Options ----------------- +------------------------- Build options include whether build binaries for CPU or GPU, which BLAS library to use etc. You may pass these settings when running cmake. -For detailed cmake tutorial please refer to `here `_ 。 +For detailed cmake tutorial please refer to `here `__ 。 You can add :code:`-D` argument to pass such options, like: @@ -217,7 +216,7 @@ keep on with latest cuDNN versions. Be sure to run with the same version of cuDN you built. Pass Compile Options -++++++++++++++ +++++++++++++++++++++++ You can pass compile options to use intended BLAS/CUDA/Cudnn libraries. When running cmake command, it will search system paths like diff --git a/doc/v2/build_and_install/docker_install_cn.rst b/doc/v2/build_and_install/docker_install_cn.rst index 79d214635a069a739060e0b79424729f6ff90387..106c86bace075764c84bc2a7f7cb09d466fa8794 100644 --- a/doc/v2/build_and_install/docker_install_cn.rst +++ b/doc/v2/build_and_install/docker_install_cn.rst @@ -73,6 +73,7 @@ 当然，您也可以进入到Docker容器中，以交互式的方式执行或调试您的代码： .. code-block:: bash + docker run -it -v $PWD:/work paddlepaddle/paddle /bin/bash cd /work python train.py @@ -97,7 +98,7 @@ PaddlePaddle Book是为用户和开发者制作的一个交互式的Jupyter Note 国内用户可以使用下面的镜像源来加速访问： - .. code-block: bash + .. code-block:: bash docker run -p 8888:8888 docker.paddlepaddlehub.com/book diff --git a/doc/v2/build_and_install/docker_install_en.rst b/doc/v2/build_and_install/docker_install_en.rst index e0e0559fb858a093db96a9b4ec1c5a45d6c71a38..25aecb8d0da9feb00006da6259b529b7011d91cb 100644 --- a/doc/v2/build_and_install/docker_install_en.rst +++ b/doc/v2/build_and_install/docker_install_en.rst @@ -80,6 +80,7 @@ Also, you can go into the container shell, run or debug your code interactively: .. code-block:: bash + docker run -it -v $PWD:/work paddlepaddle/paddle /bin/bash cd /work python train.py @@ -104,7 +105,7 @@ We provide a packaged book image, simply issue the command: For users in China, we provide a faster mirror: - .. code-block: bash + .. code-block:: bash docker run -p 8888:8888 docker.paddlepaddlehub.com/book diff --git a/doc/v2/build_and_install/index_cn.rst b/doc/v2/build_and_install/index_cn.rst index e079bb661f3a5141a09dfbc6893d1bf945697bc9..1a9305ac4b6578c14a962f223c647a71e3b8a72b 100644 --- a/doc/v2/build_and_install/index_cn.rst +++ b/doc/v2/build_and_install/index_cn.rst @@ -6,7 +6,7 @@ PaddlePaddle针对不同的用户群体提供了多种安装方式。 专注深度学习模型开发 ------------------ +-------------------- PaddlePaddle提供了多种python wheel包，可通过pip一键安装： @@ -18,7 +18,7 @@ PaddlePaddle提供了多种python wheel包，可通过pip一键安装： 这是最便捷的安装方式，请根据机器配置和系统选择对应的安装包。 关注底层框架 ----------- +------------- PaddlePaddle提供了基于Docker的安装方式，请参照以下教程： @@ -45,7 +45,7 @@ PaddlePaddle提供了基于Docker的安装方式，请参照以下教程： 常见问题汇总 ------------ +-------------- 如果在安装过程中遇到了问题，请先尝试在下面的页面寻找答案： diff --git a/doc/v2/build_and_install/index_en.rst b/doc/v2/build_and_install/index_en.rst index 5b3de0f8c3e5496060646b5ddb080d0d338a8bfa..7990bacbd6966e88e8763e9c5709e410f7e9fed4 100644 --- a/doc/v2/build_and_install/index_en.rst +++ b/doc/v2/build_and_install/index_en.rst @@ -1,12 +1,12 @@ install and Compile -========== +====================== .. _install_steps: PaddlePaddle provides various methods of installation for many different users Focus on Deep Learning Model Development ------------------ +---------------------------------------- PaddlePaddle provides lots of packages of python wheel , that pip can install: @@ -18,7 +18,7 @@ PaddlePaddle provides lots of packages of python wheel , that pip can install: This is the most convenient way of installation. Please choose the right installation package with machine configure and system. Follow the Bottom Frame ----------- +------------------------ PaddlePaddle also supports installation using Docker. Please refer to the tutorial below: diff --git a/doc/v2/build_and_install/pip_install_cn.rst b/doc/v2/build_and_install/pip_install_cn.rst index b3d882743785e8ee301b71b696230531d2b7ba58..853bdb21bbcf07ae1742d2196dbcfe4668828b7b 100644 --- a/doc/v2/build_and_install/pip_install_cn.rst +++ b/doc/v2/build_and_install/pip_install_cn.rst @@ -10,20 +10,38 @@ PaddlePaddle可以使用常用的Python包管理工具 使用pip安装 ------------------------------ - -执行下面的命令即可在当前机器上安装PaddlePaddle的运行时环境，并自动下载安装依赖软件，版本为cpu_avx_openblas。 +执行下面的命令即可在当前机器上安装PaddlePaddle的运行时环境，并自动下载安装依赖软件。 .. code-block:: bash pip install paddlepaddle +当前的默认版本为0.12.0，cpu_avx_openblas，您可以通过指定版本号来安装其它版本，例如: + + .. code-block:: bash + + pip install paddlepaddle==0.11.0 + -如果需要安装支持GPU的版本（cuda7.5_cudnn5_avx_openblas），需要执行： +如果需要安装支持GPU的版本（cuda8.0_cudnn5_avx_openblas），需要执行： .. code-block:: bash pip install paddlepaddle-gpu +当前的默认版本也是0.12.0，PaddlePaddle针对不同需求提供了更多版本的安装包，部分列表如下： + +================================= ======================================== +版本号 版本说明 +================================= ======================================== +paddlepaddle-gpu==0.12.0 使用CUDA 8.0和cuDNN 5编译的0.12.0版本 +paddlepaddle-gpu==0.11.0.post87 使用CUDA 8.0和cuDNN 7编译的0.11.0版本 +paddlepaddle-gpu==0.11.0.post8 使用CUDA 8.0和cuDNN 5编译的0.11.0版本 +paddlepaddle-gpu==0.11.0 使用CUDA 7.5和cuDNN 5编译的0.11.0版本 +================================= ======================================== + +您可以在 `Release History `_ 中找到paddlepaddle-gpu的各个发行版本。 + 如果需要获取并安装最新的（开发分支）PaddlePaddle，可以从我们的CI系统中下载最新的whl安装包和c-api开发包并安装， 您可以从下面的表格中找到需要的版本： @@ -37,12 +55,11 @@ PaddlePaddle可以使用常用的Python包管理工具 :header: "版本说明", "cp27-cp27mu", "cp27-cp27m" :widths: 1, 3, 3 - "cpu_avx_mkl", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl `_" - "cpu_avx_openblas", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl `_" - "cpu_noavx_openblas", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl `_" - "cuda7.5_cudnn5_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl `_" - "cuda8.0_cudnn5_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl `_" - "cuda8.0_cudnn7_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl `_" + "cpu_avx_mkl", "`paddlepaddle-latest-cp27-cp27mu-linux_x86_64.whl `__", "`paddlepaddle-latest-cp27-cp27m-linux_x86_64.whl `__" + "cpu_avx_openblas", "`paddlepaddle-latest-cp27-cp27mu-linux_x86_64.whl `__", "`paddlepaddle-latest-cp27-cp27m-linux_x86_64.whl `__" + "cpu_noavx_openblas", "`paddlepaddle-latest-cp27-cp27mu-linux_x86_64.whl `__", "`paddlepaddle-latest-cp27-cp27m-linux_x86_64.whl `_" + "cuda8.0_cudnn5_avx_mkl", "`paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl `__", "`paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl `__" + "cuda8.0_cudnn7_avx_mkl", "`paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl `__", "`paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl `__" .. _pip_dependency: @@ -69,7 +86,7 @@ PaddlePaddle发布的安装包会尽量对齐 `manylinux1 9.0.0) 才可以安装。可以使用下面的命令更新您的pip： .. code-block:: bash diff --git a/doc/v2/build_and_install/pip_install_en.rst b/doc/v2/build_and_install/pip_install_en.rst index 1e409d86b9775094998f72f92954f4bbc1013ea1..fecf6d3712feac3265100a6121901ba784f7d5cc 100644 --- a/doc/v2/build_and_install/pip_install_en.rst +++ b/doc/v2/build_and_install/pip_install_en.rst @@ -12,20 +12,38 @@ Install using pip ------------------------------ Run the following command to install PaddlePaddle on the current -machine, it will also download requirements, the version is cpu_avx_openblas. +machine, it will also download requirements. .. code-block:: bash pip install paddlepaddle +the default version is 0.12.0, cpu_avx_openblas, you can specify the versions to satisfy your demands, like: -If you wish to install GPU version (cuda7.5_cudnn5_avx_openblas), just run: + .. code-block:: bash + + pip install paddlepaddle==0.11.0 + +If you need to install a GPU-enabled version (cuda8.0_cudnn5_avx_openblas), you need to run: .. code-block:: bash pip install paddlepaddle-gpu -If you wish to install the latest develop branch PaddlePaddle, +The default version is also 0.12.0, PaddlePaddle provides several versions of packages for different needs, as shown in the table: + +================================= ======================================== +版本号 版本说明 +================================= ======================================== +paddlepaddle-gpu==0.12.0 0.12.0 built with CUDA 8.0 and cuDNN 5 +paddlepaddle-gpu==0.11.0.post87 0.11.0 built with CUDA 8.0 and cuDNN 7 +paddlepaddle-gpu==0.11.0.post8 0.11.0 built with CUDA 8.0 and cuDNN 5 +paddlepaddle-gpu==0.11.0 0.11.0 built with CUDA 7.5 and cuDNN 5 +================================= ======================================== + +You can find all versions released of paddlepaddle-gpu in `Release History `_ . + +If you wish to install the latest develop branch PaddlePaddle, you can download the latest whl package from our CI system. Access the below links, log in as guest, then click at the "Artifact" tab, you'll find the download link of whl packages. @@ -40,12 +58,11 @@ If the links below shows up the login form, just click "Log in as guest" to star :header: "version", "cp27-cp27mu", "cp27-cp27m" :widths: 1, 3, 3 - "cpu_avx_mkl", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl `_" - "cpu_avx_openblas", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl `_" - "cpu_noavx_openblas", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl `_" - "cuda7.5_cudnn5_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl `_" - "cuda8.0_cudnn5_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl `_" - "cuda8.0_cudnn7_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl `_" + "cpu_avx_mkl", "`paddlepaddle-latest-cp27-cp27mu-linux_x86_64.whl `__", "`paddlepaddle-latest-cp27-cp27m-linux_x86_64.whl `__" + "cpu_avx_openblas", "`paddlepaddle-latest-cp27-cp27mu-linux_x86_64.whl `__", "`paddlepaddle-latest-cp27-cp27m-linux_x86_64.whl `__" + "cpu_noavx_openblas", "`paddlepaddle-latest-cp27-cp27mu-linux_x86_64.whl `__", "`paddlepaddle-latest-cp27-cp27m-linux_x86_64.whl `__" + "cuda8.0_cudnn5_avx_mkl", "`paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl `__", "`paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl `__" + "cuda8.0_cudnn7_avx_mkl", "`paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl `__", "`paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl `__" .. _pip_dependency: @@ -79,7 +96,7 @@ FAQ ------------------------------ - paddlepaddle*.whl is not a supported wheel on this platform. - + The main cause of this issue is that your current platform is not supported. Please check that you are using Python 2.7 series. Besides, pypi only supports manylinux1 standard, you'll need to diff --git a/doc/v2/design/mkl/mkldnn.md b/doc/v2/design/mkl/mkldnn.md index 1bd2e7bc34ee79eb753b3520d97e5e7beca89b0b..bd5bcf6f67168c21cebb046a629b948d1661e75c 100644 --- a/doc/v2/design/mkl/mkldnn.md +++ b/doc/v2/design/mkl/mkldnn.md @@ -5,7 +5,7 @@ 充分展现英特尔平台的优势，有效提升PaddlePaddle在英特尔架构上的性能。
-
+
Figure 1. PaddlePaddle on IA
@@ -42,16 +42,43 @@ Figure 1. PaddlePaddle on IA MKL，MKLML以及MKL-DNN三者关系如下表： -| Name | Open Source | License | Descriptions | -| :---------- | :--------------- | :---------- | :------------ | -| MKL | No | Proprietary | Accelerate math processing routines | -| MKLML | No | Proprietary | Small package of MKL, especially for Machine Learning | -| MKL-DNN | Yes | Apache 2.0 | Accelerate primitives processing routines especially for Deep Neural Networks | + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
NameOpen SourceLicenseDescriptions
MKLNoProprietaryAccelerate math processing routines
MKLMLNoProprietarySmall package of MKL, especially for Machine Learning
MKL-DNNYesApache 2.0Accelerate primitives processing routines especially for Deep Neural Networks
MKLML可以与MKL-DNN共同使用，以此达到最好的性能。
-
+
Figure 2. PaddlePaddle with MKL Engines
@@ -103,7 +130,7 @@ MKL-DNN的库目前只有动态库`libmkldnn.so`。 所以我们定义了一个`MKLDNNMatrix`用于管理MKL-DNN数据的不同格式以及相互之间的转换。
-
+
Figure 3. MKLDNNMatrix
@@ -113,7 +140,7 @@ Figure 3. MKLDNNMatrix 子类只需要使用定义好的接口，实现具体的函数功能即可。
-
+
Figure 4. MKLDNNLayer
@@ -150,7 +177,7 @@ Figure 4. MKLDNNLayer 所以整体上，在实现每个子类的时候就不需要关心分支的事情了。
-
+
Figure 5. Merge Gradients
diff --git a/doc/v2/dev/contribute_to_paddle_cn.md b/doc/v2/dev/contribute_to_paddle_cn.md index d8bf093e09b53b302225739fa67146adc7976e4b..add06e42f1bbd221b48eb83e4e84d4a7c89e7483 100644 --- a/doc/v2/dev/contribute_to_paddle_cn.md +++ b/doc/v2/dev/contribute_to_paddle_cn.md @@ -51,6 +51,8 @@ Paddle 开发人员使用 [pre-commit](http://pre-commit.com/) 工具来管理 G Paddle 使用 `clang-format` 来调整 C/C++ 源代码格式，请确保 `clang-format` 版本在 3.8 以上。 +注：通过`pip install pre-commit`和`conda install -c conda-forge pre-commit`安装的`yapf`稍有不同的，Paddle 开发人员使用的是`pip install pre-commit`。 + ## 开始开发 在本例中，我删除了 README.md 中的一行，并创建了一个新文件。 diff --git a/doc/v2/howto/capi/workflow_of_capi_cn.md b/doc/v2/howto/capi/workflow_of_capi_cn.md index 1968c1099ac5734cd68b437f2f7aa428d7b5265e..3acdbae28e9b35f8a9104a89c9a5799f8c892334 100644 --- a/doc/v2/howto/capi/workflow_of_capi_cn.md +++ b/doc/v2/howto/capi/workflow_of_capi_cn.md @@ -59,7 +59,7 @@ 代码示例如下： ```python - from paddle.utils.merge_model import merge_v2_modelss + from paddle.utils.merge_model import merge_v2_model from mnist_v2 import network net = network(is_infer=True) diff --git a/doc/v2/howto/cluster/multi_cluster/index_en.rst b/doc/v2/howto/cluster/multi_cluster/index_en.rst index b69bd5b2dbf1967d65558da06812d76f431c1d5a..9bc1eb2e3796d95dd69b165e916e263ea34b87f6 100644 --- a/doc/v2/howto/cluster/multi_cluster/index_en.rst +++ b/doc/v2/howto/cluster/multi_cluster/index_en.rst @@ -8,28 +8,28 @@ The user's cluster environment is not the same. To facilitate everyone's deploym .. toctree:: :maxdepth: 1 - k8s_cn.md - k8s_distributed_cn.md + k8s_en.md + k8s_distributed_en.md `OpenMPI `_ is a mature high-performance parallel computing framework, which is widely used in the field of HPC. The following guide describes how to use OpenMPI to build PaddlePaddle's cluster training task: .. toctree:: :maxdepth: 1 - openmpi_cn.md + openmpi_en.md `Fabric `_ is a convenient tool for program deployment and management. We provide a way to deploy and manage with Fabric. If you want to know more about it, please read the following guidelines: .. toctree:: :maxdepth: 1 - fabric_cn.md + fabric_en.md We also support the deployment of PaddlePaddle on AWS. Learn more about: .. toctree:: :maxdepth: 1 - k8s_aws_cn.md + k8s_aws_en.md -The examples can be found under `cluster_train_v2 `_ . \ No newline at end of file +The examples can be found under `cluster_train_v2 `_ . diff --git a/doc/v2/howto/cluster/multi_cluster/k8s_distributed_en.md b/doc/v2/howto/cluster/multi_cluster/k8s_distributed_en.md index dee1b7554f97af17989c3f7739d8feea3b6b8e3f..b2dc4da8451af317df76c5b3df328b6f58429610 100644 --- a/doc/v2/howto/cluster/multi_cluster/k8s_distributed_en.md +++ b/doc/v2/howto/cluster/multi_cluster/k8s_distributed_en.md @@ -41,7 +41,7 @@ Training docker image needs to package the paddle pserver and paddle trainer run - Generating the initialization arguments for `Paddle PServer` and `Paddle Training` processes. Since the paddlepaddle official docker image already has the runtimes we need, we'll take it as the base image and pack some additional scripts for the processes mentioned above to build our training image. for more detail, please find from the following link: -- https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/usage/cluster/src/k8s_train/Dockerfile +- https://github.com/PaddlePaddle/Paddle/tree/develop/doc/v2/howto/cluster/multi_cluster/src/k8s_train/Dockerfile ```bash @@ -62,7 +62,7 @@ represent the Docker Image which built in this step. ### Prepare Training Data We can download and split the training job by creating a Kubernetes Job, or custom your image -by editing [k8s_train](./src/k8s_train/). +by editing [k8s_train](https://github.com/PaddlePaddle/Paddle/tree/develop/doc/v2/howto/cluster/multi_cluster/src/k8s_train). Before creating a Job, we need to bind a [persistenVolumeClaim](https://kubernetes.io/docs/user-guide/persistent-volumes) by the different type of the different file system, the generated dataset would be saved on this volume. diff --git a/doc/v2/images/FullyConnected.jpg b/doc/v2/images/FullyConnected.jpg new file mode 100644 index 0000000000000000000000000000000000000000..b2241f401434e527f95ee4e0e541a3f2ff78fd1e Binary files /dev/null and b/doc/v2/images/FullyConnected.jpg differ diff --git a/doc/v2/images/add_security_group.png b/doc/v2/images/add_security_group.png new file mode 100644 index 0000000000000000000000000000000000000000..bd34f46c9b0ada7027fd53e553e7d033255d25fc Binary files /dev/null and b/doc/v2/images/add_security_group.png differ diff --git a/doc/v2/images/bi_lstm.jpg b/doc/v2/images/bi_lstm.jpg new file mode 100644 index 0000000000000000000000000000000000000000..adec1606d64d6e35ffe7e62abfa9a09309b05c84 Binary files /dev/null and b/doc/v2/images/bi_lstm.jpg differ diff --git a/doc/v2/images/checkpointing.png b/doc/v2/images/checkpointing.png new file mode 100644 index 0000000000000000000000000000000000000000..c221e8474f90f37e31416cbb19c9452207a0d14c Binary files /dev/null and b/doc/v2/images/checkpointing.png differ diff --git a/doc/v2/images/create_efs.png b/doc/v2/images/create_efs.png new file mode 100644 index 0000000000000000000000000000000000000000..e5f1526033d1daf401700989af1d25919bcb7675 Binary files /dev/null and b/doc/v2/images/create_efs.png differ diff --git a/doc/v2/images/csr.png b/doc/v2/images/csr.png new file mode 100644 index 0000000000000000000000000000000000000000..3dc10b8de4f6d3f517624956b1694b689405a031 Binary files /dev/null and b/doc/v2/images/csr.png differ diff --git a/doc/v2/images/data_dispatch.png b/doc/v2/images/data_dispatch.png new file mode 100644 index 0000000000000000000000000000000000000000..5bdcc24d6a6d193cb014f8c38b362451fded5e54 Binary files /dev/null and b/doc/v2/images/data_dispatch.png differ diff --git a/doc/v2/images/dataset.graffle b/doc/v2/images/dataset.graffle new file mode 100644 index 0000000000000000000000000000000000000000..c10a423ed16a23229a9ee33d11bfc82bb59646c8 Binary files /dev/null and b/doc/v2/images/dataset.graffle differ diff --git a/doc/v2/images/dataset.png b/doc/v2/images/dataset.png new file mode 100644 index 0000000000000000000000000000000000000000..2fb7f1cce3b6dd21489392557826e95a9f207c34 Binary files /dev/null and b/doc/v2/images/dataset.png differ diff --git a/doc/v2/images/doc_en.png b/doc/v2/images/doc_en.png new file mode 100644 index 0000000000000000000000000000000000000000..ed6b9178fba91a3bdf45ae797a9924f84146fbc8 Binary files /dev/null and b/doc/v2/images/doc_en.png differ diff --git a/doc/v2/images/efs_mount.png b/doc/v2/images/efs_mount.png new file mode 100644 index 0000000000000000000000000000000000000000..0f9e3cab98445707e5e9baa18ddabe15cdf04576 Binary files /dev/null and b/doc/v2/images/efs_mount.png differ diff --git a/doc/v2/images/encoder-decoder-attention-model.png b/doc/v2/images/encoder-decoder-attention-model.png new file mode 100644 index 0000000000000000000000000000000000000000..79f911d4ba12ac0c0d1a936c9df639c302786914 Binary files /dev/null and b/doc/v2/images/encoder-decoder-attention-model.png differ diff --git a/doc/v2/images/engine.png b/doc/v2/images/engine.png new file mode 100644 index 0000000000000000000000000000000000000000..1f5f65c2cc765a514a3ba9e7b7f468e1dc4b0c3b Binary files /dev/null and b/doc/v2/images/engine.png differ diff --git a/doc/v2/images/file_storage.graffle b/doc/v2/images/file_storage.graffle new file mode 100644 index 0000000000000000000000000000000000000000..50a17e70fa255495337c529a3bf12a5c0024a5be Binary files /dev/null and b/doc/v2/images/file_storage.graffle differ diff --git a/doc/v2/images/file_storage.png b/doc/v2/images/file_storage.png new file mode 100644 index 0000000000000000000000000000000000000000..fccb4e3e7e738224c7f1584326bd5f351ce799aa Binary files /dev/null and b/doc/v2/images/file_storage.png differ diff --git a/doc/v2/images/glossary_rnn.dot b/doc/v2/images/glossary_rnn.dot new file mode 100644 index 0000000000000000000000000000000000000000..2cd0fb1820c44b0e8e0b869f9d39fcad27efa758 --- /dev/null +++ b/doc/v2/images/glossary_rnn.dot @@ -0,0 +1,42 @@ +digraph G{ + subgraph cluster_timestep0 { + label="recurrent timestep i-1" + bgcolor=lightgray + node [style=filled,color=white] + fc0_0 [label="fc 0"] + fc0_1 [label="fc 1"] + fc0_2 [label="fc 2"] + + fc0_0 -> fc0_1 + fc0_1 -> fc0_2 + } + + subgraph cluster_timestep1 { + label="recurrent timestep i" + node [style=filled]; + fc1_0 [label="fc 0"] + fc1_1 [label="fc 1"] + fc1_2 [label="fc 2"] + color=blue + + fc1_0 -> fc1_1 + fc1_1 -> fc1_2 + } + + subgraph cluster_timestep2 { + label="recurrent timestep i+1" + bgcolor=lightgray + node [style=filled,color=white] + fc2_0 [label="fc 0"] + fc2_1 [label="fc 1"] + fc2_2 [label="fc 2"] + + fc2_0 -> fc2_1 + fc2_1 -> fc2_2 + } + + + fc0_1 -> fc1_1 [style="dotted" constraint=false] + fc1_1 -> fc2_1 [style="dotted" constraint=false] + +} \ No newline at end of file diff --git a/doc/v2/images/glossary_rnn_with_memory.dot b/doc/v2/images/glossary_rnn_with_memory.dot new file mode 100644 index 0000000000000000000000000000000000000000..0f101ec2d8f15aec76c57f328046b6b55cf0c7eb --- /dev/null +++ b/doc/v2/images/glossary_rnn_with_memory.dot @@ -0,0 +1,48 @@ +digraph G{ + subgraph cluster_timestep0 { + label="recurrent timestep i-1" + bgcolor=lightgray + node [style=filled,color=white] + fc0_0 [label="fc 0"] + fc0_1 [label="fc 1"] + fc0_2 [label="fc 2"] + m0 [label="memory"] + fc0_0 -> fc0_1 + fc0_1 -> fc0_2 + fc0_1 -> m0 + m0 -> fc0_1 + } + + subgraph cluster_timestep1 { + label="recurrent timestep i" + node [style=filled]; + fc1_0 [label="fc 0"] + fc1_1 [label="fc 1"] + fc1_2 [label="fc 2"] + m1 [label="memory"] + color=blue + fc1_0 -> fc1_1 + fc1_1 -> fc1_2 + fc1_1 -> m1 + m1 -> fc1_1 + } + + subgraph cluster_timestep2 { + label="recurrent timestep i+1" + bgcolor=lightgray + node [style=filled,color=white] + fc2_0 [label="fc 0"] + fc2_1 [label="fc 1"] + fc2_2 [label="fc 2"] + m2 [label="memory"] + fc2_0 -> fc2_1 + fc2_1 -> fc2_2 + fc2_1 -> m2 + m2 -> fc2_1 + } + + + m0 -> m1 [style="dotted" constraint=false] + m1 -> m2 [style="dotted" constraint=false] + +} \ No newline at end of file diff --git a/doc/v2/images/gradients.png b/doc/v2/images/gradients.png new file mode 100644 index 0000000000000000000000000000000000000000..f031bcf8e4cec14e63075b8b9d2c7bbd9f1b1a3c Binary files /dev/null and b/doc/v2/images/gradients.png differ diff --git a/doc/v2/images/init_lock.graffle b/doc/v2/images/init_lock.graffle new file mode 100644 index 0000000000000000000000000000000000000000..fa9149f21b1311eed48ef72ec55e556559d0fc94 Binary files /dev/null and b/doc/v2/images/init_lock.graffle differ diff --git a/doc/v2/images/init_lock.png b/doc/v2/images/init_lock.png new file mode 100644 index 0000000000000000000000000000000000000000..92404ee6d6c0f9a7727952bae3c869ba338ecd7f Binary files /dev/null and b/doc/v2/images/init_lock.png differ diff --git a/doc/v2/images/k8s-paddle-arch.png b/doc/v2/images/k8s-paddle-arch.png new file mode 100644 index 0000000000000000000000000000000000000000..b3800c4fe81302d35e49f7dbacb9221c4dfa5cde Binary files /dev/null and b/doc/v2/images/k8s-paddle-arch.png differ diff --git a/doc/v2/images/layers.png b/doc/v2/images/layers.png new file mode 100644 index 0000000000000000000000000000000000000000..306f79b7a844610915eb8944128f57d2b7a3065a Binary files /dev/null and b/doc/v2/images/layers.png differ diff --git a/doc/v2/images/managed_policy.png b/doc/v2/images/managed_policy.png new file mode 100644 index 0000000000000000000000000000000000000000..c7ecda555b81d7750e9292a9ab72d2f517f76a2a Binary files /dev/null and b/doc/v2/images/managed_policy.png differ diff --git a/doc/v2/images/matrix.png b/doc/v2/images/matrix.png new file mode 100644 index 0000000000000000000000000000000000000000..c33ce9cf0335e47cc8c1253304d0fe179186e6f2 Binary files /dev/null and b/doc/v2/images/matrix.png differ diff --git a/doc/v2/images/nvvp1.png b/doc/v2/images/nvvp1.png new file mode 100644 index 0000000000000000000000000000000000000000..1af23ac3c52929b2b0645d2f9fa4d4c6db1f6e77 Binary files /dev/null and b/doc/v2/images/nvvp1.png differ diff --git a/doc/v2/images/nvvp2.png b/doc/v2/images/nvvp2.png new file mode 100644 index 0000000000000000000000000000000000000000..177c9db708da6863d1075f3e615f5962dbe18b29 Binary files /dev/null and b/doc/v2/images/nvvp2.png differ diff --git a/doc/v2/images/nvvp3.png b/doc/v2/images/nvvp3.png new file mode 100644 index 0000000000000000000000000000000000000000..d8f393667d6569b6f1e61ffccac43fae5888b6db Binary files /dev/null and b/doc/v2/images/nvvp3.png differ diff --git a/doc/v2/images/nvvp4.png b/doc/v2/images/nvvp4.png new file mode 100644 index 0000000000000000000000000000000000000000..51f2f3e183295de6cf8ddaf2b3b8a0862aa35f01 Binary files /dev/null and b/doc/v2/images/nvvp4.png differ diff --git a/doc/v2/images/overview.png b/doc/v2/images/overview.png new file mode 100644 index 0000000000000000000000000000000000000000..8fb7bbb9dd654bf363d701d0c8cd4a557043d188 Binary files /dev/null and b/doc/v2/images/overview.png differ diff --git a/doc/v2/images/paddle-cloud-in-data-center.png b/doc/v2/images/paddle-cloud-in-data-center.png new file mode 100644 index 0000000000000000000000000000000000000000..da5d1a77562480ad1d886f5f21dbd84001d3d508 Binary files /dev/null and b/doc/v2/images/paddle-cloud-in-data-center.png differ diff --git a/doc/v2/images/paddle-etcd.graffle b/doc/v2/images/paddle-etcd.graffle new file mode 100644 index 0000000000000000000000000000000000000000..f973dc9b9dbf72e9bc31e2d32822916cd281f8d9 Binary files /dev/null and b/doc/v2/images/paddle-etcd.graffle differ diff --git a/doc/v2/images/paddle-etcd.png b/doc/v2/images/paddle-etcd.png new file mode 100644 index 0000000000000000000000000000000000000000..57981ceb4b94f0f7d6dfa63f3d28c0402bf9cc31 Binary files /dev/null and b/doc/v2/images/paddle-etcd.png differ diff --git a/doc/v2/images/paddle-model-sharding.graffle b/doc/v2/images/paddle-model-sharding.graffle new file mode 100644 index 0000000000000000000000000000000000000000..fba30f0ca2b47f0d202a432821d95e55aac37ec8 Binary files /dev/null and b/doc/v2/images/paddle-model-sharding.graffle differ diff --git a/doc/v2/images/paddle-model-sharding.png b/doc/v2/images/paddle-model-sharding.png new file mode 100644 index 0000000000000000000000000000000000000000..8c3f6724ef46c6527e63a4cd8cb0b50fe0167124 Binary files /dev/null and b/doc/v2/images/paddle-model-sharding.png differ diff --git a/doc/v2/images/paddle-ps-0.png b/doc/v2/images/paddle-ps-0.png new file mode 100644 index 0000000000000000000000000000000000000000..47ef32806f182cab003da77f1556823b3f6d1721 Binary files /dev/null and b/doc/v2/images/paddle-ps-0.png differ diff --git a/doc/v2/images/paddle-ps-1.png b/doc/v2/images/paddle-ps-1.png new file mode 100644 index 0000000000000000000000000000000000000000..f3125db73096c52bac6e7c60e1675552857c0774 Binary files /dev/null and b/doc/v2/images/paddle-ps-1.png differ diff --git a/doc/v2/images/paddle-ps.graffle b/doc/v2/images/paddle-ps.graffle new file mode 100644 index 0000000000000000000000000000000000000000..0e536ffdd91cd696008b4c01bad3cb53edebdc16 Binary files /dev/null and b/doc/v2/images/paddle-ps.graffle differ diff --git a/doc/v2/images/paddle-task-queues.graffle b/doc/v2/images/paddle-task-queues.graffle new file mode 100644 index 0000000000000000000000000000000000000000..4263ed8bfd2ef0e55058828bf23f2fac3595e5fd Binary files /dev/null and b/doc/v2/images/paddle-task-queues.graffle differ diff --git a/doc/v2/images/paddle-task-queues.png b/doc/v2/images/paddle-task-queues.png new file mode 100644 index 0000000000000000000000000000000000000000..5f980266795776752cebd0c346b85c4a75a47780 Binary files /dev/null and b/doc/v2/images/paddle-task-queues.png differ diff --git a/doc/v2/images/paddle-task-states.graffle b/doc/v2/images/paddle-task-states.graffle new file mode 100644 index 0000000000000000000000000000000000000000..cf1a0b9246d9386a949d2dbb8c32fe84f72eea83 Binary files /dev/null and b/doc/v2/images/paddle-task-states.graffle differ diff --git a/doc/v2/images/paddle-task-states.png b/doc/v2/images/paddle-task-states.png new file mode 100644 index 0000000000000000000000000000000000000000..4ae43cb66c071aee9eb90d875e2373b29af9c3e0 Binary files /dev/null and b/doc/v2/images/paddle-task-states.png differ diff --git a/doc/v2/images/ps_cn.png b/doc/v2/images/ps_cn.png new file mode 100644 index 0000000000000000000000000000000000000000..f9525739cc8bc6506adde642aafa0a85ae3ebebc Binary files /dev/null and b/doc/v2/images/ps_cn.png differ diff --git a/doc/v2/images/ps_en.png b/doc/v2/images/ps_en.png new file mode 100644 index 0000000000000000000000000000000000000000..6537d3d56589ca9f19a77a50a970e4b5275e6ce0 Binary files /dev/null and b/doc/v2/images/ps_en.png differ diff --git a/doc/v2/images/pserver_and_trainer.png b/doc/v2/images/pserver_and_trainer.png new file mode 100644 index 0000000000000000000000000000000000000000..f41fe48920590333ad332bb51eb18e03dc251541 Binary files /dev/null and b/doc/v2/images/pserver_and_trainer.png differ diff --git a/doc/v2/images/pserver_init.graffle b/doc/v2/images/pserver_init.graffle new file mode 100644 index 0000000000000000000000000000000000000000..5f3f1f52be8aa7f9049a8fcd6b7c93c8560c1676 Binary files /dev/null and b/doc/v2/images/pserver_init.graffle differ diff --git a/doc/v2/images/pserver_init.png b/doc/v2/images/pserver_init.png new file mode 100644 index 0000000000000000000000000000000000000000..dfe491ff98dd7db1c336093c80964a260df2cd90 Binary files /dev/null and b/doc/v2/images/pserver_init.png differ diff --git a/doc/v2/images/route53_create_recordset.png b/doc/v2/images/route53_create_recordset.png new file mode 100644 index 0000000000000000000000000000000000000000..34e476c7beac30fcdde13fccc4cc8d08b4be3d35 Binary files /dev/null and b/doc/v2/images/route53_create_recordset.png differ diff --git a/doc/v2/images/route53_create_zone.png b/doc/v2/images/route53_create_zone.png new file mode 100644 index 0000000000000000000000000000000000000000..25b7ddb831c5cba97f4b2edddd27da3234d621af Binary files /dev/null and b/doc/v2/images/route53_create_zone.png differ diff --git a/doc/v2/images/sequence_data.png b/doc/v2/images/sequence_data.png new file mode 100644 index 0000000000000000000000000000000000000000..6e47a46b8955dfe977e85898fe3c9f33ed28de7e Binary files /dev/null and b/doc/v2/images/sequence_data.png differ diff --git a/doc/v2/images/simple_full_hierarchical_recurrent.dot b/doc/v2/images/simple_full_hierarchical_recurrent.dot new file mode 100644 index 0000000000000000000000000000000000000000..ff278a0323bb2c3ef07bf6f016a3a8df05783581 --- /dev/null +++ b/doc/v2/images/simple_full_hierarchical_recurrent.dot @@ -0,0 +1,30 @@ +digraph G { + rankdir=LR; + + subgraph cluster_t0 { + a [label="4"] + b [label="5"] + c [label="2"] + } + + subgraph cluster_t1 { + d [label="0"] + e [label="9"] + } + + subgraph cluster_t2 { + f [label="8"] + g [label="1"] + h [label="4"] + } + + a -> b; + b -> c; + c -> d [constraint=false]; + + d -> e; + e -> f [constraint=false]; + + f -> g; + g -> h; +} \ No newline at end of file diff --git a/doc/v2/images/simple_full_recurrent.dot b/doc/v2/images/simple_full_recurrent.dot new file mode 100644 index 0000000000000000000000000000000000000000..cee281fbac993afbd0cc3416570f95965cdf0a59 --- /dev/null +++ b/doc/v2/images/simple_full_recurrent.dot @@ -0,0 +1,19 @@ +digraph G { + rankdir=LR; + a [label="4"] + b [label="5"] + c [label="2"] + d [label="0"] + e [label="9"] + f [label="8"] + g [label="1"] + h [label="4"] + + a -> b; + b -> c; + c -> d; + d -> e; + e -> f; + f -> g; + g -> h; +} \ No newline at end of file diff --git a/doc/v2/images/submit-job.graffle b/doc/v2/images/submit-job.graffle new file mode 100644 index 0000000000000000000000000000000000000000..677cdfb6d9a32168bf71729eb841fa1ca0dd31d6 Binary files /dev/null and b/doc/v2/images/submit-job.graffle differ diff --git a/doc/v2/images/submit-job.png b/doc/v2/images/submit-job.png new file mode 100644 index 0000000000000000000000000000000000000000..3046a460a7ba708079e88a560debaa215a694680 Binary files /dev/null and b/doc/v2/images/submit-job.png differ diff --git a/doc/v2/images/trainer.graffle b/doc/v2/images/trainer.graffle new file mode 100644 index 0000000000000000000000000000000000000000..43415ed8cf61a5acfa34f8e56b9577f338dbf254 Binary files /dev/null and b/doc/v2/images/trainer.graffle differ diff --git a/doc/v2/images/trainer.png b/doc/v2/images/trainer.png new file mode 100644 index 0000000000000000000000000000000000000000..6537d3d56589ca9f19a77a50a970e4b5275e6ce0 Binary files /dev/null and b/doc/v2/images/trainer.png differ diff --git a/doc/v2/images/trainer_cn.png b/doc/v2/images/trainer_cn.png new file mode 100644 index 0000000000000000000000000000000000000000..f9525739cc8bc6506adde642aafa0a85ae3ebebc Binary files /dev/null and b/doc/v2/images/trainer_cn.png differ diff --git a/doc/v2/images/worker_security_group.png b/doc/v2/images/worker_security_group.png new file mode 100644 index 0000000000000000000000000000000000000000..57eb0265a34ad4223b69600d2a3dd355482e0bf5 Binary files /dev/null and b/doc/v2/images/worker_security_group.png differ diff --git a/doc/v2/images/workflow_of_CAPI.png b/doc/v2/images/workflow_of_CAPI.png new file mode 100644 index 0000000000000000000000000000000000000000..a4399ade048b3fe10d2d9c714bc34333ca068edb Binary files /dev/null and b/doc/v2/images/workflow_of_CAPI.png differ diff --git a/go/pserver/client/c/test/CMakeLists.txt b/go/pserver/client/c/test/CMakeLists.txt index 411dc50332672143d7a1f7bd0556ae86dc37f6f3..4500b1f288372ed0e2d9d383234df97ae976c60b 100644 --- a/go/pserver/client/c/test/CMakeLists.txt +++ b/go/pserver/client/c/test/CMakeLists.txt @@ -13,4 +13,3 @@ # limitations under the License. # cc_test(test_cclient SRCS test_cclient.c DEPS paddle_pserver_cclient paddle_go_optimizer) -add_style_check_target(test_cclient test_cclient.c) diff --git a/paddle/.gitignore b/paddle/.gitignore index 1c1c0c2c829f088d7e3f52ca007fcb8f33a16a36..01904aa6ef2057afee95ddd6e30cde064b06c52e 100644 --- a/paddle/.gitignore +++ b/paddle/.gitignore @@ -11,7 +11,6 @@ GTAGS *.pb.cc *.pb.h *_pb2.py -paddle_* output/ google/ Makefile diff --git a/paddle/CMakeLists.txt b/paddle/CMakeLists.txt index 8b1ca5e16548334ed0c9a6d31b88e0805304579e..d722eec1892206ac44c49e7a12d92be0c54df8c0 100644 --- a/paddle/CMakeLists.txt +++ b/paddle/CMakeLists.txt @@ -24,6 +24,6 @@ if(NOT WITH_FLUID_ONLY) endif() add_subdirectory(testing) -if(NOT MOBILE_INFERENCE AND NOT RPI) +if(NOT MOBILE_INFERENCE AND NOT RPI AND NOT WITH_C_API) add_subdirectory(fluid) endif() diff --git a/paddle/api/GradientMachine.cpp b/paddle/api/GradientMachine.cpp index a3d6f0f080abcf1f45d9bc5fbdb39bb6b6ca1553..0d9ad30de9c1f3f8f58c856a748abdc050ff8740 100644 --- a/paddle/api/GradientMachine.cpp +++ b/paddle/api/GradientMachine.cpp @@ -94,7 +94,7 @@ void UpdateCallback::apply(Parameter* p) { } class UpdateCallbackWrapper { -public: + public: explicit UpdateCallbackWrapper(const UpdateCallback& callback) : callback(const_cast(callback)) {} @@ -105,7 +105,7 @@ public: delete p; } -private: + private: UpdateCallback& callback; }; diff --git a/paddle/api/PaddleAPI.h b/paddle/api/PaddleAPI.h index 67368d1a99d980b248789d24a2ea4f466255687a..7866122006a996cbe5201c661cab9c81aa82a219 100644 --- a/paddle/api/PaddleAPI.h +++ b/paddle/api/PaddleAPI.h @@ -59,9 +59,10 @@ class RangeError {}; /// Not support Error, such as access GPU memory directly, etc. class UnsupportError : public std::runtime_error { -public: - UnsupportError() : std::runtime_error(" "){}; - UnsupportError(const std::string& message) : std::runtime_error(message){}; + public: + UnsupportError() : std::runtime_error(" ") {} + explicit UnsupportError(const std::string& message) + : std::runtime_error(message) {} }; /// This type will map to python's list of float. @@ -105,7 +106,7 @@ class Matrix { DISABLE_COPY(Matrix); static Matrix* createByPaddleMatrixPtr(void* sharedPtr); -public: + public: virtual ~Matrix(); /** @@ -231,7 +232,7 @@ public: bool isGpu() const; -private: + private: void* getSharedPtr() const; MatrixPrivate* m; @@ -248,7 +249,7 @@ class Vector { void* getSharedPtr(); -public: + public: ~Vector(); /// Create Vector filled with zero. @@ -310,10 +311,10 @@ public: /// __len__ in python size_t getSize() const; -private: + private: VectorPrivate* m; -private: + private: friend class Parameter; friend class ParameterOptimizer; friend struct ParameterTraverseCallbackPrivate; @@ -325,7 +326,7 @@ class IVector { DISABLE_COPY(IVector); static IVector* createByPaddleVectorPtr(void* ptr); -public: + public: /// Create IVector filled with zero static IVector* createZero(size_t sz, bool useGpu = isUsingGpu()); @@ -389,7 +390,7 @@ public: /// This method will map to python __len__(); size_t getSize() const; -private: + private: void* getSharedPtr() const; friend class Arguments; @@ -400,11 +401,11 @@ struct ArgumentsPrivate; /// The Arguments is actual a std::vector in paddle. class Arguments { -private: + private: Arguments(); // Internal Create. DISABLE_COPY(Arguments); -public: + public: /** * Create a arguments with size. * Note that it can be zero. @@ -475,12 +476,12 @@ public: float sum() const; -private: + private: static Arguments* createByPaddleArgumentVector(void* ptr); static Arguments* createByPaddleArgument(const void* ptr); void* getInternalArgumentsPtr() const; -private: + private: ArgumentsPrivate* m; friend class Trainer; friend class GradientMachine; @@ -507,7 +508,7 @@ class ParameterConfig { static ParameterConfig* createParameterConfigFromParameterPtr(void* ptr); void* getRawPtr(); -public: + public: ~ParameterConfig(); /** @@ -515,10 +516,10 @@ public: */ std::string toProtoString() const; -private: + private: ParameterConfigPrivate* m; -private: + private: friend class Parameter; friend class ParameterOptimizer; friend struct ParameterTraverseCallbackPrivate; @@ -529,7 +530,7 @@ class OptimizationConfig { DISABLE_COPY(OptimizationConfig); OptimizationConfig(); -public: + public: static OptimizationConfig* createFromProtoString(const std::string& str); ~OptimizationConfig(); @@ -538,7 +539,7 @@ public: */ std::string toProtoString(); -private: + private: OptimizationConfigPrivate* m; friend class TrainerConfig; @@ -549,11 +550,11 @@ private: struct ParameterPrivate; class Parameter { -private: + private: Parameter(); DISABLE_COPY(Parameter); -public: + public: virtual ~Parameter(); /** @@ -580,11 +581,11 @@ public: size_t getSize() const; -private: + private: static Parameter* createFromRawPtr(void* ptr); static Parameter* createFromSharedPtr(void* ptr); -private: + private: ParameterPrivate* m; friend class UpdateCallbackWrapper; friend class GradientMachine; @@ -598,14 +599,14 @@ struct ModelConfigPrivate; * It is used by GradientMachine. */ class ModelConfig { -private: + private: ModelConfig(); DISABLE_COPY(ModelConfig); -public: + public: virtual ~ModelConfig(); -private: + private: ModelConfigPrivate* m; friend class TrainerConfig; friend struct TrainerConfigPrivate; @@ -619,11 +620,11 @@ struct TrainerConfigPrivate; * It is used by GradientMachine. */ class TrainerConfig { -private: + private: TrainerConfig(); DISABLE_COPY(TrainerConfig); -public: + public: virtual ~TrainerConfig(); static TrainerConfig* createFromTrainerConfigFile( @@ -634,7 +635,7 @@ public: OptimizationConfig* getOptimizationConfig() const; -private: + private: TrainerConfigPrivate* m; friend class Trainer; }; @@ -654,7 +655,7 @@ private: * @endcode */ class UpdateCallback { -public: + public: virtual ~UpdateCallback(); virtual void apply(Parameter* p); }; @@ -664,14 +665,14 @@ class ParameterTraverseCallback { DISABLE_COPY(ParameterTraverseCallback); ParameterTraverseCallback(); -public: + public: ~ParameterTraverseCallback(); void apply(const std::vector& vecs, const ParameterConfig& config, size_t sparseId); -private: + private: ParameterTraverseCallbackPrivate* m; friend class ParameterOptimizer; }; @@ -686,7 +687,7 @@ class ParameterOptimizer { DISABLE_COPY(ParameterOptimizer); ParameterOptimizer(); -public: + public: static ParameterOptimizer* create(OptimizationConfig* config); ~ParameterOptimizer(); @@ -710,7 +711,7 @@ public: ParameterTraverseCallback* needSpecialTraversal( const ParameterConfig& config) const; -private: + private: ParameterOptimizerPrivate* m; }; @@ -718,11 +719,11 @@ class SequenceGenerator; class Evaluator; struct GradientMachinePrivate; class GradientMachine { -private: + private: GradientMachine(); DISABLE_COPY(GradientMachine); -public: + public: virtual ~GradientMachine(); /** @@ -817,7 +818,7 @@ public: void eval(Evaluator* evaluator); -private: + private: GradientMachinePrivate* m; static GradientMachine* createFromPaddleModelPtr( @@ -833,10 +834,10 @@ private: struct ParameterUpdaterPrivate; class ParameterUpdater { -private: + private: ParameterUpdater(); -public: + public: static ParameterUpdater* createLocalUpdater(OptimizationConfig* config); static ParameterUpdater* createRemoteUpdater(OptimizationConfig* config, int passCount, @@ -911,17 +912,17 @@ public: */ void catchUpWith(); -private: + private: ParameterUpdaterPrivate* m; }; struct EvaluatorPrivate; class Evaluator { -private: + private: Evaluator(); DISABLE_COPY(Evaluator); -public: + public: ~Evaluator(); /** @@ -945,7 +946,7 @@ public: double getValue(const std::string name) const; -private: + private: EvaluatorPrivate* m; friend class GradientMachine; @@ -953,13 +954,13 @@ private: struct TrainerPrivate; class Trainer { -private: + private: TrainerPrivate* m; Trainer(); Trainer(TrainerConfig* optConfig, GradientMachine* gm); DISABLE_COPY(Trainer); -public: + public: virtual ~Trainer(); /// Create A Trainer By TrainerConfig. using paddle command line. @@ -1002,7 +1003,7 @@ public: /// the N-Best results generated from one input sequence. class ISequenceResults { -public: + public: virtual ~ISequenceResults(); /// Number of result. @@ -1026,7 +1027,7 @@ class SequenceGenerator { DISABLE_COPY(SequenceGenerator); SequenceGenerator(); -public: + public: virtual ~SequenceGenerator(); /** @@ -1044,10 +1045,10 @@ public: void setMaxLength(size_t maxlength); void setBeamSize(size_t beamSize); -private: + private: static SequenceGenerator* createByGradientMachineSharedPtr(void* ptr); friend class GradientMachine; -private: + private: SequenceGeneratorPrivate* m; }; diff --git a/paddle/api/SequenceGenerator.cpp b/paddle/api/SequenceGenerator.cpp index 1b30aec8f6b6b73764886a7c7274be67851e4815..1446c3084238859a759669f3a32c7efde67dcc2b 100644 --- a/paddle/api/SequenceGenerator.cpp +++ b/paddle/api/SequenceGenerator.cpp @@ -138,7 +138,7 @@ struct SequenceGeneratorPrivate { maxLength(0UL), feedback(__create_feedback__()) {} -private: + private: static paddle::Argument __create_feedback__() { paddle::Argument feedback; feedback.ids = paddle::IVector::create(/* size= */ 1, FLAGS_use_gpu); @@ -157,7 +157,7 @@ SequenceGenerator::~SequenceGenerator() { delete m; } class PathSequenceResults : public ISequenceResults { // ISequenceResults interface -public: + public: PathSequenceResults(const std::shared_ptr>& path, const std::shared_ptr>& dict) : path_(path), dict_(dict) {} @@ -196,7 +196,7 @@ public: } } -private: + private: std::shared_ptr> path_; std::shared_ptr> dict_; }; diff --git a/paddle/capi/CMakeLists.txt b/paddle/capi/CMakeLists.txt index e06e9a2b363d1ffc6876b98bcb7304b0a54dbcaa..957b1a3e6b07b058a76605992da387b43657146a 100644 --- a/paddle/capi/CMakeLists.txt +++ b/paddle/capi/CMakeLists.txt @@ -33,9 +33,6 @@ add_library(paddle_capi STATIC ${CAPI_HEADERS} ${CAPI_PRIVATE_HEADER} target_include_directories(paddle_capi PUBLIC ${CMAKE_CURRENT_BINARY_DIR}) -add_style_check_target(paddle_capi ${CAPI_SOURCES} ${CAPI_HEADER} - ${CAPI_PRIVATE_HEADER}) - add_dependencies(paddle_capi paddle_proto paddle_gserver) # TODO: paddle_capi_whole will be removed. diff --git a/paddle/capi/Matrix.cpp b/paddle/capi/Matrix.cpp index 24b0020636c0141b87dc80f5079f7342ec28157c..733d49cacfda17ad19b7bd7918be73c1fd14a64f 100644 --- a/paddle/capi/Matrix.cpp +++ b/paddle/capi/Matrix.cpp @@ -108,7 +108,7 @@ paddle_error paddle_matrix_get_row(paddle_matrix mat, paddle_error paddle_matrix_get_shape(paddle_matrix mat, uint64_t* height, uint64_t* width) { - if (mat == nullptr) return kPD_NULLPTR; + if (mat == nullptr || cast(mat)->mat == nullptr) return kPD_NULLPTR; if (height != nullptr) { *height = cast(mat)->mat->getHeight(); } diff --git a/paddle/capi/gradient_machine.cpp b/paddle/capi/gradient_machine.cpp index ea9aab00e3d05f1e2ef0c91eab93b67e0a3d5f37..8c3f504e5a2d807c0cc664af486ebab4a82ddec3 100644 --- a/paddle/capi/gradient_machine.cpp +++ b/paddle/capi/gradient_machine.cpp @@ -26,7 +26,7 @@ enum GradientMatchineCreateMode { namespace paddle { class MyNeuralNetwork : public NeuralNetwork { -public: + public: MyNeuralNetwork(const std::string& name, NeuralNetwork* network) : NeuralNetwork(name, network) {} }; diff --git a/paddle/contrib/CMakeLists.txt b/paddle/contrib/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..4b19256ef4533a09162edf907f6cd51146517e46 --- /dev/null +++ b/paddle/contrib/CMakeLists.txt @@ -0,0 +1,16 @@ +# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# + +add_subdirectory(inference) diff --git a/paddle/contrib/float16/.gitignore b/paddle/contrib/float16/.gitignore new file mode 100644 index 0000000000000000000000000000000000000000..dd28d354f4160b4be68b46a7bebcdf2097d5811a --- /dev/null +++ b/paddle/contrib/float16/.gitignore @@ -0,0 +1 @@ +*.inference.model diff --git a/paddle/contrib/float16/README.md b/paddle/contrib/float16/README.md new file mode 100644 index 0000000000000000000000000000000000000000..58b4a50666bfb622af8acbce29355f2a4a870a82 --- /dev/null +++ b/paddle/contrib/float16/README.md @@ -0,0 +1,171 @@ +# Float16 Inference in PaddlePaddle Fluid + +Kexin Zhao + +## Introduction +Deep learning is usually a two-stage work: training and inference. The training stage estimates model parameters (weights) from data. The inference stage loads the weights and uses them to interpret inputs. Typically, weights are 32-bit float values (float32). Some new devices, including NVIDIA Volta GPUs, support higher speed computation using 16-bit float values (float16). + +This article explains our efforts with PaddlePaddle to train using float32 and to inference using float16. We describe a [*transpiler*](https://github.com/PaddlePaddle/Paddle/blob/a4d3de0071e1f3912230c3ab3f9ac74cf06b093a/doc/fluid/design/motivation/fluid_compiler.md), which converts a PaddlePaddle Fluid model, which, to be precise, should be called a [Fluid *program*](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/program.md), into the inference program, and converts the weights from float32 into float16. + + +## What is float16? +float16 (or FP16) is a half-precision floating-point format that uses 16 bits in memory to represent a value. The advantage over 32-bit single-precision floating-point format (commonly known as float or float32 data type) is that it requires half the storage and bandwidth at the expense of precision and range. Fortunately, DNN inference has a high tolerance for the loss of precision and range when using float16 to represent the weights, and the inference accuracy will only be minimally affected in most cases, which gives us the opportunity to use float16 data type to speed up the inference. + +Interested readers can refer to our [design doc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/data_type/float16.md) and [code](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/platform/float16.h) for more details on how we implement the float16 data type. + +## Why float16? +The trend in today's deep learning community is to use bigger and deeper model, which translates to larger memory footprint, higher computation demands, and as a result higher energy consumption on computing devices. The advantages of float16 over float32 are correspondingly three-fold: + +1. We only need half the memory size to load the same model using float16 representations. Moreover, most of the intermediate results generated during float16 inference are also of the float16 data type. As a result, the whole memory footprint of float16 inference is roughly half of its float counterpart, which is especially useful when deploying inference on mobile devices with limited available memory. Also given the same available memory, the maximum batch size for float16 inference is about twice that for float inference. + +2. Because float16 occupies less memory than float, in theory, hardware devices can achieve much higher floating point operators per second (FLOPS) for float16 data than float data. Right now, NVIDIA's latest Volta GPUs, including Tesla V100 and Titan V, can deliver significantly higher FLOPS for float16 using Tensor Cores. Moreover, float16 takes less time to read from or write to memory, and hence float16 can make inference more efficient especially in memory-bound applications where the performance is mostly affected by how fast it is to read and write data. + +3. From the energy efficiency perspective, the energy needed to read, write, and compute float16 data is much less than its float counterpart, which can significantly reduce the battery power consumption on mobile devices or the total cost of ownership (TCO) of data centers. + +## Fluid implementation of float16 inference +### Overview +Fluid use [Program](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#program) instead of computation graph to describe a neural network model and the optimization procedure. Fluid program is a python wrapper around a protobuf message called [ProgramDesc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/program.md). Similar to programming languages, the basic structure of a Fluid program is some nested [blocks](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#block), where each block consists of some [variable](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#variable) definitions and a sequence of [operators](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#operator). An [executor](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/executor.md) will run a given program by sequentially executing the operators in the entrance block. + +### Basic requirement +When an executor runs an operator, it uses a kernel to perform computations on tensors contained in the input variables, and then writes the results to the tensors in the output variables. Each operator has multiple kernels for different combinations of data types, devices, and library types, respectively. The operator will select the appropriate kernel to run based on, among other things, the data type of the input tensors. By default, every Fluid operator has a kernel for float data type that takes float inputs and generates float outputs. + +If we provide float input to the first operator in a program, then each operator will use float kernel to compute float output and send it as input to the next operator to trigger its float kernel. This chain effect will make the program run in float mode and gives us a final output of float data type. + +The same principle applies if we want a program to run in float16 mode. We provide input variable of the float16 data type to the first operator, and every subsequent operator will invoke the float16 kernel until we get the final output in float16. So the preliminary requirements for float16 inference are to add float16 kernels to operators that are needed in a specific kind of neural networks. Our current focus is on Convolutional Neural Networks (CNN) and hence we have added float16 kernels to the following operators: convolution, pooling, GEMM, elementwise addition, batch norm, dropout, various activations including relu and tanh, and softmax. + +### float16 transpiler +Furthermore, we need a transpiler to write float16 inference code similar to the following: + +```python +# Get the float32 inference program and load the associated float32 weights +[inference_program, feed_target_names, + fetch_targets] = fluid.io.load_inference_model(save_dirname, exe) + +# Prepare the float input data +batch_size = 1 +tensor_img = numpy.random.rand(batch_size, 3, 32, 32).astype(numpy.float32) + +# Running inference_program in float mode +float_results = exe.run(inference_program, + feed={feed_target_names[0]: tensor_img}, + fetch_list=fetch_targets) + +# Use float16 transpiler to speedup +float16_inference_program = float_inference_program.clone() +t = Float16Transpiler() +t.transpile(float16_inference_program, GPUPlace) + +# Running float16_inference_program in float16 mode using the same input data +float16_results = exe.run(float16_inference_program, + feed={feed_target_names[0]: tensor_img}, + fetch_list=fetch_targets) + +# Do some tests to verify the correctness of float16 inference +... +np.testing.assert_almost_equal(float_results, float16_results, ...) +... + +# Save the float16 inference program and float16 weights for future deployment +fluid.io.save_inference_model(fp16_save_dirname, feed_target_names, + fetch_targets, exe, + float16_inference_program) +``` + +In this scenario, we already have a float32 inference program and some associated float32 weights. We can simply use the `transpile` method of the `Float16Transpiler` class to do certain modifications to the existing program and weights so that we have a new float16 program and the associated float16 weights. + +We can then run various inference experiments in float16 mode and save the float16 program and weights on disk for future deployment. To enhance the code usability, we maintain a consistent API so that user can use the same float32 input data to run inference program in either float32 and float16 mode and obtain output data both of float32 data type. Consequently, we need to add cast operators in the float16 inference program for conversions between the float16 tensor and float32 tensor. + +The float16 transpiler is implemented to fulfill the requirements mentioned above. The details of the float16 transpiler can be found [here](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/data_type/float16.md#float16-inference). + +### Experiment results +Simply running the following commands to reproduce the experiment results presented in this section: + +```bash +git clone https://github.com/PaddlePaddle/Paddle.git +cd Paddle +# This line will generate a paddle development docker image with cuda 8 and cudnn 7 +# If you want test on cuda 9 instead, change the line 5 in Paddle/Dockerfile +# from `FROM nvidia/cuda:8.0-cudnn7-devel-ubuntu16.04` +# to `FROM nvidia/cuda:9.0-cudnn7-devel-ubuntu16.04` and similarly for other configurations +nvidia-docker build -t paddle:float16 . +# After running this, different results will be written to different log files in Paddle/contrib/float16/ +nvidia-docker run -it -v $PWD:/paddle paddle:float16 /paddle/paddle/contrib/float16/run_float16_demo.sh +``` + +#### Accuracy +As is mentioned before, DNN inference has been found to be tolerant against the loss of precision and range incurred by float16, and we want to see how good this tolerance is. + +We train a resnet32 model using cifar10 data set, save it when test set accuracy is above 60%, and then test the inference accuracy on the 10000 examples of the cifar10 test set in float16 and float32 mode, respectively. + +We repeat the test ten times and get the following results: + +| | float16 | float32 | +|--------|--------:|--------: | +| # 1 | 62.75% | 62.72% | +| # 2 | 61.27% | 61.28% | +| # 3 | 62.24% | 62.23% | +| # 4 | 64.16% | 64.17% | +| # 5 | 60.75% | 60.77% | +| # 6 | 63.25% | 63.24% | +| # 7 | 62.15% | 62.13% | +| # 8 | 62.05% | 62.02% | +| # 9 | 65.19% | 65.20% | +| #10 | 62.53% | 62.48% | +| average| 62.63% | 62.62% | + +We can see that the accuracy of float16 inference is very close to that of float32 inference in every experiment (within 0.05% difference) and is overall 0.01% better than its float32 counterpart averaged over ten tests. + +#### Performance benchmark +Currently, Fluid only supports float16 inference on NVIDIA GPUs. There is no motivation to support float16 inference on non-ARM CPUs where float16 is not natively supported, and float16 calculation will only be slower than its float32 counterpart. + +NVIDIA started to support its native float16 data type (which has the same internal memory representation as Fluid's float16 class) on CUDA 7.5. Moreover, float16 speedups on computationally intensive tasks including GEMM (general matrix-matrix multiplication) and convolution are supported since cuBLAS 7.5 and cuDNN 5.0. + +Recently, the introduction of [Tensor Core](https://devblogs.nvidia.com/programming-tensor-cores-cuda-9/) in Volta architecture GPUs and the support of Tensor Core computation in CUDA 9.0 and cuDNN 7 make float16 genuinely superior to float in some deep learning applications. + +We thus benchmark the float16 inference performance on a single NVIDIA Tesla V100 GPU (Volta architecture and with Tensor Cores) and compare it with its float32 counterpart. All the following results are in ms (millisecond) averaged over 1000 mini-batches with respective to different mini-batch(mb) sizes. + +Average inference time for one mini-batch on Vgg16 model tested on ImageNet dataset: + +| total | mb=1 | mb=2 | mb=4 | mb=8 | mb=16 | mb=32 | mb=64 | +|-------|-----: |-----: |-----: |-----: |------: |------:|-------:| +|float32| 14.01 | 9.70 | 22.99 | 28.26 | 53.87 | 84.42 | 178.95 | +|float16| 3.32 | 4.11 | 5.88 | 9.41 | 16.54 | 30.47 | 60.23 | +|Speedup| 4.22 | 2.36  | 3.91 | 3.00 | 3.26  | 2.77 | 2.97 | + +We can see that float16 inference provides **2x ~ 4x** speedup on different batch sizes. + +Convolution operation is ususally the computational bottleneck of CNN, so we also check the average time spent on the Fluid convolution operators for one mini-batch as follows: + +|conv op| mb=1 | mb=2 | mb=4 | mb=8 | mb=16 | mb=32 | mb=64 | +|-------|-----: |-----: |-----: |-----: |------: |------:|-------:| +|float32| 11.95 | 6.96 | 18.65 | 21.42 | 41.35 | 60.58 | 130.11 | +|float16| 1.78 | 2.10 | 2.93 | 4.55 | 7.99 | 14.63 | 28.67 | +|Speedup| 6.71 | 3.31  | 6.37 | 4.71 | 5.18  | 4.14 | 4.54 | + +Fluid convolution operator uses cuDNN 7 to implement the kernel, and we can see that with the help of Tensor Core, float16 convolution is significantly faster than its float32 counterpart, which makes the overall float16 inference performance much better. + +Similarly, we also list the benchmark results of Resnet50 model tested on the ImageNet dataset: + +| total | mb=1 | mb=2 | mb=4 | mb=8 | mb=16 | mb=32 | mb=64 | mb=128 | +|-------|-----: |-----: |-----: |-----: |------: |------:|-------:|-------:| +|float32| 7.03 | 7.41 | 9.16 | 12.55 | 21.13 | 38.27 | 67.93 | 127.02 | +|float16| 6.13 | 6.32 | 6.24 | 7.40 | 10.90 | 18.18 | 33.20 | 64.52 | +|Speedup| 1.15 | 1.17  | 1.47  | 1.70 | 1.94  | 2.11 | 2.05 | 1.97 | + +|conv op| mb=1 | mb=2 | mb=4 | mb=8 | mb=16 | mb=32 | mb=64 | mb=128 | +|-------|-----: |-----: |-----: |-----: |------: |------:|-------:|-------:| +|float32| 5.43 | 5.46 | 6.50 | 8.36 | 13.80 | 24.45 | 41.21 | 73.44 | +|float16| 4.19 | 4.30 | 3.96 | 4.21 | 5.63 | 8.77 | 15.24 | 28.40 | +|Speedup| 1.30 | 1.27  | 1.64  | 1.99 | 2.45  | 2.79 | 2.70 | 2.59 | + +We find that the speedup provided by float16 inference starts relatively small at 1.15x for batch size 1 and gradually increases to about 2x for larger batch sizes. A similar trend can be found for the time spent on the convolution operator. Note that right now Tensor Cores will only be utilized in the convolution operation when the input data and filter meet specific dimensional requirements. The speedup by float16 inference for Resnet50 is smaller than the Vgg16 counterpart partially because the convolution operation in Resnet is much simpler than its Vgg counterpart and this makes the tensor core less utilized in Resnet than in Vgg. + +We also did the same benchmark on a single NVIDIA GeForce GTX 1080 Ti GPU that does not support Tensor Core. The results show that for Vgg16, float16 inference provides consistent small speedup (around 1.15x) for all mini-batch sizes, while for Resnet50, float16 inference is slower than its float32 counterpart in small batch sizes (mb = 1 and 2) and then delivers around 1.15x speedup for all larger batch sizes. By comparing the benchmarks on 1080 Ti and V100, we find that Tensor Core, which is specialized for float16 computations, is a critical component of high performance float16 inference. + +Please refer to [here](https://github.com/PaddlePaddle/Paddle/blob/develop/contrib/float16/float16_benchmark.md) for complete benchmark results. + +### Summary +1. Fluid is now able to run inference in float16 mode via a float16 transpiler. We currently support CNN programs, including Vgg and Resnet, to run in float16 inference mode. +2. The accuracy of float16 inference is verified to be almost identical to its float32 counterpart at least on CNN models. +3. float16 inference provides a significant speedup on large and computationally intensive Vgg16 model on ImageNet dataset. For the much smaller and simpler Resnet50 model, the speedup provided by float16 inference is less significant than for Vgg16 model but still favorable, especially for large batch sizes. +4. We cannot achieve the superior float16 inference performance without the help of the newly introduced Tensor Cores on NVIDIA Volta architecture GPUs. diff --git a/paddle/contrib/float16/float16_benchmark.md b/paddle/contrib/float16/float16_benchmark.md new file mode 100644 index 0000000000000000000000000000000000000000..b51d6bde92fa04d2268afa36b9c4bd18bc28fe73 --- /dev/null +++ b/paddle/contrib/float16/float16_benchmark.md @@ -0,0 +1,97 @@ +# float16 benchmark + +## Description +We want to compare the inference benchmark of float16 vs float32 on the "image_classification" example on Nvidia Tesla V100 GPU, where we can enable the tensor core computation for float16 mode. We test Vgg16 and Resnet50 on the imagenet data set, and Vgg16 and Resnet32 on the cifar10 data set. For completeness, we also add the inference benchmark of Vgg16 and Resnet50 on imagenet data set tested on Nvidia GeForce GTX 1080 Ti GPU. + +For more details about tensor core, please refer to https://devblogs.nvidia.com/programming-tensor-cores-cuda-9/ + +## Test environment +- GPU: single Nvidia Tesla V100 or single Nvidia GeForce GTX 1080 Ti +- CUDNN: 7.1.1 +- CUDA: 9.0 +- Code: https://github.com/PaddlePaddle/Paddle/pull/10331 (Tensor core is enabled in float16 mode) + +## Benchmark on V100 +All times are in ms (millisecond) averaged over 1000 iterations tested on a single Nvidia V100 GPU with respective to different mini-batch(mb) sizes. + +### Vgg16 on imagenet (flowers data set: image.shape = [3, 224, 224]): + +Total inference time for one batch: + +| | mb=1 | mb=2 | mb=4 | mb=8 | mb=16 | mb=32 | mb=64 | +|-------|-----: |-----: |-----: |-----: |------: |------:|-------:| +|float32| 14.01 | 9.70 | 22.99 | 28.26 | 53.87 | 84.42 | 178.95 | +|float16| 3.32 | 4.11 | 5.88 | 9.41 | 16.54 | 30.47 | 60.23 | +|Speedup| 4.22 | 2.36  | 3.91 | 3.00 | 3.26  | 2.77 | 2.97 | + +Total time spent on conv op for one batch: + +| | mb=1 | mb=2 | mb=4 | mb=8 | mb=16 | mb=32 | mb=64 | +|-------|-----: |-----: |-----: |-----: |------: |------:|-------:| +|float32| 11.95 | 6.96 | 18.65 | 21.42 | 41.35 | 60.58 | 130.11 | +|float16| 1.78 | 2.10 | 2.93 | 4.55 | 7.99 | 14.63 | 28.67 | +|Speedup| 6.71 | 3.31  | 6.37 | 4.71 | 5.18  | 4.14 | 4.54 | + + +### Resnet50 on imagenet (flowers data set: image.shape = [3, 224, 224]): + +Total inference time for one batch: + +|       | mb=1 | mb=2 | mb=4 | mb=8 | mb=16 | mb=32 | mb=64 | mb=128 | +|-------|-----: |-----: |-----: |-----: |------: |------:|-------:|-------:| +|float32| 7.03 | 7.41 | 9.16 | 12.55 | 21.13 | 38.27 | 67.93 | 127.02 | +|float16| 6.13 | 6.32 | 6.24 | 7.40 | 10.90 | 18.18 | 33.20 | 64.52 | +|Speedup| 1.15 | 1.17  | 1.47  | 1.70 | 1.94  | 2.11 | 2.05 | 1.97 | + +Total time spent on conv op for one batch: + +|       | mb=1 | mb=2 | mb=4 | mb=8 | mb=16 | mb=32 | mb=64 | mb=128 | +|-------|-----: |-----: |-----: |-----: |------: |------:|-------:|-------:| +|float32| 5.43 | 5.46 | 6.50 | 8.36 | 13.80 | 24.45 | 41.21 | 73.44 | +|float16| 4.19 | 4.30 | 3.96 | 4.21 | 5.63 | 8.77 | 15.24 | 28.40 | +|Speedup| 1.30 | 1.27  | 1.64  | 1.99 | 2.45  | 2.79 | 2.70 | 2.59 | + + +### Vgg16 on cifar10 (image.shape = [3, 32, 32]): + +Total inference time for one batch: + +| | mb=1 | mb=2 | mb=4 | mb=8 | mb=16 | mb=32 | mb=64 | mb=128 | mb=256 | mb=512 | +|-------|-----:|-----:|-----:|-----:|------:|------:|------:|-------:|-------:|-------:| +|float32| 3.13 | 3.17 | 3.19 | 3.58 | 3.98 | 6.23 | 8.42 | 13.44 | 24.19 | 44.97 | +|float16| 2.72 | 2.77 | 2.76 | 2,88 | 2.96 | 3.24 | 4.01 | 5.78 | 9.65 | 17.37 | +|Speedup| 1.15 | 1.14 | 1.16 | 1.24 | 1.34 | 1.92  | 2.10 | 2.33  | 2.51 | 2.59 | + + +### Resnet32 on cifar10 (image.shape = [3, 32, 32]): + +Total inference time for one batch: + +| | mb=1 | mb=2 | mb=4 | mb=8 | mb=16 | mb=32 | mb=64 | mb=128 | mb=256 | mb=512 | +|-------|-----:|-----:|-----:|-----:|------:|------:|------:|-------:|-------:|-------:| +|float32| 3.11 | 3.14 | 2.99 | 3.04 | 3.10 | 3.28 | 4.47 | 6.86 | 11.63 | 21.16 | +|float16| 3.70 | 3.81 | 3.75 | 3.83 | 3.77 | 3.97 | 3.92 | 4.15 | 6.41 | 11.02 | +|Speedup|     |     |     |     |       | | 1.14  | 1.65 | 1.81 | 1.92 | + + +## Benchmark on 1080 Ti +All times are in ms (millisecond) averaged over 1000 iterations tested on a single Nvidia GeForce GTX 1080 Ti GPU with respective to different mini-batch(mb) sizes. + +### Vgg16 on imagenet (flowers data set: image.shape = [3, 224, 224]): +Total inference time for one batch: + +| | mb=1 | mb=2 | mb=4 | mb=8 | mb=16 | mb=32 | +|-------|-----: |-----: |-----: |-----: |------: |-------:| +|float32| 5.60 | 9.38 | 15.86 | 29.79 | 57.60 | 117.73 | +|float16| 4.99 | 7.79 | 13.47 | 26.02 | 52.30 | 102.34 | +|Speedup| 1.12 | 1.20  | 1.18 | 1.15 | 1.10  | 1.15 | + + +### Resnet50 on imagenet (flowers data set: image.shape = [3, 224, 224]): +Total inference time for one batch: + +| | mb=1 | mb=2 | mb=4 | mb=8 | mb=16 | mb=32 | mb=64 | +|-------|-----: |-----: |-----: |-----: |------: |-------:|-------:| +|float32| 5.63 | 6.23 | 8.85 | 14.71 | 26.07 | 52.86 | 108.95 | +|float16| 5.89 | 6.44 | 7.94 | 12.57 | 22.03 | 45.06 | 92.68 | +|Speedup| |  | 1.12  | 1.17 | 1.18  | 1.17 | 1.18 | diff --git a/paddle/contrib/float16/float16_inference_demo.py b/paddle/contrib/float16/float16_inference_demo.py new file mode 100644 index 0000000000000000000000000000000000000000..063227d5d2586d66ad4091133a8edf014da839f8 --- /dev/null +++ b/paddle/contrib/float16/float16_inference_demo.py @@ -0,0 +1,362 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import print_function +from float16_transpiler import Float16Transpiler + +import argparse +import paddle +import paddle.fluid as fluid +import contextlib +import math +import sys +import numpy as np +import os + +parser = argparse.ArgumentParser( + 'Float16 inference accuracy test and benchmark.') +parser.add_argument( + '--train_batch_size', type=int, default=16, help="Batch size for training.") +parser.add_argument( + '--inf_batch_size', type=int, default=32, help="Batch size for inference.") +parser.add_argument( + '--repeat', type=int, default=1, help="How many times to run the test.") +parser.add_argument( + '--data_set', + type=str, + default='cifar10', + choices=['cifar10', 'imagenet'], + help="Optional dataset for benchmark.") +parser.add_argument( + '--model', + type=str, + default='vgg', + choices=['vgg', 'resnet'], + help="Optional model for benchmark.") +parser.add_argument( + '--threshold', + type=float, + default=0.005, + help='Save inference model when test accuracy reach this threshold.') +parser.add_argument('--learning_rate', type=float, default=0.001) +args = parser.parse_args() + + +def conv_bn_layer(input, ch_out, filter_size, stride, padding, act='relu'): + conv1 = fluid.layers.conv2d( + input=input, + filter_size=filter_size, + num_filters=ch_out, + stride=stride, + padding=padding, + act=None, + bias_attr=False) + return fluid.layers.batch_norm(input=conv1, act=act) + + +def shortcut(input, ch_out, stride): + ch_in = input.shape[1] + if ch_in != ch_out: + return conv_bn_layer(input, ch_out, 1, stride, 0, None) + else: + return input + + +def basicblock(input, ch_out, stride): + short = shortcut(input, ch_out, stride) + conv1 = conv_bn_layer(input, ch_out, 3, stride, 1) + conv2 = conv_bn_layer(conv1, ch_out, 3, 1, 1, act=None) + return fluid.layers.elementwise_add(x=short, y=conv2, act='relu') + + +def bottleneck(input, ch_out, stride): + short = shortcut(input, ch_out * 4, stride) + conv1 = conv_bn_layer(input, ch_out, 1, stride, 0) + conv2 = conv_bn_layer(conv1, ch_out, 3, 1, 1) + conv3 = conv_bn_layer(conv2, ch_out * 4, 1, 1, 0, act=None) + return fluid.layers.elementwise_add(x=short, y=conv3, act='relu') + + +def layer_warp(block_func, input, ch_out, count, stride): + res_out = block_func(input, ch_out, stride) + for i in range(1, count): + res_out = block_func(res_out, ch_out, 1) + return res_out + + +def resnet_imagenet(input, depth=50): + cfg = { + 18: ([2, 2, 2, 1], basicblock), + 34: ([3, 4, 6, 3], basicblock), + 50: ([3, 4, 6, 3], bottleneck), + 101: ([3, 4, 23, 3], bottleneck), + 152: ([3, 8, 36, 3], bottleneck) + } + stages, block_func = cfg[depth] + conv1 = conv_bn_layer(input, ch_out=64, filter_size=7, stride=2, padding=3) + pool1 = fluid.layers.pool2d( + input=conv1, pool_type='avg', pool_size=3, pool_stride=2) + res1 = layer_warp(block_func, pool1, 64, stages[0], 1) + res2 = layer_warp(block_func, res1, 128, stages[1], 2) + res3 = layer_warp(block_func, res2, 256, stages[2], 2) + res4 = layer_warp(block_func, res3, 512, stages[3], 2) + pool2 = fluid.layers.pool2d( + input=res4, + pool_size=7, + pool_type='avg', + pool_stride=1, + global_pooling=True) + return pool2 + + +def resnet_cifar10(input, depth=32): + assert (depth - 2) % 6 == 0 + + n = (depth - 2) // 6 + + conv1 = conv_bn_layer( + input=input, ch_out=16, filter_size=3, stride=1, padding=1) + res1 = layer_warp(basicblock, conv1, 16, n, 1) + res2 = layer_warp(basicblock, res1, 32, n, 2) + res3 = layer_warp(basicblock, res2, 64, n, 2) + pool = fluid.layers.pool2d( + input=res3, pool_size=8, pool_type='avg', pool_stride=1) + return pool + + +def vgg16(input): + def conv_block(input, num_filter, groups, dropouts): + return fluid.nets.img_conv_group( + input=input, + pool_size=2, + pool_stride=2, + conv_num_filter=[num_filter] * groups, + conv_filter_size=3, + conv_act='relu', + conv_with_batchnorm=True, + conv_batchnorm_drop_rate=dropouts, + pool_type='max') + + conv1 = conv_block(input, 64, 2, [0.3, 0]) + conv2 = conv_block(conv1, 128, 2, [0.4, 0]) + conv3 = conv_block(conv2, 256, 3, [0.4, 0.4, 0]) + conv4 = conv_block(conv3, 512, 3, [0.4, 0.4, 0]) + conv5 = conv_block(conv4, 512, 3, [0.4, 0.4, 0]) + + drop = fluid.layers.dropout(x=conv5, dropout_prob=0.5) + fc1 = fluid.layers.fc(input=drop, size=4096, act=None) + bn = fluid.layers.batch_norm(input=fc1, act='relu') + drop2 = fluid.layers.dropout(x=bn, dropout_prob=0.5) + fc2 = fluid.layers.fc(input=drop2, size=4096, act=None) + return fc2 + + +def train(place, save_dirname): + if args.data_set == "cifar10": + class_dim = 10 + data_shape = [3, 32, 32] + elif args.data_set == "imagenet": + class_dim = 102 + data_shape = [3, 224, 224] + else: + raise ValueError("%s dataset is not supported" % data_set) + + images = fluid.layers.data(name='pixel', shape=data_shape, dtype='float32') + label = fluid.layers.data(name='label', shape=[1], dtype='int64') + + if args.model == "vgg": + print("train vgg") + net = vgg16(images) + elif args.model == "resnet": + print("train resnet") + if args.data_set == "cifar10": + net = resnet_cifar10(images) + elif args.data_set == "imagenet": + net = resnet_imagenet(images) + else: + raise ValueError("%s dataset is not supported" % args.data_set) + else: + raise ValueError("%s network is not supported" % args.model) + + predict = fluid.layers.fc(input=net, size=class_dim, act='softmax') + cost = fluid.layers.cross_entropy(input=predict, label=label) + avg_cost = fluid.layers.mean(x=cost) + acc = fluid.layers.accuracy(input=predict, label=label) + + #Test program + test_program = fluid.default_main_program().clone(for_test=True) + optimizer = fluid.optimizer.Adam(learning_rate=args.learning_rate) + optimizer.minimize(avg_cost) + + BATCH_SIZE = args.train_batch_size + PASS_NUM = 100 + + train_reader = paddle.batch( + paddle.reader.shuffle( + paddle.dataset.flowers.train() + if args.data_set == 'imagenet' else paddle.dataset.cifar.train10(), + buf_size=128 * 10), + batch_size=args.train_batch_size) + + test_reader = paddle.batch( + paddle.dataset.flowers.test() + if args.data_set == 'imagenet' else paddle.dataset.cifar.test10(), + batch_size=args.inf_batch_size) + + exe = fluid.Executor(place) + feeder = fluid.DataFeeder(place=place, feed_list=[images, label]) + + exe.run(fluid.default_startup_program()) + main_program = fluid.default_main_program() + + for pass_id in range(PASS_NUM): + for batch_id, data in enumerate(train_reader()): + train_image = np.array( + map(lambda x: x[0].reshape(data_shape), data)).astype("float32") + train_label = np.array(map(lambda x: x[1], data)).astype("int64") + train_label = train_label.reshape([-1, 1]) + + exe.run(main_program, + feed={'pixel': train_image, + 'label': train_label}) + + if (batch_id % 100) == 0: + acc_list = [] + avg_loss_list = [] + for tid, test_data in enumerate(test_reader()): + test_image = np.array( + map(lambda x: x[0].reshape(data_shape), + test_data)).astype("float32") + test_label = np.array(map(lambda x: x[1], + test_data)).astype("int64") + test_label = test_label.reshape([-1, 1]) + + loss_t, acc_t = exe.run( + program=test_program, + feed={"pixel": test_image, + "label": test_label}, + fetch_list=[avg_cost, acc]) + if math.isnan(float(loss_t)): + sys.exit("got NaN loss, training failed.") + acc_list.append(float(acc_t)) + avg_loss_list.append(float(loss_t)) + + acc_value = np.array(acc_list).mean() + avg_loss_value = np.array(avg_loss_list).mean() + + print( + 'PassID {0:1}, BatchID {1:04}, Test Loss {2:2.2}, Accuracy {3:2.2}'. + format(pass_id, batch_id + 1, + float(avg_loss_value), float(acc_value))) + + if acc_value > args.threshold: + print( + 'Save inference model with test accuracy of {0} at {1}'. + format(float(acc_value), save_dirname)) + fluid.io.save_inference_model(save_dirname, ["pixel"], + [predict], exe) + return + + +def test_accuracy(executor, inference_program, feed_target_names, + fetch_targets): + if args.data_set == "cifar10": + data_shape = [3, 32, 32] + elif args.data_set == "imagenet": + data_shape = [3, 224, 224] + else: + raise ValueError("%s dataset is not supported" % data_set) + + test_reader = paddle.batch( + paddle.dataset.cifar.test10() + if args.data_set == "cifar10" else paddle.dataset.flowers.test(), + batch_size=args.inf_batch_size) + + test_num = 0 + correct_num = 0 + + for test_data in test_reader(): + test_image = np.array( + map(lambda x: x[0].reshape(data_shape), test_data)).astype( + "float32") + test_label = np.array(map(lambda x: x[1], test_data)).astype("int64") + test_label = test_label.reshape([-1, 1]) + + results = executor.run(program=inference_program, + feed={feed_target_names[0]: test_image}, + fetch_list=fetch_targets) + + prediction = np.argmax(results[0], axis=1).reshape([-1, 1]) + correct_num += np.sum(prediction == test_label) + test_num += test_label.size + + print("{0} out of {1} predictions are correct.".format(correct_num, + test_num)) + print("Test accuray is {0}.".format(float(correct_num) / float(test_num))) + + +def infer(place, save_dirname): + exe = fluid.Executor(place) + inference_scope = fluid.core.Scope() + + with fluid.scope_guard(inference_scope): + # Use fluid.io.load_inference_model to obtain the inference program desc, + # the feed_target_names (the names of variables that will be feeded + # data using feed operators), and the fetch_targets (variables that + # we want to obtain data from using fetch operators). + print("Load inference model from {0}".format(save_dirname)) + [inference_program, feed_target_names, + fetch_targets] = fluid.io.load_inference_model(save_dirname, exe) + + print("The test set accuracy of inference in float mode is:") + test_accuracy(exe, inference_program, feed_target_names, fetch_targets) + + float16_inference_program = inference_program.clone() + t = Float16Transpiler() + t.transpile(float16_inference_program, place) + + print("The test set accuracy of inference in float16 mode is:") + test_accuracy(exe, float16_inference_program, feed_target_names, + fetch_targets) + + fp16_save_dirname = "float16_" + save_dirname + fluid.io.save_inference_model(fp16_save_dirname, feed_target_names, + fetch_targets, exe, + float16_inference_program) + + +@contextlib.contextmanager +def scope_prog_guard(): + prog = fluid.Program() + startup_prog = fluid.Program() + scope = fluid.core.Scope() + with fluid.scope_guard(scope): + with fluid.program_guard(prog, startup_prog): + yield + + +if __name__ == "__main__": + if not fluid.core.is_compiled_with_cuda(): + raise Exception("This test requires CUDA GPUs!") + + place = fluid.CUDAPlace(0) + if not fluid.core.is_float16_supported(place): + raise Exception( + "This test requires compute capability of CUDA GPU >= 5.3!") + + for i in range(args.repeat): + with scope_prog_guard(): + save_dirname = "image_classification_" + args.data_set + "_" + args.model + ".inference.model" + train(place, save_dirname) + infer(place, save_dirname) diff --git a/paddle/contrib/float16/float16_transpiler.py b/paddle/contrib/float16/float16_transpiler.py new file mode 100644 index 0000000000000000000000000000000000000000..91ba101edb65cd45bd5e37a0c6ad25e515593a81 --- /dev/null +++ b/paddle/contrib/float16/float16_transpiler.py @@ -0,0 +1,256 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import numpy as np +import paddle.fluid.core as core +from paddle.fluid.framework import Program +from paddle.fluid.executor import global_scope + + +class Float16Transpiler: + def transpile(self, program, place, scope=None): + ''' + Transpile the program desc and cast the weights to float16 data type to + enable float16 inference. + + Since the operator in a program desc will automatically choose the + right compute kernel to run based on the data type of the input tensor. + We actually don't need to change the program desc to run in float16 mode. + + However, in this way, users who are used to feeding and fetching tensors + of float32 data type when running typical inference may find it confusing + and difficult to run inference in float16 mode as they need to convert + input data to float16 dtype and then convert the results back to float32 + dtype to match the rest of code. + + So this function appends cast ops to the program desc where necessary so + that users are able to run inference in float16 mode while providing input + tensor (feed_holder) of float data type and obtaining output tensor + (fetch_holder) of float data type. + + Moreover, it is desired that when we have the scope and program desc to run + inference in float32 mode, we can use a single API to do the necessary + modification and then user can run float16 inference on the fly. To make + this happen, this function also create new parameters in the scope to have the + converted float16 weights and change the operators in program desc to use + these new parameters. + + :param program: program to transpile + :type program: Program + :param place: inference place + :type place: Place + :param scope: inference scope + :type scope: Scope + ''' + if not isinstance(program, Program): + raise TypeError("program should be as Program type") + if not isinstance(place, core.CPUPlace) and not isinstance( + place, core.CUDAPlace): + raise TypeError("place should be as CPUPlace/CUDAPlace type") + if scope is None: + scope = global_scope() + if not isinstance(scope, core.Scope): + raise TypeError("scope should be as Scope type or None") + + self.scope = scope + self.place = place + self.block = program.block(0) + self.input_map = {} # store the input names should be adjusted + + self._modify_feed_fetch() + self._convert_param_to_float16() + self._adjust_input(skip=True) + self._remove_unused_var() + + # TODO(luotao): use clone() method to flush the program.desc in force, + # since some large program.desc will not be flushed immediately. + # And a better solution will be considered later. + program = program.clone() + + # ====================== private transpiler functions ===================== + def _adjust_input(self, skip=False): + ''' + Change the input variable name in operators. + + When we are in the process of modifying a program desc, we usually + replace some variables with some other variables, where we create + a dictionary input_map to record the one-to-one correspondence + between each old variable and the new one. + + After that, this function will search all the operators that use the + old variables and change the info in op to use the new variables. There + maybe some exceptions to this rule when we are using the float16 transpiler + and insert cast ops to cast float32 variable to float16 one. After we + insert the cast op to cast var_1 to var_1_fp16, we don't want to change + the input of cast op to var_1_fp16 after using this function. + ''' + skip_ops = {"cast"} + for i in range(len(self.block.ops)): + current_op = self.block.ops[i] + if skip and current_op.type in skip_ops: + continue + for input_arg in current_op.input_arg_names: + if input_arg in self.input_map: + current_op.rename_input(input_arg, + self.input_map[input_arg]) + + def _remove_unused_var(self): + ''' + remove unused varibles in program + ''' + args = [] + for i in range(len(self.block.ops)): + current_op = self.block.ops[i] + args += current_op.input_arg_names + args += current_op.output_arg_names + args = list(set(args)) # unique the input and output arguments + + for var in self.block.vars.keys(): + if var not in args: + self.block.remove_var(var) + + def _modify_feed_fetch(self): + ''' + Modify feed fetch op/vars for float16 inference. + + For each feed op: + feed_op->feed_target_var + + Change it to: + feed_op->feed_target_var->cast_op(from other dtype to float16)->tmp_var + + For each fetch op: + fetch_target_var->fetch_op + + Change it to: + tmp_var->cast_op(from float16 to other dtype)->fetch_target_var->fetch_op + + :return: None + ''' + + def find_op(var): + # It is possible that var.op is not up to date after some + # modifications to program desc. Here we force to make it up to date. + var.op = None + for op in self.block.ops: + if var.name in op.output_arg_names: + var.op = op + break + + if var.op is None: + raise ValueError("The target variable must have an " + "associated operator that generates it.") + + i = 0 + while i < len(self.block.ops): + cur_op = self.block.ops[i] + if cur_op.type == "feed": + var_name = cur_op.output("Out")[0] + tmp_var_name = var_name + ".fp16" + var = self.block.vars[var_name] + tmp_var = self.block.create_var( + name=tmp_var_name.encode('ascii'), + type=var.type, + dtype=core.VarDesc.VarType.FP16, + shape=var.shape, + persistable=var.persistable) + self.block.insert_op( + i + 1, + type="cast", + inputs={"X": var}, + outputs={"Out": tmp_var}, + attrs={ + 'in_dtype': int(var.dtype), + 'out_dtype': int(tmp_var.dtype) + }) + self.input_map[var_name] = tmp_var_name + i = i + 1 + elif cur_op.type == "fetch": + var_name = cur_op.input("X")[0] + tmp_var_name = var_name + ".fp16" + var = self.block.vars[var_name] + tmp_var = self.block.create_var( + name=tmp_var_name.encode('ascii'), + type=var.type, + dtype=core.VarDesc.VarType.FP16, + shape=var.shape, + persistable=var.persistable) + find_op(var) + var.op.rename_output(var_name, tmp_var_name) + self.block.insert_op( + i, + type="cast", + inputs={"X": tmp_var}, + outputs={"Out": var}, + attrs={ + 'in_dtype': int(tmp_var.dtype), + 'out_dtype': int(var.dtype) + }) + i = i + 1 + i = i + 1 + + def _convert_param_to_float16(self): + def _get_no_fp16_conversion_var_names(): + ''' + Get the set of input variable names that shouldn't be converted to float16. + + When we want to run inference in float16 mode, most parameters need to be + firstly converted to float16. However, there are some parameters that + shouldn't be converted to float16 because the corresponding operator + requires float32 parameters even in float16 mode (when the input data is + of float16 data type). Currently, the only operator that has this exclusion + is the batch norm op. + + :return: set of input variable names + :type var_names: set + ''' + op_names = {'batch_norm'} + var_names = [] + for op in self.block.ops: + if op.type in op_names: + var_names += op.input_arg_names + return set(var_names) + + def _should_be_converted(var): + return var.persistable and \ + var.name not in self.no_conversion_vars and \ + var.type != core.VarDesc.VarType.FEED_MINIBATCH and \ + var.type != core.VarDesc.VarType.FETCH_LIST + + self.no_conversion_vars = _get_no_fp16_conversion_var_names() + conversion_var_list = filter(_should_be_converted, + self.block.vars.values()) + for var in conversion_var_list: + fp16_var_name = var.name + ".fp16" + fp16_var = self.block.create_parameter( + name=fp16_var_name.encode('ascii'), + type=var.type, + dtype=core.VarDesc.VarType.FP16, + shape=var.shape) + + # cast the data in the tensor of the original var to float16 + # data type and store it in the tensor of the new float16 var + self.scope.var(fp16_var_name) + fp16_tensor = self.scope.find_var(fp16_var_name).get_tensor() + tensor = np.array(self.scope.find_var(var.name).get_tensor()) + # After the old tensor data is converted to np.float16, view(np.uint16) + # is used so that the internal memory of the numpy array will be + # reinterpreted to be of np.uint16 data type, which is binded to fluid + # float16 data type via the help of pybind in tensor_py.h. + fp16_tensor.set( + tensor.astype(np.float16).view(np.uint16), self.place) + + # old var will be replaced by the fp16 var in program desc + self.input_map[var.name] = fp16_var_name + self.block.remove_var(var.name) diff --git a/paddle/contrib/float16/run_float16_demo.sh b/paddle/contrib/float16/run_float16_demo.sh new file mode 100755 index 0000000000000000000000000000000000000000..031225a85dabb26e5d9ea06f58909c049e7f0c08 --- /dev/null +++ b/paddle/contrib/float16/run_float16_demo.sh @@ -0,0 +1,113 @@ +#!/bin/bash + +BUILD_PATH=/paddle/fp16_build +WHEEL_PATH=$BUILD_PATH/python/dist +INFER_PATH=$BUILD_PATH/paddle/fluid/inference/tests/book +DEMO_PATH=/paddle/paddle/contrib/float16 + +# Use the single most powerful CUDA GPU on your machine +export CUDA_VISIBLE_DEVICES=0 + +# Build the PaddlePaddle Fluid wheel package and install it. +mkdir -p $BUILD_PATH && cd $BUILD_PATH +cmake .. -DWITH_AVX=OFF \ + -DWITH_MKL=OFF \ + -DWITH_GPU=ON \ + -DWITH_TESTING=ON \ + -DWITH_TIMER=ON \ + -DWITH_PROFILER=ON \ + -DWITH_FLUID_ONLY=ON +make -j `nproc` +pip install -U "$WHEEL_PATH/$(ls $WHEEL_PATH)" + +cd $DEMO_PATH +# Clear previous log results +rm -f *.log + +# Test the float16 inference accuracy of resnet32 on cifar10 data set +stdbuf -oL python float16_inference_demo.py \ + --data_set=cifar10 \ + --model=resnet \ + --threshold=0.6 \ + --repeat=10 \ + 2>&1 | tee -a float16_inference_accuracy.log + +# Sleep to cool down the GPU for consistent benchmarking +sleep 2m + +# benchmarking parameters +REPEAT=1000 +MAXIMUM_BATCH_SIZE=512 + +for ((batch_size = 1; batch_size <= MAXIMUM_BATCH_SIZE; batch_size *= 2)); +do + + # Test inference benchmark of vgg16 on imagenet + stdbuf -oL python float16_inference_demo.py \ + --data_set=imagenet \ + --model=vgg \ + --threshold=0.001 \ + --repeat=1 \ + + $INFER_PATH/test_inference_image_classification_vgg \ + --dirname=$DEMO_PATH/image_classification_imagenet_vgg.inference.model \ + --fp16_dirname=$DEMO_PATH/float16_image_classification_imagenet_vgg.inference.model \ + --repeat=$REPEAT \ + --batch_size=$batch_size \ + --skip_cpu=true \ + 2>&1 | tee -a imagenet_vgg16_benchmark.log + + sleep 2m + + # Test inference benchmark of resnet50 on imagenet + stdbuf -oL python float16_inference_demo.py \ + --data_set=imagenet \ + --model=resnet \ + --threshold=0.001 \ + --repeat=1 \ + + $INFER_PATH/test_inference_image_classification_resnet \ + --dirname=$DEMO_PATH/image_classification_imagenet_resnet.inference.model \ + --fp16_dirname=$DEMO_PATH/float16_image_classification_imagenet_resnet.inference.model \ + --repeat=$REPEAT \ + --batch_size=$batch_size \ + --skip_cpu=true \ + 2>&1 | tee -a imagenet_resnet50_benchmark.log + + sleep 2m + + # Test inference benchmark of vgg16 on cifar10 + stdbuf -oL python float16_inference_demo.py \ + --data_set=cifar10 \ + --model=vgg \ + --threshold=0.001 \ + --repeat=1 \ + + $INFER_PATH/test_inference_image_classification_vgg \ + --dirname=$DEMO_PATH/image_classification_cifar10_vgg.inference.model \ + --fp16_dirname=$DEMO_PATH/float16_image_classification_cifar10_vgg.inference.model \ + --repeat=$REPEAT \ + --batch_size=$batch_size \ + --skip_cpu=true \ + 2>&1 | tee -a cifar10_vgg16_benchmark.log + + sleep 1m + + # Test inference benchmark of resnet32 on cifar10 + stdbuf -oL python float16_inference_demo.py \ + --data_set=cifar10 \ + --model=resnet \ + --threshold=0.001 \ + --repeat=1 \ + + $INFER_PATH/test_inference_image_classification_vgg \ + --dirname=$DEMO_PATH/image_classification_cifar10_resnet.inference.model \ + --fp16_dirname=$DEMO_PATH/float16_image_classification_cifar10_resnet.inference.model \ + --repeat=$REPEAT \ + --batch_size=$batch_size \ + --skip_cpu=true \ + 2>&1 | tee -a cifar10_resnet32_benchmark.log + + sleep 1m + +done diff --git a/paddle/contrib/inference/CMakeLists.txt b/paddle/contrib/inference/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..3beb93c4e7fd1ce4dd2131cb53cb6e89e0f10ebd --- /dev/null +++ b/paddle/contrib/inference/CMakeLists.txt @@ -0,0 +1,58 @@ +# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# + +if(APPLE) + set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-error=pessimizing-move") +endif(APPLE) + +function(inference_api_test TARGET_NAME TEST_SRC) + set(options "") + set(oneValueArgs "") + set(multiValueArgs ARGS) + cmake_parse_arguments(inference_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) + + set(PYTHON_TESTS_DIR ${PADDLE_BINARY_DIR}/python/paddle/fluid/tests) + set(arg_list "") + if(inference_test_ARGS) + foreach(arg ${inference_test_ARGS}) + list(APPEND arg_list "_${arg}") + endforeach() + else() + list(APPEND arg_list "_") + endif() + foreach(arg ${arg_list}) + string(REGEX REPLACE "^_$" "" arg "${arg}") + cc_test(${TARGET_NAME} + SRCS ${TEST_SRC} + DEPS paddle_fluid_api paddle_inference_api + ARGS --dirname=${PYTHON_TESTS_DIR}/book/) + # TODO(panyx0178): Figure out how to add word2vec and image_classification + # as deps. + # set_tests_properties(${TARGET_NAME} + # PROPERTIES DEPENDS ${DEP_TEST}) + endforeach() +endfunction(inference_api_test) + + +cc_library(paddle_inference_api + SRCS paddle_inference_api.cc paddle_inference_api_impl.cc + DEPS ${FLUID_CORE_MODULES} ${GLOB_OP_LIB}) + +cc_test(test_paddle_inference_api + SRCS test_paddle_inference_api.cc + DEPS paddle_inference_api) + +inference_api_test(test_paddle_inference_api_impl + test_paddle_inference_api_impl.cc) diff --git a/paddle/contrib/inference/README.md b/paddle/contrib/inference/README.md new file mode 100644 index 0000000000000000000000000000000000000000..20969fac6c8f894ffb4a02b48f795e2a0dcbd096 --- /dev/null +++ b/paddle/contrib/inference/README.md @@ -0,0 +1,27 @@ +# Embed Paddle Inference in Your Application + +Paddle inference offers the APIs in `C` and `C++` languages. + +One can easily deploy a model trained by Paddle following the steps as below: + +1. Optimize the native model; +2. Write some codes for deployment. + + +Let's explain the steps in detail. + +## Optimize the native Fluid Model + +The native model that get from the training phase needs to be optimized for that. + +- Clean the noise such as the cost operators that do not need inference; +- Prune unnecessary computation fork that has nothing to do with the output; +- Remove extraneous variables; +- Memory reuse for native Fluid executor; +- Translate the model storage format to some third-party engine's, so that the inference API can utilize the engine for acceleration; + +We have an official tool to do the optimization, call `paddle_inference_optimize --help` for more information. + +## Write some codes + +Read `paddle_inference_api.h` for more information. diff --git a/paddle/contrib/inference/paddle_inference_api.cc b/paddle/contrib/inference/paddle_inference_api.cc new file mode 100644 index 0000000000000000000000000000000000000000..d67e1e7667800d6dd00cb8915b0d6dc7c664970b --- /dev/null +++ b/paddle/contrib/inference/paddle_inference_api.cc @@ -0,0 +1,15 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/contrib/inference/paddle_inference_api.h" diff --git a/paddle/contrib/inference/paddle_inference_api.h b/paddle/contrib/inference/paddle_inference_api.h new file mode 100644 index 0000000000000000000000000000000000000000..b4c7f9bef4d2e83038ff223614a89e1b0493fc6f --- /dev/null +++ b/paddle/contrib/inference/paddle_inference_api.h @@ -0,0 +1,103 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +/* + * This file contains the definition of a simple Inference API for Paddle. + * + * ATTENTION: It requires some C++ features, for lower version C++ or C, we + * might release another API. + */ + +#pragma once + +#include +#include +#include + +namespace paddle { + +enum PaddleDType { + FLOAT32, + INT64, +}; + +struct PaddleBuf { + void* data; // pointer to the data memory. + size_t length; // number of memory bytes. +}; + +struct PaddleTensor { + std::string name; // variable name. + std::vector shape; + PaddleBuf data; // blob of data. + PaddleDType dtype; +}; + +/* + * A simple Inference API for Paddle. Currently this API can be used by + * non-sequence scenerios. + * TODO(Superjomn) Support another API for NLP-related usages. + */ +class PaddlePredictor { + public: + struct Config; + PaddlePredictor() = default; + PaddlePredictor(const PaddlePredictor&) = delete; + + // Predict an record. + // The caller should be responsible for allocating and releasing the memory of + // `inputs`. `inputs` should be alive until Run returns. caller should be + // responsible for releasing the memory of `output_data`. + virtual bool Run(const std::vector& inputs, + std::vector* output_data) = 0; + + // Clone a predictor that share the model weights, the Cloned predictor should + // be thread-safe. + virtual std::unique_ptr Clone() = 0; + + // Destroy the Predictor. + virtual ~PaddlePredictor() {} + + enum class EngineKind { + kNative = -1, // Use the native Fluid facility. + // TODO(Superjomn) support latter. + // kAnakin, // Use Anakin for inference. + // kTensorRT, // Use TensorRT for inference. + // kAutoMixedAnakin, // Automatically mix Fluid with Anakin. + // kAutoMixedTensorRT, // Automatically mix Fluid with TensorRT. + }; + + // The common configs for all the predictors. + struct Config { + std::string model_dir; // path to the model directory. + bool enable_engine{false}; // Enable to execute (part of) the model on + }; +}; + +struct NativeConfig : public PaddlePredictor::Config { + bool use_gpu{false}; + int device; + float fraction_of_gpu_memory; + std::string prog_file; + std::string param_file; + bool share_variables; +}; + +// A factory to help create difference predictor. +template < + typename ConfigT, + PaddlePredictor::EngineKind engine = PaddlePredictor::EngineKind::kNative> +std::unique_ptr CreatePaddlePredictor(const ConfigT& config); + +} // namespace paddle diff --git a/paddle/contrib/inference/paddle_inference_api_impl.cc b/paddle/contrib/inference/paddle_inference_api_impl.cc new file mode 100644 index 0000000000000000000000000000000000000000..989252f69e42778dfd791cdee02c550f2aa78803 --- /dev/null +++ b/paddle/contrib/inference/paddle_inference_api_impl.cc @@ -0,0 +1,269 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include +#include +#include +#include +#include +#include +#include +#include + +#include "paddle/contrib/inference/paddle_inference_api_impl.h" + +namespace paddle { +namespace { + +// Timer for timer +class Timer { + public: + double start; + double startu; + void tic() { + struct timeval tp; + gettimeofday(&tp, NULL); + start = tp.tv_sec; + startu = tp.tv_usec; + } + double toc() { + struct timeval tp; + gettimeofday(&tp, NULL); + double used_time_ms = + (tp.tv_sec - start) * 1000.0 + (tp.tv_usec - startu) / 1000.0; + return used_time_ms; + } +}; + +template +std::string num2str(T a) { + std::stringstream istr; + istr << a; + return istr.str(); +} +} // namespace + +bool NativePaddlePredictor::Init() { + VLOG(3) << "Predictor::init()"; + + // TODO(panyx0718): Should CPU vs GPU device be decided by id? + if (config_.device >= 0) { + place_ = paddle::platform::CUDAPlace(config_.device); + } else { + place_ = paddle::platform::CPUPlace(); + } + paddle::framework::InitDevices(false); + executor_.reset(new paddle::framework::Executor(place_)); + scope_.reset(new paddle::framework::Scope()); + + // Initialize the inference program + if (!config_.model_dir.empty()) { + // Parameters are saved in separate files sited in + // the specified `dirname`. + inference_program_ = paddle::inference::Load( + executor_.get(), scope_.get(), config_.model_dir); + } else if (!config_.prog_file.empty() && !config_.param_file.empty()) { + // All parameters are saved in a single file. + // The file names should be consistent with that used + // in Python API `fluid.io.save_inference_model`. + inference_program_ = paddle::inference::Load( + executor_.get(), scope_.get(), config_.prog_file, config_.param_file); + } else { + LOG(ERROR) << "fail to load inference model."; + return false; + } + ctx_ = executor_->Prepare(*inference_program_, 0); + + // Create variables + // TODO(panyx0718): Why need to test share_variables here? + if (config_.share_variables) { + executor_->CreateVariables(*inference_program_, scope_.get(), 0); + } + // Get the feed_target_names and fetch_target_names + feed_target_names_ = inference_program_->GetFeedTargetNames(); + fetch_target_names_ = inference_program_->GetFetchTargetNames(); + return true; +} + +bool NativePaddlePredictor::Run(const std::vector &inputs, + std::vector *output_data) { + VLOG(3) << "Predictor::predict"; + Timer timer; + timer.tic(); + // set feed variable + std::map feed_targets; + std::vector feeds; + if (!SetFeed(inputs, &feeds)) { + LOG(ERROR) << "fail to set feed"; + return false; + } + for (size_t i = 0; i < feed_target_names_.size(); ++i) { + feed_targets[feed_target_names_[i]] = &feeds[i]; + } + // get fetch variable + std::map fetch_targets; + std::vector fetchs; + fetchs.resize(fetch_target_names_.size()); + for (size_t i = 0; i < fetch_target_names_.size(); ++i) { + fetch_targets[fetch_target_names_[i]] = &fetchs[i]; + } + // Run the inference program + // if share variables, we need not create variables + executor_->RunPreparedContext(ctx_.get(), + scope_.get(), + &feed_targets, + &fetch_targets, + !config_.share_variables); + if (!GetFetch(fetchs, output_data)) { + LOG(ERROR) << "fail to get fetchs"; + return false; + } + VLOG(3) << "predict cost: " << timer.toc() << "ms"; + return true; +} + +std::unique_ptr NativePaddlePredictor::Clone() { + VLOG(3) << "Predictor::clone"; + std::unique_ptr cls(new NativePaddlePredictor(config_)); + + if (!dynamic_cast(cls.get())->Init()) { + LOG(ERROR) << "fail to call Init"; + return nullptr; + } + // fix manylinux compile error. + return std::move(cls); +} + +bool NativePaddlePredictor::SetFeed(const std::vector &inputs, + std::vector *feeds) { + VLOG(3) << "Predictor::set_feed"; + if (inputs.size() != feed_target_names_.size()) { + LOG(ERROR) << "wrong feed input size."; + return false; + } + for (size_t i = 0; i < feed_target_names_.size(); ++i) { + framework::LoDTensor input; + framework::DDim ddim = framework::make_ddim(inputs[i].shape); + void *input_ptr; + if (inputs[i].dtype == PaddleDType::INT64) { + input_ptr = input.mutable_data(ddim, platform::CPUPlace()); + } else if (inputs[i].dtype == PaddleDType::FLOAT32) { + input_ptr = input.mutable_data(ddim, platform::CPUPlace()); + } else { + LOG(ERROR) << "unsupported feed type " << inputs[i].dtype; + return false; + } + + // TODO(panyx0718): Init LoDTensor from existing memcpy to save a copy. + std::memcpy(static_cast(input_ptr), + inputs[i].data.data, + inputs[i].data.length); + feeds->push_back(input); + } + return true; +} + +bool NativePaddlePredictor::GetFetch( + const std::vector &fetchs, + std::vector *outputs) { + VLOG(3) << "Predictor::get_fetch"; + outputs->resize(fetchs.size()); + for (size_t i = 0; i < fetchs.size(); ++i) { + // TODO(panyx0718): Support fetch of other types. + if (fetchs[i].type() != typeid(float)) { + LOG(ERROR) << "only support fetching float now."; + return false; + } + std::vector shape; + auto dims_i = fetchs[i].dims(); + auto lod = fetchs[i].lod(); + const float *output_ptr = fetchs[i].data(); + // const int64_t* output_ptr = fetchs[i].data(); + auto num = fetchs[i].numel(); + std::vector data; + if (0 == lod.size()) { + std::copy(output_ptr, output_ptr + num, std::back_inserter(data)); + for (int j = 0; j < dims_i.size(); ++j) { + shape.push_back(dims_i[j]); + } + } else { + // for batch detection + // image[0] -> output[0] shape {145, 6} + // image[1] -> output[1] shape {176, 6} + // then, + // the batch output shape {321, 6} + // the lod {{0, 145, 321}} + // so we should append output[0] to {176, 6} + size_t max_dim = 0; + for (size_t j = 1; j < lod[0].size(); j++) { + max_dim = std::max(max_dim, lod[0][j] - lod[0][j - 1]); + } + size_t common_dim = lod[0].back() == 0 ? 0 : num / lod[0].back(); + if (max_dim > 0) { + data.resize((lod[0].size() - 1) * max_dim * common_dim, 0); + } + for (size_t j = 1; j < lod[0].size(); j++) { + size_t start = lod[0][j - 1] * common_dim; + size_t end = lod[0][j] * common_dim; + if (end > start) { + std::copy(output_ptr + start, + output_ptr + end, + data.begin() + (j - 1) * max_dim * common_dim); + } + } + shape.push_back(lod[0].size() - 1); + shape.push_back(max_dim); + for (int j = 1; j < dims_i.size(); ++j) { + shape.push_back(dims_i[j]); + } + } + + outputs->at(i).shape = shape; + outputs->at(i).data.length = sizeof(float) * data.size(); + outputs->at(i).data.data = malloc(outputs->at(i).data.length); + std::memcpy( + outputs->at(i).data.data, data.data(), outputs->at(i).data.length); + outputs->at(i).dtype = PaddleDType::FLOAT32; + // TODO(panyx0718): support other types? fill tensor name? avoid a copy. + } + return true; +} + +template <> +std::unique_ptr +CreatePaddlePredictor( + const NativeConfig &config) { + VLOG(3) << "create NativePaddlePredictor"; + if (config.use_gpu) { + // 1. GPU memeroy + std::vector flags; + if (config.fraction_of_gpu_memory >= 0.0f || + config.fraction_of_gpu_memory <= 0.95f) { + flags.push_back("dummpy"); + std::string flag = "--fraction_of_gpu_memory_to_use=" + + num2str(config.fraction_of_gpu_memory); + flags.push_back(flag); + VLOG(3) << "set flag: " << flag; + framework::InitGflags(flags); + } + } + + std::unique_ptr predictor(new NativePaddlePredictor(config)); + if (!dynamic_cast(predictor.get())->Init()) { + return nullptr; + } + return std::move(predictor); +} + +} // namespace paddle diff --git a/paddle/contrib/inference/paddle_inference_api_impl.h b/paddle/contrib/inference/paddle_inference_api_impl.h new file mode 100644 index 0000000000000000000000000000000000000000..84707e223d7aa3d1ebca933923e932b3973613ae --- /dev/null +++ b/paddle/contrib/inference/paddle_inference_api_impl.h @@ -0,0 +1,62 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once + +#include +#include +#include +#include + +#include "paddle/contrib/inference/paddle_inference_api.h" + +#include "paddle/fluid/framework/ddim.h" +#include "paddle/fluid/framework/init.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/inference/io.h" +#include "paddle/fluid/platform/profiler.h" + +namespace paddle { + +class NativePaddlePredictor : public PaddlePredictor { + public: + explicit NativePaddlePredictor(const NativeConfig &config) + : config_(config) {} + + bool Init(); + + bool Run(const std::vector &inputs, + std::vector *output_data) override; + + std::unique_ptr Clone() override; + + ~NativePaddlePredictor() override{}; + + private: + bool SetFeed(const std::vector &input_datas, + std::vector *feeds); + bool GetFetch(const std::vector &fetchs, + std::vector *output_data); + + NativeConfig config_; + platform::Place place_; + std::unique_ptr executor_; + std::unique_ptr scope_; + std::unique_ptr ctx_; + std::unique_ptr inference_program_; + std::vector feed_target_names_; + std::vector fetch_target_names_; +}; + +} // namespace paddle diff --git a/paddle/contrib/inference/test_paddle_inference_api.cc b/paddle/contrib/inference/test_paddle_inference_api.cc new file mode 100644 index 0000000000000000000000000000000000000000..bc7faab6e208a66d7a56e41a56bd743c7644eea2 --- /dev/null +++ b/paddle/contrib/inference/test_paddle_inference_api.cc @@ -0,0 +1,64 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/contrib/inference/paddle_inference_api.h" + +#include +#include + +namespace paddle { + +/* + * Do not use this, just a demo indicating how to customize a config for a + * specific predictor. + */ +struct DemoConfig : public PaddlePredictor::Config { + float other_config; +}; + +/* + * Do not use this, just a demo indicating how to customize a Predictor. + */ +class DemoPredictor : public PaddlePredictor { + public: + explicit DemoPredictor(const DemoConfig &config) { + LOG(INFO) << "I get other_config " << config.other_config; + } + bool Run(const std::vector &inputs, + std::vector *output_data) override { + LOG(INFO) << "Run"; + return false; + } + + std::unique_ptr Clone() override { return nullptr; } + + ~DemoPredictor() override {} +}; + +template <> +std::unique_ptr CreatePaddlePredictor( + const DemoConfig &config) { + std::unique_ptr x(new DemoPredictor(config)); + return x; +} + +TEST(paddle_inference_api, demo) { + DemoConfig config; + config.other_config = 1.7; + auto predictor = CreatePaddlePredictor(config); + std::vector outputs; + predictor->Run({}, &outputs); +} + +} // namespace paddle diff --git a/paddle/contrib/inference/test_paddle_inference_api_impl.cc b/paddle/contrib/inference/test_paddle_inference_api_impl.cc new file mode 100644 index 0000000000000000000000000000000000000000..5240fc2f20211ac5d38c57b71db31d04a6dc536a --- /dev/null +++ b/paddle/contrib/inference/test_paddle_inference_api_impl.cc @@ -0,0 +1,153 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include + +#include "gflags/gflags.h" +#include "paddle/contrib/inference/paddle_inference_api_impl.h" +#include "paddle/fluid/inference/tests/test_helper.h" + +DEFINE_string(dirname, "", "Directory of the inference model."); + +namespace paddle { + +PaddleTensor LodTensorToPaddleTensor(framework::LoDTensor* t) { + PaddleTensor pt; + pt.data.data = t->data(); + + if (t->type() == typeid(int64_t)) { + pt.data.length = t->numel() * sizeof(int64_t); + pt.dtype = PaddleDType::INT64; + } else if (t->type() == typeid(float)) { + pt.data.length = t->numel() * sizeof(float); + pt.dtype = PaddleDType::FLOAT32; + } else { + LOG(FATAL) << "unsupported type."; + } + pt.shape = framework::vectorize2int(t->dims()); + return pt; +} + +NativeConfig GetConfig() { + NativeConfig config; + config.model_dir = FLAGS_dirname + "word2vec.inference.model"; + LOG(INFO) << "dirname " << config.model_dir; + config.fraction_of_gpu_memory = 0.15; + config.use_gpu = true; + config.device = 0; + config.share_variables = true; + return config; +} + +TEST(paddle_inference_api_impl, word2vec) { + NativeConfig config = GetConfig(); + auto predictor = CreatePaddlePredictor(config); + + framework::LoDTensor first_word, second_word, third_word, fourth_word; + framework::LoD lod{{0, 1}}; + int64_t dict_size = 2073; // The size of dictionary + + SetupLoDTensor(&first_word, lod, static_cast(0), dict_size - 1); + SetupLoDTensor(&second_word, lod, static_cast(0), dict_size - 1); + SetupLoDTensor(&third_word, lod, static_cast(0), dict_size - 1); + SetupLoDTensor(&fourth_word, lod, static_cast(0), dict_size - 1); + + std::vector paddle_tensor_feeds; + paddle_tensor_feeds.push_back(LodTensorToPaddleTensor(&first_word)); + paddle_tensor_feeds.push_back(LodTensorToPaddleTensor(&second_word)); + paddle_tensor_feeds.push_back(LodTensorToPaddleTensor(&third_word)); + paddle_tensor_feeds.push_back(LodTensorToPaddleTensor(&fourth_word)); + + std::vector outputs; + ASSERT_TRUE(predictor->Run(paddle_tensor_feeds, &outputs)); + ASSERT_EQ(outputs.size(), 1UL); + size_t len = outputs[0].data.length; + float* data = static_cast(outputs[0].data.data); + for (int j = 0; j < len / sizeof(float); ++j) { + ASSERT_LT(data[j], 1.0); + ASSERT_GT(data[j], -1.0); + } + + std::vector cpu_feeds; + cpu_feeds.push_back(&first_word); + cpu_feeds.push_back(&second_word); + cpu_feeds.push_back(&third_word); + cpu_feeds.push_back(&fourth_word); + + framework::LoDTensor output1; + std::vector cpu_fetchs1; + cpu_fetchs1.push_back(&output1); + + TestInference(config.model_dir, cpu_feeds, cpu_fetchs1); + + float* lod_data = output1.data(); + for (size_t i = 0; i < output1.numel(); ++i) { + EXPECT_LT(lod_data[i] - data[i], 1e-3); + EXPECT_GT(lod_data[i] - data[i], -1e-3); + } + + free(outputs[0].data.data); +} + +TEST(paddle_inference_api_impl, image_classification) { + int batch_size = 2; + bool use_mkldnn = false; + bool repeat = false; + NativeConfig config = GetConfig(); + config.model_dir = + FLAGS_dirname + "image_classification_resnet.inference.model"; + + const bool is_combined = false; + std::vector> feed_target_shapes = + GetFeedTargetShapes(config.model_dir, is_combined); + + framework::LoDTensor input; + // Use normilized image pixels as input data, + // which should be in the range [0.0, 1.0]. + feed_target_shapes[0][0] = batch_size; + framework::DDim input_dims = framework::make_ddim(feed_target_shapes[0]); + SetupTensor( + &input, input_dims, static_cast(0), static_cast(1)); + std::vector cpu_feeds; + cpu_feeds.push_back(&input); + + framework::LoDTensor output1; + std::vector cpu_fetchs1; + cpu_fetchs1.push_back(&output1); + + TestInference(config.model_dir, + cpu_feeds, + cpu_fetchs1, + repeat, + is_combined, + use_mkldnn); + + auto predictor = CreatePaddlePredictor(config); + std::vector paddle_tensor_feeds; + paddle_tensor_feeds.push_back(LodTensorToPaddleTensor(&input)); + + std::vector outputs; + ASSERT_TRUE(predictor->Run(paddle_tensor_feeds, &outputs)); + ASSERT_EQ(outputs.size(), 1UL); + size_t len = outputs[0].data.length; + float* data = static_cast(outputs[0].data.data); + float* lod_data = output1.data(); + for (size_t j = 0; j < len / sizeof(float); ++j) { + EXPECT_NEAR(lod_data[j], data[j], 1e-3); + } + free(data); +} + +} // namespace paddle diff --git a/paddle/cuda/CMakeLists.txt b/paddle/cuda/CMakeLists.txt index efd1b7a73e1655f95eb83a5e2f59e82cbf7eba16..9bbb8de78e09829d24faf42c360811084981578f 100755 --- a/paddle/cuda/CMakeLists.txt +++ b/paddle/cuda/CMakeLists.txt @@ -87,8 +87,3 @@ else() endif() add_dependencies(paddle_cuda paddle_proto ${external_project_dependencies}) - -add_style_check_target(paddle_cuda - ${CUDA_SOURCES} - ${CUDA_HEADERS} - ${CUDA_CXX_SOURCES}) diff --git a/paddle/cuda/include/hl_activation_functions.h b/paddle/cuda/include/hl_activation_functions.h index 29ec248420058db08bd1932f702d26074d49f38c..66a69db545b541409f895820ad621a2a9a684e20 100644 --- a/paddle/cuda/include/hl_activation_functions.h +++ b/paddle/cuda/include/hl_activation_functions.h @@ -31,7 +31,7 @@ namespace hppl { */ template class Active { -public: + public: typedef T (*forward)(T); typedef T (*backward)(T, T); }; diff --git a/paddle/cuda/include/hl_base.h b/paddle/cuda/include/hl_base.h index 6c4f09dacb47c431db2e2610a3e61390a82dcba0..77f5d82dbe2cad183491033736bac85961b6d320 100644 --- a/paddle/cuda/include/hl_base.h +++ b/paddle/cuda/include/hl_base.h @@ -12,8 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#ifndef HL_BASE_H_ -#define HL_BASE_H_ +#pragma once #include @@ -207,8 +206,8 @@ typedef struct { #ifdef __NVCC__ -#include "cuda_runtime.h" -#include "hl_cuda.h" +#include +#include "paddle/cuda/include/hl_cuda.h" #include "paddle/utils/Logging.h" extern __thread bool g_sync_flag; @@ -228,6 +227,24 @@ extern __thread cudaStream_t default_stream; << "CUDA error: " << hl_get_device_error_string((size_t)err); \ } -#endif /* __NVCC__ */ +// __shfl has been deprecated as of CUDA 9.0. +#if CUDA_VERSION < 9000 +template +__forceinline__ __device__ T __shfl_down_sync(unsigned, T val, int delta) { + return __shfl_down(val, delta); +} -#endif /* HL_BASE_H_ */ +template +__forceinline__ __device__ T +__shfl_sync(unsigned, T val, int src_line, int width) { + return __shfl(val, src_line, width); +} + +#define CREATE_SHFL_MASK(mask, predicate) mask = 0u; +#else +#define FULL_WARP_MASK 0xFFFFFFFF +#define CREATE_SHFL_MASK(mask, predicate) \ + mask = __ballot_sync(FULL_WARP_MASK, (predicate)) +#endif + +#endif // __NVCC__ diff --git a/paddle/cuda/include/hl_tensor_ops.h b/paddle/cuda/include/hl_tensor_ops.h index 85a022ff5e26daab97be52b7ea9814c6b8078561..bc5e5da53d5c6ac2bae3b0067f46e39accd1b9d8 100644 --- a/paddle/cuda/include/hl_tensor_ops.h +++ b/paddle/cuda/include/hl_tensor_ops.h @@ -23,128 +23,128 @@ namespace unary { template class add_scale { -private: + private: const T p; -public: + public: INLINE add_scale(const T s) : p(s) {} INLINE T operator()(const T a) const { return a + p; } }; template class sub_scale { -private: + private: const T p; -public: + public: INLINE sub_scale(const T s) : p(s) {} INLINE T operator()(const T a) const { return a - p; } }; template class mul_scale { -private: + private: const T p; -public: + public: INLINE mul_scale(const T s) : p(s) {} INLINE T operator()(const T a) const { return a * p; } }; template class div_scale { -private: + private: const T p; -public: + public: INLINE div_scale(const T s) : p(s) {} INLINE T operator()(const T a) const { return a / p; } }; template class neg { -public: + public: INLINE T operator()(const T a) const { return -a; } }; template class exp_op { -public: + public: INLINE T operator()(const T a) const { return std::exp(a); } }; template class log_op { -public: + public: INLINE T operator()(const T a) const { return std::log(a); } }; template class sqrt_op { -public: + public: INLINE T operator()(const T a) const { return std::sqrt(a); } }; template class square { -public: + public: INLINE T operator()(const T a) const { return a * a; } }; template class reciprocal { -public: + public: INLINE T operator()(const T a) const { return T(1) / a; } }; template class abs { -public: + public: INLINE T operator()(const T a) const { return a > 0 ? a : -a; } }; template class sign { -public: + public: INLINE T operator()(const T a) const { return (a > 0) - (a < 0); } }; template class min { -private: + private: const T p; -public: + public: INLINE min(const T s) : p(s) {} INLINE T operator()(const T a) const { return a > p ? p : a; } }; template class max { -private: + private: const T p; -public: + public: INLINE max(const T s) : p(s) {} INLINE T operator()(const T a) const { return a < p ? p : a; } }; template class pow_op { -private: + private: const T p; -public: + public: INLINE pow_op(const T s) : p(s) {} INLINE T operator()(const T a) const { return std::pow(a, p); } }; template class constant { -private: + private: const T p; -public: + public: INLINE constant(const T s) : p(s) {} INLINE T operator()(int i) const { return p; } INLINE T operator()(int i, int j) const { return p; } @@ -152,80 +152,80 @@ public: template class cmp_eq { -private: + private: const T p; -public: + public: INLINE cmp_eq(const T s) : p(s) {} INLINE bool operator()(const T a) const { return a == p; } }; template class cmp_ne { -private: + private: const T p; -public: + public: INLINE cmp_ne(const T s) : p(s) {} INLINE bool operator()(const T a) const { return a != p; } }; template class cmp_le { -private: + private: const T p; -public: + public: INLINE cmp_le(const T s) : p(s) {} INLINE bool operator()(const T a) const { return a <= p; } }; template class cmp_lt { -private: + private: const T p; -public: + public: INLINE cmp_lt(const T s) : p(s) {} INLINE bool operator()(const T a) const { return a < p; } }; template class cmp_ge { -private: + private: const T p; -public: + public: INLINE cmp_ge(const T s) : p(s) {} INLINE bool operator()(const T a) const { return a >= p; } }; template class cmp_gt { -private: + private: const T p; -public: + public: INLINE cmp_gt(const T s) : p(s) {} INLINE bool operator()(const T a) const { return a > p; } }; template class and_op { -private: + private: const T p; -public: + public: INLINE and_op(const T s) : p(s) {} INLINE bool operator()(const T a) const { return a && p; } }; template class or_op { -private: + private: const T p; -public: + public: INLINE or_op(const T s) : p(s) {} INLINE bool operator()(const T a) const { return a || p; } }; @@ -235,96 +235,96 @@ public: namespace binary { template class add { -public: + public: INLINE T operator()(const T a, const T b) const { return a + b; } }; template class add_scale { -private: + private: const T p1; const T p2; -public: + public: INLINE add_scale(const T s1, const T s2) : p1(s1), p2(s2) {} INLINE T operator()(const T a, const T b) const { return p1 * a + p2 * b; } }; template class sub { -public: + public: INLINE T operator()(const T a, const T b) const { return a - b; } }; template class mul { -public: + public: INLINE T operator()(const T a, const T b) const { return a * b; } }; template class div { -public: + public: INLINE T operator()(const T a, const T b) const { return a / b; } }; template class cmp_eq { -public: + public: INLINE bool operator()(const T a, const T b) const { return a == b; } }; template class cmp_ne { -public: + public: INLINE bool operator()(const T a, const T b) const { return a != b; } }; template class cmp_le { -public: + public: INLINE bool operator()(const T a, const T b) const { return a <= b; } }; template class cmp_lt { -public: + public: INLINE bool operator()(const T a, const T b) const { return a < b; } }; template class cmp_ge { -public: + public: INLINE bool operator()(const T a, const T b) const { return a >= b; } }; template class cmp_gt { -public: + public: INLINE bool operator()(const T a, const T b) const { return a > b; } }; template class and_op { -public: + public: INLINE bool operator()(const T a, const T b) const { return a && b; } }; template class or_op { -public: + public: INLINE bool operator()(const T a, const T b) const { return a || b; } }; template class min { -public: + public: INLINE T operator()(const T a, const T b) const { return a > b ? b : a; } }; template class max { -public: + public: INLINE T operator()(const T a, const T b) const { return a < b ? b : a; } }; @@ -332,7 +332,7 @@ public: #ifndef PADDLE_TYPE_DOUBLE template <> class add<__m128> { -public: + public: INLINE __m128 operator()(const __m128 a, const __m128 b) const { return _mm_add_ps(a, b); } @@ -340,11 +340,11 @@ public: template <> class add_scale<__m128> { -private: + private: const __m128 p1; const __m128 p2; -public: + public: INLINE add_scale(const __m128 s1, const __m128 s2) : p1(s1), p2(s2) {} INLINE __m128 operator()(const __m128 a, const __m128 b) const { return _mm_add_ps(_mm_mul_ps(p1, a), _mm_mul_ps(p2, b)); @@ -353,7 +353,7 @@ public: template <> class sub<__m128> { -public: + public: INLINE __m128 operator()(const __m128 a, const __m128 b) const { return _mm_sub_ps(a, b); } @@ -361,7 +361,7 @@ public: template <> class mul<__m128> { -public: + public: INLINE __m128 operator()(const __m128 a, const __m128 b) const { return _mm_mul_ps(a, b); } @@ -369,7 +369,7 @@ public: template <> class div<__m128> { -public: + public: INLINE __m128 operator()(const __m128 a, const __m128 b) const { return _mm_div_ps(a, b); } @@ -377,7 +377,7 @@ public: template <> class min<__m128> { -public: + public: INLINE __m128 operator()(const __m128 a, const __m128 b) const { return _mm_min_ps(a, b); } @@ -385,7 +385,7 @@ public: template <> class max<__m128> { -public: + public: INLINE __m128 operator()(const __m128 a, const __m128 b) const { return _mm_max_ps(a, b); } @@ -393,7 +393,7 @@ public: #else template <> class add<__m128d> { -public: + public: INLINE __m128d operator()(const __m128d a, const __m128d b) const { return _mm_add_pd(a, b); } @@ -401,11 +401,11 @@ public: template <> class add_scale<__m128d> { -private: + private: const __m128d p1; const __m128d p2; -public: + public: INLINE add_scale(const __m128d s1, const __m128d s2) : p1(s1), p2(s2) {} INLINE __m128d operator()(const __m128d a, const __m128d b) const { return _mm_add_pd(_mm_mul_pd(p1, a), _mm_mul_pd(p2, b)); @@ -414,7 +414,7 @@ public: template <> class sub<__m128d> { -public: + public: INLINE __m128d operator()(const __m128d a, const __m128d b) const { return _mm_sub_pd(a, b); } @@ -422,7 +422,7 @@ public: template <> class mul<__m128d> { -public: + public: INLINE __m128d operator()(const __m128d a, const __m128d b) const { return _mm_mul_pd(a, b); } @@ -430,7 +430,7 @@ public: template <> class div<__m128d> { -public: + public: INLINE __m128d operator()(const __m128d a, const __m128d b) const { return _mm_div_pd(a, b); } @@ -438,7 +438,7 @@ public: template <> class min<__m128d> { -public: + public: INLINE __m128d operator()(const __m128d a, const __m128d b) const { return _mm_min_pd(a, b); } @@ -446,7 +446,7 @@ public: template <> class max<__m128d> { -public: + public: INLINE __m128d operator()(const __m128d a, const __m128d b) const { return _mm_max_pd(a, b); } @@ -458,7 +458,7 @@ public: #ifndef PADDLE_TYPE_DOUBLE template <> class add { -public: + public: INLINE float32x4_t operator()(const float32x4_t a, const float32x4_t b) const { return vaddq_f32(a, b); @@ -467,11 +467,11 @@ public: template <> class add_scale { -private: + private: const float32x4_t p1; const float32x4_t p2; -public: + public: INLINE add_scale(const float32x4_t s1, const float32x4_t s2) : p1(s1), p2(s2) {} INLINE float32x4_t operator()(const float32x4_t a, @@ -482,7 +482,7 @@ public: template <> class sub { -public: + public: INLINE float32x4_t operator()(const float32x4_t a, const float32x4_t b) const { return vsubq_f32(a, b); @@ -491,7 +491,7 @@ public: template <> class mul { -public: + public: INLINE float32x4_t operator()(const float32x4_t a, const float32x4_t b) const { return vmulq_f32(a, b); @@ -500,7 +500,7 @@ public: template <> class div { -public: + public: INLINE float32x4_t operator()(const float32x4_t a, const float32x4_t b) const { float32x4_t tmp = vrecpeq_f32(b); @@ -510,7 +510,7 @@ public: template <> class min { -public: + public: INLINE float32x4_t operator()(const float32x4_t a, const float32x4_t b) const { return vminq_f32(a, b); @@ -519,7 +519,7 @@ public: template <> class max { -public: + public: INLINE float32x4_t operator()(const float32x4_t a, const float32x4_t b) const { return vmaxq_f32(a, b); diff --git a/paddle/cuda/src/hl_cuda_lstm.cu b/paddle/cuda/src/hl_cuda_lstm.cu index 21c0c26b6ef0420b1a719736a66eeb8114ed9680..b8c4e433a118fb1c5af753751f91c34543b1114c 100644 --- a/paddle/cuda/src/hl_cuda_lstm.cu +++ b/paddle/cuda/src/hl_cuda_lstm.cu @@ -30,7 +30,7 @@ bool hl_lstm_sequence_parallel(int frameSize) { } class frameValue { -public: + public: real *value_; __device__ frameValue(real *value) : value_(value) {} template @@ -341,12 +341,15 @@ void hl_lstm_parallel_forward(real *gateValue, } __device__ __forceinline__ void transpose_32x32(real a[], const int idx) { - int addr = idx % 32; + const int warp_size = 32; + int addr = idx % warp_size; + unsigned mask = 0u; + CREATE_SHFL_MASK(mask, addr < warp_size); #pragma unroll for (int k = 1; k < 32; k++) { - // rSrc[k] = __shfl(rSrc[k], (threadIdx.x + k) % 32, 32); - addr = __shfl(addr, (idx + 1) % 32, 32); - a[k] = __shfl(a[k], addr, 32); + // rSrc[k] = __shfl_sync(rSrc[k], (threadIdx.x + k) % 32, 32); + addr = __shfl_sync(mask, addr, (idx + 1) % 32, 32); + a[k] = __shfl_sync(mask, a[k], addr, 32); } #pragma unroll @@ -360,10 +363,11 @@ __device__ __forceinline__ void transpose_32x32(real a[], const int idx) { } addr = (32 - idx) % 32; + CREATE_SHFL_MASK(mask, idx % 32 < warp_size); #pragma unroll for (int k = 0; k < 32; k++) { - a[k] = __shfl(a[k], addr, 32); - addr = __shfl(addr, (idx + 31) % 32, 32); + a[k] = __shfl_sync(mask, a[k], addr, 32); + addr = __shfl_sync(mask, addr, (idx + 31) % 32, 32); } } diff --git a/paddle/cuda/src/hl_top_k.cu b/paddle/cuda/src/hl_top_k.cu index fea8712a773b1524022f4bba626cf5044edebef6..b17290557c4f635a963d88525409b9373b057a4b 100644 --- a/paddle/cuda/src/hl_top_k.cu +++ b/paddle/cuda/src/hl_top_k.cu @@ -12,9 +12,9 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include "hl_base.h" -#include "hl_sparse.ph" -#include "hl_top_k.h" +#include "paddle/cuda/include/hl_base.h" +#include "paddle/cuda/include/hl_sparse.ph" +#include "paddle/cuda/include/hl_top_k.h" #include "paddle/utils/Logging.h" // using namespace hppl; @@ -244,13 +244,17 @@ __device__ __forceinline__ void blockReduce(Pair* shTopK, if (--beamSize == 0) break; __syncthreads(); + // NOTE(zcd): temporary solution + unsigned mask = 0u; + CREATE_SHFL_MASK(mask, true); + if (tid == maxId[0]) { if (beam < maxLength) { shTopK[tid] = topK[beam]; } } if (maxId[0] / 32 == warp) { - if (__shfl(beam, (maxId[0]) % 32, 32) == maxLength) break; + if (__shfl_sync(mask, beam, (maxId[0]) % 32, 32) == maxLength) break; } } } diff --git a/paddle/fluid/framework/CMakeLists.txt b/paddle/fluid/framework/CMakeLists.txt index 340b891e41671df7e61a4a66ec538d4603bb9842..ed1e70c6460b513c1d2e1add18ac037f71d36944 100644 --- a/paddle/fluid/framework/CMakeLists.txt +++ b/paddle/fluid/framework/CMakeLists.txt @@ -5,11 +5,11 @@ proto_library(framework_proto SRCS framework.proto) cc_library(ddim SRCS ddim.cc DEPS eigen3 boost) cc_test(ddim_test SRCS ddim_test.cc DEPS ddim) nv_test(dim_test SRCS dim_test.cu DEPS ddim) - +cc_library(data_type SRCS data_type.cc DEPS framework_proto ddim device_context) if(WITH_GPU) - nv_library(tensor SRCS tensor.cc tensor_util.cu DEPS ddim place memory device_context framework_proto) + nv_library(tensor SRCS tensor.cc tensor_util.cu DEPS place memory data_type) else() - cc_library(tensor SRCS tensor.cc tensor_util.cc DEPS ddim place memory device_context framework_proto) + cc_library(tensor SRCS tensor.cc tensor_util.cc DEPS place memory data_type) endif() cc_test(tensor_test SRCS tensor_test.cc DEPS tensor) @@ -57,7 +57,7 @@ cc_library(data_transform SRCS data_transform.cc DEPS math_function tensor cc_library(attribute SRCS attribute.cc DEPS framework_proto boost) cc_test(program_desc_test SRCS program_desc_test.cc DEPS proto_desc device_context) -cc_library(op_proto_maker SRCS op_proto_maker.cc DEPS framework_proto attribute) +cc_library(op_proto_maker SRCS op_proto_maker.cc DEPS framework_proto attribute glog) cc_test(op_proto_maker_test SRCS op_proto_maker_test.cc DEPS op_proto_maker) cc_library(op_info SRCS op_info.cc DEPS attribute framework_proto) cc_library(shape_inference SRCS shape_inference.cc DEPS ddim attribute device_context) diff --git a/paddle/fluid/framework/block_desc.cc b/paddle/fluid/framework/block_desc.cc index 9f753478d8ecf12441d4b1745a9f6750a1038e31..e83d1a218d32cfc3f81d2b4d61da7215a3237d06 100644 --- a/paddle/fluid/framework/block_desc.cc +++ b/paddle/fluid/framework/block_desc.cc @@ -134,6 +134,11 @@ OpDesc *BlockDesc::PrependOp() { return ops_.front().get(); } +void BlockDesc::PrependAllocatedOp(std::unique_ptr &&op_desc) { + need_update_ = true; + ops_.emplace_front(std::move(op_desc)); +} + OpDesc *BlockDesc::InsertOp(size_t index) { need_update_ = true; auto it = ops_.begin() + index; @@ -143,7 +148,7 @@ OpDesc *BlockDesc::InsertOp(size_t index) { } void BlockDesc::RemoveOp(size_t s, size_t e) { - if (ops_.begin() + s == ops_.end() || ops_.begin() + e == ops_.end()) { + if (ops_.begin() + s >= ops_.end() || ops_.begin() + e > ops_.end()) { return; } need_update_ = true; @@ -164,17 +169,13 @@ void BlockDesc::Flush() { } if (need_update_) { - auto &op_field = *this->desc_->mutable_ops(); - this->ClearPBOps(); - op_field.Reserve(static_cast(ops_.size())); + this->desc_->mutable_ops()->Clear(); for (auto &op_desc : ops_) { - op_field.AddAllocated(op_desc->Proto()); + this->desc_->mutable_ops()->Add()->CopyFrom(*op_desc->Proto()); } - auto &var_field = *this->desc_->mutable_vars(); - this->ClearPBVars(); - var_field.Reserve(static_cast(vars_.size())); + this->desc_->mutable_vars()->Clear(); for (auto &var_desc : vars_) { - var_field.AddAllocated(var_desc.second->Proto()); + this->desc_->mutable_vars()->Add()->CopyFrom(*var_desc.second->Proto()); } need_update_ = false; } @@ -212,22 +213,6 @@ BlockDesc::BlockDesc(const BlockDesc &other, proto::BlockDesc *desc, } } -void BlockDesc::ClearPBOps() { - auto ops = this->desc_->mutable_ops(); - while (!ops->empty()) { - // we do not own the OpDesc, so release the ownership. - ops->ReleaseLast(); - } -} - -void BlockDesc::ClearPBVars() { - auto vars = this->desc_->mutable_vars(); - while (!vars->empty()) { - // we do not own the VarDesc, so release the ownership. - vars->ReleaseLast(); - } -} - void BlockDesc::SetForwardBlockID(int32_t forward_block_id) { PADDLE_ENFORCE(!desc_->has_forward_block_idx(), "Parent block ID has been set to %d. Cannot set to %d", diff --git a/paddle/fluid/framework/block_desc.h b/paddle/fluid/framework/block_desc.h index eef19c4f09c60b9df18f154c85c421f5bff9413f..524a5ba176c92a4498922e35e1c5bb072cdbea89 100644 --- a/paddle/fluid/framework/block_desc.h +++ b/paddle/fluid/framework/block_desc.h @@ -41,11 +41,6 @@ class BlockDesc { BlockDesc(const BlockDesc &other, proto::BlockDesc *desc, ProgramDesc *prog); - ~BlockDesc() { - this->ClearPBVars(); - this->ClearPBOps(); - } - int32_t ID() const { return desc_->idx(); } int32_t Parent() const { return desc_->parent_idx(); } @@ -88,6 +83,8 @@ class BlockDesc { OpDesc *PrependOp(); + void PrependAllocatedOp(std::unique_ptr &&op_desc); + OpDesc *InsertOp(size_t index); /* @@ -111,10 +108,6 @@ class BlockDesc { ProgramDesc *Program() const { return this->prog_; } - private: - void ClearPBOps(); - void ClearPBVars(); - private: ProgramDesc *prog_; // not_own proto::BlockDesc *desc_; // not_own diff --git a/paddle/fluid/framework/data_device_transform.cc b/paddle/fluid/framework/data_device_transform.cc index 85dbb39e6fba735471446b5e5e71a612282c498a..6bcfc6cd55f02f0d4f0f6e3170e7cc19ce666a28 100644 --- a/paddle/fluid/framework/data_device_transform.cc +++ b/paddle/fluid/framework/data_device_transform.cc @@ -16,29 +16,25 @@ limitations under the License. */ namespace paddle { namespace framework { -static const platform::DeviceContext* GetDeviceContext( - const platform::Place& src_place, const platform::Place& dst_place) { - platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); - - if (platform::is_gpu_place(src_place) && platform::is_cpu_place(dst_place)) { - return pool.Get(src_place); - } else if (platform::is_cpu_place(src_place) && - platform::is_gpu_place(dst_place)) { - return pool.Get(dst_place); - } else { - PADDLE_THROW( - "Currently, model parallelism is only supported between CPU and CUDA"); - } -} - -void TransDataDevice(const Tensor& in, const platform::Place& dst_place, - Tensor* out) { +void TransDataDevice(const Tensor &in, const platform::Place &dst_place, + Tensor *out) { VLOG(3) << "DeviceTransform in, src_place " << in.place() << " dst_place: " << dst_place; - auto* dev_ctx = GetDeviceContext(in.place(), dst_place); - dev_ctx->Wait(); - TensorCopy(in, dst_place, *dev_ctx, out); - dev_ctx->Wait(); + + PADDLE_ENFORCE_NE( + in.place().which(), dst_place.which(), + "Currently, model parallelism is only supported between CPU and CUDA"); + + // FIXME(zcd): TransDataDevice is used to transform data from GPU to CPU and + // the enforced checkings have been done in GetDeviceContext, so the + // `dev_ctx->Wait()` is necessary. But `dev_ctx->Wait()` will make the program + // slow, especially when the number of elements is little, for example, + // the elements of learning rate are one and it's CPU side. + // One solution is to use a CUDA kernel to complete the copy operation when + // the transforming is from CPU to GPU and the number of elements is little. + // But the embarrassment is that this solution this solution makes training + // slower. + TensorCopySync(in, dst_place, out); } } // namespace framework diff --git a/paddle/fluid/framework/data_device_transform_test.cu b/paddle/fluid/framework/data_device_transform_test.cu index df4caa45eba2470f7528d2fbd99cca39cae0b596..a91fe5c99d397ef1bf04f6d22e988b6d3f33e500 100644 --- a/paddle/fluid/framework/data_device_transform_test.cu +++ b/paddle/fluid/framework/data_device_transform_test.cu @@ -32,8 +32,7 @@ struct AddFunctor { class OpKernelTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker { public: - OpKernelTestProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() { AddInput("input", "input1 of test op"); AddOutput("output", "output of test op"); AddAttr("use_gpu", "force to use gpu kernel").SetDefault(false); diff --git a/paddle/fluid/framework/data_type.cc b/paddle/fluid/framework/data_type.cc new file mode 100644 index 0000000000000000000000000000000000000000..b6b93cf422a60c1d8e9cb8b477efd562f9fe4758 --- /dev/null +++ b/paddle/fluid/framework/data_type.cc @@ -0,0 +1,102 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/framework/data_type.h" +#include +#include +#include + +namespace paddle { +namespace framework { + +struct DataTypeMap { + std::unordered_map cpp_to_proto_; + std::unordered_map proto_to_cpp_; + std::unordered_map proto_to_str_; + std::unordered_map cpp_to_size_; +}; + +static DataTypeMap* InitDataTypeMap(); +static DataTypeMap& gDataTypeMap() { + static DataTypeMap* g_data_type_map_ = InitDataTypeMap(); + return *g_data_type_map_; +} + +template +static inline void RegisterType(DataTypeMap* map, + proto::VarType::Type proto_type, + const std::string& name) { + map->proto_to_cpp_.emplace(static_cast(proto_type), typeid(T)); + map->cpp_to_proto_.emplace(typeid(T), proto_type); + map->proto_to_str_.emplace(static_cast(proto_type), name); + map->cpp_to_size_.emplace(typeid(T), sizeof(T)); +} + +static DataTypeMap* InitDataTypeMap() { + auto retv = new DataTypeMap(); + +#define RegType(cc_type, proto_type) \ + RegisterType(retv, proto_type, #cc_type) + + // NOTE: Add your customize type here. + RegType(platform::float16, proto::VarType::FP16); + RegType(float, proto::VarType::FP32); + RegType(double, proto::VarType::FP64); + RegType(int, proto::VarType::INT32); + RegType(int64_t, proto::VarType::INT64); + RegType(bool, proto::VarType::BOOL); + RegType(size_t, proto::VarType::SIZE_T); + RegType(int16_t, proto::VarType::INT16); + RegType(uint8_t, proto::VarType::UINT8); + +#undef RegType + return retv; +} + +proto::VarType::Type ToDataType(std::type_index type) { + auto it = gDataTypeMap().cpp_to_proto_.find(type); + if (it != gDataTypeMap().cpp_to_proto_.end()) { + return it->second; + } + PADDLE_THROW("Not support %s as tensor type", type.name()); +} + +std::type_index ToTypeIndex(proto::VarType::Type type) { + auto it = gDataTypeMap().proto_to_cpp_.find(static_cast(type)); + if (it != gDataTypeMap().proto_to_cpp_.end()) { + return it->second; + } + PADDLE_THROW("Not support proto::VarType::Type(%d) as tensor type", + static_cast(type)); +} + +std::string DataTypeToString(const proto::VarType::Type type) { + auto it = gDataTypeMap().proto_to_str_.find(static_cast(type)); + if (it != gDataTypeMap().proto_to_str_.end()) { + return it->second; + } + PADDLE_THROW("Not support proto::VarType::Type(%d) as tensor type", + static_cast(type)); +} + +size_t SizeOfType(std::type_index type) { + auto it = gDataTypeMap().cpp_to_size_.find(type); + if (it != gDataTypeMap().cpp_to_size_.end()) { + return it->second; + } + PADDLE_THROW("Not support %s as tensor type", type.name()); +} + +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/data_type.h b/paddle/fluid/framework/data_type.h index 2a528eb3aa562568c92059250f2c9bc5a75ec103..491413db8c8d66fd907801131e89d9303bdef9f2 100644 --- a/paddle/fluid/framework/data_type.h +++ b/paddle/fluid/framework/data_type.h @@ -17,51 +17,14 @@ limitations under the License. */ #include #include "paddle/fluid/framework/framework.pb.h" #include "paddle/fluid/platform/enforce.h" + #include "paddle/fluid/platform/float16.h" namespace paddle { namespace framework { -inline proto::VarType::Type ToDataType(std::type_index type) { - if (typeid(platform::float16).hash_code() == type.hash_code()) { - return proto::VarType::FP16; - } else if (typeid(const float).hash_code() == type.hash_code()) { - // CPPLint complains Using C-style cast. Use static_cast() instead - // One fix to this is to replace float with const float because - // typeid(T) == typeid(const T) - // http://en.cppreference.com/w/cpp/language/typeid - return proto::VarType::FP32; - } else if (typeid(const double).hash_code() == type.hash_code()) { - return proto::VarType::FP64; - } else if (typeid(const int).hash_code() == type.hash_code()) { - return proto::VarType::INT32; - } else if (typeid(const int64_t).hash_code() == type.hash_code()) { - return proto::VarType::INT64; - } else if (typeid(const bool).hash_code() == type.hash_code()) { - return proto::VarType::BOOL; - } else { - PADDLE_THROW("Not supported"); - } -} - -inline std::type_index ToTypeIndex(proto::VarType::Type type) { - switch (type) { - case proto::VarType::FP16: - return typeid(platform::float16); - case proto::VarType::FP32: - return typeid(float); - case proto::VarType::FP64: - return typeid(double); - case proto::VarType::INT32: - return typeid(int); - case proto::VarType::INT64: - return typeid(int64_t); - case proto::VarType::BOOL: - return typeid(bool); - default: - PADDLE_THROW("Not support type %d", type); - } -} +extern proto::VarType::Type ToDataType(std::type_index type); +extern std::type_index ToTypeIndex(proto::VarType::Type type); template inline void VisitDataType(proto::VarType::Type type, Visitor visitor) { @@ -84,37 +47,23 @@ inline void VisitDataType(proto::VarType::Type type, Visitor visitor) { case proto::VarType::BOOL: visitor.template operator()(); break; - default: - PADDLE_THROW("Not supported"); - } -} - -inline std::string DataTypeToString(const proto::VarType::Type type) { - switch (type) { - case proto::VarType::FP16: - return "float16"; - case proto::VarType::FP32: - return "float32"; - case proto::VarType::FP64: - return "float64"; + case proto::VarType::UINT8: + visitor.template operator()(); + break; case proto::VarType::INT16: - return "int16"; - case proto::VarType::INT32: - return "int32"; - case proto::VarType::INT64: - return "int64"; - case proto::VarType::BOOL: - return "bool"; + visitor.template operator()(); + break; default: - PADDLE_THROW("Not support type %d", type); + PADDLE_THROW("Not supported %d", type); } } +extern std::string DataTypeToString(const proto::VarType::Type type); +extern size_t SizeOfType(std::type_index type); inline std::ostream& operator<<(std::ostream& out, const proto::VarType::Type& type) { out << DataTypeToString(type); return out; } - } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/data_type_transform.cc b/paddle/fluid/framework/data_type_transform.cc index c0523f3c795b103c0c27081ec5dc717f6a0f11e0..5a57ec20585c26dbcd4251464718fc819148a7a5 100644 --- a/paddle/fluid/framework/data_type_transform.cc +++ b/paddle/fluid/framework/data_type_transform.cc @@ -91,6 +91,12 @@ void TransDataType(const OpKernelType& kernel_type_for_var, case proto::VarType::BOOL: framework::VisitDataType(dst_type, CastDataType(in, out, ctx)); break; + case proto::VarType::INT16: + framework::VisitDataType(dst_type, CastDataType(in, out, ctx)); + break; + case proto::VarType::UINT8: + framework::VisitDataType(dst_type, CastDataType(in, out, ctx)); + break; default: PADDLE_THROW("Not support type %d", src_type); } diff --git a/paddle/fluid/framework/details/CMakeLists.txt b/paddle/fluid/framework/details/CMakeLists.txt index 96c181f983a33961e3d5fb8745740f2fdbb210de..1bcd8412eb2d618b923bcd0557d118af62271f4a 100644 --- a/paddle/fluid/framework/details/CMakeLists.txt +++ b/paddle/fluid/framework/details/CMakeLists.txt @@ -3,7 +3,7 @@ cc_library(op_handle_base SRCS op_handle_base.cc DEPS var_handle device_context cc_library(scale_loss_grad_op_handle SRCS scale_loss_grad_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory) cc_library(fetch_op_handle SRCS fetch_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory) cc_library(computation_op_handle SRCS computation_op_handle.cc DEPS framework_proto scope place operator op_registry) -cc_library(send_op_handle SRCS send_op_handle.cc DEPS framework_proto scope place operator op_registry) +cc_library(rpc_op_handle SRCS rpc_op_handle.cc DEPS framework_proto scope place operator op_registry) cc_library(ssa_graph SRCS ssa_graph.cc DEPS var_handle op_handle_base) cc_library(ssa_graph_builder SRCS ssa_graph_builder.cc DEPS ssa_graph) @@ -15,16 +15,18 @@ if(WITH_GPU) dynload_cuda) set(multi_devices_graph_builder_deps nccl_all_reduce_op_handle) nv_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope ddim dynload_cuda) + nv_library(broadcast_op_handle SRCS broadcast_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor dynload_cuda) + else() set(multi_devices_graph_builder_deps) cc_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope ddim) + cc_library(broadcast_op_handle SRCS broadcast_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor) endif() -cc_library(broadcast_op_handle SRCS broadcast_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor) cc_library(gather_op_handle SRCS gather_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor) cc_library(multi_devices_graph_builder SRCS multi_devices_graph_builder.cc DEPS ssa_graph_builder computation_op_handle - scale_loss_grad_op_handle send_op_handle ${multi_devices_graph_builder_deps} reduce_op_handle broadcast_op_handle) + scale_loss_grad_op_handle rpc_op_handle ${multi_devices_graph_builder_deps} reduce_op_handle broadcast_op_handle) cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS ssa_graph framework_proto) cc_library(threaded_ssa_graph_executor SRCS threaded_ssa_graph_executor.cc DEPS fetch_op_handle ssa_graph_executor scope @@ -34,5 +36,5 @@ cc_test(broadcast_op_test SRCS broadcast_op_handle_test.cc DEPS var_handle op_ha device_context broadcast_op_handle) cc_test(gather_op_test SRCS gather_op_handle_test.cc DEPS var_handle op_handle_base scope ddim memory device_context gather_op_handle) -cc_test(reduce_op_handle_test SRCS reduce_op_handle_test.cc DEPS var_handle op_handle_base scope ddim memory - device_context reduce_op_handle ) +#cc_test(reduce_op_handle_test SRCS reduce_op_handle_test.cc DEPS var_handle op_handle_base scope ddim memory +# device_context reduce_op_handle ) diff --git a/paddle/fluid/framework/details/broadcast_op_handle.cc b/paddle/fluid/framework/details/broadcast_op_handle.cc index 33e02ab65a251a338225ee621ff14acbb0631992..d5ca061944f33939cea59a5275e691b1966194fa 100644 --- a/paddle/fluid/framework/details/broadcast_op_handle.cc +++ b/paddle/fluid/framework/details/broadcast_op_handle.cc @@ -19,14 +19,12 @@ namespace paddle { namespace framework { namespace details { -BroadcastOpHandle::BroadcastOpHandle(const std::vector &local_scopes, - const std::vector &places) - : local_scopes_(local_scopes), places_(places) {} void BroadcastOpHandle::RunImpl() { - // the input and output may have dummy var. - VarHandle *in_var_handle; + if (places_.size() == 1) return; + // The input and output may have dummy vars. + VarHandle *in_var_handle; { auto in_var_handles = DynamicCast(inputs_); PADDLE_ENFORCE_EQ(in_var_handles.size(), 1, @@ -40,9 +38,7 @@ void BroadcastOpHandle::RunImpl() { out_var_handles.size(), places_.size(), "The number of output should equal to the number of places."); - // Wait input done, this Wait is asynchronous operation platform::Place - // &in_place; - WaitInputVarGenerated(*in_var_handle); + WaitInputVarGenerated(); std::vector var_scopes; for (auto *s : local_scopes_) { @@ -52,38 +48,114 @@ void BroadcastOpHandle::RunImpl() { auto *in_var = var_scopes.at(in_var_handle->scope_idx_)->FindVar(in_var_handle->name_); PADDLE_ENFORCE_NOT_NULL(in_var); - Tensor &in_tensor = VariableVisitor::GetMutableTensor(in_var); - for (auto *out : out_var_handles) { - if (*out == *in_var_handle) { - continue; + InitOutputValue(*in_var_handle, out_var_handles); + + if (platform::is_cpu_place(in_tensor.place())) { + for (auto *out_var_handle : out_var_handles) { + if (out_var_handle->IsTheSameVar(*in_var_handle)) { + continue; + } + auto &out_p = out_var_handle->place_; + auto *out_var = var_scopes.at(out_var_handle->scope_idx_) + ->FindVar(out_var_handle->name_); + + RunAndRecordEvent(out_p, [in_tensor, out_var] { + paddle::framework::TensorCopy( + in_tensor, platform::CPUPlace(), + &VariableVisitor::GetMutableTensor(out_var)); + }); + } + } else { +#ifdef PADDLE_WITH_CUDA + VarHandle *out_handle = nullptr; + int root_id = boost::get(in_tensor.place()).device; + std::vector> broadcast_calls; + + for (auto out_var_handle : out_var_handles) { + Variable *out_var = var_scopes.at(out_var_handle->scope_idx_) + ->FindVar(out_var_handle->name_); + + int dst_id = + boost::get(out_var_handle->place_).device; + + auto &nccl_ctx = nccl_ctxs_->at(dst_id); + + void *send_recv_buffer = nullptr; + if (root_id == dst_id) { + send_recv_buffer = const_cast(in_tensor.data()); + out_handle = out_var_handle; + } else { + send_recv_buffer = + VariableVisitor::GetMutableTensor(out_var).mutable_data( + out_var_handle->place_); + } + + int type = platform::ToNCCLDataType(in_tensor.type()); + size_t numel = static_cast(in_tensor.numel()); + broadcast_calls.emplace_back( + [send_recv_buffer, numel, type, root_id, &nccl_ctx] { + PADDLE_ENFORCE(platform::dynload::ncclBcast( + send_recv_buffer, numel, static_cast(type), + root_id, nccl_ctx.comm_, nccl_ctx.stream())); + }); } - auto &out_p = out->place_; - auto *out_var = var_scopes.at(out->scope_idx_)->FindVar(out->name_); - PADDLE_ENFORCE_NOT_NULL(out_var); - PADDLE_ENFORCE_EQ(out_p.which(), in_var_handle->place_.which(), - "Places must be all on CPU or all on CUDA."); - - VariableVisitor::ShareDimsAndLoD(*in_var, out_var); - VariableVisitor::GetMutableTensor(out_var).mutable_data(out_p, - in_tensor.type()); - - auto dev_ctx = dev_ctxes_.at(out_p); - RunAndRecordEvent(out_p, [in_tensor, out_var, dev_ctx, out_p] { - paddle::framework::TensorCopy( - in_tensor, out_p, *(dev_ctx), - &VariableVisitor::GetMutableTensor(out_var)); + this->RunAndRecordEvent([&] { + { + platform::NCCLGroupGuard guard; + for (auto &call : broadcast_calls) { + call(); + } + } + + if (!out_handle->IsTheSameVar(*in_var_handle)) { + auto out_var = var_scopes.at(in_var_handle->scope_idx_) + ->FindVar(out_var_handles[0]->name_); + paddle::framework::TensorCopy( + in_tensor, in_var_handle->place_, + *(dev_ctxes_.at(in_var_handle->place_)), + &VariableVisitor::GetMutableTensor(out_var)); + } }); +#else + PADDLE_THROW("CUDA is not enabled."); +#endif } } -void BroadcastOpHandle::WaitInputVarGenerated(const VarHandle &in_var) { - if (in_var.generated_op_) { - for (auto &pair : dev_ctxes_) { - in_var.generated_op_->Wait(pair.second); +void BroadcastOpHandle::InitOutputValue( + const VarHandle &in_var_handle, + const std::vector &out_var_handles) const { + std::vector var_scopes; + for (auto *s : local_scopes_) { + var_scopes.emplace_back(s->FindVar(kLocalExecScopeName)->Get()); + } + auto *in_var = + var_scopes.at(in_var_handle.scope_idx_)->FindVar(in_var_handle.name_); + + Tensor &in_tensor = VariableVisitor::GetMutableTensor(in_var); + + // NOTE: The tensors' Place of input and output must be all on GPU or all on + // CPU. + for (auto *out_var_handle : out_var_handles) { + if (out_var_handle->IsTheSameVar(in_var_handle)) { + continue; + } + auto t_out_p = out_var_handle->place_; + auto *out_var = var_scopes.at(out_var_handle->scope_idx_) + ->FindVar(out_var_handle->name_); + PADDLE_ENFORCE_NOT_NULL(out_var); + if (is_gpu_place(in_tensor.place())) { + PADDLE_ENFORCE(platform::is_gpu_place(t_out_p), + "Places of input and output must be all on GPU."); + } else { + t_out_p = platform::CPUPlace(); } + VariableVisitor::ShareDimsAndLoD(*in_var, out_var); + VariableVisitor::GetMutableTensor(out_var).mutable_data(t_out_p, + in_tensor.type()); } } diff --git a/paddle/fluid/framework/details/broadcast_op_handle.h b/paddle/fluid/framework/details/broadcast_op_handle.h index 92420f10ac5972b7924d83b43bb28234079e5072..629aa00cb817c4b1446e7b750ca62a7c6b1db670 100644 --- a/paddle/fluid/framework/details/broadcast_op_handle.h +++ b/paddle/fluid/framework/details/broadcast_op_handle.h @@ -24,14 +24,32 @@ #include "paddle/fluid/framework/selected_rows.h" #include "paddle/fluid/platform/device_context.h" +#ifdef PADDLE_WITH_CUDA +#include "paddle/fluid/platform/nccl_helper.h" +#endif + namespace paddle { namespace framework { namespace details { struct BroadcastOpHandle : public OpHandleBase { public: +#ifdef PADDLE_WITH_CUDA + BroadcastOpHandle(const std::vector &local_scopes, + const std::vector &places, + const platform::NCCLContextMap *nccl_ctxs) + : local_scopes_(local_scopes), places_(places), nccl_ctxs_(nccl_ctxs) { + if (nccl_ctxs_) { + for (auto &p_ctx : nccl_ctxs_->contexts_) { + dev_ctxes_[platform::CUDAPlace(p_ctx.first)] = p_ctx.second.ctx_.get(); + } + } + } +#else BroadcastOpHandle(const std::vector &local_scopes, - const std::vector &places); + const std::vector &places) + : local_scopes_(local_scopes), places_(places) {} +#endif std::string Name() const override; @@ -39,11 +57,16 @@ struct BroadcastOpHandle : public OpHandleBase { protected: void RunImpl() override; - void WaitInputVarGenerated(const VarHandle &in_var); private: const std::vector &local_scopes_; const std::vector &places_; +#ifdef PADDLE_WITH_CUDA + const platform::NCCLContextMap *nccl_ctxs_; +#endif + + void InitOutputValue(const VarHandle &in_var_handle, + const std::vector &out_var_handles) const; }; } // namespace details } // namespace framework diff --git a/paddle/fluid/framework/details/broadcast_op_handle_test.cc b/paddle/fluid/framework/details/broadcast_op_handle_test.cc index 3f2dcde3e9597287d72046dd4f8b07faab1ede25..c6e923ef77ff03413eefe4f26457a5322747618e 100644 --- a/paddle/fluid/framework/details/broadcast_op_handle_test.cc +++ b/paddle/fluid/framework/details/broadcast_op_handle_test.cc @@ -35,15 +35,25 @@ struct TestBroadcastOpHandle { std::unique_ptr op_handle_; std::vector> vars_; std::vector gpu_list_; + bool use_gpu_; +#ifdef PADDLE_WITH_CUDA + std::unique_ptr nccl_ctxs_; +#endif void WaitAll() { for (size_t j = 0; j < ctxs_.size(); ++j) { ctxs_[j]->Wait(); } +#ifdef PADDLE_WITH_CUDA + if (nccl_ctxs_) { + nccl_ctxs_->WaitAll(); + } +#endif } void InitCtxOnGpu(bool use_gpu) { - if (use_gpu) { + use_gpu_ = use_gpu; + if (use_gpu_) { #ifdef PADDLE_WITH_CUDA int count = p::GetCUDADeviceCount(); if (count <= 1) { @@ -57,6 +67,7 @@ struct TestBroadcastOpHandle { gpu_list_.push_back(p); ctxs_.emplace_back(new p::CUDADeviceContext(p)); } + nccl_ctxs_.reset(new platform::NCCLContextMap(gpu_list_)); #else PADDLE_THROW("CUDA is not support."); #endif @@ -67,6 +78,9 @@ struct TestBroadcastOpHandle { gpu_list_.push_back(p); ctxs_.emplace_back(new p::CPUDeviceContext(p)); } +#ifdef PADDLE_WITH_CUDA + nccl_ctxs_.reset(nullptr); +#endif } } @@ -82,7 +96,21 @@ struct TestBroadcastOpHandle { } param_scopes_[input_scope_idx]->Var("input"); - op_handle_.reset(new BroadcastOpHandle(local_scopes_, gpu_list_)); + if (use_gpu_) { +#ifdef PADDLE_WITH_CUDA + op_handle_.reset( + new BroadcastOpHandle(local_scopes_, gpu_list_, nccl_ctxs_.get())); +#else + PADDLE_THROW("CUDA is not support."); +#endif + } else { +#ifdef PADDLE_WITH_CUDA + op_handle_.reset( + new BroadcastOpHandle(local_scopes_, gpu_list_, nccl_ctxs_.get())); +#else + op_handle_.reset(new BroadcastOpHandle(local_scopes_, gpu_list_)); +#endif + } auto* in_var_handle = new VarHandle(1, input_scope_idx, "input", gpu_list_[input_scope_idx]); @@ -97,7 +125,9 @@ struct TestBroadcastOpHandle { op_handle_->AddInput(dummy_var_handle); for (size_t j = 0; j < gpu_list_.size(); ++j) { - op_handle_->SetDeviceContext(gpu_list_[j], ctxs_[j].get()); + if (!use_gpu_) { + op_handle_->SetDeviceContext(gpu_list_[j], ctxs_[j].get()); + } VarHandle* out_var_handle = new VarHandle(2, j, "out", gpu_list_[j]); vars_.emplace_back(out_var_handle); op_handle_->AddOutput(out_var_handle); @@ -139,7 +169,7 @@ struct TestBroadcastOpHandle { PADDLE_ENFORCE_EQ(out_tensor.lod(), lod, "lod is not equal."); f::Tensor result_tensor; - f::TensorCopy(out_tensor, cpu_place, *(ctxs_[j]), &result_tensor); + f::TensorCopySync(out_tensor, cpu_place, &result_tensor); float* ct = result_tensor.mutable_data(cpu_place); for (int64_t i = 0; i < f::product(kDims); ++i) { @@ -185,7 +215,7 @@ struct TestBroadcastOpHandle { } f::Tensor result_tensor; - f::TensorCopy(rt, cpu_place, *(ctxs_[j]), &result_tensor); + f::TensorCopySync(rt, cpu_place, &result_tensor); float* ct = result_tensor.data(); for (int64_t i = 0; i < f::product(kDims); ++i) { diff --git a/paddle/fluid/framework/details/build_strategy.h b/paddle/fluid/framework/details/build_strategy.h new file mode 100644 index 0000000000000000000000000000000000000000..91bdfe6134ffbd1404336c9d6d1222a505084b2b --- /dev/null +++ b/paddle/fluid/framework/details/build_strategy.h @@ -0,0 +1,36 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +namespace paddle { +namespace framework { +namespace details { + +struct BuildStrategy { + enum class ReduceStrategy { kAllReduce = 0, kReduce = 1 }; + + enum class GradientScaleStrategy { + kCoeffNumDevice = 0, + kOne = 1, + kCustomized = 2, + }; + + ReduceStrategy reduce_{ReduceStrategy::kAllReduce}; + GradientScaleStrategy gradient_scale_{GradientScaleStrategy::kCoeffNumDevice}; +}; + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/computation_op_handle.cc b/paddle/fluid/framework/details/computation_op_handle.cc index 7ff0efe09387b7e5d7cfe0dfe5e129ca9914d90b..df05bb06333d6b964f2f5434c3d43214e5d2cb7a 100644 --- a/paddle/fluid/framework/details/computation_op_handle.cc +++ b/paddle/fluid/framework/details/computation_op_handle.cc @@ -26,20 +26,20 @@ ComputationOpHandle::ComputationOpHandle(const OpDesc &op_desc, Scope *scope, place_(place) {} void ComputationOpHandle::RunImpl() { - auto *cur_ctx = dev_ctxes_[place_]; - for (auto *in : inputs_) { - bool need_wait = in->generated_op_ && - in->generated_op_->DeviceContext(place_) != cur_ctx; - if (need_wait) { - in->generated_op_->Wait(cur_ctx); - } - } + WaitInputVarGenerated(place_); this->RunAndRecordEvent([this] { op_->Run(*scope_->FindVar(kLocalExecScopeName)->Get(), place_); }); } +bool ComputationOpHandle::NeedWait(VarHandleBase *in_var) { + bool need_wait = + in_var && in_var->generated_op_ && + in_var->generated_op_->DeviceContext(place_) != dev_ctxes_[place_]; + return need_wait; +} + std::string ComputationOpHandle::Name() const { return op_->Type(); } } // namespace details } // namespace framework diff --git a/paddle/fluid/framework/details/computation_op_handle.h b/paddle/fluid/framework/details/computation_op_handle.h index c363b973d9abbae6bea76c2458fbe82a37a342ca..f048f973fdeb6cf7d1485cda8cea7d530d9ba465 100644 --- a/paddle/fluid/framework/details/computation_op_handle.h +++ b/paddle/fluid/framework/details/computation_op_handle.h @@ -36,6 +36,8 @@ struct ComputationOpHandle : public OpHandleBase { protected: void RunImpl() override; + bool NeedWait(VarHandleBase *in_var) override; + private: std::unique_ptr op_; Scope *scope_; diff --git a/paddle/fluid/framework/details/execution_strategy.h b/paddle/fluid/framework/details/execution_strategy.h new file mode 100644 index 0000000000000000000000000000000000000000..e8d510ec955602b5a3f73ca06caa121886eb150b --- /dev/null +++ b/paddle/fluid/framework/details/execution_strategy.h @@ -0,0 +1,29 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +namespace paddle { +namespace framework { +namespace details { + +struct ExecutionStrategy { + size_t num_threads_{0}; + bool use_event_{true}; + bool allow_op_delay_{false}; +}; + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/fetch_op_handle.cc b/paddle/fluid/framework/details/fetch_op_handle.cc index 423449abff97dbf70d81314f852d9135e25f243f..224e8e1f6efd7a894591ac51c929517cae7539ce 100644 --- a/paddle/fluid/framework/details/fetch_op_handle.cc +++ b/paddle/fluid/framework/details/fetch_op_handle.cc @@ -31,7 +31,7 @@ FetchOpHandle::~FetchOpHandle() { } } -void FetchOpHandle::Wait(platform::DeviceContext *waited_dev) { +void FetchOpHandle::RecordWaitEventOnCtx(platform::DeviceContext *waited_ctx) { PADDLE_THROW("Nobody should wait FetchOp. Unexpceted Error"); } @@ -45,29 +45,25 @@ void FetchOpHandle::WaitAndMergeCPUTensors() const { } void FetchOpHandle::RunImpl() { - auto cpu_ctx = - platform::DeviceContextPool::Instance().Get(platform::CPUPlace()); - for (auto *input : inputs_) { - auto *var = static_cast(input); - var->generated_op_->Wait(cpu_ctx); - } + WaitInputVarGenerated(platform::CPUPlace()); + tensors_.resize(inputs_.size()); - auto *var_handle = static_cast(inputs_[0]); - auto &var_name = var_handle->name_; platform::CPUPlace cpu; auto &scopes = *local_scopes_; - for (size_t i = 0; i < scopes.size(); ++i) { - auto &scope = scopes[i]; - auto *var = - scope->FindVar(kLocalExecScopeName)->Get()->FindVar(var_name); + for (size_t i = 0; i < inputs_.size(); ++i) { + auto *var_handle = static_cast(inputs_[i]); + auto &scope = scopes.at(var_handle->scope_idx_); + auto *var = scope->FindVar(kLocalExecScopeName) + ->Get() + ->FindVar(var_handle->name_); PADDLE_ENFORCE_NOT_NULL(var, "Cannot find variable %s in execution scope", - var_name); + var_handle->name_); + auto &t = var->Get(); if (platform::is_gpu_place(t.place())) { #ifdef PADDLE_WITH_CUDA - TensorCopy(t, cpu, *dev_ctxes_[t.place()], &tensors_[i], true); - dev_ctxes_.at(t.place())->Wait(); + TensorCopySync(t, cpu, &tensors_[i]); #endif } else { tensors_[i].ShareDataWith(t); @@ -78,6 +74,15 @@ void FetchOpHandle::RunImpl() { this->WaitAndMergeCPUTensors(); } +void FetchOpHandle::WaitInputVarGenerated(const platform::Place &place) { + auto cpu_ctx = platform::DeviceContextPool::Instance().Get(place); + for (auto *input : inputs_) { + if (input->generated_op_) { + input->generated_op_->RecordWaitEventOnCtx(cpu_ctx); + } + } +} + std::string FetchOpHandle::Name() const { return "Fetch"; } } // namespace details diff --git a/paddle/fluid/framework/details/fetch_op_handle.h b/paddle/fluid/framework/details/fetch_op_handle.h index b49f3df338dc11310a4a0c27c8aaae3602373fcc..e09bdd1d3338bb175c1ddae35b53f98197b68e9a 100644 --- a/paddle/fluid/framework/details/fetch_op_handle.h +++ b/paddle/fluid/framework/details/fetch_op_handle.h @@ -33,7 +33,7 @@ struct FetchOpHandle : public OpHandleBase { ~FetchOpHandle(); - void Wait(platform::DeviceContext *waited_dev) override; + void RecordWaitEventOnCtx(platform::DeviceContext *waited_ctx) override; void WaitAndMergeCPUTensors() const; @@ -42,6 +42,8 @@ struct FetchOpHandle : public OpHandleBase { protected: void RunImpl() override; + void WaitInputVarGenerated(const platform::Place &place) override; + private: FeedFetchList *data_; size_t offset_; diff --git a/paddle/fluid/framework/details/gather_op_handle.cc b/paddle/fluid/framework/details/gather_op_handle.cc index 3ed7723919fc3a547b15c28b846de758a8155e66..2be02304566cf5dbe348fa01fc4171990eafd158 100644 --- a/paddle/fluid/framework/details/gather_op_handle.cc +++ b/paddle/fluid/framework/details/gather_op_handle.cc @@ -25,6 +25,7 @@ GatherOpHandle::GatherOpHandle(const std::vector &local_scopes, : local_scopes_(local_scopes), places_(places) {} void GatherOpHandle::RunImpl() { + if (places_.size() == 1) return; // the input and output may have dummy var. auto in_var_handles = DynamicCast(inputs_); @@ -35,7 +36,6 @@ void GatherOpHandle::RunImpl() { VarHandle *out_var_handle; { auto out_var_handles = DynamicCast(outputs_); - PADDLE_ENFORCE_EQ(out_var_handles.size(), 1, "The number of output should be one."); out_var_handle = out_var_handles.front(); @@ -50,84 +50,67 @@ void GatherOpHandle::RunImpl() { auto pre_in_var = var_scopes.at(in_0_handle->scope_idx_)->FindVar(in_0_handle->name_); PADDLE_ENFORCE_NOT_NULL(pre_in_var); + PADDLE_ENFORCE(pre_in_var->IsType(), "Currently, gather_op only can gather SelectedRows."); - auto pre_place = in_0_handle->place_; - PADDLE_ENFORCE_EQ(out_var_handle->place_.which(), pre_place.which(), - "The place of input and output should be the same."); - // Wait input done, this Wait is asynchronous operation - WaitInputVarGenerated(in_var_handles); + WaitInputVarGenerated(); + auto &pre_in_value = pre_in_var->Get(); std::vector out_rows; std::vector in_tensors; - std::vector in_places; - auto &pre_in = pre_in_var->Get(); - // gather the inputs + // Gather the inputs for (auto *in_handle : in_var_handles) { - auto in_p = in_handle->place_; - in_places.push_back(in_p); - PADDLE_ENFORCE_EQ(in_p.which(), pre_place.which(), - "Places must be all on CPU or all on CUDA."); auto *in_var = var_scopes.at(in_handle->scope_idx_)->FindVar(in_handle->name_); - auto &in_sr = in_var->Get(); + PADDLE_ENFORCE_NOT_NULL(in_var); + VariableVisitor::EnforceShapeAndDTypeEQ(*in_var, *pre_in_var); - PADDLE_ENFORCE_EQ(in_sr.value().type(), pre_in.value().type(), - "The type of input is not consistent."); - PADDLE_ENFORCE_EQ(pre_in.height(), in_sr.height(), - "The height of inputs is not consistent."); - PADDLE_ENFORCE_EQ(pre_in.GetCompleteDims(), in_sr.GetCompleteDims(), - "The dims of inputs is not consistent."); + auto &in_sr_value = in_var->Get(); - auto &in_sr_rows = in_sr.rows(); + auto &in_sr_rows = in_sr_value.rows(); out_rows.insert(out_rows.end(), in_sr_rows.begin(), in_sr_rows.end()); - - in_tensors.emplace_back(in_sr.value()); + in_tensors.emplace_back(in_sr_value.value()); } - // write the output - auto &out_place = out_var_handle->place_; - auto out_scope_idx = out_var_handle->scope_idx_; - auto out_var = var_scopes.at(out_scope_idx)->FindVar(out_var_handle->name_); + // NOTE: The Places of all input tensor must be all on CPU or all on GPU. + platform::Place t_out_p = out_var_handle->place_; + if (platform::is_gpu_place(pre_in_value.place())) { + PADDLE_ENFORCE(platform::is_gpu_place(t_out_p), + "Places of input and output must be all on GPU."); + } else { + t_out_p = platform::CPUPlace(); + } - auto out = out_var->GetMutable(); - out->set_height(pre_in.height()); - out->set_rows(out_rows); + auto out_var = + var_scopes.at(out_var_handle->scope_idx_)->FindVar(out_var_handle->name_); + PADDLE_ENFORCE_NOT_NULL(out_var); + auto out_value = out_var->GetMutable(); + out_value->set_height(pre_in_value.height()); + out_value->set_rows(out_rows); size_t rows = out_rows.size(); - DDim out_dim = pre_in.GetCompleteDims(); + DDim out_dim = pre_in_value.GetCompleteDims(); out_dim[0] = static_cast(rows); - out->mutable_value()->Resize(out_dim); - out->mutable_value()->mutable_data(out_place, pre_in.value().type()); - Tensor *out_tensor = out->mutable_value(); + out_value->mutable_value()->Resize(out_dim).mutable_data( + t_out_p, pre_in_value.value().type()); + Tensor *out_tensor = out_value->mutable_value(); // copy - auto dev_ctx = dev_ctxes_[out_place]; - RunAndRecordEvent(out_place, [in_tensors, out_tensor, dev_ctx, out_place] { + auto dev_ctx = dev_ctxes_[out_var_handle->place_]; + RunAndRecordEvent(out_var_handle->place_, [in_tensors, out_tensor, &dev_ctx, + t_out_p] { int s = 0, e = 0; for (size_t j = 0; j < in_tensors.size(); ++j) { e += in_tensors[j].dims()[0]; auto sub_out = out_tensor->Slice(s, e); - paddle::framework::TensorCopy(in_tensors[j], out_place, *(dev_ctx), - &sub_out); + paddle::framework::TensorCopy(in_tensors[j], t_out_p, *dev_ctx, &sub_out); s = e; } }); } -void GatherOpHandle::WaitInputVarGenerated( - const std::vector &in_var_handles) { - for (auto *in : in_var_handles) { - if (in->generated_op_) { - for (auto pair : dev_ctxes_) { - in->generated_op_->Wait(pair.second); - } - } - } -} - std::string GatherOpHandle::Name() const { return "gather"; } } // namespace details } // namespace framework diff --git a/paddle/fluid/framework/details/gather_op_handle.h b/paddle/fluid/framework/details/gather_op_handle.h index c394dd7a14b07cb956aa1aedfc0df4fa25744dd7..d11ef8556aa8840949ca8dc7aa176413f70b9f22 100644 --- a/paddle/fluid/framework/details/gather_op_handle.h +++ b/paddle/fluid/framework/details/gather_op_handle.h @@ -39,7 +39,6 @@ struct GatherOpHandle : public OpHandleBase { protected: void RunImpl() override; - void WaitInputVarGenerated(const std::vector &in_var_handles); private: const std::vector &local_scopes_; diff --git a/paddle/fluid/framework/details/multi_devices_graph_builder.cc b/paddle/fluid/framework/details/multi_devices_graph_builder.cc index 3413467b149539bcff42d78a9a6fe315d6558bb4..d8e711994c5dba15ce0a1c237558b121888902e3 100644 --- a/paddle/fluid/framework/details/multi_devices_graph_builder.cc +++ b/paddle/fluid/framework/details/multi_devices_graph_builder.cc @@ -11,11 +11,15 @@ // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. // See the License for the specific language governing permissions and // limitations under the License. - #include "paddle/fluid/framework/details/multi_devices_graph_builder.h" +#include +#include +#include "paddle/fluid/framework/details/broadcast_op_handle.h" #include "paddle/fluid/framework/details/computation_op_handle.h" +#include "paddle/fluid/framework/details/reduce_op_handle.h" +#include "paddle/fluid/framework/details/rpc_op_handle.h" #include "paddle/fluid/framework/details/scale_loss_grad_op_handle.h" -#include "paddle/fluid/framework/details/send_op_handle.h" +#include "paddle/fluid/framework/op_info.h" #include "paddle/fluid/framework/scope.h" #ifdef PADDLE_WITH_CUDA @@ -25,6 +29,10 @@ #include #include +DEFINE_string(ssa_graph_path, "/tmp/ssa_graph.dot", + "the ssa graph path only print with GLOG_v=10," + "default /tmp/graph.dot"); + namespace paddle { namespace framework { namespace details { @@ -34,79 +42,117 @@ MultiDevSSAGraphBuilder::MultiDevSSAGraphBuilder( const std::vector &places, const std::string &loss_var_name, const std::unordered_set ¶ms, - const std::vector &local_scopes, bool skip_scale_loss, - platform::NCCLContextMap *nccl_ctxs) + const std::vector &local_scopes, + platform::NCCLContextMap *nccl_ctxs, const BuildStrategy &strategy) : loss_var_name_(loss_var_name), places_(places), local_scopes_(local_scopes), - nccl_ctxs_(nccl_ctxs) { + nccl_ctxs_(nccl_ctxs), + strategy_(strategy) { #else MultiDevSSAGraphBuilder::MultiDevSSAGraphBuilder( const std::vector &places, const std::string &loss_var_name, const std::unordered_set ¶ms, - const std::vector &local_scopes, bool skip_scale_loss) + const std::vector &local_scopes, const BuildStrategy &strategy) : loss_var_name_(loss_var_name), places_(places), - local_scopes_(local_scopes) { + local_scopes_(local_scopes), + strategy_(strategy) { #endif for (auto &p : params) { grad_names_.insert(GradVarName(p)); } - skip_scale_loss_ = skip_scale_loss; } void MultiDevSSAGraphBuilder::CreateOpHandleIOs(SSAGraph *result, const OpDesc &op, - const platform::Place &p, - const size_t &i) const { + size_t place_id) const { + auto p = places_[place_id]; auto *op_handle = result->ops_.back().get(); op_handle->SetDeviceContext(p, platform::DeviceContextPool::Instance().Get(p)); - auto var_names = op.InputArgumentNames(); - - for (auto &each_var_name : var_names) { - VarHandle *var = CreateOrGetLatestVarHandle(result, each_var_name, p, i); + for (auto &each_var_name : op.InputArgumentNames()) { + VarHandle *var = + CreateOrGetLatestVarHandle(result, each_var_name, p, place_id); op_handle->AddInput(var); } - var_names = op.OutputArgumentNames(); + for (auto &each_var_name : op.OutputArgumentNames()) { + CreateOpOutput(result, op_handle, each_var_name, p, place_id); + } +} - for (auto &each_var_name : var_names) { - CreateOpOutput(result, op_handle, each_var_name, p, i); +std::vector MultiDevSSAGraphBuilder::FindDistTrainSendVars( + const ProgramDesc &program) const { + std::vector send_vars; + // since parameters are all in block 0, + // it's enough to only scan send ops in block 0 + for (auto *op : program.Block(0).AllOps()) { + // TODO(Yancey1989): use a graceful method to find send op, + // instead of the the hard code string + if (op->Type() == "send_vars") { + auto op_vars = op->InputArgumentNames(); + send_vars.reserve(send_vars.size() + + std::distance(op_vars.begin(), op_vars.end())); + send_vars.insert(send_vars.end(), op_vars.begin(), op_vars.end()); + } } + return send_vars; } -bool MultiDevSSAGraphBuilder::IsDistTrainOp(const OpDesc &op, - OpDesc *send_op) const { - if (send_op == nullptr) { +std::vector MultiDevSSAGraphBuilder::FindDistTrainRecvVars( + const ProgramDesc &program) const { + std::vector recv_vars; + for (auto *op : program.Block(0).AllOps()) { + // TODO(Yancey1989): use a graceful method to find recv op, + // instead of the hard code string + if (op->Type() == "recv") { + auto op_vars = op->OutputArgumentNames(); + recv_vars.reserve(recv_vars.size() + + std::distance(op_vars.begin(), op_vars.end())); + recv_vars.insert(recv_vars.end(), op_vars.begin(), op_vars.end()); + } + } + return recv_vars; +} + +bool MultiDevSSAGraphBuilder::IsDistTrainOp( + const OpDesc &op, const std::vector &send_vars, + const std::vector &recv_vars) const { + if (send_vars.size() == 0 || recv_vars.size() == 0) { return false; } - auto checker = [&](const std::vector opvars, - const std::vector sendvars) -> bool { - bool is_dist_train_op = false; + /** + * Check any of opvars contains `.block` and in sendvars + */ + auto checker = [](const std::vector &opvars, + const std::vector &rpc_vars) -> bool { for (auto &var : opvars) { + // a variable name with the suffix `.block` means it's a splited + // variable by (DistributeTranspiler) + // [python/paddle/fluid/transpiler/distribute_transpiler.py] if (var.find(".block") != std::string::npos && - std::find(sendvars.begin(), sendvars.end(), var) != sendvars.end()) { - is_dist_train_op = true; - break; + std::find(rpc_vars.begin(), rpc_vars.end(), var) != rpc_vars.end()) { + return true; } } - return is_dist_train_op; + return false; }; - if (op.Type() == "split") { - return checker(op.OutputArgumentNames(), send_op->InputArgumentNames()); - } else if (op.Type() == "concat") { - return checker(op.InputArgumentNames(), send_op->OutputArgumentNames()); - } - return false; + return checker(op.OutputArgumentNames(), send_vars) || + checker(op.InputArgumentNames(), recv_vars); } std::unique_ptr MultiDevSSAGraphBuilder::Build( const ProgramDesc &program) const { + std::unordered_map var_types; + for (auto *var : program.Block(0).AllVars()) { + var_types[var->Name()] = var->GetType(); + } + auto graph = new SSAGraph(); SSAGraph &result = *graph; std::unordered_set og_has_been_broadcast; @@ -116,45 +162,93 @@ std::unique_ptr MultiDevSSAGraphBuilder::Build( std::unordered_map>>>( places_.size()); - // Find "send" op first for split is in front of send. - OpDesc *send_op = nullptr; - for (auto *op : program.Block(0).AllOps()) { - if (op->Type() == "send") { - send_op = op; - break; - } - } + // find send/recv vars so that we can place the distributed training + // realted op in the place 0 + auto send_vars = FindDistTrainSendVars(program); + auto recv_vars = FindDistTrainRecvVars(program); + + size_t cur_device_id = 0; + std::vector> var_name_on_devices; + std::vector> bcast_var_name_set; + var_name_on_devices.resize(places_.size()); + bcast_var_name_set.resize(places_.size()); bool is_forwarding = true; for (auto *op : program.Block(0).AllOps()) { - if (op->Type() == "send") { - // append send op if program is distributed trainer main program. + if (boost::get( + op->GetAttr(OpProtoAndCheckerMaker::OpRoleAttrName())) == + static_cast(OpRole::kRPC)) { + // append rpc op if program is distributed trainer main program. // always use the first device - CreateSendOp(&result, *op); - } else if (IsDistTrainOp(*op, send_op)) { - CreateComputationalOps(&result, *op, 1); + CreateRPCOp(&result, *op); + } else if (IsDistTrainOp(*op, send_vars, recv_vars)) { + CreateDistTrainOp(&result, *op); } else if (IsScaleLossOp(*op)) { - if (!skip_scale_loss_) { + // user can customize loss@grad if not use_default_grad_scale_ + if (strategy_.gradient_scale_ != + BuildStrategy::GradientScaleStrategy::kCustomized) { CreateScaleLossGradOp(&result); } is_forwarding = false; } else { - CreateComputationalOps(&result, *op, places_.size()); - if (!is_forwarding) { + int op_dev_id = GetOpDeviceID(var_name_on_devices, *op); + if (op_dev_id == -1) { // var on all device + CreateComputationalOps(&result, *op, places_.size()); + } else { + CreateComputationalOp(&result, *op, op_dev_id); + for (auto &var_name : op->OutputArgumentNames()) { + var_name_on_devices[op_dev_id].emplace(var_name); + } + } + if (!is_forwarding && places_.size() > 1) { // Currently, we assume that once gradient is generated, it can be - // broadcast, and each gradient is only broadcast once. But there are no - // other cases, for example, we need to adjust the gradient according to - // the input when we get the gradient, which is not considered at - // present. - for (auto &og : op->OutputArgumentNames()) { - if (IsParameterGradientOnce(og, &og_has_been_broadcast)) { - InsertNCCLAllReduceOp(&result, og); + // broadcast, and each gradient is only broadcast once. + if (static_cast(boost::get(op->GetAttr( + OpProtoAndCheckerMaker::OpRoleAttrName())) & + static_cast(OpRole::kBackward))) { + try { + auto backward_vars = + boost::get>(op->GetNullableAttr( + OpProtoAndCheckerMaker::OpRoleVarAttrName())); + + PADDLE_ENFORCE_EQ(backward_vars.size() % 2, 0); + + for (size_t i = 0; i < backward_vars.size(); i += 2) { + auto &p_name = backward_vars[i]; + auto &g_name = backward_vars[i + 1]; + VLOG(10) << "Bcast " << g_name << " for parameter " << p_name; + + switch (strategy_.reduce_) { + case BuildStrategy::ReduceStrategy::kReduce: + CreateReduceOp(&result, g_name, cur_device_id); + var_name_on_devices[cur_device_id].emplace(g_name); + bcast_var_name_set[cur_device_id].emplace(p_name); + cur_device_id = (cur_device_id + 1) % places_.size(); + break; + case BuildStrategy::ReduceStrategy::kAllReduce: + if (IsSparseGradient(var_types, g_name)) { + CreateReduceOp(&result, g_name, 0); + CreateBroadcastOp(&result, g_name, 0); + } else { + InsertNCCLAllReduceOp(&result, g_name); + } + break; + } + } + } catch (boost::bad_get e) { } } } } } + // Insert BCast Ops + for (size_t dev_id = 0; dev_id < bcast_var_name_set.size(); ++dev_id) { + auto &to_bcast_set = bcast_var_name_set[dev_id]; + for (auto &bcast_name : to_bcast_set) { + CreateBroadcastOp(&result, bcast_name, dev_id); + } + } /* Dependency graph has been constructed. However, there are still data harzaeds need to be handled. @@ -167,14 +261,57 @@ std::unique_ptr MultiDevSSAGraphBuilder::Build( AddOutputToLeafOps(&result); if (VLOG_IS_ON(10)) { - std::ostringstream sout; - PrintGraphviz(*graph, sout); - VLOG(10) << sout.str(); + std::ofstream fout(FLAGS_ssa_graph_path); + PrintGraphviz(*graph, fout); } return std::unique_ptr(graph); } +bool MultiDevSSAGraphBuilder::IsSparseGradient( + const std::unordered_map &var_types, + const std::string &og) const { + PADDLE_ENFORCE(var_types.count(og) != 0); + if (var_types.at(og) == proto::VarType::SELECTED_ROWS) { + return true; + } + return false; +} + +void MultiDevSSAGraphBuilder::CreateBroadcastOp(SSAGraph *result, + const std::string &p_name, + size_t src_dev_id) const { +#ifdef PADDLE_WITH_CUDA + auto *op_handle = new BroadcastOpHandle(local_scopes_, places_, nccl_ctxs_); +#else + auto *op_handle = new BroadcastOpHandle(local_scopes_, places_); +#endif + + result->ops_.emplace_back(op_handle); + auto *in = result->vars_.at(src_dev_id).at(p_name).back().get(); + op_handle->AddInput(in); + + for (size_t i = 0; i < places_.size(); ++i) { + auto &vars = result->vars_.at(i).at(p_name); + auto &p = places_[i]; + auto *out_var = new VarHandle(vars.size(), i, p_name, p); + vars.emplace_back(out_var); + op_handle->AddOutput(out_var); +#ifndef ADDLE_WITH_CUDA + op_handle->SetDeviceContext(p, + platform::DeviceContextPool::Instance().Get(p)); +#endif + } +} + +void MultiDevSSAGraphBuilder::CreateComputationalOp(SSAGraph *result, + const OpDesc &op, + int dev_id) const { + result->ops_.emplace_back( + new ComputationOpHandle(op, local_scopes_[dev_id], places_[dev_id])); + CreateOpHandleIOs(result, op, dev_id); +} + void MultiDevSSAGraphBuilder::InsertNCCLAllReduceOp( SSAGraph *result, const std::string &og) const { #ifdef PADDLE_WITH_CUDA @@ -210,6 +347,26 @@ bool MultiDevSSAGraphBuilder::IsParameterGradientOnce( return is_pg_once; } +int MultiDevSSAGraphBuilder::GetOpDeviceID( + const std::vector> &var_name_on_devices, + const OpDesc &op) const { + if (strategy_.reduce_ != BuildStrategy::ReduceStrategy::kReduce) { + return -1; + } + + int var_dev_id = -1; + for (auto &var_name : op.InputArgumentNames()) { + if (var_dev_id != -1) break; + for (size_t i = 0; i < var_name_on_devices.size(); ++i) { + if (var_name_on_devices[i].count(var_name)) { + var_dev_id = static_cast(i); + break; + } + } + } + return var_dev_id; +} + void MultiDevSSAGraphBuilder::CreateScaleLossGradOp(SSAGraph *result) const { for (size_t i = 0; i < places_.size(); ++i) { // Insert ScaleCost OpHandle @@ -243,25 +400,91 @@ void MultiDevSSAGraphBuilder::CreateComputationalOps(SSAGraph *result, auto p = places_[scope_idx]; auto s = local_scopes_[scope_idx]; result->ops_.emplace_back(new ComputationOpHandle(op, s, p)); - CreateOpHandleIOs(result, op, p, scope_idx); + CreateOpHandleIOs(result, op, scope_idx); + } +} + +VarHandle *MultiDevSSAGraphBuilder::CreateReduceOp(SSAGraph *result, + const std::string &og, + int dst_dev_id) const { +#ifdef PADDLE_WITH_CUDA + result->ops_.emplace_back( + new ReduceOpHandle(local_scopes_, places_, nccl_ctxs_)); +#else + result->ops_.emplace_back(new ReduceOpHandle(local_scopes_, places_)); +#endif + auto *op_handle = result->ops_.back().get(); + + for (size_t i = 0; i < places_.size(); ++i) { + auto &vars = result->vars_[i][og]; +#ifndef PADDLE_WITH_CUDA + auto &p = places_[i]; + op_handle->SetDeviceContext(p, + platform::DeviceContextPool::Instance().Get(p)); +#endif + PADDLE_ENFORCE(!vars.empty()); + auto &prev_grad = vars.back(); + op_handle->AddInput(prev_grad.get()); + } + auto &vars = result->vars_[dst_dev_id][og]; + auto var = + new VarHandle(vars.size() - 1, dst_dev_id, og, places_[dst_dev_id]); + vars.emplace_back(var); + op_handle->AddOutput(var); + return var; +} + +void MultiDevSSAGraphBuilder::ConnectOp(SSAGraph *result, OpHandleBase *op, + const std::string &prev_op_name) const { + for (auto &prev_op : result->ops_) { + if (prev_op->Name() == prev_op_name) { + auto *dep_var = new DummyVarHandle(); + prev_op->AddOutput(dep_var); + result->dep_vars_.emplace(dep_var); + op->AddInput(dep_var); + } } } -void MultiDevSSAGraphBuilder::CreateSendOp(SSAGraph *result, - const OpDesc &op) const { +void MultiDevSSAGraphBuilder::CreateDistTrainOp(SSAGraph *result, + const OpDesc &op) const { + CreateComputationalOp(result, op, 0); + if (op.Type() == "concat") { + ConnectOp(result, result->ops_.back().get(), "fetch_barrier"); + } +} + +void MultiDevSSAGraphBuilder::CreateRPCOp(SSAGraph *result, + const OpDesc &op) const { auto &p = places_[0]; auto *s = local_scopes_[0]; - // FIXME(wuyi): send op always copy from GPU 0 - result->ops_.emplace_back(new SendOpHandle(op, s, p)); - // Create inputs for output on original place and no ssa output - // is created for send op. - CreateOpHandleIOs(result, op, p, 0); + result->ops_.emplace_back(new RPCOpHandle(op, s, p, op.Type())); + + if (op.Type() == "send_barrier") { + ConnectOp(result, result->ops_.back().get(), "send_vars"); + } else if (op.Type() == "recv") { + ConnectOp(result, result->ops_.back().get(), "send_barrier"); + } else if (op.Type() == "fetch_barrier") { + ConnectOp(result, result->ops_.back().get(), "recv"); + } else if (op.Type() == "send_vars") { + // do nothing + } else { + PADDLE_THROW( + "rpc op should be in [" + "send_vars, send_barrier. recv, fetch_barrier]"); + } + + // TODO(Yancey1989): schedule rpc op on different place may + // increate throughput + CreateOpHandleIOs(result, op, 0); } bool MultiDevSSAGraphBuilder::IsScaleLossOp(const OpDesc &op) const { - // FIXME(yy): Do not hard code like this - return op.OutputArgumentNames().size() == 1 && - op.OutputArgumentNames()[0] == GradVarName(loss_var_name_); + return boost::get( + op.GetAttr(OpProtoAndCheckerMaker::OpRoleAttrName())) == + (static_cast(OpRole::kBackward) | + static_cast(OpRole::kLoss)) && + !loss_var_name_.empty(); // If loss_var is empty. This is test mode } } // namespace details } // namespace framework diff --git a/paddle/fluid/framework/details/multi_devices_graph_builder.h b/paddle/fluid/framework/details/multi_devices_graph_builder.h index dc3da70eda2abaa1a312c25aedf94fa7e427c78a..e07597dbd80889c366babe79455beb12c9eb80d9 100644 --- a/paddle/fluid/framework/details/multi_devices_graph_builder.h +++ b/paddle/fluid/framework/details/multi_devices_graph_builder.h @@ -13,10 +13,11 @@ // limitations under the License. #pragma once - #include +#include #include +#include "paddle/fluid/framework/details/build_strategy.h" #include "paddle/fluid/framework/details/ssa_graph_builder.h" namespace paddle { @@ -27,6 +28,7 @@ class NCCLContextMap; namespace framework { class Scope; namespace details { + class MultiDevSSAGraphBuilder : public SSAGraphBuilder { public: #ifdef PADDLE_WITH_CUDA @@ -34,21 +36,21 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder { const std::string &loss_var_name, const std::unordered_set ¶ms, const std::vector &local_scopes, - bool skip_scale_loss, - platform::NCCLContextMap *nccl_ctxs); + platform::NCCLContextMap *nccl_ctxs, + const BuildStrategy &strategy); #else MultiDevSSAGraphBuilder(const std::vector &places, const std::string &loss_var_name, const std::unordered_set ¶ms, const std::vector &local_scopes, - bool skip_scale_loss); + const BuildStrategy &strategy); #endif std::unique_ptr Build(const ProgramDesc &program) const override; private: void CreateOpHandleIOs(SSAGraph *result, const OpDesc &op, - const platform::Place &p, const size_t &i) const; + size_t place_id) const; private: std::string loss_var_name_; @@ -59,24 +61,56 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder { #ifdef PADDLE_WITH_CUDA platform::NCCLContextMap *nccl_ctxs_; #endif - bool skip_scale_loss_; bool IsScaleLossOp(const OpDesc &op) const; - void CreateSendOp(SSAGraph *result, const OpDesc &op) const; + void CreateRPCOp(SSAGraph *result, const OpDesc &op) const; + void CreateDistTrainOp(SSAGraph *result, const OpDesc &op) const; + + /** + * Is this operator as the end-point operator before/after send operator. + */ + bool IsDistTrainOp(const OpDesc &op, + const std::vector &send_vars, + const std::vector &recv_vars) const; + + std::vector FindDistTrainSendVars( + const ProgramDesc &program) const; + + std::vector FindDistTrainRecvVars( + const ProgramDesc &program) const; - bool IsDistTrainOp(const OpDesc &op, OpDesc *send_op) const; + void ConnectOp(SSAGraph *result, OpHandleBase *op, + const std::string &prev_op_name) const; void CreateComputationalOps(SSAGraph *result, const OpDesc &op, size_t num_places) const; void CreateScaleLossGradOp(SSAGraph *result) const; + VarHandle *CreateReduceOp(SSAGraph *result, const std::string &og, + int dst_dev_id) const; + void CreateComputationalOp(SSAGraph *result, const OpDesc &op, + int dev_id) const; bool IsParameterGradientOnce( const std::string &og, std::unordered_set *og_has_been_broadcast) const; + int GetOpDeviceID( + const std::vector> &var_name_on_devices, + const OpDesc &op) const; + void InsertNCCLAllReduceOp(SSAGraph *result, const std::string &og) const; + + void CreateBroadcastOp(SSAGraph *result, const std::string &p_name, + size_t src_dev_id) const; + + bool IsSparseGradient( + const std::unordered_map &var_types, + const std::string &og) const; + + private: + BuildStrategy strategy_; }; } // namespace details } // namespace framework diff --git a/paddle/fluid/framework/details/nccl_all_reduce_op_handle.cc b/paddle/fluid/framework/details/nccl_all_reduce_op_handle.cc index b055bb48f608c9fd9cc671d175cb463d25dc489b..95aa599cd3e403e9cc66b2b5ad35d0d214d1ab5b 100644 --- a/paddle/fluid/framework/details/nccl_all_reduce_op_handle.cc +++ b/paddle/fluid/framework/details/nccl_all_reduce_op_handle.cc @@ -34,10 +34,7 @@ void NCCLAllReduceOpHandle::RunImpl() { return; // No need to all reduce when GPU count = 1; } else { // Wait input done - for (auto *in : inputs_) { - auto &p = static_cast(in)->place_; - in->generated_op_->Wait(dev_ctxes_[p]); - } + WaitInputVarGenerated(); auto &var_name = static_cast(this->inputs_[0])->name_; int dtype = -1; diff --git a/paddle/fluid/framework/details/op_handle_base.cc b/paddle/fluid/framework/details/op_handle_base.cc index 534d77860f87be08c8834efd373d90eb199ed6a2..6b064650b4f09737836bda4a43fa421720077929 100644 --- a/paddle/fluid/framework/details/op_handle_base.cc +++ b/paddle/fluid/framework/details/op_handle_base.cc @@ -56,15 +56,15 @@ void OpHandleBase::Run(bool use_event) { RunImpl(); } -void OpHandleBase::Wait(platform::DeviceContext *waited_dev) { +void OpHandleBase::RecordWaitEventOnCtx(platform::DeviceContext *waited_ctx) { #ifdef PADDLE_WITH_CUDA - if (platform::is_cpu_place(waited_dev->GetPlace()) || events_.empty()) { + if (platform::is_cpu_place(waited_ctx->GetPlace()) || events_.empty()) { for (auto &dev_ctx : dev_ctxes_) { dev_ctx.second->Wait(); } } else { auto stream = - static_cast(waited_dev)->stream(); + static_cast(waited_ctx)->stream(); for (auto &ev : events_) { PADDLE_ENFORCE(cudaStreamWaitEvent(stream, ev.second, 0)); } @@ -86,6 +86,28 @@ void OpHandleBase::AddOutput(VarHandleBase *out) { out->generated_op_ = this; } +void OpHandleBase::WaitInputVarGenerated() { + for (auto in_var : inputs_) { + if (NeedWait(in_var)) { + for (auto &pair : dev_ctxes_) { + in_var->generated_op_->RecordWaitEventOnCtx(pair.second); + } + } + } +} + +void OpHandleBase::WaitInputVarGenerated(const platform::Place &place) { + for (auto *in : inputs_) { + if (NeedWait(in)) { + in->generated_op_->RecordWaitEventOnCtx(dev_ctxes_[place]); + } + } +} + +bool OpHandleBase::NeedWait(VarHandleBase *in_var) { + return in_var && in_var->generated_op_; +} + void OpHandleBase::RunAndRecordEvent(const std::function &callback) { #ifdef PADDLE_WITH_CUDA if (!events_.empty()) { // Use event diff --git a/paddle/fluid/framework/details/op_handle_base.h b/paddle/fluid/framework/details/op_handle_base.h index 00f213f3ed294adcce7c540e3ff346de8e2be7fb..8f94206a87dbae8a81727ca48718886bbabbe25c 100644 --- a/paddle/fluid/framework/details/op_handle_base.h +++ b/paddle/fluid/framework/details/op_handle_base.h @@ -38,12 +38,24 @@ class OpHandleBase { void Run(bool use_event); - virtual void Wait(platform::DeviceContext *waited_dev); + virtual void RecordWaitEventOnCtx(platform::DeviceContext *waited_ctx); void AddInput(VarHandleBase *in); void AddOutput(VarHandleBase *out); + // This method adds the wait events of all the input on all the device + // context. + // NODE: This Wait is asynchronous operation. + virtual void WaitInputVarGenerated(); + + // This method adds the wait events of all the input on the specified device + // context. + // NODE: This Wait is asynchronous operation. + virtual void WaitInputVarGenerated(const platform::Place &place); + + virtual bool NeedWait(VarHandleBase *in_var); + // If the Op involves data transfer of multiple devices that // will likely block other computations. virtual bool IsMultiDeviceTransfer() { return false; } @@ -58,6 +70,14 @@ class OpHandleBase { const std::vector &Inputs() const { return inputs_; } + size_t NoDupInputSize() const { + std::unordered_set res; + for (auto *var : inputs_) { + res.emplace(var); + } + return res.size(); + } + const std::vector &Outputs() const { return outputs_; } protected: diff --git a/paddle/fluid/framework/details/op_registry.h b/paddle/fluid/framework/details/op_registry.h index 06603db31e0092382c0cc05482a038473d647ef1..eea7e712f8f6e187cdceedce77cc76d1d4ca2101 100644 --- a/paddle/fluid/framework/details/op_registry.h +++ b/paddle/fluid/framework/details/op_registry.h @@ -95,8 +95,8 @@ struct OpInfoFiller { void operator()(const char* op_type, OpInfo* info) const { info->proto_ = new proto::OpProto; info->checker_ = new OpAttrChecker(); - auto maker = T(info->proto_, info->checker_); - maker.Validate(); + T maker; + maker(info->proto_, info->checker_); info->proto_->set_type(op_type); PADDLE_ENFORCE( info->proto_->IsInitialized(), diff --git a/paddle/fluid/framework/details/reduce_op_handle.cc b/paddle/fluid/framework/details/reduce_op_handle.cc index 409e8f72b841de03dcb50e62de447ae9895df2c0..7160e346dad0615e2fd32b70c096880af0359e1a 100644 --- a/paddle/fluid/framework/details/reduce_op_handle.cc +++ b/paddle/fluid/framework/details/reduce_op_handle.cc @@ -22,6 +22,7 @@ namespace framework { namespace details { void ReduceOpHandle::RunImpl() { + if (places_.size() == 1) return; // the input and output may have dummy var. auto in_var_handles = DynamicCast(inputs_); @@ -50,45 +51,51 @@ void ReduceOpHandle::RunImpl() { PADDLE_ENFORCE_NOT_NULL(pre_in_var); // Wait input done, this Wait is asynchronous operation - WaitInputVarGenerated(in_var_handles); - auto pre_place = in_0_handle->place_; - std::vector in_places; - auto pre_in_tensor = VariableVisitor::GetMutableTensor(pre_in_var); - for (auto *in_handle : in_var_handles) { - auto in_p = in_handle->place_; - PADDLE_ENFORCE_EQ(in_p.which(), pre_place.which(), - "Places must be all on CPU or all on CUDA."); - in_places.emplace_back(in_p); + WaitInputVarGenerated(); + // NOTE: The Places of all input tensor must be all on CPU or all on GPU. + std::vector in_places; // used to get dev_ctx + for (auto *in_handle : in_var_handles) { + in_places.emplace_back(in_handle->place_); auto in_var = var_scopes.at(in_handle->scope_idx_)->FindVar(in_handle->name_); PADDLE_ENFORCE_NOT_NULL(in_var); - - auto in_tensor = VariableVisitor::GetMutableTensor(in_var); - PADDLE_ENFORCE_EQ(in_tensor.type(), pre_in_tensor.type(), - "The type of input is not consistent."); + VariableVisitor::EnforceShapeAndDTypeEQ(*pre_in_var, *in_var); } auto out_var = var_scopes.at(out_var_handle->scope_idx_)->FindVar(out_var_handle->name_); PADDLE_ENFORCE_NOT_NULL(out_var); - if (pre_in_var->IsType()) { - std::vector in_selected_rows = - GetInputValues(in_var_handles, var_scopes); + // NOTE: The tensors' Place of input and output must be all on GPU or all on + // CPU. + auto in_p = VariableVisitor::GetMutableTensor(pre_in_var).place(); + platform::Place t_out_p; + if (platform::is_gpu_place(in_p)) { + PADDLE_ENFORCE(platform::is_gpu_place(out_var_handle->place_), + "Places of input and output must be all on GPU."); + t_out_p = out_var_handle->place_; + } else { + t_out_p = platform::CPUPlace(); + } - GatherSelectedRows(in_selected_rows, in_places, dev_ctxes_, - out_var_handle->place_, - out_var->GetMutable()); + if (pre_in_var->IsType()) { + this->RunAndRecordEvent([&] { + std::vector in_selected_rows = + GetInputValues(in_var_handles, var_scopes); + GatherSelectedRows(in_selected_rows, in_places, dev_ctxes_, t_out_p, + out_var->GetMutable()); + }); } else { std::vector lod_tensors = GetInputValues(in_var_handles, var_scopes); - - if (paddle::platform::is_cpu_place(pre_place)) { - ReduceLoDTensor func(lod_tensors, - out_var->GetMutable()); - VisitDataType(ToDataType(lod_tensors[0]->type()), func); - } else if (paddle::platform::is_gpu_place(pre_place)) { + if (paddle::platform::is_cpu_place(lod_tensors[0]->place())) { + this->RunAndRecordEvent([&] { + ReduceLoDTensor func(lod_tensors, + out_var->GetMutable()); + VisitDataType(ToDataType(lod_tensors[0]->type()), func); + }); + } else if (paddle::platform::is_gpu_place(lod_tensors[0]->place())) { #ifdef PADDLE_WITH_CUDA auto pre_in = pre_in_var->Get(); VariableVisitor::ShareDimsAndLoD(*pre_in_var, out_var); @@ -96,7 +103,7 @@ void ReduceOpHandle::RunImpl() { out_var_handle->place_, pre_in.type()); auto out_p = out_var_handle->place_; - int root = boost::get(out_p).device; + int root_id = boost::get(out_p).device; std::vector> all_reduce_calls; for (size_t i = 0; i < var_scopes.size(); ++i) { auto &p = in_places[i]; @@ -104,23 +111,23 @@ void ReduceOpHandle::RunImpl() { int dev_id = boost::get(p).device; auto &nccl_ctx = nccl_ctxs_->at(dev_id); - auto stream = nccl_ctx.stream(); - auto comm = nccl_ctx.comm_; void *buffer = const_cast(lod_tensor.data()); void *recvbuffer = nullptr; - if (root == dev_id) { + if (root_id == dev_id) { recvbuffer = out_var->GetMutable()->mutable_data( out_var_handle->place_); } int type = platform::ToNCCLDataType(lod_tensor.type()); - all_reduce_calls.emplace_back([=] { - PADDLE_ENFORCE(platform::dynload::ncclReduce( - buffer, recvbuffer, static_cast(lod_tensor.numel()), - static_cast(type), ncclSum, root, comm, stream)); - }); + size_t numel = static_cast(lod_tensor.numel()); + all_reduce_calls.emplace_back( + [buffer, recvbuffer, type, numel, root_id, &nccl_ctx] { + PADDLE_ENFORCE(platform::dynload::ncclReduce( + buffer, recvbuffer, numel, static_cast(type), + ncclSum, root_id, nccl_ctx.comm_, nccl_ctx.stream())); + }); } this->RunAndRecordEvent([&] { @@ -130,7 +137,7 @@ void ReduceOpHandle::RunImpl() { } }); #else - PADDLE_THROW("CUDA is not support."); + PADDLE_THROW("CUDA is not enabled."); #endif } else { PADDLE_THROW("Place should be CPUPlace or CUDAPlace."); @@ -152,17 +159,6 @@ std::vector ReduceOpHandle::GetInputValues( return in_selected_rows; } -void ReduceOpHandle::WaitInputVarGenerated( - const std::vector &in_var_handles) { - for (auto *in : in_var_handles) { - if (in->generated_op_) { - for (auto pair : dev_ctxes_) { - in->generated_op_->Wait(pair.second); - } - } - } -} - std::string ReduceOpHandle::Name() const { return "reduce"; } } // namespace details } // namespace framework diff --git a/paddle/fluid/framework/details/reduce_op_handle.h b/paddle/fluid/framework/details/reduce_op_handle.h index 9746b3bdbde14d24a83a27a593c5f1ebfec201ff..c652a2f4eb0f9b73cb19ebbd9d0809210b280ad3 100644 --- a/paddle/fluid/framework/details/reduce_op_handle.h +++ b/paddle/fluid/framework/details/reduce_op_handle.h @@ -55,13 +55,11 @@ struct ReduceOpHandle : public OpHandleBase { std::string Name() const override; - bool IsMultiDeviceTransfer() override { return false; }; + bool IsMultiDeviceTransfer() override { return true; }; protected: void RunImpl() override; - void WaitInputVarGenerated(const std::vector &in_var_handles); - template std::vector GetInputValues( const std::vector &in_var_handles, diff --git a/paddle/fluid/framework/details/reduce_op_handle_test.cc b/paddle/fluid/framework/details/reduce_op_handle_test.cc index c17aabee53680fba10eac289cf8f8bd5f7d419e8..ffdd7c14eb5097cc8285da090e4a72e1e3f43d86 100644 --- a/paddle/fluid/framework/details/reduce_op_handle_test.cc +++ b/paddle/fluid/framework/details/reduce_op_handle_test.cc @@ -194,7 +194,7 @@ struct TestReduceOpHandle { } f::Tensor result_tensor; - f::TensorCopy(rt, cpu_place, *(ctxs_[output_scope_idx]), &result_tensor); + f::TensorCopySync(rt, cpu_place, &result_tensor); float *ct = result_tensor.data(); for (int64_t j = 0; j < f::product(result_tensor.dims()); ++j) { @@ -239,7 +239,7 @@ struct TestReduceOpHandle { auto &rt = out_var->Get(); f::Tensor result_tensor; - f::TensorCopy(rt, cpu_place, *(ctxs_[output_scope_idx]), &result_tensor); + f::TensorCopySync(rt, cpu_place, &result_tensor); float *ct = result_tensor.data(); for (int64_t j = 0; j < f::product(result_tensor.dims()); ++j) { diff --git a/paddle/fluid/framework/details/send_op_handle.cc b/paddle/fluid/framework/details/rpc_op_handle.cc similarity index 66% rename from paddle/fluid/framework/details/send_op_handle.cc rename to paddle/fluid/framework/details/rpc_op_handle.cc index 0763f92171e7813ec0ee8ca4f3aa42b76205130a..7f4da4c01de1010467d839ee5490c5e0d02d8c24 100644 --- a/paddle/fluid/framework/details/send_op_handle.cc +++ b/paddle/fluid/framework/details/rpc_op_handle.cc @@ -12,27 +12,32 @@ // See the License for the specific language governing permissions and // limitations under the License. -#include "paddle/fluid/framework/details/send_op_handle.h" +#include "paddle/fluid/framework/details/rpc_op_handle.h" namespace paddle { namespace framework { namespace details { -SendOpHandle::SendOpHandle(const framework::OpDesc &op_desc, - const Scope *local_scope, - const platform::Place &place) +RPCOpHandle::RPCOpHandle(const framework::OpDesc &op_desc, + const Scope *local_scope, const platform::Place &place, + const std::string &name) : op_(framework::OpRegistry::CreateOp(op_desc)), local_scope_(local_scope), - place_(place) {} + place_(place), + name_(name) {} -void SendOpHandle::RunImpl() { +void RPCOpHandle::RunImpl() { + // TODO(wuyi): need further analysis whether wait VarDummyHandle. // Wait input done for (auto *in : inputs_) { auto &p = static_cast(in)->place_; + // FIXME(Yancey1989): need a better solution instead of use DebugString() if (in->DebugString() == "dummy") { // HACK continue; } - in->generated_op_->Wait(dev_ctxes_[p]); + if (in->generated_op_) { + in->generated_op_->RecordWaitEventOnCtx(dev_ctxes_[p]); + } } auto &tmp_scope = local_scope_->FindVar(kLocalExecScopeName)->Get(); // FIXME(wuyi): can not use RunAndRecordEvent here, for it will cause dead @@ -40,7 +45,7 @@ void SendOpHandle::RunImpl() { op_->Run(*tmp_scope, place_); } -std::string SendOpHandle::Name() const { return "send"; } +std::string RPCOpHandle::Name() const { return name_; } } // namespace details } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/details/send_op_handle.h b/paddle/fluid/framework/details/rpc_op_handle.h similarity index 87% rename from paddle/fluid/framework/details/send_op_handle.h rename to paddle/fluid/framework/details/rpc_op_handle.h index 2f78811fad50642b5e45776c41910df6f4cc48f6..d28b7721720d808a8d81701c3811eae16121fb41 100644 --- a/paddle/fluid/framework/details/send_op_handle.h +++ b/paddle/fluid/framework/details/rpc_op_handle.h @@ -27,9 +27,9 @@ namespace paddle { namespace framework { namespace details { -struct SendOpHandle : public OpHandleBase { - SendOpHandle(const framework::OpDesc& op_desc, const Scope* local_scope, - const platform::Place& place); +struct RPCOpHandle : public OpHandleBase { + RPCOpHandle(const framework::OpDesc& op_desc, const Scope* local_scope, + const platform::Place& place, const std::string& name); std::string Name() const override; @@ -44,6 +44,7 @@ struct SendOpHandle : public OpHandleBase { std::unique_ptr op_; const Scope* local_scope_; const platform::Place& place_; + const std::string name_; }; } // namespace details diff --git a/paddle/fluid/framework/details/scale_loss_grad_op_handle.cc b/paddle/fluid/framework/details/scale_loss_grad_op_handle.cc index 7a65ee62c9bfc0dad2ebee3be21de825fa405d73..d9c387e79dc71288e7330597fed57171d447f31b 100644 --- a/paddle/fluid/framework/details/scale_loss_grad_op_handle.cc +++ b/paddle/fluid/framework/details/scale_loss_grad_op_handle.cc @@ -29,6 +29,7 @@ ScaleLossGradOpHandle::ScaleLossGradOpHandle(size_t num_dev, Scope *scope, ScaleLossGradOpHandle::~ScaleLossGradOpHandle() {} void ScaleLossGradOpHandle::RunImpl() { + // Doesn't wait any event std::string var_name = static_cast(this->outputs_[0])->name_; auto &local_scope = *scope_->FindVar(kLocalExecScopeName)->Get(); @@ -46,6 +47,7 @@ void ScaleLossGradOpHandle::RunImpl() { ->stream(); memory::Copy(boost::get(place_), tmp, platform::CPUPlace(), &coeff_, sizeof(float), stream); + VLOG(1) << place_ << "RUN Scale loss grad op"; }); #endif } diff --git a/paddle/fluid/framework/details/ssa_graph.h b/paddle/fluid/framework/details/ssa_graph.h index 72684e7f97f1324d0efba960903cf9f2acb618a4..e996a00c162186e47e77d007503ac67caa9f8024 100644 --- a/paddle/fluid/framework/details/ssa_graph.h +++ b/paddle/fluid/framework/details/ssa_graph.h @@ -25,12 +25,22 @@ namespace paddle { namespace framework { namespace details { +// A SSA graph used by parallel executor. struct SSAGraph { + // all variable in each devices. + // The outside vector is the device vector. Each element of this vector is a + // map from variable name to variables. The variables, who have the same name, + // will have a different version. The offset in the + // `std::vector>` is the version of varaibles. std::vector< std::unordered_map>>> vars_; + // aux variables to represent dependency. Useful to resolve data hazard. std::unordered_set> dep_vars_; + + // all operators. NOTE that even we use a vector here, the operators is + // unordered. std::vector> ops_; }; diff --git a/paddle/fluid/framework/details/ssa_graph_builder.h b/paddle/fluid/framework/details/ssa_graph_builder.h index be1f0460e45402806b18835f054a7195df1374cc..64e5d93081eb76c56898bbeb530e37364619fdbb 100644 --- a/paddle/fluid/framework/details/ssa_graph_builder.h +++ b/paddle/fluid/framework/details/ssa_graph_builder.h @@ -48,6 +48,8 @@ class SSAGraphBuilder { const platform::Place &place, size_t place_offset); + // Add an output variable (each_var_name, place, place_offset) to op_handle, + // which belongs to graph static void CreateOpOutput(SSAGraph *graph, OpHandleBase *op_handle, const std::string &each_var_name, const platform::Place &place, size_t place_offset); diff --git a/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc b/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc index 5e6ed5cb7cdc534332d402380458f39aecd841b8..815f739371e77d953a28be99b38ec1b8ff26506c 100644 --- a/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc +++ b/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc @@ -14,24 +14,21 @@ #include "paddle/fluid/framework/details/threaded_ssa_graph_executor.h" -#include "paddle/fluid/framework/details/fetch_op_handle.h" - namespace paddle { namespace framework { namespace details { ThreadedSSAGraphExecutor::ThreadedSSAGraphExecutor( - size_t num_threads, bool use_event, - const std::vector &local_scopes, + const ExecutionStrategy &strategy, const std::vector &local_scopes, const std::vector &places, - std::unique_ptr &&graph, bool allow_op_delay) + std::unique_ptr &&graph) : SSAGraphExecutor(std::move(graph)), - pool_(num_threads >= 2 ? new ::ThreadPool(num_threads) : nullptr), + pool_(strategy.num_threads_ >= 2 ? new ::ThreadPool(strategy.num_threads_) + : nullptr), local_scopes_(local_scopes), places_(places), fetch_ctxs_(places), - use_event_(use_event), running_ops_(0), - allow_op_delay_(allow_op_delay) {} + strategy_(strategy) {} FeedFetchList ThreadedSSAGraphExecutor::Run( const std::vector &fetch_tensors) { @@ -45,73 +42,33 @@ FeedFetchList ThreadedSSAGraphExecutor::Run( // Should revisit it if overlapping is available. std::unordered_set delayed_ops; - auto InsertPendingVar = [&pending_vars, &ready_vars](VarHandleBase &var) { - pending_vars.insert(&var); - if (var.generated_op_ == nullptr) { - ready_vars.Push(&var); - } - }; - - auto InsertPendingOp = [&pending_ops](OpHandleBase &op_instance) { - pending_ops.insert({&op_instance, op_instance.Inputs().size()}); - }; - // Transform SSAGraph to pending_ops & pending_vars for (auto &var_map : graph_->vars_) { for (auto &name_pair : var_map) { for (auto &version_pair : name_pair.second) { - InsertPendingVar(*version_pair); + InsertPendingVar(&pending_vars, &ready_vars, version_pair.get()); } } } for (auto &var : graph_->dep_vars_) { - InsertPendingVar(*var); + InsertPendingVar(&pending_vars, &ready_vars, var.get()); } for (auto &op : graph_->ops_) { if (op->Inputs().empty()) { // Special case, Op has no input. ready_ops.insert(op.get()); } else { - InsertPendingOp(*op); + InsertPendingOp(&pending_ops, op.get()); } } // Step 2. Insert FetchOps std::vector> fetch_ops; - FeedFetchList fetch_data(fetch_tensors.size()); - - std::unordered_map> fetched_vars; - - for (auto &fetch_var_name : fetch_tensors) { - for (auto &var_map : graph_->vars_) { - auto it = var_map.find(fetch_var_name); - if (it != var_map.end()) { - fetched_vars[fetch_var_name].push_back(it->second.rbegin()->get()); - } - } - } - std::unordered_set> fetch_dependencies; - for (size_t i = 0; i < fetch_tensors.size(); ++i) { - auto &var_name = fetch_tensors[i]; - auto &vars = fetched_vars.at(var_name); - auto *op = new FetchOpHandle(&fetch_data, i, &local_scopes_); - fetch_ops.emplace_back(op); - - for (auto &p : places_) { - op->SetDeviceContext(p, fetch_ctxs_.Get(p)); - } - - for (auto *var : vars) { - op->AddInput(var); - } + FeedFetchList fetch_data(fetch_tensors.size()); - auto *fetch_dummy = new DummyVarHandle(); - op->AddOutput(fetch_dummy); - fetch_dependencies.emplace(fetch_dummy); - InsertPendingVar(*fetch_dummy); - InsertPendingOp(*op); - } + InsertFetchOps(fetch_tensors, &fetch_ops, &fetch_dependencies, &pending_ops, + &pending_vars, &ready_vars, &fetch_data); auto run_all_ops = [&](std::unordered_set &set) { for (auto *op : set) { @@ -128,7 +85,7 @@ FeedFetchList ThreadedSSAGraphExecutor::Run( // // NOTE: DelayedOps have a lower priority. It will be scheduled after all // ready_ops have been performed. - if (ready_ops.empty() && allow_op_delay_ && running_ops_ == 0) { + if (ready_ops.empty() && strategy_.allow_op_delay_ && running_ops_ == 0) { run_all_ops(delayed_ops); } else { run_all_ops(ready_ops); @@ -155,7 +112,7 @@ FeedFetchList ThreadedSSAGraphExecutor::Run( auto &deps = pending_ops[op]; --deps; if (deps == 0) { - if (op->IsMultiDeviceTransfer() && allow_op_delay_) { + if (op->IsMultiDeviceTransfer() && strategy_.allow_op_delay_) { delayed_ops.insert(op); } else { ready_ops.insert(op); @@ -174,12 +131,66 @@ FeedFetchList ThreadedSSAGraphExecutor::Run( return fetch_data; } +void ThreadedSSAGraphExecutor::InsertFetchOps( + const std::vector &fetch_tensors, + std::vector> *fetch_ops, + std::unordered_set> *fetch_dependencies, + std::unordered_map *pending_ops, + std::unordered_set *pending_vars, + BlockingQueue *ready_vars, FeedFetchList *fetch_data) { + std::unordered_map> fetched_vars; + + for (auto &fetch_var_name : fetch_tensors) { + for (auto &var_map : graph_->vars_) { + auto it = var_map.find(fetch_var_name); + if (it != var_map.end()) { + fetched_vars[fetch_var_name].push_back(it->second.rbegin()->get()); + } + } + } + + for (size_t i = 0; i < fetch_tensors.size(); ++i) { + auto &var_name = fetch_tensors[i]; + auto &vars = fetched_vars.at(var_name); + auto *op = new FetchOpHandle(fetch_data, i, &local_scopes_); + fetch_ops->emplace_back(op); + + for (auto &p : places_) { + op->SetDeviceContext(p, fetch_ctxs_.Get(p)); + } + + for (auto *var : vars) { + op->AddInput(var); + } + + auto *fetch_dummy = new DummyVarHandle(); + op->AddOutput(fetch_dummy); + fetch_dependencies->emplace(fetch_dummy); + this->InsertPendingVar(pending_vars, ready_vars, fetch_dummy); + this->InsertPendingOp(pending_ops, op); + } +} + +void ThreadedSSAGraphExecutor::InsertPendingOp( + std::unordered_map *pending_ops, + OpHandleBase *op_instance) const { + pending_ops->insert({op_instance, op_instance->NoDupInputSize()}); +} + +void ThreadedSSAGraphExecutor::InsertPendingVar( + std::unordered_set *pending_vars, + BlockingQueue *ready_vars, VarHandleBase *var) const { + pending_vars->insert(var); + if (var->generated_op_ == nullptr) { + ready_vars->Push(var); + } +} void ThreadedSSAGraphExecutor::RunOp( BlockingQueue *ready_var_q, details::OpHandleBase *op) { auto op_run = [ready_var_q, op, this] { try { VLOG(10) << op << " " << op->Name() << " : " << op->DebugString(); - op->Run(use_event_); + op->Run(strategy_.use_event_); VLOG(10) << op << " " << op->Name() << " Done "; running_ops_--; ready_var_q->Extend(op->Outputs()); diff --git a/paddle/fluid/framework/details/threaded_ssa_graph_executor.h b/paddle/fluid/framework/details/threaded_ssa_graph_executor.h index d089b79d91327e38408439a8019ec5189ff6d189..1f7f88d75218e757e4555ad093f3cd6558f624dd 100644 --- a/paddle/fluid/framework/details/threaded_ssa_graph_executor.h +++ b/paddle/fluid/framework/details/threaded_ssa_graph_executor.h @@ -23,6 +23,8 @@ #include #include "ThreadPool.h" // ThreadPool in thrird party #include "paddle/fluid/framework/blocking_queue.h" +#include "paddle/fluid/framework/details/execution_strategy.h" +#include "paddle/fluid/framework/details/fetch_op_handle.h" #include "paddle/fluid/framework/details/ssa_graph_executor.h" namespace paddle { @@ -33,11 +35,10 @@ namespace details { class ThreadedSSAGraphExecutor : public SSAGraphExecutor { public: - ThreadedSSAGraphExecutor(size_t num_threads, bool use_event, + ThreadedSSAGraphExecutor(const ExecutionStrategy &strategy, const std::vector &local_scopes, const std::vector &places, - std::unique_ptr &&graph, - bool allow_op_delay); + std::unique_ptr &&graph); // Run a SSAGraph by a thread pool // Use topological sort algorithm @@ -54,10 +55,26 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor { std::vector local_scopes_; std::vector places_; platform::DeviceContextPool fetch_ctxs_; - const bool use_event_; std::unique_ptr exception_; std::atomic running_ops_; - bool allow_op_delay_; + + void InsertPendingOp(std::unordered_map *pending_ops, + OpHandleBase *op_instance) const; + + void InsertPendingVar(std::unordered_set *pending_vars, + BlockingQueue *ready_vars, + VarHandleBase *var) const; + + void InsertFetchOps( + const std::vector &fetch_tensors, + std::vector> *fetch_ops, + std::unordered_set> *fetch_dependencies, + std::unordered_map *pending_ops, + std::unordered_set *pending_vars, + BlockingQueue *ready_vars, FeedFetchList *fetch_data); + + private: + ExecutionStrategy strategy_; }; } // namespace details diff --git a/paddle/fluid/framework/details/var_handle.h b/paddle/fluid/framework/details/var_handle.h index 9f7fd69e64fe9d7ef0bf3037bea7f686cb2eee0b..cae9af7217660fb7e4b8535ee8e022fb3a127668 100644 --- a/paddle/fluid/framework/details/var_handle.h +++ b/paddle/fluid/framework/details/var_handle.h @@ -62,7 +62,7 @@ struct VarHandle : public VarHandleBase { std::string name_; platform::Place place_; - bool operator==(const VarHandle& o) const { + bool IsTheSameVar(const VarHandle& o) const { return o.generated_op_ == generated_op_ && o.name_ == name_ && o.scope_idx_ == scope_idx_; } diff --git a/paddle/fluid/framework/details/variable_visitor.cc b/paddle/fluid/framework/details/variable_visitor.cc index 10bac0fae9504215fab11dd8cca7c278feaa4bda..3dfd14419d94379a0bf79f55d7a139acd77cbd7e 100644 --- a/paddle/fluid/framework/details/variable_visitor.cc +++ b/paddle/fluid/framework/details/variable_visitor.cc @@ -88,6 +88,52 @@ void VariableVisitor::ShareDimsAndLoD(const Variable& src, Variable* trg) { VisitVariable(src, &visitor); } +struct EnforceShapeAndDTypeEQVisitor { + const Variable* trg_; + + void operator()(const LoDTensor& src) { + auto& tensor = trg_->Get(); + PADDLE_ENFORCE_EQ( + src.place().which(), tensor.place().which(), + "The Places of the two Variable must be all on CPU or all on GPU."); + PADDLE_ENFORCE_EQ(src.type(), tensor.type(), + "The dtype of the two Variable is not equal."); + PADDLE_ENFORCE_EQ(src.dims(), tensor.dims(), + "The dims of the two Variable is not equal."); + PADDLE_ENFORCE_EQ(src.lod(), tensor.lod(), + "The lod of the two Variable is not equal."); + PADDLE_ENFORCE_EQ(src.layout(), tensor.layout(), + "The layout of the two Variable's tensor is not equal."); + } + + void operator()(const SelectedRows& src) { + auto& selected_rows = trg_->Get(); + PADDLE_ENFORCE_EQ( + src.place().which(), selected_rows.place().which(), + "The Places of the two Variable must be all on CPU or all on GPU."); + PADDLE_ENFORCE_EQ(src.value().type(), selected_rows.value().type(), + "The dtype of the two Variable is not equal."); + PADDLE_ENFORCE_EQ(src.value().layout(), selected_rows.value().layout(), + "The layout of the two Variable's tensor is not equal."); + PADDLE_ENFORCE_EQ(src.height(), selected_rows.height(), + "The height of the two Variable is not equal."); + PADDLE_ENFORCE_EQ(src.GetCompleteDims(), selected_rows.GetCompleteDims(), + "The dims of the two Variable is not equal."); + } + + template + void operator()(const T&) { + PADDLE_ENFORCE("EnforceShapeAndDTypeEQ is not supported by type %s", + typeid(T).name()); + } +}; + +void VariableVisitor::EnforceShapeAndDTypeEQ(const Variable& var1, + const Variable& var2) { + EnforceShapeAndDTypeEQVisitor visitor{&var1}; + VisitVariable(var2, &visitor); +} + } // namespace details } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/details/variable_visitor.h b/paddle/fluid/framework/details/variable_visitor.h index 67baa1895e4513738fa73d49c46660da92279b9d..ca9a19bdcf1be7bf0e1d2b0de560a38f528a2d2c 100644 --- a/paddle/fluid/framework/details/variable_visitor.h +++ b/paddle/fluid/framework/details/variable_visitor.h @@ -26,6 +26,9 @@ class VariableVisitor { static Tensor &GetMutableTensor(Variable *var); static void ShareDimsAndLoD(const Variable &src, Variable *trg); + + static void EnforceShapeAndDTypeEQ(const Variable &var1, + const Variable &var2); }; } // namespace details diff --git a/paddle/fluid/framework/executor.cc b/paddle/fluid/framework/executor.cc index 766bf0ab0c1c50146ad3f6e048738209428707b9..863053c32b190f4e8497b16f3edd76cb2f76168b 100644 --- a/paddle/fluid/framework/executor.cc +++ b/paddle/fluid/framework/executor.cc @@ -24,9 +24,6 @@ limitations under the License. */ #include "paddle/fluid/platform/profiler.h" DECLARE_bool(benchmark); -DEFINE_bool(check_nan_inf, false, - "Checking whether operator produce NAN/INF or not. It will be " - "extremely slow so please use this flag wisely."); namespace paddle { namespace framework { @@ -78,21 +75,6 @@ void InitializeVariable(Variable* var, proto::VarType::Type var_type) { } } -static void CheckTensorNANOrInf(const std::string& name, - const framework::Tensor& tensor) { - if (tensor.memory_size() == 0) { - return; - } - if (tensor.type().hash_code() != typeid(float).hash_code() && // NOLINT - tensor.type().hash_code() != typeid(double).hash_code()) { // NOLINT - return; - } - PADDLE_ENFORCE(!framework::TensorContainsInf(tensor), - "Tensor %s contains Inf", name); - PADDLE_ENFORCE(!framework::TensorContainsNAN(tensor), - "Tensor %s contains NAN", name); -} - void Executor::CreateVariables(const ProgramDesc& pdesc, Scope* scope, int block_id) { auto& global_block = pdesc.Block(block_id); @@ -228,7 +210,8 @@ static bool has_fetch_operators( void Executor::Run(const ProgramDesc& program, Scope* scope, std::map* feed_targets, std::map* fetch_targets, - bool create_vars, const std::string& feed_holder_name, + bool create_local_scope, bool create_vars, + const std::string& feed_holder_name, const std::string& fetch_holder_name) { platform::RecordBlock b(kProgramId); bool has_feed_ops = @@ -290,8 +273,9 @@ void Executor::Run(const ProgramDesc& program, Scope* scope, } auto ctx = Prepare(*copy_program, 0); - RunPreparedContext(ctx.get(), scope, feed_targets, fetch_targets, create_vars, - feed_holder_name, fetch_holder_name); + RunPreparedContext(ctx.get(), scope, feed_targets, fetch_targets, + create_local_scope, create_vars, feed_holder_name, + fetch_holder_name); } std::unique_ptr Executor::Prepare( @@ -338,18 +322,13 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope, VLOG(2) << "Memory used after operator " + op->Type() + " running: " << memory::memory_usage(place_); } - if (FLAGS_check_nan_inf) { - for (auto& vname : op->OutputVars(true)) { - auto* var = local_scope->FindVar(vname); - if (var == nullptr) continue; - if (var->IsType()) { - CheckTensorNANOrInf(vname, var->Get()); - } - } - } } + platform::DeviceContextPool::Instance().Get(place_)->Wait(); if (create_vars && create_local_scope) { scope->DeleteScope(local_scope); + } else { + // Delete the local scopes created in operators. + scope->DropKids(); } if (FLAGS_benchmark) { VLOG(2) << "-------------------------------------------------------"; @@ -362,8 +341,9 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope, void Executor::RunPreparedContext( ExecutorPrepareContext* ctx, Scope* scope, std::map* feed_targets, - std::map* fetch_targets, bool create_vars, - const std::string& feed_holder_name, const std::string& fetch_holder_name) { + std::map* fetch_targets, bool create_local_scope, + bool create_vars, const std::string& feed_holder_name, + const std::string& fetch_holder_name) { auto& global_block = ctx->prog_.Block(ctx->block_id_); PADDLE_ENFORCE( @@ -383,7 +363,7 @@ void Executor::RunPreparedContext( } } - RunPreparedContext(ctx, scope, create_vars, create_vars); + RunPreparedContext(ctx, scope, create_local_scope, create_vars); // obtain the data of fetch_targets from fetch_holder for (auto* op : global_block.AllOps()) { diff --git a/paddle/fluid/framework/executor.h b/paddle/fluid/framework/executor.h index 4a3d637e2d79f8cbd83412eea2d73e4b497ef1e7..0c3c23611d95e0da67cabfb8fb2755a4a52c991b 100644 --- a/paddle/fluid/framework/executor.h +++ b/paddle/fluid/framework/executor.h @@ -57,7 +57,7 @@ class Executor { void Run(const ProgramDesc& program, Scope* scope, std::map* feed_targets, std::map* fetch_targets, - bool create_vars = true, + bool create_local_scope = true, bool create_vars = true, const std::string& feed_holder_name = "feed", const std::string& fetch_holder_name = "fetch"); @@ -76,6 +76,7 @@ class Executor { void RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope, std::map* feed_targets, std::map* fetch_targets, + bool create_local_scope = true, bool create_vars = true, const std::string& feed_holder_name = "feed", const std::string& fetch_holder_name = "fetch"); diff --git a/paddle/fluid/framework/framework.proto b/paddle/fluid/framework/framework.proto index 96f53dc1bc8747e1b8ea84166614f98ff363ae5e..d35125fe8c3c8018c38650dc87b2b1474ded6058 100644 --- a/paddle/fluid/framework/framework.proto +++ b/paddle/fluid/framework/framework.proto @@ -101,6 +101,9 @@ message VarType { FP16 = 4; FP32 = 5; FP64 = 6; + // Tensor is used in C++. + SIZE_T = 19; + UINT8 = 20; // Other types that may need additional descriptions LOD_TENSOR = 7; diff --git a/paddle/fluid/framework/init.cc b/paddle/fluid/framework/init.cc index b30f276b4b7c61fda1b40273ce6ccfa19738da41..85beae775b96c3b7e08a2795bcd0ec79b24faeb4 100644 --- a/paddle/fluid/framework/init.cc +++ b/paddle/fluid/framework/init.cc @@ -15,7 +15,6 @@ limitations under the License. */ #include #include #include -#include #include "paddle/fluid/framework/init.h" #include "paddle/fluid/framework/operator.h" @@ -31,6 +30,7 @@ std::once_flag p2p_init_flag; void InitGflags(std::vector argv) { std::call_once(gflags_init_flag, [&]() { + argv.insert(argv.begin(), "dummy"); int argc = argv.size(); char **arr = new char *[argv.size()]; std::string line; @@ -44,20 +44,23 @@ void InitGflags(std::vector argv) { }); } -void InitP2P(int count) { +void InitP2P(std::vector devices) { #ifdef PADDLE_WITH_CUDA std::call_once(p2p_init_flag, [&]() { + int count = devices.size(); for (int i = 0; i < count; ++i) { for (int j = 0; j < count; ++j) { - if (i == j) continue; + if (devices[i] == devices[j]) continue; int can_acess = -1; - PADDLE_ENFORCE(cudaDeviceCanAccessPeer(&can_acess, i, j), - "Failed to test P2P access."); + PADDLE_ENFORCE( + cudaDeviceCanAccessPeer(&can_acess, devices[i], devices[j]), + "Failed to test P2P access."); if (can_acess != 1) { - LOG(WARNING) << "Cannot enable P2P access from " << i << " to " << j; + LOG(WARNING) << "Cannot enable P2P access from " << devices[i] + << " to " << devices[j]; } else { - cudaSetDevice(i); - cudaDeviceEnablePeerAccess(j, 0); + cudaSetDevice(devices[i]); + cudaDeviceEnablePeerAccess(devices[j], 0); } } } @@ -67,11 +70,26 @@ void InitP2P(int count) { void InitDevices(bool init_p2p) { /*Init all available devices by default */ + std::vector devices; +#ifdef PADDLE_WITH_CUDA + try { + int count = platform::GetCUDADeviceCount(); + for (int i = 0; i < count; ++i) { + devices.push_back(i); + } + } catch (const std::exception &exp) { + LOG(WARNING) << "Compiled with WITH_GPU, but no GPU found in runtime."; + } +#else + LOG(WARNING) + << "'CUDA' is not supported, Please re-compile with WITH_GPU option"; +#endif + InitDevices(init_p2p, devices); +} +void InitDevices(bool init_p2p, const std::vector devices) { std::vector places; - places.emplace_back(platform::CPUPlace()); int count = 0; - #ifdef PADDLE_WITH_CUDA try { count = platform::GetCUDADeviceCount(); @@ -83,12 +101,17 @@ void InitDevices(bool init_p2p) { << "'CUDA' is not supported, Please re-compile with WITH_GPU option"; #endif - for (int i = 0; i < count; ++i) { - places.emplace_back(platform::CUDAPlace(i)); + for (size_t i = 0; i < devices.size(); ++i) { + if (devices[i] >= count || devices[i] < 0) { + LOG(WARNING) << "Invalid devices id."; + continue; + } + places.emplace_back(platform::CUDAPlace(devices[i])); } if (init_p2p) { - InitP2P(count); + InitP2P(devices); } + places.emplace_back(platform::CPUPlace()); platform::DeviceContextPool::Init(places); } diff --git a/paddle/fluid/framework/init.h b/paddle/fluid/framework/init.h index 1155ca36049dc66e7ee40e8eca87285d7a728299..0e30594672927253cc8083dcb88bb867d63ec729 100644 --- a/paddle/fluid/framework/init.h +++ b/paddle/fluid/framework/init.h @@ -28,5 +28,7 @@ void InitGLOG(const std::string &prog_name); void InitDevices(bool init_p2p); +void InitDevices(bool init_p2p, const std::vector devices); + } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/lod_tensor_test.cc b/paddle/fluid/framework/lod_tensor_test.cc index 97ab98f09b1a902a942d9667bc7716a28b98d54c..2ceffc93319359683e87e7fec2d18784c9bf02f3 100644 --- a/paddle/fluid/framework/lod_tensor_test.cc +++ b/paddle/fluid/framework/lod_tensor_test.cc @@ -228,11 +228,12 @@ TEST(LoD, CheckAbsLoD) { ASSERT_FALSE(CheckAbsLoD(abs_lod0)); } -TEST(LoDTensor, RecordIO) { +template +static void TestRecordIO() { LoDTensor tensor; - int* tmp = tensor.mutable_data(make_ddim({4, 5}), platform::CPUPlace()); + T* tmp = tensor.mutable_data(make_ddim({4, 5}), platform::CPUPlace()); for (int i = 0; i < 20; ++i) { - tmp[i] = i; + tmp[i] = static_cast(i); } std::stringstream* stream = new std::stringstream(); @@ -247,7 +248,7 @@ TEST(LoDTensor, RecordIO) { auto assert_tensor_ok = [](const LoDTensor& tensor) { for (int i = 0; i < 20; ++i) { - ASSERT_EQ(tensor.data()[i], i); + ASSERT_EQ(tensor.data()[i], static_cast(i)); } }; @@ -255,15 +256,23 @@ TEST(LoDTensor, RecordIO) { std::unique_ptr stream_ptr(stream); recordio::Scanner scanner(std::move(stream_ptr)); auto tensors = ReadFromRecordIO(&scanner, ctx); - ASSERT_EQ(tensors.size(), 2); + ASSERT_EQ(tensors.size(), static_cast(2)); assert_tensor_ok(tensors[0]); assert_tensor_ok(tensors[1]); tensors = ReadFromRecordIO(&scanner, ctx); - ASSERT_EQ(tensors.size(), 2); + ASSERT_EQ(tensors.size(), static_cast(2)); assert_tensor_ok(tensors[0]); assert_tensor_ok(tensors[1]); } } +TEST(LoDTensor, RecordIO) { + TestRecordIO(); + TestRecordIO(); + TestRecordIO(); + TestRecordIO(); + TestRecordIO(); +} + } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/op_desc.cc b/paddle/fluid/framework/op_desc.cc index 076c45713015797f86a3611dd333132bae40044d..09b67e5a1741c68c5f5487340e8fc86ff31e00a4 100644 --- a/paddle/fluid/framework/op_desc.cc +++ b/paddle/fluid/framework/op_desc.cc @@ -20,6 +20,7 @@ limitations under the License. */ #include #include "glog/logging.h" #include "paddle/fluid/framework/block_desc.h" +#include "paddle/fluid/framework/op_proto_maker.h" #include "paddle/fluid/framework/operator.h" #include "paddle/fluid/framework/program_desc.h" #include "paddle/fluid/framework/shape_inference.h" @@ -222,6 +223,15 @@ Attribute OpDesc::GetAttr(const std::string &name) const { return it->second; } +Attribute OpDesc::GetNullableAttr(const std::string &name) const { + auto it = attrs_.find(name); + if (it != attrs_.end()) { + return it->second; + } else { + return Attribute(); + } +} + int OpDesc::GetBlockAttr(const std::string &name) const { auto it = attrs_.find(name); PADDLE_ENFORCE(it != attrs_.end(), "Attribute %s is not found", name); @@ -233,13 +243,8 @@ const std::unordered_map &OpDesc::GetAttrMap() const { } void OpDesc::Rename(const std::string &old_name, const std::string &new_name) { - for (auto &input : inputs_) { - std::replace(input.second.begin(), input.second.end(), old_name, new_name); - } - for (auto &output : outputs_) { - std::replace(output.second.begin(), output.second.end(), old_name, - new_name); - } + RenameInput(old_name, new_name); + RenameOutput(old_name, new_name); need_update_ = true; } @@ -249,6 +254,13 @@ void OpDesc::RenameOutput(const std::string &old_name, std::replace(output.second.begin(), output.second.end(), old_name, new_name); } + + auto it = attrs_.find(framework::OpProtoAndCheckerMaker::OpRoleVarAttrName()); + if (it != attrs_.end()) { + auto &op_vars = boost::get>(it->second); + std::replace(op_vars.begin(), op_vars.end(), old_name, new_name); + } + need_update_ = true; } @@ -257,6 +269,13 @@ void OpDesc::RenameInput(const std::string &old_name, for (auto &input : inputs_) { std::replace(input.second.begin(), input.second.end(), old_name, new_name); } + + auto it = attrs_.find(framework::OpProtoAndCheckerMaker::OpRoleVarAttrName()); + if (it != attrs_.end()) { + auto &op_vars = boost::get>(it->second); + std::replace(op_vars.begin(), op_vars.end(), old_name, new_name); + } + need_update_ = true; } diff --git a/paddle/fluid/framework/op_desc.h b/paddle/fluid/framework/op_desc.h index 3ee36a47c156da67a9ff70852665fbbd464bea17..1a330db7cc5555a939950043ac90a321573b292d 100644 --- a/paddle/fluid/framework/op_desc.h +++ b/paddle/fluid/framework/op_desc.h @@ -78,6 +78,8 @@ class OpDesc { Attribute GetAttr(const std::string &name) const; + Attribute GetNullableAttr(const std::string &name) const; + int GetBlockAttr(const std::string &name) const; void Rename(const std::string &old_name, const std::string &new_name); diff --git a/paddle/fluid/framework/op_kernel_type_test.cc b/paddle/fluid/framework/op_kernel_type_test.cc index d37ce149ce3df63692b41289bb03448d54e392f5..db95861c510b52a5b52229541434e6437d3fb9f4 100644 --- a/paddle/fluid/framework/op_kernel_type_test.cc +++ b/paddle/fluid/framework/op_kernel_type_test.cc @@ -27,7 +27,7 @@ TEST(OpKernelType, ToString) { LibraryType::kCUDNN); ASSERT_EQ(paddle::framework::KernelTypeToString(op_kernel_type), - "data_type[float32]:data_layout[NCHW]:place[CPUPlace]:library_type[" + "data_type[float]:data_layout[NCHW]:place[CPUPlace]:library_type[" "CUDNN]"); } diff --git a/paddle/fluid/framework/op_proto_maker.cc b/paddle/fluid/framework/op_proto_maker.cc index c479d7617cfa34cd381d84d15d5e214d57af52d0..ae9f4efd44acdcdff2806deea6826e4089459a78 100644 --- a/paddle/fluid/framework/op_proto_maker.cc +++ b/paddle/fluid/framework/op_proto_maker.cc @@ -13,6 +13,7 @@ limitations under the License. */ #include "paddle/fluid/framework/op_proto_maker.h" #include +#include namespace paddle { namespace framework { @@ -55,5 +56,28 @@ void OpProtoAndCheckerMaker::CheckNoDuplicatedInOutAttrs() { } } +void OpProtoAndCheckerMaker::operator()(proto::OpProto* proto, + OpAttrChecker* attr_checker) { + proto_ = proto; + op_checker_ = attr_checker; + Make(); + + AddAttr(OpRoleAttrName(), "The role of this operator") + .InEnum( + {static_cast(OpRole::kForward), + static_cast(OpRole::kBackward), + static_cast(OpRole::kOptimize), static_cast(OpRole::kRPC), + static_cast(OpRole::kLoss) | static_cast(OpRole::kForward), + static_cast(OpRole::kLoss) | + static_cast(OpRole::kBackward), + static_cast(OpRole::kNotSpecified)}) + .SetDefault(static_cast(OpRole::kNotSpecified)); + AddAttr>(OpRoleVarAttrName(), + "Optimized for variable") + .SetDefault({}); + + Validate(); +} + } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/op_proto_maker.h b/paddle/fluid/framework/op_proto_maker.h index 0beb57ce1609d2e90c05d3255647bd321bc1f6a9..8493b9d8b326c71a33b95bf95e5fc1743c686eb7 100644 --- a/paddle/fluid/framework/op_proto_maker.h +++ b/paddle/fluid/framework/op_proto_maker.h @@ -14,56 +14,68 @@ limitations under the License. */ #pragma once #include +#include "glog/logging.h" #include "paddle/fluid/framework/attribute.h" #include "paddle/fluid/framework/framework.pb.h" - namespace paddle { namespace framework { +enum class OpRole { + kForward = 0x0000, + kBackward = 0x0001, + kOptimize = 0x0002, + kRPC = 0x0003, + + kLoss = 0x0100, + // The default value of op's role. This should be only used for unittests and + // CreateOp inside a operator. + kNotSpecified = 0x1000, +}; + // this class not only make proto but also init attribute checkers. class OpProtoAndCheckerMaker { public: - using OpProto = proto::OpProto; - using OpAttrChecker = framework::OpAttrChecker; - OpProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker) - : proto_(proto), op_checker_(op_checker) {} + static const char *OpRoleAttrName() { return "op_role"; } + static const char *OpRoleVarAttrName() { return "op_role_var"; } + + void operator()(proto::OpProto *proto, OpAttrChecker *attr_checker); + + virtual void Make() = 0; virtual ~OpProtoAndCheckerMaker() { - PADDLE_ENFORCE(validated_, "should call Validate after build"); + CHECK(validated_) << "should call Validate after build"; } - void Validate(); - protected: struct VariableBuilder { - OpProto::Var* var_; + proto::OpProto::Var *var_; - VariableBuilder& AsDuplicable() { + VariableBuilder &AsDuplicable() { var_->set_duplicable(true); return *this; } - VariableBuilder& AsIntermediate() { + VariableBuilder &AsIntermediate() { var_->set_intermediate(true); return *this; } - VariableBuilder& AsDispensable() { + VariableBuilder &AsDispensable() { var_->set_dispensable(true); return *this; } }; - VariableBuilder AddInput(const std::string& name, const std::string& comment); + VariableBuilder AddInput(const std::string &name, const std::string &comment); - VariableBuilder AddOutput(const std::string& name, - const std::string& comment); + VariableBuilder AddOutput(const std::string &name, + const std::string &comment); template - TypedAttrChecker& AddAttr(const std::string& name, - const std::string& comment, + TypedAttrChecker &AddAttr(const std::string &name, + const std::string &comment, bool generated = false) { - auto* attr = proto_->add_attrs(); + auto *attr = proto_->add_attrs(); attr->set_name(name); attr->set_comment(comment); attr->set_generated(generated); @@ -71,21 +83,15 @@ class OpProtoAndCheckerMaker { return op_checker_->AddAttrChecker(name); } - void AddComment(const std::string& comment) { proto_->set_comment(comment); } + void AddComment(const std::string &comment) { proto_->set_comment(comment); } private: void CheckNoDuplicatedInOutAttrs(); + void Validate(); - OpProto* proto_; - OpAttrChecker* op_checker_; + proto::OpProto *proto_; + OpAttrChecker *op_checker_; bool validated_{false}; }; - -class NOPMaker : public OpProtoAndCheckerMaker { - public: - NOPMaker(OpProto* proto, framework::OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) {} -}; - } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/op_proto_maker_test.cc b/paddle/fluid/framework/op_proto_maker_test.cc index a8d8c6386af940d4a14016b30de344e1c7877b22..a8030d377fdb4d4aef74b315e21792dad10fac96 100644 --- a/paddle/fluid/framework/op_proto_maker_test.cc +++ b/paddle/fluid/framework/op_proto_maker_test.cc @@ -18,9 +18,7 @@ limitations under the License. */ class TestAttrProtoMaker : public paddle::framework::OpProtoAndCheckerMaker { public: - TestAttrProtoMaker(paddle::framework::proto::OpProto* proto, - paddle::framework::OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() { AddAttr("scale", "scale of test op"); AddAttr("scale", "scale of test op"); } @@ -29,15 +27,14 @@ class TestAttrProtoMaker : public paddle::framework::OpProtoAndCheckerMaker { TEST(ProtoMaker, DuplicatedAttr) { paddle::framework::proto::OpProto op_proto; paddle::framework::OpAttrChecker op_checker; - auto proto_maker = TestAttrProtoMaker(&op_proto, &op_checker); - ASSERT_THROW(proto_maker.Validate(), paddle::platform::EnforceNotMet); + TestAttrProtoMaker proto_maker; + ASSERT_THROW(proto_maker(&op_proto, &op_checker), + paddle::platform::EnforceNotMet); } class TestInOutProtoMaker : public paddle::framework::OpProtoAndCheckerMaker { public: - TestInOutProtoMaker(paddle::framework::proto::OpProto* proto, - paddle::framework::OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() { AddInput("input", "input of test op"); AddInput("input", "input of test op"); } @@ -46,6 +43,7 @@ class TestInOutProtoMaker : public paddle::framework::OpProtoAndCheckerMaker { TEST(ProtoMaker, DuplicatedInOut) { paddle::framework::proto::OpProto op_proto; paddle::framework::OpAttrChecker op_checker; - auto proto_maker = TestInOutProtoMaker(&op_proto, &op_checker); - ASSERT_THROW(proto_maker.Validate(), paddle::platform::EnforceNotMet); + TestAttrProtoMaker proto_maker; + ASSERT_THROW(proto_maker(&op_proto, &op_checker), + paddle::platform::EnforceNotMet); } diff --git a/paddle/fluid/framework/op_registry_test.cc b/paddle/fluid/framework/op_registry_test.cc index 6dc4cf261bad3c004aa53fba5502fe166e3a47f7..18b1649cc71d5edd5b07740bbad1fe8f81128898 100644 --- a/paddle/fluid/framework/op_registry_test.cc +++ b/paddle/fluid/framework/op_registry_test.cc @@ -33,8 +33,7 @@ class CosineOp : public OperatorBase { class CosineOpProtoAndCheckerMaker : public OpProtoAndCheckerMaker { public: - CosineOpProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() { AddInput("input", "input of cosine op"); AddOutput("output", "output of cosine op"); AddAttr("scale", "scale of cosine op") @@ -55,8 +54,7 @@ class MyTestOp : public OperatorBase { class MyTestOpProtoAndCheckerMaker : public OpProtoAndCheckerMaker { public: - MyTestOpProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() { AddInput("input", "input of cosine op").AsDuplicable(); AddOutput("output", "output of cosine op").AsIntermediate(); auto my_checker = [](int i) { @@ -212,10 +210,7 @@ namespace framework { class OpKernelTestMaker : public OpProtoAndCheckerMaker { public: - OpKernelTestMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { - AddComment("NoGradOp, same input output. no Grad"); - } + void Make() { AddComment("NoGradOp, same input output. no Grad"); } }; class OpWithKernelTest : public OperatorWithKernel { @@ -275,9 +270,9 @@ TEST(OperatorRegistrar, CUDA) { static int op_test_value = 0; -using paddle::platform::DeviceContext; using paddle::platform::CPUDeviceContext; using paddle::platform::CUDADeviceContext; +using paddle::platform::DeviceContext; namespace paddle { namespace framework { diff --git a/paddle/fluid/framework/operator.cc b/paddle/fluid/framework/operator.cc index 32576423a62a1a12f085d565e7ff267145bf979c..f87d5521492418d2daf5b7fba1500c4bb31e10f5 100644 --- a/paddle/fluid/framework/operator.cc +++ b/paddle/fluid/framework/operator.cc @@ -24,6 +24,9 @@ limitations under the License. */ #include "paddle/fluid/platform/profiler.h" DECLARE_bool(benchmark); +DEFINE_bool(check_nan_inf, false, + "Checking whether operator produce NAN/INF or not. It will be " + "extremely slow so please use this flag wisely."); namespace paddle { namespace framework { @@ -93,6 +96,14 @@ void OperatorBase::Run(const Scope& scope, const platform::Place& place) { RunImpl(scope, place); } +bool OperatorBase::HasInputs(const std::string& name) const { + if (inputs_.find(name) != inputs_.end()) { + return true; + } else { + return false; + } +} + std::string OperatorBase::Input(const std::string& name) const { auto& ins = Inputs(name); PADDLE_ENFORCE_LE(ins.size(), 1UL, @@ -109,6 +120,14 @@ const std::vector& OperatorBase::Inputs( return it->second; } +bool OperatorBase::HasOutputs(const std::string& name) const { + if (outputs_.find(name) != outputs_.end()) { + return true; + } else { + return false; + } +} + std::string OperatorBase::Output(const std::string& name) const { auto& outs = Outputs(name); PADDLE_ENFORCE_LE(outs.size(), 1UL, @@ -220,13 +239,18 @@ void OperatorBase::CheckAllInputOutputSet() const { if (op_info == nullptr || op_info->proto_ == nullptr) return; for (auto& in : op_info->Proto().inputs()) { - PADDLE_ENFORCE(inputs_.find(in.name()) != inputs_.end(), - "Type %s's input %s is not set", Type(), in.name()); + if (!in.dispensable()) { + PADDLE_ENFORCE(inputs_.find(in.name()) != inputs_.end(), + "Operator %s's input, %s, is not set", Type(), in.name()); + } } for (auto& out : op_info->Proto().outputs()) { - PADDLE_ENFORCE(outputs_.find(out.name()) != outputs_.end(), - "Type %s's output %s is not set", Type(), out.name()); + if (!out.dispensable()) { + PADDLE_ENFORCE(outputs_.find(out.name()) != outputs_.end(), + "Operator %s's output, %s, is not set", Type(), + out.name()); + } } } @@ -332,6 +356,9 @@ class RuntimeInferShapeContext : public InferShapeContext { : op_(op), scope_(scope) {} bool HasInput(const std::string& name) const override { + if (!op_.HasInputs(name)) { + return false; + } auto& ins = Inputs(name); size_t length = ins.size(); if (length == 0) { @@ -345,6 +372,9 @@ class RuntimeInferShapeContext : public InferShapeContext { } bool HasOutput(const std::string& name) const override { + if (!op_.HasOutputs(name)) { + return false; + } auto& outs = Outputs(name); size_t length = outs.size(); if (length == 0) { @@ -358,6 +388,9 @@ class RuntimeInferShapeContext : public InferShapeContext { } bool HasInputs(const std::string& name) const override { + if (!op_.HasInputs(name)) { + return false; + } auto inputs = op_.Inputs(name); if (inputs.empty()) { return false; @@ -371,6 +404,9 @@ class RuntimeInferShapeContext : public InferShapeContext { } bool HasOutputs(const std::string& name) const override { + if (!op_.HasOutputs(name)) { + return false; + } auto outputs = op_.Outputs(name); if (outputs.empty()) { return false; @@ -433,6 +469,7 @@ class RuntimeInferShapeContext : public InferShapeContext { protected: DDim GetDim(const std::string& name) const override { Variable* var = scope_.FindVar(name); + PADDLE_ENFORCE_NOT_NULL(var); if (var->IsType()) { return var->Get().dims(); } else if (var->IsType()) { @@ -480,6 +517,21 @@ class RuntimeInferShapeContext : public InferShapeContext { const Scope& scope_; }; +static void CheckTensorNANOrInf(const std::string& name, + const framework::Tensor& tensor) { + if (tensor.memory_size() == 0) { + return; + } + if (tensor.type().hash_code() != typeid(float).hash_code() && // NOLINT + tensor.type().hash_code() != typeid(double).hash_code()) { // NOLINT + return; + } + PADDLE_ENFORCE(!framework::TensorContainsInf(tensor), + "Tensor %s contains Inf", name); + PADDLE_ENFORCE(!framework::TensorContainsNAN(tensor), + "Tensor %s contains NAN", name); +} + void OperatorWithKernel::RunImpl(const Scope& scope, const platform::Place& place) const { RuntimeInferShapeContext infer_shape_ctx(*this, scope); @@ -564,6 +616,16 @@ void OperatorWithKernel::RunImpl(const Scope& scope, if (FLAGS_benchmark) { new_dev_ctx->Wait(); } + + if (FLAGS_check_nan_inf) { + for (auto& vname : OutputVars(true)) { + auto* var = new_scope.FindVar(vname); + if (var == nullptr) continue; + if (var->IsType()) { + CheckTensorNANOrInf(vname, var->Get()); + } + } + } } proto::VarType::Type OperatorWithKernel::IndicateDataType( diff --git a/paddle/fluid/framework/operator.h b/paddle/fluid/framework/operator.h index 826cc57b725ab4b52e5d67ab82e939cbd62a8460..2f480e00c100d579e100de17d3feb957f5ef6167 100644 --- a/paddle/fluid/framework/operator.h +++ b/paddle/fluid/framework/operator.h @@ -33,7 +33,6 @@ limitations under the License. */ #include "paddle/fluid/framework/tensor.h" #include "paddle/fluid/platform/device_context.h" #include "paddle/fluid/platform/variant.h" -#include "paddle/utils/Error.h" namespace paddle { namespace framework { @@ -105,6 +104,7 @@ class OperatorBase { const VariableNameMap& Inputs() const { return inputs_; } const VariableNameMap& Outputs() const { return outputs_; } + bool HasInputs(const std::string& name) const; //! Get a input with argument's name described in `op_proto` std::string Input(const std::string& name) const; //! Get a input which has multiple variables. @@ -112,6 +112,7 @@ class OperatorBase { //! Get all inputs variable names std::vector InputVars() const; + bool HasOutputs(const std::string& name) const; //! Get a output with argument's name described in `op_proto` std::string Output(const std::string& name) const; //! Get an output which has multiple variables. @@ -190,6 +191,10 @@ class ExecutionContext { return op_.Attr(name); } + bool HasInput(const std::string& name) const { return op_.HasInputs(name); } + + bool HasOutput(const std::string& name) const { return op_.HasOutputs(name); } + size_t InputSize(const std::string& name) const { return op_.Inputs(name).size(); } diff --git a/paddle/fluid/framework/operator_test.cc b/paddle/fluid/framework/operator_test.cc index 1bf8c81469bb4afdd00921cfa0acf6089dedbbaa..74043b5d7990178976baf2fad991ae03f9c8dd25 100644 --- a/paddle/fluid/framework/operator_test.cc +++ b/paddle/fluid/framework/operator_test.cc @@ -46,8 +46,7 @@ class OpWithoutKernelTest : public OperatorBase { class OpWithoutKernelCheckerMaker : public OpProtoAndCheckerMaker { public: - OpWithoutKernelCheckerMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() { AddInput("input", "input of test op"); AddOutput("output", "output of test op"); AddAttr("scale", "scale of cosine op"); @@ -98,8 +97,7 @@ namespace framework { class OpKernelTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker { public: - OpKernelTestProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() { AddInput("x", "input of test op"); AddOutput("y", "output of test op"); AddAttr("scale", "scale of cosine op") @@ -137,9 +135,7 @@ class CPUKernelTest : public OpKernel { class OpKernelTestMultiInputsProtoAndCheckerMaker : public OpProtoAndCheckerMaker { public: - OpKernelTestMultiInputsProtoAndCheckerMaker(OpProto* proto, - OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() { AddInput("xs", "inputs of test op").AsDuplicable(); AddInput("k", "input of test op"); AddOutput("ys", "outputs of test op").AsDuplicable(); diff --git a/paddle/fluid/framework/parallel_executor.cc b/paddle/fluid/framework/parallel_executor.cc index de644e851999920251c762a75c050e8182e950c6..50c3468d556bfe05d6c41906cf35cb671f711b1e 100644 --- a/paddle/fluid/framework/parallel_executor.cc +++ b/paddle/fluid/framework/parallel_executor.cc @@ -52,13 +52,13 @@ std::vector &ParallelExecutor::GetLocalScopes() { } ParallelExecutor::ParallelExecutor( - size_t num_threads, bool use_event, const std::vector &places, const std::unordered_set ¶ms, const std::unordered_set &bcast_vars, const ProgramDesc &main_program, const std::string &loss_var_name, - Scope *scope, const std::vector &local_scopes, bool allow_op_delay, - bool customize_scale_loss) + Scope *scope, const std::vector &local_scopes, + const ExecutionStrategy &exec_strategy, const BuildStrategy &build_strategy, + size_t num_trainers, size_t trainer_id) : member_(new ParallelExecutorPrivate(places)) { member_->global_scope_ = scope; @@ -80,7 +80,13 @@ ParallelExecutor::ParallelExecutor( // Bcast Parameters to all GPUs #ifdef PADDLE_WITH_CUDA - member_->nccl_ctxs_.reset(new platform::NCCLContextMap(member_->places_)); + auto *nccl_id_var = scope->FindVar(NCCL_ID_VARNAME); + ncclUniqueId *nccl_id = nullptr; + if (nccl_id_var != nullptr) { + nccl_id = nccl_id_var->GetMutable(); + } + member_->nccl_ctxs_.reset(new platform::NCCLContextMap( + member_->places_, nccl_id, num_trainers, trainer_id)); #endif if (platform::is_gpu_place(places[0]) && member_->local_scopes_.size() != 1 && local_scopes.empty()) { // Is CUDA @@ -93,17 +99,16 @@ ParallelExecutor::ParallelExecutor( #ifdef PADDLE_WITH_CUDA details::MultiDevSSAGraphBuilder builder( member_->places_, loss_var_name, params, member_->local_scopes_, - customize_scale_loss, member_->nccl_ctxs_.get()); + member_->nccl_ctxs_.get(), build_strategy); #else details::MultiDevSSAGraphBuilder builder(member_->places_, loss_var_name, params, member_->local_scopes_, - customize_scale_loss); + build_strategy); #endif auto graph = builder.Build(main_program); member_->executor_.reset(new details::ThreadedSSAGraphExecutor( - num_threads, use_event, member_->local_scopes_, places, std::move(graph), - allow_op_delay)); + exec_strategy, member_->local_scopes_, places, std::move(graph))); // Step 3. Create vars in each scope; for (auto *var : main_program.Block(0).AllVars()) { diff --git a/paddle/fluid/framework/parallel_executor.h b/paddle/fluid/framework/parallel_executor.h index 49da123d98181c3d3abcdd64d14c5583142eba58..5247e790649e76567f4527d54499d6e95dac5c27 100644 --- a/paddle/fluid/framework/parallel_executor.h +++ b/paddle/fluid/framework/parallel_executor.h @@ -14,55 +14,60 @@ limitations under the License. */ #pragma once +#include #include #include #include +#include "paddle/fluid/framework/details/execution_strategy.h" #include "paddle/fluid/framework/executor.h" #include "paddle/fluid/framework/op_info.h" #include "paddle/fluid/framework/program_desc.h" #include "paddle/fluid/framework/scope.h" #include "paddle/fluid/framework/tensor.h" #include "paddle/fluid/platform/device_context.h" - namespace paddle { namespace framework { class ParallelExecutorPrivate; +using details::BuildStrategy; +using details::ExecutionStrategy; + class ParallelExecutor { DISABLE_COPY_AND_ASSIGN(ParallelExecutor); public: - explicit ParallelExecutor(size_t num_threads, bool use_event, - const std::vector& places, - const std::unordered_set& params, - const std::unordered_set& bcast_vars, - const ProgramDesc& main_program, - const std::string& loss_var_name, Scope* scope, - const std::vector& local_scopes, - bool allow_op_delay, bool customize_scale_loss); + explicit ParallelExecutor(const std::vector &places, + const std::unordered_set ¶ms, + const std::unordered_set &bcast_vars, + const ProgramDesc &main_program, + const std::string &loss_var_name, Scope *scope, + const std::vector &local_scopes, + const ExecutionStrategy &exec_strategy, + const BuildStrategy &build_strategy, + size_t num_trainers = 1, size_t trainer_id = 0); ~ParallelExecutor(); - std::vector& GetLocalScopes(); + std::vector &GetLocalScopes(); /** * Feed tensors to local scopes. The size of tensors should be equal to the * size of local scopes. */ void FeedTensorsIntoLocalScopes( - const std::vector>& tensors); + const std::vector> &tensors); void FeedAndSplitTensorIntoLocalScopes( - const std::unordered_map& tensors); + const std::unordered_map &tensors); - void Run(const std::vector& fetch_tensors, - const std::string& fetched_var_name); + void Run(const std::vector &fetch_tensors, + const std::string &fetched_var_name); - void BCastParamsToGPUs(const std::unordered_set& vars) const; + void BCastParamsToGPUs(const std::unordered_set &vars) const; private: - ParallelExecutorPrivate* member_; + ParallelExecutorPrivate *member_; }; } // namespace framework diff --git a/paddle/fluid/framework/reader.cc b/paddle/fluid/framework/reader.cc index 76126f3dc64d71770d13f9d66bb30f176c112629..0b36f1116d15004b355e854e101abb9ad3297836 100644 --- a/paddle/fluid/framework/reader.cc +++ b/paddle/fluid/framework/reader.cc @@ -25,8 +25,10 @@ void FileReader::ReadNext(std::vector *out) { if (out->empty()) { return; } + + PADDLE_ENFORCE_EQ(out->size(), dims_.size()); for (size_t i = 0; i < dims_.size(); ++i) { - auto &actual = out->at(i).dims(); + auto &actual = (*out)[i].dims(); auto &expect = dims_[i]; PADDLE_ENFORCE_EQ(actual.size(), expect.size()); diff --git a/paddle/fluid/framework/selected_rows.cc b/paddle/fluid/framework/selected_rows.cc index 794e7f743413b068119afd5df232bfc2bb91a8c7..06ed87e7e8a2d5324b48a466b05207042ec1b7fa 100644 --- a/paddle/fluid/framework/selected_rows.cc +++ b/paddle/fluid/framework/selected_rows.cc @@ -18,8 +18,8 @@ namespace paddle { namespace framework { struct ReAllocateVisitor { - ReAllocateVisitor(framework::Tensor* tensor, const framework::DDim& dims) - : tensor_(tensor), dims_(dims) {} + ReAllocateVisitor(const framework::DDim& dims, framework::Tensor* tensor) + : dims_(dims), tensor_(tensor) {} template void operator()() const { @@ -34,8 +34,8 @@ struct ReAllocateVisitor { tensor_->ShareDataWith(cpu_tensor); } - framework::Tensor* tensor_; framework::DDim dims_; + framework::Tensor* tensor_; }; struct TensorCopyVisitor { @@ -120,28 +120,33 @@ bool SelectedRows::HasKey(int64_t key) const { : true; } -std::vector SelectedRows::Get(std::vector keys, - framework::Tensor* value) const { +std::vector> SelectedRows::Get( + const std::vector& keys, framework::Tensor* value) const { PADDLE_ENFORCE(value->IsInitialized(), "The value tensor should be initialized."); - std::vector non_keys; - int64_t value_width = value_->numel() / value_->dims()[0]; - PADDLE_ENFORCE_EQ(value_width, value->numel() / value->dims()[0], - "output tensor should have the same shape with table " - "execpt the dims[0]."); - - for (size_t i = 0; i < keys.size(); ++i) { - int64_t index = Index(keys[i]); - if (index == -1) { - non_keys.push_back(keys[i]); - } else { - framework::VisitDataType( - framework::ToDataType(value_->type()), - TensorCopyVisitor(value, i * value_width, *value_.get(), - index * value_width, value_width)); + std::vector> non_keys_pair; + if (keys.empty()) { + VLOG(3) << "keys is empty, please check data!"; + } else { + int64_t value_width = value_->numel() / value_->dims()[0]; + PADDLE_ENFORCE_EQ(value_width, value->numel() / value->dims()[0], + "output tensor should have the same shape with table " + "except the dims[0]."); + + for (size_t i = 0; i < keys.size(); ++i) { + int64_t index = Index(keys[i]); + if (index == -1) { + non_keys_pair.push_back( + std::make_pair(keys[i], static_cast(i))); + } else { + framework::VisitDataType( + framework::ToDataType(value_->type()), + TensorCopyVisitor(value, i * value_width, *value_.get(), + index * value_width, value_width)); + } } } - return non_keys; + return non_keys_pair; } bool SelectedRows::Set(int64_t key, const framework::Tensor& value) { @@ -153,6 +158,7 @@ bool SelectedRows::Set(int64_t key, const framework::Tensor& value) { } PADDLE_ENFORCE_EQ(value.dims()[0], static_cast(1), "The first dim of value should be 1."); + std::lock_guard lock(*auto_grown_mutex_.get()); auto index = Index(key); bool is_new_key = false; if (index == -1) { @@ -164,7 +170,7 @@ bool SelectedRows::Set(int64_t key, const framework::Tensor& value) { auto dims = value_->dims(); dims[0] = (dims[0] + 1) << 1; framework::VisitDataType(framework::ToDataType(value.type()), - ReAllocateVisitor(value_.get(), dims)); + ReAllocateVisitor(dims, value_.get())); } } diff --git a/paddle/fluid/framework/selected_rows.h b/paddle/fluid/framework/selected_rows.h index d6c9507b1681855e759a4b1b9d3dddf6fcb2fc13..7160670ddd204c20021ea87cdd67ee4721d03451 100644 --- a/paddle/fluid/framework/selected_rows.h +++ b/paddle/fluid/framework/selected_rows.h @@ -15,6 +15,9 @@ limitations under the License. */ #pragma once #include +#include +#include // NOLINT +#include #include #include "paddle/fluid/framework/lod_tensor.h" @@ -45,11 +48,13 @@ class SelectedRows { SelectedRows(const std::vector& rows, const int64_t& height) : rows_(rows), height_(height) { value_.reset(new Tensor()); + auto_grown_mutex_.reset(new std::mutex); } SelectedRows() { height_ = 0; value_.reset(new Tensor()); + auto_grown_mutex_.reset(new std::mutex); } platform::Place place() const { return value_->place(); } @@ -78,10 +83,11 @@ class SelectedRows { /* * @brief Get value by the key list, if the * - * @return a list of keys which does not exists in table + * @return a list of pair which contains the non-exists key and the index in + * the value */ - std::vector Get(std::vector keys, - framework::Tensor* tensor) const; + std::vector> Get(const std::vector& keys, + framework::Tensor* value) const; /* * @brief Set a key-value pair into the table. @@ -123,6 +129,7 @@ class SelectedRows { Vector rows_; std::unique_ptr value_{nullptr}; int64_t height_; + std::unique_ptr auto_grown_mutex_{nullptr}; }; /* diff --git a/paddle/fluid/framework/selected_rows_test.cc b/paddle/fluid/framework/selected_rows_test.cc index 39fe6d92940606084c28eec1a4d6486cb58844ce..eefcaa5672c5a3debf162f5c8eda653408dcf221 100644 --- a/paddle/fluid/framework/selected_rows_test.cc +++ b/paddle/fluid/framework/selected_rows_test.cc @@ -59,7 +59,7 @@ TEST_F(SelectedRowsTester, SerializeAndDeseralize) { ASSERT_EQ(selected_rows_->GetCompleteDims(), dst_tensor.GetCompleteDims()); } -TEST_F(SelectedRowsTester, Table) { +TEST_F(SelectedRowsTester, SparseTable) { platform::CPUPlace cpu; SelectedRows table; // initialize a sparse table @@ -87,11 +87,11 @@ TEST_F(SelectedRowsTester, Table) { framework::Tensor get_value; get_value.mutable_data(framework::make_ddim({2, 100}), cpu); std::vector keys({non_key, key}); - auto non_keys = table.Get(keys, &get_value); + auto non_key_pairs = table.Get(keys, &get_value); ASSERT_EQ(get_value.data()[100], static_cast(10)); - ASSERT_EQ(non_keys.size(), static_cast(1)); - ASSERT_EQ(non_keys[0], non_key); + ASSERT_EQ(non_key_pairs.size(), static_cast(1)); + ASSERT_EQ(non_key_pairs[0].first, non_key); } } // namespace framework diff --git a/paddle/fluid/framework/shape_inference.h b/paddle/fluid/framework/shape_inference.h index 46c8feec001584a872f7f62682080e0e72c06f50..5f497cafa0f75f7c23d550ef767d55274de7c900 100644 --- a/paddle/fluid/framework/shape_inference.h +++ b/paddle/fluid/framework/shape_inference.h @@ -63,6 +63,7 @@ class InferShapeContext { std::vector GetInputVarPtrs(const std::string &name); std::vector GetOutputVarPtrs(const std::string &name); + virtual InferShapeVarPtr GetVarPtr(const std::string &name) = 0; // Note: In while op, we need this to be public void SetDims(const std::vector &names, @@ -81,8 +82,6 @@ class InferShapeContext { const std::vector &names) const; virtual proto::VarType::Type GetVarType(const std::string &name) const = 0; - - virtual InferShapeVarPtr GetVarPtr(const std::string &name) = 0; }; } // namespace framework diff --git a/paddle/fluid/framework/tensor_impl.h b/paddle/fluid/framework/tensor_impl.h index f49d1a47a325b2aac6185073203df124be18b54d..2f19ec0f0a9338e2b96d1f64eac45387bae4d1eb 100644 --- a/paddle/fluid/framework/tensor_impl.h +++ b/paddle/fluid/framework/tensor_impl.h @@ -13,54 +13,14 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once +#include "paddle/fluid/framework/data_type.h" #include "paddle/fluid/memory/memcpy.h" #include "paddle/fluid/platform/enforce.h" #include "paddle/fluid/platform/float16.h" namespace paddle { namespace framework { - -template -struct SizeOfTypeFunctor; - -template -struct SizeOfTypeFunctor { - size_t operator()(std::type_index type) const { - if (typeid(T).hash_code() == type.hash_code()) { - return sizeof(T); - } else { - return 0UL; - } - } -}; - -template <> -struct SizeOfTypeFunctor<> { - size_t operator()(std::type_index type) const { return 0UL; } -}; - -template -struct SizeOfTypeFunctor { - size_t operator()(std::type_index type) const { - SizeOfTypeFunctor head; - size_t head_size = head(type); - if (head_size != 0) { - return head_size; - } - SizeOfTypeFunctor tail; - return tail(type); - } -}; - -static inline size_t SizeOfType(std::type_index type) { - SizeOfTypeFunctor - functor; - size_t size = functor(type); - PADDLE_ENFORCE(size != 0UL, "Cannot get size of type %s", type.name()); - return size; -} - +extern size_t SizeOfType(std::type_index type); inline void Tensor::check_memory_size() const { PADDLE_ENFORCE_NOT_NULL( holder_, "Tensor holds no memory. Call Tensor::mutable_data first."); @@ -79,7 +39,7 @@ template inline const T* Tensor::data() const { check_memory_size(); PADDLE_ENFORCE(std::is_same::value || - holder_->type().hash_code() == typeid(T).hash_code(), + holder_->type() == std::type_index(typeid(T)), "Tensor holds the wrong type, it holds %s", this->holder_->type().name()); @@ -93,7 +53,7 @@ template inline T* Tensor::data() { check_memory_size(); PADDLE_ENFORCE(std::is_same::value || - holder_->type().hash_code() == typeid(T).hash_code(), + holder_->type() == std::type_index(typeid(T)), "Tensor holds the wrong type, it holds %s", this->holder_->type().name()); return reinterpret_cast(reinterpret_cast(holder_->ptr()) + diff --git a/paddle/fluid/framework/tensor_util.cc b/paddle/fluid/framework/tensor_util.cc index d2e60ab1dd16758a91d22ef6872edc5053ef88b3..e5bc74755f46449296a153e8b330968e6d9f1e1d 100644 --- a/paddle/fluid/framework/tensor_util.cc +++ b/paddle/fluid/framework/tensor_util.cc @@ -20,7 +20,7 @@ namespace paddle { namespace framework { void TensorCopy(const Tensor& src, const platform::Place& dst_place, - const platform::DeviceContext& ctx, Tensor* dst, bool sync) { + const platform::DeviceContext& ctx, Tensor* dst) { VLOG(3) << "TensorCopy " << src.dims() << " from " << src.place() << " to " << dst_place; src.check_memory_size(); @@ -48,9 +48,7 @@ void TensorCopy(const Tensor& src, const platform::Place& dst_place, auto ctx_gpu_place = boost::get(ctx_place); PADDLE_ENFORCE_EQ(src_gpu_place, ctx_gpu_place); auto stream = - sync ? nullptr - : reinterpret_cast(ctx) - .stream(); + reinterpret_cast(ctx).stream(); memory::Copy(dst_cpu_place, dst_ptr, src_gpu_place, src_ptr, size, stream); } else if (platform::is_cpu_place(src_place) && platform::is_gpu_place(dst_place)) { @@ -61,9 +59,7 @@ void TensorCopy(const Tensor& src, const platform::Place& dst_place, auto ctx_gpu_place = boost::get(ctx_place); PADDLE_ENFORCE_EQ(dst_gpu_place, ctx_gpu_place); auto stream = - sync ? nullptr - : reinterpret_cast(ctx) - .stream(); + reinterpret_cast(ctx).stream(); memory::Copy(dst_gpu_place, dst_ptr, src_cpu_place, src_ptr, size, stream); } else if (platform::is_gpu_place(src_place) && platform::is_gpu_place(dst_place)) { @@ -72,9 +68,7 @@ void TensorCopy(const Tensor& src, const platform::Place& dst_place, auto ctx_place = ctx.GetPlace(); PADDLE_ENFORCE(platform::is_gpu_place(ctx_place)); auto stream = - sync ? nullptr - : reinterpret_cast(ctx) - .stream(); + reinterpret_cast(ctx).stream(); memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size, stream); } #endif @@ -92,6 +86,41 @@ void TensorCopy(const Tensor& src, const platform::Place& dst_place, TensorCopy(src, dst_place, *dev_ctx, dst); } +void TensorCopySync(const Tensor& src, const platform::Place& dst_place, + Tensor* dst) { + VLOG(3) << "TensorCopySync " << src.dims() << " from " << src.place() + << " to " << dst_place; + src.check_memory_size(); + dst->Resize(src.dims()); + dst->set_layout(src.layout()); + auto src_place = src.place(); + auto src_ptr = src.data(); + auto dst_ptr = dst->mutable_data(dst_place, src.type()); + auto size = src.numel() * SizeOfType(src.type()); + if (platform::is_cpu_place(src_place) && platform::is_cpu_place(dst_place)) { + memory::Copy(boost::get(dst_place), dst_ptr, + boost::get(src_place), src_ptr, size); + } +#ifdef PADDLE_WITH_CUDA + else if (platform::is_gpu_place(src_place) && // NOLINT + platform::is_cpu_place(dst_place)) { + auto src_gpu_place = boost::get(src_place); + auto dst_cpu_place = boost::get(dst_place); + memory::Copy(dst_cpu_place, dst_ptr, src_gpu_place, src_ptr, size, nullptr); + } else if (platform::is_cpu_place(src_place) && + platform::is_gpu_place(dst_place)) { + auto src_cpu_place = boost::get(src_place); + auto dst_gpu_place = boost::get(dst_place); + memory::Copy(dst_gpu_place, dst_ptr, src_cpu_place, src_ptr, size, nullptr); + } else if (platform::is_gpu_place(src_place) && + platform::is_gpu_place(dst_place)) { + auto src_gpu_place = boost::get(src_place); + auto dst_gpu_place = boost::get(dst_place); + memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size, nullptr); + } +#endif +} + template struct AnyDTypeVisitor { Predicate predicate_; diff --git a/paddle/fluid/framework/tensor_util.h b/paddle/fluid/framework/tensor_util.h index 3af68402dc56230171e858bf8f8f8c89c2bfe760..dca279b69382b80e055f661cefe84b81326704b5 100644 --- a/paddle/fluid/framework/tensor_util.h +++ b/paddle/fluid/framework/tensor_util.h @@ -24,10 +24,11 @@ namespace paddle { namespace framework { void TensorCopy(const Tensor& src, const platform::Place& dst_place, - const platform::DeviceContext& ctx, Tensor* dst, - bool sync = false); + const platform::DeviceContext& ctx, Tensor* dst); void TensorCopy(const Tensor& src, const platform::Place& dst_place, Tensor* dst); +void TensorCopySync(const Tensor& src, const platform::Place& dst_place, + Tensor* dst); template void TensorFromVector(const std::vector& src, diff --git a/paddle/fluid/framework/var_type_inference_test.cc b/paddle/fluid/framework/var_type_inference_test.cc index 9e33003b442762210c990b35f30bc3524963b8b4..14b81ddfecb8c996ae8709910c022a074e91eb3c 100644 --- a/paddle/fluid/framework/var_type_inference_test.cc +++ b/paddle/fluid/framework/var_type_inference_test.cc @@ -24,8 +24,7 @@ namespace framework { class SumOpMaker : public OpProtoAndCheckerMaker { public: - SumOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() { AddInput("X", "").AsDuplicable(); AddOutput("Out", ""); AddComment(""); diff --git a/paddle/fluid/inference/CMakeLists.txt b/paddle/fluid/inference/CMakeLists.txt index 50f635a41a99b2ae292d13afde5637a3bf4e6f8c..ec16a1c600a3bafc1c4cbbd920360253c106e3a1 100644 --- a/paddle/fluid/inference/CMakeLists.txt +++ b/paddle/fluid/inference/CMakeLists.txt @@ -1,17 +1,23 @@ set(FLUID_CORE_MODULES proto_desc memory lod_tensor executor init) +# TODO(panyx0718): Should this be called paddle_fluid_inference_api_internal? cc_library(paddle_fluid_api SRCS io.cc DEPS ${FLUID_CORE_MODULES} ${GLOB_OP_LIB}) -# Create static library get_property(fluid_modules GLOBAL PROPERTY FLUID_MODULES) -cc_library(paddle_fluid DEPS ${fluid_modules}) +if(WITH_CONTRIB) + set(fluid_modules "${fluid_modules}" paddle_inference_api) +endif() + +# Create static library +cc_library(paddle_fluid DEPS ${fluid_modules} paddle_fluid_api) # Create shared library cc_library(paddle_fluid_shared SHARED SRCS io.cc - DEPS ${fluid_modules}) + DEPS ${fluid_modules} paddle_fluid_api) + set_target_properties(paddle_fluid_shared PROPERTIES OUTPUT_NAME paddle_fluid) if(NOT APPLE) # TODO(liuyiqun): Temporarily disable the link flag because it is not support on Mac. @@ -20,7 +26,9 @@ if(NOT APPLE) endif() if(WITH_TESTING) + # both tests/book and analysis depends the models that generated by python/paddle/fluid/tests/book add_subdirectory(tests/book) + add_subdirectory(analysis) endif() if (TENSORRT_FOUND) diff --git a/paddle/fluid/inference/analysis/CMakeLists.txt b/paddle/fluid/inference/analysis/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..9faf5bb3036775a2ba0c08d3d6ea17ffa73753c6 --- /dev/null +++ b/paddle/fluid/inference/analysis/CMakeLists.txt @@ -0,0 +1,17 @@ +set(FLUID_CORE_MODULES proto_desc memory lod_tensor executor init) +cc_library(analysis SRCS dot.cc node.cc data_flow_graph.cc graph_traits.cc subgraph_splitter.cc fluid_to_data_flow_graph_pass.cc + DEPS paddle_fluid) +cc_test(test_node SRCS node_tester.cc DEPS analysis) +cc_test(test_dot SRCS dot_tester.cc DEPS analysis) + +set(PYTHON_TESTS_DIR ${PADDLE_BINARY_DIR}/python/paddle/fluid/tests) + +cc_test(test_data_flow_graph SRCS data_flow_graph_tester.cc DEPS analysis ${FLUID_CORE_MODULES} paddle_fluid + ARGS --inference_model_dir=${PYTHON_TESTS_DIR}/book/word2vec.inference.model) +set_tests_properties(test_data_flow_graph PROPERTIES DEPENDS test_word2vec) + +cc_test(test_subgraph_splitter + SRCS subgraph_splitter_tester.cc + DEPS analysis paddle_fluid tensor + ARGS --inference_model_dir=${PYTHON_TESTS_DIR}/book/word2vec.inference.model) +set_tests_properties(test_subgraph_splitter PROPERTIES DEPENDS test_word2vec) diff --git a/paddle/fluid/inference/analysis/data_flow_graph.cc b/paddle/fluid/inference/analysis/data_flow_graph.cc new file mode 100644 index 0000000000000000000000000000000000000000..4220451e3caee62caa51af5bc33d6dd3fd891018 --- /dev/null +++ b/paddle/fluid/inference/analysis/data_flow_graph.cc @@ -0,0 +1,205 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/inference/analysis/data_flow_graph.h" +#include "paddle/fluid/inference/analysis/dot.h" + +namespace paddle { +namespace inference { +namespace analysis { + +// It is a better idea that the inputs and outputs of this graph is set manully +// before, but there must be a Pass that helps to prune the unnecessary ops that +// do not contribute to the given targets, so in this pass, analysis and get the +// inputs and outputs is OK. +void DataFlowGraph::Build() { + inputs.clear(); + outputs.clear(); + std::unordered_set ins; + std::unordered_set outs; + for (auto &node : nodes.nodes()) { + for (auto *in : node->inlinks) { + ins.insert(in); + } + for (auto *out : node->outlinks) { + outs.insert(out); + } + } + + // The nodes that in ins but not in outs is the graph's inputs + // similarly, the nodes that in outs but not in ins is the graphs' outputs + for (auto *in : ins) { + if (!outs.count(in)) { + inputs.push_back(in); + } + } + for (auto *out : outs) { + if (!outs.count(out)) { + outputs.push_back(out); + } + } +} + +std::string DataFlowGraph::DotString() const { + Dot dot; + + // Add nodes + for (size_t i = 0; i < nodes.size(); i++) { + const Node &node = nodes.Get(i); + switch (node.type()) { + case Node::Type::kValue: + dot.AddNode(node.repr(), node.dot_attrs()); + break; + case Node::Type::kFunction: + dot.AddNode(node.repr(), node.dot_attrs()); + break; + case Node::Type::kFunctionBlock: + dot.AddNode(node.repr(), node.dot_attrs()); + break; + default: + PADDLE_THROW("unsupported Node type %d", static_cast(node.type())); + } + } + + // Add edges + for (size_t i = 0; i < nodes.size(); i++) { + const Node &node = nodes.Get(i); + for (auto &in : node.inlinks) { + dot.AddEdge(in->repr(), node.repr(), {}); + } + } + return dot.Build(); +} + +// +// NodesBFSIterator +// + +GraphTraits::NodesBFSIterator::NodesBFSIterator( + const std::vector &source) + : queue_(source.begin(), source.end()) {} + +// GraphTraits::NodesBFSIterator::NodesBFSIterator( +// GraphTraits::NodesBFSIterator &&other) noexcept +// : queue_(std::move(other.queue_)), +// visited_(std::move(other.visited_)) {} + +GraphTraits::NodesBFSIterator::NodesBFSIterator( + const GraphTraits::NodesBFSIterator &other) + : queue_(other.queue_), visited_(other.visited_) {} + +Node &GraphTraits::NodesBFSIterator::operator*() { + PADDLE_ENFORCE(!queue_.empty()); + return *queue_.front(); +} + +Node *GraphTraits::NodesBFSIterator::operator->() { + PADDLE_ENFORCE(!queue_.empty()); + return queue_.front(); +} + +GraphTraits::NodesBFSIterator & +GraphTraits::NodesBFSIterator::operator=( + const GraphTraits::NodesBFSIterator &other) { + queue_ = other.queue_; + visited_ = other.visited_; + return *this; +} + +GraphTraits::NodesBFSIterator + &GraphTraits::NodesBFSIterator::operator++() { + PADDLE_ENFORCE(!queue_.empty()); + auto *cur = queue_.front(); + visited_.insert(cur); + queue_.pop_front(); + for (auto *output : cur->outlinks) { + if (!visited_.count(output)) { + queue_.push_back(output); + visited_.insert(output); + } + } + return *this; +} + +bool GraphTraits::NodesBFSIterator::operator==( + const GraphTraits::NodesBFSIterator &other) { + if (queue_.empty()) return other.queue_.empty(); + if ((!queue_.empty()) && (!other.queue_.empty())) { + return queue_.front() == other.queue_.front() && + visited_.size() == other.visited_.size(); // here need to check the + // equality of queue and + // visited. Just a light but week implementation. + } + return false; +} + +// +// NodesDFSIterator +// +GraphTraits::NodesDFSIterator::NodesDFSIterator( + const std::vector &source) { + for (auto *x : source) stack_.push(x); +} + +// GraphTraits::NodesDFSIterator::NodesDFSIterator( +// GraphTraits::NodesDFSIterator &&other) noexcept +// : stack_(std::move(other.stack_)), +// visited_(std::move(other.visited_)) {} + +GraphTraits::NodesDFSIterator::NodesDFSIterator( + const GraphTraits::NodesDFSIterator &other) + : stack_(other.stack_), visited_(other.visited_) {} + +Node &GraphTraits::NodesDFSIterator::operator*() { + PADDLE_ENFORCE(!stack_.empty()); + return *stack_.top(); +} + +GraphTraits::NodesDFSIterator + &GraphTraits::NodesDFSIterator::operator++() { + if (stack_.empty()) return *this; + visited_.insert(stack_.top()); + auto *cur = stack_.top(); + stack_.pop(); + for (auto *x : cur->outlinks) { + if (!visited_.count(x)) { + stack_.push(x); + visited_.insert(x); + } + } + return *this; +} +bool GraphTraits::NodesDFSIterator::operator==( + const GraphTraits::NodesDFSIterator &other) { + if (stack_.empty()) return other.stack_.empty(); + if ((!stack_.empty()) && (!other.stack_.empty())) { + return stack_.top() == other.stack_.top(); + } + return false; +} + +GraphTraits::NodesDFSIterator & +GraphTraits::NodesDFSIterator::operator=( + const GraphTraits::NodesDFSIterator &other) { + stack_ = other.stack_; + visited_ = other.visited_; + return *this; +} +Node *GraphTraits::NodesDFSIterator::operator->() { + return stack_.top(); +} + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/data_flow_graph.h b/paddle/fluid/inference/analysis/data_flow_graph.h new file mode 100644 index 0000000000000000000000000000000000000000..9f6ce40ede25248a4f779b379c132806a4ec06ba --- /dev/null +++ b/paddle/fluid/inference/analysis/data_flow_graph.h @@ -0,0 +1,159 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +/* + * Data flow graph is an pass that build the basic graph. It contains a graph + * and the iterators that enable the iteration over the graph. + */ + +#pragma once + +#include +#include +#include + +#include "paddle/fluid/inference/analysis/graph_traits.h" +#include "paddle/fluid/inference/analysis/node.h" +#include "paddle/fluid/platform/enforce.h" + +namespace paddle { +namespace inference { +namespace analysis { + +/* + * DataFlowGraph - A container of Value and Function Nodes. + */ +struct DataFlowGraph { + NodeMap nodes; + std::vector inputs; + std::vector outputs; + + // Extract inputs and outputs of the graph. + void Build(); + + // Output a DOT graph file for debug. + std::string DotString() const; +}; + +/* + * An graph trait help to traverse the graph using BFS. + * The BFS start from a graph's inputs, the graph should be fully-connected, so + * that the iterator can reach the end. + */ +template <> +struct GraphTraits { + // BFS iterator on nodes. + struct NodesBFSIterator + : public std::iterator { + NodesBFSIterator() = default; + explicit NodesBFSIterator(const std::vector &source); + // NodesBFSIterator(NodesBFSIterator &&other) noexcept; + // NOTE Heavy to use. + NodesBFSIterator(const NodesBFSIterator &other); + + Node &operator*(); + NodesBFSIterator &operator++(); + Node *operator->(); + // TODO(Superjomn) current implementation just compare the first + // element, need to compare the graph and all the elements in the queue and + // set. + NodesBFSIterator &operator=(const NodesBFSIterator &other); + bool operator==(const NodesBFSIterator &other); + bool operator!=(const NodesBFSIterator &other) { return !(*this == other); } + + private: + std::deque queue_; + std::unordered_set visited_; + }; + + // DFS iterator on nodes. + struct NodesDFSIterator + : public std::iterator { + NodesDFSIterator() = default; + explicit NodesDFSIterator(const std::vector &source); + // NodesDFSIterator(NodesDFSIterator &&other) noexcept; + NodesDFSIterator(const NodesDFSIterator &other); + + Node &operator*(); + NodesDFSIterator &operator++(); + // TODO(Superjomn) current implementation just compare the first + // element, need to compare the graph and all the elements in the queue and + // set. + NodesDFSIterator &operator=(const NodesDFSIterator &other); + bool operator==(const NodesDFSIterator &other); + bool operator!=(const NodesDFSIterator &other) { return !(*this == other); } + Node *operator->(); + + private: + std::stack stack_; + std::unordered_set visited_; + }; + + explicit GraphTraits(DataFlowGraph *graph) : graph_(graph) {} + + // default use BFS to visit the nodes. + iterator_range nodes() { + return iterator_range(nodes_bfs_begin(), nodes_bfs_end()); + } + iterator_range nodes_in_BFS() { + return iterator_range(nodes_bfs_begin(), nodes_bfs_end()); + } + iterator_range nodes_in_DFS() { + return iterator_range(nodes_dfs_begin(), nodes_dfs_end()); + } + + private: + NodesBFSIterator nodes_bfs_begin() { + return NodesBFSIterator(graph_->inputs); + } + NodesBFSIterator nodes_bfs_end() { return NodesBFSIterator(); } + NodesDFSIterator nodes_dfs_begin() { + return NodesDFSIterator(graph_->inputs); + } + NodesDFSIterator nodes_dfs_end() { return NodesDFSIterator(); } + + private: + DataFlowGraph *graph_; +}; + +// Extract the inputs and outputs of a graph. The inputs and outputs of a +// sub-graph is the inputs nodes and output nodes that doesn't inside the +// sub-graph. +std::pair< + std::vector, + std::vector< + Node *>> static ExtractInputAndOutputOfSubGraph(std::vector + &graph) { + std::unordered_set nodes(graph.begin(), graph.end()); + std::unordered_set inputs; + std::unordered_set outputs; + for (auto &node : graph) { + for (auto *in : node->inlinks) { + if (!nodes.count(in) && in->type() == Node::Type::kValue) { + inputs.insert(in); + } + } + for (auto *out : node->outlinks) { + if (!nodes.count(out) && out->type() == Node::Type::kValue) { + outputs.insert(out); + } + } + } + return std::make_pair(std::vector(inputs.begin(), inputs.end()), + std::vector(outputs.begin(), outputs.end())); +} + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/data_flow_graph_tester.cc b/paddle/fluid/inference/analysis/data_flow_graph_tester.cc new file mode 100644 index 0000000000000000000000000000000000000000..9d7cceeb65888b8ba3fdf39e88fc2877abd82d11 --- /dev/null +++ b/paddle/fluid/inference/analysis/data_flow_graph_tester.cc @@ -0,0 +1,62 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/inference/analysis/data_flow_graph.h" +#include "paddle/fluid/inference/analysis/ut_helper.h" + +namespace paddle { +namespace inference { +namespace analysis { + +TEST(DataFlowGraph, BFS) { + auto desc = LoadProgramDesc(); + auto dfg = ProgramDescToDFG(desc); + dfg.Build(); + + for (auto* in : dfg.inputs) { + LOG(INFO) << "inputs: " << in->name() << " " + << static_cast(in->type()); + } + for (auto* out : dfg.outputs) { + LOG(INFO) << "outputs: " << out->name() << " " + << static_cast(out->type()); + } + + GraphTraits trait(&dfg); + auto nodes = trait.nodes(); + size_t count = 0; + for (auto it = nodes.begin(); it != nodes.end(); ++it) { + LOG(INFO) << "visiting " << it->name(); + ++count; + } + ASSERT_EQ(count, dfg.nodes.size()); +} + +TEST(DataFlowGraph, DFS) { + auto desc = LoadProgramDesc(); + auto dfg = ProgramDescToDFG(desc); + dfg.Build(); + GraphTraits trait(&dfg); + auto nodes = trait.nodes_in_DFS(); + size_t count = 0; + for (auto it = nodes.begin(); it != nodes.end(); ++it) { + LOG(INFO) << "visiting " << it->name(); + ++count; + } + ASSERT_EQ(count, dfg.nodes.size()); +} + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass_tester.cc b/paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass_tester.cc new file mode 100644 index 0000000000000000000000000000000000000000..60f159da9140516284449a0274906df004b23ac5 --- /dev/null +++ b/paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass_tester.cc @@ -0,0 +1,49 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.h" + +#include +#include +#include +#include "paddle/fluid/framework/executor.h" +#include "paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.h" +#include "paddle/fluid/inference/analysis/ut_helper.h" +#include "paddle/fluid/inference/io.h" + +namespace paddle { +namespace inference { +namespace analysis { + +TEST_F(DFG_Tester, Test) { + framework::proto::ProgramDesc new_desc; + DataFlowGraph graph; + + FluidToDataFlowGraphPass pass0; + DataFlowGraphToFluidPass pass1; + pass0.Initialize(desc); + pass1.Initialize(&new_desc); + + pass0.Run(&graph); + pass1.Run(&graph); + + pass0.Finalize(); + pass1.Finalize(); + + LOG(INFO) << graph.nodes.size(); +} + +} // analysis +} // inference +} // paddle diff --git a/paddle/fluid/inference/analysis/device.h b/paddle/fluid/inference/analysis/device.h new file mode 100644 index 0000000000000000000000000000000000000000..585c9923291e5f9cb6e50dbc4bcd28c374191048 --- /dev/null +++ b/paddle/fluid/inference/analysis/device.h @@ -0,0 +1,24 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ +#pragma once + +namespace paddle { +namespace inference { +namespace analysis { + +enum class Device { CPU, GPU }; + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/dot.cc b/paddle/fluid/inference/analysis/dot.cc new file mode 100644 index 0000000000000000000000000000000000000000..d5471ffcb594a6915e9e65c0fee5adc5f5bdf40c --- /dev/null +++ b/paddle/fluid/inference/analysis/dot.cc @@ -0,0 +1,23 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/inference/analysis/dot.h" + +namespace paddle { +namespace inference { +namespace analysis { +size_t Dot::counter = 0; +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/dot.h b/paddle/fluid/inference/analysis/dot.h new file mode 100644 index 0000000000000000000000000000000000000000..4bf1840fdda8508b52d7274a338c5b1c95baf354 --- /dev/null +++ b/paddle/fluid/inference/analysis/dot.h @@ -0,0 +1,155 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +/* + * This file implements some helper classes and methods for DOT programming + * support. It will give a visualization of the graph and that helps to debug + * the logics of each Pass. + */ +#pragma once + +#include +#include +#include +#include +#include + +namespace paddle { +namespace inference { +namespace analysis { + +/* + * A Dot template that helps to build a DOT graph definition. + */ +class Dot { + public: + static size_t counter; + + struct Attr { + std::string key; + std::string value; + + Attr(const std::string& key, const std::string& value) + : key(key), value(value) {} + + std::string repr() const { + std::stringstream ss; + ss << key << "=" << '"' << value << '"'; + return ss.str(); + } + }; + + struct Node { + std::string name; + std::vector attrs; + + Node(const std::string& name, const std::vector& attrs) + : name(name), + attrs(attrs), + id_("node_" + std::to_string(Dot::counter++)) {} + + std::string id() const { return id_; } + + std::string repr() const { + std::stringstream ss; + CHECK(!name.empty()); + ss << id_; + for (size_t i = 0; i < attrs.size(); i++) { + if (i == 0) { + ss << "[label=" << '"' << name << '"' << " "; + } + ss << attrs[i].repr(); + ss << ((i < attrs.size() - 1) ? " " : "]"); + } + return ss.str(); + } + + private: + std::string id_; + }; + + struct Edge { + std::string source; + std::string target; + std::vector attrs; + + Edge(const std::string& source, const std::string& target, + const std::vector& attrs) + : source(source), target(target), attrs(attrs) {} + + std::string repr() const { + std::stringstream ss; + CHECK(!source.empty()); + CHECK(!target.empty()); + ss << source << "->" << target; + for (size_t i = 0; i < attrs.size(); i++) { + if (i == 0) { + ss << "["; + } + ss << attrs[i].repr(); + ss << ((i < attrs.size() - 1) ? " " : "]"); + } + return ss.str(); + } + }; + + Dot() = default; + + explicit Dot(const std::vector& attrs) : attrs_(attrs) {} + + void AddNode(const std::string& name, const std::vector& attrs) { + CHECK(!nodes_.count(name)) << "duplicate Node '" << name << "'"; + nodes_.emplace(name, Node{name, attrs}); + } + + void AddEdge(const std::string& source, const std::string& target, + const std::vector& attrs) { + CHECK(!source.empty()); + CHECK(!target.empty()); + auto sid = nodes_.at(source).id(); + auto tid = nodes_.at(target).id(); + edges_.emplace_back(sid, tid, attrs); + } + + // Compile to DOT language codes. + std::string Build() const { + std::stringstream ss; + const std::string indent = " "; + ss << "digraph G {" << '\n'; + + // Add graph attrs + for (const auto& attr : attrs_) { + ss << indent << attr.repr() << '\n'; + } + // add nodes + for (auto& item : nodes_) { + ss << indent << item.second.repr() << '\n'; + } + // add edges + for (auto& edge : edges_) { + ss << indent << edge.repr() << '\n'; + } + ss << "} // end G"; + return ss.str(); + } + + private: + std::unordered_map nodes_; + std::vector edges_; + std::vector attrs_; +}; + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/dot_tester.cc b/paddle/fluid/inference/analysis/dot_tester.cc new file mode 100644 index 0000000000000000000000000000000000000000..56ceb9bd5d6f41a601d66f6124fb7b4099c9337e --- /dev/null +++ b/paddle/fluid/inference/analysis/dot_tester.cc @@ -0,0 +1,62 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/inference/analysis/dot.h" + +#include +#include +#include "paddle/fluid/inference/analysis/data_flow_graph.h" + +namespace paddle { +namespace inference { +namespace analysis { + +class DotTester : public ::testing::Test { + protected: + void SetUp() override { + std::vector attrs({{"title", "hello"}}); + dot.reset(new Dot(attrs)); + dot->AddNode("a", {Dot::Attr{"shape", "box"}, Dot::Attr("color", "blue")}); + dot->AddNode("b", {}); + dot->AddNode("c", {}); + dot->AddEdge("a", "b", {}); + dot->AddEdge("b", "c", {}); + dot->AddEdge("a", "c", {}); + } + + std::unique_ptr dot; +}; + +TEST_F(DotTester, Build) { + auto codes = dot->Build(); + // Output the DOT language code, the generated codes are too long to compare + // the string. + // + // The output is + // + // digraph G { + // title="hello" + // node_1 + // node_2 + // node_0[label="a" shape="box" color="blue"] + // node_0->node_1 + // node_1->node_2 + // node_0->node_2 + // } // end G + LOG(INFO) << '\n' << codes; +} + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.cc b/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.cc new file mode 100644 index 0000000000000000000000000000000000000000..f848a7d1add79c3032da7defc34a406dccf29d2e --- /dev/null +++ b/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.cc @@ -0,0 +1,83 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.h" +#include + +namespace paddle { +namespace inference { +namespace analysis { + +FluidToDataFlowGraphPass::FluidToDataFlowGraphPass() {} + +bool FluidToDataFlowGraphPass::Initialize() { return Pass::Initialize(); } + +bool FluidToDataFlowGraphPass::Initialize( + const framework::proto::ProgramDesc &desc) { + desc_ = &desc; + return true; +} + +bool FluidToDataFlowGraphPass::Finalize() { return Pass::Finalize(); } + +void FluidToDataFlowGraphPass::Run(DataFlowGraph *graph) { + // insert vars + std::unordered_map var2id; + auto &main_block = desc_->blocks(framework::kRootBlockIndex); + for (int i = 0; i < main_block.vars_size(); i++) { + const auto &var = main_block.vars(i); + auto *v = graph->nodes.Create(Node::Type::kValue); + v->SetName(var.name()); + v->SetExtraInfo(const_cast(static_cast(&var))); + var2id[var.name()] = v->id(); + } + for (int i = 0; i < main_block.ops_size(); i++) { + const auto &op = main_block.ops(i); + auto *o = graph->nodes.Create(Node::Type::kFunction); + o->SetName(op.type()); + static_cast(o)->SetFuncType(op.type()); + // Link to the original protobuf message's memory, make it easier to + // generate from a data flow graph to fluid ProgramDesc. + o->SetExtraInfo(const_cast(static_cast(&op))); + // set inputs and outputs + // TODO(Superjomn) make sure the InputNames is the real variable name. + for (int j = 0; j < op.inputs_size(); j++) { + auto &in_var = op.inputs(j); + for (int k = 0; k < in_var.arguments_size(); k++) { + auto *in = graph->nodes.GetMutable(var2id.at(in_var.arguments(k))); + in->outlinks.push_back(o); + o->inlinks.push_back(in); + } + } + for (int j = 0; j < op.outputs_size(); j++) { + auto &out_var = op.outputs(j); + for (int k = 0; k < out_var.arguments_size(); k++) { + auto *out = graph->nodes.GetMutable(var2id[out_var.arguments(k)]); + out->inlinks.push_back(o); + o->outlinks.push_back(out); + } + } + } + // Analysis and extract the inputs and outputs of this graph. + graph->Build(); +} + +Pass *FluidToDataFlowGraphPass::CreatePrinterPass( + std::ostream &os, const std::string &banner) const { + return nullptr; +} + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.h b/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.h new file mode 100644 index 0000000000000000000000000000000000000000..cd0d4fabaafe844bcc5bb8bfc2586971197d9167 --- /dev/null +++ b/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.h @@ -0,0 +1,51 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +/* + * This file implements the transformation from data flow graph to fluid + * ProgramDesc. + */ + +#pragma once + +#include "paddle/fluid/framework/program_desc.h" +#include "paddle/fluid/inference/analysis/data_flow_graph.h" +#include "paddle/fluid/inference/analysis/pass.h" + +namespace paddle { +namespace inference { +namespace analysis { + +/* + * Transform a FluidDesc to a data flow graph. + */ +class FluidToDataFlowGraphPass final : public DataFlowGraphPass { + public: + FluidToDataFlowGraphPass(); + bool Initialize() override; + bool Initialize(const framework::proto::ProgramDesc &desc) override; + bool Finalize() override; + + void Run(DataFlowGraph *graph) override; + + Pass *CreatePrinterPass(std::ostream &os, + const std::string &banner) const override; + + private: + framework::proto::ProgramDesc const *desc_; +}; + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass_tester.cc b/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass_tester.cc new file mode 100644 index 0000000000000000000000000000000000000000..851c98bef305fa9e20dced5f7c26e9d1b6ddf4f2 --- /dev/null +++ b/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass_tester.cc @@ -0,0 +1,37 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.h" + +#include +#include "paddle/fluid/inference/analysis/ut_helper.h" + +namespace paddle { +namespace inference { +namespace analysis { + +TEST_F(DFG_Tester, Init) { + FluidToDataFlowGraphPass pass; + pass.Initialize(); + pass.Initialize(desc); + DataFlowGraph graph; + pass.Run(&graph); + ASSERT_GT(graph.nodes.size(), 0); + pass.Finalize(); + LOG(INFO) << '\n' << graph.DotString(); +} + +} // analysis +} // inference +} // paddle diff --git a/paddle/fluid/inference/analysis/graph_traits.cc b/paddle/fluid/inference/analysis/graph_traits.cc new file mode 100644 index 0000000000000000000000000000000000000000..2ea70a1d2060e03769d67060dc6f008207342b52 --- /dev/null +++ b/paddle/fluid/inference/analysis/graph_traits.cc @@ -0,0 +1,15 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/inference/analysis/graph_traits.h" diff --git a/paddle/fluid/inference/analysis/graph_traits.h b/paddle/fluid/inference/analysis/graph_traits.h new file mode 100644 index 0000000000000000000000000000000000000000..aed2b1e8e27d94b430201d70ecf09d4acc33c8fa --- /dev/null +++ b/paddle/fluid/inference/analysis/graph_traits.h @@ -0,0 +1,63 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +/* + * This file defines the GraphTraits template class that should be specified + * by classes that want to be iteratable by generic graph iterators. + * + * This file also defines the marker class Inverse that is used to iterate over + * graphs in a graph defined, inverse ordering... + */ + +#pragma once + +#include "paddle/fluid/inference/analysis/helper.h" + +namespace paddle { +namespace inference { +namespace analysis { + +/* + * This class should be specialized by different graph types... + * That's why the base class is empty. + */ +template +struct GraphTraits { + // using NodesBFSIterator = xxx + + // NodesBFSIterator nodes_begin(); + // NodesBFSIterator nodes_end(); +}; + +/* + * Inverse - This class is used as a marker class to tell the graph iterator to + * iterate in a graph defined Inverse order. + */ +template +struct Inverse { + const GraphType &graph; + + explicit Inverse(const GraphType &graph) : graph(graph) {} +}; + +/* + * Provide a partial specialization of GraphTraits so that the inverse of an + * inverse turns into the original graph. + */ +template +struct GraphTraits>> : GraphTraits {}; + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/helper.h b/paddle/fluid/inference/analysis/helper.h new file mode 100644 index 0000000000000000000000000000000000000000..24ea9a4bae7132eb1692b0ffb02f8ab5e02b21a9 --- /dev/null +++ b/paddle/fluid/inference/analysis/helper.h @@ -0,0 +1,114 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include +#include + +#include "paddle/fluid/platform/enforce.h" + +namespace paddle { +namespace inference { +namespace analysis { + +template +int AccuDims(Vec &&vec, int size) { + int res = 1; + for (int i = 0; i < size; i++) { + res *= std::forward(vec)[i]; + } + return res; +} + +#define SET_TYPE(type__) dic_[typeid(type__).hash_code()] = #type__; +/* + * Map typeid to representation. + */ +struct DataTypeNamer { + static const DataTypeNamer &Global() { + static auto *x = new DataTypeNamer(); + return *x; + } + + template + const std::string &repr() const { + auto x = typeid(T).hash_code(); + PADDLE_ENFORCE(dic_.count(x), "unknown type for representation"); + return dic_.at(x); + } + + const std::string &repr(size_t &hash) const { + PADDLE_ENFORCE(dic_.count(hash), "unknown type for representation"); + return dic_.at(hash); + } + + private: + DataTypeNamer() { + SET_TYPE(int); + SET_TYPE(bool); + SET_TYPE(float); + } + + std::unordered_map dic_; +}; +#undef SET_TYPE + +template +class iterator_range { + IteratorT begin_, end_; + + public: + template + explicit iterator_range(Container &&c) : begin_(c.begin()), end_(c.end()) {} + + iterator_range(const IteratorT &begin, const IteratorT &end) + : begin_(begin), end_(end) {} + + const IteratorT &begin() const { return begin_; } + const IteratorT &end() const { return end_; } +}; + +/* + * An registry helper class, with its records keeps the order they registers. + */ +template +class OrderedRegistry { + public: + T *Register(const std::string &name, T *x) { + PADDLE_ENFORCE(!dic_.count(name)); + dic_[name] = data_.size(); + data_.emplace_back(std::unique_ptr(x)); + return data_.back().get(); + } + + T *Lookup(const std::string &name) { + auto it = dic_.find(name); + if (it == dic_.end()) return nullptr; + return data_[it->second].get(); + } + + protected: + std::unordered_map dic_; + std::vector> data_; +}; + +} // namespace analysis +} // namespace inference +} // namespace paddle + +#define PADDLE_DISALLOW_COPY_AND_ASSIGN(type__) \ + type__(const type__ &) = delete; \ + void operator=(const type__ &) = delete; diff --git a/paddle/fluid/inference/analysis/node.cc b/paddle/fluid/inference/analysis/node.cc new file mode 100644 index 0000000000000000000000000000000000000000..fe060526080b1ee01aa98f2ff06fb2191eddf9da --- /dev/null +++ b/paddle/fluid/inference/analysis/node.cc @@ -0,0 +1,67 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/inference/analysis/node.h" +#include "glog/logging.h" +#include "paddle/fluid/platform/enforce.h" + +namespace paddle { +namespace inference { +namespace analysis { + +std::vector Value::dot_attrs() const { + return std::vector({Dot::Attr("style", "filled,rounded"), + Dot::Attr("shape", "box"), + Dot::Attr("fillcolor", "red")}); +} + +std::vector Function::dot_attrs() const { + return std::vector({Dot::Attr("style", "filled,rounded"), + Dot::Attr("shape", "diamond"), + Dot::Attr("fillcolor", "yellow")}); +} + +Node *NodeMap::Create(Node::Type type) { + switch (type) { + case Node::Type::kFunction: + nodes_.emplace_back(new Function); + break; + case Node::Type::kValue: + nodes_.emplace_back(new Value); + break; + default: + PADDLE_THROW("Not supported node type."); + } + nodes_.back()->id_ = size() - 1; + return nodes_.back().get(); +} + +Node *NodeMap::GetMutable(size_t id) { + PADDLE_ENFORCE_GT(size(), id); + return nodes_[id].get(); +} + +const Node &NodeMap::Get(size_t id) const { + PADDLE_ENFORCE_GT(size(), id); + return *nodes_[id].get(); +} + +void NodeMap::Delete(size_t id) { + PADDLE_ENFORCE_LT(id, size()); + nodes_[id]->SetDeleted(); +} + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/node.h b/paddle/fluid/inference/analysis/node.h new file mode 100644 index 0000000000000000000000000000000000000000..7972ca25c92186a8c55a76de645f4fdbb089e8d3 --- /dev/null +++ b/paddle/fluid/inference/analysis/node.h @@ -0,0 +1,242 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +/* + * This file defines the Node class and its subclasses. A Node is the basis + * analysis element in a computation graph. + * There are basically two kinds of nodes, the function node and value node. + */ +#pragma once + +#include +#include +#include +#include +#include + +#include "paddle/fluid/inference/analysis/device.h" +#include "paddle/fluid/inference/analysis/dot.h" +#include "paddle/fluid/inference/analysis/helper.h" + +namespace paddle { +namespace inference { +namespace analysis { + +class NodeMap; + +/* + * Node Representation. + * + * This is a very important class for analysis. It is the base class of all + * nodes computed by a program that may be used as operands to other nodes. + * Node is the super class of other important classes such as Function and + * Value, some nodes can have a name. + */ +class Node { + public: + // Node type. NOTE the new node types should add here. + enum class Type { kNone = -1, kFunction, kValue, kFunctionBlock }; + + Node() = default; + + struct Attr; + + // Cast to a subclass type, Function for example. + template + Subclass &As() { + return *dynamic_cast(this); + } + + // Formatted representation of this Node. + virtual std::string repr() const { + return name() + "(" + std::to_string(id()) + ")"; + } + + // DOT node representation. One Node type can customize its own node + // representation. + virtual std::vector dot_attrs() const { + return std::vector({Dot::Attr("style", "filled")}); + } + + // Get an additional attribute and convert it to T data type. NOTE this will + // silently create a new attribute if not exists. + Attr &attr(const std::string &name) { return attrs_[name]; } + + int id() const { return id_; } + + bool deleted() const { return deleted_; } + void SetDeleted() { deleted_ = true; } + + void SetName(const std::string &name) { name_ = name; } + const std::string &name() const { return name_; } + + void SetType(Type type) { type_ = type; } + Type type() const { return type_; } + + void *extra_info() const { return extra_info_; } + void SetExtraInfo(void *extra_info) { extra_info_ = extra_info; } + + // Input links. + std::vector inlinks; + // Output links. + std::vector outlinks; + + // A helper class to maintain the status from Pass. + // TODO(superjomn) add a checker here to ensure the T is primary. + struct Attr { + // NOTE T should be a primary type or a struct combined by several primary + // types. + // NOTE the STL containers should not use here. + // Some usages + // Attr attr; + // T data; + // attr.data.assign((char*)data, sizeof(data)); + + bool &Bool() { return As(); } + float &Float() { return As(); } + int32_t &Int32() { return As(); } + int64_t &Int64() { return As(); } + + private: + template + T &As() { + // init storage in the first usage. + if (data_.empty()) { + VLOG(4) << "resize data to " << sizeof(T); + type_hash_ = typeid(T).hash_code(); + data_.resize(sizeof(T)); + } + PADDLE_ENFORCE(type_hash_ == typeid(T).hash_code(), + "type not matched, origin is %s, want %s", + DataTypeNamer::Global().repr(type_hash_), + DataTypeNamer::Global().repr()); + PADDLE_ENFORCE_EQ(data_.size(), sizeof(T), "Node attr type recast error"); + return *reinterpret_cast(&data_[0]); + } + + private: + std::string data_; + size_t type_hash_{std::numeric_limits::max()}; + }; + + bool IsFunction() const { return type_ == Node::Type::kFunction; } + bool IsValue() const { return type_ == Node::Type::kValue; } + bool IsFunctionBlock() const { return type_ == Node::Type::kFunctionBlock; } + + virtual ~Node() {} + + friend class NodeMap; + + PADDLE_DISALLOW_COPY_AND_ASSIGN(Node); + + protected: + // The id number not the name is a node's unique identifier in the computation + // graph. + int id_{-1}; + std::string name_; + Type type_{Type::kNone}; + // Mark this node is deleted by some pass. + bool deleted_{false}; + + void *extra_info_; + + mutable std::unordered_map attrs_; +}; + +class Function; +/* + * Value represents a value node, it has some attributes including dims, data + * type and so on. + */ +class Value : public Node { + public: + enum class DataType { kInt32, kInt64, kFloat32, kFloat64 }; + using Dims = std::vector; + + void SetDataType(DataType data_type) { data_type_ = data_type; } + DataType data_type() const { return data_type_; } + + void SetDims(const Dims &dims) { dims_ = dims; } + const Dims &dims() const { return dims_; } + + Device device() const { return device_; } + void SetDevice(Device device) { device_ = device; } + + std::vector dot_attrs() const override; + + PADDLE_DISALLOW_COPY_AND_ASSIGN(Value); + + protected: + Value() { SetType(Node::Type::kValue); } + friend class NodeMap; + + private: + DataType data_type_; + Dims dims_; + Device device_; +}; + +/* + * Function represents any kind of executable concepts that takes several Values + * as input, and outputs several Values. + */ +class Function : public Node { + public: + std::vector dot_attrs() const override; + + // Get the operator's type from Desc. + const std::string &func_type() const { return func_type_; } + // Set the operator's type. + void SetFuncType(const std::string &func_type) { func_type_ = func_type; } + + PADDLE_DISALLOW_COPY_AND_ASSIGN(Function); + + protected: + std::string func_type_; + Function() { SetType(Node::Type::kFunction); } + friend class NodeMap; +}; + +/* + * FunctionBlock is a Node that contains a sub-graph multiple Node. + */ +struct FunctionBlock : public Node { + std::string repr() const override { return "block-" + std::to_string(id()); } + std::vector subgraph; +}; + +class NodeMap { + public: + // Create a new node with type. + Node *Create(Node::Type type); + + // Get a node by its id. + Node *GetMutable(size_t id); + + const Node &Get(size_t id) const; + + void Delete(size_t id); + + const std::vector> &nodes() { return nodes_; } + + size_t size() const { return nodes_.size(); } + + private: + std::vector> nodes_; + std::unordered_map map_; +}; + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/node_tester.cc b/paddle/fluid/inference/analysis/node_tester.cc new file mode 100644 index 0000000000000000000000000000000000000000..ea832a3a7e47758be9b6bd59a4325ddb576ec446 --- /dev/null +++ b/paddle/fluid/inference/analysis/node_tester.cc @@ -0,0 +1,34 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/inference/analysis/node.h" + +#include + +namespace paddle { +namespace inference { +namespace analysis { + +TEST(Node, Attr) { + // Node is an abstract class, use Value instead for they share the same Attr + // logic. + NodeMap nodes; + auto* node = nodes.Create(Node::Type::kValue); + node->attr("v0").Int32() = 2008; + ASSERT_EQ(node->attr("v0").Int32(), 2008); +} + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/pass.cc b/paddle/fluid/inference/analysis/pass.cc new file mode 100644 index 0000000000000000000000000000000000000000..121b72c0a0aa9a0c568b04f7ee9a5bc5c1d6f5f8 --- /dev/null +++ b/paddle/fluid/inference/analysis/pass.cc @@ -0,0 +1,15 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/inference/analysis/pass.h" diff --git a/paddle/fluid/inference/analysis/pass.h b/paddle/fluid/inference/analysis/pass.h new file mode 100644 index 0000000000000000000000000000000000000000..5c89b1304d84abc9a4942f12da46b4bfe76f44f5 --- /dev/null +++ b/paddle/fluid/inference/analysis/pass.h @@ -0,0 +1,90 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include + +#include "paddle/fluid/framework/framework.pb.h" +#include "paddle/fluid/inference/analysis/data_flow_graph.h" +#include "paddle/fluid/inference/analysis/helper.h" +#include "paddle/fluid/inference/analysis/node.h" + +namespace paddle { +namespace inference { +namespace analysis { + +class Pass { + public: + Pass() = default; + virtual ~Pass() {} + // Virtual method overridden by subclasses to do only necessary initialization + // before any pass is run. + virtual bool Initialize() { return false; } + // There is some passes such as FlowToDataFlowGraphPass that needs a + // ProgramDesc. Here use the native ProgramDesc ProtoBuf message, so that it + // only couple with the proto file. + virtual bool Initialize(const framework::proto::ProgramDesc &desc) { + return false; + } + // There are some Passes such as DataFlowGraphToFluidPass that will output a + // ProgramDesc. + virtual bool Initialize(framework::proto::ProgramDesc *desc) { return false; } + + // Virtual method overriden by subclasses to do any necessary clean up after + // all passes have run. + virtual bool Finalize() { return false; } + + // Get a Pass appropriate to print the Node this pass operates on. + virtual Pass *CreatePrinterPass(std::ostream &os, + const std::string &banner) const = 0; + + // Run on a single Node. + virtual void Run(Node *x) { LOG(FATAL) << "not valid"; } + // Run on a single Function. + virtual void Run(Function *x) { LOG(FATAL) << "not valid"; } + // Run on a single FunctionBlock. + virtual void Run(FunctionBlock *x) { LOG(FATAL) << "not valid"; } + // Run on a single DataFlowGraph. + virtual void Run(DataFlowGraph *x) { LOG(FATAL) << "not valid"; } +}; + +// NodePass process on any Node types. +class NodePass : public Pass { + public: + virtual void Run(Node *node) = 0; +}; + +// NodePass process on any Function node types. +class FunctionPass : public Pass { + public: + virtual void Run(Function *node) = 0; +}; + +// NodePass process on any FunctionBlock node types. +class FunctionBlockPass : public Pass { + public: + virtual void Run(FunctionBlock *node) = 0; +}; + +// GraphPass processes on any GraphType. +class DataFlowGraphPass : public Pass { + public: + virtual void Run(DataFlowGraph *graph) = 0; +}; + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/subgraph_splitter.cc b/paddle/fluid/inference/analysis/subgraph_splitter.cc new file mode 100644 index 0000000000000000000000000000000000000000..43ccac96c84e987ad1f494af3e314c810fc1ffe3 --- /dev/null +++ b/paddle/fluid/inference/analysis/subgraph_splitter.cc @@ -0,0 +1,154 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/inference/analysis/subgraph_splitter.h" + +namespace paddle { +namespace inference { +namespace analysis { + +const char *SubGraphSplitter::kMarkerAttrName = + "_sub_graph_splitter_inside_sub_graph"; + +std::vector> SubGraphSplitter::operator()() { + MarkNodesInsideSubGraph(); + return ExtractSubGraphs(); +} + +// Mark the output variables inside a subgraph with the func. +inline void MarkOutLinksInSubGraph(const Function *func) { + for (auto *var : func->outlinks) { + var->attr(SubGraphSplitter::kMarkerAttrName).Bool() = true; + } +} + +void SubGraphSplitter::MarkNodesInsideSubGraph() { + for (auto &node : GraphTraits(graph_).nodes()) { + if (node_inside_subgraph_teller_(&node)) { + node.attr(kMarkerAttrName).Bool() = true; + if (node.type() == Node::Type::kFunction) { + // If a function is inside the sub-graph, mark all the output variables + // to be inside too, so that two marked functions will be inside a same + // sub-graph, lets take a example: A_function->var->B_function, if + // A_function is marked, var should also be marked, so that B_function + // will be in the same sub-graph with A_function if B_function is + // marked. + MarkOutLinksInSubGraph(static_cast(&node)); + } + } + } +} + +const char *kUnionFindParent = "_sub_graph_splitter_union_find_parent_"; + +// Use the Union Find(UF) algorithm to find fully connected sub-graphs, if node +// a's output is node b, that is a and b is in the same sub-graph. The UF +// algorithm will group them to the same cluster. +using node_map_t = std::unordered_map; +// Find the ancestor id of a node. +int UnionFindGetAncestor(const node_map_t &node_map, size_t id) { + int tmp = id; + do { + tmp = node_map.at(tmp)->attr(kUnionFindParent).Int32(); + } while (node_map.at(tmp)->attr(kUnionFindParent).Int32() != tmp); + return tmp; +} +// Make this two node share the same ancestor. +// TODO(Superjom) bad performance, make a balanced tree latter. +void UnionFindCombine(const node_map_t &node_map, size_t a, size_t b) { + int a_ancestor = UnionFindGetAncestor(node_map, a); + int b_ancestor = UnionFindGetAncestor(node_map, b); + node_map.at(b_ancestor)->attr(kUnionFindParent).Int32() = a_ancestor; + node_map.at(a)->attr(kUnionFindParent).Int32() = a_ancestor; + node_map.at(b)->attr(kUnionFindParent).Int32() = a_ancestor; +} + +std::vector> SubGraphSplitter::ExtractSubGraphs() { + std::vector marked_nodes; + for (auto &node : GraphTraits(graph_).nodes()) { + if (node.attr(kMarkerAttrName).Bool()) { + marked_nodes.push_back(&node); + } + } + // extract sub-graphs in the marked node set, use Union Find algorithm. + node_map_t node_map; // id to ptr + for (auto *n : marked_nodes) { + // n's parent == n.id means it is the ancestor + n->attr(kUnionFindParent).Int32() = n->id(); + node_map[n->id()] = n; + } + std::unordered_set visited; + for (auto *n : marked_nodes) { + for (auto *out : n->outlinks) { + if (node_map.count(out->id())) { + UnionFindCombine(node_map, n->id(), out->id()); + } + } + } + + std::unordered_map> clusters; + for (auto *n : marked_nodes) { + if (n->type() == Node::Type::kFunction) { + clusters[UnionFindGetAncestor(node_map, + n->attr(kUnionFindParent).Int32())] + .push_back(n); + } + } + std::vector> result; + std::for_each(clusters.begin(), clusters.end(), + [&](const decltype(clusters)::value_type &it) { + result.push_back(it.second); + }); + + return result; +} + +void SubGraphFuse::operator()() { ReplaceNodesWithSubGraphs(); } + +void SubGraphFuse::ReplaceNodesWithSubGraphs() { + auto subgraphs = SubGraphSplitter(graph_, node_inside_subgraph_teller_)(); + for (auto &subgraph : subgraphs) { + // replace this sub-graph with the first node. Two steps: 1. Create a Block + // Node that contains this subgraph 2. Mark the nodes inside the sub-graph + // as deleted. 3. Replace the deleted node with the new Block Node. + auto *block_node = graph_->nodes.Create(Node::Type::kFunctionBlock); + auto io = ExtractInputAndOutputOfSubGraph(subgraph); + block_node->inlinks = std::move(io.first); + block_node->outlinks = std::move(io.second); + for (auto *node : subgraph) { + // TODO(Superjomn) need a unified mechanism to treat deleted node in each + // pass. + node->SetDeleted(); + } + + std::unordered_map + delelte_node_map; // deleted node to BlockNode + for (auto *n : block_node->inlinks) { + n->inlinks.clear(); + } + for (auto *n : block_node->outlinks) { + n->outlinks.clear(); + } + for (auto *n : block_node->inlinks) { + n->outlinks.push_back(block_node); + } + for (auto *n : block_node->outlinks) { + n->inlinks.push_back(n); + } + } +} + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/subgraph_splitter.h b/paddle/fluid/inference/analysis/subgraph_splitter.h new file mode 100644 index 0000000000000000000000000000000000000000..ed90a0dcf31e154c4d82be08ce35e2f11d11c139 --- /dev/null +++ b/paddle/fluid/inference/analysis/subgraph_splitter.h @@ -0,0 +1,81 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +/* + * This file defines the the class to partition a graph. + */ + +#pragma once + +#include "paddle/fluid/inference/analysis/data_flow_graph.h" +#include "paddle/fluid/inference/analysis/node.h" + +namespace paddle { +namespace inference { +namespace analysis { + +/* + * Detect the nodes in a sub-graph that meet some conditions. This class doesn't + * modify the graph. + */ +class SubGraphSplitter { + public: + static const char *kMarkerAttrName; + // Tell whether a node is inside a sub-graph. + using NodeInsideSubgraphTeller = std::function; + + SubGraphSplitter(DataFlowGraph *graph, const NodeInsideSubgraphTeller &teller) + : graph_(graph), node_inside_subgraph_teller_(teller) {} + + std::vector> operator()(); + + protected: + // Mark the nodes inside the accepted sub-graph using + // node_inside_subgraph_teller. + void MarkNodesInsideSubGraph(); + + // Merge the marked nodes into sub-graphs and return the sub-graphs. + std::vector> ExtractSubGraphs(); + + private: + DataFlowGraph *graph_; + NodeInsideSubgraphTeller node_inside_subgraph_teller_; +}; + +/* + * SubGraphFuse - Replace some nodes with the sub-graph node they are inside. To + * some extent, the TensorRT engine is just a fusion op for a model. + */ +class SubGraphFuse { + public: + using NodeInsideSubgraphTeller = SubGraphSplitter::NodeInsideSubgraphTeller; + + SubGraphFuse(DataFlowGraph *graph, const NodeInsideSubgraphTeller &teller) + : graph_(graph), node_inside_subgraph_teller_(teller) {} + + // The main method which run all the logic. + void operator()(); + + protected: + // Remove the nodes inside sub-graphs and replace with the SubGraphNode. + void ReplaceNodesWithSubGraphs(); + + private: + DataFlowGraph *graph_; + NodeInsideSubgraphTeller node_inside_subgraph_teller_; +}; + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/subgraph_splitter_tester.cc b/paddle/fluid/inference/analysis/subgraph_splitter_tester.cc new file mode 100644 index 0000000000000000000000000000000000000000..0644c0db12e3daabba76dbaad33847f5624b157a --- /dev/null +++ b/paddle/fluid/inference/analysis/subgraph_splitter_tester.cc @@ -0,0 +1,67 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/inference/analysis/subgraph_splitter.h" +#include "paddle/fluid/inference/analysis/ut_helper.h" + +namespace paddle { +namespace inference { +namespace analysis { + +TEST_F(DFG_Tester, Split) { + auto desc = LoadProgramDesc(); + auto dfg = ProgramDescToDFG(desc); + LOG(INFO) << "spliter\n" << dfg.DotString(); + + SubGraphSplitter::NodeInsideSubgraphTeller teller = [](const Node* node) { + if (node->type() != Node::Type::kFunction) return false; + const auto* func = static_cast(node); + if (func->func_type() == "elementwise_add" || func->func_type() == "relu" || + func->func_type() == "conv2d" || func->func_type() == "mul" || + func->func_type() == "sigmoid" || func->func_type() == "softmax") { + LOG(INFO) << "sub-graph marked " << node->repr(); + return true; + } + return false; + }; + ASSERT_GT(dfg.nodes.size(), 5UL); + + auto subgraphs = SubGraphSplitter(&dfg, teller)(); + + // Check the number of the marked nodes. + int marked_nodes = 0; + for (auto& node : dfg.nodes.nodes()) { + if (node->IsFunction() && + node->attr(SubGraphSplitter::kMarkerAttrName).Bool()) { + ++marked_nodes; + } + } + EXPECT_EQ(marked_nodes, 6); + + // For human debug. + for (auto& subgraph : subgraphs) { + LOG(INFO) << "subgraph size " << subgraph.size(); + for (auto* node : subgraph) { + LOG(INFO) << "node " << node->repr(); + } + } + + ASSERT_EQ(subgraphs.size(), 1UL); + // The last sub-graph has 5 Functions. + ASSERT_EQ(subgraphs.back().size(), 6UL); +} + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/ut_helper.h b/paddle/fluid/inference/analysis/ut_helper.h new file mode 100644 index 0000000000000000000000000000000000000000..c86083d12153921672e15c172b874f77a8b46cde --- /dev/null +++ b/paddle/fluid/inference/analysis/ut_helper.h @@ -0,0 +1,58 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include +#include +#include "paddle/fluid/framework/executor.h" +#include "paddle/fluid/inference/analysis/data_flow_graph.h" +#include "paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.h" +#include "paddle/fluid/inference/analysis/ut_helper.h" +#include "paddle/fluid/inference/io.h" + +namespace paddle { +namespace inference { +namespace analysis { + +DEFINE_string(inference_model_dir, "", "inference test model dir"); + +static framework::proto::ProgramDesc LoadProgramDesc( + const std::string& model_dir = FLAGS_inference_model_dir) { + paddle::platform::CPUPlace place; + paddle::framework::Executor executor(place); + paddle::framework::Scope scope; + auto program = Load(&executor, &scope, model_dir); + return *program->Proto(); +} + +static DataFlowGraph ProgramDescToDFG( + const framework::proto::ProgramDesc& desc) { + DataFlowGraph graph; + FluidToDataFlowGraphPass pass; + pass.Initialize(desc); + pass.Run(&graph); + pass.Finalize(); + return graph; +} + +class DFG_Tester : public ::testing::Test { + protected: + void SetUp() override { desc = LoadProgramDesc(FLAGS_inference_model_dir); } + + framework::proto::ProgramDesc desc; +}; + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/engine.h b/paddle/fluid/inference/engine.h index 0633c052e4dc61d495b9acacb9a7ab6c941a267a..ce2b8161715a3fa2278ce950dbac82c6d0042bef 100644 --- a/paddle/fluid/inference/engine.h +++ b/paddle/fluid/inference/engine.h @@ -14,11 +14,15 @@ limitations under the License. */ #pragma once +#include #include "paddle/fluid/framework/framework.pb.h" namespace paddle { namespace inference { +struct Buffer; +enum class DeviceType { UNK = -1, CPU, GPU }; + /* * EngineBase is the base class of all inference engines. An inference engine * takes a paddle program as input, and outputs the result in fluid Tensor @@ -45,9 +49,20 @@ class EngineBase { // Execute the engine, that will run the inference network. virtual void Execute(int batch_size) = 0; - virtual ~EngineBase() {} + // Return the IO buffer that allocated in engine. One can read/write directly + // on the buffer. If the buffer's buffer is nullptr, one can also allocate + // memory and maintain it outside the engine. + virtual Buffer& buffer(const std::string& name) = 0; + virtual ~EngineBase() {} }; // class EngineBase +struct Buffer { + void* buffer{nullptr}; // buffer should be allocated only once. + size_t max_size; // buffer allocated space. + size_t size; // data size. + DeviceType device{DeviceType::UNK}; // tells which device this buffer is on. +}; + } // namespace inference } // namespace paddle diff --git a/paddle/fluid/inference/io.cc b/paddle/fluid/inference/io.cc index 78d2f16746cf478c4424df929bd1f62b08f8a67c..65db7c7b5008dcb301e741ec17c3623715e10bb8 100644 --- a/paddle/fluid/inference/io.cc +++ b/paddle/fluid/inference/io.cc @@ -16,17 +16,29 @@ limitations under the License. */ #include #include +#include #include "paddle/fluid/framework/block_desc.h" #include "paddle/fluid/framework/feed_fetch_type.h" #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/pybind/pybind.h" +DEFINE_string(devices, "", "The devices to be used which is joined by comma."); +DEFINE_bool(init_p2p, false, "Whether to init p2p."); + namespace paddle { namespace inference { -// Temporarily add this function for exposing framework::InitDevices() when -// linking the inference shared library. -void Init(bool init_p2p) { framework::InitDevices(init_p2p); } +void Init(const std::vector argv) { + framework::InitGflags(argv); + // init devices + std::vector devices; + std::string token; + std::istringstream tokenStream(FLAGS_devices); + while (std::getline(tokenStream, token, ',')) { + devices.push_back(std::stoi(token)); + } + framework::InitDevices(FLAGS_init_p2p, devices); +} void ReadBinaryFile(const std::string& filename, std::string* contents) { std::ifstream fin(filename, std::ios::in | std::ios::binary); diff --git a/paddle/fluid/inference/io.h b/paddle/fluid/inference/io.h index ba3e45099ae7c1626bf11d9527d4fa4c7f772fec..caf599b1a68783f155cd134c2a29e9ffa49a0895 100644 --- a/paddle/fluid/inference/io.h +++ b/paddle/fluid/inference/io.h @@ -25,7 +25,7 @@ limitations under the License. */ namespace paddle { namespace inference { -void Init(bool init_p2p); +void Init(const std::vector argv); void LoadPersistables(framework::Executor* executor, framework::Scope* scope, const framework::ProgramDesc& main_program, diff --git a/paddle/fluid/inference/tensorrt/CMakeLists.txt b/paddle/fluid/inference/tensorrt/CMakeLists.txt index 4b5866ad5dd5c2e33830b61e575ab92954cecb39..b52d083f280e5e7713600a7b748dedd37aca0a1e 100644 --- a/paddle/fluid/inference/tensorrt/CMakeLists.txt +++ b/paddle/fluid/inference/tensorrt/CMakeLists.txt @@ -1,4 +1,4 @@ -if(WITH_TESTING) - nv_test(test_tensorrt SRCS test_tensorrt.cc DEPS dynload_cuda device_context dynamic_loader) - nv_test(test_tensorrt_engine SRCS test_engine.cc engine.cc DEPS dynload_cuda) -endif() +nv_library(tensorrt_engine SRCS engine.cc DEPS framework_proto) +nv_test(test_tensorrt SRCS test_tensorrt.cc DEPS dynload_cuda device_context dynamic_loader) +nv_test(test_tensorrt_engine SRCS test_engine.cc DEPS dynload_cuda tensorrt_engine) +add_subdirectory(convert) diff --git a/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt b/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..5ada1d631269209e912e2d4817382ea2c6c67353 --- /dev/null +++ b/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt @@ -0,0 +1,10 @@ +# Add TRT tests +nv_test(test_op_converter SRCS test_op_converter.cc mul_op.cc conv2d_op.cc DEPS ${FLUID_CORE_MODULES} tensorrt_engine) +# This test is not stable +# See https://paddleci.ngrok.io/viewLog.html?tab=buildLog&buildTypeId=Paddle_PrCi2&buildId=36834&_focus=8828 +#nv_test(test_trt_activation_op SRCS test_activation_op.cc activation_op.cc io_converter.cc +# DEPS ${FLUID_CORE_MODULES} activation_op tensorrt_engine +# SERIAL) +nv_test(test_io_converter SRCS test_io_converter.cc io_converter.cc DEPS dynload_cuda dynamic_loader lod_tensor) +nv_test(test_trt_mul_op SRCS test_mul_op.cc mul_op.cc + DEPS ${FLUID_CORE_MODULES} tensorrt_engine mul_op SERIAL) diff --git a/paddle/fluid/inference/tensorrt/convert/activation_op.cc b/paddle/fluid/inference/tensorrt/convert/activation_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..6297051e5a30f1daa512d25d5aa3ab3b2f79f1d1 --- /dev/null +++ b/paddle/fluid/inference/tensorrt/convert/activation_op.cc @@ -0,0 +1,43 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/inference/tensorrt/convert/op_converter.h" + +namespace paddle { +namespace inference { +namespace tensorrt { + +class ReluOpConverter : public OpConverter { + public: + ReluOpConverter() {} + void operator()(const framework::proto::OpDesc& op) override { + // Here the two nullptr looks strange, that's because the + // framework::OpDesc's constructor is strange. + framework::OpDesc op_desc(op, nullptr, nullptr); + LOG(INFO) << "convert a fluid relu op to tensorrt activation layer whose " + "type is Relu"; + const nvinfer1::ITensor* input_tensor = + engine_->GetITensor(op_desc.Input("X")[0]); + nvinfer1::IActivationLayer* layer = TRT_ENGINE_ADD_LAYER( + engine_, Activation, *const_cast(input_tensor), + nvinfer1::ActivationType::kRELU); + engine_->SetITensor(op_desc.Output("Out")[0], layer->getOutput(0)); + } +}; + +REGISTER_TRT_OP_CONVERTER(relu, ReluOpConverter); + +} // namespace tensorrt +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/tensorrt/convert/conv2d_op.cc b/paddle/fluid/inference/tensorrt/convert/conv2d_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..209936c3bafb0d31546856dc36c1b48053a0634b --- /dev/null +++ b/paddle/fluid/inference/tensorrt/convert/conv2d_op.cc @@ -0,0 +1,34 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/inference/tensorrt/convert/op_converter.h" + +namespace paddle { +namespace inference { +namespace tensorrt { + +class Conv2dOpConverter : public OpConverter { + public: + Conv2dOpConverter() {} + void operator()(const framework::proto::OpDesc& op) override { + LOG(INFO) + << "convert a fluid conv2d op to tensorrt conv layer without bias"; + } +}; + +REGISTER_TRT_OP_CONVERTER(conv2d, Conv2dOpConverter); + +} // namespace tensorrt +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/tensorrt/convert/io_converter.cc b/paddle/fluid/inference/tensorrt/convert/io_converter.cc new file mode 100644 index 0000000000000000000000000000000000000000..854f434d93e81237dc85c5df62debcf3b3824b78 --- /dev/null +++ b/paddle/fluid/inference/tensorrt/convert/io_converter.cc @@ -0,0 +1,74 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/inference/tensorrt/convert/io_converter.h" +#include +#include "paddle/fluid/platform/enforce.h" + +namespace paddle { +namespace inference { +namespace tensorrt { + +using platform::is_gpu_place; +using platform::is_cpu_place; + +class DefaultIOConverter : public EngineIOConverter { + public: + DefaultIOConverter() {} + // NOTE out is GPU memory. + virtual void operator()(const LoDTensor& in, void* out, + size_t max_size) override { + PADDLE_ENFORCE(out != nullptr); + PADDLE_ENFORCE(stream_ != nullptr); + const auto& place = in.place(); + size_t size = in.memory_size(); + PADDLE_ENFORCE_LE(size, max_size); + if (is_cpu_place(place)) { + PADDLE_ENFORCE_EQ(0, cudaMemcpyAsync(out, in.data(), size, + cudaMemcpyHostToDevice, *stream_)); + } else if (is_gpu_place(place)) { + PADDLE_ENFORCE_EQ(0, cudaMemcpyAsync(out, in.data(), size, + cudaMemcpyDeviceToDevice, *stream_)); + } else { + PADDLE_THROW("Unknown device for converter"); + } + cudaStreamSynchronize(*stream_); + } + // NOTE in is GPU memory. + virtual void operator()(const void* in, LoDTensor* out, + size_t max_size) override { + PADDLE_ENFORCE(in != nullptr); + PADDLE_ENFORCE(stream_ != nullptr); + const auto& place = out->place(); + size_t size = out->memory_size(); + PADDLE_ENFORCE_LE(size, max_size); + if (is_cpu_place(place)) { + PADDLE_ENFORCE_EQ(0, cudaMemcpyAsync(out->data(), in, size, + cudaMemcpyDeviceToHost, *stream_)); + } else if (is_gpu_place(place)) { + PADDLE_ENFORCE_EQ(0, cudaMemcpyAsync(out->data(), in, size, + cudaMemcpyDeviceToDevice, *stream_)); + } else { + PADDLE_THROW("Unknown device for converter"); + } + cudaStreamSynchronize(*stream_); + } +}; + +// fluid LodTensor <-> tensorrt ITensor +REGISTER_TENSORRT_IO_CONVERTER(default, DefaultIOConverter); + +} // namespace tensorrt +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/tensorrt/convert/io_converter.h b/paddle/fluid/inference/tensorrt/convert/io_converter.h new file mode 100644 index 0000000000000000000000000000000000000000..71c48e085d25d2bc6720d93735f661f9e3af7b40 --- /dev/null +++ b/paddle/fluid/inference/tensorrt/convert/io_converter.h @@ -0,0 +1,82 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/inference/utils/singleton.h" + +namespace paddle { +namespace inference { +namespace tensorrt { + +using framework::LoDTensor; + +/* + * Convert Input from Fluid to TensorRT Engine. + * Convert Output from TensorRT Engine to Fluid. + * + * Note that TensorRT's ITensor follows row major, NCHW. Fluid is also row + * major, + * so in the default case just need to copy the data. + */ +class EngineIOConverter { + public: + EngineIOConverter() {} + + virtual void operator()(const LoDTensor& in, void* out, size_t max_size) {} + virtual void operator()(const void* in, LoDTensor* out, size_t max_size) {} + + void SetStream(cudaStream_t* stream) { stream_ = stream; } + + static void ConvertInput(const std::string& op_type, const LoDTensor& in, + void* out, size_t max_size, cudaStream_t* stream) { + PADDLE_ENFORCE(stream != nullptr); + auto* converter = Registry::Lookup( + op_type, "default" /* default_type */); + PADDLE_ENFORCE_NOT_NULL(converter); + converter->SetStream(stream); + (*converter)(in, out, max_size); + } + + static void ConvertOutput(const std::string& op_type, const void* in, + LoDTensor* out, size_t max_size, + cudaStream_t* stream) { + PADDLE_ENFORCE(stream != nullptr); + auto* converter = Registry::Lookup( + op_type, "default" /* default_type */); + PADDLE_ENFORCE_NOT_NULL(converter); + converter->SetStream(stream); + (*converter)(in, out, max_size); + } + + virtual ~EngineIOConverter() {} + + protected: + cudaStream_t* stream_{nullptr}; +}; + +#define REGISTER_TENSORRT_IO_CONVERTER(op_type__, Converter__) \ + struct trt_io_##op_type__##_converter { \ + trt_io_##op_type__##_converter() { \ + Registry::Register(#op_type__); \ + } \ + }; \ + trt_io_##op_type__##_converter trt_io_##op_type__##_converter__; + +} // namespace tensorrt +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/tensorrt/convert/mul_op.cc b/paddle/fluid/inference/tensorrt/convert/mul_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..ed09f54bde00d12aaec829ba90cc08ebfef57e92 --- /dev/null +++ b/paddle/fluid/inference/tensorrt/convert/mul_op.cc @@ -0,0 +1,47 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/inference/tensorrt/convert/op_converter.h" + +namespace paddle { +namespace inference { +namespace tensorrt { + +/* + * MulOp, IMatrixMultiplyLayer in TRT. This Layer doesn't has weights. + */ +class MulOpConverter : public OpConverter { + public: + MulOpConverter() {} + void operator()(const framework::proto::OpDesc& op) override { + VLOG(4) << "convert a fluid mul op to tensorrt fc layer without bias"; + + framework::OpDesc op_desc(op, nullptr, nullptr); + // Declare inputs + auto* input1 = engine_->GetITensor(op_desc.Input("X")[0]); + auto* input2 = engine_->GetITensor(op_desc.Input("Y")[0]); + // Both the input1 and input2 do not need transpose. + auto* layer = TRT_ENGINE_ADD_LAYER( + engine_, MatrixMultiply, *const_cast(input1), false, + *const_cast(input2), false); + + engine_->DeclareOutput(layer, 0, op_desc.Output("Out")[0]); + } +}; + +REGISTER_TRT_OP_CONVERTER(mul, MulOpConverter); + +} // namespace tensorrt +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/tensorrt/convert/op_converter.h b/paddle/fluid/inference/tensorrt/convert/op_converter.h new file mode 100644 index 0000000000000000000000000000000000000000..1cd3ed9a00acead2599420f88499bd0d74c2974b --- /dev/null +++ b/paddle/fluid/inference/tensorrt/convert/op_converter.h @@ -0,0 +1,83 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include +#include "paddle/fluid/framework/block_desc.h" +#include "paddle/fluid/framework/scope.h" +#include "paddle/fluid/inference/tensorrt/engine.h" +#include "paddle/fluid/inference/utils/singleton.h" + +namespace paddle { +namespace inference { +namespace tensorrt { + +/* + * Convert Op from Fluid to TensorRT Engine. + */ +class OpConverter { + public: + OpConverter() {} + virtual void operator()(const framework::proto::OpDesc& op) {} + + void Run(const framework::proto::OpDesc& op, TensorRTEngine* engine) { + std::string type = op.type(); + auto* it = Registry::Lookup(type); + PADDLE_ENFORCE_NOT_NULL(it, "no OpConverter for optype [%s]", type); + it->SetEngine(engine); + (*it)(op); + } + + // convert fluid op to tensorrt layer + void ConvertOp(const framework::proto::OpDesc& op, TensorRTEngine* engine) { + OpConverter::Run(op, engine); + } + + // convert fluid block to tensorrt network + void ConvertBlock(const framework::proto::BlockDesc& block, + TensorRTEngine* engine) { + for (int i = 0; i < block.ops_size(); i++) { + const auto& op = block.ops(i); + OpConverter::Run(op, engine); + } + } + + void SetEngine(TensorRTEngine* engine) { engine_ = engine; } + + virtual ~OpConverter() {} + + // TensorRT engine + TensorRTEngine* engine_{nullptr}; + + private: + // registered op converter map, whose key is the fluid op type, and value is + // the pointer position of corresponding OpConverter class. + std::unordered_map converters_; + // fluid inference scope + framework::Scope* scope_{nullptr}; +}; + +#define REGISTER_TRT_OP_CONVERTER(op_type__, Converter__) \ + struct trt_##op_type__##_converter { \ + trt_##op_type__##_converter() { \ + Registry::Register(#op_type__); \ + } \ + }; \ + trt_##op_type__##_converter trt_##op_type__##_converter__; + +} // namespace tensorrt +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/tensorrt/convert/test_activation_op.cc b/paddle/fluid/inference/tensorrt/convert/test_activation_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..86ca2ca08eb14265e1bfe7abd5eb6af5c83b8a5c --- /dev/null +++ b/paddle/fluid/inference/tensorrt/convert/test_activation_op.cc @@ -0,0 +1,106 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/framework/program_desc.h" +#include "paddle/fluid/inference/tensorrt/convert/io_converter.h" +#include "paddle/fluid/inference/tensorrt/convert/op_converter.h" +#include "paddle/fluid/platform/device_context.h" +#include "paddle/fluid/platform/place.h" + +USE_OP(relu); + +namespace paddle { +namespace inference { +namespace tensorrt { + +void Compare(const std::string op_type, float input, float expect) { + framework::Scope scope; + platform::CUDAPlace place; + platform::CUDADeviceContext ctx(place); + + // init fluid op and variable + auto x_var = scope.Var("X"); + auto x_tensor = x_var->GetMutable(); + x_tensor->Resize({1, 1}); + x_tensor->mutable_data(place); + std::vector init; + init.push_back(input); + framework::TensorFromVector(init, ctx, x_tensor); + + auto out_var = scope.Var("Out"); + auto out_tensor = out_var->GetMutable(); + out_tensor->Resize({1, 1}); + out_tensor->mutable_data(place); + + framework::OpDesc op_desc; + op_desc.SetType(op_type); + op_desc.SetInput("X", {"X"}); + op_desc.SetOutput("Out", {"Out"}); + + auto op = framework::OpRegistry::CreateOp(*op_desc.Proto()); + + // run fluid op + op->Run(scope, place); + // get fluid output + std::vector out1; + framework::TensorToVector(*out_tensor, ctx, &out1); + + // init tensorrt op + cudaStream_t stream; + ASSERT_EQ(0, cudaStreamCreate(&stream)); + TensorRTEngine* engine = new TensorRTEngine(1, 1 << 10, &stream); + engine->InitNetwork(); + engine->DeclareInput("X", nvinfer1::DataType::kFLOAT, + nvinfer1::DimsCHW{1, 1, 1}); + // convert op + OpConverter op_converter; + op_converter.ConvertOp(*op_desc.Proto(), engine); + + engine->DeclareOutput("Out"); + engine->FreezeNetwork(); + + // convert LoDTensor to ITensor + size_t size = x_tensor->memory_size(); + EngineIOConverter::ConvertInput(op_type, *x_tensor, + engine->buffer("X").buffer, size, &stream); + // run tensorrt Outp + engine->Execute(1); + // convert ITensor to LoDTensor + EngineIOConverter::ConvertOutput(op_type, engine->buffer("Out").buffer, + out_tensor, size, &stream); + // get tensorrt output + std::vector out2; + framework::TensorToVector(*out_tensor, ctx, &out2); + + // compare + ASSERT_EQ(out1[0], out2[0]); + ASSERT_EQ(out1[0], expect); + + delete engine; + cudaStreamDestroy(stream); +} + +TEST(OpConverter, ConvertRelu) { + Compare("relu", 1, 1); // relu(1) = 1 + Compare("relu", -5, 0); // relu(-5) = 0 +} + +} // namespace tensorrt +} // namespace inference +} // namespace paddle + +USE_OP(activation); diff --git a/paddle/fluid/inference/tensorrt/convert/test_io_converter.cc b/paddle/fluid/inference/tensorrt/convert/test_io_converter.cc new file mode 100644 index 0000000000000000000000000000000000000000..8f91309a0a00d5131268f026c319e25ba3cb964a --- /dev/null +++ b/paddle/fluid/inference/tensorrt/convert/test_io_converter.cc @@ -0,0 +1,76 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/inference/tensorrt/convert/io_converter.h" + +namespace paddle { +namespace inference { +namespace tensorrt { + +void IOConverterTester(const platform::DeviceContext& ctx) { + cudaStream_t stream; + ASSERT_EQ(0, cudaStreamCreate(&stream)); + + // init fluid in_tensor + framework::LoDTensor in_tensor; + in_tensor.Resize({10, 10}); + auto place = ctx.GetPlace(); + in_tensor.mutable_data(place); + std::vector init; + for (int64_t i = 0; i < 10 * 10; ++i) { + init.push_back(i); + } + framework::TensorFromVector(init, ctx, &in_tensor); + + // init tensorrt buffer + void* buffer; + size_t size = in_tensor.memory_size(); + ASSERT_EQ(cudaMalloc(&buffer, size), 0); + + // convert fluid in_tensor to tensorrt buffer + EngineIOConverter::ConvertInput("test", in_tensor, buffer, size, &stream); + + // convert tensorrt buffer to fluid out_tensor + framework::LoDTensor out_tensor; + out_tensor.Resize({10, 10}); + out_tensor.mutable_data(place); + EngineIOConverter::ConvertOutput("test", buffer, &out_tensor, size, &stream); + + // compare in_tensor and out_tensor + std::vector result; + framework::TensorToVector(out_tensor, ctx, &result); + EXPECT_EQ(init.size(), result.size()); + for (size_t i = 0; i < init.size(); i++) { + EXPECT_EQ(init[i], result[i]); + } + cudaStreamDestroy(stream); +} + +TEST(EngineIOConverterTester, DefaultCPU) { + platform::CPUPlace place; + platform::CPUDeviceContext ctx(place); + IOConverterTester(ctx); +} + +TEST(EngineIOConverterTester, DefaultGPU) { + platform::CUDAPlace place; + platform::CUDADeviceContext ctx(place); + IOConverterTester(ctx); +} + +} // namespace tensorrt +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/tensorrt/convert/test_mul_op.cc b/paddle/fluid/inference/tensorrt/convert/test_mul_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..d8b61d5f08ffd071c112b4677fcb6f6f50784bbc --- /dev/null +++ b/paddle/fluid/inference/tensorrt/convert/test_mul_op.cc @@ -0,0 +1,47 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/inference/tensorrt/convert/ut_helper.h" + +namespace paddle { +namespace inference { +namespace tensorrt { + +TEST(MulOpConverter, main) { + TRTConvertValidation validator(10, 1000); + validator.DeclInputVar("mul-X", nvinfer1::Dims2(10, 6)); + validator.DeclInputVar("mul-Y", nvinfer1::Dims2(6, 10)); + validator.DeclOutputVar("mul-Out", nvinfer1::Dims2(10, 10)); + + // Prepare Op description + framework::OpDesc desc; + desc.SetType("mul"); + desc.SetInput("X", {"mul-X"}); + desc.SetInput("Y", {"mul-Y"}); + desc.SetOutput("Out", {"mul-Out"}); + + LOG(INFO) << "set OP"; + validator.SetOp(*desc.Proto()); + LOG(INFO) << "execute"; + + validator.Execute(10); +} + +} // namespace tensorrt +} // namespace inference +} // namespace paddle + +USE_OP(mul); diff --git a/paddle/fluid/inference/tensorrt/convert/test_op_converter.cc b/paddle/fluid/inference/tensorrt/convert/test_op_converter.cc new file mode 100644 index 0000000000000000000000000000000000000000..9ae7de9cbfa656fbcbb48557bd4b548115897c6d --- /dev/null +++ b/paddle/fluid/inference/tensorrt/convert/test_op_converter.cc @@ -0,0 +1,35 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include "paddle/fluid/framework/program_desc.h" +#include "paddle/fluid/inference/tensorrt/convert/op_converter.h" + +namespace paddle { +namespace inference { +namespace tensorrt { + +TEST(OpConverter, ConvertBlock) { + framework::ProgramDesc prog; + auto* block = prog.MutableBlock(0); + auto* conv2d_op = block->AppendOp(); + conv2d_op->SetType("conv2d"); + + OpConverter converter; + converter.ConvertBlock(*block->Proto(), nullptr /*TensorRTEngine*/); +} + +} // namespace tensorrt +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/tensorrt/convert/ut_helper.h b/paddle/fluid/inference/tensorrt/convert/ut_helper.h new file mode 100644 index 0000000000000000000000000000000000000000..37fcb5c50309db0ad0924a057a6b481750665531 --- /dev/null +++ b/paddle/fluid/inference/tensorrt/convert/ut_helper.h @@ -0,0 +1,156 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +/* + * This file implements a UT framework to make the validation of transforming + * Fluid Op to TRT Layer. + */ + +#pragma once + +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/inference/analysis/helper.h" +#include "paddle/fluid/inference/tensorrt/convert/op_converter.h" +#include "paddle/fluid/inference/tensorrt/engine.h" + +namespace paddle { +namespace inference { +namespace tensorrt { + +/* + * Get a random float value between [low, high] + */ +float random(float low, float high) { + static std::random_device rd; + static std::mt19937 mt(rd()); + std::uniform_real_distribution dist(1.0, 10.0); + return dist(mt); +} + +void RandomizeTensor(framework::LoDTensor* tensor, const platform::Place& place, + const platform::DeviceContext& ctx) { + auto dims = tensor->dims(); + size_t num_elements = analysis::AccuDims(dims, dims.size()); + PADDLE_ENFORCE_GT(num_elements, 0); + auto* data = tensor->mutable_data(place); + for (size_t i = 0; i < num_elements; i++) { + *(data + i) = random(0., 1.); + } +} + +/* + * Help to validate the correctness between Fluid Op and the corresponding TRT + * layer. + */ +class TRTConvertValidation { + public: + TRTConvertValidation() = delete; + + TRTConvertValidation(int batch_size, int workspace_size = 1 << 10) { + // create engine. + engine_.reset(new TensorRTEngine(10, 1 << 10, &stream_)); + engine_->InitNetwork(); + + PADDLE_ENFORCE_EQ(cudaStreamCreate(&stream_), 0); + } + + // Declare a Variable as input with random initialization. + void DeclInputVar(const std::string& name, const nvinfer1::Dims& dims) { + DeclVar(name, dims); + // Declare TRT inputs. + engine_->DeclareInput(name, nvinfer1::DataType::kFLOAT, dims); + } + + void DeclOutputVar(const std::string& name, const nvinfer1::Dims& dims) { + DeclVar(name, dims); + } + + void DeclVar(const std::string& name, const nvinfer1::Dims& dims) { + platform::CPUPlace place; + platform::CPUDeviceContext ctx(place); + + // Init Fluid tensor. + std::vector dim_vec(dims.nbDims); + for (int i = 0; i < dims.nbDims; i++) { + dim_vec[i] = dims.d[i]; + } + auto* x = scope_.Var(name); + auto* x_tensor = x->GetMutable(); + x_tensor->Resize(framework::make_ddim(dim_vec)); + RandomizeTensor(x_tensor, place, ctx); + } + + void SetOp(const framework::proto::OpDesc& desc) { + op_ = framework::OpRegistry::CreateOp(desc); + + OpConverter op_converter; + op_converter.ConvertOp(desc, engine_.get()); + + engine_->FreezeNetwork(); + + // Declare outputs. + op_desc_.reset(new framework::OpDesc(desc, nullptr, nullptr)); + + // Set Inputs. + for (const auto& input : op_desc_->InputArgumentNames()) { + auto* var = scope_.FindVar(input); + PADDLE_ENFORCE(var); + auto tensor = var->GetMutable(); + engine_->SetInputFromCPU( + input, static_cast(tensor->data()), + sizeof(float) * + analysis::AccuDims(tensor->dims(), tensor->dims().size())); + } + } + + void Execute(int batch_size) { + // Execute Fluid Op + // Execute TRT + platform::CPUPlace place; + platform::CPUDeviceContext ctx(place); + engine_->Execute(batch_size); + + op_->Run(scope_, place); + + ASSERT_FALSE(op_desc_->OutputArgumentNames().empty()); + for (const auto& output : op_desc_->OutputArgumentNames()) { + std::vector fluid_out; + std::vector trt_out(200); + engine_->GetOutputInCPU(output, &trt_out[0], 200 * sizeof(float)); + + auto* var = scope_.FindVar(output); + auto tensor = var->GetMutable(); + framework::TensorToVector(*tensor, ctx, &fluid_out); + // Compare two output + ASSERT_FALSE(fluid_out.empty()); + for (size_t i = 0; i < fluid_out.size(); i++) { + EXPECT_LT(std::abs(fluid_out[i] - trt_out[i]), 0.001); + } + } + } + + framework::Scope& scope() { return scope_; } + + private: + std::unique_ptr engine_; + cudaStream_t stream_; + framework::Scope scope_; + std::unique_ptr op_; + std::unique_ptr op_desc_; +}; + +} // namespace tensorrt +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/tensorrt/engine.cc b/paddle/fluid/inference/tensorrt/engine.cc index 03a25f8e8b52b452cb621aeddf3563cc21a033df..a88236ae98e1816fc43796ead596c432b798d7de 100644 --- a/paddle/fluid/inference/tensorrt/engine.cc +++ b/paddle/fluid/inference/tensorrt/engine.cc @@ -18,6 +18,7 @@ limitations under the License. */ #include #include #include +#include "paddle/fluid/inference/analysis/helper.h" #include "paddle/fluid/inference/tensorrt/helper.h" #include "paddle/fluid/platform/enforce.h" @@ -30,16 +31,24 @@ void TensorRTEngine::Build(const DescType& paddle_model) { } void TensorRTEngine::Execute(int batch_size) { - infer_context_->enqueue(batch_size, buffers_.data(), *stream_, nullptr); + std::vector buffers; + for (auto& buf : buffers_) { + PADDLE_ENFORCE_NOT_NULL(buf.buffer, "buffer should be allocated"); + PADDLE_ENFORCE_GT(buf.max_size, 0); + PADDLE_ENFORCE(buf.device == DeviceType::GPU); + buffers.push_back(buf.buffer); + } + infer_context_->enqueue(batch_size, buffers.data(), *stream_, nullptr); cudaStreamSynchronize(*stream_); } TensorRTEngine::~TensorRTEngine() { // clean buffer - for (auto& buffer : buffers_) { - if (buffer != nullptr) { - PADDLE_ENFORCE_EQ(0, cudaFree(buffer)); - buffer = nullptr; + for (auto& buf : buffers_) { + if (buf.buffer != nullptr) { + PADDLE_ENFORCE_EQ(0, cudaFree(buf.buffer)); + buf.buffer = nullptr; + buf.max_size = 0; } } } @@ -59,29 +68,35 @@ void TensorRTEngine::FreezeNetwork() { infer_context_.reset(infer_engine_->createExecutionContext()); // allocate GPU buffers. - buffers_.resize(buffer_sizes_.size(), nullptr); + buffers_.resize(buffer_sizes_.size()); for (auto& item : buffer_sizes_) { if (item.second == 0) { auto slot_offset = infer_engine_->getBindingIndex(item.first.c_str()); + auto dims = infer_engine_->getBindingDimensions(slot_offset); item.second = kDataTypeSize[static_cast( infer_engine_->getBindingDataType(slot_offset))] * - AccumDims(infer_engine_->getBindingDimensions(slot_offset)); + analysis::AccuDims(dims.d, dims.nbDims); } - PADDLE_ENFORCE_EQ(0, cudaMalloc(&buffer(item.first), item.second)); + auto& buf = buffer(item.first); + CHECK(buf.buffer == nullptr); // buffer should be allocated only once. + PADDLE_ENFORCE_EQ(0, cudaMalloc(&buf.buffer, item.second)); + buf.size = buf.max_size = item.second; + buf.device = DeviceType::GPU; } } nvinfer1::ITensor* TensorRTEngine::DeclareInput(const std::string& name, nvinfer1::DataType dtype, - const nvinfer1::Dims& dim) { + const nvinfer1::Dims& dims) { PADDLE_ENFORCE_EQ(0, buffer_sizes_.count(name), "duplicate input name %s", name); PADDLE_ENFORCE(infer_network_ != nullptr, "should initnetwork first"); - auto* input = infer_network_->addInput(name.c_str(), dtype, dim); + auto* input = infer_network_->addInput(name.c_str(), dtype, dims); PADDLE_ENFORCE(input, "infer network add input %s failed", name); - - buffer_sizes_[name] = kDataTypeSize[static_cast(dtype)] * AccumDims(dim); + buffer_sizes_[name] = kDataTypeSize[static_cast(dtype)] * + analysis::AccuDims(dims.d, dims.nbDims); + TensorRTEngine::SetITensor(name, input); return input; } @@ -99,23 +114,51 @@ void TensorRTEngine::DeclareOutput(const nvinfer1::ILayer* layer, int offset, buffer_sizes_[name] = 0; } +void TensorRTEngine::DeclareOutput(const std::string& name) { + PADDLE_ENFORCE_EQ(0, buffer_sizes_.count(name), "duplicate output name %s", + name); + + auto* output = TensorRTEngine::GetITensor(name); + PADDLE_ENFORCE(output != nullptr); + output->setName(name.c_str()); + infer_network_->markOutput(*output); + // output buffers' size can only be decided latter, set zero here to mark this + // and will reset latter. + buffer_sizes_[name] = 0; +} + void* TensorRTEngine::GetOutputInGPU(const std::string& name) { - return buffer(name); + return buffer(name).buffer; } -void TensorRTEngine::GetOutputInCPU(const std::string& name, void* dst, +void TensorRTEngine::GetOutputInGPU(const std::string& name, void* dst, size_t max_size) { // determine data size auto it = buffer_sizes_.find(name); PADDLE_ENFORCE(it != buffer_sizes_.end()); PADDLE_ENFORCE_GT(it->second, 0); PADDLE_ENFORCE_GE(max_size, it->second); + auto& buf = buffer(name); + PADDLE_ENFORCE_NOT_NULL(buf.buffer, "buffer should be allocated before"); + PADDLE_ENFORCE_EQ(cudaMemcpyAsync(dst, buf.buffer, it->second, + cudaMemcpyDeviceToDevice, *stream_), + 0); +} - PADDLE_ENFORCE_EQ(0, cudaMemcpyAsync(dst, buffer(name), it->second, +void TensorRTEngine::GetOutputInCPU(const std::string& name, void* dst, + size_t max_size) { + // determine data size + auto it = buffer_sizes_.find(name); + PADDLE_ENFORCE(it != buffer_sizes_.end()); + PADDLE_ENFORCE_GT(it->second, 0); + PADDLE_ENFORCE_GE(max_size, it->second); + auto& buf = buffer(name); + PADDLE_ENFORCE_NOT_NULL(buf.buffer, "buffer should be allocated before"); + PADDLE_ENFORCE_EQ(0, cudaMemcpyAsync(dst, buf.buffer, it->second, cudaMemcpyDeviceToHost, *stream_)); } -void*& TensorRTEngine::buffer(const std::string& name) { +Buffer& TensorRTEngine::buffer(const std::string& name) { PADDLE_ENFORCE(infer_engine_ != nullptr, "call FreezeNetwork first."); auto it = buffer_sizes_.find(name); PADDLE_ENFORCE(it != buffer_sizes_.end()); @@ -123,11 +166,37 @@ void*& TensorRTEngine::buffer(const std::string& name) { return buffers_[slot_offset]; } -void TensorRTEngine::SetInputFromCPU(const std::string& name, void* data, +void TensorRTEngine::SetInputFromCPU(const std::string& name, const void* data, size_t size) { - void* buf = buffer(name); - PADDLE_ENFORCE_EQ( - 0, cudaMemcpyAsync(buf, data, size, cudaMemcpyHostToDevice, *stream_)); + auto& buf = buffer(name); + PADDLE_ENFORCE_NOT_NULL(buf.buffer); + PADDLE_ENFORCE_LE(size, buf.max_size, "buffer is too small"); + PADDLE_ENFORCE(buf.device == DeviceType::GPU); + PADDLE_ENFORCE_EQ(0, cudaMemcpyAsync(buf.buffer, data, size, + cudaMemcpyHostToDevice, *stream_)); +} + +void TensorRTEngine::SetInputFromGPU(const std::string& name, const void* data, + size_t size) { + auto& buf = buffer(name); + PADDLE_ENFORCE_NOT_NULL(buf.buffer); + PADDLE_ENFORCE_LE(size, buf.max_size, "buffer is too small"); + PADDLE_ENFORCE(buf.device == DeviceType::GPU); + PADDLE_ENFORCE_EQ(0, cudaMemcpyAsync(buf.buffer, data, size, + cudaMemcpyDeviceToDevice, *stream_)); +} + +void TensorRTEngine::SetITensor(const std::string& name, + nvinfer1::ITensor* tensor) { + PADDLE_ENFORCE(tensor != nullptr); + PADDLE_ENFORCE_EQ(0, itensor_map_.count(name), "duplicate ITensor name %s", + name); + itensor_map_[name] = tensor; +} + +nvinfer1::ITensor* TensorRTEngine::GetITensor(const std::string& name) { + PADDLE_ENFORCE(itensor_map_.count(name), "no ITensor %s", name); + return itensor_map_[name]; } } // namespace tensorrt diff --git a/paddle/fluid/inference/tensorrt/engine.h b/paddle/fluid/inference/tensorrt/engine.h index ff853455b8bf0c9db36353f3cbe6be0ed4948fb3..d9d3163b66d4c4c302d12edcc42f00e1cdfa5a30 100644 --- a/paddle/fluid/inference/tensorrt/engine.h +++ b/paddle/fluid/inference/tensorrt/engine.h @@ -16,7 +16,9 @@ limitations under the License. */ #include #include +#include #include +#include #include "paddle/fluid/inference/engine.h" #include "paddle/fluid/inference/tensorrt/helper.h" @@ -56,14 +58,14 @@ class TensorRTEngine : public EngineBase { virtual ~TensorRTEngine(); // TODO(Superjomn) implement it later when graph segmentation is supported. - virtual void Build(const DescType& paddle_model) override; + void Build(const DescType& paddle_model) override; - virtual void Execute(int batch_size) override; + void Execute(int batch_size) override; // Initialize the inference network, so that TensorRT layers can add to this // network. void InitNetwork() { - infer_builder_.reset(createInferBuilder(logger_)); + infer_builder_.reset(createInferBuilder(&logger_)); infer_network_.reset(infer_builder_->createNetwork()); } // After finishing adding ops, freeze this network and creates the executation @@ -78,24 +80,34 @@ class TensorRTEngine : public EngineBase { // name. void DeclareOutput(const nvinfer1::ILayer* layer, int offset, const std::string& name); + // Set the itensor_map_[name] as the network's output, and set its name. + void DeclareOutput(const std::string& name); // GPU memory address for an ITensor with specific name. One can operate on // these memory directly for acceleration, for example, output the converted // data directly to the buffer to save data copy overhead. // NOTE this should be used after calling `FreezeNetwork`. - void*& buffer(const std::string& name); + Buffer& buffer(const std::string& name) override; + + cudaStream_t* stream() { return stream_; } // Fill an input from CPU memory with name and size. - void SetInputFromCPU(const std::string& name, void* data, size_t size); + void SetInputFromCPU(const std::string& name, const void* data, size_t size); // TODO(Superjomn) is this method necessary given that buffer(xxx) can be // accessed directly. Fill an input from GPU memory with name and size. - void SetInputFromGPU(const std::string& name, void* data, size_t size); + void SetInputFromGPU(const std::string& name, const void* data, size_t size); // Get an output called name, the output of tensorrt is in GPU, so this method - // will just return the output's GPU memory address. + // Return the output's GPU memory address without copy. void* GetOutputInGPU(const std::string& name); + // Copy data into dst inside the GPU device. + void GetOutputInGPU(const std::string& name, void* dst, size_t max_size); // LOW EFFICENCY! Get output to CPU, this will trigger a memory copy from GPU // to CPU. void GetOutputInCPU(const std::string& name, void* dst, size_t max_size); + // Fill an ITensor into map itensor_map_. + void SetITensor(const std::string& name, nvinfer1::ITensor* tensor); + // Get an ITensor called name. + nvinfer1::ITensor* GetITensor(const std::string& name); nvinfer1::ICudaEngine* engine() { return infer_engine_.get(); } nvinfer1::INetworkDefinition* network() { return infer_network_.get(); } @@ -108,9 +120,11 @@ class TensorRTEngine : public EngineBase { cudaStream_t* stream_; nvinfer1::ILogger& logger_; - std::vector buffers_; + std::vector buffers_; // max data size for the buffers. std::unordered_map buffer_sizes_; + std::unordered_map + itensor_map_; // TensorRT related internal members template diff --git a/paddle/fluid/inference/tensorrt/helper.h b/paddle/fluid/inference/tensorrt/helper.h index 796283d325ceb84c733eff5c119b808300bca069..b6e7968108403c9c9c192759c44eac040d1c5073 100644 --- a/paddle/fluid/inference/tensorrt/helper.h +++ b/paddle/fluid/inference/tensorrt/helper.h @@ -26,15 +26,6 @@ namespace tensorrt { namespace dy = paddle::platform::dynload; -static size_t AccumDims(nvinfer1::Dims dims) { - size_t num = dims.nbDims == 0 ? 0 : 1; - for (int i = 0; i < dims.nbDims; i++) { - PADDLE_ENFORCE_GT(dims.d[i], 0); - num *= dims.d[i]; - } - return num; -} - // TensorRT data type to size const int kDataTypeSize[] = { 4, // kFLOAT @@ -46,13 +37,13 @@ const int kDataTypeSize[] = { // The following two API are implemented in TensorRT's header file, cannot load // from the dynamic library. So create our own implementation and directly // trigger the method from the dynamic library. -static nvinfer1::IBuilder* createInferBuilder(nvinfer1::ILogger& logger) { +static nvinfer1::IBuilder* createInferBuilder(nvinfer1::ILogger* logger) { return static_cast( - dy::createInferBuilder_INTERNAL(&logger, NV_TENSORRT_VERSION)); + dy::createInferBuilder_INTERNAL(logger, NV_TENSORRT_VERSION)); } -static nvinfer1::IRuntime* createInferRuntime(nvinfer1::ILogger& logger) { +static nvinfer1::IRuntime* createInferRuntime(nvinfer1::ILogger* logger) { return static_cast( - dy::createInferRuntime_INTERNAL(&logger, NV_TENSORRT_VERSION)); + dy::createInferRuntime_INTERNAL(logger, NV_TENSORRT_VERSION)); } // A logger for create TensorRT infer builder. @@ -80,7 +71,7 @@ class NaiveLogger : public nvinfer1::ILogger { return *x; } - virtual ~NaiveLogger() override {} + ~NaiveLogger() override {} }; } // namespace tensorrt diff --git a/paddle/fluid/inference/tensorrt/test_engine.cc b/paddle/fluid/inference/tensorrt/test_engine.cc index f3dbdf11f217ca35bc541b9bdb95bf0c06377fd3..e635f0f87d577a1f1ac74687ee60f762be525418 100644 --- a/paddle/fluid/inference/tensorrt/test_engine.cc +++ b/paddle/fluid/inference/tensorrt/test_engine.cc @@ -12,13 +12,12 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include "paddle/fluid/inference/tensorrt/engine.h" - #include #include #include #include +#include "paddle/fluid/inference/tensorrt/engine.h" #include "paddle/fluid/platform/enforce.h" namespace paddle { @@ -65,12 +64,12 @@ TEST_F(TensorRTEngineTest, add_layer) { // fill in real data float x_v = 1234; - engine_->SetInputFromCPU("x", (void*)&x_v, 1 * sizeof(float)); + engine_->SetInputFromCPU("x", reinterpret_cast(&x_v), + 1 * sizeof(float)); LOG(INFO) << "to execute"; engine_->Execute(1); LOG(INFO) << "to get output"; - // void* y_v = float y_cpu; engine_->GetOutputInCPU("y", &y_cpu, sizeof(float)); @@ -78,6 +77,37 @@ TEST_F(TensorRTEngineTest, add_layer) { ASSERT_EQ(y_cpu, x_v * 2 + 3); } +TEST_F(TensorRTEngineTest, add_layer_multi_dim) { + // Weight in CPU memory. + // It seems tensorrt FC use col-major: [[1.0, 3.3], [1.1, 4.4]] + // instead of row-major, which is [[1.0, 1.1], [3.3, 4.4]] + float raw_weight[4] = {1.0, 1.1, 3.3, 4.4}; + float raw_bias[2] = {1.3, 2.4}; + + TensorRTEngine::Weight weight(nvinfer1::DataType::kFLOAT, raw_weight, 4); + TensorRTEngine::Weight bias(nvinfer1::DataType::kFLOAT, raw_bias, 2); + auto* x = engine_->DeclareInput("x", nvinfer1::DataType::kFLOAT, + nvinfer1::DimsCHW{1, 2, 1}); + auto* fc_layer = TRT_ENGINE_ADD_LAYER(engine_, FullyConnected, *x, 2, + weight.get(), bias.get()); + PADDLE_ENFORCE(fc_layer != nullptr); + + engine_->DeclareOutput(fc_layer, 0, "y"); + engine_->FreezeNetwork(); + ASSERT_EQ(engine_->engine()->getNbBindings(), 2); + + float x_v[2] = {1.0, 2.0}; + engine_->SetInputFromCPU("x", reinterpret_cast(&x_v), + 2 * sizeof(float)); + engine_->Execute(1); + + LOG(INFO) << "to get output"; + float y_cpu[2] = {-1., -1.}; + engine_->GetOutputInCPU("y", &y_cpu[0], sizeof(float) * 2); + ASSERT_EQ(y_cpu[0], 4.5); + ASSERT_EQ(y_cpu[1], 14.5); +} + } // namespace tensorrt } // namespace inference } // namespace paddle diff --git a/paddle/fluid/inference/tensorrt/test_tensorrt.cc b/paddle/fluid/inference/tensorrt/test_tensorrt.cc index aed5b5e1a22cbed1256d4f28d0a8a4c29c6cc744..a07537985738ab0ad4092b794f3b62ba53dfa866 100644 --- a/paddle/fluid/inference/tensorrt/test_tensorrt.cc +++ b/paddle/fluid/inference/tensorrt/test_tensorrt.cc @@ -12,11 +12,11 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ +#include +#include #include #include #include "NvInfer.h" -#include "cuda.h" -#include "cuda_runtime_api.h" #include "paddle/fluid/platform/dynload/tensorrt.h" namespace dy = paddle::platform::dynload; @@ -43,7 +43,7 @@ class Logger : public nvinfer1::ILogger { class ScopedWeights { public: - ScopedWeights(float value) : value_(value) { + explicit ScopedWeights(float value) : value_(value) { w.type = nvinfer1::DataType::kFLOAT; w.values = &value_; w.count = 1; @@ -58,13 +58,13 @@ class ScopedWeights { // The following two API are implemented in TensorRT's header file, cannot load // from the dynamic library. So create our own implementation and directly // trigger the method from the dynamic library. -nvinfer1::IBuilder* createInferBuilder(nvinfer1::ILogger& logger) { +nvinfer1::IBuilder* createInferBuilder(nvinfer1::ILogger* logger) { return static_cast( - dy::createInferBuilder_INTERNAL(&logger, NV_TENSORRT_VERSION)); + dy::createInferBuilder_INTERNAL(logger, NV_TENSORRT_VERSION)); } -nvinfer1::IRuntime* createInferRuntime(nvinfer1::ILogger& logger) { +nvinfer1::IRuntime* createInferRuntime(nvinfer1::ILogger* logger) { return static_cast( - dy::createInferRuntime_INTERNAL(&logger, NV_TENSORRT_VERSION)); + dy::createInferRuntime_INTERNAL(logger, NV_TENSORRT_VERSION)); } const char* kInputTensor = "input"; @@ -74,7 +74,7 @@ const char* kOutputTensor = "output"; nvinfer1::IHostMemory* CreateNetwork() { Logger logger; // Create the engine. - nvinfer1::IBuilder* builder = createInferBuilder(logger); + nvinfer1::IBuilder* builder = createInferBuilder(&logger); ScopedWeights weights(2.); ScopedWeights bias(3.); @@ -103,9 +103,9 @@ nvinfer1::IHostMemory* CreateNetwork() { return model; } -void Execute(nvinfer1::IExecutionContext& context, const float* input, +void Execute(nvinfer1::IExecutionContext* context, const float* input, float* output) { - const nvinfer1::ICudaEngine& engine = context.getEngine(); + const nvinfer1::ICudaEngine& engine = context->getEngine(); // Two binds, input and output ASSERT_EQ(engine.getNbBindings(), 2); const int input_index = engine.getBindingIndex(kInputTensor); @@ -119,7 +119,7 @@ void Execute(nvinfer1::IExecutionContext& context, const float* input, // Copy the input to the GPU, execute the network, and copy the output back. ASSERT_EQ(0, cudaMemcpyAsync(buffers[input_index], input, sizeof(float), cudaMemcpyHostToDevice, stream)); - context.enqueue(1, buffers, stream, nullptr); + context->enqueue(1, buffers, stream, nullptr); ASSERT_EQ(0, cudaMemcpyAsync(output, buffers[output_index], sizeof(float), cudaMemcpyDeviceToHost, stream)); cudaStreamSynchronize(stream); @@ -136,7 +136,7 @@ TEST(TensorrtTest, BasicFunction) { // Use the model to create an engine and an execution context. Logger logger; - nvinfer1::IRuntime* runtime = createInferRuntime(logger); + nvinfer1::IRuntime* runtime = createInferRuntime(&logger); nvinfer1::ICudaEngine* engine = runtime->deserializeCudaEngine(model->data(), model->size(), nullptr); model->destroy(); @@ -145,7 +145,7 @@ TEST(TensorrtTest, BasicFunction) { // Execute the network. float input = 1234; float output; - Execute(*context, &input, &output); + Execute(context, &input, &output); EXPECT_EQ(output, input * 2 + 3); // Destroy the engine. diff --git a/paddle/fluid/inference/tests/book/CMakeLists.txt b/paddle/fluid/inference/tests/book/CMakeLists.txt index 97d9f03f88ad3e851a2dd4256d34e8ca76fdfb01..dbb81462b8273bd701e9c9f530eaf69817abd6a1 100644 --- a/paddle/fluid/inference/tests/book/CMakeLists.txt +++ b/paddle/fluid/inference/tests/book/CMakeLists.txt @@ -24,6 +24,11 @@ function(inference_test TARGET_NAME) endforeach() endfunction(inference_test) +#################### +# Inference tests here depend on fluid/tests/book. If users want to run +# individual test with ctest, they need to run tests in fluid/tests/book +# first to generate saved model. +#################### # This unittest is buggy! #inference_test(fit_a_line) inference_test(image_classification ARGS vgg resnet) @@ -31,5 +36,5 @@ inference_test(label_semantic_roles) inference_test(recognize_digits ARGS mlp conv) inference_test(recommender_system) #inference_test(rnn_encoder_decoder) -inference_test(understand_sentiment ARGS conv) +#inference_test(understand_sentiment ARGS conv) inference_test(word2vec) diff --git a/paddle/fluid/inference/tests/book/test_inference_image_classification.cc b/paddle/fluid/inference/tests/book/test_inference_image_classification.cc index 1e6555bb02033a28dedd2a1d1962981dfcc97cc2..987da18116cc6f4902bd66ae317f2470a8bc5057 100644 --- a/paddle/fluid/inference/tests/book/test_inference_image_classification.cc +++ b/paddle/fluid/inference/tests/book/test_inference_image_classification.cc @@ -17,8 +17,11 @@ limitations under the License. */ #include "paddle/fluid/inference/tests/test_helper.h" DEFINE_string(dirname, "", "Directory of the inference model."); +DEFINE_string(fp16_dirname, "", "Directory of the float16 inference model."); DEFINE_int32(batch_size, 1, "Batch size of input data"); DEFINE_int32(repeat, 1, "Running the inference program repeat times"); +DEFINE_bool(skip_cpu, false, "Skip the cpu test"); +DEFINE_bool(use_mkldnn, false, "Use MKLDNN to run inference"); TEST(inference, image_classification) { if (FLAGS_dirname.empty() || FLAGS_batch_size < 1 || FLAGS_repeat < 1) { @@ -32,23 +35,36 @@ TEST(inference, image_classification) { // 0. Call `paddle::framework::InitDevices()` initialize all the devices // In unittests, this is done in paddle/testing/paddle_gtest_main.cc + const bool is_combined = false; + std::vector> feed_target_shapes = + GetFeedTargetShapes(dirname, is_combined); + paddle::framework::LoDTensor input; // Use normilized image pixels as input data, // which should be in the range [0.0, 1.0]. - SetupTensor(&input, {FLAGS_batch_size, 3, 32, 32}, - static_cast(0), static_cast(1)); + feed_target_shapes[0][0] = FLAGS_batch_size; + paddle::framework::DDim input_dims = + paddle::framework::make_ddim(feed_target_shapes[0]); + LOG(INFO) << input_dims; + SetupTensor(&input, input_dims, static_cast(0), + static_cast(1)); std::vector cpu_feeds; cpu_feeds.push_back(&input); paddle::framework::LoDTensor output1; - std::vector cpu_fetchs1; - cpu_fetchs1.push_back(&output1); + if (!FLAGS_skip_cpu) { + std::vector cpu_fetchs1; + cpu_fetchs1.push_back(&output1); - // Run inference on CPU - LOG(INFO) << "--- CPU Runs: ---"; - TestInference( - dirname, cpu_feeds, cpu_fetchs1, FLAGS_repeat); - LOG(INFO) << output1.dims(); + // Run inference on CPU + LOG(INFO) << "--- CPU Runs: ---"; + LOG(INFO) << "Batch size is " << FLAGS_batch_size; + LOG(INFO) << "FLAGS_use_mkldnn: " << FLAGS_use_mkldnn; + TestInference( + dirname, cpu_feeds, cpu_fetchs1, FLAGS_repeat, is_combined, + FLAGS_use_mkldnn); + LOG(INFO) << output1.dims(); + } #ifdef PADDLE_WITH_CUDA paddle::framework::LoDTensor output2; @@ -57,10 +73,29 @@ TEST(inference, image_classification) { // Run inference on CUDA GPU LOG(INFO) << "--- GPU Runs: ---"; + LOG(INFO) << "Batch size is " << FLAGS_batch_size; TestInference( - dirname, cpu_feeds, cpu_fetchs2, FLAGS_repeat); + dirname, cpu_feeds, cpu_fetchs2, FLAGS_repeat, is_combined); LOG(INFO) << output2.dims(); - CheckError(output1, output2); + if (!FLAGS_skip_cpu) { + CheckError(output1, output2); + } + + // float16 inference requires cuda GPUs with >= 5.3 compute capability + if (!FLAGS_fp16_dirname.empty() && + paddle::platform::GetCUDAComputeCapability(0) >= 53) { + paddle::framework::LoDTensor output3; + std::vector cpu_fetchs3; + cpu_fetchs3.push_back(&output3); + + LOG(INFO) << "--- GPU Runs in float16 mode: ---"; + LOG(INFO) << "Batch size is " << FLAGS_batch_size; + + TestInference( + FLAGS_fp16_dirname, cpu_feeds, cpu_fetchs3, FLAGS_repeat); + + CheckError(output2, output3); + } #endif } diff --git a/paddle/fluid/inference/tests/test_helper.h b/paddle/fluid/inference/tests/test_helper.h index af2a7a5620487a10c1df6152fc4e4bf67b150752..01b8dc0be662da22fe15a79cd9abfe5fa92c9577 100644 --- a/paddle/fluid/inference/tests/test_helper.h +++ b/paddle/fluid/inference/tests/test_helper.h @@ -89,11 +89,68 @@ void CheckError(const paddle::framework::LoDTensor& output1, EXPECT_EQ(count, 0U) << "There are " << count << " different elements."; } +std::unique_ptr InitProgram( + paddle::framework::Executor* executor, paddle::framework::Scope* scope, + const std::string& dirname, const bool is_combined = false) { + std::unique_ptr inference_program; + if (is_combined) { + // All parameters are saved in a single file. + // Hard-coding the file names of program and parameters in unittest. + // The file names should be consistent with that used in Python API + // `fluid.io.save_inference_model`. + std::string prog_filename = "__model_combined__"; + std::string param_filename = "__params_combined__"; + inference_program = + paddle::inference::Load(executor, scope, dirname + "/" + prog_filename, + dirname + "/" + param_filename); + } else { + // Parameters are saved in separate files sited in the specified + // `dirname`. + inference_program = paddle::inference::Load(executor, scope, dirname); + } + return inference_program; +} + +std::vector> GetFeedTargetShapes( + const std::string& dirname, const bool is_combined = false) { + auto place = paddle::platform::CPUPlace(); + auto executor = paddle::framework::Executor(place); + auto* scope = new paddle::framework::Scope(); + + auto inference_program = InitProgram(&executor, scope, dirname, is_combined); + auto& global_block = inference_program->Block(0); + + const std::vector& feed_target_names = + inference_program->GetFeedTargetNames(); + std::vector> feed_target_shapes; + for (size_t i = 0; i < feed_target_names.size(); ++i) { + auto* var = global_block.FindVar(feed_target_names[i]); + std::vector var_shape = var->GetShape(); + feed_target_shapes.push_back(var_shape); + } + + delete scope; + return feed_target_shapes; +} + +void EnableMKLDNN( + const std::unique_ptr& program) { + for (size_t bid = 0; bid < program->Size(); ++bid) { + auto* block = program->MutableBlock(bid); + for (auto* op : block->AllOps()) { + if (op->HasAttr("use_mkldnn")) { + op->SetAttr("use_mkldnn", true); + } + } + } +} + template void TestInference(const std::string& dirname, const std::vector& cpu_feeds, const std::vector& cpu_fetchs, - const int repeat = 1, const bool is_combined = false) { + const int repeat = 1, const bool is_combined = false, + const bool use_mkldnn = false) { // 1. Define place, executor, scope auto place = Place(); auto executor = paddle::framework::Executor(place); @@ -105,7 +162,7 @@ void TestInference(const std::string& dirname, state = paddle::platform::ProfilerState::kCPU; } else { #ifdef PADDLE_WITH_CUDA - state = paddle::platform::ProfilerState::kCUDA; + state = paddle::platform::ProfilerState::kAll; // The default device_id of paddle::platform::CUDAPlace is 0. // Users can get the device_id using: // int device_id = place.GetDeviceId(); @@ -124,26 +181,14 @@ void TestInference(const std::string& dirname, paddle::platform::RecordEvent record_event( "init_program", paddle::platform::DeviceContextPool::Instance().Get(place)); - - if (is_combined) { - // All parameters are saved in a single file. - // Hard-coding the file names of program and parameters in unittest. - // The file names should be consistent with that used in Python API - // `fluid.io.save_inference_model`. - std::string prog_filename = "__model_combined__"; - std::string param_filename = "__params_combined__"; - inference_program = paddle::inference::Load( - &executor, scope, dirname + "/" + prog_filename, - dirname + "/" + param_filename); - } else { - // Parameters are saved in separate files sited in the specified - // `dirname`. - inference_program = paddle::inference::Load(&executor, scope, dirname); + inference_program = InitProgram(&executor, scope, dirname, is_combined); + if (use_mkldnn) { + EnableMKLDNN(inference_program); } } // Disable the profiler and print the timing information paddle::platform::DisableProfiler(paddle::platform::EventSortingKey::kDefault, - "load_program_profiler.txt"); + "load_program_profiler"); paddle::platform::ResetProfiler(); // 3. Get the feed_target_names and fetch_target_names @@ -179,10 +224,10 @@ void TestInference(const std::string& dirname, if (PrepareContext) { ctx = executor.Prepare(*inference_program, 0); executor.RunPreparedContext(ctx.get(), scope, &feed_targets, - &fetch_targets, CreateVars); + &fetch_targets, true, CreateVars); } else { executor.Run(*inference_program, scope, &feed_targets, &fetch_targets, - CreateVars); + true, CreateVars); } // Enable the profiler @@ -207,8 +252,7 @@ void TestInference(const std::string& dirname, // Disable the profiler and print the timing information paddle::platform::DisableProfiler( - paddle::platform::EventSortingKey::kDefault, - "run_inference_profiler.txt"); + paddle::platform::EventSortingKey::kDefault, "run_inference_profiler"); paddle::platform::ResetProfiler(); } diff --git a/paddle/fluid/inference/utils/singleton.h b/paddle/fluid/inference/utils/singleton.h new file mode 100644 index 0000000000000000000000000000000000000000..cfb89e704457a11a3cd6e89dba5efad5acae0bce --- /dev/null +++ b/paddle/fluid/inference/utils/singleton.h @@ -0,0 +1,80 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include +#include "paddle/fluid/platform/enforce.h" + +namespace paddle { +namespace inference { + +// NOTE not thread-safe. +template +struct Singleton { + static T& Global() { + static T* x = new T; + return *x; + } + + Singleton() = delete; + Singleton& operator=(const Singleton&) = delete; +}; + +/* + * An registor for any type. + * NOTE not thread-safe. + */ +template +struct Registry { + static Registry& Global() { + static auto* x = new Registry; + return *x; + } + + template + static void Register(const std::string& name) { + PADDLE_ENFORCE_EQ(items_.count(name), 0); + items_[name] = new ItemChild; + } + + static ItemParent* Lookup(const std::string& name, + const std::string& default_name = "") { + auto it = items_.find(name); + if (it == items_.end()) { + if (default_name == "") + return nullptr; + else + return items_.find(default_name)->second; + } + return it->second; + } + + ~Registry() { + for (auto& item : items_) { + delete item.second; + } + } + + private: + Registry() = default; + static std::unordered_map items_; +}; + +template +std::unordered_map Registry::items_; + +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/operators/CMakeLists.txt b/paddle/fluid/operators/CMakeLists.txt index 256aded8ca234a24229e11f27b9e3e25728ad293..de6ff29c6f8edbcf930546ff157a1c226e1311db 100644 --- a/paddle/fluid/operators/CMakeLists.txt +++ b/paddle/fluid/operators/CMakeLists.txt @@ -166,6 +166,10 @@ function(op_library TARGET) # NOTE(*): activation use macro to regist the kernels, set use_op manually. if(${TARGET} STREQUAL "activation") file(APPEND ${pybind_file} "USE_OP(relu);\n") + elseif(${TARGET} STREQUAL "reduce") + file(APPEND ${pybind_file} "USE_OP(reduce_sum);\n") + elseif(${TARGET} STREQUAL "fake_dequantize") + file(APPEND ${pybind_file} "USE_OP(fake_dequantize_max_abs);\n") else() file(APPEND ${pybind_file} "USE_OP(${TARGET});\n") endif() @@ -184,6 +188,7 @@ endif() add_subdirectory(detail) if(WITH_DISTRIBUTE) + set(DISTRIBUTE_DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib protobuf) set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor") op_library(send_op DEPS ${DISTRIBUTE_DEPS}) @@ -197,17 +202,34 @@ if(WITH_DISTRIBUTE) op_library(send_vars_op DEPS ${DISTRIBUTE_DEPS}) set_source_files_properties(send_vars_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) op_library(send_barrier_op DEPS ${DISTRIBUTE_DEPS}) + op_library(fetch_barrier_op DEPS ${DISTRIBUTE_DEPS}) set_source_files_properties(send_barrier_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) - set_source_files_properties(send_recv_op_test.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) - cc_test(test_send_recv SRCS send_recv_op_test.cc DEPS prefetch_op send_op listen_and_serv_op sum_op executor) + set_source_files_properties(fetch_barrier_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) + #set_source_files_properties(send_recv_op_test.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) + #cc_test(test_send_recv SRCS send_recv_op_test.cc DEPS prefetch_op send_op + # listen_and_serv_op sum_op executor SERIAL) + if(WITH_GPU) + set_source_files_properties(test_send_nccl_id.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) + cc_test(test_send_nccl_id SRCS test_send_nccl_id.cc DEPS send_op + listen_and_serv_op executor SERIAL) + op_library(gen_nccl_id_op DEPS nccl_common sendrecvop_grpc) + set_source_files_properties(gen_nccl_id_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) + else() + set(DEPS_OPS ${DEPS_OPS} gen_nccl_id_op) + endif() else() - set(DEPS_OPS ${DEPS_OPS} send_op prefetch_op recv_op listen_and_serv_op send_vars_op send_barrier_op) + set(DEPS_OPS ${DEPS_OPS} send_op prefetch_op recv_op listen_and_serv_op send_vars_op send_barrier_op fetch_barrier_op gen_nccl_id_op) endif() op_library(cross_entropy_op DEPS cross_entropy) op_library(softmax_with_cross_entropy_op DEPS cross_entropy softmax) op_library(softmax_op DEPS softmax) op_library(sequence_softmax_op DEPS softmax) +if (WITH_GPU AND TENSORRT_FOUND) + op_library(tensorrt_engine_op DEPS tensorrt_engine) +else() + set(DEPS_OPS ${DEPS_OPS} tensorrt_engine_op) +endif() op_library(sum_op DEPS selected_rows_functor) op_library(sgd_op DEPS selected_rows_functor) op_library(print_op DEPS lod_tensor) @@ -268,6 +290,11 @@ foreach(src ${READER_LIBRARY}) set(OP_LIBRARY ${src} ${OP_LIBRARY}) endforeach() +add_subdirectory(detection) +foreach(src ${DETECTION_LIBRARY}) + set(OP_LIBRARY ${src} ${OP_LIBRARY}) +endforeach() + set(GLOB_OP_LIB ${OP_LIBRARY} CACHE INTERNAL "Global OP library") cc_test(gather_test SRCS gather_test.cc DEPS tensor) diff --git a/paddle/fluid/operators/accuracy_op.cc b/paddle/fluid/operators/accuracy_op.cc index ac10d759fecb56635d1303fd383a5f9ea18f0a4d..42fcace17926641b5caf677eb3c8ba5222e37190 100644 --- a/paddle/fluid/operators/accuracy_op.cc +++ b/paddle/fluid/operators/accuracy_op.cc @@ -63,8 +63,7 @@ class AccuracyOp : public framework::OperatorWithKernel { class AccuracyOpMaker : public framework::OpProtoAndCheckerMaker { public: - AccuracyOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { // TODO(typhoonzero): support both inference value and indices. AddInput("Out", "The network output of topk (inferences)"); AddInput("Indices", "The the network output of topk (indices)"); diff --git a/paddle/fluid/operators/accuracy_op.cu b/paddle/fluid/operators/accuracy_op.cu index 630a4a2df2ca8f6afe81be3c455d255a0693fcc3..23b48c6fdf427348879de07c671c65327d6436d7 100644 --- a/paddle/fluid/operators/accuracy_op.cu +++ b/paddle/fluid/operators/accuracy_op.cu @@ -15,7 +15,7 @@ limitations under the License. */ #include #include #include "paddle/fluid/operators/accuracy_op.h" -#include "paddle/fluid/platform/cuda_helper.h" +#include "paddle/fluid/platform/cuda_primitives.h" #include "paddle/fluid/platform/gpu_info.h" namespace paddle { diff --git a/paddle/fluid/operators/activation_mkldnn_op.cc b/paddle/fluid/operators/activation_mkldnn_op.cc index ab7c61227114fe7a0ce2ff2515dd560706058b64..b892ac77d9ed60210ddadaecb1a4f214e5a25180 100644 --- a/paddle/fluid/operators/activation_mkldnn_op.cc +++ b/paddle/fluid/operators/activation_mkldnn_op.cc @@ -15,6 +15,7 @@ #include "mkldnn.hpp" #include "paddle/fluid/operators/activation_op.h" #include "paddle/fluid/operators/mkldnn_activation_op.h" +#include "paddle/fluid/platform/mkldnn_helper.h" namespace paddle { namespace operators { @@ -23,6 +24,18 @@ using paddle::framework::Tensor; using paddle::platform::MKLDNNDeviceContext; namespace { +std::string gethash(const mkldnn::memory::dims &operand_dims, + const mkldnn::algorithm algorithm) { + auto dim2str = [](const mkldnn::memory::dims &operand_dims) { + std::string dstr = ""; + for (size_t i = 0; i < operand_dims.size(); ++i) { + dstr += std::to_string(operand_dims[i]) + "-"; + } + return dstr; + }; + return dim2str(operand_dims) + std::to_string(algorithm); +} + template void eltwise_forward(const ExecContext &ctx, mkldnn::algorithm algorithm, const T alpha = 0, const T beta = 0) { @@ -37,42 +50,70 @@ void eltwise_forward(const ExecContext &ctx, mkldnn::algorithm algorithm, const auto *src_data = src->template data(); auto *dst = ctx.template Output("Out"); - const T *dst_data = dst->template mutable_data(ctx.GetPlace()); + T *dst_data = dst->template mutable_data(ctx.GetPlace()); // get memory dim PADDLE_ENFORCE(src->dims().size() == 2 || src->dims().size() == 4, "Input dim must be with 2 or 4"); std::vector src_tz = framework::vectorize2int(src->dims()); - // create memory description - auto data_md = src_tz.size() == 2 - ? platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32, - mkldnn::memory::format::nc) - : platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32, - mkldnn::memory::format::nchw); - - // create memory primitives - auto src_memory = - mkldnn::memory({data_md, mkldnn_engine}, - static_cast(const_cast(src_data))); - auto dst_memory = - mkldnn::memory({data_md, mkldnn_engine}, - static_cast(const_cast(dst_data))); - - auto forward_desc = mkldnn::eltwise_forward::desc( - mkldnn::prop_kind::forward_training, algorithm, data_md, alpha, beta); - - // save prim desc into global device context to be referred in backward path - const std::string key = ctx.op().Output("Out"); - const std::string key_eltwise_pd = key + "@eltwise_pd"; - auto forward_pd = std::make_shared( - forward_desc, mkldnn_engine); - dev_ctx.SetBlob(key_eltwise_pd, forward_pd); - - auto eltwise = mkldnn::eltwise_forward(*forward_pd, src_memory, dst_memory); + const std::string key = gethash(src_tz, algorithm); + const std::string key_src_data = + key + ctx.op().Output("Out") + "@eltwise_fwd_src_data"; + const std::string key_src_mem = key + "@eltwise_fwd_src_mem"; + const std::string key_dst_mem = key + "@eltwise_fwd_dst_mem"; + const std::string key_fwd = key + "@eltwise_fwd"; + + auto p_fwd = std::static_pointer_cast( + dev_ctx.GetBlob(key_fwd)); + + // save input data to be referred in backward path + auto p_src_data = std::make_shared(src_data); + dev_ctx.SetBlob(key_src_data, p_src_data); + + if (p_fwd == nullptr) { + // create memory description + auto data_md = src_tz.size() == 2 + ? platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32, + mkldnn::memory::format::nc) + : platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32, + mkldnn::memory::format::nchw); + + // create memory primitives + auto p_src_mem = std::make_shared(mkldnn::memory( + {data_md, mkldnn_engine}, platform::to_void_cast(src_data))); + dev_ctx.SetBlob(key_src_mem, p_src_mem); + + auto p_dst_mem = std::make_shared(mkldnn::memory( + {data_md, mkldnn_engine}, platform::to_void_cast(dst_data))); + dev_ctx.SetBlob(key_dst_mem, p_dst_mem); + + auto fwd_desc = mkldnn::eltwise_forward::desc( + mkldnn::prop_kind::forward_training, algorithm, data_md, alpha, beta); + auto p_fwd_pd = std::make_shared( + fwd_desc, mkldnn_engine); + const std::string key_fwd_pd = key + "eltwise_fwd_pd"; + dev_ctx.SetBlob(key_fwd_pd, p_fwd_pd); + p_fwd = std::make_shared( + *p_fwd_pd, *(p_src_mem.get()), *(p_dst_mem.get())); + dev_ctx.SetBlob(key_fwd, p_fwd); + } else { + // primitives already exist + auto p_src_mem = + std::static_pointer_cast(dev_ctx.GetBlob(key_src_mem)); + PADDLE_ENFORCE(p_src_mem != nullptr, + "Fail to find eltwise p_src_mem in device context."); + auto p_dst_mem = + std::static_pointer_cast(dev_ctx.GetBlob(key_dst_mem)); + PADDLE_ENFORCE(p_dst_mem != nullptr, + "Fail to find eltwise p_src_mem in device context."); + + p_src_mem->set_data_handle(platform::to_void_reinterpret_cast(src_data)); + p_dst_mem->set_data_handle(dst_data); + } // push primitive to stream and wait until it's executed - std::vector pipeline = {eltwise}; + std::vector pipeline = {*(p_fwd.get())}; mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); } @@ -83,8 +124,7 @@ void eltwise_grad(const ExecContext &ctx, mkldnn::algorithm algorithm, const auto &mkldnn_engine = dev_ctx.GetEngine(); // get buffers - const auto *x = ctx.template Input("X"); - const auto *src = x->template data(); + const auto *out = ctx.template Input("Out"); auto *dout = ctx.template Input(framework::GradVarName("Out")); const auto *diff_dst = dout->template data(); @@ -94,45 +134,73 @@ void eltwise_grad(const ExecContext &ctx, mkldnn::algorithm algorithm, const T *diff_src = dx->template mutable_data(ctx.GetPlace()); // get memory dim - std::vector src_tz = framework::vectorize2int(x->dims()); - - // create memory description - auto data_md = src_tz.size() == 2 - ? platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32, - mkldnn::memory::format::nc) - : platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32, - mkldnn::memory::format::nchw); - - // create memory primitives - auto src_memory = mkldnn::memory( - {data_md, mkldnn_engine}, static_cast(const_cast(src))); - auto diff_src_memory = - mkldnn::memory({data_md, mkldnn_engine}, - static_cast(const_cast(diff_src))); - auto diff_dst_memory = - mkldnn::memory({data_md, mkldnn_engine}, - static_cast(const_cast(diff_dst))); - - auto backward_desc = - mkldnn::eltwise_backward::desc(algorithm, data_md, data_md, alpha, beta); - - // retrieve eltwise primitive desc from device context - const std::string key = ctx.op().Input("Out"); - const std::string key_eltwise_pd = key + "@eltwise_pd"; - const std::shared_ptr forward_pd = dev_ctx.GetBlob(key_eltwise_pd); - PADDLE_ENFORCE(forward_pd != nullptr, - "Fail to find eltwise_pd in device context"); - auto *p_forward_pd = - static_cast(forward_pd.get()); - - auto eltwise_bwd_prim_desc = mkldnn::eltwise_backward::primitive_desc( - backward_desc, mkldnn_engine, *p_forward_pd); - - auto eltwise_bwd = mkldnn::eltwise_backward(eltwise_bwd_prim_desc, src_memory, - diff_dst_memory, diff_src_memory); + std::vector src_tz = framework::vectorize2int(out->dims()); + + const std::string key = gethash(src_tz, algorithm); + const std::string key_diff_src_mem = key + "@eltwise_diff_src_mem"; + const std::string key_diff_dst_mem = key + "@eltwise_diff_dst_mem"; + const std::string key_grad = key + "@eltwise_grad"; + + const std::string key_src_data = + key + ctx.op().Input("Out") + "@eltwise_fwd_src_data"; + const auto p_src_data = + std::static_pointer_cast(dev_ctx.GetBlob(key_src_data)); + + const std::string key_src_mem = key + "@eltwise_fwd_src_mem"; + auto p_src_mem = + std::static_pointer_cast(dev_ctx.GetBlob(key_src_mem)); + p_src_mem->set_data_handle(*p_src_data.get()); + + auto p_grad = std::static_pointer_cast( + dev_ctx.GetBlob(key_grad)); + + if (p_grad == nullptr) { + // create memory description + auto data_md = src_tz.size() == 2 + ? platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32, + mkldnn::memory::format::nc) + : platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32, + mkldnn::memory::format::nchw); + + // create memory primitives + std::shared_ptr p_diff_src_mem = + std::make_shared(mkldnn::memory( + {data_md, mkldnn_engine}, platform::to_void_cast(diff_src))); + dev_ctx.SetBlob(key_diff_src_mem, p_diff_src_mem); + std::shared_ptr p_diff_dst_mem = + std::make_shared(mkldnn::memory( + {data_md, mkldnn_engine}, platform::to_void_cast(diff_dst))); + dev_ctx.SetBlob(key_diff_dst_mem, p_diff_dst_mem); + + auto bwd_desc = mkldnn::eltwise_backward::desc(algorithm, data_md, data_md, + alpha, beta); + + const std::string key_fwd_pd = key + "eltwise_fwd_pd"; + auto *p_fwd_pd = static_cast( + dev_ctx.GetBlob(key_fwd_pd).get()); + + auto eltwise_bwd_prim_desc = mkldnn::eltwise_backward::primitive_desc( + bwd_desc, mkldnn_engine, *p_fwd_pd); + + p_grad = std::make_shared( + eltwise_bwd_prim_desc, *static_cast(p_src_mem.get()), + *(static_cast(p_diff_dst_mem.get())), + *(static_cast(p_diff_src_mem.get()))); + } else { + // primitives already exist + auto p_diff_src_mem = std::static_pointer_cast( + dev_ctx.GetBlob(key_diff_src_mem)); + auto p_diff_dst_mem = std::static_pointer_cast( + dev_ctx.GetBlob(key_diff_dst_mem)); + + p_diff_src_mem->set_data_handle( + platform::to_void_reinterpret_cast(diff_src)); + p_diff_dst_mem->set_data_handle( + platform::to_void_reinterpret_cast(diff_dst)); + } // push primitive to stream and wait until it's executed - std::vector pipeline = {eltwise_bwd}; + std::vector pipeline = {*(p_grad.get())}; mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); } } // anonymous namespace diff --git a/paddle/fluid/operators/activation_op.cc b/paddle/fluid/operators/activation_op.cc index 87ef55c50b0be46492a695928625d140345d415d..dd71c66a75a039429f6e4b1771bb31508bb6b56d 100644 --- a/paddle/fluid/operators/activation_op.cc +++ b/paddle/fluid/operators/activation_op.cc @@ -19,19 +19,18 @@ limitations under the License. */ namespace paddle { namespace operators { -#define REGISTER_ACTIVATION_OP_MAKER(OP_NAME, OP_COMMENT) \ - class OP_NAME##OpMaker \ - : public ::paddle::framework::OpProtoAndCheckerMaker { \ - public: \ - OP_NAME##OpMaker(OpProto *proto, OpAttrChecker *op_checker) \ - : ::paddle::framework::OpProtoAndCheckerMaker(proto, op_checker) { \ - AddInput("X", "Input of " #OP_NAME "operator"); \ - AddOutput("Out", "Output of" #OP_NAME "operator"); \ - AddAttr("use_mkldnn", \ - "(bool, default false) Only used in mkldnn kernel") \ - .SetDefault(false); \ - AddComment(#OP_COMMENT); \ - } \ +#define REGISTER_ACTIVATION_OP_MAKER(OP_NAME, OP_COMMENT) \ + class OP_NAME##OpMaker \ + : public ::paddle::framework::OpProtoAndCheckerMaker { \ + public: \ + void Make() override { \ + AddInput("X", "Input of " #OP_NAME "operator"); \ + AddOutput("Out", "Output of" #OP_NAME "operator"); \ + AddAttr("use_mkldnn", \ + "(bool, default false) Only used in mkldnn kernel") \ + .SetDefault(false); \ + AddComment(#OP_COMMENT); \ + } \ } #define REGISTER_ACTIVATION_OP_GRAD_MAKER(OP_NAME, KERNEL_TYPE) \ @@ -42,7 +41,7 @@ namespace operators { \ protected: \ std::unique_ptr<::paddle::framework::OpDesc> Apply() const override { \ - auto *op = new ::paddle::framework::OpDesc(); \ + auto* op = new ::paddle::framework::OpDesc(); \ op->SetType(#KERNEL_TYPE "_grad"); \ op->SetInput("Out", Output("Out")); \ op->SetInput(::paddle::framework::GradVarName("Out"), \ @@ -55,23 +54,50 @@ namespace operators { } \ } +framework::OpKernelType GetKernelType(const framework::ExecutionContext& ctx, + const framework::OperatorWithKernel& oper, + const std::string& name) { + framework::LibraryType library{framework::LibraryType::kPlain}; +#ifdef PADDLE_WITH_MKLDNN + auto it = oper.Attrs().find("use_mkldnn"); + if (library == framework::LibraryType::kPlain && it != oper.Attrs().end() && + platform::CanMKLDNNBeUsed(ctx)) { + library = framework::LibraryType::kMKLDNN; + } +#endif + framework::DataLayout layout = framework::DataLayout::kAnyLayout; + return framework::OpKernelType( + framework::ToDataType(ctx.Input(name)->type()), + ctx.GetPlace(), layout, library); +} + class ActivationOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; - void InferShape(framework::InferShapeContext *ctx) const override { + void InferShape(framework::InferShapeContext* ctx) const override { ctx->SetOutputDim("Out", ctx->GetInputDim("X")); ctx->ShareLoD("X", /*->*/ "Out"); } + + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + return GetKernelType(ctx, *this, "X"); + } }; class ActivationOpGrad : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; - void InferShape(framework::InferShapeContext *ctx) const override { + void InferShape(framework::InferShapeContext* ctx) const override { ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("Out")); } + + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + return GetKernelType(ctx, *this, "Out"); + } }; __attribute__((unused)) constexpr char SigmoidDoc[] = R"DOC( @@ -204,8 +230,7 @@ $$out = \frac{x}{1 + |x|}$$ class LeakyReluOpMaker : public framework::OpProtoAndCheckerMaker { public: - LeakyReluOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "Input of LeakyRelu operator"); AddOutput("Out", "Output of LeakyRelu operator"); AddAttr("alpha", "The small negative slope").SetDefault(0.02f); @@ -220,8 +245,7 @@ $out = \max(x, \alpha * x)$ class SoftShrinkOpMaker : public framework::OpProtoAndCheckerMaker { public: - SoftShrinkOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "Input of Softshrink operator"); AddOutput("Out", "Output of Softshrink operator"); AddAttr("lambda", "non-negative offset").SetDefault(0.5f); @@ -242,8 +266,7 @@ $$ class HardShrinkOpMaker : public framework::OpProtoAndCheckerMaker { public: - HardShrinkOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "Input of HardShrink operator"); AddOutput("Out", "Output of HardShrink operator"); AddAttr("threshold", "The value of threshold for HardShrink") @@ -265,8 +288,7 @@ $$ class BReluOpMaker : public framework::OpProtoAndCheckerMaker { public: - BReluOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "Input of BRelu operator"); AddOutput("Out", "Output of BRelu operator"); AddAttr("t_min", "The min marginal value of BRelu") @@ -284,8 +306,7 @@ $out = \max(\min(x, t_{min}), t_{max})$ class SoftReluOpMaker : public framework::OpProtoAndCheckerMaker { public: - SoftReluOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "Input of SoftRelu operator"); AddOutput("Out", "Output of SoftRelu operator"); AddAttr("threshold", "The threshold value of SoftRelu") @@ -301,8 +322,7 @@ $out = \ln(1 + \exp(\max(\min(x, threshold), threshold))$ class ELUOpMaker : public framework::OpProtoAndCheckerMaker { public: - ELUOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "Input of ELU operator"); AddOutput("Out", "Output of ELU operator"); AddAttr("alpha", "The alpha value of ELU").SetDefault(1.0f); @@ -320,8 +340,7 @@ $out = \max(0, x) + \min(0, \alpha * (e^x - 1))$ class Relu6OpMaker : public framework::OpProtoAndCheckerMaker { public: - Relu6OpMaker(OpProto *proto, OpAttrChecker *op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "Input of Relu6 operator"); AddOutput("Out", "Output of Relu6 operator"); AddAttr("threshold", "The threshold value of Relu6") @@ -337,8 +356,7 @@ $out = \min(\max(0, x), 6)$ class PowOpMaker : public framework::OpProtoAndCheckerMaker { public: - PowOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "Input of Pow operator"); AddOutput("Out", "Output of Pow operator"); AddAttr("factor", "The exponential factor of Pow").SetDefault(1.0f); @@ -353,8 +371,7 @@ $out = x^{factor}$ class STanhOpMaker : public framework::OpProtoAndCheckerMaker { public: - STanhOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "Input of STanh operator"); AddOutput("Out", "Output of STanh operator"); AddAttr("scale_a", "The scale parameter of a for the input") @@ -372,8 +389,7 @@ $$out = b * \frac{e^{a * x} - e^{-a * x}}{e^{a * x} + e^{-a * x}}$$ class ThresholdedReluOpMaker : public framework::OpProtoAndCheckerMaker { public: - ThresholdedReluOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "Input of ThresholdedRelu operator"); AddOutput("Out", "Output of ThresholdedRelu operator"); AddAttr("threshold", "The threshold location of activation") @@ -394,8 +410,7 @@ $$ class HardSigmoidOpMaker : public framework::OpProtoAndCheckerMaker { public: - HardSigmoidOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "Input of HardSigmoid operator"); AddOutput("Out", "Output of HardSigmoid operator"); AddAttr("slope", "Slope for linear approximation of sigmoid") @@ -420,8 +435,7 @@ It is recommended to use the defaults for this activation. class SwishOpMaker : public framework::OpProtoAndCheckerMaker { public: - SwishOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "Input of Swish operator"); AddOutput("Out", "Output of Swish operator"); AddAttr("beta", "Constant beta of swish operator").SetDefault(1.0f); diff --git a/paddle/fluid/operators/adadelta_op.cc b/paddle/fluid/operators/adadelta_op.cc index c9ed221a6e662e8c213fe1d34ff85a3f77483a3c..d1970515f58969948b1d2db5847e4344112f77f9 100644 --- a/paddle/fluid/operators/adadelta_op.cc +++ b/paddle/fluid/operators/adadelta_op.cc @@ -17,6 +17,7 @@ limitations under the License. */ namespace paddle { namespace operators { +using Tensor = framework::Tensor; class AdadeltaOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; @@ -55,12 +56,17 @@ class AdadeltaOp : public framework::OperatorWithKernel { ctx->SetOutputDim("AvgSquaredGradOut", param_dim); ctx->SetOutputDim("AvgSquaredUpdateOut", param_dim); } + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext &ctx) const override { + auto input_data_type = + framework::ToDataType(ctx.Input("Param")->type()); + return framework::OpKernelType(input_data_type, ctx.GetPlace()); + } }; class AdadeltaOpMaker : public framework::OpProtoAndCheckerMaker { public: - AdadeltaOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Param", "(Tensor) Input parameter"); AddInput("Grad", "(Tensor) Input gradient"); AddInput("AvgSquaredGrad", "(Tensor) Input average of squared gradient"); diff --git a/paddle/fluid/operators/adagrad_op.cc b/paddle/fluid/operators/adagrad_op.cc index 0153e1253b00ded21a7a14e37faf5a76d904d8d1..a3ef9ad9f91f1f626bd33876693ecc17ad76b96b 100644 --- a/paddle/fluid/operators/adagrad_op.cc +++ b/paddle/fluid/operators/adagrad_op.cc @@ -23,6 +23,7 @@ limitations under the License. */ namespace paddle { namespace operators { +using Tensor = framework::Tensor; class AdagradOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; @@ -56,12 +57,17 @@ class AdagradOp : public framework::OperatorWithKernel { ctx->SetOutputDim("ParamOut", param_dims); ctx->SetOutputDim("MomentOut", param_dims); } + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + auto input_data_type = + framework::ToDataType(ctx.Input("Param")->type()); + return framework::OpKernelType(input_data_type, ctx.GetPlace()); + } }; class AdagradOpMaker : public framework::OpProtoAndCheckerMaker { public: - AdagradOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Param", "(Tensor) Input parameter"); AddInput("Grad", "(Tensor) Input gradient"); AddInput("Moment", "(Tensor) Second moment"); diff --git a/paddle/fluid/operators/adagrad_op.cu b/paddle/fluid/operators/adagrad_op.cu index e798101ca6a3a44de749a2d2219295bd8911dfac..b25268786d622bc7a94117849763833e528bef48 100644 --- a/paddle/fluid/operators/adagrad_op.cu +++ b/paddle/fluid/operators/adagrad_op.cu @@ -16,7 +16,7 @@ limitations under the License. */ #include "paddle/fluid/operators/adagrad_op.h" #include "paddle/fluid/operators/math/math_function.h" #include "paddle/fluid/operators/math/selected_rows_functor.h" -#include "paddle/fluid/platform/cuda_helper.h" +#include "paddle/fluid/platform/cuda_primitives.h" namespace paddle { namespace operators { diff --git a/paddle/fluid/operators/adam_op.cc b/paddle/fluid/operators/adam_op.cc index 267dcab8104c337c8590180c8093098c756ab27d..99b0239855d6241b064a5883c2be3d58078b3b61 100644 --- a/paddle/fluid/operators/adam_op.cc +++ b/paddle/fluid/operators/adam_op.cc @@ -17,6 +17,7 @@ limitations under the License. */ namespace paddle { namespace operators { +using Tensor = framework::Tensor; class AdamOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; @@ -69,12 +70,17 @@ class AdamOp : public framework::OperatorWithKernel { ctx->SetOutputDim("Moment1Out", param_dims); ctx->SetOutputDim("Moment2Out", param_dims); } + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext &ctx) const override { + auto input_data_type = + framework::ToDataType(ctx.Input("Param")->type()); + return framework::OpKernelType(input_data_type, ctx.GetPlace()); + } }; class AdamOpMaker : public framework::OpProtoAndCheckerMaker { public: - AdamOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Param", "(Tensor) Input parameter"); AddInput("Grad", "(Tensor) Input gradient"); AddInput("LearningRate", "(Tensor) Learning rate"); diff --git a/paddle/fluid/operators/adam_op.h b/paddle/fluid/operators/adam_op.h index b332b6716369ca355454dc57fbd6cc2cbc71c658..f82ff47b52490c354f383515d430d14e24cbf6af 100644 --- a/paddle/fluid/operators/adam_op.h +++ b/paddle/fluid/operators/adam_op.h @@ -14,6 +14,7 @@ limitations under the License. */ #pragma once #include // for sqrt in CPU and CUDA +#include #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/operators/detail/safe_ref.h" #include "paddle/fluid/operators/math/selected_rows_functor.h" @@ -24,8 +25,14 @@ namespace operators { namespace scatter = paddle::operators::math::scatter; +struct GPUAdam; +struct CPUAdam; + +template +struct AdamFunctor; + template -struct AdamFunctor { +struct AdamFunctor { T beta1_; T beta2_; T epsilon_; @@ -71,6 +78,7 @@ struct AdamFunctor { // Calculation lr *= sqrt(1 - beta2_pow) / (1 - beta1_pow); + mom1 = beta1_ * mom1 + (1 - beta1_) * g; mom2 = beta2_ * mom2 + (1 - beta2_) * g * g; p -= lr * (mom1 / (sqrt(mom2) + epsilon_)); @@ -82,6 +90,71 @@ struct AdamFunctor { } }; +template +struct AdamFunctor { + T beta1_; + T beta2_; + T epsilon_; + + const T* beta1_pow_; + const T* beta2_pow_; + const T* moment1_; + T* moment1_out_; + const T* moment2_; + T* moment2_out_; + const T* lr_; + const T* grad_; + const T* param_; + T* param_out_; + + AdamFunctor(T beta1, T beta2, T epsilon, const T* beta1_pow, + const T* beta2_pow, const T* mom1, T* mom1_out, const T* mom2, + T* mom2_out, const T* lr, const T* grad, const T* param, + T* param_out) + : beta1_(beta1), + beta2_(beta2), + epsilon_(epsilon), + beta1_pow_(beta1_pow), + beta2_pow_(beta2_pow), + moment1_(mom1), + moment1_out_(mom1_out), + moment2_(mom2), + moment2_out_(mom2_out), + lr_(lr), + grad_(grad), + param_(param), + param_out_(param_out) {} + + void operator()(size_t numel) const { + Eigen::Map> g{ + grad_, static_cast(numel)}; + Eigen::Map> mom1{ + moment1_, static_cast(numel)}; + Eigen::Map> mom2{ + moment2_, static_cast(numel)}; + Eigen::Map> param{ + param_, static_cast(numel)}; + + Eigen::Map> param_out{ + param_out_, static_cast(numel)}; + Eigen::Map> moment1_out{ + moment1_out_, static_cast(numel)}; + Eigen::Map> moment2_out{ + moment2_out_, static_cast(numel)}; + + T lr = *lr_; + T beta1_pow = *beta1_pow_; + T beta2_pow = *beta2_pow_; + + // Calculation + lr *= sqrt(1 - beta2_pow) / (1 - beta1_pow); + + moment1_out = beta1_ * mom1 + (1 - beta1_) * g; + moment2_out = beta2_ * mom2 + (1 - beta2_) * g * g; + param_out = param - lr * (moment1_out / (moment2_out.sqrt() + epsilon_)); + } +}; + template struct SparseAdamFunctor { T beta1_; @@ -134,6 +207,7 @@ struct SparseAdamFunctor { T p = param_[rows_[i] * row_numel_ + j]; lr *= sqrt(1 - beta2_pow) / (1 - beta1_pow); + mom1 = beta1_ * mom1 + (1 - beta1_) * g; mom2 = beta2_ * mom2 + (1 - beta2_) * g * g; p -= lr * (mom1 / (sqrt(mom2) + epsilon_)); @@ -177,19 +251,34 @@ class AdamOpKernel : public framework::OpKernel { if (grad_var->IsType()) { auto& grad = Ref(ctx.Input("Grad"), "Must set Grad"); - AdamFunctor functor( - beta1, beta2, epsilon, beta1_pow.template data(), - beta2_pow.template data(), mom1.template data(), - mom1_out.template mutable_data(ctx.GetPlace()), - mom2.template data(), - mom2_out.template mutable_data(ctx.GetPlace()), - lr.template data(), grad.template data(), - param.template data(), - param_out.template mutable_data(ctx.GetPlace())); - platform::ForRange for_range( - static_cast(ctx.device_context()), - param.numel()); - for_range(functor); + + if (platform::is_cpu_place(ctx.GetPlace())) { + AdamFunctor functor( + beta1, beta2, epsilon, beta1_pow.template data(), + beta2_pow.template data(), mom1.template data(), + mom1_out.template mutable_data(ctx.GetPlace()), + mom2.template data(), + mom2_out.template mutable_data(ctx.GetPlace()), + lr.template data(), grad.template data(), + param.template data(), + param_out.template mutable_data(ctx.GetPlace())); + functor(param.numel()); + } else if (platform::is_gpu_place(ctx.GetPlace())) { + AdamFunctor functor( + beta1, beta2, epsilon, beta1_pow.template data(), + beta2_pow.template data(), mom1.template data(), + mom1_out.template mutable_data(ctx.GetPlace()), + mom2.template data(), + mom2_out.template mutable_data(ctx.GetPlace()), + lr.template data(), grad.template data(), + param.template data(), + param_out.template mutable_data(ctx.GetPlace())); + + platform::ForRange for_range( + static_cast(ctx.device_context()), + param.numel()); + for_range(functor); + } } else if (grad_var->IsType()) { auto& grad = Ref(ctx.Input("Grad"), "Must set Grad"); diff --git a/paddle/fluid/operators/adamax_op.cc b/paddle/fluid/operators/adamax_op.cc index 7e2f1cc66ebf8b7deebf55057a27129844129d5d..32062574bcf71ff96e451eaa6865b6bbfc3b1c80 100644 --- a/paddle/fluid/operators/adamax_op.cc +++ b/paddle/fluid/operators/adamax_op.cc @@ -17,6 +17,7 @@ limitations under the License. */ namespace paddle { namespace operators { +using Tensor = framework::Tensor; class AdamaxOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; @@ -63,12 +64,17 @@ class AdamaxOp : public framework::OperatorWithKernel { ctx->SetOutputDim("MomentOut", param_dims); ctx->SetOutputDim("InfNormOut", param_dims); } + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext &ctx) const override { + auto input_data_type = + framework::ToDataType(ctx.Input("Param")->type()); + return framework::OpKernelType(input_data_type, ctx.GetPlace()); + } }; class AdamaxOpMaker : public framework::OpProtoAndCheckerMaker { public: - AdamaxOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Param", "(Tensor) Input parameter"); AddInput("Grad", "(Tensor) Input gradient"); AddInput("LearningRate", "(Tensor) Learning rate"); diff --git a/paddle/fluid/operators/array_to_lod_tensor_op.cc b/paddle/fluid/operators/array_to_lod_tensor_op.cc index 5db2e4540ef170079328f24ac8d30f7b1901fa1e..149226e92d4d08a25c211bce686ff03c5d7ddf40 100644 --- a/paddle/fluid/operators/array_to_lod_tensor_op.cc +++ b/paddle/fluid/operators/array_to_lod_tensor_op.cc @@ -123,8 +123,7 @@ class ArrayToLoDTensorOp : public framework::OperatorBase { class ArrayToLoDTensorOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - ArrayToLoDTensorOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(std::vector) A vector of tensors that is going to " "be casted to a big LoDTensor."); diff --git a/paddle/fluid/operators/assign_op.cc b/paddle/fluid/operators/assign_op.cc index d372213e1b6008b0c4227103dd40730f86a84301..d9294048a9e89662958fd5c6af4fcbe5da3814c2 100644 --- a/paddle/fluid/operators/assign_op.cc +++ b/paddle/fluid/operators/assign_op.cc @@ -94,8 +94,7 @@ class AssignOp : public framework::OperatorBase { class AssignOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - AssignOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(LoDTensor, SelectedRows or LoDTensorArray) The input variable " "could be LoDTensor, SelectedRows or LoDTensorArray.") diff --git a/paddle/fluid/operators/assign_value_op.cc b/paddle/fluid/operators/assign_value_op.cc index 993610fdedde4bafd99f59a0adeeeef4526eb089..4ad6f3443db33fd14b67091d14fd877b951730ff 100644 --- a/paddle/fluid/operators/assign_value_op.cc +++ b/paddle/fluid/operators/assign_value_op.cc @@ -45,8 +45,7 @@ class AssignValueOp : public framework::OperatorWithKernel { class AssignValueOpMaker : public framework::OpProtoAndCheckerMaker { public: - AssignValueOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddOutput("Out", "(Tensor) Output tensor of assign_value operator."); AddAttr>("shape", "(vector) " diff --git a/paddle/fluid/operators/auc_op.cc b/paddle/fluid/operators/auc_op.cc index a168eaeab56128b75bbe97d7ccf843a081b5dced..c9871a9fe6b3b0d0cf671c2d155715f92c94fd8f 100644 --- a/paddle/fluid/operators/auc_op.cc +++ b/paddle/fluid/operators/auc_op.cc @@ -50,8 +50,7 @@ class AucOp : public framework::OperatorWithKernel { class AucOpMaker : public framework::OpProtoAndCheckerMaker { public: - AucOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Out", "A floating point 2D tensor, values are in the range [0, 1]." "Each row is sorted in descending order. This input should be the" diff --git a/paddle/fluid/operators/average_accumulates_op.cc b/paddle/fluid/operators/average_accumulates_op.cc index b21deaf9258567c05a8816b14ac7d6462964e8ba..25864e95d7e290c7f684501893e99c828c511979 100644 --- a/paddle/fluid/operators/average_accumulates_op.cc +++ b/paddle/fluid/operators/average_accumulates_op.cc @@ -111,8 +111,7 @@ class AverageAccumulatesOp : public framework::OperatorWithKernel { class AverageAccumulatesOpMaker : public framework::OpProtoAndCheckerMaker { public: - AverageAccumulatesOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("param", "(Tensor), The parameter to be accumulated."); AddInput("in_sum_1", "(Tensor), A tensor used to store the parameter " diff --git a/paddle/fluid/operators/batch_norm_mkldnn_op.cc b/paddle/fluid/operators/batch_norm_mkldnn_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..0e4a56d4a45a732cfcf43b09228bc0c44df5924c --- /dev/null +++ b/paddle/fluid/operators/batch_norm_mkldnn_op.cc @@ -0,0 +1,325 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "mkldnn.hpp" +#include "paddle/fluid/operators/batch_norm_op.h" +#include "paddle/fluid/platform/mkldnn_helper.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; +using paddle::platform::MKLDNNDeviceContext; +using paddle::platform::MKLDNNMemDesc; +using mkldnn::memory; + +template +using EigenArrayMap = + Eigen::Map>; +template +using ConstEigenArrayMap = + Eigen::Map>; +template +using EigenVectorArrayMap = Eigen::Map>; +template +using ConstEigenVectorArrayMap = + Eigen::Map>; + +namespace { +template +struct bn_type_traits { + using op_type = T; + using op_desc = typename op_type::desc; + using op_prim = typename op_type::primitive_desc; +}; + +template +void copy_to_weights(T scale_begin, T scale_end, T shift_begin, T shift_end, + Container *c) { + auto it = std::begin(*c); + + std::copy(scale_begin, scale_end, std::inserter(*c, it)); + std::copy( + shift_begin, shift_end, + std::inserter(*c, std::next(it, std::distance(scale_begin, scale_end)))); +} + +template +void run_batch_norm_op(Args &&... args) { + Op batch_norm_op{args...}; + + std::vector pipeline; + pipeline.push_back(batch_norm_op); + mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); +} + +template +inline void *cast_const_to_void(const T *t) { + return static_cast(const_cast(t)); +} +} // namespace + +template +class BatchNormMKLDNNOpKernel : public paddle::framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &ctx) const override { + auto data_layout_str = ctx.Attr("data_layout"); + auto data_layout = framework::StringToDataLayout(data_layout_str); + PADDLE_ENFORCE(data_layout == framework::DataLayout::kNCHW, + "MKLDNN batch normalization handles only NCHW data layout"); + + const float epsilon = ctx.Attr("epsilon"); + const float momentum = ctx.Attr("momentum"); + const bool is_test = ctx.Attr("is_test"); + + const auto *x = ctx.Input("X"); + const auto *mean = ctx.Input("Mean"); + const auto *variance = ctx.Input("Variance"); + + auto &dev_ctx = ctx.template device_context(); + auto mkldnn_engine = dev_ctx.GetEngine(); + + auto *y = ctx.Output("Y"); + auto *mean_out = ctx.Output("MeanOut"); + auto *variance_out = ctx.Output("VarianceOut"); + auto *batch_mean = ctx.Output("SavedMean"); + auto *batch_variance = ctx.Output("SavedVariance"); + + const auto *scale = ctx.Input("Scale"); + const auto *shift = ctx.Input("Bias"); + + y->mutable_data(ctx.GetPlace()); + mean_out->mutable_data(ctx.GetPlace()); + variance_out->mutable_data(ctx.GetPlace()); + + if (!is_test) { + batch_mean->mutable_data(ctx.GetPlace()); + batch_variance->mutable_data(ctx.GetPlace()); + } + + auto propagation = is_test == true ? mkldnn::prop_kind::forward_scoring + : mkldnn::prop_kind::forward_training; + + auto dims = paddle::framework::vectorize2int(x->dims()); + + auto src_md = + MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw); + auto dst_md = + MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw); + + auto src_pd = mkldnn::memory::primitive_desc{src_md, mkldnn_engine}; + auto dst_pd = mkldnn::memory::primitive_desc{dst_md, mkldnn_engine}; + + auto src = mkldnn::memory{src_pd, cast_const_to_void(x->data())}; + auto dst = mkldnn::memory{dst_pd, y->data()}; + + unsigned flags = mkldnn::use_scale_shift; + if (is_test) flags |= mkldnn::use_global_stats; + + using bn_fwd_types = bn_type_traits; + auto batch_norm_fwd_desc = + bn_fwd_types::op_desc{propagation, src_md, epsilon, flags}; + auto batch_norm_fwd_pd = + bn_fwd_types::op_prim{batch_norm_fwd_desc, mkldnn_engine}; + + const unsigned int ic = dims[1]; + + // MKLDNN requires a single piece of memory for scale and shift/bias data + const size_t scaleshift_size = 2 * ic; + std::vector scaleshift_data; + scaleshift_data.reserve(scaleshift_size); + + copy_to_weights(scale->data(), scale->data() + ic, shift->data(), + shift->data() + ic, &scaleshift_data); + + auto scaleshift_memory = mkldnn::memory{ + batch_norm_fwd_pd.weights_primitive_desc(), scaleshift_data.data()}; + + if (is_test) { + auto mean_memory = mkldnn::memory{batch_norm_fwd_pd.mean_primitive_desc(), + cast_const_to_void(mean->data())}; + + auto variance_memory = + mkldnn::memory{batch_norm_fwd_pd.variance_primitive_desc(), + cast_const_to_void(variance->data())}; + + run_batch_norm_op( + batch_norm_fwd_pd, src, (const mkldnn::primitive::at &)mean_memory, + (const mkldnn::primitive::at &)variance_memory, scaleshift_memory, + dst); + } else { + auto mean_memory = + mkldnn::memory{batch_norm_fwd_pd.mean_primitive_desc(), + cast_const_to_void(batch_mean->data())}; + + auto variance_memory = + mkldnn::memory{batch_norm_fwd_pd.variance_primitive_desc(), + cast_const_to_void(batch_variance->data())}; + + run_batch_norm_op(batch_norm_fwd_pd, src, + scaleshift_memory, dst, + mean_memory, variance_memory); + } + + if (!is_test) { + const unsigned int in = dims[0]; + const unsigned int sample_size = x->numel() / in / ic; + + // saved_xx is use just in this batch of data + EigenVectorArrayMap saved_mean_e( + batch_mean->mutable_data(ctx.GetPlace()), ic); + EigenVectorArrayMap saved_variance_e( + batch_variance->mutable_data(ctx.GetPlace()), ic); + saved_mean_e.setZero(); + saved_variance_e.setZero(); + + const unsigned int x_arr_size = in * ic; + ConstEigenArrayMap x_arr(x->data(), sample_size, x_arr_size); + for (unsigned int nc = 0; nc < x_arr_size; ++nc) { + saved_mean_e(nc % ic) += x_arr.col(nc).sum(); + } + saved_mean_e /= in * sample_size; + for (unsigned int nc = 0; nc < x_arr_size; ++nc) { + saved_variance_e(nc % ic) += + (x_arr.col(nc) - saved_mean_e(nc % ic)).matrix().squaredNorm(); + } + saved_variance_e /= in * sample_size; + + ConstEigenVectorArrayMap mean_arr{mean->data(), ic}; + ConstEigenVectorArrayMap variance_arr{variance->data(), ic}; + + EigenVectorArrayMap running_mean_arr( + mean_out->mutable_data(ctx.GetPlace()), ic); + EigenVectorArrayMap running_var_arr( + variance_out->mutable_data(ctx.GetPlace()), ic); + + auto one_minus_momentum = 1. - momentum; + running_mean_arr = + mean_arr * momentum + saved_mean_e * one_minus_momentum; + running_var_arr = + variance_arr * momentum + saved_variance_e * one_minus_momentum; + } + } +}; + +template +class BatchNormMKLDNNGradOpKernel : public paddle::framework::OpKernel { + public: + void Compute(const paddle::framework::ExecutionContext &ctx) const override { + auto data_layout_str = ctx.Attr("data_layout"); + auto data_layout = framework::StringToDataLayout(data_layout_str); + PADDLE_ENFORCE(data_layout == framework::DataLayout::kNCHW, + "MKLDNN batch normalization handles only NCHW data layout"); + + auto &dev_ctx = ctx.template device_context(); + auto mkldnn_engine = dev_ctx.GetEngine(); + + const float epsilon = ctx.Attr("epsilon"); + + const auto *x = ctx.Input("X"); + const auto *scale = ctx.Input("Scale"); + const auto *shift = ctx.Input("Bias"); + const auto *batch_mean = ctx.Input("SavedMean"); + const auto *batch_variance = ctx.Input("SavedVariance"); + + const auto *diff_y = ctx.Input(framework::GradVarName("Y")); + auto *diff_x = ctx.Output(framework::GradVarName("X")); + auto *diff_scale = ctx.Output(framework::GradVarName("Scale")); + auto *diff_shift = ctx.Output(framework::GradVarName("Bias")); + + diff_x->mutable_data(ctx.GetPlace()); + diff_scale->mutable_data(ctx.GetPlace()); + diff_shift->mutable_data(ctx.GetPlace()); + + auto dims = paddle::framework::vectorize2int(x->dims()); + unsigned flags = mkldnn::use_scale_shift | !mkldnn::use_global_stats; + + auto src_md = + MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw); + auto dst_md = + MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw); + auto diff_src_md = + MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw); + auto diff_dst_md = + MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw); + + using bn_bwd_types = bn_type_traits; + using bn_fwd_types = bn_type_traits; + + auto batch_norm_fwd_desc = bn_fwd_types::op_desc{ + mkldnn::prop_kind::forward_training, src_md, epsilon, flags}; + auto batch_norm_fwd_pd = + bn_fwd_types::op_prim{batch_norm_fwd_desc, mkldnn_engine}; + + auto batch_norm_bwd_desc = bn_bwd_types::op_desc{ + mkldnn::prop_kind::backward, diff_dst_md, dst_md, epsilon, flags}; + auto batch_norm_bwd_pd = bn_bwd_types::op_prim{ + batch_norm_bwd_desc, mkldnn_engine, batch_norm_fwd_pd}; + + auto src = mkldnn::memory{{src_md, mkldnn_engine}, + cast_const_to_void(x->data())}; + + auto mean = mkldnn::memory{batch_norm_bwd_pd.mean_primitive_desc(), + cast_const_to_void(batch_mean->data())}; + + auto variance = + mkldnn::memory{batch_norm_bwd_pd.variance_primitive_desc(), + cast_const_to_void(batch_variance->data())}; + + auto diff_dst = mkldnn::memory{{diff_dst_md, mkldnn_engine}, + cast_const_to_void(diff_y->data())}; + + const unsigned int ic = dims[1]; + + const size_t scaleshift_size = 2 * ic; + + std::vector scaleshift_data; + scaleshift_data.reserve(scaleshift_size); + copy_to_weights(scale->data(), scale->data() + ic, shift->data(), + shift->data() + ic, &scaleshift_data); + + auto scaleshift_memory = mkldnn::memory{ + batch_norm_bwd_pd.weights_primitive_desc(), scaleshift_data.data()}; + + std::vector diff_scaleshift_data; + diff_scaleshift_data.reserve(scaleshift_size); + copy_to_weights(diff_scale->data(), diff_scale->data() + ic, + diff_shift->data(), diff_shift->data() + ic, + &diff_scaleshift_data); + + auto diff_scaleshift_memory = + mkldnn::memory{batch_norm_bwd_pd.diff_weights_primitive_desc(), + diff_scaleshift_data.data()}; + + auto diff_src = mkldnn::memory{{diff_src_md, mkldnn_engine}, + static_cast(diff_x->data())}; + + run_batch_norm_op( + batch_norm_bwd_pd, src, mean, variance, diff_dst, scaleshift_memory, + diff_src, diff_scaleshift_memory); + + auto it = std::begin(diff_scaleshift_data); + std::copy(it, std::next(it, ic), diff_scale->data()); + std::copy(std::next(it, ic), std::end(diff_scaleshift_data), + diff_shift->data()); + } +}; +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP_KERNEL(batch_norm, MKLDNN, paddle::platform::CPUPlace, + ops::BatchNormMKLDNNOpKernel); +REGISTER_OP_KERNEL(batch_norm_grad, MKLDNN, paddle::platform::CPUPlace, + ops::BatchNormMKLDNNGradOpKernel); diff --git a/paddle/fluid/operators/batch_norm_op.cc b/paddle/fluid/operators/batch_norm_op.cc index c9939e8602ed341d37784ca292a55326899e8e65..6ec8c9d18b466142acdb46b0f46826a2aca7a47e 100644 --- a/paddle/fluid/operators/batch_norm_op.cc +++ b/paddle/fluid/operators/batch_norm_op.cc @@ -15,6 +15,9 @@ limitations under the License. */ #include "paddle/fluid/operators/batch_norm_op.h" #include #include "paddle/fluid/framework/data_layout.h" +#ifdef PADDLE_WITH_MKLDNN +#include "paddle/fluid/platform/mkldnn_helper.h" +#endif namespace paddle { namespace operators { @@ -87,9 +90,13 @@ class BatchNormOp : public framework::OperatorWithKernel { const framework::ExecutionContext &ctx) const override { auto input_data_type = framework::ToDataType(ctx.Input("X")->type()); - // For float or float16 input tensor, the type of the scale, bias, mean, - // and var tensors should both be float. + // By default, the type of the scale, bias, mean, + // and var tensors should both be float. (For float or float16 input tensor) + // or double (For double input tensor). auto bn_param_type = framework::proto::VarType::FP32; + if (input_data_type == framework::proto::VarType::FP64) { + bn_param_type = framework::proto::VarType::FP64; + } PADDLE_ENFORCE_EQ(bn_param_type, framework::ToDataType(ctx.Input("Scale")->type()), "Scale input should be of float type"); @@ -102,14 +109,24 @@ class BatchNormOp : public framework::OperatorWithKernel { PADDLE_ENFORCE_EQ(bn_param_type, framework::ToDataType( ctx.Input("Variance")->type()), "Variance input should be of float type"); - return framework::OpKernelType(input_data_type, ctx.GetPlace()); + + framework::LibraryType library_{framework::LibraryType::kPlain}; +#ifdef PADDLE_WITH_MKLDNN + if (library_ == framework::LibraryType::kPlain && + platform::CanMKLDNNBeUsed(ctx)) { + library_ = framework::LibraryType::kMKLDNN; + } +#endif + // TODO(pzelazko-intel): enable MKLDNN layout when it's ready + framework::DataLayout layout = framework::DataLayout::kAnyLayout; + return framework::OpKernelType(input_data_type, ctx.GetPlace(), layout, + library_); } }; class BatchNormOpMaker : public framework::OpProtoAndCheckerMaker { public: - BatchNormOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddAttr("is_test", "").SetDefault(false); AddAttr("momentum", "").SetDefault(0.9); AddAttr("epsilon", "") @@ -147,6 +164,9 @@ class BatchNormOpMaker : public framework::OpProtoAndCheckerMaker { "Variance of the current mini batch, " "will apply to output when training") .AsIntermediate(); + AddAttr("use_mkldnn", + "(bool, default false) Only used in mkldnn kernel") + .SetDefault(false); AddComment(R"DOC( Batch Normalization. @@ -345,8 +365,19 @@ class BatchNormGradOp : public framework::OperatorWithKernel { if (t == nullptr) { PADDLE_THROW("can't find Y@GRAD"); } - return framework::OpKernelType(framework::ToDataType(t->type()), - ctx.GetPlace()); + + framework::LibraryType library_{framework::LibraryType::kPlain}; +#ifdef PADDLE_WITH_MKLDNN + if (library_ == framework::LibraryType::kPlain && + platform::CanMKLDNNBeUsed(ctx)) { + library_ = framework::LibraryType::kMKLDNN; + } +#endif + // TODO(pzelazko-intel): enable MKLDNN layout when it's ready + framework::DataLayout layout = framework::DataLayout::kAnyLayout; + return framework::OpKernelType( + framework::ToDataType(ctx.Input("X")->type()), ctx.GetPlace(), + layout, library_); } }; @@ -470,6 +501,7 @@ class BatchNormGradMaker : public framework::SingleGradOpDescMaker { op->SetInput(framework::GradVarName("Y"), OutputGrad("Y")); op->SetInput("Scale", Input("Scale")); + op->SetInput("Bias", Input("Bias")); op->SetInput("SavedMean", Output("SavedMean")); op->SetInput("SavedVariance", Output("SavedVariance")); @@ -492,8 +524,9 @@ REGISTER_OPERATOR(batch_norm, ops::BatchNormOp, ops::BatchNormOpMaker, REGISTER_OPERATOR(batch_norm_grad, ops::BatchNormGradOp); REGISTER_OP_CPU_KERNEL( - batch_norm, - ops::BatchNormKernel); + batch_norm, ops::BatchNormKernel, + ops::BatchNormKernel); REGISTER_OP_CPU_KERNEL( batch_norm_grad, - ops::BatchNormGradKernel); + ops::BatchNormGradKernel, + ops::BatchNormGradKernel); diff --git a/paddle/fluid/operators/batch_norm_op.cu.cc b/paddle/fluid/operators/batch_norm_op.cu.cc index cb1927bc0f2eb735f0a3184df5f0f8fada2f9dca..550dd32d36767f90e880415bfffaf01aeb623609 100644 --- a/paddle/fluid/operators/batch_norm_op.cu.cc +++ b/paddle/fluid/operators/batch_norm_op.cu.cc @@ -287,6 +287,8 @@ namespace ops = paddle::operators; namespace plat = paddle::platform; REGISTER_OP_CUDA_KERNEL( batch_norm, ops::BatchNormKernel, + ops::BatchNormKernel, ops::BatchNormKernel); REGISTER_OP_CUDA_KERNEL( - batch_norm_grad, ops::BatchNormGradKernel); + batch_norm_grad, ops::BatchNormGradKernel, + ops::BatchNormGradKernel); diff --git a/paddle/fluid/operators/batch_size_like.h b/paddle/fluid/operators/batch_size_like.h index dd51a11fbe6ad5e528197b67536518c4b31fa355..483c9f8c2191fa4eb98b91112f9d6753e2fbddc3 100644 --- a/paddle/fluid/operators/batch_size_like.h +++ b/paddle/fluid/operators/batch_size_like.h @@ -53,8 +53,7 @@ class BatchSizeLikeOp : public framework::OperatorWithKernel { class BatchSizeLikeOpMaker : public framework::OpProtoAndCheckerMaker { public: - BatchSizeLikeOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() final { AddInput("Input", "(Tensor) Tensor " "whose input_dim_idx'th dimension specifies the batch_size"); @@ -68,7 +67,11 @@ class BatchSizeLikeOpMaker : public framework::OpProtoAndCheckerMaker { AddAttr("output_dim_idx", "(int, default 0) The index of output's batch size dimension") .SetDefault(0); + Apply(); } + + protected: + virtual void Apply() = 0; }; } // namespace operators diff --git a/paddle/fluid/operators/beam_search_decode_op.cc b/paddle/fluid/operators/beam_search_decode_op.cc index 4a8dfd4b54227070c2143b180f8ab92753885550..c3dd22119ddab8ecf9213ee274e4cbd4f05e78fd 100644 --- a/paddle/fluid/operators/beam_search_decode_op.cc +++ b/paddle/fluid/operators/beam_search_decode_op.cc @@ -23,16 +23,54 @@ struct BeamSearchDecodeFunctor { BeamSearchDecodeFunctor(const LoDTensorArray& step_ids, const LoDTensorArray& step_scores, LoDTensor* id_tensor, LoDTensor* score_tensor) - : step_ids_(step_ids), - step_scores_(step_scores), + : step_ids_origin_(step_ids), + step_scores_origin_(step_scores), id_tensor_(id_tensor), - score_tensor_(score_tensor) {} + score_tensor_(score_tensor) { + tensor_on_gpu_ = false; + // First make a copy of GPU data on CPU + if (platform::is_gpu_place(step_ids_origin_[0].place())) { + tensor_on_gpu_ = true; + platform::DeviceContextPool& pool = + platform::DeviceContextPool::Instance(); + auto* dev_ctx = pool.Get(step_ids_origin_[0].place()); + // Copy all tensors in the input tensor array + for (auto& step_id : step_ids_origin_) { + framework::LoDTensor out; + dev_ctx->Wait(); + framework::TensorCopy(step_id, platform::CPUPlace(), *dev_ctx, &out); + dev_ctx->Wait(); + + out.set_lod(step_id.lod()); + step_ids_.push_back(out); + } + } + if (platform::is_gpu_place(step_scores_origin_[0].place())) { + tensor_on_gpu_ = true; + platform::DeviceContextPool& pool = + platform::DeviceContextPool::Instance(); + auto* dev_ctx = pool.Get(step_scores_origin_[0].place()); + // Copy all tensors in the input tensor array + for (auto& step_score : step_scores_origin_) { + framework::LoDTensor out; + dev_ctx->Wait(); + framework::TensorCopy(step_score, platform::CPUPlace(), *dev_ctx, &out); + dev_ctx->Wait(); + + out.set_lod(step_score.lod()); + step_scores_.push_back(out); + } + } + } template void operator()() const; - const LoDTensorArray& step_ids_; - const LoDTensorArray& step_scores_; + bool tensor_on_gpu_; + const LoDTensorArray& step_ids_origin_; + const LoDTensorArray& step_scores_origin_; + LoDTensorArray step_ids_ = LoDTensorArray(); + LoDTensorArray step_scores_ = LoDTensorArray(); LoDTensor* id_tensor_; LoDTensor* score_tensor_; }; @@ -40,8 +78,14 @@ struct BeamSearchDecodeFunctor { template void BeamSearchDecodeFunctor::operator()() const { BeamSearchDecoder beam_search_decoder; - beam_search_decoder.PackAllSteps(step_ids_, step_scores_, id_tensor_, - score_tensor_); + // Check if the tensor is on GPU. If so, use the CPU copy instead + if (tensor_on_gpu_) { + beam_search_decoder.PackAllSteps(step_ids_, step_scores_, id_tensor_, + score_tensor_); + } else { + beam_search_decoder.PackAllSteps(step_ids_origin_, step_scores_origin_, + id_tensor_, score_tensor_); + } } template <> @@ -90,8 +134,7 @@ class BeamSearchDecodeOp : public framework::OperatorBase { class BeamSearchDecodeOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - BeamSearchDecodeOpProtoMaker(OpProto* proto, OpAttrChecker* op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Ids", "(LodTensorArray)" "score of the candidate words in each step"); diff --git a/paddle/fluid/operators/beam_search_op.cc b/paddle/fluid/operators/beam_search_op.cc index fdab4e92f47c7c8f241d93268a73dcb8c2eb2dc6..df0b50881f4e3ec6f57bdb2b63033931059c486e 100644 --- a/paddle/fluid/operators/beam_search_op.cc +++ b/paddle/fluid/operators/beam_search_op.cc @@ -195,11 +195,9 @@ std::string ItemToString(const BeamSearch::Item &item) { return stream.str(); } -class BeamSearchProtoAndCheckerMaker - : public framework::OpProtoAndCheckerMaker { +class BeamSearchOpMaker : public framework::OpProtoAndCheckerMaker { public: - BeamSearchProtoAndCheckerMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { // inputs and outputs stored in proto AddInput("pre_ids", "ids in previous step"); AddInput("ids", "a LoDTensor of shape of [None,k]"); @@ -222,20 +220,32 @@ class BeamSearchProtoAndCheckerMaker } }; -class BeamSearchInferShape : public framework::InferShapeBase { +class BeamSearchOp : public framework::OperatorWithKernel { public: - void operator()(framework::InferShapeContext *context) const override { + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContext *ctx) const override { for (const std::string &arg : std::vector({"pre_ids", "ids", "scores"})) { - PADDLE_ENFORCE(context->HasInput(arg), - "BeamSearch need input argument '%s'", arg); + PADDLE_ENFORCE(ctx->HasInput(arg), "BeamSearch need input argument '%s'", + arg); } for (const std::string &arg : std::vector({"selected_ids", "selected_scores"})) { - PADDLE_ENFORCE(context->HasOutput(arg), + PADDLE_ENFORCE(ctx->HasOutput(arg), "BeamSearch need output argument '%s'", arg); } } + + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext &ctx) const override { + framework::OpKernelType kt = framework::OpKernelType( + framework::ToDataType( + ctx.Input("pre_ids")->type()), + platform::CPUPlace()); + return kt; + } }; class BeamSearchInferVarType : public framework::VarTypeInference { @@ -254,8 +264,13 @@ class BeamSearchInferVarType : public framework::VarTypeInference { } // namespace operators } // namespace paddle -REGISTER_OPERATOR(beam_search, paddle::operators::BeamSearchOp, - paddle::operators::BeamSearchProtoAndCheckerMaker, - paddle::operators::BeamSearchInferShape, - paddle::operators::BeamSearchInferVarType, - paddle::framework::EmptyGradOpMaker); +namespace ops = paddle::operators; + +REGISTER_OPERATOR(beam_search, ops::BeamSearchOp, ops::BeamSearchOpMaker, + ops::BeamSearchInferVarType); +REGISTER_OP_CPU_KERNEL( + beam_search, + ops::BeamSearchOpKernel, + ops::BeamSearchOpKernel, + ops::BeamSearchOpKernel, + ops::BeamSearchOpKernel); diff --git a/paddle/fluid/operators/beam_search_op.h b/paddle/fluid/operators/beam_search_op.h index 0a481a85ce6fbb582b8c0e12710455aaaac72aa1..46bc4f6f936929050276e8b3b93f1eddd62ac638 100644 --- a/paddle/fluid/operators/beam_search_op.h +++ b/paddle/fluid/operators/beam_search_op.h @@ -14,10 +14,6 @@ limitations under the License. */ #pragma once -#ifdef PADDLE_WITH_TESTING -#include "gtest/gtest.h" -#endif - #include #include #include "paddle/fluid/framework/lod_tensor.h" @@ -192,49 +188,29 @@ std::ostream& operator<<(std::ostream& os, const BeamSearch::Item& item); std::string ItemToString(const BeamSearch::Item& item); -class BeamSearchOp : public framework::OperatorBase { +template +class BeamSearchOpKernel : public framework::OpKernel { public: - BeamSearchOp(const std::string& type, - const framework::VariableNameMap& inputs, - const framework::VariableNameMap& outputs, - const framework::AttributeMap& attrs) - : OperatorBase(type, inputs, outputs, attrs) {} - - BeamSearchOp(const BeamSearchOp& o) - : framework::OperatorBase( - static_cast(o)) { - PADDLE_THROW("Not Implemented"); - } - - private: - void RunImpl(const framework::Scope& scope, - const platform::Place& dev_place) const override { - auto ids_var = scope.FindVar(Input("ids")); - auto scores_var = scope.FindVar(Input("scores")); - auto pre_ids_var = scope.FindVar(Input("pre_ids")); + void Compute(const framework::ExecutionContext& context) const override { + auto* ids_var = context.Input("ids"); + auto* scores_var = context.Input("scores"); + auto* pre_ids_var = context.Input("pre_ids"); PADDLE_ENFORCE_NOT_NULL(ids_var); PADDLE_ENFORCE_NOT_NULL(scores_var); PADDLE_ENFORCE_NOT_NULL(pre_ids_var); - auto& ids = ids_var->Get(); - auto& scores = scores_var->Get(); - auto& pre_ids = pre_ids_var->Get(); - size_t level = Attr("level"); - size_t beam_size = Attr("beam_size"); - int end_id = Attr("end_id"); - BeamSearch alg(ids, scores, level, beam_size, end_id); - - auto selected_ids_var = scope.FindVar(Output("selected_ids")); - auto selected_scores_var = scope.FindVar(Output("selected_scores")); + size_t level = context.Attr("level"); + size_t beam_size = context.Attr("beam_size"); + int end_id = context.Attr("end_id"); + BeamSearch alg(*ids_var, *scores_var, level, beam_size, end_id); + auto selected_ids_var = + context.Output("selected_ids"); + auto selected_scores_var = + context.Output("selected_scores"); PADDLE_ENFORCE_NOT_NULL(selected_ids_var); PADDLE_ENFORCE_NOT_NULL(selected_scores_var); - auto& selected_ids_tensor = - *selected_ids_var->GetMutable(); - auto& selected_scores_tensor = - *selected_scores_var->GetMutable(); - alg(pre_ids, &selected_ids_tensor, &selected_scores_tensor); + alg(*pre_ids_var, selected_ids_var, selected_scores_var); } }; - } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/bilinear_interp_op.cc b/paddle/fluid/operators/bilinear_interp_op.cc index 69f79bf93be8ac7df9cab43b84cf755f2f3dfeaa..d46fda54e7a9d5bc737a7ec2116daca33ffa015f 100644 --- a/paddle/fluid/operators/bilinear_interp_op.cc +++ b/paddle/fluid/operators/bilinear_interp_op.cc @@ -41,8 +41,7 @@ class BilinearInterpOp : public framework::OperatorWithKernel { class BilinearInterpOpMaker : public framework::OpProtoAndCheckerMaker { public: - BilinearInterpOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor) The input tensor of bilinear interpolation, " "This is a 4-D tensor with shape of (N x C x h x w)"); diff --git a/paddle/fluid/operators/bilinear_interp_op.cu b/paddle/fluid/operators/bilinear_interp_op.cu index 82eb9e83bd84e6ec6881facbb2fac0aebce93d55..510190f1aaf02960284216a1bedd409011088499 100644 --- a/paddle/fluid/operators/bilinear_interp_op.cu +++ b/paddle/fluid/operators/bilinear_interp_op.cu @@ -10,7 +10,7 @@ limitations under the License. */ #include "paddle/fluid/operators/bilinear_interp_op.h" -#include "paddle/fluid/platform/cuda_helper.h" +#include "paddle/fluid/platform/cuda_primitives.h" namespace paddle { namespace operators { diff --git a/paddle/fluid/operators/bilinear_tensor_product_op.cc b/paddle/fluid/operators/bilinear_tensor_product_op.cc index e910ad92d1051aa89fdb3290a977ff376378a227..8d261a118a75ee16027faf60341cefd30c3cdbba 100644 --- a/paddle/fluid/operators/bilinear_tensor_product_op.cc +++ b/paddle/fluid/operators/bilinear_tensor_product_op.cc @@ -65,8 +65,7 @@ class BilinearTensorProductOp : public framework::OperatorWithKernel { class BilinearTensorProductOpMaker : public framework::OpProtoAndCheckerMaker { public: - BilinearTensorProductOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The first input of bilinear_tensor_product operator."); AddInput("Y", "The second input of bilinear_tensor_product operator."); AddInput("Weight", diff --git a/paddle/fluid/operators/bilinear_tensor_product_op.h b/paddle/fluid/operators/bilinear_tensor_product_op.h index ca80e6085c4f8b242cd26803202475cd50474bcd..f23336f7b98d6d71d155373cff3515a8463aecbe 100644 --- a/paddle/fluid/operators/bilinear_tensor_product_op.h +++ b/paddle/fluid/operators/bilinear_tensor_product_op.h @@ -16,7 +16,7 @@ limitations under the License. */ #include "paddle/fluid/framework/eigen.h" #include "paddle/fluid/framework/op_registry.h" -#include "paddle/fluid/operators/math/math_function.h" +#include "paddle/fluid/operators/math/blas.h" namespace paddle { namespace operators { @@ -61,9 +61,9 @@ class BilinearTensorProductKernel : public framework::OpKernel { auto output_col_vec = output_mat.chip(i, 1); Tensor weight_mat = weight->Slice(i, i + 1).Resize(framework::make_ddim({x_dim, y_dim})); - math::gemm(dev_ctx, CblasNoTrans, CblasNoTrans, - batch_size, y_dim, x_dim, 1, x->data(), - weight_mat.data(), 0, left_mul.data()); + math::GetBlas(dev_ctx).GEMM( + CblasNoTrans, CblasNoTrans, batch_size, y_dim, x_dim, 1, x->data(), + weight_mat.data(), 0, left_mul.data()); output_col_vec.device(place) = (left_mul_mat * y_mat).sum(Eigen::DSizes(1)); } @@ -125,6 +125,8 @@ class BilinearTensorProductGradKernel : public framework::OpKernel { set_zero(dev_ctx, d_y, static_cast(0)); } + auto blas = math::GetBlas(ctx); + // Caculate the Output(X@Grad) and Output(Y@Grad). if (d_x || d_y) { Eigen::DSizes bcast_for_x(1, y_dim); @@ -138,18 +140,16 @@ class BilinearTensorProductGradKernel : public framework::OpKernel { output_vec.reshape(Eigen::DSizes(batch_size, 1)) .broadcast(bcast_for_x) * y_mat; - math::gemm( - dev_ctx, CblasNoTrans, CblasTrans, batch_size, x_dim, y_dim, 1, - y_scale.data(), weight_i.data(), 1, d_x->data()); + blas.GEMM(CblasNoTrans, CblasTrans, batch_size, x_dim, y_dim, 1, + y_scale.data(), weight_i.data(), 1, d_x->data()); } if (d_y) { x_scale_mat.device(place) = output_vec.reshape(Eigen::DSizes(batch_size, 1)) .broadcast(bcast_for_y) * x_mat; - math::gemm( - dev_ctx, CblasNoTrans, CblasNoTrans, batch_size, y_dim, x_dim, 1, - x_scale.data(), weight_i.data(), 1, d_y->data()); + blas.GEMM(CblasNoTrans, CblasNoTrans, batch_size, y_dim, x_dim, 1, + x_scale.data(), weight_i.data(), 1, d_y->data()); } } } @@ -166,9 +166,8 @@ class BilinearTensorProductGradKernel : public framework::OpKernel { output_vec.reshape(Eigen::DSizes(batch_size, 1)) .broadcast(bcast_for_weight) * x_mat; - math::gemm(dev_ctx, CblasTrans, CblasNoTrans, x_dim, - y_dim, batch_size, 1, x_scale.data(), - y->data(), 0, d_weight_i.data()); + blas.GEMM(CblasTrans, CblasNoTrans, x_dim, y_dim, batch_size, 1, + x_scale.data(), y->data(), 0, d_weight_i.data()); } } diff --git a/paddle/fluid/operators/cast_op.cc b/paddle/fluid/operators/cast_op.cc index dd0068d571f72c9c22334e523cd091fe4c8da5a6..8d6a498dc941e44688ec8a2b49a6e080608f9b85 100644 --- a/paddle/fluid/operators/cast_op.cc +++ b/paddle/fluid/operators/cast_op.cc @@ -21,8 +21,7 @@ namespace operators { class CastOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - CastOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The input tensor of cast op"); AddOutput("Out", "The output tensor of cast op"); AddAttr("out_dtype", "output data type"); @@ -90,4 +89,5 @@ REGISTER_OP_CPU_KERNEL(cast, ops::CastOpKernel, ops::CastOpKernel, ops::CastOpKernel, ops::CastOpKernel, + ops::CastOpKernel, ops::CastOpKernel); diff --git a/paddle/fluid/operators/cast_op.cu b/paddle/fluid/operators/cast_op.cu index c486c5850e25fcf4370f02cb145c244743a4cc4b..657d162878c108760585ca9bd58e2fd34bf1fef3 100644 --- a/paddle/fluid/operators/cast_op.cu +++ b/paddle/fluid/operators/cast_op.cu @@ -21,5 +21,5 @@ using CastOpKernel = REGISTER_OP_CUDA_KERNEL(cast, CastOpKernel, CastOpKernel, CastOpKernel, CastOpKernel, - CastOpKernel, + CastOpKernel, CastOpKernel, CastOpKernel); diff --git a/paddle/fluid/operators/channel_close_op.cc b/paddle/fluid/operators/channel_close_op.cc index 5892650c49e2e9d7345fb94465d124cff57f0a6f..8e2db250a069c488ee98f618bc03df6485022456 100644 --- a/paddle/fluid/operators/channel_close_op.cc +++ b/paddle/fluid/operators/channel_close_op.cc @@ -50,8 +50,7 @@ class ChannelCloseOpOpInferShape : public framework::InferShapeBase { class ChannelCloseOpMaker : public framework::OpProtoAndCheckerMaker { public: - ChannelCloseOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput(kChannel, "The Channel Variable that should be closed by" " the ChannelClose Op."); diff --git a/paddle/fluid/operators/channel_create_op.cc b/paddle/fluid/operators/channel_create_op.cc index b2fdfd0e1f24ed071bb57b7de8f99b2d5e1d3196..a7f59e4088e3fb328e5b5a83eed65f0f90edb9f0 100644 --- a/paddle/fluid/operators/channel_create_op.cc +++ b/paddle/fluid/operators/channel_create_op.cc @@ -91,8 +91,7 @@ class ChannelCreateOpOpInferShape : public framework::InferShapeBase { class ChannelCreateOpMaker : public framework::OpProtoAndCheckerMaker { public: - ChannelCreateOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddOutput(kOutput, "The object of a Channel type created by ChannelCreate Op."); AddAttr("capacity", "The size of the buffer of Channel.") diff --git a/paddle/fluid/operators/channel_recv_op.cc b/paddle/fluid/operators/channel_recv_op.cc index 25c5c3c95ef6899589c98570df6ecbf9b3241d89..101015e837e28b504b71d919abd5f908a102c812 100644 --- a/paddle/fluid/operators/channel_recv_op.cc +++ b/paddle/fluid/operators/channel_recv_op.cc @@ -72,8 +72,7 @@ class ChannelRecvOp : public framework::OperatorBase { class ChannelRecvOpMaker : public framework::OpProtoAndCheckerMaker { public: - ChannelRecvOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput(Channel, "(Channel) A variable which \"receives\" the a value sent" "to it by a channel_send op.") diff --git a/paddle/fluid/operators/channel_send_op.cc b/paddle/fluid/operators/channel_send_op.cc index 66d33617ede5bef8a95de14f5b447c0910fe3eb4..67d6deb511d883ac69426ddd34be2199367cd4c7 100644 --- a/paddle/fluid/operators/channel_send_op.cc +++ b/paddle/fluid/operators/channel_send_op.cc @@ -57,8 +57,7 @@ class ChannelSendOp : public framework::OperatorBase { class ChannelSendOpMaker : public framework::OpProtoAndCheckerMaker { public: - ChannelSendOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput(Channel, "(Channel) A variable which \"sends\" the passed in value to " "a listening receiver.") diff --git a/paddle/fluid/operators/chunk_eval_op.cc b/paddle/fluid/operators/chunk_eval_op.cc index 95440ff89e883e754795c67cd58a08f1131df368..62636bb2f9078768180ab1e0016e3565617d24d2 100644 --- a/paddle/fluid/operators/chunk_eval_op.cc +++ b/paddle/fluid/operators/chunk_eval_op.cc @@ -66,8 +66,7 @@ class ChunkEvalOp : public framework::OperatorWithKernel { class ChunkEvalOpMaker : public framework::OpProtoAndCheckerMaker { public: - ChunkEvalOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Inference", "(Tensor, default: Tensor). " "Predictions from the network."); diff --git a/paddle/fluid/operators/clip_by_norm_op.cc b/paddle/fluid/operators/clip_by_norm_op.cc index f43726b4793f284f14226f90c94ac6eebf632bd5..c87bded034e382c981d119e8499d6780e288031f 100644 --- a/paddle/fluid/operators/clip_by_norm_op.cc +++ b/paddle/fluid/operators/clip_by_norm_op.cc @@ -37,8 +37,7 @@ class ClipByNormOp : public framework::OperatorWithKernel { class ClipByNormOpMaker : public framework::OpProtoAndCheckerMaker { public: - ClipByNormOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor) The input of clip_by_norm op." "The number of dimensions must be between [1, 9]."); diff --git a/paddle/fluid/operators/clip_op.cc b/paddle/fluid/operators/clip_op.cc index c71139fc7c01a696299296e43d06cf195fb3d03f..a679f7e2536a0a44148193f423f5ffe11b5e35fc 100644 --- a/paddle/fluid/operators/clip_op.cc +++ b/paddle/fluid/operators/clip_op.cc @@ -38,8 +38,7 @@ class ClipOp : public framework::OperatorWithKernel { template class ClipOpMaker : public framework::OpProtoAndCheckerMaker { public: - ClipOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor)The input of clip op." "The number of dimensions must be between [1, 9]."); diff --git a/paddle/fluid/operators/compare_op.cc b/paddle/fluid/operators/compare_op.cc index 3a6a357e81949014a70e5bae1ee0e1c8b9d0c2ce..3a4819f3dec9704a4a7c8910dd22e80fda082335 100644 --- a/paddle/fluid/operators/compare_op.cc +++ b/paddle/fluid/operators/compare_op.cc @@ -21,8 +21,7 @@ namespace operators { template class CompareOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - CompareOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { OpComment comment; AddInput("X", string::Sprintf("(LoDTensor) the left hand operand of %s operator", diff --git a/paddle/fluid/operators/concat_op.cc b/paddle/fluid/operators/concat_op.cc index 3bb3bd4eb15881afb5ae42beb944b76b5e8207cb..38337f9aa52435c445420047957500d21069506a 100644 --- a/paddle/fluid/operators/concat_op.cc +++ b/paddle/fluid/operators/concat_op.cc @@ -63,8 +63,7 @@ class ConcatOp : public framework::OperatorWithKernel { class ConcatOpMaker : public framework::OpProtoAndCheckerMaker { public: - ConcatOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "Input tensors of concat operator.").AsDuplicable(); AddOutput("Out", "Output tensor of concat operator."); AddAttr("axis", diff --git a/paddle/fluid/operators/concat_op.h b/paddle/fluid/operators/concat_op.h index 92c8ab6d9ff11ec6acd46a39877eb67d624748a9..1b1b8bf5ed959dd9c2ce8c9f5c905a75b81865fd 100644 --- a/paddle/fluid/operators/concat_op.h +++ b/paddle/fluid/operators/concat_op.h @@ -87,7 +87,7 @@ class ConcatGradKernel : public framework::OpKernel { auto& dev_ctx = ctx.template device_context(); paddle::operators::math::ConcatGradFunctor concat_grad_functor; - concat_grad_functor(dev_ctx, *in, static_cast(axis), outputs); + concat_grad_functor(dev_ctx, *in, static_cast(axis), &outputs); } } }; diff --git a/paddle/fluid/operators/concurrency/channel_util.cc b/paddle/fluid/operators/concurrency/channel_util.cc index 246c99489c45efec16babb1d3980606318236605..fba4abf1897bceea615222b2438700085ed8e551 100644 --- a/paddle/fluid/operators/concurrency/channel_util.cc +++ b/paddle/fluid/operators/concurrency/channel_util.cc @@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include "channel_util.h" +#include "paddle/fluid/operators/concurrency/channel_util.h" #include "paddle/fluid/framework/var_type.h" namespace poc = paddle::operators::concurrency; diff --git a/paddle/fluid/operators/conditional_block_op.cc b/paddle/fluid/operators/conditional_block_op.cc index 27f74a789beef02d31ebceb9b909e97ebd68232a..5984f80d04bdeb232f8e24264ae979725af24ef4 100644 --- a/paddle/fluid/operators/conditional_block_op.cc +++ b/paddle/fluid/operators/conditional_block_op.cc @@ -108,8 +108,7 @@ class ConditionalBlockOp : public ConditionalOp { class ConditionalBlockOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - ConditionalBlockOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The conditional variable of this operator. If X is empty, the " "whole sub-block will not be executed.") diff --git a/paddle/fluid/operators/conv_cudnn_op.cu.cc b/paddle/fluid/operators/conv_cudnn_op.cu.cc index c70e3cc3c9198008d9eca5f462000aa67ff7e5ba..7a7b8b76e43b1f91a3ba2767c217993cc39f26b6 100644 --- a/paddle/fluid/operators/conv_cudnn_op.cu.cc +++ b/paddle/fluid/operators/conv_cudnn_op.cu.cc @@ -20,6 +20,11 @@ limitations under the License. */ #include "paddle/fluid/platform/cudnn_helper.h" #include "paddle/fluid/platform/float16.h" +DEFINE_bool(cudnn_algo_use_autotune, true, + "Whether allow using an autotuning algorithm for convolution " + "operator. The autotuning algorithm may be non-deterministic. If " + "false, the algorithm is deterministic."); + namespace paddle { namespace operators { @@ -267,17 +272,23 @@ class CUDNNConvGradOpKernel : public framework::OpKernel { auto& dev_ctx = ctx.template device_context(); auto handle = dev_ctx.cudnn_handle(); if (input_grad) { - PADDLE_ENFORCE( - platform::dynload::cudnnGetConvolutionBackwardDataAlgorithm( - handle, cudnn_filter_desc, - // dyDesc: Handle to the previously initialized input differential - // tensor descriptor. - cudnn_output_grad_desc, cudnn_conv_desc, - // dxDesc: Handle to the previously initialized output tensor - // descriptor. - cudnn_input_desc, - CUDNN_CONVOLUTION_BWD_DATA_SPECIFY_WORKSPACE_LIMIT, - workspace_size_limit, &data_algo)); + if (FLAGS_cudnn_algo_use_autotune) { + PADDLE_ENFORCE( + platform::dynload::cudnnGetConvolutionBackwardDataAlgorithm( + handle, cudnn_filter_desc, + // dyDesc: Handle to the previously initialized input + // differential + // tensor descriptor. + cudnn_output_grad_desc, cudnn_conv_desc, + // dxDesc: Handle to the previously initialized output tensor + // descriptor. + cudnn_input_desc, + CUDNN_CONVOLUTION_BWD_DATA_SPECIFY_WORKSPACE_LIMIT, + workspace_size_limit, &data_algo)); + } else { + data_algo = CUDNN_CONVOLUTION_BWD_DATA_ALGO_1; + } + PADDLE_ENFORCE( platform::dynload::cudnnGetConvolutionBackwardDataWorkspaceSize( handle, cudnn_filter_desc, cudnn_output_grad_desc, @@ -286,12 +297,16 @@ class CUDNNConvGradOpKernel : public framework::OpKernel { } if (filter_grad) { - PADDLE_ENFORCE( - platform::dynload::cudnnGetConvolutionBackwardFilterAlgorithm( - handle, cudnn_input_desc, cudnn_output_grad_desc, cudnn_conv_desc, - cudnn_filter_desc, - CUDNN_CONVOLUTION_BWD_FILTER_SPECIFY_WORKSPACE_LIMIT, - workspace_size_limit, &filter_algo)); + if (FLAGS_cudnn_algo_use_autotune) { + PADDLE_ENFORCE( + platform::dynload::cudnnGetConvolutionBackwardFilterAlgorithm( + handle, cudnn_input_desc, cudnn_output_grad_desc, + cudnn_conv_desc, cudnn_filter_desc, + CUDNN_CONVOLUTION_BWD_FILTER_SPECIFY_WORKSPACE_LIMIT, + workspace_size_limit, &filter_algo)); + } else { + filter_algo = CUDNN_CONVOLUTION_BWD_FILTER_ALGO_1; + } PADDLE_ENFORCE( platform::dynload::cudnnGetConvolutionBackwardFilterWorkspaceSize( @@ -351,7 +366,8 @@ REGISTER_OP_KERNEL(conv2d_grad, CUDNN, plat::CUDAPlace, REGISTER_OP_KERNEL(conv3d, CUDNN, plat::CUDAPlace, paddle::operators::CUDNNConvOpKernel, - paddle::operators::CUDNNConvOpKernel); + paddle::operators::CUDNNConvOpKernel, + paddle::operators::CUDNNConvOpKernel); REGISTER_OP_KERNEL(conv3d_grad, CUDNN, plat::CUDAPlace, paddle::operators::CUDNNConvGradOpKernel, paddle::operators::CUDNNConvGradOpKernel); diff --git a/paddle/fluid/operators/conv_op.cc b/paddle/fluid/operators/conv_op.cc index 92748993c32ffb93ae25db8d9916798e657cc804..697d91484257984b104a13b0572cf19b16f8d37e 100644 --- a/paddle/fluid/operators/conv_op.cc +++ b/paddle/fluid/operators/conv_op.cc @@ -106,8 +106,7 @@ framework::OpKernelType ConvOp::GetExpectedKernelType( library); } -Conv2DOpMaker::Conv2DOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { +void Conv2DOpMaker::Make() { AddInput( "Input", "(Tensor) The input tensor of convolution operator. " @@ -200,8 +199,7 @@ $$ )DOC"); } -Conv3DOpMaker::Conv3DOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { +void Conv3DOpMaker::Make() { AddInput( "Input", "(Tensor) The input tensor of convolution operator. " diff --git a/paddle/fluid/operators/conv_op.h b/paddle/fluid/operators/conv_op.h index d6f86a5c88e37970379da0afe2a1d46e18b653f4..b3140116dfe6a17a400bb88219ff43b249ecb32a 100644 --- a/paddle/fluid/operators/conv_op.h +++ b/paddle/fluid/operators/conv_op.h @@ -17,9 +17,9 @@ limitations under the License. */ #include #include "paddle/fluid/framework/eigen.h" #include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/math/blas.h" #include "paddle/fluid/operators/math/depthwise_conv.h" #include "paddle/fluid/operators/math/im2col.h" -#include "paddle/fluid/operators/math/math_function.h" #include "paddle/fluid/operators/math/vol2col.h" namespace paddle { @@ -60,12 +60,12 @@ inline bool IsExpand(const std::vector& filter_dim, // operator implementations can reuse the code. class Conv2DOpMaker : public framework::OpProtoAndCheckerMaker { public: - Conv2DOpMaker(OpProto* proto, OpAttrChecker* op_checker); + void Make() override; }; class Conv3DOpMaker : public framework::OpProtoAndCheckerMaker { public: - Conv3DOpMaker(OpProto* proto, OpAttrChecker* op_checker); + void Make() override; }; class ConvOp : public framework::OperatorWithKernel { @@ -161,6 +161,7 @@ class GemmConvKernel : public framework::OpKernel { math::Im2ColFunctor im2col; auto& dev_ctx = context.template device_context(); + auto blas = math::GetBlas(dev_ctx); for (int i = 0; i < batch_size; i++) { Tensor in_batch = input->Slice(i, i + 1).Resize(input_shape); Tensor out_batch = output->Slice(i, i + 1).Resize(output_matrix_shape); @@ -186,8 +187,8 @@ class GemmConvKernel : public framework::OpKernel { // gemm Tensor out_slice = out_batch.Slice(g * out_step, (g + 1) * out_step); Tensor filter_slice = filter.Slice(g * out_step, (g + 1) * out_step); - math::matmul(dev_ctx, filter_slice, false, col_matrix, - false, T(1.0), &out_slice, T(0.0)); + blas.MatMul(filter_slice, false, col_matrix, false, T(1.0), &out_slice, + T(0.0)); } } } @@ -274,6 +275,7 @@ class GemmConvGradKernel : public framework::OpKernel { math::SetConstant set_zero; auto& dev_ctx = context.template device_context(); + auto blas = math::GetBlas(dev_ctx); if (input_grad) { input_grad->mutable_data(context.GetPlace()); @@ -303,9 +305,8 @@ class GemmConvGradKernel : public framework::OpKernel { col_matrix.ShareDataWith(in_grad_slice); col_matrix.Resize(col_matrix_shape); } - math::matmul(dev_ctx, filter_slice, true, - out_grad_slice, false, T(1.0), - &col_matrix, T(0.0)); + blas.MatMul(filter_slice, true, out_grad_slice, false, T(1.0), + &col_matrix, T(0.0)); if (is_expand && data_dim == 2U) { col2im(dev_ctx, col, dilations, strides, @@ -352,9 +353,8 @@ class GemmConvGradKernel : public framework::OpKernel { // gemm Tensor filter_grad_slice = filter_grad_.Slice(g * out_step, (g + 1) * out_step); - math::matmul(dev_ctx, out_grad_slice, false, - col_matrix, true, T(1.0), - &filter_grad_slice, T(1.0)); + blas.MatMul(out_grad_slice, false, col_matrix, true, T(1.0), + &filter_grad_slice, T(1.0)); } } } diff --git a/paddle/fluid/operators/conv_shift_op.cc b/paddle/fluid/operators/conv_shift_op.cc index 82fdd308207adb159632dbb9decd67fd2d1c4646..f2549e814d6f3b5674fe2eec1139f1c3dc6fa0b4 100644 --- a/paddle/fluid/operators/conv_shift_op.cc +++ b/paddle/fluid/operators/conv_shift_op.cc @@ -75,8 +75,7 @@ class ConvShiftGradOp : public framework::OperatorWithKernel { class ConvShiftOpMaker : public framework::OpProtoAndCheckerMaker { public: - ConvShiftOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor, default Tensor), a 2-D tensor with shape B x M, " "where B is the batch size and M is the data dimension."); diff --git a/paddle/fluid/operators/conv_shift_op.cu b/paddle/fluid/operators/conv_shift_op.cu index 344bbade7055aa8e0aede61dd31dab246bddd169..314d33310588ed960eecaf1a0319ebf56d925c55 100644 --- a/paddle/fluid/operators/conv_shift_op.cu +++ b/paddle/fluid/operators/conv_shift_op.cu @@ -14,7 +14,7 @@ limitations under the License. */ #include "paddle/fluid/operators/conv_shift_op.h" #include "paddle/fluid/operators/math/math_function.h" -#include "paddle/fluid/platform/cuda_helper.h" +#include "paddle/fluid/platform/cuda_primitives.h" namespace paddle { namespace operators { diff --git a/paddle/fluid/operators/conv_transpose_cudnn_op.cu.cc b/paddle/fluid/operators/conv_transpose_cudnn_op.cu.cc index 901682edbb01c563be6ea407228336b14f942778..038ea8999072f562104c5386ed18b6b275816345 100644 --- a/paddle/fluid/operators/conv_transpose_cudnn_op.cu.cc +++ b/paddle/fluid/operators/conv_transpose_cudnn_op.cu.cc @@ -44,6 +44,7 @@ class CUDNNConvTransposeOpKernel : public framework::OpKernel { std::vector paddings = ctx.Attr>("paddings"); // cudnn v5 does not support dilations std::vector dilations = ctx.Attr>("dilations"); + int groups = ctx.Attr("groups"); int user_workspace_size = ctx.Attr("workspace_size_MB"); const T* input_data = input->data(); @@ -64,13 +65,13 @@ class CUDNNConvTransposeOpKernel : public framework::OpKernel { // (N, M, H, W) or (N, M, D, H, W) cudnnTensorDescriptor_t cudnn_input_desc = input_desc.descriptor( - layout, framework::vectorize2int(input->dims())); + layout, framework::vectorize2int(input->dims()), groups); // (N, C, O_h, O_w) or (N, C, O_d, O_h, O_w) cudnnTensorDescriptor_t cudnn_output_desc = output_desc.descriptor( - layout, framework::vectorize2int(output->dims())); + layout, framework::vectorize2int(output->dims()), groups); // (M, C, K_h, K_w) or (M, C, K_d, K_h, K_w) cudnnFilterDescriptor_t cudnn_filter_desc = filter_desc.descriptor( - layout, framework::vectorize2int(filter->dims())); + layout, framework::vectorize2int(filter->dims()), groups); cudnnConvolutionDescriptor_t cudnn_conv_desc = conv_desc.descriptor(paddings, strides, dilations); @@ -104,11 +105,17 @@ class CUDNNConvTransposeOpKernel : public framework::OpKernel { cudnn_workspace = paddle::memory::Alloc(gpu, workspace_size_in_bytes); // ------------------- cudnn conv transpose forward --------------------- + int input_offset = input->numel() / input->dims()[0] / groups; + int output_offset = output->numel() / output->dims()[0] / groups; + int filter_offset = filter->numel() / groups; T alpha = 1.0f, beta = 0.0f; - PADDLE_ENFORCE(platform::dynload::cudnnConvolutionBackwardData( - handle, &alpha, cudnn_filter_desc, filter_data, cudnn_input_desc, - input_data, cudnn_conv_desc, algo, cudnn_workspace, - workspace_size_in_bytes, &beta, cudnn_output_desc, output_data)); + for (int g = 0; g < groups; g++) { + PADDLE_ENFORCE(platform::dynload::cudnnConvolutionBackwardData( + handle, &alpha, cudnn_filter_desc, filter_data + filter_offset * g, + cudnn_input_desc, input_data + input_offset * g, cudnn_conv_desc, + algo, cudnn_workspace, workspace_size_in_bytes, &beta, + cudnn_output_desc, output_data + output_offset * g)); + } // Release the cudnn workspace paddle::memory::Free(gpu, cudnn_workspace); @@ -134,6 +141,7 @@ class CUDNNConvTransposeGradOpKernel : public framework::OpKernel { std::vector paddings = ctx.Attr>("paddings"); // cudnn v5 does not support dilations std::vector dilations = ctx.Attr>("dilations"); + int groups = ctx.Attr("groups"); int user_workspace_size = ctx.Attr("workspace_size_MB"); // ------------------- cudnn descriptors --------------------- @@ -145,13 +153,13 @@ class CUDNNConvTransposeGradOpKernel : public framework::OpKernel { // Input: (N, M, H, W) or (N, M, D, H, W) cudnnTensorDescriptor_t cudnn_input_desc = input_desc.descriptor( - layout, framework::vectorize2int(input->dims())); + layout, framework::vectorize2int(input->dims()), groups); // Output: (N, C, O_h, O_w) or (N, C, O_d, O_h, O_w) cudnnTensorDescriptor_t cudnn_output_desc = output_desc.descriptor( - layout, framework::vectorize2int(output_grad->dims())); + layout, framework::vectorize2int(output_grad->dims()), groups); // Filter (M, C, K_h, K_w) or (M, C, K_d K_h, K_w) cudnnFilterDescriptor_t cudnn_filter_desc = filter_desc.descriptor( - layout, framework::vectorize2int(filter->dims())); + layout, framework::vectorize2int(filter->dims()), groups); cudnnConvolutionDescriptor_t cudnn_conv_desc = conv_desc.descriptor(paddings, strides, dilations); @@ -205,15 +213,22 @@ class CUDNNConvTransposeGradOpKernel : public framework::OpKernel { cudnn_workspace = paddle::memory::Alloc(gpu, workspace_size_in_bytes); // ------------------- cudnn conv backward data --------------------- // FIXME(typhoonzero): template type T may not be the same as cudnn call. + int input_offset = input->numel() / input->dims()[0] / groups; + int output_grad_offset = + output_grad->numel() / output_grad->dims()[0] / groups; + int filter_offset = filter->numel() / groups; T alpha = 1.0f, beta = 0.0f; if (input_grad) { T* input_grad_data = input_grad->mutable_data(ctx.GetPlace()); // Because beta is zero, it is unnecessary to reset input_grad. - PADDLE_ENFORCE(platform::dynload::cudnnConvolutionForward( - handle, &alpha, cudnn_output_desc, output_grad_data, - cudnn_filter_desc, filter_data, cudnn_conv_desc, data_algo, - cudnn_workspace, workspace_size_in_bytes, &beta, cudnn_input_desc, - input_grad_data)); + for (int g = 0; g < groups; g++) { + PADDLE_ENFORCE(platform::dynload::cudnnConvolutionForward( + handle, &alpha, cudnn_output_desc, + output_grad_data + output_grad_offset * g, cudnn_filter_desc, + filter_data + filter_offset * g, cudnn_conv_desc, data_algo, + cudnn_workspace, workspace_size_in_bytes, &beta, cudnn_input_desc, + input_grad_data + input_offset * g)); + } } // ------------------- cudnn conv backward filter --------------------- @@ -221,11 +236,16 @@ class CUDNNConvTransposeGradOpKernel : public framework::OpKernel { T* filter_grad_data = filter_grad->mutable_data(ctx.GetPlace()); // Because beta is zero, it is unnecessary to reset filter_grad. // Gradient with respect to the filter - PADDLE_ENFORCE(platform::dynload::cudnnConvolutionBackwardFilter( - handle, &alpha, cudnn_output_desc, output_grad_data, cudnn_input_desc, - input_data, cudnn_conv_desc, filter_algo, cudnn_workspace, - workspace_size_in_bytes, &beta, cudnn_filter_desc, filter_grad_data)); + for (int g = 0; g < groups; g++) { + PADDLE_ENFORCE(platform::dynload::cudnnConvolutionBackwardFilter( + handle, &alpha, cudnn_output_desc, + output_grad_data + output_grad_offset * g, cudnn_input_desc, + input_data + input_offset * g, cudnn_conv_desc, filter_algo, + cudnn_workspace, workspace_size_in_bytes, &beta, cudnn_filter_desc, + filter_grad_data + filter_offset * g)); + } } + // Release the cudnn workspace paddle::memory::Free(gpu, cudnn_workspace); } diff --git a/paddle/fluid/operators/conv_transpose_op.cc b/paddle/fluid/operators/conv_transpose_op.cc index d699dcafa4e2c7e0a3ffb62ec3985e4961fa2133..0b363f5c43f9fc191790e5cca629ffc46eb9388c 100644 --- a/paddle/fluid/operators/conv_transpose_op.cc +++ b/paddle/fluid/operators/conv_transpose_op.cc @@ -32,6 +32,7 @@ void ConvTransposeOp::InferShape(framework::InferShapeContext* ctx) const { std::vector strides = ctx->Attrs().Get>("strides"); std::vector paddings = ctx->Attrs().Get>("paddings"); std::vector dilations = ctx->Attrs().Get>("dilations"); + int groups = ctx->Attrs().Get("groups"); PADDLE_ENFORCE(in_dims.size() == 4 || in_dims.size() == 5, "ConvTransposeOp intput should be 4-D or 5-D tensor."); @@ -48,10 +49,10 @@ void ConvTransposeOp::InferShape(framework::InferShapeContext* ctx) const { "ConvTransposeOp paddings dimension and dilations " "dimension should be the same."); PADDLE_ENFORCE_EQ(in_dims[1], filter_dims[0], - "In ConvTransposeOp, The input channel should be the same " - "as the number of filters."); + "In ConvTransposeOp, The number of input channels should " + "be equal to the number of filter's channels."); - std::vector output_shape({in_dims[0], filter_dims[1]}); + std::vector output_shape({in_dims[0], filter_dims[1] * groups}); for (size_t i = 0; i < strides.size(); ++i) { auto filter_extent = dilations[i] * (filter_dims[i + 2] - 1) + 1; output_shape.push_back((in_dims[i + 2] - 1) * strides[i] - 2 * paddings[i] + @@ -84,9 +85,7 @@ framework::OpKernelType ConvTransposeOp::GetExpectedKernelType( layout_, library_); } -Conv2DTransposeOpMaker::Conv2DTransposeOpMaker(OpProto* proto, - OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { +void Conv2DTransposeOpMaker::Make() { AddInput( "Input", "(Tensor) The input tensor of convolution transpose operator. " @@ -104,7 +103,10 @@ Conv2DTransposeOpMaker::Conv2DTransposeOpMaker(OpProto* proto, AddOutput("Output", "(Tensor) The output tensor of convolution transpose operator. " "The format of output tensor is also NCHW."); - + AddAttr("groups", + "(int default:1), the groups number of the convolution " + "transpose operator. ") + .SetDefault(1); AddAttr>("dilations", "(vector default:{1, 1}), the " "dilations(h_dilation, w_dilation) of convolution " @@ -168,9 +170,7 @@ Example: )DOC"); } -Conv3DTransposeOpMaker::Conv3DTransposeOpMaker(OpProto* proto, - OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { +void Conv3DTransposeOpMaker::Make() { AddInput("Input", "(Tensor) The input tensor of convolution transpose operator." "The format of input tensor is NCDHW. Where N is batch size, C is " @@ -208,6 +208,10 @@ Conv3DTransposeOpMaker::Conv3DTransposeOpMaker(OpProto* proto, "(vector default:{0, 0, 0}), paddings(d_pad, " "h_pad, w_pad) of convolution transpose operator.") .SetDefault({0, 0, 0}); + AddAttr("groups", + "(int default:1), the groups number of the convolution3d " + "transpose operator. ") + .SetDefault(1); AddAttr( "use_cudnn", "(bool, default false) Only used in cudnn kernel, need install cudnn") diff --git a/paddle/fluid/operators/conv_transpose_op.h b/paddle/fluid/operators/conv_transpose_op.h index bfc0177c2a0da1627fbca532764fdae8167b6b2a..1dcfc651fdd79aed50736d05d38ec8576b183d41 100644 --- a/paddle/fluid/operators/conv_transpose_op.h +++ b/paddle/fluid/operators/conv_transpose_op.h @@ -16,8 +16,8 @@ limitations under the License. */ #include #include "paddle/fluid/framework/eigen.h" #include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/math/blas.h" #include "paddle/fluid/operators/math/im2col.h" -#include "paddle/fluid/operators/math/math_function.h" #include "paddle/fluid/operators/math/vol2col.h" namespace paddle { @@ -30,12 +30,12 @@ using DDim = framework::DDim; // operator implementations can reuse the code. class Conv2DTransposeOpMaker : public framework::OpProtoAndCheckerMaker { public: - Conv2DTransposeOpMaker(OpProto* proto, OpAttrChecker* op_checker); + void Make() override; }; class Conv3DTransposeOpMaker : public framework::OpProtoAndCheckerMaker { public: - Conv3DTransposeOpMaker(OpProto* proto, OpAttrChecker* op_checker); + void Make() override; }; class ConvTransposeOp : public framework::OperatorWithKernel { @@ -70,7 +70,7 @@ class GemmConvTransposeKernel : public framework::OpKernel { std::vector strides = context.Attr>("strides"); std::vector paddings = context.Attr>("paddings"); std::vector dilations = context.Attr>("dilations"); - // groups will alway be disabled in conv2dtranspose. + int groups = context.Attr("groups"); const int batch_size = static_cast(input->dims()[0]); @@ -81,10 +81,10 @@ class GemmConvTransposeKernel : public framework::OpKernel { // use col_shape in the im2col and col2im (or vol2col and col2vol) // calculation - // col_shape_vec: {c, k_h, k_w, h, w} or {c, k_d, k_h, k_w, d, h, w} + // col_shape_vec: {c/g, k_h, k_w, h, w} or {c/g, k_d, k_h, k_w, d, h, w} size_t data_dim = filter_shape_vec.size() - 2; std::vector col_shape_vec(1 + 2 * data_dim); - col_shape_vec[0] = output->dims()[1]; + col_shape_vec[0] = output->dims()[1] / groups; for (size_t j = 0; j < data_dim; ++j) { col_shape_vec[j + 1] = filter_shape_vec[j + 2]; col_shape_vec[j + 1 + data_dim] = input_shape_vec[j + 2]; @@ -92,7 +92,7 @@ class GemmConvTransposeKernel : public framework::OpKernel { DDim col_shape(framework::make_ddim(col_shape_vec)); // use col_matrix_shape in the gemm calculation - // size: (c * k_h * k_w, h * w) or (c * k_d * k_h * k_w, d * h * w) + // size: (c/g * k_h * k_w, h * w) or (c/g * k_d * k_h * k_w, d * h * w) DDim col_matrix_shape = framework::flatten_to_2d(col_shape, data_dim + 1); Tensor col; @@ -111,15 +111,18 @@ class GemmConvTransposeKernel : public framework::OpKernel { // input matrix size: (m, h * w) or (m, d * h * w) DDim input_matrix_shape = {input->dims()[1], col_matrix_shape[1]}; - // filter size: (m, c * k_h * k_w) or (m, c * k_d * k_h * k_w) + // filter size: (m, c/g * k_h * k_w) or (m, c/g * k_d * k_h * k_w) DDim filter_matrix_shape = {input->dims()[1], col_matrix_shape[0]}; filter.Resize(filter_matrix_shape); output->mutable_data(context.GetPlace()); math::SetConstant set_zero; auto& dev_ctx = context.template device_context(); + auto blas = math::GetBlas(dev_ctx); set_zero(dev_ctx, output, static_cast(0)); + int in_step = static_cast(input->dims()[1]) / groups; + int out_step = static_cast(output->dims()[1]) / groups; math::Col2ImFunctor col2im; math::Col2VolFunctor col2vol; @@ -132,23 +135,29 @@ class GemmConvTransposeKernel : public framework::OpKernel { // output size: (c, o_h, o_w) or (c, o_d, o_h, o_w) Tensor output_batch = output->Slice(i, i + 1).Resize(output_shape); - // col_matrix = filter * input_batch - // of shape (c * k_h * k_w, h * w) or (c * k_d * k_h * k_w, d * h * w) - math::matmul(dev_ctx, filter, true, input_batch, false, - static_cast(1.0), &col_matrix, - static_cast(0.0)); - - if (data_dim == 2U) { - // col2im: col_matrix -> dy - // from (c * k_h * k_w, h * w) to (c, o_h, o_w) - col2im(dev_ctx, col, dilations, strides, - std::vector{paddings[0], paddings[1], paddings[0], - paddings[1]}, - &output_batch); - } else if (data_dim == 3U) { - // col2vol: col_matrix -> dy - // from (c * k_d * k_h * k_w, d * h * w) to (c, o_d, o_h, o_w) - col2vol(dev_ctx, col, dilations, strides, paddings, &output_batch); + for (int g = 0; g < groups; g++) { + Tensor in_slice = input_batch.Slice(g * in_step, (g + 1) * in_step); + Tensor filter_slice = filter.Slice(g * in_step, (g + 1) * in_step); + Tensor out_slice = output_batch.Slice(g * out_step, (g + 1) * out_step); + + // col_matrix = filter_slice * input_slice + // of shape (c/g * k_h * k_w, h * w) + // or (c/g * k_d * k_h * k_w, d * h * w) + blas.MatMul(filter_slice, true, in_slice, false, static_cast(1.0), + &col_matrix, static_cast(0.0)); + + if (data_dim == 2U) { + // col2im: col_matrix -> dy + // from (c/g * k_h * k_w, h * w) to (c/g, o_h, o_w) + col2im(dev_ctx, col, dilations, strides, + std::vector{paddings[0], paddings[1], paddings[0], + paddings[1]}, + &out_slice); + } else if (data_dim == 3U) { + // col2vol: col_matrix -> dy + // from (c/g * k_d * k_h * k_w, d * h * w) to (c/g, o_d, o_h, o_w) + col2vol(dev_ctx, col, dilations, strides, paddings, &out_slice); + } } } } @@ -174,6 +183,7 @@ class GemmConvTransposeGradKernel : public framework::OpKernel { std::vector strides = context.Attr>("strides"); std::vector paddings = context.Attr>("paddings"); std::vector dilations = context.Attr>("dilations"); + int groups = context.Attr("groups"); const int batch_size = static_cast(input->dims()[0]); @@ -205,14 +215,17 @@ class GemmConvTransposeGradKernel : public framework::OpKernel { // input matrix size: (m, h * w) or (m, d * h * w) DDim input_matrix_shape = {input->dims()[1], col_matrix_shape[1]}; - // filter size: (m, c * k_h * k_w) or (m, c * k_d * k_h * k_w) - DDim filter_matrix_shape = {input->dims()[1], col_matrix_shape[0]}; + // filter size: (m, c/g * k_h * k_w) or (m, c/g * k_d * k_h * k_w) + DDim filter_matrix_shape = {input->dims()[1], col_matrix_shape[0] / groups}; filter.Resize(filter_matrix_shape); + int in_step = static_cast(input->dims()[1]) / groups; + int col_step = static_cast(col_matrix_shape[0]) / groups; // convolution transpose grad on input: // im2col + gemm (similar to conv-forward) // input need to compute gradient auto& dev_ctx = context.template device_context(); + auto blas = math::GetBlas(dev_ctx); if (input_grad || filter_grad) { Tensor col; col.mutable_data(col_shape, context.GetPlace()); @@ -232,7 +245,7 @@ class GemmConvTransposeGradKernel : public framework::OpKernel { if (input_grad) { input_grad->mutable_data(context.GetPlace()); } - if (filter_grad) { // filter size (m, c, k_h, k_w) + if (filter_grad) { // filter size (m, c/g, k_h, k_w) filter_grad->mutable_data(context.GetPlace()); set_zero(dev_ctx, filter_grad, static_cast(0)); filter_grad_ = *filter_grad; @@ -267,9 +280,17 @@ class GemmConvTransposeGradKernel : public framework::OpKernel { // or // (m, c * k_d * k_h * k_w) * (c * k_d * k_h * k_w, d * h * w) -> (m, // d, h, w) - math::matmul( - dev_ctx, filter, false, col_matrix, false, static_cast(1.0), - &input_grad_batch, static_cast(0.0)); + for (int g = 0; g < groups; g++) { + Tensor input_grad_slice = + input_grad_batch.Slice(g * in_step, (g + 1) * in_step); + Tensor filter_slice = filter.Slice(g * in_step, (g + 1) * in_step); + Tensor col_matrix_slice = + col_matrix.Slice(g * col_step, (g + 1) * col_step); + + blas.MatMul(filter_slice, false, col_matrix_slice, false, + static_cast(1.0), &input_grad_slice, + static_cast(0.0)); + } } if (filter_grad) { // input batch @@ -279,9 +300,17 @@ class GemmConvTransposeGradKernel : public framework::OpKernel { // or // (m, d * h * w) * (d * h * w, c * k_d * k_h * k_w) -> (m, c * k_d * // k_h * k_w) - math::matmul(dev_ctx, in_batch, false, col_matrix, - true, static_cast(1.0), - &filter_grad_, static_cast(1.0)); + for (int g = 0; g < groups; g++) { + Tensor in_batch_slice = + in_batch.Slice(g * in_step, (g + 1) * in_step); + Tensor filter_grad_slice = + filter_grad_.Slice(g * in_step, (g + 1) * in_step); + Tensor col_matrix_slice = + col_matrix.Slice(g * col_step, (g + 1) * col_step); + blas.MatMul(in_batch_slice, false, col_matrix_slice, true, + static_cast(1.0), &filter_grad_slice, + static_cast(1.0)); + } } } } diff --git a/paddle/fluid/operators/cos_sim_op.cc b/paddle/fluid/operators/cos_sim_op.cc index 04ca878e687f9b8e5239d8c4aad7e5f262fda0fa..046dd11910bb0ff46b567c3b89883582782205d3 100644 --- a/paddle/fluid/operators/cos_sim_op.cc +++ b/paddle/fluid/operators/cos_sim_op.cc @@ -62,8 +62,7 @@ class CosSimOp : public framework::OperatorWithKernel { class CosSimOpMaker : public framework::OpProtoAndCheckerMaker { public: - CosSimOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The 1st input of cos_sim op."); AddInput("Y", "The 2nd input of cos_sim op."); AddOutput("Out", "The output of cos_sim op."); diff --git a/paddle/fluid/operators/crf_decoding_op.cc b/paddle/fluid/operators/crf_decoding_op.cc index a83013c428a77a0ead545d87852e1017bc927edf..40f43936db662f2b18ffa540da4794755b5d6fc7 100644 --- a/paddle/fluid/operators/crf_decoding_op.cc +++ b/paddle/fluid/operators/crf_decoding_op.cc @@ -18,8 +18,7 @@ namespace paddle { namespace operators { class CRFDecodingOpMaker : public framework::OpProtoAndCheckerMaker { public: - CRFDecodingOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Emission", "(LoDTensor, default: LoDTensor). A LoDTensor with shape " "[N x D] where N is the size of the mini-batch and D is the total " diff --git a/paddle/fluid/operators/crop_op.cc b/paddle/fluid/operators/crop_op.cc index a8f1fbd529c71d1915c75fa90b7e4e8239d2fa3f..669b3bbe9df4cae1aa381184092dfa51157ab6a3 100644 --- a/paddle/fluid/operators/crop_op.cc +++ b/paddle/fluid/operators/crop_op.cc @@ -52,8 +52,7 @@ class CropOp : public framework::OperatorWithKernel { class CropOpMaker : public framework::OpProtoAndCheckerMaker { public: - CropOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The input of pad op. " "The input should be a k-D tensor(k > 0 and k < 7)."); diff --git a/paddle/fluid/operators/cross_entropy_op.cc b/paddle/fluid/operators/cross_entropy_op.cc index 0e0622e290f42811c83c354d749ef32a2d9dcadb..a3bec3da45136bca5cb2763e7ffd6b67703a1813 100644 --- a/paddle/fluid/operators/cross_entropy_op.cc +++ b/paddle/fluid/operators/cross_entropy_op.cc @@ -111,8 +111,7 @@ class CrossEntropyGradientOp : public framework::OperatorWithKernel { class CrossEntropyOpMaker : public framework::OpProtoAndCheckerMaker { public: - CrossEntropyOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor, default Tensor), a 2-D tensor with shape [N x D]," " where N is the batch size and D is the number of classes. " @@ -164,11 +163,13 @@ or not. But the output only shares the LoD information with input X. } // namespace paddle namespace ops = paddle::operators; +using CPUCtx = paddle::platform::CPUDeviceContext; + REGISTER_OPERATOR(cross_entropy, ops::CrossEntropyOp, ops::CrossEntropyOpMaker, paddle::framework::DefaultGradOpDescMaker); REGISTER_OPERATOR(cross_entropy_grad, ops::CrossEntropyGradientOp); -REGISTER_OP_CPU_KERNEL(cross_entropy, ops::CrossEntropyOpKernel, - ops::CrossEntropyOpKernel); +REGISTER_OP_CPU_KERNEL(cross_entropy, ops::CrossEntropyOpKernel, + ops::CrossEntropyOpKernel); REGISTER_OP_CPU_KERNEL(cross_entropy_grad, - ops::CrossEntropyGradientOpKernel, - ops::CrossEntropyGradientOpKernel); + ops::CrossEntropyGradientOpKernel, + ops::CrossEntropyGradientOpKernel); diff --git a/paddle/fluid/operators/cross_entropy_op.cu b/paddle/fluid/operators/cross_entropy_op.cu index 6449149d4b55962e84baafffc0c2c03f8108866f..30dbd5bd3d39dd2992c3dd91364003bb7715a2eb 100644 --- a/paddle/fluid/operators/cross_entropy_op.cu +++ b/paddle/fluid/operators/cross_entropy_op.cu @@ -14,98 +14,11 @@ limitations under the License. */ #include "paddle/fluid/operators/cross_entropy_op.h" -namespace paddle { -namespace operators { - -namespace { - -template -__global__ void CrossEntropyGradientKernel(T* dX, const T* dY, const T* X, - const int64_t* label, const int N, - const int D) { - for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < N; - i += blockDim.x * gridDim.x) { - int idx = i * D + label[i]; - dX[idx] = -dY[i] / X[idx]; - } -} - -template -__global__ void SoftCrossEntropyGradientKernel(T* dX, const T* dY, const T* X, - const T* label, const int N, - const int D) { - int ids = blockIdx.x * blockDim.x + threadIdx.x; - if (ids < N * D) { - int row_ids = ids / D; - dX[ids] = -label[ids] * dY[row_ids] / X[ids]; - } -} -} // namespace - -template -class CrossEntropyOpCUDAKernel : public framework::OpKernel { - public: - void Compute(const framework::ExecutionContext& ctx) const override { - PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), - "This kernel only runs on GPU device."); - const Tensor* x = ctx.Input("X"); - const Tensor* label = ctx.Input("Label"); - Tensor* y = ctx.Output("Y"); - y->mutable_data(ctx.GetPlace()); - - math::CrossEntropyFunctor()( - ctx.template device_context(), y, x, label, - ctx.Attr("soft_label")); - } -}; - -template -class CrossEntropyGradientOpCUDAKernel : public framework::OpKernel { - public: - void Compute(const framework::ExecutionContext& ctx) const override { - PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), - "This kernel only runs on GPU device."); - - const Tensor* x = ctx.Input("X"); - const Tensor* label = ctx.Input("Label"); - Tensor* dx = ctx.Output(framework::GradVarName("X")); - dx->mutable_data(ctx.GetPlace()); - - const T* dy_data = - ctx.Input(framework::GradVarName("Y"))->data(); - T* dx_data = dx->mutable_data(ctx.GetPlace()); - const T* x_data = x->data(); - - int64_t batch_size = x->dims()[0]; - int64_t class_num = x->dims()[1]; - - int block = 512; - int grid = (batch_size * class_num + block - 1) / block; - - auto& dev_ctx = ctx.template device_context(); - auto stream = dev_ctx.stream(); - - if (ctx.Attr("soft_label")) { - auto* label_data = label->data(); - SoftCrossEntropyGradientKernel<<>>( - dx_data, dy_data, x_data, label_data, batch_size, class_num); - } else { - math::SetConstant functor; - functor(dev_ctx, dx, 0); - auto* label_data = label->data(); - grid = (batch_size + block - 1) / block; - CrossEntropyGradientKernel<<>>( - dx_data, dy_data, x_data, label_data, batch_size, class_num); - } - } -}; - -} // namespace operators -} // namespace paddle - namespace ops = paddle::operators; -REGISTER_OP_CUDA_KERNEL(cross_entropy, ops::CrossEntropyOpCUDAKernel, - ops::CrossEntropyOpCUDAKernel); +using CUDACtx = paddle::platform::CUDADeviceContext; +REGISTER_OP_CUDA_KERNEL(cross_entropy, + ops::CrossEntropyOpKernel, + ops::CrossEntropyOpKernel); REGISTER_OP_CUDA_KERNEL(cross_entropy_grad, - ops::CrossEntropyGradientOpCUDAKernel, - ops::CrossEntropyGradientOpCUDAKernel); + ops::CrossEntropyGradientOpKernel, + ops::CrossEntropyGradientOpKernel); diff --git a/paddle/fluid/operators/cross_entropy_op.h b/paddle/fluid/operators/cross_entropy_op.h index 6da3a24dc89a85fe432b6350d3af7b0e84337c9d..19a2aec92b267ece94685ce34604b7d1cfa5d209 100644 --- a/paddle/fluid/operators/cross_entropy_op.h +++ b/paddle/fluid/operators/cross_entropy_op.h @@ -17,69 +17,106 @@ limitations under the License. */ #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/operators/math/cross_entropy.h" #include "paddle/fluid/operators/math/math_function.h" +#include "paddle/fluid/platform/for_range.h" namespace paddle { namespace operators { using Tensor = framework::Tensor; -template -using EigenMatrix = framework::EigenMatrix; -template +template class CrossEntropyOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { - PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()), - "This kernel only runs on CPU."); - const Tensor* x = ctx.Input("X"); - const Tensor* labels = ctx.Input("Label"); - Tensor* y = ctx.Output("Y"); + auto* x = ctx.Input("X"); + auto* labels = ctx.Input("Label"); + auto* y = ctx.Output("Y"); y->mutable_data(ctx.GetPlace()); - math::CrossEntropyFunctor()( - ctx.template device_context(), y, x, labels, + math::CrossEntropyFunctor()( + ctx.template device_context(), y, x, labels, ctx.Attr("soft_label")); } }; template +class XeSoftlabelGradFunctor { + public: + XeSoftlabelGradFunctor(T* dx, + const T* dy, // NOLINT + const T* x, // NOLINT + const T* label, // NOLINT + size_t num_classes) + : dx_(dx), dy_(dy), x_(x), label_(label), num_classes_(num_classes) {} + + HOSTDEVICE void operator()(size_t i) { + auto row_ids = i / num_classes_; + dx_[i] = -label_[i] * dy_[row_ids] / x_[i]; + } + + private: + T* dx_; + const T* dy_; + const T* x_; + const T* label_; + size_t num_classes_; +}; + +template +class XeGradFunctor { + public: + XeGradFunctor(T* dx, + const T* dy, // NOLINT + const T* x, // NOLINT + const int64_t* label, // NOLINT + size_t num_classes) + : dx_(dx), dy_(dy), x_(x), label_(label), num_classes_(num_classes) {} + + HOSTDEVICE void operator()(size_t sample_id) { + auto x_is_true_offset = sample_id * num_classes_ + label_[sample_id]; + for (size_t x_offset = sample_id * num_classes_; + x_offset < (sample_id + 1) * num_classes_; ++x_offset) { + dx_[x_offset] = x_offset != x_is_true_offset + ? static_cast(0) + : -dy_[sample_id] / x_[x_offset]; + } + } + + private: + T* dx_; + const T* dy_; + const T* x_; + const int64_t* label_; + size_t num_classes_; +}; + +template class CrossEntropyGradientOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { - PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()), - "This kernel only runs on CPU."); - const Tensor* x = ctx.Input("X"); - const Tensor* dy = ctx.Input(framework::GradVarName("Y")); - const Tensor* label = ctx.Input("Label"); - Tensor* dx = ctx.Output(framework::GradVarName("X")); - T* dx_data = dx->mutable_data(ctx.GetPlace()); + auto* x = ctx.Input("X"); + auto* dy = ctx.Input(framework::GradVarName("Y")); + auto* label = ctx.Input("Label"); + auto* dx = ctx.Output(framework::GradVarName("X")); + auto* dx_data = dx->mutable_data(ctx.GetPlace()); int64_t class_num = x->dims()[1]; if (ctx.Attr("soft_label")) { - auto x_mat = EigenMatrix::From(*x); - auto dy_mat = EigenMatrix::From(*dy); - auto lbl_mat = EigenMatrix::From(*label); - auto dx_mat = EigenMatrix::From(*dx); - - dx_mat.device(*ctx.template device_context() - .eigen_device()) = - -(lbl_mat * - dy_mat.broadcast(Eigen::DSizes(1, class_num)) / x_mat); + XeSoftlabelGradFunctor functor(dx_data, dy->data(), x->data(), + label->data(), + static_cast(class_num)); + platform::ForRange for_range( + ctx.template device_context(), + static_cast(dx->numel())); + for_range(functor); } else { - int64_t batch_size = x->dims()[0]; - const T* dy_data = dy->data(); - const T* x_data = x->data(); - const int64_t* label_data = label->data(); - - math::SetConstant functor; - functor(ctx.template device_context(), dx, 0); - - for (int64_t i = 0; i < batch_size; ++i) { - PADDLE_ASSERT(label_data[i] >= 0 || label_data[i] < class_num); - int64_t index = i * class_num + label_data[i]; - dx_data[index] = math::TolerableValue()(-dy_data[i] / x_data[index]); - } + XeGradFunctor functor(dx_data, dy->data(), x->data(), + label->data(), + static_cast(class_num)); + platform::ForRange for_range( + ctx.template device_context(), + static_cast(dy->numel())); + for_range(functor); } } }; diff --git a/paddle/fluid/operators/ctc_align_op.cc b/paddle/fluid/operators/ctc_align_op.cc index 19e7649660edd0bc90bc6a9537b1cdbb2e7e8ebc..d2b440d9d2e50340af7a7bb4e76e55beea1bcb46 100644 --- a/paddle/fluid/operators/ctc_align_op.cc +++ b/paddle/fluid/operators/ctc_align_op.cc @@ -44,8 +44,7 @@ class CTCAlignOp : public framework::OperatorWithKernel { class CTCAlignOpMaker : public framework::OpProtoAndCheckerMaker { public: - CTCAlignOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Input", "(LodTensor, default: LoDTensor), Its shape is " "[Lp, 1], where Lp is the sum of all input sequences' length."); diff --git a/paddle/fluid/operators/cumsum_op.cc b/paddle/fluid/operators/cumsum_op.cc index f7c516a0ba375a68e3adeb44c99f2808dc0418bb..92bb835e8f18e17ae1355fdec29f43b8ffb70460 100644 --- a/paddle/fluid/operators/cumsum_op.cc +++ b/paddle/fluid/operators/cumsum_op.cc @@ -29,8 +29,7 @@ class CumOp : public framework::OperatorWithKernel { class CumsumOpMaker : public framework::OpProtoAndCheckerMaker { public: - CumsumOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "Input of Cumsum operator"); AddOutput("Out", "Output of Cumsum operator"); AddAttr("axis", diff --git a/paddle/fluid/operators/decayed_adagrad_op.cc b/paddle/fluid/operators/decayed_adagrad_op.cc index 5eeb3dee095e330174f35fa8eebd418bf764b132..c0f2b49a04d9e88502c4b63bca493cd2b7ad1c5c 100644 --- a/paddle/fluid/operators/decayed_adagrad_op.cc +++ b/paddle/fluid/operators/decayed_adagrad_op.cc @@ -17,6 +17,7 @@ limitations under the License. */ namespace paddle { namespace operators { +using Tensor = framework::Tensor; class DecayedAdagradOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; @@ -51,12 +52,17 @@ class DecayedAdagradOp : public framework::OperatorWithKernel { ctx->SetOutputDim("ParamOut", param_dims); ctx->SetOutputDim("MomentOut", param_dims); } + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext &ctx) const override { + auto input_data_type = + framework::ToDataType(ctx.Input("Param")->type()); + return framework::OpKernelType(input_data_type, ctx.GetPlace()); + } }; class DecayedAdagradOpMaker : public framework::OpProtoAndCheckerMaker { public: - DecayedAdagradOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Param", "(Tensor) Input parameter"); AddInput("Grad", "(Tensor) Input gradient"); AddInput("Moment", "(Tensor) Second moment"); diff --git a/paddle/fluid/operators/delete_var_op.cc b/paddle/fluid/operators/delete_var_op.cc index 1fe9404c00335edbe3594486f8c403e69f2ab08f..d7a9bfbc437dbf4c723b9c87ff62ec6b62c38638 100644 --- a/paddle/fluid/operators/delete_var_op.cc +++ b/paddle/fluid/operators/delete_var_op.cc @@ -34,8 +34,7 @@ class DeleteVarOp : public framework::OperatorBase { class DeleteVarOpInfoMaker : public framework::OpProtoAndCheckerMaker { public: - DeleteVarOpInfoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The input of delete op").AsDuplicable(); AddComment(R"DOC( Delete Operator. diff --git a/paddle/fluid/operators/detail/CMakeLists.txt b/paddle/fluid/operators/detail/CMakeLists.txt index 719a7465b8d58ef8588ff1e83c2b971eb6fbb00f..b9a66474c9afc27462f9c47af1a0465e2cec70bc 100644 --- a/paddle/fluid/operators/detail/CMakeLists.txt +++ b/paddle/fluid/operators/detail/CMakeLists.txt @@ -4,6 +4,8 @@ if(WITH_DISTRIBUTE) set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor") set_source_files_properties(serde_test.cc grpc_server_test.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) cc_test(serde_test SRCS serde_test.cc variable_response.cc DEPS grpc++_unsecure grpc_unsecure gpr - cares zlib protobuf sendrecvop_grpc) - cc_test(grpc_server_test SRCS grpc_server_test.cc DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib protobuf executor proto_desc lookup_table_op) + cares zlib protobuf sendrecvop_grpc SERIAL) + cc_test(grpc_server_test SRCS grpc_server_test.cc DEPS sendrecvop_grpc + grpc++_unsecure grpc_unsecure gpr cares zlib protobuf executor + proto_desc lookup_table_op SERIAL) endif() diff --git a/paddle/fluid/operators/detail/grpc_client.cc b/paddle/fluid/operators/detail/grpc_client.cc index 661dfa69fe1580ff3890f12defcd124225be0c06..f7ce7786874285795878b655365974f082c00b44 100644 --- a/paddle/fluid/operators/detail/grpc_client.cc +++ b/paddle/fluid/operators/detail/grpc_client.cc @@ -19,11 +19,27 @@ limitations under the License. */ #include #include "paddle/fluid/framework/threadpool.h" +#include "paddle/fluid/platform/profiler.h" namespace paddle { namespace operators { namespace detail { +std::once_flag RPCClient::init_flag_; + +std::unique_ptr RPCClient::rpc_client_(nullptr); + +RPCClient* RPCClient::GetInstance() { + std::call_once(init_flag_, &RPCClient::Init); + return rpc_client_.get(); +} + +void RPCClient::Init() { + if (rpc_client_.get() == nullptr) { + rpc_client_.reset(new RPCClient()); + } +} + bool RPCClient::AsyncSendVariable(const std::string& ep, const platform::DeviceContext& ctx, const framework::Scope& scope, @@ -52,14 +68,13 @@ bool RPCClient::AsyncSendVariable(const std::string& ep, // stub context SendProcessor* s = new SendProcessor(ch); s->Prepare(var_h, time_out); - s->response_call_back_ = NULL; + s->response_call_back_ = nullptr; auto call = s->stub_g_.PrepareUnaryCall( s->context_.get(), "/sendrecv.SendRecvService/SendVariable", req, &cq_); call->StartCall(); call->Finish(&s->reply_, &s->status_, reinterpret_cast(s)); }); - req_count_++; return true; @@ -196,9 +211,14 @@ bool RPCClient::Wait() { const size_t kReqCnt = req_count_; bool a[kReqCnt]; std::vector> waits(req_count_); + std::mutex mu; for (int i = 0; i < req_count_; i++) { - waits[i] = framework::AsyncIO([i, &a, this] { a[i] = Proceed(); }); + waits[i] = framework::AsyncIO([i, &a, &mu, this] { + bool ret = Proceed(); + std::lock_guard l(mu); + a[i] = ret; + }); } for (int i = 0; i < req_count_; i++) { @@ -243,8 +263,9 @@ bool RPCClient::Proceed() { delete c; return true; } - std::shared_ptr RPCClient::GetChannel(const std::string& ep) { + // TODO(Yancey1989): make grpc client completely thread-safe + std::unique_lock lock(mutex_); auto it = channels_.find(ep); if (it != channels_.end()) { return it->second; @@ -257,7 +278,6 @@ std::shared_ptr RPCClient::GetChannel(const std::string& ep) { auto ch = grpc::CreateCustomChannel(ep, grpc::InsecureChannelCredentials(), args); - channels_[ep] = ch; return ch; } diff --git a/paddle/fluid/operators/detail/grpc_client.h b/paddle/fluid/operators/detail/grpc_client.h index f6229b71bc01a6de51f50f5fe880ada6e15e74dd..449d5105afb8c02294a0ef57610e7de1b1631b35 100644 --- a/paddle/fluid/operators/detail/grpc_client.h +++ b/paddle/fluid/operators/detail/grpc_client.h @@ -21,6 +21,7 @@ limitations under the License. */ #include #include #include +#include // NOLINT #include #include @@ -35,6 +36,7 @@ limitations under the License. */ #include "paddle/fluid/framework/scope.h" #include "paddle/fluid/framework/selected_rows.h" #include "paddle/fluid/operators/detail/sendrecvop_utils.h" +#include "paddle/fluid/platform/macros.h" // for DISABLE_COPY_AND_ASSIGN namespace paddle { namespace operators { @@ -57,7 +59,9 @@ void ProcGetResponse(const VarHandle& var_h, const grpc::ByteBuffer& msg); class BaseProcessor { public: - explicit BaseProcessor(std::shared_ptr ch) { context_ = NULL; } + explicit BaseProcessor(std::shared_ptr ch) { + context_ = nullptr; + } virtual ~BaseProcessor() {} @@ -105,7 +109,7 @@ class SendProcessor : public BaseProcessor { ::grpc::GenericStub stub_g_; ::grpc::ByteBuffer reply_; - RequestSendCallBack response_call_back_ = NULL; + RequestSendCallBack response_call_back_ = nullptr; }; typedef std::function @@ -159,6 +163,10 @@ class FetchBarrierProcessor : public BaseProcessor { class RPCClient { public: + RPCClient() {} + + static RPCClient* GetInstance(); + bool AsyncSendVariable(const std::string& ep, const platform::DeviceContext& ctx, const framework::Scope& scope, @@ -189,11 +197,17 @@ class RPCClient { private: bool Proceed(); std::shared_ptr GetChannel(const std::string& ep); + // Init is called by GetInstance. + static void Init(); private: grpc::CompletionQueue cq_; std::map> channels_; - int64_t req_count_ = 0; + std::atomic req_count_{0}; + std::mutex mutex_; + static std::unique_ptr rpc_client_; + static std::once_flag init_flag_; + DISABLE_COPY_AND_ASSIGN(RPCClient); }; } // namespace detail diff --git a/paddle/fluid/operators/detail/grpc_server.cc b/paddle/fluid/operators/detail/grpc_server.cc index 95f4738b4ff50852d9591719133ca650533bf848..361cc24b5ba11e2654f1282327730befaeca9f55 100644 --- a/paddle/fluid/operators/detail/grpc_server.cc +++ b/paddle/fluid/operators/detail/grpc_server.cc @@ -19,10 +19,16 @@ limitations under the License. */ using ::grpc::ServerAsyncResponseWriter; +DEFINE_int32(rpc_server_handle_send_threads, 20, + "Number of threads used to handle send at rpc server."); +DEFINE_int32(rpc_server_handle_get_threads, 20, + "Number of threads used to handle get at rpc server."); +DEFINE_int32(rpc_server_handle_prefetch_threads, 1, + "Number of threads used to handle prefetch at rpc server."); + namespace paddle { namespace operators { namespace detail { - enum CallStatus { PROCESS = 0, FINISH }; // reference: @@ -63,18 +69,20 @@ class RequestSend final : public RequestBase { explicit RequestSend(GrpcService::AsyncService* service, ::grpc::ServerCompletionQueue* cq, bool sync_mode, framework::Scope* scope, ReceivedQueue* queue, - const platform::DeviceContext* dev_ctx) + const platform::DeviceContext* dev_ctx, int req_id) : RequestBase(service, cq, sync_mode, dev_ctx), queue_(queue), - responder_(&ctx_) { + responder_(&ctx_), + req_id_(req_id) { if (sync_mode_) { request_.reset(new VariableResponse(scope, dev_ctx_, false)); } else { request_.reset(new VariableResponse(scope, dev_ctx_, true)); } int method_id = static_cast(detail::GrpcMethod::kSendVariable); - service_->RequestAsyncUnary(method_id, &ctx_, request_.get(), &responder_, - cq_, cq_, this); + service_->RequestAsyncUnary( + method_id, &ctx_, request_.get(), &responder_, cq_, cq_, + reinterpret_cast(static_cast(req_id))); } virtual ~RequestSend() {} @@ -82,17 +90,21 @@ class RequestSend final : public RequestBase { virtual std::string GetReqName() { return request_->Varname(); } virtual void Process() { - queue_->Push(std::make_pair(request_->Varname(), request_)); + std::string var_name = GetReqName(); + VLOG(3) << "RequestSend " << var_name; + queue_->Push(std::make_pair(var_name, request_)); - sendrecv::VoidMessage reply; - responder_.Finish(reply, ::grpc::Status::OK, this); status_ = FINISH; + responder_.Finish(reply_, ::grpc::Status::OK, + reinterpret_cast(static_cast(req_id_))); } protected: + sendrecv::VoidMessage reply_; std::shared_ptr request_; ReceivedQueue* queue_; ServerAsyncResponseWriter responder_; + int req_id_; }; class RequestGet final : public RequestBase { @@ -101,14 +113,17 @@ class RequestGet final : public RequestBase { ::grpc::ServerCompletionQueue* cq, bool sync_mode, framework::Scope* scope, const platform::DeviceContext* dev_ctx, - framework::BlockingQueue* queue) + framework::BlockingQueue* queue, + int req_id) : RequestBase(service, cq, sync_mode, dev_ctx), responder_(&ctx_), scope_(scope), - queue_(queue) { - int method_id = static_cast(detail::GrpcMethod::kGetVariable); - service_->RequestAsyncUnary(method_id, &ctx_, &request_, &responder_, cq_, - cq_, this); + queue_(queue), + req_id_(req_id) { + auto method_id = static_cast(detail::GrpcMethod::kGetVariable); + service_->RequestAsyncUnary( + method_id, &ctx_, &request_, &responder_, cq_, cq_, + reinterpret_cast(static_cast(req_id_))); } virtual ~RequestGet() {} @@ -118,15 +133,16 @@ class RequestGet final : public RequestBase { virtual void Process() { // proc request. std::string var_name = request_.varname(); + VLOG(3) << "RequestGet " << var_name; auto* var = scope_->FindVar(var_name); - ::grpc::ByteBuffer reply; if (var_name != FETCH_BARRIER_MESSAGE) { - SerializeToByteBuffer(var_name, var, *dev_ctx_, &reply); + SerializeToByteBuffer(var_name, var, *dev_ctx_, &reply_); } - responder_.Finish(reply, ::grpc::Status::OK, this); status_ = FINISH; + responder_.Finish(reply_, ::grpc::Status::OK, + reinterpret_cast(static_cast(req_id_))); if (var_name == FETCH_BARRIER_MESSAGE) { sendrecv::VariableMessage msg; @@ -137,9 +153,11 @@ class RequestGet final : public RequestBase { protected: sendrecv::VariableMessage request_; + ::grpc::ByteBuffer reply_; ServerAsyncResponseWriter<::grpc::ByteBuffer> responder_; framework::Scope* scope_; framework::BlockingQueue* queue_; + int req_id_; }; class RequestPrefetch final : public RequestBase { @@ -150,21 +168,21 @@ class RequestPrefetch final : public RequestBase { const platform::DeviceContext* dev_ctx, framework::Executor* executor, framework::ProgramDesc* program, - framework::ExecutorPrepareContext* prefetch_ctx) + framework::ExecutorPrepareContext* prefetch_ctx, + int req_id) : RequestBase(service, cq, sync_mode, dev_ctx), responder_(&ctx_), scope_(scope), executor_(executor), program_(program), - prefetch_ctx_(prefetch_ctx) { - if (sync_mode_) { - request_.reset(new VariableResponse(scope, dev_ctx_, false)); - } else { - request_.reset(new VariableResponse(scope, dev_ctx_, true)); - } + prefetch_ctx_(prefetch_ctx), + req_id_(req_id) { + // prefetch always create a new sub scope + request_.reset(new VariableResponse(scope, dev_ctx_, true)); int method_id = static_cast(detail::GrpcMethod::kPrefetchVariable); - service_->RequestAsyncUnary(method_id, &ctx_, request_.get(), &responder_, - cq_, cq_, this); + service_->RequestAsyncUnary( + method_id, &ctx_, request_.get(), &responder_, cq_, cq_, + reinterpret_cast(static_cast(req_id_))); } virtual ~RequestPrefetch() {} @@ -173,29 +191,31 @@ class RequestPrefetch final : public RequestBase { virtual void Process() { // prefetch process... - ::grpc::ByteBuffer reply; std::string var_name = request_->OutVarname(); - VLOG(3) << "prefetch var " << var_name; + VLOG(3) << "RequestPrefetch " << var_name; auto var_desc = program_->Block(0).FindVar(var_name); - framework::Scope* local_scope = &scope_->NewScope(); + framework::Scope* local_scope = request_->GetMutableLocalScope(); auto* var = local_scope->FindVar(var_name); InitializeVariable(var, var_desc->GetType()); - executor_->RunPreparedContext(prefetch_ctx_, scope_, false, false); + executor_->RunPreparedContext(prefetch_ctx_, local_scope); - SerializeToByteBuffer(var_name, var, *dev_ctx_, &reply); + SerializeToByteBuffer(var_name, var, *dev_ctx_, &reply_); - responder_.Finish(reply, ::grpc::Status::OK, this); status_ = FINISH; + responder_.Finish(reply_, ::grpc::Status::OK, + reinterpret_cast(static_cast(req_id_))); } protected: std::shared_ptr request_; + ::grpc::ByteBuffer reply_; ServerAsyncResponseWriter<::grpc::ByteBuffer> responder_; framework::Scope* scope_; framework::Executor* executor_; framework::ProgramDesc* program_; framework::ExecutorPrepareContext* prefetch_ctx_; + int req_id_; }; void AsyncGRPCServer::WaitClientGet(int count) { @@ -208,6 +228,11 @@ void AsyncGRPCServer::WaitClientGet(int count) { } } +void AsyncGRPCServer::WaitServerReady() { + std::unique_lock lock(this->mutex_ready_); + condition_ready_.wait(lock, [=] { return this->ready_ == 1; }); +} + void AsyncGRPCServer::RunSyncUpdate() { ::grpc::ServerBuilder builder; builder.AddListeningPort(address_, ::grpc::InsecureServerCredentials(), @@ -224,28 +249,55 @@ void AsyncGRPCServer::RunSyncUpdate() { LOG(INFO) << "Server listening on " << address_ << " selected port: " << selected_port_; - std::function send_register = - std::bind(&AsyncGRPCServer::TryToRegisterNewSendOne, this); - std::function get_register = - std::bind(&AsyncGRPCServer::TryToRegisterNewGetOne, this); - std::function prefetch_register = - std::bind(&AsyncGRPCServer::TryToRegisterNewPrefetchOne, this); - - // TODO(wuyi): Run these "HandleRequest" in thread pool - t_send_.reset( - new std::thread(std::bind(&AsyncGRPCServer::HandleRequest, this, - cq_send_.get(), "cq_send", send_register))); - t_get_.reset( - new std::thread(std::bind(&AsyncGRPCServer::HandleRequest, this, - cq_get_.get(), "cq_get", get_register))); - t_prefetch_.reset(new std::thread( - std::bind(&AsyncGRPCServer::HandleRequest, this, cq_prefetch_.get(), - "cq_prefetch", prefetch_register))); + std::function send_register = std::bind( + &AsyncGRPCServer::TryToRegisterNewSendOne, this, std::placeholders::_1); + std::function get_register = std::bind( + &AsyncGRPCServer::TryToRegisterNewGetOne, this, std::placeholders::_1); + std::function prefetch_register = + std::bind(&AsyncGRPCServer::TryToRegisterNewPrefetchOne, this, + std::placeholders::_1); + + for (int i = 0; i < kSendReqsBufSize; ++i) { + TryToRegisterNewSendOne(i); + } + for (int i = 0; i < kGetReqsBufSize; ++i) { + TryToRegisterNewGetOne(i); + } + for (int i = 0; i < kPrefetchReqsBufSize; ++i) { + TryToRegisterNewPrefetchOne(i); + } + + for (int i = 0; i < FLAGS_rpc_server_handle_send_threads; ++i) { + t_sends_.emplace_back( + new std::thread(std::bind(&AsyncGRPCServer::HandleRequest, this, + cq_send_.get(), "cq_send", send_register))); + } + for (int i = 0; i < FLAGS_rpc_server_handle_get_threads; ++i) { + t_gets_.emplace_back( + new std::thread(std::bind(&AsyncGRPCServer::HandleRequest, this, + cq_get_.get(), "cq_get", get_register))); + } + for (int i = 0; i < FLAGS_rpc_server_handle_prefetch_threads; ++i) { + t_prefetchs_.emplace_back(new std::thread( + std::bind(&AsyncGRPCServer::HandleRequest, this, cq_prefetch_.get(), + "cq_prefetch", prefetch_register))); + } + { + std::lock_guard lock(this->mutex_ready_); + ready_ = 1; + } + condition_ready_.notify_all(); // wait server server_->Wait(); - t_send_->join(); - t_get_->join(); - t_prefetch_->join(); + for (int i = 0; i < FLAGS_rpc_server_handle_send_threads; ++i) { + t_sends_[i]->join(); + } + for (int i = 0; i < FLAGS_rpc_server_handle_get_threads; ++i) { + t_gets_[i]->join(); + } + for (int i = 0; i < FLAGS_rpc_server_handle_prefetch_threads; ++i) { + t_prefetchs_[i]->join(); + } } void AsyncGRPCServer::ShutdownQueue() { @@ -262,66 +314,78 @@ void AsyncGRPCServer::ShutDown() { server_->Shutdown(); } -void AsyncGRPCServer::TryToRegisterNewSendOne() { +void AsyncGRPCServer::TryToRegisterNewSendOne(int i) { std::unique_lock lock(cq_mutex_); if (is_shut_down_) { VLOG(3) << "shutdown, do not TryToRegisterNewSendOne"; return; } RequestSend* send = new RequestSend(&service_, cq_send_.get(), sync_mode_, - scope_, &var_recv_queue_, dev_ctx_); + scope_, &var_recv_queue_, dev_ctx_, i); + send_reqs_[i] = static_cast(send); VLOG(4) << "Create RequestSend status:" << send->Status(); } -void AsyncGRPCServer::TryToRegisterNewGetOne() { +void AsyncGRPCServer::TryToRegisterNewGetOne(int req_id) { std::unique_lock lock(cq_mutex_); if (is_shut_down_) { VLOG(3) << "shutdown, do not TryToRegisterNewGetOne"; return; } RequestGet* get = new RequestGet(&service_, cq_get_.get(), sync_mode_, scope_, - dev_ctx_, &var_get_queue_); + dev_ctx_, &var_get_queue_, req_id); + get_reqs_[req_id] = static_cast(get); VLOG(4) << "Create RequestGet status:" << get->Status(); } -void AsyncGRPCServer::TryToRegisterNewPrefetchOne() { +void AsyncGRPCServer::TryToRegisterNewPrefetchOne(int req_id) { std::unique_lock lock(cq_mutex_); if (is_shut_down_) { VLOG(3) << "shutdown, do not TryToRegisterNewPrefetchOne"; return; } - RequestPrefetch* prefetch = - new RequestPrefetch(&service_, cq_prefetch_.get(), sync_mode_, scope_, - dev_ctx_, executor_, program_, prefetch_ctx_); + RequestPrefetch* prefetch = new RequestPrefetch( + &service_, cq_prefetch_.get(), sync_mode_, scope_, dev_ctx_, executor_, + program_, prefetch_ctx_.get(), req_id); + prefetch_reqs_[req_id] = static_cast(prefetch); VLOG(4) << "Create RequestPrefetch status:" << prefetch->Status(); } // FIXME(typhoonzero): change cq_name to enum. -void AsyncGRPCServer::HandleRequest(::grpc::ServerCompletionQueue* cq, - const std::string& cq_name, - std::function TryToRegisterNewOne) { - TryToRegisterNewOne(); - +void AsyncGRPCServer::HandleRequest( + ::grpc::ServerCompletionQueue* cq, const std::string& cq_name, + std::function TryToRegisterNewOne) { void* tag = NULL; bool ok = false; while (true) { - VLOG(3) << "HandleRequest for " << cq_name << " while in"; + VLOG(3) << "HandleRequest for " << cq_name << " wait Next"; if (!cq->Next(&tag, &ok)) { LOG(INFO) << cq_name << " CompletionQueue shutdown!"; break; } - VLOG(3) << "HandleRequest for " << cq_name << " while after Next"; + VLOG(3) << "HandleRequest for " << cq_name << " get Next"; + int req_id = static_cast(reinterpret_cast(tag)); - PADDLE_ENFORCE(tag); if (sync_mode_) { // FIXME(typhoonzero): de-couple the barriers with recv_op if (!is_shut_down_ && cq_name == "cq_get") WaitCond(1); if (!is_shut_down_ && cq_name == "cq_send") WaitCond(0); + VLOG(3) << "HandleRequest for " << cq_name << " after WaitCond"; } - RequestBase* base = reinterpret_cast(tag); + RequestBase* base = nullptr; + { + std::lock_guard l(cq_mutex_); + if (cq_name == "cq_get") { + base = get_reqs_[req_id]; + } else if (cq_name == "cq_send") { + base = send_reqs_[req_id]; + } else if (cq_name == "cq_prefetch") { + base = prefetch_reqs_[req_id]; + } + } // reference: // https://github.com/tensorflow/tensorflow/issues/5596 // https://groups.google.com/forum/#!topic/grpc-io/xftlRy-IQwM @@ -329,19 +393,19 @@ void AsyncGRPCServer::HandleRequest(::grpc::ServerCompletionQueue* cq, if (!ok) { LOG(WARNING) << cq_name << " recv no regular event:argument name[" << base->GetReqName() << "]"; - TryToRegisterNewOne(); + TryToRegisterNewOne(req_id); delete base; continue; } switch (base->Status()) { case PROCESS: { - VLOG(4) << cq_name << " PROCESS status:" << base->Status(); - TryToRegisterNewOne(); base->Process(); + VLOG(4) << cq_name << " PROCESS status:" << base->Status(); break; } case FINISH: { + TryToRegisterNewOne(req_id); VLOG(4) << cq_name << " FINISH status:" << base->Status(); delete base; break; diff --git a/paddle/fluid/operators/detail/grpc_server.h b/paddle/fluid/operators/detail/grpc_server.h index 99b87b8c6cb3e597778b88c395e4abf400d82c39..bdff9801a928699f8391bfb68c1c7bd2d75aa642 100644 --- a/paddle/fluid/operators/detail/grpc_server.h +++ b/paddle/fluid/operators/detail/grpc_server.h @@ -17,6 +17,7 @@ limitations under the License. */ #include #include // NOLINT #include +#include #include "grpc++/grpc++.h" #include "paddle/fluid/framework/blocking_queue.h" @@ -30,6 +31,7 @@ limitations under the License. */ #include "paddle/fluid/operators/detail/send_recv.grpc.pb.h" #include "paddle/fluid/operators/detail/send_recv.pb.h" #include "paddle/fluid/operators/detail/sendrecvop_utils.h" +#include "paddle/fluid/platform/profiler.h" namespace paddle { namespace operators { @@ -45,8 +47,10 @@ class RequestBase; class AsyncGRPCServer final { public: explicit AsyncGRPCServer(const std::string &address, bool sync_mode) - : address_(address), sync_mode_(sync_mode) {} + : address_(address), sync_mode_(sync_mode), ready_(0) {} + ~AsyncGRPCServer() {} + void WaitServerReady(); void RunSyncUpdate(); // functions to sync server barrier status. @@ -62,8 +66,9 @@ class AsyncGRPCServer final { void SetExecutor(framework::Executor *executor) { executor_ = executor; } - void SetPrefetchPreparedCtx(framework::ExecutorPrepareContext *prepared) { - prefetch_ctx_ = prepared; + void SetPrefetchPreparedCtx( + std::unique_ptr prepared) { + prefetch_ctx_.reset(prepared.release()); } int GetSelectedPort() const { return selected_port_; } @@ -79,19 +84,27 @@ class AsyncGRPCServer final { protected: void HandleRequest(::grpc::ServerCompletionQueue *cq, const std::string &cq_name, - std::function TryToRegisterNewOne); - void TryToRegisterNewSendOne(); - void TryToRegisterNewGetOne(); - void TryToRegisterNewPrefetchOne(); + std::function TryToRegisterNewOne); + void TryToRegisterNewSendOne(int req_id); + void TryToRegisterNewGetOne(int req_id); + void TryToRegisterNewPrefetchOne(int req_id); void ShutdownQueue(); private: + static const int kSendReqsBufSize = 100; + static const int kGetReqsBufSize = 100; + static const int kPrefetchReqsBufSize = 10; + std::mutex cq_mutex_; volatile bool is_shut_down_ = false; std::unique_ptr<::grpc::ServerCompletionQueue> cq_send_; std::unique_ptr<::grpc::ServerCompletionQueue> cq_get_; std::unique_ptr<::grpc::ServerCompletionQueue> cq_prefetch_; + RequestBase *send_reqs_[kSendReqsBufSize]; + RequestBase *get_reqs_[kGetReqsBufSize]; + RequestBase *prefetch_reqs_[kPrefetchReqsBufSize]; + GrpcService::AsyncService service_; std::unique_ptr<::grpc::Server> server_; @@ -110,14 +123,20 @@ class AsyncGRPCServer final { mutable int barrier_cond_step_; std::condition_variable barrier_condition_; - std::unique_ptr t_send_; - std::unique_ptr t_get_; + std::vector> t_sends_; + std::vector> t_gets_; + std::vector> t_prefetchs_; + std::unique_ptr t_prefetch_; - framework::ExecutorPrepareContext *prefetch_ctx_; + std::unique_ptr prefetch_ctx_; framework::ProgramDesc *program_; framework::Executor *executor_; int selected_port_; + + std::mutex mutex_ready_; + std::condition_variable condition_ready_; + int ready_; }; }; // namespace detail diff --git a/paddle/fluid/operators/detail/grpc_server_test.cc b/paddle/fluid/operators/detail/grpc_server_test.cc index 25b95d608d10d6e456d5f563ce9fbe35d812cb0f..350a7ee1234da5b88d09ea955ce14b7c161d804e 100644 --- a/paddle/fluid/operators/detail/grpc_server_test.cc +++ b/paddle/fluid/operators/detail/grpc_server_test.cc @@ -100,7 +100,7 @@ void StartServer(const std::string& endpoint) { InitTensorsOnServer(&scope, &place, 10); rpc_service_->SetProgram(&program); - rpc_service_->SetPrefetchPreparedCtx(prepared.get()); + rpc_service_->SetPrefetchPreparedCtx(std::move(prepared)); rpc_service_->SetDevCtx(&ctx); rpc_service_->SetScope(&scope); rpc_service_->SetExecutor(&exe); @@ -108,7 +108,7 @@ void StartServer(const std::string& endpoint) { rpc_service_->RunSyncUpdate(); } -TEST(PREFETCH, CPU) { +TEST(PREFETCH, DISABLED_CPU) { // start up a server instance backend std::thread server_thread(StartServer, "127.0.0.1:8889"); sleep(2); @@ -121,10 +121,10 @@ TEST(PREFETCH, CPU) { std::string in_var_name("ids"); std::string out_var_name("out"); - detail::RPCClient client; - client.AsyncPrefetchVariable("127.0.0.1:8889", ctx, scope, in_var_name, - out_var_name); - client.Wait(); + auto client = detail::RPCClient::GetInstance(); + client->AsyncPrefetchVariable("127.0.0.1:8889", ctx, scope, in_var_name, + out_var_name); + client->Wait(); auto var = scope.Var(out_var_name); auto value = var->GetMutable()->value(); diff --git a/paddle/fluid/operators/detail/grpc_service.h b/paddle/fluid/operators/detail/grpc_service.h index e6dab2f5a3a4280f3979417c3ca2d884a0b8ff2f..e0505c2b9d0903837713d7e0032b01ab091c2e04 100644 --- a/paddle/fluid/operators/detail/grpc_service.h +++ b/paddle/fluid/operators/detail/grpc_service.h @@ -25,6 +25,8 @@ #include #include "paddle/fluid/operators/detail/variable_response.h" +#include "paddle/fluid/platform/profiler.h" + // NOTE: This method was originally created by tensorflow // (https://github.com/tensorflow/tensorflow/) we borrow this // method and did some modifications so that we can parse gRPC diff --git a/paddle/fluid/operators/detail/send_recv.proto b/paddle/fluid/operators/detail/send_recv.proto index 02bb2b9cebb87b83aa1cbef0c644f969b4d17284..a244afc46f3247c7e6e8481b09b5c729a2a569f7 100644 --- a/paddle/fluid/operators/detail/send_recv.proto +++ b/paddle/fluid/operators/detail/send_recv.proto @@ -32,6 +32,7 @@ service SendRecvService { enum VarType { LOD_TENSOR = 0; SELECTED_ROWS = 1; + NCCL_ID = 2; } // NOTICE(gongwb):don't modify this proto if you are not @@ -69,6 +70,10 @@ message VariableMessage { bytes rows = 9; // Look up table block execution output variable name. string out_varname = 10; + // If 1, the ps server will start profiling, the ps + // server stops profiling and generates a profile to /tmp/profile_ps_* + // when profile switches from 1 to 2. + int64 profile = 11; } message VoidMessage {} diff --git a/paddle/fluid/operators/detail/sendrecvop_utils.cc b/paddle/fluid/operators/detail/sendrecvop_utils.cc index 766bcf1ac5e06628638fcc8a305c00ab2795bbf2..507b465435609a91ebca97dd70b176c3b79bee02 100644 --- a/paddle/fluid/operators/detail/sendrecvop_utils.cc +++ b/paddle/fluid/operators/detail/sendrecvop_utils.cc @@ -14,6 +14,9 @@ limitations under the License. */ #include "paddle/fluid/operators/detail/sendrecvop_utils.h" +#ifdef PADDLE_WITH_CUDA +#include +#endif #include #include // NOLINT @@ -23,127 +26,164 @@ limitations under the License. */ #include "paddle/fluid/operators/detail/bytebuffer_stream.h" #include "paddle/fluid/operators/detail/proto_encoder_helper.h" #include "paddle/fluid/operators/detail/variable_response.h" +#include "paddle/fluid/platform/profiler.h" namespace paddle { namespace operators { namespace detail { +using VarMsg = sendrecv::VariableMessage; + +void GetTensorPayload(framework::Variable* var, + const platform::DeviceContext& ctx, VarMsg* request, + void** payload, size_t* payload_size) { + auto tensor = var->Get(); + // FIXME(wuyi): data types in send_recv.proto is copied from + // framework.proto + request->set_data_type( + static_cast(framework::ToDataType(tensor.type()))); + for (auto& dim : framework::vectorize(tensor.dims())) { + request->add_dims(dim); + } + const framework::LoD lod = tensor.lod(); + if (lod.size() > 0) { + request->set_lod_level(lod.size()); + for (auto& each : lod) { + VarMsg::LodData* lod_inner = request->add_lod(); + for (auto& d : each) { + lod_inner->add_lod_data(d); + } + } + } + if (platform::is_gpu_place(ctx.GetPlace())) { +#ifdef PADDLE_WITH_CUDA + PADDLE_ENFORCE(platform::is_gpu_place(tensor.place())); + platform::CUDAPinnedPlace cuda_pinned; + auto& gpu_dev_ctx = static_cast(ctx); + auto copy_size = tensor.numel() * framework::SizeOfType(tensor.type()); + *payload = memory::Alloc(cuda_pinned, copy_size); + + memory::Copy(cuda_pinned, *payload, + boost::get(tensor.place()), + reinterpret_cast(tensor.data()), copy_size, + gpu_dev_ctx.stream()); + ctx.Wait(); +#endif + } else { + *payload = tensor.data(); + } + *payload_size = tensor.numel() * framework::SizeOfType(tensor.type()); +} + +void GetSelectedRowsPayload(framework::Variable* var, + const platform::DeviceContext& ctx, VarMsg* request, + void** payload, size_t* payload_size) { + auto* slr = var->GetMutable(); + request->set_data_type( + static_cast(framework::ToDataType(slr->value().type()))); + request->set_lod_level(0); + request->set_slr_height(slr->height()); + + for (auto& dim : framework::vectorize(slr->value().dims())) { + request->add_dims(dim); + } + + auto* tensor = slr->mutable_value(); + if (platform::is_gpu_place(ctx.GetPlace())) { +#ifdef PADDLE_WITH_CUDA + platform::CUDAPinnedPlace cuda_pinned; + auto& gpu_dev_ctx = static_cast(ctx); + auto copy_size = tensor->numel() * framework::SizeOfType(tensor->type()); + *payload = memory::Alloc(cuda_pinned, copy_size); + memory::Copy(cuda_pinned, *payload, + boost::get(tensor->place()), + reinterpret_cast(tensor->data()), copy_size, + gpu_dev_ctx.stream()); + ctx.Wait(); +#endif + } else { + *payload = slr->mutable_value()->data(); + } + *payload_size = tensor->numel() * framework::SizeOfType(tensor->type()); +} + void SerializeToByteBuffer(const std::string& name, framework::Variable* var, const platform::DeviceContext& ctx, ::grpc::ByteBuffer* msg, const std::string& out_name) { - using VarMsg = sendrecv::VariableMessage; - // When using GPU, need to free the copied CPU buffer - // when the ByteBuffer destroies - // TODO(typhoonzero): add unref here, if we have dependent - // parallelism execution, need to know when to free the tensor. + // Default DestroyCallback does nothing, When using GPU + // the CPU buffer need to be freed. DestroyCallback destroy_callback = [](void* backing) {}; - - auto buffer = std::unique_ptr(new char[1024]); - void* buf = buffer.get(); - + VarMsg request; void* payload = nullptr; size_t payload_size; - ProtoEncodeHelper e(static_cast(buf), 1024); - e.WriteString(VarMsg::kVarnameFieldNumber, name); - if (var->IsType()) { - e.WriteUint64(VarMsg::kTypeFieldNumber, 0); - } else if (var->IsType()) { - e.WriteUint64(VarMsg::kTypeFieldNumber, 1); - } + request.set_varname(name); + // Note: normally the profiler is enabled in 1 trainer, hence only + // 1 trainer returns true for ShouldSendProfileState(). It tells PS + // servers the trainer's profiling state so that PS can follow the + // trainer. + if (platform::ShouldSendProfileState()) { + if (platform::IsProfileEnabled()) { + request.set_profile(platform::kEnableProfiler); + } else { + request.set_profile(platform::kDisableProfiler); + } + } if (!out_name.empty()) { - e.WriteString(VarMsg::kOutVarnameFieldNumber, out_name); + request.set_out_varname(out_name); } - switch (framework::ToVarType(var->Type())) { - case framework::proto::VarType_Type_LOD_TENSOR: { - auto tensor = var->Get(); - e.WriteUint64(VarMsg::kDataTypeFieldNumber, - framework::ToDataType(tensor.type())); - for (auto& dim : framework::vectorize(tensor.dims())) { - e.WriteUint64(VarMsg::kDimsFieldNumber, dim); - } - auto lod = tensor.lod(); // std::vector> - if (lod.size() > 0) { - e.WriteUint64(VarMsg::kLodLevelFieldNumber, lod.size()); - - for (auto& each : lod) { - e.WriteVarlengthBeginning(VarMsg::kLodFieldNumber, - 2 + // tag + varintlength of submessage - 1 + // kLodDataFieldNumber - each.size()); - // auto copied from GPU - for (auto& d : each) { - e.WriteUint64(VarMsg::LodData::kLodDataFieldNumber, d); - } - } - } - if (platform::is_gpu_place(ctx.GetPlace())) { + if (var->IsType()) { + request.set_type(::sendrecv::LOD_TENSOR); + GetTensorPayload(var, ctx, &request, &payload, &payload_size); + } else if (var->IsType()) { + request.set_type(::sendrecv::SELECTED_ROWS); + GetSelectedRowsPayload(var, ctx, &request, &payload, &payload_size); #ifdef PADDLE_WITH_CUDA - PADDLE_ENFORCE(platform::is_gpu_place(tensor.place())); - platform::CPUPlace cpu; - auto& gpu_dev_ctx = - static_cast(ctx); - auto copy_size = tensor.numel() * framework::SizeOfType(tensor.type()); - payload = memory::Alloc(cpu, copy_size); - - memory::Copy(cpu, payload, - boost::get(tensor.place()), - reinterpret_cast(tensor.data()), - copy_size, gpu_dev_ctx.stream()); - ctx.Wait(); - destroy_callback = [](void* backing) { - platform::CPUPlace cpu; - memory::Free(cpu, backing); - }; - + } else if (var->IsType()) { + request.set_type(::sendrecv::NCCL_ID); #endif - } else { - payload = tensor.data(); - } - payload_size = tensor.numel() * framework::SizeOfType(tensor.type()); - e.WriteVarlengthBeginning(VarMsg::kSerializedFieldNumber, payload_size); - } break; - case framework::proto::VarType_Type_SELECTED_ROWS: { - // TODO(typhoonzero): selectedrows implement should not use unique_ptr - auto* slr = var->GetMutable(); - e.WriteUint64(VarMsg::kDataTypeFieldNumber, - framework::ToDataType(slr->value().type())); - for (auto& dim : framework::vectorize(slr->value().dims())) { - e.WriteUint64(VarMsg::kDimsFieldNumber, dim); - } - e.WriteUint64(VarMsg::kLodLevelFieldNumber, 0); - e.WriteUint64(VarMsg::kSlrHeightFieldNumber, slr->height()); - auto* tensor = slr->mutable_value(); - if (platform::is_gpu_place(ctx.GetPlace())) { + } else { + PADDLE_THROW("Serialize does not support type: %s", + typeid(var->Type()).name()); + } + + if (platform::is_gpu_place(ctx.GetPlace())) { #ifdef PADDLE_WITH_CUDA - platform::CPUPlace cpu; - auto& gpu_dev_ctx = - static_cast(ctx); - auto copy_size = - tensor->numel() * framework::SizeOfType(tensor->type()); - payload = memory::Alloc(cpu, copy_size); - memory::Copy(cpu, payload, - boost::get(tensor->place()), - reinterpret_cast(tensor->data()), - copy_size, gpu_dev_ctx.stream()); - ctx.Wait(); - destroy_callback = [](void* backing) { - platform::CPUPlace cpu; - memory::Free(cpu, backing); - }; + // GPU data is copied to CPU buffer when sending, + // free the buffer when possible. + destroy_callback = [](void* backing) { + platform::CUDAPinnedPlace cuda_pinned; + memory::Free(cuda_pinned, backing); + }; #endif - } else { - payload = slr->mutable_value()->data(); - } - payload_size = tensor->numel() * framework::SizeOfType(tensor->type()); - e.WriteVarlengthBeginning(VarMsg::kSerializedFieldNumber, payload_size); - } break; - default: - PADDLE_THROW("Serialize does not support type: %s", - typeid(var->Type()).name()); - break; } + + std::string header; + request.AppendToString(&header); + auto buffer = std::unique_ptr(new char[1024]); + void* buf = buffer.get(); + ProtoEncodeHelper e(static_cast(buf), 1024); + e.WriteRawBytes(std::string(header.data(), header.size())); +// NCCLID is copied directly to the message, return bytebuffer +// with only one slice if serializing NCCLID. +#ifdef PADDLE_WITH_CUDA + if (var->IsType()) { + e.WriteVarlengthBeginning(VarMsg::kSerializedFieldNumber, + NCCL_UNIQUE_ID_BYTES); + const ncclUniqueId& uid = var->Get(); + e.WriteRawBytes(std::string(uid.internal, NCCL_UNIQUE_ID_BYTES)); + + // for serialize NCCL_ID + ::grpc::Slice slices(e.size()); + memcpy(const_cast(slices.begin()), e.data(), e.size()); + ::grpc::ByteBuffer tmp(&slices, 1); + msg->Swap(&tmp); + return; + } +#endif + + e.WriteVarlengthBeginning(VarMsg::kSerializedFieldNumber, payload_size); // steal reference of tensor data ::grpc::Slice slices[4]; // metadata, tensor, rows meta, rows int num_slices = 2; // only SelectedRows have rows buffer @@ -154,12 +194,9 @@ void SerializeToByteBuffer(const std::string& name, framework::Variable* var, static_cast(payload)), ::grpc::Slice::STEAL_REF); - if (framework::ToVarType(var->Type()) == - framework::proto::VarType_Type_SELECTED_ROWS) { + if (var->IsType()) { auto* slr = var->GetMutable(); - ProtoEncodeHelper e2(static_cast(buf), 128); - // NOTE: rows is of type int64_t size_t rows_memory_size = slr->rows().size() * framework::SizeOfType(typeid(int64_t)); e2.WriteVarlengthBeginning(VarMsg::kRowsFieldNumber, rows_memory_size); @@ -170,10 +207,7 @@ void SerializeToByteBuffer(const std::string& name, framework::Variable* var, grpc_slice_new_with_user_data( const_cast( reinterpret_cast(slr->rows().data())), - rows_memory_size, - [](void* backing) { - // TODO(typhoonzero): add unref here, same as above. - }, + rows_memory_size, [](void* backing) {}, const_cast( reinterpret_cast(slr->rows().data()))), ::grpc::Slice::STEAL_REF); diff --git a/paddle/fluid/operators/detail/serde_test.cc b/paddle/fluid/operators/detail/serde_test.cc index 221d2f4c5b30aef022a5d6b54cd657d1dec1f5a2..15892295e6901fe649788c9e34604008fc8cbdfa 100644 --- a/paddle/fluid/operators/detail/serde_test.cc +++ b/paddle/fluid/operators/detail/serde_test.cc @@ -108,7 +108,7 @@ void RunSerdeTestSelectedRows(platform::Place place) { EXPECT_FLOAT_EQ(tensor_data2[i], 32.7); } for (size_t i = 0; i < rows2->size(); ++i) { - EXPECT_EQ(rows_data2[i], i); + EXPECT_EQ(rows_data2[i], static_cast(i)); } EXPECT_EQ(slr2->height(), 1000); } @@ -117,11 +117,11 @@ void RunTestLodTensor(platform::Place place, int from_type = 0) { // serialize var to ByteBuffer framework::Variable var; auto* tensor = var.GetMutable(); - tensor->Resize(framework::make_ddim({4, 8, 4, 2})); + tensor->Resize(framework::make_ddim({512, 8, 4, 2})); framework::LoD lod; lod.push_back(framework::Vector({1, 3, 8})); tensor->set_lod(lod); - int tensor_numel = 4 * 8 * 4 * 2; + int tensor_numel = 512 * 8 * 4 * 2; platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); auto& ctx = *pool.Get(place); tensor->mutable_data(place); @@ -142,7 +142,7 @@ void RunTestLodTensor(platform::Place place, int from_type = 0) { EXPECT_TRUE(varmsg.ParseFromString(tmp)); EXPECT_EQ(varmsg.varname(), "myvar"); EXPECT_EQ(varmsg.type(), 0); - EXPECT_EQ(varmsg.dims()[0], 4); + EXPECT_EQ(varmsg.dims()[0], 512); EXPECT_EQ(varmsg.dims()[1], 8); EXPECT_EQ(varmsg.dims()[2], 4); EXPECT_EQ(varmsg.dims()[3], 2); diff --git a/paddle/fluid/operators/detail/variable_response.cc b/paddle/fluid/operators/detail/variable_response.cc index fbef8d02a4d765052fccf3792ebe0373d46b1ef6..24cb91a3bb820a0e5d51aaa49154434919080f69 100644 --- a/paddle/fluid/operators/detail/variable_response.cc +++ b/paddle/fluid/operators/detail/variable_response.cc @@ -17,6 +17,10 @@ #include #include #include +#ifdef PADDLE_WITH_CUDA +#include +#endif +#include "paddle/fluid/platform/profiler.h" #include "paddle/fluid/operators/detail/send_recv.pb.h" #include "paddle/fluid/operators/detail/sendrecvop_utils.h" @@ -209,15 +213,15 @@ bool ParseLodData(::google::protobuf::io::CodedInputStream* input, } if (wt == WIRETYPE_LENGTH_DELIMITED) { - int length = 0; - if (!input->ReadVarintSizeAsInt(&length)) { + int num_bytes = 0; + if (!input->ReadVarintSizeAsInt(&num_bytes)) { return tag; } - - for (int i = 0; i < length; i++) { + int start_pos = input->CurrentPosition(); + while (input->CurrentPosition() - start_pos < num_bytes) { uint64_t v; if (!input->ReadVarint64(&v)) { - return false; + return tag; } lod->push_back(v); } @@ -274,8 +278,8 @@ int VariableResponse::Parse(Source* source) { break; } case sendrecv::VariableMessage::kTypeFieldNumber: { - uint64_t v; - if ((wt != WIRETYPE_VARINT) || !input.ReadVarint64(&v)) { + uint32_t v; + if ((wt != WIRETYPE_VARINT) || !input.ReadVarint32(&v)) { return tag; } @@ -283,8 +287,8 @@ int VariableResponse::Parse(Source* source) { break; } case sendrecv::VariableMessage::kDataTypeFieldNumber: { - uint64_t v = 0; - if ((wt != WIRETYPE_VARINT) || !input.ReadVarint64(&v)) { + uint32_t v = 0; + if ((wt != WIRETYPE_VARINT) || !input.ReadVarint32(&v)) { return tag; } @@ -304,11 +308,12 @@ int VariableResponse::Parse(Source* source) { // packed if (wt == WIRETYPE_LENGTH_DELIMITED) { - int length = 0; - if (!input.ReadVarintSizeAsInt(&length)) { + int num_bytes = 0; + if (!input.ReadVarintSizeAsInt(&num_bytes)) { return tag; } - for (int i = 0; i < length; i++) { + int start_pos = input.CurrentPosition(); + while (input.CurrentPosition() - start_pos < num_bytes) { uint64_t v; if (!input.ReadVarint64(&v)) { return tag; @@ -317,7 +322,6 @@ int VariableResponse::Parse(Source* source) { } break; } - return tag; } case sendrecv::VariableMessage::kLodLevelFieldNumber: { @@ -367,28 +371,45 @@ int VariableResponse::Parse(Source* source) { } case sendrecv::VariableMessage::kSerializedFieldNumber: { PADDLE_ENFORCE((meta_.type() == sendrecv::SELECTED_ROWS || - meta_.type() == sendrecv::LOD_TENSOR) && + meta_.type() == sendrecv::LOD_TENSOR || + meta_.type() == sendrecv::NCCL_ID) && meta_.varname() != "", "meta info should be got first!"); - int length = 0; + int num_bytes = 0; if (wt != WIRETYPE_LENGTH_DELIMITED || - !ReadVarintSizeAsInt(&input, &length)) { + !ReadVarintSizeAsInt(&input, &num_bytes)) { return tag; } + if (meta_.type() == sendrecv::NCCL_ID) { +#ifdef PADDLE_WITH_CUDA + auto* var = scope_->FindVar(meta_.varname()); + if (var != nullptr) { + ncclUniqueId* id = var->GetMutable(); + if (!ReadRaw(&input, *dev_ctx_, platform::CPUPlace(), id->internal, + num_bytes)) { + return tag; + } + } + break; +#else + PADDLE_THROW("Not compiled with CUDA!"); +#endif + } + framework::DDim dims = GetDims(meta_.dims()); if (meta_.type() == sendrecv::LOD_TENSOR) { PADDLE_ENFORCE(meta_.lod_size() >= 0, "lod info should be got first!"); - if (!CopyLodTensorData(&input, *dev_ctx_, dims, length)) { + if (!CopyLodTensorData(&input, *dev_ctx_, dims, num_bytes)) { return tag; } break; } if (meta_.type() == sendrecv::SELECTED_ROWS) { - if (!CopySelectRowsTensorData(&input, *dev_ctx_, dims, length)) { + if (!CopySelectRowsTensorData(&input, *dev_ctx_, dims, num_bytes)) { return tag; } break; @@ -402,13 +423,13 @@ int VariableResponse::Parse(Source* source) { meta_.varname() != "", "meta info should be got first!"); - int length = 0; + int num_bytes = 0; if (wt != WIRETYPE_LENGTH_DELIMITED || - !ReadVarintSizeAsInt(&input, &length)) { + !ReadVarintSizeAsInt(&input, &num_bytes)) { return tag; } - if (!CopySelectRowsData(&input, *dev_ctx_, length)) { + if (!CopySelectRowsData(&input, *dev_ctx_, num_bytes)) { return tag; } break; @@ -427,7 +448,28 @@ int VariableResponse::Parse(Source* source) { meta_.set_out_varname(temp); break; } - + case sendrecv::VariableMessage::kProfileFieldNumber: { + uint64_t profiling = 0; + if (!input.ReadVarint64(&profiling)) { + return tag; + } + meta_.set_profile(profiling); + int64_t listener_id = platform::ListenerId(); + if (listener_id <= 0) { + break; + } + if (profiling == platform::kEnableProfiler && + !platform::IsProfileEnabled()) { + platform::EnableProfiler(platform::ProfilerState::kCPU); + } else if (profiling == platform::kDisableProfiler && + platform::IsProfileEnabled()) { + // TODO(panyx0718): Should we allow to customize file dir. + platform::DisableProfiler( + platform::EventSortingKey::kDefault, + string::Sprintf("/tmp/profile_ps_%lld", listener_id)); + } + break; + } default: { // Unknown tag, return unknown error. return -1; diff --git a/paddle/fluid/operators/detection/CMakeLists.txt b/paddle/fluid/operators/detection/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..20d960f9fee1eae42b2241fb96c163e15db5e24d --- /dev/null +++ b/paddle/fluid/operators/detection/CMakeLists.txt @@ -0,0 +1,31 @@ +set(LOCAL_DETECTION_LIBS) + +function(detection_library TARGET_NAME) + set(oneValueArgs "") + set(multiValueArgs SRCS DEPS) + set(options "") + set(common_deps op_registry) + set(pybind_flag 0) + cmake_parse_arguments(detection_library "${options}" "${oneValueArgs}" + "${multiValueArgs}" ${ARGN}) + op_library(${TARGET_NAME} SRCS ${detection_library_SRCS} DEPS ${common_deps} ${detection_library_DEPS}) + set(LOCAL_DETECTION_LIBS + ${TARGET_NAME} + ${LOCAL_DETECTION_LIBS} + PARENT_SCOPE) +endfunction() + +detection_library(bipartite_match_op SRCS bipartite_match_op.cc) +detection_library(box_coder_op SRCS box_coder_op.cc box_coder_op.cu) +detection_library(iou_similarity_op SRCS iou_similarity_op.cc +iou_similarity_op.cu) +detection_library(mine_hard_examples_op SRCS mine_hard_examples_op.cc) +detection_library(multiclass_nms_op SRCS multiclass_nms_op.cc) +detection_library(prior_box_op SRCS prior_box_op.cc prior_box_op.cu) +detection_library(target_assign_op SRCS target_assign_op.cc +target_assign_op.cu) +detection_library(polygon_box_transform_op SRCS polygon_box_transform_op.cc + polygon_box_transform_op.cu) + +# Export local libraries to parent +set(DETECTION_LIBRARY ${LOCAL_DETECTION_LIBS} PARENT_SCOPE) diff --git a/paddle/fluid/operators/bipartite_match_op.cc b/paddle/fluid/operators/detection/bipartite_match_op.cc similarity index 98% rename from paddle/fluid/operators/bipartite_match_op.cc rename to paddle/fluid/operators/detection/bipartite_match_op.cc index 1218d9fdc1e6101d17bc09a4ae769f5fbf8e7b15..d437ad5c19828331c749244404ba80d0f3acda2a 100644 --- a/paddle/fluid/operators/bipartite_match_op.cc +++ b/paddle/fluid/operators/detection/bipartite_match_op.cc @@ -182,8 +182,7 @@ class BipartiteMatchKernel : public framework::OpKernel { class BipartiteMatchOpMaker : public framework::OpProtoAndCheckerMaker { public: - BipartiteMatchOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput( "DistMat", "(LoDTensor or Tensor) this input is a 2-D LoDTensor with shape " diff --git a/paddle/fluid/operators/box_coder_op.cc b/paddle/fluid/operators/detection/box_coder_op.cc similarity index 97% rename from paddle/fluid/operators/box_coder_op.cc rename to paddle/fluid/operators/detection/box_coder_op.cc index ec416f725e75fae57484751ee8a066c0b9da8a70..74848005d0bea6e5c818ff999727aa2b8ad51d84 100644 --- a/paddle/fluid/operators/box_coder_op.cc +++ b/paddle/fluid/operators/detection/box_coder_op.cc @@ -9,7 +9,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include "paddle/fluid/operators/box_coder_op.h" +#include "paddle/fluid/operators/detection/box_coder_op.h" namespace paddle { namespace operators { @@ -60,8 +60,7 @@ class BoxCoderOp : public framework::OperatorWithKernel { class BoxCoderOpMaker : public framework::OpProtoAndCheckerMaker { public: - BoxCoderOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput( "PriorBox", "(Tensor, default Tensor) " diff --git a/paddle/fluid/operators/box_coder_op.cu b/paddle/fluid/operators/detection/box_coder_op.cu similarity index 98% rename from paddle/fluid/operators/box_coder_op.cu rename to paddle/fluid/operators/detection/box_coder_op.cu index 0944e9c95d4a66cc4a51751a8c70cd7a3fefaf1a..8cef8e03439df4ca5b0fa94176a21a36f9eb9f70 100644 --- a/paddle/fluid/operators/box_coder_op.cu +++ b/paddle/fluid/operators/detection/box_coder_op.cu @@ -9,8 +9,8 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include "paddle/fluid/operators/box_coder_op.h" -#include "paddle/fluid/platform/cuda_helper.h" +#include "paddle/fluid/operators/detection/box_coder_op.h" +#include "paddle/fluid/platform/cuda_primitives.h" namespace paddle { namespace operators { diff --git a/paddle/fluid/operators/box_coder_op.h b/paddle/fluid/operators/detection/box_coder_op.h similarity index 100% rename from paddle/fluid/operators/box_coder_op.h rename to paddle/fluid/operators/detection/box_coder_op.h diff --git a/paddle/fluid/operators/iou_similarity_op.cc b/paddle/fluid/operators/detection/iou_similarity_op.cc similarity index 95% rename from paddle/fluid/operators/iou_similarity_op.cc rename to paddle/fluid/operators/detection/iou_similarity_op.cc index 4b78ec510d1fb73592ee8af9a641622f4d713f8d..8e58605fcea04f9ffa97ce8cca53c073e7068aaf 100644 --- a/paddle/fluid/operators/iou_similarity_op.cc +++ b/paddle/fluid/operators/detection/iou_similarity_op.cc @@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include "paddle/fluid/operators/iou_similarity_op.h" +#include "paddle/fluid/operators/detection/iou_similarity_op.h" namespace paddle { namespace operators { @@ -42,8 +42,7 @@ class IOUSimilarityOp : public framework::OperatorWithKernel { class IOUSimilarityOpMaker : public framework::OpProtoAndCheckerMaker { public: - IOUSimilarityOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(LoDTensor, default LoDTensor) " "Box list X is a 2-D LoDTensor with shape [N, 4] holds N boxes, " diff --git a/paddle/fluid/operators/iou_similarity_op.cu b/paddle/fluid/operators/detection/iou_similarity_op.cu similarity index 92% rename from paddle/fluid/operators/iou_similarity_op.cu rename to paddle/fluid/operators/detection/iou_similarity_op.cu index f40a388d62e66a110656ebb71094d46b5ac147eb..8342b4138c87e6ea1803146bac6d6954a569ef5f 100644 --- a/paddle/fluid/operators/iou_similarity_op.cu +++ b/paddle/fluid/operators/detection/iou_similarity_op.cu @@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include "paddle/fluid/operators/iou_similarity_op.h" +#include "paddle/fluid/operators/detection/iou_similarity_op.h" namespace ops = paddle::operators; REGISTER_OP_CUDA_KERNEL( diff --git a/paddle/fluid/operators/iou_similarity_op.h b/paddle/fluid/operators/detection/iou_similarity_op.h similarity index 100% rename from paddle/fluid/operators/iou_similarity_op.h rename to paddle/fluid/operators/detection/iou_similarity_op.h diff --git a/paddle/fluid/operators/mine_hard_examples_op.cc b/paddle/fluid/operators/detection/mine_hard_examples_op.cc similarity index 99% rename from paddle/fluid/operators/mine_hard_examples_op.cc rename to paddle/fluid/operators/detection/mine_hard_examples_op.cc index 277901cff493445e1e85e92e22ea0ada0e1cba43..d4a09bae3a98e4518f9885c1e9182f7033a0d262 100644 --- a/paddle/fluid/operators/mine_hard_examples_op.cc +++ b/paddle/fluid/operators/detection/mine_hard_examples_op.cc @@ -253,8 +253,7 @@ class MineHardExamplesOp : public framework::OperatorWithKernel { class MineHardExamplesOpMaker : public framework::OpProtoAndCheckerMaker { public: - MineHardExamplesOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput( "ClsLoss", "(Tensor, default Tensor), The classification loss with shape " diff --git a/paddle/fluid/operators/multiclass_nms_op.cc b/paddle/fluid/operators/detection/multiclass_nms_op.cc similarity index 99% rename from paddle/fluid/operators/multiclass_nms_op.cc rename to paddle/fluid/operators/detection/multiclass_nms_op.cc index a12b975326519c776c9f4a1d9f2894b4028c2440..60b93efdce810f8552374449fe5a6fc79b1a92c1 100644 --- a/paddle/fluid/operators/multiclass_nms_op.cc +++ b/paddle/fluid/operators/detection/multiclass_nms_op.cc @@ -309,8 +309,7 @@ class MultiClassNMSKernel : public framework::OpKernel { class MultiClassNMSOpMaker : public framework::OpProtoAndCheckerMaker { public: - MultiClassNMSOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("BBoxes", "(Tensor) A 3-D Tensor with shape [N, M, 4] represents the " "predicted locations of M bounding bboxes, N is the batch size. " diff --git a/paddle/fluid/operators/detection/polygon_box_transform_op.cc b/paddle/fluid/operators/detection/polygon_box_transform_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..335e8dd470f851d8c5f6bdbc94cfc343da269034 --- /dev/null +++ b/paddle/fluid/operators/detection/polygon_box_transform_op.cc @@ -0,0 +1,105 @@ +/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/framework/op_registry.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; + +template +class PolygonBoxTransformCPUKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()), + "It must use CUDAPlace."); + auto* in = ctx.Input("Input"); + auto in_dims = in->dims(); + const T* in_data = in->data(); + auto* out = ctx.Output("Output"); + T* out_data = out->mutable_data(ctx.GetPlace()); + + int batch_size = in_dims[0]; + int geo_channel = in_dims[1]; + int height = in_dims[2]; + int width = in_dims[3]; + int id = 0; + for (int id_n = 0; id_n < batch_size * geo_channel; ++id_n) { + for (int id_h = 0; id_h < height; ++id_h) { + for (int id_w = 0; id_w < width; ++id_w) { + id = id_n * height * width + width * id_h + id_w; + if (id_n % 2 == 0) { + out_data[id] = id_w - in_data[id]; + } else { + out_data[id] = id_h - in_data[id]; + } + } + } + } + } +}; + +class PolygonBoxTransformOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE( + ctx->HasInput("Input"), + "Input (Input) of polygon_box transform op should not be null."); + PADDLE_ENFORCE( + ctx->HasOutput("Output"), + "Output (Output) of polygon_box transform op should not be null."); + + auto in_dim = ctx->GetInputDim("Input"); + + PADDLE_ENFORCE_EQ(in_dim.size(), 4, "input's rank must be 4."); + PADDLE_ENFORCE_EQ(in_dim[1] % 2, 0, + "input's second dimension must be even."); + + ctx->SetOutputDim("Output", in_dim); + } +}; + +class PolygonBoxTransformOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput( + "Input", + "The input with shape [batch_size, geometry_channels, height, width]"); + AddOutput("Output", "The output with the same shape as input"); + + AddComment(R"DOC( +PolygonBoxTransform Operator. +The input is the final geometry output in detection network. +We use 2*n numbers to denote the coordinate shift from n corner vertices of +the polygon_box to the pixel location. As each distance offset contains two numbers (xi, yi), +the geometry output contains 2*n channels. +PolygonBoxTransform Operator is used to transform the coordinate shift to the real coordinate. +)DOC"); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OPERATOR(polygon_box_transform, ops::PolygonBoxTransformOp, + ops::PolygonBoxTransformOpMaker, + paddle::framework::EmptyGradOpMaker); +REGISTER_OP_CPU_KERNEL( + polygon_box_transform, + ops::PolygonBoxTransformCPUKernel, + ops::PolygonBoxTransformCPUKernel); diff --git a/paddle/fluid/operators/detection/polygon_box_transform_op.cu b/paddle/fluid/operators/detection/polygon_box_transform_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..6187ac6622c65d2bbc525c3fe2cb397cf74ac612 --- /dev/null +++ b/paddle/fluid/operators/detection/polygon_box_transform_op.cu @@ -0,0 +1,76 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/platform/cuda_primitives.h" +#include "paddle/fluid/platform/gpu_info.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; +using platform::PADDLE_CUDA_NUM_THREADS; +#define CUDA_BLOCK_SIZE 16 + +template +__global__ void PolygonBoxTransformKernel(const int n, const int h, const int w, + const T* input, T* output) { + int id_n = threadIdx.x + blockDim.x * blockIdx.x; + int id_h = threadIdx.y + blockDim.y * blockIdx.y; + int id_w = threadIdx.z + blockDim.z * blockIdx.z; + if (id_n < n && id_h < h && id_w < w) { + int id = id_n * h * w + w * id_h + id_w; + if (id_n % 2 == 0) { + output[id] = id_w - input[id]; + } else { + output[id] = id_h - input[id]; + } + } +} + +template +class PolygonBoxTransformOpCUDAKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), + "It must use CUDAPlace."); + auto* in = ctx.Input("Input"); + auto in_dims = in->dims(); + const T* in_data = in->data(); + auto* out = ctx.Output("Output"); + T* out_data = out->mutable_data(ctx.GetPlace()); + + int batch_size = in_dims[0]; + int geo_channels = in_dims[1]; + int height = in_dims[2]; + int width = in_dims[3]; + dim3 threadsPerBlock( + PADDLE_CUDA_NUM_THREADS / (CUDA_BLOCK_SIZE * CUDA_BLOCK_SIZE), + CUDA_BLOCK_SIZE, CUDA_BLOCK_SIZE); + dim3 numBlocks((batch_size * geo_channels) / threadsPerBlock.x, + (height + threadsPerBlock.y - 1) / threadsPerBlock.y, + (width + threadsPerBlock.z - 1) / threadsPerBlock.z); + auto stream = ctx.cuda_device_context().stream(); + PolygonBoxTransformKernel<<>>( + batch_size * geo_channels, height, width, in_data, out_data); + } +}; + +} // namespace operators +} // namespace paddle + +REGISTER_OP_CUDA_KERNEL( + polygon_box_transform, + paddle::operators::PolygonBoxTransformOpCUDAKernel, + paddle::operators::PolygonBoxTransformOpCUDAKernel); diff --git a/paddle/fluid/operators/prior_box_op.cc b/paddle/fluid/operators/detection/prior_box_op.cc similarity index 97% rename from paddle/fluid/operators/prior_box_op.cc rename to paddle/fluid/operators/detection/prior_box_op.cc index 058b13eeb872aaa77a88da37db64a6d59fbdd1cf..4e35c38e4e03d4d0f00601812fdc4803519b89ae 100644 --- a/paddle/fluid/operators/prior_box_op.cc +++ b/paddle/fluid/operators/detection/prior_box_op.cc @@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include "paddle/fluid/operators/prior_box_op.h" +#include "paddle/fluid/operators/detection/prior_box_op.h" namespace paddle { namespace operators { @@ -79,8 +79,7 @@ class PriorBoxOp : public framework::OperatorWithKernel { class PriorBoxOpMaker : public framework::OpProtoAndCheckerMaker { public: - PriorBoxOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Input", "(Tensor, default Tensor), " "the input feature data of PriorBoxOp, The layout is NCHW."); diff --git a/paddle/fluid/operators/prior_box_op.cu b/paddle/fluid/operators/detection/prior_box_op.cu similarity index 99% rename from paddle/fluid/operators/prior_box_op.cu rename to paddle/fluid/operators/detection/prior_box_op.cu index 0ea8909296f8f52d252b0ec258666cf32d69a8bb..f67e6ca91c0852b5a3be35d23246884d1157caa4 100644 --- a/paddle/fluid/operators/prior_box_op.cu +++ b/paddle/fluid/operators/detection/prior_box_op.cu @@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include "paddle/fluid/operators/prior_box_op.h" +#include "paddle/fluid/operators/detection/prior_box_op.h" namespace paddle { namespace operators { diff --git a/paddle/fluid/operators/prior_box_op.h b/paddle/fluid/operators/detection/prior_box_op.h similarity index 100% rename from paddle/fluid/operators/prior_box_op.h rename to paddle/fluid/operators/detection/prior_box_op.h diff --git a/paddle/fluid/operators/target_assign_op.cc b/paddle/fluid/operators/detection/target_assign_op.cc similarity index 97% rename from paddle/fluid/operators/target_assign_op.cc rename to paddle/fluid/operators/detection/target_assign_op.cc index 33ff967e5e8f5afbaa62ba39ce596687ae0a71cd..367001939251114a9cf442fd85c734958ccb2da8 100644 --- a/paddle/fluid/operators/target_assign_op.cc +++ b/paddle/fluid/operators/detection/target_assign_op.cc @@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include "paddle/fluid/operators/target_assign_op.h" +#include "paddle/fluid/operators/detection/target_assign_op.h" namespace paddle { namespace operators { @@ -65,8 +65,7 @@ class TargetAssignOp : public framework::OperatorWithKernel { class TargetAssignOpMaker : public framework::OpProtoAndCheckerMaker { public: - TargetAssignOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(LoDTensor), This input is a 3D LoDTensor with shape [M, P, K]. " "Some elements in X will be assigned to Out based on the " diff --git a/paddle/fluid/operators/target_assign_op.cu b/paddle/fluid/operators/detection/target_assign_op.cu similarity index 97% rename from paddle/fluid/operators/target_assign_op.cu rename to paddle/fluid/operators/detection/target_assign_op.cu index 24664f99b20f92108220d27ec58e8fdf3ba6193c..ddf6889942355457fb281b6c33430ab8337db3ed 100644 --- a/paddle/fluid/operators/target_assign_op.cu +++ b/paddle/fluid/operators/detection/target_assign_op.cu @@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include "paddle/fluid/operators/target_assign_op.h" +#include "paddle/fluid/operators/detection/target_assign_op.h" namespace paddle { namespace operators { diff --git a/paddle/fluid/operators/target_assign_op.h b/paddle/fluid/operators/detection/target_assign_op.h similarity index 100% rename from paddle/fluid/operators/target_assign_op.h rename to paddle/fluid/operators/detection/target_assign_op.h diff --git a/paddle/fluid/operators/detection_map_op.cc b/paddle/fluid/operators/detection_map_op.cc index 38f43b6d031372948bd82c686a2d9ce5f8ecd07c..716c8625d35308f98582e6802e90d99d643e188b 100644 --- a/paddle/fluid/operators/detection_map_op.cc +++ b/paddle/fluid/operators/detection_map_op.cc @@ -51,7 +51,8 @@ class DetectionMAPOp : public framework::OperatorWithKernel { PADDLE_ENFORCE_EQ(label_dims.size(), 2, "The rank of Input(Label) must be 2, " "the shape is [N, 6]."); - PADDLE_ENFORCE_EQ(label_dims[1], 6, "The shape is of Input(Label) [N, 6]."); + PADDLE_ENFORCE(label_dims[1] == 6 || label_dims[1] == 5, + "The shape of Input(Label) is [N, 6] or [N, 5]."); if (ctx->HasInput("PosCount")) { PADDLE_ENFORCE(ctx->HasInput("TruePos"), @@ -78,8 +79,7 @@ class DetectionMAPOp : public framework::OperatorWithKernel { class DetectionMAPOpMaker : public framework::OpProtoAndCheckerMaker { public: - DetectionMAPOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("DetectRes", "(LoDTensor) A 2-D LoDTensor with shape [M, 6] represents the " "detections. Each row has 6 values: " @@ -89,9 +89,10 @@ class DetectionMAPOpMaker : public framework::OpProtoAndCheckerMaker { "offset is N + 1, if LoD[i + 1] - LoD[i] == 0, means there is " "no detected data."); AddInput("Label", - "(LoDTensor) A 2-D LoDTensor with shape[N, 6] represents the" + "(LoDTensor) A 2-D LoDTensor represents the" "Labeled ground-truth data. Each row has 6 values: " - "[label, is_difficult, xmin, ymin, xmax, ymax], N is the total " + "[label, xmin, ymin, xmax, ymax, is_difficult] or 5 values: " + "[label, xmin, ymin, xmax, ymax], where N is the total " "number of ground-truth data in this mini-batch. For each " "instance, the offsets in first dimension are called LoD, " "the number of offset is N + 1, if LoD[i + 1] - LoD[i] == 0, " diff --git a/paddle/fluid/operators/detection_map_op.h b/paddle/fluid/operators/detection_map_op.h index 431812e2bfcf926cadf8d7be6a7d1a79e78c7762..dd1ab85fd8d0c8170afcd9dd2a49ee55c41dc8be 100644 --- a/paddle/fluid/operators/detection_map_op.h +++ b/paddle/fluid/operators/detection_map_op.h @@ -72,7 +72,7 @@ class DetectionMAPOpKernel : public framework::OpKernel { auto* out_false_pos = ctx.Output("AccumFalsePos"); float overlap_threshold = ctx.Attr("overlap_threshold"); - float evaluate_difficult = ctx.Attr("evaluate_difficult"); + bool evaluate_difficult = ctx.Attr("evaluate_difficult"); auto ap_type = GetAPType(ctx.Attr("ap_type")); int class_num = ctx.Attr("class_num"); @@ -175,14 +175,20 @@ class DetectionMAPOpKernel : public framework::OpKernel { for (int n = 0; n < batch_size; ++n) { std::map> boxes; for (size_t i = label_index[n]; i < label_index[n + 1]; ++i) { - Box box(labels(i, 2), labels(i, 3), labels(i, 4), labels(i, 5)); int label = labels(i, 0); - auto is_difficult = labels(i, 1); - if (std::abs(is_difficult - 0.0) < 1e-6) - box.is_difficult = false; - else - box.is_difficult = true; - boxes[label].push_back(box); + if (input_label.dims()[1] == 6) { + Box box(labels(i, 2), labels(i, 3), labels(i, 4), labels(i, 5)); + auto is_difficult = labels(i, 1); + if (std::abs(is_difficult - 0.0) < 1e-6) + box.is_difficult = false; + else + box.is_difficult = true; + boxes[label].push_back(box); + } else { + PADDLE_ENFORCE_EQ(input_label.dims()[1], 5); + Box box(labels(i, 1), labels(i, 2), labels(i, 3), labels(i, 4)); + boxes[label].push_back(box); + } } gt_boxes->push_back(boxes); } diff --git a/paddle/fluid/operators/dropout_op.cc b/paddle/fluid/operators/dropout_op.cc index 4ed1b548840fabd2383632beb5f35fa6aa096443..07322e720f26213ea777be3cd22f2fead28507f0 100644 --- a/paddle/fluid/operators/dropout_op.cc +++ b/paddle/fluid/operators/dropout_op.cc @@ -37,8 +37,7 @@ class DropoutOp : public framework::OperatorWithKernel { class DropoutOpMaker : public framework::OpProtoAndCheckerMaker { public: - DropoutOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The input of dropout op."); AddOutput("Out", "The output of dropout op."); AddOutput("Mask", "The random sampled dropout mask.").AsIntermediate(); diff --git a/paddle/fluid/operators/edit_distance_op.cc b/paddle/fluid/operators/edit_distance_op.cc index c7f037d2df4372d0c4e3a261c0dff1fd6704d182..de25a3dab53492e38a92fbcf07ccbe43f7546950 100644 --- a/paddle/fluid/operators/edit_distance_op.cc +++ b/paddle/fluid/operators/edit_distance_op.cc @@ -49,8 +49,7 @@ class EditDistanceOp : public framework::OperatorWithKernel { class EditDistanceOpMaker : public framework::OpProtoAndCheckerMaker { public: - EditDistanceOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Hyps", "(2-D LoDTensor, 2nd dim. equal to 1) " "The indices for hypothesis strings."); diff --git a/paddle/fluid/operators/edit_distance_op.cu b/paddle/fluid/operators/edit_distance_op.cu index 913a9145420dae7c4f6a4df10c0330636b5796b0..c25b7d2f9ec32bcef44db239de43feefd855bfe5 100644 --- a/paddle/fluid/operators/edit_distance_op.cu +++ b/paddle/fluid/operators/edit_distance_op.cu @@ -16,7 +16,7 @@ limitations under the License. */ #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/operators/edit_distance_op.h" #include "paddle/fluid/operators/math/math_function.h" -#include "paddle/fluid/platform/cuda_helper.h" +#include "paddle/fluid/platform/cuda_primitives.h" #include "paddle/fluid/platform/gpu_info.h" namespace paddle { diff --git a/paddle/fluid/operators/elementwise_add_op.cc b/paddle/fluid/operators/elementwise_add_op.cc index 4aab54f60236ecc5fa7f70e22f1553c3bfe68198..d2c20537136fc3ac9d1bece24a2238f26215c922 100644 --- a/paddle/fluid/operators/elementwise_add_op.cc +++ b/paddle/fluid/operators/elementwise_add_op.cc @@ -14,26 +14,8 @@ limitations under the License. */ #include "paddle/fluid/operators/elementwise_add_op.h" #include "paddle/fluid/operators/elementwise_op.h" - -namespace paddle { -namespace operators { -class ElementwiseAddOpMaker : public ElementwiseOpMaker { - public: - ElementwiseAddOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : ElementwiseOpMaker(proto, op_checker) { - SetComment("Add", "Out = X + Y"); - AddComment(comment_); - } -}; -} // namespace operators -} // namespace paddle - namespace ops = paddle::operators; -REGISTER_OPERATOR(elementwise_add, ops::ElementwiseOp, - ops::ElementwiseAddOpMaker, ops::ElementwiseOpInferVarType, - paddle::framework::DefaultGradOpDescMaker); -REGISTER_OPERATOR(elementwise_add_grad, ops::ElementwiseOpGrad); - +REGISTER_ELEMWISE_OP(elementwise_add, "Add", "Out = X + Y"); REGISTER_OP_CPU_KERNEL( elementwise_add, ops::ElementwiseAddKernel, diff --git a/paddle/fluid/operators/elementwise_add_op.h b/paddle/fluid/operators/elementwise_add_op.h index 253964562c8d34e0fda3b4760761206895f749aa..baf04c30b17cb333fc8a6544afd6c479442f835b 100644 --- a/paddle/fluid/operators/elementwise_add_op.h +++ b/paddle/fluid/operators/elementwise_add_op.h @@ -14,7 +14,9 @@ limitations under the License. */ #pragma once +#include "paddle/fluid/framework/eigen.h" #include "paddle/fluid/operators/elementwise_op_function.h" +#include "paddle/fluid/operators/math/blas.h" namespace paddle { namespace operators { @@ -24,19 +26,57 @@ struct AddFunctor { inline HOSTDEVICE T operator()(T a, T b) const { return a + b; } }; +template +void default_elementwise_add(const framework::ExecutionContext& ctx, + const framework::Tensor* x, + const framework::Tensor* y, framework::Tensor* z) { + int axis = ctx.Attr("axis"); + ElementwiseComputeEx, DeviceContext, T>(ctx, x, y, axis, + AddFunctor(), z); +} + +template +typename std::enable_if< + std::is_floating_point::value && + std::is_same::value>::type +elementwise_add(const framework::ExecutionContext& ctx, + const framework::Tensor* x, const framework::Tensor* y, + framework::Tensor* z) { + auto eigen_x = framework::EigenVector::Flatten(*x); + auto eigen_y = framework::EigenVector::Flatten(*y); + auto eigen_z = framework::EigenVector::Flatten(*z); + + auto blas = math::GetBlas(ctx); + blas.VADD(x->numel(), eigen_x.data(), eigen_y.data(), eigen_z.data()); +} + +template +typename std::enable_if< + !std::is_floating_point::value || + !std::is_same::value>::type +elementwise_add(const framework::ExecutionContext& ctx, + const framework::Tensor* x, const framework::Tensor* y, + framework::Tensor* z) { + default_elementwise_add(ctx, x, y, z); +} + template class ElementwiseAddKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { using Tensor = framework::Tensor; - auto* x = ctx.Input("X"); - auto* y = ctx.Input("Y"); - auto* z = ctx.Output("Out"); + const auto x = ctx.Input("X"); + const auto y = ctx.Input("Y"); + auto z = ctx.Output("Out"); z->mutable_data(ctx.GetPlace()); - int axis = ctx.Attr("axis"); - ElementwiseComputeEx, DeviceContext, T>(ctx, x, y, axis, - AddFunctor(), z); + + auto dims_equal = x->dims() == y->dims(); + if (dims_equal) { + elementwise_add(ctx, x, y, z); + } else { + default_elementwise_add(ctx, x, y, z); + } } }; @@ -45,6 +85,55 @@ struct IdentityGrad { HOSTDEVICE T operator()(T x, T y, T out, T dout) const { return dout; } }; +template +void default_elementwise_add_grad(const framework::ExecutionContext& ctx, + const framework::Tensor* x, + const framework::Tensor* y, + const framework::Tensor* out, + const framework::Tensor* dout, + framework::Tensor* dx, + framework::Tensor* dy) { + int axis = ctx.Attr("axis"); + + ElemwiseGradCompute, IdentityGrad>( + ctx, *x, *y, *out, *dout, axis, dx, dy, IdentityGrad(), + IdentityGrad()); +} + +template +typename std::enable_if< + std::is_floating_point::value && + std::is_same::value>::type +elementwise_add_grad(const framework::ExecutionContext& ctx, + const framework::Tensor* x, const framework::Tensor* y, + const framework::Tensor* out, + const framework::Tensor* dout, framework::Tensor* dx, + framework::Tensor* dy) { + auto blas = math::GetBlas(ctx); + + if (dx) { + blas.VCOPY(dout->numel(), dout->data(), + dx->mutable_data(ctx.GetPlace())); + } + + if (dy) { + blas.VCOPY(dout->numel(), dout->data(), + dy->mutable_data(ctx.GetPlace())); + } +} + +template +typename std::enable_if< + !std::is_floating_point::value || + !std::is_same::value>::type +elementwise_add_grad(const framework::ExecutionContext& ctx, + const framework::Tensor* x, const framework::Tensor* y, + const framework::Tensor* out, + const framework::Tensor* dout, framework::Tensor* dx, + framework::Tensor* dy) { + default_elementwise_add_grad(ctx, x, y, out, dout, dx, dy); +} + template class ElementwiseAddGradKernel : public framework::OpKernel { public: @@ -57,10 +146,13 @@ class ElementwiseAddGradKernel : public framework::OpKernel { auto* dout = ctx.Input(framework::GradVarName("Out")); auto* dx = ctx.Output(framework::GradVarName("X")); auto* dy = ctx.Output(framework::GradVarName("Y")); - int axis = ctx.Attr("axis"); - ElemwiseGradCompute, IdentityGrad>( - ctx, *x, *y, *out, *dout, axis, dx, dy, IdentityGrad(), - IdentityGrad()); + + if (platform::is_cpu_place(ctx.GetPlace()) && (x->dims() == y->dims())) { + elementwise_add_grad(ctx, x, y, out, dout, dx, dy); + } else { + default_elementwise_add_grad(ctx, x, y, out, dout, dx, + dy); + } } }; diff --git a/paddle/fluid/operators/elementwise_div_op.cc b/paddle/fluid/operators/elementwise_div_op.cc index c7ddafcad1d1f6c14791fde665f43881d6b49836..824b1221e5a77c8799dc34820b7f0db180c2439e 100644 --- a/paddle/fluid/operators/elementwise_div_op.cc +++ b/paddle/fluid/operators/elementwise_div_op.cc @@ -14,26 +14,8 @@ limitations under the License. */ #include "paddle/fluid/operators/elementwise_div_op.h" #include "paddle/fluid/operators/elementwise_op.h" - -namespace paddle { -namespace operators { -class ElementwiseDivOpMaker : public ElementwiseOpMaker { - public: - ElementwiseDivOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : ElementwiseOpMaker(proto, op_checker) { - SetComment("Div", "Out = X / Y"); - AddComment(comment_); - } -}; - -} // namespace operators -} // namespace paddle - namespace ops = paddle::operators; -REGISTER_OPERATOR(elementwise_div, ops::ElementwiseOp, - ops::ElementwiseDivOpMaker, - paddle::framework::DefaultGradOpDescMaker); -REGISTER_OPERATOR(elementwise_div_grad, ops::ElementwiseOpGrad); +REGISTER_ELEMWISE_OP(elementwise_div, "Div", "Out = X / Y"); REGISTER_OP_CPU_KERNEL( elementwise_div, ops::ElementwiseDivKernel, diff --git a/paddle/fluid/operators/elementwise_max_op.cc b/paddle/fluid/operators/elementwise_max_op.cc index a4fe386bb1907bf7c0099d2b1109077b21146948..411671335a19ae2283ca9db8b8f6bcbb6a6b630a 100644 --- a/paddle/fluid/operators/elementwise_max_op.cc +++ b/paddle/fluid/operators/elementwise_max_op.cc @@ -14,25 +14,8 @@ limitations under the License. */ #include "paddle/fluid/operators/elementwise_max_op.h" #include "paddle/fluid/operators/elementwise_op.h" - -namespace paddle { -namespace operators { -class ElementwiseMaxOpMaker : public ElementwiseOpMaker { - public: - ElementwiseMaxOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : ElementwiseOpMaker(proto, op_checker) { - SetComment("Max", "Out = max(X, Y)"); - AddComment(comment_); - } -}; -} // namespace operators -} // namespace paddle - namespace ops = paddle::operators; -REGISTER_OPERATOR(elementwise_max, ops::ElementwiseOp, - ops::ElementwiseMaxOpMaker, - paddle::framework::DefaultGradOpDescMaker); -REGISTER_OPERATOR(elementwise_max_grad, ops::ElementwiseOpGrad); +REGISTER_ELEMWISE_OP(elementwise_max, "Max", "Out = max(X, Y)"); REGISTER_OP_CPU_KERNEL( elementwise_max, ops::ElementwiseMaxKernel, diff --git a/paddle/fluid/operators/elementwise_min_op.cc b/paddle/fluid/operators/elementwise_min_op.cc index 68cd6ddb4a938b2b1c33e3f89c6d1151acb27f48..816192083d2275b26e6dd9afc76f2c021a01cf73 100644 --- a/paddle/fluid/operators/elementwise_min_op.cc +++ b/paddle/fluid/operators/elementwise_min_op.cc @@ -14,25 +14,8 @@ limitations under the License. */ #include "paddle/fluid/operators/elementwise_min_op.h" #include "paddle/fluid/operators/elementwise_op.h" - -namespace paddle { -namespace operators { -class ElementwiseMinOpMaker : public ElementwiseOpMaker { - public: - ElementwiseMinOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : ElementwiseOpMaker(proto, op_checker) { - SetComment("Max", "Out = min(X, Y)"); - AddComment(comment_); - } -}; -} // namespace operators -} // namespace paddle - namespace ops = paddle::operators; -REGISTER_OPERATOR(elementwise_min, ops::ElementwiseOp, - ops::ElementwiseMinOpMaker, - paddle::framework::DefaultGradOpDescMaker); -REGISTER_OPERATOR(elementwise_min_grad, ops::ElementwiseOpGrad); +REGISTER_ELEMWISE_OP(elementwise_min, "Min", "Out = min(X, Y)"); REGISTER_OP_CPU_KERNEL( elementwise_min, ops::ElementwiseMinKernel, diff --git a/paddle/fluid/operators/elementwise_mul_op.cc b/paddle/fluid/operators/elementwise_mul_op.cc index 2dec27136ad57ea032d5abb51799bd04ccc0b2e3..ba343909bb87b4f2efa56c0a4ff664b278e90c60 100644 --- a/paddle/fluid/operators/elementwise_mul_op.cc +++ b/paddle/fluid/operators/elementwise_mul_op.cc @@ -14,27 +14,8 @@ limitations under the License. */ #include "paddle/fluid/operators/elementwise_mul_op.h" #include "paddle/fluid/operators/elementwise_op.h" - -namespace paddle { -namespace operators { - -class ElementwiseMulOpMaker : public ElementwiseOpMaker { - public: - ElementwiseMulOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : ElementwiseOpMaker(proto, op_checker) { - SetComment("Mul", "Out = X \\odot\\ Y"); - AddComment(comment_); - } -}; - -} // namespace operators -} // namespace paddle - namespace ops = paddle::operators; -REGISTER_OPERATOR(elementwise_mul, ops::ElementwiseOp, - ops::ElementwiseMulOpMaker, - paddle::framework::DefaultGradOpDescMaker); -REGISTER_OPERATOR(elementwise_mul_grad, ops::ElementwiseOpGrad); +REGISTER_ELEMWISE_OP(elementwise_mul, "Mul", "Out = X \\odot\\ Y"); REGISTER_OP_CPU_KERNEL( elementwise_mul, ops::ElementwiseMulKernel, diff --git a/paddle/fluid/operators/elementwise_op.h b/paddle/fluid/operators/elementwise_op.h index a33634ab2503f988a8a692682ddb238d4794a3c0..f4cec8ad971abebe8d6dff1a384c8414269148a5 100644 --- a/paddle/fluid/operators/elementwise_op.h +++ b/paddle/fluid/operators/elementwise_op.h @@ -46,16 +46,17 @@ class ElementwiseOpInferVarType : public framework::VarTypeInference { public: void operator()(const framework::OpDesc& op_desc, framework::BlockDesc* block) const override { - auto x_var = op_desc.Input("X")[0]; - auto out_var = op_desc.Output("Out")[0]; - block->Var(out_var)->SetType(block->Var(x_var)->GetType()); + auto x_name = op_desc.Input("X")[0]; + auto out_name = op_desc.Output("Out")[0]; + auto& x = block->FindRecursiveOrCreateVar(x_name); + auto& out = block->FindRecursiveOrCreateVar(out_name); + out.SetType(x.GetType()); } }; class ElementwiseOpMaker : public framework::OpProtoAndCheckerMaker { public: - ElementwiseOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() final { AddInput("X", "(Tensor), The first input tensor of elementwise op."); AddInput("Y", "(Tensor), The second input tensor of elementwise op."); AddOutput("Out", "The output of elementwise op."); @@ -64,12 +65,12 @@ class ElementwiseOpMaker : public framework::OpProtoAndCheckerMaker { "for broadcasting Y onto X.") .SetDefault(-1) .EqualGreaterThan(-1); - comment_ = R"DOC( -Limited Elementwise {name} Operator. + AddComment(string::Sprintf(R"DOC( +Limited Elementwise %s Operator. The equation is: -$${equation}$$ +$$%s$$ $X$ is a tensor of any dimension and the dimensions of tensor $Y$ must be smaller than or equal to the dimensions of $X$. @@ -100,26 +101,13 @@ For example Either of the inputs $X$ and $Y$ or none can carry the LoD (Level of Details) information. However, the output only shares the LoD information with input $X$. -)DOC"; - AddComment(comment_); +)DOC", + GetName(), GetEquation())); } protected: - std::string comment_; - - void Replace(std::string* src, std::string from, std::string to) { - std::size_t len_from = std::strlen(from.c_str()); - std::size_t len_to = std::strlen(to.c_str()); - for (std::size_t pos = src->find(from); pos != std::string::npos; - pos = src->find(from, pos + len_to)) { - src->replace(pos, len_from, to); - } - } - - void SetComment(std::string name, std::string equation) { - Replace(&comment_, "{name}", name); - Replace(&comment_, "{equation}", equation); - } + virtual std::string GetName() const = 0; + virtual std::string GetEquation() const = 0; }; class ElementwiseOpGrad : public framework::OperatorWithKernel { @@ -152,3 +140,16 @@ class ElementwiseOpGrad : public framework::OperatorWithKernel { }; } // namespace operators } // namespace paddle + +#define REGISTER_ELEMWISE_OP(op_type, op_name, equation) \ + class __ElemwiseOp##op_type##Maker__ \ + : public ::paddle::operators::ElementwiseOpMaker { \ + protected: \ + virtual std::string GetName() const { return op_name; } \ + virtual std::string GetEquation() const { return equation; } \ + }; \ + REGISTER_OPERATOR(op_type, ::paddle::operators::ElementwiseOp, \ + __ElemwiseOp##op_type##Maker__, \ + ::paddle::operators::ElementwiseOpInferVarType, \ + ::paddle::framework::DefaultGradOpDescMaker); \ + REGISTER_OPERATOR(op_type##_grad, ::paddle::operators::ElementwiseOpGrad) diff --git a/paddle/fluid/operators/elementwise_op_function.h b/paddle/fluid/operators/elementwise_op_function.h index f0362ec606c994d69f31c7a2e1e9ad0d0108b621..8b052611f80ddf874ca48c1c58e13346528a834e 100644 --- a/paddle/fluid/operators/elementwise_op_function.h +++ b/paddle/fluid/operators/elementwise_op_function.h @@ -22,6 +22,8 @@ limitations under the License. */ #ifdef __NVCC__ #include #include +#include "paddle/fluid/platform/cuda_device_function.h" +#include "paddle/fluid/platform/cuda_primitives.h" constexpr int ELEMWISE_MAX_BLOCK_DIM = 1024; #endif @@ -333,57 +335,8 @@ static void ElemwiseGradBroadcast1CPU(const T* x, const T* y, const T* out, } } } -#ifdef __NVCC__ - -// __shfl_down has been deprecated as of CUDA 9.0. -#if CUDA_VERSION < 9000 -template -__forceinline__ __device__ T __shfl_down_sync(unsigned, T val, int delta) { - return __shfl_down(val, delta); -} -#define CREATE_SHFL_MASK(mask, predicate) mask = 0u; -#else -#define FULL_WARP_MASK 0xFFFFFFFF -#define CREATE_SHFL_MASK(mask, predicate) \ - mask = __ballot_sync(FULL_WARP_MASK, (predicate)) -#endif - -template -__device__ T reduceSum(T val, int tid, int len) { - // TODO(zcd): The warp size should be taken from the - // parameters of the GPU but not specified as 32 simply. - // To make the reduceSum more efficiently, - // I use Warp-Level Parallelism and assume the Warp size - // is 32 which may be different for different GPU, - // but most card's warp size is 32. - const int warpSize = 32; - __shared__ T shm[warpSize]; - unsigned mask = 0u; - CREATE_SHFL_MASK(mask, tid < len); - - for (int offset = warpSize / 2; offset > 0; offset /= 2) - val += __shfl_down_sync(mask, val, offset); - - if (tid < warpSize) shm[tid] = 0; - - __syncthreads(); - - if (tid % warpSize == 0) { - shm[tid / warpSize] = val; - } - __syncthreads(); - - CREATE_SHFL_MASK(mask, tid < warpSize); - - if (tid < warpSize) { - val = shm[tid]; - for (int offset = warpSize / 2; offset > 0; offset /= 2) - val += __shfl_down_sync(mask, val, offset); - } - - return val; -} +#ifdef __NVCC__ template static __global__ void ElemwiseGradBroadcast1CUDAKernel( const T* x, const T* y, const T* out, const T* dout, int h, int w, @@ -406,7 +359,7 @@ static __global__ void ElemwiseGradBroadcast1CUDAKernel( if (dy) { h = h > ELEMWISE_MAX_BLOCK_DIM ? ELEMWISE_MAX_BLOCK_DIM : h; - val = reduceSum(val, tid, h); + val = paddle::platform::reduceSum(val, tid, h); if (threadIdx.x == 0) { dy[j] = val; } @@ -483,7 +436,7 @@ static __global__ void ElemwiseGradBroadcast2CUDAKernel( if (dy) { int h = pre * post; h = h > ELEMWISE_MAX_BLOCK_DIM ? ELEMWISE_MAX_BLOCK_DIM : h; - val = reduceSum(val, tid, h); + val = paddle::platform::reduceSum(val, tid, h); if (threadIdx.x == 0) { dy[j] = val; } diff --git a/paddle/fluid/operators/elementwise_pow_op.cc b/paddle/fluid/operators/elementwise_pow_op.cc index 60302c5e59f8ce595861405713045b05d90002e3..5fd6bde9ba0930e29f2161f1ff23ff9f5e7dc85d 100644 --- a/paddle/fluid/operators/elementwise_pow_op.cc +++ b/paddle/fluid/operators/elementwise_pow_op.cc @@ -13,17 +13,15 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/operators/elementwise_pow_op.h" +#include #include "paddle/fluid/operators/elementwise_op.h" namespace paddle { namespace operators { class ElementwisePowOpMaker : public ElementwiseOpMaker { - public: - ElementwisePowOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : ElementwiseOpMaker(proto, op_checker) { - SetComment("Pow", "Out = X ^ Y"); - AddComment(comment_); - } + protected: + std::string GetName() const override { return "Pow"; } + std::string GetEquation() const override { return "Out = X ^ Y"; } }; } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/elementwise_sub_op.cc b/paddle/fluid/operators/elementwise_sub_op.cc index 9d0598fc39a3922fa830f18729d90a7dac6a890b..a7562b166b373ee2a8c9b6f379431d88d3e45fcb 100644 --- a/paddle/fluid/operators/elementwise_sub_op.cc +++ b/paddle/fluid/operators/elementwise_sub_op.cc @@ -14,25 +14,8 @@ limitations under the License. */ #include "paddle/fluid/operators/elementwise_sub_op.h" #include "paddle/fluid/operators/elementwise_op.h" - -namespace paddle { -namespace operators { -class ElementwiseSubOpMaker : public ElementwiseOpMaker { - public: - ElementwiseSubOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : ElementwiseOpMaker(proto, op_checker) { - SetComment("Sub", "Out = X - Y"); - AddComment(comment_); - } -}; -} // namespace operators -} // namespace paddle - namespace ops = paddle::operators; -REGISTER_OPERATOR(elementwise_sub, ops::ElementwiseOp, - ops::ElementwiseSubOpMaker, - paddle::framework::DefaultGradOpDescMaker); -REGISTER_OPERATOR(elementwise_sub_grad, ops::ElementwiseOpGrad); +REGISTER_ELEMWISE_OP(elementwise_sub, "Sub", "Out = X - Y"); REGISTER_OP_CPU_KERNEL( elementwise_sub, ops::ElementwiseSubKernel, diff --git a/paddle/fluid/operators/expand_op.cc b/paddle/fluid/operators/expand_op.cc index 4ae91d074d3df8b910a7f5d816a22b6f1d51dff6..5ad0ec251328cc1ba580026bb47bf05316e7dc77 100644 --- a/paddle/fluid/operators/expand_op.cc +++ b/paddle/fluid/operators/expand_op.cc @@ -56,8 +56,7 @@ class ExpandOp : public framework::OperatorWithKernel { class ExpandOpMaker : public framework::OpProtoAndCheckerMaker { public: - ExpandOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor, default Tensor). A tensor with rank in [1, 6]." "X is the input to be expanded."); diff --git a/paddle/fluid/operators/fake_dequantize_op.cc b/paddle/fluid/operators/fake_dequantize_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..43f949111104ee56efc8625bdd609e412ef7f37d --- /dev/null +++ b/paddle/fluid/operators/fake_dequantize_op.cc @@ -0,0 +1,76 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/fake_dequantize_op.h" +#include + +namespace paddle { +namespace operators { + +class FakeDequantizeMaxAbsOp : public framework::OperatorWithKernel { + public: + FakeDequantizeMaxAbsOp(const std::string &type, + const framework::VariableNameMap &inputs, + const framework::VariableNameMap &outputs, + const framework::AttributeMap &attrs) + : OperatorWithKernel(type, inputs, outputs, attrs) {} + + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of FakeDequantizeMaxAbsOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of FakeDequantizeMaxAbsOp should not be null."); + ctx->SetOutputDim("Out", ctx->GetInputDim("X")); + ctx->ShareLoD("X", /*->*/ "Out"); + } +}; + +class FakeDequantizeMaxAbsOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("X", + "(Tensor) The input with float-32/64 type is the " + "low precision tensor."); + AddOutput("Out", + "(Tensor) The output is the dequantized high " + "precision tensor."); + AddAttr("num_bits", + "(int) `num_bits` is the quantization level bits, " + "such as 2, 5, 8."); + AddAttr("scale", + "(float) The maximum absolute value of low precision tensor." + "It is usually calculated by the fake_quantize_max_abs_op."); + AddComment(R"DOC( +FakeDequantizeMaxAbsOp operator. + +This calculation is an opposite operation of FakeQuantizeMaxAbsOp: + +$$Out = \frac{scale*X}{2^{num_bits} - 1}$$ + +)DOC"); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +using CPU = paddle::platform::CPUDeviceContext; + +REGISTER_OPERATOR(fake_dequantize_max_abs, ops::FakeDequantizeMaxAbsOp, + ops::FakeDequantizeMaxAbsOpMaker, + paddle::framework::EmptyGradOpMaker); +REGISTER_OP_CPU_KERNEL(fake_dequantize_max_abs, + ops::FakeDequantizeMaxAbsKernel, + ops::FakeDequantizeMaxAbsKernel); diff --git a/paddle/fluid/operators/matmul_op.cu.cc b/paddle/fluid/operators/fake_dequantize_op.cu similarity index 68% rename from paddle/fluid/operators/matmul_op.cu.cc rename to paddle/fluid/operators/fake_dequantize_op.cu index e021bbe645399e410cde5c3ff7035d4d68c71744..1bd38d1bd2c3a6f90d2fbad415d61efaead3afe9 100644 --- a/paddle/fluid/operators/matmul_op.cu.cc +++ b/paddle/fluid/operators/fake_dequantize_op.cu @@ -12,11 +12,10 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include "paddle/fluid/operators/matmul_op.h" +#include "paddle/fluid/operators/fake_dequantize_op.h" namespace ops = paddle::operators; -REGISTER_OP_CUDA_KERNEL( - matmul, ops::MatMulKernel); -REGISTER_OP_CUDA_KERNEL( - matmul_grad, - ops::MatMulGradKernel); +using CUDA = paddle::platform::CUDADeviceContext; +REGISTER_OP_CUDA_KERNEL(fake_dequantize_max_abs, + ops::FakeDequantizeMaxAbsKernel, + ops::FakeDequantizeMaxAbsKernel); diff --git a/paddle/fluid/operators/fake_dequantize_op.h b/paddle/fluid/operators/fake_dequantize_op.h new file mode 100644 index 0000000000000000000000000000000000000000..0901e68b3761159c3cc9c6684567bee38ec3f16d --- /dev/null +++ b/paddle/fluid/operators/fake_dequantize_op.h @@ -0,0 +1,42 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "paddle/fluid/framework/eigen.h" +#include "paddle/fluid/framework/op_registry.h" + +namespace paddle { +namespace operators { +template +class FakeDequantizeMaxAbsKernel : public framework::OpKernel { + public: + virtual void Compute(const framework::ExecutionContext& ctx) const { + auto* in = ctx.Input("X"); + auto* out = ctx.Output("Out"); + out->mutable_data(in->place()); + + int num_bits = ctx.Attr("num_bits"); + T scale = static_cast(ctx.Attr("scale")); + int range = std::pow(2, num_bits) - 1; + + auto eigen_out = framework::EigenVector::Flatten(*out); + auto eigen_in = framework::EigenVector::Flatten(*in); + auto& dev = *ctx.template device_context().eigen_device(); + eigen_out.device(dev) = (scale / range) * eigen_in; + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/fc_op.cc b/paddle/fluid/operators/fc_op.cc index 45e4d5b2b863a55ae0aa0414ff8697141fd2aa6f..8843a1c44b7004ba5d7935f75d3c99d9c30fc6c0 100644 --- a/paddle/fluid/operators/fc_op.cc +++ b/paddle/fluid/operators/fc_op.cc @@ -72,8 +72,7 @@ framework::OpKernelType FCOpGrad::GetExpectedKernelType( layout, library); } -FCOpMaker::FCOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { +void FCOpMaker::Make() { AddInput("Input", "(Tensor) The input tensor of fully connected operator. "); AddInput("W", "(Tensor), The second input tensor of fc op."); AddOutput("Out", "(Tensor) The output tensor of fully connected operator. "); diff --git a/paddle/fluid/operators/fc_op.h b/paddle/fluid/operators/fc_op.h index 70fa96440d344397a7427c1338afee85bde923d4..e1b780fc0c401fbf34a9db03aa31137cbc016939 100644 --- a/paddle/fluid/operators/fc_op.h +++ b/paddle/fluid/operators/fc_op.h @@ -45,7 +45,7 @@ class FCOpGrad : public framework::OperatorWithKernel { class FCOpMaker : public framework::OpProtoAndCheckerMaker { public: - FCOpMaker(OpProto* proto, OpAttrChecker* op_checker); + void Make() override; }; } // namespace operators diff --git a/paddle/fluid/operators/feed_op.cc b/paddle/fluid/operators/feed_op.cc index debacf07c360b9aa69000a0d891f04239ed08807..bcb3e63ed7dbc775c1de6c4522f0548ea48a6cf0 100644 --- a/paddle/fluid/operators/feed_op.cc +++ b/paddle/fluid/operators/feed_op.cc @@ -66,8 +66,7 @@ class FeedOp : public framework::OperatorBase { class FeedOpInfoMaker : public framework::OpProtoAndCheckerMaker { public: - FeedOpInfoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The input of feed op"); AddOutput("Out", "The output of feed op"); AddAttr("col", "(int) The column of feed"); diff --git a/paddle/fluid/operators/fetch_barrier_op.cc b/paddle/fluid/operators/fetch_barrier_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..79ec02f52094121d01c6bda2a5d99d2211893e89 --- /dev/null +++ b/paddle/fluid/operators/fetch_barrier_op.cc @@ -0,0 +1,87 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include // NOLINT +#include + +#include "paddle/fluid/framework/data_type.h" +#include "paddle/fluid/framework/framework.pb.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/op_registry.h" + +#include "paddle/fluid/operators/detail/grpc_client.h" +#include "paddle/fluid/platform/profiler.h" + +namespace paddle { +namespace operators { + +class FetchBarrierOp : public framework::OperatorBase { + public: + FetchBarrierOp(const std::string& type, + const framework::VariableNameMap& inputs, + const framework::VariableNameMap& outputs, + const framework::AttributeMap& attrs) + : OperatorBase(type, inputs, outputs, attrs) {} + + void RunImpl(const framework::Scope& scope, + const platform::Place& place) const override { + std::vector eps = Attr>("endpoints"); + + platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); + auto& ctx = *pool.Get(place); + // For profiling + platform::RecordEvent record_event(Type(), &ctx); + + auto rpc_client = detail::RPCClient::GetInstance(); + + PADDLE_ENFORCE(rpc_client->Wait()); + + for (auto& ep : eps) { + VLOG(3) << "fetch barrier, ep: " << ep; + rpc_client->AsyncSendFetchBarrier(ep); + } + PADDLE_ENFORCE(rpc_client->Wait()); + } +}; + +class FetchBarrierOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() { + AddComment(R"DOC( +SendBarrier operator + +This operator will send a send barrier signal to list_and_serv op, so that +the Parameter Server would knew all variables have been sent. +)DOC"); + + AddAttr>("endpoints", + "(string vector, default 127.0.0.1:6164)" + "Server endpoints to send variables to.") + .SetDefault({"127.0.0.1:6164"}); + } +}; + +class FetchBarrierOpShapeInference : public framework::InferShapeBase { + public: + void operator()(framework::InferShapeContext* ctx) const override {} +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; + +REGISTER_OPERATOR(fetch_barrier, ops::FetchBarrierOp, + paddle::framework::EmptyGradOpMaker, ops::FetchBarrierOpMaker, + ops::FetchBarrierOpShapeInference); diff --git a/paddle/fluid/operators/fetch_op.cc b/paddle/fluid/operators/fetch_op.cc index 7c7f3e9059fbb1e3f2cca4f04edfff55c9452761..1640a2a22c69a0e3ab81a2889d6105b2cf4162b7 100644 --- a/paddle/fluid/operators/fetch_op.cc +++ b/paddle/fluid/operators/fetch_op.cc @@ -57,10 +57,7 @@ class FetchOp : public framework::OperatorBase { // FIXME(yuyang18): Should we assume the fetch operator always generate // CPU outputs? - auto &dev_ctx = *pool.Get(src_item.place()); - - TensorCopy(src_item, platform::CPUPlace(), dev_ctx, &dst_item); - dev_ctx.Wait(); + TensorCopySync(src_item, platform::CPUPlace(), &dst_item); dst_item.set_lod(src_item.lod()); VLOG(3) << "Fetch variable " << fetch_var_name << " to " << out_name; @@ -69,8 +66,7 @@ class FetchOp : public framework::OperatorBase { class FetchOpInfoMaker : public framework::OpProtoAndCheckerMaker { public: - FetchOpInfoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The input of fetch op"); AddOutput("Out", "The output of fetch op"); AddAttr("col", "(int) The column of fetch"); diff --git a/paddle/fluid/operators/fill_constant_batch_size_like_op.cc b/paddle/fluid/operators/fill_constant_batch_size_like_op.cc index 72da80baaf9bb3286f09b7ae5fcf24326b391906..1ae78675a0cac8a72aeaef1227b631a41e4a10b2 100644 --- a/paddle/fluid/operators/fill_constant_batch_size_like_op.cc +++ b/paddle/fluid/operators/fill_constant_batch_size_like_op.cc @@ -30,9 +30,8 @@ class FillConstantBatchSizeLikeOp : public BatchSizeLikeOp { }; class FillConstantBatchSizeLikeOpMaker : public BatchSizeLikeOpMaker { - public: - FillConstantBatchSizeLikeOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : BatchSizeLikeOpMaker(proto, op_checker) { + protected: + void Apply() override { AddAttr("dtype", "(int, default 5 (FP32)) " "Output data type") diff --git a/paddle/fluid/operators/fill_constant_batch_size_like_op.h b/paddle/fluid/operators/fill_constant_batch_size_like_op.h index 2a7df149a9f4b03676f172da980c927d7fa5e8a4..63ea60678f80708f5a8340edd22588553b9ec139 100644 --- a/paddle/fluid/operators/fill_constant_batch_size_like_op.h +++ b/paddle/fluid/operators/fill_constant_batch_size_like_op.h @@ -24,6 +24,14 @@ class FillConstantBatchSizeLikeOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { auto* out = ctx.Output("Out"); + auto* in = ctx.Input("Input"); + if (in->lod().size() && ctx.Attr("input_dim_idx") == 0) { + // set the correct batch size for the LoDTensor. + auto odims = out->dims(); + int output_dim_idx = ctx.Attr("output_dim_idx"); + odims[output_dim_idx] = static_cast(in->lod().back().size()) - 1; + out->mutable_data(odims, ctx.GetPlace()); + } out->mutable_data(ctx.GetPlace()); auto value = ctx.Attr("value"); diff --git a/paddle/fluid/operators/fill_constant_op.cc b/paddle/fluid/operators/fill_constant_op.cc index 07e0a80f8d644d4d011f2821785d49ece6cecfb5..130f18dde4f979a6a9925ede9cbf745fcec14d48 100644 --- a/paddle/fluid/operators/fill_constant_op.cc +++ b/paddle/fluid/operators/fill_constant_op.cc @@ -59,8 +59,7 @@ class FillConstantOp : public framework::OperatorBase { class FillConstantOpMaker : public framework::OpProtoAndCheckerMaker { public: - FillConstantOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddAttr("dtype", "(int, default 5 (FP32)) " "Output data type") diff --git a/paddle/fluid/operators/fill_op.cc b/paddle/fluid/operators/fill_op.cc index ee8a2fc353f86cdabd35459a9195c3aa35f63e31..925dc19061e2196a40411f415eb6e5ad59ab52ff 100644 --- a/paddle/fluid/operators/fill_op.cc +++ b/paddle/fluid/operators/fill_op.cc @@ -82,8 +82,7 @@ class FillOp : public framework::OperatorBase { class FillOpMaker : public framework::OpProtoAndCheckerMaker { public: - FillOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddComment(R"DOC(Fill operator Fill an tensor with `value` and `shape`. The type of the tensor is specify by diff --git a/paddle/fluid/operators/fill_zeros_like_op.cc b/paddle/fluid/operators/fill_zeros_like_op.cc index 58c814ba6413626a48310da595a13238994f5ef1..d67bec36b3248be8602da562a88aeb58f5effe39 100644 --- a/paddle/fluid/operators/fill_zeros_like_op.cc +++ b/paddle/fluid/operators/fill_zeros_like_op.cc @@ -33,8 +33,7 @@ class FillZerosLikeOp : public framework::OperatorWithKernel { class FillZerosLikeOpMaker : public framework::OpProtoAndCheckerMaker { public: - FillZerosLikeOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The input of fill-zeros-like op."); AddOutput("Out", "The variable will be filled up with zeros."); AddComment(R"DOC( diff --git a/paddle/fluid/operators/ftrl_op.cc b/paddle/fluid/operators/ftrl_op.cc index 0a456f0981e5753a7de5a6f2ba029681beb347a5..70ba25c213046cc934f46be067080d5fdbb42f9e 100644 --- a/paddle/fluid/operators/ftrl_op.cc +++ b/paddle/fluid/operators/ftrl_op.cc @@ -17,6 +17,7 @@ limitations under the License. */ namespace paddle { namespace operators { +using Tensor = framework::Tensor; class FTRLOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; @@ -53,12 +54,17 @@ class FTRLOp : public framework::OperatorWithKernel { ctx->SetOutputDim("SquaredAccumOut", param_dim); ctx->SetOutputDim("LinearAccumOut", param_dim); } + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext &ctx) const override { + auto input_data_type = + framework::ToDataType(ctx.Input("Param")->type()); + return framework::OpKernelType(input_data_type, ctx.GetPlace()); + } }; class FTRLOpMaker : public framework::OpProtoAndCheckerMaker { public: - FTRLOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Param", "(Tensor, default Tensor) " "Input parameter value that has to be updated."); diff --git a/paddle/fluid/operators/gather_op.cc b/paddle/fluid/operators/gather_op.cc index 4c82f5c429038504d9876ee240a705911feb0b7a..e21b57258928856a10d6e86c3e2c6e81fb241ee3 100644 --- a/paddle/fluid/operators/gather_op.cc +++ b/paddle/fluid/operators/gather_op.cc @@ -67,8 +67,7 @@ class GatherGradOp : public framework::OperatorWithKernel { class GatherOpMaker : public framework::OpProtoAndCheckerMaker { public: - GatherOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The source input of gather op"); AddInput("Index", "The index input of gather op"); AddOutput("Out", "The output of gather op"); diff --git a/paddle/fluid/operators/gaussian_random_batch_size_like_op.cc b/paddle/fluid/operators/gaussian_random_batch_size_like_op.cc index 53c706a83e5bfb9e93d485141314e8b652d73593..8050f61d4546f3351645f23ddcc63b2c49f17929 100644 --- a/paddle/fluid/operators/gaussian_random_batch_size_like_op.cc +++ b/paddle/fluid/operators/gaussian_random_batch_size_like_op.cc @@ -32,9 +32,8 @@ class GaussianRandomBatchSizeLikeOp : public BatchSizeLikeOp { }; class GaussianRandomBatchSizeLikeOpMaker : public BatchSizeLikeOpMaker { - public: - GaussianRandomBatchSizeLikeOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : BatchSizeLikeOpMaker(proto, op_checker) { + protected: + void Apply() override { AddAttr("mean", "(float, default 0.0) " "mean of random tensor.") diff --git a/paddle/fluid/operators/gaussian_random_op.cc b/paddle/fluid/operators/gaussian_random_op.cc index 4d197637b3f49f7e63f5b1a5cba212d1bf774f7e..815c1bb50988be49ca9996e368a59344c6583d58 100644 --- a/paddle/fluid/operators/gaussian_random_op.cc +++ b/paddle/fluid/operators/gaussian_random_op.cc @@ -70,8 +70,7 @@ class GaussianRandomOp : public framework::OperatorWithKernel { class GaussianRandomOpMaker : public framework::OpProtoAndCheckerMaker { public: - GaussianRandomOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddOutput("Out", "Output matrix of gaussian random op"); AddAttr>("shape", diff --git a/paddle/fluid/operators/gen_nccl_id_op.cc b/paddle/fluid/operators/gen_nccl_id_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..a5678f63466d368b3dd59380c18f9625cabd368b --- /dev/null +++ b/paddle/fluid/operators/gen_nccl_id_op.cc @@ -0,0 +1,128 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include +#include +#include + +#include "paddle/fluid/framework/executor.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/framework/threadpool.h" +#include "paddle/fluid/operators/detail/grpc_client.h" +#include "paddle/fluid/operators/detail/grpc_server.h" +#include "paddle/fluid/platform/nccl_helper.h" + +namespace paddle { +namespace operators { + +class GenNCCLIdOp : public framework::OperatorBase { + public: + GenNCCLIdOp(const std::string& type, const framework::VariableNameMap& inputs, + const framework::VariableNameMap& outputs, + const framework::AttributeMap& attrs) + : OperatorBase(type, inputs, outputs, attrs) {} + + void RunImpl(const framework::Scope& scope, + const platform::Place& dev_place) const override { + platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); + // put nccl id in CPUPlace + auto& dev_ctx = *pool.Get(platform::CPUPlace()); + int trainer_id = Attr("trainer_id"); + framework::Scope& local_scope = scope.NewScope(); + + if (trainer_id == 0) { + GenerateAndSend(&local_scope, dev_ctx); + } else { + GetIdByServer(&local_scope, dev_ctx); + } + } + + private: + void GenerateAndSend(framework::Scope* scope, + const platform::DeviceContext& dev_ctx) const { + auto var = scope->FindVar(NCCL_ID_VARNAME); + PADDLE_ENFORCE_NOT_NULL(var); + auto id = var->GetMutable(); + PADDLE_ENFORCE(platform::dynload::ncclGetUniqueId(id)); + + std::vector endpoint_list = + Attr>("endpoint_list"); + detail::RPCClient client; + for (auto& ep : endpoint_list) { + VLOG(3) << "sending nccl id to " << ep; + client.AsyncSendVariable(ep, dev_ctx, *scope, NCCL_ID_VARNAME); + } + client.Wait(); + VLOG(3) << "sending completed..."; + } + + void GetIdByServer(framework::Scope* scope, + const platform::DeviceContext& dev_ctx) const { + std::string endpoint = Attr("endpoint"); + // NOTE: Can not use unique_ptr here because the default + // deleter will call GRPC Server's base class's dtor and + // that will cause a wired crash. + detail::AsyncGRPCServer rpc_service(endpoint, true); + framework::ProgramDesc empty_program; + framework::Executor executor(dev_ctx.GetPlace()); + rpc_service.SetScope(scope); + rpc_service.SetDevCtx(&dev_ctx); + rpc_service.SetProgram(&empty_program); + rpc_service.SetExecutor(&executor); + + std::thread server_thread( + std::bind(&detail::AsyncGRPCServer::RunSyncUpdate, &rpc_service)); + rpc_service.SetCond(0); + VLOG(3) << "start getting nccl id from trainer 0..."; + auto recv = rpc_service.Get(); + VLOG(3) << "got nccl id and stop server..."; + rpc_service.ShutDown(); + VLOG(3) << "rpc server stopped"; + server_thread.join(); + } +}; + +class GenNCCLIdOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddOutput("NCCLID", "Raw variable contains a NCCL UniqueId instaces."); + AddComment(R"DOC( +GenNCCLId operator + +For trainer 0: generate a new UniqueId and send it to all the other trainers. +For trainer 1~n: start a gRPC server to get the UniqueId, once got, stop the server. +)DOC"); + AddAttr("endpoint", + "(string), e.g. 127.0.0.1:6175 " + "current listen endpoint"); + AddAttr>( + "endpoint_list", + "['trainer1_ip:port', 'trainer2_ip:port', ...] " + "list of trainer endpoints start from trainer 1") + .SetDefault({}); + AddAttr("trainer_id", + "(int default 0) " + "The index of the trainer in distributed training.") + .SetDefault(0); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; + +REGISTER_OPERATOR(gen_nccl_id, ops::GenNCCLIdOp, ops::GenNCCLIdOpMaker); diff --git a/paddle/fluid/operators/get_places_op.cc b/paddle/fluid/operators/get_places_op.cc index 0d7219ac5c624236b85916d5faf6810dbed2198a..eafc364a15fa17cc5107bba737b0b44e712b0bef 100644 --- a/paddle/fluid/operators/get_places_op.cc +++ b/paddle/fluid/operators/get_places_op.cc @@ -78,8 +78,7 @@ class GetPlacesOp : public framework::OperatorBase { class GetPlacesOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - GetPlacesOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddOutput("Out", "vector of Place"); AddAttr("device_count", "device count").SetDefault(0); AddAttr("device_type", "device type") diff --git a/paddle/fluid/operators/go_op.cc b/paddle/fluid/operators/go_op.cc index b8e1556c23a3b7357ed56d1b83c09622559040a4..48f9d967adc90838dc4c7a09bfaf5a5a1ac9c99b 100644 --- a/paddle/fluid/operators/go_op.cc +++ b/paddle/fluid/operators/go_op.cc @@ -89,8 +89,7 @@ class GoOp : public framework::OperatorBase { class GoOpMaker : public framework::OpProtoAndCheckerMaker { public: - GoOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput(kX, "A set of variables, which are required by operators inside the " "block of Go Op.") diff --git a/paddle/fluid/operators/gru_op.cc b/paddle/fluid/operators/gru_op.cc index 0a524c914d305661745c5d85cbbee2edb57c97ba..5c746878823b3dcde2573feec00d3d9dac5ceab8 100644 --- a/paddle/fluid/operators/gru_op.cc +++ b/paddle/fluid/operators/gru_op.cc @@ -71,8 +71,7 @@ class GRUOp : public framework::OperatorWithKernel { class GRUOpMaker : public framework::OpProtoAndCheckerMaker { public: - GRUOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Input", "(LoDTensor) The first input is a LodTensor, which supports " "variable-time length input sequence. The underlying tensor in " diff --git a/paddle/fluid/operators/gru_op.h b/paddle/fluid/operators/gru_op.h index 53f844a6607bd2e98c53b53c23422f6b48e2ced6..3b0d93e54b72910de1429ddf41eb6b0fe9646942 100644 --- a/paddle/fluid/operators/gru_op.h +++ b/paddle/fluid/operators/gru_op.h @@ -34,7 +34,7 @@ inline void ReorderInitState(const DeviceContext& ctx, framework::Tensor* dst, bool indexed_src) { math::CopyMatrixRowsFunctor row_shuffle; dst->mutable_data(src.dims(), ctx.GetPlace()); - row_shuffle(ctx, src, index_lod, *dst, indexed_src); + row_shuffle(ctx, src, index_lod, dst, indexed_src); } template @@ -61,7 +61,7 @@ class GRUKernel : public framework::OpKernel { bool is_reverse = context.Attr("is_reverse"); math::LoDTensor2BatchFunctor to_batch; auto& dev_ctx = context.template device_context(); - to_batch(dev_ctx, *input, *batch_gate, true, is_reverse); + to_batch(dev_ctx, *input, batch_gate, true, is_reverse); if (bias) { math::RowwiseAdd add_bias; @@ -113,7 +113,7 @@ class GRUKernel : public framework::OpKernel { math::Batch2LoDTensorFunctor to_seq; batch_hidden->set_lod(batch_gate->lod()); - to_seq(dev_ctx, *batch_hidden, *hidden); + to_seq(dev_ctx, *batch_hidden, hidden); } void Compute(const framework::ExecutionContext& context) const override { @@ -174,7 +174,7 @@ class GRUGradKernel : public framework::OpKernel { bool is_reverse = context.Attr("is_reverse"); batch_hidden_grad.set_lod(batch_hidden->lod()); - to_batch(dev_ctx, *hidden_grad, batch_hidden_grad, false, is_reverse); + to_batch(dev_ctx, *hidden_grad, &batch_hidden_grad, false, is_reverse); math::GRUMetaValue gru_value; gru_value.gate_weight = const_cast(weight_data); @@ -236,7 +236,7 @@ class GRUGradKernel : public framework::OpKernel { input_grad->mutable_data(context.GetPlace()); math::Batch2LoDTensorFunctor to_seq; batch_gate_grad.set_lod(batch_gate->lod()); - to_seq(dev_ctx, batch_gate_grad, *input_grad); + to_seq(dev_ctx, batch_gate_grad, input_grad); } if (bias_grad) { bias_grad->mutable_data(context.GetPlace()); diff --git a/paddle/fluid/operators/gru_unit_op.cc b/paddle/fluid/operators/gru_unit_op.cc index f8d1d44b5423dd09fe5aad11434911af6f14fe77..82a808b01e99ec33b0ca00a065fb301d3c633b19 100644 --- a/paddle/fluid/operators/gru_unit_op.cc +++ b/paddle/fluid/operators/gru_unit_op.cc @@ -71,8 +71,7 @@ class GRUUnitOp : public framework::OperatorWithKernel { class GRUUnitOpMaker : public framework::OpProtoAndCheckerMaker { public: - GRUUnitOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Input", "(Tensor) Matrix with shape [batch_size, frame_size * 3] for the " "input."); diff --git a/paddle/fluid/operators/gru_unit_op.h b/paddle/fluid/operators/gru_unit_op.h index 15d91ca30593871e2b343eb0e5c0b76aa8055968..2d9faed648aef78da60706e13db3862080c96514 100644 --- a/paddle/fluid/operators/gru_unit_op.h +++ b/paddle/fluid/operators/gru_unit_op.h @@ -14,11 +14,10 @@ limitations under the License. */ #pragma once -#include "paddle/fluid/operators/activation_op.h" -#include "paddle/fluid/operators/math/math_function.h" - #include "paddle/fluid/framework/eigen.h" #include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/activation_op.h" +#include "paddle/fluid/operators/math/blas.h" namespace paddle { namespace operators { @@ -87,10 +86,10 @@ class GRUUnitKernel : public framework::OpKernel { const T* weight_data = weight->data(); T* gate_data = gate->data(); T* reset_hidden_prev_data = reset_hidden_prev->data(); - math::gemm( - context.template device_context(), false, false, - batch_size, 2 * frame_size, frame_size, 1, hidden_prev_data, frame_size, - weight_data, frame_size * 2, 1, gate_data, frame_size * 3); + auto blas = math::GetBlas(context); + blas.GEMM(false, false, batch_size, 2 * frame_size, frame_size, 1, + hidden_prev_data, frame_size, weight_data, frame_size * 2, 1, + gate_data, frame_size * 3); // calculate activited gate Eigen::array extents({{batch_size, frame_size}}); @@ -103,11 +102,10 @@ class GRUUnitKernel : public framework::OpKernel { g.slice(r_offsets, extents), g.slice(r_offsets, extents)); auto r = g.slice(r_offsets, extents); // reset gate r_h_p.device(place) = r * h_p; // reset previous hidden state - math::gemm( - context.template device_context(), false, false, - batch_size, frame_size, frame_size, 1, reset_hidden_prev_data, - frame_size, weight_data + frame_size * frame_size * 2, frame_size, 1, - gate_data + frame_size * 2, frame_size * 3); + blas.GEMM(false, false, batch_size, frame_size, frame_size, 1, + reset_hidden_prev_data, frame_size, + weight_data + frame_size * frame_size * 2, frame_size, 1, + gate_data + frame_size * 2, frame_size * 3); Eigen::array c_offsets({{0, frame_size * 2}}); ActCompute(context.Attr("activation"), place, @@ -188,11 +186,11 @@ class GRUUnitGradKernel : public framework::OpKernel { ActGradCompute(context.Attr("activation"), place, c, c, d_g.slice(c_offsets, extents), d_h * u); // backward for reset_hidden_prev - math::gemm( - context.template device_context(), false, true, - batch_size, frame_size, frame_size, 1, gate_grad_data + frame_size * 2, - frame_size * 3, weight_data + frame_size * frame_size * 2, frame_size, - 0, reset_hidden_prev_grad_data, frame_size); + auto blas = math::GetBlas(context); + blas.GEMM(false, true, batch_size, frame_size, frame_size, 1, + gate_grad_data + frame_size * 2, frame_size * 3, + weight_data + frame_size * frame_size * 2, frame_size, 0, + reset_hidden_prev_grad_data, frame_size); // backward for unactivated reset gate ActGradCompute(context.Attr("gate_activation"), place, r, r, d_g.slice(r_offsets, extents), d_r_h_p * h_p); @@ -200,18 +198,15 @@ class GRUUnitGradKernel : public framework::OpKernel { if (weight_grad) { T* weight_grad_data = weight_grad->mutable_data(context.GetPlace()); // backward for state_weight - math::gemm( - context.template device_context(), true, false, - frame_size, frame_size, batch_size, 1, reset_hidden_prev_data, - frame_size, gate_grad_data + frame_size * 2, frame_size * 3, 0, - weight_grad_data + frame_size * frame_size * 2, frame_size); + blas.GEMM(true, false, frame_size, frame_size, batch_size, 1, + reset_hidden_prev_data, frame_size, + gate_grad_data + frame_size * 2, frame_size * 3, 0, + weight_grad_data + frame_size * frame_size * 2, frame_size); // backward for update_gate_weight and reset_gate_weight - math::gemm( - context.template device_context(), true, false, - frame_size, frame_size * 2, batch_size, 1, hidden_prev_data, - frame_size, gate_grad_data, frame_size * 3, 0, weight_grad_data, - frame_size * 2); + blas.GEMM(true, false, frame_size, frame_size * 2, batch_size, 1, + hidden_prev_data, frame_size, gate_grad_data, frame_size * 3, 0, + weight_grad_data, frame_size * 2); } // backward for hidden_prev if (hidden_prev_grad) { @@ -219,11 +214,9 @@ class GRUUnitGradKernel : public framework::OpKernel { hidden_prev_grad->mutable_data(context.GetPlace()); auto d_h_p = EigenMatrix::From(*hidden_prev_grad); d_h_p.device(place) = d_r_h_p * r + d_h * (u.constant(T(1)) - u); - math::gemm( - context.template device_context(), false, true, - batch_size, frame_size, frame_size * 2, 1, gate_grad_data, - frame_size * 3, weight_data, frame_size * 2, 1, hidden_prev_grad_data, - frame_size); + blas.GEMM(false, true, batch_size, frame_size, frame_size * 2, 1, + gate_grad_data, frame_size * 3, weight_data, frame_size * 2, 1, + hidden_prev_grad_data, frame_size); } // backward for input if (input_grad) { diff --git a/paddle/fluid/operators/hinge_loss_op.cc b/paddle/fluid/operators/hinge_loss_op.cc index 086b5a97dec9a3d5b8f91b802b92d64ca73bf57c..69e7fa4490b892373d85898b13b976a474a6096a 100644 --- a/paddle/fluid/operators/hinge_loss_op.cc +++ b/paddle/fluid/operators/hinge_loss_op.cc @@ -46,8 +46,7 @@ class HingeLossOp : public framework::OperatorWithKernel { template class HingeLossOpMaker : public framework::OpProtoAndCheckerMaker { public: - HingeLossOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Logits", "The input value (Logits) of Hinge loss op." "Logits is a 2-D tensor with shape [batch_size, 1]."); diff --git a/paddle/fluid/operators/huber_loss_op.cc b/paddle/fluid/operators/huber_loss_op.cc index 74d8e0e2b76adc7a3e69649f277a8c0df6f38056..4ecd8634ff41ff4eba6b5ed1d0fc78068190dce5 100644 --- a/paddle/fluid/operators/huber_loss_op.cc +++ b/paddle/fluid/operators/huber_loss_op.cc @@ -45,8 +45,7 @@ class HuberLossOp : public framework::OperatorWithKernel { template class HuberLossOpMaker : public framework::OpProtoAndCheckerMaker { public: - HuberLossOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The input value of huber loss op." "X is a 2-D tensor with shape [batch_size, 1]."); diff --git a/paddle/fluid/operators/im2sequence_op.cc b/paddle/fluid/operators/im2sequence_op.cc index 8c120eec86601146500721bbb4249bc458190093..0669661d225c664010fce97f0a526b62988b92c5 100644 --- a/paddle/fluid/operators/im2sequence_op.cc +++ b/paddle/fluid/operators/im2sequence_op.cc @@ -54,8 +54,7 @@ class Im2SequenceOp : public framework::OperatorWithKernel { class Im2SequenceOpMaker : public framework::OpProtoAndCheckerMaker { public: - Im2SequenceOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor) The input tensor has NCHW format." "N: batch size" diff --git a/paddle/fluid/operators/increment_op.cc b/paddle/fluid/operators/increment_op.cc index d8c97b27b328b1470bece4a6c1872b5ccc75115e..f0ffc9706689f5afe4546c3483114b38bc2b7872 100644 --- a/paddle/fluid/operators/increment_op.cc +++ b/paddle/fluid/operators/increment_op.cc @@ -47,8 +47,7 @@ class IncrementOp : public framework::OperatorWithKernel { class IncrementOpMaker : public framework::OpProtoAndCheckerMaker { public: - IncrementOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor) The input tensor of increment operator"); AddOutput("Out", "(Tensor) The output tensor of increment operator."); AddAttr("step", diff --git a/paddle/fluid/operators/is_empty_op.cc b/paddle/fluid/operators/is_empty_op.cc index 2a7be90dab1cc23ffe5e1c296c37a4bbeacb7d8e..29b73951bbddd9bfd73c932d7801797590de5e8e 100644 --- a/paddle/fluid/operators/is_empty_op.cc +++ b/paddle/fluid/operators/is_empty_op.cc @@ -12,46 +12,41 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ +#include "paddle/fluid/operators/is_empty_op.h" #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/operator.h" namespace paddle { namespace operators { -constexpr char kInput[] = "X"; -constexpr char kOutput[] = "Out"; - -class IsEmptyOp : public framework::OperatorBase { +class IsEmptyOp : public framework::OperatorWithKernel { public: - IsEmptyOp(const std::string &type, const framework::VariableNameMap &inputs, - const framework::VariableNameMap &outputs, - const framework::AttributeMap &attrs) - : OperatorBase(type, inputs, outputs, attrs) {} + using framework::OperatorWithKernel::OperatorWithKernel; - private: - void RunImpl(const framework::Scope &scope, - const platform::Place &place) const override { - // get input - auto *var = scope.FindVar(Input(kInput)); - PADDLE_ENFORCE_NOT_NULL(var); - auto &tensor = var->Get(); - // get output - auto *out = scope.FindVar(Output(kOutput)); - PADDLE_ENFORCE_NOT_NULL(out); - auto *out_tensor = out->GetMutable(); + protected: + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of IsEmptyOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of IsEmptyOp should not be null."); + ctx->SetOutputDim("Out", {1}); + } - out_tensor->Resize({1}); - out_tensor->mutable_data(platform::CPUPlace())[0] = - framework::product(tensor.dims()) == 0; + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext &ctx) const override { + framework::OpKernelType kt = framework::OpKernelType( + framework::ToDataType(ctx.Input("X")->type()), + platform::CPUPlace()); + return kt; } }; -class IsEmptyOpProtoMaker : public framework::OpProtoAndCheckerMaker { +class IsEmptyOpMaker : public framework::OpProtoAndCheckerMaker { public: - IsEmptyOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput(kInput, "(Tensor) Tensor which is to be checked."); - AddOutput(kOutput, "(Tensor) a boolean Tensor that indicate empty or not."); + void Make() override { + AddInput("X", "(LoDTensor) Tensor which is to be checked."); + AddOutput("Out", + "(LoDTensor) a boolean Tensor that indicate empty or not."); AddComment(R"DOC( IsEmpty Operator which checks whether a tensor is empty. @@ -63,5 +58,12 @@ It will just return product(tensor.ddims()) > 0; } // namespace operators } // namespace paddle -REGISTER_OP_WITHOUT_GRADIENT(is_empty, paddle::operators::IsEmptyOp, - paddle::operators::IsEmptyOpProtoMaker); +namespace ops = paddle::operators; + +REGISTER_OPERATOR(is_empty, ops::IsEmptyOp, ops::IsEmptyOpMaker, + paddle::framework::EmptyGradOpMaker); +REGISTER_OP_CPU_KERNEL( + is_empty, ops::IsEmptyOpKernel, + ops::IsEmptyOpKernel, + ops::IsEmptyOpKernel, + ops::IsEmptyOpKernel); diff --git a/paddle/fluid/operators/is_empty_op.h b/paddle/fluid/operators/is_empty_op.h new file mode 100644 index 0000000000000000000000000000000000000000..3e3af22fa8d842b6a1e67418446f1a40949e046b --- /dev/null +++ b/paddle/fluid/operators/is_empty_op.h @@ -0,0 +1,37 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/framework/operator.h" + +namespace paddle { +namespace operators { + +template +class IsEmptyOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + // get input + auto* input_tensor = context.Input("X"); + // get output + auto* output_tensor = context.Output("Out"); + + output_tensor->mutable_data(platform::CPUPlace())[0] = + framework::product(input_tensor->dims()) == 0; + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/l1_norm_op.cc b/paddle/fluid/operators/l1_norm_op.cc index 0c143b7c8aed13a202e2597632d17d8bccc8b66d..bc115090acb473ac3175999ca96c5e00c0aeaeae 100644 --- a/paddle/fluid/operators/l1_norm_op.cc +++ b/paddle/fluid/operators/l1_norm_op.cc @@ -48,8 +48,7 @@ class L1NormGradOp : public framework::OperatorWithKernel { class L1NormOpMaker : public framework::OpProtoAndCheckerMaker { public: - L1NormOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor) The input of l1_norm op."); AddOutput("Out", "(Scalar) The output of l1_norm op."); AddComment(R"DOC( diff --git a/paddle/fluid/operators/label_smooth_op.cc b/paddle/fluid/operators/label_smooth_op.cc index a73c626032f3bf6e97ac5974424e76bacb9a0799..da59bd53bce010d0d6ad2ab14acaffb9cc2f99e6 100644 --- a/paddle/fluid/operators/label_smooth_op.cc +++ b/paddle/fluid/operators/label_smooth_op.cc @@ -47,8 +47,7 @@ class LabelSmoothOp : public framework::OperatorWithKernel { class LabelSmoothOpMaker : public framework::OpProtoAndCheckerMaker { public: - LabelSmoothOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(LoDTensor) The input labels of LabelSmooth operator. This " "input can be batched labels in one-hot encoding or output from " diff --git a/paddle/fluid/operators/layer_norm_op.cc b/paddle/fluid/operators/layer_norm_op.cc index de1056aef7bfa2f53f8a92b262e7d15aa7c2b75c..ab097d31e9ab5eafa788539170e7e405df697625 100644 --- a/paddle/fluid/operators/layer_norm_op.cc +++ b/paddle/fluid/operators/layer_norm_op.cc @@ -61,8 +61,7 @@ class LayerNormOp : public framework::OperatorWithKernel { class LayerNormOpMaker : public framework::OpProtoAndCheckerMaker { public: - LayerNormOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(LoDTensor) The input tensor."); AddInput("Scale", "(Tensor, optional) Scale is a 1-dimensional tensor of size " diff --git a/paddle/fluid/operators/layer_norm_op.h b/paddle/fluid/operators/layer_norm_op.h index 7b84ba0a7daf10e9e636f62eea6bd759ebec9541..2e54bb497dec11eaeda03a1aa6acfd4cc261dbfe 100644 --- a/paddle/fluid/operators/layer_norm_op.h +++ b/paddle/fluid/operators/layer_norm_op.h @@ -15,8 +15,8 @@ limitations under the License. */ #pragma once #include "paddle/fluid/framework/eigen.h" #include "paddle/fluid/framework/op_registry.h" - #include "paddle/fluid/operators/elementwise_op_function.h" +#include "paddle/fluid/operators/math/blas.h" #include "paddle/fluid/operators/math/math_function.h" namespace paddle { @@ -46,9 +46,9 @@ class RowwiseMean2D { } void operator()(const platform::CUDADeviceContext& context, const framework::Tensor& input, framework::Tensor* out) { - math::gemv( - context, false, left_, right_, 1., input.data(), divisor_.data(), - 0., out->data()); + math::GetBlas(context).GEMV( + false, left_, right_, 1., input.data(), divisor_.data(), 0., + out->data()); } private: @@ -93,9 +93,9 @@ class ColwiseSum2D { void operator()(const platform::CUDADeviceContext& context, const framework::Tensor& input, framework::Tensor* out) { - math::gemv( - context, true, left_, right_, 1., input.data(), divisor_.data(), - 0., out->data()); + math::GetBlas(context).GEMV( + true, left_, right_, 1., input.data(), divisor_.data(), 0., + out->data()); } private: diff --git a/paddle/fluid/operators/linear_chain_crf_op.cc b/paddle/fluid/operators/linear_chain_crf_op.cc index 2f29e377fdada918f2c9dca8c2d94eb06278320d..e38525cd7f44de020f364ffd16e71a439048347f 100644 --- a/paddle/fluid/operators/linear_chain_crf_op.cc +++ b/paddle/fluid/operators/linear_chain_crf_op.cc @@ -19,8 +19,7 @@ namespace operators { class LinearChainCRFOpMaker : public framework::OpProtoAndCheckerMaker { public: - LinearChainCRFOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Emission", "(LoDTensor, default LoDTensor) " "A 2-D LoDTensor with shape [N x D], where N is the size of the " diff --git a/paddle/fluid/operators/listen_and_serv_op.cc b/paddle/fluid/operators/listen_and_serv_op.cc index 57cff680ab89f2df7e71af4056ee06cdf330bbab..df5f229acd75ee3df55d46444a63d9f1915f9d22 100644 --- a/paddle/fluid/operators/listen_and_serv_op.cc +++ b/paddle/fluid/operators/listen_and_serv_op.cc @@ -12,12 +12,15 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ +#include // for removing the port file +#include +#include #include -#include #include // NOLINT #include #include "paddle/fluid/operators/listen_and_serv_op.h" +#include "paddle/fluid/platform/profiler.h" namespace paddle { namespace operators { @@ -26,7 +29,6 @@ void RunServer(std::shared_ptr service) { service->RunSyncUpdate(); VLOG(4) << "RunServer thread end"; } - static void split(const std::string &str, char sep, std::vector *pieces) { pieces->clear(); @@ -45,20 +47,6 @@ static void split(const std::string &str, char sep, } } -static void AsyncExecuteBlock(framework::Executor *executor, - framework::ExecutorPrepareContext *prepared, - framework::Scope *scope) { - std::future future = framework::Async([&executor, &prepared, &scope]() { - try { - executor->RunPreparedContext(prepared, scope, false, false); - } catch (std::exception &e) { - LOG(ERROR) << "run sub program error " << e.what(); - } - }); - // TODO(qiao) maybe we can remove this - future.wait(); -} - static void ParallelExecuteBlocks( const std::vector ¶llel_blkids, framework::Executor *executor, const std::vector> @@ -70,9 +58,8 @@ static void ParallelExecuteBlocks( framework::Async([&executor, &prepared, &program, &scope, idx]() { int run_block = idx; // thread local try { - executor->RunPreparedContext(prepared[run_block].get(), scope, - false, false); - } catch (std::exception &e) { + executor->RunPreparedContext(prepared[run_block].get(), scope); + } catch (const std::exception &e) { LOG(ERROR) << "run sub program error " << e.what(); } })); @@ -80,12 +67,7 @@ static void ParallelExecuteBlocks( for (size_t i = 0; i < fs.size(); ++i) fs[i].wait(); } -static void SavePort(std::shared_ptr rpc_service) { - std::ofstream port_file; - port_file.open("/tmp/paddle.selected_port"); - port_file << rpc_service->GetSelectedPort(); - port_file.close(); -} +std::atomic_int ListenAndServOp::selected_port_{0}; ListenAndServOp::ListenAndServOp(const std::string &type, const framework::VariableNameMap &inputs, @@ -93,13 +75,30 @@ ListenAndServOp::ListenAndServOp(const std::string &type, const framework::AttributeMap &attrs) : OperatorBase(type, inputs, outputs, attrs) {} -int ListenAndServOp::GetSelectedPort() const { - return rpc_service_->GetSelectedPort(); -} +ListenAndServOp::~ListenAndServOp() { Stop(); } void ListenAndServOp::Stop() { rpc_service_->Push(LISTEN_TERMINATE_MESSAGE); + rpc_service_->ShutDown(); server_thread_->join(); + auto file_path = string::Sprintf("/tmp/paddle.%d.port", ::getpid()); + remove(file_path.c_str()); +} + +void ListenAndServOp::SavePort() const { + // NOTE: default write file to /tmp/paddle.selected_port + selected_port_ = rpc_service_->GetSelectedPort(); + auto file_path = string::Sprintf("/tmp/paddle.%d.port", ::getpid()); + std::ofstream port_file; + port_file.open(file_path); + port_file << selected_port_.load(); + port_file.close(); + VLOG(4) << "selected port written to " << file_path; +} + +void ListenAndServOp::WaitServerReady() { + while (selected_port_.load() == 0) { + } } void ListenAndServOp::RunSyncLoop(framework::Executor *executor, @@ -126,7 +125,7 @@ void ListenAndServOp::RunSyncLoop(framework::Executor *executor, // Record received sparse variables, so that // we could reset those after execute optimize program std::vector sparse_vars; - while (!exit_flag) { + while (!exit_flag && !SignalHandler::IsProgramExit()) { // Get from multiple trainers, we don't care about the order in which // the gradients arrives, just add suffix 0~n and merge the gradient. rpc_service_->SetCond(0); @@ -191,9 +190,10 @@ void ListenAndServOp::RunSyncLoop(framework::Executor *executor, // mini-batch. // TODO(Yancey1989): move the reset action into an operator, we couldn't // have any hide logic in the operator. - for (auto &var : sparse_vars) { + for (framework::Variable *var : sparse_vars) { var->GetMutable()->mutable_rows()->clear(); } + rpc_service_->SetCond(1); // FIXME(typhoonzero): use another condition to sync wait clients get. rpc_service_->WaitClientGet(fan_in); @@ -201,18 +201,49 @@ void ListenAndServOp::RunSyncLoop(framework::Executor *executor, } // while(true) } +static void AsyncUpdateThread( + const std::string &var_name, const bool &exit_flag, + const std::shared_ptr &queue, + framework::Executor *executor, + framework::ExecutorPrepareContext *prepared) { + VLOG(3) << "update thread for " << var_name << " started"; + while (!exit_flag && !SignalHandler::IsProgramExit()) { + const detail::ReceivedMessage v = queue->Pop(); + if (SignalHandler::IsProgramExit()) { + VLOG(3) << "update thread for " << var_name << " exit"; + break; + } + auto recv_var_name = v.first; + VLOG(4) << "async update " << recv_var_name; + auto var = v.second->GetVar(); + if (var == nullptr) { + LOG(ERROR) << "Can not find server side var: " << recv_var_name; + PADDLE_THROW("Can not find server side var"); + } + auto fs = framework::Async([var_name, &executor, &v, prepared] { + try { + executor->RunPreparedContext(prepared, + v.second->GetMutableLocalScope()); + } catch (const std::exception &e) { + LOG(ERROR) << "run sub program error " << e.what(); + } + }); + fs.wait(); + } +} + void ListenAndServOp::RunAsyncLoop(framework::Executor *executor, - framework::ProgramDesc *program, - framework::Scope *recv_scope, - framework::BlockDesc *prefetch_block) const { + framework::ProgramDesc *program) const { VLOG(3) << "RunAsyncLoop in"; // grad name to block id std::unordered_map grad_to_block_id; std::unordered_map id_to_grad; + std::unordered_map> + grad_to_queue; auto grad_to_block_id_str = Attr>("grad_to_block_id"); - for (auto &grad_and_id : grad_to_block_id_str) { + for (const auto &grad_and_id : grad_to_block_id_str) { std::vector pieces; split(grad_and_id, ':', &pieces); VLOG(3) << "after split, grad = " << pieces[0] << ", id=" << pieces[1]; @@ -220,6 +251,11 @@ void ListenAndServOp::RunAsyncLoop(framework::Executor *executor, PADDLE_ENFORCE_EQ(grad_to_block_id.count(pieces[0]), 0); int block_id = std::stoi(pieces[1]); grad_to_block_id[pieces[0]] = block_id; + std::shared_ptr queue = + std::make_shared(); + grad_to_queue[pieces[0]] = queue; + // record blocking queue in SignalHandler + SignalHandler::RegisterBlockingQueue(queue); id_to_grad[block_id] = pieces[0]; } size_t num_blocks = program->Size(); @@ -238,9 +274,21 @@ void ListenAndServOp::RunAsyncLoop(framework::Executor *executor, grad_to_prepared_ctx[id_to_grad[block_list[i]]] = optimize_prepared[i]; } - VLOG(3) << "RunAsyncLoop into while"; bool exit_flag = false; - while (!exit_flag) { + + VLOG(3) << "start async optimize threads"; + std::vector> fs; + for (auto iter = grad_to_queue.begin(); iter != grad_to_queue.end(); iter++) { + std::string grad_name = iter->first; + VLOG(3) << "create async update thread for " << grad_name; + fs.push_back(framework::AsyncIO([grad_name, &exit_flag, &executor, + &grad_to_queue, &grad_to_prepared_ctx]() { + AsyncUpdateThread(grad_name, exit_flag, grad_to_queue[grad_name], + executor, grad_to_prepared_ctx[grad_name].get()); + })); + } + VLOG(3) << "RunAsyncLoop into while"; + while (!exit_flag && !SignalHandler::IsProgramExit()) { const detail::ReceivedMessage v = rpc_service_->Get(); auto recv_var_name = v.first; if (recv_var_name == LISTEN_TERMINATE_MESSAGE) { @@ -249,13 +297,7 @@ void ListenAndServOp::RunAsyncLoop(framework::Executor *executor, break; } else { VLOG(3) << "received grad: " << recv_var_name; - auto var = v.second->GetVar(); - if (var == nullptr) { - LOG(ERROR) << "Can not find server side var: " << recv_var_name; - PADDLE_THROW("Can not find server side var"); - } - AsyncExecuteBlock(executor, grad_to_prepared_ctx[recv_var_name].get(), - v.second->GetMutableLocalScope()); + grad_to_queue[recv_var_name]->Push(v); } if (exit_flag) { @@ -267,6 +309,8 @@ void ListenAndServOp::RunAsyncLoop(framework::Executor *executor, void ListenAndServOp::RunImpl(const framework::Scope &scope, const platform::Place &dev_place) const { + // Mark this as PS that it should decide profiling by listening from trainer. + platform::SetProfileListener(); platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance(); auto &dev_ctx = *pool.Get(dev_place); framework::Scope &recv_scope = scope.NewScope(); @@ -292,33 +336,35 @@ void ListenAndServOp::RunImpl(const framework::Scope &scope, // prepare for prefetch VLOG(3) << "prefetch block id is " << prefetch_block->ID(); auto prefetch_prepared = executor.Prepare(*program, prefetch_block->ID()); - rpc_service_->SetPrefetchPreparedCtx(prefetch_prepared.get()); - prefetch_prepared.release(); + rpc_service_->SetPrefetchPreparedCtx(std::move(prefetch_prepared)); // start the server listening after all member initialized. server_thread_.reset(new std::thread(RunServer, rpc_service_)); VLOG(3) << "wait server thread to become ready..."; - sleep(5); + rpc_service_->WaitServerReady(); + + // register SIGINT(from ctrl+C) and SIGTERM(from kill) signal handlers + signal(SIGINT, SignalHandler::StopAndExit); + signal(SIGTERM, SignalHandler::StopAndExit); + // Write to a file of server selected port for python use. - SavePort(rpc_service_); + std::string file_path = string::Sprintf("/tmp/paddle.%d.selected_port", + static_cast(::getpid())); + SavePort(); if (sync_mode) { RunSyncLoop(&executor, program, &recv_scope, prefetch_block); } else { - RunAsyncLoop(&executor, program, &recv_scope, prefetch_block); + RunAsyncLoop(&executor, program); } } class ListenAndServOpMaker : public framework::OpProtoAndCheckerMaker { public: - ListenAndServOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() { AddInput("X", "(Tensor) Variables that server recv.").AsDuplicable(); - AddComment(R"DOC( -ListenAndServ operator - -This operator will start a RPC server which can receive variables -from send_op and send back variables to recv_op. -)DOC"); + AddComment(R"DOC(" + "ListenAndServ operator" + "\n" + "This operator" + +" will start a RPC server which can receive variables from send_op and send" + +"back variables to recv_op.)DOC"); AddAttr("endpoint", "(string, default 127.0.0.1:6164)" "IP address to listen on.") @@ -339,6 +385,29 @@ from send_op and send back variables to recv_op. } }; +bool SignalHandler::program_exit_flag_ = false; + +SignalHandler::BlockingQueueSet SignalHandler::blocking_queue_set_{}; + +void SignalHandler::StopAndExit(int signal_num) { + VLOG(3) << "Catch interrupt signal: " << signal_num << ", program will exit"; + + program_exit_flag_ = true; + + // awake all blocking queues + for (BlockingQueueSet::iterator iter = blocking_queue_set_.begin(); + iter != blocking_queue_set_.end(); iter++) { + iter->get()->Push( + std::make_pair(std::string(LISTEN_TERMINATE_MESSAGE), nullptr)); + } + + exit(EXIT_SUCCESS); +} + +void SignalHandler::RegisterBlockingQueue(BlockingQueue &queue) { + blocking_queue_set_.insert(queue); +} + } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/listen_and_serv_op.h b/paddle/fluid/operators/listen_and_serv_op.h index 3cc0f3047733bea94daa310cd39cb0a4f44bef85..6f868369dcf2067fd71f4107d20c79ead0cf9f56 100644 --- a/paddle/fluid/operators/listen_and_serv_op.h +++ b/paddle/fluid/operators/listen_and_serv_op.h @@ -15,7 +15,8 @@ limitations under the License. */ #pragma once #include -#include +#include +#include #include #include "paddle/fluid/framework/executor.h" @@ -39,7 +40,7 @@ class ListenAndServOp : public framework::OperatorBase { const framework::VariableNameMap& outputs, const framework::AttributeMap& attrs); - int GetSelectedPort() const; + virtual ~ListenAndServOp(); void RunSyncLoop(framework::Executor* executor, framework::ProgramDesc* program, @@ -47,18 +48,46 @@ class ListenAndServOp : public framework::OperatorBase { framework::BlockDesc* prefetch_block) const; void RunAsyncLoop(framework::Executor* executor, - framework::ProgramDesc* program, - framework::Scope* recv_scope, - framework::BlockDesc* prefetch_block) const; + framework::ProgramDesc* program) const; + + void SavePort() const; + + void WaitServerReady(); + + int GetSelectedPort() { return selected_port_; } void Stop() override; void RunImpl(const framework::Scope& scope, const platform::Place& dev_place) const override; + static void ResetPort() { selected_port_ = 0; } + protected: mutable std::shared_ptr rpc_service_; mutable std::shared_ptr server_thread_; + // FIXME(wuyi): it's static so that the operator can be cloned. + static std::atomic_int selected_port_; +}; + +class SignalHandler { + public: + typedef std::shared_ptr BlockingQueue; + typedef std::unordered_set BlockingQueueSet; + + public: + static void StopAndExit(int signal_num); + + static void RegisterBlockingQueue(BlockingQueue&); + + static inline bool IsProgramExit() { return program_exit_flag_; } + + private: + static bool program_exit_flag_; + + static BlockingQueueSet blocking_queue_set_; + + DISABLE_COPY_AND_ASSIGN(SignalHandler); }; } // namespace operators diff --git a/paddle/fluid/operators/load_combine_op.cc b/paddle/fluid/operators/load_combine_op.cc index e5353144e91455fc71460459e6e799b54f750f71..0522a94195786c767194ec727d982a60451e7c62 100644 --- a/paddle/fluid/operators/load_combine_op.cc +++ b/paddle/fluid/operators/load_combine_op.cc @@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ #include - +#include "paddle/fluid/framework/data_type_transform.h" #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/platform/device_context.h" @@ -31,6 +31,7 @@ class LoadCombineOp : public framework::OperatorBase { void RunImpl(const framework::Scope &scope, const platform::Place &place) const override { auto filename = Attr("file_path"); + auto load_as_fp16 = Attr("load_as_fp16"); std::ifstream fin(filename); PADDLE_ENFORCE(static_cast(fin), @@ -59,17 +60,25 @@ class LoadCombineOp : public framework::OperatorBase { // Get data from fin to tensor DeserializeFromStream(fin, tensor, dev_ctx); - if (platform::is_gpu_place(place)) { - // copy CPU to GPU - framework::LoDTensor cpu_tensor; - cpu_tensor.ShareDataWith(*tensor); - cpu_tensor.set_lod(tensor->lod()); - - // reset tensor + auto in_dtype = framework::ToDataType(tensor->type()); + auto out_dtype = + load_as_fp16 ? framework::proto::VarType::FP16 : in_dtype; + + if (in_dtype != out_dtype) { + // convert to float16 tensor + auto in_kernel_type = framework::OpKernelType(in_dtype, place); + auto out_kernel_type = framework::OpKernelType(out_dtype, place); + framework::LoDTensor fp16_tensor; + // copy LoD info to the new tensor + fp16_tensor.set_lod(tensor->lod()); + framework::TransDataType(in_kernel_type, out_kernel_type, *tensor, + &fp16_tensor); + + // reset output tensor out_var->Clear(); tensor = out_var->GetMutable(); - tensor->set_lod(cpu_tensor.lod()); - TensorCopy(cpu_tensor, place, dev_ctx, tensor); + tensor->set_lod(fp16_tensor.lod()); + tensor->ShareDataWith(fp16_tensor); } } } @@ -77,12 +86,18 @@ class LoadCombineOp : public framework::OperatorBase { class LoadCombineOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - LoadCombineOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddOutput( "Out", "(vector) The output LoDTensors that will be read from the input file.") .AsDuplicable(); + AddAttr( + "load_as_fp16", + "(boolean, default false)" + "If true, the tensor will be first loaded and then " + "converted to float16 data type. Otherwise, the tensor will be " + "directly loaded without data type conversion.") + .SetDefault(false); AddAttr("file_path", "(string) " "LoDTensors will be loaded from \"file_path\".") diff --git a/paddle/fluid/operators/load_op.cc b/paddle/fluid/operators/load_op.cc index 6ffe0bec5e38432676ecadfa1abbbe70a1425bb1..93f45cff8a26201b1fbb1c44141e125a67c44037 100644 --- a/paddle/fluid/operators/load_op.cc +++ b/paddle/fluid/operators/load_op.cc @@ -13,6 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #include +#include "paddle/fluid/framework/data_type_transform.h" #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/platform/device_context.h" #include "paddle/fluid/platform/profiler.h" @@ -47,26 +48,40 @@ class LoadOp : public framework::OperatorBase { DeserializeFromStream(fin, tensor, *dev_ctx); - if (platform::is_gpu_place(place)) { - // copy CPU to GPU - framework::LoDTensor cpu_tensor; - cpu_tensor.ShareDataWith(*tensor); - cpu_tensor.set_lod(tensor->lod()); - - // reset tensor + auto load_as_fp16 = Attr("load_as_fp16"); + auto in_dtype = framework::ToDataType(tensor->type()); + auto out_dtype = load_as_fp16 ? framework::proto::VarType::FP16 : in_dtype; + + if (in_dtype != out_dtype) { + // convert to float16 tensor + auto in_kernel_type = framework::OpKernelType(in_dtype, place); + auto out_kernel_type = framework::OpKernelType(out_dtype, place); + framework::LoDTensor fp16_tensor; + // copy LoD info to the new tensor + fp16_tensor.set_lod(tensor->lod()); + framework::TransDataType(in_kernel_type, out_kernel_type, *tensor, + &fp16_tensor); + + // reset output tensor out_var->Clear(); tensor = out_var->GetMutable(); - tensor->set_lod(cpu_tensor.lod()); - TensorCopy(cpu_tensor, place, *dev_ctx, tensor); + tensor->set_lod(fp16_tensor.lod()); + tensor->ShareDataWith(fp16_tensor); } } }; class LoadOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - LoadOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddOutput("Out", "(Tensor) The tensor need to be loaded"); + AddAttr( + "load_as_fp16", + "(boolean, default false)" + "If true, the tensor will be first loaded and then " + "converted to float16 data type. Otherwise, the tensor will be " + "directly loaded without data type conversion.") + .SetDefault(false); AddAttr("file_path", "(string) " "Variable will be loaded from \"file_path\".") diff --git a/paddle/fluid/operators/lod_array_length_op.cc b/paddle/fluid/operators/lod_array_length_op.cc index e6212405770093455ec89bde9dc0a092b956fc83..e4551b8ba681fe92ac5f21bb0b509f43439f6b66 100644 --- a/paddle/fluid/operators/lod_array_length_op.cc +++ b/paddle/fluid/operators/lod_array_length_op.cc @@ -40,8 +40,7 @@ class LoDArrayLengthOp : public framework::OperatorBase { class LoDArrayLengthProtoMaker : public framework::OpProtoAndCheckerMaker { public: - LoDArrayLengthProtoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(LoDTensorArray) The input tensor array."); AddOutput("Out", "(Tensor) 1x1 CPU Tensor of length, int64_t"); AddComment(R"DOC( diff --git a/paddle/fluid/operators/lod_rank_table_op.cc b/paddle/fluid/operators/lod_rank_table_op.cc index 590b44e14f518c3c60c141c9a0dfe7f2b96f69c6..166952fe23192799443ef9c9d1f7ba5056d19290 100644 --- a/paddle/fluid/operators/lod_rank_table_op.cc +++ b/paddle/fluid/operators/lod_rank_table_op.cc @@ -38,8 +38,7 @@ class LoDRankTableOp : public framework::OperatorBase { class LoDRankTableOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - LoDRankTableOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(LoDTensor) input lod tensor, must contain lod information."); AddOutput("Out", "(LoDRankTable) The rank table of specific level."); diff --git a/paddle/fluid/operators/lod_reset_op.cc b/paddle/fluid/operators/lod_reset_op.cc index 92ebfc274b84f738f5bd688a9a6d9f437b6318aa..0d4e84e85083399e3803d0648dc7a10aa276d536 100644 --- a/paddle/fluid/operators/lod_reset_op.cc +++ b/paddle/fluid/operators/lod_reset_op.cc @@ -47,8 +47,7 @@ class LoDResetOp : public framework::OperatorWithKernel { class LoDResetOpMaker : public framework::OpProtoAndCheckerMaker { public: - LoDResetOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor, LoDTensor) Input variable of LoDResetOp which " "could be a Tensor or LoDTensor, where the data of output " diff --git a/paddle/fluid/operators/lod_reset_op.h b/paddle/fluid/operators/lod_reset_op.h index bd19d8908e35e51872d324ea5aa6bb02110d5a92..d36aa0ce025a1c0f717913131fcc75040d16afac 100644 --- a/paddle/fluid/operators/lod_reset_op.h +++ b/paddle/fluid/operators/lod_reset_op.h @@ -46,8 +46,7 @@ class LoDResetKernel : public framework::OpKernel { auto* lod = lod_t->data(); if (platform::is_gpu_place(ctx.GetPlace())) { framework::Tensor lod_cpu; - framework::TensorCopy(*lod_t, platform::CPUPlace(), - ctx.device_context(), &lod_cpu); + framework::TensorCopySync(*lod_t, platform::CPUPlace(), &lod_cpu); lod = lod_cpu.data(); } level0 = std::vector(lod, lod + lod_t->numel()); diff --git a/paddle/fluid/operators/lod_tensor_to_array_op.cc b/paddle/fluid/operators/lod_tensor_to_array_op.cc index 543495ce4e66c0955c9ce1b0db480088069b36db..00ba5ce8ee5e4084c8af204cfc37fe80c437f0d7 100644 --- a/paddle/fluid/operators/lod_tensor_to_array_op.cc +++ b/paddle/fluid/operators/lod_tensor_to_array_op.cc @@ -105,8 +105,7 @@ class LoDTensorToArrayOp : public framework::OperatorBase { class LoDTensorToArrayOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - LoDTensorToArrayOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", ""); AddInput("RankTable", ""); AddOutput("Out", ""); diff --git a/paddle/fluid/operators/log_loss_op.cc b/paddle/fluid/operators/log_loss_op.cc index a8258a1afd70574c174abe8d5630ade5d4ac3de6..9d248e03218b83a65b9786cb317aafbe3dbb67ee 100644 --- a/paddle/fluid/operators/log_loss_op.cc +++ b/paddle/fluid/operators/log_loss_op.cc @@ -46,8 +46,7 @@ class LogLossOp : public framework::OperatorWithKernel { template class LogLossOpMaker : public framework::OpProtoAndCheckerMaker { public: - LogLossOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Predicted", "The input value (Predicted) of Log loss op." "Predicted is a 2-D tensor with shape [batch_size, 1]."); diff --git a/paddle/fluid/operators/logical_op.cc b/paddle/fluid/operators/logical_op.cc index 41aa00ee8ac10e0776c066fc3c37f97b0dd40cc3..db109f5cd053d84718ac85bd4693ecece12ce172 100644 --- a/paddle/fluid/operators/logical_op.cc +++ b/paddle/fluid/operators/logical_op.cc @@ -21,8 +21,7 @@ namespace operators { template class BinaryLogicalOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - BinaryLogicalOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { OpComment comment; AddInput("X", string::Sprintf("(LoDTensor) Left hand operand of %s operator", @@ -45,8 +44,7 @@ Each element of Out is calculated by %s template class UnaryLogicalOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - UnaryLogicalOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { OpComment comment; AddInput("X", string::Sprintf("(LoDTensor) Operand of %s operator", comment.type)); diff --git a/paddle/fluid/operators/lookup_sparse_table_op.cc b/paddle/fluid/operators/lookup_sparse_table_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..2ce11e712fb1a8aa9748313ec7cf4e895a931465 --- /dev/null +++ b/paddle/fluid/operators/lookup_sparse_table_op.cc @@ -0,0 +1,164 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include + +#include "paddle/fluid/framework/data_type.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/math/math_function.h" +#include "paddle/fluid/platform/device_context.h" + +namespace paddle { +namespace operators { + +constexpr int64_t kNoPadding = -1; + +class LookupSparseTableInferShape : public framework::InferShapeBase { + public: + void operator()(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of LookupSparseTableOp should not be null."); + auto shape_w = ctx->GetInputDim("W"); + auto shape_ids = ctx->GetInputDim("Ids"); + shape_w[0] = shape_ids.size(); + ctx->SetOutputDim("Out", shape_w); + } +}; + +class LookupSparseTableOp : public framework::OperatorBase { + public: + using framework::OperatorBase::OperatorBase; + + private: + void RunImpl(const framework::Scope &scope, + const platform::Place &dev_place) const override { + auto out_var = scope.FindVar(Output("Out")); + auto w_var = scope.FindVar(Input("W")); + auto ids_var = scope.FindVar(Input("Ids")); + unsigned int seed = static_cast(Attr("seed")); + float min = Attr("min"); + float max = Attr("max"); + bool auto_grown_table = Attr("auto_grown_table"); + + PADDLE_ENFORCE(out_var->IsType(), + "The type of Out var should be LodTensor."); + PADDLE_ENFORCE(w_var->IsType(), + "The type of W var should be SelectedRows."); + PADDLE_ENFORCE(ids_var->IsType(), + "The type of Ids var should be LoDTensor."); + auto &ids_t = ids_var->Get(); + auto out_t = out_var->GetMutable(); + auto w_t = w_var->GetMutable(); + std::vector keys; + keys.resize(ids_t.numel()); + for (int64_t i = 0; i < ids_t.numel(); ++i) { + keys[i] = ids_t.data()[i]; + } + + // TODO(Yancey1989): support CUDA Place for the sparse table + platform::CPUPlace cpu; + auto out_shape = w_t->value().dims(); + out_shape[0] = keys.size(); + out_t->Resize(out_shape); + out_t->mutable_data(cpu, w_t->value().type()); + PADDLE_ENFORCE_EQ(framework::ToDataType(w_t->value().type()), + framework::proto::VarType::FP32, + "The sparse table only support FP32"); + auto non_keys_pair = w_t->Get(keys, out_t); + if (!auto_grown_table) { + PADDLE_ENFORCE_EQ(non_keys_pair.size(), static_cast(0), + "there is some keys does exists in the sparse table."); + } + auto value_shape = w_t->value().dims(); + value_shape[0] = 1; + for (const auto &it : non_keys_pair) { + const auto key = it.first; + const auto index = it.second; + framework::Tensor value; + value.Resize(value_shape); + auto data = value.mutable_data(cpu); + + std::minstd_rand engine; + engine.seed(seed); + std::uniform_real_distribution dist(min, max); + int64_t size = value.numel(); + for (int64_t i = 0; i < size; ++i) { + data[i] = dist(engine); + } + w_t->Set(key, value); + memory::Copy(cpu, out_t->mutable_data(cpu) + index * value.numel(), + cpu, value.data(), value.numel() * sizeof(float)); + } + } +}; + +class LookupSparseTableOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("W", + "(SelectedRows) The input represents embedding table, " + "which is a learnable parameter."); + AddInput("Ids", + "(LoDTensor) Ids's type should be LoDTensor" + "THe ids to be looked up in W."); + AddOutput("Out", + "(LoDTensor) The lookup results, which have the " + "same type as W."); + AddAttr("padding_idx", + "(int64, default -1) " + "If the value is -1, it makes no effect to lookup. " + "Otherwise the given value indicates padding the output " + "with zeros whenever lookup encounters it in Ids.") + .SetDefault(kNoPadding); + AddAttr("min", + "(float, default -1.0) " + "Minimum value of uniform random") + .SetDefault(-1.0f); + AddAttr("max", + "(float, default 1.0) " + "Maximum value of uniform random") + .SetDefault(1.0f); + AddAttr("seed", + "(int, default 0) " + "Random seed used for generating samples. " + "0 means use a seed generated by the system." + "Note that if seed is not 0, this operator will always " + "generate the same random numbers every time.") + .SetDefault(0); + AddAttr("auto_grown_table", + "(bool default false)" + "Whether create new value if for nonexistent key.") + .SetDefault(true); + AddComment(R"DOC( +Lookup Sprase Tablel Operator. + +This operator is used to perform lookup on parameter W, +then concatenated into a sparse tensor. + +The type of Ids(Input) is SelectedRows, the rows of Ids contains +the ids to be looked up in W; +if the Id is not in the sparse table, this operator will return a +random value and set the value into the table for the next looking up. + +)DOC"); + } +}; +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OPERATOR(lookup_sparse_table, ops::LookupSparseTableOp, + ops::LookupSparseTableInferShape, + ops::LookupSparseTableOpMaker, + paddle::framework::EmptyGradOpMaker); diff --git a/paddle/fluid/operators/lookup_table_op.cc b/paddle/fluid/operators/lookup_table_op.cc index 5e59bd1b178ad1803f6f70c5f3f9fd7af495ac3c..bda499432214b8841c8dfc406ee45ca0367920e7 100644 --- a/paddle/fluid/operators/lookup_table_op.cc +++ b/paddle/fluid/operators/lookup_table_op.cc @@ -58,8 +58,7 @@ class LookupTableOp : public framework::OperatorWithKernel { class LookupTableOpMaker : public framework::OpProtoAndCheckerMaker { public: - LookupTableOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("W", "(Tensor) The input represents embedding tensors, " "which is a learnable parameter."); diff --git a/paddle/fluid/operators/lookup_table_op.cu b/paddle/fluid/operators/lookup_table_op.cu index 6d81fccd2059c511f71d403229e04587e553e93d..77722c50d39003d9342afb04a61ae3aaf6b21100 100644 --- a/paddle/fluid/operators/lookup_table_op.cu +++ b/paddle/fluid/operators/lookup_table_op.cu @@ -16,7 +16,7 @@ limitations under the License. */ #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/operators/lookup_table_op.h" #include "paddle/fluid/platform/assert.h" -#include "paddle/fluid/platform/cuda_helper.h" +#include "paddle/fluid/platform/cuda_primitives.h" namespace paddle { namespace operators { diff --git a/paddle/fluid/operators/lrn_op.cc b/paddle/fluid/operators/lrn_op.cc index f5c0e47fda913b4635833c31496644b60a0a8504..52b9cd7fb7019b738098a8649f23277afd40e938 100644 --- a/paddle/fluid/operators/lrn_op.cc +++ b/paddle/fluid/operators/lrn_op.cc @@ -169,8 +169,7 @@ class LRNOp : public framework::OperatorWithKernel { template class LRNOpMaker : public framework::OpProtoAndCheckerMaker { public: - LRNOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor) The input of LRN operator. " "It must be a 4D tenor with NCHW format."); diff --git a/paddle/fluid/operators/lstm_op.cc b/paddle/fluid/operators/lstm_op.cc index 084ee1cfe602af3622ef2a3f35f2892d5540cec7..4751e3e8025e51a687f8fcfd25e603b61e762f6d 100644 --- a/paddle/fluid/operators/lstm_op.cc +++ b/paddle/fluid/operators/lstm_op.cc @@ -103,8 +103,7 @@ class LSTMOp : public framework::OperatorWithKernel { class LSTMOpMaker : public framework::OpProtoAndCheckerMaker { public: - LSTMOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Input", "(LoDTensor) the first input is a LodTensor, which support " "variable-time length input sequence. The underlying tensor in " diff --git a/paddle/fluid/operators/lstm_op.h b/paddle/fluid/operators/lstm_op.h index a1ef0eb278dea7205cd8052bbe006b0ae4e3a466..7d62d2d020ec2e3a29ad8720a8f04fead3a90a63 100644 --- a/paddle/fluid/operators/lstm_op.h +++ b/paddle/fluid/operators/lstm_op.h @@ -15,9 +15,9 @@ limitations under the License. */ #pragma once #include #include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/math/blas.h" #include "paddle/fluid/operators/math/detail/activation_functions.h" #include "paddle/fluid/operators/math/lstm_compute.h" -#include "paddle/fluid/operators/math/math_function.h" #include "paddle/fluid/operators/math/sequence2batch.h" namespace paddle { @@ -33,7 +33,7 @@ inline void ReorderInitState(const DeviceContext& ctx, framework::Tensor* dst, bool indexed_src) { math::CopyMatrixRowsFunctor row_shuffle; dst->mutable_data(src.dims(), ctx.GetPlace()); - row_shuffle(ctx, src, index_lod, *dst, indexed_src); + row_shuffle(ctx, src, index_lod, dst, indexed_src); } template @@ -57,7 +57,7 @@ class LSTMKernel : public framework::OpKernel { bool is_reverse = ctx.Attr("is_reverse"); math::LoDTensor2BatchFunctor to_batch; auto& device_ctx = ctx.template device_context(); - to_batch(device_ctx, *input, *batch_gate, true, is_reverse); + to_batch(device_ctx, *input, batch_gate, true, is_reverse); auto in_dims = input->dims(); int frame_size = static_cast(in_dims[1] / 4); @@ -114,6 +114,7 @@ class LSTMKernel : public framework::OpKernel { auto cand_act = math::detail::GetActivationType( ctx.Attr("candidate_activation")); + auto blas = math::GetBlas(device_ctx); for (size_t n = 0; n < num_batch; n++) { int bstart = static_cast(batch_starts[n]); int bend = static_cast(batch_starts[n + 1]); @@ -129,9 +130,8 @@ class LSTMKernel : public framework::OpKernel { int pre_h_start = static_cast(batch_starts[n - 1]); int pre_h_end = pre_h_start + cur_batch_size; auto pre_hidden_t = batch_hidden.Slice(pre_h_start, pre_h_end); - math::matmul(device_ctx, pre_hidden_t, false, *weight, - false, static_cast(1.0), &gate_t, - static_cast(1.0)); + blas.MatMul(pre_hidden_t, false, *weight, false, static_cast(1.0), + &gate_t, static_cast(1.0)); } else if (hidden_t0) { // If n == 0 and there is no initialized hidden state, that is to say // the H0 is zeros, the calculation W_h * H0 will be skiped. @@ -143,9 +143,8 @@ class LSTMKernel : public framework::OpKernel { Tensor ordered_h0; ReorderInitState(device_ctx, *hidden_t0, order, &ordered_h0, true); - math::matmul(device_ctx, ordered_h0, false, *weight, - false, static_cast(1.0), &gate_t, - static_cast(1.0)); + blas.MatMul(ordered_h0, false, *weight, false, static_cast(1.0), + &gate_t, static_cast(1.0)); } lstm_value.gate_value = gate_t.data(); @@ -161,11 +160,11 @@ class LSTMKernel : public framework::OpKernel { math::Batch2LoDTensorFunctor to_seq; batch_hidden.set_lod(batch_gate->lod()); // restore the output hidden in LoDTensor from the batch hidden - to_seq(device_ctx, batch_hidden, *hidden_out); + to_seq(device_ctx, batch_hidden, hidden_out); batch_cell.set_lod(batch_gate->lod()); // restore the output cell state in LoDTensor from the batch cell - to_seq(device_ctx, batch_cell, *cell_out); + to_seq(device_ctx, batch_cell, cell_out); } }; @@ -257,7 +256,7 @@ class LSTMGradKernel : public framework::OpKernel { const framework::DDim& dims, framework::LoDTensor& dst) { dst.mutable_data(dims, ctx.GetPlace()); dst.set_lod(batch_gate->lod()); - to_batch(ctx, src, dst, false); + to_batch(ctx, src, &dst, false); }; LoDTensor batch_hidden, batch_hidden_g, batch_cell; @@ -282,6 +281,7 @@ class LSTMGradKernel : public framework::OpKernel { auto batch_starts = batch_gate->lod()[0]; size_t num_batch = batch_starts.size() - 1; + auto blas = math::GetBlas(device_ctx); for (int n = static_cast(num_batch) - 1; n >= 0; n--) { int bstart = static_cast(batch_starts[n]); int bend = static_cast(batch_starts[n + 1]); @@ -320,29 +320,25 @@ class LSTMGradKernel : public framework::OpKernel { int pre_h_start = static_cast(batch_starts[n - 1]); int pre_h_end = pre_h_start + cur_batch_size; auto pre_hidden_g = batch_hidden_g.Slice(pre_h_start, pre_h_end); - math::matmul(device_ctx, gate_g, false, *weight, true, - static_cast(1.0), &pre_hidden_g, - static_cast(1.0)); + blas.MatMul(gate_g, false, *weight, true, static_cast(1.0), + &pre_hidden_g, static_cast(1.0)); if (weight_g) { /* backward weight */ auto pre_hidden = batch_hidden.Slice(pre_h_start, pre_h_end); - math::matmul(device_ctx, pre_hidden, true, gate_g, - false, static_cast(1.0), weight_g, - static_cast(1.0)); + blas.MatMul(pre_hidden, true, gate_g, false, static_cast(1.0), + weight_g, static_cast(1.0)); } } else { if (h0 && weight_g) { ReorderInitState(device_ctx, *h0, order, &ordered_h0, true); - math::matmul(device_ctx, ordered_h0, true, gate_g, - false, static_cast(1.0), weight_g, - static_cast(1.0)); + blas.MatMul(ordered_h0, true, gate_g, false, static_cast(1.0), + weight_g, static_cast(1.0)); } if (h0 && h0_g) { ordered_h0_g.mutable_data(h0_g->dims(), ctx.GetPlace()); - math::matmul(device_ctx, gate_g, false, *weight, - true, static_cast(1.0), - &ordered_h0_g, static_cast(0.0)); + blas.MatMul(gate_g, false, *weight, true, static_cast(1.0), + &ordered_h0_g, static_cast(0.0)); } } } @@ -351,7 +347,7 @@ class LSTMGradKernel : public framework::OpKernel { if (in_g) { /* backward data */ in_g->mutable_data(ctx.GetPlace()); - to_seq(device_ctx, batch_gate_g, *in_g); + to_seq(device_ctx, batch_gate_g, in_g); } if (bias && bias_g) { /* backward bias */ diff --git a/paddle/fluid/operators/lstm_unit_op.cc b/paddle/fluid/operators/lstm_unit_op.cc index e1157ef6c640be17e7f48abe1ab972cf88504526..0895c58f5f58afd444000ebeac7a92e3eb7778d3 100644 --- a/paddle/fluid/operators/lstm_unit_op.cc +++ b/paddle/fluid/operators/lstm_unit_op.cc @@ -48,8 +48,7 @@ class LstmUnitOp : public framework::OperatorWithKernel { class LstmUnitOpMaker : public framework::OpProtoAndCheckerMaker { public: - LstmUnitOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "Lstm unit only applies non-linear activations, please make sure" "that linear tranformation has already been applied to `X`. " diff --git a/paddle/fluid/operators/lstmp_op.cc b/paddle/fluid/operators/lstmp_op.cc index f9261323f0f50c78b3b4b66a9fa8abcdf5ba27e9..e398b51480f6fc0c6c568770b3b2a9746360744e 100644 --- a/paddle/fluid/operators/lstmp_op.cc +++ b/paddle/fluid/operators/lstmp_op.cc @@ -120,8 +120,7 @@ class LSTMPOp : public framework::OperatorWithKernel { class LSTMPOpMaker : public framework::OpProtoAndCheckerMaker { public: - LSTMPOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Input", "(LoDTensor) the input for sequence data, which supports " "variable-time length input sequence. The underlying tensor in " diff --git a/paddle/fluid/operators/lstmp_op.h b/paddle/fluid/operators/lstmp_op.h index 172db548960135fbc1841cf58b73894d4f74d838..370dd04d1449a8e211febf9a4f9e90e6f5008e20 100644 --- a/paddle/fluid/operators/lstmp_op.h +++ b/paddle/fluid/operators/lstmp_op.h @@ -14,15 +14,14 @@ limitations under the License. */ #pragma once #include +#include "paddle/fluid/framework/eigen.h" +#include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/operators/activation_op.h" +#include "paddle/fluid/operators/math/blas.h" #include "paddle/fluid/operators/math/detail/activation_functions.h" #include "paddle/fluid/operators/math/lstm_compute.h" -#include "paddle/fluid/operators/math/math_function.h" #include "paddle/fluid/operators/math/sequence2batch.h" -#include "paddle/fluid/framework/eigen.h" -#include "paddle/fluid/framework/op_registry.h" - namespace paddle { namespace operators { @@ -40,7 +39,7 @@ inline void ReorderInitState(const DeviceContext& ctx, framework::Tensor* dst, bool indexed_src) { math::CopyMatrixRowsFunctor row_shuffle; dst->mutable_data(src.dims(), ctx.GetPlace()); - row_shuffle(ctx, src, index, *dst, indexed_src); + row_shuffle(ctx, src, index, dst, indexed_src); } template @@ -81,7 +80,7 @@ class LSTMPKernel : public framework::OpKernel { bool is_reverse = ctx.Attr("is_reverse"); math::LoDTensor2BatchFunctor to_batch; auto& device_ctx = ctx.template device_context(); - to_batch(device_ctx, *input, *batch_gate, true, is_reverse); + to_batch(device_ctx, *input, batch_gate, true, is_reverse); auto in_dims = input->dims(); int frame_size = static_cast(in_dims[1] / 4); @@ -143,7 +142,7 @@ class LSTMPKernel : public framework::OpKernel { auto proj_act = math::detail::GetActivationType( ctx.Attr("proj_activation")); auto& place = *ctx.template device_context().eigen_device(); - + auto blas = math::GetBlas(device_ctx); for (size_t n = 0; n < num_batch; n++) { int bstart = static_cast(batch_starts[n]); int bend = static_cast(batch_starts[n + 1]); @@ -160,9 +159,8 @@ class LSTMPKernel : public framework::OpKernel { int pre_h_start = static_cast(batch_starts[n - 1]); int pre_h_end = pre_h_start + cur_batch_size; auto pre_proj_t = batch_proj.Slice(pre_h_start, pre_h_end); - math::matmul(device_ctx, pre_proj_t, false, *weight, - false, static_cast(1.0), &gate_t, - static_cast(1.0)); + blas.MatMul(pre_proj_t, false, *weight, false, static_cast(1.0), + &gate_t, static_cast(1.0)); } else if (hidden_t0) { // If n == 0 and there is no initialized hidden state, that is to say // the H0 is zeros, the calculation W_h * H0 will be skiped. @@ -176,16 +174,14 @@ class LSTMPKernel : public framework::OpKernel { ordered_proj0->mutable_data(ctx.GetPlace()); ReorderInitState(device_ctx, *hidden_t0, order, &ordered_h0, true); - math::matmul(device_ctx, ordered_h0, false, - *proj_weight, false, static_cast(1.0), - ordered_proj0, static_cast(0.0)); + blas.MatMul(ordered_h0, false, *proj_weight, false, static_cast(1.0), + ordered_proj0, static_cast(0.0)); if (proj_act != math::detail::ActivationType::kIdentity) { auto proj0_dev = EigenMatrix::From(*ordered_proj0); ActCompute(cell_act, place, proj0_dev, proj0_dev); } - math::matmul(device_ctx, *ordered_proj0, false, - *weight, false, static_cast(1.0), - &gate_t, static_cast(1.0)); + blas.MatMul(*ordered_proj0, false, *weight, false, static_cast(1.0), + &gate_t, static_cast(1.0)); } lstmp_value.gate_value = gate_t.data(); @@ -196,9 +192,8 @@ class LSTMPKernel : public framework::OpKernel { device_ctx, lstmp_value, frame_size, cur_batch_size, gate_act, cell_act, cand_act); lstmp_value.prev_state_value = lstmp_value.state_value; - math::matmul(device_ctx, hidden_t, false, *proj_weight, - false, static_cast(1.0), &proj_t, - static_cast(0.0)); + blas.MatMul(hidden_t, false, *proj_weight, false, static_cast(1.0), + &proj_t, static_cast(0.0)); if (proj_act != math::detail::ActivationType::kIdentity) { auto proj_t_dev = EigenMatrix::From(proj_t); ActCompute(cell_act, place, proj_t_dev, proj_t_dev); @@ -208,11 +203,11 @@ class LSTMPKernel : public framework::OpKernel { math::Batch2LoDTensorFunctor to_seq; batch_proj.set_lod(batch_gate->lod()); // restore the output hidden in LoDTensor from the batch hidden - to_seq(device_ctx, batch_proj, *proj_out); + to_seq(device_ctx, batch_proj, proj_out); batch_cell.set_lod(batch_gate->lod()); // restore the output cell state in LoDTensor from the batch cell - to_seq(device_ctx, batch_cell, *cell_out); + to_seq(device_ctx, batch_cell, cell_out); } }; @@ -332,7 +327,7 @@ class LSTMPGradKernel : public framework::OpKernel { const framework::DDim& dims, framework::LoDTensor& dst) { dst.mutable_data(dims, ctx.GetPlace()); dst.set_lod(batch_gate->lod()); - to_batch(ctx, src, dst, false); + to_batch(ctx, src, &dst, false); }; LoDTensor batch_hidden_g, batch_proj, batch_proj_g, batch_cell; @@ -361,6 +356,7 @@ class LSTMPGradKernel : public framework::OpKernel { auto batch_starts = batch_gate->lod()[0]; size_t num_batch = batch_starts.size() - 1; + auto blas = math::GetBlas(device_ctx); for (int n = static_cast(num_batch) - 1; n >= 0; n--) { int bstart = static_cast(batch_starts[n]); int bend = static_cast(batch_starts[n + 1]); @@ -375,15 +371,13 @@ class LSTMPGradKernel : public framework::OpKernel { } /* hidden state backwarad */ Tensor out_g = batch_hidden_g.Slice(bstart, bend); - math::matmul(device_ctx, proj_g, false, *proj_weight, - true, static_cast(1.0), &out_g, - static_cast(0.0)); + blas.MatMul(proj_g, false, *proj_weight, true, static_cast(1.0), + &out_g, static_cast(0.0)); /* projection weight backward*/ if (proj_weight_g) { Tensor hidden_t = batch_hidden->Slice(bstart, bend); - math::matmul(device_ctx, hidden_t, true, proj_g, - false, static_cast(1.0), - proj_weight_g, static_cast(1.0)); + blas.MatMul(hidden_t, true, proj_g, false, static_cast(1.0), + proj_weight_g, static_cast(1.0)); } Tensor gate = batch_gate->Slice(bstart, bend); @@ -419,24 +413,21 @@ class LSTMPGradKernel : public framework::OpKernel { int pre_h_start = static_cast(batch_starts[n - 1]); int pre_h_end = pre_h_start + cur_batch_size; auto pre_proj_g = batch_proj_g.Slice(pre_h_start, pre_h_end); - math::matmul(device_ctx, gate_g, false, *weight, true, - static_cast(1.0), &pre_proj_g, - static_cast(1.0)); + blas.MatMul(gate_g, false, *weight, true, static_cast(1.0), + &pre_proj_g, static_cast(1.0)); if (weight_g) { /* weight backward*/ auto pre_proj = batch_proj.Slice(pre_h_start, pre_h_end); - math::matmul(device_ctx, pre_proj, true, gate_g, - false, static_cast(1.0), weight_g, - static_cast(1.0)); + blas.MatMul(pre_proj, true, gate_g, false, static_cast(1.0), + weight_g, static_cast(1.0)); } } else { if (h0 && weight_g) { ReorderInitState(device_ctx, *h0, order, &ordered_h0, true); if (weight_g) { - math::matmul(device_ctx, *ordered_proj0, true, - gate_g, false, static_cast(1.0), - weight_g, static_cast(1.0)); + blas.MatMul(*ordered_proj0, true, gate_g, false, + static_cast(1.0), weight_g, static_cast(1.0)); } } if (h0 && (h0_g || proj_weight_g)) { @@ -444,9 +435,8 @@ class LSTMPGradKernel : public framework::OpKernel { Tensor proj0_g; proj0_g.Resize({in_dims[0], proj_weight->dims()[1]}); proj0_g.mutable_data(ctx.GetPlace()); - math::matmul(device_ctx, gate_g, false, *weight, - true, static_cast(1.0), &proj0_g, - static_cast(0.0)); + blas.MatMul(gate_g, false, *weight, true, static_cast(1.0), + &proj0_g, static_cast(0.0)); if (proj_act != math::detail::ActivationType::kIdentity) { auto proj0_dev = EigenMatrix::From(*ordered_proj0); auto proj0_g_dev = EigenMatrix::From(proj0_g); @@ -454,14 +444,12 @@ class LSTMPGradKernel : public framework::OpKernel { proj0_g_dev); } if (h0_g) { - math::matmul( - device_ctx, proj0_g, false, *proj_weight, true, - static_cast(1.0), &ordered_h0_g, static_cast(0.0)); + blas.MatMul(proj0_g, false, *proj_weight, true, static_cast(1.0), + &ordered_h0_g, static_cast(0.0)); } if (proj_weight_g) { - math::matmul(device_ctx, ordered_h0, true, - proj0_g, false, static_cast(1.0), - proj_weight_g, static_cast(1.0)); + blas.MatMul(ordered_h0, true, proj0_g, false, static_cast(1.0), + proj_weight_g, static_cast(1.0)); } } } @@ -471,7 +459,7 @@ class LSTMPGradKernel : public framework::OpKernel { if (in_g) { /* backward data */ in_g->mutable_data(ctx.GetPlace()); - to_seq(device_ctx, batch_gate_g, *in_g); + to_seq(device_ctx, batch_gate_g, in_g); } if (bias && bias_g) { /* backward bias */ diff --git a/paddle/fluid/operators/margin_rank_loss_op.cc b/paddle/fluid/operators/margin_rank_loss_op.cc index 0b41a3e1ffdb32d248bb55651aba242336307e74..b643ba9d7fa61d758e871ebe7a463c22e937fa2c 100644 --- a/paddle/fluid/operators/margin_rank_loss_op.cc +++ b/paddle/fluid/operators/margin_rank_loss_op.cc @@ -42,8 +42,7 @@ class MarginRankLossOp : public framework::OperatorWithKernel { template class MarginRankLossOpMaker : public framework::OpProtoAndCheckerMaker { public: - MarginRankLossOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X1", "(2-D tensor with shape [batch_size x 1]) The score for " "one item X1 to be ranked, from pairwise ranking model."); diff --git a/paddle/fluid/operators/math/CMakeLists.txt b/paddle/fluid/operators/math/CMakeLists.txt index ee0e91132bce52998e9c45b37335618e4354e1cd..53a478c1ac0bdf8c0a3f3721161779ef10cb14f8 100644 --- a/paddle/fluid/operators/math/CMakeLists.txt +++ b/paddle/fluid/operators/math/CMakeLists.txt @@ -41,7 +41,8 @@ math_library(depthwise_conv) math_library(gru_compute DEPS activation_functions math_function) math_library(im2col) math_library(lstm_compute DEPS activation_functions) -math_library(math_function DEPS cblas) +cc_library(blas SRCS blas.cc DEPS cblas framework_proto device_context) +math_library(math_function DEPS blas) math_library(maxouting) math_library(pooling) math_library(selected_rows_functor DEPS selected_rows math_function) diff --git a/paddle/fluid/operators/math/blas.cc b/paddle/fluid/operators/math/blas.cc new file mode 100644 index 0000000000000000000000000000000000000000..6a143b3c056455595fdedc131b0c5f4ee756e1e0 --- /dev/null +++ b/paddle/fluid/operators/math/blas.cc @@ -0,0 +1,52 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/operators/math/blas.h" + +#include +namespace paddle { +namespace operators { +namespace math { +MatDescriptor CreateMatrixDescriptor(const framework::DDim &tensor_dim, + int num_flatten_cols, bool trans) { + PADDLE_ENFORCE_GT(tensor_dim.size(), 1); + MatDescriptor retv; + if (num_flatten_cols > 1) { + auto flatten_dim = framework::flatten_to_2d(tensor_dim, num_flatten_cols); + retv.height_ = flatten_dim[0]; + retv.width_ = flatten_dim[1]; + } else { + if (tensor_dim.size() == 2) { + retv.height_ = tensor_dim[0]; + retv.width_ = tensor_dim[1]; + } else { + auto dim_vec = framework::vectorize(tensor_dim); + retv.batch_size_ = 1; + for (size_t i = 0; i < dim_vec.size() - 2; ++i) { + retv.batch_size_ *= dim_vec[i]; + } + retv.height_ = dim_vec[dim_vec.size() - 2]; + retv.width_ = dim_vec[dim_vec.size() - 1]; + retv.stride_ = retv.height_ * retv.width_; + } + } + if (trans) { + std::swap(retv.width_, retv.height_); + } + retv.trans_ = trans; + return retv; +} +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/math/blas.h b/paddle/fluid/operators/math/blas.h new file mode 100644 index 0000000000000000000000000000000000000000..1a37cb39d56066b8380338b9710a441e41518c39 --- /dev/null +++ b/paddle/fluid/operators/math/blas.h @@ -0,0 +1,217 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include "paddle/fluid/framework/operator.h" +#include "paddle/fluid/framework/tensor.h" + +#ifdef PADDLE_WITH_MKLML +#include +#include +#include +#endif + +#ifdef PADDLE_USE_OPENBLAS +#include +#include +#endif + +#ifndef LAPACK_FOUND +extern "C" { +#include // NOLINT +int LAPACKE_sgetrf(int matrix_layout, int m, int n, float* a, int lda, + int* ipiv); +int LAPACKE_dgetrf(int matrix_layout, int m, int n, double* a, int lda, + int* ipiv); +int LAPACKE_sgetri(int matrix_layout, int n, float* a, int lda, + const int* ipiv); +int LAPACKE_dgetri(int matrix_layout, int n, double* a, int lda, + const int* ipiv); +} +#endif + +namespace paddle { +namespace operators { +namespace math { + +/** + * Matrix Descriptor of a memory buffer. + * + * It is used for Blas::MatMul. MatMul operator can be batched. + * if Mat A is [BatchSize, H, W], Mat B is [BatchSize, H, W]. It will be a + * `batch_size` times of GEMM. The batched GEMM could be faster base on the + * implementation of the blas library. The batch size could be zero. If any + * matrix of `matmul` has a batch size, the will be a batched GEMM, too. e.g., + * Mat A is [BatchSize, H1, W2], and Mat B [H2, W2], The result matrix wil be + * [BatchSize, H1, W2] + * + * The boolean flag, `trans`, describe the memory is the transpose of matrix or + * not. If the trans is true, the last two dims of matrix are transposed. The + * memory layout of the matrix is [Width, Height] or [BatchSize, Width, Height]. + * + * The MatDescriptor is not only the dimension or shape of a matrix, it also + * contains the layout, stride of matrix. It is clearer to have a structure than + * reuse `DDim`. + */ +struct MatDescriptor { + int64_t height_; + int64_t width_; + int64_t stride_{0}; + int64_t batch_size_{0}; + bool trans_; +}; + +/** + * Create Matrix Descriptor from a tensor dim, num_flatten_cols, and transpose + * flag + * + * @param tensor_dim: The dimension of the tensor. The rank of this dimension + * must larger than 1. + * + * @param num_flatten_cols: Reshape a tensor to a matrix. The matrix's first + * dimension(column length) will be the product of tensor's first `num_col_dims` + * dimensions. If num_flatten_cols is zero, the first N-2 dimension will be the + * batch_size of descriptor. + * + * @param trans: True if the matrix is transposed. + */ +extern MatDescriptor CreateMatrixDescriptor(const framework::DDim& tensor_dim, + int num_flatten_cols, bool trans); + +template +class Blas { + public: + explicit Blas(const DeviceContext& context) : context_(context) {} + + template + void GEMM(CBLAS_TRANSPOSE transA, CBLAS_TRANSPOSE transB, int M, int N, int K, + T alpha, const T* A, const T* B, T beta, T* C) const; + + template + void GEMM(bool transA, bool transB, int M, int N, int K, T alpha, const T* A, + int lda, const T* B, int ldb, T beta, T* C, int ldc) const; + + template + void MatMul(const framework::Tensor& mat_a, bool trans_a, + const framework::Tensor& mat_b, bool trans_b, T alpha, + framework::Tensor* mat_out, T beta) const; + + template + void MatMul(const framework::Tensor& mat_a, bool trans_a, + const framework::Tensor& mat_b, bool trans_b, + framework::Tensor* mat_out) const { + MatMul(mat_a, trans_a, mat_b, trans_b, static_cast(1.0), mat_out, + static_cast(0.0)); + } + + template + void MatMul(const framework::Tensor& mat_a, const framework::Tensor& mat_b, + framework::Tensor* mat_out) const { + this->template MatMul(mat_a, false, mat_b, false, mat_out); + } + + template + void AXPY(int n, T alpha, const T* x, T* y) const; + + template + void VADD(int n, const T* x, const T* y, T* z) const; + + template + void VCOPY(int n, const T* x, T* y) const; + + template + void GEMV(bool trans_a, int M, int N, T alpha, const T* A, const T* B, T beta, + T* C) const; + + template + void BatchedGEMM(CBLAS_TRANSPOSE transA, CBLAS_TRANSPOSE transB, int M, int N, + int K, T alpha, const T* A, const T* B, T beta, T* C, + int batchCount, int64_t strideA, int64_t strideB) const; + + template + void MatMul(const framework::Tensor& mat_a, const MatDescriptor& dim_a, + const framework::Tensor& mat_b, const MatDescriptor& dim_b, + T alpha, framework::Tensor* mat_out, T beta) const; + + private: + const DeviceContext& context_; +}; + +template +class BlasT : private Blas { + public: + using Blas::Blas; + + template + void GEMM(ARGS... args) const { + Base()->template GEMM(args...); + } + + template + void MatMul(ARGS... args) const { + Base()->template MatMul(args...); + } + + template + void AXPY(ARGS... args) const { + Base()->template AXPY(args...); + } + + template + void VADD(ARGS... args) const { + Base()->template VADD(args...); + } + + template + void VCOPY(ARGS... args) const { + Base()->template VCOPY(args...); + } + + template + void GEMV(ARGS... args) const { + Base()->template GEMV(args...); + } + + template + void BatchedGEMM(ARGS... args) const { + Base()->template BatchedGEMM(args...); + } + + private: + const Blas* Base() const { + return static_cast*>(this); + } +}; + +template +inline BlasT GetBlas( + const framework::ExecutionContext& exe_ctx) { + return BlasT( + exe_ctx.template device_context()); +} + +template +inline BlasT GetBlas(const DeviceContext& dev_ctx) { + return BlasT(dev_ctx); +} + +} // namespace math +} // namespace operators +} // namespace paddle + +#include "paddle/fluid/operators/math/blas_impl.h" +#ifdef PADDLE_WITH_CUDA +#include "paddle/fluid/operators/math/blas_impl.cu.h" +#endif diff --git a/paddle/fluid/operators/math/blas_impl.cu.h b/paddle/fluid/operators/math/blas_impl.cu.h new file mode 100644 index 0000000000000000000000000000000000000000..d84c88cb3bc1a13acb83b3444dbd1bfca3cba503 --- /dev/null +++ b/paddle/fluid/operators/math/blas_impl.cu.h @@ -0,0 +1,248 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include "paddle/fluid/operators/math/math_function.h" +#include "paddle/fluid/platform/dynload/cublas.h" + +namespace paddle { +namespace operators { +namespace math { + +template +struct CUBlas; + +template <> +struct CUBlas { + template + static void GEMM(ARGS... args) { + PADDLE_ENFORCE(platform::dynload::cublasSgemm(args...)); + } + + template + static void AXPY(ARGS... args) { + PADDLE_ENFORCE(platform::dynload::cublasSaxpy(args...)); + } + + template + static void GEMV(ARGS... args) { + PADDLE_ENFORCE(platform::dynload::cublasSgemv(args...)); + } + + template + static void GEMM_BATCH(ARGS... args) { +#if CUDA_VERSION >= 8000 + PADDLE_ENFORCE(platform::dynload::cublasSgemmStridedBatched(args...)); +#else + PADDLE_THROW("SgemmStridedBatched is not supported on cuda <= 7.5"); +#endif + } +}; + +template <> +struct CUBlas { + template + static void GEMM(ARGS... args) { + PADDLE_ENFORCE(platform::dynload::cublasDgemm(args...)); + } + + template + static void AXPY(ARGS... args) { + PADDLE_ENFORCE(platform::dynload::cublasDaxpy(args...)); + } + + template + static void GEMV(ARGS... args) { + PADDLE_ENFORCE(platform::dynload::cublasDgemv(args...)); + } + + template + static void GEMM_BATCH(ARGS... args) { +#if CUDA_VERSION >= 8000 + PADDLE_ENFORCE(platform::dynload::cublasDgemmStridedBatched(args...)); +#else + PADDLE_THROW("DgemmStridedBatched is not supported on cuda <= 7.5"); +#endif + } +}; + +template <> +struct CUBlas { + using float16 = platform::float16; + + static void GEMM(cublasHandle_t handle, cublasOperation_t transa, + cublasOperation_t transb, int m, int n, int k, + const float16 *alpha, const float16 *A, int lda, + const float16 *B, int ldb, const float16 *beta, float16 *C, + int ldc) { + PADDLE_ENFORCE( + platform::dynload::cublasHgemm(handle, transa, transb, m, n, k, + reinterpret_cast(alpha), + reinterpret_cast(A), lda, + reinterpret_cast(B), ldb, + reinterpret_cast(beta), + reinterpret_cast<__half *>(C), ldc)); + } + + static void GEMM_BATCH(cublasHandle_t handle, cublasOperation_t transa, + cublasOperation_t transb, int m, int n, int k, + const float16 *alpha, const float16 *A, int lda, + long long int strideA, const float16 *B, // NOLINT + int ldb, long long int strideB, // NOLINT + const float16 *beta, float16 *C, int ldc, + long long int strideC, // NOLINT + int batchCount) { +#if CUDA_VERSION >= 8000 + PADDLE_ENFORCE(platform::dynload::cublasHgemmStridedBatched( + handle, transa, transb, m, n, k, + reinterpret_cast(alpha), + reinterpret_cast(A), lda, strideA, + reinterpret_cast(B), ldb, strideB, + reinterpret_cast(beta), reinterpret_cast<__half *>(C), + ldc, strideC, batchCount)); +#else + PADDLE_THROW("HgemmStridedBatched is not supported on cuda <= 7.5"); +#endif + } +}; + +template <> +template +void Blas::GEMM(CBLAS_TRANSPOSE transA, + CBLAS_TRANSPOSE transB, int M, + int N, int K, T alpha, const T *A, + const T *B, T beta, T *C) const { + // Note that cublas follows fortran order, so the order is different from + // the cblas convention. + int lda = (transA == CblasNoTrans) ? K : M; + int ldb = (transB == CblasNoTrans) ? N : K; + cublasOperation_t cuTransA = + (transA == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T; + cublasOperation_t cuTransB = + (transB == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T; + + CUBlas::GEMM(context_.cublas_handle(), cuTransB, cuTransA, N, M, K, &alpha, + B, ldb, A, lda, &beta, C, N); +} + +template <> +template <> +inline void Blas::GEMM( + CBLAS_TRANSPOSE transA, CBLAS_TRANSPOSE transB, int M, int N, int K, + platform::float16 alpha, const platform::float16 *A, + const platform::float16 *B, platform::float16 beta, + platform::float16 *C) const { + // Note that cublas follows fortran order, so the order is different from + // the cblas convention. + int lda = (transA == CblasNoTrans) ? K : M; + int ldb = (transB == CblasNoTrans) ? N : K; + cublasOperation_t cuTransA = + (transA == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T; + cublasOperation_t cuTransB = + (transB == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T; + + // TODO(kexinzhao): add processing code for compute capability < 53 case + PADDLE_ENFORCE_GE(context_.GetComputeCapability(), 53, + "cublas fp16 gemm requires GPU compute capability >= 53"); + +#if CUDA_VERSION >= 8000 + float h_alpha = static_cast(alpha); + float h_beta = static_cast(beta); + + cublasGemmAlgo_t algo = CUBLAS_GEMM_DFALT; +#if CUDA_VERSION >= 9000 + if (context_.GetComputeCapability() >= 70) { + PADDLE_ENFORCE(platform::dynload::cublasSetMathMode( + context_.cublas_handle(), CUBLAS_TENSOR_OP_MATH)); + algo = CUBLAS_GEMM_DFALT_TENSOR_OP; + } else { + PADDLE_ENFORCE(platform::dynload::cublasSetMathMode( + context_.cublas_handle(), CUBLAS_DEFAULT_MATH)); + } +#endif // CUDA_VERSION >= 9000 + + // cublasHgemm does true FP16 computation which is slow for non-Volta + // GPUs. So use cublasGemmEx instead which does pesudo FP16 computation: + // input/output in fp16, computation in fp32, which can also be accelerated + // using tensor cores in volta GPUs. + PADDLE_ENFORCE(platform::dynload::cublasGemmEx( + context_.cublas_handle(), cuTransB, cuTransA, N, M, K, &h_alpha, B, + CUDA_R_16F, ldb, A, CUDA_R_16F, lda, &h_beta, C, CUDA_R_16F, N, + CUDA_R_32F, algo)); +#else + // CUDA 7.5 does not support cublasGemmEx, hence we fall back to use hgemm + CUBlas::GEMM(context_.cublas_handle(), cuTransB, cuTransA, + N, M, K, &h_alpha, h_B, ldb, h_A, lda, + &h_beta, h_C, N); +#endif // CUDA_VERSION >= 8000 +} + +template <> +template +void Blas::GEMM(bool transA, bool transB, int M, + int N, int K, T alpha, const T *A, + int lda, const T *B, int ldb, + T beta, T *C, int ldc) const { + // Note that cublas follows fortran order, so the order is different from + // the cblas convention. + cublasOperation_t cuTransA = transA ? CUBLAS_OP_T : CUBLAS_OP_N; + cublasOperation_t cuTransB = transB ? CUBLAS_OP_T : CUBLAS_OP_N; + CUBlas::GEMM(context_.cublas_handle(), cuTransB, cuTransA, N, M, K, &alpha, + B, ldb, A, lda, &beta, C, ldc); +} + +template <> +template +void Blas::AXPY(int n, T alpha, const T *x, + T *y) const { + CUBlas::AXPY(context_.cublas_handle(), n, &alpha, x, 1, y, 1); +} + +template <> +template +void Blas::GEMV(bool trans_a, int M, int N, + T alpha, const T *A, const T *B, + T beta, T *C) const { + cublasOperation_t cuTransA = !trans_a ? CUBLAS_OP_T : CUBLAS_OP_N; + + CUBlas::GEMV(context_.cublas_handle(), cuTransA, N, M, &alpha, A, N, B, 1, + &beta, C, 1); +} + +template <> +template +void Blas::BatchedGEMM( + CBLAS_TRANSPOSE transA, CBLAS_TRANSPOSE transB, int M, int N, int K, + T alpha, const T *A, const T *B, T beta, T *C, int batchCount, + int64_t strideA, int64_t strideB) const { + // Note that cublas follows fortran order, so the order is different from + // the cblas convention. + int lda = (transA == CblasNoTrans) ? K : M; + int ldb = (transB == CblasNoTrans) ? N : K; + int ldc = N; + cublasOperation_t cuTransA = + (transA == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T; + cublasOperation_t cuTransB = + (transB == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T; + const int64_t strideC = M * N; + + CUBlas::GEMM_BATCH(context_.cublas_handle(), cuTransB, cuTransA, N, M, K, + &alpha, B, ldb, strideB, A, lda, strideA, &beta, C, ldc, + strideC, batchCount); +} + +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/math/blas_impl.h b/paddle/fluid/operators/math/blas_impl.h new file mode 100644 index 0000000000000000000000000000000000000000..ae20406bc21d5e08359be8295cd98495dda7813b --- /dev/null +++ b/paddle/fluid/operators/math/blas_impl.h @@ -0,0 +1,252 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. +#pragma once +#include +#include "paddle/fluid/operators/math/math_function.h" + +namespace paddle { +namespace operators { +namespace math { + +template +struct CBlas; + +template <> +struct CBlas { + template + static void GEMM(ARGS... args) { + cblas_sgemm(args...); + } + + template + static void AXPY(ARGS... args) { + cblas_saxpy(args...); + } + +#ifdef PADDLE_WITH_MKLML + template + static void VADD(ARGS... args) { + vsAdd(args...); + } +#endif + + template + static void VCOPY(ARGS... args) { + cblas_scopy(args...); + } + + template + static void GEMV(ARGS... args) { + cblas_sgemv(args...); + } + +#ifdef PADDLE_WITH_MKLML + template + static void GEMM_BATCH(ARGS... args) { + cblas_sgemm_batch(args...); + } +#endif +}; + +template <> +struct CBlas { + template + static void GEMM(ARGS... args) { + cblas_dgemm(args...); + } + + template + static void AXPY(ARGS... args) { + cblas_daxpy(args...); + } + +#ifdef PADDLE_WITH_MKLML + template + static void VADD(ARGS... args) { + vdAdd(args...); + } +#endif + + template + static void VCOPY(ARGS... args) { + cblas_dcopy(args...); + } + + template + static void GEMV(ARGS... args) { + cblas_dgemv(args...); + } + +#ifdef PADDLE_WITH_MKLML + template + static void GEMM_BATCH(ARGS... args) { + cblas_dgemm_batch(args...); + } +#endif +}; + +template <> +struct CBlas { + static void GEMM(...) { PADDLE_THROW("float16 GEMM not supported on CPU"); } +#ifdef PADDLE_WITH_MKLML + static void GEMM_BATCH(...) { + PADDLE_THROW("float16 GEMM_BATCH not supported on CPU"); + } +#endif +}; + +template <> +template +void Blas::GEMM(CBLAS_TRANSPOSE transA, + CBLAS_TRANSPOSE transB, int M, + int N, int K, T alpha, const T *A, + const T *B, T beta, T *C) const { + int lda = (transA == CblasNoTrans) ? K : M; + int ldb = (transB == CblasNoTrans) ? N : K; + int ldc = N; + CBlas::GEMM(CblasRowMajor, transA, transB, M, N, K, alpha, A, lda, B, ldb, + beta, C, ldc); +} + +template <> +template +void Blas::GEMM(bool transA, bool transB, int M, + int N, int K, T alpha, const T *A, + int lda, const T *B, int ldb, + T beta, T *C, int ldc) const { + CBlas::GEMM(CblasRowMajor, transA == false ? CblasNoTrans : CblasTrans, + transB == false ? CblasNoTrans : CblasTrans, M, N, K, alpha, A, + lda, B, ldb, beta, C, ldc); +} + +template +template +void Blas::MatMul(const framework::Tensor &mat_a, bool trans_a, + const framework::Tensor &mat_b, bool trans_b, + T alpha, framework::Tensor *mat_out, + T beta) const { + auto dim_a = mat_a.dims(); + auto dim_b = mat_b.dims(); + auto dim_out = mat_out->dims(); + PADDLE_ENFORCE(dim_a.size() == 2 && dim_b.size() == 2 && dim_out.size() == 2, + "The input and output of matmul be matrix"); + PADDLE_ENFORCE( + mat_a.place() == mat_b.place() && mat_a.place() == mat_out->place(), + "The places of matrices must be same"); + + int M = dim_out[0]; + int N = dim_out[1]; + int K = !trans_a ? dim_a[1] : dim_a[0]; + + CBLAS_TRANSPOSE transA = !trans_a ? CblasNoTrans : CblasTrans; + CBLAS_TRANSPOSE transB = !trans_b ? CblasNoTrans : CblasTrans; + + this->GEMM(transA, transB, M, N, K, alpha, mat_a.data(), mat_b.data(), + beta, mat_out->data()); +} + +template <> +template +void Blas::AXPY(int n, T alpha, const T *x, + T *y) const { + CBlas::AXPY(n, alpha, x, 1, y, 1); +} + +template <> +template +void Blas::VCOPY(int n, const T *x, T *y) const { + CBlas::VCOPY(n, x, 1, y, 1); +} + +template <> +template +void Blas::VADD(int n, const T *x, const T *y, + T *z) const { +#ifdef PADDLE_WITH_MKLML + CBlas::VADD(n, x, y, z); +#else + this->template VCOPY(n, y, z); + this->template AXPY(n, 1., x, z); +#endif +} + +template <> +template +void Blas::GEMV(bool trans_a, int M, int N, T alpha, + const T *A, const T *B, T beta, + T *C) const { + CBLAS_TRANSPOSE transA = !trans_a ? CblasNoTrans : CblasTrans; + CBlas::GEMV(CblasRowMajor, transA, M, N, alpha, A, N, B, 1, beta, C, 1); +} + +template <> +template +void Blas::BatchedGEMM( + CBLAS_TRANSPOSE transA, CBLAS_TRANSPOSE transB, int M, int N, int K, + T alpha, const T *A, const T *B, T beta, T *C, int batchCount, + int64_t strideA, int64_t strideB) const { +#ifdef PADDLE_WITH_MKLML + int lda = (transA == CblasNoTrans) ? K : M; + int ldb = (transB == CblasNoTrans) ? N : K; + int ldc = N; + auto a_array = std::vector(batchCount); + auto b_array = std::vector(batchCount); + auto c_array = std::vector(batchCount); + for (int k = 0; k < batchCount; ++k) { + a_array[k] = &A[k * strideA]; + b_array[k] = &B[k * strideB]; + c_array[k] = &C[k * M * N]; + } + + CBlas::GEMM_BATCH(CblasRowMajor, &transA, &transB, &M, &N, &K, &alpha, + a_array.data(), &lda, b_array.data(), &ldb, &beta, + c_array.data(), &ldc, 1 /* group_count */, &batchCount); +#else + for (int k = 0; k < batchCount; ++k) { + auto *Ak = &A[k * strideA]; + auto *Bk = &B[k * strideB]; + auto *Ck = &C[k * M * N]; + this->template GEMM(transA, transB, M, N, K, alpha, Ak, Bk, beta, Ck); + } +#endif +} + +template +template +void Blas::MatMul(const framework::Tensor &mat_a, + const MatDescriptor &dim_a, + const framework::Tensor &mat_b, + const MatDescriptor &dim_b, T alpha, + framework::Tensor *mat_out, T beta) const { + PADDLE_ENFORCE_EQ(dim_a.width_, dim_b.height_); + CBLAS_TRANSPOSE transA = !dim_a.trans_ ? CblasNoTrans : CblasTrans; + CBLAS_TRANSPOSE transB = !dim_b.trans_ ? CblasNoTrans : CblasTrans; + if (dim_a.batch_size_ == 0 && dim_b.batch_size_ == 0) { + this->template GEMM(transA, transB, dim_a.height_, dim_b.width_, + dim_a.width_, alpha, mat_a.data(), + mat_b.data(), beta, mat_out->data()); + } else { + PADDLE_ENFORCE(dim_a.batch_size_ == dim_b.batch_size_ || + dim_a.batch_size_ == 0 || dim_b.batch_size_ == 0); + this->template BatchedGEMM( + transA, transB, dim_a.height_, dim_b.width_, dim_a.width_, alpha, + mat_a.data(), mat_b.data(), beta, mat_out->data(), + dim_a.batch_size_ == 0 ? dim_b.batch_size_ : dim_a.batch_size_, + dim_a.stride_, dim_b.stride_); + } +} + +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/math/concat.cc b/paddle/fluid/operators/math/concat.cc index bfce56f9fdcafa0800c9742b9fae41fd6a572b40..cc69212466b72f3fa82e8f5f58b4f3229dab28ec 100644 --- a/paddle/fluid/operators/math/concat.cc +++ b/paddle/fluid/operators/math/concat.cc @@ -13,6 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/operators/math/concat.h" +#include namespace paddle { namespace operators { @@ -70,20 +71,20 @@ class ConcatGradFunctor { public: void operator()(const platform::CPUDeviceContext& context, const framework::Tensor& input, const int axis, - std::vector& outputs) { + std::vector* outputs) { // TODO(zcd): Add input data validity checking - int num = outputs.size(); + int num = outputs->size(); int input_rows = 1; - auto dim_0 = outputs[0].dims(); + auto dim_0 = outputs->at(0).dims(); for (int i = 0; i < axis; ++i) { input_rows *= dim_0[i]; } int input_cols = 0; - std::vector output_cols(outputs.size()); + std::vector output_cols(outputs->size()); for (int i = 0; i < num; ++i) { - int t_cols = outputs[i].numel() / input_rows; + int t_cols = outputs->at(i).numel() / input_rows; input_cols += t_cols; output_cols[i] = t_cols; } @@ -95,7 +96,7 @@ class ConcatGradFunctor { int col_idx = 0; for (int j = 0; j < num; ++j) { int col_len = output_cols[j]; - T* dst_ptr = outputs[j].data() + k * col_len; + T* dst_ptr = outputs->at(j).data() + k * col_len; memory::Copy(cpu_place, dst_ptr, cpu_place, src_ptr + col_idx, sizeof(T) * col_len); col_idx += col_len; diff --git a/paddle/fluid/operators/math/concat.cu b/paddle/fluid/operators/math/concat.cu index c0786757b34195d47c3b1cadc938f0e9fcfd6038..4285d38dcd6a4124543cdd2246c82a8203f5a281 100644 --- a/paddle/fluid/operators/math/concat.cu +++ b/paddle/fluid/operators/math/concat.cu @@ -12,9 +12,11 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ +#include +#include #include "paddle/fluid/framework/mixed_vector.h" #include "paddle/fluid/operators/math/concat.h" -#include "paddle/fluid/platform/cuda_helper.h" +#include "paddle/fluid/platform/cuda_primitives.h" namespace paddle { namespace operators { @@ -202,16 +204,16 @@ class ConcatGradFunctor { public: void operator()(const platform::CUDADeviceContext& context, const framework::Tensor& input, const int axis, - std::vector& outputs) { + std::vector* outputs) { // TODO(zcd): Add input data validity checking - int o_num = outputs.size(); + int o_num = outputs->size(); int out_row = 1; - auto dim_0 = outputs[0].dims(); + auto dim_0 = outputs->at(0).dims(); for (int i = 0; i < axis; ++i) { out_row *= dim_0[i]; } - int out_col = outputs[0].numel() / out_row; + int out_col = outputs->at(0).numel() / out_row; int in_col = 0, in_row = out_row; bool sameShape = true; @@ -221,13 +223,13 @@ class ConcatGradFunctor { outputs_cols[0] = 0; for (int i = 0; i < o_num; ++i) { - int t_col = outputs[i].numel() / out_row; + int t_col = outputs->at(i).numel() / out_row; if (sameShape) { if (t_col != out_col) sameShape = false; } in_col += t_col; outputs_cols[i + 1] = in_col; - outputs_ptr[i] = outputs[i].data(); + outputs_ptr[i] = outputs->at(i).data(); } T** dev_out_gpu_data = diff --git a/paddle/fluid/operators/math/concat.h b/paddle/fluid/operators/math/concat.h index c0e983e4aa7abbdd87649f5a3147d2a464993bce..041ce8bf8a2e9528a004c076ead4471a3837c1a6 100644 --- a/paddle/fluid/operators/math/concat.h +++ b/paddle/fluid/operators/math/concat.h @@ -13,6 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once +#include #include "paddle/fluid/framework/data_type.h" #include "paddle/fluid/framework/tensor.h" @@ -56,7 +57,7 @@ template class ConcatGradFunctor { public: void operator()(const DeviceContext& context, const framework::Tensor& input, - const int axis, std::vector& outputs); + const int axis, std::vector* outputs); }; } // namespace math diff --git a/paddle/fluid/operators/math/concat_test.cc b/paddle/fluid/operators/math/concat_test.cc index 1741af8148bb90863f294ba4930006a58b5ddbf9..a46f2d51ca64501a622b5b48b424dffa16efc5b4 100644 --- a/paddle/fluid/operators/math/concat_test.cc +++ b/paddle/fluid/operators/math/concat_test.cc @@ -17,17 +17,14 @@ limitations under the License. */ #include #include "paddle/fluid/framework/tensor_util.h" -using namespace paddle::framework; -using namespace paddle::platform; - template void testConcat() { - Tensor input_a_cpu; - Tensor input_b_cpu; - Tensor out_cpu; - Tensor input_a; - Tensor input_b; - Tensor out; + paddle::framework::Tensor input_a_cpu; + paddle::framework::Tensor input_b_cpu; + paddle::framework::Tensor out_cpu; + paddle::framework::Tensor input_a; + paddle::framework::Tensor input_b; + paddle::framework::Tensor out; DeviceContext* context = new DeviceContext(Place()); // DeviceContext context(Place()); @@ -40,18 +37,18 @@ void testConcat() { * output: * out.shape: [5, 3, 4] */ - auto dim_a = make_ddim({2, 3, 4}); - auto dim_b = make_ddim({3, 3, 4}); - auto dim_out = make_ddim({5, 3, 4}); + auto dim_a = paddle::framework::make_ddim({2, 3, 4}); + auto dim_b = paddle::framework::make_ddim({3, 3, 4}); + auto dim_out = paddle::framework::make_ddim({5, 3, 4}); input_a.mutable_data(dim_a, Place()); input_b.mutable_data(dim_b, Place()); out.mutable_data(dim_out, Place()); if (paddle::platform::is_gpu_place(Place())) { - input_a_cpu.mutable_data(dim_a, CPUPlace()); - input_b_cpu.mutable_data(dim_b, CPUPlace()); - out_cpu.mutable_data(dim_out, CPUPlace()); + input_a_cpu.mutable_data(dim_a, paddle::platform::CPUPlace()); + input_b_cpu.mutable_data(dim_b, paddle::platform::CPUPlace()); + out_cpu.mutable_data(dim_out, paddle::platform::CPUPlace()); } int* a_ptr; @@ -72,11 +69,11 @@ void testConcat() { } if (paddle::platform::is_gpu_place(Place())) { - TensorCopy(input_a_cpu, Place(), *context, &input_a); - TensorCopy(input_b_cpu, Place(), *context, &input_b); + paddle::framework::TensorCopySync(input_a_cpu, Place(), &input_a); + paddle::framework::TensorCopySync(input_b_cpu, Place(), &input_b); } - std::vector input; + std::vector input; input.push_back(input_a); input.push_back(input_b); @@ -89,7 +86,8 @@ void testConcat() { int* out_ptr; if (paddle::platform::is_gpu_place(Place())) { - TensorCopy(out, CPUPlace(), *context, &out_cpu); + paddle::framework::TensorCopySync(out, paddle::platform::CPUPlace(), + &out_cpu); out_ptr = out_cpu.data(); } else { out_ptr = out.data(); @@ -115,9 +113,9 @@ void testConcat() { * output: * out.shape: [2, 7, 4] */ - dim_a = make_ddim({2, 3, 4}); - dim_b = make_ddim({2, 4, 4}); - dim_out = make_ddim({2, 7, 4}); + dim_a = paddle::framework::make_ddim({2, 3, 4}); + dim_b = paddle::framework::make_ddim({2, 4, 4}); + dim_out = paddle::framework::make_ddim({2, 7, 4}); input_a.Resize(dim_a); input_b.Resize(dim_b); @@ -144,8 +142,8 @@ void testConcat() { } if (paddle::platform::is_gpu_place(Place())) { - TensorCopy(input_a_cpu, Place(), *context, &input_a); - TensorCopy(input_b_cpu, Place(), *context, &input_b); + paddle::framework::TensorCopySync(input_a_cpu, Place(), &input_a); + paddle::framework::TensorCopySync(input_b_cpu, Place(), &input_b); } input.clear(); @@ -159,7 +157,8 @@ void testConcat() { PADDLE_ENFORCE_EQ(input_b.dims(), dim_b); if (paddle::platform::is_gpu_place(Place())) { - TensorCopy(out, CPUPlace(), *context, &out_cpu); + paddle::framework::TensorCopySync(out, paddle::platform::CPUPlace(), + &out_cpu); out_ptr = out_cpu.data(); } else { out_ptr = out.data(); @@ -187,9 +186,9 @@ void testConcat() { * output: * out.shape: [2, 3, 9] */ - dim_a = make_ddim({2, 3, 4}); - dim_b = make_ddim({2, 3, 5}); - dim_out = make_ddim({2, 3, 9}); + dim_a = paddle::framework::make_ddim({2, 3, 4}); + dim_b = paddle::framework::make_ddim({2, 3, 5}); + dim_out = paddle::framework::make_ddim({2, 3, 9}); input_a.Resize(dim_a); input_b.Resize(dim_b); @@ -216,8 +215,8 @@ void testConcat() { } if (paddle::platform::is_gpu_place(Place())) { - TensorCopy(input_a_cpu, Place(), *context, &input_a); - TensorCopy(input_b_cpu, Place(), *context, &input_b); + paddle::framework::TensorCopySync(input_a_cpu, Place(), &input_a); + paddle::framework::TensorCopySync(input_b_cpu, Place(), &input_b); } input.clear(); @@ -231,7 +230,8 @@ void testConcat() { PADDLE_ENFORCE_EQ(input_b.dims(), dim_b); if (paddle::platform::is_gpu_place(Place())) { - TensorCopy(out, CPUPlace(), *context, &out_cpu); + paddle::framework::TensorCopySync(out, paddle::platform::CPUPlace(), + &out_cpu); out_ptr = out_cpu.data(); } else { out_ptr = out.data(); @@ -261,9 +261,9 @@ void testConcat() { * output: * out.shape: [2, 6, 4] */ - dim_a = make_ddim({2, 3, 4}); - dim_b = make_ddim({2, 3, 4}); - dim_out = make_ddim({2, 6, 4}); + dim_a = paddle::framework::make_ddim({2, 3, 4}); + dim_b = paddle::framework::make_ddim({2, 3, 4}); + dim_out = paddle::framework::make_ddim({2, 6, 4}); input_a.Resize(dim_a); input_b.Resize(dim_b); @@ -290,8 +290,8 @@ void testConcat() { } if (paddle::platform::is_gpu_place(Place())) { - TensorCopy(input_a_cpu, Place(), *context, &input_a); - TensorCopy(input_b_cpu, Place(), *context, &input_b); + paddle::framework::TensorCopySync(input_a_cpu, Place(), &input_a); + paddle::framework::TensorCopySync(input_b_cpu, Place(), &input_b); } input.clear(); @@ -305,7 +305,8 @@ void testConcat() { PADDLE_ENFORCE_EQ(input_b.dims(), dim_b); if (paddle::platform::is_gpu_place(Place())) { - TensorCopy(out, CPUPlace(), *context, &out_cpu); + paddle::framework::TensorCopySync(out, paddle::platform::CPUPlace(), + &out_cpu); out_ptr = out_cpu.data(); } else { out_ptr = out.data(); diff --git a/paddle/fluid/operators/math/context_project.h b/paddle/fluid/operators/math/context_project.h index 4da94383af6c927a2db3337be6e82ca95a8cb036..bc0df3f3551c7a100d5d285cab585bb81c07fc5e 100644 --- a/paddle/fluid/operators/math/context_project.h +++ b/paddle/fluid/operators/math/context_project.h @@ -14,9 +14,11 @@ limitations under the License. */ #pragma once +#include +#include #include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/operators/math/blas.h" #include "paddle/fluid/operators/math/im2col.h" -#include "paddle/fluid/operators/math/math_function.h" namespace paddle { namespace operators { @@ -209,6 +211,7 @@ class ContextProjectGradFunctor { int input_row_begin, input_row_end; int sequence_height, sequence_width; sequence_width = in.dims()[1]; + auto blas = math::GetBlas(context); if (input_grad) { for (int i = 0; i < static_cast(lod_level_0.size()) - 1; ++i) { @@ -260,8 +263,8 @@ class ContextProjectGradFunctor { Tensor out_t_sub = out_t.Slice(k * context_length, k * context_length + padding_size); Tensor w_sub = padding_data->Slice(k, k + padding_size); - axpy(context, w_sub.numel(), static_cast(1), - out_t_sub.data(), w_sub.data()); + blas.AXPY(w_sub.numel(), static_cast(1), out_t_sub.data(), + w_sub.data()); } } if (down_pad > 0) { @@ -292,8 +295,8 @@ class ContextProjectGradFunctor { (down_pad_begin_row + t) * context_length); Tensor w_sub = padding_data->Slice( up_pad + padding_idx, up_pad + padding_idx + padding_size); - axpy(context, w_sub.numel(), static_cast(1), - out_t_sub.data(), w_sub.data()); + blas.AXPY(w_sub.numel(), static_cast(1), out_t_sub.data(), + w_sub.data()); } } out_t.Resize({sequence_height, context_length * sequence_width}); diff --git a/paddle/fluid/operators/math/cos_sim_functor.cu b/paddle/fluid/operators/math/cos_sim_functor.cu index 55c1e726335dfe010e39847ac90b84cc49955360..4e6ff5ee0a449b42762748ba1a103876beee01f2 100644 --- a/paddle/fluid/operators/math/cos_sim_functor.cu +++ b/paddle/fluid/operators/math/cos_sim_functor.cu @@ -13,7 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/operators/math/cos_sim_functor.h" -#include "paddle/fluid/platform/cuda_helper.h" +#include "paddle/fluid/platform/cuda_primitives.h" namespace paddle { namespace operators { diff --git a/paddle/fluid/operators/math/cross_entropy.cc b/paddle/fluid/operators/math/cross_entropy.cc index fc0fca5ad3370633b2f60db65fdb7c01c417dc50..caff35e03ae3a144f799d982c859ded62cb3e93d 100644 --- a/paddle/fluid/operators/math/cross_entropy.cc +++ b/paddle/fluid/operators/math/cross_entropy.cc @@ -46,7 +46,10 @@ class CrossEntropyFunctor { const int64_t* label_data = labels->data(); for (int i = 0; i < batch_size; ++i) { - int index = i * class_num + label_data[i]; + int lbl = label_data[i]; + PADDLE_ENFORCE_GE(lbl, 0); + PADDLE_ENFORCE_LT(lbl, class_num); + int index = i * class_num + lbl; loss_data[i] = -math::TolerableValue()(std::log(prob_data[index])); } } diff --git a/paddle/fluid/operators/math/cross_entropy.cu b/paddle/fluid/operators/math/cross_entropy.cu index f4935c2813c9f84699f1182df6a9adb613190506..0de58d5fddd84d33f708c4c73e5a19dc2fe8a86b 100644 --- a/paddle/fluid/operators/math/cross_entropy.cu +++ b/paddle/fluid/operators/math/cross_entropy.cu @@ -13,6 +13,8 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/operators/math/cross_entropy.h" +#include "paddle/fluid/platform/cuda_device_function.h" +#include "paddle/fluid/platform/cuda_primitives.h" namespace paddle { namespace operators { @@ -29,66 +31,22 @@ __global__ void CrossEntropyKernel(T* Y, const T* X, const int64_t* label, } } -template -__device__ __forceinline__ T sum_single_warp(T val) { - val += __shfl_down(val, 16); - val += __shfl_down(val, 8); - val += __shfl_down(val, 4); - val += __shfl_down(val, 2); - val += __shfl_down(val, 1); - return val; -} - -// CUDA do not support dynamic arrary in template -// https://stackoverflow.com/questions/20497209 -template -struct SharedMemory { - // Ensure that we won't compile any un-specialized types - __device__ T* GetPointer() { return NULL; } -}; - -template <> -struct SharedMemory { - __device__ float* GetPointer() { - extern __shared__ float s_float[]; - return s_float; - } -}; - -template <> -struct SharedMemory { - __device__ double* GetPointer() { - extern __shared__ double s_double[]; - return s_double; - } -}; - template __global__ void SoftCrossEntropyKernel(T* Y, const T* X, const T* label, const int class_num) { int tid = threadIdx.x; - SharedMemory d_sum_shared; - T* d_sum = d_sum_shared.GetPointer(); - d_sum[tid] = 0; + T val = 0; - int cur_idx = tid; - int next_idx = blockIdx.x * class_num + tid; - while (cur_idx < class_num) { - d_sum[tid] += - math::TolerableValue()(std::log(X[next_idx])) * label[next_idx]; - next_idx += blockDim.x; - cur_idx += blockDim.x; + int idx = blockIdx.x * class_num + tid; + int end = blockIdx.x * class_num + class_num; + for (; idx < end; idx += blockDim.x) { + val += math::TolerableValue()(std::log(X[idx])) * label[idx]; } - __syncthreads(); - for (unsigned int stride = blockDim.x >> 1; stride >= 32; stride >>= 1) { - if (tid < stride) d_sum[tid] += d_sum[tid + stride]; - __syncthreads(); + val = paddle::platform::reduceSum(val, tid, blockDim.x); + if (threadIdx.x == 0) { + Y[blockIdx.x] = -val; } - - T val = d_sum[tid]; - val = sum_single_warp(val); - if (tid == 0) Y[blockIdx.x] = -val; } } // namespace @@ -108,11 +66,11 @@ class CrossEntropyFunctor { if (softLabel) { const T* label_data = labels->data(); - int block = class_num > 512 ? 512 : pow(2, int(std::log2(class_num))); + int block = class_num > 512 + ? 512 + : pow(2, static_cast(std::log2(class_num))); - SoftCrossEntropyKernel<<< - batch_size, block, block * sizeof(T), - reinterpret_cast(ctx).stream()>>>( + SoftCrossEntropyKernel<<>>( loss_data, prob_data, label_data, class_num); } else { const int64_t* label_data = labels->data(); diff --git a/paddle/fluid/operators/math/depthwise_conv.cu b/paddle/fluid/operators/math/depthwise_conv.cu index d360728484a73ce844b4a36fbffd7dc631f8e786..027e2de48d229761f12f974dc73625c8ea1b3567 100644 --- a/paddle/fluid/operators/math/depthwise_conv.cu +++ b/paddle/fluid/operators/math/depthwise_conv.cu @@ -14,7 +14,7 @@ limitations under the License. */ #include #include "paddle/fluid/operators/math/depthwise_conv.h" -#include "paddle/fluid/platform/cuda_helper.h" +#include "paddle/fluid/platform/cuda_primitives.h" namespace paddle { namespace operators { diff --git a/paddle/fluid/operators/math/detail/gru_cpu_kernel.h b/paddle/fluid/operators/math/detail/gru_cpu_kernel.h index 1e5ff8ef46db960ddf88ebf03041893b176c1950..b6f4ab93777f2b5eb13a4a3172028d87c4546017 100644 --- a/paddle/fluid/operators/math/detail/gru_cpu_kernel.h +++ b/paddle/fluid/operators/math/detail/gru_cpu_kernel.h @@ -43,8 +43,8 @@ void hl_naive_gru_forward_reset_output(OpResetOutput op_reset_output, r_prev_out = prev_output_value[i]; } - op_reset_output(r_value_update_gate, r_value_reset_gate, r_prev_out, - r_value_reset_output, active_gate); + op_reset_output(&r_value_update_gate, &r_value_reset_gate, &r_prev_out, + &r_value_reset_output, active_gate); update_gate[i] = r_value_update_gate; reset_gate[i] = r_value_reset_gate; @@ -71,8 +71,8 @@ void hl_naive_gru_forward_final_output(OpFinalOutput op_final_output, r_prev_out = prev_output_value[i]; } - op_final_output(r_value_update_gate, r_value_frame_state, r_prev_out, - r_output, active_node); + op_final_output(&r_value_update_gate, &r_value_frame_state, &r_prev_out, + &r_output, active_node); frame_state[i] = r_value_frame_state; output_value[i] = r_output; @@ -89,22 +89,22 @@ void hl_avx_gru_forward_reset_output(OpResetOutput op_reset_output, __m256 r_value_reset_gate; __m256 r_value_reset_output; __m256 r_prev_out = _mm256_set1_ps(0.0f); - __m256 *update_gate = (__m256 *)gate_value; - __m256 *reset_gate = (__m256 *)(gate_value + frame_size); + __m256 *update_gate = reinterpret_cast<__m256 *>(gate_value); + __m256 *reset_gate = reinterpret_cast<__m256 *>(gate_value + frame_size); for (int i = 0; i < frame_size / 8; i++) { r_value_update_gate = update_gate[i]; r_value_reset_gate = reset_gate[i]; if (prev_output_value) { - r_prev_out = ((__m256 *)prev_output_value)[i]; + r_prev_out = (reinterpret_cast<__m256 *>(prev_output_value))[i]; } - op_reset_output(r_value_update_gate, r_value_reset_gate, r_prev_out, - r_value_reset_output, active_gate); + op_reset_output(&r_value_update_gate, &r_value_reset_gate, &r_prev_out, + &r_value_reset_output, active_gate); update_gate[i] = r_value_update_gate; reset_gate[i] = r_value_reset_gate; - ((__m256 *)reset_output_value)[i] = r_value_reset_output; + (reinterpret_cast<__m256 *>(reset_output_value))[i] = r_value_reset_output; } #endif } @@ -119,21 +119,21 @@ void hl_avx_gru_forward_final_output(OpFinalOutput op_final_output, __m256 r_value_frame_state; __m256 r_prev_out = _mm256_set1_ps(0.0f); __m256 r_output; - __m256 *update_gate = (__m256 *)gate_value; - __m256 *frame_state = (__m256 *)(gate_value + frame_size * 2); + __m256 *update_gate = reinterpret_cast<__m256 *>(gate_value); + __m256 *frame_state = reinterpret_cast<__m256 *>(gate_value + frame_size * 2); for (int i = 0; i < frame_size / 8; i++) { r_value_update_gate = update_gate[i]; r_value_frame_state = frame_state[i]; if (prev_output_value) { - r_prev_out = ((__m256 *)prev_output_value)[i]; + r_prev_out = (reinterpret_cast<__m256 *>(prev_output_value))[i]; } - op_final_output(r_value_update_gate, r_value_frame_state, r_prev_out, - r_output, active_node); + op_final_output(&r_value_update_gate, &r_value_frame_state, &r_prev_out, + &r_output, active_node); frame_state[i] = r_value_frame_state; - ((__m256 *)output_value)[i] = r_output; + (reinterpret_cast<__m256 *>(output_value))[i] = r_output; } #endif } @@ -213,9 +213,9 @@ void hl_naive_gru_backward_state_grad(OpStateGrad op_state_grad, T *gate_value, r_prev_out_grad = prev_out_grad[i]; } - op_state_grad(r_update_gate_value, r_update_gate_grad, r_frame_state_value, - r_frame_state_grad, r_prev_out_value, r_prev_out_grad, - r_out_grad, active_node); + op_state_grad(&r_update_gate_value, &r_update_gate_grad, + &r_frame_state_value, &r_frame_state_grad, &r_prev_out_value, + &r_prev_out_grad, &r_out_grad, active_node); update_gate_grad[i] = r_update_gate_grad; frame_state_grad[i] = r_frame_state_grad; @@ -258,9 +258,9 @@ void hl_naive_gru_backward_reset_grad(OpResetGrad op_reset_grad, T *gate_value, r_prev_out_grad = prev_out_grad[i]; } - op_reset_grad(r_update_gate_value, r_update_gate_grad, r_reset_gate_value, - r_reset_gate_grad, r_prev_out_value, r_prev_out_grad, - r_reset_output_grad, active_gate); + op_reset_grad(&r_update_gate_value, &r_update_gate_grad, + &r_reset_gate_value, &r_reset_gate_grad, &r_prev_out_value, + &r_prev_out_grad, &r_reset_output_grad, active_gate); update_gate_grad[i] = r_update_gate_grad; reset_gate_grad[i] = r_reset_gate_grad; @@ -284,30 +284,32 @@ void hl_avx_gru_backward_state_grad(OpStateGrad op_state_grad, T *gate_value, __m256 r_out_grad; __m256 r_prev_out_value = _mm256_set1_ps(0.0f); __m256 r_prev_out_grad = _mm256_set1_ps(0.0f); - __m256 *update_gate_value = (__m256 *)gate_value; - __m256 *update_gate_grad = (__m256 *)gate_grad; - __m256 *frame_state_value = (__m256 *)(gate_value + frame_size * 2); - __m256 *frame_state_grad = (__m256 *)(gate_grad + frame_size * 2); + __m256 *update_gate_value = reinterpret_cast<__m256 *>(gate_value); + __m256 *update_gate_grad = reinterpret_cast<__m256 *>(gate_grad); + __m256 *frame_state_value = + reinterpret_cast<__m256 *>(gate_value + frame_size * 2); + __m256 *frame_state_grad = + reinterpret_cast<__m256 *>(gate_grad + frame_size * 2); for (int i = 0; i < frame_size / 8; i++) { r_update_gate_value = update_gate_value[i]; r_frame_state_value = frame_state_value[i]; - r_out_grad = ((__m256 *)output_grad)[i]; + r_out_grad = (reinterpret_cast<__m256 *>(output_grad))[i]; if (prev_out_value) { - r_prev_out_value = ((__m256 *)prev_out_value)[i]; + r_prev_out_value = (reinterpret_cast<__m256 *>(prev_out_value))[i]; } if (prev_out_grad) { - r_prev_out_grad = ((__m256 *)prev_out_grad)[i]; + r_prev_out_grad = (reinterpret_cast<__m256 *>(prev_out_grad))[i]; } - op_state_grad(r_update_gate_value, r_update_gate_grad, r_frame_state_value, - r_frame_state_grad, r_prev_out_value, r_prev_out_grad, - r_out_grad, active_node); + op_state_grad(&r_update_gate_value, &r_update_gate_grad, + &r_frame_state_value, &r_frame_state_grad, &r_prev_out_value, + &r_prev_out_grad, &r_out_grad, active_node); update_gate_grad[i] = r_update_gate_grad; frame_state_grad[i] = r_frame_state_grad; if (prev_out_grad) { - ((__m256 *)prev_out_grad)[i] = r_prev_out_grad; + (reinterpret_cast<__m256 *>(prev_out_grad))[i] = r_prev_out_grad; } } #endif @@ -327,10 +329,11 @@ void hl_avx_gru_backward_reset_grad(OpResetGrad op_reset_grad, T *gate_value, __m256 r_reset_output_grad = _mm256_set1_ps(0.0f); __m256 r_prev_out_value = _mm256_set1_ps(0.0f); __m256 r_prev_out_grad = _mm256_set1_ps(0.0f); - __m256 *update_gate_value = (__m256 *)gate_value; - __m256 *update_gate_grad = (__m256 *)gate_grad; - __m256 *reset_gate_value = (__m256 *)(gate_value + frame_size); - __m256 *reset_gate_grad = (__m256 *)(gate_grad + frame_size); + __m256 *update_gate_value = reinterpret_cast<__m256 *>(gate_value); + __m256 *update_gate_grad = reinterpret_cast<__m256 *>(gate_grad); + __m256 *reset_gate_value = + reinterpret_cast<__m256 *>(gate_value + frame_size); + __m256 *reset_gate_grad = reinterpret_cast<__m256 *>(gate_grad + frame_size); for (int i = 0; i < frame_size / 8; i++) { r_update_gate_value = update_gate_value[i]; @@ -338,23 +341,23 @@ void hl_avx_gru_backward_reset_grad(OpResetGrad op_reset_grad, T *gate_value, r_reset_gate_value = reset_gate_value[i]; if (prev_out_value && prev_out_grad) { - r_reset_output_grad = ((__m256 *)reset_output_grad)[i]; + r_reset_output_grad = (reinterpret_cast<__m256 *>(reset_output_grad))[i]; } if (prev_out_value) { - r_prev_out_value = ((__m256 *)prev_out_value)[i]; + r_prev_out_value = (reinterpret_cast<__m256 *>(prev_out_value))[i]; } if (prev_out_grad) { - r_prev_out_grad = ((__m256 *)prev_out_grad)[i]; + r_prev_out_grad = (reinterpret_cast<__m256 *>(prev_out_grad))[i]; } - op_reset_grad(r_update_gate_value, r_update_gate_grad, r_reset_gate_value, - r_reset_gate_grad, r_prev_out_value, r_prev_out_grad, - r_reset_output_grad, active_gate); + op_reset_grad(&r_update_gate_value, &r_update_gate_grad, + &r_reset_gate_value, &r_reset_gate_grad, &r_prev_out_value, + &r_prev_out_grad, &r_reset_output_grad, active_gate); update_gate_grad[i] = r_update_gate_grad; reset_gate_grad[i] = r_reset_gate_grad; if (prev_out_grad) { - ((__m256 *)prev_out_grad)[i] = r_prev_out_grad; + (reinterpret_cast<__m256 *>(prev_out_grad))[i] = r_prev_out_grad; } } #endif diff --git a/paddle/fluid/operators/math/detail/gru_gpu_kernel.h b/paddle/fluid/operators/math/detail/gru_gpu_kernel.h index 657652562780ae9932a4394335bfa3c3b397bb80..813d69f6aba722609a0523a5be71d32f91f76d59 100644 --- a/paddle/fluid/operators/math/detail/gru_gpu_kernel.h +++ b/paddle/fluid/operators/math/detail/gru_gpu_kernel.h @@ -16,7 +16,7 @@ limitations under the License. */ #include #include "paddle/fluid/operators/math/detail/activation_functions.h" #include "paddle/fluid/operators/math/gru_compute.h" -#include "paddle/fluid/platform/cuda_helper.h" +#include "paddle/fluid/platform/cuda_primitives.h" #include "paddle/fluid/platform/device_context.h" namespace paddle { @@ -55,8 +55,8 @@ __global__ void KeGruForwardResetOutput(OpResetOutput op_reset_output, r_prev_out = prev_output_value[frame_idx]; } - op_reset_output(r_value_update_gate, r_value_reset_gate, r_prev_out, - r_value_reset_output, active_gate); + op_reset_output(&r_value_update_gate, &r_value_reset_gate, &r_prev_out, + &r_value_reset_output, active_gate); gate_value[frame_idx + frame_size * 0] = r_value_update_gate; gate_value[frame_idx + frame_size * 1] = r_value_reset_gate; @@ -93,8 +93,8 @@ __global__ void KeGruForwardFinalOutput(OpFinalOutput op_final_output, r_prev_out = prev_output_value[frame_idx]; } - op_final_output(r_value_update_gate, r_value_frame_state, r_prev_out, - r_output, active_node); + op_final_output(&r_value_update_gate, &r_value_frame_state, &r_prev_out, + &r_output, active_node); gate_value[frame_idx + frame_size * 2] = r_value_frame_state; output_value[frame_idx] = r_output; @@ -137,9 +137,9 @@ __global__ void KeGruBackwardStateGrad(OpStateGrad op_state_grad, T *gate_value, r_prev_out_grad = prev_out_grad[frame_idx]; } - op_state_grad(r_update_gate_value, r_update_gate_grad, r_frame_state_value, - r_frame_state_grad, r_prev_out_value, r_prev_out_grad, - r_out_grad, active_node); + op_state_grad(&r_update_gate_value, &r_update_gate_grad, &r_frame_state_value, + &r_frame_state_grad, &r_prev_out_value, &r_prev_out_grad, + &r_out_grad, active_node); gate_grad[frame_idx + frame_size * 0] = r_update_gate_grad; gate_grad[frame_idx + frame_size * 2] = r_frame_state_grad; @@ -185,9 +185,9 @@ __global__ void KeGruBackwardResetGrad(OpResetGrad op_reset_grad, T *gate_value, r_reset_output_grad = reset_output_grad[frame_idx]; } - op_reset_grad(r_update_gate_value, r_update_gate_grad, r_reset_gate_value, - r_reset_gate_grad, r_prev_out_value, r_prev_out_grad, - r_reset_output_grad, active_gate); + op_reset_grad(&r_update_gate_value, &r_update_gate_grad, &r_reset_gate_value, + &r_reset_gate_grad, &r_prev_out_value, &r_prev_out_grad, + &r_reset_output_grad, active_gate); gate_grad[frame_idx + frame_size * 0] = r_update_gate_grad; gate_grad[frame_idx + frame_size * 1] = r_reset_gate_grad; diff --git a/paddle/fluid/operators/math/detail/gru_kernel.h b/paddle/fluid/operators/math/detail/gru_kernel.h index 991f2e758c2c3b4acc93c1eb84dbd196d623b5cc..f6d192358bd84eb56a2e01eb36f28d8832ef271f 100644 --- a/paddle/fluid/operators/math/detail/gru_kernel.h +++ b/paddle/fluid/operators/math/detail/gru_kernel.h @@ -12,11 +12,11 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ +#pragma once +#include #include "paddle/fluid/operators/math/detail/activation_functions.h" #include "paddle/fluid/platform/hostdevice.h" -#include - // TODO(guosheng): refine code style in gru_kernel namespace paddle { namespace operators { @@ -28,25 +28,25 @@ namespace forward { template class gru_resetOutput { public: - HOSTDEVICE void operator()(T &value_update_gate, T &value_reset_gate, - T &prev_out, T &value_reset_output, + HOSTDEVICE void operator()(T *value_update_gate, T *value_reset_gate, + T *prev_out, T *value_reset_output, ActivationType act_gate) { - value_update_gate = activation(value_update_gate, act_gate); - value_reset_gate = activation(value_reset_gate, act_gate); - value_reset_output = prev_out * value_reset_gate; + *value_update_gate = activation(*value_update_gate, act_gate); + *value_reset_gate = activation(*value_reset_gate, act_gate); + *value_reset_output = (*prev_out) * (*value_reset_gate); } #ifndef __NVCC__ #ifndef __AVX__ static const bool avx = false; #else static const bool avx = true; - HOSTDEVICE void operator()(__m256 &value_update_gate, - __m256 &value_reset_gate, __m256 &prev_out, - __m256 &value_reset_output, + HOSTDEVICE void operator()(__m256 *value_update_gate, + __m256 *value_reset_gate, __m256 *prev_out, + __m256 *value_reset_output, ActivationType act_gate) { - value_update_gate = activation(value_update_gate, act_gate); - value_reset_gate = activation(value_reset_gate, act_gate); - value_reset_output = _mm256_mul_ps(prev_out, value_reset_gate); + *value_update_gate = activation(*value_update_gate, act_gate); + *value_reset_gate = activation(*value_reset_gate, act_gate); + *value_reset_output = _mm256_mul_ps(*prev_out, *value_reset_gate); } #endif #endif @@ -55,25 +55,25 @@ class gru_resetOutput { template class gru_finalOutput { public: - HOSTDEVICE void operator()(T &value_update_gate, T &value_frame_state, - T &prev_out, T &value_output, + HOSTDEVICE void operator()(T *value_update_gate, T *value_frame_state, + T *prev_out, T *value_output, ActivationType act_input) { - value_frame_state = activation(value_frame_state, act_input); - value_output = prev_out - (value_update_gate * prev_out) + - (value_update_gate * value_frame_state); + *value_frame_state = activation(*value_frame_state, act_input); + *value_output = *prev_out - ((*value_update_gate) * (*prev_out)) + + ((*value_update_gate) * (*value_frame_state)); } #ifndef __NVCC__ #ifndef __AVX__ static const bool avx = false; #else static const bool avx = true; - HOSTDEVICE void operator()(__m256 &value_update_gate, - __m256 &value_frame_state, __m256 &prev_out, - __m256 &value_output, ActivationType act_input) { - value_frame_state = activation(value_frame_state, act_input); - value_output = _mm256_add_ps( - _mm256_sub_ps(prev_out, _mm256_mul_ps(value_update_gate, prev_out)), - _mm256_mul_ps(value_update_gate, value_frame_state)); + HOSTDEVICE void operator()(__m256 *value_update_gate, + __m256 *value_frame_state, __m256 *prev_out, + __m256 *value_output, ActivationType act_input) { + *value_frame_state = activation(*value_frame_state, act_input); + *value_output = _mm256_add_ps( + _mm256_sub_ps(*prev_out, _mm256_mul_ps(*value_update_gate, *prev_out)), + _mm256_mul_ps(*value_update_gate, *value_frame_state)); } #endif #endif @@ -85,37 +85,38 @@ namespace backward { template class gru_stateGrad { public: - HOSTDEVICE void operator()(T &value_update_gate, T &grad_update_gate, - T &value_frame_state, T &grad_frame_state, - T &value_prev_out, T &grad_prev_out, - T &grad_output, ActivationType act_input) { - grad_update_gate = (grad_output * value_frame_state); - grad_update_gate -= (grad_output * value_prev_out); - grad_prev_out -= (grad_output * value_update_gate); - grad_prev_out += grad_output; - grad_frame_state = activation(grad_output * value_update_gate, - value_frame_state, act_input); + HOSTDEVICE void operator()(T *value_update_gate, T *grad_update_gate, + T *value_frame_state, T *grad_frame_state, + T *value_prev_out, T *grad_prev_out, + T *grad_output, ActivationType act_input) { + *grad_update_gate = (*grad_output * (*value_frame_state)); + *grad_update_gate -= (*grad_output * (*value_prev_out)); + *grad_prev_out -= (*grad_output * (*value_update_gate)); + *grad_prev_out += *grad_output; + *grad_frame_state = activation(*grad_output * (*value_update_gate), + *value_frame_state, act_input); } #ifndef __NVCC__ #ifndef __AVX__ static const bool avx = false; #else static const bool avx = true; - HOSTDEVICE void operator()(__m256 &value_update_gate, - __m256 &grad_update_gate, - __m256 &value_frame_state, - __m256 &grad_frame_state, __m256 &value_prev_out, - __m256 &grad_prev_out, __m256 &grad_output, + HOSTDEVICE void operator()(__m256 *value_update_gate, + __m256 *grad_update_gate, + __m256 *value_frame_state, + __m256 *grad_frame_state, __m256 *value_prev_out, + __m256 *grad_prev_out, __m256 *grad_output, ActivationType act_input) { - grad_update_gate = _mm256_mul_ps(grad_output, value_frame_state); - grad_update_gate = _mm256_sub_ps( - grad_update_gate, _mm256_mul_ps(grad_output, value_prev_out)); - grad_prev_out = _mm256_add_ps( - _mm256_sub_ps(grad_prev_out, - _mm256_mul_ps(grad_output, value_update_gate)), - grad_output); - grad_frame_state = activation(_mm256_mul_ps(grad_output, value_update_gate), - value_frame_state, act_input); + *grad_update_gate = _mm256_mul_ps(*grad_output, *value_frame_state); + *grad_update_gate = _mm256_sub_ps( + *grad_update_gate, _mm256_mul_ps(*grad_output, *value_prev_out)); + *grad_prev_out = _mm256_add_ps( + _mm256_sub_ps(*grad_prev_out, + _mm256_mul_ps(*grad_output, *value_update_gate)), + *grad_output); + *grad_frame_state = + activation(_mm256_mul_ps(*grad_output, *value_update_gate), + *value_frame_state, act_input); } #endif #endif @@ -124,32 +125,34 @@ class gru_stateGrad { template class gru_resetGrad { public: - HOSTDEVICE void operator()(T &value_update_gate, T &grad_update_gate, - T &value_reset_gate, T &grad_reset_gate, - T &value_prev_out, T &grad_prev_out, - T &grad_reset_output, ActivationType act_gate) { - grad_reset_gate = (grad_reset_output * value_prev_out); - grad_prev_out += (grad_reset_output * value_reset_gate); - grad_update_gate = - activation(grad_update_gate, value_update_gate, act_gate); - grad_reset_gate = activation(grad_reset_gate, value_reset_gate, act_gate); + HOSTDEVICE void operator()(T *value_update_gate, T *grad_update_gate, + T *value_reset_gate, T *grad_reset_gate, + T *value_prev_out, T *grad_prev_out, + T *grad_reset_output, ActivationType act_gate) { + *grad_reset_gate = (*grad_reset_output * (*value_prev_out)); + *grad_prev_out += (*grad_reset_output * (*value_reset_gate)); + *grad_update_gate = + activation(*grad_update_gate, *value_update_gate, act_gate); + *grad_reset_gate = + activation(*grad_reset_gate, *value_reset_gate, act_gate); } #ifndef __NVCC__ #ifndef __AVX__ static const bool avx = false; #else static const bool avx = true; - HOSTDEVICE void operator()(__m256 &value_update_gate, - __m256 &grad_update_gate, __m256 &value_reset_gate, - __m256 &grad_reset_gate, __m256 &value_prev_out, - __m256 &grad_prev_out, __m256 &grad_reset_output, + HOSTDEVICE void operator()(__m256 *value_update_gate, + __m256 *grad_update_gate, __m256 *value_reset_gate, + __m256 *grad_reset_gate, __m256 *value_prev_out, + __m256 *grad_prev_out, __m256 *grad_reset_output, ActivationType act_gate) { - grad_reset_gate = _mm256_mul_ps(grad_reset_output, value_prev_out); - grad_prev_out = _mm256_add_ps( - grad_prev_out, _mm256_mul_ps(grad_reset_output, value_reset_gate)); - grad_update_gate = - activation(grad_update_gate, value_update_gate, act_gate); - grad_reset_gate = activation(grad_reset_gate, value_reset_gate, act_gate); + *grad_reset_gate = _mm256_mul_ps(*grad_reset_output, *value_prev_out); + *grad_prev_out = _mm256_add_ps( + *grad_prev_out, _mm256_mul_ps(*grad_reset_output, *value_reset_gate)); + *grad_update_gate = + activation(*grad_update_gate, *value_update_gate, act_gate); + *grad_reset_gate = + activation(*grad_reset_gate, *value_reset_gate, act_gate); } #endif #endif diff --git a/paddle/fluid/operators/math/detail/lstm_cpu_kernel.h b/paddle/fluid/operators/math/detail/lstm_cpu_kernel.h index 6ad77830fd7a9809c4922878cc8ccdff1e8e0ef7..ccbd05c82ad6a880d21269092088be9656b35c99 100644 --- a/paddle/fluid/operators/math/detail/lstm_cpu_kernel.h +++ b/paddle/fluid/operators/math/detail/lstm_cpu_kernel.h @@ -59,9 +59,9 @@ void naive_lstm_forward_one_sequence(Op op, LstmMetaValue value, r_prev_state = value.prev_state_value[i]; } - op(r_value_in, r_value_ig, r_value_fg, r_value_og, r_prev_state, r_state, - r_state_atv, r_out, r_checkI, r_checkF, r_checkO, active_node, - active_gate, active_state); + op(&r_value_in, &r_value_ig, &r_value_fg, &r_value_og, &r_prev_state, + &r_state, &r_state_atv, &r_out, &r_checkI, &r_checkF, &r_checkO, + active_node, active_gate, active_state); value_in[i] = r_value_in; value_ig[i] = r_value_ig; @@ -125,11 +125,11 @@ void naive_lstm_backward_one_sequence(Op op, LstmMetaValue value, r_prev_state = value.prev_state_value[i]; } - op(r_value_in, r_value_ig, r_value_fg, r_value_og, r_grad_in, r_grad_ig, - r_grad_fg, r_grad_og, r_prev_state, r_prev_state_grad, r_state, - r_state_grad, r_state_atv, r_output_grad, r_checkI, r_checkF, r_checkO, - r_checkIGrad, r_checkFGrad, r_checkOGrad, active_node, active_gate, - active_state); + op(&r_value_in, &r_value_ig, &r_value_fg, &r_value_og, &r_grad_in, + &r_grad_ig, &r_grad_fg, &r_grad_og, &r_prev_state, &r_prev_state_grad, + &r_state, &r_state_grad, &r_state_atv, &r_output_grad, &r_checkI, + &r_checkF, &r_checkO, &r_checkIGrad, &r_checkFGrad, &r_checkOGrad, + active_node, active_gate, active_state); grad_in[i] = r_grad_in; grad_ig[i] = r_grad_ig; @@ -164,10 +164,12 @@ void avx_lstm_forward_one_sequence(Op op, LstmMetaValue value, __m256 r_state_atv; __m256 r_out; - __m256 *value_in = (__m256 *)value.gate_value; - __m256 *value_ig = (__m256 *)(value.gate_value + frame_size); - __m256 *value_fg = (__m256 *)(value.gate_value + frame_size * 2); - __m256 *value_og = (__m256 *)(value.gate_value + frame_size * 3); + __m256 *value_in = reinterpret_cast<__m256 *>(value.gate_value); + __m256 *value_ig = reinterpret_cast<__m256 *>(value.gate_value + frame_size); + __m256 *value_fg = + reinterpret_cast<__m256 *>(value.gate_value + frame_size * 2); + __m256 *value_og = + reinterpret_cast<__m256 *>(value.gate_value + frame_size * 3); for (int i = 0; i < frame_size / 8; i++) { r_value_in = value_in[i]; @@ -175,26 +177,26 @@ void avx_lstm_forward_one_sequence(Op op, LstmMetaValue value, r_value_fg = value_fg[i]; r_value_og = value_og[i]; if (value.check_ig) { - r_checkI = ((__m256 *)value.check_ig)[i]; - r_checkF = ((__m256 *)value.check_fg)[i]; - r_checkO = ((__m256 *)value.check_og)[i]; + r_checkI = (reinterpret_cast<__m256 *>(value.check_ig))[i]; + r_checkF = (reinterpret_cast<__m256 *>(value.check_fg))[i]; + r_checkO = (reinterpret_cast<__m256 *>(value.check_og))[i]; } if (value.prev_state_value) { - r_prev_state = ((__m256 *)value.prev_state_value)[i]; + r_prev_state = (reinterpret_cast<__m256 *>(value.prev_state_value))[i]; } - op(r_value_in, r_value_ig, r_value_fg, r_value_og, r_prev_state, r_state, - r_state_atv, r_out, r_checkI, r_checkF, r_checkO, active_node, - active_gate, active_state); + op(&r_value_in, &r_value_ig, &r_value_fg, &r_value_og, &r_prev_state, + &r_state, &r_state_atv, &r_out, &r_checkI, &r_checkF, &r_checkO, + active_node, active_gate, active_state); value_in[i] = r_value_in; value_ig[i] = r_value_ig; value_fg[i] = r_value_fg; value_og[i] = r_value_og; - ((__m256 *)value.state_value)[i] = r_state; - ((__m256 *)value.state_active_value)[i] = r_state_atv; - ((__m256 *)value.output_value)[i] = r_out; + (reinterpret_cast<__m256 *>(value.state_value))[i] = r_state; + (reinterpret_cast<__m256 *>(value.state_active_value))[i] = r_state_atv; + (reinterpret_cast<__m256 *>(value.output_value))[i] = r_out; } #endif } @@ -227,14 +229,16 @@ void avx_lstm_backward_one_sequence(Op op, LstmMetaValue value, __m256 r_checkFGrad; __m256 r_checkOGrad; - __m256 *value_in = (__m256 *)value.gate_value; - __m256 *value_ig = (__m256 *)(value.gate_value + frame_size); - __m256 *value_fg = (__m256 *)(value.gate_value + frame_size * 2); - __m256 *value_og = (__m256 *)(value.gate_value + frame_size * 3); - __m256 *grad_in = (__m256 *)grad.gate_grad; - __m256 *grad_ig = (__m256 *)(grad.gate_grad + frame_size); - __m256 *grad_fg = (__m256 *)(grad.gate_grad + frame_size * 2); - __m256 *grad_og = (__m256 *)(grad.gate_grad + frame_size * 3); + __m256 *value_in = reinterpret_cast<__m256 *>(value.gate_value); + __m256 *value_ig = reinterpret_cast<__m256 *>(value.gate_value + frame_size); + __m256 *value_fg = + reinterpret_cast<__m256 *>(value.gate_value + frame_size * 2); + __m256 *value_og = + reinterpret_cast<__m256 *>(value.gate_value + frame_size * 3); + __m256 *grad_in = reinterpret_cast<__m256 *>(grad.gate_grad); + __m256 *grad_ig = reinterpret_cast<__m256 *>(grad.gate_grad + frame_size); + __m256 *grad_fg = reinterpret_cast<__m256 *>(grad.gate_grad + frame_size * 2); + __m256 *grad_og = reinterpret_cast<__m256 *>(grad.gate_grad + frame_size * 3); for (int i = 0; i < frame_size / 8; i++) { r_value_in = value_in[i]; @@ -242,37 +246,40 @@ void avx_lstm_backward_one_sequence(Op op, LstmMetaValue value, r_value_fg = value_fg[i]; r_value_og = value_og[i]; if (value.check_ig) { - r_checkI = ((__m256 *)value.check_ig)[i]; - r_checkF = ((__m256 *)value.check_fg)[i]; - r_checkO = ((__m256 *)value.check_og)[i]; + r_checkI = (reinterpret_cast<__m256 *>(value.check_ig))[i]; + r_checkF = (reinterpret_cast<__m256 *>(value.check_fg))[i]; + r_checkO = (reinterpret_cast<__m256 *>(value.check_og))[i]; } - r_state = ((__m256 *)value.state_value)[i]; - r_state_atv = ((__m256 *)value.state_active_value)[i]; - r_output_grad = ((__m256 *)grad.output_grad)[i]; - r_state_grad = ((__m256 *)grad.state_grad)[i]; + r_state = (reinterpret_cast<__m256 *>(value.state_value))[i]; + r_state_atv = (reinterpret_cast<__m256 *>(value.state_active_value))[i]; + r_output_grad = (reinterpret_cast<__m256 *>(grad.output_grad))[i]; + r_state_grad = (reinterpret_cast<__m256 *>(grad.state_grad))[i]; if (value.prev_state_value) { - r_prev_state = ((__m256 *)value.prev_state_value)[i]; + r_prev_state = (reinterpret_cast<__m256 *>(value.prev_state_value))[i]; } - op(r_value_in, r_value_ig, r_value_fg, r_value_og, r_grad_in, r_grad_ig, - r_grad_fg, r_grad_og, r_prev_state, r_prev_state_grad, r_state, - r_state_grad, r_state_atv, r_output_grad, r_checkI, r_checkF, r_checkO, - r_checkIGrad, r_checkFGrad, r_checkOGrad, active_node, active_gate, - active_state); + op(&r_value_in, &r_value_ig, &r_value_fg, &r_value_og, &r_grad_in, + &r_grad_ig, &r_grad_fg, &r_grad_og, &r_prev_state, &r_prev_state_grad, + &r_state, &r_state_grad, &r_state_atv, &r_output_grad, &r_checkI, + &r_checkF, &r_checkO, &r_checkIGrad, &r_checkFGrad, &r_checkOGrad, + active_node, active_gate, active_state); grad_in[i] = r_grad_in; grad_ig[i] = r_grad_ig; grad_fg[i] = r_grad_fg; grad_og[i] = r_grad_og; - ((__m256 *)grad.state_grad)[i] = r_state_grad; + (reinterpret_cast<__m256 *>(grad.state_grad))[i] = r_state_grad; if (grad.prev_state_grad) - ((__m256 *)grad.prev_state_grad)[i] = r_prev_state_grad; + (reinterpret_cast<__m256 *>(grad.prev_state_grad))[i] = r_prev_state_grad; if (value.prev_state_value) { - if (grad.check_ig_grad) ((__m256 *)grad.check_ig_grad)[i] += r_checkIGrad; - if (grad.check_fg_grad) ((__m256 *)grad.check_fg_grad)[i] += r_checkFGrad; + if (grad.check_ig_grad) + (reinterpret_cast<__m256 *>(grad.check_ig_grad))[i] += r_checkIGrad; + if (grad.check_fg_grad) + (reinterpret_cast<__m256 *>(grad.check_fg_grad))[i] += r_checkFGrad; } - if (grad.check_og_grad) ((__m256 *)grad.check_og_grad)[i] += r_checkOGrad; + if (grad.check_og_grad) + (reinterpret_cast<__m256 *>(grad.check_og_grad))[i] += r_checkOGrad; } #endif } diff --git a/paddle/fluid/operators/math/detail/lstm_gpu_kernel.h b/paddle/fluid/operators/math/detail/lstm_gpu_kernel.h index ee7b16da4187e0f7f7839eff5c8753d2eb4f9c6d..2aecb69237fdf344ebc0bfe72d9c7c147f06358d 100644 --- a/paddle/fluid/operators/math/detail/lstm_gpu_kernel.h +++ b/paddle/fluid/operators/math/detail/lstm_gpu_kernel.h @@ -13,13 +13,13 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once +#include + #include "paddle/fluid/operators/math/detail/activation_functions.h" #include "paddle/fluid/operators/math/lstm_compute.h" -#include "paddle/fluid/platform/cuda_helper.h" +#include "paddle/fluid/platform/cuda_primitives.h" #include "paddle/fluid/platform/device_context.h" -#include - namespace paddle { namespace operators { namespace math { @@ -70,9 +70,9 @@ __global__ void KeLstmForward(Op op, LstmMetaValue value, int frame_size, r_prev_state = value.prev_state_value[frame_idx]; } - op(r_value_in, r_value_ig, r_value_fg, r_value_og, r_prev_state, r_state, - r_state_atv, r_out, r_checkI, r_checkF, r_checkO, active_node, active_gate, - active_state); + op(&r_value_in, &r_value_ig, &r_value_fg, &r_value_og, &r_prev_state, + &r_state, &r_state_atv, &r_out, &r_checkI, &r_checkF, &r_checkO, + active_node, active_gate, active_state); value.gate_value[frame_idx] = r_value_in; value.gate_value[frame_idx + frame_size] = r_value_ig; @@ -145,11 +145,11 @@ __global__ void KeLstmBackward(Op op, LstmMetaValue value, r_prev_state = value.prev_state_value[frame_idx]; } - op(r_value_in, r_value_ig, r_value_fg, r_value_og, r_grad_in, r_grad_ig, - r_grad_fg, r_grad_og, r_prev_state, r_prev_state_grad, r_state, - r_state_grad, r_state_atv, r_output_grad, r_checkI, r_checkF, r_checkO, - r_checkIGrad, r_checkFGrad, r_checkOGrad, active_node, active_gate, - active_state); + op(&r_value_in, &r_value_ig, &r_value_fg, &r_value_og, &r_grad_in, &r_grad_ig, + &r_grad_fg, &r_grad_og, &r_prev_state, &r_prev_state_grad, &r_state, + &r_state_grad, &r_state_atv, &r_output_grad, &r_checkI, &r_checkF, + &r_checkO, &r_checkIGrad, &r_checkFGrad, &r_checkOGrad, active_node, + active_gate, active_state); grad.gate_grad[frame_idx] = r_grad_in; grad.gate_grad[frame_idx + frame_size] = r_grad_ig; diff --git a/paddle/fluid/operators/math/detail/lstm_kernel.h b/paddle/fluid/operators/math/detail/lstm_kernel.h index 9080634f2b3fc122a420e049314f53abd50376e0..cbe73d62938d7c4c03a2c8731665260624417fd7 100644 --- a/paddle/fluid/operators/math/detail/lstm_kernel.h +++ b/paddle/fluid/operators/math/detail/lstm_kernel.h @@ -12,11 +12,11 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ +#pragma once +#include #include "paddle/fluid/operators/math/detail/activation_functions.h" #include "paddle/fluid/platform/hostdevice.h" -#include - namespace paddle { namespace operators { namespace math { @@ -27,19 +27,19 @@ namespace forward { template class lstm { public: - HOSTDEVICE void operator()(T &value_in, T &value_ig, T &value_fg, T &value_og, - T &prev_state, T &state, T &state_atv, T &output, - T &checkI, T &checkF, T &checkO, + HOSTDEVICE void operator()(T *value_in, T *value_ig, T *value_fg, T *value_og, + T *prev_state, T *state, T *state_atv, T *output, + T *checkI, T *checkF, T *checkO, ActivationType active_node, ActivationType active_gate, ActivationType active_state) { - value_in = activation(value_in, active_node); - value_ig = activation(value_ig + prev_state * checkI, active_gate); - value_fg = activation(value_fg + prev_state * checkF, active_gate); - state = value_in * value_ig + prev_state * value_fg; - value_og = activation(value_og + state * checkO, active_gate); - state_atv = activation(state, active_state); - output = value_og * state_atv; + *value_in = activation(*value_in, active_node); + *value_ig = activation(*value_ig + (*prev_state) * (*checkI), active_gate); + *value_fg = activation(*value_fg + (*prev_state) * (*checkF), active_gate); + *state = (*value_in) * (*value_ig) + (*prev_state) * (*value_fg); + *value_og = activation(*value_og + (*state) * (*checkO), active_gate); + *state_atv = activation(*state, active_state); + *output = (*value_og) * (*state_atv); } #ifndef __NVCC__ #ifndef __AVX__ // If not compiled with AVX instructs. Disable AVX by default @@ -48,27 +48,27 @@ class lstm { // Only float support AVX optimization static const bool avx = std::is_same::value; - HOSTDEVICE void operator()(__m256 &value_in, __m256 &value_ig, - __m256 &value_fg, __m256 &value_og, - __m256 &prev_state, __m256 &state, - __m256 &state_atv, __m256 &output, __m256 &checkI, - __m256 &checkF, __m256 &checkO, + HOSTDEVICE void operator()(__m256 *value_in, __m256 *value_ig, + __m256 *value_fg, __m256 *value_og, + __m256 *prev_state, __m256 *state, + __m256 *state_atv, __m256 *output, __m256 *checkI, + __m256 *checkF, __m256 *checkO, ActivationType active_node, ActivationType active_gate, ActivationType active_state) { - value_in = activation(value_in, active_node); - value_ig = - activation(_mm256_add_ps(value_ig, _mm256_mul_ps(prev_state, checkI)), - active_gate); - value_fg = - activation(_mm256_add_ps(value_fg, _mm256_mul_ps(prev_state, checkF)), - active_gate); - state = _mm256_add_ps(_mm256_mul_ps(value_in, value_ig), - _mm256_mul_ps(prev_state, value_fg)); - value_og = activation(_mm256_add_ps(value_og, _mm256_mul_ps(state, checkO)), - active_gate); - state_atv = activation(state, active_state); - output = _mm256_mul_ps(value_og, state_atv); + *value_in = activation(*value_in, active_node); + *value_ig = activation( + _mm256_add_ps(*value_ig, _mm256_mul_ps(*prev_state, *checkI)), + active_gate); + *value_fg = activation( + _mm256_add_ps(*value_fg, _mm256_mul_ps(*prev_state, *checkF)), + active_gate); + *state = _mm256_add_ps(_mm256_mul_ps(*value_in, *value_ig), + _mm256_mul_ps(*prev_state, *value_fg)); + *value_og = activation( + _mm256_add_ps(*value_og, _mm256_mul_ps(*state, *checkO)), active_gate); + *state_atv = activation(*state, active_state); + *output = _mm256_mul_ps(*value_og, *state_atv); } #endif #endif @@ -81,26 +81,29 @@ namespace backward { template class lstm { public: - HOSTDEVICE void operator()(T &value_in, T &value_ig, T &value_fg, T &value_og, - T &grad_in, T &grad_ig, T &grad_fg, T &grad_og, - T &prev_state, T &prev_state_grad, T &state, - T &state_grad, T &state_atv, T &output_grad, - T &checkI, T &checkF, T &checkO, T &checkIGrad, - T &checkFGrad, T &checkOGrad, + HOSTDEVICE void operator()(T *value_in, T *value_ig, T *value_fg, T *value_og, + T *grad_in, T *grad_ig, T *grad_fg, T *grad_og, + T *prev_state, T *prev_state_grad, T *state, + T *state_grad, T *state_atv, T *output_grad, + T *checkI, T *checkF, T *checkO, T *checkIGrad, + T *checkFGrad, T *checkOGrad, ActivationType active_node, ActivationType active_gate, ActivationType active_state) { - grad_og = activation(output_grad * state_atv, value_og, active_gate); - state_grad += activation(output_grad * value_og, state_atv, active_state) + - grad_og * checkO; - grad_in = activation(state_grad * value_ig, value_in, active_node); - grad_ig = activation(state_grad * value_in, value_ig, active_gate); - grad_fg = activation(state_grad * prev_state, value_fg, active_gate); - prev_state_grad = - grad_ig * checkI + grad_fg * checkF + state_grad * value_fg; - checkIGrad = grad_ig * prev_state; - checkFGrad = grad_fg * prev_state; - checkOGrad = grad_og * state; + *grad_og = + activation((*output_grad) * (*state_atv), *value_og, active_gate); + *state_grad += + activation((*output_grad) * (*value_og), *state_atv, active_state) + + (*grad_og) * (*checkO); + *grad_in = activation((*state_grad) * (*value_ig), *value_in, active_node); + *grad_ig = activation((*state_grad) * (*value_in), *value_ig, active_gate); + *grad_fg = + activation((*state_grad) * (*prev_state), *value_fg, active_gate); + *prev_state_grad = (*grad_ig) * (*checkI) + (*grad_fg) * (*checkF) + + (*state_grad) * (*value_fg); + *checkIGrad = (*grad_ig) * (*prev_state); + *checkFGrad = (*grad_fg) * (*prev_state); + *checkOGrad = (*grad_og) * (*state); } #ifndef __NVCC__ #ifndef __AVX__ // If not compiled with AVX instructs. Disable AVX by default @@ -109,32 +112,33 @@ class lstm { // Only float support AVX optimization static const bool avx = std::is_same::value; HOSTDEVICE void operator()( - __m256 &value_in, __m256 &value_ig, __m256 &value_fg, __m256 &value_og, - __m256 &grad_in, __m256 &grad_ig, __m256 &grad_fg, __m256 &grad_og, - __m256 &prev_state, __m256 &prev_state_grad, __m256 &state, - __m256 &state_grad, __m256 &state_atv, __m256 &output_grad, - __m256 &checkI, __m256 &checkF, __m256 &checkO, __m256 &checkIGrad, - __m256 &checkFGrad, __m256 &checkOGrad, ActivationType active_node, + __m256 *value_in, __m256 *value_ig, __m256 *value_fg, __m256 *value_og, + __m256 *grad_in, __m256 *grad_ig, __m256 *grad_fg, __m256 *grad_og, + __m256 *prev_state, __m256 *prev_state_grad, __m256 *state, + __m256 *state_grad, __m256 *state_atv, __m256 *output_grad, + __m256 *checkI, __m256 *checkF, __m256 *checkO, __m256 *checkIGrad, + __m256 *checkFGrad, __m256 *checkOGrad, ActivationType active_node, ActivationType active_gate, ActivationType active_state) { - grad_og = activation(_mm256_mul_ps(output_grad, state_atv), value_og, - active_gate); - state_grad = _mm256_add_ps(activation(_mm256_mul_ps(output_grad, value_og), - state_atv, active_state), - state_grad); - state_grad = _mm256_add_ps(_mm256_mul_ps(grad_og, checkO), state_grad); - grad_in = - activation(_mm256_mul_ps(state_grad, value_ig), value_in, active_node); - grad_ig = - activation(_mm256_mul_ps(state_grad, value_in), value_ig, active_gate); - grad_fg = activation(_mm256_mul_ps(state_grad, prev_state), value_fg, - active_gate); - prev_state_grad = _mm256_add_ps(_mm256_mul_ps(grad_ig, checkI), - _mm256_mul_ps(grad_fg, checkF)); - prev_state_grad = - _mm256_add_ps(_mm256_mul_ps(state_grad, value_fg), prev_state_grad); - checkIGrad = _mm256_mul_ps(grad_ig, prev_state); - checkFGrad = _mm256_mul_ps(grad_fg, prev_state); - checkOGrad = _mm256_mul_ps(grad_og, state); + *grad_og = activation(_mm256_mul_ps(*output_grad, *state_atv), *value_og, + active_gate); + *state_grad = + _mm256_add_ps(activation(_mm256_mul_ps(*output_grad, *value_og), + *state_atv, active_state), + *state_grad); + *state_grad = _mm256_add_ps(_mm256_mul_ps(*grad_og, *checkO), *state_grad); + *grad_in = activation(_mm256_mul_ps(*state_grad, *value_ig), *value_in, + active_node); + *grad_ig = activation(_mm256_mul_ps(*state_grad, *value_in), *value_ig, + active_gate); + *grad_fg = activation(_mm256_mul_ps(*state_grad, *prev_state), *value_fg, + active_gate); + *prev_state_grad = _mm256_add_ps(_mm256_mul_ps(*grad_ig, *checkI), + _mm256_mul_ps(*grad_fg, *checkF)); + *prev_state_grad = + _mm256_add_ps(_mm256_mul_ps(*state_grad, *value_fg), *prev_state_grad); + *checkIGrad = _mm256_mul_ps(*grad_ig, *prev_state); + *checkFGrad = _mm256_mul_ps(*grad_fg, *prev_state); + *checkOGrad = _mm256_mul_ps(*grad_og, *state); } #endif #endif diff --git a/paddle/fluid/operators/math/gru_compute.cc b/paddle/fluid/operators/math/gru_compute.cc index 3f044b775138c495052bec3d19121bf26c37cb37..0e15b81deef43a932d4b2d3f545393b0ad9e080c 100644 --- a/paddle/fluid/operators/math/gru_compute.cc +++ b/paddle/fluid/operators/math/gru_compute.cc @@ -10,9 +10,9 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/operators/math/gru_compute.h" +#include "paddle/fluid/operators/math/blas.h" #include "paddle/fluid/operators/math/detail/gru_cpu_kernel.h" #include "paddle/fluid/operators/math/detail/gru_kernel.h" -#include "paddle/fluid/operators/math/math_function.h" namespace paddle { namespace operators { @@ -25,21 +25,21 @@ struct GRUUnitFunctor { const detail::ActivationType active_node, const detail::ActivationType active_gate) { #ifndef __NVCC__ + auto blas = math::GetBlas(context); if (value.prev_out_value) { - math::gemm( - context, false, false, batch_size, frame_size * 2, frame_size, 1, - value.prev_out_value, frame_size, value.gate_weight, frame_size * 2, - 1, value.gate_value, frame_size * 3); + blas.GEMM(false, false, batch_size, frame_size * 2, frame_size, 1, + value.prev_out_value, frame_size, value.gate_weight, + frame_size * 2, 1, value.gate_value, frame_size * 3); } detail::forward_reset_output(detail::forward::gru_resetOutput(), value, frame_size, batch_size, active_gate); if (value.prev_out_value) { - math::gemm( - context, false, false, batch_size, frame_size, frame_size, 1, - value.reset_output_value, frame_size, value.state_weight, frame_size, - 1, value.gate_value + frame_size * 2, frame_size * 3); + blas.GEMM(false, false, batch_size, frame_size, frame_size, 1, + value.reset_output_value, frame_size, value.state_weight, + frame_size, 1, value.gate_value + frame_size * 2, + frame_size * 3); } detail::forward_final_output(detail::forward::gru_finalOutput(), value, @@ -58,36 +58,32 @@ struct GRUUnitGradFunctor { #ifndef __NVCC__ detail::backward_state_grad(detail::backward::gru_stateGrad(), value, grad, frame_size, batch_size, active_node); - + auto blas = math::GetBlas(context); if (value.prev_out_value && grad.prev_out_grad) { - math::gemm( - context, false, true, batch_size, frame_size, frame_size, 1, - grad.gate_grad + frame_size * 2, frame_size * 3, value.state_weight, - frame_size, 0, grad.reset_output_grad, frame_size); + blas.GEMM(false, true, batch_size, frame_size, frame_size, 1, + grad.gate_grad + frame_size * 2, frame_size * 3, + value.state_weight, frame_size, 0, grad.reset_output_grad, + frame_size); if (grad.state_weight_grad) { - math::gemm( - context, true, false, frame_size, frame_size, batch_size, 1, - value.reset_output_value, frame_size, - grad.gate_grad + frame_size * 2, frame_size * 3, 1, - grad.state_weight_grad, frame_size); + blas.GEMM(true, false, frame_size, frame_size, batch_size, 1, + value.reset_output_value, frame_size, + grad.gate_grad + frame_size * 2, frame_size * 3, 1, + grad.state_weight_grad, frame_size); } } detail::backward_reset_grad(detail::backward::gru_resetGrad(), value, grad, frame_size, batch_size, active_gate); - if (grad.prev_out_grad && value.prev_out_value) { - math::gemm( - context, false, true, batch_size, frame_size, frame_size * 2, 1, - grad.gate_grad, frame_size * 3, value.gate_weight, frame_size * 2, 1, - grad.prev_out_grad, frame_size); + blas.GEMM(false, true, batch_size, frame_size, frame_size * 2, 1, + grad.gate_grad, frame_size * 3, value.gate_weight, + frame_size * 2, 1, grad.prev_out_grad, frame_size); if (grad.gate_weight_grad) { - math::gemm( - context, true, false, frame_size, frame_size * 2, batch_size, 1, - value.prev_out_value, frame_size, grad.gate_grad, frame_size * 3, 1, - grad.gate_weight_grad, frame_size * 2); + blas.GEMM(true, false, frame_size, frame_size * 2, batch_size, 1, + value.prev_out_value, frame_size, grad.gate_grad, + frame_size * 3, 1, grad.gate_weight_grad, frame_size * 2); } } #endif diff --git a/paddle/fluid/operators/math/gru_compute.cu b/paddle/fluid/operators/math/gru_compute.cu index 27caf3383dd6cd94779391b722bba1d6b74772c0..1327d914952d57aab6e5d17090d0ea976a6d4755 100644 --- a/paddle/fluid/operators/math/gru_compute.cu +++ b/paddle/fluid/operators/math/gru_compute.cu @@ -9,10 +9,11 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ +#include +#include "paddle/fluid/operators/math/blas.h" #include "paddle/fluid/operators/math/detail/gru_gpu_kernel.h" #include "paddle/fluid/operators/math/detail/gru_kernel.h" #include "paddle/fluid/operators/math/gru_compute.h" -#include "paddle/fluid/operators/math/math_function.h" namespace paddle { namespace operators { @@ -36,12 +37,11 @@ struct GRUUnitFunctor { threads = dim3(32, 32); grid = dim3((frame_size + 32 - 1) / 32, (batch_size + 32 - 1) / 32); } - + auto blas = math::GetBlas(context); if (value.prev_out_value) { - math::gemm( - context, false, false, batch_size, frame_size * 2, frame_size, 1, - value.prev_out_value, frame_size, value.gate_weight, frame_size * 2, - 1, value.gate_value, frame_size * 3); + blas.GEMM(false, false, batch_size, frame_size * 2, frame_size, 1, + value.prev_out_value, frame_size, value.gate_weight, + frame_size * 2, 1, value.gate_value, frame_size * 3); } if (batch_size == 1) { @@ -61,10 +61,10 @@ struct GRUUnitFunctor { } if (value.prev_out_value) { - math::gemm( - context, false, false, batch_size, frame_size, frame_size, 1, - value.reset_output_value, frame_size, value.state_weight, frame_size, - 1, value.gate_value + frame_size * 2, frame_size * 3); + blas.GEMM(false, false, batch_size, frame_size, frame_size, 1, + value.reset_output_value, frame_size, value.state_weight, + frame_size, 1, value.gate_value + frame_size * 2, + frame_size * 3); } if (batch_size == 1) { @@ -121,18 +121,19 @@ struct GRUUnitGradFunctor { grad.output_grad, frame_size, batch_size, active_node); } + auto blas = math::GetBlas(context); + if (value.prev_out_value && grad.prev_out_grad) { - math::gemm( - context, false, true, batch_size, frame_size, frame_size, 1, - grad.gate_grad + frame_size * 2, frame_size * 3, value.state_weight, - frame_size, 0, grad.reset_output_grad, frame_size); + blas.GEMM(false, true, batch_size, frame_size, frame_size, 1, + grad.gate_grad + frame_size * 2, frame_size * 3, + value.state_weight, frame_size, 0, grad.reset_output_grad, + frame_size); if (grad.state_weight_grad) { - math::gemm( - context, true, false, frame_size, frame_size, batch_size, 1, - value.reset_output_value, frame_size, - grad.gate_grad + frame_size * 2, frame_size * 3, 1, - grad.state_weight_grad, frame_size); + blas.GEMM(true, false, frame_size, frame_size, batch_size, 1, + value.reset_output_value, frame_size, + grad.gate_grad + frame_size * 2, frame_size * 3, 1, + grad.state_weight_grad, frame_size); } } @@ -153,16 +154,14 @@ struct GRUUnitGradFunctor { } if (grad.prev_out_grad && value.prev_out_value) { - math::gemm( - context, false, true, batch_size, frame_size, frame_size * 2, 1, - grad.gate_grad, frame_size * 3, value.gate_weight, frame_size * 2, 1, - grad.prev_out_grad, frame_size); + blas.GEMM(false, true, batch_size, frame_size, frame_size * 2, 1, + grad.gate_grad, frame_size * 3, value.gate_weight, + frame_size * 2, 1, grad.prev_out_grad, frame_size); if (grad.gate_weight_grad) { - math::gemm( - context, true, false, frame_size, frame_size * 2, batch_size, 1, - value.prev_out_value, frame_size, grad.gate_grad, frame_size * 3, 1, - grad.gate_weight_grad, frame_size * 2); + blas.GEMM(true, false, frame_size, frame_size * 2, batch_size, 1, + value.prev_out_value, frame_size, grad.gate_grad, + frame_size * 3, 1, grad.gate_weight_grad, frame_size * 2); } } } diff --git a/paddle/fluid/operators/math/im2col.cc b/paddle/fluid/operators/math/im2col.cc index 123e10586f60f65d6c31a7c1f837bf32610d4ea5..336d6febc2ce3a55e82ed613bbc1081101f822f0 100644 --- a/paddle/fluid/operators/math/im2col.cc +++ b/paddle/fluid/operators/math/im2col.cc @@ -13,6 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/operators/math/im2col.h" +#include namespace paddle { namespace operators { diff --git a/paddle/fluid/operators/math/im2col.cu b/paddle/fluid/operators/math/im2col.cu index f41c78140fb601cd27a4e9997a736f46e903efc7..eecb233d22cea06da016b2671fd606b70eddf5a5 100644 --- a/paddle/fluid/operators/math/im2col.cu +++ b/paddle/fluid/operators/math/im2col.cu @@ -12,8 +12,10 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ +#include +#include #include "paddle/fluid/operators/math/im2col.h" -#include "paddle/fluid/platform/cuda_helper.h" +#include "paddle/fluid/platform/cuda_primitives.h" namespace paddle { namespace operators { diff --git a/paddle/fluid/operators/math/im2col.h b/paddle/fluid/operators/math/im2col.h index 451ec9d53498caa6d53a28dbfc764775d3ed3ecf..26d94e0f2e6163eb7452cf1fbea5966b4344ace1 100644 --- a/paddle/fluid/operators/math/im2col.h +++ b/paddle/fluid/operators/math/im2col.h @@ -14,6 +14,7 @@ limitations under the License. */ #pragma once +#include #include "paddle/fluid/framework/tensor.h" #include "paddle/fluid/framework/tensor_util.h" #include "paddle/fluid/platform/device_context.h" diff --git a/paddle/fluid/operators/math/im2col_test.cc b/paddle/fluid/operators/math/im2col_test.cc index b497c2a66146eed7b4b39ead6b05a9977cd4be21..8e3f0f286823c383bb0c44d0e7887040ec9b20a0 100644 --- a/paddle/fluid/operators/math/im2col_test.cc +++ b/paddle/fluid/operators/math/im2col_test.cc @@ -63,7 +63,7 @@ void testIm2col() { if (paddle::platform::is_cpu_place(*place)) { input = input_tmp; } else { - TensorCopy(input_tmp, *place, *context, &input); + TensorCopySync(input_tmp, *place, &input); } output_cfo.mutable_data( {1, filter_size, filter_size, output_height, output_width}, *place); @@ -88,7 +88,7 @@ void testIm2col() { if (paddle::platform::is_cpu_place(*place)) { out_cfo_ptr = output_cfo.data(); } else { - TensorCopy(output_cfo, paddle::platform::CPUPlace(), *context, &output_tmp); + TensorCopySync(output_cfo, paddle::platform::CPUPlace(), &output_tmp); out_cfo_ptr = output_tmp.data(); } for (int i = 0; i < 6; ++i) { @@ -99,7 +99,7 @@ void testIm2col() { if (paddle::platform::is_cpu_place(*place)) { out_ocf_ptr = output_ocf.data(); } else { - TensorCopy(output_ocf, paddle::platform::CPUPlace(), *context, &output_tmp); + TensorCopySync(output_ocf, paddle::platform::CPUPlace(), &output_tmp); out_ocf_ptr = output_tmp.data(); } @@ -120,7 +120,7 @@ void testIm2col() { if (paddle::platform::is_cpu_place(*place)) { input = input_tmp; } else { - TensorCopy(input_tmp, *place, *context, &input); + TensorCopySync(input_tmp, *place, &input); } col2im(*context, output_cfo, dilation, stride, padding, &input); @@ -129,7 +129,7 @@ void testIm2col() { if (paddle::platform::is_cpu_place(*place)) { in_ptr = input.data(); } else { - TensorCopy(input, paddle::platform::CPUPlace(), *context, &input_tmp); + TensorCopySync(input, paddle::platform::CPUPlace(), &input_tmp); in_ptr = input_tmp.data(); } for (int i = 0; i < 6; ++i) { @@ -141,7 +141,7 @@ void testIm2col() { if (paddle::platform::is_cpu_place(*place)) { input = input_tmp; } else { - TensorCopy(input_tmp, *place, *context, &input); + TensorCopySync(input_tmp, *place, &input); } col2im_ocf(*context, output_ocf, dilation, stride, padding, &input); @@ -149,7 +149,7 @@ void testIm2col() { if (paddle::platform::is_cpu_place(*place)) { in_ptr = input.data(); } else { - TensorCopy(input, paddle::platform::CPUPlace(), *context, &input_tmp); + TensorCopySync(input, paddle::platform::CPUPlace(), &input_tmp); in_ptr = input_tmp.data(); } for (int i = 0; i < 6; ++i) { diff --git a/paddle/fluid/operators/math/math_function.cc b/paddle/fluid/operators/math/math_function.cc index b5ae41c8f9d7aeb8e410b795fb9fbbd57ec69d4b..d39154c6f88d6d17c1719eb9a5b048211f4bb52b 100644 --- a/paddle/fluid/operators/math/math_function.cc +++ b/paddle/fluid/operators/math/math_function.cc @@ -24,266 +24,6 @@ namespace math { using float16 = paddle::platform::float16; -template <> -void gemm( - const platform::CPUDeviceContext& context, const CBLAS_TRANSPOSE transA, - const CBLAS_TRANSPOSE transB, const int M, const int N, const int K, - const float16 alpha, const float16* A, const float16* B, const float16 beta, - float16* C) { - PADDLE_THROW("float16 GEMM not supported on CPU"); -} - -template <> -void gemm( - const platform::CPUDeviceContext& context, const CBLAS_TRANSPOSE transA, - const CBLAS_TRANSPOSE transB, const int M, const int N, const int K, - const float alpha, const float* A, const float* B, const float beta, - float* C) { - int lda = (transA == CblasNoTrans) ? K : M; - int ldb = (transB == CblasNoTrans) ? N : K; - int ldc = N; - cblas_sgemm(CblasRowMajor, transA, transB, M, N, K, alpha, A, lda, B, ldb, - beta, C, ldc); -} - -template <> -void gemm( - const platform::CPUDeviceContext& context, const CBLAS_TRANSPOSE transA, - const CBLAS_TRANSPOSE transB, const int M, const int N, const int K, - const double alpha, const double* A, const double* B, const double beta, - double* C) { - int lda = (transA == CblasNoTrans) ? K : M; - int ldb = (transB == CblasNoTrans) ? N : K; - int ldc = N; - cblas_dgemm(CblasRowMajor, transA, transB, M, N, K, alpha, A, lda, B, ldb, - beta, C, ldc); -} - -template <> -void gemm( - const platform::CPUDeviceContext& context, const bool transA, - const bool transB, const int M, const int N, const int K, - const float16 alpha, const float16* A, const int lda, const float16* B, - const int ldb, const float16 beta, float16* C, const int ldc) { - PADDLE_THROW("float16 GEMM not supported on CPU"); -} - -template <> -void gemm( - const platform::CPUDeviceContext& context, const bool transA, - const bool transB, const int M, const int N, const int K, const float alpha, - const float* A, const int lda, const float* B, const int ldb, - const float beta, float* C, const int ldc) { - cblas_sgemm(CblasRowMajor, transA == false ? CblasNoTrans : CblasTrans, - transB == false ? CblasNoTrans : CblasTrans, M, N, K, alpha, A, - lda, B, ldb, beta, C, ldc); -} - -template <> -void gemm( - const platform::CPUDeviceContext& context, const bool transA, - const bool transB, const int M, const int N, const int K, - const double alpha, const double* A, const int lda, const double* B, - const int ldb, const double beta, double* C, const int ldc) { - cblas_dgemm(CblasRowMajor, transA == false ? CblasNoTrans : CblasTrans, - transB == false ? CblasNoTrans : CblasTrans, M, N, K, alpha, A, - lda, B, ldb, beta, C, ldc); -} - -template <> -void matmul( - const platform::CPUDeviceContext& context, - const framework::Tensor& matrix_a, bool trans_a, - const framework::Tensor& matrix_b, bool trans_b, float16 alpha, - framework::Tensor* matrix_out, float16 beta) { - PADDLE_THROW("float16 matmul not supported on CPU"); -} - -template <> -void matmul( - const platform::CPUDeviceContext& context, - const framework::Tensor& matrix_a, bool trans_a, - const framework::Tensor& matrix_b, bool trans_b, float alpha, - framework::Tensor* matrix_out, float beta) { - auto dim_a = matrix_a.dims(); - auto dim_b = matrix_b.dims(); - auto dim_out = matrix_out->dims(); - PADDLE_ENFORCE(dim_a.size() == 2 && dim_b.size() == 2 && dim_out.size() == 2, - "The input and output of matmul be matrix"); - - PADDLE_ENFORCE(platform::is_cpu_place(matrix_a.place()) && - platform::is_cpu_place(matrix_b.place()) && - platform::is_cpu_place(matrix_out->place()), - "Matrix must all be in CPUPlace"); - - int M = dim_out[0]; - int N = dim_out[1]; - int K = (trans_a == false) ? dim_a[1] : dim_a[0]; - - CBLAS_TRANSPOSE transA = (trans_a == false) ? CblasNoTrans : CblasTrans; - CBLAS_TRANSPOSE transB = (trans_b == false) ? CblasNoTrans : CblasTrans; - - gemm( - context, transA, transB, M, N, K, alpha, matrix_a.data(), - matrix_b.data(), beta, matrix_out->data()); -} - -template <> -void matmul( - const platform::CPUDeviceContext& context, - const framework::Tensor& matrix_a, bool trans_a, - const framework::Tensor& matrix_b, bool trans_b, double alpha, - framework::Tensor* matrix_out, double beta) { - auto dim_a = matrix_a.dims(); - auto dim_b = matrix_b.dims(); - auto dim_out = matrix_out->dims(); - PADDLE_ENFORCE(dim_a.size() == 2 && dim_b.size() == 2 && dim_out.size() == 2, - "The input and output of matmul be matrix"); - - PADDLE_ENFORCE(platform::is_cpu_place(matrix_a.place()) && - platform::is_cpu_place(matrix_b.place()) && - platform::is_cpu_place(matrix_out->place()), - "Matrix must all be in CPUPlace"); - - int M = dim_out[0]; - int N = dim_out[1]; - int K = (trans_a == false) ? dim_a[1] : dim_a[0]; - - CBLAS_TRANSPOSE transA = (trans_a == false) ? CblasNoTrans : CblasTrans; - CBLAS_TRANSPOSE transB = (trans_b == false) ? CblasNoTrans : CblasTrans; - - gemm( - context, transA, transB, M, N, K, alpha, matrix_a.data(), - matrix_b.data(), beta, matrix_out->data()); -} - -template <> -void batched_gemm( - const platform::CPUDeviceContext& context, const CBLAS_TRANSPOSE transA, - const CBLAS_TRANSPOSE transB, const int M, const int N, const int K, - const float16 alpha, const float16* A, const float16* B, const float16 beta, - float16* C, const int batchCount, const int64_t strideA, - const int64_t strideB) { - PADDLE_THROW("float16 batched_gemm not supported on CPU"); -} - -#ifdef PADDLE_WITH_MKLML -// Use cblas_{s,d}gemm_batched if available: Run with 1 group of size batchSize. -template <> -void batched_gemm( - const platform::CPUDeviceContext& context, const CBLAS_TRANSPOSE transA, - const CBLAS_TRANSPOSE transB, const int M, const int N, const int K, - const float alpha, const float* A, const float* B, const float beta, - float* C, const int batchCount, const int64_t strideA, - const int64_t strideB) { - int lda = (transA == CblasNoTrans) ? K : M; - int ldb = (transB == CblasNoTrans) ? N : K; - int ldc = N; - auto a_array = std::vector(batchCount); - auto b_array = std::vector(batchCount); - auto c_array = std::vector(batchCount); - for (int k = 0; k < batchCount; ++k) { - a_array[k] = &A[k * strideA]; - b_array[k] = &B[k * strideB]; - c_array[k] = &C[k * M * N]; - } - cblas_sgemm_batch(CblasRowMajor, &transA, &transB, &M, &N, &K, &alpha, - a_array.data(), &lda, b_array.data(), &ldb, &beta, - c_array.data(), &ldc, 1 /* group_count */, &batchCount); -} - -template <> -void batched_gemm( - const platform::CPUDeviceContext& context, const CBLAS_TRANSPOSE transA, - const CBLAS_TRANSPOSE transB, const int M, const int N, const int K, - const double alpha, const double* A, const double* B, const double beta, - double* C, const int batchCount, const int64_t strideA, - const int64_t strideB) { - int lda = (transA == CblasNoTrans) ? K : M; - int ldb = (transB == CblasNoTrans) ? N : K; - int ldc = N; - auto a_array = std::vector(batchCount); - auto b_array = std::vector(batchCount); - auto c_array = std::vector(batchCount); - for (int k = 0; k < batchCount; ++k) { - a_array[k] = &A[k * strideA]; - b_array[k] = &B[k * strideB]; - c_array[k] = &C[k * M * N]; - } - cblas_dgemm_batch(CblasRowMajor, &transA, &transB, &M, &N, &K, &alpha, - a_array.data(), &lda, b_array.data(), &ldb, &beta, - c_array.data(), &ldc, 1 /* group_count */, &batchCount); -} -#else -// The below is a naive but correct serial implementation that just loops -// over the batch dimension. This is a fallback for when the batched gemm -// functions of Intel MKL are not available. In the future, this computation -// should be parallelized. -template <> -void batched_gemm( - const platform::CPUDeviceContext& context, const CBLAS_TRANSPOSE transA, - const CBLAS_TRANSPOSE transB, const int M, const int N, const int K, - const float alpha, const float* A, const float* B, const float beta, - float* C, const int batchCount, const int64_t strideA, - const int64_t strideB) { - for (int k = 0; k < batchCount; ++k) { - const float* Ak = &A[k * strideA]; - const float* Bk = &B[k * strideB]; - float* Ck = &C[k * M * N]; - gemm(context, transA, transB, M, N, K, - alpha, Ak, Bk, beta, Ck); - } -} - -template <> -void batched_gemm( - const platform::CPUDeviceContext& context, const CBLAS_TRANSPOSE transA, - const CBLAS_TRANSPOSE transB, const int M, const int N, const int K, - const double alpha, const double* A, const double* B, const double beta, - double* C, const int batchCount, const int64_t strideA, - const int64_t strideB) { - for (int k = 0; k < batchCount; ++k) { - const double* Ak = &A[k * strideA]; - const double* Bk = &B[k * strideB]; - double* Ck = &C[k * M * N]; - gemm(context, transA, transB, M, N, K, - alpha, Ak, Bk, beta, Ck); - } -} -#endif - -template <> -void gemv( - const platform::CPUDeviceContext& context, const bool trans_a, const int M, - const int N, const float alpha, const float* A, const float* B, - const float beta, float* C) { - CBLAS_TRANSPOSE transA = (trans_a == false) ? CblasNoTrans : CblasTrans; - cblas_sgemv(CblasRowMajor, transA, M, N, alpha, A, N, B, 1, beta, C, 1); -} - -template <> -void gemv( - const platform::CPUDeviceContext& context, const bool trans_a, const int M, - const int N, const double alpha, const double* A, const double* B, - const double beta, double* C) { - CBLAS_TRANSPOSE transA = (trans_a == false) ? CblasNoTrans : CblasTrans; - cblas_dgemv(CblasRowMajor, transA, M, N, alpha, A, N, B, 1, beta, C, 1); -} - -template <> -void axpy( - const platform::CPUDeviceContext& context, const int n, const float alpha, - const float* x, float* y) { - cblas_saxpy(n, alpha, x, 1, y, 1); -} - -template <> -void axpy( - const platform::CPUDeviceContext& context, const int n, const double alpha, - const double* x, double* y) { - cblas_daxpy(n, alpha, x, 1, y, 1); -} - template struct SetConstant; template struct SetConstant; template struct SetConstant; @@ -298,7 +38,9 @@ template struct SetConstant; template struct Transpose; \ template struct Transpose; \ template struct Transpose; \ - template struct Transpose; + template struct Transpose; \ + template struct Transpose; \ + template struct Transpose; DEFINE_CPU_TRANS(1); DEFINE_CPU_TRANS(2); diff --git a/paddle/fluid/operators/math/math_function.cu b/paddle/fluid/operators/math/math_function.cu index 2aa819625e0f5213a6001908e715bcc73d4747c3..d5af718723e8d44da0971ea7756b8c36e771cca2 100644 --- a/paddle/fluid/operators/math/math_function.cu +++ b/paddle/fluid/operators/math/math_function.cu @@ -15,6 +15,7 @@ limitations under the License. */ #define EIGEN_USE_GPU #include #include "paddle/fluid/framework/data_type.h" +#include "paddle/fluid/operators/math/blas.h" #include "paddle/fluid/operators/math/math_function.h" #include "paddle/fluid/operators/math/math_function_impl.h" #include "paddle/fluid/platform/float16.h" @@ -25,374 +26,6 @@ namespace math { using float16 = paddle::platform::float16; -template <> -void gemm( - const platform::CUDADeviceContext& context, const CBLAS_TRANSPOSE transA, - const CBLAS_TRANSPOSE transB, const int M, const int N, const int K, - const float16 alpha, const float16* A, const float16* B, const float16 beta, - float16* C) { - // Note that cublas follows fortran order, so the order is different from - // the cblas convention. - int lda = (transA == CblasNoTrans) ? K : M; - int ldb = (transB == CblasNoTrans) ? N : K; - cublasOperation_t cuTransA = - (transA == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T; - cublasOperation_t cuTransB = - (transB == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T; - - // TODO(kexinzhao): add processing code for compute capability < 53 case - PADDLE_ENFORCE_GE(context.GetComputeCapability(), 53, - "cublas fp16 gemm requires GPU compute capability >= 53"); - -#if CUDA_VERSION >= 8000 - float h_alpha = static_cast(alpha); - float h_beta = static_cast(beta); - - cublasGemmAlgo_t algo = CUBLAS_GEMM_DFALT; -#if CUDA_VERSION >= 9000 - if (context.GetComputeCapability() >= 70) { - PADDLE_ENFORCE(platform::dynload::cublasSetMathMode(context.cublas_handle(), - CUBLAS_TENSOR_OP_MATH)); - algo = CUBLAS_GEMM_DFALT_TENSOR_OP; - } else { - PADDLE_ENFORCE(platform::dynload::cublasSetMathMode(context.cublas_handle(), - CUBLAS_DEFAULT_MATH)); - } -#endif // CUDA_VERSION >= 9000 - - // cublasHgemm does true FP16 computation which is slow for non-Volta - // GPUs. So use cublasGemmEx instead which does pesudo FP16 computation: - // input/output in fp16, computation in fp32, which can also be accelerated - // using tensor cores in volta GPUs. - PADDLE_ENFORCE(platform::dynload::cublasGemmEx( - context.cublas_handle(), cuTransB, cuTransA, N, M, K, &h_alpha, B, - CUDA_R_16F, ldb, A, CUDA_R_16F, lda, &h_beta, C, CUDA_R_16F, N, - CUDA_R_32F, algo)); -#else - // CUDA 7.5 does not support cublasGemmEx, hence we fall back to use hgemm - const half h_alpha = static_cast(alpha); - const half h_beta = static_cast(beta); - const half* h_A = reinterpret_cast(A); - const half* h_B = reinterpret_cast(B); - half* h_C = reinterpret_cast(C); - - PADDLE_ENFORCE(platform::dynload::cublasHgemm( - context.cublas_handle(), cuTransB, cuTransA, N, M, K, &h_alpha, h_B, ldb, - h_A, lda, &h_beta, h_C, N)); -#endif // CUDA_VERSION >= 8000 -} - -template <> -void gemm( - const platform::CUDADeviceContext& context, const CBLAS_TRANSPOSE transA, - const CBLAS_TRANSPOSE transB, const int M, const int N, const int K, - const float alpha, const float* A, const float* B, const float beta, - float* C) { - // Note that cublas follows fortran order, so the order is different from - // the cblas convention. - int lda = (transA == CblasNoTrans) ? K : M; - int ldb = (transB == CblasNoTrans) ? N : K; - cublasOperation_t cuTransA = - (transA == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T; - cublasOperation_t cuTransB = - (transB == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T; - - PADDLE_ENFORCE(platform::dynload::cublasSgemm( - context.cublas_handle(), cuTransB, cuTransA, N, M, K, &alpha, B, ldb, A, - lda, &beta, C, N)); -} - -template <> -void gemm( - const platform::CUDADeviceContext& context, const CBLAS_TRANSPOSE transA, - const CBLAS_TRANSPOSE transB, const int M, const int N, const int K, - const double alpha, const double* A, const double* B, const double beta, - double* C) { - // Note that cublas follows fortran order, so the order is different from - // the cblas convention. - int lda = (transA == CblasNoTrans) ? K : M; - int ldb = (transB == CblasNoTrans) ? N : K; - cublasOperation_t cuTransA = - (transA == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T; - cublasOperation_t cuTransB = - (transB == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T; - PADDLE_ENFORCE(platform::dynload::cublasDgemm( - context.cublas_handle(), cuTransB, cuTransA, N, M, K, &alpha, B, ldb, A, - lda, &beta, C, N)); -} - -template <> -void gemm( - const platform::CUDADeviceContext& context, const bool transA, - const bool transB, const int M, const int N, const int K, - const float16 alpha, const float16* A, const int lda, const float16* B, - const int ldb, const float16 beta, float16* C, const int ldc) { - // Note that cublas follows fortran order, so the order is different from - // the cblas convention. - cublasOperation_t cuTransA = transA == false ? CUBLAS_OP_N : CUBLAS_OP_T; - cublasOperation_t cuTransB = transB == false ? CUBLAS_OP_N : CUBLAS_OP_T; - - const half h_alpha = static_cast(alpha); - const half h_beta = static_cast(beta); - const half* h_A = reinterpret_cast(A); - const half* h_B = reinterpret_cast(B); - half* h_C = reinterpret_cast(C); - - // TODO(kexinzhao): add processing code for compute capability < 53 case - PADDLE_ENFORCE_GE(context.GetComputeCapability(), 53, - "cublas Hgemm requires GPU compute capability >= 53"); - PADDLE_ENFORCE(platform::dynload::cublasHgemm( - context.cublas_handle(), cuTransB, cuTransA, N, M, K, &h_alpha, h_B, ldb, - h_A, lda, &h_beta, h_C, ldc)); -} - -template <> -void gemm( - const platform::CUDADeviceContext& context, const bool transA, - const bool transB, const int M, const int N, const int K, const float alpha, - const float* A, const int lda, const float* B, const int ldb, - const float beta, float* C, const int ldc) { - // Note that cublas follows fortran order, so the order is different from - // the cblas convention. - cublasOperation_t cuTransA = transA == false ? CUBLAS_OP_N : CUBLAS_OP_T; - cublasOperation_t cuTransB = transB == false ? CUBLAS_OP_N : CUBLAS_OP_T; - PADDLE_ENFORCE(platform::dynload::cublasSgemm( - context.cublas_handle(), cuTransB, cuTransA, N, M, K, &alpha, B, ldb, A, - lda, &beta, C, ldc)); -} - -template <> -void gemm( - const platform::CUDADeviceContext& context, const bool transA, - const bool transB, const int M, const int N, const int K, - const double alpha, const double* A, const int lda, const double* B, - const int ldb, const double beta, double* C, const int ldc) { - // Note that cublas follows fortran order, so the order is different from - // the cblas convention. - cublasOperation_t cuTransA = transA == false ? CUBLAS_OP_N : CUBLAS_OP_T; - cublasOperation_t cuTransB = transB == false ? CUBLAS_OP_N : CUBLAS_OP_T; - PADDLE_ENFORCE(platform::dynload::cublasDgemm( - context.cublas_handle(), cuTransB, cuTransA, N, M, K, &alpha, B, ldb, A, - lda, &beta, C, ldc)); -} - -template <> -void matmul( - const platform::CUDADeviceContext& context, - const framework::Tensor& matrix_a, bool trans_a, - const framework::Tensor& matrix_b, bool trans_b, float16 alpha, - framework::Tensor* matrix_out, float16 beta) { - auto dim_a = matrix_a.dims(); - auto dim_b = matrix_b.dims(); - auto dim_out = matrix_out->dims(); - PADDLE_ENFORCE(dim_a.size() == 2 && dim_b.size() == 2 && dim_out.size() == 2, - "The input and output of matmul be matrix"); - - PADDLE_ENFORCE(platform::is_gpu_place(matrix_a.place()) && - platform::is_gpu_place(matrix_b.place()) && - platform::is_gpu_place(matrix_out->place()), - "Matrix must all be in CUDAPlace"); - - int M = dim_out[0]; - int N = dim_out[1]; - int K = (trans_a == false) ? dim_a[1] : dim_a[0]; - - CBLAS_TRANSPOSE transA = (trans_a == false) ? CblasNoTrans : CblasTrans; - CBLAS_TRANSPOSE transB = (trans_b == false) ? CblasNoTrans : CblasTrans; - - gemm( - context, transA, transB, M, N, K, alpha, matrix_a.data(), - matrix_b.data(), beta, matrix_out->data()); -} - -template <> -void matmul( - const platform::CUDADeviceContext& context, - const framework::Tensor& matrix_a, bool trans_a, - const framework::Tensor& matrix_b, bool trans_b, float alpha, - framework::Tensor* matrix_out, float beta) { - auto dim_a = matrix_a.dims(); - auto dim_b = matrix_b.dims(); - auto dim_out = matrix_out->dims(); - PADDLE_ENFORCE(dim_a.size() == 2 && dim_b.size() == 2 && dim_out.size() == 2, - "The input and output of matmul be matrix"); - - PADDLE_ENFORCE(platform::is_gpu_place(matrix_a.place()) && - platform::is_gpu_place(matrix_b.place()) && - platform::is_gpu_place(matrix_out->place()), - "Matrix must all be in CUDAPlace"); - - int M = dim_out[0]; - int N = dim_out[1]; - int K = (trans_a == false) ? dim_a[1] : dim_a[0]; - - CBLAS_TRANSPOSE transA = (trans_a == false) ? CblasNoTrans : CblasTrans; - CBLAS_TRANSPOSE transB = (trans_b == false) ? CblasNoTrans : CblasTrans; - - gemm( - context, transA, transB, M, N, K, alpha, matrix_a.data(), - matrix_b.data(), beta, matrix_out->data()); -} - -template <> -void matmul( - const platform::CUDADeviceContext& context, - const framework::Tensor& matrix_a, bool trans_a, - const framework::Tensor& matrix_b, bool trans_b, double alpha, - framework::Tensor* matrix_out, double beta) { - auto dim_a = matrix_a.dims(); - auto dim_b = matrix_b.dims(); - auto dim_out = matrix_out->dims(); - PADDLE_ENFORCE(dim_a.size() == 2 && dim_b.size() == 2 && dim_out.size() == 2, - "The input and output of matmul be matrix"); - - PADDLE_ENFORCE(platform::is_gpu_place(matrix_a.place()) && - platform::is_gpu_place(matrix_b.place()) && - platform::is_gpu_place(matrix_out->place()), - "Matrix must all be in CUDAPlace"); - - int M = dim_out[0]; - int N = dim_out[1]; - int K = (trans_a == false) ? dim_a[1] : dim_a[0]; - - CBLAS_TRANSPOSE transA = (trans_a == false) ? CblasNoTrans : CblasTrans; - CBLAS_TRANSPOSE transB = (trans_b == false) ? CblasNoTrans : CblasTrans; - - gemm( - context, transA, transB, M, N, K, alpha, matrix_a.data(), - matrix_b.data(), beta, matrix_out->data()); -} - -template <> -void batched_gemm( - const platform::CUDADeviceContext& context, const CBLAS_TRANSPOSE transA, - const CBLAS_TRANSPOSE transB, const int M, const int N, const int K, - const float16 alpha, const float16* A, const float16* B, const float16 beta, - float16* C, const int batchCount, const int64_t strideA, - const int64_t strideB) { -#if CUDA_VERSION >= 8000 - // Note that cublas follows fortran order, so the order is different from - // the cblas convention. - int lda = (transA == CblasNoTrans) ? K : M; - int ldb = (transB == CblasNoTrans) ? N : K; - int ldc = N; - cublasOperation_t cuTransA = - (transA == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T; - cublasOperation_t cuTransB = - (transB == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T; - const int64_t strideC = M * N; - - const half h_alpha = static_cast(alpha); - const half h_beta = static_cast(beta); - const half* h_A = reinterpret_cast(A); - const half* h_B = reinterpret_cast(B); - half* h_C = reinterpret_cast(C); - - // TODO(kexinzhao): add processing code for compute capability < 53 case - PADDLE_ENFORCE_GE(context.GetComputeCapability(), 53, - "cublas Hgemm requires GPU compute capability >= 53"); - - PADDLE_ENFORCE(platform::dynload::cublasHgemmStridedBatched( - context.cublas_handle(), cuTransB, cuTransA, N, M, K, &h_alpha, h_B, ldb, - strideB, h_A, lda, strideA, &h_beta, h_C, ldc, strideC, batchCount)); -#else - PADDLE_ENFORCE(false, "HgemmStridedBatched is not supported on cuda <= 7.5"); -#endif -} - -template <> -void batched_gemm( - const platform::CUDADeviceContext& context, const CBLAS_TRANSPOSE transA, - const CBLAS_TRANSPOSE transB, const int M, const int N, const int K, - const float alpha, const float* A, const float* B, const float beta, - float* C, const int batchCount, const int64_t strideA, - const int64_t strideB) { -#if CUDA_VERSION >= 8000 - // Note that cublas follows fortran order, so the order is different from - // the cblas convention. - int lda = (transA == CblasNoTrans) ? K : M; - int ldb = (transB == CblasNoTrans) ? N : K; - int ldc = N; - cublasOperation_t cuTransA = - (transA == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T; - cublasOperation_t cuTransB = - (transB == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T; - const int64_t strideC = M * N; - - PADDLE_ENFORCE(platform::dynload::cublasSgemmStridedBatched( - context.cublas_handle(), cuTransB, cuTransA, N, M, K, &alpha, B, ldb, - strideB, A, lda, strideA, &beta, C, ldc, strideC, batchCount)); -#else - PADDLE_ENFORCE(false, "SgemmStridedBatched is not supported on cuda <= 7.5"); -#endif -} - -template <> -void batched_gemm( - const platform::CUDADeviceContext& context, const CBLAS_TRANSPOSE transA, - const CBLAS_TRANSPOSE transB, const int M, const int N, const int K, - const double alpha, const double* A, const double* B, const double beta, - double* C, const int batchCount, const int64_t strideA, - const int64_t strideB) { -#if CUDA_VERSION >= 8000 - // Note that cublas follows fortran order, so the order is different from - // the cblas convention. - int lda = (transA == CblasNoTrans) ? K : M; - int ldb = (transB == CblasNoTrans) ? N : K; - int ldc = N; - cublasOperation_t cuTransA = - (transA == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T; - cublasOperation_t cuTransB = - (transB == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T; - const int64_t strideC = M * N; - - PADDLE_ENFORCE(platform::dynload::cublasDgemmStridedBatched( - context.cublas_handle(), cuTransB, cuTransA, N, M, K, &alpha, B, ldb, - strideB, A, lda, strideA, &beta, C, ldc, strideC, batchCount)); -#else - PADDLE_ENFORCE(false, "DgemmStridedBatched is not supported on cuda <= 7.5"); -#endif -} - -template <> -void gemv( - const platform::CUDADeviceContext& context, const bool trans_a, const int M, - const int N, const float alpha, const float* A, const float* B, - const float beta, float* C) { - cublasOperation_t cuTransA = (trans_a == false) ? CUBLAS_OP_T : CUBLAS_OP_N; - - PADDLE_ENFORCE(platform::dynload::cublasSgemv(context.cublas_handle(), - cuTransA, N, M, &alpha, A, N, B, - 1, &beta, C, 1)); -} - -template <> -void gemv( - const platform::CUDADeviceContext& context, const bool trans_a, const int M, - const int N, const double alpha, const double* A, const double* B, - const double beta, double* C) { - cublasOperation_t cuTransA = (trans_a == false) ? CUBLAS_OP_T : CUBLAS_OP_N; - PADDLE_ENFORCE(platform::dynload::cublasDgemv(context.cublas_handle(), - cuTransA, N, M, &alpha, A, N, B, - 1, &beta, C, 1)); -} - -template <> -void axpy( - const platform::CUDADeviceContext& context, const int n, const float alpha, - const float* x, float* y) { - PADDLE_ENFORCE(platform::dynload::cublasSaxpy(context.cublas_handle(), n, - &alpha, x, 1, y, 1)); -} - -template <> -void axpy( - const platform::CUDADeviceContext& context, const int n, const double alpha, - const double* x, double* y) { - PADDLE_ENFORCE(platform::dynload::cublasDaxpy(context.cublas_handle(), n, - &alpha, x, 1, y, 1)); -} - template struct SetConstant; template struct SetConstant; template struct SetConstant; @@ -400,9 +33,10 @@ template struct SetConstant; template struct SetConstant; template struct SetConstant; -#define DEFINE_GPU_TRANS(RANK) \ - template struct Transpose; \ - template struct Transpose; +#define DEFINE_GPU_TRANS(RANK) \ + template struct Transpose; \ + template struct Transpose; \ + template struct Transpose; DEFINE_GPU_TRANS(1); DEFINE_GPU_TRANS(2); @@ -484,10 +118,9 @@ void ColwiseSum::operator()( one.mutable_data({in_dims[0]}, context.GetPlace()); SetConstant set; set(context, &one, static_cast(1.0)); - gemv( - context, true, static_cast(in_dims[0]), static_cast(in_dims[1]), - 1.0, input.data(), one.data(), 0.0, - vector->data()); + GetBlas(context).GEMV( + true, static_cast(in_dims[0]), static_cast(in_dims[1]), 1.0, + input.data(), one.data(), 0.0, vector->data()); } template struct RowwiseSum; @@ -506,10 +139,9 @@ void RowwiseSum::operator()( one.mutable_data({size}, context.GetPlace()); SetConstant set; set(context, &one, static_cast(1.0)); - gemv( - context, true, static_cast(in_dims[1]), static_cast(in_dims[0]), - 1.0, one.data(), input.data(), 0.0, - vector->data()); + GetBlas(context).GEMV( + true, static_cast(in_dims[1]), static_cast(in_dims[0]), 1.0, + one.data(), input.data(), 0.0, vector->data()); } template struct RowwiseMean; diff --git a/paddle/fluid/operators/math/math_function.h b/paddle/fluid/operators/math/math_function.h index cdd02974722045457aacdfa517c147751185f332..d4b0e17ed44da61e2633b9bd97faeb62f9967c3c 100644 --- a/paddle/fluid/operators/math/math_function.h +++ b/paddle/fluid/operators/math/math_function.h @@ -42,6 +42,7 @@ int LAPACKE_dgetri(int matrix_layout, int n, double* a, int lda, #include #include "paddle/fluid/framework/eigen.h" +#include "paddle/fluid/framework/operator.h" #include "paddle/fluid/framework/tensor.h" #include "paddle/fluid/framework/tensor_util.h" #include "paddle/fluid/platform/device_context.h" @@ -50,47 +51,6 @@ int LAPACKE_dgetri(int matrix_layout, int n, double* a, int lda, namespace paddle { namespace operators { namespace math { - -// Support continuous memory now -// If transA = N, and transB = N -// Then matrixA: M * K, matrixB: K * N, matrixC : M * N -// For more detailed info, please refer to -// http://www.netlib.org/lapack/explore-html/d4/de2/sgemm_8f.html -template -void gemm(const DeviceContext& context, const CBLAS_TRANSPOSE transA, - const CBLAS_TRANSPOSE transB, const int M, const int N, const int K, - const T alpha, const T* A, const T* B, const T beta, T* C); - -// gemm wrapper with stride args for matrix uncontinuous in memory -template -void gemm(const DeviceContext& context, const bool transA, const bool transB, - const int M, const int N, const int K, const T alpha, const T* A, - const int lda, const T* B, const int ldb, const T beta, T* C, - const int ldc); - -// matrix multiply with continuous memory -template -void matmul(const DeviceContext& context, const framework::Tensor& matrix_a, - bool trans_a, const framework::Tensor& matrix_b, bool trans_b, - T alpha, framework::Tensor* matrix_out, T beta); - -// Batched gemm -template -void batched_gemm(const DeviceContext& context, const CBLAS_TRANSPOSE transA, - const CBLAS_TRANSPOSE transB, const int M, const int N, - const int K, const T alpha, const T* A, const T* B, - const T beta, T* C, const int batchCount, - const int64_t strideA, const int64_t strideB); - -template -void gemv(const DeviceContext& context, const bool trans_a, const int M, - const int N, const T alpha, const T* A, const T* B, const T beta, - T* C); - -template -void axpy(const DeviceContext& context, const int n, const T alpha, const T* x, - T* y); - template struct Transpose { void operator()(const DeviceContext& context, const framework::Tensor& in, diff --git a/paddle/fluid/operators/math/math_function_impl.h b/paddle/fluid/operators/math/math_function_impl.h index f9d4e4532428e3387ae39621d155434f05eb89d3..b9bd49d77d935e985705f78402ffe1ea90f24cb3 100644 --- a/paddle/fluid/operators/math/math_function_impl.h +++ b/paddle/fluid/operators/math/math_function_impl.h @@ -13,6 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once +#include #include "paddle/fluid/framework/data_type.h" #include "paddle/fluid/operators/math/math_function.h" diff --git a/paddle/fluid/operators/math/math_function_test.cc b/paddle/fluid/operators/math/math_function_test.cc index 25a9d0111eee45b28adff012b705cbfa2407d2b6..3719a264e90ea7d1a99eb9589ce4fd0d8e074781 100644 --- a/paddle/fluid/operators/math/math_function_test.cc +++ b/paddle/fluid/operators/math/math_function_test.cc @@ -13,6 +13,14 @@ // limitations under the License. #include "paddle/fluid/operators/math/math_function.h" #include "gtest/gtest.h" +#include "paddle/fluid/operators/math/blas.h" + +template +inline paddle::operators::math::BlasT +GetBlas(const paddle::platform::CPUDeviceContext& context) { + return paddle::operators::math::GetBlas(context); +} TEST(math_function, gemm_notrans_cblas) { paddle::framework::Tensor input1; @@ -34,9 +42,8 @@ TEST(math_function, gemm_notrans_cblas) { memcpy(input3_ptr, arr3, 8 * sizeof(float)); paddle::platform::CPUDeviceContext context(*cpu_place); - paddle::operators::math::gemm( - context, false, false, m, n, k, 1, input1_ptr, 3, input2_ptr + 1, 4, 1, - input3_ptr + 1, 4); + GetBlas(context).GEMM(false, false, m, n, k, 1, input1_ptr, 3, + input2_ptr + 1, 4, 1, input3_ptr + 1, 4); EXPECT_EQ(input3_ptr[0], 0); EXPECT_EQ(input3_ptr[1], 24); @@ -68,9 +75,8 @@ TEST(math_function, gemm_trans_clbas) { memcpy(input3_ptr, arr3, 8 * sizeof(float)); paddle::platform::CPUDeviceContext context(*cpu_place); - paddle::operators::math::gemm( - context, false, true, m, n, k, 1, input1_ptr, 3, input2_ptr + 3, 3, 1, - input3_ptr + 1, 4); + GetBlas(context).GEMM(false, true, m, n, k, 1, input1_ptr, 3, + input2_ptr + 3, 3, 1, input3_ptr + 1, 4); EXPECT_EQ(input3_ptr[0], 0); EXPECT_EQ(input3_ptr[1], 24); @@ -124,9 +130,8 @@ void GemvTest(int m, int n, bool trans) { } paddle::platform::CPUDeviceContext context(*cpu_place); - paddle::operators::math::gemv( - context, trans, static_cast(m), static_cast(n), 1., data_a, - data_b, 0., data_c); + GetBlas(context).GEMV(trans, static_cast(m), static_cast(n), 1., + data_a, data_b, 0., data_c); if (!trans) { for (int i = 0; i < m; ++i) { diff --git a/paddle/fluid/operators/math/math_function_test.cu b/paddle/fluid/operators/math/math_function_test.cu index 8982d9d066165a9da0461288685baa0c60e5f114..bcbb4a8274f149240b9f0990f38d9f38bdd0e5b1 100644 --- a/paddle/fluid/operators/math/math_function_test.cu +++ b/paddle/fluid/operators/math/math_function_test.cu @@ -12,7 +12,9 @@ // See the License for the specific language governing permissions and // limitations under the License. #include "gtest/gtest.h" +#include "paddle/fluid/operators/math/blas.h" #include "paddle/fluid/operators/math/math_function.h" +#include "paddle/fluid/platform/device_context.h" void fill_fp16_data(paddle::platform::float16* in_ptr, size_t size, const std::vector& data) { @@ -22,33 +24,36 @@ void fill_fp16_data(paddle::platform::float16* in_ptr, size_t size, } } -TEST(math_function, notrans_mul_trans_fp32) { - using namespace paddle::framework; - using namespace paddle::platform; +template +inline paddle::operators::math::BlasT +GetBlas(const paddle::platform::CUDADeviceContext& context) { + return paddle::operators::math::GetBlas(context); +} - Tensor input1; - Tensor input1_gpu; - Tensor input2_gpu; - Tensor out_gpu; - Tensor out; +TEST(math_function, notrans_mul_trans_fp32) { + paddle::framework::Tensor input1; + paddle::framework::Tensor input1_gpu; + paddle::framework::Tensor input2_gpu; + paddle::framework::Tensor out_gpu; + paddle::framework::Tensor out; - CPUPlace cpu_place; - CUDAPlace gpu_place(0); - CUDADeviceContext context(gpu_place); + paddle::platform::CPUPlace cpu_place; + paddle::platform::CUDAPlace gpu_place(0); + paddle::platform::CUDADeviceContext context(gpu_place); float* input1_ptr = input1.mutable_data({2, 3}, cpu_place); float arr[6] = {0, 1, 2, 3, 4, 5}; memcpy(input1_ptr, arr, 6 * sizeof(float)); - TensorCopy(input1, gpu_place, context, &input1_gpu); - TensorCopy(input1, gpu_place, context, &input2_gpu); + paddle::framework::TensorCopySync(input1, gpu_place, &input1_gpu); + paddle::framework::TensorCopySync(input1, gpu_place, &input2_gpu); out_gpu.mutable_data({2, 2}, gpu_place); + GetBlas(context).MatMul(input1_gpu, false, input2_gpu, true, 1, + &out_gpu, 0); - paddle::operators::math::matmul( - context, input1_gpu, false, input2_gpu, true, 1, &out_gpu, 0); - - TensorCopy(out_gpu, cpu_place, context, &out); + paddle::framework::TensorCopySync(out_gpu, cpu_place, &out); float* out_ptr = out.data(); context.Wait(); @@ -59,39 +64,37 @@ TEST(math_function, notrans_mul_trans_fp32) { } TEST(math_function, notrans_mul_trans_fp16) { - using namespace paddle::framework; - using namespace paddle::platform; - - Tensor input1; - Tensor input1_gpu; - Tensor input2_gpu; - Tensor out_gpu; - Tensor out; + paddle::framework::Tensor input1; + paddle::framework::Tensor input1_gpu; + paddle::framework::Tensor input2_gpu; + paddle::framework::Tensor out_gpu; + paddle::framework::Tensor out; - CPUPlace cpu_place; - CUDAPlace gpu_place(0); - CUDADeviceContext context(gpu_place); + paddle::platform::CPUPlace cpu_place; + paddle::platform::CUDAPlace gpu_place(0); + paddle::platform::CUDADeviceContext context(gpu_place); // fp16 GEMM in cublas requires GPU compute capability >= 53 if (context.GetComputeCapability() < 53) { return; } - float16* input1_ptr = input1.mutable_data({2, 3}, cpu_place); + paddle::platform::float16* input1_ptr = + input1.mutable_data({2, 3}, cpu_place); fill_fp16_data(input1_ptr, input1.numel(), {0, 1, 2, 3, 4, 5}); - TensorCopy(input1, gpu_place, context, &input1_gpu); - TensorCopy(input1, gpu_place, context, &input2_gpu); + paddle::framework::TensorCopySync(input1, gpu_place, &input1_gpu); + paddle::framework::TensorCopySync(input1, gpu_place, &input2_gpu); - out_gpu.mutable_data({2, 2}, gpu_place); + out_gpu.mutable_data({2, 2}, gpu_place); - paddle::operators::math::matmul( - context, input1_gpu, false, input2_gpu, true, float16(1), &out_gpu, - float16(0)); + GetBlas(context).MatMul( + input1_gpu, false, input2_gpu, true, paddle::platform::float16(1), + &out_gpu, paddle::platform::float16(0)); - TensorCopy(out_gpu, cpu_place, context, &out); + paddle::framework::TensorCopySync(out_gpu, cpu_place, &out); - float16* out_ptr = out.data(); + paddle::platform::float16* out_ptr = out.data(); context.Wait(); EXPECT_EQ(static_cast(out_ptr[0]), 5); EXPECT_EQ(static_cast(out_ptr[1]), 14); @@ -100,32 +103,29 @@ TEST(math_function, notrans_mul_trans_fp16) { } TEST(math_function, trans_mul_notrans_fp32) { - using namespace paddle::framework; - using namespace paddle::platform; - - Tensor input1; - Tensor input1_gpu; - Tensor input2_gpu; - Tensor out_gpu; - Tensor out; + paddle::framework::Tensor input1; + paddle::framework::Tensor input1_gpu; + paddle::framework::Tensor input2_gpu; + paddle::framework::Tensor out_gpu; + paddle::framework::Tensor out; - CPUPlace cpu_place; - CUDAPlace gpu_place(0); - CUDADeviceContext context(gpu_place); + paddle::platform::CPUPlace cpu_place; + paddle::platform::CUDAPlace gpu_place(0); + paddle::platform::CUDADeviceContext context(gpu_place); float* input1_ptr = input1.mutable_data({2, 3}, cpu_place); float arr[6] = {0, 1, 2, 3, 4, 5}; memcpy(input1_ptr, arr, 6 * sizeof(float)); - TensorCopy(input1, gpu_place, context, &input1_gpu); - TensorCopy(input1, gpu_place, context, &input2_gpu); + paddle::framework::TensorCopySync(input1, gpu_place, &input1_gpu); + paddle::framework::TensorCopySync(input1, gpu_place, &input2_gpu); out_gpu.mutable_data({3, 3}, gpu_place); - paddle::operators::math::matmul( - context, input1_gpu, true, input2_gpu, false, 1, &out_gpu, 0); + GetBlas(context).MatMul(input1_gpu, true, input2_gpu, false, 1, + &out_gpu, 0); - TensorCopy(out_gpu, cpu_place, context, &out); + paddle::framework::TensorCopySync(out_gpu, cpu_place, &out); float* out_ptr = out.data(); context.Wait(); @@ -141,39 +141,37 @@ TEST(math_function, trans_mul_notrans_fp32) { } TEST(math_function, trans_mul_notrans_fp16) { - using namespace paddle::framework; - using namespace paddle::platform; + paddle::framework::Tensor input1; + paddle::framework::Tensor input1_gpu; + paddle::framework::Tensor input2_gpu; + paddle::framework::Tensor out_gpu; + paddle::framework::Tensor out; - Tensor input1; - Tensor input1_gpu; - Tensor input2_gpu; - Tensor out_gpu; - Tensor out; - - CPUPlace cpu_place; - CUDAPlace gpu_place(0); - CUDADeviceContext context(gpu_place); + paddle::platform::CPUPlace cpu_place; + paddle::platform::CUDAPlace gpu_place(0); + paddle::platform::CUDADeviceContext context(gpu_place); // fp16 GEMM in cublas requires GPU compute capability >= 53 if (context.GetComputeCapability() < 53) { return; } - float16* input1_ptr = input1.mutable_data({2, 3}, cpu_place); + paddle::platform::float16* input1_ptr = + input1.mutable_data({2, 3}, cpu_place); fill_fp16_data(input1_ptr, input1.numel(), {0, 1, 2, 3, 4, 5}); - TensorCopy(input1, gpu_place, context, &input1_gpu); - TensorCopy(input1, gpu_place, context, &input2_gpu); + paddle::framework::TensorCopySync(input1, gpu_place, &input1_gpu); + paddle::framework::TensorCopySync(input1, gpu_place, &input2_gpu); - out_gpu.mutable_data({3, 3}, gpu_place); + out_gpu.mutable_data({3, 3}, gpu_place); - paddle::operators::math::matmul( - context, input1_gpu, true, input2_gpu, false, float16(1), &out_gpu, - float16(0)); + GetBlas(context).MatMul( + input1_gpu, true, input2_gpu, false, paddle::platform::float16(1), + &out_gpu, paddle::platform::float16(0)); - TensorCopy(out_gpu, cpu_place, context, &out); + paddle::framework::TensorCopySync(out_gpu, cpu_place, &out); - float16* out_ptr = out.data(); + paddle::platform::float16* out_ptr = out.data(); context.Wait(); EXPECT_EQ(static_cast(out_ptr[0]), 9); EXPECT_EQ(static_cast(out_ptr[1]), 12); @@ -187,19 +185,16 @@ TEST(math_function, trans_mul_notrans_fp16) { } TEST(math_function, gemm_notrans_cublas_fp32) { - using namespace paddle::framework; - using namespace paddle::platform; - - Tensor input1; - Tensor input2; - Tensor input3; - Tensor input1_gpu; - Tensor input2_gpu; - Tensor input3_gpu; + paddle::framework::Tensor input1; + paddle::framework::Tensor input2; + paddle::framework::Tensor input3; + paddle::framework::Tensor input1_gpu; + paddle::framework::Tensor input2_gpu; + paddle::framework::Tensor input3_gpu; - CPUPlace cpu_place; - CUDAPlace gpu_place(0); - CUDADeviceContext context(gpu_place); + paddle::platform::CPUPlace cpu_place; + paddle::platform::CUDAPlace gpu_place(0); + paddle::platform::CUDADeviceContext context(gpu_place); int m = 2; int n = 3; @@ -214,17 +209,17 @@ TEST(math_function, gemm_notrans_cublas_fp32) { float arr3[8] = {0, 1, 2, 3, 4, 5, 6, 7}; memcpy(input3_ptr, arr3, 8 * sizeof(float)); - TensorCopy(input1, gpu_place, context, &input1_gpu); - TensorCopy(input2, gpu_place, context, &input2_gpu); - TensorCopy(input3, gpu_place, context, &input3_gpu); + paddle::framework::TensorCopySync(input1, gpu_place, &input1_gpu); + paddle::framework::TensorCopySync(input2, gpu_place, &input2_gpu); + paddle::framework::TensorCopySync(input3, gpu_place, &input3_gpu); float* a = input1_gpu.data(); float* b = input2_gpu.data(); float* c = input3_gpu.mutable_data(gpu_place); - paddle::operators::math::gemm( - context, false, false, m, n, k, 1, a, 3, b + 1, 4, 1, c + 1, 4); + GetBlas(context).GEMM(false, false, m, n, k, 1, a, 3, b + 1, 4, 1, + c + 1, 4); - TensorCopy(input3_gpu, cpu_place, context, &input3); + paddle::framework::TensorCopySync(input3_gpu, cpu_place, &input3); // numpy code: // a = np.arange(6).reshape(2, 3) @@ -244,19 +239,16 @@ TEST(math_function, gemm_notrans_cublas_fp32) { } TEST(math_function, gemm_notrans_cublas_fp16) { - using namespace paddle::framework; - using namespace paddle::platform; - - Tensor input1; - Tensor input2; - Tensor input3; - Tensor input1_gpu; - Tensor input2_gpu; - Tensor input3_gpu; + paddle::framework::Tensor input1; + paddle::framework::Tensor input2; + paddle::framework::Tensor input3; + paddle::framework::Tensor input1_gpu; + paddle::framework::Tensor input2_gpu; + paddle::framework::Tensor input3_gpu; - CPUPlace cpu_place; - CUDAPlace gpu_place(0); - CUDADeviceContext context(gpu_place); + paddle::platform::CPUPlace cpu_place; + paddle::platform::CUDAPlace gpu_place(0); + paddle::platform::CUDADeviceContext context(gpu_place); // fp16 GEMM in cublas requires GPU compute capability >= 53 if (context.GetComputeCapability() < 53) { @@ -266,26 +258,30 @@ TEST(math_function, gemm_notrans_cublas_fp16) { int m = 2; int n = 3; int k = 3; - float16* input1_ptr = input1.mutable_data({2, 3}, cpu_place); + paddle::platform::float16* input1_ptr = + input1.mutable_data({2, 3}, cpu_place); fill_fp16_data(input1_ptr, input1.numel(), {0, 1, 2, 3, 4, 5}); - float16* input2_ptr = input2.mutable_data({3, 4}, cpu_place); + paddle::platform::float16* input2_ptr = + input2.mutable_data({3, 4}, cpu_place); fill_fp16_data(input2_ptr, input2.numel(), {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11}); - float16* input3_ptr = input3.mutable_data({2, 4}, cpu_place); + paddle::platform::float16* input3_ptr = + input3.mutable_data({2, 4}, cpu_place); fill_fp16_data(input3_ptr, input3.numel(), {0, 1, 2, 3, 4, 5, 6, 7}); - TensorCopy(input1, gpu_place, context, &input1_gpu); - TensorCopy(input2, gpu_place, context, &input2_gpu); - TensorCopy(input3, gpu_place, context, &input3_gpu); - float16* a = input1_gpu.data(); - float16* b = input2_gpu.data(); - float16* c = input3_gpu.mutable_data(gpu_place); + paddle::framework::TensorCopySync(input1, gpu_place, &input1_gpu); + paddle::framework::TensorCopySync(input2, gpu_place, &input2_gpu); + paddle::framework::TensorCopySync(input3, gpu_place, &input3_gpu); + paddle::platform::float16* a = input1_gpu.data(); + paddle::platform::float16* b = input2_gpu.data(); + paddle::platform::float16* c = + input3_gpu.mutable_data(gpu_place); - paddle::operators::math::gemm( - context, false, false, m, n, k, float16(1), a, 3, b + 1, 4, float16(1), - c + 1, 4); + GetBlas(context).GEMM( + false, false, m, n, k, static_cast(1), a, 3, + b + 1, 4, static_cast(1), c + 1, 4); - TensorCopy(input3_gpu, cpu_place, context, &input3); + paddle::framework::TensorCopySync(input3_gpu, cpu_place, &input3); // numpy code: // a = np.arange(6).reshape(2, 3) @@ -305,19 +301,16 @@ TEST(math_function, gemm_notrans_cublas_fp16) { } TEST(math_function, gemm_trans_cublas_fp32) { - using namespace paddle::framework; - using namespace paddle::platform; + paddle::framework::Tensor input1; + paddle::framework::Tensor input2; + paddle::framework::Tensor input3; + paddle::framework::Tensor input1_gpu; + paddle::framework::Tensor input2_gpu; + paddle::framework::Tensor input3_gpu; - Tensor input1; - Tensor input2; - Tensor input3; - Tensor input1_gpu; - Tensor input2_gpu; - Tensor input3_gpu; - - CPUPlace cpu_place; - CUDAPlace gpu_place(0); - CUDADeviceContext context(gpu_place); + paddle::platform::CPUPlace cpu_place; + paddle::platform::CUDAPlace gpu_place(0); + paddle::platform::CUDADeviceContext context(gpu_place); int m = 2; int n = 3; @@ -332,17 +325,17 @@ TEST(math_function, gemm_trans_cublas_fp32) { float arr3[8] = {0, 1, 2, 3, 4, 5, 6, 7}; memcpy(input3_ptr, arr3, 8 * sizeof(float)); - TensorCopy(input1, gpu_place, context, &input1_gpu); - TensorCopy(input2, gpu_place, context, &input2_gpu); - TensorCopy(input3, gpu_place, context, &input3_gpu); + paddle::framework::TensorCopySync(input1, gpu_place, &input1_gpu); + paddle::framework::TensorCopySync(input2, gpu_place, &input2_gpu); + paddle::framework::TensorCopySync(input3, gpu_place, &input3_gpu); float* a = input1_gpu.data(); float* b = input2_gpu.data(); float* c = input3_gpu.mutable_data(gpu_place); - paddle::operators::math::gemm( - context, false, true, m, n, k, 1, a, 3, b + 3, 3, 1, c + 1, 4); + GetBlas(context).GEMM(false, true, m, n, k, 1, a, 3, b + 3, 3, 1, + c + 1, 4); - TensorCopy(input3_gpu, cpu_place, context, &input3); + paddle::framework::TensorCopySync(input3_gpu, cpu_place, &input3); context.Wait(); EXPECT_EQ(input3_ptr[0], 0); @@ -356,19 +349,16 @@ TEST(math_function, gemm_trans_cublas_fp32) { } TEST(math_function, gemm_trans_cublas_fp16) { - using namespace paddle::framework; - using namespace paddle::platform; - - Tensor input1; - Tensor input2; - Tensor input3; - Tensor input1_gpu; - Tensor input2_gpu; - Tensor input3_gpu; + paddle::framework::Tensor input1; + paddle::framework::Tensor input2; + paddle::framework::Tensor input3; + paddle::framework::Tensor input1_gpu; + paddle::framework::Tensor input2_gpu; + paddle::framework::Tensor input3_gpu; - CPUPlace cpu_place; - CUDAPlace gpu_place(0); - CUDADeviceContext context(gpu_place); + paddle::platform::CPUPlace cpu_place; + paddle::platform::CUDAPlace gpu_place(0); + paddle::platform::CUDADeviceContext context(gpu_place); // fp16 GEMM in cublas requires GPU compute capability >= 53 if (context.GetComputeCapability() < 53) { @@ -378,26 +368,30 @@ TEST(math_function, gemm_trans_cublas_fp16) { int m = 2; int n = 3; int k = 3; - float16* input1_ptr = input1.mutable_data({2, 3}, cpu_place); + paddle::platform::float16* input1_ptr = + input1.mutable_data({2, 3}, cpu_place); fill_fp16_data(input1_ptr, input1.numel(), {0, 1, 2, 3, 4, 5}); - float16* input2_ptr = input2.mutable_data({4, 3}, cpu_place); + paddle::platform::float16* input2_ptr = + input2.mutable_data({4, 3}, cpu_place); fill_fp16_data(input2_ptr, input2.numel(), {0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11}); - float16* input3_ptr = input3.mutable_data({2, 4}, cpu_place); + paddle::platform::float16* input3_ptr = + input3.mutable_data({2, 4}, cpu_place); fill_fp16_data(input3_ptr, input3.numel(), {0, 1, 2, 3, 4, 5, 6, 7}); - TensorCopy(input1, gpu_place, context, &input1_gpu); - TensorCopy(input2, gpu_place, context, &input2_gpu); - TensorCopy(input3, gpu_place, context, &input3_gpu); - float16* a = input1_gpu.data(); - float16* b = input2_gpu.data(); - float16* c = input3_gpu.mutable_data(gpu_place); + paddle::framework::TensorCopySync(input1, gpu_place, &input1_gpu); + paddle::framework::TensorCopySync(input2, gpu_place, &input2_gpu); + paddle::framework::TensorCopySync(input3, gpu_place, &input3_gpu); + paddle::platform::float16* a = input1_gpu.data(); + paddle::platform::float16* b = input2_gpu.data(); + paddle::platform::float16* c = + input3_gpu.mutable_data(gpu_place); - paddle::operators::math::gemm( - context, false, true, m, n, k, float16(1), a, 3, b + 3, 3, float16(1), - c + 1, 4); + GetBlas(context).GEMM( + false, true, m, n, k, static_cast(1), a, 3, + b + 3, 3, static_cast(1), c + 1, 4); - TensorCopy(input3_gpu, cpu_place, context, &input3); + paddle::framework::TensorCopySync(input3_gpu, cpu_place, &input3); context.Wait(); EXPECT_EQ(static_cast(input3_ptr[0]), 0); @@ -412,24 +406,21 @@ TEST(math_function, gemm_trans_cublas_fp16) { template void GemvTest(int m, int n, bool trans) { - using namespace paddle::framework; - using namespace paddle::platform; - - Tensor mat_a; - Tensor vec_b; - Tensor vec_c; + paddle::framework::Tensor mat_a; + paddle::framework::Tensor vec_b; + paddle::framework::Tensor vec_c; - CPUPlace cpu_place; - CUDAPlace gpu_place(0); - CUDADeviceContext context(gpu_place); + paddle::platform::CPUPlace cpu_place; + paddle::platform::CUDAPlace gpu_place(0); + paddle::platform::CUDADeviceContext context(gpu_place); T* data_a = mat_a.mutable_data({m, n}, cpu_place); T* data_b = vec_b.mutable_data({trans ? m : n}, cpu_place); T* data_c = vec_c.mutable_data({trans ? n : m}, cpu_place); - Tensor g_mat_a; - Tensor g_vec_b; - Tensor g_vec_c; + paddle::framework::Tensor g_mat_a; + paddle::framework::Tensor g_vec_b; + paddle::framework::Tensor g_vec_c; T* g_data_a = g_mat_a.mutable_data(mat_a.dims(), gpu_place); T* g_data_b = g_vec_b.mutable_data(vec_b.dims(), gpu_place); T* g_data_c = g_vec_c.mutable_data(vec_c.dims(), gpu_place); @@ -441,14 +432,13 @@ void GemvTest(int m, int n, bool trans) { data_b[i] = static_cast(i); } - TensorCopy(mat_a, gpu_place, context, &g_mat_a); - TensorCopy(vec_b, gpu_place, context, &g_vec_b); + paddle::framework::TensorCopySync(mat_a, gpu_place, &g_mat_a); + paddle::framework::TensorCopySync(vec_b, gpu_place, &g_vec_b); - paddle::operators::math::gemv( - context, trans, static_cast(m), static_cast(n), 1., g_data_a, - g_data_b, 0., g_data_c); + GetBlas(context).GEMV(trans, static_cast(m), static_cast(n), 1., + g_data_a, g_data_b, 0., g_data_c); - TensorCopy(g_vec_c, cpu_place, context, &vec_c); + paddle::framework::TensorCopySync(g_vec_c, cpu_place, &vec_c); if (!trans) { for (int i = 0; i < m; ++i) { diff --git a/paddle/fluid/operators/math/matmul.h b/paddle/fluid/operators/math/matmul.h deleted file mode 100644 index 0006c5062f3639da589eea44d47917d879933615..0000000000000000000000000000000000000000 --- a/paddle/fluid/operators/math/matmul.h +++ /dev/null @@ -1,147 +0,0 @@ -/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once -#include -#include -#include "paddle/fluid/operators/math/math_function.h" - -namespace paddle { -namespace operators { -namespace math { - -// Implements the logic of numpy matmul: -// https://docs.scipy.org/doc/numpy-1.13.0/reference/generated/numpy.matmul.html -// -// but allowing also for a, b to be transposed -// -// Both a & b can be 1- to 3-dimensional. Higher rank tensors are not supported -// yet. -template -class MatMulFunctor { - public: - void operator()(const DeviceContext& context, const framework::Tensor& a, - bool trans_a, const framework::Tensor& b, bool trans_b, - T alpha, framework::Tensor* out, T beta) { - auto dim_a = a.dims(); - auto dim_b = b.dims(); - - PADDLE_ENFORCE(a.place() == b.place() && b.place() == out->place(), - "Tensors must all be in the same place."); - PADDLE_ENFORCE_GE(dim_a.size(), 1, - "Input tensor a must be at least 1-dimensional."); - PADDLE_ENFORCE_GE(dim_b.size(), 1, - "Input tensor b must be at least 1-dimensional."); - - std::vector out_dim; - int64_t batch_count = 1; - if (dim_a.size() > 3) { - PADDLE_ENFORCE(dim_b.size() == dim_a.size(), - "The dimensions of X and Y must be the same, and both of " - "them should be %d-dimensional.", - dim_b.size()); - // The first rank-2 dimensions are accumulated on the batch_count, and the - // last two dimensions are used for matrix multiplication. - for (int j = 0; j < dim_a.size() - 2; ++j) { - PADDLE_ENFORCE_EQ(dim_b[j], dim_a[j], - "The %d-th dimension of X and Y must be the same.", - j); - out_dim.push_back(dim_a[j]); - batch_count *= dim_a[j]; - } - } - - int M = 0, N = 0, kA = 0, kB = 0, batchCountA = 0, batchCountB = 0, - strideA = 0, strideB = 0; - - switch (dim_a.size()) { - case 1: - // similar to np.matmul: - // prepend dimension 1 (no transpose) or append dimension 1 (transpose) - M = trans_a ? dim_a[0] : 1; - kA = trans_a ? 1 : dim_a[0]; - break; - case 2: - M = trans_a ? dim_a[1] : dim_a[0]; - kA = trans_a ? dim_a[0] : dim_a[1]; - break; - case 3: - batchCountA = dim_a[0]; - M = trans_a ? dim_a[2] : dim_a[1]; - kA = trans_a ? dim_a[1] : dim_a[2]; - strideA = M * kA; - break; - default: - batchCountA = batch_count; - size_t mat_s = dim_a.size() - 2; - M = trans_a ? dim_a[mat_s + 1] : dim_a[mat_s]; - kA = trans_a ? dim_a[mat_s] : dim_a[mat_s + 1]; - strideA = M * kA; - } - - switch (dim_b.size()) { - case 1: - // similar to np.matmul: - // append dimension 1 (no transpose) or prepend dimension 1 (transpose) - kB = trans_b ? 1 : dim_b[0]; - N = trans_b ? dim_b[0] : 1; - break; - case 2: - kB = trans_b ? dim_b[1] : dim_b[0]; - N = trans_b ? dim_b[0] : dim_b[1]; - break; - case 3: - batchCountB = dim_b[0]; - kB = trans_b ? dim_b[2] : dim_b[1]; - N = trans_b ? dim_b[1] : dim_b[2]; - strideB = kB * N; - break; - default: - batchCountB = batch_count; - size_t mat_s = dim_b.size() - 2; - kB = trans_b ? dim_b[mat_s + 1] : dim_b[mat_s]; - N = trans_b ? dim_b[mat_s] : dim_b[mat_s + 1]; - strideB = kB * N; - } - - PADDLE_ENFORCE_EQ( - kA, kB, - "First matrix's width must be equal with second matrix's height."); - if (batchCountA && batchCountB) { - PADDLE_ENFORCE_EQ( - batchCountA, batchCountB, - "When input tensors a and b are both batched, they must have the " - "same batch dimension."); - } - int batchCount = std::max(batchCountA, batchCountB); - - CBLAS_TRANSPOSE transA = (trans_a == false) ? CblasNoTrans : CblasTrans; - CBLAS_TRANSPOSE transB = (trans_b == false) ? CblasNoTrans : CblasTrans; - - if (!batchCount) { - // regular matrix multiplication - gemm(context, transA, transB, M, N, kA, alpha, - a.data(), b.data(), beta, out->data()); - } else { - // batched matrix multiplication - batched_gemm( - context, transA, transB, M, N, kA, alpha, a.data(), b.data(), - beta, out->data(), batchCount, strideA, strideB); - } - } -}; - -} // namespace math -} // namespace operators -} // namespace paddle diff --git a/paddle/fluid/operators/math/maxouting.cu b/paddle/fluid/operators/math/maxouting.cu index 1e1a6a221c71c9d9cb9fda468360cb502c5ea52f..d9a23299a4d5750fc8c7fe3e5d1f8cd94bcb9cae 100644 --- a/paddle/fluid/operators/math/maxouting.cu +++ b/paddle/fluid/operators/math/maxouting.cu @@ -13,7 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/operators/math/maxouting.h" -#include "paddle/fluid/platform/cuda_helper.h" +#include "paddle/fluid/platform/cuda_primitives.h" namespace paddle { namespace operators { diff --git a/paddle/fluid/operators/math/pooling.cc b/paddle/fluid/operators/math/pooling.cc index 97a2e81c84c060a8be57db6274839ee39edf466c..b871851798e48e6b598cb4ab8e2e42db478a3820 100644 --- a/paddle/fluid/operators/math/pooling.cc +++ b/paddle/fluid/operators/math/pooling.cc @@ -11,8 +11,9 @@ distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ - #include "paddle/fluid/operators/math/pooling.h" +#include +#include namespace paddle { namespace operators { @@ -27,9 +28,10 @@ template class Pool2dFunctor { public: void operator()(const platform::CPUDeviceContext& context, - const framework::Tensor& input, std::vector& ksize, - std::vector& strides, std::vector& paddings, - PoolProcess pool_process, framework::Tensor* output) { + const framework::Tensor& input, const std::vector& ksize, + const std::vector& strides, + const std::vector& paddings, PoolProcess pool_process, + framework::Tensor* output) { const int batch_size = input.dims()[0]; const int input_height = input.dims()[2]; const int input_width = input.dims()[3]; @@ -63,11 +65,11 @@ class Pool2dFunctor { T ele = pool_process.initial(); for (int h = hstart; h < hend; ++h) { for (int w = wstart; w < wend; ++w) { - pool_process.compute(ele, input_data[h * input_width + w]); + pool_process.compute(input_data[h * input_width + w], &ele); } } int pool_size = (hend - hstart) * (wend - wstart); - pool_process.finalize(ele, (static_cast(pool_size))); + pool_process.finalize(static_cast(pool_size), &ele); output_data[ph * output_width + pw] = ele; } } @@ -86,13 +88,12 @@ class Pool2dFunctor { template class Pool2dGradFunctor { public: - void operator()(const platform::CPUDeviceContext& context, - const framework::Tensor& input, - const framework::Tensor& output, - const framework::Tensor& output_grad, std::vector& ksize, - std::vector& strides, std::vector& paddings, - PoolProcess pool_grad_process, - framework::Tensor* input_grad) { + void operator()( + const platform::CPUDeviceContext& context, const framework::Tensor& input, + const framework::Tensor& output, const framework::Tensor& output_grad, + const std::vector& ksize, const std::vector& strides, + const std::vector& paddings, PoolProcess pool_grad_process, + framework::Tensor* input_grad) { const int batch_size = input.dims()[0]; const int input_height = input.dims()[2]; const int input_width = input.dims()[3]; @@ -131,8 +132,8 @@ class Pool2dGradFunctor { input_data[h * input_width + w], output_data[ph * output_width + pw], output_grad_data[ph * output_width + pw], - input_grad_data[h * input_width + w], - static_cast(scale)); + static_cast(scale), + input_grad_data + h * input_width + w); } } } @@ -154,12 +155,11 @@ class Pool2dGradFunctor { template class MaxPool2dGradFunctor { public: - void operator()(const platform::CPUDeviceContext& context, - const framework::Tensor& input, - const framework::Tensor& output, - const framework::Tensor& output_grad, std::vector& ksize, - std::vector& strides, std::vector& paddings, - framework::Tensor* input_grad) { + void operator()( + const platform::CPUDeviceContext& context, const framework::Tensor& input, + const framework::Tensor& output, const framework::Tensor& output_grad, + const std::vector& ksize, const std::vector& strides, + const std::vector& paddings, framework::Tensor* input_grad) { const int batch_size = input.dims()[0]; const int input_height = input.dims()[2]; const int input_width = input.dims()[3]; @@ -246,9 +246,10 @@ template class Pool3dFunctor { public: void operator()(const platform::CPUDeviceContext& context, - const framework::Tensor& input, std::vector& ksize, - std::vector& strides, std::vector& paddings, - PoolProcess pool_process, framework::Tensor* output) { + const framework::Tensor& input, const std::vector& ksize, + const std::vector& strides, + const std::vector& paddings, PoolProcess pool_process, + framework::Tensor* output) { const int batch_size = input.dims()[0]; const int input_depth = input.dims()[2]; const int input_height = input.dims()[3]; @@ -293,14 +294,14 @@ class Pool3dFunctor { for (int h = hstart; h < hend; ++h) { for (int w = wstart; w < wend; ++w) { pool_process.compute( - ele, - input_data[(d * input_height + h) * input_width + w]); + input_data[(d * input_height + h) * input_width + w], + &ele); } } } int pool_size = (dend - dstart) * (hend - hstart) * (wend - wstart); - pool_process.finalize(ele, static_cast(pool_size)); + pool_process.finalize(static_cast(pool_size), &ele); output_data[output_idx] = ele; } } @@ -320,13 +321,12 @@ class Pool3dFunctor { template class Pool3dGradFunctor { public: - void operator()(const platform::CPUDeviceContext& context, - const framework::Tensor& input, - const framework::Tensor& output, - const framework::Tensor& output_grad, std::vector& ksize, - std::vector& strides, std::vector& paddings, - PoolProcess pool_grad_process, - framework::Tensor* input_grad) { + void operator()( + const platform::CPUDeviceContext& context, const framework::Tensor& input, + const framework::Tensor& output, const framework::Tensor& output_grad, + const std::vector& ksize, const std::vector& strides, + const std::vector& paddings, PoolProcess pool_grad_process, + framework::Tensor* input_grad) { const int batch_size = input.dims()[0]; const int input_depth = input.dims()[2]; const int input_height = input.dims()[3]; @@ -379,8 +379,8 @@ class Pool3dGradFunctor { (pd * output_height + ph) * output_width + pw; pool_grad_process.compute( input_data[input_idx], output_data[output_idx], - output_grad_data[output_idx], - input_grad_data[input_idx], static_cast(scale)); + output_grad_data[output_idx], static_cast(scale), + input_grad_data + input_idx); } } } @@ -404,12 +404,11 @@ class Pool3dGradFunctor { template class MaxPool3dGradFunctor { public: - void operator()(const platform::CPUDeviceContext& context, - const framework::Tensor& input, - const framework::Tensor& output, - const framework::Tensor& output_grad, std::vector& ksize, - std::vector& strides, std::vector& paddings, - framework::Tensor* input_grad) { + void operator()( + const platform::CPUDeviceContext& context, const framework::Tensor& input, + const framework::Tensor& output, const framework::Tensor& output_grad, + const std::vector& ksize, const std::vector& strides, + const std::vector& paddings, framework::Tensor* input_grad) { const int batch_size = input.dims()[0]; const int input_depth = input.dims()[2]; const int input_height = input.dims()[3]; @@ -510,9 +509,10 @@ template class MaxPool2dWithIndexFunctor { public: void operator()(const platform::CPUDeviceContext& context, - const framework::Tensor& input, std::vector& ksize, - std::vector& strides, std::vector& paddings, - framework::Tensor* output, framework::Tensor* mask) { + const framework::Tensor& input, const std::vector& ksize, + const std::vector& strides, + const std::vector& paddings, framework::Tensor* output, + framework::Tensor* mask) { const int batch_size = input.dims()[0]; const int input_height = input.dims()[2]; const int input_width = input.dims()[3]; @@ -576,8 +576,9 @@ class MaxPool2dWithIndexGradFunctor { public: void operator()(const platform::CPUDeviceContext& context, const framework::Tensor& output_grad, - const framework::Tensor& mask, std::vector& ksize, - std::vector& strides, std::vector& paddings, + const framework::Tensor& mask, const std::vector& ksize, + const std::vector& strides, + const std::vector& paddings, framework::Tensor* input_grad) { const int batch_size = input_grad->dims()[0]; const int input_height = input_grad->dims()[2]; @@ -628,9 +629,10 @@ template class MaxPool3dWithIndexFunctor { public: void operator()(const platform::CPUDeviceContext& context, - const framework::Tensor& input, std::vector& ksize, - std::vector& strides, std::vector& paddings, - framework::Tensor* output, framework::Tensor* mask) { + const framework::Tensor& input, const std::vector& ksize, + const std::vector& strides, + const std::vector& paddings, framework::Tensor* output, + framework::Tensor* mask) { const int batch_size = input.dims()[0]; const int input_depth = input.dims()[2]; const int input_height = input.dims()[3]; @@ -708,8 +710,9 @@ class MaxPool3dWithIndexGradFunctor { public: void operator()(const platform::CPUDeviceContext& context, const framework::Tensor& output_grad, - const framework::Tensor& mask, std::vector& ksize, - std::vector& strides, std::vector& paddings, + const framework::Tensor& mask, const std::vector& ksize, + const std::vector& strides, + const std::vector& paddings, framework::Tensor* input_grad) { const int batch_size = input_grad->dims()[0]; const int input_depth = input_grad->dims()[2]; diff --git a/paddle/fluid/operators/math/pooling.cu b/paddle/fluid/operators/math/pooling.cu index 274263c69c535249fceee11075c5948b1fc34358..b1c76350d1724629bae175abf47e6671a1532242 100644 --- a/paddle/fluid/operators/math/pooling.cu +++ b/paddle/fluid/operators/math/pooling.cu @@ -12,8 +12,10 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ +#include +#include #include "paddle/fluid/operators/math/pooling.h" -#include "paddle/fluid/platform/cuda_helper.h" +#include "paddle/fluid/platform/cuda_primitives.h" namespace paddle { namespace operators { @@ -47,11 +49,11 @@ __global__ void KernelPool2D(const int nthreads, const T* input_data, T ele = pool_process.initial(); for (int h = hstart; h < hend; ++h) { for (int w = wstart; w < wend; ++w) { - pool_process.compute(ele, input_data[h * input_width + w]); + pool_process.compute(input_data[h * input_width + w], &ele); } } int pool_size = (hend - hstart) * (wend - wstart); - pool_process.finalize(ele, (static_cast(pool_size))); + pool_process.finalize(static_cast(pool_size), &ele); output_data[index] = ele; } } @@ -96,8 +98,8 @@ __global__ void KernelPool2DGrad( int pool_size = (hend - hstart) * (wend - wstart); int output_sub_idx = ph * output_width + pw; pool_process.compute(input, output_data[output_sub_idx], - output_grad[output_sub_idx], gradient, - static_cast(1.0 / pool_size)); + output_grad[output_sub_idx], + static_cast(1.0 / pool_size), &gradient); } } input_grad[index] = gradient; @@ -158,9 +160,10 @@ template class Pool2dFunctor { public: void operator()(const platform::CUDADeviceContext& context, - const framework::Tensor& input, std::vector& ksize, - std::vector& strides, std::vector& paddings, - PoolProcess pool_process, framework::Tensor* output) { + const framework::Tensor& input, const std::vector& ksize, + const std::vector& strides, + const std::vector& paddings, PoolProcess pool_process, + framework::Tensor* output) { const int batch_size = input.dims()[0]; const int input_channels = input.dims()[1]; const int input_height = input.dims()[2]; @@ -201,9 +204,11 @@ class Pool2dGradFunctor { void operator()(const platform::CUDADeviceContext& context, const framework::Tensor& input, const framework::Tensor& output, - const framework::Tensor& output_grad, std::vector& ksize, - std::vector& strides, std::vector& paddings, - PoolProcess pool_process, framework::Tensor* input_grad) { + const framework::Tensor& output_grad, + const std::vector& ksize, + const std::vector& strides, + const std::vector& paddings, PoolProcess pool_process, + framework::Tensor* input_grad) { const int batch_size = input.dims()[0]; const int input_channels = input.dims()[1]; const int input_height = input.dims()[2]; @@ -246,8 +251,10 @@ class MaxPool2dGradFunctor { void operator()(const platform::CUDADeviceContext& context, const framework::Tensor& input, const framework::Tensor& output, - const framework::Tensor& output_grad, std::vector& ksize, - std::vector& strides, std::vector& paddings, + const framework::Tensor& output_grad, + const std::vector& ksize, + const std::vector& strides, + const std::vector& paddings, framework::Tensor* input_grad) { const int batch_size = input.dims()[0]; const int input_channels = input.dims()[1]; @@ -340,12 +347,12 @@ __global__ void KernelPool3D(const int nthreads, const T* input_data, for (int h = hstart; h < hend; ++h) { for (int w = wstart; w < wend; ++w) { pool_process.compute( - ele, input_data[(d * input_height + h) * input_width + w]); + input_data[(d * input_height + h) * input_width + w], &ele); } } } int pool_size = (dend - dstart) * (hend - hstart) * (wend - wstart); - pool_process.finalize(ele, static_cast(pool_size)); + pool_process.finalize(static_cast(pool_size), &ele); output_data[index] = ele; } } @@ -405,8 +412,8 @@ __global__ void KernelPool3DGrad( int pool_size = (dend - dstart) * (hend - hstart) * (wend - wstart); int output_sub_idx = (pd * output_height + ph) * output_width + pw; pool_process.compute(input, output_data[output_sub_idx], - output_grad[output_sub_idx], gradient, - static_cast(1.0 / pool_size)); + output_grad[output_sub_idx], + static_cast(1.0 / pool_size), &gradient); } } } @@ -474,9 +481,10 @@ template class Pool3dFunctor { public: void operator()(const platform::CUDADeviceContext& context, - const framework::Tensor& input, std::vector& ksize, - std::vector& strides, std::vector& paddings, - PoolProcess pool_process, framework::Tensor* output) { + const framework::Tensor& input, const std::vector& ksize, + const std::vector& strides, + const std::vector& paddings, PoolProcess pool_process, + framework::Tensor* output) { const int batch_size = input.dims()[0]; const int input_channels = input.dims()[1]; const int input_depth = input.dims()[2]; @@ -525,9 +533,11 @@ class Pool3dGradFunctor { void operator()(const platform::CUDADeviceContext& context, const framework::Tensor& input, const framework::Tensor& output, - const framework::Tensor& output_grad, std::vector& ksize, - std::vector& strides, std::vector& paddings, - PoolProcess pool_process, framework::Tensor* input_grad) { + const framework::Tensor& output_grad, + const std::vector& ksize, + const std::vector& strides, + const std::vector& paddings, PoolProcess pool_process, + framework::Tensor* input_grad) { const int batch_size = input.dims()[0]; const int input_channels = input.dims()[1]; const int input_depth = input.dims()[2]; @@ -578,8 +588,10 @@ class MaxPool3dGradFunctor { void operator()(const platform::CUDADeviceContext& context, const framework::Tensor& input, const framework::Tensor& output, - const framework::Tensor& output_grad, std::vector& ksize, - std::vector& strides, std::vector& paddings, + const framework::Tensor& output_grad, + const std::vector& ksize, + const std::vector& strides, + const std::vector& paddings, framework::Tensor* input_grad) { const int batch_size = input.dims()[0]; const int input_channels = input.dims()[1]; @@ -736,9 +748,10 @@ template class MaxPool2dWithIndexFunctor { public: void operator()(const platform::CUDADeviceContext& context, - const framework::Tensor& input, std::vector& ksize, - std::vector& strides, std::vector& paddings, - framework::Tensor* output, framework::Tensor* mask) { + const framework::Tensor& input, const std::vector& ksize, + const std::vector& strides, + const std::vector& paddings, framework::Tensor* output, + framework::Tensor* mask) { const int batch_size = input.dims()[0]; const int input_channels = input.dims()[1]; const int input_height = input.dims()[2]; @@ -779,8 +792,9 @@ class MaxPool2dWithIndexGradFunctor { public: void operator()(const platform::CUDADeviceContext& context, const framework::Tensor& output_grad, - const framework::Tensor& mask, std::vector& ksize, - std::vector& strides, std::vector& paddings, + const framework::Tensor& mask, const std::vector& ksize, + const std::vector& strides, + const std::vector& paddings, framework::Tensor* input_grad) { const int batch_size = input_grad->dims()[0]; const int input_channels = input_grad->dims()[1]; @@ -937,9 +951,10 @@ template class MaxPool3dWithIndexFunctor { public: void operator()(const platform::CUDADeviceContext& context, - const framework::Tensor& input, std::vector& ksize, - std::vector& strides, std::vector& paddings, - framework::Tensor* output, framework::Tensor* mask) { + const framework::Tensor& input, const std::vector& ksize, + const std::vector& strides, + const std::vector& paddings, framework::Tensor* output, + framework::Tensor* mask) { const int batch_size = input.dims()[0]; const int input_channels = input.dims()[1]; const int input_depth = input.dims()[2]; @@ -987,8 +1002,9 @@ class MaxPool3dWithIndexGradFunctor { public: void operator()(const platform::CUDADeviceContext& context, const framework::Tensor& output_grad, - const framework::Tensor& mask, std::vector& ksize, - std::vector& strides, std::vector& paddings, + const framework::Tensor& mask, const std::vector& ksize, + const std::vector& strides, + const std::vector& paddings, framework::Tensor* input_grad) { const int batch_size = input_grad->dims()[0]; const int input_channels = input_grad->dims()[1]; diff --git a/paddle/fluid/operators/math/pooling.h b/paddle/fluid/operators/math/pooling.h index 74cb42f0d02086a6776b22d57832757ae3ffc470..2538d739cce95d1b2fc5b3f905af5e6d94cf7af5 100644 --- a/paddle/fluid/operators/math/pooling.h +++ b/paddle/fluid/operators/math/pooling.h @@ -13,6 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once +#include #include "paddle/fluid/framework/eigen.h" #include "paddle/fluid/framework/tensor.h" #include "paddle/fluid/platform/device_context.h" @@ -23,8 +24,8 @@ namespace operators { namespace math { #define FLT_MAX \ - __FLT_MAX__ // It might need to be placed in another file, but I'm still - // wondering where to put it. + __FLT_MAX__ // TODO(zcd) :It might need to be placed in another file, but I'm + // still wondering where to put it. /* * \brief Extracting simple operations from pooling. @@ -40,33 +41,33 @@ template class MaxPool { public: DEVICE inline T initial() { return static_cast(-FLT_MAX); } - DEVICE inline void compute(T& y, const T& x) { y = y > x ? y : x; } - DEVICE inline void finalize(T& y, const T& pool_field) {} + DEVICE inline void compute(const T& x, T* y) { *y = *y > x ? *y : x; } + DEVICE inline void finalize(const T& pool_field, T* y) {} }; template class AvgPool { public: DEVICE inline T initial() { return static_cast(0); } - DEVICE inline void compute(T& y, const T& x) { y += x; } - DEVICE inline void finalize(T& y, const T& pool_field) { y /= pool_field; } + DEVICE inline void compute(const T& x, T* y) { *y += x; } + DEVICE inline void finalize(const T& pool_field, T* y) { *y /= pool_field; } }; template class MaxPoolGrad { public: - DEVICE inline void compute(const T& x, const T& y, const T& dy, T& dx, - T scale) { - dx += dy * (x == y); + DEVICE inline void compute(const T& x, const T& y, const T& dy, T scale, + T* dx) { + *dx += dy * (x == y); } }; template class AvgPoolGrad { public: - DEVICE inline void compute(const T& x, const T& y, const T& dy, T& dx, - T scale) { - dx += (scale * dy); + DEVICE inline void compute(const T& x, const T& y, const T& dy, T scale, + T* dx) { + *dx += (scale * dy); } }; @@ -88,8 +89,9 @@ template class Pool2dFunctor { public: void operator()(const DeviceContext& context, const framework::Tensor& input, - std::vector& ksize, std::vector& strides, - std::vector& paddings, PoolProcess pool_compute, + const std::vector& ksize, + const std::vector& strides, + const std::vector& paddings, PoolProcess pool_compute, framework::Tensor* output); }; @@ -98,9 +100,11 @@ class Pool2dGradFunctor { public: void operator()(const DeviceContext& context, const framework::Tensor& input, const framework::Tensor& output, - const framework::Tensor& output_grad, std::vector& ksize, - std::vector& strides, std::vector& paddings, - PoolProcess pool_compute, framework::Tensor* input_grad); + const framework::Tensor& output_grad, + const std::vector& ksize, + const std::vector& strides, + const std::vector& paddings, PoolProcess pool_compute, + framework::Tensor* input_grad); }; template @@ -108,8 +112,10 @@ class MaxPool2dGradFunctor { public: void operator()(const DeviceContext& context, const framework::Tensor& input, const framework::Tensor& output, - const framework::Tensor& output_grad, std::vector& ksize, - std::vector& strides, std::vector& paddings, + const framework::Tensor& output_grad, + const std::vector& ksize, + const std::vector& strides, + const std::vector& paddings, framework::Tensor* input_grad); }; @@ -117,8 +123,9 @@ template class Pool3dFunctor { public: void operator()(const DeviceContext& context, const framework::Tensor& input, - std::vector& ksize, std::vector& strides, - std::vector& paddings, PoolProcess pool_compute, + const std::vector& ksize, + const std::vector& strides, + const std::vector& paddings, PoolProcess pool_compute, framework::Tensor* output); }; @@ -127,9 +134,11 @@ class Pool3dGradFunctor { public: void operator()(const DeviceContext& context, const framework::Tensor& input, const framework::Tensor& output, - const framework::Tensor& output_grad, std::vector& ksize, - std::vector& strides, std::vector& paddings, - PoolProcess pool_compute, framework::Tensor* input_grad); + const framework::Tensor& output_grad, + const std::vector& ksize, + const std::vector& strides, + const std::vector& paddings, PoolProcess pool_compute, + framework::Tensor* input_grad); }; template @@ -137,8 +146,10 @@ class MaxPool3dGradFunctor { public: void operator()(const DeviceContext& context, const framework::Tensor& input, const framework::Tensor& output, - const framework::Tensor& output_grad, std::vector& ksize, - std::vector& strides, std::vector& paddings, + const framework::Tensor& output_grad, + const std::vector& ksize, + const std::vector& strides, + const std::vector& paddings, framework::Tensor* input_grad); }; @@ -153,8 +164,9 @@ template class MaxPool2dWithIndexFunctor { public: void operator()(const DeviceContext& context, const framework::Tensor& input, - std::vector& ksize, std::vector& strides, - std::vector& paddings, framework::Tensor* output, + const std::vector& ksize, + const std::vector& strides, + const std::vector& paddings, framework::Tensor* output, framework::Tensor* mask); }; @@ -163,8 +175,9 @@ class MaxPool2dWithIndexGradFunctor { public: void operator()(const DeviceContext& context, const framework::Tensor& output_grad, - const framework::Tensor& mask, std::vector& ksize, - std::vector& strides, std::vector& paddings, + const framework::Tensor& mask, const std::vector& ksize, + const std::vector& strides, + const std::vector& paddings, framework::Tensor* input_grad); }; @@ -172,8 +185,9 @@ template class MaxPool3dWithIndexFunctor { public: void operator()(const DeviceContext& context, const framework::Tensor& input, - std::vector& ksize, std::vector& strides, - std::vector& paddings, framework::Tensor* output, + const std::vector& ksize, + const std::vector& strides, + const std::vector& paddings, framework::Tensor* output, framework::Tensor* mask); }; @@ -182,8 +196,9 @@ class MaxPool3dWithIndexGradFunctor { public: void operator()(const DeviceContext& context, const framework::Tensor& output_grad, - const framework::Tensor& mask, std::vector& ksize, - std::vector& strides, std::vector& paddings, + const framework::Tensor& mask, const std::vector& ksize, + const std::vector& strides, + const std::vector& paddings, framework::Tensor* input_grad); }; diff --git a/paddle/fluid/operators/math/sampler.h b/paddle/fluid/operators/math/sampler.h index 9d6a6c28c4304019d0347a30be605e1374c169ee..b82691f269c5d0f267ca98c78646efe9b26f0b34 100644 --- a/paddle/fluid/operators/math/sampler.h +++ b/paddle/fluid/operators/math/sampler.h @@ -13,9 +13,9 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once +#include #include #include -typedef long int64; namespace paddle { namespace operators { namespace math { @@ -27,25 +27,25 @@ namespace math { */ class Sampler { public: - explicit Sampler(int64 range) : range_(range) { + explicit Sampler(int64_t range) : range_(range) { PADDLE_ENFORCE_GT(range, 0); std::random_device r; seed_ = r(); } - explicit Sampler(int64 range, unsigned int seed) + explicit Sampler(int64_t range, unsigned int seed) : range_(range), seed_(seed) { PADDLE_ENFORCE_GT(range, 0); } virtual ~Sampler(); // Sample a single value - virtual int64 Sample() const = 0; + virtual int64_t Sample() const = 0; // The probability that a single call to Sample() returns the given value. - virtual float Probability(int64 value) const = 0; + virtual float Probability(int64_t value) const = 0; - int64 range() { return range_; }; + int64 range() { return range_; } protected: - const int64 range_; + const int64_t range_; unsigned int seed_; }; @@ -56,15 +56,15 @@ class Sampler { */ class UniformSampler : public Sampler { public: - explicit UniformSampler(int64 range); + explicit UniformSampler(int64_t range); - explicit UniformSampler(int64 range, unsigned int seed); + explicit UniformSampler(int64_t range, unsigned int seed); ~UniformSampler() override {} int64 Sample() const override; - float Probability(int64 value) const override; + float Probability(int64_t value) const override; private: const float inv_range_; @@ -79,15 +79,15 @@ class UniformSampler : public Sampler { */ class LogUniformSampler : public Sampler { public: - explicit LogUniformSampler(int64 range); + explicit LogUniformSampler(int64_t range); - explicit LogUniformSampler(int64 range, unsigned int seed); + explicit LogUniformSampler(int64_t range, unsigned int seed); ~LogUniformSampler() override {} int64 Sample() const override; - float Probability(int64 value) const override; + float Probability(int64_t value) const override; private: const float log_range_; @@ -95,6 +95,6 @@ class LogUniformSampler : public Sampler { std::shared_ptr> dist_; }; -} // math +} // namespace math } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/math/selected_rows_functor.cc b/paddle/fluid/operators/math/selected_rows_functor.cc index 5da3d15277cff8b413a116819b17f50061632b5d..a830dc5250a6aea7e622da4046b512d0c7c5d6f9 100644 --- a/paddle/fluid/operators/math/selected_rows_functor.cc +++ b/paddle/fluid/operators/math/selected_rows_functor.cc @@ -13,6 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #include +#include #include "paddle/fluid/operators/math/math_function.h" #include "paddle/fluid/operators/math/selected_rows_functor.h" diff --git a/paddle/fluid/operators/math/selected_rows_functor.cu b/paddle/fluid/operators/math/selected_rows_functor.cu index 5d78fd9d213556204d56087128dc84fe6a91e97d..a92762c7fea865fad2c7784736cce93a8af21892 100644 --- a/paddle/fluid/operators/math/selected_rows_functor.cu +++ b/paddle/fluid/operators/math/selected_rows_functor.cu @@ -13,10 +13,11 @@ See the License for the specific language governing permissions and limitations under the License. */ #include +#include #include "paddle/fluid/operators/math/math_function.h" #include "paddle/fluid/operators/math/selected_rows_functor.h" -#include "paddle/fluid/platform/cuda_helper.h" +#include "paddle/fluid/platform/cuda_primitives.h" namespace paddle { namespace operators { diff --git a/paddle/fluid/operators/math/selected_rows_functor_test.cu b/paddle/fluid/operators/math/selected_rows_functor_test.cu index 942d9b13fc1a8f578da779351be9ba9de7fcce33..e89b27855bdeba3a5189feff94eb063ddfb9b9b8 100644 --- a/paddle/fluid/operators/math/selected_rows_functor_test.cu +++ b/paddle/fluid/operators/math/selected_rows_functor_test.cu @@ -12,43 +12,52 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ +#include #include "gtest/gtest.h" #include "paddle/fluid/operators/math/math_function.h" #include "paddle/fluid/operators/math/selected_rows_functor.h" TEST(selected_rows_functor, gpu_add) { - using namespace paddle::framework; - using namespace paddle::platform; - using namespace paddle::operators::math; - - CUDAPlace gpu_place(0); - CPUPlace cpu_place; - CUDADeviceContext ctx(gpu_place); - SetConstant functor; + paddle::platform::CUDAPlace gpu_place(0); + paddle::platform::CPUPlace cpu_place; + paddle::platform::CUDADeviceContext ctx(gpu_place); + paddle::operators::math::SetConstant + functor; int64_t height = 10; int64_t row_numel = 10; std::vector rows1{0, 4, 7}; - std::unique_ptr selected_rows1{new SelectedRows(rows1, height)}; + std::unique_ptr selected_rows1{ + new paddle::framework::SelectedRows(rows1, height)}; auto* in1_value = selected_rows1->mutable_value(); in1_value->mutable_data( - make_ddim({static_cast(rows1.size()), row_numel}), gpu_place); + paddle::framework::make_ddim( + {static_cast(rows1.size()), row_numel}), + gpu_place); functor(ctx, in1_value, 1.0); std::vector rows2{0, 5, 7, 9}; - std::unique_ptr selected_rows2{new SelectedRows(rows2, height)}; + std::unique_ptr selected_rows2{ + new paddle::framework::SelectedRows(rows2, height)}; auto* in2_value = selected_rows2->mutable_value(); in2_value->mutable_data( - make_ddim({static_cast(rows2.size()), row_numel}), gpu_place); + paddle::framework::make_ddim( + {static_cast(rows2.size()), row_numel}), + gpu_place); functor(ctx, in2_value, 2.0); - std::unique_ptr output{new SelectedRows()}; + std::unique_ptr output{ + new paddle::framework::SelectedRows()}; auto* out_value = output->mutable_value(); - // simplely concat two SelectedRows - out_value->mutable_data(make_ddim({7, 10}), gpu_place); + // simply concat two SelectedRows + out_value->mutable_data(paddle::framework::make_ddim({7, 10}), + gpu_place); - SelectedRowsAdd add_functor; + paddle::operators::math::SelectedRowsAdd + add_functor; add_functor(ctx, *selected_rows1, *selected_rows2, output.get()); auto out_height = output->height(); @@ -66,8 +75,8 @@ TEST(selected_rows_functor, gpu_add) { EXPECT_EQ(out_rows[5], 7); EXPECT_EQ(out_rows[6], 9); - Tensor out_cpu; - TensorCopy(*out_value, cpu_place, ctx, &out_cpu); + paddle::framework::Tensor out_cpu; + paddle::framework::TensorCopy(*out_value, cpu_place, ctx, &out_cpu); ctx.Wait(); auto* out_cpu_data = out_cpu.data(); @@ -83,18 +92,24 @@ TEST(selected_rows_functor, gpu_add) { EXPECT_EQ(out_cpu_data[5 * row_numel + 7], 2.0); EXPECT_EQ(out_cpu_data[6 * row_numel + 9], 2.0); - std::unique_ptr tensor1{new Tensor()}; - tensor1->mutable_data(make_ddim({height, row_numel}), gpu_place); + std::unique_ptr tensor1{ + new paddle::framework::Tensor()}; + tensor1->mutable_data( + paddle::framework::make_ddim({height, row_numel}), gpu_place); functor(ctx, tensor1.get(), 3.0); - std::unique_ptr tensor2{new Tensor()}; - tensor2->mutable_data(make_ddim({height, row_numel}), gpu_place); + std::unique_ptr tensor2{ + new paddle::framework::Tensor()}; + tensor2->mutable_data( + paddle::framework::make_ddim({height, row_numel}), gpu_place); - SelectedRowsAddTensor add_tensor_functor; + paddle::operators::math::SelectedRowsAddTensor< + paddle::platform::CUDADeviceContext, float> + add_tensor_functor; add_tensor_functor(ctx, *output, *tensor1, tensor2.get()); - Tensor tensor2_cpu; - TensorCopy(*tensor2, cpu_place, ctx, &tensor2_cpu); + paddle::framework::Tensor tensor2_cpu; + paddle::framework::TensorCopy(*tensor2, cpu_place, ctx, &tensor2_cpu); ctx.Wait(); auto* tensor2_cpu_data = tensor2_cpu.data(); @@ -115,39 +130,47 @@ TEST(selected_rows_functor, gpu_add) { } TEST(selected_rows_functor, gpu_add_to) { - using namespace paddle::framework; - using namespace paddle::platform; - using namespace paddle::operators::math; - - CUDAPlace gpu_place(0); - CPUPlace cpu_place; - CUDADeviceContext ctx(gpu_place); - SetConstant functor; + paddle::platform::CUDAPlace gpu_place(0); + paddle::platform::CPUPlace cpu_place; + paddle::platform::CUDADeviceContext ctx(gpu_place); + paddle::operators::math::SetConstant + functor; int64_t height = 10; int64_t row_numel = 10; std::vector rows1{0, 4, 7}; - std::unique_ptr selected_rows1{new SelectedRows(rows1, height)}; + std::unique_ptr selected_rows1{ + new paddle::framework::SelectedRows(rows1, height)}; auto* in1_value = selected_rows1->mutable_value(); in1_value->mutable_data( - make_ddim({static_cast(rows1.size()), row_numel}), gpu_place); + paddle::framework::make_ddim( + {static_cast(rows1.size()), row_numel}), + gpu_place); functor(ctx, in1_value, 1.0); std::vector rows2{0, 5, 7, 9}; - std::unique_ptr selected_rows2{new SelectedRows(rows2, height)}; + std::unique_ptr selected_rows2{ + new paddle::framework::SelectedRows(rows2, height)}; auto* in2_value = selected_rows2->mutable_value(); in2_value->mutable_data( - make_ddim({static_cast(rows2.size()), row_numel}), gpu_place); + paddle::framework::make_ddim( + {static_cast(rows2.size()), row_numel}), + gpu_place); functor(ctx, in2_value, 2.0); - std::unique_ptr output{new SelectedRows()}; + std::unique_ptr output{ + new paddle::framework::SelectedRows()}; output->set_height(height); auto* out_value = output->mutable_value(); - // simplely concat two SelectedRows - out_value->mutable_data(make_ddim({7, 10}), gpu_place); + // simply concat two SelectedRows + out_value->mutable_data(paddle::framework::make_ddim({7, 10}), + gpu_place); - SelectedRowsAddTo add_to_functor; + paddle::operators::math::SelectedRowsAddTo< + paddle::platform::CUDADeviceContext, float> + add_to_functor; add_to_functor(ctx, *selected_rows1, 0, output.get()); add_to_functor(ctx, *selected_rows2, in1_value->numel(), output.get()); @@ -166,8 +189,8 @@ TEST(selected_rows_functor, gpu_add_to) { EXPECT_EQ(out_rows[5], 7); EXPECT_EQ(out_rows[6], 9); - Tensor out_cpu; - TensorCopy(*out_value, cpu_place, ctx, &out_cpu); + paddle::framework::Tensor out_cpu; + paddle::framework::TensorCopy(*out_value, cpu_place, ctx, &out_cpu); ctx.Wait(); auto* out_cpu_data = out_cpu.data(); @@ -183,15 +206,19 @@ TEST(selected_rows_functor, gpu_add_to) { EXPECT_EQ(out_cpu_data[5 * row_numel + 7], 2.0); EXPECT_EQ(out_cpu_data[6 * row_numel + 9], 2.0); - std::unique_ptr tensor1{new Tensor()}; - tensor1->mutable_data(make_ddim({height, row_numel}), gpu_place); + std::unique_ptr tensor1{ + new paddle::framework::Tensor()}; + tensor1->mutable_data( + paddle::framework::make_ddim({height, row_numel}), gpu_place); functor(ctx, tensor1.get(), 3.0); - SelectedRowsAddToTensor add_to_tensor_functor; + paddle::operators::math::SelectedRowsAddToTensor< + paddle::platform::CUDADeviceContext, float> + add_to_tensor_functor; add_to_tensor_functor(ctx, *output, tensor1.get()); - Tensor tensor1_cpu; - TensorCopy(*tensor1, cpu_place, ctx, &tensor1_cpu); + paddle::framework::Tensor tensor1_cpu; + paddle::framework::TensorCopy(*tensor1, cpu_place, ctx, &tensor1_cpu); ctx.Wait(); auto* tensor1_cpu_data = tensor1_cpu.data(); diff --git a/paddle/fluid/operators/math/sequence2batch.cc b/paddle/fluid/operators/math/sequence2batch.cc index 8899abff360ea867872d3433722cdb37ef358500..b546b8728217ed6013247555dcd5d7180ddeae74 100644 --- a/paddle/fluid/operators/math/sequence2batch.cc +++ b/paddle/fluid/operators/math/sequence2batch.cc @@ -23,11 +23,11 @@ class CopyMatrixRowsFunctor { public: void operator()(const platform::CPUDeviceContext& context, const framework::Tensor& src, - framework::Vector index_lod, framework::Tensor& dst, + framework::Vector index_lod, framework::Tensor* dst, bool is_src_index) { size_t* index = index_lod.data(); auto src_dims = src.dims(); - auto dst_dims = dst.dims(); + auto dst_dims = dst->dims(); PADDLE_ENFORCE_EQ(src_dims.size(), 2UL, "The src must be matrix with rank 2."); PADDLE_ENFORCE_EQ(dst_dims.size(), 2UL, @@ -37,7 +37,7 @@ class CopyMatrixRowsFunctor { auto height = dst_dims[0]; auto width = dst_dims[1]; auto* src_data = src.data(); - auto* dst_data = dst.data(); + auto* dst_data = dst->data(); for (int i = 0; i < height; ++i) { if (is_src_index) { memcpy(dst_data + i * width, src_data + index[i] * width, diff --git a/paddle/fluid/operators/math/sequence2batch.cu b/paddle/fluid/operators/math/sequence2batch.cu index 3185f10d4180437ab5a3f78df8583613edd9ed43..be73adfc0cbe37ed8831b5ad34e66bc95e342e9d 100644 --- a/paddle/fluid/operators/math/sequence2batch.cu +++ b/paddle/fluid/operators/math/sequence2batch.cu @@ -43,10 +43,10 @@ class CopyMatrixRowsFunctor { public: void operator()(const platform::CUDADeviceContext& context, const framework::Tensor& src, - framework::Vector index_lod, framework::Tensor& dst, + framework::Vector index_lod, framework::Tensor* dst, bool is_src_index) { auto src_dims = src.dims(); - auto dst_dims = dst.dims(); + auto dst_dims = dst->dims(); PADDLE_ENFORCE_EQ(src_dims.size(), 2, "The src must be matrix with rank 2."); PADDLE_ENFORCE_EQ(dst_dims.size(), 2, @@ -56,7 +56,7 @@ class CopyMatrixRowsFunctor { auto height = dst_dims[0]; auto width = dst_dims[1]; auto* src_data = src.data(); - auto* dst_data = dst.data(); + auto* dst_data = dst->data(); dim3 threads(128, 8); dim3 grid(8, 1); diff --git a/paddle/fluid/operators/math/sequence2batch.h b/paddle/fluid/operators/math/sequence2batch.h index e78aafd37d1dda91a035f3ed850537e80f188cb2..62e6307ae9f4236a38c49daaf09fc05c54268159 100644 --- a/paddle/fluid/operators/math/sequence2batch.h +++ b/paddle/fluid/operators/math/sequence2batch.h @@ -13,6 +13,8 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once +#include +#include #include "paddle/fluid/framework/eigen.h" #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/tensor.h" @@ -35,7 +37,7 @@ class CopyMatrixRowsFunctor { // copy the input src to the indexed rows of output dst. // The indexed rows are based on the input index. void operator()(const DeviceContext& context, const framework::Tensor& src, - framework::Vector index_lod, framework::Tensor& dst, + framework::Vector index_lod, framework::Tensor* dst, bool is_src_index); }; @@ -58,22 +60,26 @@ class LoDTensor2BatchFunctor { public: void operator()(const DeviceContext& context, const framework::LoDTensor& lod_tensor, - framework::LoDTensor& batch, bool is_cal_batch_lod, + framework::LoDTensor* batch, bool is_cal_batch_lod, bool is_reverse = false) const { if (!is_cal_batch_lod) { - auto lods = batch.lod(); - PADDLE_ENFORCE_GT(lods.size(), 2UL); - PADDLE_ENFORCE_EQ(lods[1].size(), - static_cast(lod_tensor.dims()[0])); + auto lods = batch->lod(); + PADDLE_ENFORCE_GT(lods.size(), 2UL, + "The LoD of LoDTensor should inlcude at least 2-level " + "sequence information."); + PADDLE_ENFORCE_EQ( + lods[1].size(), static_cast(lod_tensor.dims()[0]), + "The LoD information should be consistent with the dims."); CopyMatrixRowsFunctor to_batch; to_batch(context, lod_tensor, lods[1], batch, true); return; } auto lods = lod_tensor.lod(); - auto lod = lods[0]; PADDLE_ENFORCE_EQ(lods.size(), 1UL, "Only support one level sequence now."); + auto lod = lods[0]; + std::vector seq_info; for (size_t seq_id = 0; seq_id < lod.size() - 1; ++seq_id) { int length = lod[seq_id + 1] - lod[seq_id]; @@ -141,7 +147,7 @@ class LoDTensor2BatchFunctor { for (size_t i = 0; i < seq_info.size(); ++i) { seq_order[i] = seq_info[i].seq_idx; } - batch.set_lod(batch_lods); + batch->set_lod(batch_lods); CopyMatrixRowsFunctor to_batch; to_batch(context, lod_tensor, batch_lods[1], batch, true); @@ -153,11 +159,14 @@ class Batch2LoDTensorFunctor { public: void operator()(const DeviceContext& context, const framework::LoDTensor& batch, - framework::LoDTensor& lod_tensor) const { + framework::LoDTensor* lod_tensor) const { auto in_lod = batch.lod(); - PADDLE_ENFORCE_GT(in_lod.size(), 2UL); - PADDLE_ENFORCE_EQ(in_lod[1].size(), - static_cast(lod_tensor.dims()[0])); + PADDLE_ENFORCE_GT(in_lod.size(), 2UL, + "The LoD of LoDTensor should inlcude at least 2-level " + "sequence information."); + PADDLE_ENFORCE_EQ( + in_lod[1].size(), static_cast(lod_tensor->dims()[0]), + "The LoD information should be consistent with the dims."); CopyMatrixRowsFunctor to_seq; to_seq(context, batch, in_lod[1], lod_tensor, false); } diff --git a/paddle/fluid/operators/math/sequence_padding.cc b/paddle/fluid/operators/math/sequence_padding.cc index 38bd3b99758555a24b3b8eb0de06cca8e424fcb2..d63c6c4ed55331235188c1c750468d4e75b9b7f2 100644 --- a/paddle/fluid/operators/math/sequence_padding.cc +++ b/paddle/fluid/operators/math/sequence_padding.cc @@ -22,7 +22,7 @@ template class PaddingLoDTensorFunctor { public: void operator()(const platform::CPUDeviceContext& context, - const framework::LoDTensor& seq, framework::Tensor& padding, + const framework::LoDTensor& seq, framework::Tensor* padding, bool norm_by_times) { auto lod = seq.lod(); PADDLE_ENFORCE_GT(lod.size(), 0UL, @@ -37,7 +37,7 @@ class PaddingLoDTensorFunctor { "The first dimension of LoDTensor seq should be " "equal to the sum of all sequences's length."); - auto padding_dims = padding.dims(); + auto padding_dims = padding->dims(); PADDLE_ENFORCE_EQ(padding_dims.size(), 3UL, "The input padding should be a 3-D Tensor of shape " "[max_sequence_length, num_sequences, sequence_width]."); @@ -58,7 +58,7 @@ class PaddingLoDTensorFunctor { "width of sequence in LoDTensor seq."); const T* seq_data = seq.data(); - T* padding_data = padding.data(); + T* padding_data = padding->data(); for (int64_t i = 0; i < max_sequence_length; ++i) { for (int64_t j = 0; j < num_sequences; ++j) { int64_t start_pos = abs_offset_lod[level][j]; @@ -84,16 +84,16 @@ template class UnpaddingLoDTensorFunctor { public: void operator()(const platform::CPUDeviceContext& context, - framework::LoDTensor& seq, const framework::Tensor& padding, + framework::LoDTensor* seq, const framework::Tensor& padding, bool norm_by_times) { - auto lod = seq.lod(); + auto lod = seq->lod(); PADDLE_ENFORCE_GT(lod.size(), 0UL, "The LoD of LoDTensor seq should not be null."); const size_t level = 0; framework::LoD abs_offset_lod = framework::ToAbsOffset(lod); - auto seq_dims = seq.dims(); + auto seq_dims = seq->dims(); PADDLE_ENFORCE_EQ(seq_dims[0], static_cast(abs_offset_lod[level].back()), "The first dimension of LoDTensor seq should be " @@ -114,13 +114,13 @@ class UnpaddingLoDTensorFunctor { "The second dimension of Tensor padding should be " "the number of sequences in LoDTensor seq."); - const int64_t sequence_width = seq.numel() / seq_dims[0]; + const int64_t sequence_width = seq->numel() / seq_dims[0]; PADDLE_ENFORCE_EQ(padding_dims[2], sequence_width, "The third dimension of Tensor padding should be the " "width of sequence in LoDTensor seq."); const T* padding_data = padding.data(); - T* seq_data = seq.data(); + T* seq_data = seq->data(); for (int64_t i = 0; i < num_sequences; ++i) { int64_t start_pos = abs_offset_lod[level][i]; int64_t sequence_length = abs_offset_lod[level][i + 1] - start_pos; diff --git a/paddle/fluid/operators/math/sequence_padding.cu b/paddle/fluid/operators/math/sequence_padding.cu index c044e6fc32bab8f72a0dce45b4abdb1174a0d72f..0956a0c17d387f4a174c7ed4e9b1b1f816dcf4ae 100644 --- a/paddle/fluid/operators/math/sequence_padding.cu +++ b/paddle/fluid/operators/math/sequence_padding.cu @@ -12,6 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ +#include #include "paddle/fluid/operators/math/sequence_padding.h" namespace paddle { @@ -61,7 +62,7 @@ template class PaddingLoDTensorFunctor { public: void operator()(const platform::CUDADeviceContext& context, - const framework::LoDTensor& seq, framework::Tensor& padding, + const framework::LoDTensor& seq, framework::Tensor* padding, bool norm_by_times) { auto lod = seq.lod(); PADDLE_ENFORCE_GT(lod.size(), 0UL, @@ -76,7 +77,7 @@ class PaddingLoDTensorFunctor { "The first dimension of LoDTensor seq should be " "equal to the sum of all sequences's length."); - auto padding_dims = padding.dims(); + auto padding_dims = padding->dims(); PADDLE_ENFORCE_EQ(padding_dims.size(), 3UL, "The input padding should be a 3-D Tensor of shape " "[max_sequence_length, num_sequences, sequence_width]."); @@ -97,8 +98,8 @@ class PaddingLoDTensorFunctor { "width of sequence in LoDTensor seq."); if (!norm_by_times && num_sequences == 1UL) { - TensorCopy(seq, context.GetPlace(), context, &padding); - padding.Resize(padding_dims); + TensorCopy(seq, context.GetPlace(), context, padding); + padding->Resize(padding_dims); return; } @@ -117,7 +118,7 @@ class PaddingLoDTensorFunctor { dim3 grid(grid_dim_x, grid_dim_y); const T* seq_data = seq.data(); - T* padding_data = padding.data(); + T* padding_data = padding->data(); if (norm_by_times) { SequencePaddingKernel<<>>( padding_data, const_cast(seq_data), @@ -136,16 +137,16 @@ template class UnpaddingLoDTensorFunctor { public: void operator()(const platform::CUDADeviceContext& context, - framework::LoDTensor& seq, const framework::Tensor& padding, + framework::LoDTensor* seq, const framework::Tensor& padding, bool norm_by_times) { - auto lod = seq.lod(); + auto lod = seq->lod(); PADDLE_ENFORCE_GT(lod.size(), 0UL, "The lod of LoDTensor seq should not be null."); const size_t level = 0; framework::LoD abs_offset_lod = framework::ToAbsOffset(lod); - auto seq_dims = seq.dims(); + auto seq_dims = seq->dims(); PADDLE_ENFORCE_EQ(seq_dims[0], static_cast(abs_offset_lod[level].back()), "The first dimension of LoDTensor seq should be " @@ -166,14 +167,14 @@ class UnpaddingLoDTensorFunctor { "The second dimension of Tensor padding should be " "the number of sequences in LoDTensor seq."); - const int64_t sequence_width = seq.numel() / seq_dims[0]; + const int64_t sequence_width = seq->numel() / seq_dims[0]; PADDLE_ENFORCE_EQ(padding_dims[2], sequence_width, "The third dimension of Tensor padding should be the " "width of sequence in LoDTensor seq."); if (!norm_by_times && num_sequences == 1UL) { - TensorCopy(padding, context.GetPlace(), context, &seq); - seq.Resize(seq_dims); + TensorCopy(padding, context.GetPlace(), context, seq); + seq->Resize(seq_dims); return; } @@ -192,7 +193,7 @@ class UnpaddingLoDTensorFunctor { dim3 grid(grid_dim_x, grid_dim_y); const T* padding_data = padding.data(); - T* seq_data = seq.data(); + T* seq_data = seq->data(); if (norm_by_times) { SequencePaddingKernel<<>>( const_cast(padding_data), seq_data, diff --git a/paddle/fluid/operators/math/sequence_padding.h b/paddle/fluid/operators/math/sequence_padding.h index 17f044b9d6667ed6a45bf5a0c2362c351d2c2beb..b56e6db1ebdac1a00561c07845c03bb8fbd8d35a 100644 --- a/paddle/fluid/operators/math/sequence_padding.h +++ b/paddle/fluid/operators/math/sequence_padding.h @@ -14,6 +14,7 @@ limitations under the License. */ #pragma once +#include #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/platform/device_context.h" @@ -64,13 +65,13 @@ template class PaddingLoDTensorFunctor { public: void operator()(const DeviceContext& context, const framework::LoDTensor& seq, - framework::Tensor& padding, bool norm_by_times); + framework::Tensor* padding, bool norm_by_times); }; template class UnpaddingLoDTensorFunctor { public: - void operator()(const DeviceContext& context, framework::LoDTensor& seq, + void operator()(const DeviceContext& context, framework::LoDTensor* seq, const framework::Tensor& padding, bool norm_by_times); }; diff --git a/paddle/fluid/operators/math/sequence_padding_test.cc b/paddle/fluid/operators/math/sequence_padding_test.cc index bece46e75374cc38512d77c9d6b2fc6584e39db2..b0c201db0ccbe81d8f57cd984d2cdfd2f6a48f25 100644 --- a/paddle/fluid/operators/math/sequence_padding_test.cc +++ b/paddle/fluid/operators/math/sequence_padding_test.cc @@ -14,6 +14,7 @@ limitations under the License. */ #include "paddle/fluid/operators/math/sequence_padding.h" #include +#include template void TestSequencePadding(const paddle::framework::LoD& lod, @@ -40,7 +41,7 @@ void TestSequencePadding(const paddle::framework::LoD& lod, if (paddle::platform::is_cpu_place(*place)) { seq = cpu_seq; } else { - TensorCopy(cpu_seq, *place, *context, &seq); + TensorCopySync(cpu_seq, *place, &seq); seq.set_lod(lod); } @@ -53,17 +54,17 @@ void TestSequencePadding(const paddle::framework::LoD& lod, static_cast(sequence_width)}); padding.mutable_data(padding_dims, *place); paddle::operators::math::PaddingLoDTensorFunctor()( - *context, seq, padding, false); + *context, seq, &padding, false); seq_back.set_lod(lod); seq_back.mutable_data(seq_dims, *place); paddle::operators::math::UnpaddingLoDTensorFunctor()( - *context, seq_back, padding, false); + *context, &seq_back, padding, false); if (paddle::platform::is_cpu_place(*place)) { cpu_seq_back = seq_back; } else { - TensorCopy(seq_back, paddle::platform::CPUPlace(), *context, &cpu_seq_back); + TensorCopySync(seq_back, paddle::platform::CPUPlace(), &cpu_seq_back); cpu_seq_back.set_lod(lod); } @@ -75,7 +76,7 @@ void TestSequencePadding(const paddle::framework::LoD& lod, delete place; delete context; -}; +} TEST(Seq2BatchPadding, CPU) { paddle::framework::LoD lod1; diff --git a/paddle/fluid/operators/math/sequence_pooling.cu b/paddle/fluid/operators/math/sequence_pooling.cu index 1935364da37e9a9881651455d2da4ecef1b1e266..97c2e69fe5327956fc2828781fe3a37b88cc1b99 100644 --- a/paddle/fluid/operators/math/sequence_pooling.cu +++ b/paddle/fluid/operators/math/sequence_pooling.cu @@ -12,9 +12,10 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ +#include #include "paddle/fluid/operators/math/math_function.h" #include "paddle/fluid/operators/math/sequence_pooling.h" -#include "paddle/fluid/platform/cuda_helper.h" +#include "paddle/fluid/platform/cuda_primitives.h" namespace paddle { namespace operators { diff --git a/paddle/fluid/operators/math/sequence_pooling.h b/paddle/fluid/operators/math/sequence_pooling.h index 38e780222955644c14e5bbbf16dee720c7758f5c..8dcbee65d0b63a137e5f422ec8667cc950641b4a 100644 --- a/paddle/fluid/operators/math/sequence_pooling.h +++ b/paddle/fluid/operators/math/sequence_pooling.h @@ -13,6 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once +#include #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/tensor.h" #include "paddle/fluid/platform/device_context.h" diff --git a/paddle/fluid/operators/math/sequence_scale.cc b/paddle/fluid/operators/math/sequence_scale.cc index 2c46d4183b5ccd6db909e4142797f97a626c43d5..ee5b22ca855b4fa26e9626aadb84fa9b93b72952 100644 --- a/paddle/fluid/operators/math/sequence_scale.cc +++ b/paddle/fluid/operators/math/sequence_scale.cc @@ -21,15 +21,15 @@ namespace math { template class ScaleLoDTensorFunctor { public: - void operator()(const platform::CPUDeviceContext& context, - framework::LoDTensor& seq, const T* scales) { + void operator()(const platform::CPUDeviceContext& context, const T* scales, + framework::LoDTensor* seq) { const size_t level = 0; - auto lod = seq.lod(); + auto lod = seq->lod(); const size_t num_seq = lod[level].size() - 1; - size_t seq_width = seq.dims()[1]; + size_t seq_width = seq->dims()[1]; framework::LoD abs_offset_lod = framework::ToAbsOffset(lod); - T* seq_data = seq.mutable_data(context.GetPlace()); + T* seq_data = seq->mutable_data(context.GetPlace()); for (size_t i = 0; i < num_seq; ++i) { for (size_t j = lod[level][i] * seq_width; j < lod[level][i + 1] * seq_width; ++j) { diff --git a/paddle/fluid/operators/math/sequence_scale.cu b/paddle/fluid/operators/math/sequence_scale.cu index 74085153c62354771f6126b58746229b5564f2d0..079338c1d3dac6a9403c5871f3face9f1f8e77d2 100644 --- a/paddle/fluid/operators/math/sequence_scale.cu +++ b/paddle/fluid/operators/math/sequence_scale.cu @@ -13,7 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/operators/math/sequence_scale.h" -#include "paddle/fluid/platform/cuda_helper.h" +#include "paddle/fluid/platform/cuda_primitives.h" namespace paddle { namespace operators { @@ -35,14 +35,14 @@ __global__ void SequenceScaleKernel(T* seq, size_t* lod, const T* scales, template class ScaleLoDTensorFunctor { public: - void operator()(const platform::CUDADeviceContext& context, - framework::LoDTensor& seq, const T* scales) { + void operator()(const platform::CUDADeviceContext& context, const T* scales, + framework::LoDTensor* seq) { const size_t level = 0; - auto lod = seq.lod(); + auto lod = seq->lod(); const size_t num_seq = lod[level].size() - 1; - const size_t seq_width = seq.numel() / seq.dims()[0]; + const size_t seq_width = seq->numel() / seq->dims()[0]; framework::LoD abs_offset_lod = framework::ToAbsOffset(lod); - T* seq_data = seq.mutable_data(context.GetPlace()); + T* seq_data = seq->mutable_data(context.GetPlace()); SequenceScaleKernel<<< num_seq, PADDLE_CUDA_NUM_THREADS, 0, context.stream()>>>( diff --git a/paddle/fluid/operators/math/sequence_scale.h b/paddle/fluid/operators/math/sequence_scale.h index 6cdcbe21cbf82881679d90de470f342f75b3e2f3..202243985c125cd518a27477eb370bf1a325fe16 100644 --- a/paddle/fluid/operators/math/sequence_scale.h +++ b/paddle/fluid/operators/math/sequence_scale.h @@ -46,8 +46,8 @@ namespace math { template class ScaleLoDTensorFunctor { public: - void operator()(const DeviceContext& context, framework::LoDTensor& seq, - const T* scales); + void operator()(const DeviceContext& context, const T* scales, + framework::LoDTensor* seq); }; } // namespace math diff --git a/paddle/fluid/operators/math/unpooling.cu b/paddle/fluid/operators/math/unpooling.cu index 367f343d51712d38edbb7eb50b41433258cf8c9d..c467ae8427d8f461b332eed8075631ed7e47b96e 100644 --- a/paddle/fluid/operators/math/unpooling.cu +++ b/paddle/fluid/operators/math/unpooling.cu @@ -13,7 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/operators/math/unpooling.h" -#include "paddle/fluid/platform/cuda_helper.h" +#include "paddle/fluid/platform/cuda_primitives.h" namespace paddle { namespace operators { diff --git a/paddle/fluid/operators/math/vol2col.cu b/paddle/fluid/operators/math/vol2col.cu index e0f3ef36879327c0592bb955dd800b44b228e721..28e1a752e34cf0171785a0341d8f0d8d3712fc7b 100644 --- a/paddle/fluid/operators/math/vol2col.cu +++ b/paddle/fluid/operators/math/vol2col.cu @@ -15,7 +15,7 @@ limitations under the License. */ #include #include #include "paddle/fluid/operators/math/vol2col.h" -#include "paddle/fluid/platform/cuda_helper.h" +#include "paddle/fluid/platform/cuda_primitives.h" namespace paddle { namespace operators { diff --git a/paddle/fluid/operators/math/vol2col.h b/paddle/fluid/operators/math/vol2col.h index dbc2ed7a6939cfb5dca61082decb8960e5f9ebda..5f59de8f02a52209a3901ca03680eb2d0dbc2658 100644 --- a/paddle/fluid/operators/math/vol2col.h +++ b/paddle/fluid/operators/math/vol2col.h @@ -14,6 +14,7 @@ limitations under the License. */ #pragma once +#include #include "paddle/fluid/framework/tensor.h" #include "paddle/fluid/framework/tensor_util.h" #include "paddle/fluid/platform/device_context.h" diff --git a/paddle/fluid/operators/math/vol2col_test.cc b/paddle/fluid/operators/math/vol2col_test.cc index b4e4186c986605b4fd14b10cadab81f07be4b27b..aa979c4f10907e604758c3e2cfb776cb994c9ceb 100644 --- a/paddle/fluid/operators/math/vol2col_test.cc +++ b/paddle/fluid/operators/math/vol2col_test.cc @@ -72,7 +72,7 @@ void testVol2col() { if (paddle::platform::is_cpu_place(*place)) { input = input_tmp; } else { - paddle::framework::TensorCopy(input_tmp, *place, *context, &input); + paddle::framework::TensorCopySync(input_tmp, *place, &input); } output.mutable_data({1, filter_size, filter_size, filter_size, output_depth, output_height, output_width}, @@ -86,7 +86,7 @@ void testVol2col() { if (paddle::platform::is_cpu_place(*place)) { out_cfo_ptr = output.data(); } else { - TensorCopy(output, paddle::platform::CPUPlace(), *context, &output_tmp); + TensorCopySync(output, paddle::platform::CPUPlace(), &output_tmp); out_cfo_ptr = output_tmp.data(); } @@ -100,7 +100,7 @@ void testVol2col() { if (paddle::platform::is_cpu_place(*place)) { input = input_tmp; } else { - TensorCopy(input_tmp, *place, *context, &input); + TensorCopySync(input_tmp, *place, &input); } paddle::operators::math::Col2VolFunctor col2vol; @@ -110,7 +110,7 @@ void testVol2col() { if (paddle::platform::is_cpu_place(*place)) { in_ptr = input.data(); } else { - TensorCopy(input, paddle::platform::CPUPlace(), *context, &input_tmp); + TensorCopySync(input, paddle::platform::CPUPlace(), &input_tmp); in_ptr = input_tmp.data(); } diff --git a/paddle/fluid/operators/matmul_op.cc b/paddle/fluid/operators/matmul_op.cc index e5d33fbc36438f97ff5b604e4efdbfbfa91fcee4..7182149164854038bb67a9f06cdbec8a4a0f1fb2 100644 --- a/paddle/fluid/operators/matmul_op.cc +++ b/paddle/fluid/operators/matmul_op.cc @@ -12,21 +12,264 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include "paddle/fluid/operators/matmul_op.h" #include +#include #include +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/detail/safe_ref.h" +#include "paddle/fluid/operators/math/blas.h" namespace paddle { namespace operators { +/** + * Get row matrix shape from a vector shape. If the rank of x_dim > 1, the + * original x_dim is returned. + */ +static framework::DDim RowMatrixFromVector(const framework::DDim &x_dim) { + if (x_dim.size() > 1) { + return x_dim; + } + return framework::make_ddim({1, x_dim[0]}); +} + +/** + * Get column matrix shape from a vector shape. If the ran of y_dim > 1, the + * original y_dim is returned. + */ +static framework::DDim ColumnMatrixFromVector(const framework::DDim &y_dim) { + if (y_dim.size() > 1) { + return y_dim; + } + return framework::make_ddim({y_dim[0], 1}); +} + +template +class MatMulKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &context) const override { + auto &x = + detail::Ref(context.Input("X"), "Cannot find X"); + auto &y = + detail::Ref(context.Input("Y"), "Cannot find Y"); + auto *out = context.Output("Out"); + out->mutable_data(context.GetPlace()); + + auto blas = math::GetBlas(context); + auto mat_dim_a = math::CreateMatrixDescriptor( + RowMatrixFromVector(x.dims()), 0, context.Attr("transpose_X")); + auto mat_dim_b = math::CreateMatrixDescriptor( + ColumnMatrixFromVector(y.dims()), 0, context.Attr("transpose_Y")); + blas.MatMul(x, mat_dim_a, y, mat_dim_b, T(1), out, T(0)); + } +}; + +// Reshape a rank-3 tensor from P x M x N to (P * M) x N. +// Identity op if the tensor is not of rank 3. +static framework::Tensor FoldInitDims(const framework::Tensor &input) { + auto output = input; + auto in_dims = input.dims(); + if (in_dims.size() == 3) { + output.Resize({in_dims[0] * in_dims[1], in_dims[2]}); + } + return output; +} + +// Reshape a rank-3 tensor from P x M x N to M x (P * N). +// (Warning: This requires transposing data and writes into new memory.) +// Identity op if the tensor is not of rank 3. +template +static framework::Tensor FoldHeadAndLastDims(const DeviceContext &context, + const framework::Tensor &input) { + auto in_dims = input.dims(); + if (in_dims.size() != 3) { + return input; + } + framework::Tensor output; + output.Resize({in_dims[1], in_dims[0], in_dims[2]}); + output.mutable_data(context.GetPlace()); + std::vector axis = {1, 0, 2}; + math::Transpose trans; + trans(context, input, &output, axis); + output.Resize({in_dims[1], in_dims[0] * in_dims[2]}); + + return output; +} + +/** + * Reshape a tensor to 3-D or 2-D tensor by matrix descriptor. + * + * The shape would be [BatchSize, H, W] or [H, W]. + * If transposed, `H,W` will be swapped. + */ +static void ReshapeTensorIntoMatrixSequence( + framework::Tensor *x, const math::MatDescriptor &descriptor) { + int64_t h, w; + h = descriptor.height_; + w = descriptor.width_; + if (descriptor.trans_) { + std::swap(w, h); + } + if (descriptor.batch_size_) { + x->Resize({descriptor.batch_size_, h, w}); + } else { + x->Resize({h, w}); + } +} + +/** + * Reshape the x,y,out tensor to 3-D or 2-D tensor by matrix descriptor + * Out = matmul(x, y) + * + * This method will first calculate X,Y matrix sequence, and then calculate + * the out shape. + * + * Assume X = [BatchSize, H1, W1], Y = [BatchSize, H2, W2] + * The out = [BatchSize, H1, W2] + * + * If there is no batch size in `X` and `Y`, the out will be [H1, W2] + * If any of `X` and `Y` has batch size BatchSize, the out will have the + * BatchSize. + */ +static void ReshapeXYOutIntoMatrixSequence(framework::Tensor *x, + framework::Tensor *y, + framework::Tensor *out, bool trans_x, + bool trans_y) { + auto x_dim = RowMatrixFromVector(x->dims()); + auto y_dim = ColumnMatrixFromVector(y->dims()); + auto mat_dim_x = math::CreateMatrixDescriptor(x_dim, 0, trans_x); + auto mat_dim_y = math::CreateMatrixDescriptor(y_dim, 0, trans_y); + if (mat_dim_x.batch_size_ == 0 && mat_dim_y.batch_size_ == 0) { + out->Resize({mat_dim_x.height_, mat_dim_y.width_}); + } else { + out->Resize({std::max(mat_dim_x.batch_size_, mat_dim_y.batch_size_), + mat_dim_x.height_, mat_dim_y.width_}); + } + + ReshapeTensorIntoMatrixSequence(x, mat_dim_x); + ReshapeTensorIntoMatrixSequence(y, mat_dim_y); +} + +// Using dimensional constraints on matrix multiplication, it is +// straight-forward to check the following table for when X and Y +// are both matrices. +// +// transpose_X | False | True | False | True +// transpose_Y | False | False | True | True +// -----------+----------+----------+----------+----------- +// dX = | dOut Y^T | Y dOut^T | dOut Y | Y^T dOut^T +// dY = | X^T dOut | X dOut | dOut^T X | dOut^T X^T +// +// When X is a vector of size K, we treat it instead as a matrix of shape +// (1, K). Similarly, when Y is a vector of size K, we treat it instead as +// a matrix of shape (K, 1). +// +// When X and Y are both 3-dimensional tensors, then the first dimension +// the batch dimension can be ignored and the exact same formulas apply +// as for two matrices. +// +// Finally, when, e.g., X is a 3-dimensional tensor but Y is a matrix, we end +// up with formulas like +// +// dY_{ij} = \sum_{p, m} X_{pmi} dOut_{pmj} +// +// To handle this sort of scenario, we reshape X : P x M x K, dOut: P x M x N +// to X: (P * M) x K, dOut: (P * M) x N. +template +class MatMulGradKernel : public framework::OpKernel { + public: + void MatMul(const framework::ExecutionContext &context, + const framework::Tensor &a, bool trans_a, + const framework::Tensor &b, bool trans_b, + framework::Tensor *out) const { + out->mutable_data(context.GetPlace()); + auto blas = math::GetBlas(context); + auto mat_dim_a = math::CreateMatrixDescriptor(a.dims(), 0, trans_a); + auto mat_dim_b = math::CreateMatrixDescriptor(b.dims(), 0, trans_b); + blas.MatMul(a, mat_dim_a, b, mat_dim_b, T(1), out, T(0)); + } + + void CalcInputGrad(const framework::ExecutionContext &context, + const framework::Tensor &a, bool trans_a, + bool is_fold_init_dims_a, const framework::Tensor &b, + bool trans_b, bool is_fold_init_dims_b, + framework::Tensor *out) const { + if (out == nullptr) return; + bool need_combine = (a.dims().size() == 3 || b.dims().size() == 3) && + out->dims().size() == 2; + if (!need_combine) { + MatMul(context, a, trans_a, b, trans_b, out); + } else { + auto &ctx = context.template device_context(); + MatMul(context, is_fold_init_dims_a + ? FoldInitDims(a) + : FoldHeadAndLastDims(ctx, a), + trans_a, is_fold_init_dims_b + ? FoldInitDims(b) + : FoldHeadAndLastDims(ctx, b), + trans_b, out); + } + } + + void Compute(const framework::ExecutionContext &context) const override { + auto x = *context.Input("X"); + auto y = *context.Input("Y"); + auto dout = + *context.Input(framework::GradVarName("Out")); + auto *dx = context.Output(framework::GradVarName("X")); + auto *dy = context.Output(framework::GradVarName("Y")); + bool transpose_x = context.Attr("transpose_X"); + bool transpose_y = context.Attr("transpose_Y"); + + ReshapeXYOutIntoMatrixSequence(&x, &y, &dout, transpose_x, transpose_y); + framework::DDim dx_dims; + if (dx) { + dx_dims = dx->dims(); + if (dx_dims != x.dims()) { + dx->Resize(x.dims()); + } + } + + framework::DDim dy_dims; + if (dy) { + dy_dims = dy->dims(); + if (dy_dims != y.dims()) { + dy->Resize(y.dims()); + } + } -using framework::Tensor; + if (transpose_x && transpose_y) { + CalcInputGrad(context, y, true, true, dout, true, false, dx); + CalcInputGrad(context, dout, true, true, x, true, false, dy); + } else if (transpose_x) { + CalcInputGrad(context, y, false, false, dout, true, false, dx); + CalcInputGrad(context, x, false, false, dout, false, true, dy); + } else if (transpose_y) { + CalcInputGrad(context, dout, false, false, y, false, true, dx); + CalcInputGrad(context, dout, true, true, x, false, true, dy); + } else { + CalcInputGrad(context, dout, false, false, y, true, false, dx); + CalcInputGrad(context, x, true, true, dout, false, true, dy); + } + + if (dx) { + if (dx_dims != x.dims()) { + dx->Resize(dx_dims); + } + } + if (dy) { + if (dy_dims != y.dims()) { + dy->Resize(dy_dims); + } + } + } +}; class MatMulOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(framework::InferShapeContext* context) const override { + void InferShape(framework::InferShapeContext *context) const override { PADDLE_ENFORCE(context->HasInput("X"), "Input(X) of MatMulOp should not be null."); PADDLE_ENFORCE(context->HasInput("Y"), @@ -36,121 +279,41 @@ class MatMulOp : public framework::OperatorWithKernel { auto dim_x = context->GetInputDim("X"); auto dim_y = context->GetInputDim("Y"); - bool transpose_x = context->Attrs().Get("transpose_X"); - bool transpose_y = context->Attrs().Get("transpose_Y"); - - PADDLE_ENFORCE_GE(dim_x.size(), 1, - "Input tensor X must be at least 1-dimensional."); - PADDLE_ENFORCE_GE(dim_y.size(), 1, - "Input tensor Y must be at least 1-dimensional."); - - std::vector out_dim; - int64_t batch_count = 1; - if (dim_x.size() > 3) { - PADDLE_ENFORCE_EQ( - dim_y.size(), dim_x.size(), - "The dimensions of X and Y must be the same, and both of " - "them should be %d-dimensional.", - dim_x.size()); - - // The first rank-2 dimensions are accumulated on the batch_count, and the - // last two dimensions are used for matrix multiplication. - for (int j = 0; j < dim_x.size() - 2; ++j) { - PADDLE_ENFORCE_EQ(dim_y[j], dim_x[j], - "The %d-th dimension of X and Y must be the same.", - j); - out_dim.push_back(dim_x[j]); - batch_count *= dim_x[j]; - } - } - int M = 0, N = 0, KX = 0, KY = 0, batchCountX = 0, batchCountY = 0; - bool remove_initial_dim = false, remove_final_dim = false; - - switch (dim_x.size()) { - case 1: - if (transpose_x) { - M = dim_x[0]; - KX = 1; - } else { - M = 1; - KX = dim_x[0]; - remove_initial_dim = true; - } - break; - case 2: - M = transpose_x ? dim_x[1] : dim_x[0]; - KX = transpose_x ? dim_x[0] : dim_x[1]; - break; - case 3: - batchCountX = dim_x[0]; - M = transpose_x ? dim_x[2] : dim_x[1]; - KX = transpose_x ? dim_x[1] : dim_x[2]; - break; - default: - batchCountX = batch_count; - size_t mat_s = dim_x.size() - 2; - M = transpose_x ? dim_x[mat_s + 1] : dim_x[mat_s]; - KX = transpose_x ? dim_x[mat_s] : dim_x[mat_s + 1]; - break; - } + auto mat_dim_x = + math::CreateMatrixDescriptor(RowMatrixFromVector(dim_x), 0, + context->Attrs().Get("transpose_X")); + auto mat_dim_y = + math::CreateMatrixDescriptor(ColumnMatrixFromVector(dim_y), 0, + context->Attrs().Get("transpose_Y")); - switch (dim_y.size()) { - case 1: - if (transpose_y) { - N = dim_y[0]; - KY = 1; - } else { - N = 1; - KY = dim_y[0]; - remove_final_dim = true; - } - break; - case 2: - KY = transpose_y ? dim_y[1] : dim_y[0]; - N = transpose_y ? dim_y[0] : dim_y[1]; - break; - case 3: - batchCountY = dim_y[0]; - KY = transpose_y ? dim_y[2] : dim_y[1]; - N = transpose_y ? dim_y[1] : dim_y[2]; - break; - default: - batchCountY = batch_count; - size_t mat_s = dim_y.size() - 2; - KY = transpose_y ? dim_y[mat_s + 1] : dim_y[mat_s]; - N = transpose_y ? dim_y[mat_s] : dim_y[mat_s + 1]; + PADDLE_ENFORCE_EQ(mat_dim_x.width_, mat_dim_y.height_); + PADDLE_ENFORCE(mat_dim_x.batch_size_ == mat_dim_y.batch_size_ || + mat_dim_x.batch_size_ == 0 || mat_dim_y.batch_size_ == 0); + std::vector dim_out; + if (mat_dim_x.batch_size_ != 0) { + dim_out = framework::vectorize(dim_x); + dim_out[dim_out.size() - 2] = mat_dim_x.height_; + dim_out[dim_out.size() - 1] = mat_dim_y.width_; + } else if (mat_dim_y.batch_size_ != 0) { + dim_out = framework::vectorize(dim_y); + dim_out[dim_out.size() - 2] = mat_dim_x.height_; + dim_out[dim_out.size() - 1] = mat_dim_y.width_; + } else { + dim_out = {mat_dim_x.height_, mat_dim_y.width_}; } - PADDLE_ENFORCE_EQ( - KX, KY, - "First matrix's width must be equal with second matrix's height."); - if (batchCountX && batchCountY) { - PADDLE_ENFORCE_EQ( - batchCountX, batchCountY, - "When Input(X) and Input(Y) are both three dimensional, they " - "must have the same batch dimension."); + if (dim_x.size() == 1 && dim_out[dim_out.size() - 2] == 1) { + std::swap(dim_out[dim_out.size() - 2], dim_out[dim_out.size() - 1]); + dim_out.resize(dim_out.size() - 1); } - int batchCount = std::max(batchCountX, batchCountY); - std::vector dim_out; - if (batchCount) { - if (dim_x.size() > 3) { - dim_out.insert(dim_out.begin(), out_dim.begin(), out_dim.end()); - } else { - dim_out.push_back(batchCount); - } + if (dim_y.size() == 1 && dim_out[dim_out.size() - 1] == 1) { + dim_out.resize(dim_out.size() - 1); } - if (!remove_initial_dim) { - dim_out.push_back(M); - } - if (!remove_final_dim) { - dim_out.push_back(N); - } - if (dim_out.size() == 0) { - // We don't support 0-dimensional Tensors (scalars), so instead - // treat the output as a Tensor of shape (1, ) in this case. - dim_out.push_back(1); + + if (dim_out.empty()) { + dim_out = {1}; } context->SetOutputDim("Out", framework::make_ddim(dim_out)); context->ShareLoD("X", /*->*/ "Out"); @@ -159,8 +322,7 @@ class MatMulOp : public framework::OperatorWithKernel { class MatMulOpMaker : public framework::OpProtoAndCheckerMaker { public: - MatMulOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The first input of MatMul op"); AddInput("Y", "The second input of MatMul op"); AddOutput("Out", "The output of MatMul op"); @@ -213,7 +375,7 @@ class MatMulOpGrad : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(framework::InferShapeContext* context) const override { + void InferShape(framework::InferShapeContext *context) const override { PADDLE_ENFORCE(context->HasInput("X"), "Input(X) should not be null"); PADDLE_ENFORCE(context->HasInput("Y"), "Input(Y) should not be null"); PADDLE_ENFORCE(context->HasInput(framework::GradVarName("Out")), @@ -233,15 +395,52 @@ class MatMulOpGrad : public framework::OperatorWithKernel { } }; +class MatMulOpGradMaker : public framework::SingleGradOpDescMaker { + public: + using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; + + protected: + std::unique_ptr Apply() const override { + auto *retv = new framework::OpDesc(); + retv->SetType("matmul_grad"); + retv->SetInput("X", Input("X")); + retv->SetInput("Y", Input("Y")); + retv->SetInput(framework::GradVarName("Out"), OutputGrad("Out")); + retv->SetOutput(framework::GradVarName("X"), InputGrad("X")); + retv->SetOutput(framework::GradVarName("Y"), InputGrad("Y")); + retv->SetAttrMap(Attrs()); + return std::unique_ptr(retv); + } +}; } // namespace operators } // namespace paddle namespace ops = paddle::operators; REGISTER_OPERATOR(matmul, ops::MatMulOp, ops::MatMulOpMaker, - paddle::framework::DefaultGradOpDescMaker); + ops::MatMulOpGradMaker); REGISTER_OPERATOR(matmul_grad, ops::MatMulOpGrad); REGISTER_OP_CPU_KERNEL( - matmul, ops::MatMulKernel); + matmul, ops::MatMulKernel, + ops::MatMulKernel, + ops::MatMulKernel); REGISTER_OP_CPU_KERNEL( matmul_grad, - ops::MatMulGradKernel); + ops::MatMulGradKernel, + ops::MatMulGradKernel, + ops::MatMulGradKernel); + +#ifdef PADDLE_WITH_CUDA +REGISTER_OP_CUDA_KERNEL( + matmul, ops::MatMulKernel, + ops::MatMulKernel, + ops::MatMulKernel); +REGISTER_OP_CUDA_KERNEL( + matmul_grad, + ops::MatMulGradKernel, + ops::MatMulGradKernel, + ops::MatMulGradKernel); +#endif diff --git a/paddle/fluid/operators/matmul_op.h b/paddle/fluid/operators/matmul_op.h deleted file mode 100644 index f2e9cfdcdbf93326ae193776a7d5f6a324373603..0000000000000000000000000000000000000000 --- a/paddle/fluid/operators/matmul_op.h +++ /dev/null @@ -1,244 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once -#include -#include -#include -#include "paddle/fluid/framework/op_registry.h" -#include "paddle/fluid/operators/math/math_function.h" -#include "paddle/fluid/operators/math/matmul.h" - -namespace paddle { -namespace operators { -namespace matmul_detail { - -using Tensor = framework::Tensor; -using DDim = framework::DDim; -using framework::make_ddim; -using framework::vectorize; - -template -class MatMulKernel : public framework::OpKernel { - public: - void Compute(const framework::ExecutionContext& context) const override { - const Tensor& x = *context.Input("X"); - const Tensor& y = *context.Input("Y"); - Tensor* out = context.Output("Out"); - out->mutable_data(context.GetPlace()); - bool transpose_x = context.Attr("transpose_X"); - bool transpose_y = context.Attr("transpose_Y"); - - math::MatMulFunctor()( - context.template device_context(), x, transpose_x, y, - transpose_y, T(1), out, T(0)); - } -}; - -template -inline Tensor Reshape(const Tensor& input, const DDim& dims) { - Tensor output; - output.ShareDataWith(input); - output.Resize(dims); - return output; -} - -// Reshape a rank-3 tensor from P x M x N to (P * M) x N. -// Identity op if the tensor is not of rank 3. -template -Tensor CombineBatchAndM(const Tensor& input) { - Tensor output; - output.ShareDataWith(input); - auto in_dims = input.dims(); - if (in_dims.size() == 3) { - std::vector out_dims = {in_dims[0] * in_dims[1], in_dims[2]}; - output.Resize(make_ddim(out_dims)); - } - return output; -} - -// Reshape a rank-3 tensor from P x M x N to M x (P * N). -// (Warning: This requires transposing data and writes into new memory.) -// Identity op if the tensor is not of rank 3. -template -Tensor CombineBatchAndN(const DeviceContext& context, const Tensor& input) { - Tensor output; - auto in_dims = input.dims(); - if (in_dims.size() == 3) { - output.Resize({in_dims[1], in_dims[0], in_dims[2]}); - output.mutable_data(context.GetPlace()); - std::vector axis = {1, 0, 2}; - math::Transpose trans; - trans(context, input, &output, axis); - std::vector out_dims = {in_dims[1], in_dims[0] * in_dims[2]}; - output.Resize({in_dims[1], in_dims[0] * in_dims[2]}); - } else { - output.ShareDataWith(input); - } - return output; -} - -// Using dimensional constraints on matrix multiplication, it is -// straight-forward to check the following table for when X and Y -// are both matrices. -// -// transpose_X | False | True | False | True -// transpose_Y | False | False | True | True -// -----------+----------+----------+----------+----------- -// dX = | dOut Y^T | Y dOut^T | dOut Y | Y^T dOut^T -// dY = | X^T dOut | X dOut | dOut^T X | dOut^T X^T -// -// When X is a vector of size K, we treat it instead as a matrix of shape -// (1, K). Similarly, when Y is a vector of size K, we treat it instead as -// a matrix of shape (K, 1). -// -// When X and Y are both 3-dimensional tensors, then the first dimension -// the batch dimension can be ignored and the exact same formulas apply -// as for two matrices. -// -// Finally, when, e.g., X is a 3-dimensional tensor but Y is a matrix, we end -// up with formulas like -// -// dY_{ij} = \sum_{p, m} X_{pmi} dOut_{pmj} -// -// To handle this sort of scenario, we reshape X : P x M x K, dOut: P x M x N -// to X: (P * M) x K, dOut: (P * M) x N. -template -class MatMulGradKernel : public framework::OpKernel { - public: - void Compute(const framework::ExecutionContext& context) const override { - const Tensor& x = *context.Input("X"); - const Tensor& y = *context.Input("Y"); - const Tensor& dout = *context.Input(framework::GradVarName("Out")); - Tensor* dx = context.Output(framework::GradVarName("X")); - Tensor* dy = context.Output(framework::GradVarName("Y")); - bool transpose_x = context.Attr("transpose_X"); - bool transpose_y = context.Attr("transpose_Y"); - - std::vector x_dims = vectorize(x.dims()); - std::vector y_dims = vectorize(y.dims()); - - // If X is a vector, reshape it to a matrix. - if (x_dims.size() == 1) { - x_dims.insert(x_dims.begin(), 1); - } - - // If Y is a vector, reshape it to a matrix. - if (y_dims.size() == 1) { - y_dims.push_back(1); - } - - int batch_count = 0; - // The first rank-2 dimensions are accumulated on the batch_count, and the - // last two dimensions are used for matrix multiplication. - if (x_dims.size() > 3) { - batch_count = accumulate(x_dims.begin(), x_dims.end() - 2, 1, - std::multiplies()); - } - // Fix the dOut dimensions. - int M = 0, N = 0, batchCountX = 0, batchCountY = 0; - - switch (x_dims.size()) { - case 2: - M = transpose_x ? x_dims[1] : x_dims[0]; - break; - case 3: - batchCountX = x_dims[0]; - M = transpose_x ? x_dims[2] : x_dims[1]; - break; - default: - batchCountX = batch_count; - size_t mat_s = x_dims.size() - 2; - M = transpose_x ? x_dims[mat_s + 1] : x_dims[mat_s]; - } - - switch (y_dims.size()) { - case 2: - N = transpose_y ? y_dims[0] : y_dims[1]; - break; - case 3: - batchCountY = y_dims[0]; - N = transpose_y ? y_dims[1] : y_dims[2]; - break; - default: - batchCountY = batch_count; - size_t mat_s = y_dims.size() - 2; - N = transpose_y ? y_dims[mat_s] : y_dims[mat_s + 1]; - } - if (batchCountX && batchCountY) { - PADDLE_ENFORCE_EQ( - batchCountX, batchCountY, - "When Input(X) and Input(Y) are both three dimensional, they " - "must have the same batch dimension."); - } - int batchCount = std::max(batchCountX, batchCountY); - std::vector dout_dims = {M, N}; - if (batchCount) { - if (x_dims.size() > 3) { - dout_dims.insert(dout_dims.begin(), x_dims.begin(), x_dims.end() - 2); - } else { - dout_dims.insert(dout_dims.begin(), batchCount); - } - } - Tensor X = Reshape(x, make_ddim(x_dims)); - Tensor Y = Reshape(y, make_ddim(y_dims)); - Tensor dOut = Reshape(dout, make_ddim(dout_dims)); - - auto& dev_ctx = context.template device_context(); - if (dx) { - dx->mutable_data(context.GetPlace()); - const Tensor& dOut_for_dX = - (x_dims.size() == 2 && y_dims.size() == 3) - ? CombineBatchAndN(dev_ctx, dOut) - : dOut; - if (x_dims.size() == 2 && y_dims.size() == 3) { - Y = transpose_y ? CombineBatchAndM(Y) - : CombineBatchAndN(dev_ctx, Y); - } - if (transpose_x) { - math::MatMulFunctor()( - dev_ctx, Y, transpose_y, dOut_for_dX, transpose_x, T(1), dx, T(0)); - } else { - math::MatMulFunctor()( - dev_ctx, dOut_for_dX, transpose_x, Y, !transpose_y, T(1), dx, T(0)); - } - } - - if (dy) { - dy->mutable_data(context.GetPlace()); - const Tensor& dOut_for_dY = (y_dims.size() == 2 && x_dims.size() == 3) - ? CombineBatchAndM(dOut) - : dOut; - if (y_dims.size() == 2 && x_dims.size() == 3) { - X = transpose_x ? CombineBatchAndN(dev_ctx, X) - : CombineBatchAndM(X); - dOut = CombineBatchAndM(dOut); - } - if (transpose_y) { - math::MatMulFunctor()( - dev_ctx, dOut_for_dY, transpose_y, X, transpose_x, T(1), dy, T(0)); - } else { - math::MatMulFunctor()( - dev_ctx, X, !transpose_x, dOut_for_dY, transpose_y, T(1), dy, T(0)); - } - } - } -}; -} // namespace matmul_detail - -using matmul_detail::MatMulKernel; -using matmul_detail::MatMulGradKernel; - -} // namespace operators -} // namespace paddle diff --git a/paddle/fluid/operators/max_sequence_len_op.cc b/paddle/fluid/operators/max_sequence_len_op.cc index 4cd7c89b48a2442ee7a5074abbf0f3dd9ea3bcb4..8e508b68eeab69a4595904dcc3ea0a541d9ab6e6 100644 --- a/paddle/fluid/operators/max_sequence_len_op.cc +++ b/paddle/fluid/operators/max_sequence_len_op.cc @@ -41,8 +41,7 @@ class MaxSeqenceLenOp : public framework::OperatorBase { class MaxSeqenceLenOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - MaxSeqenceLenOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("RankTable", "The lod_rank_table."); AddOutput("Out", "The max sequence length."); AddComment( diff --git a/paddle/fluid/operators/maxout_op.cc b/paddle/fluid/operators/maxout_op.cc index e2bcba5a5e15d4d5f10ae4ae64b5262f750137ab..058115cb624627d81b31d0903f7d615d19708c77 100644 --- a/paddle/fluid/operators/maxout_op.cc +++ b/paddle/fluid/operators/maxout_op.cc @@ -22,8 +22,7 @@ using framework::Tensor; class MaxOutOpMaker : public framework::OpProtoAndCheckerMaker { public: - MaxOutOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput( "X", "(Tensor) The input tensor of maxout operator. " diff --git a/paddle/fluid/operators/mean_op.cc b/paddle/fluid/operators/mean_op.cc index a134796bfcaa9dea2483ace9f5045e257916daba..74477eb439dc202c3f5f17fdf3e1647bc5c23512 100644 --- a/paddle/fluid/operators/mean_op.cc +++ b/paddle/fluid/operators/mean_op.cc @@ -32,8 +32,7 @@ class MeanOp : public framework::OperatorWithKernel { class MeanOpMaker : public framework::OpProtoAndCheckerMaker { public: - MeanOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The input of mean op"); AddOutput("Out", "The output of mean op"); AddComment(R"DOC( diff --git a/paddle/fluid/operators/merge_lod_tensor_op.cc b/paddle/fluid/operators/merge_lod_tensor_op.cc index 4ebf20cbba69bee09dfddb8e928ddc95665e4731..a16861b3b77fc980ab932b9d88859b38ec36108b 100644 --- a/paddle/fluid/operators/merge_lod_tensor_op.cc +++ b/paddle/fluid/operators/merge_lod_tensor_op.cc @@ -121,8 +121,7 @@ class MergeLoDTensorOp : public framework::OperatorBase { class MergeLoDTensorOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - MergeLoDTensorOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The input LoDTensor, contains complete lod information to " "construct the output"); diff --git a/paddle/fluid/operators/minus_op.cc b/paddle/fluid/operators/minus_op.cc index a302b24560e680076d62d02b422c6410467deb1d..34571a38a14795a98ac8454cec606077727b5ffa 100644 --- a/paddle/fluid/operators/minus_op.cc +++ b/paddle/fluid/operators/minus_op.cc @@ -48,8 +48,7 @@ class MinusOp : public framework::OperatorWithKernel { class MinusOpMaker : public framework::OpProtoAndCheckerMaker { public: - MinusOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The left tensor of minus operator."); AddInput("Y", "The right tensor of minus operator."); AddOutput("Out", "The output tensor of minus operator."); diff --git a/paddle/fluid/operators/mkldnn_activation_op.h b/paddle/fluid/operators/mkldnn_activation_op.h index f26a165b5a59f01f864d62bbf798f4cbffa65371..85664623d7330e9473286d995bec67879510dbd7 100644 --- a/paddle/fluid/operators/mkldnn_activation_op.h +++ b/paddle/fluid/operators/mkldnn_activation_op.h @@ -13,6 +13,8 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once +#include + #include "paddle/fluid/framework/eigen.h" #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/operators/detail/safe_ref.h" @@ -60,52 +62,5 @@ class MKLDNNActivationGradKernel } }; -namespace { // NOLINT -framework::OpKernelType GetKernelType( - const framework::ExecutionContext& ctx, - const framework::OperatorWithKernel& oper) { - framework::LibraryType library{framework::LibraryType::kPlain}; -#ifdef PADDLE_WITH_MKLDNN - if (library == framework::LibraryType::kPlain && - platform::CanMKLDNNBeUsed(ctx)) { - library = framework::LibraryType::kMKLDNN; - } -#endif - framework::DataLayout layout = framework::DataLayout::kAnyLayout; - return framework::OpKernelType( - framework::ToDataType(ctx.Input("X")->type()), - ctx.GetPlace(), layout, library); -} -} // anonymous namespace - -class ActivationWithMKLDNNOp : public framework::OperatorWithKernel { - public: - using framework::OperatorWithKernel::OperatorWithKernel; - - void InferShape(framework::InferShapeContext* ctx) const override { - ctx->SetOutputDim("Out", ctx->GetInputDim("X")); - ctx->ShareLoD("X", /*->*/ "Out"); - } - - framework::OpKernelType GetExpectedKernelType( - const framework::ExecutionContext& ctx) const override { - return GetKernelType(ctx, *this); - } -}; - -class ActivationWithMKLDNNOpGrad : public framework::OperatorWithKernel { - public: - using framework::OperatorWithKernel::OperatorWithKernel; - - void InferShape(framework::InferShapeContext* ctx) const override { - ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("Out")); - } - - framework::OpKernelType GetExpectedKernelType( - const framework::ExecutionContext& ctx) const override { - return GetKernelType(ctx, *this); - } -}; - } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/modified_huber_loss_op.cc b/paddle/fluid/operators/modified_huber_loss_op.cc index 3a0fc74584391d0441105a8ac7d7ac292e10fb8d..35db4c1ad1f6c6481eca397e99fc8c1f0bc7164c 100644 --- a/paddle/fluid/operators/modified_huber_loss_op.cc +++ b/paddle/fluid/operators/modified_huber_loss_op.cc @@ -39,8 +39,7 @@ class ModifiedHuberLossOp : public framework::OperatorWithKernel { class ModifiedHuberLossOpMaker : public framework::OpProtoAndCheckerMaker { public: - ModifiedHuberLossOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The input tensor of modified huber loss op. " "X is 2-D tensor with shape [batch_size, 1]."); diff --git a/paddle/fluid/operators/momentum_op.cc b/paddle/fluid/operators/momentum_op.cc index 6c70970e15f0d63ebe2134c6bc8163339ba30e75..dcd73e3c3e40f80e07b73944d1f0cc57fea010d3 100644 --- a/paddle/fluid/operators/momentum_op.cc +++ b/paddle/fluid/operators/momentum_op.cc @@ -17,6 +17,8 @@ limitations under the License. */ namespace paddle { namespace operators { +using Tensor = framework::Tensor; + class MomentumOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; @@ -50,12 +52,17 @@ class MomentumOp : public framework::OperatorWithKernel { ctx->SetOutputDim("ParamOut", param_dim); ctx->SetOutputDim("VelocityOut", param_dim); } + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext &ctx) const override { + auto input_data_type = + framework::ToDataType(ctx.Input("Param")->type()); + return framework::OpKernelType(input_data_type, ctx.GetPlace()); + } }; class MomentumOpMaker : public framework::OpProtoAndCheckerMaker { public: - MomentumOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Param", "(Tensor, default Tensor) " "Input parameter that has to be updated"); diff --git a/paddle/fluid/operators/mul_mkldnn_op.cc b/paddle/fluid/operators/mul_mkldnn_op.cc deleted file mode 100644 index a5f3a98f678a870d30eebfc4cf329de7c93266ee..0000000000000000000000000000000000000000 --- a/paddle/fluid/operators/mul_mkldnn_op.cc +++ /dev/null @@ -1,197 +0,0 @@ -/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "mkldnn.hpp" -#include "paddle/fluid/framework/tensor.h" -#include "paddle/fluid/operators/mul_op.h" -#include "paddle/fluid/platform/device_context.h" -#include "paddle/fluid/platform/mkldnn_helper.h" - -namespace paddle { -namespace operators { - -using paddle::framework::Tensor; -using paddle::platform::MKLDNNDeviceContext; - -template -mkldnn::memory::desc type(const std::vector& dims, Format&& f) { - return platform::MKLDNNMemDesc(dims, mkldnn::memory::data_type::f32, f); -} - -template -class MulMKLDNNOpKernel : public paddle::framework::OpKernel { - void Compute(const paddle::framework::ExecutionContext& ctx) const override { - PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()), - "It must use CPUPlace."); - - auto& dev_ctx = ctx.template device_context(); - auto mkldnn_engine = dev_ctx.GetEngine(); - - auto input = ctx.Input("X"); - auto weight = ctx.Input("Y"); - - PADDLE_ENFORCE(input->dims().size() & (2 | 4), - "Input must be with 2 or 4 dimensions, i.e. NC or NCHW"); - PADDLE_ENFORCE(weight->dims().size() & (2 | 4), - "Weights must be with 2 or 4 dimensions, i.e. OI or OIHW"); - - std::vector w_tz = paddle::framework::vectorize2int(weight->dims()); - std::vector src_tz = paddle::framework::vectorize2int(input->dims()); - - auto src_md = - src_tz.size() != 2 - ? type(src_tz, mkldnn::memory::format::nchw) - : type({src_tz[0], src_tz[1]}, mkldnn::memory::format::nc); - - auto dst_md = type({src_tz[0], w_tz[1]}, mkldnn::memory::format::nc); - - auto weights_md = - src_tz.size() != 2 - ? type({w_tz[1], src_tz[1], src_tz[2], src_tz[3]}, - mkldnn::memory::format::oihw) - : type({w_tz[1], src_tz[1]}, mkldnn::memory::format::oi); - - auto output = ctx.Output("Out"); - T* output_data = output->mutable_data(ctx.GetPlace()); - - const std::string key = ctx.op().Output("Out"); - const std::string key_fc_pd = key + "@mul_pd"; - - const T* input_data = input->data(); - const T* w_data = weight->data(); - - auto dst_memory = mkldnn::memory({dst_md, mkldnn_engine}, output_data); - - auto src_memory = mkldnn::memory({src_md, mkldnn_engine}, - platform::to_void_cast(input_data)); - - auto weights_memory = mkldnn::memory({weights_md, mkldnn_engine}, - platform::to_void_cast(w_data)); - - auto pd = platform::MKLDNNFwdPrimitiveDesc( - mkldnn_engine, src_md, weights_md, dst_md); - - dev_ctx.SetBlob(key_fc_pd, pd); - - auto forward = mkldnn::inner_product_forward(*pd, src_memory, - weights_memory, dst_memory); - - std::vector pipeline = {forward}; - mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); - } -}; - -template -class MulMKLDNNGradOpKernel : public paddle::framework::OpKernel { - public: - void Compute(const paddle::framework::ExecutionContext& ctx) const override { - PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()), - "It must use CPUPlace."); - - auto& dev_ctx = ctx.template device_context(); - auto mkldnn_engine = dev_ctx.GetEngine(); - - const Tensor* input = ctx.Input("X"); - const Tensor* w = ctx.Input("Y"); - - const Tensor* out_grad = ctx.Input(framework::GradVarName("Out")); - Tensor* input_grad = ctx.Output(framework::GradVarName("X")); - Tensor* w_grad = ctx.Output(framework::GradVarName("Y")); - - const std::string key = ctx.op().Input("Out"); - const std::string key_fc_pd = key + "@mul_pd"; - - const T* input_data = input->data(); - const T* w_data = w->data(); - const T* out_grad_data = out_grad->data(); - T* input_grad_data = nullptr; - T* w_grad_data = nullptr; - - if (input_grad) { - input_grad_data = input_grad->mutable_data(ctx.GetPlace()); - } - if (w_grad) { - w_grad_data = w_grad->mutable_data(ctx.GetPlace()); - } - - std::vector src_tz = paddle::framework::vectorize2int(input->dims()); - std::vector w_tz = paddle::framework::vectorize2int(w->dims()); - - auto src_md = - src_tz.size() != 2 - ? type(src_tz, mkldnn::memory::format::nchw) - : type({src_tz[0], src_tz[1]}, mkldnn::memory::format::nc); - - auto dst_md = type({src_tz[0], w_tz[1]}, mkldnn::memory::format::nc); - - auto weights_md = - src_tz.size() != 2 - ? type({w_tz[1], src_tz[1], src_tz[2], src_tz[3]}, - mkldnn::memory::format::oihw) - : type({w_tz[1], src_tz[1]}, mkldnn::memory::format::oi); - - auto src_memory = mkldnn::memory({src_md, mkldnn_engine}, - platform::to_void_cast(input_data)); - - auto dst_memory = mkldnn::memory({dst_md, mkldnn_engine}, - platform::to_void_cast(out_grad_data)); - - auto weight_memory = mkldnn::memory({weights_md, mkldnn_engine}, - platform::to_void_cast(w_data)); - - auto pd = - std::static_pointer_cast( - dev_ctx.GetBlob(key_fc_pd)); - - PADDLE_ENFORCE(pd != nullptr, "Fail to find pd in device context"); - - if (w_grad) { - auto weights_grad_memory = mkldnn::memory( - {weights_md, mkldnn_engine}, platform::to_void_cast(w_grad_data)); - - auto bwd_weight_pd = platform::MKLDNNBwdPrimitiveDesc< - mkldnn::inner_product_backward_weights>(mkldnn_engine, *pd, src_md, - weights_md, dst_md); - - auto bwd_weights_prim = mkldnn::inner_product_backward_weights( - bwd_weight_pd, src_memory, dst_memory, weights_grad_memory); - - std::vector pipeline{bwd_weights_prim}; - mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); - } - - if (input_grad) { - auto src_grad_memory = mkldnn::memory( - {src_md, mkldnn_engine}, platform::to_void_cast(input_grad_data)); - - auto bwd_data_pd = - platform::MKLDNNBwdPrimitiveDesc( - mkldnn_engine, *pd, src_md, weights_md, dst_md); - - auto bwd_data_prim = mkldnn::inner_product_backward_data( - bwd_data_pd, dst_memory, weight_memory, src_grad_memory); - - std::vector pipeline{bwd_data_prim}; - mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); - } - } -}; -} // namespace operators -} // namespace paddle - -REGISTER_OP_KERNEL(mul, MKLDNN, ::paddle::platform::CPUPlace, - paddle::operators::MulMKLDNNOpKernel); - -REGISTER_OP_KERNEL(mul_grad, MKLDNN, ::paddle::platform::CPUPlace, - paddle::operators::MulMKLDNNGradOpKernel); diff --git a/paddle/fluid/operators/mul_op.cc b/paddle/fluid/operators/mul_op.cc index c9fabc8d485b3bba2c8ae14b3616d0bdcae058a7..51993398bd3427e1f0da155918395bc50fa65e45 100644 --- a/paddle/fluid/operators/mul_op.cc +++ b/paddle/fluid/operators/mul_op.cc @@ -16,10 +16,6 @@ limitations under the License. */ #include #include -#ifdef PADDLE_WITH_MKLDNN -#include "paddle/fluid/platform/mkldnn_helper.h" -#endif - namespace paddle { namespace operators { @@ -76,28 +72,11 @@ class MulOp : public framework::OperatorWithKernel { ctx->SetOutputDim("Out", framework::make_ddim(output_dims)); ctx->ShareLoD("X", /*->*/ "Out"); } - - private: - framework::OpKernelType GetExpectedKernelType( - const framework::ExecutionContext& ctx) const override { - framework::LibraryType library{framework::LibraryType::kPlain}; -#ifdef PADDLE_WITH_MKLDNN - if (library == framework::LibraryType::kPlain && - platform::CanMKLDNNBeUsed(ctx)) { - library = framework::LibraryType::kMKLDNN; - } -#endif - framework::DataLayout layout{framework::DataLayout::kAnyLayout}; - return framework::OpKernelType( - framework::ToDataType(ctx.Input("X")->type()), ctx.GetPlace(), - layout, library); - } }; class MulOpMaker : public framework::OpProtoAndCheckerMaker { public: - MulOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor), The first input tensor of mul op."); AddInput("Y", "(Tensor), The second input tensor of mul op."); AddOutput("Out", "(Tensor), The output tensor of mul op."); @@ -121,9 +100,6 @@ class MulOpMaker : public framework::OpProtoAndCheckerMaker { )DOC") .SetDefault(1) .EqualGreaterThan(1); - AddAttr("use_mkldnn", - "(bool, default false) Only used in mkldnn kernel") - .SetDefault(false); AddAttr( "y_num_col_dims", R"DOC((int, default 1), The mul_op can take tensors with more than two, @@ -178,22 +154,6 @@ class MulGradOp : public framework::OperatorWithKernel { ctx->SetOutputDim(y_grad_name, y_dims); } } - - private: - framework::OpKernelType GetExpectedKernelType( - const framework::ExecutionContext& ctx) const override { - framework::LibraryType library{framework::LibraryType::kPlain}; -#ifdef PADDLE_WITH_MKLDNN - if (library == framework::LibraryType::kPlain && - platform::CanMKLDNNBeUsed(ctx)) { - library = framework::LibraryType::kMKLDNN; - } -#endif - framework::DataLayout layout{framework::DataLayout::kAnyLayout}; - return framework::OpKernelType( - framework::ToDataType(ctx.Input("X")->type()), ctx.GetPlace(), - layout, library); - } }; } // namespace operators @@ -204,6 +164,8 @@ REGISTER_OPERATOR(mul, ops::MulOp, ops::MulOpMaker, paddle::framework::DefaultGradOpDescMaker); REGISTER_OPERATOR(mul_grad, ops::MulGradOp); REGISTER_OP_CPU_KERNEL( - mul, ops::MulKernel); + mul, ops::MulKernel, + ops::MulKernel); REGISTER_OP_CPU_KERNEL( - mul_grad, ops::MulGradKernel); + mul_grad, ops::MulGradKernel, + ops::MulGradKernel); diff --git a/paddle/fluid/operators/mul_op.cu.cc b/paddle/fluid/operators/mul_op.cu.cc index 757f9c3ee2665c7ac654659416fe8dd727dca16d..81f3e42bf412fa4d2cb48405f2f8ee49b6aa0b67 100644 --- a/paddle/fluid/operators/mul_op.cu.cc +++ b/paddle/fluid/operators/mul_op.cu.cc @@ -18,6 +18,8 @@ limitations under the License. */ namespace ops = paddle::operators; namespace plat = paddle::platform; REGISTER_OP_CUDA_KERNEL(mul, ops::MulKernel, + ops::MulKernel, ops::MulKernel); REGISTER_OP_CUDA_KERNEL(mul_grad, - ops::MulGradKernel); + ops::MulGradKernel, + ops::MulGradKernel); diff --git a/paddle/fluid/operators/mul_op.h b/paddle/fluid/operators/mul_op.h index b1260d36ebe11f65529ac274c959479dcb38ee5f..15dd975e3bbf80b2e616e6628555e812d025f70a 100644 --- a/paddle/fluid/operators/mul_op.h +++ b/paddle/fluid/operators/mul_op.h @@ -14,9 +14,9 @@ limitations under the License. */ #pragma once -#include "paddle/fluid/operators/math/math_function.h" - #include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/math/blas.h" +#include "paddle/fluid/operators/math/math_function.h" namespace paddle { namespace operators { @@ -46,9 +46,10 @@ class MulKernel : public framework::OpKernel { if (z_dim.size() != 2) { z->Resize({x_matrix.dims()[0], y_matrix.dims()[1]}); } - math::matmul( - context.template device_context(), x_matrix, false, - y_matrix, false, static_cast(1), z, static_cast(0)); + + auto blas = math::GetBlas(context); + + blas.MatMul(x_matrix, y_matrix, z); if (z_dim.size() != 2) { z->Resize(z_dim); } @@ -79,6 +80,7 @@ class MulGradKernel : public framework::OpKernel { Tensor* dx = ctx.Output(framework::GradVarName("X")); Tensor* dy = ctx.Output(framework::GradVarName("Y")); auto& dev_ctx = ctx.template device_context(); + auto blas = math::GetBlas(dev_ctx); if (dx) { dx->mutable_data(ctx.GetPlace()); Tensor dx_matrix = dx->dims().size() > 2 @@ -86,8 +88,7 @@ class MulGradKernel : public framework::OpKernel { : *dx; // dx = dout * y'. dx: M x K, dout : M x N, y : K x N - math::matmul(dev_ctx, dout_mat, false, y_matrix, true, - 1, &dx_matrix, 0); + blas.MatMul(dout_mat, false, y_matrix, true, &dx_matrix); } if (dy) { dy->mutable_data(ctx.GetPlace()); @@ -95,8 +96,7 @@ class MulGradKernel : public framework::OpKernel { ? framework::ReshapeToMatrix(*dy, y_num_col_dims) : *dy; // dy = x' * dout. dy K x N, dout : M x N, x : M x K - math::matmul(dev_ctx, x_matrix, true, dout_mat, false, - 1, &dy_matrix, 0); + blas.MatMul(x_matrix, true, dout_mat, false, &dy_matrix); } } }; diff --git a/paddle/fluid/operators/multiplex_op.cc b/paddle/fluid/operators/multiplex_op.cc index b698c1bf8a05e053db07db34712a13c8074ee4d0..a4363fd25d57edb5c2509904a1f55634832613be 100644 --- a/paddle/fluid/operators/multiplex_op.cc +++ b/paddle/fluid/operators/multiplex_op.cc @@ -61,8 +61,7 @@ class MultiplexOp : public framework::OperatorWithKernel { class MultiplexOpMaker : public framework::OpProtoAndCheckerMaker { public: - MultiplexOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Ids", "The index tensor of multiplex operator."); AddInput("X", "The candidate tensors of multiplex operator.") .AsDuplicable(); diff --git a/paddle/fluid/operators/multiplex_op.cu b/paddle/fluid/operators/multiplex_op.cu index 45a2550793511f7cb8c20644ac79e9e88629ce7b..2f8a602f3c5c0a7c262235f99943ce336e20a7b4 100644 --- a/paddle/fluid/operators/multiplex_op.cu +++ b/paddle/fluid/operators/multiplex_op.cu @@ -33,7 +33,7 @@ class MultiplexGPUKernel : public framework::OpKernel { auto cols = ins[0]->numel() / rows; // copy index to cpu Tensor index_t_cpu; - TensorCopy(*ids, platform::CPUPlace(), ctx.device_context(), &index_t_cpu); + TensorCopySync(*ids, platform::CPUPlace(), &index_t_cpu); auto* index = index_t_cpu.data(); auto stream = ctx.cuda_device_context().stream(); platform::CUDAPlace place = boost::get(ctx.GetPlace()); @@ -69,7 +69,7 @@ class MultiplexGradGPUKernel : public framework::OpKernel { auto cols = ins[0]->numel() / rows; // copy index to cpu Tensor index_t_cpu; - TensorCopy(*ids, platform::CPUPlace(), ctx.device_context(), &index_t_cpu); + TensorCopySync(*ids, platform::CPUPlace(), &index_t_cpu); auto* index = index_t_cpu.data(); auto stream = ctx.cuda_device_context().stream(); diff --git a/paddle/fluid/operators/nccl/nccl_gpu_common.h b/paddle/fluid/operators/nccl/nccl_gpu_common.h index 113f93e346681e568524f9fb6a0ab9a56de8569e..558ff4cc09603eebbcd95a234ff1aa63ada7fbb2 100644 --- a/paddle/fluid/operators/nccl/nccl_gpu_common.h +++ b/paddle/fluid/operators/nccl/nccl_gpu_common.h @@ -15,9 +15,9 @@ limitations under the License. */ #pragma once #include -#include +#include // NOLINT #include -#include +#include // NOLINT #include #include #include diff --git a/paddle/fluid/operators/nccl_op.cc b/paddle/fluid/operators/nccl_op.cc index 5e4ed886b10bd48bf991ce84a9099611cf5d1d26..0018139cb06fe0573565c920849843e674df6f4c 100644 --- a/paddle/fluid/operators/nccl_op.cc +++ b/paddle/fluid/operators/nccl_op.cc @@ -76,8 +76,7 @@ class NCCLInitOpShapeInference : public framework::InferShapeBase { class NCCLInitOpMaker : public framework::OpProtoAndCheckerMaker { public: - NCCLInitOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput(kParallelScopes, "The working place of parallel do."); AddOutput("Communicator", "Create Communicator for communicating between gpus"); @@ -118,8 +117,7 @@ class NCCLAllReduceOp : public framework::OperatorWithKernel { // AllReduceOp class NCCLAllReduceOpMaker : public framework::OpProtoAndCheckerMaker { public: - NCCLAllReduceOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The input of AllReduce op"); AddInput("Communicator", "Communicator for communicating between gpus"); AddOutput("Out", "The output of AllReduce op"); @@ -165,8 +163,7 @@ class NCCLReduceOp : public framework::OperatorWithKernel { // ReduceOp class NCCLReduceOpMaker : public framework::OpProtoAndCheckerMaker { public: - NCCLReduceOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The input of Reduce op"); AddInput("Communicator", "Communicator for communicating between gpus"); AddOutput("Out", "The output of Reduce op"); @@ -214,8 +211,7 @@ class NCCLBcastOp : public framework::OperatorWithKernel { // BcastOp class NCCLBcastOpMaker : public framework::OpProtoAndCheckerMaker { public: - NCCLBcastOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The input of BcastSend op"); AddInput("Communicator", "Communicator for communicating between gpus"); AddOutput("Out", "The output of Bcast"); diff --git a/paddle/fluid/operators/nccl_op_test.cu.cc b/paddle/fluid/operators/nccl_op_test.cu.cc index 20b8a5c98ab16ac8121cb2fd01deb8ecc1966d44..ef54d79fdf2becde98c68044d14bd4347773b975 100644 --- a/paddle/fluid/operators/nccl_op_test.cu.cc +++ b/paddle/fluid/operators/nccl_op_test.cu.cc @@ -228,10 +228,8 @@ TEST_F(NCCLTester, ncclReduceOp) { result_tensor->Resize(kDims); auto *ct = result_tensor->mutable_data(cpu_place); - paddle::memory::Copy( - cpu_place, ct, p::CUDAPlace(gpu_list_[kRoot]), rt, - recv_tensor.numel() * sizeof(float), - static_cast(dev_ctxs_[kRoot])->stream()); + paddle::memory::Copy(cpu_place, ct, p::CUDAPlace(gpu_list_[kRoot]), rt, + recv_tensor.numel() * sizeof(float), nullptr); for (int64_t j = 0; j < f::product(kDims); ++j) { ASSERT_NEAR(ct[j], expected_result, 1e-5); diff --git a/paddle/fluid/operators/nce_op.cc b/paddle/fluid/operators/nce_op.cc index 192bdf8ea553f3a82066f8562458d286ee15a6ee..06092e680a1efbef379ccf40fdf476769f820429 100644 --- a/paddle/fluid/operators/nce_op.cc +++ b/paddle/fluid/operators/nce_op.cc @@ -75,8 +75,7 @@ class NCEOp : public framework::OperatorWithKernel { class NCEOpMaker : public framework::OpProtoAndCheckerMaker { public: - NCEOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Input", "(Tensor) A tensor of shape [batch_size, dim]."); AddInput( "Label", diff --git a/paddle/fluid/operators/norm_op.cc b/paddle/fluid/operators/norm_op.cc index 30a991224fa184257a8e59af5e6a27a0b0a4da86..cdbc975c02214721ceae3a338741101ef32d7ee9 100644 --- a/paddle/fluid/operators/norm_op.cc +++ b/paddle/fluid/operators/norm_op.cc @@ -19,8 +19,7 @@ namespace operators { template class NormOpMaker : public framework::OpProtoAndCheckerMaker { public: - NormOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput( "X", "(Tensor) The input tensor of norm operator. " diff --git a/paddle/fluid/operators/one_hot_op.cc b/paddle/fluid/operators/one_hot_op.cc index 1d42dfdd765166c9596abc08ce8abd534453bc63..4fcb1d69935175c3f643db7a4da04db34492f8fb 100644 --- a/paddle/fluid/operators/one_hot_op.cc +++ b/paddle/fluid/operators/one_hot_op.cc @@ -46,8 +46,7 @@ class OneHotOp : public framework::OperatorWithKernel { class OneHotOpMaker : public framework::OpProtoAndCheckerMaker { public: - OneHotOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(LoDTensor, LoDTensor) Input variable with rank at least 2. " "The last dimension of X should be 1. Each value of X is an index " diff --git a/paddle/fluid/operators/one_hot_op.cu b/paddle/fluid/operators/one_hot_op.cu index 240ac895e2c8391322411d347384f4834995eb7c..625065692c1f32c89d9e566d00051e237ac9a3af 100644 --- a/paddle/fluid/operators/one_hot_op.cu +++ b/paddle/fluid/operators/one_hot_op.cu @@ -13,7 +13,7 @@ // limitations under the License. #include "paddle/fluid/operators/one_hot_op.h" -#include "paddle/fluid/platform/cuda_helper.h" +#include "paddle/fluid/platform/cuda_primitives.h" #include "paddle/fluid/platform/gpu_info.h" namespace paddle { diff --git a/paddle/fluid/operators/pad_op.cc b/paddle/fluid/operators/pad_op.cc index d2a0106f80144e3550d73ea22f8e012426eb01ae..d4b631a6f5bf9332f4ed1d1a4bda529fbb6ada0a 100644 --- a/paddle/fluid/operators/pad_op.cc +++ b/paddle/fluid/operators/pad_op.cc @@ -48,8 +48,7 @@ class PadOp : public framework::OperatorWithKernel { class PadOpMaker : public framework::OpProtoAndCheckerMaker { public: - PadOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The input of pad op. " "The input should be a k-D tensor(k > 0 and k < 7)"); diff --git a/paddle/fluid/operators/parallel_do_op.cc b/paddle/fluid/operators/parallel_do_op.cc index ae34fe2184b43cc104c14672dec30efd3b0e9f3b..1012640d5e2052e4f347ad458cea9072a004f334 100644 --- a/paddle/fluid/operators/parallel_do_op.cc +++ b/paddle/fluid/operators/parallel_do_op.cc @@ -196,8 +196,7 @@ class ParallelDoOp : public framework::OperatorBase { class ParallelDoOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - ParallelDoOpProtoMaker(OpProto *proto, framework::OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput(kInputs, "").AsDuplicable(); AddInput(kParameters, "").AsDuplicable(); AddInput(kPlaces, ""); diff --git a/paddle/fluid/operators/pool_cudnn_op.cu.cc b/paddle/fluid/operators/pool_cudnn_op.cu.cc index 39c862b03ad497dca5c38ccecff20be510ab60e5..d60a99994edc926456706ff6a3ba998a3e5e7dd5 100644 --- a/paddle/fluid/operators/pool_cudnn_op.cu.cc +++ b/paddle/fluid/operators/pool_cudnn_op.cu.cc @@ -174,7 +174,8 @@ REGISTER_OP_KERNEL(pool2d_grad, CUDNN, plat::CUDAPlace, REGISTER_OP_KERNEL(pool3d, CUDNN, plat::CUDAPlace, ops::PoolCUDNNOpKernel, - ops::PoolCUDNNOpKernel); + ops::PoolCUDNNOpKernel, + ops::PoolCUDNNOpKernel); REGISTER_OP_KERNEL(pool3d_grad, CUDNN, plat::CUDAPlace, ops::PoolCUDNNGradOpKernel, ops::PoolCUDNNGradOpKernel); diff --git a/paddle/fluid/operators/pool_mkldnn_op.cc b/paddle/fluid/operators/pool_mkldnn_op.cc index 63eaaedcd5fc3df17902511dc02b25bf43ccd241..60e936298defe7c6ce8a33bdc7de05b52eb950e7 100644 --- a/paddle/fluid/operators/pool_mkldnn_op.cc +++ b/paddle/fluid/operators/pool_mkldnn_op.cc @@ -18,6 +18,26 @@ limitations under the License. */ namespace paddle { namespace operators { +using mkldnn::memory; // Note: paddle has also "memory" namespace +using mkldnn::pooling_forward; +using mkldnn::pooling_backward; + +// Generate keys for storing/retriving primitives for this operator +// TODO(jczaja): Make hashing function more optimial +static std::string gethash(memory::dims& input_dims, std::string& pooling_type, + std::vector& ksize, std::vector& strides, + std::vector& paddings, std::string suffix) { + auto dims2str = [](memory::dims& operand_dims) { + std::string dstr = ""; + for (size_t i = 0; i < operand_dims.size(); ++i) { + dstr += std::to_string(operand_dims[i]) + "-"; + } + return dstr; + }; + return dims2str(input_dims) + dims2str(ksize) + dims2str(strides) + + dims2str(paddings) + pooling_type + suffix; +} + template class PoolMKLDNNOpKernel : public paddle::framework::OpKernel { public: @@ -34,10 +54,6 @@ class PoolMKLDNNOpKernel : public paddle::framework::OpKernel { // Get an unique name from "argument" name of "Out" variable // This name will be used as key when saving info into device context - const std::string key = ctx.op().Output("Out"); - const std::string key_pool_pd = key + "@pool_pd"; - const std::string key_pool_workspace_memory = - key + "@pool_workspace_memory"; std::string pooling_type = ctx.Attr("pooling_type"); std::vector ksize = ctx.Attr>("ksize"); @@ -63,37 +79,71 @@ class PoolMKLDNNOpKernel : public paddle::framework::OpKernel { std::vector src_tz = paddle::framework::vectorize2int(input->dims()); std::vector dst_tz = paddle::framework::vectorize2int(output->dims()); - // TODO(pzelazko-intel): support more formats - auto src_md = platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32, - mkldnn::memory::format::nchw); - auto dst_md = platform::MKLDNNMemDesc(dst_tz, mkldnn::memory::f32, - mkldnn::memory::format::nchw); - - std::shared_ptr pool_pd = - CreatePrimitiveDesc(src_md, dst_md, strides, paddings, ksize, - pooling_type, mkldnn_engine); - - // save pool_pd into global device context to be referred in backward path - dev_ctx.SetBlob(key_pool_pd, pool_pd); - - std::shared_ptr workspace_memory = - CreateWorkspaceMemory(pool_pd, pooling_type, mkldnn_engine); - - // save pool_workspace_memory to be referred in backward path - dev_ctx.SetBlob(key_pool_workspace_memory, workspace_memory); - - auto src_memory = - mkldnn::memory({src_md, mkldnn_engine}, - static_cast(const_cast(input_data))); - auto dst_memory = - mkldnn::memory({dst_md, mkldnn_engine}, - static_cast(const_cast(output_data))); + const std::string key = gethash(src_tz, pooling_type, ksize, strides, + paddings, ctx.op().Output("Out")); + const std::string key_pool_p = key + "@pool_p"; + const std::string key_pool_pd = key + "@pool_pd"; + const std::string key_pool_src_mem_p = key + "@pool_src_mem_p"; + const std::string key_pool_dst_mem_p = key + "@pool_dst_mem_p"; + const std::string key_pool_workspace_memory = + key + "@pool_workspace_memory"; - auto pool_prim = mkldnn::pooling_forward(*pool_pd, src_memory, dst_memory, - *workspace_memory); + auto pool_p = + std::static_pointer_cast(dev_ctx.GetBlob(key_pool_p)); + if (pool_p == nullptr) { + // TODO(pzelazko-intel): support more formats + + auto src_md = + platform::MKLDNNMemDesc(src_tz, platform::MKLDNNGetDataType(), + mkldnn::memory::format::nchw); + auto dst_md = + platform::MKLDNNMemDesc(dst_tz, platform::MKLDNNGetDataType(), + mkldnn::memory::format::nchw); + + std::shared_ptr pool_pd = + CreatePrimitiveDesc(src_md, dst_md, strides, paddings, ksize, + pooling_type, mkldnn_engine); + + // save pool_pd into global device context to be referred in backward path + dev_ctx.SetBlob(key_pool_pd, pool_pd); + + std::shared_ptr workspace_memory = + CreateWorkspaceMemory(pool_pd, pooling_type, mkldnn_engine); + + // save pool_workspace_memory to be referred in backward path + dev_ctx.SetBlob(key_pool_workspace_memory, workspace_memory); + + auto pool_src_memory_p = std::make_shared( + memory::primitive_desc{src_md, mkldnn_engine}, + static_cast(const_cast(input_data))); + dev_ctx.SetBlob(key_pool_src_mem_p, pool_src_memory_p); + + auto pool_dst_memory_p = std::make_shared( + memory::primitive_desc{dst_md, mkldnn_engine}, + static_cast(output_data)); + dev_ctx.SetBlob(key_pool_dst_mem_p, pool_dst_memory_p); + + pool_p = std::make_shared( + *pool_pd, *(pool_src_memory_p.get()), *(pool_dst_memory_p.get()), + *workspace_memory); + dev_ctx.SetBlob(key_pool_p, pool_p); + } else { + // Primitives already exist + auto pool_src_memory_p = + std::static_pointer_cast(dev_ctx.GetBlob(key_pool_src_mem_p)); + PADDLE_ENFORCE(pool_src_memory_p != nullptr, + "Fail to find pooling src mem_p in device context"); + auto pool_dst_memory_p = + std::static_pointer_cast(dev_ctx.GetBlob(key_pool_dst_mem_p)); + PADDLE_ENFORCE(pool_dst_memory_p != nullptr, + "Fail to find pooling dst mem_p in device context"); + pool_src_memory_p->set_data_handle( + reinterpret_cast(const_cast(input_data))); + pool_dst_memory_p->set_data_handle(output_data); + } // push primitive to stream and wait until it's executed - std::vector pipeline{pool_prim}; + std::vector pipeline{*(pool_p.get())}; mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); } @@ -120,9 +170,10 @@ class PoolMKLDNNOpKernel : public paddle::framework::OpKernel { mkldnn::memory::primitive_desc workspace_md = pooling_type == "max" ? pool_pd->workspace_primitive_desc() - : mkldnn::memory::primitive_desc( - {{}, mkldnn::memory::f32, mkldnn::memory::format::nchw}, - engine); + : mkldnn::memory::primitive_desc({{}, + platform::MKLDNNGetDataType(), + mkldnn::memory::format::nchw}, + engine); auto p_workspace_memory = new mkldnn::memory(workspace_md); return std::unique_ptr(p_workspace_memory); @@ -140,13 +191,6 @@ class PoolMKLDNNGradOpKernel : public paddle::framework::OpKernel { const Tensor* out_grad = ctx.Input(framework::GradVarName("Out")); Tensor* in_x_grad = ctx.Output(framework::GradVarName("X")); - // Get an unique name from "argument" name of "Out" variable - // This name will be used as key when referring info from device context - const std::string key = ctx.op().Input("Out"); - const std::string key_pool_pd = key + "@pool_pd"; - const std::string key_pool_workspace_memory = - key + "@pool_workspace_memory"; - std::string pooling_type = ctx.Attr("pooling_type"); std::vector ksize = ctx.Attr>("ksize"); std::vector strides = ctx.Attr>("strides"); @@ -171,43 +215,76 @@ class PoolMKLDNNGradOpKernel : public paddle::framework::OpKernel { std::vector diff_dst_tz = paddle::framework::vectorize2int(out_grad->dims()); - auto diff_src_md = platform::MKLDNNMemDesc(diff_src_tz, mkldnn::memory::f32, - mkldnn::memory::format::nchw); - auto diff_dst_md = platform::MKLDNNMemDesc(diff_dst_tz, mkldnn::memory::f32, - mkldnn::memory::format::nchw); - - // Retrieve pool_pd/pool_workspace_memory from device context - auto pool_pd = - std::static_pointer_cast( - dev_ctx.GetBlob(key_pool_pd)); - PADDLE_ENFORCE(pool_pd != nullptr, - "Fail to find pool_pd in device context"); - - auto workspace_memory = std::static_pointer_cast( - dev_ctx.GetBlob(key_pool_workspace_memory)); - PADDLE_ENFORCE(workspace_memory != nullptr, - "Fail to find workspace_memory in device context"); - - auto pool_bwd_desc = mkldnn::pooling_backward::desc( - pooling_type == "max" ? mkldnn::algorithm::pooling_max - : mkldnn::algorithm::pooling_avg, - diff_src_md, diff_dst_md, strides, ksize, paddings, paddings, - mkldnn::padding_kind::zero); - auto pool_bwd_pd = mkldnn::pooling_backward::primitive_desc( - pool_bwd_desc, mkldnn_engine, *pool_pd); - - auto diff_src_memory = - mkldnn::memory({diff_src_md, mkldnn_engine}, - static_cast(const_cast(in_x_grad_data))); - auto diff_dst_memory = - mkldnn::memory({diff_dst_md, mkldnn_engine}, - static_cast(const_cast(out_grad_data))); + // Get an unique name from "argument" name of "Out" variable + // This name will be used as key when referring info from device context + const std::string key = gethash(diff_src_tz, pooling_type, ksize, strides, + paddings, ctx.op().Input("Out")); + const std::string key_pool_bwd_p = key + "@pool_bwd_p"; + const std::string key_pool_diff_src_mem_p = key + "@pool_diff_src_mem_p"; + const std::string key_pool_diff_dst_mem_p = key + "@pool_diff_dst_mem_p"; + const std::string key_pool_pd = key + "@pool_pd"; + const std::string key_pool_workspace_memory = + key + "@pool_workspace_memory"; - auto bwd_prim = mkldnn::pooling_backward( - pool_bwd_pd, diff_dst_memory, *workspace_memory, diff_src_memory); + auto pool_bwd_p = std::static_pointer_cast( + dev_ctx.GetBlob(key_pool_bwd_p)); + if (pool_bwd_p == nullptr) { + auto diff_src_md = + platform::MKLDNNMemDesc(diff_src_tz, platform::MKLDNNGetDataType(), + mkldnn::memory::format::nchw); + auto diff_dst_md = + platform::MKLDNNMemDesc(diff_dst_tz, platform::MKLDNNGetDataType(), + mkldnn::memory::format::nchw); + // Retrieve pool_pd/pool_workspace_memory from device context + auto pool_pd = + std::static_pointer_cast( + dev_ctx.GetBlob(key_pool_pd)); + PADDLE_ENFORCE(pool_pd != nullptr, + "Fail to find pool_pd in device context"); + + auto workspace_memory = std::static_pointer_cast( + dev_ctx.GetBlob(key_pool_workspace_memory)); + PADDLE_ENFORCE(workspace_memory != nullptr, + "Fail to find workspace_memory in device context"); + + auto pool_diff_src_memory_p = std::make_shared(memory( + {diff_src_md, mkldnn_engine}, static_cast(in_x_grad_data))); + dev_ctx.SetBlob(key_pool_diff_src_mem_p, pool_diff_src_memory_p); + + auto pool_diff_dst_memory_p = std::make_shared( + memory({diff_dst_md, mkldnn_engine}, + static_cast(const_cast(out_grad_data)))); + dev_ctx.SetBlob(key_pool_diff_dst_mem_p, pool_diff_dst_memory_p); + + auto pool_bwd_desc = mkldnn::pooling_backward::desc( + pooling_type == "max" ? mkldnn::algorithm::pooling_max + : mkldnn::algorithm::pooling_avg, + diff_src_md, diff_dst_md, strides, ksize, paddings, paddings, + mkldnn::padding_kind::zero); + auto pool_bwd_pd = mkldnn::pooling_backward::primitive_desc( + pool_bwd_desc, mkldnn_engine, *pool_pd); + + pool_bwd_p = std::make_shared( + pool_bwd_pd, *(pool_diff_dst_memory_p.get()), *workspace_memory, + *(pool_diff_src_memory_p)); + dev_ctx.SetBlob(key_pool_bwd_p, pool_bwd_p); + } else { + // Primitives already exist + auto pool_diff_src_memory_p = std::static_pointer_cast( + dev_ctx.GetBlob(key_pool_diff_src_mem_p)); + PADDLE_ENFORCE(pool_diff_src_memory_p != nullptr, + "Fail to find pooling src mem_p in device context"); + auto pool_diff_dst_memory_p = std::static_pointer_cast( + dev_ctx.GetBlob(key_pool_diff_dst_mem_p)); + PADDLE_ENFORCE(pool_diff_dst_memory_p != nullptr, + "Fail to find pooling dst mem_p in device context"); + pool_diff_src_memory_p->set_data_handle( + reinterpret_cast(in_x_grad_data)); + pool_diff_dst_memory_p->set_data_handle(const_cast(out_grad_data)); + } // push primitive to stream and wait until it's executed - std::vector pipeline{bwd_prim}; + std::vector pipeline{*(pool_bwd_p.get())}; mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); } // Compute() }; diff --git a/paddle/fluid/operators/pool_op.cc b/paddle/fluid/operators/pool_op.cc index f2de075e0d82fc5bd0ac41b481ac80314f3857a3..f4fb2b132fe8d59cb50f5a1f7359240ac50445fe 100644 --- a/paddle/fluid/operators/pool_op.cc +++ b/paddle/fluid/operators/pool_op.cc @@ -135,8 +135,7 @@ framework::OpKernelType PoolOpGrad::GetExpectedKernelType( library_); } -Pool2dOpMaker::Pool2dOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { +void Pool2dOpMaker::Make() { AddInput( "X", "(Tensor) The input tensor of pooling operator. " @@ -229,8 +228,7 @@ Example: )DOC"); } -Pool3dOpMaker::Pool3dOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { +void Pool3dOpMaker::Make() { AddInput("X", "(Tensor) The input tensor of pooling operator. " "The format of input tensor is NCDHW, where N is batch size, C is " diff --git a/paddle/fluid/operators/pool_op.h b/paddle/fluid/operators/pool_op.h index a48127ea6983d3d4ea12ec4925f30af233002ef2..a63963ca926bb94ff99e5cfe6dbcb2b15075bcb8 100644 --- a/paddle/fluid/operators/pool_op.h +++ b/paddle/fluid/operators/pool_op.h @@ -50,12 +50,12 @@ class PoolOpGrad : public framework::OperatorWithKernel { class Pool2dOpMaker : public framework::OpProtoAndCheckerMaker { public: - Pool2dOpMaker(OpProto* proto, OpAttrChecker* op_checker); + void Make() override; }; class Pool3dOpMaker : public framework::OpProtoAndCheckerMaker { public: - Pool3dOpMaker(OpProto* proto, OpAttrChecker* op_checker); + void Make() override; }; template diff --git a/paddle/fluid/operators/pool_with_index_op.cc b/paddle/fluid/operators/pool_with_index_op.cc index 848cd61b23c2389d3fe11f585b256d55c1ff177f..873706593e4c856f0079738654a9e7e59a1c0cd8 100644 --- a/paddle/fluid/operators/pool_with_index_op.cc +++ b/paddle/fluid/operators/pool_with_index_op.cc @@ -100,8 +100,7 @@ class MaxPoolWithIndexOpGrad : public framework::OperatorWithKernel { class MaxPool2dWithIndexOpMaker : public framework::OpProtoAndCheckerMaker { public: - MaxPool2dWithIndexOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput( "X", "(Tensor) The input tensor of pooling operator. " @@ -177,8 +176,7 @@ Example: class MaxPool3dWithIndexOpMaker : public framework::OpProtoAndCheckerMaker { public: - MaxPool3dWithIndexOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor) The input tensor of pooling operator. " "The format of input tensor is NCDHW, where N is batch size, C is " diff --git a/paddle/fluid/operators/positive_negative_pair_op.cc b/paddle/fluid/operators/positive_negative_pair_op.cc index d237da25a00de13057e009b6705d3241b8b26539..4d865b7f17b050ac6f04addc9949f3f65da06ded 100644 --- a/paddle/fluid/operators/positive_negative_pair_op.cc +++ b/paddle/fluid/operators/positive_negative_pair_op.cc @@ -95,8 +95,7 @@ class PositiveNegativePairOp : public framework::OperatorWithKernel { class PositiveNegativePairOpMaker : public framework::OpProtoAndCheckerMaker { public: - PositiveNegativePairOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Score", "(Tensor, float) Model Score on an item (with " "respect to QueryID). It's a 2-D tensor with shape [batch_size, " diff --git a/paddle/fluid/operators/precision_recall_op.cc b/paddle/fluid/operators/precision_recall_op.cc index c34b0d072bdb2f5b97dd4615ff9338d98f2bfbe5..e7ce16f33fb5052ffb41fc05bd1538e2f0dc35be 100644 --- a/paddle/fluid/operators/precision_recall_op.cc +++ b/paddle/fluid/operators/precision_recall_op.cc @@ -90,8 +90,7 @@ class PrecisionRecallOp : public framework::OperatorWithKernel { class PrecisionRecallOpMaker : public framework::OpProtoAndCheckerMaker { public: - PrecisionRecallOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("MaxProbs", "(Tensor, default Tensor) A 2-D tensor with shape N x 1, " "where N is the batch size. Each row contains the max probability " diff --git a/paddle/fluid/operators/prefetch_op.cc b/paddle/fluid/operators/prefetch_op.cc index f9ae01ab5d2972d2a74b36ae6035985d1d874bb6..e0a9b24ac8978418a1a4ece62286e022bec8b834 100644 --- a/paddle/fluid/operators/prefetch_op.cc +++ b/paddle/fluid/operators/prefetch_op.cc @@ -41,12 +41,7 @@ class PrefetchOp : public framework::OperatorBase { platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); auto& ctx = *pool.Get(place); - auto client_var_name = Output("RPCClient"); - PADDLE_ENFORCE_NOT_NULL(scope.FindVar(client_var_name), - "Can not find variable '%s' in the scope.", - client_var_name); - auto* client_var = scope.FindVar(client_var_name); - detail::RPCClient* rpc_client = client_var->GetMutable(); + auto rpc_client = detail::RPCClient::GetInstance(); for (size_t i = 0; i < ins.size(); i++) { if (NeedSend(scope, ins[i])) { @@ -64,12 +59,8 @@ class PrefetchOp : public framework::OperatorBase { class PrefetchOpMaker : public framework::OpProtoAndCheckerMaker { public: - PrefetchOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() { AddInput("X", "(LoDTensor) Input Id variables to be sent").AsDuplicable(); - AddOutput("RPCClient", - "(RPCClient) The RPC client object which will be" - "initialized at most once."); AddOutput("Out", "(LoDTensor) result " "to be fetched from parameter server") @@ -88,17 +79,6 @@ the parameter server and fetch result back. } }; -class PrefetchOpVarTypeInference : public framework::VarTypeInference { - public: - void operator()(const framework::OpDesc& op_desc, - framework::BlockDesc* block) const override { - auto out_var_name = op_desc.Output("RPCClient").front(); - auto& out_var = block->FindRecursiveOrCreateVar(out_var_name); - auto var_type = framework::proto::VarType::RAW; - out_var.SetType(var_type); - } -}; - class PrefetchOpShapeInference : public framework::InferShapeBase { public: void operator()(framework::InferShapeContext* ctx) const override {} @@ -111,5 +91,4 @@ namespace ops = paddle::operators; REGISTER_OPERATOR(prefetch, ops::PrefetchOp, paddle::framework::EmptyGradOpMaker, ops::PrefetchOpMaker, - ops::PrefetchOpVarTypeInference, ops::PrefetchOpShapeInference); diff --git a/paddle/fluid/operators/prelu_op.cc b/paddle/fluid/operators/prelu_op.cc index a066b3e06e5eca2661827425b5b2d0059d5bcc3c..db040509bc08c3f6ad031c5b97c93574e31337e0 100644 --- a/paddle/fluid/operators/prelu_op.cc +++ b/paddle/fluid/operators/prelu_op.cc @@ -38,8 +38,7 @@ class PReluOp : public framework::OperatorWithKernel { class PReluOpMaker : public framework::OpProtoAndCheckerMaker { public: - PReluOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The input tensor of prelu operator."); AddInput("Alpha", "The alpha weight of prelu operator."); AddOutput("Out", "The output tensor of prelu operator."); diff --git a/paddle/fluid/operators/print_op.cc b/paddle/fluid/operators/print_op.cc index fafc7e54d7a44d6bb2dadf67135537dc16430e76..db7634918a5179a61304315ecd08350d23fb4642 100644 --- a/paddle/fluid/operators/print_op.cc +++ b/paddle/fluid/operators/print_op.cc @@ -209,8 +209,7 @@ class TensorPrintOp : public framework::OperatorBase { class PrintOpProtoAndCheckMaker : public framework::OpProtoAndCheckerMaker { public: - PrintOpProtoAndCheckMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("In", "Input tensor to be displayed."); AddAttr("first_n", "Only log `first_n` number of times."); AddAttr("message", "A string message to print as a prefix."); diff --git a/paddle/fluid/operators/proximal_adagrad_op.cc b/paddle/fluid/operators/proximal_adagrad_op.cc index 38cd97c17b16a4cc64f7e6d52150fae392df6036..8d8075d76111928ec9855eb0b70fe6dbd90a979b 100644 --- a/paddle/fluid/operators/proximal_adagrad_op.cc +++ b/paddle/fluid/operators/proximal_adagrad_op.cc @@ -17,6 +17,7 @@ limitations under the License. */ namespace paddle { namespace operators { +using Tensor = framework::Tensor; class ProximalAdagradOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; @@ -55,12 +56,17 @@ class ProximalAdagradOp : public framework::OperatorWithKernel { ctx->SetOutputDim("ParamOut", param_dim); ctx->SetOutputDim("MomentOut", param_dim); } + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext &ctx) const override { + auto input_data_type = + framework::ToDataType(ctx.Input("Param")->type()); + return framework::OpKernelType(input_data_type, ctx.GetPlace()); + } }; class ProximalAdagradOpMaker : public framework::OpProtoAndCheckerMaker { public: - ProximalAdagradOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Param", "(Tensor, default Tensor) " "Input parameter that has to be updated."); diff --git a/paddle/fluid/operators/proximal_gd_op.cc b/paddle/fluid/operators/proximal_gd_op.cc index efb4e1ac204ce79bfad7d77038f342be09e8f0e8..baf9cbcba2ed89f62afc9816e0ab9e0f112e6008 100644 --- a/paddle/fluid/operators/proximal_gd_op.cc +++ b/paddle/fluid/operators/proximal_gd_op.cc @@ -17,6 +17,7 @@ limitations under the License. */ namespace paddle { namespace operators { +using Tensor = framework::Tensor; class ProximalGDOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; @@ -43,12 +44,17 @@ class ProximalGDOp : public framework::OperatorWithKernel { ctx->SetOutputDim("ParamOut", param_dim); } + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext &ctx) const override { + auto input_data_type = + framework::ToDataType(ctx.Input("Param")->type()); + return framework::OpKernelType(input_data_type, ctx.GetPlace()); + } }; class ProximalGDOpMaker : public framework::OpProtoAndCheckerMaker { public: - ProximalGDOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Param", "(Tensor, default Tensor) " "Input parameter value that has to be updated."); diff --git a/paddle/fluid/operators/random_crop_op.cc b/paddle/fluid/operators/random_crop_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..b14b559e31dd422f8ebe4002988a9746dfdf28a2 --- /dev/null +++ b/paddle/fluid/operators/random_crop_op.cc @@ -0,0 +1,81 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. +#include "paddle/fluid/operators/random_crop_op.h" + +namespace paddle { +namespace operators { + +class RandomCropOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + return framework::OpKernelType( + framework::ToDataType(ctx.Input("X")->type()), + ctx.device_context()); + } +}; + +class RandomCropOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("X", "A batch of instances to random crop."); + AddInput("Seed", "The random seed."); + AddOutput("Out", "The cropped instance batch."); + AddOutput("SeedOut", "The random seed after random cropping.") + .AsDispensable(); + AddAttr>("shape", "The shape of a cropped instance."); + AddComment(R"DOC( + This operator takes a batch of instance, and do random cropping on each instance. + It means that cropping positions differs on each instance, which is determined + by an uniform random generator. All cropped instances have the same shape, which + is determined by the operator's attribute 'shape'. + )DOC"); + } +}; + +class RandomCropOpInferShape : public framework::InferShapeBase { + public: + void operator()(framework::InferShapeContext* ctx) const override { + auto seed_dim = ctx->GetInputDim("Seed"); + PADDLE_ENFORCE(seed_dim.size() == 1 && seed_dim[0] == 1); + auto shape = ctx->Attrs().Get>("shape"); + auto x_dim = ctx->GetInputDim("X"); + PADDLE_ENFORCE_GT(x_dim.size(), static_cast(shape.size())); + auto out_dim = framework::vectorize2int(x_dim); + for (size_t i = 1; i <= shape.size(); ++i) { + size_t x_i = x_dim.size() - i; + size_t shape_i = shape.size() - i; + PADDLE_ENFORCE_GE(x_dim[x_i], shape[shape_i]); + out_dim[x_i] = shape[shape_i]; + } + ctx->SetOutputDim("Out", framework::make_ddim(out_dim)); + ctx->SetOutputDim("SeedOut", framework::make_ddim({1})); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +namespace f = paddle::framework; +REGISTER_OPERATOR(random_crop, ops::RandomCropOp, ops::RandomCropOpMaker, + ops::RandomCropOpInferShape, f::EmptyGradOpMaker); + +template +using Kernel = ops::RandomCropKernel; +REGISTER_OP_CPU_KERNEL(random_crop, Kernel, Kernel, Kernel, + Kernel, Kernel); diff --git a/paddle/fluid/operators/random_crop_op.cu b/paddle/fluid/operators/random_crop_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..6fc9bedc55b4d349ddf3d109c7f9049113235f0c --- /dev/null +++ b/paddle/fluid/operators/random_crop_op.cu @@ -0,0 +1,21 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/operators/random_crop_op.h" + +namespace ops = paddle::operators; +template +using Kernel = ops::RandomCropKernel; +REGISTER_OP_CUDA_KERNEL(random_crop, Kernel, Kernel, Kernel, + Kernel, Kernel); diff --git a/paddle/fluid/operators/random_crop_op.h b/paddle/fluid/operators/random_crop_op.h new file mode 100644 index 0000000000000000000000000000000000000000..f3261cbdc986b0cc724315c1eb92b8b84e18c742 --- /dev/null +++ b/paddle/fluid/operators/random_crop_op.h @@ -0,0 +1,181 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/detail/safe_ref.h" +#include "paddle/fluid/platform/device_context.h" +#include "paddle/fluid/platform/for_range.h" +#ifdef PADDLE_WITH_CUDA +#include +#endif + +namespace paddle { +namespace operators { + +template +struct Random; + +template <> +struct Random { + using Engine = std::minstd_rand; + + template + using UniformIntDist = std::uniform_int_distribution; +}; + +#ifdef PADDLE_WITH_CUDA +template <> +struct Random { + using Engine = thrust::minstd_rand; + + template + using UniformIntDist = thrust::uniform_int_distribution; +}; +#endif + +template +HOSTDEVICE inline void StridedMemcpy(const T* x, const size_t* x_dims, T* out, + const size_t* out_dims, int i, int rank, + size_t prod_x_remain, + size_t prod_out_remain, + const size_t* offsets) { + size_t x_dim_i = x_dims[i]; + size_t out_dim_i = out_dims[i]; + size_t x_stride = prod_x_remain / x_dim_i; + size_t out_stride = prod_out_remain / out_dim_i; + size_t offset_i = offsets[i]; + + if (i == rank - 1) { + PADDLE_ASSERT(x_stride == 1 && out_stride == 1); + x += offset_i; + for (size_t j = 0; j < out_dim_i; ++j) { + *out++ = *x++; + } + } else { + x += offset_i * x_stride; + for (size_t j = 0; j < out_dim_i; ++j) { + StridedMemcpy(x, x_dims, out, out_dims, i + 1, rank, x_stride, + out_stride, offsets); + x += x_stride; + out += out_stride; + } + } +} + +template +struct RandomCropFunctor { + const T* x_; + T* out_; + size_t x_dims_[9]; + size_t out_dims_[9]; + int num_batchsize_dims_; + int rank_; + int64_t seed_; + + size_t prod_batchsize_dims_; + size_t prod_x_ins_dims_; + size_t prod_out_ins_dims_; + + RandomCropFunctor(const T* x, T* out, const framework::DDim& x_dims, + const framework::DDim& out_dims, int num_batchsize_dims, + int64_t seed) + : x_(x), + out_(out), + num_batchsize_dims_(num_batchsize_dims), + rank_(x_dims.size()), + seed_(seed) { + PADDLE_ENFORCE_EQ(x_dims.size(), out_dims.size()); + PADDLE_ENFORCE_GT(rank_, num_batchsize_dims_); + prod_batchsize_dims_ = 1; + prod_x_ins_dims_ = 1; + prod_out_ins_dims_ = 1; + for (size_t i = 0; i < static_cast(rank_); ++i) { + size_t x_dim_i = x_dims[i]; + size_t out_dim_i = out_dims[i]; + x_dims_[i] = x_dim_i; + out_dims_[i] = out_dim_i; + if (i < static_cast(num_batchsize_dims_)) { + PADDLE_ENFORCE_EQ(x_dim_i, out_dim_i); + prod_batchsize_dims_ *= x_dim_i; + } else { + prod_x_ins_dims_ *= x_dim_i; + prod_out_ins_dims_ *= out_dim_i; + } + } + } + + HOSTDEVICE void operator()(size_t ins_idx) { + typename Random::Engine engine(seed_); + engine.discard(ins_idx * (rank_ - num_batchsize_dims_)); + size_t offsets[9]; + for (int i = num_batchsize_dims_; i < rank_; ++i) { + typename Random::template UniformIntDist dist( + 0, x_dims_[i] - out_dims_[i]); + offsets[i - num_batchsize_dims_] = dist(engine); + } + + const T* x = x_ + ins_idx * prod_x_ins_dims_; + T* out = out_ + ins_idx * prod_out_ins_dims_; + + StridedMemcpy(x, x_dims_ + num_batchsize_dims_, out, + out_dims_ + num_batchsize_dims_, 0, + rank_ - num_batchsize_dims_, prod_x_ins_dims_, + prod_out_ins_dims_, offsets); + } +}; + +template +class RandomCropKernel : public framework::OpKernel { + public: + virtual void Compute(const framework::ExecutionContext& ctx) const { + auto& seed_tensor = detail::Ref(ctx.Input("Seed")); + int64_t seed = 0; + if (platform::is_cpu_place(seed_tensor.place())) { + seed = *seed_tensor.data(); + } else { + LOG(WARNING) << "It is slow to place seed in GPU memory. Please verify " + "your program"; + framework::LoDTensor cpu_seed; + framework::TensorCopySync(seed_tensor, platform::CPUPlace(), &cpu_seed); + seed = *cpu_seed.data(); + } + auto shape = ctx.Attr>("shape"); + auto& x = detail::Ref(ctx.Input("X")); + auto& out = detail::Ref(ctx.Output("Out")); + + int num_batchsize_dims = x.dims().size() - shape.size(); + RandomCropFunctor functor( + x.data(), out.mutable_data(ctx.GetPlace()), x.dims(), out.dims(), + num_batchsize_dims, seed); + platform::ForRange for_range( + ctx.template device_context(), + functor.prod_batchsize_dims_); + + for_range(functor); + + Random::Engine engine(seed); + engine.discard(functor.prod_batchsize_dims_ * + (functor.rank_ - functor.num_batchsize_dims_)); + *ctx.Output("SeedOut")->mutable_data( + platform::CPUPlace()) = engine(); + } +}; + +// TODO(fengjiayi): Backward of random crop op + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/rank_loss_op.cc b/paddle/fluid/operators/rank_loss_op.cc index eb9ff8de3e4b37ef0bbf7477c1bb62856bdb6310..313cf01541dd88a0f4f8bf54fe4436984c2cbcf8 100644 --- a/paddle/fluid/operators/rank_loss_op.cc +++ b/paddle/fluid/operators/rank_loss_op.cc @@ -46,8 +46,7 @@ class RankLossOp : public framework::OperatorWithKernel { class RankLossOpMaker : public framework::OpProtoAndCheckerMaker { public: - RankLossOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Label", "(2-D Tensor with shape [batch_size x 1]) " "The label indicating A ranked higher than B or not."); diff --git a/paddle/fluid/operators/read_op.cc b/paddle/fluid/operators/read_op.cc index bf02b9958927580608b95d6b8ecfddc7231a02d4..72a27d43584d55cd0859c63577ae85ff0f5fdfa8 100644 --- a/paddle/fluid/operators/read_op.cc +++ b/paddle/fluid/operators/read_op.cc @@ -79,8 +79,7 @@ class ReadOp : public framework::OperatorBase { class ReadOpMaker : public framework::OpProtoAndCheckerMaker { public: - ReadOpMaker(OpProto* op_proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(op_proto, op_checker) { + void Make() override { AddInput("Reader", "(ReaderHolder) The executed reader."); AddOutput("Out", "(LoDTensor) The output data.").AsDuplicable(); AddComment(R"DOC( diff --git a/paddle/fluid/operators/reader/CMakeLists.txt b/paddle/fluid/operators/reader/CMakeLists.txt index 3106978eb0149b14849dfd1aaad8bbe76791f2f6..62532036f86bfb82465ccd9e0ec526299489932a 100644 --- a/paddle/fluid/operators/reader/CMakeLists.txt +++ b/paddle/fluid/operators/reader/CMakeLists.txt @@ -23,6 +23,7 @@ reader_library(create_recordio_file_reader_op SRCS create_recordio_file_reader_o reader_library(create_double_buffer_reader_op SRCS create_double_buffer_reader_op.cc) reader_library(create_multi_pass_reader_op SRCS create_multi_pass_reader_op.cc) reader_library(create_threaded_reader_op SRCS create_threaded_reader_op.cc) +reader_library(create_custom_reader_op SRCS create_custom_reader_op.cc) cc_test(reader_blocking_queue_test SRCS reader_blocking_queue_test.cc) # Export local libraries to parent diff --git a/paddle/fluid/operators/reader/create_batch_reader_op.cc b/paddle/fluid/operators/reader/create_batch_reader_op.cc index 04c5872bef4600e30ba572a025cc5f0a5e9839ca..4cc7cbc6e89b0712faf9ad9c51480bce00da15f5 100644 --- a/paddle/fluid/operators/reader/create_batch_reader_op.cc +++ b/paddle/fluid/operators/reader/create_batch_reader_op.cc @@ -52,9 +52,8 @@ class CreateBatchReaderOp : public framework::OperatorBase { }; class CreateBatchReaderOpMaker : public DecoratedReaderMakerBase { - public: - CreateBatchReaderOpMaker(OpProto* op_proto, OpAttrChecker* op_checker) - : DecoratedReaderMakerBase(op_proto, op_checker) { + protected: + void Apply() override { AddAttr("batch_size", "How many instances the batch reader yields each time.") .GreaterThan(0); diff --git a/paddle/fluid/operators/reader/create_custom_reader_op.cc b/paddle/fluid/operators/reader/create_custom_reader_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..331224a59899b4a7d517ca4f7141fb5b8f4f5168 --- /dev/null +++ b/paddle/fluid/operators/reader/create_custom_reader_op.cc @@ -0,0 +1,187 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/framework/executor.h" +#include "paddle/fluid/operators/detail/safe_ref.h" +#include "paddle/fluid/operators/reader/reader_op_registry.h" + +namespace paddle { +namespace operators { +namespace reader { + +class CustomReader : public framework::DecoratedReader { + public: + CustomReader(ReaderBase* reader, const framework::BlockDesc& sub_block, + const std::vector& source_var_names, + const std::vector& sink_var_names) + : DecoratedReader(reader), + program_(*sub_block.Program()), + sub_block_id_(sub_block.ID()), + exe_(framework::Executor(platform::CPUPlace())), + source_var_names_(source_var_names), + sink_var_names_(sink_var_names) {} + + void ReadNext(std::vector* out) override; + + private: + const framework::ProgramDesc program_; + int sub_block_id_; + framework::Executor exe_; + + std::vector source_var_names_; + std::vector sink_var_names_; +}; + +class CreateCustomReaderOp : public framework::OperatorBase { + public: + using framework::OperatorBase::OperatorBase; + + private: + void RunImpl(const framework::Scope& scope, + const platform::Place& dev_place) const override { + auto* out = scope.FindVar(Output("Out")) + ->template GetMutable(); + auto* sub_block = Attr("sub_block"); + if (out->Get() != nullptr) { + return; + } + const auto& underlying_reader = scope.FindVar(Input("UnderlyingReader")) + ->Get(); + out->Reset( + new CustomReader(underlying_reader.Get(), *sub_block, + Attr>("source_var_names"), + Attr>("sink_var_names"))); + } +}; + +class CreateCustomReaderOpMaker : public DecoratedReaderMakerBase { + protected: + void Apply() override { + AddAttr( + "sub_block", "The block to hold all preprocessing operators."); + AddAttr>( + "source_var_names", + "Source variables are starting points of data preprocessing. They hold " + "preprocessing's input tensors. Each source variable corresponds to " + "one of underlying reader's output datas."); + AddAttr>( + "sink_var_names", + "Sink variables are ending points of data preprocessing. They hold " + "preprocessing's output tensors. Each sink variable corresponds to " + "one of custom reader's output datas."); + AddComment(R"DOC( + CreateCustomReader Operator + + A custom reader can be used for input data preprocessing. + A custom reader holds its own sub-block, which will be executed in CPU + in its 'ReadNext()' function. Users can configurate their own + preprocessing pipelines by inserting operators into custom reader's + sub-block. + )DOC"); + } +}; + +class CustomReaderInferShape : public framework::InferShapeBase { + public: + void operator()(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(!ctx->IsRuntime(), + "'CustomReaderInferShape' should only be invoked during " + "compile time."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "The output decorated reader should not be null."); + const auto* sub_block = + ctx->Attrs().Get("sub_block"); + const auto sink_var_names = + ctx->Attrs().Get>("sink_var_names"); + std::vector> res_dims; + std::vector res_lod_levels; + for (const std::string& var_name : sink_var_names) { + auto* sink_var = sub_block->FindVar(var_name); + PADDLE_ENFORCE_NOT_NULL(sink_var); + res_dims.emplace_back(sink_var->GetShape()); + res_lod_levels.push_back(sink_var->GetLoDLevel()); + } + auto* out_reader = + boost::get(ctx->GetOutputVarPtrs("Out")[0]); + out_reader->SetShapes(res_dims); + out_reader->SetLoDLevels(res_lod_levels); + } +}; + +class CustomReaderInferVarType : public framework::VarTypeInference { + public: + void operator()(const framework::OpDesc& op_desc, + framework::BlockDesc* block) const override { + framework::VarDesc* out_reader = block->FindVar(op_desc.Output("Out")[0]); + PADDLE_ENFORCE_NOT_NULL(out_reader); + out_reader->SetType(framework::proto::VarType::READER); + + auto sink_var_names = + boost::get>(op_desc.GetAttr("sink_var_names")); + const auto* sub_block = + boost::get(op_desc.GetAttr("sub_block")); + std::vector res_data_types; + for (const std::string& var_name : sink_var_names) { + framework::VarDesc* var = sub_block->FindVar(var_name); + PADDLE_ENFORCE_NOT_NULL(var); + res_data_types.emplace_back(var->GetDataType()); + } + out_reader->SetDataTypes(res_data_types); + } +}; + +void CustomReader::ReadNext(std::vector* out) { + out->clear(); + std::vector underlying_outs; + reader_->ReadNext(&underlying_outs); + if (underlying_outs.empty()) { + // There is not next data. + return; + } + PADDLE_ENFORCE(source_var_names_.size() == underlying_outs.size(), + "The size of source_var_names(%d) and the size of " + "underlying_outs(%d) are not consistent. Each feeding element " + "must have its own source variable.", + source_var_names_.size(), underlying_outs.size()); + // The scope for CustomReader's sub-block should be independent and shouldn't + // be any other computation scope's child. Otherwise, data preprocessing and + // compution cannot be concurrent. + framework::Scope scope; + // 1. Copy LoDTensors from underlying reader's output to source variables. + for (size_t i = 0; i < source_var_names_.size(); ++i) { + framework::Variable* var = scope.Var(source_var_names_[i]); + framework::LoDTensor* tensor = var->GetMutable(); + tensor->ShareDataWith(underlying_outs[i]); + tensor->set_lod(underlying_outs[i].lod()); + } + // 2. Run the sub-block. + exe_.Run(program_, &scope, sub_block_id_, false, true); + // 3. Copy LoDTensors from sink variables to out. + out->resize(sink_var_names_.size()); + for (size_t i = 0; i < sink_var_names_.size(); ++i) { + const auto& tensor = detail::Ref(scope.FindVar(sink_var_names_[i])) + .Get(); + framework::TensorCopySync(tensor, platform::CPUPlace(), &(*out)[i]); + } +} + +} // namespace reader +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators::reader; +REGISTER_OPERATOR(create_custom_reader, ops::CreateCustomReaderOp, + ops::CreateCustomReaderOpMaker, ops::CustomReaderInferShape, + ops::CustomReaderInferVarType, + paddle::framework::EmptyGradOpMaker) diff --git a/paddle/fluid/operators/reader/create_double_buffer_reader_op.cc b/paddle/fluid/operators/reader/create_double_buffer_reader_op.cc index 3fdc31dfa5242b6487c308d395d70d7ff348bc73..bc830a2b72e657f79f4c94e24428d38ff2b7c42e 100644 --- a/paddle/fluid/operators/reader/create_double_buffer_reader_op.cc +++ b/paddle/fluid/operators/reader/create_double_buffer_reader_op.cc @@ -113,14 +113,13 @@ class CreateDoubleBufferReaderOp : public framework::OperatorBase { }; class CreateDoubleBufferReaderOpMaker : public DecoratedReaderMakerBase { - public: - CreateDoubleBufferReaderOpMaker(OpProto* op_proto, OpAttrChecker* op_checker) - : DecoratedReaderMakerBase(op_proto, op_checker) { + protected: + void Apply() override { AddComment(R"DOC( CreateDoubleBufferReader Operator A double buffer reader takes another reader as its 'underlying reader'. - It launches another thread to execute the 'underlying reader' asynchronously, + It launches another thread to execute the 'underlying reader' asynchronously, which prevents reading process from blocking subsequent training. )DOC"); std::unordered_set enum_range; @@ -168,11 +167,10 @@ void DoubleBufferReader::PrefetchThreadFunc() { } if (platform::is_gpu_place(place_)) { auto& gpu_batch = gpu_tensor_cache_[cached_tensor_id]; - auto* gpu_ctx = ctxs_[cached_tensor_id].get(); gpu_batch.resize(cpu_batch.size()); for (size_t i = 0; i < cpu_batch.size(); ++i) { - framework::TensorCopy(cpu_batch[i], place_, *gpu_ctx, &gpu_batch[i], - true); + // TODO(fengjiayi): Use asynchronous TensorCopy instead + framework::TensorCopySync(cpu_batch[i], place_, &gpu_batch[i]); gpu_batch[i].set_lod(cpu_batch[i].lod()); } } diff --git a/paddle/fluid/operators/reader/create_multi_pass_reader_op.cc b/paddle/fluid/operators/reader/create_multi_pass_reader_op.cc index 0573345ba502b6a9af35710840d5acf7634f332f..249b0b7c6dbc8b8104bce95562e6e9b2a28c77f8 100644 --- a/paddle/fluid/operators/reader/create_multi_pass_reader_op.cc +++ b/paddle/fluid/operators/reader/create_multi_pass_reader_op.cc @@ -65,20 +65,19 @@ class CreateMultiPassReaderOp : public framework::OperatorBase { }; class CreateMultiPassReaderOpMaker : public DecoratedReaderMakerBase { - public: - CreateMultiPassReaderOpMaker(OpProto* op_proto, OpAttrChecker* op_checker) - : DecoratedReaderMakerBase(op_proto, op_checker) { + protected: + void Apply() override { AddAttr("pass_num", "The number of pass to run.").GreaterThan(0); AddComment(R"DOC( CreateMultiPassReader Operator - This operator creates a multi-pass reader. A multi-pass reader - is used to yield data for several pass training continuously. + This operator creates a multi-pass reader. A multi-pass reader + is used to yield data for several pass training continuously. It takes the number of passes to run as one of its attributes - ('pass_num'), and maintains a pass counter to record how many - passes it has completed. When the underlying reader reaches the - EOF, the multi-pass reader checks whether it has completed training - of the given number of pass. If not, the underlying reader will + ('pass_num'), and maintains a pass counter to record how many + passes it has completed. When the underlying reader reaches the + EOF, the multi-pass reader checks whether it has completed training + of the given number of pass. If not, the underlying reader will be re-initialized and starts a new pass automatically. )DOC"); } diff --git a/paddle/fluid/operators/reader/create_random_data_generator_op.cc b/paddle/fluid/operators/reader/create_random_data_generator_op.cc index d1cb8e47da70cab784858caea7e791151fc104dd..5b7e8a063a034f0be056065826fca0fe807bc9a7 100644 --- a/paddle/fluid/operators/reader/create_random_data_generator_op.cc +++ b/paddle/fluid/operators/reader/create_random_data_generator_op.cc @@ -21,14 +21,15 @@ namespace reader { template class RandomDataGenerator : public framework::ReaderBase { public: - RandomDataGenerator(const std::vector& shapes, float min, - float max) - : framework::ReaderBase(), min_(min), max_(max), shapes_(shapes) { - PADDLE_ENFORCE_LE( - min, max, "'min' shouldn't be greater than 'max'.(%f vs %f)", min, max); + RandomDataGenerator(const std::vector& shapes, float low, + float high) + : framework::ReaderBase(), low_(low), high_(high), shapes_(shapes) { + PADDLE_ENFORCE_LE(low, high, + "'low' shouldn't be greater than 'high'.(%f vs %f)", low, + high); unsigned int seed = std::random_device()(); engine_.seed(seed); - dist_ = std::uniform_real_distribution(min_, max_); + dist_ = std::uniform_real_distribution(low_, high_); } void ReadNext(std::vector* out) override { @@ -53,8 +54,8 @@ class RandomDataGenerator : public framework::ReaderBase { void ReInit() override { return; } private: - float min_; - float max_; + float low_; + float high_; std::minstd_rand engine_; std::uniform_real_distribution dist_; std::vector shapes_; @@ -78,23 +79,22 @@ class CreateRandomDataGeneratorOp : public framework::OperatorBase { std::vector shapes = RestoreShapes(shape_concat, ranks); auto* out = scope.FindVar(Output("Out")) ->template GetMutable(); - out->Reset(new RandomDataGenerator(shapes, Attr("min"), - Attr("max"))); + out->Reset(new RandomDataGenerator(shapes, Attr("low"), + Attr("high"))); } }; class CreateRandomDataGeneratorOpMaker : public FileReaderMakerBase { - public: - CreateRandomDataGeneratorOpMaker(OpProto* op_proto, OpAttrChecker* op_checker) - : FileReaderMakerBase(op_proto, op_checker) { - AddAttr("min", "The lower bound of reader's uniform distribution."); - AddAttr("max", "The upper bound of reader's uniform distribution."); + protected: + void Apply() override { + AddAttr("low", "The lower bound of reader's uniform distribution."); + AddAttr("high", "The upper bound of reader's uniform distribution."); AddComment(R"DOC( CreateRandomDataGenerator Operator This Op creates a random reader. The reader generates random data instead of really reading from files. - Generated data follow an uniform distribution between 'min' and 'max'. + Generated data follow an uniform distribution between 'low' and 'high'. )DOC"); } }; diff --git a/paddle/fluid/operators/reader/create_recordio_file_reader_op.cc b/paddle/fluid/operators/reader/create_recordio_file_reader_op.cc index 2ae29725561769ebe6428002c9983246b8eec724..282ec3f36b98e7aa62d71fb04f72721a5464e21c 100644 --- a/paddle/fluid/operators/reader/create_recordio_file_reader_op.cc +++ b/paddle/fluid/operators/reader/create_recordio_file_reader_op.cc @@ -76,9 +76,8 @@ class CreateRecordIOReaderOp : public framework::OperatorBase { }; class CreateRecordIOReaderOpMaker : public FileReaderMakerBase { - public: - CreateRecordIOReaderOpMaker(OpProto* op_proto, OpAttrChecker* op_checker) - : FileReaderMakerBase(op_proto, op_checker) { + protected: + void Apply() override { AddAttr("filename", "The filename of record io reader"); AddComment(R"DOC( CreateRecordIOReader Operator diff --git a/paddle/fluid/operators/reader/create_shuffle_reader_op.cc b/paddle/fluid/operators/reader/create_shuffle_reader_op.cc index 13825d65913be95f4f444bd9d5271a036ec8b1e2..fd233be945932eee9f9a3c0c578a43d5b7cc83aa 100644 --- a/paddle/fluid/operators/reader/create_shuffle_reader_op.cc +++ b/paddle/fluid/operators/reader/create_shuffle_reader_op.cc @@ -92,9 +92,8 @@ class CreateShuffleReaderOp : public framework::OperatorBase { }; class CreateShuffleReaderOpMaker : public DecoratedReaderMakerBase { - public: - CreateShuffleReaderOpMaker(OpProto* op_proto, OpAttrChecker* op_checker) - : DecoratedReaderMakerBase(op_proto, op_checker) { + protected: + void Apply() override { AddAttr("buffer_size", "The shuffle buffer size.").GreaterThan(0); AddComment(R"DOC( CreateShuffleReader Operator diff --git a/paddle/fluid/operators/reader/create_threaded_reader_op.cc b/paddle/fluid/operators/reader/create_threaded_reader_op.cc index 1cb9bd36455a2287b8ba4fb4ca14a4c5338da098..1db70f3e9699dba604569c36dc35025dfe2c94fe 100644 --- a/paddle/fluid/operators/reader/create_threaded_reader_op.cc +++ b/paddle/fluid/operators/reader/create_threaded_reader_op.cc @@ -53,17 +53,16 @@ class CreateThreadedReaderOp : public framework::OperatorBase { }; class CreateThreadedReaderOpMaker : public DecoratedReaderMakerBase { - public: - CreateThreadedReaderOpMaker(OpProto* op_proto, OpAttrChecker* op_checker) - : DecoratedReaderMakerBase(op_proto, op_checker) { + protected: + void Apply() override { AddComment(R"DOC( CreateThreadedReader Operator - This operator creates a threaded reader. A threaded reader's - 'ReadNext()' can be invoked by several threads at the same - time. - When the attribute 'safe_mode' is true, the threaded reader's - 'ReInit()' is disabled to avoid unexpected bugs in multi-thread + This operator creates a threaded reader. A threaded reader's + 'ReadNext()' can be invoked by several threads at the same + time. + When the attribute 'safe_mode' is true, the threaded reader's + 'ReInit()' is disabled to avoid unexpected bugs in multi-thread environment. )DOC"); } diff --git a/paddle/fluid/operators/reader/open_files_op.cc b/paddle/fluid/operators/reader/open_files_op.cc index 91ad7d56583446ee4686e74187de166f387125df..8c0dac65dd691954b112bfa61622d399b2b9c3e5 100644 --- a/paddle/fluid/operators/reader/open_files_op.cc +++ b/paddle/fluid/operators/reader/open_files_op.cc @@ -185,9 +185,8 @@ class OpenFilesOp : public framework::OperatorBase { }; class OpenFilesOpMaker : public FileReaderMakerBase { - public: - OpenFilesOpMaker(OpProto* op_proto, OpAttrChecker* op_checker) - : FileReaderMakerBase(op_proto, op_checker) { + protected: + void Apply() override { AddAttr>("file_names", "Files to be read."); AddAttr("thread_num", "The maximal concurrent prefetch thread number.") .GreaterThan(0); @@ -196,7 +195,7 @@ class OpenFilesOpMaker : public FileReaderMakerBase { AddComment(R"DOC( OpenFiles Operator - An OpenFilesOp creates a MultiFileReader, which is able to + An OpenFilesOp creates a MultiFileReader, which is able to read data multi-threaded from multiple files. )DOC"); } diff --git a/paddle/fluid/operators/reader/reader_op_registry.cc b/paddle/fluid/operators/reader/reader_op_registry.cc index 3ff4536819b128d9c593b97f4942a0292a3b6b36..612e1f5eca3a4836db1fd167fc6bb63400d20177 100644 --- a/paddle/fluid/operators/reader/reader_op_registry.cc +++ b/paddle/fluid/operators/reader/reader_op_registry.cc @@ -53,10 +53,7 @@ std::unique_ptr CreateReaderByFileName( return std::unique_ptr(reader); } -FileReaderMakerBase::FileReaderMakerBase( - framework::OpProtoAndCheckerMaker::OpProto* op_proto, - framework::OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(op_proto, op_checker) { +void FileReaderMakerBase::Make() { AddOutput("Out", "(ReaderHolder) The created random reader.").AsDuplicable(); AddAttr>("shape_concat", "The concat of all data's shapes."); AddAttr>( @@ -68,6 +65,7 @@ FileReaderMakerBase::FileReaderMakerBase( "It means the reader will generate two data each time," "whose shapes are [2,3,4] and [5,6] respectively."); AddAttr>("lod_levels", "The LoD levels of each data."); + Apply(); } void FileReaderInferShape::operator()(framework::InferShapeContext* ctx) const { @@ -117,6 +115,7 @@ void DecoratedReaderInferShape::operator()( boost::get(ctx->GetOutputVarPtrs("Out")[0]); out_reader->SetLoDLevels(in_reader->GetLoDLevels()); } + void DecoratedReaderInferVarType::operator()( const framework::OpDesc& op_desc, framework::BlockDesc* block) const { std::string in_reader_name = op_desc.Input("UnderlyingReader")[0]; @@ -127,13 +126,11 @@ void DecoratedReaderInferVarType::operator()( out_reader->SetDataTypes(in_reader->GetDataTypes()); } -DecoratedReaderMakerBase::DecoratedReaderMakerBase( - framework::OpProtoAndCheckerMaker::OpProto* op_proto, - framework::OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(op_proto, op_checker) { +void DecoratedReaderMakerBase::Make() { AddInput("UnderlyingReader", "(ReaderHolder) The underlying reader for creating a batch reader."); AddOutput("Out", "(ReaderHolder) The created batch reader."); + Apply(); } } // namespace reader diff --git a/paddle/fluid/operators/reader/reader_op_registry.h b/paddle/fluid/operators/reader/reader_op_registry.h index 929d32ad8b367865e33530f8517343c513ee9878..244bf15f068a47efc29ee54492cdbdeb10025020 100644 --- a/paddle/fluid/operators/reader/reader_op_registry.h +++ b/paddle/fluid/operators/reader/reader_op_registry.h @@ -14,6 +14,8 @@ #pragma once +#include +#include #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/reader.h" @@ -45,7 +47,10 @@ extern std::vector RestoreShapes( class FileReaderMakerBase : public framework::OpProtoAndCheckerMaker { public: - FileReaderMakerBase(OpProto* op_proto, OpAttrChecker* op_checker); + void Make() final; + + protected: + virtual void Apply() = 0; }; class FileReaderInferShape : public framework::InferShapeBase { @@ -74,7 +79,10 @@ class DecoratedReaderInferVarType : public framework::VarTypeInference { class DecoratedReaderMakerBase : public framework::OpProtoAndCheckerMaker { public: - DecoratedReaderMakerBase(OpProto* op_proto, OpAttrChecker* op_checker); + void Make() final; + + protected: + virtual void Apply() = 0; }; } // namespace reader diff --git a/paddle/fluid/operators/recurrent_op.cc b/paddle/fluid/operators/recurrent_op.cc index 72c2905872c528a7ed05820744f4031799ad9e46..9c1cee7022a9b9a98f026f7602f0f7badc44a49b 100644 --- a/paddle/fluid/operators/recurrent_op.cc +++ b/paddle/fluid/operators/recurrent_op.cc @@ -508,8 +508,7 @@ class RecurrentGradOp : public RecurrentBase { class RecurrentOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - RecurrentOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput(kInputs, "rnn inputs").AsDuplicable(); AddInput(kInitialStates, "rnn initial states").AsDuplicable(); AddInput(kParameters, diff --git a/paddle/fluid/operators/recv_op.cc b/paddle/fluid/operators/recv_op.cc index a4dcf704a63ae3bad6567ddb042ea23513bccff7..d8ddb7b448910b5e0e6e71742eb2fdc6a225c919 100644 --- a/paddle/fluid/operators/recv_op.cc +++ b/paddle/fluid/operators/recv_op.cc @@ -21,6 +21,7 @@ limitations under the License. */ #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/operators/detail/grpc_client.h" +#include "paddle/fluid/platform/profiler.h" namespace paddle { namespace operators { @@ -36,25 +37,28 @@ class RecvOp : public framework::OperatorBase { const platform::Place& place) const override { auto outs = Outputs("Out"); std::vector epmap = Attr>("epmap"); + int sync_mode = Attr("sync_mode"); platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); auto& ctx = *pool.Get(place); + // For profiling + platform::RecordEvent record_event(Type(), &ctx); + + auto rpc_client = detail::RPCClient::GetInstance(); for (size_t i = 0; i < outs.size(); i++) { - VLOG(3) << "getting " << outs[i]; - client_.AsyncGetVariable(epmap[i], ctx, scope, outs[i]); + VLOG(3) << "getting " << outs[i] << " from " << epmap[i]; + rpc_client->AsyncGetVariable(epmap[i], ctx, scope, outs[i]); + } + if (sync_mode) { + PADDLE_ENFORCE(rpc_client->Wait()); } - PADDLE_ENFORCE(client_.Wait()); } - - private: - mutable detail::RPCClient client_; }; class RecvOpMaker : public framework::OpProtoAndCheckerMaker { public: - RecvOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() { AddOutput("Out", "(Tensor) Variables to get from server.").AsDuplicable(); AddComment(R"DOC( Recv operator @@ -66,6 +70,10 @@ This operator can get variables from server side. "Server endpoints in the order of input " "variables for mapping") .SetDefault({}); + AddAttr("sync_mode", + "(int, default 0)" + "sync recv or async recv.") + .SetDefault(0); } }; diff --git a/paddle/fluid/operators/reduce_op.cc b/paddle/fluid/operators/reduce_op.cc index 093db966472cf100b2f1e4159ce20399cee1f481..e293fd5e410b2a34b3c71ea674607ba9d7654535 100644 --- a/paddle/fluid/operators/reduce_op.cc +++ b/paddle/fluid/operators/reduce_op.cc @@ -14,6 +14,7 @@ limitations under the License. */ #include "paddle/fluid/operators/reduce_op.h" +#include #include #include @@ -34,11 +35,14 @@ class ReduceOp : public framework::OperatorWithKernel { auto x_dims = ctx->GetInputDim("X"); auto x_rank = x_dims.size(); PADDLE_ENFORCE_LE(x_rank, 6, "Tensors with rank at most 6 are supported."); - int dim = ctx->Attrs().Get("dim"); - if (dim < 0) dim = x_rank + dim; - PADDLE_ENFORCE_LT( - dim, x_rank, - "The dim should be in the range [-rank(input), rank(input))."); + auto dims = ctx->Attrs().Get>("dim"); + for (size_t i = 0; i < dims.size(); ++i) { + if (dims[i] < 0) dims[i] = x_rank + dims[i]; + PADDLE_ENFORCE_LT( + dims[i], x_rank, + "The dim should be in the range [-rank(input), rank(input))."); + } + sort(dims.begin(), dims.end()); bool reduce_all = ctx->Attrs().Get("reduce_all"); bool keep_dim = ctx->Attrs().Get("keep_dim"); if (reduce_all) { @@ -49,14 +53,22 @@ class ReduceOp : public framework::OperatorWithKernel { ctx->SetOutputDim("Out", {1}); } else { auto dims_vector = vectorize(x_dims); - if (keep_dim || x_rank == 1) { - dims_vector[dim] = 1; + if (keep_dim) { + for (size_t i = 0; i < dims.size(); ++i) { + dims_vector[dims[i]] = 1; + } } else { - dims_vector.erase(dims_vector.begin() + dim); + const int kDelFlag = -2; + for (size_t i = 0; i < dims.size(); ++i) { + dims_vector[dims[i]] = kDelFlag; + } + dims_vector.erase( + remove(dims_vector.begin(), dims_vector.end(), kDelFlag), + dims_vector.end()); } auto out_dims = framework::make_ddim(dims_vector); ctx->SetOutputDim("Out", out_dims); - if (dim != 0) { + if (dims[0] != 0) { // Only pass LoD when not reducing on the first dim. ctx->ShareLoD("X", /*->*/ "Out"); } @@ -75,11 +87,14 @@ class ReduceGradOp : public framework::OperatorWithKernel { auto x_dims = ctx->GetInputDim("X"); auto x_rank = x_dims.size(); PADDLE_ENFORCE_LE(x_rank, 6, "Tensors with rank at most 6 are supported."); - int dim = ctx->Attrs().Get("dim"); - if (dim < 0) dim = x_rank + dim; - PADDLE_ENFORCE_LT( - dim, x_rank, - "The dim should be in the range [-rank(input), rank(input))."); + auto dims = ctx->Attrs().Get>("dim"); + for (size_t i = 0; i < dims.size(); ++i) { + if (dims[i] < 0) dims[i] = x_rank + dims[i]; + PADDLE_ENFORCE_LT( + dims[i], x_rank, + "The dim should be in the range [-rank(input), rank(input))."); + } + sort(dims.begin(), dims.end()); auto x_grad_name = framework::GradVarName("X"); if (ctx->HasOutput(x_grad_name)) { ctx->SetOutputDim(x_grad_name, x_dims); @@ -90,19 +105,18 @@ class ReduceGradOp : public framework::OperatorWithKernel { class ReduceOpMaker : public framework::OpProtoAndCheckerMaker { public: - ReduceOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() final { AddInput("X", "(Tensor) The input tensor. Tensors with rank at most 6 are " "supported."); AddOutput("Out", "(Tensor) The result tensor."); - AddAttr( + AddAttr>( "dim", - "(int, default 0) The dimension to reduce. " + "(list, default {0}) The dimensions to reduce. " "Must be in the range [-rank(input), rank(input)). " - "If `dim < 0`, the dim to reduce is `rank + dim`. " + "If `dim[i] < 0`, the dims[i] to reduce is `rank + dims[i]`. " "Note that reducing on the first dim will make the LoD info lost.") - .SetDefault(0); + .SetDefault({0}); AddAttr("keep_dim", "(bool, default false) " "If true, retain the reduced dimension with length 1.") @@ -111,78 +125,20 @@ class ReduceOpMaker : public framework::OpProtoAndCheckerMaker { "(bool, default false) " "If true, output a scalar reduced along all dimensions.") .SetDefault(false); - comment_ = R"DOC( -{ReduceOp} Operator. + AddComment(string::Sprintf(R"DOC( +%s Operator. -This operator computes the {reduce} of input tensor along the given dimension. +This operator computes the %s of input tensor along the given dimension. The result tensor has 1 fewer dimension than the input unless keep_dim is true. If reduce_all is true, just reduce along all dimensions and output a scalar. -)DOC"; - AddComment(comment_); +)DOC", + GetOpType(), GetName())); } protected: - std::string comment_; - - void Replace(std::string *src, std::string from, std::string to) { - std::size_t len_from = std::strlen(from.c_str()); - std::size_t len_to = std::strlen(to.c_str()); - for (std::size_t pos = src->find(from); pos != std::string::npos; - pos = src->find(from, pos + len_to)) { - src->replace(pos, len_from, to); - } - } - - void SetComment(std::string name, std::string op) { - Replace(&comment_, "{ReduceOp}", name); - Replace(&comment_, "{reduce}", op); - } -}; - -class ReduceSumOpMaker : public ReduceOpMaker { - public: - ReduceSumOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : ReduceOpMaker(proto, op_checker) { - SetComment("ReduceSum", "sum"); - AddComment(comment_); - } -}; - -class ReduceMeanOpMaker : public ReduceOpMaker { - public: - ReduceMeanOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : ReduceOpMaker(proto, op_checker) { - SetComment("ReduceMean", "mean"); - AddComment(comment_); - } -}; - -class ReduceMaxOpMaker : public ReduceOpMaker { - public: - ReduceMaxOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : ReduceOpMaker(proto, op_checker) { - SetComment("ReduceMax", "max"); - AddComment(comment_); - } -}; - -class ReduceMinOpMaker : public ReduceOpMaker { - public: - ReduceMinOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : ReduceOpMaker(proto, op_checker) { - SetComment("ReduceMin", "min"); - AddComment(comment_); - } -}; - -class ReduceProdOpMaker : public ReduceOpMaker { - public: - ReduceProdOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : ReduceOpMaker(proto, op_checker) { - SetComment("ReduceProd", "production"); - AddComment(comment_); - } + virtual std::string GetName() const = 0; + virtual std::string GetOpType() const = 0; }; } // namespace operators @@ -190,25 +146,21 @@ class ReduceProdOpMaker : public ReduceOpMaker { namespace ops = paddle::operators; -REGISTER_OPERATOR(reduce_sum, ops::ReduceOp, ops::ReduceSumOpMaker, - paddle::framework::DefaultGradOpDescMaker); -REGISTER_OPERATOR(reduce_sum_grad, ops::ReduceGradOp); - -REGISTER_OPERATOR(reduce_mean, ops::ReduceOp, ops::ReduceMeanOpMaker, - paddle::framework::DefaultGradOpDescMaker); -REGISTER_OPERATOR(reduce_mean_grad, ops::ReduceGradOp); - -REGISTER_OPERATOR(reduce_max, ops::ReduceOp, ops::ReduceMaxOpMaker, - paddle::framework::DefaultGradOpDescMaker); -REGISTER_OPERATOR(reduce_max_grad, ops::ReduceGradOp); - -REGISTER_OPERATOR(reduce_min, ops::ReduceOp, ops::ReduceMinOpMaker, - paddle::framework::DefaultGradOpDescMaker); -REGISTER_OPERATOR(reduce_min_grad, ops::ReduceGradOp); - -REGISTER_OPERATOR(reduce_prod, ops::ReduceOp, ops::ReduceProdOpMaker, - paddle::framework::DefaultGradOpDescMaker); -REGISTER_OPERATOR(reduce_prod_grad, ops::ReduceGradOp); +#define REGISTER_REDUCE_OP(op_name) \ + class __##op_name##Maker__ : public ops::ReduceOpMaker { \ + protected: \ + virtual std::string GetName() const { return #op_name; } \ + virtual std::string GetOpType() const { return "Reduce " #op_name; } \ + }; \ + REGISTER_OPERATOR(reduce_##op_name, ops::ReduceOp, __##op_name##Maker__, \ + paddle::framework::DefaultGradOpDescMaker); \ + REGISTER_OPERATOR(reduce_##op_name##_grad, ops::ReduceGradOp) + +REGISTER_REDUCE_OP(sum); +REGISTER_REDUCE_OP(mean); +REGISTER_REDUCE_OP(max); +REGISTER_REDUCE_OP(min); +REGISTER_REDUCE_OP(prod); #define REGISTER_REDUCE_CPU_KERNEL(reduce_type, functor, grad_functor) \ REGISTER_OP_CPU_KERNEL(reduce_type, \ diff --git a/paddle/fluid/operators/reduce_op.h b/paddle/fluid/operators/reduce_op.h index e42b4bfe42df05346020d4f48519fecf39aa37d2..cd19cc1460a6b4d4201f21f6f27f988c1547b88a 100644 --- a/paddle/fluid/operators/reduce_op.h +++ b/paddle/fluid/operators/reduce_op.h @@ -14,6 +14,7 @@ limitations under the License. */ #pragma once +#include #include "glog/logging.h" #include "paddle/fluid/framework/eigen.h" #include "paddle/fluid/framework/op_registry.h" @@ -109,6 +110,11 @@ struct ProdGradFunctor { } }; +#define HANDLE_DIM(NDIM, RDIM) \ + if (ndim == NDIM && rdim == RDIM) { \ + ReduceCompute(context); \ + } + template class ReduceKernel : public framework::OpKernel { public: @@ -127,32 +133,29 @@ class ReduceKernel : public framework::OpKernel { Functor functor; functor(place, &x, &out, reduce_dim); } else { - int rank = context.Input("X")->dims().size(); - switch (rank) { - case 1: - ReduceCompute<1>(context); - break; - case 2: - ReduceCompute<2>(context); - break; - case 3: - ReduceCompute<3>(context); - break; - case 4: - ReduceCompute<4>(context); - break; - case 5: - ReduceCompute<5>(context); - break; - case 6: - ReduceCompute<6>(context); - break; - } + int ndim = context.Input("X")->dims().size(); + int rdim = context.Attr>("dim").size(); + HANDLE_DIM(6, 5); + HANDLE_DIM(6, 4); + HANDLE_DIM(6, 3); + HANDLE_DIM(6, 2); + HANDLE_DIM(6, 1); + HANDLE_DIM(5, 4); + HANDLE_DIM(5, 3); + HANDLE_DIM(5, 2); + HANDLE_DIM(5, 1); + HANDLE_DIM(4, 3); + HANDLE_DIM(4, 2); + HANDLE_DIM(4, 1); + HANDLE_DIM(3, 2); + HANDLE_DIM(3, 1); + HANDLE_DIM(2, 1); + HANDLE_DIM(1, 1); } } private: - template + template void ReduceCompute(const framework::ExecutionContext& context) const { auto* input = context.Input("X"); auto* output = context.Output("Out"); @@ -160,18 +163,26 @@ class ReduceKernel : public framework::OpKernel { auto x = EigenTensor::From(*input); auto x_rank = static_cast(x.dimensions().size()); - int dim = static_cast(context.Attr("dim")); - if (dim < 0) dim = x_rank + dim; - auto reduce_dim = Eigen::array({{dim}}); + auto dims = context.Attr>("dim"); + auto reduce_dim = Eigen::array(); + for (size_t i = 0; i < dims.size(); ++i) { + if (dims[i] < 0) dims[i] = x_rank + dims[i]; + reduce_dim[i] = dims[i]; + } // construct the squeezed output tensor bool keep_dim = context.Attr("keep_dim"); - DDim dims = output->dims(); - auto dims_vector = vectorize(dims); + DDim out_dims = output->dims(); if (keep_dim && x_rank > 1) { - dims_vector.erase(dims_vector.begin() + dim); - dims = framework::make_ddim(dims_vector); + const int kDelFlag = -2; + auto dims_vector = vectorize(out_dims); + for (size_t i = 0; i < dims.size(); ++i) { + dims_vector[dims[i]] = kDelFlag; + } + dims_vector.erase( + remove(dims_vector.begin(), dims_vector.end(), kDelFlag), + dims_vector.end()); + out_dims = framework::make_ddim(dims_vector); } - auto& place = *context.template device_context().eigen_device(); Functor functor; @@ -180,7 +191,7 @@ class ReduceKernel : public framework::OpKernel { auto out = EigenScalar::From(*output); functor(place, &x, &out, reduce_dim); } else { - auto out = EigenTensor::From(*output, dims); + auto out = EigenTensor::From(*output, out_dims); functor(place, &x, &out, reduce_dim); } } @@ -245,21 +256,29 @@ class ReduceGradKernel : public framework::OpKernel { auto x = EigenTensor::From(*input0); auto x_grad = EigenTensor::From(*output); auto x_rank = static_cast(x.dimensions().size()); - int dim = static_cast(context.Attr("dim")); - if (dim < 0) dim = x_rank + dim; - DDim dims = input0->dims(); - dims[dim] = 1; - auto x_reduce = EigenTensor::From(*input1, dims); - auto x_reduce_grad = EigenTensor::From(*input2, dims); - + auto dims = context.Attr>("dim"); + auto x_dims = input0->dims(); + auto reduced_dims_v = vectorize(x_dims); Eigen::array broadcast_dim; for (size_t i = 0; i < D; ++i) broadcast_dim[i] = 1; - broadcast_dim[dim] = input0->dims()[dim]; + + int broad_cats_times = 1; + for (size_t i = 0; i < dims.size(); ++i) { + if (dims[i] < 0) dims[i] = x_rank + dims[i]; + reduced_dims_v[dims[i]] = 1; + broadcast_dim[dims[i]] = x_dims[dims[i]]; + broad_cats_times *= x_dims[dims[i]]; + } + auto reduced_dims = framework::make_ddim(reduced_dims_v); + auto x_reduce = EigenTensor::From(*input1, reduced_dims); + auto x_reduce_grad = EigenTensor::From(*input2, reduced_dims); + auto& place = *context.template device_context().eigen_device(); + Functor functor; functor(place, &x, &x_reduce, &x_grad, &x_reduce_grad, broadcast_dim, - broadcast_dim[dim]); + broad_cats_times); } }; diff --git a/paddle/fluid/operators/reorder_lod_tensor_by_rank_op.cc b/paddle/fluid/operators/reorder_lod_tensor_by_rank_op.cc index 5c3e1f5678df0270c837ed407d1e6cc662276880..e4f4fe358e0e8cd2080525227f14a3d40f3c1411 100644 --- a/paddle/fluid/operators/reorder_lod_tensor_by_rank_op.cc +++ b/paddle/fluid/operators/reorder_lod_tensor_by_rank_op.cc @@ -23,9 +23,7 @@ namespace operators { class ReorderLoDTensorByRankTableOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - ReorderLoDTensorByRankTableOpProtoMaker(OpProto *proto, - OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(LoDTensor), the input lod tensor to be reordered according to " "Input(RankTable)."); diff --git a/paddle/fluid/operators/reshape_op.cc b/paddle/fluid/operators/reshape_op.cc index 5e5ccc3ded95d57dfed37c1ac9c7eae61d36b8c0..7f743f577fbcdaf6f62e01031e25ef09a842c2e9 100644 --- a/paddle/fluid/operators/reshape_op.cc +++ b/paddle/fluid/operators/reshape_op.cc @@ -22,8 +22,7 @@ namespace operators { class ReshapeOpMaker : public framework::OpProtoAndCheckerMaker { public: - ReshapeOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor). The input tensor of reshape operator."); AddInput("Shape", "(Tensor, optional). If provided, reshape according to " diff --git a/paddle/fluid/operators/reshape_op.h b/paddle/fluid/operators/reshape_op.h index 8320c257c9ab15efec29eabe99eca5b6f74c9e31..3dd8c7c11eca241e747bfa129962032d882ce44c 100644 --- a/paddle/fluid/operators/reshape_op.h +++ b/paddle/fluid/operators/reshape_op.h @@ -92,9 +92,17 @@ class ReshapeOp : public framework::OperatorWithKernel { } if (unk_dim_idx != -1) { - output_shape[unk_dim_idx] = -in_size / capacity; - PADDLE_ENFORCE_EQ(output_shape[unk_dim_idx] * capacity, -in_size, - "Invalid shape is given."); + if (in_size > 0) { + // in_size < 0 and is un-determinate in compile time, skip the check, + // for example, in_dims = [-1, 8, 1, 1], shape = [-1, 3, 8], + // capacity = -24, in_size = -8, output_shape[0] = 0 + // the following check will fail. + output_shape[unk_dim_idx] = -in_size / capacity; + PADDLE_ENFORCE_EQ(output_shape[unk_dim_idx] * capacity, -in_size, + "Invalid shape is given."); + } else { + output_shape[unk_dim_idx] = -1; + } } else { PADDLE_ENFORCE_EQ(capacity, in_size, "Invalid shape is given."); } @@ -116,7 +124,10 @@ class ReshapeKernel : public framework::OpKernel { void Compute(const framework::ExecutionContext &ctx) const { auto *out = ctx.Output("Out"); auto *in = ctx.Input("X"); - auto *shape_tensor = ctx.Input("Shape"); + + auto *shape_tensor = ctx.HasInput("Shape") + ? ctx.Input("Shape") + : nullptr; framework::DDim out_dims = out->dims(); @@ -124,10 +135,8 @@ class ReshapeKernel : public framework::OpKernel { auto *shape_data = shape_tensor->data(); framework::Tensor cpu_shape_tensor; if (platform::is_gpu_place(ctx.GetPlace())) { - TensorCopy(*shape_tensor, platform::CPUPlace(), ctx.device_context(), - &cpu_shape_tensor); + TensorCopySync(*shape_tensor, platform::CPUPlace(), &cpu_shape_tensor); shape_data = cpu_shape_tensor.data(); - ctx.device_context().Wait(); } auto shape = std::vector(shape_data, shape_data + shape_tensor->numel()); @@ -146,9 +155,7 @@ class ReshapeKernel : public framework::OpKernel { out->Resize(out_dims); if (!inplace) { out->mutable_data(ctx.GetPlace()); - framework::TensorCopy(*in, ctx.GetPlace(), ctx.device_context(), out); - ctx.device_context().Wait(); - // TensorCopy will resize to in_dims. + framework::TensorCopySync(*in, ctx.GetPlace(), out); out->Resize(out_dims); } else { out->ShareDataWith(*in); diff --git a/paddle/fluid/operators/rmsprop_op.cc b/paddle/fluid/operators/rmsprop_op.cc index a8855b3ccd1686c75953e762ce42cc27b26202e6..919ebe48ca38040274bd2052b95ef96eccff4db6 100644 --- a/paddle/fluid/operators/rmsprop_op.cc +++ b/paddle/fluid/operators/rmsprop_op.cc @@ -63,8 +63,7 @@ class RmspropOp : public framework::OperatorWithKernel { class RmspropOpMaker : public framework::OpProtoAndCheckerMaker { public: - RmspropOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Param", "(Tensor, default Tensor) " "Input parameter value that has to be updated."); diff --git a/paddle/fluid/operators/rnn_memory_helper_op.cc b/paddle/fluid/operators/rnn_memory_helper_op.cc index 70f205d887ef710aeed02905713200ce32988987..23e5fc1112d0b1e634d0ab288721cbba57b3ffe5 100644 --- a/paddle/fluid/operators/rnn_memory_helper_op.cc +++ b/paddle/fluid/operators/rnn_memory_helper_op.cc @@ -59,8 +59,7 @@ class RNNMemoryHelperOpShapeInference : public framework::InferShapeBase { class RNNMemoryHelperOpInfoMaker : public framework::OpProtoAndCheckerMaker { public: - RNNMemoryHelperOpInfoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", ""); AddOutput("Out", ""); AddAttr("dtype", @@ -117,8 +116,7 @@ class RNNMemoryHelperGradOp : public framework::OperatorBase { class RNNMemoryHelperGradOpInfoMaker : public framework::OpProtoAndCheckerMaker { public: - RNNMemoryHelperGradOpInfoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput(framework::GradVarName("Out"), ""); AddInput("X", ""); AddInput("Out", ""); diff --git a/paddle/fluid/operators/roi_pool_op.cc b/paddle/fluid/operators/roi_pool_op.cc index 224ec93d28ec75c52848d7c8400e684df0d69209..293abb0ea4f1ac03c3889ce2937ef8fa0845db73 100644 --- a/paddle/fluid/operators/roi_pool_op.cc +++ b/paddle/fluid/operators/roi_pool_op.cc @@ -18,8 +18,7 @@ namespace paddle { namespace operators { using Tensor = framework::Tensor; - -static constexpr int kROISize = 5; +using LoDTensor = framework::LoDTensor; class ROIPoolOp : public framework::OperatorWithKernel { public: @@ -40,11 +39,11 @@ class ROIPoolOp : public framework::OperatorWithKernel { PADDLE_ENFORCE(input_dims.size() == 4, "The format of input tensor is NCHW."); PADDLE_ENFORCE(rois_dims.size() == 2, - "ROIs should be a 2-D tensor of shape (num_rois, 5)" - "given as [[batch_id, x1, y1, x2, y2], …]."); + "ROIs should be a 2-D LoDTensor of shape (num_rois, 4)" + "given as [[x1, y1, x2, y2], …]."); PADDLE_ENFORCE(rois_dims[1] == kROISize, - "ROIs should be a 2-D tensor of shape (num_rois, 5)" - "given as [[batch_id, x1, y1, x2, y2], …]."); + "ROIs should be a 2-D LoDTensor of shape (num_rois, 4)" + "given as [[x1, y1, x2, y2], …]."); int pooled_height = ctx->Attrs().Get("pooled_height"); int pooled_width = ctx->Attrs().Get("pooled_width"); @@ -99,8 +98,7 @@ class ROIPoolGradOp : public framework::OperatorWithKernel { class ROIPoolOpMaker : public framework::OpProtoAndCheckerMaker { public: - ROIPoolOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor), " "the input of ROIPoolOp. " @@ -109,10 +107,10 @@ class ROIPoolOpMaker : public framework::OpProtoAndCheckerMaker { "H is the height of the feature, and " "W is the width of the feature."); AddInput("ROIs", - "(Tensor), " + "(LoDTensor), " "ROIs (Regions of Interest) to pool over. " - "should be a 2-D tensor of shape (num_rois, 5)" - "given as [[batch_id, x1, y1, x2, y2], …]. " + "should be a 2-D LoDTensor of shape (num_rois, 4)" + "given as [[x1, y1, x2, y2], …]. " "Where batch_id is the id of the data, " "(x1, y1) is the top left coordinates, and " "(x2, y2) is the bottom right coordinates."); diff --git a/paddle/fluid/operators/roi_pool_op.cu b/paddle/fluid/operators/roi_pool_op.cu index 1931629d1340758edb6664a5e3ffdba126b33717..50450b62f7b1c0b2b5abf01a43581a0e2d2cd01e 100644 --- a/paddle/fluid/operators/roi_pool_op.cu +++ b/paddle/fluid/operators/roi_pool_op.cu @@ -13,16 +13,16 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/operators/roi_pool_op.h" -#include "paddle/fluid/platform/cuda_helper.h" +#include "paddle/fluid/platform/cuda_primitives.h" namespace paddle { namespace operators { using Tensor = framework::Tensor; +using LoDTensor = framework::LoDTensor; static constexpr int kNumCUDAThreads = 512; static constexpr int kNumMaxinumNumBlocks = 4096; -static constexpr int kROISize = 5; static inline int NumBlocks(const int N) { return std::min((N + kNumCUDAThreads - 1) / kNumCUDAThreads, @@ -30,38 +30,41 @@ static inline int NumBlocks(const int N) { } template -__global__ void GPUROIPoolForward(const int nthreads, const T* input_data, - const int64_t* input_rois, - const float spatial_scale, const int channels, - const int height, const int width, - const int pooled_height, - const int pooled_width, T* output_data, - int64_t* argmax_data) { +__global__ void GPUROIPoolForward( + const int nthreads, const T* input_data, const int64_t* input_rois, + const float spatial_scale, const int channels, const int height, + const int width, const int pooled_height, const int pooled_width, + int* roi_batch_id_data, T* output_data, int64_t* argmax_data) { int index = blockIdx.x * blockDim.x + threadIdx.x; int offset = blockDim.x * gridDim.x; for (size_t i = index; i < nthreads; i += offset) { - int pw = index % pooled_width; - int ph = (index / pooled_width) % pooled_height; - int c = (index / pooled_width / pooled_height) % channels; - int n = index / pooled_width / pooled_height / channels; + int pw = i % pooled_width; + int ph = (i / pooled_width) % pooled_height; + int c = (i / pooled_width / pooled_height) % channels; + int n = i / pooled_width / pooled_height / channels; const int64_t* offset_input_rois = input_rois + n * kROISize; - int roi_batch_ind = offset_input_rois[0]; - int roi_start_w = round(offset_input_rois[1] * spatial_scale); - int roi_start_h = round(offset_input_rois[2] * spatial_scale); - int roi_end_w = round(offset_input_rois[3] * spatial_scale); - int roi_end_h = round(offset_input_rois[4] * spatial_scale); + int roi_batch_ind = roi_batch_id_data[n]; + int roi_start_w = round(offset_input_rois[0] * spatial_scale); + int roi_start_h = round(offset_input_rois[1] * spatial_scale); + int roi_end_w = round(offset_input_rois[2] * spatial_scale); + int roi_end_h = round(offset_input_rois[3] * spatial_scale); int roi_width = max(roi_end_w - roi_start_w + 1, 1); int roi_height = max(roi_end_h - roi_start_h + 1, 1); - T bin_size_h = static_cast(roi_height) / static_cast(pooled_height); - T bin_size_w = static_cast(roi_width) / static_cast(pooled_width); - - int hstart = static_cast(floor(static_cast(ph) * bin_size_h)); - int wstart = static_cast(floor(static_cast(pw) * bin_size_w)); - int hend = static_cast(ceil(static_cast(ph + 1) * bin_size_h)); - int wend = static_cast(ceil(static_cast(pw + 1) * bin_size_w)); + int hstart = static_cast(floor(static_cast(ph) * + static_cast(roi_height) / + static_cast(pooled_height))); + int wstart = static_cast(floor(static_cast(pw) * + static_cast(roi_width) / + static_cast(pooled_width))); + int hend = static_cast(ceil(static_cast(ph + 1) * + static_cast(roi_height) / + static_cast(pooled_height))); + int wend = static_cast(ceil(static_cast(pw + 1) * + static_cast(roi_width) / + static_cast(pooled_width))); hstart = min(max(hstart + roi_start_h, 0), height); hend = min(max(hend + roi_start_h, 0), height); wstart = min(max(wstart + roi_start_w, 0), width); @@ -81,9 +84,9 @@ __global__ void GPUROIPoolForward(const int nthreads, const T* input_data, } } } - output_data[index] = maxval; + output_data[i] = maxval; if (argmax_data) { - argmax_data[index] = maxidx; + argmax_data[i] = maxidx; } } } @@ -93,17 +96,17 @@ __global__ void GPUROIPoolBackward( const int nthreads, const int64_t* input_rois, const T* output_grad, const int64_t* argmax_data, const int num_rois, const float spatial_scale, const int channels, const int height, const int width, - const int pooled_height, const int pooled_width, T* input_grad) { + const int pooled_height, const int pooled_width, int* roi_batch_id_data, + T* input_grad) { int index = blockIdx.x * blockDim.x + threadIdx.x; int offset = blockDim.x * gridDim.x; for (int i = index; i < nthreads; i += offset) { - int pw = index % pooled_width; - int ph = (index / pooled_width) % pooled_height; - int c = (index / pooled_width / pooled_height) % channels; - int n = index / pooled_width / pooled_height / channels; + int pw = i % pooled_width; + int ph = (i / pooled_width) % pooled_height; + int c = (i / pooled_width / pooled_height) % channels; + int n = i / pooled_width / pooled_height / channels; - const int64_t* offset_input_rois = input_rois + n * kROISize; - int roi_batch_ind = offset_input_rois[0]; + int roi_batch_ind = roi_batch_id_data[n]; int input_offset = (roi_batch_ind * channels + c) * height * width; int output_offset = (n * channels + c) * pooled_height * pooled_width; const T* offset_output_grad = output_grad + output_offset; @@ -124,7 +127,7 @@ class GPUROIPoolOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { auto* in = ctx.Input("X"); - auto* rois = ctx.Input("ROIs"); + auto* rois = ctx.Input("ROIs"); auto* out = ctx.Output("Out"); auto* argmax = ctx.Output("Argmax"); @@ -133,23 +136,47 @@ class GPUROIPoolOpKernel : public framework::OpKernel { auto spatial_scale = ctx.Attr("spatial_scale"); auto in_dims = in->dims(); + int batch_size = in_dims[0]; auto in_stride = framework::stride(in_dims); int channels = in_dims[1]; int height = in_dims[2]; int width = in_dims[3]; - size_t rois_num = rois->dims()[0]; + int rois_num = rois->dims()[0]; + if (rois_num == 0) return; int output_size = out->numel(); int blocks = NumBlocks(output_size); int threads = kNumCUDAThreads; + framework::Tensor roi_batch_id_list; + roi_batch_id_list.Resize({rois_num}); + int* roi_batch_id_data = + roi_batch_id_list.mutable_data(platform::CPUPlace()); + auto rois_lod = rois->lod().back(); + int rois_batch_size = rois_lod.size() - 1; + PADDLE_ENFORCE_EQ( + rois_batch_size, batch_size, + "The rois_batch_size and imgs batch_size must be the same."); + int rois_num_with_lod = rois_lod[rois_batch_size]; + PADDLE_ENFORCE_EQ(rois_num, rois_num_with_lod, + "The rois_num from input and lod must be the same."); + for (int n = 0; n < rois_batch_size; ++n) { + for (size_t i = rois_lod[n]; i < rois_lod[n + 1]; ++i) { + roi_batch_id_data[i] = n; + } + } + + framework::Tensor roi_batch_id_list_gpu; + framework::TensorCopy(roi_batch_id_list, ctx.GetPlace(), + ctx.device_context(), &roi_batch_id_list_gpu); + GPUROIPoolForward< T><<>>( output_size, in->data(), rois->data(), spatial_scale, channels, height, width, pooled_height, pooled_width, - out->mutable_data(ctx.GetPlace()), + roi_batch_id_list_gpu.data(), out->mutable_data(ctx.GetPlace()), argmax->mutable_data(ctx.GetPlace())); } }; @@ -159,7 +186,7 @@ class GPUROIPoolGradOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { auto* in = ctx.Input("X"); - auto* rois = ctx.Input("ROIs"); + auto* rois = ctx.Input("ROIs"); auto* argmax = ctx.Input("Argmax"); auto* out_grad = ctx.Input(framework::GradVarName("Out")); @@ -169,12 +196,27 @@ class GPUROIPoolGradOpKernel : public framework::OpKernel { auto pooled_width = ctx.Attr("pooled_width"); auto spatial_scale = ctx.Attr("spatial_scale"); - size_t rois_num = rois->dims()[0]; + int rois_num = rois->dims()[0]; int channels = in->dims()[1]; int height = in->dims()[2]; int width = in->dims()[3]; if (x_grad) { + framework::Tensor roi_batch_id_list; + roi_batch_id_list.Resize({rois_num}); + int* roi_batch_id_data = + roi_batch_id_list.mutable_data(platform::CPUPlace()); + auto rois_lod = rois->lod().back(); + int rois_batch_size = rois_lod.size() - 1; + for (int n = 0; n < rois_batch_size; ++n) { + for (size_t i = rois_lod[n]; i < rois_lod[n + 1]; ++i) { + roi_batch_id_data[i] = n; + } + } + framework::Tensor roi_batch_id_list_gpu; + framework::TensorCopy(roi_batch_id_list, ctx.GetPlace(), + ctx.device_context(), &roi_batch_id_list_gpu); + x_grad->mutable_data(ctx.GetPlace()); math::SetConstant set_zero; set_zero(ctx.cuda_device_context(), x_grad, static_cast(0)); @@ -189,6 +231,7 @@ class GPUROIPoolGradOpKernel : public framework::OpKernel { output_grad_size, rois->data(), out_grad->data(), argmax->data(), rois_num, spatial_scale, channels, height, width, pooled_height, pooled_width, + roi_batch_id_list_gpu.data(), x_grad->mutable_data(ctx.GetPlace())); } } diff --git a/paddle/fluid/operators/roi_pool_op.h b/paddle/fluid/operators/roi_pool_op.h index 54e07490319cf1da749bd33449a7b51efd6c3d65..c4f739b2c6b2d62ebebcc15fd627ebad040e7b3f 100644 --- a/paddle/fluid/operators/roi_pool_op.h +++ b/paddle/fluid/operators/roi_pool_op.h @@ -21,12 +21,14 @@ limitations under the License. */ namespace paddle { namespace operators { +static constexpr int kROISize = 4; + template class CPUROIPoolOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { auto* in = ctx.Input("X"); - auto* rois = ctx.Input("ROIs"); + auto* rois = ctx.Input("ROIs"); auto* out = ctx.Output("Out"); auto* argmax = ctx.Output("Argmax"); @@ -47,24 +49,36 @@ class CPUROIPoolOpKernel : public framework::OpKernel { auto out_stride = framework::stride(out->dims()); const T* input_data = in->data(); - const int64_t* rois_data = rois->data(); - T* output_data = out->mutable_data(ctx.GetPlace()); - int64_t* argmax_data = argmax->mutable_data(ctx.GetPlace()); - for (int n = 0; n < rois_num; ++n) { - int roi_batch_id = rois_data[0]; - PADDLE_ENFORCE_GE(roi_batch_id, 0); - PADDLE_ENFORCE_LT(roi_batch_id, batch_size); - rois_data += roi_stride[0]; + framework::Tensor roi_batch_id_list; + roi_batch_id_list.Resize({rois_num}); + int* roi_batch_id_data = + roi_batch_id_list.mutable_data(ctx.GetPlace()); + + auto rois_lod = rois->lod().back(); + int rois_batch_size = rois_lod.size() - 1; + PADDLE_ENFORCE_EQ( + rois_batch_size, batch_size, + "The rois_batch_size and imgs batch_size must be the same."); + int rois_num_with_lod = rois_lod[rois_batch_size]; + PADDLE_ENFORCE_EQ(rois_num, rois_num_with_lod, + "The rois_num from input and lod must be the same."); + for (int n = 0; n < rois_batch_size; ++n) { + for (size_t i = rois_lod[n]; i < rois_lod[n + 1]; ++i) { + roi_batch_id_data[i] = n; + } } - rois_data = rois->data(); + T* output_data = out->mutable_data(ctx.GetPlace()); + int64_t* argmax_data = argmax->mutable_data(ctx.GetPlace()); + + const int64_t* rois_data = rois->data(); for (int n = 0; n < rois_num; ++n) { - int roi_batch_id = rois_data[0]; - int roi_start_w = round(rois_data[1] * spatial_scale); - int roi_start_h = round(rois_data[2] * spatial_scale); - int roi_end_w = round(rois_data[3] * spatial_scale); - int roi_end_h = round(rois_data[4] * spatial_scale); + int roi_batch_id = roi_batch_id_data[n]; + int roi_start_w = round(rois_data[0] * spatial_scale); + int roi_start_h = round(rois_data[1] * spatial_scale); + int roi_end_w = round(rois_data[2] * spatial_scale); + int roi_end_h = round(rois_data[3] * spatial_scale); // Force malformed ROIs to be 1x1 int roi_height = std::max(roi_end_h - roi_start_h + 1, 1); @@ -133,7 +147,7 @@ class CPUROIPoolGradOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { auto* in = ctx.Input("X"); - auto* rois = ctx.Input("ROIs"); + auto* rois = ctx.Input("ROIs"); auto* argmax = ctx.Input("Argmax"); auto* out_grad = ctx.Input(framework::GradVarName("Out")); @@ -143,6 +157,20 @@ class CPUROIPoolGradOpKernel : public framework::OpKernel { auto pooled_width = ctx.Attr("pooled_width"); if (in_grad) { + int rois_num = rois->dims()[0]; + framework::Tensor roi_batch_id_list; + roi_batch_id_list.Resize({rois_num}); + int* roi_batch_id_data = + roi_batch_id_list.mutable_data(ctx.GetPlace()); + + auto rois_lod = rois->lod().back(); + int rois_batch_size = rois_lod.size() - 1; + for (int n = 0; n < rois_batch_size; ++n) { + for (size_t i = rois_lod[n]; i < rois_lod[n + 1]; ++i) { + roi_batch_id_data[i] = n; + } + } + const int64_t* rois_data = rois->data(); const T* out_grad_data = out_grad->data(); const int64_t* argmax_data = argmax->data(); @@ -156,11 +184,10 @@ class CPUROIPoolGradOpKernel : public framework::OpKernel { auto roi_stride = framework::stride(rois->dims()); auto out_stride = framework::stride(out_grad->dims()); - int rois_num = rois->dims()[0]; int channels = in->dims()[1]; for (int n = 0; n < rois_num; ++n) { - int roi_batch_idx = rois_data[0]; + int roi_batch_idx = roi_batch_id_data[n]; T* batch_grad_data = in_grad_data + roi_batch_idx * in_stride[0]; for (int c = 0; c < channels; ++c) { for (int ph = 0; ph < pooled_height; ++ph) { diff --git a/paddle/fluid/operators/row_conv_op.cc b/paddle/fluid/operators/row_conv_op.cc index 23f720da0b68cd2fd4c9b51182bf82f72078a906..20f140f962c3aac364a1239a663d5f340bbeb6b2 100644 --- a/paddle/fluid/operators/row_conv_op.cc +++ b/paddle/fluid/operators/row_conv_op.cc @@ -76,8 +76,7 @@ class RowConvGradOp : public framework::OperatorWithKernel { class RowConvOpMaker : public framework::OpProtoAndCheckerMaker { public: - RowConvOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(LoDTensor), the input(X) is a LodTensor, which supports " "variable time-length input sequences. The underlying tensor " diff --git a/paddle/fluid/operators/row_conv_op.cu b/paddle/fluid/operators/row_conv_op.cu index 67083455a7579a4bbb6d9598a77b68a8375cf815..9ae80da6550bcef39c07f05e35d4153c24738f09 100644 --- a/paddle/fluid/operators/row_conv_op.cu +++ b/paddle/fluid/operators/row_conv_op.cu @@ -14,7 +14,7 @@ limitations under the License. */ #include "paddle/fluid/operators/math/math_function.h" #include "paddle/fluid/operators/row_conv_op.h" -#include "paddle/fluid/platform/cuda_helper.h" +#include "paddle/fluid/platform/cuda_device_function.h" namespace paddle { namespace operators { @@ -189,6 +189,10 @@ __global__ void RowConvGradFilterImproved(const T *in, const T *dout, } __syncthreads(); + // NOTE(zcd): temporary solution + unsigned mask = 0u; + CREATE_SHFL_MASK(mask, true); + for (int i = 0; i < num_sequence; i++) { int start = static_cast(batch_indices[i]); int end = static_cast(batch_indices[i + 1]); @@ -220,7 +224,7 @@ __global__ void RowConvGradFilterImproved(const T *in, const T *dout, for (int offset = 16; offset > 0; offset = offset / 2) { // blockDim.x is 32. - val += __shfl_down(val, offset); + val += platform::CudaShuffleDownSync(mask, val, offset); } __syncthreads(); @@ -251,6 +255,10 @@ __global__ void RowConvGradFilter(const T *in, const T *dout, int num_sequence, T *sh_in = mem; T *sh_dout = &mem[block_x * block_y]; + // NOTE(zcd): temporary solution + unsigned mask = 0u; + CREATE_SHFL_MASK(mask, true); + for (int i = 0; i < num_sequence; i++) { int start = static_cast(batch_indices[i]); int end = static_cast(batch_indices[i + 1]); @@ -276,7 +284,7 @@ __global__ void RowConvGradFilter(const T *in, const T *dout, int num_sequence, for (int offset = 16; offset > 0; offset = offset / 2) { // blockDim.x is 32. - val += __shfl_down(val, offset); + val += platform::CudaShuffleDownSync(mask, val, offset); } __syncthreads(); diff --git a/paddle/fluid/operators/save_combine_op.cc b/paddle/fluid/operators/save_combine_op.cc index 94703393bfa53124d16e34ae4373773eece5f11f..cfee9207083b46f7c27354f22e82a7d3c38a027c 100644 --- a/paddle/fluid/operators/save_combine_op.cc +++ b/paddle/fluid/operators/save_combine_op.cc @@ -18,6 +18,7 @@ limitations under the License. */ #include #include #include "paddle/fluid/framework/data_type.h" +#include "paddle/fluid/framework/data_type_transform.h" #include "paddle/fluid/framework/framework.pb.h" #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/op_registry.h" @@ -69,6 +70,7 @@ class SaveCombineOp : public framework::OperatorBase { const platform::Place &place) const override { auto filename = Attr("file_path"); auto overwrite = Attr("overwrite"); + auto save_as_fp16 = Attr("save_as_fp16"); bool is_present = FileExists(filename); if (is_present && !overwrite) { @@ -100,8 +102,24 @@ class SaveCombineOp : public framework::OperatorBase { inp_var_names[i]); auto &tensor = var->Get(); - // Serialize tensor - framework::SerializeToStream(fout, tensor, dev_ctx); + // Serialize tensors one by one + + // Check types to see if a fp16 transformation is required + auto in_dtype = framework::ToDataType(tensor.type()); + auto out_dtype = + save_as_fp16 ? framework::proto::VarType::FP16 : in_dtype; + + if (in_dtype != out_dtype) { + auto in_kernel_type = framework::OpKernelType(in_dtype, place); + auto out_kernel_type = framework::OpKernelType(out_dtype, place); + framework::LoDTensor out; + // copy LoD info to the new tensor + out.set_lod(tensor.lod()); + framework::TransDataType(in_kernel_type, out_kernel_type, tensor, &out); + framework::SerializeToStream(fout, out, dev_ctx); + } else { + framework::SerializeToStream(fout, tensor, dev_ctx); + } } fout.close(); } @@ -109,8 +127,7 @@ class SaveCombineOp : public framework::OperatorBase { class SaveCombineOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - SaveCombineOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput( "X", "(vector) Input LoDTensors that need to be saved together in a file.") @@ -125,6 +142,12 @@ to a file on disk. "(boolean, default true)" "Overwrite the output file if it exists.") .SetDefault(true); + AddAttr("save_as_fp16", + "(boolean, default false)" + "If true, the tensor will be converted to float16 data " + "type and then saved. Otherwise, the tensor will be " + "directly saved without data type conversion.") + .SetDefault(false); AddAttr( "file_path", "(string)" diff --git a/paddle/fluid/operators/save_load_combine_op_test.cc b/paddle/fluid/operators/save_load_combine_op_test.cc index 2773c32a0a10269e28c24e12527711e3c5b8f869..4743e0d9499b111d8baa921dbb245431713fd7a8 100644 --- a/paddle/fluid/operators/save_load_combine_op_test.cc +++ b/paddle/fluid/operators/save_load_combine_op_test.cc @@ -17,15 +17,17 @@ limitations under the License. */ #include #include "gtest/gtest.h" #include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/platform/float16.h" USE_NO_KERNEL_OP(save_combine); USE_NO_KERNEL_OP(load_combine); -int* CreateForSaveCombineOp(int x, int y, const std::vector& lod_info, - std::string var_name, - const paddle::platform::CPUPlace& place, - paddle::framework::Scope* scope, - paddle::framework::LoD* expect_lod) { +template +T* CreateForSaveCombineOp(int x, int y, const std::vector& lod_info, + std::string var_name, + const paddle::platform::CPUPlace& place, + paddle::framework::Scope* scope, + paddle::framework::LoD* expect_lod) { auto var = scope->Var(var_name); auto tensor = var->GetMutable(); tensor->Resize({x, y}); @@ -34,9 +36,10 @@ int* CreateForSaveCombineOp(int x, int y, const std::vector& lod_info, (*expect_lod)[0].push_back(lod_info[i]); } tensor->set_lod(*expect_lod); - int* expect = tensor->mutable_data(place); + T* expect = tensor->mutable_data(place); for (int64_t i = 0; i < tensor->numel(); ++i) { - expect[i] = static_cast(i); + expect[i] = static_cast( + static_cast(i)); // For FP16, we intend to do float(float16(i)) } return expect; } @@ -48,18 +51,20 @@ paddle::framework::LoDTensor* GeneratePlaceholderBeforeLoad( return target; } -int* GetValuesAfterLoadCombineOp(paddle::framework::LoDTensor* target, - const paddle::framework::Scope& scope, - paddle::framework::LoD* actual_lod) { - int* actual = target->data(); +template +T* GetValuesAfterLoadCombineOp(paddle::framework::LoDTensor* target, + const paddle::framework::Scope& scope, + paddle::framework::LoD* actual_lod) { + T* actual = target->data(); *actual_lod = target->lod(); return actual; } -void CheckValues(int* expect, int* actual, paddle::framework::LoD expect_lod, - paddle::framework::LoD actual_lod, const int& numel) { - for (int64_t i = 0; i < numel; ++i) { - EXPECT_EQ(expect[i], actual[i]); +template +void CheckValues(T* expect, U* actual, const paddle::framework::LoD& expect_lod, + const paddle::framework::LoD& actual_lod, const int& numel) { + for (int i = 0; i < numel; ++i) { + EXPECT_EQ(expect[i], static_cast(actual[i])); } EXPECT_EQ(expect_lod.size(), actual_lod.size()); for (size_t i = 0; i < expect_lod.size(); ++i) { @@ -78,26 +83,26 @@ TEST(SaveLoadCombineOp, CPU) { std::vector lod1 = {0, 1, 2, 3, 10}; int numel1 = 100; paddle::framework::LoD expect_lod1; - int* expect1 = CreateForSaveCombineOp(10, 10, lod1, "test_var1", place, - &scope, &expect_lod1); + int* expect1 = CreateForSaveCombineOp(10, 10, lod1, "test_var1", + place, &scope, &expect_lod1); std::vector lod2 = {0, 2, 5, 10}; int numel2 = 200; paddle::framework::LoD expect_lod2; - int* expect2 = CreateForSaveCombineOp(10, 20, lod2, "test_var2", place, - &scope, &expect_lod2); + int* expect2 = CreateForSaveCombineOp(10, 20, lod2, "test_var2", + place, &scope, &expect_lod2); std::vector lod3 = {0, 2, 3, 20}; int numel3 = 4000; paddle::framework::LoD expect_lod3; - int* expect3 = CreateForSaveCombineOp(20, 200, lod3, "test_var3", place, - &scope, &expect_lod3); + int* expect3 = CreateForSaveCombineOp(20, 200, lod3, "test_var3", + place, &scope, &expect_lod3); std::vector lod4 = {0, 1, 20}; int numel4 = 1000; paddle::framework::LoD expect_lod4; - int* expect4 = CreateForSaveCombineOp(20, 50, lod4, "test_var4", place, - &scope, &expect_lod4); + int* expect4 = CreateForSaveCombineOp(20, 50, lod4, "test_var4", + place, &scope, &expect_lod4); // Set attributes std::string filename = "check_tensor.ls"; @@ -123,15 +128,176 @@ TEST(SaveLoadCombineOp, CPU) { load_combine_op->Run(scope, place); paddle::framework::LoD actual_lod1, actual_lod2, actual_lod3, actual_lod4; - int* actual1 = GetValuesAfterLoadCombineOp(target1, scope, &actual_lod1); - int* actual2 = GetValuesAfterLoadCombineOp(target2, scope, &actual_lod2); - int* actual3 = GetValuesAfterLoadCombineOp(target3, scope, &actual_lod3); - int* actual4 = GetValuesAfterLoadCombineOp(target4, scope, &actual_lod4); - - CheckValues(expect1, actual1, expect_lod1, actual_lod1, numel1); - CheckValues(expect2, actual2, expect_lod2, actual_lod2, numel2); - CheckValues(expect3, actual3, expect_lod3, actual_lod3, numel3); - CheckValues(expect4, actual4, expect_lod4, actual_lod4, numel4); + int* actual1 = GetValuesAfterLoadCombineOp(target1, scope, &actual_lod1); + int* actual2 = GetValuesAfterLoadCombineOp(target2, scope, &actual_lod2); + int* actual3 = GetValuesAfterLoadCombineOp(target3, scope, &actual_lod3); + int* actual4 = GetValuesAfterLoadCombineOp(target4, scope, &actual_lod4); + + CheckValues(expect1, actual1, expect_lod1, actual_lod1, numel1); + CheckValues(expect2, actual2, expect_lod2, actual_lod2, numel2); + CheckValues(expect3, actual3, expect_lod3, actual_lod3, numel3); + CheckValues(expect4, actual4, expect_lod4, actual_lod4, numel4); +} + +// FP16 version of SaveLoadCombineOp Test, only altering the saving aspect +// to save as FP16. +TEST(SaveCombineFP16Op, CPU) { + paddle::framework::Scope scope; + paddle::platform::CPUPlace place; + + std::vector lod1 = {0, 1, 2, 3, 10}; + int numel1 = 100; + paddle::framework::LoD expect_lod1; + float* expect1 = CreateForSaveCombineOp( + 10, 10, lod1, "test_var1", place, &scope, &expect_lod1); + + std::vector lod2 = {0, 2, 5, 10}; + int numel2 = 200; + paddle::framework::LoD expect_lod2; + float* expect2 = CreateForSaveCombineOp( + 10, 20, lod2, "test_var2", place, &scope, &expect_lod2); + + std::vector lod3 = {0, 20}; + int numel3 = 4000; + paddle::framework::LoD expect_lod3; + float* expect3 = CreateForSaveCombineOp( + 20, 200, lod3, "test_var3", place, &scope, &expect_lod3); + + std::vector lod4 = {0, 1, 20}; + int numel4 = 1000; + paddle::framework::LoD expect_lod4; + float* expect4 = CreateForSaveCombineOp( + 20, 50, lod4, "test_var4", place, &scope, &expect_lod4); + + // Set attributes + std::string filename = "check_tensor_fp16_save.ls"; + paddle::framework::AttributeMap attrs; + attrs.insert({"file_path", std::string(filename)}); + attrs.insert({"save_as_fp16", true}); + + // Run the save_combine_op + auto save_combine_op = paddle::framework::OpRegistry::CreateOp( + "save_combine", + {{"X", {"test_var1", "test_var2", "test_var3", "test_var4"}}}, {}, attrs); + save_combine_op->Run(scope, place); + + // Set up output vars + auto target1 = GeneratePlaceholderBeforeLoad("out_var1", &scope); + auto target2 = GeneratePlaceholderBeforeLoad("out_var2", &scope); + auto target3 = GeneratePlaceholderBeforeLoad("out_var3", &scope); + auto target4 = GeneratePlaceholderBeforeLoad("out_var4", &scope); + + // Run the load_combine_op + auto load_combine_op = paddle::framework::OpRegistry::CreateOp( + "load_combine", {}, + {{"Out", {"out_var1", "out_var2", "out_var3", "out_var4"}}}, attrs); + load_combine_op->Run(scope, place); + + paddle::framework::LoD actual_lod1, actual_lod2, actual_lod3, actual_lod4; + paddle::platform::float16* actual1 = + GetValuesAfterLoadCombineOp(target1, scope, + &actual_lod1); + paddle::platform::float16* actual2 = + GetValuesAfterLoadCombineOp(target2, scope, + &actual_lod2); + paddle::platform::float16* actual3 = + GetValuesAfterLoadCombineOp(target3, scope, + &actual_lod3); + paddle::platform::float16* actual4 = + GetValuesAfterLoadCombineOp(target4, scope, + &actual_lod4); + + CheckValues(expect1, actual1, expect_lod1, + actual_lod1, numel1); + CheckValues(expect2, actual2, expect_lod2, + actual_lod2, numel2); + CheckValues(expect3, actual3, expect_lod3, + actual_lod3, numel3); + CheckValues(expect4, actual4, expect_lod4, + actual_lod4, numel4); +} + +// FP16 version of SaveLoadCombineOp Test, only altering the loading aspect +// to load tensors with FP16 precision. +TEST(LoadCombineFP16Op, CPU) { + paddle::framework::Scope scope; + paddle::platform::CPUPlace place; + + std::vector lod1 = {0, 1, 2, 3, 10}; + int numel1 = 100; + paddle::framework::LoD expect_lod1; + float* expect1 = CreateForSaveCombineOp( + 10, 10, lod1, "test_var1", place, &scope, &expect_lod1); + + std::vector lod2 = {0, 2, 5, 10}; + int numel2 = 200; + paddle::framework::LoD expect_lod2; + float* expect2 = CreateForSaveCombineOp( + 10, 20, lod2, "test_var2", place, &scope, &expect_lod2); + + std::vector lod3 = {0, 20}; + int numel3 = 4000; + paddle::framework::LoD expect_lod3; + float* expect3 = CreateForSaveCombineOp( + 20, 200, lod3, "test_var3", place, &scope, &expect_lod3); + + std::vector lod4 = {0, 1, 20}; + int numel4 = 1000; + paddle::framework::LoD expect_lod4; + float* expect4 = CreateForSaveCombineOp( + 20, 50, lod4, "test_var4", place, &scope, &expect_lod4); + + // Set attributes + std::string filename = "check_tensor_fp16_load.ls"; + paddle::framework::AttributeMap attrs; + attrs.insert({"file_path", std::string(filename)}); + + // Run the save_combine_op + auto save_combine_op = paddle::framework::OpRegistry::CreateOp( + "save_combine", + {{"X", {"test_var1", "test_var2", "test_var3", "test_var4"}}}, {}, attrs); + save_combine_op->Run(scope, place); + + // Set up output vars + auto load_var1 = scope.Var("out_var1"); + auto load_var2 = scope.Var("out_var2"); + auto load_var3 = scope.Var("out_var3"); + auto load_var4 = scope.Var("out_var4"); + + attrs.insert({"load_as_fp16", true}); + // Run the load_combine_op + auto load_combine_op = paddle::framework::OpRegistry::CreateOp( + "load_combine", {}, + {{"Out", {"out_var1", "out_var2", "out_var3", "out_var4"}}}, attrs); + load_combine_op->Run(scope, place); + + auto* target1 = load_var1->GetMutable(); + auto* target2 = load_var2->GetMutable(); + auto* target3 = load_var3->GetMutable(); + auto* target4 = load_var4->GetMutable(); + + paddle::framework::LoD actual_lod1, actual_lod2, actual_lod3, actual_lod4; + paddle::platform::float16* actual1 = + GetValuesAfterLoadCombineOp(target1, scope, + &actual_lod1); + paddle::platform::float16* actual2 = + GetValuesAfterLoadCombineOp(target2, scope, + &actual_lod2); + paddle::platform::float16* actual3 = + GetValuesAfterLoadCombineOp(target3, scope, + &actual_lod3); + paddle::platform::float16* actual4 = + GetValuesAfterLoadCombineOp(target4, scope, + &actual_lod4); + + CheckValues(expect1, actual1, expect_lod1, + actual_lod1, numel1); + CheckValues(expect2, actual2, expect_lod2, + actual_lod2, numel2); + CheckValues(expect3, actual3, expect_lod3, + actual_lod3, numel3); + CheckValues(expect4, actual4, expect_lod4, + actual_lod4, numel4); } // Test with original SaveLoadTest @@ -141,7 +307,7 @@ TEST(SaveLoadTestWithCombineOp, CPU) { auto var = scope.Var("test_var"); auto tensor = var->GetMutable(); - tensor->Resize({3, 10}); + tensor->Resize({3, 4000}); paddle::framework::LoD expect_lod; expect_lod.resize(1); expect_lod[0].push_back(0); diff --git a/paddle/fluid/operators/save_load_op_test.cc b/paddle/fluid/operators/save_load_op_test.cc index a7ba1e0ae1d22a22cf2943c9aaf0c394ef4ae326..c4fcc61af4b75e6dc7d5c31e20c5fff358637af5 100644 --- a/paddle/fluid/operators/save_load_op_test.cc +++ b/paddle/fluid/operators/save_load_op_test.cc @@ -14,6 +14,7 @@ limitations under the License. */ #include "gtest/gtest.h" #include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/platform/float16.h" USE_NO_KERNEL_OP(save); USE_NO_KERNEL_OP(load); @@ -61,3 +62,98 @@ TEST(SaveLoadOp, CPU) { } } } + +TEST(SaveFP16Op, CPU) { + paddle::framework::Scope scope; + paddle::platform::CPUPlace place; + + auto var = scope.Var("test_var"); + auto tensor = var->GetMutable(); + tensor->Resize({3, 10}); + paddle::framework::LoD expect_lod; + expect_lod.resize(1); + expect_lod[0].push_back(0); + expect_lod[0].push_back(1); + expect_lod[0].push_back(2); + expect_lod[0].push_back(3); + + tensor->set_lod(expect_lod); + float* expect = tensor->mutable_data(place); + for (int64_t i = 0; i < tensor->numel(); ++i) { + expect[i] = static_cast(paddle::platform::float16(i)); + } + + paddle::framework::AttributeMap attrs; + attrs.insert({"file_path", std::string("tensor.save")}); + attrs.insert({"save_as_fp16", true}); + + auto save_op = paddle::framework::OpRegistry::CreateOp( + "save", {{"X", {"test_var"}}}, {}, attrs); + save_op->Run(scope, place); + + auto load_var = scope.Var("out_var"); + auto target = load_var->GetMutable(); + auto load_op = paddle::framework::OpRegistry::CreateOp( + "load", {}, {{"Out", {"out_var"}}}, attrs); + load_op->Run(scope, place); + paddle::platform::float16* actual = target->data(); + for (int64_t i = 0; i < tensor->numel(); ++i) { + EXPECT_EQ(expect[i], static_cast(actual[i])); + } + auto& actual_lod = target->lod(); + EXPECT_EQ(expect_lod.size(), actual_lod.size()); + for (size_t i = 0; i < expect_lod.size(); ++i) { + for (size_t j = 0; j < expect_lod[i].size(); ++j) { + EXPECT_EQ(expect_lod[i][j], actual_lod[i][j]); + } + } +} + +TEST(LoadFP16Op, CPU) { + paddle::framework::Scope scope; + paddle::platform::CPUPlace place; + + auto var = scope.Var("test_var"); + auto tensor = var->GetMutable(); + tensor->Resize({3, 10}); + + paddle::framework::LoD expect_lod; + expect_lod.resize(1); + expect_lod[0].push_back(0); + expect_lod[0].push_back(1); + expect_lod[0].push_back(2); + expect_lod[0].push_back(3); + + tensor->set_lod(expect_lod); + float* expect = tensor->mutable_data(place); + for (int64_t i = 0; i < tensor->numel(); ++i) { + expect[i] = static_cast(paddle::platform::float16(i)); + } + + paddle::framework::AttributeMap attrs; + attrs.insert({"file_path", std::string("tensor.save")}); + attrs.insert({"load_as_fp16", true}); + + auto save_op = paddle::framework::OpRegistry::CreateOp( + "save", {{"X", {"test_var"}}}, {}, attrs); + save_op->Run(scope, place); + + auto load_var = scope.Var("out_var"); + auto load_op = paddle::framework::OpRegistry::CreateOp( + "load", {}, {{"Out", {"out_var"}}}, attrs); + load_op->Run(scope, place); + + auto target = load_var->Get(); + paddle::platform::float16* actual = target.data(); + for (int64_t i = 0; i < tensor->numel(); ++i) { + EXPECT_EQ(expect[i], static_cast(actual[i])); + } + + auto& actual_lod = target.lod(); + EXPECT_EQ(expect_lod.size(), actual_lod.size()); + for (size_t i = 0; i < expect_lod.size(); ++i) { + for (size_t j = 0; j < expect_lod[i].size(); ++j) { + EXPECT_EQ(expect_lod[i][j], actual_lod[i][j]); + } + } +} diff --git a/paddle/fluid/operators/save_op.cc b/paddle/fluid/operators/save_op.cc index 4a715c4baab2da7b7af86ada22ee88a16b05a814..e6d27e2dedd7668b93bd8ddc330a897d1c6fa732 100644 --- a/paddle/fluid/operators/save_op.cc +++ b/paddle/fluid/operators/save_op.cc @@ -18,6 +18,7 @@ limitations under the License. */ #include #include "paddle/fluid/framework/data_type.h" +#include "paddle/fluid/framework/data_type_transform.h" #include "paddle/fluid/framework/framework.pb.h" #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/op_registry.h" @@ -68,6 +69,7 @@ class SaveOp : public framework::OperatorBase { const platform::Place &place) const override { auto filename = Attr("file_path"); auto overwrite = Attr("overwrite"); + auto save_as_fp16 = Attr("save_as_fp16"); if (FileExists(filename) && !overwrite) { PADDLE_THROW("%s is existed, cannot save to it when overwrite=false", @@ -96,14 +98,26 @@ class SaveOp : public framework::OperatorBase { platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance(); auto &dev_ctx = *pool.Get(place); - framework::SerializeToStream(fout, tensor, dev_ctx); + auto in_dtype = framework::ToDataType(tensor.type()); + auto out_dtype = save_as_fp16 ? framework::proto::VarType::FP16 : in_dtype; + + if (in_dtype != out_dtype) { + auto in_kernel_type = framework::OpKernelType(in_dtype, place); + auto out_kernel_type = framework::OpKernelType(out_dtype, place); + framework::LoDTensor out; + framework::TransDataType(in_kernel_type, out_kernel_type, tensor, &out); + // copy LoD info to the new tensor + out.set_lod(tensor.lod()); + framework::SerializeToStream(fout, out, dev_ctx); + } else { + framework::SerializeToStream(fout, tensor, dev_ctx); + } } }; class SaveOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - SaveOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor ) Input tensor to be saved"); AddComment(R"DOC( Save operator @@ -114,6 +128,12 @@ This operator will serialize and write a tensor variable to file on disk. "(boolean, default true)" "Overwrite the output file if exist") .SetDefault(true); + AddAttr("save_as_fp16", + "(boolean, default false)" + "If true, the tensor will be converted to float16 data " + "type and then saved. Otherwise, the tensor will be " + "directly saved without data type conversion.") + .SetDefault(false); AddAttr("file_path", "(string)" "The \"file_path\" where the variable will be saved.") diff --git a/paddle/fluid/operators/scale_op.cc b/paddle/fluid/operators/scale_op.cc index 1e938638c9182972a2ae2436166ff0aa49efd4be..4687e21e7155fc7309fb28c881c0d47152df9ad5 100644 --- a/paddle/fluid/operators/scale_op.cc +++ b/paddle/fluid/operators/scale_op.cc @@ -35,11 +35,9 @@ class ScaleOp : public framework::OperatorWithKernel { } }; -template class ScaleOpMaker : public framework::OpProtoAndCheckerMaker { public: - ScaleOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor) Input tensor of scale operator."); AddOutput("Out", "(Tensor) Output tensor of scale operator."); AddComment(R"DOC( @@ -47,9 +45,9 @@ Scale operator $$Out = scale*X$$ )DOC"); - AddAttr("scale", - "(float, default 1.0)" - "The scaling factor of the scale operator.") + AddAttr("scale", + "(float, default 1.0)" + "The scaling factor of the scale operator.") .SetDefault(1.0); } }; @@ -73,8 +71,7 @@ class ScaleGradMaker : public framework::SingleGradOpDescMaker { namespace ops = paddle::operators; -REGISTER_OPERATOR(scale, ops::ScaleOp, ops::ScaleOpMaker, - ops::ScaleGradMaker); +REGISTER_OPERATOR(scale, ops::ScaleOp, ops::ScaleOpMaker, ops::ScaleGradMaker); REGISTER_OP_CPU_KERNEL( scale, ops::ScaleKernel, ops::ScaleKernel, diff --git a/paddle/fluid/operators/scatter_op.cc b/paddle/fluid/operators/scatter_op.cc index 95b12455ea4996f00bab8a353ccd425b2c37aed1..bf5e0d864495ce3a651a31c9d5a7664fe9eb2396 100644 --- a/paddle/fluid/operators/scatter_op.cc +++ b/paddle/fluid/operators/scatter_op.cc @@ -78,8 +78,7 @@ class ScatterGradOp : public framework::OperatorWithKernel { class ScatterOpMaker : public framework::OpProtoAndCheckerMaker { public: - ScatterOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The source input of scatter op"); AddInput("Ids", "The index input of scatter op where X will be updated"); AddInput("Updates", "The updated value of updates op"); diff --git a/paddle/fluid/operators/select_op.cc b/paddle/fluid/operators/select_op.cc index 876d8acf0d880a7ef806514014d297f98e04c53d..e71841d4d1815d50cd9800910c9db34e121beffc 100644 --- a/paddle/fluid/operators/select_op.cc +++ b/paddle/fluid/operators/select_op.cc @@ -380,8 +380,7 @@ class SelectOp : public framework::OperatorBase { class SelectOpMaker : public framework::OpProtoAndCheckerMaker { public: - SelectOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput(kX, "A set of variables, which are required by operators inside the " "cases of Select Op") diff --git a/paddle/fluid/operators/send_barrier_op.cc b/paddle/fluid/operators/send_barrier_op.cc index 12b844daaa33162b86b7daffa2e4c49785701662..2c77ee2e2792d6fdd76bacd68b6c3b4a296b2e3a 100644 --- a/paddle/fluid/operators/send_barrier_op.cc +++ b/paddle/fluid/operators/send_barrier_op.cc @@ -21,6 +21,7 @@ limitations under the License. */ #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/operators/detail/grpc_client.h" +#include "paddle/fluid/platform/profiler.h" namespace paddle { namespace operators { @@ -36,32 +37,30 @@ class SendBarrierOp : public framework::OperatorBase { void RunImpl(const framework::Scope& scope, const platform::Place& place) const override { std::vector eps = Attr>("endpoints"); + bool sync_mode = Attr("sync_mode"); - auto client_var_name = Output("RPCClient"); - PADDLE_ENFORCE_NOT_NULL(scope.FindVar(client_var_name), - "Can not find variable '%s' in the scope.", - client_var_name); - auto* client_var = scope.FindVar(client_var_name); - detail::RPCClient* rpc_client = client_var->GetMutable(); + platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); + auto& ctx = *pool.Get(place); + // For profiling + platform::RecordEvent record_event(Type(), &ctx); + + auto rpc_client = detail::RPCClient::GetInstance(); // need to wait before sending send_barrier message PADDLE_ENFORCE(rpc_client->Wait()); - - for (auto& ep : eps) { - VLOG(3) << "send barrier, ep: " << ep; - rpc_client->AsyncSendBatchBarrier(ep); + if (sync_mode) { + for (auto& ep : eps) { + VLOG(3) << "send barrier, ep: " << ep; + rpc_client->AsyncSendBatchBarrier(ep); + } + PADDLE_ENFORCE(rpc_client->Wait()); } - PADDLE_ENFORCE(rpc_client->Wait()); } }; class SendBarrierOpMaker : public framework::OpProtoAndCheckerMaker { public: - SendBarrierOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { - AddOutput("RPCClient", - "(RPCClient) The RPC client object which is" - "initialized at most once."); + void Make() { AddComment(R"DOC( SendBarrier operator @@ -73,17 +72,7 @@ the Parameter Server would knew all variables have been sent. "(string vector, default 127.0.0.1:6164)" "Server endpoints to send variables to.") .SetDefault({"127.0.0.1:6164"}); - } -}; - -class SendBarrierOpVarTypeInference : public framework::VarTypeInference { - public: - void operator()(const framework::OpDesc& op_desc, - framework::BlockDesc* block) const override { - auto out_var_name = op_desc.Output("RPCClient").front(); - auto& out_var = block->FindRecursiveOrCreateVar(out_var_name); - auto var_type = framework::proto::VarType::RAW; - out_var.SetType(var_type); + AddAttr("sync_mode", "work in sync_mode or not").SetDefault(true); } }; @@ -99,5 +88,4 @@ namespace ops = paddle::operators; REGISTER_OPERATOR(send_barrier, ops::SendBarrierOp, paddle::framework::EmptyGradOpMaker, ops::SendBarrierOpMaker, - ops::SendBarrierOpVarTypeInference, ops::SendBarrierOpShapeInference); diff --git a/paddle/fluid/operators/send_op.cc b/paddle/fluid/operators/send_op.cc index e4386b640a298cd216bb60104653f20c4a96e7dc..a5150f242ca3b0befafa2443f0bc466e2aea85e4 100644 --- a/paddle/fluid/operators/send_op.cc +++ b/paddle/fluid/operators/send_op.cc @@ -49,12 +49,7 @@ class SendOp : public framework::OperatorBase { // For profiling platform::RecordEvent record_event(Type(), &ctx); - auto client_var_name = Output("RPCClient"); - PADDLE_ENFORCE_NOT_NULL(scope.FindVar(client_var_name), - "Can not find variable '%s' in the scope.", - client_var_name); - auto* client_var = scope.FindVar(client_var_name); - detail::RPCClient* rpc_client = client_var->GetMutable(); + auto rpc_client = detail::RPCClient::GetInstance(); for (size_t i = 0; i < ins.size(); i++) { if (NeedSend(scope, ins[i])) { @@ -92,14 +87,10 @@ class SendOp : public framework::OperatorBase { class SendOpMaker : public framework::OpProtoAndCheckerMaker { public: - SendOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() { AddInput("X", "(Tensor) Input tensor to be sent").AsDuplicable(); AddOutput("Out", "(Tensor) Output tensor to be received from server") .AsDuplicable(); - AddOutput("RPCClient", - "(RPCClient) The RPC client object which is" - "initialized at most once."); AddComment(R"DOC( Send operator @@ -120,17 +111,6 @@ This operator will send tensor to recv_op at the parameter server. } }; -class SendOpVarTypeInference : public framework::VarTypeInference { - public: - void operator()(const framework::OpDesc& op_desc, - framework::BlockDesc* block) const override { - auto out_var_name = op_desc.Output("RPCClient").front(); - auto& out_var = block->FindRecursiveOrCreateVar(out_var_name); - auto var_type = framework::proto::VarType::RAW; - out_var.SetType(var_type); - } -}; - class SendOpShapeInference : public framework::InferShapeBase { public: void operator()(framework::InferShapeContext* ctx) const override {} @@ -142,5 +122,4 @@ class SendOpShapeInference : public framework::InferShapeBase { namespace ops = paddle::operators; REGISTER_OPERATOR(send, ops::SendOp, paddle::framework::EmptyGradOpMaker, - ops::SendOpMaker, ops::SendOpVarTypeInference, - ops::SendOpShapeInference); + ops::SendOpMaker, ops::SendOpShapeInference); diff --git a/paddle/fluid/operators/send_recv_op_test.cc b/paddle/fluid/operators/send_recv_op_test.cc index d2e1f3cb2ff9c8254cd4815a0f8750966a6e161c..e550552b195b768d68ec64e9c3b5889b56ca719f 100644 --- a/paddle/fluid/operators/send_recv_op_test.cc +++ b/paddle/fluid/operators/send_recv_op_test.cc @@ -92,12 +92,16 @@ void InitSelectedRowsInScope(const p::CPUPlace &place, f::Scope *scope) { void AddOp(const std::string &type, const f::VariableNameMap &inputs, const f::VariableNameMap &outputs, f::AttributeMap attrs, - f::BlockDesc *block) { + f::BlockDesc *block, bool is_sparse) { // insert output for (auto kv : outputs) { for (auto v : kv.second) { auto var = block->Var(v); var->SetDataType(f::proto::VarType::FP32); + var->SetPersistable(true); + if (is_sparse) { + var->SetType(f::proto::VarType::SELECTED_ROWS); + } } } @@ -113,23 +117,23 @@ void AddOp(const std::string &type, const f::VariableNameMap &inputs, op->SetAttrMap(attrs); } -void StartServerNet(bool is_sparse) { +void StartServerNet(bool is_sparse, std::atomic *initialized) { f::Scope scope; p::CPUPlace place; + VLOG(4) << "before init tensor"; if (is_sparse) { InitSelectedRowsInScope(place, &scope); } else { InitTensorsInScope(place, &scope); } - // sub program run in listen_and_serv_op, for simple test we use sum f::ProgramDesc program; const auto &root_block = program.Block(0); auto *optimize_block = program.AppendBlock(root_block); auto *prefetch_block = program.AppendBlock(root_block); // X for server side tensors, RX for received tensors, must be of same shape. - AddOp("sum", {{"X", {"x0", "x1"}}}, {{"Out", {"Out"}}}, {}, optimize_block); - + AddOp("sum", {{"X", {"x0", "x1"}}}, {{"Out", {"Out"}}}, {}, optimize_block, + is_sparse); f::AttributeMap attrs; attrs.insert({"endpoint", std::string("127.0.0.1:0")}); attrs.insert({"Fanin", 1}); @@ -139,15 +143,23 @@ void StartServerNet(bool is_sparse) { attrs.insert({"PrefetchBlock", prefetch_block}); attrs.insert({"grad_to_block_id", std::vector({""})}); attrs.insert({"sync_mode", true}); + VLOG(4) << "before init op"; listen_and_serv_op = f::OpRegistry::CreateOp("listen_and_serv", {{"X", {"x1"}}}, {}, attrs); + *initialized = true; listen_and_serv_op->Run(scope, place); LOG(INFO) << "server exit"; } TEST(SendRecvOp, CPUDense) { - std::thread server_thread(StartServerNet, false); - sleep(5); // wait server to start + std::atomic initialized{false}; + std::thread server_thread(StartServerNet, false, &initialized); + while (!initialized) { + } + + static_cast(listen_and_serv_op.get()) + ->WaitServerReady(); + // local net f::Scope scope; p::CPUPlace place; @@ -156,15 +168,18 @@ TEST(SendRecvOp, CPUDense) { scope.Var("RPC_CLIENT_VAR"); f::AttributeMap attrs; - selected_port = static_cast( - listen_and_serv_op.get()) - ->GetSelectedPort(); + auto *listen_and_serv_op_ptr = + static_cast( + listen_and_serv_op.get()); + ASSERT_TRUE(listen_and_serv_op_ptr != nullptr); + selected_port = listen_and_serv_op_ptr->GetSelectedPort(); std::string endpoint = paddle::string::Sprintf("127.0.0.1:%d", selected_port); attrs.insert({"endpoints", std::vector({endpoint})}); attrs.insert({"epmap", std::vector({endpoint})}); - auto send_op = f::OpRegistry::CreateOp( - "send", {{"X", {"x1"}}}, - {{"Out", {"Out"}}, {"RPCClient", {"RPC_CLIENT_VAR"}}}, attrs); + const f::VariableNameMap &inputs = {{"X", {"x1"}}}; + const f::VariableNameMap &outputs = {{"Out", {"Out"}}}; + + auto send_op = f::OpRegistry::CreateOp("send", inputs, outputs, attrs); send_op->Run(scope, place); auto in_var = scope.Var("x1"); @@ -181,11 +196,21 @@ TEST(SendRecvOp, CPUDense) { listen_and_serv_op->Stop(); server_thread.join(); listen_and_serv_op.reset(nullptr); + paddle::operators::ListenAndServOp::ResetPort(); } TEST(SendRecvOp, CPUSparse) { - std::thread server_thread(StartServerNet, true); - sleep(3); // wait server to start + std::atomic initialized; + initialized = false; + std::thread server_thread(StartServerNet, true, &initialized); + while (!initialized) { + } + auto *listen_and_serv_op_ptr = + static_cast( + listen_and_serv_op.get()); + ASSERT_TRUE(listen_and_serv_op_ptr != nullptr); + listen_and_serv_op_ptr->WaitServerReady(); + // local net f::Scope scope; p::CPUPlace place; @@ -193,15 +218,12 @@ TEST(SendRecvOp, CPUSparse) { InitSelectedRowsInScope(place, &scope); scope.Var("RPC_CLIENT_VAR"); f::AttributeMap attrs; - selected_port = static_cast( - listen_and_serv_op.get()) - ->GetSelectedPort(); + selected_port = listen_and_serv_op_ptr->GetSelectedPort(); std::string endpoint = paddle::string::Sprintf("127.0.0.1:%d", selected_port); attrs.insert({"endpoints", std::vector({endpoint})}); attrs.insert({"epmap", std::vector({endpoint})}); - auto send_op = f::OpRegistry::CreateOp( - "send", {{"X", {"x1"}}}, - {{"Out", {"Out"}}, {"RPCClient", {"RPC_CLIENT_VAR"}}}, attrs); + auto send_op = f::OpRegistry::CreateOp("send", {{"X", {"x1"}}}, + {{"Out", {"Out"}}}, attrs); send_op->Run(scope, place); auto x0 = scope.Var("x0")->GetMutable(); @@ -226,4 +248,5 @@ TEST(SendRecvOp, CPUSparse) { listen_and_serv_op->Stop(); server_thread.join(); listen_and_serv_op.reset(); + paddle::operators::ListenAndServOp::ResetPort(); } diff --git a/paddle/fluid/operators/send_recv_util.h b/paddle/fluid/operators/send_recv_util.h index 113513eb6b327773ab4a1c062fb8a3f06fddfbca..deab005149027caffa962783df944fad7110382f 100644 --- a/paddle/fluid/operators/send_recv_util.h +++ b/paddle/fluid/operators/send_recv_util.h @@ -20,6 +20,9 @@ namespace operators { inline bool NeedSend(const framework::Scope& scope, const std::string& varname) { + // dummy variable is only used in parallel executor to represent + // some dependency relationship, we don't need to send/recv it. + if (varname == "dummy") return false; auto* var = scope.FindVar(varname); PADDLE_ENFORCE_NOT_NULL(var, "Can not find variable '%s' in the send side.", varname); diff --git a/paddle/fluid/operators/send_vars_op.cc b/paddle/fluid/operators/send_vars_op.cc index 56b3713d6af28d0787e114a672a503e86cbd85fd..fe839dab6924618c8a4c39868d9bf86056a0be40 100644 --- a/paddle/fluid/operators/send_vars_op.cc +++ b/paddle/fluid/operators/send_vars_op.cc @@ -20,6 +20,7 @@ limitations under the License. */ #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/operators/detail/grpc_client.h" #include "paddle/fluid/operators/send_recv_util.h" +#include "paddle/fluid/platform/profiler.h" namespace paddle { namespace operators { @@ -41,12 +42,10 @@ class SendVarsOp : public framework::OperatorBase { platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); auto& ctx = *pool.Get(place); - auto client_var_name = Output("RPCClient"); - PADDLE_ENFORCE_NOT_NULL(scope.FindVar(client_var_name), - "Can not find variable '%s' in the scope.", - client_var_name); - auto* client_var = scope.FindVar(client_var_name); - detail::RPCClient* rpc_client = client_var->GetMutable(); + // For profiling + platform::RecordEvent record_event(Type(), &ctx); + + auto rpc_client = detail::RPCClient::GetInstance(); for (size_t i = 0; i < ins.size(); i++) { if (NeedSend(scope, ins[i])) { @@ -66,13 +65,9 @@ class SendVarsOp : public framework::OperatorBase { class SendVarsOpMaker : public framework::OpProtoAndCheckerMaker { public: - SendVarsOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() { AddInput("X", "(Tensor, SelectedRows) Input variables to be sent") .AsDuplicable(); - AddOutput("RPCClient", - "(RPCClient) The RPC client object which will be" - "initialized at most once."); AddComment(R"DOC( Send operator @@ -90,17 +85,6 @@ This operator will send variables to listen_and_serve op at the parameter server } }; -class SendVarsOpVarTypeInference : public framework::VarTypeInference { - public: - void operator()(const framework::OpDesc& op_desc, - framework::BlockDesc* block) const override { - auto out_var_name = op_desc.Output("RPCClient").front(); - auto& out_var = block->FindRecursiveOrCreateVar(out_var_name); - auto var_type = framework::proto::VarType::RAW; - out_var.SetType(var_type); - } -}; - class SendVarsOpShapeInference : public framework::InferShapeBase { public: void operator()(framework::InferShapeContext* ctx) const override {} @@ -113,5 +97,4 @@ namespace ops = paddle::operators; REGISTER_OPERATOR(send_vars, ops::SendVarsOp, paddle::framework::EmptyGradOpMaker, ops::SendVarsOpMaker, - ops::SendVarsOpVarTypeInference, ops::SendVarsOpShapeInference); diff --git a/paddle/fluid/operators/sequence_concat_op.cc b/paddle/fluid/operators/sequence_concat_op.cc index 3c21903e3a08dcfb55c6c07370a117d0ad633e69..077b9a5f7d935a39706ef3c2b710522bf1b713ed 100644 --- a/paddle/fluid/operators/sequence_concat_op.cc +++ b/paddle/fluid/operators/sequence_concat_op.cc @@ -43,8 +43,7 @@ class SequenceConcatOp : public framework::OperatorWithKernel { class SequenceConcatOpMaker : public framework::OpProtoAndCheckerMaker { public: - SequenceConcatOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(LodTensorArray) Input is a vector of LoDTensor, " "each of which is a variable-length sequence or nested sequence.") diff --git a/paddle/fluid/operators/sequence_conv_op.cc b/paddle/fluid/operators/sequence_conv_op.cc index 94f4b49b0018fdbff6e67c3c081aa5706ccb2e66..ec6cb24350ae276724aae339590d40be1e9ea400 100644 --- a/paddle/fluid/operators/sequence_conv_op.cc +++ b/paddle/fluid/operators/sequence_conv_op.cc @@ -102,8 +102,7 @@ class SequenceConvGradOp : public framework::OperatorWithKernel { class SequenceConvOpMaker : public framework::OpProtoAndCheckerMaker { public: - SequenceConvOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput( "X", "(LoDTensor) the input(X) is a LodTensor, which supports " diff --git a/paddle/fluid/operators/sequence_conv_op.h b/paddle/fluid/operators/sequence_conv_op.h index 3916cdbb6a69c5a18f7a21ec60bad2732b4c3e58..ee70281d51673b94a1451f636e607fad3404863b 100644 --- a/paddle/fluid/operators/sequence_conv_op.h +++ b/paddle/fluid/operators/sequence_conv_op.h @@ -58,17 +58,15 @@ class SequenceConvKernel : public framework::OpKernel { // Because if padding_trainable is false, padding data should be zeros. math::SetConstant set_zero; auto& dev_ctx = context.template device_context(); + auto blas = math::GetBlas(dev_ctx); set_zero(dev_ctx, &col, static_cast(0)); - math::ContextProjectFunctor seq_project_functor; seq_project_functor(dev_ctx, *in, *padding_data, padding_trainable, context_start, context_length, context_stride, up_pad, down_pad, &col); - math::matmul(dev_ctx, col, false, filter, false, - static_cast(1.0), out, - static_cast(0.0)); + blas.MatMul(col, filter, out); } }; @@ -99,6 +97,7 @@ class SequenceConvGradKernel : public framework::OpKernel { math::SetConstant set_zero; auto& dev_ctx = context.template device_context(); + auto blas = math::GetBlas(dev_ctx); // use col_shape in the im2col calculation framework::DDim col_shape = {in->dims()[0], sequence_width * context_length}; @@ -108,8 +107,7 @@ class SequenceConvGradKernel : public framework::OpKernel { col.mutable_data(col_shape, context.GetPlace()); // Because if padding_trainable is false, padding data should be zeros. set_zero(dev_ctx, &col, static_cast(0)); - math::matmul(dev_ctx, *out_g, false, *filter, true, - T(1.0), &col, T(1.0)); + blas.MatMul(*out_g, false, *filter, true, &col); } math::ContextProjectFunctor seq_project_functor; math::ContextProjectGradFunctor seq_project_grad_functor; @@ -150,8 +148,7 @@ class SequenceConvGradKernel : public framework::OpKernel { context_start, context_length, context_stride, up_pad, down_pad, &col); - math::matmul(dev_ctx, col, true, out_grad, false, - T(1.0), &filter_grad, T(1.0)); + blas.MatMul(col, true, out_grad, false, &filter_grad); } } }; diff --git a/paddle/fluid/operators/sequence_erase_op.cc b/paddle/fluid/operators/sequence_erase_op.cc index 73c0e89512972cda002bd902ee0c78b4b77d8502..1c86486157a02c3b78ed61e840fd8e452b9cb452 100644 --- a/paddle/fluid/operators/sequence_erase_op.cc +++ b/paddle/fluid/operators/sequence_erase_op.cc @@ -37,8 +37,7 @@ class SequenceEraseOp : public framework::OperatorWithKernel { class SequenceEraseOpMaker : public framework::OpProtoAndCheckerMaker { public: - SequenceEraseOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(2-D LoDTensor with the 2nd dim. equal to 1) " "Input LoDTensor of SequenceEraseOp."); diff --git a/paddle/fluid/operators/sequence_erase_op.cu b/paddle/fluid/operators/sequence_erase_op.cu index fc9b91c351defb92246e0966b9993fd1e288aaac..3a58e47f1132cd1ac85584b2470e8c6cddcfb28a 100644 --- a/paddle/fluid/operators/sequence_erase_op.cu +++ b/paddle/fluid/operators/sequence_erase_op.cu @@ -15,7 +15,7 @@ limitations under the License. */ #include #include #include "paddle/fluid/operators/sequence_erase_op.h" -#include "paddle/fluid/platform/cuda_helper.h" +#include "paddle/fluid/platform/cuda_primitives.h" namespace paddle { namespace operators { diff --git a/paddle/fluid/operators/sequence_expand_op.cc b/paddle/fluid/operators/sequence_expand_op.cc index 84a35d7172a567a3f6505559fa45a32290288533..944c7f85e5f43679e1875fcce813382be2ba5526 100644 --- a/paddle/fluid/operators/sequence_expand_op.cc +++ b/paddle/fluid/operators/sequence_expand_op.cc @@ -94,8 +94,7 @@ class SequenceExpandOp : public framework::OperatorWithKernel { class SequenceExpandOpMaker : public framework::OpProtoAndCheckerMaker { public: - SequenceExpandOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(LoDTensor, default LoDTensor) A 2-D LoDTensor whose lod " "level is at most 1."); diff --git a/paddle/fluid/operators/sequence_expand_op.cu b/paddle/fluid/operators/sequence_expand_op.cu index c00765e5d59af068e5682b39ebace5f3d7a62250..550677b22694085059e914678a5361d914b455bc 100644 --- a/paddle/fluid/operators/sequence_expand_op.cu +++ b/paddle/fluid/operators/sequence_expand_op.cu @@ -14,7 +14,7 @@ limitations under the License. */ #include #include "paddle/fluid/operators/sequence_expand_op.h" -#include "paddle/fluid/platform/cuda_helper.h" +#include "paddle/fluid/platform/cuda_primitives.h" namespace paddle { namespace operators { diff --git a/paddle/fluid/operators/sequence_pool_op.cc b/paddle/fluid/operators/sequence_pool_op.cc index 933c8c26239d49221819a583f999389ed6fb6cb6..5c6fd13d42e43e3502a1cab85a56e019420c708d 100644 --- a/paddle/fluid/operators/sequence_pool_op.cc +++ b/paddle/fluid/operators/sequence_pool_op.cc @@ -38,8 +38,7 @@ class SequencePoolOp : public framework::OperatorWithKernel { class SequencePoolOpMaker : public framework::OpProtoAndCheckerMaker { public: - SequencePoolOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(LoDTensor) The variable-length input of SequencePoolOp"); AddOutput("Out", "(Tensor) The output of SequencePoolOp does not contain LoD " diff --git a/paddle/fluid/operators/sequence_reshape_op.cc b/paddle/fluid/operators/sequence_reshape_op.cc index a2999650b8903f9d819a8e8011421349e098b219..ef5e6f3210234d59298fcf04c812390643c693d0 100644 --- a/paddle/fluid/operators/sequence_reshape_op.cc +++ b/paddle/fluid/operators/sequence_reshape_op.cc @@ -42,8 +42,7 @@ class SequenceReshapeOp : public framework::OperatorWithKernel { class SequenceReshapeOpMaker : public framework::OpProtoAndCheckerMaker { public: - SequenceReshapeOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(LoDTensor, default LoDTensor) A 2-D LoDTensor with shape " "being [N, M]."); diff --git a/paddle/fluid/operators/sequence_slice_op.cc b/paddle/fluid/operators/sequence_slice_op.cc index 7cd620af07fa9b5f8fcee3c0f88207ef2800c4a1..df9243dc04c584d70dfa6ca78d5fac8423796466 100644 --- a/paddle/fluid/operators/sequence_slice_op.cc +++ b/paddle/fluid/operators/sequence_slice_op.cc @@ -79,8 +79,7 @@ class SequenceSliceGradOp : public framework::OperatorWithKernel { class SequenceSliceOpMaker : public framework::OpProtoAndCheckerMaker { public: - SequenceSliceOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(LoDTensor), " "the input of SequenceSliceOp."); diff --git a/paddle/fluid/operators/sequence_slice_op.h b/paddle/fluid/operators/sequence_slice_op.h index b9c565cac9581a2e830697c1136919062eef345c..b5ea6ff49bbb29571f9a6ef6358ef881acd9be9e 100644 --- a/paddle/fluid/operators/sequence_slice_op.h +++ b/paddle/fluid/operators/sequence_slice_op.h @@ -66,13 +66,11 @@ class SequenceSliceOpKernel : public framework::OpKernel { if (platform::is_gpu_place(ctx.GetPlace())) { offset_cpu.mutable_data(offset->dims(), platform::CPUPlace()); - framework::TensorCopy(*offset, platform::CPUPlace(), ctx.device_context(), - &offset_cpu); + framework::TensorCopySync(*offset, platform::CPUPlace(), &offset_cpu); offset_data = offset_cpu.data(); length_cpu.mutable_data(length->dims(), platform::CPUPlace()); - framework::TensorCopy(*length, platform::CPUPlace(), ctx.device_context(), - &length_cpu); + framework::TensorCopySync(*length, platform::CPUPlace(), &length_cpu); length_data = length_cpu.data(); } @@ -127,13 +125,11 @@ class SequenceSliceGradOpKernel : public framework::OpKernel { if (platform::is_gpu_place(ctx.GetPlace())) { offset_cpu.mutable_data(offset->dims(), platform::CPUPlace()); - framework::TensorCopy(*offset, platform::CPUPlace(), ctx.device_context(), - &offset_cpu); + framework::TensorCopySync(*offset, platform::CPUPlace(), &offset_cpu); offset_data = offset_cpu.data(); length_cpu.mutable_data(length->dims(), platform::CPUPlace()); - framework::TensorCopy(*length, platform::CPUPlace(), ctx.device_context(), - &length_cpu); + framework::TensorCopySync(*length, platform::CPUPlace(), &length_cpu); length_data = length_cpu.data(); } diff --git a/paddle/fluid/operators/sequence_softmax_op.cc b/paddle/fluid/operators/sequence_softmax_op.cc index a0d47c12ba606eb62bbbea4d5ea793ce915e8100..c44f8206eb5079fef969e3e527552512eebd0f1a 100644 --- a/paddle/fluid/operators/sequence_softmax_op.cc +++ b/paddle/fluid/operators/sequence_softmax_op.cc @@ -57,8 +57,7 @@ class SequenceSoftmaxOp : public framework::OperatorWithKernel { class SequenceSoftmaxOpMaker : public framework::OpProtoAndCheckerMaker { public: - SequenceSoftmaxOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(LoDTensor) 1-D or 2-D input LoDTensor with the 2-nd dimension " "of length 1."); diff --git a/paddle/fluid/operators/sgd_op.cc b/paddle/fluid/operators/sgd_op.cc index 06cb0550ad7d4ad0241a4f439ea9ac16d9714c38..7a2bdeac09d61603f437ff10d58d0542bb3c3689 100644 --- a/paddle/fluid/operators/sgd_op.cc +++ b/paddle/fluid/operators/sgd_op.cc @@ -48,10 +48,27 @@ class SGDOp : public framework::OperatorWithKernel { } }; +class SGDOpInferVarType : public framework::VarTypeInference { + public: + void operator()(const framework::OpDesc& op_desc, + framework::BlockDesc* block) const override { + auto input_var = op_desc.Input("Param")[0]; + for (auto& out_var : op_desc.Output("ParamOut")) { + if (block->FindRecursiveOrCreateVar(input_var).GetType() == + framework::proto::VarType::SELECTED_ROWS) { + block->FindRecursiveOrCreateVar(out_var).SetType( + framework::proto::VarType::SELECTED_ROWS); + } else { + block->FindRecursiveOrCreateVar(out_var).SetType( + framework::proto::VarType::LOD_TENSOR); + } + } + } +}; + class SGDOpMaker : public framework::OpProtoAndCheckerMaker { public: - SGDOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Param", "(Tensor or SelectedRows) Input parameter"); AddInput("LearningRate", "(Tensor) Learning rate of SGD"); AddInput("Grad", "(Tensor or SelectedRows) Input gradient"); @@ -74,5 +91,6 @@ $$param\_out = param - learning\_rate * grad$$ } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP_WITHOUT_GRADIENT(sgd, ops::SGDOp, ops::SGDOpMaker); +REGISTER_OPERATOR(sgd, ops::SGDOp, ops::SGDOpMaker, + paddle::framework::EmptyGradOpMaker, ops::SGDOpInferVarType); REGISTER_OP_CPU_KERNEL(sgd, ops::SGDOpKernel, ops::SGDOpKernel); diff --git a/paddle/fluid/operators/sgd_op.cu b/paddle/fluid/operators/sgd_op.cu index 9d211541c0bf729393b8190edb18e101d5e07d1a..4722be7a666d3e8f3c25c9499f88ddda835f60e3 100644 --- a/paddle/fluid/operators/sgd_op.cu +++ b/paddle/fluid/operators/sgd_op.cu @@ -14,7 +14,7 @@ limitations under the License. */ #define EIGEN_USE_GPU #include "paddle/fluid/operators/sgd_op.h" -#include "paddle/fluid/platform/cuda_helper.h" +#include "paddle/fluid/platform/cuda_primitives.h" namespace paddle { namespace operators { diff --git a/paddle/fluid/operators/sgd_op.h b/paddle/fluid/operators/sgd_op.h index f3e88b0a0b05ef792b2cc8e880bdfddb6e6124d1..f9e0596191d0b86686e0fa36265806111c774b38 100644 --- a/paddle/fluid/operators/sgd_op.h +++ b/paddle/fluid/operators/sgd_op.h @@ -96,8 +96,12 @@ class SGDOpKernel : public framework::OpKernel { return; } - size_t param_row_width = param.value().numel() / param.rows().size(); - size_t grad_row_width = grad.value().numel() / grad.rows().size(); + auto param_row_width = param.value().dims()[1]; + auto grad_row_width = grad.value().dims()[1]; + VLOG(4) << " param rows: " << param.rows().size() + << " param memory rows: " << param.value().dims()[0] + << " grad rows: " << grad.rows().size() + << " grad memory rows: " << grad.value().dims()[0]; PADDLE_ENFORCE_EQ(param_row_width, grad_row_width, "param_row should have the same size with grad_row"); diff --git a/paddle/fluid/operators/shrink_rnn_memory_op.cc b/paddle/fluid/operators/shrink_rnn_memory_op.cc index a1871a8e7fb27d351f9d333966baa63c6f32ae01..8146c5f56104b7dec86b1c4491ed10fc2e94b58b 100644 --- a/paddle/fluid/operators/shrink_rnn_memory_op.cc +++ b/paddle/fluid/operators/shrink_rnn_memory_op.cc @@ -69,8 +69,7 @@ class ShrinkRNNMemoryOp : public ArrayOp { class ShrinkRNNMemoryOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - ShrinkRNNMemoryOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(LoDTensor) The RNN step memory to be shrinked."); AddInput("RankTable", "(LoDRankTable) The lod_rank_table of dynamic RNN."); AddInput("I", diff --git a/paddle/fluid/operators/sigmoid_cross_entropy_with_logits_op.cc b/paddle/fluid/operators/sigmoid_cross_entropy_with_logits_op.cc index 5db77d0493fc0abaa0a696cb559c3ca0534d4101..135e2a6f7f877c9ef159a4542b834d5627649e81 100644 --- a/paddle/fluid/operators/sigmoid_cross_entropy_with_logits_op.cc +++ b/paddle/fluid/operators/sigmoid_cross_entropy_with_logits_op.cc @@ -86,9 +86,7 @@ class SigmoidCrossEntropyWithLogitsGradOp class SigmoidCrossEntropyWithLogitsOpMaker : public framework::OpProtoAndCheckerMaker { public: - SigmoidCrossEntropyWithLogitsOpMaker(OpProto* proto, - OpAttrChecker* op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor, default Tensor), a 2-D tensor with shape N x D, " "where N is the batch size and D is the number of classes. " diff --git a/paddle/fluid/operators/sign_op.cc b/paddle/fluid/operators/sign_op.cc index 8f8b7abd03212c12ca351e551621e63b4c7148c2..f3985dcc027f974e0213a73ea9a21e268d77615f 100644 --- a/paddle/fluid/operators/sign_op.cc +++ b/paddle/fluid/operators/sign_op.cc @@ -34,8 +34,7 @@ class SignOp : public framework::OperatorWithKernel { template class SignOpMaker : public framework::OpProtoAndCheckerMaker { public: - SignOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor) Input tensor of sign operator."); AddOutput("Out", "(Tensor) Output tensor of sign operator."); AddComment(R"DOC( diff --git a/paddle/fluid/operators/smooth_l1_loss_op.cc b/paddle/fluid/operators/smooth_l1_loss_op.cc index 322581fdef27b12a06704abc9c3b8772adf002f2..622420c1c33a62994c81ad9534c4fa37a4a1fa1a 100644 --- a/paddle/fluid/operators/smooth_l1_loss_op.cc +++ b/paddle/fluid/operators/smooth_l1_loss_op.cc @@ -46,8 +46,7 @@ class SmoothL1LossOp : public framework::OperatorWithKernel { class SmoothL1LossOpMaker : public framework::OpProtoAndCheckerMaker { public: - SmoothL1LossOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor, default Tensor) A tensor with rank at least 2. " "The input value of smooth l1 loss op with shape " @@ -106,7 +105,7 @@ class SmoothL1LossGradOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; void InferShape(framework::InferShapeContext* ctx) const override { - auto in_dims = ctx->GetInputDim("X"); + auto in_dims = ctx->GetInputDim("Diff"); auto out_dims = ctx->GetInputDim(framework::GradVarName("Out")); PADDLE_ENFORCE_GE(out_dims.size(), 2, @@ -128,12 +127,33 @@ class SmoothL1LossGradOp : public framework::OperatorWithKernel { } }; +class SmoothL1LossGradMaker : public framework::SingleGradOpDescMaker { + public: + using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; + + protected: + std::unique_ptr Apply() const override { + auto* op = new framework::OpDesc(); + op->SetType("smooth_l1_loss_grad"); + op->SetInput("InsideWeight", Input("InsideWeight")); + op->SetInput("OutsideWeight", Input("OutsideWeight")); + op->SetInput("Diff", Output("Diff")); + op->SetInput(framework::GradVarName("Out"), OutputGrad("Out")); + + op->SetAttrMap(Attrs()); + + op->SetOutput(framework::GradVarName("X"), InputGrad("X")); + op->SetOutput(framework::GradVarName("Y"), InputGrad("Y")); + return std::unique_ptr(op); + } +}; + } // namespace operators } // namespace paddle namespace ops = paddle::operators; REGISTER_OPERATOR(smooth_l1_loss, ops::SmoothL1LossOp, ops::SmoothL1LossOpMaker, - paddle::framework::DefaultGradOpDescMaker); + ops::SmoothL1LossGradMaker); REGISTER_OPERATOR(smooth_l1_loss_grad, ops::SmoothL1LossGradOp); REGISTER_OP_CPU_KERNEL( smooth_l1_loss, diff --git a/paddle/fluid/operators/softmax_mkldnn_op.cc b/paddle/fluid/operators/softmax_mkldnn_op.cc index 71b541d98f6e0d3e12601c9988ca6ffb8bb7554d..14b57b11fefb2b726531cb164dbf479f8df26b24 100644 --- a/paddle/fluid/operators/softmax_mkldnn_op.cc +++ b/paddle/fluid/operators/softmax_mkldnn_op.cc @@ -53,25 +53,60 @@ class SoftmaxMKLDNNKernel : public paddle::framework::OpKernel { "Softmax input and output dimensions should match"); // Same memory descriptor to be used for input and output memory::dims softmax_tz = {src_tz[0], src_tz[1]}; - // Currently only supports NC data format - // TODO(jczaja-intel): support more formats - auto softmax_md = - MKLDNNMemDesc({softmax_tz}, memory::f32, memory::format::nc); - // Normalization is made after innermost dimension eg. C out of NC - auto softmax_desc = softmax_forward::desc(prop_kind::forward_scoring, - softmax_md, 1 /*dim: C*/); - // create memory primitives - auto softmax_src_memory = - memory({softmax_md, mkldnn_engine}, - static_cast(const_cast(input_data))); - auto softmax_dst_memory = - memory({softmax_md, mkldnn_engine}, - static_cast(const_cast(output_data))); - auto softmax_prim_desc = - softmax_forward::primitive_desc(softmax_desc, mkldnn_engine); - auto softmax = softmax_forward(softmax_prim_desc, softmax_src_memory, - softmax_dst_memory); - std::vector pipeline{softmax}; + // Generate keys for storing/retriving primitives for this operator + // TODO(jczaja): Each MKLDNN operator may have diffrent hashing function + auto gethash = [](memory::dims& operand_dims) { + return std::string(std::to_string(operand_dims[0]) + "-" + + std::to_string(operand_dims[1])); + }; + const std::string key = gethash(softmax_tz); + const std::string key_softmax_p = key + "@softmax_p"; + const std::string key_softmax_src_mem_p = key + "@softmax_src_mem_p"; + const std::string key_softmax_dst_mem_p = key + "@softmax_dst_mem_p"; + + std::shared_ptr softmax_p = dev_ctx.GetBlob(key_softmax_p); + if (softmax_p == nullptr) { + // Currently only NC data format is supported + auto softmax_md = + MKLDNNMemDesc({softmax_tz}, memory::f32, memory::format::nc); + // Normalization is made after innermost dimension eg. C out of NC + auto softmax_desc = softmax_forward::desc(prop_kind::forward_scoring, + softmax_md, 1 /*dim: C*/); + // create memory primitives + auto softmax_src_memory_p = std::make_shared( + memory::primitive_desc{softmax_md, mkldnn_engine}, + static_cast(const_cast(input_data))); + dev_ctx.SetBlob(key_softmax_src_mem_p, softmax_src_memory_p); + auto softmax_dst_memory_p = std::make_shared( + memory::primitive_desc{softmax_md, mkldnn_engine}, + static_cast(output_data)); + dev_ctx.SetBlob(key_softmax_dst_mem_p, softmax_dst_memory_p); + + auto softmax_forward_pd = + std::make_shared(softmax_desc, + mkldnn_engine); + softmax_p = std::make_shared( + *(softmax_forward_pd.get()), + *(static_cast(softmax_src_memory_p.get())), + *(static_cast(softmax_dst_memory_p.get()))); + dev_ctx.SetBlob(key_softmax_p, softmax_p); + } else { + // Primitives already exist + auto src_memory_p = std::static_pointer_cast( + dev_ctx.GetBlob(key_softmax_src_mem_p)); + PADDLE_ENFORCE(src_memory_p != nullptr, + "Fail to find softmax src mem_p in device context"); + auto dst_memory_p = std::static_pointer_cast( + dev_ctx.GetBlob(key_softmax_dst_mem_p)); + PADDLE_ENFORCE(dst_memory_p != nullptr, + "Fail to find softmax dst mem_p in device context"); + src_memory_p->set_data_handle( + reinterpret_cast(const_cast(input_data))); + dst_memory_p->set_data_handle(output_data); + } + + std::vector pipeline{ + *(static_cast(softmax_p.get()))}; stream(stream::kind::eager).submit(pipeline).wait(); const bool is_test = ctx.Attr("is_test"); diff --git a/paddle/fluid/operators/softmax_op.cc b/paddle/fluid/operators/softmax_op.cc index 2741ba95bcfc1db3d74e0fb8c3f6fddf7d5a2caa..cc256aa627bdda0609f496cab93a2dec7d95f348 100644 --- a/paddle/fluid/operators/softmax_op.cc +++ b/paddle/fluid/operators/softmax_op.cc @@ -77,8 +77,7 @@ class SoftmaxOp : public framework::OperatorWithKernel { class SoftmaxOpMaker : public framework::OpProtoAndCheckerMaker { public: - SoftmaxOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The input tensor of softmax. " "2-D with shape [batch_size, input_feature_dimensions]."); @@ -164,7 +163,9 @@ REGISTER_OPERATOR(softmax, ops::SoftmaxOp, ops::SoftmaxOpMaker, paddle::framework::DefaultGradOpDescMaker); REGISTER_OPERATOR(softmax_grad, ops::SoftmaxOpGrad); REGISTER_OP_CPU_KERNEL( - softmax, ops::SoftmaxKernel); + softmax, ops::SoftmaxKernel, + ops::SoftmaxKernel); REGISTER_OP_CPU_KERNEL( softmax_grad, - ops::SoftmaxGradKernel); + ops::SoftmaxGradKernel, + ops::SoftmaxGradKernel); diff --git a/paddle/fluid/operators/softmax_op.cu.cc b/paddle/fluid/operators/softmax_op.cu.cc index 0c1f7cef7ab7b66358d80f6f0670e0d07536128c..5fb4f011d9b47cebc4a23bcce47eada825263343 100644 --- a/paddle/fluid/operators/softmax_op.cu.cc +++ b/paddle/fluid/operators/softmax_op.cu.cc @@ -19,6 +19,8 @@ namespace ops = paddle::operators; namespace plat = paddle::platform; REGISTER_OP_CUDA_KERNEL( softmax, ops::SoftmaxKernel, + ops::SoftmaxKernel, ops::SoftmaxKernel); -REGISTER_OP_CUDA_KERNEL(softmax_grad, - ops::SoftmaxGradKernel); +REGISTER_OP_CUDA_KERNEL( + softmax_grad, ops::SoftmaxGradKernel, + ops::SoftmaxGradKernel); diff --git a/paddle/fluid/operators/softmax_with_cross_entropy_op.cc b/paddle/fluid/operators/softmax_with_cross_entropy_op.cc index 857e5733573497b56520daa7860f4feb4e01cda7..53cb716a979229c99fcbdc12f1aeab4e21b320f3 100644 --- a/paddle/fluid/operators/softmax_with_cross_entropy_op.cc +++ b/paddle/fluid/operators/softmax_with_cross_entropy_op.cc @@ -20,8 +20,7 @@ namespace operators { class SoftmaxWithCrossEntropyOpMaker : public framework::OpProtoAndCheckerMaker { public: - SoftmaxWithCrossEntropyOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Logits", "(Tensor, default: Tensor), The unscaled log probabilities " "which is a 2-D tensor with shape [N x K]. N is the batch_size, " diff --git a/paddle/fluid/operators/split_byref_op.cc b/paddle/fluid/operators/split_byref_op.cc index 7413ce3e9ce60ed733bb4d27e9ec205e5f0a7e1b..bc998e1abbd7131a7497288cc9d66315a6fedc85 100644 --- a/paddle/fluid/operators/split_byref_op.cc +++ b/paddle/fluid/operators/split_byref_op.cc @@ -64,8 +64,7 @@ class SplitByrefOp : public framework::OperatorWithKernel { class SplitByrefOpMaker : public framework::OpProtoAndCheckerMaker { public: - SplitByrefOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor) Input tensor of the split operator."); AddOutput("Out", "(Tensor) Output tensors of the split operator.") .AsDuplicable(); diff --git a/paddle/fluid/operators/split_ids_op.cc b/paddle/fluid/operators/split_ids_op.cc index a53cbc8ac5199061dafdc7f4cf560b9e4fc577ab..c867c46873ae7ddbdbda280351e4ab28235bcc08 100644 --- a/paddle/fluid/operators/split_ids_op.cc +++ b/paddle/fluid/operators/split_ids_op.cc @@ -19,8 +19,7 @@ namespace operators { class SplitIdsOpMaker : public framework::OpProtoAndCheckerMaker { public: - SplitIdsOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Ids", "(LoDTensor) the input ids with shape{batch_num, 1}"); AddOutput("Out", "(LoDTensor) The outputs of the input Ids.") .AsDuplicable(); diff --git a/paddle/fluid/operators/split_lod_tensor_op.cc b/paddle/fluid/operators/split_lod_tensor_op.cc index 3222cce239988b170501f2b99e9f1253036b7fbc..767449cde981e5925b7144ff1038560c67651f3e 100644 --- a/paddle/fluid/operators/split_lod_tensor_op.cc +++ b/paddle/fluid/operators/split_lod_tensor_op.cc @@ -125,8 +125,7 @@ class SplitLoDTensorOp : public framework::OperatorBase { class SplitLoDTensorOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - SplitLoDTensorOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The input LoDTensor"); AddInput("Mask", "A bool column vector which mask the input"); AddOutput("OutTrue", "True branch of input LoDTensor"); diff --git a/paddle/fluid/operators/split_op.cc b/paddle/fluid/operators/split_op.cc index a4398df36bcc2d3b8bbe8949f27f5d6508861d95..5e2b2a994534c2fb1e053c067b36651d358b9da8 100644 --- a/paddle/fluid/operators/split_op.cc +++ b/paddle/fluid/operators/split_op.cc @@ -70,8 +70,7 @@ class SplitOp : public framework::OperatorWithKernel { class SplitOpMaker : public framework::OpProtoAndCheckerMaker { public: - SplitOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor) Input tensor of the split operator."); AddOutput("Out", "(Tensor) Output tensors of the split operator.") .AsDuplicable(); diff --git a/paddle/fluid/operators/split_selected_rows_op.cc b/paddle/fluid/operators/split_selected_rows_op.cc index e1ce3d0c1bf11e9a623e4e9adc8f08f5069f4d94..76615a9405d7a8e3fa9dba8d01a956209e02ae8f 100644 --- a/paddle/fluid/operators/split_selected_rows_op.cc +++ b/paddle/fluid/operators/split_selected_rows_op.cc @@ -19,8 +19,7 @@ namespace operators { class SplitSelectedRowsOpMaker : public framework::OpProtoAndCheckerMaker { public: - SplitSelectedRowsOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The input SelectedRows."); AddOutput("Out", "The outputs of the input SelectedRows.").AsDuplicable(); AddAttr>("height_sections", diff --git a/paddle/fluid/operators/spp_op.cc b/paddle/fluid/operators/spp_op.cc index 1cada95501a76da27081d533b451ce7f6a384a49..a2a96b72f09df86790ad1f90ead9189ff9bd581c 100644 --- a/paddle/fluid/operators/spp_op.cc +++ b/paddle/fluid/operators/spp_op.cc @@ -20,8 +20,7 @@ namespace operators { class SppOpMaker : public framework::OpProtoAndCheckerMaker { public: - SppOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput( "X", "(Tensor) The input tensor of spp operator. " diff --git a/paddle/fluid/operators/squared_l2_distance_op.cc b/paddle/fluid/operators/squared_l2_distance_op.cc index c32f575b541d6a6441cc1b6e999496eacef421a5..42532a294b2ef9ffdb240fac8596278047daf7fe 100644 --- a/paddle/fluid/operators/squared_l2_distance_op.cc +++ b/paddle/fluid/operators/squared_l2_distance_op.cc @@ -56,8 +56,7 @@ class SquaredL2DistanceOp : public framework::OperatorWithKernel { class SquaredL2DistanceOpMaker : public framework::OpProtoAndCheckerMaker { public: - SquaredL2DistanceOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor) Input of SquaredL2DistanceOp."); AddInput("Y", "(Tensor) Target of SquaredL2DistanceOp."); AddOutput("sub_result", diff --git a/paddle/fluid/operators/squared_l2_norm_op.cc b/paddle/fluid/operators/squared_l2_norm_op.cc index 4ce51259da3530367d91b5da34f06fbe5d969fce..7bd82e0ce4add6d4434e1defaee43da178a6f309 100644 --- a/paddle/fluid/operators/squared_l2_norm_op.cc +++ b/paddle/fluid/operators/squared_l2_norm_op.cc @@ -48,8 +48,7 @@ class SquaredL2NormGradOp : public framework::OperatorWithKernel { class SquaredL2NormOpMaker : public framework::OpProtoAndCheckerMaker { public: - SquaredL2NormOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor) The input of squared_l2_norm op."); AddOutput("Out", "(Scalar) The output of squared_l2_norm op."); AddComment(R"DOC( diff --git a/paddle/fluid/operators/sum_op.cc b/paddle/fluid/operators/sum_op.cc index 108f26fafe7af76eaa613d77ed77748ee43ea234..bcc5e22d4a77349e7cde9a43b83f23d4c867d994 100644 --- a/paddle/fluid/operators/sum_op.cc +++ b/paddle/fluid/operators/sum_op.cc @@ -112,8 +112,7 @@ class SumOp : public framework::OperatorWithKernel { class SumOpMaker : public framework::OpProtoAndCheckerMaker { public: - SumOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(vector) The input tensors of sum operator.") .AsDuplicable(); AddOutput("Out", "(Tensor) The output tensor of sum operator."); diff --git a/paddle/fluid/operators/tensor_array_read_write_op.cc b/paddle/fluid/operators/tensor_array_read_write_op.cc index 2636812c42985536e7ca3475c03bbd8d1638ece6..c703d11eeccf8418250f00c801f47418ee9c85ae 100644 --- a/paddle/fluid/operators/tensor_array_read_write_op.cc +++ b/paddle/fluid/operators/tensor_array_read_write_op.cc @@ -57,8 +57,7 @@ class WriteToArrayOp : public ArrayOp { class WriteToArrayOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - WriteToArrayOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(LoDTensor) the tensor will be written to tensor array"); AddInput( "I", @@ -148,8 +147,7 @@ class ReadFromArrayOp : public ArrayOp { class ReadFromArrayProtoMaker : public framework::OpProtoAndCheckerMaker { public: - ReadFromArrayProtoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(TensorArray) the array will be read from."); AddInput("I", "(Tensor) the subscript index in tensor array. The number of " diff --git a/paddle/fluid/operators/tensorrt_engine_op.cc b/paddle/fluid/operators/tensorrt_engine_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..83e768b4dc9c607b0f73d7183462d772ae7ab994 --- /dev/null +++ b/paddle/fluid/operators/tensorrt_engine_op.cc @@ -0,0 +1,70 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#ifdef PADDLE_WITH_CUDA + +#include "paddle/fluid/operators/tensorrt_engine_op.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/inference/tensorrt/convert/op_converter.h" +#include "paddle/fluid/inference/utils/singleton.h" + +namespace paddle { +namespace operators { + +template +void paddle::operators::TensorRTEngineKernel::Prepare( + const framework::ExecutionContext &context) const { + // Get the ProgramDesc and pass to convert. + const auto &block = context.Attr("subgraph"); + max_batch_ = context.Attr("max_batch"); + auto max_workspace = context.Attr("max_workspace"); + engine_.reset(new inference::tensorrt::TensorRTEngine( + max_batch_, max_workspace, nullptr)); + inference::Singleton::Global().ConvertBlock( + block, engine_.get()); + engine_->FreezeNetwork(); +} + +class TensorRTEngineOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("Xs", "A list of inputs.").AsDuplicable(); + AddOutput("Ys", "A list of outputs").AsDuplicable(); + AddAttr("subgraph", "the subgraph"); + AddComment("TensorRT engine operator."); + } +}; + +class TensorRTEngineInferVarType : public framework::VarTypeInference { + public: + void operator()(const framework::OpDesc &op_desc, + framework::BlockDesc *block) const override {} +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; + +REGISTER_OPERATOR(tensorrt_engine, ops::TensorRTEngineOp, + ops::TensorRTEngineOpMaker, ops::TensorRTEngineOpMaker); + +REGISTER_OP_CPU_KERNEL( + tensorrt_engine, + ops::TensorRTEngineKernel, + ops::TensorRTEngineKernel, + ops::TensorRTEngineKernel, + ops::TensorRTEngineKernel); + +#endif // PADDLE_WITH_CUDA diff --git a/paddle/fluid/operators/tensorrt_engine_op.h b/paddle/fluid/operators/tensorrt_engine_op.h new file mode 100644 index 0000000000000000000000000000000000000000..fe273d386c529be3df05a955f492e2c39d4d8812 --- /dev/null +++ b/paddle/fluid/operators/tensorrt_engine_op.h @@ -0,0 +1,110 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once + +#ifdef PADDLE_WITH_CUDA + +#include "paddle/fluid/framework/operator.h" +#include "paddle/fluid/inference/analysis/helper.h" +#include "paddle/fluid/inference/tensorrt/engine.h" + +namespace paddle { +namespace operators { + +class TensorRTEngineOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContext* ctx) const override {} + + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + framework::OpKernelType kt = framework::OpKernelType( + framework::ToDataType( + ctx.Input("pre_ids")->type()), + platform::CPUPlace()); + return kt; + } +}; + +template +class TensorRTEngineKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + if (!engine_) { + Prepare(context); + } + auto input_names = context.op().Inputs("Xs"); + PADDLE_ENFORCE(!input_names.empty(), "should pass more than one inputs"); + // Try to determine a batch_size + auto* tensor0 = context.Input(input_names.front()); + PADDLE_ENFORCE_NOT_NULL(tensor0); + int batch_size = tensor0->dims()[0]; + PADDLE_ENFORCE_LE(batch_size, max_batch_); + + // Convert input tensor from fluid to engine. + for (const auto& x : context.Inputs("Xs")) { + // convert input and copy to TRT engine's buffer + auto* v = context.scope().FindVar(x); + PADDLE_ENFORCE_NOT_NULL(v, "no variable called %s", x); + auto& t = v->Get(); + if (platform::is_cpu_place(t.place())) { + engine_->SetInputFromCPU(x, static_cast(t.data()), + t.memory_size()); + } else { + engine_->SetInputFromGPU(x, static_cast(t.data()), + t.memory_size()); + } + } + // Execute the engine. + PADDLE_ENFORCE_GT(batch_size, 0); + engine_->Execute(batch_size); + // Convert output tensor from engine to fluid + for (const auto& y : context.Outputs("Ys")) { + // convert output and copy to fluid. + nvinfer1::ITensor* trt_t = engine_->GetITensor(y); + auto dims = trt_t->getDimensions(); + // Use the output ITensor's dims to reshape the Fluid Tensor. + std::vector ddim(dims.d, dims.d + dims.nbDims); + + auto* fluid_v = context.scope().FindVar(y); + PADDLE_ENFORCE_NOT_NULL(fluid_v, "no output variable called %s", y); + auto* fluid_t = fluid_v->GetMutable(); + fluid_t->Resize(framework::make_ddim(ddim)); + auto size = inference::analysis::AccuDims(dims.d, dims.nbDims); + if (platform::is_cpu_place(fluid_t->place())) { + engine_->GetOutputInCPU( + y, fluid_t->mutable_data(platform::CPUPlace()), size); + } else { + engine_->GetOutputInGPU( + y, fluid_t->mutable_data(platform::CUDAPlace()), size); + } + } + } + + protected: + // Build the engine. + void Prepare(const framework::ExecutionContext& context) const; + + private: + mutable std::unique_ptr engine_; + mutable int max_batch_{0}; +}; + +} // namespace operators +} // namespace paddle + +#endif // PADDLE_WITH_CUDA diff --git a/paddle/fluid/operators/test_send_nccl_id.cc b/paddle/fluid/operators/test_send_nccl_id.cc new file mode 100644 index 0000000000000000000000000000000000000000..719f039a0f5fcd7445bf1589a683f122e6d62ba0 --- /dev/null +++ b/paddle/fluid/operators/test_send_nccl_id.cc @@ -0,0 +1,94 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include +#include // NOLINT + +#include "gtest/gtest.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/framework/operator.h" +#include "paddle/fluid/framework/program_desc.h" +#include "paddle/fluid/operators/detail/grpc_client.h" +#include "paddle/fluid/operators/listen_and_serv_op.h" +#include "paddle/fluid/operators/math/math_function.h" +#include "paddle/fluid/operators/math/selected_rows_functor.h" +#include "paddle/fluid/platform/nccl_helper.h" +#include "paddle/fluid/string/printf.h" + +USE_NO_KERNEL_OP(listen_and_serv); + +namespace f = paddle::framework; +namespace p = paddle::platform; +namespace m = paddle::operators::math; +namespace detail = paddle::operators::detail; +namespace string = paddle::string; + +std::unique_ptr rpc_service; + +void StartServer(std::atomic* initialized) { + f::Scope scope; + p::CPUPlace place; + scope.Var(NCCL_ID_VARNAME); + p::DeviceContextPool& pool = p::DeviceContextPool::Instance(); + auto& dev_ctx = *pool.Get(p::CPUPlace()); + + rpc_service.reset(new detail::AsyncGRPCServer("127.0.0.1:0", true)); + + f::ProgramDesc empty_program; + f::Executor executor(dev_ctx.GetPlace()); + rpc_service->SetScope(&scope); + rpc_service->SetDevCtx(&dev_ctx); + rpc_service->SetProgram(&empty_program); + rpc_service->SetExecutor(&executor); + + std::thread server_thread( + std::bind(&detail::AsyncGRPCServer::RunSyncUpdate, rpc_service.get())); + *initialized = true; + rpc_service->SetCond(0); + auto recv = rpc_service->Get(); + LOG(INFO) << "got nccl id and stop server..."; + rpc_service->ShutDown(); + server_thread.join(); +} + +TEST(SendNcclId, DISABLED_Normal) { + std::atomic initialized{false}; + std::thread server_thread(StartServer, &initialized); + while (!initialized) { + } + // wait server to start + // sleep(2); + rpc_service->WaitServerReady(); + + f::Scope scope; + p::CPUPlace place; + p::DeviceContextPool& pool = p::DeviceContextPool::Instance(); + auto& dev_ctx = *pool.Get(p::CPUPlace()); + + auto var = scope.Var(NCCL_ID_VARNAME); + // var->SetType(f::proto::VarType_Type_RAW); + auto id = var->GetMutable(); + p::dynload::ncclGetUniqueId(id); + + int port = rpc_service->GetSelectedPort(); + std::string ep = string::Sprintf("127.0.0.1:%d", port); + detail::RPCClient client; + + client.AsyncSendVariable(ep, dev_ctx, scope, NCCL_ID_VARNAME); + client.Wait(); + server_thread.join(); + auto* ptr = rpc_service.release(); + delete ptr; +} diff --git a/paddle/fluid/operators/top_k_op.cc b/paddle/fluid/operators/top_k_op.cc index 2e4e8caed5327f4ca9038c376de2ec831354917e..c17d1afc309c65035063348d4934ea1783b018ed 100644 --- a/paddle/fluid/operators/top_k_op.cc +++ b/paddle/fluid/operators/top_k_op.cc @@ -48,8 +48,7 @@ class TopkOp : public framework::OperatorWithKernel { class TopkOpMaker : public framework::OpProtoAndCheckerMaker { public: - TopkOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor) The input of Topk op"); AddOutput("Out", "(Tensor) The output tensor of Topk op"); AddOutput("Indices", "(Tensor) The indices of Topk elements of input"); @@ -75,4 +74,5 @@ namespace ops = paddle::operators; REGISTER_OPERATOR(top_k, ops::TopkOp, ops::TopkOpMaker, paddle::framework::EmptyGradOpMaker); REGISTER_OP_CPU_KERNEL(top_k, - ops::TopkKernel); + ops::TopkKernel, + ops::TopkKernel); diff --git a/paddle/fluid/operators/top_k_op.cu b/paddle/fluid/operators/top_k_op.cu index d7f4d383ce0d9e1ff42fc12c96aaf0ceb532e5db..9da8551eb2d7ea66ad434c42b54522432095ce29 100644 --- a/paddle/fluid/operators/top_k_op.cu +++ b/paddle/fluid/operators/top_k_op.cu @@ -15,6 +15,7 @@ limitations under the License. */ #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/operators/top_k_op.h" #include "paddle/fluid/platform/assert.h" +#include "paddle/fluid/platform/cuda_device_function.h" namespace paddle { namespace operators { @@ -235,8 +236,14 @@ __device__ __forceinline__ void BlockReduce(Pair* sh_topk, int* maxid, sh_topk[tid] = topk[*beam]; } } + // NOTE(zcd): temporary solution + unsigned mask = 0u; + CREATE_SHFL_MASK(mask, true); + if (maxid[0] / 32 == warp) { - if (__shfl(*beam, (maxid[0]) % 32, 32) == MaxLength) break; + if (platform::CudaShuffleSync(mask, *beam, (maxid[0]) % 32, 32) == + MaxLength) + break; } } } @@ -318,4 +325,5 @@ class TopkOpCUDAKernel : public framework::OpKernel { } // namespace operators } // namespace paddle -REGISTER_OP_CUDA_KERNEL(top_k, paddle::operators::TopkOpCUDAKernel); +REGISTER_OP_CUDA_KERNEL(top_k, paddle::operators::TopkOpCUDAKernel, + paddle::operators::TopkOpCUDAKernel); diff --git a/paddle/fluid/operators/top_k_op.h b/paddle/fluid/operators/top_k_op.h index d44eeae8e6ff9ac87ab093d04e3f5427743f0c08..7ddb82ef6ff063868a4b9b603b8ab89700b9dd13 100644 --- a/paddle/fluid/operators/top_k_op.h +++ b/paddle/fluid/operators/top_k_op.h @@ -55,6 +55,9 @@ class TopkKernel : public framework::OpKernel { // NOTE: eigen shape doesn't affect paddle tensor. eg_input.reshape(flat2dims); +#ifdef PADDLE_WITH_MKLML +#pragma omp parallel for +#endif for (size_t i = 0; i < row; i++) { std::vector> vec; for (size_t j = 0; j < col; j++) { diff --git a/paddle/fluid/operators/transpose_op.cc b/paddle/fluid/operators/transpose_op.cc index 3555cb68cab97c0cf983f1173c3b4ca9307e4f7d..60556a564c25c08612447ebd47a4b432b8a12d29 100644 --- a/paddle/fluid/operators/transpose_op.cc +++ b/paddle/fluid/operators/transpose_op.cc @@ -56,8 +56,7 @@ class TransposeOp : public framework::OperatorWithKernel { class TransposeOpMaker : public framework::OpProtoAndCheckerMaker { public: - TransposeOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput( "X", "(Tensor) The input tensor, tensors with rank up to 6 are supported."); diff --git a/paddle/fluid/operators/uniform_random_batch_size_like_op.cc b/paddle/fluid/operators/uniform_random_batch_size_like_op.cc index 00f00bb403db5e40939a1502b2219fb4d36d58e5..78fee77df8151221459b0afa0d6789bfe82cfda5 100644 --- a/paddle/fluid/operators/uniform_random_batch_size_like_op.cc +++ b/paddle/fluid/operators/uniform_random_batch_size_like_op.cc @@ -32,9 +32,8 @@ class UniformRandomBatchSizeLikeOp : public BatchSizeLikeOp { }; class UniformRandomBatchSizeLikeOpMaker : public BatchSizeLikeOpMaker { - public: - UniformRandomBatchSizeLikeOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : BatchSizeLikeOpMaker(proto, op_checker) { + protected: + void Apply() override { AddComment(R"DOC( Uniform random operator diff --git a/paddle/fluid/operators/uniform_random_op.cc b/paddle/fluid/operators/uniform_random_op.cc index acaefaacdaa593c090d81084fdc1b3665314833f..137ea91caedabc3167146d91b063dbe9e2e2b931 100644 --- a/paddle/fluid/operators/uniform_random_op.cc +++ b/paddle/fluid/operators/uniform_random_op.cc @@ -85,8 +85,7 @@ class UniformRandomOp : public framework::OperatorWithKernel { class UniformRandomOpMaker : public framework::OpProtoAndCheckerMaker { public: - UniformRandomOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddOutput("Out", "(Tensor) The output tensor of uniform random op"); AddComment(R"DOC( Uniform random operator. @@ -116,11 +115,31 @@ uniform distribution. .SetDefault(framework::proto::VarType::FP32); } }; + +class UniformRandomOpVarTypeInference : public framework::VarTypeInference { + public: + void operator()(const framework::OpDesc& op_desc, + framework::BlockDesc* block) const override { + auto out_var_name = op_desc.Output("Out").front(); + if (block->FindRecursiveOrCreateVar(out_var_name).GetType() == + framework::proto::VarType::SELECTED_ROWS) { + block->FindRecursiveOrCreateVar(out_var_name) + .SetType(framework::proto::VarType::SELECTED_ROWS); + } else { + block->FindRecursiveOrCreateVar(out_var_name) + .SetType(framework::proto::VarType::LOD_TENSOR); + } + } +}; + } // namespace operators } // namespace paddle -REGISTER_OP_WITHOUT_GRADIENT(uniform_random, paddle::operators::UniformRandomOp, - paddle::operators::UniformRandomOpMaker); +REGISTER_OPERATOR(uniform_random, paddle::operators::UniformRandomOp, + paddle::operators::UniformRandomOpMaker, + paddle::framework::EmptyGradOpMaker, + paddle::operators::UniformRandomOpVarTypeInference); + REGISTER_OP_CPU_KERNEL(uniform_random, paddle::operators::CPUUniformRandomKernel, paddle::operators::CPUUniformRandomKernel); diff --git a/paddle/fluid/operators/unpool_op.cc b/paddle/fluid/operators/unpool_op.cc index b3cd87efa21115565b32659cb35fee4b5bed2d4f..1d441b43b14ea194152095874645f8133c423efd 100644 --- a/paddle/fluid/operators/unpool_op.cc +++ b/paddle/fluid/operators/unpool_op.cc @@ -20,8 +20,7 @@ namespace operators { class Unpool2dOpMaker : public framework::OpProtoAndCheckerMaker { public: - Unpool2dOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput( "X", "(Tensor) The input tensor of unpool operator. " diff --git a/paddle/fluid/operators/warpctc_op.cc b/paddle/fluid/operators/warpctc_op.cc index 6835a5dd6286ece20c4ce6f3e951ed4b0057012c..e06c8c962f45a4e91b7efed7431571f0fc6870a3 100644 --- a/paddle/fluid/operators/warpctc_op.cc +++ b/paddle/fluid/operators/warpctc_op.cc @@ -53,8 +53,7 @@ class WarpCTCOp : public framework::OperatorWithKernel { class WarpCTCOpMaker : public framework::OpProtoAndCheckerMaker { public: - WarpCTCOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Logits", "(LodTensor, default: LoDTensor), the unscaled " "probabilities of variable-length sequences, which is a 2-D " diff --git a/paddle/fluid/operators/warpctc_op.h b/paddle/fluid/operators/warpctc_op.h index afbfe69973830bde93ec0af8d1c844580a786663..ab70c1f0592d122ba248a101db487e64c0bdae6f 100644 --- a/paddle/fluid/operators/warpctc_op.h +++ b/paddle/fluid/operators/warpctc_op.h @@ -162,7 +162,7 @@ class WarpCTCKernel : public framework::OpKernel { static_cast(sequence_width)}); warpctc_logits.mutable_data(warpctc_logits_dims, ctx.GetPlace()); math::PaddingLoDTensorFunctor()( - ctx.template device_context(), *logits, warpctc_logits, + ctx.template device_context(), *logits, &warpctc_logits, false); const T* warpctc_logits_data = warpctc_logits.data(); @@ -186,8 +186,7 @@ class WarpCTCKernel : public framework::OpKernel { // warpctc accesses labels in CPU memory Tensor warpctc_label; - TensorCopy(*label, platform::CPUPlace(), ctx.device_context(), - &warpctc_label); + TensorCopySync(*label, platform::CPUPlace(), &warpctc_label); const int* warpctc_label_data = warpctc_label.data(); // warpctc stores loss in CPU memory Tensor warpctc_loss; @@ -217,13 +216,13 @@ class WarpCTCGradKernel : public framework::OpKernel { logits_grad->mutable_data(ctx.GetPlace()); bool norm_by_times = ctx.Attr("norm_by_times"); math::UnpaddingLoDTensorFunctor()( - ctx.template device_context(), *logits_grad, + ctx.template device_context(), logits_grad, *warpctc_grad, norm_by_times); const T* loss_grad_data = loss_grad->data(); math::ScaleLoDTensorFunctor()( - ctx.template device_context(), *logits_grad, - loss_grad_data); + ctx.template device_context(), loss_grad_data, + logits_grad); } }; diff --git a/paddle/fluid/operators/while_op.cc b/paddle/fluid/operators/while_op.cc index 710cc9fc2e716da2e4fd067562a34d312e48b1a1..175c3ac5d79f24e47d21417df8e3eaeb4d5b2335 100644 --- a/paddle/fluid/operators/while_op.cc +++ b/paddle/fluid/operators/while_op.cc @@ -68,8 +68,7 @@ class WhileOp : public framework::OperatorBase { class WhileOpMaker : public framework::OpProtoAndCheckerMaker { public: - WhileOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput(kX, "A set of variables, which are required by operators inside the " "block of While Op.") diff --git a/paddle/fluid/platform/CMakeLists.txt b/paddle/fluid/platform/CMakeLists.txt index 598fd4d419078a973647f2f8f20e8a12c8115a8b..b29035bafd34fa81dc6b59691142fe74439202b8 100644 --- a/paddle/fluid/platform/CMakeLists.txt +++ b/paddle/fluid/platform/CMakeLists.txt @@ -1,4 +1,4 @@ -proto_library(profiler_proto SRCS profiler.proto) +proto_library(profiler_proto SRCS profiler.proto DEPS framework_proto) py_proto_compile(profiler_py_proto SRCS profiler.proto) add_custom_target(profiler_py_proto_init ALL COMMAND ${CMAKE_COMMAND} -E touch __init__.py) @@ -49,7 +49,7 @@ nv_test(device_context_test SRCS device_context_test.cu DEPS device_context gpu_ nv_test(cudnn_helper_test SRCS cudnn_helper_test.cc DEPS dynload_cuda) nv_test(transform_test SRCS transform_test.cu DEPS memory place device_context) -cc_library(device_tracer SRCS device_tracer.cc DEPS boost profiler_proto ${GPU_CTX_DEPS}) +cc_library(device_tracer SRCS device_tracer.cc DEPS boost profiler_proto framework_proto ${GPU_CTX_DEPS}) cc_library(profiler SRCS profiler.cc DEPS device_context device_tracer) cc_test(profiler_test SRCS profiler_test.cc DEPS profiler) diff --git a/paddle/fluid/platform/cuda_device_function.h b/paddle/fluid/platform/cuda_device_function.h new file mode 100644 index 0000000000000000000000000000000000000000..ecec4178f2d9937920e52eb74bf9068b84e741a0 --- /dev/null +++ b/paddle/fluid/platform/cuda_device_function.h @@ -0,0 +1,84 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include + +namespace paddle { +namespace platform { + +#if CUDA_VERSION < 9000 +#define CREATE_SHFL_MASK(mask, predicate) mask = 0u; +#else +#define FULL_WARP_MASK 0xFFFFFFFF +#define CREATE_SHFL_MASK(mask, predicate) \ + mask = __ballot_sync(FULL_WARP_MASK, (predicate)) +#endif + +template +__forceinline__ __device__ T CudaShuffleDownSync(unsigned mask, T val, + int delta, int width = 32) { +#if CUDA_VERSION < 9000 + return __shfl_down(val, delta, width); +#else + return __shfl_down_sync(mask, val, delta, width); +#endif +} + +template +__forceinline__ __device__ T CudaShuffleSync(unsigned mask, T val, int src_line, + int width = 32) { +#if CUDA_VERSION < 9000 + return __shfl(val, src_line, width); +#else + return __shfl_sync(mask, val, src_line, width); +#endif +} + +template +__device__ T reduceSum(T val, int tid, int len) { + // NOTE(zcd): The warp size should be taken from the + // parameters of the GPU but not specified as 32 simply. + // To make the reduceSum more efficiently, + // I use Warp-Level Parallelism and assume the Warp size + // is 32 which may be different for different GPU, + // but most card's warp size is 32. + const int warpSize = 32; + __shared__ T shm[warpSize]; + unsigned mask = 0u; + CREATE_SHFL_MASK(mask, tid < len); + + for (int offset = warpSize / 2; offset > 0; offset /= 2) + val += platform::CudaShuffleDownSync(mask, val, offset); + + if (tid < warpSize) shm[tid] = 0; + __syncthreads(); + + if (tid % warpSize == 0) { + shm[tid / warpSize] = val; + } + __syncthreads(); + + CREATE_SHFL_MASK(mask, tid < warpSize); + + if (tid < warpSize) { + val = shm[tid]; + for (int offset = warpSize / 2; offset > 0; offset /= 2) + val += platform::CudaShuffleDownSync(mask, val, offset); + } + return val; +} + +} // namespace platform +} // namespace paddle diff --git a/paddle/fluid/platform/cuda_helper.h b/paddle/fluid/platform/cuda_primitives.h similarity index 99% rename from paddle/fluid/platform/cuda_helper.h rename to paddle/fluid/platform/cuda_primitives.h index 8758af0804ae08fec6fa64d7387f197f046ce20e..d535ed2f89df6a0b311ec068ecd92c8e3183cee7 100644 --- a/paddle/fluid/platform/cuda_helper.h +++ b/paddle/fluid/platform/cuda_primitives.h @@ -65,6 +65,5 @@ CUDA_ATOMIC_WRAPPER(Add, double) { return __longlong_as_double(old); } #endif - } // namespace platform } // namespace paddle diff --git a/paddle/fluid/platform/device_tracer.cc b/paddle/fluid/platform/device_tracer.cc index c9e10631680a6ea3876f555a3a6e6c12f79b39d5..1a9be044e024e4b1dda5ef7d515c65f3a7513710 100644 --- a/paddle/fluid/platform/device_tracer.cc +++ b/paddle/fluid/platform/device_tracer.cc @@ -245,7 +245,6 @@ class DeviceTracerImpl : public DeviceTracer { void Enable() { std::lock_guard l(trace_mu_); if (enabled_) { - fprintf(stderr, "DeviceTracer already enabled\n"); return; } EnableActivity(); diff --git a/paddle/fluid/platform/mkldnn_helper.h b/paddle/fluid/platform/mkldnn_helper.h index 23f1d615daab91f0e4b353bc7d9a3ca7f5cec5ae..f1187620d81ff3bc1deef2106edb54d6199fa927 100644 --- a/paddle/fluid/platform/mkldnn_helper.h +++ b/paddle/fluid/platform/mkldnn_helper.h @@ -38,6 +38,11 @@ void* to_void_cast(const Type* t) { return static_cast(const_cast(t)); } +template +void* to_void_reinterpret_cast(const Type* t) { + return reinterpret_cast(const_cast(t)); +} + template using tf_desc = typename Type::desc; @@ -71,5 +76,15 @@ inline bool CanMKLDNNBeUsed(const framework::ExecutionContext& ctx) { return use_mkldnn && platform::is_cpu_place(ctx.GetPlace()); } +template +mkldnn::memory::data_type MKLDNNGetDataType() { + return mkldnn::memory::data_undef; +} + +template <> +inline mkldnn::memory::data_type MKLDNNGetDataType() { + return mkldnn::memory::f32; +} + } // namespace platform } // namespace paddle diff --git a/paddle/fluid/platform/nccl_helper.h b/paddle/fluid/platform/nccl_helper.h index 0013597fd516d15c7d502370eec77e1a6a5dca88..09367889a9517956ad01ad2847c31e2633cc643d 100644 --- a/paddle/fluid/platform/nccl_helper.h +++ b/paddle/fluid/platform/nccl_helper.h @@ -14,12 +14,15 @@ #pragma once +#include #include // NOLINT #include #include #include "paddle/fluid/platform/dynload/nccl.h" #include "paddle/fluid/platform/enforce.h" +#define NCCL_ID_VARNAME "NCCLID" + namespace paddle { namespace platform { @@ -50,7 +53,7 @@ class NCCLGroupGuard { } inline ~NCCLGroupGuard() { - PADDLE_ENFORCE(dynload::ncclGroupEnd()); + CHECK_EQ(dynload::ncclGroupEnd(), ncclSuccess); NCCLMutex().unlock(); } }; @@ -73,7 +76,9 @@ struct NCCLContextMap { std::unordered_map contexts_; std::vector order_; - explicit NCCLContextMap(const std::vector &places) { + explicit NCCLContextMap(const std::vector &places, + ncclUniqueId *nccl_id = nullptr, + size_t num_trainers = 1, size_t trainer_id = 0) { PADDLE_ENFORCE(!places.empty()); order_.reserve(places.size()); for (auto &p : places) { @@ -85,18 +90,34 @@ struct NCCLContextMap { order_.size(), contexts_.size(), "NCCL Context Map does not support contain two or more same device"); - if (places.size() > 1) { - std::unique_ptr comms(new ncclComm_t[order_.size()]); + if (places.size() <= 1) { + return; + } + std::unique_ptr comms(new ncclComm_t[order_.size()]); + // if pass nccl_id here, can assume we are doing multi node training + if (nccl_id == nullptr) { + std::lock_guard guard(NCCLGroupGuard::NCCLMutex()); + PADDLE_ENFORCE(platform::dynload::ncclCommInitAll( + comms.get(), static_cast(order_.size()), order_.data())); + } else { + PADDLE_ENFORCE_GT(num_trainers, 1); + // TODO(wuyi): need to ensure each node have same number of GPUs { - std::lock_guard guard(NCCLGroupGuard::NCCLMutex()); - PADDLE_ENFORCE(platform::dynload::ncclCommInitAll( - comms.get(), static_cast(order_.size()), order_.data())); - } - int i = 0; - for (auto &dev_id : order_) { - contexts_.at(dev_id).comm_ = comms[i++]; + int nranks = num_trainers * order_.size(); + NCCLGroupGuard gurad; + for (auto &gpu_id : order_) { + int rank = trainer_id * order_.size() + gpu_id; + VLOG(3) << "init nccl rank: " << rank << " nranks: " << nranks; + PADDLE_ENFORCE(cudaSetDevice(gpu_id)); + PADDLE_ENFORCE(platform::dynload::ncclCommInitRank( + comms.get() + gpu_id, nranks, *nccl_id, rank)); + } } } + int i = 0; + for (auto &dev_id : order_) { + contexts_.at(dev_id).comm_ = comms[i++]; + } } NCCLContextMap(const NCCLContextMap &other) = delete; diff --git a/paddle/fluid/platform/profiler.cc b/paddle/fluid/platform/profiler.cc index 412cdda286c3a77af002fdc5eb6a5ae440606b82..3d8d64e4c2758675067834810ebb9aee1e88fdb9 100644 --- a/paddle/fluid/platform/profiler.cc +++ b/paddle/fluid/platform/profiler.cc @@ -13,12 +13,15 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/platform/profiler.h" + #include #include #include #include +#include #include #include // NOLINT +#include #include #ifdef PADDLE_WITH_CUDA #include @@ -33,6 +36,10 @@ namespace platform { struct EventList; +static int64_t profiler_lister_id = 0; +static bool should_send_profile_state = false; +std::mutex profiler_mu; + // The profiler state, the initial value is ProfilerState::kDisabled static ProfilerState g_state = ProfilerState::kDisabled; // The thread local event list only can be accessed by the specific thread @@ -167,8 +174,9 @@ void PopEvent(const std::string& name, const DeviceContext* dev_ctx) { } RecordEvent::RecordEvent(const std::string& name, const DeviceContext* dev_ctx) - : start_ns_(PosixInNsec()) { + : is_enabled_(false), start_ns_(PosixInNsec()) { if (g_state == ProfilerState::kDisabled) return; + is_enabled_ = true; dev_ctx_ = dev_ctx; name_ = name; PushEvent(name_, dev_ctx_); @@ -177,7 +185,7 @@ RecordEvent::RecordEvent(const std::string& name, const DeviceContext* dev_ctx) } RecordEvent::~RecordEvent() { - if (g_state == ProfilerState::kDisabled) return; + if (g_state == ProfilerState::kDisabled || !is_enabled_) return; DeviceTracer* tracer = GetDeviceTracer(); if (tracer) { tracer->AddCPURecords(CurAnnotation(), start_ns_, PosixInNsec(), @@ -187,14 +195,16 @@ RecordEvent::~RecordEvent() { PopEvent(name_, dev_ctx_); } -RecordBlock::RecordBlock(int block_id) : start_ns_(PosixInNsec()) { +RecordBlock::RecordBlock(int block_id) + : is_enabled_(false), start_ns_(PosixInNsec()) { if (g_state == ProfilerState::kDisabled) return; + is_enabled_ = true; SetCurBlock(block_id); name_ = string::Sprintf("block_%d", block_id); } RecordBlock::~RecordBlock() { - if (g_state == ProfilerState::kDisabled) return; + if (g_state == ProfilerState::kDisabled || !is_enabled_) return; DeviceTracer* tracer = GetDeviceTracer(); if (tracer) { // We try to put all blocks at the same nested depth in the @@ -219,13 +229,14 @@ void EnableProfiler(ProfilerState state) { PADDLE_ENFORCE(state != ProfilerState::kDisabled, "Can't enbale profling, since the input state is ", "ProfilerState::kDisabled"); - PADDLE_ENFORCE(g_state == ProfilerState::kDisabled, - "The profiling state should be disabled when calling ", - "EnableProfiler."); - g_state = state; - if (g_state == ProfilerState::kAll) { - GetDeviceTracer()->Enable(); + + std::lock_guard l(profiler_mu); + if (state == g_state) { + return; } + g_state = state; + should_send_profile_state = true; + GetDeviceTracer()->Enable(); #ifdef PADDLE_WITH_CUDA if (g_state == ProfilerState::kCUDA) { // Generate some dummy events first to reduce the startup overhead. @@ -287,7 +298,7 @@ void PrintProfiler(const std::vector>& events_table, } else if (g_state == ProfilerState::kAll) { place = "All"; } else { - PADDLE_THROW("Invalid profiler state"); + PADDLE_THROW("Invalid profiler state", g_state); } std::cout << "Place: " << place << std::endl; @@ -435,8 +446,8 @@ void ParseEvents(const std::vector>& events, void DisableProfiler(EventSortingKey sorted_key, const std::string& profile_path) { - PADDLE_ENFORCE(g_state != ProfilerState::kDisabled, - "Can't disable profiling, since it's not starting."); + std::lock_guard l(profiler_mu); + if (g_state == ProfilerState::kDisabled) return; // Mark the profiling stop. Mark("_stop_profiler_", nullptr); @@ -444,12 +455,25 @@ void DisableProfiler(EventSortingKey sorted_key, ParseEvents(all_events, sorted_key); ResetProfiler(); DeviceTracer* tracer = GetDeviceTracer(); - if (g_state == ProfilerState::kAll && tracer && tracer->IsEnabled()) { + if (tracer->IsEnabled()) { tracer->Disable(); tracer->GenProfile(profile_path); } g_state = ProfilerState::kDisabled; + should_send_profile_state = true; +} + +bool IsProfileEnabled() { return g_state != ProfilerState::kDisabled; } +bool ShouldSendProfileState() { return should_send_profile_state; } + +void SetProfileListener() { + std::mt19937 rng; + rng.seed(std::random_device()()); + std::uniform_int_distribution dist6( + 1, std::numeric_limits::max()); + profiler_lister_id = dist6(rng); } +int64_t ListenerId() { return profiler_lister_id; } } // namespace platform } // namespace paddle diff --git a/paddle/fluid/platform/profiler.h b/paddle/fluid/platform/profiler.h index b07427c8f6903e0100ca9a478881444d86501bcc..bf43925373a12cd9ff2155d68c42d0266ba4df60 100644 --- a/paddle/fluid/platform/profiler.h +++ b/paddle/fluid/platform/profiler.h @@ -18,7 +18,6 @@ limitations under the License. */ #include #include #include "paddle/fluid/platform/device_context.h" -#include "paddle/fluid/platform/profiler.pb.h" namespace paddle { namespace platform { @@ -75,6 +74,7 @@ struct RecordEvent { ~RecordEvent(); + bool is_enabled_; uint64_t start_ns_; // The device context is used by Event to get the current cuda stream. const DeviceContext* dev_ctx_; @@ -90,6 +90,7 @@ struct RecordBlock { ~RecordBlock(); private: + bool is_enabled_; std::string name_; uint64_t start_ns_; }; @@ -115,5 +116,15 @@ void ResetProfiler(); void DisableProfiler(EventSortingKey sorted_key, const std::string& profile_path); +const int kEnableProfiler = 1; +const int kDisableProfiler = 2; +// Test if the profiler is currently enabled. +bool IsProfileEnabled(); +// Whether the trainer should send profiling state to PS. +bool ShouldSendProfileState(); +// Mark current process as PS by assigning a lister id. +void SetProfileListener(); +int64_t ListenerId(); + } // namespace platform } // namespace paddle diff --git a/paddle/fluid/pybind/const_value.cc b/paddle/fluid/pybind/const_value.cc index 3f28e616494ad1322708ad6403aaf50b22d724e6..76aa7d2010682416f68e982e9b89da9813abb078 100644 --- a/paddle/fluid/pybind/const_value.cc +++ b/paddle/fluid/pybind/const_value.cc @@ -13,6 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/pybind/const_value.h" +#include #include "paddle/fluid/framework/operator.h" namespace paddle { @@ -23,6 +24,22 @@ void BindConstValue(pybind11::module* m) { m->def("kTempVarName", [] { return framework::kTempVarName; }); m->def("kGradVarSuffix", [] { return framework::kGradVarSuffix; }); m->def("kZeroVarSuffix", [] { return framework::kZeroVarSuffix; }); + + auto op_proto_and_checker_maker = + m->def_submodule("op_proto_and_checker_maker"); + + pybind11::enum_(op_proto_and_checker_maker, "OpRole") + .value("Forward", framework::OpRole::kForward) + .value("Backward", framework::OpRole::kBackward) + .value("Optimize", framework::OpRole::kOptimize) + .value("Loss", framework::OpRole::kLoss) + .value("RPC", framework::OpRole::kRPC); + + op_proto_and_checker_maker.def( + "kOpRoleAttrName", framework::OpProtoAndCheckerMaker::OpRoleAttrName); + op_proto_and_checker_maker.def( + "kOpRoleVarAttrName", + framework::OpProtoAndCheckerMaker::OpRoleVarAttrName); } } // namespace pybind diff --git a/paddle/fluid/pybind/protobuf.cc b/paddle/fluid/pybind/protobuf.cc index 6471eb3ab7bf05365c0bb2bf68bb74ef9044c527..bcf6d4dd3087060c016e53722cde80704ef2e834 100644 --- a/paddle/fluid/pybind/protobuf.cc +++ b/paddle/fluid/pybind/protobuf.cc @@ -238,6 +238,7 @@ void BindVarDsec(pybind11::module *m) { pybind11::enum_(var_desc, "VarType", "") .value("BOOL", pd::proto::VarType::BOOL) + .value("UINT8", pd::proto::VarType::UINT8) .value("INT16", pd::proto::VarType::INT16) .value("INT32", pd::proto::VarType::INT32) .value("INT64", pd::proto::VarType::INT64) diff --git a/paddle/fluid/pybind/pybind.cc b/paddle/fluid/pybind/pybind.cc index b20b514fcdd0b41fefa0933bc2d22645e7d4b6d6..3af8941be69fe507bc105e26b608ec768e4b5998 100644 --- a/paddle/fluid/pybind/pybind.cc +++ b/paddle/fluid/pybind/pybind.cc @@ -117,6 +117,7 @@ PYBIND11_PLUGIN(core) { .def("set", PyCPUTensorSetFromArray) .def("set", PyCPUTensorSetFromArray) .def("set", PyCPUTensorSetFromArray) + .def("set", PyCPUTensorSetFromArray) #ifdef PADDLE_WITH_CUDA .def("set", PyCUDATensorSetFromArray) .def("set", PyCUDATensorSetFromArray) @@ -124,12 +125,14 @@ PYBIND11_PLUGIN(core) { .def("set", PyCUDATensorSetFromArray) .def("set", PyCUDATensorSetFromArray) .def("set", PyCUDATensorSetFromArray) + .def("set", PyCUDATensorSetFromArray) .def("set", PyCUDAPinnedTensorSetFromArray) .def("set", PyCUDAPinnedTensorSetFromArray) .def("set", PyCUDAPinnedTensorSetFromArray) .def("set", PyCUDAPinnedTensorSetFromArray) .def("set", PyCUDAPinnedTensorSetFromArray) .def("set", PyCUDAPinnedTensorSetFromArray) + .def("set", PyCUDAPinnedTensorSetFromArray) #endif .def("shape", [](Tensor &self) { return vectorize(self.dims()); }) .def("set_float_element", TensorSetElement) @@ -492,22 +495,64 @@ All parameter, weight, gradient are variables in Paddle. m.def("enable_profiler", platform::EnableProfiler); m.def("disable_profiler", platform::DisableProfiler); + m.def("is_profiler_enabled", platform::IsProfileEnabled); m.def("reset_profiler", platform::ResetProfiler); - py::class_(m, "ParallelExecutor") - .def("__init__", - [](ParallelExecutor &self, size_t num_threads, bool use_event, - const std::vector &places, - const std::unordered_set ¶ms, - const std::unordered_set &bcast_vars, - const ProgramDesc &main_program, const std::string &loss_var_name, - Scope *scope, std::vector &local_scopes, - bool allow_op_delay, bool customize_loss_grad) { - new (&self) ParallelExecutor(num_threads, use_event, places, - params, bcast_vars, main_program, - loss_var_name, scope, local_scopes, - allow_op_delay, customize_loss_grad); - }) + // -- python binds for parallel executor. + py::class_ pe(m, "ParallelExecutor"); + py::class_(pe, "ExecutionStrategy") + .def(py::init()) + .def_property( + "num_threads", + [](const ExecutionStrategy &self) { return self.num_threads_; }, + [](ExecutionStrategy &self, size_t num_threads) { + self.num_threads_ = num_threads; + }) + .def_property( + "use_event", + [](const ExecutionStrategy &self) { return self.use_event_; }, + [](ExecutionStrategy &self, bool use_event) { + self.use_event_ = use_event; + }) + .def_property( + "allow_op_delay", + [](const ExecutionStrategy &self) { return self.allow_op_delay_; }, + [](ExecutionStrategy &self, bool allow_op_delay) { + self.allow_op_delay_ = allow_op_delay; + }); + py::class_ build_strategy(pe, "BuildStrategy"); + + py::enum_(build_strategy, "ReduceStrategy") + .value("Reduce", BuildStrategy::ReduceStrategy::kReduce) + .value("AllReduce", BuildStrategy::ReduceStrategy::kAllReduce); + py::enum_(build_strategy, + "GradientScaleStrategy") + .value("CoeffNumDevice", + BuildStrategy::GradientScaleStrategy::kCoeffNumDevice) + .value("One", BuildStrategy::GradientScaleStrategy::kOne) + .value("Customized", BuildStrategy::GradientScaleStrategy::kCustomized); + + build_strategy.def(py::init()) + .def_property( + "reduce_strategy", + [](const BuildStrategy &self) { return self.reduce_; }, + [](BuildStrategy &self, BuildStrategy::ReduceStrategy strategy) { + self.reduce_ = strategy; + }) + .def_property( + "gradient_scale_strategy", + [](const BuildStrategy &self) { return self.gradient_scale_; }, + [](BuildStrategy &self, + BuildStrategy::GradientScaleStrategy strategy) { + self.gradient_scale_ = strategy; + }); + + pe.def(py::init &, + const std::unordered_set &, + const std::unordered_set &, const ProgramDesc &, + const std::string &, Scope *, std::vector &, + const ExecutionStrategy &, const BuildStrategy &, size_t, + size_t>()) .def("bcast_params", &ParallelExecutor::BCastParamsToGPUs) // NOTE: even we return a vec* to Python use reference policy. // We still cannot get local_scope from this vector, since the element diff --git a/paddle/fluid/pybind/tensor_py.h b/paddle/fluid/pybind/tensor_py.h index 159d1d5f4e70033fabf93514bd63b38f83675bff..93b09ed6922b32a5531224acc470daf0d97f95bd 100644 --- a/paddle/fluid/pybind/tensor_py.h +++ b/paddle/fluid/pybind/tensor_py.h @@ -63,15 +63,9 @@ struct CastToPyBufferImpl { auto *dst_ptr = static_cast(dst_tensor.mutable_data( tensor.dims(), platform::CPUPlace())); - platform::DeviceContextPool &pool = - platform::DeviceContextPool::Instance(); - auto dev_ctx = static_cast( - pool.Get(tensor.place())); - - paddle::platform::GpuMemcpyAsync( - dst_ptr, src_ptr, sizeof(CUR_TYPE) * tensor.numel(), - cudaMemcpyDeviceToHost, dev_ctx->stream()); - dev_ctx->Wait(); + paddle::platform::GpuMemcpySync(dst_ptr, src_ptr, + sizeof(CUR_TYPE) * tensor.numel(), + cudaMemcpyDeviceToHost); #else PADDLE_THROW("'CUDAPlace' is not supported in CPU only device."); #endif @@ -113,7 +107,7 @@ T TensorGetElement(const framework::Tensor &self, size_t offset) { return self.data()[offset]; } else { std::shared_ptr dst(new framework::Tensor); - framework::TensorCopy(self, platform::CPUPlace(), dst.get()); + framework::TensorCopySync(self, platform::CPUPlace(), dst.get()); return dst->data()[offset]; } } @@ -123,9 +117,9 @@ template void TensorSetElement(framework::Tensor *self, size_t offset, T elem) { if (platform::is_gpu_place(self->place())) { std::shared_ptr dst(new framework::Tensor); - framework::TensorCopy(*self, platform::CPUPlace(), dst.get()); + framework::TensorCopySync(*self, platform::CPUPlace(), dst.get()); dst->data()[offset] = elem; - framework::TensorCopy(*dst.get(), self->place(), self); + framework::TensorCopySync(*dst.get(), self->place(), self); } else if (platform::is_cpu_place(self->place())) { self->data()[offset] = elem; @@ -184,17 +178,8 @@ void PyCUDATensorSetFromArray( self->Resize(framework::make_ddim(dims)); auto *dst = self->mutable_data(place); - - platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance(); - auto dev_ctx = - static_cast(pool.Get(place)); - paddle::platform::GpuMemcpyAsync(dst, array.data(), sizeof(T) * array.size(), - cudaMemcpyHostToDevice, dev_ctx->stream()); - // NOTE: For safety, here wait the copy complete. - // It because the CPU array.data() could be destroyed after this method. - // If we make this method async, it could be copied data from a memory buffer - // that has been freed. - dev_ctx->Wait(); + paddle::platform::GpuMemcpySync(dst, array.data(), sizeof(T) * array.size(), + cudaMemcpyHostToDevice); } template <> @@ -214,18 +199,9 @@ void PyCUDATensorSetFromArray( self->Resize(framework::make_ddim(dims)); auto *dst = self->mutable_data(place); - - platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance(); - auto dev_ctx = - static_cast(pool.Get(place)); - paddle::platform::GpuMemcpyAsync(dst, array.data(), - sizeof(uint16_t) * array.size(), - cudaMemcpyHostToDevice, dev_ctx->stream()); - // NOTE: For safety, here wait the copy complete. - // It because the CPU array.data() could be destroyed after this method. - // If we make this method async, it could be copied data from a memory buffer - // that has been freed. - dev_ctx->Wait(); + paddle::platform::GpuMemcpySync(dst, array.data(), + sizeof(uint16_t) * array.size(), + cudaMemcpyHostToDevice); } template diff --git a/paddle/fluid/train/demo/CMakeLists.txt b/paddle/fluid/train/demo/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..78d6e5ff554b9cd9facae85be166a697e0b75337 --- /dev/null +++ b/paddle/fluid/train/demo/CMakeLists.txt @@ -0,0 +1,66 @@ +cmake_minimum_required(VERSION 3.0) + +project(cpp_train_demo CXX C) + +set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++11") + +if(NOT DEFINED PADDLE_LIB) + message(FATAL_ERROR "please set PADDLE_LIB with -DPADDLE_LIB=/paddle/lib/dir") +endif() + +option(WITH_MKLDNN "Compile PaddlePaddle with MKLDNN" OFF) +option(WITH_MKL "Compile PaddlePaddle with MKL support, default use openblas." OFF) + +include_directories("${PADDLE_LIB}") +include_directories("${PADDLE_LIB}/third_party/install/protobuf/include") +include_directories("${PADDLE_LIB}/third_party/install/glog/include") +include_directories("${PADDLE_LIB}/third_party/install/gflags/include") +include_directories("${PADDLE_LIB}/third_party/install/snappy/include") +include_directories("${PADDLE_LIB}/third_party/install/snappystream/include") +include_directories("${PADDLE_LIB}/third_party/install/zlib/include") + +include_directories("${PADDLE_LIB}/third_party/boost") +include_directories("${PADDLE_LIB}/third_party/eigen3") + +link_directories("${PADDLE_LIB}/third_party/install/snappy/lib") +link_directories("${PADDLE_LIB}/third_party/install/snappystream/lib") +link_directories("${PADDLE_LIB}/third_party/install/protobuf/lib") +link_directories("${PADDLE_LIB}/third_party/install/glog/lib") +link_directories("${PADDLE_LIB}/third_party/install/gflags/lib") +link_directories("${PADDLE_LIB}/third_party/install/zlib/lib") + +add_executable(demo_trainer demo_trainer.cc) + +if(WITH_MKLDNN) + include_directories("${PADDLE_LIB}/third_party/install/mkldnn/include") + set(MKLDNN_LIB ${PADDLE_LIB}/third_party/install/mkldnn/lib/libmkldnn.so.0) +endif() + +if(WITH_MKL) + include_directories("${PADDLE_LIB}/third_party/install/mklml/include") + set(MATH_LIB ${PADDLE_LIB}/third_party/install/mklml/lib/libmklml_intel.so) +else() + if(APPLE) + set(MATH_LIB cblas) + else(APPLE) + set(MATH_LIB ${PADDLE_LIB}/third_party/install/openblas/lib/libopenblas.a) + endif(APPLE) +endif() + +if(APPLE) + set(MACOS_LD_FLAGS "-undefined dynamic_lookup -Wl,-all_load -framework CoreFoundation -framework Security") +else(APPLE) + set(ARCHIVE_START "-Wl,--whole-archive") + set(ARCHIVE_END "-Wl,--no-whole-archive") + set(EXTERNAL_LIB "-lrt -ldl -lpthread") +endif(APPLE) + +target_link_libraries(demo_trainer + ${MACOS_LD_FLAGS} + ${ARCHIVE_START} + ${PADDLE_LIB}/paddle/fluid/inference/libpaddle_fluid.a + ${ARCHIVE_END} + ${MATH_LIB} + ${MKLDNN_LIB} + glog gflags protobuf snappystream snappy z + ${EXTERNAL_LIB}) diff --git a/paddle/fluid/train/demo/README.md b/paddle/fluid/train/demo/README.md new file mode 100644 index 0000000000000000000000000000000000000000..41b01d33828f750f67bba5f82cb7ed6fe4d4ea0a --- /dev/null +++ b/paddle/fluid/train/demo/README.md @@ -0,0 +1,66 @@ + +### step 1. build paddle lib + +``` + +# WITH_MKL=ON|OFF +# WITH_MKLDNN=ON|OFF + +PADDLE_LIB=/paddle/lib/dir +cmake .. -DFLUID_INSTALL_DIR=$PADDLE_LIB \ + -DCMAKE_BUILD_TYPE=Release \ + -DWITH_FLUID_ONLY=ON \ + -DWITH_GPU=OFF \ + -DWITH_STYLE_CHECK=OFF \ + -DWITH_MKL=OFF \ + -DWITH_MKLDNN=OFF +make -j8 +make -j8 inference_lib_dist +``` + +### step 2. generate program desc +``` +# please install paddle before run this scripe +pip install --upgrade paddlepaddle-*.whl +python demo_network.py +``` + +This will generate two program desc files: + - startup_program: used to init all parameters + - main_program: main logic of the network + +### step 3. build demo_trainer and run it. + + +``` +# Make a build dir at the same dir of this README.md document. +# The demo dir can be put anywhere. +mkdir build +cd build + +# WITH_MKL=ON|OFF +# WITH_MKLDNN=ON|OFF +PADDLE_LIB=/paddle/lib/dir + +# PADDLE_LIB is the same with FLUID_INSTALL_DIR when building the lib +cmake .. -DPADDLE_LIB=$PADDLE_LIB \ + -DWITH_MKLDNN=OFF \ + -DWITH_MKL=OFF +make + +# copy startup_program and main_program to this dir +cp ../startup_program . +cp ../main_program . + +# run demo cpp trainer +./demo_trainer + +``` + +The output will be: +``` +step: 0 loss: 1069.02 +step: 1 loss: 1069.02 +step: 2 loss: 1069.02 +.... +``` diff --git a/paddle/fluid/train/demo/demo_network.py b/paddle/fluid/train/demo/demo_network.py new file mode 100644 index 0000000000000000000000000000000000000000..41e98c6a24a750a9300b5c2a6d370303cc0e59c5 --- /dev/null +++ b/paddle/fluid/train/demo/demo_network.py @@ -0,0 +1,47 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import paddle.fluid as fluid +import paddle.fluid.framework as framework + + +def train_network(with_optimize): + x = fluid.layers.data(name='x', shape=[13], dtype='float32') + y_predict = fluid.layers.fc(input=x, size=1, act=None) + + y = fluid.layers.data(name='y', shape=[1], dtype='float32') + cost = fluid.layers.square_error_cost(input=y_predict, label=y) + avg_cost = fluid.layers.mean(cost) + + if with_optimize: + sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.00001) + sgd_optimizer.minimize(avg_cost) + else: + fluid.backward.append_backward(avg_cost) + + +def save_program_desc(network_func): + startup_program = framework.Program() + train_program = framework.Program() + + with framework.program_guard(train_program, startup_program): + network_func(with_optimize=False) + + with open("startup_program", "w") as f: + f.write(startup_program.desc.serialize_to_string()) + with open("main_program", "w") as f: + f.write(train_program.desc.serialize_to_string()) + + +save_program_desc(train_network) diff --git a/paddle/fluid/train/demo/demo_trainer.cc b/paddle/fluid/train/demo/demo_trainer.cc new file mode 100644 index 0000000000000000000000000000000000000000..813d8386868558bd62a9d5670d540ddeddb2b77d --- /dev/null +++ b/paddle/fluid/train/demo/demo_trainer.cc @@ -0,0 +1,103 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include + +#include "paddle/fluid/framework/executor.h" +#include "paddle/fluid/framework/init.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/framework/program_desc.h" +#include "paddle/fluid/framework/tensor_util.h" +#include "paddle/fluid/platform/device_context.h" +#include "paddle/fluid/platform/place.h" + +namespace paddle { +namespace train { + +void ReadBinaryFile(const std::string& filename, std::string* contents) { + std::ifstream fin(filename, std::ios::in | std::ios::binary); + PADDLE_ENFORCE(static_cast(fin), "Cannot open file %s", filename); + fin.seekg(0, std::ios::end); + contents->clear(); + contents->resize(fin.tellg()); + fin.seekg(0, std::ios::beg); + fin.read(&(contents->at(0)), contents->size()); + fin.close(); +} + +std::unique_ptr Load( + paddle::framework::Executor* executor, const std::string& model_filename) { + VLOG(3) << "loading model from " << model_filename; + std::string program_desc_str; + ReadBinaryFile(model_filename, &program_desc_str); + + std::unique_ptr main_program( + new paddle::framework::ProgramDesc(program_desc_str)); + return main_program; +} + +} // namespace train +} // namespace paddle + +int main() { + paddle::framework::InitDevices(false); + + const auto cpu_place = paddle::platform::CPUPlace(); + + paddle::framework::Executor executor(cpu_place); + paddle::framework::Scope scope; + auto startup_program = paddle::train::Load(&executor, "startup_program"); + auto train_program = paddle::train::Load(&executor, "main_program"); + + std::string loss_name = ""; + for (auto op_desc : train_program->Block(0).AllOps()) { + if (op_desc->Type() == "mean") { + loss_name = op_desc->Output("Out")[0]; + break; + } + } + + PADDLE_ENFORCE_NE(loss_name, "", "loss not found"); + + // init all parameters + executor.Run(*startup_program.get(), &scope, 0); + + // prepare data + auto x_var = scope.Var("x"); + auto x_tensor = x_var->GetMutable(); + x_tensor->Resize({2, 13}); + + auto x_data = x_tensor->mutable_data(cpu_place); + for (int i = 0; i < 2 * 13; ++i) { + x_data[i] = static_cast(i); + } + + auto y_var = scope.Var("y"); + auto y_tensor = y_var->GetMutable(); + y_tensor->Resize({2, 1}); + auto y_data = y_tensor->mutable_data(cpu_place); + for (int i = 0; i < 2 * 1; ++i) { + y_data[i] = static_cast(i); + } + + auto loss_var = scope.Var(loss_name); + + for (int i = 0; i < 10; ++i) { + executor.Run(*train_program.get(), &scope, 0, false, true); + std::cout << "step: " << i << " loss: " + << loss_var->Get().data()[0] + << std::endl; + } + return 0; +} diff --git a/paddle/function/BlockExpandOp.cpp b/paddle/function/BlockExpandOp.cpp index aa53853e08716ff0dd8dce7c73766d9543bed2b9..f01f89a7277acc5fe494b92a3e7ca3ca18498c97 100644 --- a/paddle/function/BlockExpandOp.cpp +++ b/paddle/function/BlockExpandOp.cpp @@ -33,7 +33,7 @@ namespace paddle { * \param outputs[0] Image data of NCHW format. */ class BlockExpandFunction : public FunctionBase { -public: + public: void init(const FuncConfig& config) override { // function arguments strides_ = config.get>("strides"); @@ -81,7 +81,7 @@ public: (size_t)blockW()}); } -protected: + protected: std::vector strides_; std::vector paddings_; std::vector blocks_; @@ -101,7 +101,7 @@ protected: template class BlockExpandForward : public BlockExpandFunction { -public: + public: void init(const FuncConfig& config) override { BlockExpandFunction::init(config); } @@ -149,7 +149,7 @@ public: template class BlockExpandBackward : public BlockExpandFunction { -public: + public: void init(const FuncConfig& config) override { BlockExpandFunction::init(config); } diff --git a/paddle/function/BufferArg.h b/paddle/function/BufferArg.h index 89ee09837db69d79bbd678312f02f6dc87e8067c..6de8c94e778c8d1439b2a2aa3c581a5a3cf70261 100644 --- a/paddle/function/BufferArg.h +++ b/paddle/function/BufferArg.h @@ -63,12 +63,12 @@ enum ArgType { ADD_TO = 2, }; class BufferArg { -public: + public: void setArgType(ArgType argType) { argType_ = argType; } ArgType getArgType() const { return argType_; } -public: + public: BufferArg(ValueType valueType, const TensorShape& shape, ArgType argType = UNSPECIFIED) @@ -169,7 +169,7 @@ public: const SequenceArg& sequence() const; const SparseMatrixArg& sparse() const; -protected: + protected: void* buf_; ValueType valueType_; TensorShape shape_; @@ -185,7 +185,7 @@ protected: // valueType_ = int32 // if a < b then value_.buf_[a] < value_.buf_[b] class SequenceIdArg : public BufferArg { -public: + public: SequenceIdArg(const TensorShape& shape, ArgType argType = UNSPECIFIED) : BufferArg(VALUE_TYPE_INT32, shape, argType) { bufferType_ = TENSOR_SEQUENCE_ID; @@ -212,7 +212,7 @@ public: size_t numSeqs() const { return numSeqs_; } -private: + private: size_t numSeqs_; }; @@ -222,7 +222,7 @@ private: // SequenceArg can be used to represent sequences that contain multiple // unequal lengths. class SequenceArg : public BufferArg { -public: + public: SequenceArg(ValueType valueType, const TensorShape& shape, ArgType argType = UNSPECIFIED) @@ -255,7 +255,7 @@ public: SequenceIdArg& getSequenceId() { return startPositions_; } const SequenceIdArg& getSequenceId() const { return startPositions_; } -private: + private: SequenceIdArg startPositions_; }; @@ -263,7 +263,7 @@ private: // valueType_ == float or double // shape_.ndims() == 2 class SparseMatrixArg : public BufferArg { -public: + public: SparseMatrixArg(void* buf, ValueType valueType, const TensorShape& shape, @@ -353,7 +353,7 @@ public: SparseDataType dataType() const { return type_; } -private: + private: BufferArg row_; BufferArg col_; size_t nnz_; diff --git a/paddle/function/CMakeLists.txt b/paddle/function/CMakeLists.txt index 9b2779b42cad324253dadf27dbff20fd8e8c8e16..29b4ac098e21ee315d5c9b2f2499521d1aa1c322 100644 --- a/paddle/function/CMakeLists.txt +++ b/paddle/function/CMakeLists.txt @@ -52,9 +52,3 @@ add_simple_unittest(Im2ColTest) add_simple_unittest(GemmConvOpTest) add_simple_unittest(DepthwiseConvOpTest) endif() - -add_style_check_target(paddle_function ${h_files}) -add_style_check_target(paddle_function ${cpp_files}) -if(WITH_GPU) - add_style_check_target(paddle_function ${cu_files}) -endif() diff --git a/paddle/function/ContextProjectionOp.cpp b/paddle/function/ContextProjectionOp.cpp index 904b0958e6f2c1b8fb8cf56f3cd7d07ad8e24f19..1187842452460ac3fd71f48150fab6467f93dc6c 100644 --- a/paddle/function/ContextProjectionOp.cpp +++ b/paddle/function/ContextProjectionOp.cpp @@ -100,7 +100,7 @@ void ContextProjectionForward(CpuMatrix& out_mat, */ template class ContextProjectionForwardFunc : public FunctionBase { -public: + public: void init(const FuncConfig& config) override { context_length_ = config.get("context_length"); context_start_ = config.get("context_start"); @@ -146,7 +146,7 @@ public: begin_pad_); } -private: + private: size_t context_length_; int context_start_; size_t begin_pad_; @@ -223,7 +223,7 @@ void ContextProjectionBackward(const CpuMatrix& out_grad_mat, */ template class ContextProjectionBackwardFunc : public FunctionBase { -public: + public: void init(const FuncConfig& config) override { context_length_ = config.get("context_length"); context_start_ = config.get("context_start"); @@ -278,7 +278,7 @@ public: total_pad_); } -private: + private: size_t context_length_; int context_start_; size_t begin_pad_; @@ -299,7 +299,7 @@ private: */ template class ContextProjectionBackwardDataFunc : public FunctionBase { -public: + public: void init(const FuncConfig& config) override { context_length_ = config.get("context_length"); context_start_ = config.get("context_start"); @@ -331,7 +331,7 @@ public: out_grad_mat, in_grad_mat, seq_vec, context_length_, context_start_); } -private: + private: size_t context_length_; int context_start_; }; @@ -348,7 +348,7 @@ private: */ template class ContextProjectionBackwardWeightFunc : public FunctionBase { -public: + public: void init(const FuncConfig& config) override { context_length_ = config.get("context_length"); context_start_ = config.get("context_start"); @@ -382,7 +382,7 @@ public: begin_pad_); } -private: + private: size_t context_length_; int context_start_; size_t begin_pad_; diff --git a/paddle/function/ConvOp.h b/paddle/function/ConvOp.h index 7d23d0079c8f62b2c8912dfcb9f191c622a60bc9..2d8437bcfe60d1d81897f1c4be1cbfecb5b27fe0 100644 --- a/paddle/function/ConvOp.h +++ b/paddle/function/ConvOp.h @@ -56,7 +56,7 @@ namespace paddle { * H and W is height and width of filter. */ class ConvFunctionBase : public FunctionBase { -public: + public: void init(const FuncConfig& config) override { // function arguments strides_ = config.get>("strides"); @@ -101,7 +101,7 @@ public: } } -protected: + protected: size_t getFilterHeight(const TensorShape& filter) const { return filter[filter.ndims() - 2]; } diff --git a/paddle/function/CosSimOp.cpp b/paddle/function/CosSimOp.cpp index 81bccc1a9c7d614763a10e3838271b57eef2c603..2c25e1af44965d30591faeccc9a181e36c7e0a0f 100644 --- a/paddle/function/CosSimOp.cpp +++ b/paddle/function/CosSimOp.cpp @@ -97,7 +97,7 @@ class CosSimForwardFunc : public FunctionBase { CosSimForward(out_mat, in1_mat, in2_mat, scale_); } -private: + private: real scale_; }; @@ -227,7 +227,7 @@ class CosSimBackwardFunc : public FunctionBase { out_grad, out_val, in1_val, in2_val, in1_grad, in2_grad, scale_); } -private: + private: real scale_; }; diff --git a/paddle/function/CropOp.cpp b/paddle/function/CropOp.cpp index 7aa527d21615e19257bd003d0563b5e26b2fcb2f..5bd98910fe838751935f8ef2387ce96e755c6df1 100644 --- a/paddle/function/CropOp.cpp +++ b/paddle/function/CropOp.cpp @@ -112,7 +112,7 @@ void CropGrad(const real* inGrad, */ template class CropFunc : public FunctionBase { -public: + public: void init(const FuncConfig& config) override { conf_ = config; } void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { @@ -130,7 +130,7 @@ public: conf_); } -private: + private: FuncConfig conf_; }; @@ -145,7 +145,7 @@ private: template class CropGradFunc : public FunctionBase { -public: + public: void init(const FuncConfig& config) override { conf_ = config; } void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { @@ -163,7 +163,7 @@ public: conf_); } -private: + private: FuncConfig conf_; }; diff --git a/paddle/function/CrossMapNormalOp.cpp b/paddle/function/CrossMapNormalOp.cpp index 75c0fc2a3d047a9162d49809a717629f2270872d..7ff9227e5c2702d9d5334db501730b57ec10bfe3 100644 --- a/paddle/function/CrossMapNormalOp.cpp +++ b/paddle/function/CrossMapNormalOp.cpp @@ -160,7 +160,7 @@ void CrossMapNormalGrad(real* inputsGrad, */ template class CrossMapNormalFunc : public FunctionBase { -public: + public: void init(const FuncConfig& config) override { // function arguments size_ = config.get("size"); @@ -220,7 +220,7 @@ public: return ops; } -private: + private: size_t size_; real scale_; real pow_; @@ -260,7 +260,7 @@ private: */ template class CrossMapNormalGradFunc : public FunctionBase { -public: + public: void init(const FuncConfig& config) override { // function arguments size_ = config.get("size"); @@ -328,7 +328,7 @@ public: return ops; } -private: + private: size_t size_; real scale_; real pow_; diff --git a/paddle/function/DepthwiseConvOp.cpp b/paddle/function/DepthwiseConvOp.cpp index 46651345b45e4ced9a3ef3373af437d939a66716..958034e08e60c9a63d1c480bde7c84b760205ae4 100644 --- a/paddle/function/DepthwiseConvOp.cpp +++ b/paddle/function/DepthwiseConvOp.cpp @@ -19,7 +19,7 @@ namespace paddle { template class DepthwiseConvFunctor { -public: + public: void operator()(const T* inputData, const T* filterData, int batchSize, @@ -43,7 +43,7 @@ public: template class DepthwiseConvGradInputFunctor { -public: + public: void operator()(const T* outputGrad, const T* filterData, int batchSize, @@ -66,7 +66,7 @@ public: template class DepthwiseConvGradFilterFunctor { -public: + public: void operator()(const T* outputGrad, const T* inputData, int batchSize, @@ -93,7 +93,7 @@ public: */ template class DepthwiseConvFunction : public ConvFunctionBase { -public: + public: void init(const FuncConfig& config) override { ConvFunctionBase::init(config); } @@ -156,7 +156,7 @@ public: */ template class DepthwiseConvGradInputFunction : public ConvFunctionBase { -public: + public: void init(const FuncConfig& config) override { ConvFunctionBase::init(config); } @@ -220,7 +220,7 @@ public: */ template class DepthwiseConvGradFilterFunction : public ConvFunctionBase { -public: + public: void init(const FuncConfig& config) override { ConvFunctionBase::init(config); } diff --git a/paddle/function/DepthwiseConvOp.h b/paddle/function/DepthwiseConvOp.h index 6700747314fa8377828dab0c436eb4b2053f46f6..7837edd1c071980592b1cf36ecb69a3b7c12cc5e 100644 --- a/paddle/function/DepthwiseConvOp.h +++ b/paddle/function/DepthwiseConvOp.h @@ -44,7 +44,7 @@ namespace paddle { */ template class DepthwiseConvFunctor { -public: + public: void operator()(const T* inputData, const T* filterData, int batchSize, @@ -89,7 +89,7 @@ public: */ template class DepthwiseConvGradInputFunctor { -public: + public: void operator()(const T* outputGrad, const T* filterData, int batchSize, @@ -135,7 +135,7 @@ public: */ template class DepthwiseConvGradFilterFunctor { -public: + public: void operator()(const T* outputGrad, const T* inputData, int batchSize, diff --git a/paddle/function/DepthwiseConvOpGpu.cu b/paddle/function/DepthwiseConvOpGpu.cu index cd1d55a416c84c6327226ffaae4d5d9d5be81038..2c0e71b19b22abac25d273d8bbeddc330e67f8b0 100644 --- a/paddle/function/DepthwiseConvOpGpu.cu +++ b/paddle/function/DepthwiseConvOpGpu.cu @@ -199,7 +199,7 @@ __global__ void ConvolutionDepthwiseFilterBackward(const int num_i, template class DepthwiseConvFunctor { -public: + public: void operator()(const T* inputData, const T* filterData, int batchSize, @@ -249,7 +249,7 @@ public: template class DepthwiseConvGradInputFunctor { -public: + public: void operator()(const T* outputGrad, const T* filterData, int batchSize, @@ -300,7 +300,7 @@ public: template class DepthwiseConvGradFilterFunctor { -public: + public: void operator()(const T* outputGrad, const T* inputData, int batchSize, diff --git a/paddle/function/EigenGemm.cpp b/paddle/function/EigenGemm.cpp index bac4659e62b107dd80ef95dd0907b3da4becffbc..8e9dbbd7a154095a7298bb2f59a82d13a60f9bd3 100644 --- a/paddle/function/EigenGemm.cpp +++ b/paddle/function/EigenGemm.cpp @@ -13,7 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #include -#include "unsupported/Eigen/CXX11/Tensor" +#include "paddle/function/EigenThreadDevice.h" namespace paddle { @@ -70,25 +70,26 @@ struct EigenBlasGemm { dims[0].first = transA ? 0 : 1; dims[0].second = transB ? 1 : 0; - Eigen::DefaultDevice device; + auto* device = EigenDeviceWarpper::device(); if (N == ldc) { if (alpha == T(1) && beta == T(0)) { - c.device(device) = a.contract(b, dims); + c.device(*device) = a.contract(b, dims); } else if (alpha == T(1) && beta == T(1)) { - c.device(device) += a.contract(b, dims); + c.device(*device) += a.contract(b, dims); } else { - c.device(device) = alpha * a.contract(b, dims) + beta * c; + c.device(*device) = alpha * a.contract(b, dims) + beta * c; } } else { if (alpha == T(1) && beta == T(0)) { - c.slice(offsetC, extentC).device(device) = a.contract(b, dims); + c.slice(offsetC, extentC).device(*device) = a.contract(b, dims); } else if (alpha == T(1) && beta == T(1)) { - c.slice(offsetC, extentC).device(device) += a.contract(b, dims); + c.slice(offsetC, extentC).device(*device) += a.contract(b, dims); } else { - c.slice(offsetC, extentC).device(device) = + c.slice(offsetC, extentC).device(*device) = alpha * a.contract(b, dims) + beta * c.slice(offsetC, extentC); } } + EigenDeviceWarpper::free_device(device); } }; diff --git a/paddle/function/EigenThreadDevice.h b/paddle/function/EigenThreadDevice.h new file mode 100644 index 0000000000000000000000000000000000000000..eb92251c827a26d55ca021c4418182bae28dd6a5 --- /dev/null +++ b/paddle/function/EigenThreadDevice.h @@ -0,0 +1,73 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. */ + +#pragma once + +#if defined(__OSX__) || defined(__APPLE__) +#include +#include +#endif +#include "unsupported/Eigen/CXX11/Tensor" + +namespace paddle { + +#if defined(__ANDROID__) +int GetCpuCount() { + FILE* fp = fopen("/sys/devices/system/cpu/possible", "r"); + if (!fp) { + return 1; + } + int rank0, rank1; + int num = fscanf(fp, "%d-%d", &rank0, &rank1); + fclose(fp); + if (num < 2) return 1; + return rank1 + 1; +} +#elif defined(__OSX__) || defined(__APPLE__) +int GetCpuCount() { + int count = 0; + size_t len = sizeof(int); + sysctlbyname("hw.ncpu", &count, &len, NULL, 0); + return count > 0 ? count : 1; +} +#else +int GetCpuCount() { return 1; } +#endif + +class EigenDeviceWarpper { + public: // NOLINT +#if EIGEN_USE_THREADS + static Eigen::ThreadPoolDevice* device() { + const int num_cpus = GetCpuCount(); + const int num_threads = (num_cpus > 2) ? 2 : num_cpus; + static Eigen::ThreadPool tp(num_threads); + static Eigen::ThreadPoolDevice* device = + new Eigen::ThreadPoolDevice(&tp, num_threads); + return device; + } + + static void free_device(Eigen::ThreadPoolDevice* device) { + // do nothing + } +#else + static Eigen::DefaultDevice* device() { + Eigen::DefaultDevice* device = new Eigen::DefaultDevice; + return device; + } + + static void free_device(Eigen::DefaultDevice* device) { delete device; } +#endif +}; + +} // namespace paddle diff --git a/paddle/function/Function.h b/paddle/function/Function.h index 01288ef92e7b59d7958e6e23daf641b30a60eed1..a6c14ef29b760faa393c37bd2357824a061c7b38 100644 --- a/paddle/function/Function.h +++ b/paddle/function/Function.h @@ -29,7 +29,7 @@ namespace paddle { * The argument type of Function::init. */ class FuncConfig { -public: + public: template T get(const std::string& key, Error* err = nullptr) const { try { @@ -59,7 +59,7 @@ public: return *this; } -protected: + protected: mutable std::unordered_map valueMap_; }; @@ -77,7 +77,7 @@ protected: * in the BufferArgs life time. */ class BufferArgs { -public: + public: BufferArgs() {} ~BufferArgs() { @@ -137,7 +137,7 @@ public: void addArg(SparseMatrixArg& arg) { args_.push_back(&arg); } -private: + private: std::vector args_; // The BufferArg object is constructed and freed by BufferArgs. std::vector _args_; @@ -163,7 +163,7 @@ private: * If Function has more than one output, each output can have different modes. */ class FunctionBase { -public: + public: virtual ~FunctionBase() {} virtual void init(const FuncConfig& config) {} @@ -192,7 +192,7 @@ public: static ClassRegistrar funcRegistrar_; -protected: + protected: // numInputs_ and numOutputs_ represents the maximum // input and output supported by Function. // Some functions are optimized for input and output, diff --git a/paddle/function/FunctionTest.h b/paddle/function/FunctionTest.h index 56c3537b6a96c8042d172f8aca2163fa18c813c1..14003d2c885c8f846f9445ad8844869c9112816e 100644 --- a/paddle/function/FunctionTest.h +++ b/paddle/function/FunctionTest.h @@ -39,7 +39,7 @@ struct Allocator { // Copy argument1 to argument2 template class CopyArgument { -public: + public: void operator()(const BufferArg& arg1, BufferArg& arg2) { CHECK_EQ(arg1.valueType(), arg2.valueType()); CHECK_LE(arg1.shape().getElements(), arg2.shape().getElements()); @@ -95,7 +95,7 @@ public: */ template class Compare2Function { -public: + public: typedef typename test::Allocator::type Allocator1; typedef typename test::Allocator::type Allocator2; typedef typename Tensor::Vector Vector1; @@ -305,7 +305,7 @@ public: std::shared_ptr getFunction2() const { return function2_; } -protected: + protected: // only init cpu argument, gpu argument copy from cpu argument. void initArg(BufferArg& arg) { Vector1 vector(arg.shape().getElements(), (real*)arg.data()); @@ -381,7 +381,7 @@ protected: } } -protected: + protected: std::shared_ptr function1_; std::shared_ptr function2_; std::vector> func1Memory_; @@ -400,7 +400,7 @@ protected: class CpuGpuFuncCompare : public Compare2Function { -public: + public: CpuGpuFuncCompare(const std::string& name, const FuncConfig& config) : Compare2Function(name + "-CPU", name + "-GPU", config) {} diff --git a/paddle/function/GemmConvOp.cpp b/paddle/function/GemmConvOp.cpp index 2b7c6f9eab223c8d6a2107ff4605ac6e60295f7d..5b023e2c10e5040a28660d555efceb0e26b40d49 100644 --- a/paddle/function/GemmConvOp.cpp +++ b/paddle/function/GemmConvOp.cpp @@ -24,7 +24,7 @@ namespace paddle { */ template class GemmConvFunction : public ConvFunctionBase { -public: + public: void init(const FuncConfig& config) override { ConvFunctionBase::init(config); } @@ -136,7 +136,7 @@ public: */ template class GemmConvMobileFunction : public ConvFunctionBase { -public: + public: void init(const FuncConfig& config) override { ConvFunctionBase::init(config); } @@ -297,7 +297,7 @@ public: */ template class GemmConvGradInputFunction : public ConvFunctionBase { -public: + public: void init(const FuncConfig& config) override { ConvFunctionBase::init(config); } @@ -404,7 +404,7 @@ public: */ template class GemmConvGradFilterFunction : public ConvFunctionBase { -public: + public: void init(const FuncConfig& config) override { ConvFunctionBase::init(config); } diff --git a/paddle/function/Im2Col.h b/paddle/function/Im2Col.h index 6a0778700037c142d62fdb99667403ade806f7c1..e0ce6918a2a5324a396ade734945cf426b81ab56 100644 --- a/paddle/function/Im2Col.h +++ b/paddle/function/Im2Col.h @@ -70,7 +70,7 @@ enum ColFormat { kCFO = 0, kOCF = 1 }; */ template class Im2ColFunctor { -public: + public: void operator()(const T* imData, const TensorShape& imShape, T* colData, @@ -85,7 +85,7 @@ public: template class Col2ImFunctor { -public: + public: void operator()(T* imData, const TensorShape& imShape, const T* colData, @@ -100,7 +100,7 @@ public: template class Im2ColMobileFunctor { -public: + public: void operator()(const T* imData, const TensorShape& imShape, T* colData, diff --git a/paddle/function/Im2ColOp.cpp b/paddle/function/Im2ColOp.cpp index ad2aed8f3c237cf9c0f7f0dcc4900cac807e25ea..55a3ff98db63ede96094a3d3fdeedf03b573294f 100644 --- a/paddle/function/Im2ColOp.cpp +++ b/paddle/function/Im2ColOp.cpp @@ -23,7 +23,7 @@ namespace paddle { */ template class Im2ColFunctor { -public: + public: void operator()(const T* imData, const TensorShape& imShape, T* colData, @@ -75,7 +75,7 @@ public: */ template class Col2ImFunctor { -public: + public: void operator()(T* imData, const TensorShape& imShape, const T* colData, @@ -130,7 +130,7 @@ template class Col2ImFunctor; */ template class Im2ColFunctor { -public: + public: void operator()(const T* imData, const TensorShape& imShape, T* colData, @@ -188,7 +188,7 @@ public: */ template class Col2ImFunctor { -public: + public: void operator()(T* imData, const TensorShape& imShape, const T* colData, diff --git a/paddle/function/Im2ColOpGpu.cu b/paddle/function/Im2ColOpGpu.cu index a944a0ee687fefc5e002096b9c5b869495554167..96dd8f528eaa38f9d174ab7c2a5ea5eb96e2a060 100644 --- a/paddle/function/Im2ColOpGpu.cu +++ b/paddle/function/Im2ColOpGpu.cu @@ -71,7 +71,7 @@ __global__ void im2col(const T* data_im, */ template class Im2ColFunctor { -public: + public: void operator()(const T* imData, const TensorShape& imShape, T* colData, @@ -184,7 +184,7 @@ __global__ void col2im(size_t n, */ template class Col2ImFunctor { -public: + public: void operator()(T* imData, const TensorShape& imShape, const T* colData, @@ -292,7 +292,7 @@ __global__ void im2colOCF(const T* imData, */ template class Im2ColFunctor { -public: + public: void operator()(const T* imData, const TensorShape& imShape, T* colData, @@ -399,7 +399,7 @@ __global__ void col2imOCF(T* imData, */ template class Col2ImFunctor { -public: + public: void operator()(T* imData, const TensorShape& imShape, const T* colData, diff --git a/paddle/function/MulOp.cpp b/paddle/function/MulOp.cpp index 90cd4a2b6d1bfb2529e1c966cf7a1fb904a844d7..7bf36c8050a8c33d836ce98dc7f3cf6d3de38d55 100644 --- a/paddle/function/MulOp.cpp +++ b/paddle/function/MulOp.cpp @@ -240,7 +240,7 @@ void MulOp(CpuMatrix& out, */ template class MulFunc : public FunctionBase { -public: + public: void init(const FuncConfig& config) override { aTrans_ = config.get("aTrans"); bTrans_ = config.get("bTrans"); @@ -335,7 +335,7 @@ public: } } -private: + private: bool aTrans_; bool bTrans_; }; diff --git a/paddle/function/NaiveConvOp.cpp b/paddle/function/NaiveConvOp.cpp index 22d3b33d0f4a730691234c6c742978abd72294a6..99c8b81acbbb16a91bc0faa1c7f2873fa94ab108 100644 --- a/paddle/function/NaiveConvOp.cpp +++ b/paddle/function/NaiveConvOp.cpp @@ -24,7 +24,7 @@ namespace paddle { */ template class NaiveConvFunctor { -public: + public: void operator()(const T* inputData, size_t batchSize, size_t inputChannels, @@ -85,7 +85,7 @@ public: template class NaiveConvFunction : public ConvFunctionBase { -public: + public: void init(const FuncConfig& config) override { ConvFunctionBase::init(config); } diff --git a/paddle/function/PadOp.cpp b/paddle/function/PadOp.cpp index db6dd518ca5df9d852e545b37f61f1141c81f57c..5d7515e8c053439b95fb18de3c8ffe70705600a3 100644 --- a/paddle/function/PadOp.cpp +++ b/paddle/function/PadOp.cpp @@ -132,7 +132,7 @@ static inline PadConf castToPadConf(const FuncConfig& conf) { template class PadFunc : public FunctionBase { -public: + public: void init(const FuncConfig& config) override { pad_ = castToPadConf(config); } void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { @@ -157,7 +157,7 @@ public: pad_); } -private: + private: PadConf pad_; }; @@ -173,7 +173,7 @@ private: template class PadGradFunc : public FunctionBase { -public: + public: void init(const FuncConfig& config) override { pad_ = castToPadConf(config); } void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { @@ -201,7 +201,7 @@ public: pad_); } -private: + private: PadConf pad_; }; diff --git a/paddle/function/RowConvOp.cpp b/paddle/function/RowConvOp.cpp index 925860346e1a53065b0fe4ccbd26853afc8898a1..129e9334582fad011c259e8ab8268b00a7fab7b6 100644 --- a/paddle/function/RowConvOp.cpp +++ b/paddle/function/RowConvOp.cpp @@ -129,7 +129,7 @@ void RowConvGrad(const CpuMatrix& outG, template class RowConvFunc : public FunctionBase { -public: + public: void init(const FuncConfig& config) override {} void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { @@ -176,7 +176,7 @@ public: template class RowConvGradFunc : public FunctionBase { // TODO(qingqing): split into RowConvDataFunc and RowConvWeightFunc -public: + public: void init(const FuncConfig& config) override {} void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { diff --git a/paddle/function/RowConvOpGpu.cu b/paddle/function/RowConvOpGpu.cu index 9d8a6d80bb2c95a9aa8eb3666fd637402c932559..f820ee9a9713ce17547aa03945dc3c291ef50a59 100644 --- a/paddle/function/RowConvOpGpu.cu +++ b/paddle/function/RowConvOpGpu.cu @@ -12,8 +12,8 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include "RowConvOp.h" -#include "hl_base.h" +#include "paddle/cuda/include/hl_base.h" +#include "paddle/function/RowConvOp.h" namespace paddle { @@ -94,7 +94,7 @@ __global__ void KeRowConv2(real* y, } template <> -void RowConv(GpuMatrix& out, +void RowConv(GpuMatrix& out, // NOLINT const GpuMatrix& in, const GpuMatrix& filter, const GpuIVector& seq) { @@ -144,6 +144,10 @@ __global__ void KeRowConvBwWeight(real* dw, } __syncthreads(); + // NOTE(zcd): temporary solution + unsigned mask = 0u; + CREATE_SHFL_MASK(mask, true); + for (int i = 0; i < numSeq; ++i) { const int start = starts[i]; const int end = starts[i + 1]; @@ -170,11 +174,10 @@ __global__ void KeRowConvBwWeight(real* dw, real val = sh_x[tidy][tidx] * sh_dy[tidy][tidx + context - 1 - t]; __syncthreads(); // warp size and blockDim.x is 32. - val += __shfl_down(val, 16); - val += __shfl_down(val, 8); - val += __shfl_down(val, 4); - val += __shfl_down(val, 2); - val += __shfl_down(val, 1); + + for (int offset = 16; offset > 0; offset /= 2) + val += __shfl_down_sync(mask, val, offset); + __syncthreads(); if (tidx == 0) { sh_dw[t][tidy] += val; @@ -205,6 +208,10 @@ __global__ void KeRowConvBwWeight2(real* dw, __shared__ real sh_x[BLOCK_H][BLOCK_W]; __shared__ real sh_dy[BLOCK_H][BLOCK_W]; + // NOTE(zcd): temporary solution + unsigned mask = 0u; + CREATE_SHFL_MASK(mask, true); + for (int i = 0; i < numSeq; ++i) { const int start = starts[i]; const int end = starts[i + 1]; @@ -230,11 +237,9 @@ __global__ void KeRowConvBwWeight2(real* dw, real val = sh_x[tidy][tidx] * sh_dy[tidy][tidx]; __syncthreads(); // warp size and blockDim.x is 32. - val += __shfl_down(val, 16); - val += __shfl_down(val, 8); - val += __shfl_down(val, 4); - val += __shfl_down(val, 2); - val += __shfl_down(val, 1); + for (int offset = 16; offset > 0; offset /= 2) + val += __shfl_down_sync(mask, val, offset); + __syncthreads(); if (tidx == 0 && (gidx + tidy) < width) { @@ -323,8 +328,8 @@ template <> void RowConvGrad(const GpuMatrix& outG, const GpuMatrix& in, const GpuMatrix& filter, - GpuMatrix& inG, - GpuMatrix& filterG, + GpuMatrix& inG, // NOLINT + GpuMatrix& filterG, // NOLINT const GpuIVector& seq) { const size_t numSeq = seq.getSize() - 1; const size_t contextLength = filter.getHeight(); diff --git a/paddle/function/ScaleSubRegionOp.cpp b/paddle/function/ScaleSubRegionOp.cpp index 6ed6eb2dba477722664ca4a29f4689114f368846..9a06ef2a96f25b5b7326049df2a708637f319561 100644 --- a/paddle/function/ScaleSubRegionOp.cpp +++ b/paddle/function/ScaleSubRegionOp.cpp @@ -92,7 +92,7 @@ void ScaleSubRegionGrad(const real* inGrad, */ template class ScaleSubRegionFunc : public FunctionBase { -public: + public: void init(const FuncConfig& config) override { conf_ = config; } void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { @@ -109,7 +109,7 @@ public: conf_); } -private: + private: FuncConfig conf_; }; @@ -124,7 +124,7 @@ private: template class ScaleSubRegionGradFunc : public FunctionBase { -public: + public: void init(const FuncConfig& config) override { conf_ = config; } void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { @@ -141,7 +141,7 @@ public: conf_); } -private: + private: FuncConfig conf_; }; diff --git a/paddle/function/SwitchOp.cpp b/paddle/function/SwitchOp.cpp index 50e1d6c04c54fed5b847aa10dbb253f00cfa42d4..750fb6bf28baf050b1f9f965a1a9b315363e5645 100644 --- a/paddle/function/SwitchOp.cpp +++ b/paddle/function/SwitchOp.cpp @@ -75,7 +75,7 @@ void NHWC2NCHW(real* outputs, */ template class NCHW2NHWCFunc : public FunctionBase { -public: + public: void init(const FuncConfig& config) override {} void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { @@ -108,7 +108,7 @@ public: */ template class NHWC2NCHWFunc : public FunctionBase { -public: + public: void init(const FuncConfig& config) override {} void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { diff --git a/paddle/function/TensorShape.h b/paddle/function/TensorShape.h index 02d38c32c007325a928910d136d48214ba5f6bc3..d4d1eae3960c333a2a7dc6099ae7a68677fdcd5f 100644 --- a/paddle/function/TensorShape.h +++ b/paddle/function/TensorShape.h @@ -22,7 +22,7 @@ namespace paddle { * TensorShape used to represent shape of normal tensor. */ class TensorShape { -public: + public: TensorShape() : ndims_(0), nelements_(0) { initDims(0); } TensorShape(size_t ndims) : ndims_(ndims), nelements_(1) { initDims(ndims); }; @@ -80,7 +80,7 @@ public: bool operator!=(const TensorShape& t) const { return !(*this == t); } -private: + private: // compute number of elements void numElements() { nelements_ = 1; diff --git a/paddle/function/neon/NeonDepthwiseConv.cpp b/paddle/function/neon/NeonDepthwiseConv.cpp index d3298c753853ca6d212a619cf8d0bd9356a8dbd7..d7ac83da41aaba5cd38b042d0381dea527f9c42d 100644 --- a/paddle/function/neon/NeonDepthwiseConv.cpp +++ b/paddle/function/neon/NeonDepthwiseConv.cpp @@ -21,7 +21,7 @@ namespace paddle { template class NeonDepthwiseConvFunction : public ConvFunctionBase { -public: + public: void init(const FuncConfig& config) override { ConvFunctionBase::init(config); } @@ -66,18 +66,18 @@ public: float* inputPadding = inputData; int padInputHeight = inputHeight + 2 * paddingH(); int padInputWidth = inputWidth + 2 * paddingW(); - if (paddingH() > 0 || paddingW() > 0) { - int newSize = batchSize * inputChannels * padInputHeight * padInputWidth; - resizeBuffer(newSize); - inputPadding = reinterpret_cast(memory_->getBuf()); - neon::Padding::run(inputData, - inputPadding, - batchSize * inputChannels, - inputHeight, - inputWidth, - padInputHeight, - padInputWidth); - } + int newSize = + batchSize * (inputChannels + 1) * padInputHeight * padInputWidth; + + resizeBuffer(newSize); + inputPadding = reinterpret_cast(memory_->getBuf()); + neon::Padding::run(inputData, + inputPadding, + batchSize * inputChannels, + inputHeight, + inputWidth, + padInputHeight, + padInputWidth); std::function diff --git a/paddle/function/neon/NeonDepthwiseConvTranspose.cpp b/paddle/function/neon/NeonDepthwiseConvTranspose.cpp index d443d3fa4902f998230651c5c64355d93c4c4f6a..1fc5daf6078bbd5b4506ff2e0832e2cc3ec48fe3 100644 --- a/paddle/function/neon/NeonDepthwiseConvTranspose.cpp +++ b/paddle/function/neon/NeonDepthwiseConvTranspose.cpp @@ -21,7 +21,7 @@ namespace paddle { template class NeonDepthwiseConvTransposeFunction : public ConvFunctionBase { -public: + public: void init(const FuncConfig& config) override { ConvFunctionBase::init(config); } diff --git a/paddle/function/nnpack/NNPACKConvOp.cpp b/paddle/function/nnpack/NNPACKConvOp.cpp index 3cdba4f2ed0dad42035fe2d0de87ad5aeeef20ca..48c997b50d8c73b25c58801c30e597c9d1f3232a 100644 --- a/paddle/function/nnpack/NNPACKConvOp.cpp +++ b/paddle/function/nnpack/NNPACKConvOp.cpp @@ -46,7 +46,7 @@ nnp_convolution_algorithm get_nnp_convolution_algorithm( template class NNPACKConvFunction : public ConvFunctionBase { -public: + public: void init(const FuncConfig& config) override { ConvFunctionBase::init(config); algorithm_ = get_nnp_convolution_algorithm(config.get("algo")); @@ -231,7 +231,7 @@ public: } } -private: + private: nnp_convolution_algorithm algorithm_; nnp_convolution_transform_strategy transform_strategy_; void* workspaceBuffer_; diff --git a/paddle/gserver/CMakeLists.txt b/paddle/gserver/CMakeLists.txt index 3d6ced713f00bd72622d8aeed3967642b6774ffe..6dc877dd90ee2ae3d99406299a9244eb3e3d7b53 100644 --- a/paddle/gserver/CMakeLists.txt +++ b/paddle/gserver/CMakeLists.txt @@ -146,8 +146,6 @@ else() ${GSERVER_SOURCES}) endif() -add_style_check_target(paddle_gserver ${GSERVER_SOURCES}) -add_style_check_target(paddle_gserver ${GSERVER_HEADER}) add_dependencies(paddle_gserver paddle_proto ${external_project_dependencies}) if(WITH_TESTING) add_subdirectory(tests) diff --git a/paddle/gserver/activations/ActivationFunction.cpp b/paddle/gserver/activations/ActivationFunction.cpp index 8d8f01234fe3859989e44fe6147105fb72b832ff..71c238fbfe9f32f3764601ebb441336931f8ef5f 100644 --- a/paddle/gserver/activations/ActivationFunction.cpp +++ b/paddle/gserver/activations/ActivationFunction.cpp @@ -44,10 +44,10 @@ static ClassRegistrar gActivationRegistrar; */ #define BEGIN_DEFINE_ACTIVATION(ACTIVATION_NAME) \ class ACTIVATION_CLASS_NAME(ACTIVATION_NAME) : public ActivationFunction { \ - private: \ + private: \ static const std::string name; \ \ - public: \ + public: \ const std::string& getName() const { return name; } /** * @def END_DEFINE_ACTIVATION @@ -70,7 +70,7 @@ static ClassRegistrar gActivationRegistrar; * Do nothing when forward/backward. */ class IdentityActivation : public ActivationFunction { -public: + public: static const std::string name; Error __must_check forward(Argument& act) { (void)act; diff --git a/paddle/gserver/activations/ActivationFunction.h b/paddle/gserver/activations/ActivationFunction.h index 0f4b0fe0abb85403d42fc8a2ac28560e10058c20..8e2e144769f2e668a9a8f02890d29c4a7fe128a3 100644 --- a/paddle/gserver/activations/ActivationFunction.h +++ b/paddle/gserver/activations/ActivationFunction.h @@ -31,7 +31,7 @@ struct Argument; * */ class ActivationFunction { -public: + public: static ActivationFunction* create(const std::string& type); static std::vector getAllRegisteredTypes(); diff --git a/paddle/gserver/activations/MKLDNNActivation.cpp b/paddle/gserver/activations/MKLDNNActivation.cpp index 56ffb839344aabe43eaae0bd46e6dbf95e4d8f20..672444c6561adbeb78c3c453f12ab6aaedeed646 100644 --- a/paddle/gserver/activations/MKLDNNActivation.cpp +++ b/paddle/gserver/activations/MKLDNNActivation.cpp @@ -35,10 +35,10 @@ static ClassRegistrar gMKLDNNActivationRegistrar; * @def END_MKLDNN_ACTIVATION */ #define END_MKLDNN_ACTIVATION(ACT_TYPE) \ -private: \ + private: \ static const std::string name; \ \ -public: \ + public: \ const std::string& getName() const { return name; } \ } \ ; \ @@ -63,11 +63,11 @@ public: \ #define DEFINE_MKLDNN_ELTWISE_ACTIVATION( \ ACT_TYPE, BASE_CLASS, ALPHA, BWD_ALPHA) \ BEGIN_MKLDNN_ACTIVATION(ACT_TYPE, BASE_CLASS) \ -private: \ + private: \ static const float alpha; \ static const float bwdAlpha; \ \ -public: \ + public: \ float getAlpha() const { return alpha; } \ float getBwdAlpha() const { return bwdAlpha; } \ END_MKLDNN_ACTIVATION(ACT_TYPE) \ diff --git a/paddle/gserver/activations/MKLDNNActivation.h b/paddle/gserver/activations/MKLDNNActivation.h index 392b32c70dae3728e13ee64f09f135c015c122cf..eece1b9c37e72624dffd119804c65f7bd36e20fb 100644 --- a/paddle/gserver/activations/MKLDNNActivation.h +++ b/paddle/gserver/activations/MKLDNNActivation.h @@ -27,7 +27,7 @@ namespace paddle { * including mkldnn_relu, mkldnn_elu, mkldnn_tanh, mkldnn_softmax */ class MKLDNNActivation : public ActivationFunction { -protected: + protected: // input value element count size_t cnt_; // should not merge the resetBwd into resetFwd, @@ -43,7 +43,7 @@ protected: std::vector pipelineFwd_; std::vector pipelineBwd_; -public: + public: MKLDNNActivation() : cnt_(0), needResetBwd_(true) {} ~MKLDNNActivation() {} static ActivationFunction* create(const std::string& type); @@ -72,7 +72,7 @@ class MKLDNNEltwiseActivation : public MKLDNNActivation { typedef mkldnn::eltwise_backward eltwise_bwd; typedef mkldnn::algorithm algorithm; -protected: + protected: // save the forward primitive desc, which can be used backward std::shared_ptr fwdPD_; // eltwise_bwd need src input value @@ -80,7 +80,7 @@ protected: // use for copy data std::shared_ptr copyInVal_; -public: + public: MKLDNNEltwiseActivation() {} ~MKLDNNEltwiseActivation() {} virtual const std::string& getName() const = 0; @@ -102,12 +102,12 @@ public: class MKLDNNSoftmaxActivation : public MKLDNNActivation { typedef mkldnn::softmax_forward softmax_fwd; -private: + private: // for backward MatrixPtr sftMaxSum_; MatrixPtr sftMaxDot_; -public: + public: MKLDNNSoftmaxActivation() {} ~MKLDNNSoftmaxActivation() {} virtual const std::string& getName() const = 0; diff --git a/paddle/gserver/dataproviders/DataProvider.h b/paddle/gserver/dataproviders/DataProvider.h index 4851168abab7179d552648c88923a529d55e6a7e..21822b10c2ebf1d353195794cf8f49e02b64c177 100644 --- a/paddle/gserver/dataproviders/DataProvider.h +++ b/paddle/gserver/dataproviders/DataProvider.h @@ -71,7 +71,7 @@ typedef std::shared_ptr BufferBatchPtr; * @brief Data for batch training a neural network */ class DataBatch { -public: + public: DataBatch() : size_(0) { data_.clear(); } /** * @brief Get batch size @@ -181,7 +181,7 @@ public: } } -protected: + protected: /** * @brief batch size */ @@ -194,7 +194,7 @@ protected: }; class BufferBatch { -public: + public: BufferBatch() { hlStream_ = HPPL_STREAM_DEFAULT; hlEvent_ = NULL; @@ -235,7 +235,7 @@ public: void swap(BufferBatch* bufBatch); void clone(DataBatch* srcBatch, bool useGpu); -protected: + protected: DataBatch* batchData_; hl_stream_t hlStream_; hl_event_t hlEvent_; @@ -247,7 +247,7 @@ typedef std::shared_ptr DataProviderPtr; typedef Queue BufferBatchQueue; class DoubleBuffer { -public: + public: DoubleBuffer(DataProvider* dataPool, bool useGpu, int64_t batchSize = 0); virtual ~DoubleBuffer(); void removeOneBatch(DataBatch* dataBatch); @@ -267,7 +267,7 @@ public: void setPending(bool pending) { pending_ = pending; } -protected: + protected: virtual void asyncLoadBatch(); void insertOneBatch(DataBatch* batch); @@ -290,7 +290,7 @@ protected: * one is for input, one is for label. */ class DataProvider { -public: + public: static ClassRegistrar registrar_; static DataProvider* create(const DataConfig& config, const ModelConfig& modelConfig, @@ -359,7 +359,7 @@ public: */ virtual int64_t getNextBatchInternal(int64_t size, DataBatch* batch) = 0; -protected: + protected: DataConfig config_; bool skipShuffle_; float usageRatio_; @@ -382,7 +382,7 @@ protected: * necessary configurations such as stream_names */ class DummyDataProvider : public DataProvider { -public: + public: DummyDataProvider(const DataConfig& config, bool useGpu) : DataProvider(config, useGpu) {} virtual void shuffle() {} @@ -399,7 +399,7 @@ public: * Data provider for one input and one integer label. */ class SimpleDataProviderBase : public DataProvider { -protected: + protected: /// sample feature dimension int64_t sampleDim_; /// the number of samples @@ -425,7 +425,7 @@ protected: RWLock lock_; -public: + public: SimpleDataProviderBase(const DataConfig& config, bool useGpu, bool withInfo); ~SimpleDataProviderBase() {} @@ -440,7 +440,7 @@ public: /// return the number of samples in the buffer int64_t fillBuffer(); -protected: + protected: /** * @brief Fill at most size samples into data and label. * @@ -458,12 +458,12 @@ protected: }; class SimpleDataProvider : public SimpleDataProviderBase { -public: + public: SimpleDataProvider(const DataConfig& config, bool useGpu); ~SimpleDataProvider(); virtual void reset(); -protected: + protected: void loadData(const std::string& fileName); void loadDataFile(const std::string& fileName); virtual int64_t fillBufferImp(real* data, @@ -471,7 +471,7 @@ protected: int* info, int64_t size); -protected: + protected: size_t currentSampleIndex_; std::vector labels_; std::vector data_; diff --git a/paddle/gserver/dataproviders/DataProviderGroup.h b/paddle/gserver/dataproviders/DataProviderGroup.h index 768e54fe82bedd6faca5ad9eb2b6f2ee0017dc3d..91c94dc986c7aeb70df25511ce14a5f9c312a159 100644 --- a/paddle/gserver/dataproviders/DataProviderGroup.h +++ b/paddle/gserver/dataproviders/DataProviderGroup.h @@ -20,7 +20,7 @@ namespace paddle { template class DataProviderGroup : public DataProvider { -protected: + protected: typedef T ProviderType; typedef std::shared_ptr ProviderPtrType; ProviderPtrType provider_; @@ -29,7 +29,7 @@ protected: std::mutex lock_; std::unique_ptr> loader_; -public: + public: DataProviderGroup(const DataConfig& config, bool useGpu); ~DataProviderGroup() {} @@ -38,7 +38,7 @@ public: virtual int64_t getSize() { return -1; } virtual int64_t getNextBatchInternal(int64_t size, DataBatch* batch); -private: + private: void startLoader(); void stopLoader(); void forceStopLoader(); diff --git a/paddle/gserver/dataproviders/MultiDataProvider.h b/paddle/gserver/dataproviders/MultiDataProvider.h index 9a863c896773d71a99e21660fc13e3dd477a0c12..baa1fc019002f86414c9c45734ad65cda916d457 100644 --- a/paddle/gserver/dataproviders/MultiDataProvider.h +++ b/paddle/gserver/dataproviders/MultiDataProvider.h @@ -19,10 +19,10 @@ limitations under the License. */ namespace paddle { class MultiDataProvider : public DataProvider { -protected: + protected: std::vector> subDataProviders_; -public: + public: MultiDataProvider(const DataConfig& config, const ModelConfig& modelConfig, bool useGpu); @@ -33,7 +33,7 @@ public: virtual int64_t getNextBatchInternal(int64_t size, DataBatch* batch); bool isTestMode() const { return isTestMode_; } -private: + private: int totalDataRatio_; bool isTestMode_; }; diff --git a/paddle/gserver/dataproviders/ProtoReader.h b/paddle/gserver/dataproviders/ProtoReader.h index 786703f4dee4802bb967f9d15fb69ebcbc15d997..08d045226e1ebb014bdd91ebf0e8f0353179b0c8 100644 --- a/paddle/gserver/dataproviders/ProtoReader.h +++ b/paddle/gserver/dataproviders/ProtoReader.h @@ -28,7 +28,7 @@ namespace paddle { * messages from/to i/ostream. */ class ProtoReader { -public: + public: explicit ProtoReader(std::istream* s, bool dataCompression = false) { CHECK(s) << "istream pointer is nullptr"; istreamInput_.reset(new google::protobuf::io::IstreamInputStream(s)); @@ -109,7 +109,7 @@ public: return true; } -protected: + protected: std::unique_ptr istreamInput_; std::unique_ptr gzipInput_; std::unique_ptr codedInput_; @@ -144,7 +144,7 @@ protected: }; class ProtoWriter { -public: + public: explicit ProtoWriter(std::ostream* s, bool dataCompression = false) { CHECK(s) << "ostream pointer is nullptr"; ostreamOutput_.reset(new google::protobuf::io::OstreamOutputStream(s)); @@ -168,7 +168,7 @@ public: return ret; } -protected: + protected: std::unique_ptr ostreamOutput_; std::unique_ptr gzipOutput_; std::unique_ptr codedOutput_; diff --git a/paddle/gserver/dataproviders/PyDataProvider.h b/paddle/gserver/dataproviders/PyDataProvider.h index e53354c9e43ea9dc58fd4bd38a533025b6f17482..da50dd4e2ebb743ef45af319bc713ed7ac3d3e10 100644 --- a/paddle/gserver/dataproviders/PyDataProvider.h +++ b/paddle/gserver/dataproviders/PyDataProvider.h @@ -23,7 +23,7 @@ limitations under the License. */ namespace paddle { class PyDataProvider : public DataProvider { -public: + public: PyDataProvider(const DataConfig& config, bool useGpu, bool loadDataAll = true); @@ -40,7 +40,7 @@ public: virtual int64_t getNextBatchInternal(int64_t size, DataBatch* batch); -protected: + protected: struct ProtoSlot; // return false if each each sample is one sequence, i.e., independent // of other samples. @@ -73,7 +73,7 @@ protected: void resetSlots(); void loadData(const std::vector& fileList); -protected: + protected: struct ProtoSlot { SlotDef::SlotType type; int dim; diff --git a/paddle/gserver/dataproviders/PyDataProvider2.cpp b/paddle/gserver/dataproviders/PyDataProvider2.cpp index b4215bb307cc31ce64bb724986b88fdc20bbbf45..54ee091e8f257f76b113d4ca6f8a7c3989c0c1df 100644 --- a/paddle/gserver/dataproviders/PyDataProvider2.cpp +++ b/paddle/gserver/dataproviders/PyDataProvider2.cpp @@ -93,7 +93,7 @@ inline std::ostream& operator<<(std::ostream& os, const SlotHeader& header) { * prepare step, fill data into argument during fill step. */ class IFieldScanner { -public: + public: DISABLE_COPY(IFieldScanner); /** * Ctor. @@ -146,7 +146,7 @@ public: */ static IFieldScanner* create(SlotHeader* header); -protected: + protected: SlotHeader* headerPtr_; }; @@ -154,7 +154,7 @@ protected: * Py Data Provider Cache Interface. */ class IPyDataProviderCache { -public: + public: virtual ~IPyDataProviderCache() {} /** @@ -193,7 +193,7 @@ public: * data. And it support cache strategies. */ class PyDataProvider2 : public DataProvider { -public: + public: /** * Ctor */ @@ -234,7 +234,7 @@ public: */ virtual ~PyDataProvider2() { resetImpl(false); } -private: + private: void createPyDataObj(const std::string& model, const std::string& className, const std::string& fileListName, @@ -435,7 +435,7 @@ private: exit_ = false; } -private: + private: std::unique_ptr loadThread_; std::atomic exit_; std::deque callingContexts_; @@ -461,7 +461,7 @@ private: static PyObjectPtr zeroTuple_; class PositionRandom { - public: + public: inline explicit PositionRandom(bool skipRand) : eng_(ThreadLocalRandomEngine::get()), skipRand_(skipRand) {} @@ -476,14 +476,14 @@ private: } } - private: + private: std::default_random_engine& eng_; std::unique_ptr> dist_; bool skipRand_; }; // DataProvider interface -public: + public: /** * Resetting the PyDataProvider. May start reading thread here. */ @@ -666,7 +666,7 @@ REGISTER_DATA_PROVIDER_EX(py2, PyDataProvider2); * Scanner for dense slot. */ class DenseScanner : public IFieldScanner { -public: + public: explicit DenseScanner(SlotHeader* ptr) : IFieldScanner(ptr), height_(0) {} /** @@ -708,7 +708,7 @@ public: ++height_; } -private: + private: size_t height_; }; @@ -716,7 +716,7 @@ private: * Scanner for index slot */ class IndexScanner : public IFieldScanner { -public: + public: explicit IndexScanner(SlotHeader* ptr) : IFieldScanner(ptr), cnt_(0) {} /** @@ -740,12 +740,12 @@ public: CHECK(ok) << "Cannot cast int " << py::repr(obj); } -private: + private: size_t cnt_; }; class SparseNonValueScanner : public IFieldScanner { -public: + public: explicit SparseNonValueScanner(SlotHeader* ptr) : IFieldScanner(ptr), nnz_(0), height_(0) {} @@ -790,7 +790,7 @@ public: ++height_; } -protected: + protected: /** * Set a single sparse index and value. * @param [out] col sparse index @@ -809,7 +809,7 @@ protected: }; class SparseValueScanner : public SparseNonValueScanner { -public: + public: explicit SparseValueScanner(SlotHeader* ptr) : SparseNonValueScanner(ptr) {} virtual void finishPrepare(Argument& argument) { @@ -817,7 +817,7 @@ public: argument.value, height_, headerPtr_->dim, nnz_, FLOAT_VALUE); } -protected: + protected: virtual void setData(int* col, real* dat, PyObject* obj) { py::SequenceHelper s(obj); SparseNonValueScanner::setData(col, dat, s[0]); @@ -829,7 +829,7 @@ protected: * Sequence Scanner. Scanner for sequence or sub-sequence. */ class SequenceScanner : public IFieldScanner { -public: + public: /** * Ctor * @param innerScanner inner scanner for each timestep or sub-sequence. @@ -902,7 +902,7 @@ public: */ virtual void finishFill(Argument& argument) { inner_->finishFill(argument); } -protected: + protected: size_t getSize(PyObject* obj) { py::SequenceHelper s(obj); auto sc = dynamic_cast(inner_.get()); @@ -917,7 +917,7 @@ protected: } } -private: + private: std::unique_ptr inner_; size_t cnt_; std::function getSeqStartPos_; @@ -969,7 +969,7 @@ IFieldScanner* IFieldScanner::create(SlotHeader* header) { * python every pass. */ class NoCacheStrategy : public IPyDataProviderCache { -public: + public: virtual bool reset() { return true; } virtual void drop(std::deque* data) { data->clear(); } @@ -984,7 +984,7 @@ public: * The rest passes, will load data from memory. */ class CacheOnePassInMemory : public IPyDataProviderCache { -public: + public: CacheOnePassInMemory() : objPool_(new std::deque()), droppedPool_(new std::deque()) {} @@ -1011,7 +1011,7 @@ public: virtual std::deque* load() { return objPool_.get(); } -private: + private: std::unique_ptr> objPool_; std::unique_ptr> droppedPool_; }; diff --git a/paddle/gserver/evaluators/CTCErrorEvaluator.cpp b/paddle/gserver/evaluators/CTCErrorEvaluator.cpp index 0f680de776f4755ca5fe83c86ea759d88f93ed01..c6cd41de9a1a22470d8659eb90d1ac2b075b2df9 100644 --- a/paddle/gserver/evaluators/CTCErrorEvaluator.cpp +++ b/paddle/gserver/evaluators/CTCErrorEvaluator.cpp @@ -22,7 +22,7 @@ namespace paddle { * calculate sequence-to-sequence edit distance */ class CTCErrorEvaluator : public Evaluator { -private: + private: MatrixPtr outActivations_; int numTimes_, numClasses_, numSequences_, blank_; real deletions_, insertions_, substitutions_; @@ -197,7 +197,7 @@ private: (real)seqClassficationError_ / numSequences_; } -public: + public: CTCErrorEvaluator() : numTimes_(0), numClasses_(0), diff --git a/paddle/gserver/evaluators/ChunkEvaluator.cpp b/paddle/gserver/evaluators/ChunkEvaluator.cpp index 755b91d05caf33745e66415e7b111ba348c575d9..a2216293b1ab3a32e9cc903b805ca0aca10d58c1 100644 --- a/paddle/gserver/evaluators/ChunkEvaluator.cpp +++ b/paddle/gserver/evaluators/ChunkEvaluator.cpp @@ -77,7 +77,7 @@ class ChunkEvaluator : public Evaluator { std::set excludedChunkTypes_; mutable std::unordered_map values_; -public: + public: virtual void init(const EvaluatorConfig& config) { Evaluator::init(config); if (config.chunk_scheme() == "IOB") { @@ -276,7 +276,7 @@ public: return "chunk"; } -private: + private: void storeLocalValues() const { CHECK_GE(numOutputSegments_, 0); CHECK_GE(numLabelSegments_, 0); diff --git a/paddle/gserver/evaluators/DetectionMAPEvaluator.cpp b/paddle/gserver/evaluators/DetectionMAPEvaluator.cpp index f43ef5dd51407236a3a36b300b33f92a9fad885a..ddb8ebca784db4a83c328ff75f5c50c7aecd7352 100644 --- a/paddle/gserver/evaluators/DetectionMAPEvaluator.cpp +++ b/paddle/gserver/evaluators/DetectionMAPEvaluator.cpp @@ -28,7 +28,7 @@ namespace paddle { * The config file api is detection_map_evaluator. */ class DetectionMAPEvaluator : public Evaluator { -public: + public: DetectionMAPEvaluator() : evaluateDifficult_(false), cpuOutput_(nullptr), cpuLabel_(nullptr) {} @@ -132,7 +132,7 @@ public: LOG(FATAL) << "Distribute detection evaluation not implemented."; } -protected: + protected: void calcTFPos(const size_t batchSize, const vector>>& allGTBBoxes, const vector>>>& @@ -287,7 +287,7 @@ protected: real getValueImpl() const { return calcMAP(); } -private: + private: real overlapThreshold_; // overlap threshold when determining whether matched bool evaluateDifficult_; // whether evaluate difficult ground truth size_t backgroundId_; // class index of background diff --git a/paddle/gserver/evaluators/Evaluator.cpp b/paddle/gserver/evaluators/Evaluator.cpp index 79478e7fac63a49c494105d53a6944b4b89e6c63..941fb8fb539d58cca22ecf563d2effa816243c3b 100644 --- a/paddle/gserver/evaluators/Evaluator.cpp +++ b/paddle/gserver/evaluators/Evaluator.cpp @@ -38,7 +38,7 @@ void Evaluator::eval(const NeuralNetwork& nn) { * The config file api is classification_error_evaluator. */ class ClassificationErrorEvaluator : public Evaluator { -public: + public: /* ClassificationErrorEvaluator() : totalScore2_(0) {} @@ -124,7 +124,7 @@ public: } // Evaluator interface -protected: + protected: std::string getTypeImpl() const { return "classification_error"; } }; @@ -135,7 +135,7 @@ protected: */ class SequenceClassificationErrorEvaluator : public ClassificationErrorEvaluator { -public: + public: virtual void updateSamplesNum(const std::vector& arguments) { numSamples_ += arguments[0].getNumSequences(); } @@ -166,7 +166,7 @@ public: } // Evaluator interface -protected: + protected: std::string getTypeImpl() const { return "seq_classification_error"; } }; REGISTER_EVALUATOR(seq_classification_error, @@ -178,7 +178,7 @@ REGISTER_EVALUATOR(seq_classification_error, * The config file api is sum_evaluator. */ class SumEvaluator : public Evaluator { -public: + public: SumEvaluator() : cpuLabel_(nullptr), cpuWeight_(nullptr) {} virtual void updateSamplesNum(const std::vector& arguments) { @@ -255,12 +255,12 @@ public: mergeResultsOfAllClients(client); } -private: + private: IVectorPtr cpuLabel_; MatrixPtr cpuWeight_; // Evaluator interface -protected: + protected: std::string getTypeImpl() const { return "sum"; } }; /** @@ -274,7 +274,7 @@ protected: * */ class ColumnSumEvaluator : public Evaluator { -public: + public: explicit ColumnSumEvaluator(int32_t colIdx) : colIdx_(colIdx), colNum_(0), sum_(nullptr) {} @@ -368,13 +368,13 @@ public: client->reduce(&numSamples_, &numSamples_, 1, FLAGS_trainer_id, 0); } -private: + private: int32_t colIdx_; size_t colNum_; MatrixPtr sum_; /* cpu matrix */ // Evaluator interface -protected: + protected: std::string getTypeImpl() const { if (colIdx_ == -1) return "last-column-sum"; @@ -1018,7 +1018,7 @@ static InitFunction __reg_type_auc_sum__([]() { * The config file api is value_printer_evaluator. */ class ValuePrinter : public NotGetableEvaluator { -public: + public: virtual void eval(const NeuralNetwork& nn) { for (const std::string& name : config_.input_layers()) { nn.getLayer(name)->getOutput().printValueString(LOG(INFO), @@ -1038,7 +1038,7 @@ REGISTER_EVALUATOR(value_printer, ValuePrinter); * The config file api is gradient_printer_evaluator. */ class GradientPrinter : public NotGetableEvaluator { -public: + public: virtual void eval(const NeuralNetwork& nn) { for (const std::string& name : config_.input_layers()) { const Argument& argu = nn.getLayer(name)->getOutput(); @@ -1061,11 +1061,11 @@ REGISTER_EVALUATOR(gradient_printer, GradientPrinter); * The config file api is maxid_printer_evaluator. */ class MaxIdPrinter : public NotGetableEvaluator { -private: + private: IVectorPtr maxIds_; MatrixPtr maxValues_; -public: + public: MaxIdPrinter() {} virtual void eval(const NeuralNetwork& nn) { @@ -1103,12 +1103,12 @@ REGISTER_EVALUATOR(max_id_printer, MaxIdPrinter); * The config file api is maxframe_printer_evaluator. */ class MaxFramePrinter : public NotGetableEvaluator { -private: + private: IVectorPtr maxIds_; MatrixPtr maxValues_; MatrixPtr value_; -public: + public: MaxFramePrinter() { value_ = Matrix::create(nullptr, /* height= */ 1, 1, /* trans= */ false, false); @@ -1190,7 +1190,7 @@ REGISTER_EVALUATOR(max_frame_printer, MaxFramePrinter); * */ class SequenceTextPrinter : public NotGetableEvaluator { -private: + private: /// dict_file, which contains a list of tokens std::vector dict_; /// result_file, which is the output file @@ -1203,7 +1203,7 @@ private: /// store the probability associated with each sequence std::vector cpuIn_; -public: + public: SequenceTextPrinter() {} virtual void init(const EvaluatorConfig& config) { @@ -1334,7 +1334,7 @@ REGISTER_EVALUATOR(seq_text_printer, SequenceTextPrinter); * The config file api is classification_error_printer_evaluator. */ class ClassificationErrorPrinter : public ClassificationErrorEvaluator { -public: + public: virtual void updateSamplesNum(const std::vector& arguments) {} virtual real evalImp(std::vector& arguments) { diff --git a/paddle/gserver/evaluators/Evaluator.h b/paddle/gserver/evaluators/Evaluator.h index be2032992c455fe2b442dbe05d84128ef8ebf82f..42948f1097d9a12600f4b11646a47e45b9bf4e96 100644 --- a/paddle/gserver/evaluators/Evaluator.h +++ b/paddle/gserver/evaluators/Evaluator.h @@ -40,7 +40,7 @@ class NeuralNetwork; * has been by a trained model. */ class Evaluator { -public: + public: static Evaluator* create(const EvaluatorConfig& config); Evaluator() : numSamples_(0), totalScore_(0) {} @@ -172,7 +172,7 @@ public: return this->getTypeImpl(); } -protected: + protected: /** * @brief getValueImpl The simplest way to define getValue result. If this * evaluator doesn't contain multiple fields, and do not throw any error, just @@ -191,7 +191,7 @@ protected: */ virtual std::string getTypeImpl() const { return "base"; } -protected: + protected: EvaluatorConfig config_; double numSamples_; double totalScore_; @@ -204,7 +204,7 @@ protected: */ class NotGetableEvaluator : public Evaluator { // Evaluator interface -public: + public: void getNames(std::vector* names) {} real getValue(const std::string& name, Error* err) const { @@ -219,7 +219,7 @@ public: }; class DummyEvaluator : public Evaluator { -public: + public: DummyEvaluator() {} virtual void init(const EvaluatorConfig&) {} virtual void start() {} @@ -232,7 +232,7 @@ public: virtual void printStats(std::ostream&) const {} // Evaluator interface -protected: + protected: std::string getTypeImpl() const; }; /** @@ -251,7 +251,7 @@ protected: * */ class AucEvaluator : public Evaluator { -public: + public: AucEvaluator(int32_t colIdx) : colIdx_(colIdx), realColumnIdx_(0), @@ -269,7 +269,7 @@ public: virtual void distributeEval(ParameterClient2* client); -private: + private: static const uint32_t kBinNum_ = (1 << 24) - 1; static const int kNegativeLabel_ = 0; double statPos_[kBinNum_ + 1]; @@ -292,7 +292,7 @@ private: double calcAuc() const; // Evaluator interface -protected: + protected: real getValueImpl() const; std::string getTypeImpl() const; }; @@ -305,7 +305,7 @@ protected: * dense value. */ class RankAucEvaluator : public Evaluator { -public: + public: // evaluate ranking AUC virtual void start(); @@ -317,7 +317,7 @@ public: mergeResultsOfAllClients(client); } -private: + private: MatrixPtr output_; MatrixPtr click_; MatrixPtr pv_; @@ -329,7 +329,7 @@ private: size_t size); // Evaluator interface -protected: + protected: std::string getTypeImpl() const; }; @@ -344,7 +344,7 @@ protected: * The config file api is precision_recall_evaluator. */ class PrecisionRecallEvaluator : public Evaluator { -public: + public: // Evaluate precision, recall and F1 score PrecisionRecallEvaluator() : isMultiBinaryLabel_(false), @@ -379,7 +379,7 @@ public: StatsInfo() : TP(0.0), TN(0.0), FP(0.0), FN(0.0) {} }; -private: + private: bool isMultiBinaryLabel_; std::vector statsInfo_; @@ -444,7 +444,7 @@ private: * The config file api is pnpair_evaluator. */ class PnpairEvaluator : public Evaluator { -public: + public: PnpairEvaluator() : cpuOutput_(nullptr), cpuLabel_(nullptr), @@ -491,7 +491,7 @@ public: << " calc total neg pair: " << pairArray_[1]; } -private: + private: static const uint32_t kPairArrayNum_ = 2; double pairArray_[kPairArrayNum_]; MatrixPtr cpuOutput_; @@ -500,7 +500,7 @@ private: MatrixPtr cpuWeight_; // Evaluator interface -protected: + protected: real getValueImpl() const { return pairArray_[0] / ((pairArray_[1] <= 0) ? 1.0 : pairArray_[1]); } diff --git a/paddle/gserver/gradientmachines/GradientMachine.h b/paddle/gserver/gradientmachines/GradientMachine.h index 60936c311d1b0119186c76d5c95b8819294446ce..22cf5d265f429ecbcea1808a54c85d7e89f8bc99 100644 --- a/paddle/gserver/gradientmachines/GradientMachine.h +++ b/paddle/gserver/gradientmachines/GradientMachine.h @@ -73,7 +73,7 @@ class GradientMachine; typedef std::shared_ptr GradientMachinePtr; class GradientMachine { -public: + public: enum CreateMode { kNormal = 0, kSgdSparseCpuTraining = 3, @@ -240,7 +240,7 @@ public: */ virtual void releaseOutput() {} -protected: + protected: virtual void onLoadParameter() {} std::vector parameters_; diff --git a/paddle/gserver/gradientmachines/GradientMachineMode.h b/paddle/gserver/gradientmachines/GradientMachineMode.h index 898b68fbbc329145109ad0ae4b97c872d4f9a37c..dd944a35f8952e354f8e4f3eb5c67b136c5f080e 100644 --- a/paddle/gserver/gradientmachines/GradientMachineMode.h +++ b/paddle/gserver/gradientmachines/GradientMachineMode.h @@ -19,14 +19,14 @@ limitations under the License. */ namespace paddle { class IGradientMachineMode { -public: + public: virtual ~IGradientMachineMode() {} -public: // interfaces - /** - * @brief create current mode's gradient machine by model config. - * @param config model config - */ + public: // interfaces + /** + * @brief create current mode's gradient machine by model config. + * @param config model config + */ virtual GradientMachine* create(const ModelConfig& config) = 0; /** @@ -55,14 +55,14 @@ public: // interfaces */ virtual bool needTrainWholeDataInOneBatch() const = 0; -public: // static methods. - /** - * @brief register a custom gradient machine mode. - * @note For user to register a custom gradient machine mode, id should >= - * kCustom. - * @param mode mode id. - * @param ptr mode description object. - */ + public: // static methods. + /** + * @brief register a custom gradient machine mode. + * @note For user to register a custom gradient machine mode, id should >= + * kCustom. + * @param mode mode id. + * @param ptr mode description object. + */ static void regGradientMachineMode( int32_t mode, std::unique_ptr&& ptr) { modes_.insert(std::make_pair(mode, std::move(ptr))); @@ -141,7 +141,7 @@ public: // static methods. } } -private: + private: static std::unordered_map> modes_; }; diff --git a/paddle/gserver/gradientmachines/MultiGradientMachine.h b/paddle/gserver/gradientmachines/MultiGradientMachine.h index 83d2651f34b3698848427f29b1a90e606e57950e..eff7d5284c6dd4898344203b50acc94ae61b4d59 100644 --- a/paddle/gserver/gradientmachines/MultiGradientMachine.h +++ b/paddle/gserver/gradientmachines/MultiGradientMachine.h @@ -166,7 +166,7 @@ struct GradBuffer { * the merged gradient to parameter server. */ class MultiGradientMachine : public GradientMachine { -public: + public: enum TaskType { TASK_FORWARD_BACKWARD = 0, TASK_FORWARD = 1, @@ -213,7 +213,7 @@ public: /// The gradietns will be copied to each thread in the computing threads. virtual void setOutputGrad(const std::vector& args); -protected: + protected: friend class TrainerThread; std::vector& getAllThreads() { return threads_; } @@ -281,7 +281,7 @@ protected: int paraMainThread(int pid) const { return paraMainThread_[pid]; } -protected: + protected: virtual void forwardImp(const std::vector& inArgs, std::vector* outArgs, PassType passType, @@ -298,7 +298,7 @@ protected: void allocGradBufs(); -protected: + protected: bool useGpu_; bool hasNonstaticCpuParamters_; @@ -342,7 +342,7 @@ protected: }; class TrainerThread { -public: + public: TrainerThread(const ModelConfig& config, int threadId, MultiGradientMachine* multiMachine); @@ -392,7 +392,7 @@ public: /// Whether the thread has input data. bool hasInputData() { return batchSize_ != 0; } -protected: + protected: void mergeCpuGradients(); void mergeGradSparse( @@ -421,7 +421,7 @@ protected: /// GradientMachine::backward void doCallback(int pid); -protected: + protected: MultiGradientMachine* multiMachine_; ModelConfig config_; /// whether the thread should stop diff --git a/paddle/gserver/gradientmachines/MultiNetwork.cpp b/paddle/gserver/gradientmachines/MultiNetwork.cpp index a1140402b8baaae20e20802ebf87462e301b60f9..5f3d09dda26772850828e6d44e8cc65635b314dc 100644 --- a/paddle/gserver/gradientmachines/MultiNetwork.cpp +++ b/paddle/gserver/gradientmachines/MultiNetwork.cpp @@ -122,7 +122,7 @@ void MultiNetwork::finish() { } class MultiCombinedEvaluator : public Evaluator { -public: + public: MultiCombinedEvaluator() {} void addEvaluator(std::unique_ptr&& evaluator) { evaluators_.emplace_back(std::move(evaluator)); @@ -167,7 +167,7 @@ public: } } -protected: + protected: std::vector> evaluators_; }; diff --git a/paddle/gserver/gradientmachines/MultiNetwork.h b/paddle/gserver/gradientmachines/MultiNetwork.h index 186a9ad0a39cd7815aea6738e6c6bc4a0c944aa9..495d5592017b5fb937fb8243bf12a5f2f30d67e7 100644 --- a/paddle/gserver/gradientmachines/MultiNetwork.h +++ b/paddle/gserver/gradientmachines/MultiNetwork.h @@ -22,7 +22,7 @@ limitations under the License. */ namespace paddle { class MultiNetwork : public NeuralNetwork { -public: + public: explicit MultiNetwork(std::string subModelName = "") : NeuralNetwork(subModelName) {} @@ -58,7 +58,7 @@ public: virtual void finish(); -protected: + protected: std::vector> subNetworks_; }; } // namespace paddle diff --git a/paddle/gserver/gradientmachines/NeuralNetwork.cpp b/paddle/gserver/gradientmachines/NeuralNetwork.cpp index a3c13df3dbad973505d8919bce8b95348527e273..ac60a3a3408d37b66cb712d893c6b93a1750f448 100644 --- a/paddle/gserver/gradientmachines/NeuralNetwork.cpp +++ b/paddle/gserver/gradientmachines/NeuralNetwork.cpp @@ -362,7 +362,7 @@ void NeuralNetwork::releaseOutput() { #ifndef PADDLE_MOBILE_INFERENCE class CombinedEvaluator : public Evaluator { -public: + public: void addEvaluator(std::unique_ptr&& evaluator) { evaluators_.emplace_back(std::move(evaluator)); } @@ -400,11 +400,11 @@ public: } } -protected: + protected: std::vector> evaluators_; // Evaluator interface -public: + public: /** * @brief getNames will return all inside evaluators' names. * @param names [out]: return names. @@ -435,7 +435,7 @@ public: }); } -private: + private: template T getMethodHelper(const std::string& name, Error* err, @@ -454,7 +454,7 @@ private: }; class SubnetEvaluator : public CombinedEvaluator { -public: + public: SubnetEvaluator(const std::string& layerName, std::unique_ptr&& evaluator) : layerName_(layerName) { @@ -473,7 +473,7 @@ public: << " in submodel " << nn.getName(); } -protected: + protected: std::string layerName_; }; diff --git a/paddle/gserver/gradientmachines/NeuralNetwork.h b/paddle/gserver/gradientmachines/NeuralNetwork.h index 5b32f844f742c07c8bee6638cb46dc00285f49b0..3e5615c8f0b30ab1283d41e025496051869289dc 100644 --- a/paddle/gserver/gradientmachines/NeuralNetwork.h +++ b/paddle/gserver/gradientmachines/NeuralNetwork.h @@ -56,7 +56,7 @@ void parameterInitNN(int paramId, std::vector* sharedParams); class NeuralNetwork : public GradientMachine { -public: + public: virtual void init(const ModelConfig& config, ParamInitCallback callback = nullptr, const std::vector& parameterTypes = @@ -144,7 +144,7 @@ public: */ void releaseOutput(); -protected: + protected: /** * The constructor of NeuralNetwork. * The sub networks can get parameters_ and parameterMap_ diff --git a/paddle/gserver/gradientmachines/ParallelNeuralNetwork.h b/paddle/gserver/gradientmachines/ParallelNeuralNetwork.h index e3b6812123141e8e0afb9368fb06f2b34f526800..c091459506ad477bed3f429a22071eccedd664bb 100644 --- a/paddle/gserver/gradientmachines/ParallelNeuralNetwork.h +++ b/paddle/gserver/gradientmachines/ParallelNeuralNetwork.h @@ -32,7 +32,7 @@ enum TaskType { * multiple threads in parallel. */ class ParallelNeuralNetwork : public NeuralNetwork { -public: + public: ParallelNeuralNetwork(std::string subModelName = "", NeuralNetwork *rootNetwork = nullptr) : NeuralNetwork(subModelName, rootNetwork) {} @@ -66,7 +66,7 @@ public: // virtual void eval(Evaluator* evaluator); -protected: + protected: bool useGpu_; /// number of gpu devices int numDevices_; @@ -74,7 +74,7 @@ protected: }; class ParallelThread { -public: + public: ParallelThread(int threadId, int deviceId, bool useGpu); ~ParallelThread(); void jobEnqueue(LayerPtr layer, TaskType task); @@ -87,10 +87,10 @@ public: } void setForwardPassType(PassType passType) { passType_ = passType; } -protected: + protected: void computeThread(); -public: + public: struct Job { LayerPtr layer_; TaskType task_; @@ -98,7 +98,7 @@ public: typedef Queue JobQueue; JobQueue queue_; -protected: + protected: /// from 0 to threads-1 int threadId_; /// the GPU device Id which the computeThread_ used diff --git a/paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp b/paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp index 2429b5d1a0a5ccf66db365b82c494c53d8e1fd4b..73ac8cda721f200c1a02cd9c1d9456df70d7b7d2 100644 --- a/paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp +++ b/paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp @@ -96,7 +96,7 @@ static InitFunction __init__diy_prob_method( std::numeric_limits::max()); class BeamSearchControlCallbacks { -public: + public: RecurrentGradientMachine::BeamSearchCandidatesAdjustCallback beamSearchCandidateAdjust; RecurrentGradientMachine::NormOrDropNodeCallback normOrDropNode; @@ -115,7 +115,7 @@ public: }; class BeamSearchStatisticsCallbacks { -public: + public: RecurrentGradientMachine::EachStepCallback onEachStepStarted; RecurrentGradientMachine::EachStepCallback onEachStepStoped; @@ -148,11 +148,11 @@ RecurrentGradientMachine::RecurrentGradientMachine( * so it's should not be placed in root network. */ class BootBiasLayer : public Layer { -protected: + protected: std::unique_ptr biases_; IVectorPtr cpuIds_; -public: + public: explicit BootBiasLayer(const LayerConfig& config) : Layer(config) {} bool init(const LayerMap& layerMap, diff --git a/paddle/gserver/gradientmachines/RecurrentGradientMachine.h b/paddle/gserver/gradientmachines/RecurrentGradientMachine.h index 0032b72cdae44588af976f1ac542149545f551f1..7e943cebd35234ba7af357c9f64fde6b0a9546ce 100644 --- a/paddle/gserver/gradientmachines/RecurrentGradientMachine.h +++ b/paddle/gserver/gradientmachines/RecurrentGradientMachine.h @@ -30,7 +30,7 @@ class BeamSearchControlCallbacks; class BeamSearchStatisticsCallbacks; class RecurrentGradientMachine : public NeuralNetwork { -public: + public: RecurrentGradientMachine(const std::string& subModelName, NeuralNetwork* rootNetwork); @@ -290,7 +290,7 @@ public: return this->finalPaths_; } -protected: + protected: std::vector commonSeqInfo_; ICpuGpuVectorPtr sequenceStartPositions_; void calcSequenceStartPositions(); @@ -447,7 +447,7 @@ protected: MatrixPtr cpuProb_; IVectorPtr cpuEos_; -private: + private: /* * @return beam size in beam search */ diff --git a/paddle/gserver/layers/AddtoLayer.h b/paddle/gserver/layers/AddtoLayer.h index 1d000630567cb1116ab0ff69e42380fc0eae6173..6ea54f4a53d466594055db2fb5167fa1a9d6c9da 100644 --- a/paddle/gserver/layers/AddtoLayer.h +++ b/paddle/gserver/layers/AddtoLayer.h @@ -33,10 +33,10 @@ namespace paddle { * The config file api is addto_layer. */ class AddtoLayer : public Layer { -protected: + protected: std::unique_ptr biases_; -public: + public: explicit AddtoLayer(const LayerConfig& config) : Layer(config) {} ~AddtoLayer() {} diff --git a/paddle/gserver/layers/AgentLayer.h b/paddle/gserver/layers/AgentLayer.h index da0ac4530836205757399ac8eb64dd003740a53f..51f346d5c9fdf9599cddf4b668c128035fd94187 100644 --- a/paddle/gserver/layers/AgentLayer.h +++ b/paddle/gserver/layers/AgentLayer.h @@ -26,11 +26,11 @@ namespace paddle { * called to set one and only one real layer */ class AgentLayer : public Layer { -protected: + protected: LayerPtr realLayer_; int numSamples_; -public: + public: explicit AgentLayer(const LayerConfig& config) : Layer(config) {} ~AgentLayer() {} @@ -55,14 +55,14 @@ public: * GatherAgentLayer collect a complete sequence. */ class GatherAgentLayer : public Layer { -protected: + protected: std::vector realLayers_; std::vector idsVec_; // we don't clear idsVec_ vector to aviod IVector alloc/free IVectorPtr allIds_; std::vector idIndex_; -public: + public: explicit GatherAgentLayer(const LayerConfig& config) : Layer(config) {} virtual ~GatherAgentLayer() {} @@ -95,7 +95,7 @@ public: * if it is, the agent will select a few ids in real layer. */ class ScatterAgentLayer : public Layer { -protected: + protected: LayerPtr realLayer_; IVectorPtr ids_; IVectorPtr cpuIds_; @@ -113,7 +113,7 @@ protected: // true for setRealLayer, false for setRealLayerAndOutput bool selectionMode_; -public: + public: explicit ScatterAgentLayer(const LayerConfig& config) : Layer(config) {} virtual ~ScatterAgentLayer() {} diff --git a/paddle/gserver/layers/AverageLayer.h b/paddle/gserver/layers/AverageLayer.h index 24602d2a9c3e08cf76f6f98b5f9e3f593118e6e1..03e2673b55ceca7a698f1b858327ad6fad739087 100644 --- a/paddle/gserver/layers/AverageLayer.h +++ b/paddle/gserver/layers/AverageLayer.h @@ -37,7 +37,7 @@ namespace paddle { * The config file api is pooling_layer. */ class AverageLayer : public SequencePoolLayer { -public: + public: enum AverageStrategy { kAverage = 0, kSum = 1, kAverageSquareRootN = 2 }; explicit AverageLayer(const LayerConfig& config) : SequencePoolLayer(config) {} @@ -48,7 +48,7 @@ public: void forward(PassType passType) override; void backward(const UpdateCallback& callback = nullptr) override; -protected: + protected: int mode_; }; } // namespace paddle diff --git a/paddle/gserver/layers/BatchNormBaseLayer.h b/paddle/gserver/layers/BatchNormBaseLayer.h index 69d642af4f12593e8db8a726310e6b1934c8e3be..5a446c0843a22adecbaf2ae09fcd526b68865ae2 100644 --- a/paddle/gserver/layers/BatchNormBaseLayer.h +++ b/paddle/gserver/layers/BatchNormBaseLayer.h @@ -40,7 +40,7 @@ namespace paddle { */ class BatchNormBaseLayer : public Layer { -public: + public: explicit BatchNormBaseLayer(const LayerConfig& config) : Layer(config) {} ~BatchNormBaseLayer() {} @@ -61,7 +61,7 @@ public: */ void calFeatureMapSize(); -protected: + protected: /// Batch normalization scale parameter, which is referred to as gamma in /// in original paper. std::unique_ptr weight_; diff --git a/paddle/gserver/layers/BatchNormalizationLayer.h b/paddle/gserver/layers/BatchNormalizationLayer.h index 95add69215e3ea0b0225d0a245fe37905c33127b..e5e4e690b6017f32de0f4d7557065c02c03d689f 100644 --- a/paddle/gserver/layers/BatchNormalizationLayer.h +++ b/paddle/gserver/layers/BatchNormalizationLayer.h @@ -27,7 +27,7 @@ namespace paddle { */ class BatchNormalizationLayer : public BatchNormBaseLayer { -public: + public: explicit BatchNormalizationLayer(const LayerConfig& config) : BatchNormBaseLayer(config), firstTest_(true) {} @@ -38,7 +38,7 @@ public: void forward(PassType passType) override; void backward(const UpdateCallback& callback = nullptr) override; -protected: + protected: /// Load pre-calculated mean and std. void setMeanAndStd(); diff --git a/paddle/gserver/layers/BilinearInterpLayer.h b/paddle/gserver/layers/BilinearInterpLayer.h index acd320420f4bbfe313f3ae77577ffc6b5cbfbfdf..8e08c2e1ce80172f55c93d8242821f683fa1a731 100644 --- a/paddle/gserver/layers/BilinearInterpLayer.h +++ b/paddle/gserver/layers/BilinearInterpLayer.h @@ -26,13 +26,13 @@ namespace paddle { * @note The config file api is bilinear_interp_layer. */ class BilinearInterpLayer : public Layer { -protected: + protected: size_t outImgH_, outImgW_; size_t inImgH_, inImgW_; real ratioH_, ratioW_; size_t numChannels_; -public: + public: explicit BilinearInterpLayer(const LayerConfig& config) : Layer(config) {} virtual ~BilinearInterpLayer() {} diff --git a/paddle/gserver/layers/BlockExpandLayer.h b/paddle/gserver/layers/BlockExpandLayer.h index 1797b64036b5cb9f97477d5a44b2f58e2d6c0cd4..9d76584f3a4eda19a9e8f806256a7b8da617cc37 100644 --- a/paddle/gserver/layers/BlockExpandLayer.h +++ b/paddle/gserver/layers/BlockExpandLayer.h @@ -40,7 +40,7 @@ namespace paddle { * The config file api is block_expand_layer. */ class BlockExpandLayer : public Layer { -protected: + protected: /** * @brief Calculate outputH_ and outputW_ and return block number which * actually is time steps. @@ -53,7 +53,7 @@ protected: TensorShape inputShape_; TensorShape outputShape_; -public: + public: explicit BlockExpandLayer(const LayerConfig& config) : Layer(config) {} ~BlockExpandLayer() {} diff --git a/paddle/gserver/layers/CRFDecodingLayer.h b/paddle/gserver/layers/CRFDecodingLayer.h index fba3cebac1a375008c58d21c458d9e0b98305ffa..018162e146fa93725fe84bdf2da9a6124f3cea6f 100644 --- a/paddle/gserver/layers/CRFDecodingLayer.h +++ b/paddle/gserver/layers/CRFDecodingLayer.h @@ -30,14 +30,14 @@ namespace paddle { * See LinearChainCRF.h for the detail of the CRF formulation. */ class CRFDecodingLayer : public CRFLayer { -public: + public: explicit CRFDecodingLayer(const LayerConfig& config) : CRFLayer(config) {} bool init(const LayerMap& layerMap, const ParameterMap& parameterMap) override; void forward(PassType passType) override; void backward(const UpdateCallback& callback) override; -protected: + protected: std::unique_ptr crf_; }; diff --git a/paddle/gserver/layers/CRFLayer.h b/paddle/gserver/layers/CRFLayer.h index cb5bd05568cc79c0093d6af0791cf0b3ce2dae47..88c2ed343ad5743068c871fe351437270d85f223 100644 --- a/paddle/gserver/layers/CRFLayer.h +++ b/paddle/gserver/layers/CRFLayer.h @@ -27,14 +27,14 @@ namespace paddle { * See class LinearChainCRF for the detail of the CRF formulation. */ class CRFLayer : public Layer { -public: + public: explicit CRFLayer(const LayerConfig& config) : Layer(config) {} bool init(const LayerMap& layerMap, const ParameterMap& parameterMap) override; void forward(PassType passType) override; void backward(const UpdateCallback& callback) override; -protected: + protected: size_t numClasses_; ParameterPtr parameter_; std::vector crfs_; diff --git a/paddle/gserver/layers/CTCLayer.h b/paddle/gserver/layers/CTCLayer.h index fcbc42565e9340903d05aca2d0ba2091ffe20be0..5d70b1f4ceb03028865378d1d01b5706b35b10de 100644 --- a/paddle/gserver/layers/CTCLayer.h +++ b/paddle/gserver/layers/CTCLayer.h @@ -20,7 +20,7 @@ limitations under the License. */ namespace paddle { class CTCLayer : public Layer { -public: + public: explicit CTCLayer(const LayerConfig& config) : Layer(config) {} bool init(const LayerMap& layerMap, const ParameterMap& parameterMap) override; @@ -31,7 +31,7 @@ public: const Argument& softmaxSeqs, const Argument& labelSeqs); -protected: + protected: size_t numClasses_; bool normByTimes_; std::vector ctcs_; diff --git a/paddle/gserver/layers/ClipLayer.cpp b/paddle/gserver/layers/ClipLayer.cpp index dbc3337499788af5a9b6f68a6016e94c2072d61b..6aa3c8fe64f5a59e82f3271baed99fd17fd6653f 100644 --- a/paddle/gserver/layers/ClipLayer.cpp +++ b/paddle/gserver/layers/ClipLayer.cpp @@ -24,11 +24,11 @@ namespace paddle { */ class ClipLayer : public Layer { -protected: + protected: double min_; double max_; -public: + public: explicit ClipLayer(const LayerConfig& config) : Layer(config) {} bool init(const LayerMap& layerMap, diff --git a/paddle/gserver/layers/ConcatenateLayer.cpp b/paddle/gserver/layers/ConcatenateLayer.cpp index f5ab29a509e45e72c71ba122c73aeba1b3b6a827..e6de329ff3f9ccfdd1cbe697c1de1a9cd8c7926a 100644 --- a/paddle/gserver/layers/ConcatenateLayer.cpp +++ b/paddle/gserver/layers/ConcatenateLayer.cpp @@ -23,7 +23,7 @@ namespace paddle { * each input as one row for the output of this layer and apply activation. */ class ConcatenateLayer : public Layer { -public: + public: explicit ConcatenateLayer(const LayerConfig& config) : Layer(config) {} ~ConcatenateLayer() {} @@ -97,7 +97,7 @@ void ConcatenateLayer::backward(const UpdateCallback& callback) { * processed by a Projection. */ class ConcatenateLayer2 : public Layer { -public: + public: explicit ConcatenateLayer2(const LayerConfig& config) : Layer(config) {} ~ConcatenateLayer2() {} @@ -108,7 +108,7 @@ public: void forward(PassType passType) override; void backward(const UpdateCallback& callback = nullptr) override; -protected: + protected: std::vector> projections_; std::vector projOutput_; std::vector> projCol_; diff --git a/paddle/gserver/layers/ContextProjection.h b/paddle/gserver/layers/ContextProjection.h index e30f98f58d2be9ac538f6385efe68990b705ac5f..9c217145419048282a9a09ad899dc970e7c9704f 100644 --- a/paddle/gserver/layers/ContextProjection.h +++ b/paddle/gserver/layers/ContextProjection.h @@ -42,7 +42,7 @@ namespace paddle { * The config file api is context_projection. */ class ContextProjection : public Projection { -public: + public: /** * Constructor. If context_start is zero and context_lenth is one, it will * set trainable_padding false. trainable_padding is an optional arguments @@ -63,7 +63,7 @@ public: virtual bool init(); -protected: + protected: std::unique_ptr weight_; /// number of extra timesteps added at the beginning size_t beginPad_; diff --git a/paddle/gserver/layers/Conv3DLayer.h b/paddle/gserver/layers/Conv3DLayer.h index 5ab5ff3d4af07449484c441958c31c8fb06de894..07b804bad02beb6ec9c3e9fd43c3cd3aa6d50b22 100644 --- a/paddle/gserver/layers/Conv3DLayer.h +++ b/paddle/gserver/layers/Conv3DLayer.h @@ -26,7 +26,7 @@ namespace paddle { * calculate convolution operation. */ class Conv3DLayer : public ConvBaseLayer { -public: + public: explicit Conv3DLayer(const LayerConfig& config) : ConvBaseLayer(config) {} ~Conv3DLayer() {} @@ -40,7 +40,7 @@ public: void bpropWeights(int i); size_t getSize(); -protected: + protected: // Figure out the dimensions for individual gemms. IntV M_; /// numFilters_ / filter_group_; IntV N_; /// channels_ * filterSizeZ_ * filterSize_ * filterSizeY_ diff --git a/paddle/gserver/layers/ConvBaseLayer.h b/paddle/gserver/layers/ConvBaseLayer.h index 93869fe68d15b1cf38296fa8e2f6197dc74f879f..801bc4f888c5a60e803c882dcf807678c64af20c 100644 --- a/paddle/gserver/layers/ConvBaseLayer.h +++ b/paddle/gserver/layers/ConvBaseLayer.h @@ -24,7 +24,7 @@ namespace paddle { */ class ConvBaseLayer : public Layer { -protected: + protected: typedef std::vector IntV; /// True if it's deconv layer, false if it's convolution layer @@ -88,7 +88,7 @@ protected: /// of output size. bool caffeMode_; -public: + public: explicit ConvBaseLayer(const LayerConfig& config) : Layer(config) {} bool init(const LayerMap& layerMap, diff --git a/paddle/gserver/layers/ConvBaseOperator.h b/paddle/gserver/layers/ConvBaseOperator.h index 27fb0362d3c9518a263eac54206e00974d08eb20..c3c647cb69da5a70eb5346737cc0092e2201c89e 100644 --- a/paddle/gserver/layers/ConvBaseOperator.h +++ b/paddle/gserver/layers/ConvBaseOperator.h @@ -29,7 +29,7 @@ namespace paddle { */ class ConvBaseOperator : public Operator { -public: + public: ConvBaseOperator(const OperatorConfig &config, bool useGpu); /** * Free workspace in device and destroy cudnn tensor descriptor. @@ -46,7 +46,7 @@ public: hl_destroy_convolution_descriptor(convDesc_); } -protected: + protected: /** * Get convolution parameters from layer config and * initialize member variables. diff --git a/paddle/gserver/layers/ConvBaseProjection.h b/paddle/gserver/layers/ConvBaseProjection.h index ba76d236d901187093a2e372a61c5d29d661e8bb..f3266ae1ab945042cde9f24b7c2673c18d37bc11 100644 --- a/paddle/gserver/layers/ConvBaseProjection.h +++ b/paddle/gserver/layers/ConvBaseProjection.h @@ -23,7 +23,7 @@ namespace paddle { * @brief Base class for ConvProjection and ConvTransProjection. */ class ConvBaseProjection : public Projection { -public: + public: /** * Constructor. */ @@ -33,7 +33,7 @@ public: ~ConvBaseProjection(); -protected: + protected: void getConvParams(); void initCudnn(); diff --git a/paddle/gserver/layers/ConvOperator.h b/paddle/gserver/layers/ConvOperator.h index fbdb7bb1cd2b81bd72912dffdc9d059c520068a8..527dbf8c270f35e19ca23acd8a3ba8197d03b988 100644 --- a/paddle/gserver/layers/ConvOperator.h +++ b/paddle/gserver/layers/ConvOperator.h @@ -29,7 +29,7 @@ namespace paddle { */ class ConvOperator : public ConvBaseOperator { -public: + public: ConvOperator(const OperatorConfig &config, bool useGpu) : ConvBaseOperator(config, useGpu) {} /** diff --git a/paddle/gserver/layers/ConvProjection.h b/paddle/gserver/layers/ConvProjection.h index e8ecb99431a421d4b52228600909568b0808649a..22a2202bb6cc256a4a5897724d8eb8a93fefb79f 100644 --- a/paddle/gserver/layers/ConvProjection.h +++ b/paddle/gserver/layers/ConvProjection.h @@ -23,7 +23,7 @@ namespace paddle { * @brief Convolution projection do the same calculation with CudnnConvLayer. */ class ConvProjection : public ConvBaseProjection { -public: + public: /** * Constructor. */ diff --git a/paddle/gserver/layers/ConvShiftLayer.cpp b/paddle/gserver/layers/ConvShiftLayer.cpp index fb877710196835e025466f37b5da27bcf80a3db4..615c3478061b591ea30cbf0b3d27ef2551c0dd28 100644 --- a/paddle/gserver/layers/ConvShiftLayer.cpp +++ b/paddle/gserver/layers/ConvShiftLayer.cpp @@ -42,7 +42,7 @@ namespace paddle { */ class ConvShiftLayer : public Layer { -public: + public: explicit ConvShiftLayer(const LayerConfig& config) : Layer(config) {} ~ConvShiftLayer() {} diff --git a/paddle/gserver/layers/ConvTransOperator.h b/paddle/gserver/layers/ConvTransOperator.h index 1bf58f2bfb78ae7dee433455ece37d908b113045..53cb7a21b49189898d09aa20cd46d04cc5c20198 100644 --- a/paddle/gserver/layers/ConvTransOperator.h +++ b/paddle/gserver/layers/ConvTransOperator.h @@ -29,7 +29,7 @@ namespace paddle { */ class ConvTransOperator : public ConvBaseOperator { -public: + public: ConvTransOperator(const OperatorConfig &config, bool useGpu) : ConvBaseOperator(config, useGpu) {} /** diff --git a/paddle/gserver/layers/ConvTransProjection.h b/paddle/gserver/layers/ConvTransProjection.h index 269b2694c82ea076102633537d7c961139a19a43..0f9ed720d3b8855a3a24ac25a1c3917c4b98e81d 100644 --- a/paddle/gserver/layers/ConvTransProjection.h +++ b/paddle/gserver/layers/ConvTransProjection.h @@ -23,7 +23,7 @@ namespace paddle { * @brief Convolution projection do the same calculation with CudnnConvLayer. */ class ConvTransProjection : public ConvBaseProjection { -public: + public: /** * Constructor. */ diff --git a/paddle/gserver/layers/ConvexCombinationLayer.cpp b/paddle/gserver/layers/ConvexCombinationLayer.cpp index dce751940c1bf1695a034a3c551412dcb9b7b8b5..31363d97c4fd318ec2c6d48f9200f6ba1f49ba11 100644 --- a/paddle/gserver/layers/ConvexCombinationLayer.cpp +++ b/paddle/gserver/layers/ConvexCombinationLayer.cpp @@ -36,7 +36,7 @@ namespace paddle { * The config file api is linear_comb_layer. */ class ConvexCombinationLayer : public Layer { -protected: + protected: /// A matrix pointer pointing to second input. MatrixPtr tmpMtx0; /// A matrix pointer pointing to first input. @@ -44,7 +44,7 @@ protected: /// A matrix pointer pointing to output. MatrixPtr tmpRow1; -public: + public: explicit ConvexCombinationLayer(const LayerConfig& config) : Layer(config) {} ~ConvexCombinationLayer() {} diff --git a/paddle/gserver/layers/CosSimLayer.h b/paddle/gserver/layers/CosSimLayer.h index 675cdb16b563faa7acf9e701096bd334ed661160..d9fe1ff270f1f76e3b246dca374ddf45445419f9 100644 --- a/paddle/gserver/layers/CosSimLayer.h +++ b/paddle/gserver/layers/CosSimLayer.h @@ -33,7 +33,7 @@ namespace paddle { * The config file api is cos_sim. */ class CosSimLayer : public Layer { -public: + public: explicit CosSimLayer(const LayerConfig& config) : Layer(config) {} ~CosSimLayer() {} diff --git a/paddle/gserver/layers/CosSimVecMatLayer.cpp b/paddle/gserver/layers/CosSimVecMatLayer.cpp index 685b4e8ef376b76b3058eeba82d803d460e7105c..230ecc768b4d7314b21ac1d76899c3c3bab12309 100644 --- a/paddle/gserver/layers/CosSimVecMatLayer.cpp +++ b/paddle/gserver/layers/CosSimVecMatLayer.cpp @@ -32,7 +32,7 @@ namespace paddle { */ class CosSimVecMatLayer : public Layer { -protected: + protected: MatrixPtr tmpMtx0; MatrixPtr tmpMtx1; MatrixPtr tmpRow0; @@ -40,7 +40,7 @@ protected: MatrixPtr tmpRow2; MatrixPtr tmpRow3; -public: + public: explicit CosSimVecMatLayer(const LayerConfig& config) : Layer(config) {} ~CosSimVecMatLayer() {} diff --git a/paddle/gserver/layers/CostLayer.cpp b/paddle/gserver/layers/CostLayer.cpp index 484f803a8387a16152c5911d7d5c72b0111283ae..1327616950a8887efa2cba410fa7ae8b5bd97da4 100644 --- a/paddle/gserver/layers/CostLayer.cpp +++ b/paddle/gserver/layers/CostLayer.cpp @@ -716,7 +716,7 @@ void HuberTwoClassification::backwardImp(Matrix& output, * \f] */ class SumCostLayer : public Layer { -public: + public: explicit SumCostLayer(const LayerConfig& config) : Layer(config) {} bool init(const LayerMap& layerMap, diff --git a/paddle/gserver/layers/CostLayer.h b/paddle/gserver/layers/CostLayer.h index 306c067ed1c040555d2b03996cc0749faf0ea68c..9bfec0e2b169fac4f235fd13347be687c4f1a222 100644 --- a/paddle/gserver/layers/CostLayer.h +++ b/paddle/gserver/layers/CostLayer.h @@ -29,7 +29,7 @@ namespace paddle { * handled by the base class. */ class CostLayer : public Layer { -public: + public: explicit CostLayer(const LayerConfig& config) : Layer(config) {} bool init(const LayerMap& layerMap, @@ -51,7 +51,7 @@ public: Argument& label, Matrix& outputGrad) = 0; -protected: + protected: LayerPtr weightLayer_; real coeff_; }; @@ -65,7 +65,7 @@ protected: * \f] */ class MultiClassCrossEntropy : public CostLayer { -public: + public: explicit MultiClassCrossEntropy(const LayerConfig& config) : CostLayer(config) {} @@ -95,7 +95,7 @@ public: * In Proceedings of the ACL 2014 Conference. */ class MultiClassCrossEntropyWithSelfNorm : public CostLayer { -public: + public: explicit MultiClassCrossEntropyWithSelfNorm(const LayerConfig& config) : CostLayer(config) {} @@ -108,7 +108,7 @@ public: Argument& label, Matrix& outputGrad) override; -protected: + protected: MatrixPtr sftMaxSum_; MatrixPtr sumInv_; }; @@ -120,7 +120,7 @@ protected: * \f] */ class SoftBinaryClassCrossEntropy : public CostLayer { -public: + public: explicit SoftBinaryClassCrossEntropy(const LayerConfig& config) : CostLayer(config) {} @@ -133,7 +133,7 @@ public: Argument& label, Matrix& outputGrad) override; -protected: + protected: MatrixPtr targetPerDim_; }; @@ -145,7 +145,7 @@ protected: * \f] */ class SumOfSquaresCostLayer : public CostLayer { -public: + public: explicit SumOfSquaresCostLayer(const LayerConfig& config) : CostLayer(config) {} @@ -171,7 +171,7 @@ public: * x = output - label */ class SmoothL1CostLayer : public CostLayer { -public: + public: explicit SmoothL1CostLayer(const LayerConfig& config) : CostLayer(config) {} bool init(const LayerMap& layerMap, @@ -197,7 +197,7 @@ public: * Rank useing Gradient Descent. */ class RankingCost : public Layer { -public: + public: explicit RankingCost(const LayerConfig& config) : Layer(config) {} bool init(const LayerMap& layerMap, @@ -225,7 +225,7 @@ public: (void)outputGrad; } -private: + private: double posPairCount_; double negPairCount_; MatrixPtr margin_; @@ -250,7 +250,7 @@ private: * with Nonsmooth Cost Functions. */ class LambdaCost : public Layer { -public: + public: explicit LambdaCost(const LayerConfig& config) : Layer(config) {} bool init(const LayerMap& layerMap, @@ -270,7 +270,7 @@ public: real* gradData, int size); -private: + private: MatrixPtr marginGrad_; int truncationSize_; int maxSortSize_; @@ -287,10 +287,10 @@ private: * \f] */ class MultiBinaryLabelCrossEntropy : public CostLayer { -protected: + protected: MatrixPtr targetPerDim_; -public: + public: explicit MultiBinaryLabelCrossEntropy(const LayerConfig& config) : CostLayer(config) {} @@ -308,7 +308,7 @@ public: * A base layer for HuberRegressionLoss and HuberTwoClassification. */ class HuberCost : public CostLayer { -public: + public: std::vector tmpCpuInput_; explicit HuberCost(const LayerConfig& config) : CostLayer(config) {} @@ -331,7 +331,7 @@ public: * Loss = delta * abs(y - f) - 0.5 * delta^2, otherwise */ class HuberRegressionLoss : public HuberCost { -public: + public: explicit HuberRegressionLoss(const LayerConfig& config) : HuberCost(config) {} bool init(const LayerMap& layerMap, @@ -343,7 +343,7 @@ public: Argument& label, Matrix& outputGrad) override; -protected: + protected: real delta_; }; @@ -356,7 +356,7 @@ protected: * Loss = 0, otherwise */ class HuberTwoClassification : public HuberCost { -public: + public: explicit HuberTwoClassification(const LayerConfig& config) : HuberCost(config) {} diff --git a/paddle/gserver/layers/CropLayer.h b/paddle/gserver/layers/CropLayer.h index 1a85911ef75e992df587a60cfc9a727eafa4cc76..ef88bc483d157406a0f5a7924c14c345ea0df8c4 100644 --- a/paddle/gserver/layers/CropLayer.h +++ b/paddle/gserver/layers/CropLayer.h @@ -28,7 +28,7 @@ namespace paddle { * crop input as this shape conf */ class CropLayer : public Layer { -public: + public: explicit CropLayer(const LayerConfig& config) : Layer(config) {} ~CropLayer() {} @@ -38,7 +38,7 @@ public: void forward(PassType passType) override; void backward(const UpdateCallback& callback = nullptr) override; -protected: + protected: void setOutDims(); void setInDims(); diff --git a/paddle/gserver/layers/CrossEntropyOverBeam.h b/paddle/gserver/layers/CrossEntropyOverBeam.h index b47a2933c255c264ba780b2d87c9fbe53cb5665d..c8702b16165eee8d552c563082ffc708ce443deb 100644 --- a/paddle/gserver/layers/CrossEntropyOverBeam.h +++ b/paddle/gserver/layers/CrossEntropyOverBeam.h @@ -44,7 +44,7 @@ struct BeamExpansion { typedef std::shared_ptr BeamExpansionPtr; class CostForOneSequence { -public: + public: CostForOneSequence() : beamSize_(0), validExpansionCount_(0), goldAsExtraPath_(false) {} void setData(const BeamExpansionPtr bPtr, size_t beamSize) { @@ -64,7 +64,7 @@ public: real forward(); void backward(); -private: + private: void calValidExpandStep(); void constructTotalExpansion(); size_t initLastExpansion(); @@ -93,14 +93,14 @@ private: }; class CrossEntropyOverBeam : public Layer { -public: + public: explicit CrossEntropyOverBeam(const LayerConfig& config) : Layer(config) {} bool init(const LayerMap& layerMap, const ParameterMap& parameterMap) override; void forward(PassType passType) override; void backward(const UpdateCallback& callback) override; -private: + private: void checkInputs(); void copyInputsToCpu(); void resizeOutput(); diff --git a/paddle/gserver/layers/CudnnBatchNormLayer.h b/paddle/gserver/layers/CudnnBatchNormLayer.h index aa279f73d66770384815cad4d9e2ee0b04a4a1ad..1bb4eff8d2372660caa4ec4a4a20a27f365bebd0 100644 --- a/paddle/gserver/layers/CudnnBatchNormLayer.h +++ b/paddle/gserver/layers/CudnnBatchNormLayer.h @@ -30,7 +30,7 @@ namespace paddle { */ class CudnnBatchNormLayer : public BatchNormBaseLayer { -public: + public: explicit CudnnBatchNormLayer(const LayerConfig& config) : BatchNormBaseLayer(config) {} @@ -46,7 +46,7 @@ public: void forward(PassType passType) override; void backward(const UpdateCallback& callback = nullptr) override; -protected: + protected: /// Epsilon value used in the batch normalization formula. /// Same epsilon value should be used in forward and backward functions. double eps_; diff --git a/paddle/gserver/layers/CudnnConvBaseLayer.h b/paddle/gserver/layers/CudnnConvBaseLayer.h index 698104e4fbd2556f426001687a581153f32773d8..1ee1aa100d8adaed04ce24ee12b5b9af52c14b13 100644 --- a/paddle/gserver/layers/CudnnConvBaseLayer.h +++ b/paddle/gserver/layers/CudnnConvBaseLayer.h @@ -31,14 +31,14 @@ namespace paddle { * The config file api is img_conv_layer. */ class CudnnConvBaseLayer : public ConvBaseLayer { -protected: + protected: std::vector> projConf_; std::vector> projections_; hl_tensor_descriptor biasDesc_; hl_tensor_descriptor outputDesc_; -public: + public: explicit CudnnConvBaseLayer(const LayerConfig& config) : ConvBaseLayer(config) {} diff --git a/paddle/gserver/layers/CudnnPoolLayer.h b/paddle/gserver/layers/CudnnPoolLayer.h index 9eb4fc6138b0bce59660406705d15291eb38af9b..fc249354d10333211691b6844bffa3c8da8a79ee 100644 --- a/paddle/gserver/layers/CudnnPoolLayer.h +++ b/paddle/gserver/layers/CudnnPoolLayer.h @@ -26,7 +26,7 @@ namespace paddle { */ class CudnnPoolLayer : public PoolLayer { -protected: + protected: int windowHeight, windowWidth; int heightPadding, widthPadding, strideHeight, strideWidth; int imageH_, imageW_, outputH_, outputW_; @@ -40,7 +40,7 @@ protected: /// A description of a pooling operation. hl_pooling_descriptor poolingDesc_; -public: + public: static bool typeCheck(const std::string& poolType, hl_pooling_mode_t* mode = nullptr); explicit CudnnPoolLayer(const LayerConfig& config); diff --git a/paddle/gserver/layers/DataLayer.h b/paddle/gserver/layers/DataLayer.h index 4b12afe0efe81843b58e459ca1e58b4f7f4a1664..d02f5a4697b9067f7d34e4d0b2d34f8c63ffe020 100644 --- a/paddle/gserver/layers/DataLayer.h +++ b/paddle/gserver/layers/DataLayer.h @@ -25,7 +25,7 @@ namespace paddle { * The config file api is data_layer. */ class DataLayer : public Layer { -public: + public: explicit DataLayer(const LayerConfig& config) : Layer(config) {} virtual void setData(const Argument& data) { data_ = data; } @@ -58,10 +58,10 @@ public: } } -private: + private: void copyDataToOutput(Argument& output); -protected: + protected: Argument data_; }; diff --git a/paddle/gserver/layers/DataNormLayer.h b/paddle/gserver/layers/DataNormLayer.h index 2a2a2a4aa76e8e315d9d66da1b738d6d615d10f2..7ae67a877b488c8d197896b8b1e3e90057fbe1c9 100644 --- a/paddle/gserver/layers/DataNormLayer.h +++ b/paddle/gserver/layers/DataNormLayer.h @@ -37,7 +37,7 @@ namespace paddle { */ class DataNormLayer : public Layer { -public: + public: enum NormalizationStrategy { kZScore = 0, kMinMax = 1, kDecimalScaling = 2 }; explicit DataNormLayer(const LayerConfig& config) : Layer(config) {} @@ -50,7 +50,7 @@ public: void forward(PassType passType) override; void backward(const UpdateCallback& callback = nullptr) override; -protected: + protected: int mode_; std::unique_ptr weight_; MatrixPtr min_; diff --git a/paddle/gserver/layers/DeConv3DLayer.h b/paddle/gserver/layers/DeConv3DLayer.h index 57d51cdec66930b9b79c0c0395da66922cd53ae4..13d1d07cf5cc6e2a6ea89768e29b1fe8cda5e81c 100644 --- a/paddle/gserver/layers/DeConv3DLayer.h +++ b/paddle/gserver/layers/DeConv3DLayer.h @@ -27,7 +27,7 @@ namespace paddle { * calculate deconvolution3D operation. */ class DeConv3DLayer : public ConvBaseLayer { -public: + public: explicit DeConv3DLayer(const LayerConfig& config) : ConvBaseLayer(config) {} ~DeConv3DLayer() {} bool init(const LayerMap& layerMap, const ParameterMap& parameterMap); @@ -40,7 +40,7 @@ public: void bpropWeights(int i); size_t getSize(); -protected: + protected: // Figure out the dimensions for individual gemms. IntV M_; /// numFilters_ / filter_group_; IntV N_; /// channels_ * filterSizeZ_ * filterSize_ * filterSizeY_ diff --git a/paddle/gserver/layers/DetectionOutputLayer.h b/paddle/gserver/layers/DetectionOutputLayer.h index 174a6e5d9acb476276b66627b4aabce2ae6c1037..b0270ed33141993665aeabdc53829600a4403643 100644 --- a/paddle/gserver/layers/DetectionOutputLayer.h +++ b/paddle/gserver/layers/DetectionOutputLayer.h @@ -33,7 +33,7 @@ namespace paddle { */ class DetectionOutputLayer : public Layer { -public: + public: explicit DetectionOutputLayer(const LayerConfig& config) : Layer(config) {} bool init(const LayerMap& layerMap, const ParameterMap& parameterMap); @@ -42,7 +42,7 @@ public: void backward(const UpdateCallback& callback = nullptr) {} -protected: + protected: inline LayerPtr getPriorBoxLayer() { return inputLayers_[0]; } inline LayerPtr getLocInputLayer(size_t index) { @@ -53,7 +53,7 @@ protected: return inputLayers_[1 + inputNum_ + index]; } -private: + private: size_t numClasses_; // number of classes size_t inputNum_; // number of input layers real nmsThreshold_; diff --git a/paddle/gserver/layers/DotMulOperator.cpp b/paddle/gserver/layers/DotMulOperator.cpp index 68db2929adee1336e52abfcb8e6495e589afa683..03d18d9b239e57dc41334462f2324ae2d0505a62 100644 --- a/paddle/gserver/layers/DotMulOperator.cpp +++ b/paddle/gserver/layers/DotMulOperator.cpp @@ -27,7 +27,7 @@ namespace paddle { * The config file api is dotmul_operator. */ class DotMulOperator : public Operator { -public: + public: DotMulOperator(const OperatorConfig& config, bool useGpu); virtual void forward(); virtual void backward(); diff --git a/paddle/gserver/layers/DotMulProjection.cpp b/paddle/gserver/layers/DotMulProjection.cpp index 86453aae84142f9f534182d085f4a96a2c7a3e15..d7780387670e83af24fa342be3d596b618b1f677 100644 --- a/paddle/gserver/layers/DotMulProjection.cpp +++ b/paddle/gserver/layers/DotMulProjection.cpp @@ -26,14 +26,14 @@ namespace paddle { * The config file api is dotmul_projection. */ class DotMulProjection : public Projection { -public: + public: DotMulProjection(const ProjectionConfig& config, const ParameterPtr& parameter, bool useGpu); virtual void forward(); virtual void backward(const UpdateCallback& callback); -protected: + protected: /// shared memory with parameter std::unique_ptr weight_; }; diff --git a/paddle/gserver/layers/DotProdLayer.cpp b/paddle/gserver/layers/DotProdLayer.cpp index 5148d93e27d199b0c373221cedd4f03d6d32c8ab..72b0c707b2131dc275ba604cd20ae0007c34a9a9 100644 --- a/paddle/gserver/layers/DotProdLayer.cpp +++ b/paddle/gserver/layers/DotProdLayer.cpp @@ -27,7 +27,7 @@ namespace paddle { */ class DotProdLayer : public Layer { -public: + public: explicit DotProdLayer(const LayerConfig& config) : Layer(config) {} ~DotProdLayer() {} diff --git a/paddle/gserver/layers/EosIdCheckLayer.cpp b/paddle/gserver/layers/EosIdCheckLayer.cpp index 470a5b8ea208ad0acb64e3067881e0d183e1dc39..04400f2836581179849a4dd1c256bbddcc82530f 100644 --- a/paddle/gserver/layers/EosIdCheckLayer.cpp +++ b/paddle/gserver/layers/EosIdCheckLayer.cpp @@ -24,7 +24,7 @@ namespace paddle { * It is used by recurrent layer group. */ class EosIdCheckLayer : public Layer { -public: + public: explicit EosIdCheckLayer(const LayerConfig& config) : Layer(config) {} bool init(const LayerMap& layerMap, diff --git a/paddle/gserver/layers/ExpandConvLayer.h b/paddle/gserver/layers/ExpandConvLayer.h index be968155efd0b8f19503c996ccd329379c6b1104..6919ef71355a4c660b9ddd60bff75fee399cfaa9 100644 --- a/paddle/gserver/layers/ExpandConvLayer.h +++ b/paddle/gserver/layers/ExpandConvLayer.h @@ -29,7 +29,7 @@ namespace paddle { */ class ExpandConvLayer : public ConvBaseLayer { -public: + public: explicit ExpandConvLayer(const LayerConfig& config) : ConvBaseLayer(config) {} ~ExpandConvLayer() {} @@ -42,7 +42,7 @@ public: size_t getOutputSize(); -protected: + protected: std::vector inputShape_; std::vector filterShape_; std::vector outputShape_; diff --git a/paddle/gserver/layers/ExpandLayer.h b/paddle/gserver/layers/ExpandLayer.h index 04bbfcbd04931fa11d11a9fcc74f0e4f19767f1b..06bd4ef05ee206628d981fee8e7eec3c91b18b7a 100644 --- a/paddle/gserver/layers/ExpandLayer.h +++ b/paddle/gserver/layers/ExpandLayer.h @@ -37,7 +37,7 @@ namespace paddle { */ class ExpandLayer : public Layer { -protected: + protected: std::unique_ptr biases_; /// if input[0] is dense data, ExpandLevel=kNonSeq; /// if input[0] is sequence data, ExpandLevel=kSeq @@ -48,7 +48,7 @@ protected: /// of input[1] ICpuGpuVectorPtr expandStartsPos_; -public: + public: explicit ExpandLayer(const LayerConfig& config) : Layer(config) {} ~ExpandLayer() {} diff --git a/paddle/gserver/layers/FactorizationMachineLayer.h b/paddle/gserver/layers/FactorizationMachineLayer.h index 684da4e65a461d46204c348b3374b0e9e00eb389..148abe238173dd44cd0fcf3f5cda732f70078706 100644 --- a/paddle/gserver/layers/FactorizationMachineLayer.h +++ b/paddle/gserver/layers/FactorizationMachineLayer.h @@ -42,7 +42,7 @@ namespace paddle { */ class FactorizationMachineLayer : public Layer { -protected: + protected: // The latent vectors, shape: (size, factorSize_) // Each row of the latentVectors_ matrix is the latent vector // corresponding to one input feature dimension @@ -50,7 +50,7 @@ protected: // The hyperparameter that defines the dimensionality of the factorization size_t factorSize_; -private: + private: // Store the square values of the letent vectors matrix MatrixPtr latentVectorsSquare_; // Store the square values of input matrix @@ -65,7 +65,7 @@ private: // Negative identity matrix MatrixPtr negOnes_; -public: + public: explicit FactorizationMachineLayer(const LayerConfig& config) : Layer(config) {} ~FactorizationMachineLayer() {} diff --git a/paddle/gserver/layers/FeatureMapExpandLayer.cpp b/paddle/gserver/layers/FeatureMapExpandLayer.cpp index 81b98da45bc4b9b8ef0723dd6ea2db809860e219..d95f0b9b3d13e8bff635373cb4d5705c2351bd97 100644 --- a/paddle/gserver/layers/FeatureMapExpandLayer.cpp +++ b/paddle/gserver/layers/FeatureMapExpandLayer.cpp @@ -38,11 +38,11 @@ namespace paddle { */ class FeatureMapExpandLayer : public Layer { -private: + private: int numFilters_; bool asRowVector_; -public: + public: explicit FeatureMapExpandLayer(const LayerConfig& config) : Layer(config) {} ~FeatureMapExpandLayer() {} diff --git a/paddle/gserver/layers/FullMatrixProjection.h b/paddle/gserver/layers/FullMatrixProjection.h index 7c4cd1a7066d427f54e1a280a956acb025e6dc16..a27aa4a12327ac39ec3418a849b1230e13f759ee 100644 --- a/paddle/gserver/layers/FullMatrixProjection.h +++ b/paddle/gserver/layers/FullMatrixProjection.h @@ -28,14 +28,14 @@ namespace paddle { * The config file api is full_matrix_projection. */ class FullMatrixProjection : public Projection { -public: + public: FullMatrixProjection(const ProjectionConfig& config, const ParameterPtr& parameter, bool useGpu); virtual void forward(); virtual void backward(const UpdateCallback& callback); -protected: + protected: std::unique_ptr weight_; }; diff --git a/paddle/gserver/layers/FullyConnectedLayer.h b/paddle/gserver/layers/FullyConnectedLayer.h index e66aeeb7334c9c871749196d77474a02ecf82b09..e0f9d6ce55fbdf73e5507032c108c735bf04597b 100644 --- a/paddle/gserver/layers/FullyConnectedLayer.h +++ b/paddle/gserver/layers/FullyConnectedLayer.h @@ -28,11 +28,11 @@ namespace paddle { */ class FullyConnectedLayer : public Layer { -protected: + protected: WeightList weights_; std::unique_ptr biases_; -public: + public: explicit FullyConnectedLayer(const LayerConfig& config) : Layer(config) {} ~FullyConnectedLayer() {} diff --git a/paddle/gserver/layers/GatedRecurrentLayer.h b/paddle/gserver/layers/GatedRecurrentLayer.h index f0a3a823018f3943b0295c172b19d0fe9d0674b4..46508dc977bf1a6fd33dc1fb024bd1aed36a0ff3 100644 --- a/paddle/gserver/layers/GatedRecurrentLayer.h +++ b/paddle/gserver/layers/GatedRecurrentLayer.h @@ -47,7 +47,7 @@ namespace paddle { */ class GatedRecurrentLayer : public Layer, public GruCompute { -public: + public: explicit GatedRecurrentLayer(const LayerConfig& config) : Layer(config) {} bool init(const LayerMap& layerMap, @@ -63,7 +63,7 @@ public: LayerStatePtr getState() override; -protected: + protected: void forwardSequence(int batchSize, size_t numSequences, const int* starts, @@ -79,7 +79,7 @@ protected: MatrixPtr inputValue); void backwardBatch(int batchSize, MatrixPtr inputGrad); -protected: + protected: std::unique_ptr weight_; std::unique_ptr gateWeight_; std::unique_ptr stateWeight_; diff --git a/paddle/gserver/layers/GetOutputLayer.cpp b/paddle/gserver/layers/GetOutputLayer.cpp index f255681f3e678e51f069522f965fd2776680b595..7c1e3c407cca374c7aa238d07e2263c4a142b6a5 100644 --- a/paddle/gserver/layers/GetOutputLayer.cpp +++ b/paddle/gserver/layers/GetOutputLayer.cpp @@ -17,7 +17,7 @@ limitations under the License. */ namespace paddle { class GetOutputLayer : public Layer { -public: + public: explicit GetOutputLayer(const LayerConfig& config) : Layer(config) {} ~GetOutputLayer() {} diff --git a/paddle/gserver/layers/GruCompute.h b/paddle/gserver/layers/GruCompute.h index fb6bc56422002b4d4080ccb8438767b27ceef064..50006325ce9969c4941aaf28604260f0aeb9b97a 100644 --- a/paddle/gserver/layers/GruCompute.h +++ b/paddle/gserver/layers/GruCompute.h @@ -21,7 +21,7 @@ limitations under the License. */ namespace paddle { class GruCompute { -public: + public: void init(LayerConfig &config); template @@ -33,7 +33,7 @@ public: int frameSize, int batchSize = 1); -public: + public: hl_activation_mode_t activeNode_; hl_activation_mode_t activeGate_; }; diff --git a/paddle/gserver/layers/GruStepLayer.cpp b/paddle/gserver/layers/GruStepLayer.cpp index 917c50250c1c04d6c8f113c8d42ef029e1028606..114f287411c2fccbc08b7da4c05462967c81b268 100644 --- a/paddle/gserver/layers/GruStepLayer.cpp +++ b/paddle/gserver/layers/GruStepLayer.cpp @@ -44,13 +44,13 @@ namespace paddle { * The config file api if gru_step_layer. */ class GruStepLayer : public Layer, public GruCompute { -protected: + protected: Argument gate_; Argument resetOutput_; std::unique_ptr weight_; std::unique_ptr bias_; -public: + public: explicit GruStepLayer(const LayerConfig& config) : Layer(config) {} ~GruStepLayer() {} diff --git a/paddle/gserver/layers/HierarchicalSigmoidLayer.h b/paddle/gserver/layers/HierarchicalSigmoidLayer.h index 10e501f1807ef6ba03d326a1bcf257ede0ee850a..73ef252fd5a5443fe065f3b7bd8c49951ae0b4bd 100644 --- a/paddle/gserver/layers/HierarchicalSigmoidLayer.h +++ b/paddle/gserver/layers/HierarchicalSigmoidLayer.h @@ -58,7 +58,7 @@ namespace paddle { * The config file api is hsigmod_layer. */ class HierarchicalSigmoidLayer : public Layer { -public: + public: explicit HierarchicalSigmoidLayer(const LayerConfig& config) : Layer(config) {} bool init(const LayerMap& layerMap, @@ -66,7 +66,7 @@ public: void forward(PassType passType) override; void backward(const UpdateCallback& callback) override; -protected: + protected: /** * The last of inputs is label layer. */ diff --git a/paddle/gserver/layers/IdentityProjection.cpp b/paddle/gserver/layers/IdentityProjection.cpp index 6c70f77acc0c890e11a4929ea013d7745d8bbed0..34e9eb90161f7942c528b70f177e30f301a8f53f 100644 --- a/paddle/gserver/layers/IdentityProjection.cpp +++ b/paddle/gserver/layers/IdentityProjection.cpp @@ -26,7 +26,7 @@ namespace paddle { * The config file api is identity_projection. */ class IdentityProjection : public Projection { -public: + public: IdentityProjection(const ProjectionConfig& config, const ParameterPtr& parameter, bool useGpu); @@ -68,7 +68,7 @@ void IdentityProjection::backward(const UpdateCallback& callback) { * The config file api is identity_projection. */ class IdentityOffsetProjection : public Projection { -public: + public: IdentityOffsetProjection(const ProjectionConfig& config, const ParameterPtr& parameter, bool useGpu); diff --git a/paddle/gserver/layers/InterpolationLayer.cpp b/paddle/gserver/layers/InterpolationLayer.cpp index 0ac92024bc7eddf05ce023708537d0aa7bab6426..509c07cf22c9bcbe9283241b38540162b3dbe26b 100644 --- a/paddle/gserver/layers/InterpolationLayer.cpp +++ b/paddle/gserver/layers/InterpolationLayer.cpp @@ -33,12 +33,12 @@ namespace paddle { */ class InterpolationLayer : public Layer { -protected: + protected: /// weightLast = 1 - weight MatrixPtr weightLast_; MatrixPtr tmpMatrix; -public: + public: explicit InterpolationLayer(const LayerConfig& config) : Layer(config) {} ~InterpolationLayer() {} diff --git a/paddle/gserver/layers/KmaxSeqScoreLayer.cpp b/paddle/gserver/layers/KmaxSeqScoreLayer.cpp index 0ea960902efc10007896b3f4ce915dea79d0d12d..7fd25954efeb9d9e672040f9909198f2ae3c0449 100644 --- a/paddle/gserver/layers/KmaxSeqScoreLayer.cpp +++ b/paddle/gserver/layers/KmaxSeqScoreLayer.cpp @@ -17,14 +17,14 @@ limitations under the License. */ namespace paddle { class KmaxSeqScoreLayer : public Layer { -private: + private: MatrixPtr scores_; size_t beamSize_; void kmaxScorePerSeq(const real* score, real* sortedRes, const ICpuGpuVectorPtr seqStartPos); -public: + public: explicit KmaxSeqScoreLayer(const LayerConfig& config) : Layer(config) {} bool init(const LayerMap& layerMap, diff --git a/paddle/gserver/layers/L2DistanceLayer.h b/paddle/gserver/layers/L2DistanceLayer.h index 97f35daf7860fb3b082ef03203327e09dca67371..44e688e1377145845033d9d5cc3f31f5594a11f6 100644 --- a/paddle/gserver/layers/L2DistanceLayer.h +++ b/paddle/gserver/layers/L2DistanceLayer.h @@ -33,7 +33,7 @@ namespace paddle { */ class L2DistanceLayer : public Layer { -public: + public: explicit L2DistanceLayer(const LayerConfig& config) : Layer(config) {} ~L2DistanceLayer() {} @@ -43,7 +43,7 @@ public: void forward(PassType passType) override; void backward(const UpdateCallback& callback = nullptr) override; -private: + private: // Store the result of subtracting Input2 from Input1 in forward computation, // which will be reused in backward computation. MatrixPtr inputSub_; diff --git a/paddle/gserver/layers/Layer.h b/paddle/gserver/layers/Layer.h index 8da342a00f72ee1196c4af24104ce92c6bbf9f5c..13e20e8316323f9082a9615041584685853aa395 100644 --- a/paddle/gserver/layers/Layer.h +++ b/paddle/gserver/layers/Layer.h @@ -60,7 +60,7 @@ enum PADDLE_DEVICE_ID { * Define necessary variables and functions for every layer. */ class Layer { -protected: + protected: /// Layer config LayerConfig config_; /// whether to use GPU @@ -112,7 +112,7 @@ protected: /// Layer backward function std::vector> backward_; -public: + public: /** * Wait until all input value ready. * Called before Layer::forward() function. @@ -137,7 +137,7 @@ public: */ virtual void markAllInputGrad(); -protected: + protected: /** * Create layer function. Function is called in forward or backward. * \param function, Layer::forward_ or Layer::backward_ @@ -252,7 +252,7 @@ protected: */ void addOutputArgument(int deviceId); -public: + public: explicit Layer(const LayerConfig& config, bool useGpu = FLAGS_use_gpu); virtual ~Layer() {} @@ -490,7 +490,7 @@ public: */ virtual void onPassEnd() {} -protected: + protected: /** * Forward of activation function. */ diff --git a/paddle/gserver/layers/LinearChainCRF.h b/paddle/gserver/layers/LinearChainCRF.h index 1ea4c7e105703b76601499bf3944648cdc98ec99..e802b701d0237bed44adc83273fe53c3e18c92ec 100644 --- a/paddle/gserver/layers/LinearChainCRF.h +++ b/paddle/gserver/layers/LinearChainCRF.h @@ -19,7 +19,7 @@ limitations under the License. */ namespace paddle { class LinearChainCRF { -public: + public: /** * The size of para must be \f$(numClasses + 2) * numClasses\f$. * The first numClasses values of para are for starting weights (\f$a\f$). @@ -71,7 +71,7 @@ public: */ MatrixPtr getXGrad() { return matGrad_; } -protected: + protected: int numClasses_; MatrixPtr a_; MatrixPtr b_; diff --git a/paddle/gserver/layers/LinearChainCTC.h b/paddle/gserver/layers/LinearChainCTC.h index 0b774277dc8cf27f48c6905168cdea047365c99d..5b325a0deb0e9d8df241175159321e52f527f6c4 100644 --- a/paddle/gserver/layers/LinearChainCTC.h +++ b/paddle/gserver/layers/LinearChainCTC.h @@ -20,7 +20,7 @@ limitations under the License. */ namespace paddle { class LinearChainCTC { -public: + public: LinearChainCTC(int numClasses, bool normByTimes); // Calculate the negative log probability as loss @@ -35,7 +35,7 @@ public: int* labelSeq, int labelSeqLen); -protected: + protected: int numClasses_, blank_, totalSegments_, totalTime_; bool normByTimes_; bool isInvalid_; diff --git a/paddle/gserver/layers/LstmCompute.h b/paddle/gserver/layers/LstmCompute.h index b7d55eb1f984d102802cab87ba12ca9c69a2f4be..80fb01cd1885151c8d62a4b5dfdb4ba08327926d 100644 --- a/paddle/gserver/layers/LstmCompute.h +++ b/paddle/gserver/layers/LstmCompute.h @@ -21,7 +21,7 @@ limitations under the License. */ namespace paddle { class LstmCompute { -public: + public: void init(LayerConfig &config); /** @@ -57,7 +57,7 @@ public: hl_lstm_grad grad, int frameSize); -public: + public: hl_activation_mode_t activeNode_; hl_activation_mode_t activeGate_; hl_activation_mode_t activeState_; diff --git a/paddle/gserver/layers/LstmLayer.h b/paddle/gserver/layers/LstmLayer.h index 4568b13ade5555e3cff703ceda1bbce3007c409d..76dfe8146bf67a0b7b4fd4835851fae6ac38d80f 100644 --- a/paddle/gserver/layers/LstmLayer.h +++ b/paddle/gserver/layers/LstmLayer.h @@ -71,7 +71,7 @@ namespace paddle { */ class LstmLayer : public Layer, public LstmCompute { -public: + public: explicit LstmLayer(const LayerConfig &config) : Layer(config) {} bool init(const LayerMap &layerMap, @@ -87,7 +87,7 @@ public: LayerStatePtr getState() override; -protected: + protected: /** * @brief Compute lstm forward one sequence by one sequence. * @param batchSize The batchSize is not equal to the batch_size in @@ -165,7 +165,7 @@ protected: */ void getPrevBatchState(size_t numSequences); -protected: + protected: /// Learned parameters, shape: (size, 4*size). /// The weight ([size, 4*size]) contains \f$W_{hi}, W_{hf}, W_{hc}, W_{ho}\f$. std::unique_ptr weight_; diff --git a/paddle/gserver/layers/LstmStepLayer.cpp b/paddle/gserver/layers/LstmStepLayer.cpp index 8faaa1c4e138fe1ec04b1911449d05528bb5b8b5..c44768ddb2b903763288465325899d86176df73a 100644 --- a/paddle/gserver/layers/LstmStepLayer.cpp +++ b/paddle/gserver/layers/LstmStepLayer.cpp @@ -22,7 +22,7 @@ namespace paddle { * LstmStepLayer used in recurrent layer group. */ class LstmStepLayer : public Layer, public LstmCompute { -protected: + protected: Argument state_; Argument gate_; Argument stateActive_; @@ -30,7 +30,7 @@ protected: MatrixPtr checkIgGrad_, checkFgGrad_, checkOgGrad_; std::unique_ptr weight_; -public: + public: explicit LstmStepLayer(const LayerConfig& config) : Layer(config) {} ~LstmStepLayer() {} diff --git a/paddle/gserver/layers/MDLstmLayer.cpp b/paddle/gserver/layers/MDLstmLayer.cpp index 7cfdb3ff25096ad06c09434cdee48b5f85d650af..22c28157c5a5b19aa54b3151a6c9a4cdcfb01765 100644 --- a/paddle/gserver/layers/MDLstmLayer.cpp +++ b/paddle/gserver/layers/MDLstmLayer.cpp @@ -19,7 +19,7 @@ limitations under the License. */ namespace paddle { class CoordIterator { -public: + public: std::vector dims_; std::vector directions_; std::vector curPos_; @@ -51,7 +51,7 @@ public: } } -public: + public: CoordIterator(std::vector dim, std::vector directions) : dims_(dim), directions_(directions), end_(false) { CHECK_EQ(dims_.size(), directions_.size()); @@ -178,7 +178,7 @@ public: * */ class MDLstmLayer : public LstmLayer { -public: + public: explicit MDLstmLayer(const LayerConfig& config) : LstmLayer(config) {} bool init(const LayerMap& layerMap, @@ -188,13 +188,13 @@ public: void backward(const UpdateCallback& callback) override; -protected: + protected: void forwardOneSequence(int start, CoordIterator& coordIter); void backwardOneSequence(int start, CoordIterator& coordIter); void forwardGate2OutputSequence(int start, CoordIterator& coordIter); void backwardGate2OutputSequence(int start, CoordIterator& coordIter); -protected: + protected: std::vector frameInputGate_; std::vector frameForgetGate_; std::vector frameOutputGate_; diff --git a/paddle/gserver/layers/MKLDNNAddtoLayer.h b/paddle/gserver/layers/MKLDNNAddtoLayer.h index e40e2f2251a1b739958773b8e6dc95a70ed58c76..0b385e804fdbc74c8612031cf415d06f15ce311a 100644 --- a/paddle/gserver/layers/MKLDNNAddtoLayer.h +++ b/paddle/gserver/layers/MKLDNNAddtoLayer.h @@ -25,7 +25,7 @@ namespace paddle { * The config file api is mkldnn_addto */ class MKLDNNAddtoLayer : public MKLDNNLayer { -protected: + protected: // layer size == ic * ih * iw == oc * oh *ow, and can not be changed size_t layerSize_; @@ -38,7 +38,7 @@ protected: std::vector> fwdBias_; std::shared_ptr bwdBias_; -public: + public: explicit MKLDNNAddtoLayer(const LayerConfig& config) : MKLDNNLayer(config) {} ~MKLDNNAddtoLayer() {} @@ -59,7 +59,7 @@ public: void updateWeights(const UpdateCallback& callback) override; -protected: + protected: void resetFwdBuffers(std::vector& inputs, MKLDNNMatrixPtr& bias, MKLDNNMatrixPtr& out); diff --git a/paddle/gserver/layers/MKLDNNBase.h b/paddle/gserver/layers/MKLDNNBase.h index d84e2859407711c13c475a19e140e2f5f51e61c2..786ceaf86086d7c04331641693181809ac019597 100644 --- a/paddle/gserver/layers/MKLDNNBase.h +++ b/paddle/gserver/layers/MKLDNNBase.h @@ -31,7 +31,7 @@ typedef enum { * */ class CPUEngine { -public: + public: static CPUEngine& Instance() { // Thread-safe in C++11. static CPUEngine myInstance; @@ -46,12 +46,12 @@ public: mkldnn::engine& getEngine() { return cpuEngine_; } -protected: + protected: CPUEngine() : cpuEngine_(mkldnn::engine::cpu, 0) {} // CPUEngine() : cpuEngine_(mkldnn::engine::cpu_lazy, 0) {} ~CPUEngine() {} -private: + private: mkldnn::engine cpuEngine_; }; @@ -60,7 +60,7 @@ private: * */ class MKLDNNStream { -public: + public: MKLDNNStream() : ready_(false) { resetState(); } virtual ~MKLDNNStream() {} @@ -89,7 +89,7 @@ public: ready_ = true; } -private: + private: bool ready_; std::shared_ptr stream_; }; diff --git a/paddle/gserver/layers/MKLDNNBatchNormLayer.h b/paddle/gserver/layers/MKLDNNBatchNormLayer.h index 93e182206a1ab1f06087cb808bb266ddea1468c9..9aa20df98f30837e1b80b4269d05d85b7d99ba76 100644 --- a/paddle/gserver/layers/MKLDNNBatchNormLayer.h +++ b/paddle/gserver/layers/MKLDNNBatchNormLayer.h @@ -27,7 +27,7 @@ typedef mkldnn::batch_normalization_backward bn_bwd; * The config file api is mkldnn_batch_norm */ class MKLDNNBatchNormLayer : public MKLDNNLayer { -protected: + protected: // save forward primitive_desc, which can be used backward std::shared_ptr fwdPD_; @@ -62,7 +62,7 @@ protected: MKLDNNMatrixPtr mean_; MKLDNNMatrixPtr var_; -public: + public: explicit MKLDNNBatchNormLayer(const LayerConfig& config) : MKLDNNLayer(config), useGlobalStats_(true), hasInitedWgt_(false) {} @@ -88,7 +88,7 @@ public: void convertWeightsFromPaddle() override; -protected: + protected: void initWeight(); /** * cal moving mean and variance. diff --git a/paddle/gserver/layers/MKLDNNConcatLayer.h b/paddle/gserver/layers/MKLDNNConcatLayer.h index f7abdabfb51df27f8db4e6d4d88c80546eeba248..d7738df6c106c68f55b313f2d119e31c6e444cbf 100644 --- a/paddle/gserver/layers/MKLDNNConcatLayer.h +++ b/paddle/gserver/layers/MKLDNNConcatLayer.h @@ -25,7 +25,7 @@ namespace paddle { * The config file api is mkldnn_concat */ class MKLDNNConcatLayer : public MKLDNNLayer { -protected: + protected: std::vector> bwds_; // input channel numbers std::vector channels_; @@ -35,7 +35,7 @@ protected: // if axis_ == 1, concat channel (default) int axis_; -public: + public: explicit MKLDNNConcatLayer(const LayerConfig& config) : MKLDNNLayer(config), axis_(1) {} @@ -75,7 +75,7 @@ public: return totalSize; } -protected: + protected: void resetFwdBuffers(std::vector& inputs, MKLDNNMatrixPtr& out); void resetFwdPD(std::shared_ptr& pd, diff --git a/paddle/gserver/layers/MKLDNNConvLayer.h b/paddle/gserver/layers/MKLDNNConvLayer.h index 29c8735fbb91e7418797874238eb87759420f181..d399035ed3ae2f411587c1fcf1799bb71c8de63e 100644 --- a/paddle/gserver/layers/MKLDNNConvLayer.h +++ b/paddle/gserver/layers/MKLDNNConvLayer.h @@ -28,7 +28,7 @@ typedef mkldnn::convolution_backward_data conv_bwdData; * The config file api is mkldnn_conv */ class MKLDNNConvLayer : public MKLDNNLayer { -protected: + protected: // padding height and width int ph_, pw_; // stride height and width @@ -59,7 +59,7 @@ protected: std::unique_ptr weight_; std::unique_ptr biases_; -public: + public: explicit MKLDNNConvLayer(const LayerConfig& config) : MKLDNNLayer(config), hasInitedWgt_(false), caffeMode_(true) {} @@ -92,7 +92,7 @@ public: << ", sw: " << sw_ << ", dh: " << dh_ << ", dw: " << dw_; } -protected: + protected: /** * load the dims settings of this conv */ diff --git a/paddle/gserver/layers/MKLDNNFcLayer.h b/paddle/gserver/layers/MKLDNNFcLayer.h index 0d41a4379d677f86f672852fec09b1241009597b..a704066cc818a6b33bd0eed4612d62b674fa72ca 100644 --- a/paddle/gserver/layers/MKLDNNFcLayer.h +++ b/paddle/gserver/layers/MKLDNNFcLayer.h @@ -28,7 +28,7 @@ typedef mkldnn::inner_product_backward_data fc_bwdData; * The config file api is mkldnn_fc */ class MKLDNNFcLayer : public MKLDNNLayer { -protected: + protected: // input layer size, can not be change after init size_t iLayerSize_; // == ic * ih * iw @@ -42,7 +42,7 @@ protected: std::unique_ptr weight_; std::unique_ptr biases_; -public: + public: explicit MKLDNNFcLayer(const LayerConfig& config) : MKLDNNLayer(config), hasInitedWgt_(false) {} @@ -68,7 +68,7 @@ public: void convertWeightsToPaddle() override; -protected: + protected: void resetFwdBuffers(MKLDNNMatrixPtr& in, MKLDNNMatrixPtr& wgt, MKLDNNMatrixPtr& bias, diff --git a/paddle/gserver/layers/MKLDNNLRNLayer.h b/paddle/gserver/layers/MKLDNNLRNLayer.h index b503ee55947294d7c44d1760058f8c26bceed142..028438f2c93b2182318c53cd348351376d491e79 100644 --- a/paddle/gserver/layers/MKLDNNLRNLayer.h +++ b/paddle/gserver/layers/MKLDNNLRNLayer.h @@ -27,7 +27,7 @@ typedef mkldnn::lrn_backward lrn_bwd; * The config file api is mkldnn_lrn */ class MKLDNNLRNLayer : public MKLDNNLayer { -protected: + protected: // save forward primitive_desc, which can be used in backward std::shared_ptr fwdPD_; // according to https://github.com/01org/mkl-dnn/blob/master/tests/gtests/ @@ -37,7 +37,7 @@ protected: int localSize_; float alpha_, beta_; // scale and pow in paddle -public: + public: explicit MKLDNNLRNLayer(const LayerConfig& config) : MKLDNNLayer(config) {} ~MKLDNNLRNLayer() {} @@ -56,7 +56,7 @@ public: std::vector& inputs, MKLDNNMatrixPtr& out) override; -protected: + protected: void resetFwdBuffers(MKLDNNMatrixPtr& in, MKLDNNMatrixPtr& out); void resetFwdPD(std::shared_ptr& pd, MKLDNNMatrixPtr in, diff --git a/paddle/gserver/layers/MKLDNNLayer.h b/paddle/gserver/layers/MKLDNNLayer.h index 4a7eb74ce3a13ed38be3548d8ce34382c594205a..2b164d0d3bc0e1446d7e4d82bb8a713195dbd927 100644 --- a/paddle/gserver/layers/MKLDNNLayer.h +++ b/paddle/gserver/layers/MKLDNNLayer.h @@ -33,7 +33,7 @@ typedef std::shared_ptr MKLDNNLayerPtr; * */ class MKLDNNLayer : public Layer { -protected: + protected: // batch size int bs_; // their sizes are always from the first input layer @@ -95,7 +95,7 @@ protected: // tmp input argument to save input grad, only used to merge grad Argument tmpInArg_; -public: + public: explicit MKLDNNLayer(const LayerConfig& config) : Layer(config), ih_(0), @@ -162,7 +162,7 @@ public: */ void addOutputArgument(int deviceId) { Layer::addOutputArgument(deviceId); } -protected: + protected: /** * Some layers may have different condition to reset the forward. * The function returns the condition that do not need reset forward. @@ -233,7 +233,7 @@ protected: */ void resetMergeGrad(MKLDNNMatrixPtr& out); -protected: + protected: /** * Set deviceId of this layer. */ @@ -340,7 +340,7 @@ protected: } } -private: + private: /** * clear all grad */ diff --git a/paddle/gserver/layers/MKLDNNPoolLayer.h b/paddle/gserver/layers/MKLDNNPoolLayer.h index 12821cda7308602dd2fe834f52c614e6112b7cea..1eb0ee4ad946f61e32b7d4f4fd376dda89d6acf7 100644 --- a/paddle/gserver/layers/MKLDNNPoolLayer.h +++ b/paddle/gserver/layers/MKLDNNPoolLayer.h @@ -27,7 +27,7 @@ typedef mkldnn::pooling_backward pool_bwd; * The config file api is mkldnn_pool */ class MKLDNNPoolLayer : public MKLDNNLayer { -protected: + protected: // padding height and width int ph_, pw_; // stride height and width @@ -44,7 +44,7 @@ protected: // test_pooling_forward.cpp, pool need workspace for backward std::shared_ptr workspace_; -public: + public: explicit MKLDNNPoolLayer(const LayerConfig& config) : MKLDNNLayer(config) {} ~MKLDNNPoolLayer() {} @@ -70,7 +70,7 @@ public: << ", sw: " << sw_; } -protected: + protected: void resetFwdBuffers(MKLDNNMatrixPtr& in, MKLDNNMatrixPtr& out); void resetFwdPD(std::shared_ptr& pd, MKLDNNMatrixPtr in, diff --git a/paddle/gserver/layers/MKLPackedRecurrentLayer.h b/paddle/gserver/layers/MKLPackedRecurrentLayer.h index 37eb362d45215edc736984f8da784fe74bb43f2b..441025a9c9d75786b17db84c74995a96b6a06ea8 100644 --- a/paddle/gserver/layers/MKLPackedRecurrentLayer.h +++ b/paddle/gserver/layers/MKLPackedRecurrentLayer.h @@ -29,7 +29,7 @@ namespace paddle { */ class MKLPackedRecurrentLayer : public RecurrentLayer { -public: + public: explicit MKLPackedRecurrentLayer(const LayerConfig& config) : RecurrentLayer(config) {} @@ -38,7 +38,7 @@ public: void backward(const UpdateCallback& callback) override; -protected: + protected: void forwardBatch(int batchSize, size_t numSequences, const int* starts) override; @@ -47,7 +47,7 @@ protected: size_t numSequences, const int* starts) override; -protected: + protected: /// packed_weight_ contains same data with /// RecurrentLayer::weight_ but is packed std::unique_ptr packed_weight_; diff --git a/paddle/gserver/layers/MKLPackedWeight.h b/paddle/gserver/layers/MKLPackedWeight.h index 28b8a7db7cc3d2be12d6ce9291de1e415cf77bbc..b01a961d007a0e2e343db7b51e50fd3ee776435e 100644 --- a/paddle/gserver/layers/MKLPackedWeight.h +++ b/paddle/gserver/layers/MKLPackedWeight.h @@ -21,7 +21,7 @@ limitations under the License. */ namespace paddle { class MKLPackedWeight { -protected: + protected: /// The pointer of weight real *weight_; /// The pointer of cblas packed gemm to weight @@ -30,7 +30,7 @@ protected: size_t width_; bool transW_; -public: + public: explicit MKLPackedWeight(MatrixPtr weight, bool transW = false) { packedWeight_ = nullptr; weight_ = weight->getData(); @@ -59,7 +59,7 @@ public: dst->getWidth()); } -protected: + protected: void pack_(real *src) { if (!packedWeight_) { packedWeight_ = cblas_sgemm_alloc(CblasBMatrix, 1, width_, height_); diff --git a/paddle/gserver/layers/MaxIdLayer.cpp b/paddle/gserver/layers/MaxIdLayer.cpp index 84e375d7441ce3ccd8a5df94df22d85d104b5d96..eecd4996e962857b09001a1bb36bc027cbaa4308 100644 --- a/paddle/gserver/layers/MaxIdLayer.cpp +++ b/paddle/gserver/layers/MaxIdLayer.cpp @@ -23,11 +23,11 @@ namespace paddle { * The config file api is maxid_layer. */ class MaxIdLayer : public Layer { -private: + private: /// a predetermined number of best states at each level size_t beamSize_; -public: + public: explicit MaxIdLayer(const LayerConfig& config) : Layer(config) {} bool init(const LayerMap& layerMap, diff --git a/paddle/gserver/layers/MaxLayer.h b/paddle/gserver/layers/MaxLayer.h index 9dbc672652dc2670a775f02ecd3a9de9919c8ae0..e46f997c342ce5d6b724629dff6950c4f1680ce8 100644 --- a/paddle/gserver/layers/MaxLayer.h +++ b/paddle/gserver/layers/MaxLayer.h @@ -39,11 +39,11 @@ namespace paddle { */ class MaxLayer : public SequencePoolLayer { -protected: + protected: // maxIndex_[i][j] = k : the value at (i, j) is from input[k]. IVectorPtr maxIndex_; -public: + public: explicit MaxLayer(const LayerConfig& config) : SequencePoolLayer(config) {} bool init(const LayerMap& layerMap, diff --git a/paddle/gserver/layers/MaxOutLayer.h b/paddle/gserver/layers/MaxOutLayer.h index 1fb371836bacb9e02cc32eabfd21bf24165b0734..0eb8674b4c4f3f58b103c6b59ad13931a6992a1b 100644 --- a/paddle/gserver/layers/MaxOutLayer.h +++ b/paddle/gserver/layers/MaxOutLayer.h @@ -29,7 +29,7 @@ namespace paddle { */ class MaxOutLayer : public Layer { -protected: + protected: size_t groups_; size_t imgSizeH_, imgSizeW_; /// outputChannels_ = channels_ / groups_ @@ -38,7 +38,7 @@ protected: size_t featLen_; IVectorPtr maxoutId_; -public: + public: /// return imgSizeH_ * imgSizeW_ * outputChannels_; size_t getSize(); diff --git a/paddle/gserver/layers/MaxPoolWithMaskLayer.h b/paddle/gserver/layers/MaxPoolWithMaskLayer.h index 74cc8acf3515b10257ffb185061344fbcc94a337..c948364f6b83b0de1ee07cc185b69346f5cb1a7e 100644 --- a/paddle/gserver/layers/MaxPoolWithMaskLayer.h +++ b/paddle/gserver/layers/MaxPoolWithMaskLayer.h @@ -23,10 +23,10 @@ namespace paddle { * @brief Basic parent layer of different kinds of pooling */ class MaxPoolWithMaskLayer : public PoolLayer { -protected: + protected: Argument mask_; -public: + public: explicit MaxPoolWithMaskLayer(const LayerConfig& config) : PoolLayer(config) {} diff --git a/paddle/gserver/layers/MixedLayer.h b/paddle/gserver/layers/MixedLayer.h index a1a43c52e4f503178a66ad8aa6c12bec89566081..43ee2bd81854f2dea837734f556c197613f6fdaf 100644 --- a/paddle/gserver/layers/MixedLayer.h +++ b/paddle/gserver/layers/MixedLayer.h @@ -30,7 +30,7 @@ namespace paddle { * The config file api is mixed_layer. */ class MixedLayer : public Layer { -public: + public: explicit MixedLayer(const LayerConfig& config) : Layer(config) {} ~MixedLayer() {} @@ -52,7 +52,7 @@ public: */ LayerStatePtr getState() override; -protected: + protected: std::vector> projections_; std::vector> operators_; /// the matrix size of projection state diff --git a/paddle/gserver/layers/MultiBoxLossLayer.h b/paddle/gserver/layers/MultiBoxLossLayer.h index 9935da56446c1508549906becfd28548d5deecde..a358cded00bb01bfe5d02f9a6d8a24e4b2e51b74 100644 --- a/paddle/gserver/layers/MultiBoxLossLayer.h +++ b/paddle/gserver/layers/MultiBoxLossLayer.h @@ -41,7 +41,7 @@ namespace paddle { */ class MultiBoxLossLayer : public CostLayer { -public: + public: explicit MultiBoxLossLayer(const LayerConfig& config) : CostLayer(config) {} bool init(const LayerMap& layerMap, const ParameterMap& parameterMap); @@ -54,7 +54,7 @@ public: void backwardImp(Matrix& outputValue, Argument& label, Matrix& outputGrad) {} -protected: + protected: inline LayerPtr getPriorBoxLayer() { return inputLayers_[0]; } inline LayerPtr getLabelLayer() { return inputLayers_[1]; } inline LayerPtr getLocInputLayer(size_t index) { @@ -64,7 +64,7 @@ protected: return inputLayers_[2 + inputNum_ + index]; } -protected: + protected: size_t numClasses_; real overlapThreshold_; real negPosRatio_; diff --git a/paddle/gserver/layers/MultinomialSampler.h b/paddle/gserver/layers/MultinomialSampler.h index 1f9e818ee5d21188e3bd39d1225912a1a2ae1598..8cbb229f157c0904e63a696f860ec6739d5167c4 100644 --- a/paddle/gserver/layers/MultinomialSampler.h +++ b/paddle/gserver/layers/MultinomialSampler.h @@ -29,7 +29,7 @@ namespace paddle { * The computational complexity of generate one sample is O(1). */ class MultinomialSampler { -public: + public: MultinomialSampler(const real* prob, int size); //! protobuf always using double. @@ -53,7 +53,7 @@ public: return gen1([&g, this]() { return rand_(g); }); } -protected: + protected: /** * @brief Generation * @param[in] rand rand is a real random number distribution diff --git a/paddle/gserver/layers/MultiplexLayer.cpp b/paddle/gserver/layers/MultiplexLayer.cpp index 82857f8c3ef3e39ec451c1f26bac4996c12350a5..43ecc48cd97fb54d8dc4eb1d87ebf60f5aa040d8 100644 --- a/paddle/gserver/layers/MultiplexLayer.cpp +++ b/paddle/gserver/layers/MultiplexLayer.cpp @@ -37,7 +37,7 @@ namespace paddle { */ class MultiplexLayer : public Layer { -protected: + protected: /** * @brief A struct is used to save the copy information, includes input * layer index and copy size. @@ -64,7 +64,7 @@ protected: /// Temporary matrix pointer to point to output data. MatrixPtr tmpDest_; -public: + public: explicit MultiplexLayer(const LayerConfig& config) : Layer(config) {} ~MultiplexLayer() {} @@ -75,7 +75,7 @@ public: void forward(PassType passType) override; void backward(const UpdateCallback& callback = nullptr) override; -private: + private: /** * @brief Calculate copy info for input layers. */ diff --git a/paddle/gserver/layers/NCELayer.cpp b/paddle/gserver/layers/NCELayer.cpp index d3d7b1fd9ac3c366d11c3060848e89c24a16a70b..cc48fe100f12446f9522078119ae2ead039a82cc 100644 --- a/paddle/gserver/layers/NCELayer.cpp +++ b/paddle/gserver/layers/NCELayer.cpp @@ -54,7 +54,7 @@ class NCELayer : public Layer { IVectorPtr labelIds_; -public: + public: explicit NCELayer(const LayerConfig& config) : Layer(config), numClasses_(config.num_classes()), diff --git a/paddle/gserver/layers/NormLayer.h b/paddle/gserver/layers/NormLayer.h index c89cbbfce9d9e35a6dd300864ee094ef8f9e283a..3807584415f99a7110170748501589dac85eac52 100644 --- a/paddle/gserver/layers/NormLayer.h +++ b/paddle/gserver/layers/NormLayer.h @@ -27,7 +27,7 @@ namespace paddle { * @note Normalize the input in local region */ class NormLayer : public Layer { -public: + public: explicit NormLayer(const LayerConfig& config) : Layer(config) {} bool init(const LayerMap& layerMap, @@ -49,12 +49,12 @@ public: * Need to implement in the futrue. */ class ResponseNormLayer : public NormLayer { -protected: + protected: size_t channels_, size_, outputX_, imgSize_, outputY_, imgSizeY_; real scale_, pow_; MatrixPtr denoms_; -public: + public: explicit ResponseNormLayer(const LayerConfig& config) : NormLayer(config) {} bool init(const LayerMap& layerMap, @@ -76,7 +76,7 @@ public: * Cheng-Yang Fu, Alexander C. Berg. SSD: Single Shot MultiBox Detector */ class CrossChannelNormLayer : public NormLayer { -public: + public: explicit CrossChannelNormLayer(const LayerConfig& config) : NormLayer(config) {} bool init(const LayerMap& layerMap, const ParameterMap& parameterMap); @@ -85,7 +85,7 @@ public: MatrixPtr createSampleMatrix(MatrixPtr data, size_t iter, size_t spatialDim); MatrixPtr createSpatialMatrix(MatrixPtr data, size_t iter, size_t spatialDim); -protected: + protected: size_t channels_; std::unique_ptr scale_; MatrixPtr scaleDiff_; diff --git a/paddle/gserver/layers/NormProjectionLayer.h b/paddle/gserver/layers/NormProjectionLayer.h index 898b5823a9011c4b66e045c54afba070dd5cf772..64803a1603599f2e393ec772a32d64f4d271fe71 100644 --- a/paddle/gserver/layers/NormProjectionLayer.h +++ b/paddle/gserver/layers/NormProjectionLayer.h @@ -28,7 +28,7 @@ class CMRProjectionNormLayer : public ResponseNormLayer { size_t imgSizeH_, imgSizeW_; size_t outputH_, outputW_; -public: + public: explicit CMRProjectionNormLayer(const LayerConfig& config) : ResponseNormLayer(config) {} @@ -41,7 +41,7 @@ public: void forward(PassType passType) override; void backward(const UpdateCallback& callback = nullptr) override; -protected: + protected: TensorShape shape_; }; } // namespace paddle diff --git a/paddle/gserver/layers/Operator.h b/paddle/gserver/layers/Operator.h index a620926cccd3004d7bef57976047a190b4b566e2..42d525ef3e4534acea7512d5ecdbe8a0e1d110d9 100644 --- a/paddle/gserver/layers/Operator.h +++ b/paddle/gserver/layers/Operator.h @@ -34,7 +34,7 @@ namespace paddle { * @note: Operator can't have parameters. */ class Operator { -public: + public: static Operator* create(const OperatorConfig& config, bool useGpu); Operator(const OperatorConfig& config, bool useGpu) @@ -81,7 +81,7 @@ public: */ virtual LayerStatePtr getState() { return nullptr; } -protected: + protected: /// Config of operator OperatorConfig config_; bool useGpu_; diff --git a/paddle/gserver/layers/OuterProdLayer.cpp b/paddle/gserver/layers/OuterProdLayer.cpp index 75f4abf93e5db11dc688f8f2e0b2a36bf70fbccc..11a910f3316114b309efe9007a156e842b3d6229 100644 --- a/paddle/gserver/layers/OuterProdLayer.cpp +++ b/paddle/gserver/layers/OuterProdLayer.cpp @@ -28,12 +28,12 @@ namespace paddle { */ class OuterProdLayer : public Layer { -protected: + protected: MatrixPtr tmpMtx0; MatrixPtr tmpRow0; MatrixPtr tmpRow1; -public: + public: explicit OuterProdLayer(const LayerConfig& config) : Layer(config) {} ~OuterProdLayer() {} diff --git a/paddle/gserver/layers/PadLayer.h b/paddle/gserver/layers/PadLayer.h index 7e09d7f8a0d4dfd5300298ad0514b69781d87016..46b8a595978489c630b3ff2429ecb19d7c12521a 100644 --- a/paddle/gserver/layers/PadLayer.h +++ b/paddle/gserver/layers/PadLayer.h @@ -24,7 +24,7 @@ namespace paddle { * the 4th dimenstion according padc_, padh_ and padw_. */ class PadLayer : public Layer { -public: + public: explicit PadLayer(const LayerConfig& config) : Layer(config) {} ~PadLayer() {} @@ -34,7 +34,7 @@ public: void forward(PassType passType) override; void backward(const UpdateCallback& callback = nullptr) override; -protected: + protected: void setOutDims(const size_t batchSize); void setTensorDim(const size_t batchSize); diff --git a/paddle/gserver/layers/ParameterReluLayer.h b/paddle/gserver/layers/ParameterReluLayer.h index 3725fa4a1199285b703590255af492ebffdaab2c..4553413fcdbecbc83e1f50e8ffbe874fdf05d828 100644 --- a/paddle/gserver/layers/ParameterReluLayer.h +++ b/paddle/gserver/layers/ParameterReluLayer.h @@ -36,7 +36,7 @@ namespace paddle { */ class ParameterReluLayer : public Layer { -protected: + protected: std::unique_ptr weight_; /** @@ -51,7 +51,7 @@ protected: */ size_t partialSum_; -public: + public: explicit ParameterReluLayer(const LayerConfig& config) : Layer(config) {} ~ParameterReluLayer() {} diff --git a/paddle/gserver/layers/Pool3DLayer.h b/paddle/gserver/layers/Pool3DLayer.h index 59ee73f7cb9fb4287c12f3c7d0cacfc812484770..32605f8b7028cfb4909c885e83017a8cffa79575 100644 --- a/paddle/gserver/layers/Pool3DLayer.h +++ b/paddle/gserver/layers/Pool3DLayer.h @@ -26,7 +26,7 @@ namespace paddle { * Pools the input within regions */ class Pool3DLayer : public Layer { -public: + public: explicit Pool3DLayer(const LayerConfig& config) : Layer(config) {} ~Pool3DLayer() {} @@ -36,7 +36,7 @@ public: void backward(const UpdateCallback& callback) override; size_t getSize(); -protected: + protected: int channels_; int sizeX_, sizeY_, sizeZ_; int strideW_, strideH_, strideD_; diff --git a/paddle/gserver/layers/PoolLayer.h b/paddle/gserver/layers/PoolLayer.h index 58d5fb0a095e8326f9b6f9cb2a97bb88022ceed8..99f8f148e2eb00f7e431e7d8c5acbf9e27574017 100644 --- a/paddle/gserver/layers/PoolLayer.h +++ b/paddle/gserver/layers/PoolLayer.h @@ -26,7 +26,7 @@ namespace paddle { * Pools the input within regions */ class PoolLayer : public Layer { -protected: + protected: size_t channels_, sizeX_, stride_, outputX_, imgSize_; int confPadding_; @@ -40,7 +40,7 @@ protected: bool excludeMode_; -public: + public: explicit PoolLayer(const LayerConfig& config) : Layer(config) {} /** diff --git a/paddle/gserver/layers/PoolProjection.h b/paddle/gserver/layers/PoolProjection.h index c99287dbf0f4503c180b9b4e9e46abafa67bf64d..8004cc1550337160b7f022c97a23ed8eb9d43ca4 100644 --- a/paddle/gserver/layers/PoolProjection.h +++ b/paddle/gserver/layers/PoolProjection.h @@ -20,7 +20,7 @@ limitations under the License. */ namespace paddle { class PoolProjection : public Projection { -protected: + protected: size_t imgSizeY_, imgSize_; size_t outputY_, outputX_; size_t strideY_, stride_; @@ -30,7 +30,7 @@ protected: std::string poolType_; bool excludeMode_; -public: + public: PoolProjection(const ProjectionConfig& config, ParameterPtr parameter, bool useGpu); @@ -45,7 +45,7 @@ public: }; class MaxPoolProjection : public PoolProjection { -public: + public: MaxPoolProjection(const ProjectionConfig& config, ParameterPtr parameter, bool useGpu) @@ -56,7 +56,7 @@ public: }; class AvgPoolProjection : public PoolProjection { -public: + public: AvgPoolProjection(const ProjectionConfig& config, ParameterPtr parameter, bool useGpu) diff --git a/paddle/gserver/layers/PoolProjectionLayer.h b/paddle/gserver/layers/PoolProjectionLayer.h index 5a97a7769aaeebcfd4fe2c10d8ac0cc8892f68e3..9ad144cc2ad426caa522bf1061a750d47e64a755 100644 --- a/paddle/gserver/layers/PoolProjectionLayer.h +++ b/paddle/gserver/layers/PoolProjectionLayer.h @@ -24,13 +24,13 @@ namespace paddle { * @brief Basic parent layer of different kinds of pooling */ class PoolProjectionLayer : public PoolLayer { -protected: + protected: size_t imgSizeH_, imgSizeW_; size_t outputH_, outputW_; std::unique_ptr poolProjection_; ProjectionConfig projectionConfig_; -public: + public: explicit PoolProjectionLayer(const LayerConfig& config) : PoolLayer(config) { PoolConfig* conf = projectionConfig_.mutable_pool_conf(); *conf = config_.inputs(0).pool_conf(); diff --git a/paddle/gserver/layers/PowerLayer.cpp b/paddle/gserver/layers/PowerLayer.cpp index 18f650fcdaded5ad7199510594b873fc18c3d7b5..7e8d60db8fe588026c6040099745c3aefd7237b5 100644 --- a/paddle/gserver/layers/PowerLayer.cpp +++ b/paddle/gserver/layers/PowerLayer.cpp @@ -32,10 +32,10 @@ namespace paddle { */ class PowerLayer : public Layer { -protected: + protected: MatrixPtr tmpMtx; -public: + public: explicit PowerLayer(const LayerConfig& config) : Layer(config) {} ~PowerLayer() {} diff --git a/paddle/gserver/layers/PrintLayer.cpp b/paddle/gserver/layers/PrintLayer.cpp index 5a527d598dd5e11ae0b74a32c9b9884e73ed45a8..6fbcc447f92208439bddd14d421d62cab30d81f4 100644 --- a/paddle/gserver/layers/PrintLayer.cpp +++ b/paddle/gserver/layers/PrintLayer.cpp @@ -17,7 +17,7 @@ limitations under the License. */ namespace paddle { class PrintLayer : public Layer { -public: + public: explicit PrintLayer(const LayerConfig& config) : Layer(config) {} void forward(PassType passType) override { diff --git a/paddle/gserver/layers/PriorBox.cpp b/paddle/gserver/layers/PriorBox.cpp index af2cc05a954b3a6857c1015104a57339282840b8..39d2c2d737fa90737635efdb209610e156c8662f 100644 --- a/paddle/gserver/layers/PriorBox.cpp +++ b/paddle/gserver/layers/PriorBox.cpp @@ -28,7 +28,7 @@ namespace paddle { */ class PriorBoxLayer : public Layer { -public: + public: // NOLINT explicit PriorBoxLayer(const LayerConfig& config) : Layer(config) {} bool init(const LayerMap& layerMap, const ParameterMap& parameterMap) override; @@ -36,7 +36,7 @@ public: void forward(PassType passType) override; void backward(const UpdateCallback& callback) override {} -protected: + protected: // NOLINT int numPriors_; std::vector minSize_; std::vector maxSize_; @@ -109,11 +109,18 @@ void PriorBoxLayer::forward(PassType passType) { real boxWidth = minSize; real boxHeight = minSize; - // priors with different aspect ratios - for (size_t r = 0; r < aspectRatio_.size(); r++) { - real ar = aspectRatio_[r]; - boxWidth = minSize * sqrt(ar); - boxHeight = minSize / sqrt(ar); + // first prior: aspect_ratio == 1.0, compatible to old logic + tmpPtr[idx++] = (centerX - boxWidth / 2.) / imageWidth; + tmpPtr[idx++] = (centerY - boxHeight / 2.) / imageHeight; + tmpPtr[idx++] = (centerX + boxWidth / 2.) / imageWidth; + tmpPtr[idx++] = (centerY + boxHeight / 2.) / imageHeight; + // set the variance. + for (int t = 0; t < 4; t++) tmpPtr[idx++] = variance_[t]; + + if (maxSize_.size() > 0) { + // square prior with size sqrt(minSize * maxSize) + real maxSize = maxSize_[s]; + boxWidth = boxHeight = sqrt(minSize * maxSize); tmpPtr[idx++] = (centerX - boxWidth / 2.) / imageWidth; tmpPtr[idx++] = (centerY - boxHeight / 2.) / imageHeight; tmpPtr[idx++] = (centerX + boxWidth / 2.) / imageWidth; @@ -122,10 +129,14 @@ void PriorBoxLayer::forward(PassType passType) { for (int t = 0; t < 4; t++) tmpPtr[idx++] = variance_[t]; } - if (maxSize_.size() > 0) { - // square prior with size sqrt(minSize * maxSize) - real maxSize = maxSize_[s]; - boxWidth = boxHeight = sqrt(minSize * maxSize); + // priors with different aspect ratios + for (size_t r = 0; r < aspectRatio_.size(); r++) { + real ar = aspectRatio_[r]; + if (fabs(ar - 1.0) < 1e-6) { + continue; + } + boxWidth = minSize * sqrt(ar); + boxHeight = minSize / sqrt(ar); tmpPtr[idx++] = (centerX - boxWidth / 2.) / imageWidth; tmpPtr[idx++] = (centerY - boxHeight / 2.) / imageHeight; tmpPtr[idx++] = (centerX + boxWidth / 2.) / imageWidth; diff --git a/paddle/gserver/layers/Projection.h b/paddle/gserver/layers/Projection.h index 1f0b96c79ec7313cd9c5ff9139a455b3269b222b..88a41355cfce711e1e9522655058d0f1198e4e76 100644 --- a/paddle/gserver/layers/Projection.h +++ b/paddle/gserver/layers/Projection.h @@ -37,7 +37,7 @@ namespace paddle { * to output Argument. */ class Projection { -public: + public: static Projection* create(const ProjectionConfig& config, ParameterPtr parameter, bool useGpu); @@ -98,7 +98,7 @@ public: */ size_t getOutputSize() const { return config_.output_size(); } -protected: + protected: /** * Create layer function. Function is called in forward or backward. * \param function, Layer::forward_ or Layer::backward_ @@ -119,7 +119,7 @@ protected: func->init(config); } -protected: + protected: /// Config of projection ProjectionConfig config_; /// Parameter of projection diff --git a/paddle/gserver/layers/ROIPoolLayer.h b/paddle/gserver/layers/ROIPoolLayer.h index b1735e9748dc3956aade010f33303b55d4f9f439..801a9b3aebe6d718ea38b76246a6056891d0b1f6 100644 --- a/paddle/gserver/layers/ROIPoolLayer.h +++ b/paddle/gserver/layers/ROIPoolLayer.h @@ -33,7 +33,7 @@ namespace paddle { */ class ROIPoolLayer : public Layer { -protected: + protected: size_t channels_; size_t width_; size_t height_; @@ -44,7 +44,7 @@ protected: // Since there is no int matrix, use real maxtrix instead. MatrixPtr maxIdxs_; -public: + public: explicit ROIPoolLayer(const LayerConfig& config) : Layer(config) {} bool init(const LayerMap& layerMap, diff --git a/paddle/gserver/layers/RecurrentLayer.h b/paddle/gserver/layers/RecurrentLayer.h index 8fd4fe6b78ae6474f3cfcec605f25b72af8295bb..94e633e65777aad540738ea67ea1b4e03dd75954 100644 --- a/paddle/gserver/layers/RecurrentLayer.h +++ b/paddle/gserver/layers/RecurrentLayer.h @@ -40,7 +40,7 @@ namespace paddle { */ class RecurrentLayer : public Layer { -public: + public: explicit RecurrentLayer(const LayerConfig& config) : Layer(config) {} bool init(const LayerMap& layerMap, @@ -56,7 +56,7 @@ public: LayerStatePtr getState() override; -protected: + protected: /** * @brief If user do not set --rnn_use_batch=true, it will * compute rnn forward one sequence by one sequence in default. @@ -110,7 +110,7 @@ protected: size_t numSequences, const int* starts); -protected: + protected: std::unique_ptr weight_; std::unique_ptr bias_; diff --git a/paddle/gserver/layers/RecurrentLayerGroup.cpp b/paddle/gserver/layers/RecurrentLayerGroup.cpp index 44b57185c5a5fa7703ca477b990a73cdad2c2aa1..6694e8f2996fdd2c98da1507e5fb3b90b271c850 100644 --- a/paddle/gserver/layers/RecurrentLayerGroup.cpp +++ b/paddle/gserver/layers/RecurrentLayerGroup.cpp @@ -27,7 +27,7 @@ namespace paddle { * between RecurrentLayerGroupBegin and RecurrentLayerGroupEnd. */ class RecurrentLayerGroup : public Layer { -public: + public: explicit RecurrentLayerGroup(const LayerConfig& config) : Layer(config) {} void initSubNetwork(NeuralNetwork* rootNetwork, @@ -58,7 +58,7 @@ public: callback(*network_); } -private: + private: std::unique_ptr network_; }; diff --git a/paddle/gserver/layers/ResizeLayer.cpp b/paddle/gserver/layers/ResizeLayer.cpp index 831f4c3b7e103bc51d870cfa44616980adca08e8..d4ae9945934a40719d253d4b53915530423448af 100644 --- a/paddle/gserver/layers/ResizeLayer.cpp +++ b/paddle/gserver/layers/ResizeLayer.cpp @@ -24,7 +24,7 @@ namespace paddle { * resize matrix: (height * width / size) * size */ class ResizeLayer : public Layer { -public: + public: explicit ResizeLayer(const LayerConfig& config) : Layer(config) {} bool init(const LayerMap& layerMap, diff --git a/paddle/gserver/layers/RotateLayer.h b/paddle/gserver/layers/RotateLayer.h index 3b619921ab741e1236a495e497e18e265bd6e110..7ecbff20167dd95f782f2d61dc34697ab3273934 100644 --- a/paddle/gserver/layers/RotateLayer.h +++ b/paddle/gserver/layers/RotateLayer.h @@ -32,7 +32,7 @@ namespace paddle { */ class RotateLayer : public Layer { -public: + public: explicit RotateLayer(const LayerConfig& config) : Layer(config) {} bool init(const LayerMap& layerMap, const ParameterMap& parameterMap); @@ -40,7 +40,7 @@ public: void forward(PassType passType); void backward(const UpdateCallback& callback = nullptr); -private: + private: int batchSize_; int size_; int height_; diff --git a/paddle/gserver/layers/RowConvLayer.h b/paddle/gserver/layers/RowConvLayer.h index ba0af1de68a5f77d9ffefac6ef5193bb9d1b4f83..3b74df0b1af5caef1a1abd3d3c5b3ae3b67c429b 100644 --- a/paddle/gserver/layers/RowConvLayer.h +++ b/paddle/gserver/layers/RowConvLayer.h @@ -22,7 +22,7 @@ namespace paddle { * \brief Row Convolution Layer. */ class RowConvLayer : public Layer { -public: + public: explicit RowConvLayer(const LayerConfig& config) : Layer(config) {} ~RowConvLayer() {} @@ -32,7 +32,7 @@ public: void forward(PassType passType) override; void backward(const UpdateCallback& callback = nullptr) override; -protected: + protected: // Row convolution weight, context_lenght_ * fan_out. // fan_out is the size of output feature. std::unique_ptr weight_; diff --git a/paddle/gserver/layers/RowL2NormLayer.cpp b/paddle/gserver/layers/RowL2NormLayer.cpp index 7ff0c9bae927cae2bc6a332bc0bde013e07edd0a..d5e6e10a0276adb74ec31c13d9e8acc77414a85b 100644 --- a/paddle/gserver/layers/RowL2NormLayer.cpp +++ b/paddle/gserver/layers/RowL2NormLayer.cpp @@ -26,12 +26,12 @@ namespace paddle { */ class RowL2NormLayer : public Layer { -protected: + protected: MatrixPtr inSquare_; MatrixPtr l2NormReciprocal_; MatrixPtr dotSum_; -public: + public: explicit RowL2NormLayer(const LayerConfig& config) : Layer(config) {} bool init(const LayerMap& layerMap, diff --git a/paddle/gserver/layers/SamplingIdLayer.cpp b/paddle/gserver/layers/SamplingIdLayer.cpp index 2edd915d226edfd7e48df1a066d5a6f51f259511..dbce63588126c012e3b9713e8be749e0001ddec7 100644 --- a/paddle/gserver/layers/SamplingIdLayer.cpp +++ b/paddle/gserver/layers/SamplingIdLayer.cpp @@ -31,7 +31,7 @@ class SamplingIdLayer : public Layer { std::uniform_real_distribution rand1_; std::vector tmpCpuInput_; -public: + public: explicit SamplingIdLayer(const LayerConfig& config) : Layer(config), rand1_(0, 1) {} diff --git a/paddle/gserver/layers/ScaleShiftLayer.cpp b/paddle/gserver/layers/ScaleShiftLayer.cpp index 799d1fe51a65da10bef637894931627315daf0a2..8af78a2e27d2b50572f8bdd6e98696f3d1967eb1 100644 --- a/paddle/gserver/layers/ScaleShiftLayer.cpp +++ b/paddle/gserver/layers/ScaleShiftLayer.cpp @@ -30,11 +30,11 @@ namespace paddle { */ class ScaleShiftLayer : public Layer { -protected: + protected: std::unique_ptr scale_; std::unique_ptr offset_; -public: + public: explicit ScaleShiftLayer(const LayerConfig& config) : Layer(config) {} bool init(const LayerMap& layerMap, diff --git a/paddle/gserver/layers/ScaleSubRegionLayer.h b/paddle/gserver/layers/ScaleSubRegionLayer.h index 6e861be4858cfc21a42ef7293652d5cdf81be5f5..fe431698bc6cd5e52e2c545756b40be8b307e644 100644 --- a/paddle/gserver/layers/ScaleSubRegionLayer.h +++ b/paddle/gserver/layers/ScaleSubRegionLayer.h @@ -29,7 +29,7 @@ namespace paddle { * region. */ class ScaleSubRegionLayer : public Layer { -public: + public: explicit ScaleSubRegionLayer(const LayerConfig& config) : Layer(config) {} ~ScaleSubRegionLayer() {} @@ -40,7 +40,7 @@ public: void backward(const UpdateCallback& callback = nullptr); -protected: + protected: TensorShape shape_; TensorShape indicesShape_; size_t imgH_; diff --git a/paddle/gserver/layers/ScalingLayer.cpp b/paddle/gserver/layers/ScalingLayer.cpp index 1d98a7373d172d40cddc9b4611cb00434f17e00b..15e07daebee194a789da52d37a192e031348300c 100644 --- a/paddle/gserver/layers/ScalingLayer.cpp +++ b/paddle/gserver/layers/ScalingLayer.cpp @@ -32,7 +32,7 @@ namespace paddle { */ class ScalingLayer : public Layer { -public: + public: explicit ScalingLayer(const LayerConfig& config) : Layer(config) {} ~ScalingLayer() {} diff --git a/paddle/gserver/layers/ScalingProjection.cpp b/paddle/gserver/layers/ScalingProjection.cpp index 99b5b68f543842d23f20b626fddd66b677ebe059..4d871cafc4d0194a61044d76a766236209c33d47 100644 --- a/paddle/gserver/layers/ScalingProjection.cpp +++ b/paddle/gserver/layers/ScalingProjection.cpp @@ -17,7 +17,7 @@ limitations under the License. */ namespace paddle { class ScalingProjection : public Projection { -public: + public: ScalingProjection(const ProjectionConfig& config, const ParameterPtr& parameter, bool useGpu) @@ -48,7 +48,7 @@ public: } } -protected: + protected: std::unique_ptr weight_; }; diff --git a/paddle/gserver/layers/SelectiveFullyConnectedLayer.h b/paddle/gserver/layers/SelectiveFullyConnectedLayer.h index 81564074185a5d9fc80d4d3a64af998098ab5472..4b32ce8b162c2a8b1a6c34adc0885a7701f5f91e 100644 --- a/paddle/gserver/layers/SelectiveFullyConnectedLayer.h +++ b/paddle/gserver/layers/SelectiveFullyConnectedLayer.h @@ -33,11 +33,11 @@ namespace paddle { * The config file api is selective_fc_layer. */ class SelectiveFullyConnectedLayer : public Layer { -protected: + protected: WeightList weights_; std::unique_ptr biases_; -private: + private: /** * Get selected columns each forward. */ @@ -60,7 +60,7 @@ private: /// if true, means output_.value is the same as Fc Layer bool fullOutput_; -public: + public: explicit SelectiveFullyConnectedLayer(const LayerConfig& config) : Layer(config), selCols_(nullptr) {} @@ -94,7 +94,7 @@ public: void forward(PassType passType) override; void backward(const UpdateCallback& callback = nullptr) override; -private: + private: /** * @brief Make SelectiveFC act as FullyConnectedLayer */ diff --git a/paddle/gserver/layers/SequenceConcatLayer.cpp b/paddle/gserver/layers/SequenceConcatLayer.cpp index cf573f3f33fcd70c6768b164f158cb1f545414fc..c84c3ce4f080cc19f4937f04585accb5b2b347f9 100644 --- a/paddle/gserver/layers/SequenceConcatLayer.cpp +++ b/paddle/gserver/layers/SequenceConcatLayer.cpp @@ -29,10 +29,10 @@ namespace paddle { */ class SequenceConcatLayer : public Layer { -protected: + protected: std::unique_ptr biases_; -public: + public: explicit SequenceConcatLayer(const LayerConfig& config) : Layer(config) {} ~SequenceConcatLayer() {} diff --git a/paddle/gserver/layers/SequenceLastInstanceLayer.cpp b/paddle/gserver/layers/SequenceLastInstanceLayer.cpp index 6c4ae775c16ac76e237fb8f8ee5ec9ed8f11802e..28d0a9296d4accd4152e886ccae12a776fdb8f7f 100644 --- a/paddle/gserver/layers/SequenceLastInstanceLayer.cpp +++ b/paddle/gserver/layers/SequenceLastInstanceLayer.cpp @@ -38,12 +38,12 @@ namespace paddle { */ class SequenceLastInstanceLayer : public SequencePoolLayer { -protected: + protected: MatrixPtr tmpSrc_; MatrixPtr tmpDest_; std::vector instanceIds_; -public: + public: explicit SequenceLastInstanceLayer(const LayerConfig& config) : SequencePoolLayer(config) {} diff --git a/paddle/gserver/layers/SequencePoolLayer.h b/paddle/gserver/layers/SequencePoolLayer.h index 254e4cc6b3aacf21565cb03e5bdb52a2beb9fea8..01183060afd58376bb718dda64d8106cce4899f9 100644 --- a/paddle/gserver/layers/SequencePoolLayer.h +++ b/paddle/gserver/layers/SequencePoolLayer.h @@ -41,7 +41,7 @@ namespace paddle { */ class SequencePoolLayer : public Layer { -protected: + protected: int type_; std::unique_ptr biases_; enum SequenceLevel { kNonSeq = 0, kSeq = 1 }; @@ -51,7 +51,7 @@ protected: // Whether the input sequence is reversed or not. bool reversed_ = false; -public: + public: explicit SequencePoolLayer(const LayerConfig& config) : Layer(config) {} bool init(const LayerMap& layerMap, diff --git a/paddle/gserver/layers/SequenceReshapeLayer.cpp b/paddle/gserver/layers/SequenceReshapeLayer.cpp index fb96669917236b98809f1cda0d023600f1e76731..319310af8c4ac3bdefd814ad05b7fde6070f2340 100644 --- a/paddle/gserver/layers/SequenceReshapeLayer.cpp +++ b/paddle/gserver/layers/SequenceReshapeLayer.cpp @@ -29,12 +29,12 @@ namespace paddle { */ class SequenceReshapeLayer : public Layer { -protected: + protected: std::unique_ptr biases_; MatrixPtr reshapedOutputGrad; -public: + public: explicit SequenceReshapeLayer(const LayerConfig& config) : Layer(config) {} bool init(const LayerMap& layerMap, diff --git a/paddle/gserver/layers/SequenceSliceLayer.cpp b/paddle/gserver/layers/SequenceSliceLayer.cpp index 1b7c33477ea64c1cdb7c8e85d7a5302b299d7552..a6d810b583aab6e44faa583795686f06e17beeb9 100644 --- a/paddle/gserver/layers/SequenceSliceLayer.cpp +++ b/paddle/gserver/layers/SequenceSliceLayer.cpp @@ -21,7 +21,7 @@ limitations under the License. */ namespace paddle { class SequenceSliceLayer : public Layer { -public: + public: explicit SequenceSliceLayer(const LayerConfig& config) : Layer(config) {} bool init(const LayerMap& layerMap, @@ -30,7 +30,7 @@ public: void forward(PassType passType) override; void backward(const UpdateCallback& callback = nullptr) override; -private: + private: /* * TODO(caoying) * In PaddePaddle, currently all matrices are real number types, diff --git a/paddle/gserver/layers/SequenceToBatch.h b/paddle/gserver/layers/SequenceToBatch.h index 8743a5ef10f61970d3d48b105b9da29bcd10ba83..5200e702d9bc947746567c19ca7d552750828131 100644 --- a/paddle/gserver/layers/SequenceToBatch.h +++ b/paddle/gserver/layers/SequenceToBatch.h @@ -39,7 +39,7 @@ namespace paddle { * */ class SequenceToBatch { -public: + public: explicit SequenceToBatch(bool useGpu) : useGpu_(useGpu) {} /* resize and calculate the batchIndex_ */ @@ -82,7 +82,7 @@ public: numBatch_ = seq2batch.numBatch_; } -protected: + protected: void sequence2BatchCopy(Matrix &batch, Matrix &sequence, IVector &seq2BatchIdx, diff --git a/paddle/gserver/layers/SliceProjection.cpp b/paddle/gserver/layers/SliceProjection.cpp index 5627ad1eb3a49a73261bc2197cbd3735489509d2..b474f2db759adfad337f9485a5a38588b6839c54 100644 --- a/paddle/gserver/layers/SliceProjection.cpp +++ b/paddle/gserver/layers/SliceProjection.cpp @@ -44,14 +44,14 @@ namespace paddle { * The config file api is slice_projection. */ class SliceProjection : public Projection { -public: + public: SliceProjection(const ProjectionConfig& config, const ParameterPtr& parameter, bool useGpu); virtual void forward(); virtual void backward(const UpdateCallback& callback); -protected: + protected: std::vector> slices_; }; diff --git a/paddle/gserver/layers/SlopeInterceptLayer.cpp b/paddle/gserver/layers/SlopeInterceptLayer.cpp index c94a07e5da7442bba1ce7e9c09c4ffea3e5cd4ac..f7f4735c1b72d4ac6540714573fd7e15ef99ea5b 100644 --- a/paddle/gserver/layers/SlopeInterceptLayer.cpp +++ b/paddle/gserver/layers/SlopeInterceptLayer.cpp @@ -36,7 +36,7 @@ namespace paddle { */ class SlopeInterceptLayer : public Layer { -public: + public: explicit SlopeInterceptLayer(const LayerConfig& config) : Layer(config) {} bool init(const LayerMap& layerMap, diff --git a/paddle/gserver/layers/SpatialPyramidPoolLayer.h b/paddle/gserver/layers/SpatialPyramidPoolLayer.h index 6cb5fdf83e2b88ce4adb392807a1fdbac253c51c..421bdfe09c46f656f500daff195c755274bf8bb7 100644 --- a/paddle/gserver/layers/SpatialPyramidPoolLayer.h +++ b/paddle/gserver/layers/SpatialPyramidPoolLayer.h @@ -29,7 +29,7 @@ namespace paddle { */ class SpatialPyramidPoolLayer : public Layer { -protected: + protected: size_t channels_; size_t imgSizeW_; size_t imgSizeH_; @@ -40,7 +40,7 @@ protected: std::vector projOutput_; std::vector> projCol_; -public: + public: explicit SpatialPyramidPoolLayer(const LayerConfig& config) : Layer(config) {} bool init(const LayerMap& layerMap, diff --git a/paddle/gserver/layers/SubNestedSequenceLayer.cpp b/paddle/gserver/layers/SubNestedSequenceLayer.cpp index db240ab0c96510263d90b291f6396ac51a73fbbd..e2bb00bbfacb26dc736a63877119b379f22b5983 100644 --- a/paddle/gserver/layers/SubNestedSequenceLayer.cpp +++ b/paddle/gserver/layers/SubNestedSequenceLayer.cpp @@ -21,7 +21,7 @@ limitations under the License. */ namespace paddle { class SubNestedSequenceLayer : public Layer { -public: + public: explicit SubNestedSequenceLayer(const LayerConfig& config) : Layer(config) {} bool init(const LayerMap& layerMap, @@ -30,7 +30,7 @@ public: void forward(PassType passType) override; void backward(const UpdateCallback& callback = nullptr) override; -private: + private: /* * This functions generates the indices of rows in a batch according to the * indices of selected sub-sequence in each sequence. diff --git a/paddle/gserver/layers/SubSequenceLayer.cpp b/paddle/gserver/layers/SubSequenceLayer.cpp index 808627f09273950bb6f52a4a6e497bcb8ea170f7..ba49f5710f9d0bb985cf1e80d5c4a972d8f046a6 100644 --- a/paddle/gserver/layers/SubSequenceLayer.cpp +++ b/paddle/gserver/layers/SubSequenceLayer.cpp @@ -27,12 +27,12 @@ namespace paddle { */ class SubSequenceLayer : public Layer { -protected: + protected: std::unique_ptr biases_; MatrixPtr tmpSrc_; MatrixPtr tmpDest_; -public: + public: explicit SubSequenceLayer(const LayerConfig& config) : Layer(config) {} bool init(const LayerMap& layerMap, diff --git a/paddle/gserver/layers/SumToOneNormLayer.cpp b/paddle/gserver/layers/SumToOneNormLayer.cpp index ffbe14925300ad1ffbd33f43a6c0afadddd231e6..00764717e8b6be30230e44626974033e929352da 100644 --- a/paddle/gserver/layers/SumToOneNormLayer.cpp +++ b/paddle/gserver/layers/SumToOneNormLayer.cpp @@ -32,13 +32,13 @@ namespace paddle { */ class SumToOneNormLayer : public Layer { -protected: + protected: /// reciprocalRowSum_ = \f$1 / \sum_{k=1}^N in[k]\f$ MatrixPtr reciprocalRowSum_; /// dotSum = output_.grad \f$.*\f$ output_.value MatrixPtr dotSum_; -public: + public: explicit SumToOneNormLayer(const LayerConfig& config) : Layer(config) {} bool init(const LayerMap& layerMap, diff --git a/paddle/gserver/layers/SwitchOrderLayer.h b/paddle/gserver/layers/SwitchOrderLayer.h index 882437f4434c2e61a5b08328d2f79c1e7f589204..8a551a2bba698374841e73dc4dbad403034dd300 100644 --- a/paddle/gserver/layers/SwitchOrderLayer.h +++ b/paddle/gserver/layers/SwitchOrderLayer.h @@ -22,7 +22,7 @@ namespace paddle { * \brief This layer calculate softmax in image channel dimension. */ class SwitchOrderLayer : public Layer { -public: + public: explicit SwitchOrderLayer(const LayerConfig& config) : Layer(config) {} ~SwitchOrderLayer() {} @@ -34,7 +34,7 @@ public: void setInDims(); void setOutDims(); -protected: + protected: std::vector> nchw2nhwc_; std::vector> nhwc2nchw_; TensorShape inDims_; diff --git a/paddle/gserver/layers/TableProjection.h b/paddle/gserver/layers/TableProjection.h index ffb05e68f068a7b9abb0db5cea6133e64300cb55..60286149f4227fbc758dca7864c6d1f67782c7ae 100644 --- a/paddle/gserver/layers/TableProjection.h +++ b/paddle/gserver/layers/TableProjection.h @@ -32,7 +32,7 @@ namespace paddle { * @note If \f$ids[i] = -1\f$, it will be ignored. */ class TableProjection : public Projection { -public: + public: TableProjection(const ProjectionConfig& config, const ParameterPtr& parameter, bool useGpu); @@ -43,7 +43,7 @@ public: virtual void forward(); virtual void backward(const UpdateCallback& callback); -protected: + protected: std::unique_ptr table_; }; diff --git a/paddle/gserver/layers/TensorLayer.h b/paddle/gserver/layers/TensorLayer.h index 8a323aa15f6f3761c45b6ca7e3be8f15621a189e..5c1ee40ceda9387138a82368ec4edcbae4bd3419 100644 --- a/paddle/gserver/layers/TensorLayer.h +++ b/paddle/gserver/layers/TensorLayer.h @@ -37,11 +37,11 @@ namespace paddle { */ class TensorLayer : public Layer { -protected: + protected: WeightList weights_; std::unique_ptr biases_; -public: + public: explicit TensorLayer(const LayerConfig& config) : Layer(config) {} bool init(const LayerMap& layerMap, diff --git a/paddle/gserver/layers/TransLayer.h b/paddle/gserver/layers/TransLayer.h index 03d094862459c80aee8899c0352ffce732db08af..1cd8fd91f785d5a43fc7d7663e657702b32fa534 100644 --- a/paddle/gserver/layers/TransLayer.h +++ b/paddle/gserver/layers/TransLayer.h @@ -29,7 +29,7 @@ namespace paddle { * The config file api is trans_layer. */ class TransLayer : public Layer { -public: + public: explicit TransLayer(const LayerConfig& config) : Layer(config) {} bool init(const LayerMap& layerMap, diff --git a/paddle/gserver/layers/TransposedFullMatrixProjection.cpp b/paddle/gserver/layers/TransposedFullMatrixProjection.cpp index 755389f7074c252c0fad396e629c6ffedc74b531..45f59779896f993aface284e3485e1e3d801f4c5 100644 --- a/paddle/gserver/layers/TransposedFullMatrixProjection.cpp +++ b/paddle/gserver/layers/TransposedFullMatrixProjection.cpp @@ -24,14 +24,14 @@ namespace paddle { * The config file api is trans_full_matrix_projection. */ class TransposedFullMatrixProjection : public Projection { -public: + public: TransposedFullMatrixProjection(const ProjectionConfig& config, ParameterPtr parameter, bool useGPu); virtual void forward(); virtual void backward(const UpdateCallback& callback); -protected: + protected: std::unique_ptr weight_; }; diff --git a/paddle/gserver/layers/UpsampleLayer.h b/paddle/gserver/layers/UpsampleLayer.h index 25efbac5e9e6e92653f7c2b2f4dca9221737e5d6..c9d079c3141c37517866bfdad10d9b2cdb89f7d5 100644 --- a/paddle/gserver/layers/UpsampleLayer.h +++ b/paddle/gserver/layers/UpsampleLayer.h @@ -30,7 +30,7 @@ namespace paddle { */ class UpsampleLayer : public Layer { -public: + public: explicit UpsampleLayer(const LayerConfig& config) : Layer(config) {} ~UpsampleLayer() {} @@ -42,7 +42,7 @@ public: size_t getOutputSize(); -protected: + protected: size_t scale_, scaleY_; size_t upsampleSize_, upsampleSizeY_; size_t padOutX_, padOutY_; diff --git a/paddle/gserver/layers/ValidationLayer.h b/paddle/gserver/layers/ValidationLayer.h index f412d685c0541537bd4318fec2dae06215c4afbe..be41128ef4530f32a63c757648c2f393fd118ea6 100644 --- a/paddle/gserver/layers/ValidationLayer.h +++ b/paddle/gserver/layers/ValidationLayer.h @@ -23,7 +23,7 @@ DECLARE_int32(trainer_id); namespace paddle { class ValidationLayer : public Layer { -public: + public: explicit ValidationLayer(const LayerConfig& config) : Layer(config) {} bool init(const LayerMap& layerMap, @@ -51,7 +51,7 @@ public: * AucValidation */ class AucValidation : public ValidationLayer { -public: + public: explicit AucValidation(const LayerConfig& config) : ValidationLayer(config), cpuOutput_(nullptr), @@ -72,7 +72,7 @@ public: }; std::vector predictArray_; -private: + private: bool passBegin_; std::unique_ptr evaluator_; MatrixPtr cpuOutput_; @@ -84,7 +84,7 @@ private: * positive-negative pair rate Validation */ class PnpairValidation : public ValidationLayer { -public: + public: explicit PnpairValidation(const LayerConfig& config) : ValidationLayer(config) {} @@ -95,7 +95,7 @@ public: void onPassEnd() override; -private: + private: bool passBegin_; std::unique_ptr evaluator_; }; diff --git a/paddle/gserver/layers/WarpCTCLayer.h b/paddle/gserver/layers/WarpCTCLayer.h index 6f6be359c0aa46a4f3775f8405e1aa51ca1ae147..3017ca794ecc14f5a3cbd0b302a4953a191a5065 100644 --- a/paddle/gserver/layers/WarpCTCLayer.h +++ b/paddle/gserver/layers/WarpCTCLayer.h @@ -26,7 +26,7 @@ namespace paddle { * The config file api is warp_ctc_layer. */ class WarpCTCLayer : public Layer { -public: + public: explicit WarpCTCLayer(const LayerConfig& config) : Layer(config) {} ~WarpCTCLayer() {} @@ -35,7 +35,7 @@ public: void forward(PassType passType) override; void backward(const UpdateCallback& callback) override; -protected: + protected: /** * sequence matrix and batch matrix copy: * sequence (s0, s0, s0, s0; s1, s1; s2, s2, s2; s3) @@ -49,7 +49,7 @@ protected: const ICpuGpuVectorPtr& seqStartPositions, bool normByTimes); -protected: + protected: size_t numClasses_; size_t blank_; size_t maxSequenceLength_; diff --git a/paddle/gserver/tests/MKLDNNTester.h b/paddle/gserver/tests/MKLDNNTester.h index c1faa6fd90e06d8c742e97c9ce51eeba3c24a550..41ac46b70ab08d4071f4e6abfca94667268015d7 100644 --- a/paddle/gserver/tests/MKLDNNTester.h +++ b/paddle/gserver/tests/MKLDNNTester.h @@ -44,7 +44,7 @@ class MKLDNNTester { std::vector paraValues; }; -protected: + protected: std::vector configs_; vector layerNames_; vector> dataLayers_; @@ -65,7 +65,7 @@ protected: /// passType, PASS_TRAIN, PASS_TEST or PASS_GC (Gradient Check pass) PassType passType_; -public: + public: explicit MKLDNNTester(size_t iter = 3, float epsilon = 1e-4) { iter_ = iter; eps_ = epsilon; @@ -75,7 +75,7 @@ public: ~MKLDNNTester() {} -public: + public: void run(const TestConfig& dnn, const TestConfig& ref, size_t batchSize, @@ -97,7 +97,7 @@ public: bool use_mkldnn, size_t iter = 2); -private: + private: void reset(const TestConfig& dnn, const TestConfig& ref, size_t batchSize); void setInputImgSize(); void runOnce(); diff --git a/paddle/gserver/tests/test_MultinomialSampler.cpp b/paddle/gserver/tests/test_MultinomialSampler.cpp index 4a295ea9d51788f988fe79f8439cc7769f661d8e..043025239e744601cbef3ca5c241509872963bd8 100644 --- a/paddle/gserver/tests/test_MultinomialSampler.cpp +++ b/paddle/gserver/tests/test_MultinomialSampler.cpp @@ -27,7 +27,7 @@ using namespace paddle; // NOLINT using namespace std; // NOLINT class MultinomialSamplerTester : public MultinomialSampler { -public: + public: MultinomialSamplerTester(real* prob, int size) : MultinomialSampler(prob, size) {} diff --git a/paddle/gserver/tests/test_RecurrentGradientMachine.cpp b/paddle/gserver/tests/test_RecurrentGradientMachine.cpp index 72324fcf29cc60867005da25b35a8075fd590a89..9770567b88a2af946b30439300540ed61694ba10 100644 --- a/paddle/gserver/tests/test_RecurrentGradientMachine.cpp +++ b/paddle/gserver/tests/test_RecurrentGradientMachine.cpp @@ -26,7 +26,7 @@ DECLARE_int32(seed); using namespace paddle; // NOLINT using namespace std; // NOLINT class TrainerForTest : public paddle::Trainer { -public: + public: void startTrain() { GradientMachine& gm = *this->trainerInternal_.getGradientMachine(); gm.start(); diff --git a/paddle/gserver/tests/test_RecurrentLayer.cpp b/paddle/gserver/tests/test_RecurrentLayer.cpp index e5ce922f15749cb18b93f64e0e08f437c5633065..b54e37b7dbf8bffeb949f709e6a4f9ec86ea13c3 100644 --- a/paddle/gserver/tests/test_RecurrentLayer.cpp +++ b/paddle/gserver/tests/test_RecurrentLayer.cpp @@ -225,7 +225,7 @@ TEST(Layer, RecurrentLayer) { #include "paddle/gserver/layers/RecurrentLayer.h" template class TestRecurrentLayer { -public: + public: LayerConfig config_; bool useGpu_; bool useBatch_; diff --git a/paddle/math/Allocator.h b/paddle/math/Allocator.h index ae60f6fe5fa142bdffeafc31b5816b8fcc94ad5c..c43a83891eb6b7eae278169736149ad1d89e950e 100644 --- a/paddle/math/Allocator.h +++ b/paddle/math/Allocator.h @@ -27,7 +27,7 @@ namespace paddle { * This is the base class of all Allocator class. */ class Allocator { -public: + public: virtual ~Allocator() {} virtual void* alloc(size_t size) = 0; virtual void free(void* ptr) = 0; @@ -38,7 +38,7 @@ public: * @brief CPU allocator implementation. */ class CpuAllocator : public Allocator { -public: + public: ~CpuAllocator() {} /** @@ -76,7 +76,7 @@ public: * @brief GPU allocator implementation. */ class GpuAllocator : public Allocator { -public: + public: ~GpuAllocator() {} /** @@ -107,7 +107,7 @@ public: * @brief CPU pinned memory allocator implementation. */ class CudaHostAllocator : public Allocator { -public: + public: ~CudaHostAllocator() {} /** diff --git a/paddle/math/BaseMatrix.h b/paddle/math/BaseMatrix.h index 00ce5a19491048f3339d608ac37669816a9ad3f5..1958629aa0354fcc332b1e5677a64c29397e0d26 100644 --- a/paddle/math/BaseMatrix.h +++ b/paddle/math/BaseMatrix.h @@ -43,7 +43,7 @@ typedef bool_constant true_type; address += row * ld + col; class MatrixOffset { -public: + public: size_t aCol_; size_t aRow_; size_t bCol_; @@ -72,14 +72,14 @@ public: template class BaseMatrixT : public TensorExpression, T> { -public: + public: size_t height_, width_; size_t stride_; T* data_; bool trans_; bool useGpu_; -public: + public: virtual ~BaseMatrixT() {} BaseMatrixT(size_t height, size_t width, T* data, bool trans, bool useGpu) : height_(height), diff --git a/paddle/math/CMakeLists.txt b/paddle/math/CMakeLists.txt index 922fb5172273da24f9c48786961a6d850b1ed7c5..3c897b5f3e09cd53ddd5b767333ce4759250da71 100644 --- a/paddle/math/CMakeLists.txt +++ b/paddle/math/CMakeLists.txt @@ -51,10 +51,6 @@ else() endif() - -add_style_check_target(paddle_math ${MATH_SOURCES}) -add_style_check_target(paddle_math ${MATH_HEADERS}) - add_dependencies(paddle_math paddle_proto ${external_project_dependencies}) # depends if(WITH_TESTING) add_subdirectory(tests) diff --git a/paddle/math/CpuSparseMatrix.h b/paddle/math/CpuSparseMatrix.h index 22b6b71688bd555cf8bf8a29088ad01b092d67cf..172792c2950ce56281715cb7f3eb076da252d77e 100644 --- a/paddle/math/CpuSparseMatrix.h +++ b/paddle/math/CpuSparseMatrix.h @@ -22,7 +22,7 @@ limitations under the License. */ namespace paddle { class CpuSparseMatrix : public Matrix { -public: + public: CpuSparseMatrix(size_t height, size_t width, size_t nnz, /* used to allocate space */ @@ -291,10 +291,10 @@ public: LOG(FATAL) << "not supported!"; } -private: + private: MatrixPtr clone(size_t height = 0, size_t width = 0, bool useGpu = false); -protected: + protected: void sparseResize(); /*for csr , record row start position, for csc, record row index for every no * zero value*/ @@ -310,10 +310,10 @@ protected: static ThreadLocal> cpuLocalMats_; // BaseMatrixT interface -public: + public: bool isSparse() const { return true; } -private: + private: using Matrix::mul; using Matrix::copyFrom; using Matrix::rowMax; @@ -329,7 +329,7 @@ private: namespace paddle { class CpuSparseMatrix : public Matrix { -public: + public: CpuSparseMatrix(size_t height, size_t width, size_t nnz, /* used to allocate space */ diff --git a/paddle/math/ExecViaCpu.h b/paddle/math/ExecViaCpu.h index 9b2a3c2b8accd384aac896e86ef8315a744633e1..ec2337545e9e3efdf31d3d786a096a67283715f2 100644 --- a/paddle/math/ExecViaCpu.h +++ b/paddle/math/ExecViaCpu.h @@ -31,17 +31,17 @@ namespace paddle { template class CopyToCpu { -public: + public: explicit CopyToCpu(Arg& arg) : arg_(arg) {} Arg& copiedArg() const { return arg_; } -private: + private: Arg& arg_; }; template <> class CopyToCpu { -public: + public: explicit CopyToCpu(Matrix& arg) : arg_(arg) { if (arg.useGpu()) { CHECK(!arg.isTransposed()) << "Not supported"; @@ -59,14 +59,14 @@ public: } Matrix& copiedArg() const { return copied_ ? *copied_ : arg_; } -private: + private: Matrix& arg_; MatrixPtr copied_; }; template <> class CopyToCpu { -public: + public: explicit CopyToCpu(const Matrix& arg) : arg_(arg) { if (arg.useGpu()) { CHECK(!arg.isTransposed()) << "Not supported"; @@ -79,14 +79,14 @@ public: } const Matrix& copiedArg() const { return copied_ ? *copied_ : arg_; } -private: + private: const Matrix& arg_; MatrixPtr copied_; }; template <> class CopyToCpu { -public: + public: explicit CopyToCpu(IVector& arg) : arg_(arg) { if (arg.useGpu()) { copied_ = IVector::create(arg.getSize(), /* useGpu= */ false); @@ -100,14 +100,14 @@ public: } IVector& copiedArg() const { return copied_ ? *copied_ : arg_; } -private: + private: IVector& arg_; IVectorPtr copied_; }; template <> class CopyToCpu { -public: + public: explicit CopyToCpu(const IVector& arg) : arg_(arg) { if (arg.useGpu()) { copied_ = IVector::create(arg.getSize(), /* useGpu= */ false); @@ -116,7 +116,7 @@ public: } const IVector& copiedArg() const { return copied_ ? *copied_ : arg_; } -private: + private: const IVector& arg_; IVectorPtr copied_; }; @@ -128,7 +128,7 @@ class GpuFuncWrapperImp; template class GpuFuncWrapperBase { -public: + public: typedef R ResultType; R operator()(F&& f, Args... args) { return f(CopyToCpu::type>(args) diff --git a/paddle/math/MKLDNNMatrix.h b/paddle/math/MKLDNNMatrix.h index e1fb81679adf4658a58ceee73c8d5da6c0b61050..d4a78f3e54b73add3c00e17f13d91359839d3d14 100644 --- a/paddle/math/MKLDNNMatrix.h +++ b/paddle/math/MKLDNNMatrix.h @@ -35,7 +35,7 @@ typedef std::shared_ptr MKLDNNMatrixPtr; * */ class MKLDNNMatrix : public CpuMatrix, public mkldnn::memory { -public: + public: MKLDNNMatrix(CpuMatrixPtr m, mkldnn::memory::primitive_desc pd) : CpuMatrix(m->getData(), m->getHeight(), m->getWidth(), false), mkldnn::memory(pd, m->getData()), @@ -107,7 +107,7 @@ public: dst.copyFrom(*m_); } -public: + public: /** * Reorder this MKLDNNMatrix from other format. * Support inplace reorder. @@ -226,7 +226,7 @@ public: */ mkldnn::engine getEngine() { return getPrimitiveDesc().get_engine(); } -protected: + protected: /** * Do reorder once. * Can support inplace. @@ -248,7 +248,7 @@ protected: set_data_handle(data); } -private: + private: // save the CpuMatrixPtr in case the buffer released outside CpuMatrixPtr m_; }; diff --git a/paddle/math/MathFunctions.cpp b/paddle/math/MathFunctions.cpp index de404cad89fba8021b8645a40e25c1f5b7e86596..f48119aa511578b21602a225277f01b4c6a9e9a8 100644 --- a/paddle/math/MathFunctions.cpp +++ b/paddle/math/MathFunctions.cpp @@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include "MathFunctions.h" +#include "paddle/math/MathFunctions.h" #include "hl_matrix_apply.cuh" #include "hl_matrix_ops.cuh" #include "paddle/utils/DynamicLoader.h" @@ -240,6 +240,36 @@ template <> void vAdd(const int n, const double* a, const double* b, double* r) { vdAdd(n, a, b, r); } + +template <> +void vTanh(const int n, const float* a, float* r) { + vsTanh(n, a, r); +} + +template <> +void vTanh(const int n, const double* a, double* r) { + vdTanh(n, a, r); +} + +template <> +void vInvSqrt(const int n, const float* a, float* r) { + vsInvSqrt(n, a, r); +} + +template <> +void vInvSqrt(const int n, const double* a, double* r) { + vdInvSqrt(n, a, r); +} + +template <> +void vLog1p(const int n, const float* a, float* r) { + vsLog1p(n, a, r); +} + +template <> +void vLog1p(const int n, const double* a, double* r) { + vdLog1p(n, a, r); +} #else DEFINE_MATRIX_BINARY_OP(vExp, b = std::exp(a)); @@ -277,17 +307,6 @@ void vAdd(const int n, const T* a, const T* b, T* r) { n); } -template void vExp(const int n, const float* a, float* r); -template void vExp(const int n, const double* a, double* r); -template void vLog(const int n, const float* a, float* r); -template void vLog(const int n, const double* a, double* r); -template void vPow(const int n, const float* a, const float b, float* r); -template void vPow(const int n, const double* a, const double b, double* r); -template void vAdd(const int n, const float* a, const float* b, float* r); -template void vAdd(const int n, const double* a, const double* b, double* r); - -#endif - DEFINE_MATRIX_BINARY_OP(vInvSqrt, b = 1.0f / std::sqrt(a)); template void vInvSqrt(const int n, const T* a, T* r) { @@ -311,11 +330,19 @@ void vTanh(const int n, const T* a, T* r) { binary::vTanh(), const_cast(a), r, 1, n, n, n); } +template void vExp(const int n, const float* a, float* r); +template void vExp(const int n, const double* a, double* r); +template void vLog(const int n, const float* a, float* r); +template void vLog(const int n, const double* a, double* r); +template void vPow(const int n, const float* a, const float b, float* r); +template void vPow(const int n, const double* a, const double b, double* r); +template void vAdd(const int n, const float* a, const float* b, float* r); +template void vAdd(const int n, const double* a, const double* b, double* r); template void vInvSqrt(const int n, const double* a, double* r); template void vInvSqrt(const int n, const float* a, float* r); template void vLog1p(const int n, const float* a, float* r); template void vLog1p(const int n, const double* a, double* r); template void vTanh(const int n, const float* a, float* r); template void vTanh(const int n, const double* a, double* r); - +#endif } // namespace paddle diff --git a/paddle/math/Matrix.cpp b/paddle/math/Matrix.cpp index 0e84cb37392839d112448b0b3c12b042e7df838e..bcd6dfe1fda6b1243007b0c26a6e0087eedcc10c 100644 --- a/paddle/math/Matrix.cpp +++ b/paddle/math/Matrix.cpp @@ -2157,26 +2157,20 @@ void CpuMatrix::maxPoolForward(Matrix& inputMat, int wend = wstart + sizeX; wstart = wstart < 0 ? 0 : wstart; wend = wend < (int)imgSizeW ? wend : (int)imgSizeW; - if (maskData == NULL) { - real tmp = -(real)FLT_MAX; - for (int h = hstart; h < hend; ++h) { - for (int w = wstart; w < wend; ++w) { - tmp = tmp < inputData[h * imgSizeW + w] - ? inputData[h * imgSizeW + w] - : tmp; - } - } - outData[ph * outputW + pw] = tmp; - } else { - for (int h = hstart; h < hend; ++h) { - for (int w = wstart; w < wend; ++w) { - if (outData[ph * outputW + pw] < inputData[h * imgSizeW + w]) { - outData[ph * outputW + pw] = inputData[h * imgSizeW + w]; - maskData[ph * outputW + pw] = h * imgSizeW + w; - } + + real maxval = -(real)FLT_MAX; + int max_index = -1; + for (int h = hstart; h < hend; ++h) { + for (int w = wstart; w < wend; ++w) { + if (maxval < inputData[h * imgSizeW + w]) { + maxval = inputData[h * imgSizeW + w]; + max_index = h * imgSizeW + w; } } } + + outData[ph * outputW + pw] = maxval; + if (maskData != NULL) maskData[ph * outputW + pw] = max_index; } } // compute offset diff --git a/paddle/math/Matrix.h b/paddle/math/Matrix.h index 04e9614eabc47c4c661ace2106e8ca96f45a1d49..4c3b2c95361065372f5969a2da73bce0eb9d123f 100644 --- a/paddle/math/Matrix.h +++ b/paddle/math/Matrix.h @@ -77,7 +77,7 @@ typedef std::shared_ptr CpuSparseMatrixPtr; * instead. */ class Matrix : public BaseMatrix { -protected: + protected: Matrix(MemoryHandlePtr memHandle, size_t height, size_t width, @@ -95,11 +95,11 @@ protected: static ThreadLocal tmpMat_; -public: + public: size_t elementCnt_; // maximal number of elements which can be held in data_ MemoryHandlePtr memoryHandle_; -public: + public: virtual ~Matrix() {} static MatrixPtr create(MemoryHandlePtr memHandle, @@ -412,7 +412,7 @@ public: LOG(FATAL) << "Not implemented"; } -public: + public: /// Only set all variables to 0 or NULL but not free them. virtual void clear() { height_ = 0; @@ -1228,7 +1228,7 @@ inline std::ostream& operator<<(std::ostream& os, const Matrix& mat) { } class GpuMatrix : public Matrix { -public: + public: GpuMatrix(); GpuMatrix(size_t height, size_t width, bool trans = false); @@ -1660,11 +1660,11 @@ public: }; class CpuMatrix : public Matrix { -private: + private: MatrixPtr sftmaxSum_; MatrixPtr sftmaxDot_; -public: + public: CpuMatrix(size_t height, size_t width, bool trans = false); CpuMatrix(real* data, size_t height, size_t width, bool trans = false) : Matrix(data, height, width, trans, false) {} @@ -1892,7 +1892,7 @@ public: real* getRow(size_t row) { return BaseMatrix::rowBuf(row); } virtual real* getRowBuf(size_t row) { return getRow(row); } -public: + public: /// add b to each sample of this. void addBias(Matrix& b, real scale); void addSharedBias(Matrix& b, real scale); @@ -2128,7 +2128,7 @@ public: }; class SharedCpuMatrix : public CpuMatrix { -public: + public: #ifndef PADDLE_MOBILE_INFERENCE /* blockNum is number of partitions of the matrix */ SharedCpuMatrix(int blockNum, size_t height, size_t width, bool trans = false) @@ -2160,12 +2160,12 @@ public: ~SharedCpuMatrix() {} -public: + public: virtual void mul(CpuSparseMatrix* a, CpuMatrix* b, real scaleAB, real scaleT); virtual void add(Matrix& b, real p1, real p2); virtual void add(real p1, real p2); -private: + private: using Matrix::mul; void initShared(int blockNum); void initBlock(int blockNum); diff --git a/paddle/math/MatrixBitCode.cpp b/paddle/math/MatrixBitCode.cpp index 61a9923bc2e6f358738f80de4a30d83c0cc00656..f7a949294b54a5a874e1239a13ca9dce3ba18e94 100644 --- a/paddle/math/MatrixBitCode.cpp +++ b/paddle/math/MatrixBitCode.cpp @@ -27,7 +27,7 @@ struct SimpleCode { inline bool calcBit(int bit) const { return c_ & (1 << bit); } inline int getLength() const { return findLastSet(c_) - 1; } -private: + private: size_t c_; }; @@ -39,7 +39,7 @@ struct SimpleCodeTable { size_t size() const { return numClasses_; } int getMaxCodeLength() const { return findLastSet(numClasses_ - 1); } -private: + private: size_t numClasses_; int maxCodeLength_; }; diff --git a/paddle/math/MemoryHandle.h b/paddle/math/MemoryHandle.h index 03ee413c1218376635c4696ebb774c584aa67aa4..516e09dbed47ac6b039ccb094614c9588eeb3cd5 100644 --- a/paddle/math/MemoryHandle.h +++ b/paddle/math/MemoryHandle.h @@ -20,16 +20,16 @@ limitations under the License. */ namespace paddle { class MemoryHandle { -protected: + protected: explicit MemoryHandle(size_t size); virtual ~MemoryHandle() {} -public: + public: void* getBuf() const { return buf_; } size_t getSize() const { return size_; } size_t getAllocSize() const { return allocSize_; } -protected: + protected: PoolAllocator* allocator_; size_t size_; // the requested size size_t allocSize_; // the allocated size @@ -43,7 +43,7 @@ protected: * The raw handle will be released at destructor */ class GpuMemoryHandle : public MemoryHandle { -public: + public: explicit GpuMemoryHandle(size_t size); virtual ~GpuMemoryHandle(); }; @@ -54,7 +54,7 @@ public: * The raw handle will be released at destructor */ class CpuMemoryHandle : public MemoryHandle { -public: + public: explicit CpuMemoryHandle(size_t size); virtual ~CpuMemoryHandle(); }; diff --git a/paddle/math/PoolAllocator.h b/paddle/math/PoolAllocator.h index 90141fef3fd43fe221874cc50e688f6db9e2dee6..7239cf1c4494e207081e325a7e6067ba26a9c852 100644 --- a/paddle/math/PoolAllocator.h +++ b/paddle/math/PoolAllocator.h @@ -27,7 +27,7 @@ namespace paddle { * @brief Memory pool allocator implementation. */ class PoolAllocator { -public: + public: /** * @brief constructor. * @param allocator a Allocator object. @@ -47,7 +47,7 @@ public: void free(void* ptr, size_t size); std::string getName() { return name_; } -private: + private: void freeAll(); void printAll(); std::unique_ptr allocator_; diff --git a/paddle/math/RowBuffer.h b/paddle/math/RowBuffer.h index 2e4d11a86bf8bd1308b2972f549bc7c201044785..6950afaa21d60615b27c06a151b0afbb296653bf 100644 --- a/paddle/math/RowBuffer.h +++ b/paddle/math/RowBuffer.h @@ -26,7 +26,7 @@ namespace paddle { * If not set memory handler, then the data could be auto growth. */ class RowBuffer { -public: + public: /** * @brief RowBuffer create a auto-growth row buffer. The row length is width. * @param width the length of each row, a.k.a matrix width. @@ -129,7 +129,7 @@ public: */ inline size_t getWidth() const { return width_; } -private: + private: //! TODO(yuyang18): Add resize method to CpuMemHandlePtr, then we can get rid //! of std::vector here. CpuMemHandlePtr preallocatedBuf_; diff --git a/paddle/math/SparseMatrix.h b/paddle/math/SparseMatrix.h index 7c525f4edf3d53544c195f8e253c27a03854a793..9181fa29233677d8f4fac503905cc31eb66cb6c1 100644 --- a/paddle/math/SparseMatrix.h +++ b/paddle/math/SparseMatrix.h @@ -25,7 +25,7 @@ namespace paddle { typedef std::shared_ptr<_hl_sparse_matrix_s> hl_sparse_matrix_s_ptr; class GpuSparseMatrix : public Matrix { -public: + public: MemoryHandlePtr sMemoryHandle_; int* rows_; int* cols_; @@ -36,7 +36,7 @@ public: SparseValueType valueType_; SparseFormat format_; -public: + public: GpuSparseMatrix(size_t height, size_t width, size_t nnz, /* used to allocate space */ @@ -73,7 +73,7 @@ public: bool trans, MemoryHandlePtr sMemoryHandle); -protected: + protected: struct Element { int row; int col; @@ -82,7 +82,7 @@ protected: : row(rowIn), col(colIn), val(valIn) {} }; -public: + public: ~GpuSparseMatrix() {} void resize(size_t newHeight, @@ -211,13 +211,13 @@ public: */ void rowMax(IVector& maxIds, Matrix& maxVal); -protected: + protected: void sparseResize(); void copyRow(int offsets, size_t colNum, const sparse_non_value_t* row); void copyRow(int offsets, size_t colNum, const sparse_float_value_t* row); -public: + public: void mul(const Matrix& a, const Matrix& b, real scaleAB, real scaleT); void copyFrom(CpuSparseMatrix& src, hl_stream_t stream); @@ -228,10 +228,10 @@ public: void trimFromCSC(const CpuSparseMatrix& src); // BaseMatrixT interface -public: + public: bool isSparse() const { return true; } -private: + private: using Matrix::mul; using Matrix::copyFrom; using Matrix::rowMax; @@ -248,7 +248,7 @@ private: namespace paddle { class GpuSparseMatrix : public Matrix { -public: + public: GpuSparseMatrix(size_t height, size_t width, size_t nnz, /* used to allocate space */ diff --git a/paddle/math/SparseRowMatrix.h b/paddle/math/SparseRowMatrix.h index 3920de32df7de925d6e22e17b93b15bff8785675..cf6779e8b0b1d6b0c13b21a08ffff5af76e57ba6 100644 --- a/paddle/math/SparseRowMatrix.h +++ b/paddle/math/SparseRowMatrix.h @@ -29,7 +29,7 @@ namespace paddle { * Sparse Row */ class SparseRowCpuMatrix : public CpuMatrix { -public: + public: struct IndexDict { // In the following, global id means the row id in the original matrix. // Local id means the row id in the local storage which only contains @@ -53,7 +53,7 @@ public: virtual ~SparseRowCpuMatrix() {} -public: + public: /** * Get the row buf * @@ -163,7 +163,7 @@ public: return indexDictHandle_->localIndices; } -protected: + protected: template void apply(Func f) { f(buf_->data(), localIndices_->size() * width_); @@ -204,7 +204,7 @@ class SyncThreadPool; /// For prefetching parameters from remote Parameter server class SparsePrefetchRowCpuMatrix : public SparseRowCpuMatrix { -public: + public: SparsePrefetchRowCpuMatrix(CpuMemHandlePtr dataHandle, size_t height, size_t width, @@ -229,13 +229,13 @@ public: */ void setupIndices(); -protected: + protected: void addRows(const unsigned int* ids, size_t len); SyncThreadPool* pool_; }; class SparseAutoGrowRowCpuMatrix : public SparseRowCpuMatrix { -public: + public: SparseAutoGrowRowCpuMatrix(size_t height, size_t width, IndexDictPtr indexDictHandle = nullptr, @@ -258,7 +258,7 @@ public: }; class CacheRowCpuMatrix : public SparseAutoGrowRowCpuMatrix { -public: + public: CacheRowCpuMatrix(size_t height, size_t width, IndexDictPtr indexDictHandle = nullptr, @@ -287,7 +287,7 @@ public: virtual void mul(CpuSparseMatrix* a, CpuMatrix* b, real scaleAB, real scaleT); -public: + public: CpuVectorPtr sourceDataVec_; real* sourceData_; }; @@ -299,7 +299,7 @@ public: * ids are hashed by worker thread id. */ class SparseRowIdsCpuMatrix : public CpuMatrix { -public: + public: SparseRowIdsCpuMatrix(CpuMemHandlePtr dataHandle, size_t height, size_t width, @@ -310,7 +310,7 @@ public: std::vector& getIds(size_t threadId) { return idsArray_[threadId]; } -private: + private: std::vector> idsArray_; }; @@ -320,13 +320,13 @@ private: namespace paddle { class SparseRowCpuMatrix : public CpuMatrix { -public: + public: void reserveStore() {} void clearIndices() {} }; class SparsePrefetchRowCpuMatrix : public SparseRowCpuMatrix { -public: + public: void setupIndices() {} void addRows(MatrixPtr input) {} void addRows(IVectorPtr ids) {} diff --git a/paddle/math/Storage.h b/paddle/math/Storage.h index ba8f4689a1e896304aa14821b40fc8ff0c304bb2..61a9aa2a07442d9e4ede80c961e17e079eb8b3ba 100644 --- a/paddle/math/Storage.h +++ b/paddle/math/Storage.h @@ -25,7 +25,7 @@ namespace paddle { * @brief Storage manager for multiple devices. */ class StorageEngine { -public: + public: /** * @return Storage singleton */ @@ -41,7 +41,7 @@ public: */ PoolAllocator* getCpuAllocator(); -protected: + protected: StorageEngine(); ~StorageEngine(); RWLock lock_; diff --git a/paddle/math/TensorApply.h b/paddle/math/TensorApply.h index 7d79cae5a11851b190afbb9ac94efdf2ba2510b7..8b642047bffa33b47dfb8ffc8e3fd2a9b7dbae3a 100644 --- a/paddle/math/TensorApply.h +++ b/paddle/math/TensorApply.h @@ -21,7 +21,7 @@ namespace paddle { */ template class TensorApply { -public: + public: explicit INLINE TensorApply(const Derived& p) : data_(p.data_), stride_(p.stride_), @@ -52,7 +52,7 @@ public: */ template class TensorApply { -public: + public: explicit INLINE TensorApply(const Derived& p) : data_(p.data_), stride_(p.stride_), @@ -77,7 +77,7 @@ public: template class TensorApply, T> { -public: + public: explicit TensorApply(const TensorExpression& expr) : expr_(expr.derived()) {} @@ -97,7 +97,7 @@ public: */ template class TensorApply, T> { -public: + public: explicit INLINE TensorApply(const TensorUnaryOp& expr) : op_(expr.op_), expr_(expr.expr_) {} @@ -118,7 +118,7 @@ public: */ template class TensorApply, T> { -public: + public: explicit INLINE TensorApply( const TensorBinaryOp& expr) : op_(expr.op_), lhs_(expr.lhs_), rhs_(expr.rhs_) { @@ -153,7 +153,7 @@ public: */ template class TensorApply, T> { -public: + public: explicit INLINE TensorApply( const TensorTernaryOp& expr) : expr1_(expr.expr1_), expr2_(expr.expr2_), expr3_(expr.expr3_) { @@ -192,7 +192,7 @@ public: */ template class TensorApply, T> { -public: + public: explicit INLINE TensorApply(const TensorConstant& expr) : op_(expr.op_), expr_(expr.expr_) {} diff --git a/paddle/math/TensorAssign.h b/paddle/math/TensorAssign.h index 113d98c16b22b06971040b1a1ce52c696f6c3c14..7d4726ddba43202970c37dd1a08f842104b24ada 100644 --- a/paddle/math/TensorAssign.h +++ b/paddle/math/TensorAssign.h @@ -25,7 +25,7 @@ namespace paddle { */ template class TensorAssignOp { -public: + public: explicit TensorAssignOp(const LhsType& lhs, const RhsType& rhs) : lhs_(lhs), rhs_(rhs) { #ifndef __CUDA_ARCH__ @@ -49,7 +49,7 @@ public: } INLINE bool useGpu() const { return lhs_.useGpu(); } -private: + private: TensorApply lhs_; TensorApply rhs_; }; diff --git a/paddle/math/TensorExpression.h b/paddle/math/TensorExpression.h index 83229ae65dd1f4ed6b885c3d6195b3758b8ba039..f6da9adfca50e49ca260e20313c8979a38e1b06b 100644 --- a/paddle/math/TensorExpression.h +++ b/paddle/math/TensorExpression.h @@ -40,7 +40,7 @@ class TensorAssignOp; */ template class TensorExpression { -public: + public: /** * Element wise unary expression. */ @@ -355,7 +355,7 @@ public: return TensorAssignOp(derived(), expr); } -protected: + protected: const Derived& derived() const { return *static_cast(this); } }; @@ -365,7 +365,7 @@ protected: template class TensorUnaryOp : public TensorExpression, T> { -public: + public: explicit TensorUnaryOp(const OP op, const ExprType& expr) : op_(op), expr_(expr) {} @@ -379,7 +379,7 @@ public: template class TensorBinaryOp : public TensorExpression, T> { -public: + public: explicit TensorBinaryOp(const OP op, const LhsType& lhs, const RhsType& rhs) : op_(op), lhs_(lhs), rhs_(rhs) {} @@ -395,7 +395,7 @@ template class TensorTernaryOp : public TensorExpression< TensorTernaryOp, T> { -public: + public: explicit TensorTernaryOp(const ExprType1& expr1, const ExprType2& expr2, const ExprType3& expr3) @@ -412,7 +412,7 @@ public: template class TensorConstant : public TensorExpression, T> { -public: + public: explicit TensorConstant(const OP op, const ExprType& expr) : op_(op), expr_(expr) {} diff --git a/paddle/math/Vector.h b/paddle/math/Vector.h index 3efbc769dff5aa1dbc9d5015b0cbac313710d70d..964b42cae52af9b487ab17103bc5e999514e4dd1 100644 --- a/paddle/math/Vector.h +++ b/paddle/math/Vector.h @@ -40,13 +40,13 @@ class Matrix; template class BaseVector : public BaseMatrixT { -public: + public: BaseVector(size_t size, T* data, bool useGpu) : BaseMatrixT(1, size, data, false, useGpu), size_(this->width_) {} ~BaseVector() {} -protected: + protected: size_t& size_; }; @@ -57,7 +57,7 @@ protected: */ template class VectorT : public BaseVector { -protected: + protected: VectorT(size_t size, MemoryHandlePtr memoryHandle, size_t offset, bool useGpu) : BaseVector(size, reinterpret_cast(memoryHandle->getBuf()) + offset, @@ -71,7 +71,7 @@ protected: VectorT(size_t size, T* data, bool useGpu) : BaseVector(size, data, useGpu) {} -public: + public: virtual ~VectorT() {} static std::shared_ptr> create(size_t size, bool useGpu); @@ -281,7 +281,7 @@ public: } } -protected: + protected: friend class GpuVectorT; friend class CpuVectorT; virtual void copyTo(CpuVectorT* dest) const = 0; @@ -297,7 +297,7 @@ std::ostream& operator<<(std::ostream& os, const VectorT& vec) { template class GpuVectorT : public VectorT { -public: + public: explicit GpuVectorT(size_t size); GpuVectorT(size_t size, GpuMemHandlePtr memHandle, size_t offset) : VectorT(size, memHandle, offset, true) {} @@ -343,14 +343,14 @@ public: TensorGpuApply(*this, expr); } -protected: + protected: virtual void copyTo(CpuVectorT* dest) const; virtual void copyTo(GpuVectorT* dest) const; }; template class CpuVectorT : public VectorT { -public: + public: explicit CpuVectorT(size_t size); CpuVectorT(size_t size, MemoryHandlePtr memoryHandle, size_t offset) : VectorT(size, memoryHandle, offset, false) {} @@ -415,7 +415,7 @@ public: template class ParallelCpuVectorT : public CpuVectorT { -public: + public: ParallelCpuVectorT(size_t size, SyncThreadPool* pool) : CpuVectorT(size), pool_(pool) {} @@ -434,7 +434,7 @@ public: virtual void exec(SyncThreadPool::JobFunc jobFunc); -private: + private: typedef std::function& vec)> ExecFunc; void parallelExec(ExecFunc func); SyncThreadPool* pool_; @@ -445,7 +445,7 @@ private: */ template class CpuGpuVectorT { -public: + public: /** * @brief An enum type of SyncedFlag using to * mark data memory is in CPU or GPU. @@ -670,7 +670,7 @@ public: setSync(flag); } -protected: + protected: void resizeOrCreate(size_t size, bool useGpu); /** diff --git a/paddle/math/tests/TensorCheck.h b/paddle/math/tests/TensorCheck.h index f4332ede36356bc666612a240448c1be71e5170e..40ac04ef5d4baa0239bb03b04c3a6cce0fcac5a5 100644 --- a/paddle/math/tests/TensorCheck.h +++ b/paddle/math/tests/TensorCheck.h @@ -32,7 +32,7 @@ using paddle::CpuVectorT; using paddle::GpuVectorT; class AssertEqual { -public: + public: AssertEqual(real err = 0) : err_(err) {} inline bool operator()(real a, real b) { @@ -51,7 +51,7 @@ public: return true; } -private: + private: real err_; }; @@ -60,71 +60,71 @@ class CopyToCpu; template <> class CopyToCpu { -public: + public: explicit CopyToCpu(const CpuMatrix& arg) : arg_(arg) {} const CpuMatrix& copiedArg() const { return arg_; } -private: + private: const CpuMatrix& arg_; }; template <> class CopyToCpu { -public: + public: explicit CopyToCpu(const GpuMatrix& arg) : arg_(arg.getHeight(), arg.getWidth()) { arg_.copyFrom(arg); } CpuMatrix& copiedArg() { return arg_; } -private: + private: CpuMatrix arg_; }; template <> class CopyToCpu { -public: + public: explicit CopyToCpu(const Matrix& arg) : arg_(arg.getHeight(), arg.getWidth()) { arg_.copyFrom(arg); } CpuMatrix& copiedArg() { return arg_; } -private: + private: CpuMatrix arg_; }; template class CopyToCpu> { -public: + public: explicit CopyToCpu(const CpuVectorT& arg) : arg_(arg) {} const CpuVectorT& copiedArg() const { return arg_; } -private: + private: const CpuVectorT& arg_; }; template class CopyToCpu> { -public: + public: explicit CopyToCpu(const GpuVectorT& arg) : arg_(arg.getSize()) { arg_.copyFrom(arg); } CpuVectorT& copiedArg() { return arg_; } -private: + private: CpuVectorT arg_; }; template class CopyToCpu> { -public: + public: explicit CopyToCpu(const VectorT& arg) : arg_(arg.getSize()) { arg_.copyFrom(arg); } CpuVectorT& copiedArg() { return arg_; } -private: + private: CpuVectorT arg_; }; diff --git a/paddle/math/tests/TestUtils.h b/paddle/math/tests/TestUtils.h index d2b9706432f84fa082e071eb09d2ffe7402a085f..e1966ec8a74747960420ec80fdfbb957f7cf177f 100644 --- a/paddle/math/tests/TestUtils.h +++ b/paddle/math/tests/TestUtils.h @@ -56,31 +56,31 @@ using paddle::GpuSparseMatrix; template class ReplaceType { -public: + public: typedef T1 type; }; template <> class ReplaceType { -public: + public: typedef CpuMatrix type; }; template <> class ReplaceType { -public: + public: typedef GpuMatrix type; }; template <> class ReplaceType { -public: + public: typedef CpuMatrix type; }; template <> class ReplaceType { -public: + public: typedef GpuMatrix type; }; @@ -180,25 +180,25 @@ R call(C& obj, R (FC::*f)(FArgs...), Args&&... args) { template class ReturnType { -public: + public: typedef T type; }; template <> class ReturnType { -public: + public: typedef GpuMatrix type; }; template <> class ReturnType { -public: + public: typedef GpuIVector type; }; template <> class ReturnType { -public: + public: typedef GpuSparseMatrix type; }; @@ -234,7 +234,7 @@ GpuSparseMatrix autoArgs(CpuSparseMatrix& v) { } class AutoCompare { -public: + public: /** * err is the allowed calculation error. * The smaller the value of err, @@ -285,7 +285,7 @@ public: TensorCheck(compare, cpu, gpu); } -protected: + protected: CpuMatrix cpu; GpuMatrix gpu; AssertEqual compare; diff --git a/paddle/math/tests/test_ExecViaCpu.cpp b/paddle/math/tests/test_ExecViaCpu.cpp index 513c7b440e0aa6f20cc8209a3624f32f4892225b..72256cb9d4c93159418d27c7ca0d4f8b9a412a64 100644 --- a/paddle/math/tests/test_ExecViaCpu.cpp +++ b/paddle/math/tests/test_ExecViaCpu.cpp @@ -39,7 +39,7 @@ real f(Matrix& mat1, } class Functor { -public: + public: real operator()(Matrix& mat1, const Matrix& mat2, IVector& vec1, @@ -49,7 +49,7 @@ public: return a_; } -private: + private: real a_; }; diff --git a/paddle/math/tests/test_TrainingAlgorithm.cpp b/paddle/math/tests/test_TrainingAlgorithm.cpp index fb146176ca8eb97a9cdbaf9ebd5c4997a8439718..fb58d26734cab5d7d7bbbbe1cf8a920e4195b4bb 100644 --- a/paddle/math/tests/test_TrainingAlgorithm.cpp +++ b/paddle/math/tests/test_TrainingAlgorithm.cpp @@ -28,14 +28,14 @@ DEFINE_double(max_diff, 1e-13, "max diff allowed"); #endif class SetMaxDiff { -public: + public: explicit SetMaxDiff(double max_diff) { max_diff_ = FLAGS_max_diff; FLAGS_max_diff = max_diff; } ~SetMaxDiff() { FLAGS_max_diff = max_diff_; } -private: + private: double max_diff_; }; diff --git a/paddle/math/tests/test_perturbation.cpp b/paddle/math/tests/test_perturbation.cpp index ef99dab60a874846d04c5ce07d38b2857640ad7b..969400666f12e4c6001f270be3ec144e7e4d0702 100644 --- a/paddle/math/tests/test_perturbation.cpp +++ b/paddle/math/tests/test_perturbation.cpp @@ -32,7 +32,7 @@ const int TGT_SIZE = 21; const int CHANNELS = 3; class PerturbationTest : public testing::Test { -protected: + protected: virtual void SetUp() { generateTestImages(gpuImages_); } virtual void TearDown() {} diff --git a/paddle/optimizer/CMakeLists.txt b/paddle/optimizer/CMakeLists.txt index 25fc35311fc63988c64a445d72fc6255e49e8d4b..7c80faa48ce960a3a7eb7d88eda4f2b09756410e 100644 --- a/paddle/optimizer/CMakeLists.txt +++ b/paddle/optimizer/CMakeLists.txt @@ -7,6 +7,10 @@ set(OPITMIZER_SRCS sgd_optimizer.cc ) -cc_library(paddle_optimizer STATIC SRCS ${OPITMIZER_SRCS} DEPS paddle_proto glog) -cc_test(serialization_test SRCS serialization_test.cc DEPS paddle_proto) -cc_test(parameter_optimizer_test SRCS parameter_optimizer_test.cc DEPS paddle_optimizer) +add_library(paddle_optimizer ${OPITMIZER_SRCS}) +target_link_libraries(paddle_optimizer paddle_proto glog) + +if (WITH_TESTING) + add_unittest(serialization_test serialization_test.cc) + add_unittest(parameter_optimizer_test parameter_optimizer_test.cc) +endif() diff --git a/paddle/optimizer/adadelta_optimizer.h b/paddle/optimizer/adadelta_optimizer.h index 74df9d54be734fedec8aeddff5f50b1d1aefb1d3..5beb62295a83ba4826e9a6b9caf21de78d2e8ced 100644 --- a/paddle/optimizer/adadelta_optimizer.h +++ b/paddle/optimizer/adadelta_optimizer.h @@ -20,7 +20,7 @@ namespace paddle { namespace optimizer { class AdadeltaOptimizer : public ParameterOptimizer { -public: + public: AdadeltaOptimizer( Tensor *parameter, LrPolicy *lr, double rho, double epsilon, double decay) : ParameterOptimizer(parameter, lr), @@ -40,7 +40,7 @@ public: std::string SerializeState(); void DeserializeState(const std::string &state); -private: + private: Tensor *accum_gradient_; Tensor *accum_delta_; Tensor *update_delta_; diff --git a/paddle/optimizer/adagrad_optimizer.h b/paddle/optimizer/adagrad_optimizer.h index 1d58402d78ff9ada8b084a472d46c96580d01e5b..b6fc06739970984cf4bbd27d3e6e1e9066bc350f 100644 --- a/paddle/optimizer/adagrad_optimizer.h +++ b/paddle/optimizer/adagrad_optimizer.h @@ -20,7 +20,7 @@ namespace paddle { namespace optimizer { class AdagradOptimizer : public ParameterOptimizer { -public: + public: AdagradOptimizer(Tensor *parameter, LrPolicy *lr, double epsilon, @@ -36,7 +36,7 @@ public: std::string SerializeState(); void DeserializeState(const std::string &state); -private: + private: Tensor *accum_gradient_; double epsilon_; double decay_; diff --git a/paddle/optimizer/adam_optimizer.h b/paddle/optimizer/adam_optimizer.h index 7977226c8602745d5733021a51fc03d932b0921a..fce10960068364b40592b26a6b439494d75cfa03 100644 --- a/paddle/optimizer/adam_optimizer.h +++ b/paddle/optimizer/adam_optimizer.h @@ -20,7 +20,7 @@ namespace paddle { namespace optimizer { class AdamOptimizer : public ParameterOptimizer { -public: + public: AdamOptimizer(Tensor *parameter, LrPolicy *lr, double beta_1, @@ -42,7 +42,7 @@ public: std::string SerializeState(); void DeserializeState(const std::string &state); -private: + private: Tensor *momentums_; Tensor *velocitys_; double beta_1_; diff --git a/paddle/optimizer/lr_policy.h b/paddle/optimizer/lr_policy.h index 14422d1f42fc45d5e9a560c45259d4003a0b3d11..d639c9f22c8ad77267f68e2c3b35257211bf90df 100644 --- a/paddle/optimizer/lr_policy.h +++ b/paddle/optimizer/lr_policy.h @@ -20,7 +20,7 @@ namespace paddle { namespace optimizer { class LrPolicy { -public: + public: virtual ~LrPolicy() {} virtual double LearningRate(const uint64_t num_sample_passed) = 0; virtual std::string SerializeState() = 0; @@ -29,7 +29,7 @@ public: // constant learning rate policy class ConstLr final : public LrPolicy { -public: + public: ConstLr(double lr) : learning_rate_(lr){}; double LearningRate(const uint64_t num_sample_passed) { return learning_rate_; @@ -45,12 +45,12 @@ public: learning_rate_ = state.learning_rate(); } -private: + private: double learning_rate_; }; class LinearLr final : public LrPolicy { -public: + public: LinearLr(double lr, double lr_decay_a, double lr_decay_b) : learning_rate_(lr), lr_decay_a_(lr_decay_a), lr_decay_b_(lr_decay_b) {} double LearningRate(const uint64_t num_sample_passed) { @@ -72,7 +72,7 @@ public: lr_decay_b_ = state.lr_decay_b(); } -private: + private: double learning_rate_; double lr_decay_a_; double lr_decay_b_; diff --git a/paddle/optimizer/parameter_optimizer.h b/paddle/optimizer/parameter_optimizer.h index c7cf8db3ee05c75c171b68bcbcb06a5ae8fa5b48..d5abca82d55c12aed0f4fca0c4c1f21d20586155 100644 --- a/paddle/optimizer/parameter_optimizer.h +++ b/paddle/optimizer/parameter_optimizer.h @@ -26,7 +26,7 @@ namespace paddle { namespace optimizer { class ParameterOptimizer { -public: + public: /** * @brief update hook for algorithm need to traverse parameter more than * once. @@ -45,7 +45,7 @@ public: virtual std::string SerializeState() = 0; virtual void DeserializeState(const std::string &state) = 0; -protected: + protected: Tensor *parameter_; // learning rate policy LrPolicy *lr_policy_; diff --git a/paddle/optimizer/parameter_optimizer_test.cc b/paddle/optimizer/parameter_optimizer_test.cc index d663e2fd007febd3b9f0f43d213d63d2b20656b8..1d9572999e9e0f10092eecbc1b41369a89629da7 100644 --- a/paddle/optimizer/parameter_optimizer_test.cc +++ b/paddle/optimizer/parameter_optimizer_test.cc @@ -38,7 +38,7 @@ paddle::optimizer::Tensor* FixedTensor(size_t size) { } class OptimizerTest : public testing::Test { -public: + public: virtual ~OptimizerTest() {} // init paddle::optimizer::Tensor shape const size_t kSize = 5; @@ -115,7 +115,7 @@ public: } } -private: + private: std::vector opts_; paddle::OptimizerConfig config_; }; diff --git a/paddle/optimizer/sgd_optimizer.h b/paddle/optimizer/sgd_optimizer.h index f504d98adb8a01fd69ff313075b4c417222c765e..a8957cde54abd6667143d2a8265d732c849294e3 100644 --- a/paddle/optimizer/sgd_optimizer.h +++ b/paddle/optimizer/sgd_optimizer.h @@ -20,7 +20,7 @@ namespace paddle { namespace optimizer { class SGDOptimizer : public ParameterOptimizer { -public: + public: SGDOptimizer(Tensor* parameter, LrPolicy* lr, double m, double d, bool n) : ParameterOptimizer(parameter, lr), momentums_(nullptr), @@ -39,7 +39,7 @@ public: std::string SerializeState(); void DeserializeState(const std::string& state); -private: + private: Tensor* momentums_; double momentum_; double decay_; diff --git a/paddle/optimizer/tensor.h b/paddle/optimizer/tensor.h index fd32398a237e7e08a198707347cd3c0a4ed77bb3..d2cef99074335be6f9852d60daa103b9b45a550d 100644 --- a/paddle/optimizer/tensor.h +++ b/paddle/optimizer/tensor.h @@ -26,7 +26,7 @@ namespace optimizer { template class TensorT { -public: + public: TensorT(size_t size) : height_(1), width_(size) { // new T[size]() initializes all element to zero value. data_ptr_ = std::shared_ptr(new T[size](), std::default_delete()); @@ -54,7 +54,7 @@ public: // TODO: replace with tensorshape size_t size() const { return this->width_ * this->height_; } -protected: + protected: size_t height_; size_t width_; std::shared_ptr data_ptr_; diff --git a/paddle/parameter/AverageOptimizer.h b/paddle/parameter/AverageOptimizer.h index 4ad3c18d56abf16d1274c5b3b8e0347b85e64dea..f0fe2fd28e4be7df8ebc52fd9b9b5540f3d76949 100644 --- a/paddle/parameter/AverageOptimizer.h +++ b/paddle/parameter/AverageOptimizer.h @@ -21,7 +21,7 @@ namespace paddle { // After Optimization, parameter values are further averaged within // time range. class AverageOptimizer : public ParameterOptimizer { -public: + public: // if *useParameterApply* set, use PARAMETER_APPLY to store averaged parameter // else use PARAMETER_VALUE, and value backup in PARAMETER_GRADIENT AverageOptimizer(const OptimizationConfig& optConfig, @@ -65,7 +65,7 @@ public: virtual void setNoDecay() { optimizer_->setNoDecay(); } -protected: + protected: std::unique_ptr optimizer_; bool useApply_; @@ -98,7 +98,7 @@ protected: // Average Optimizer with Sparse support. class AverageSparseOptimizer : public AverageOptimizer { -public: + public: AverageSparseOptimizer(const OptimizationConfig& optConfig, ParameterOptimizer* optimizer, bool useParameterApply) @@ -130,7 +130,7 @@ public: t0Vec_.assign(t0Vec_.size(), 0); } -protected: + protected: /** * counting batches, clear after catch up with * t(timer_) is current time, diff --git a/paddle/parameter/CMakeLists.txt b/paddle/parameter/CMakeLists.txt index d2ae1c16c6b7316f1a6facdef4b933693d6ba818..19ae07e077e2b8f55ce4050566c9cf6aaa0efa0a 100644 --- a/paddle/parameter/CMakeLists.txt +++ b/paddle/parameter/CMakeLists.txt @@ -5,8 +5,6 @@ file(GLOB PARAMETERS_SOURCES . *.cpp) add_library(paddle_parameter STATIC ${PARAMETERS_SOURCES}) -add_style_check_target(paddle_parameter ${PARAMETERS_SOURCES}) -add_style_check_target(paddle_parameter ${PARAMETERS_HEADERS}) add_dependencies(paddle_parameter paddle_proto ${external_project_dependencies}) if(WITH_TESTING) add_subdirectory(tests) diff --git a/paddle/parameter/FirstOrderOptimizer.h b/paddle/parameter/FirstOrderOptimizer.h index 047989fcad52afc1d4d4c347258d0fb2f069f3d4..86b9a591aff7a58aafa194c64cb09cd6636d0454 100644 --- a/paddle/parameter/FirstOrderOptimizer.h +++ b/paddle/parameter/FirstOrderOptimizer.h @@ -22,7 +22,7 @@ namespace paddle { // Plain SGD optimization. class SgdOptimizer : public ParameterOptimizer { -public: + public: explicit SgdOptimizer(const OptimizationConfig& optConfig) : ParameterOptimizer(optConfig) { addParameterType(PARAMETER_MOMENTUM); @@ -77,7 +77,7 @@ class SparseMomentumParameterOptimizer : public ParameterOptimizer { \gamma_t: learning rate at the t'th step */ -public: + public: explicit SparseMomentumParameterOptimizer( const OptimizationConfig& optConfig); virtual void init(size_t numRows, const ParameterConfig* config); @@ -89,7 +89,7 @@ public: const ParameterConfig& config) const; virtual void finishBatch(); -private: + private: real alpha_; real beta_; real tau_; @@ -98,7 +98,7 @@ private: real momentum_; real decayRate_; -protected: + protected: int64_t timer_; mutable std::vector t0Vec_; bool isParameterSparse_; @@ -109,7 +109,7 @@ protected: * http://www.magicbroom.info/Papers/DuchiHaSi10.pdf */ class AdagradParameterOptimizer : public ParameterOptimizer { -public: + public: explicit AdagradParameterOptimizer(const OptimizationConfig& optConfig) : ParameterOptimizer(optConfig) { addParameterType(PARAMETER_MOMENTUM); @@ -129,7 +129,7 @@ public: virtual TraverseCallback needSpecialTraversal( const ParameterConfig& config) const; -protected: + protected: int64_t numUpdates_; static const int64_t kMaxNumAccumulates = 16384; }; @@ -139,7 +139,7 @@ protected: * http://www.matthewzeiler.com/pubs/googleTR2012/googleTR2012.pdf */ class AdaDeltaParameterOptimizer : public ParameterOptimizer { -public: + public: explicit AdaDeltaParameterOptimizer(const OptimizationConfig& optConfig) : ParameterOptimizer(optConfig) { addParameterType(PARAMETER_MOMENTUM); @@ -158,14 +158,14 @@ public: const ParameterConfig& config, size_t sparseId) const; -protected: + protected: real rou_; real epsilon_; }; // RMSProp Parameter Optimization. class RMSPropParameterOptimizer : public ParameterOptimizer { -public: + public: explicit RMSPropParameterOptimizer(const OptimizationConfig& optConfig) : ParameterOptimizer(optConfig) { addParameterType(PARAMETER_MOMENTUM); @@ -191,7 +191,7 @@ public: const ParameterConfig& config, size_t sparseId) const; -protected: + protected: real rou_; real epsilon_; @@ -208,7 +208,7 @@ protected: // Decayed AdaGrad Optimization. class DecayedAdagradParameterOptimizer : public ParameterOptimizer { -public: + public: explicit DecayedAdagradParameterOptimizer(const OptimizationConfig& optConfig) : ParameterOptimizer(optConfig) { addParameterType(PARAMETER_MOMENTUM); @@ -233,7 +233,7 @@ public: const ParameterConfig& config, size_t sparseId) const; -protected: + protected: real rou_; real epsilon_; @@ -253,7 +253,7 @@ protected: * Reference Paper: http://arxiv.org/abs/1412.6980 Algorithm 1 */ class AdamParameterOptimizer : public ParameterOptimizer { -public: + public: explicit AdamParameterOptimizer(const OptimizationConfig& optConfig) : ParameterOptimizer(optConfig), beta1_(optConfig.adam_beta1()), @@ -275,7 +275,7 @@ public: const ParameterConfig& config, size_t sparseId) const; -protected: + protected: real beta1_; real beta2_; real epsilon_; @@ -288,7 +288,7 @@ protected: * Reference Paper: http://arxiv.org/abs/1412.6980 Algorithm 2 */ class AdamaxParameterOptimizer : public ParameterOptimizer { -public: + public: explicit AdamaxParameterOptimizer(const OptimizationConfig& optConfig) : ParameterOptimizer(optConfig), beta1_(optConfig.adam_beta1()), @@ -305,7 +305,7 @@ public: const ParameterConfig& config, size_t sparseId) const; -protected: + protected: real beta1_; real beta2_; int64_t step_; @@ -315,7 +315,7 @@ protected: // Used in pserver, // when PARAMETER_DELTA stores in PARAMETER_GRADIENT. class AddOptimizer : public ParameterOptimizer { -public: + public: explicit AddOptimizer(const OptimizationConfig& optConfig) : ParameterOptimizer(optConfig) {} @@ -333,7 +333,7 @@ public: // A optimizer which does nothing. class DummyOptimizer : public ParameterOptimizer { -public: + public: explicit DummyOptimizer(const OptimizationConfig& optConfig) : ParameterOptimizer(optConfig) {} @@ -344,7 +344,7 @@ public: // Do gradient clipping before sgd update class OptimizerWithGradientClipping : public ParameterOptimizer { -public: + public: OptimizerWithGradientClipping(const OptimizationConfig& optConfig, ParameterOptimizer* optimizer) : ParameterOptimizer(optConfig), optimizer_(optimizer) { @@ -374,7 +374,7 @@ public: virtual void setNoDecay() { optimizer_->setNoDecay(); } -protected: + protected: std::unique_ptr optimizer_; }; diff --git a/paddle/parameter/LearningRateScheduler.cpp b/paddle/parameter/LearningRateScheduler.cpp index b6b58e3ddad6a0e8811bf56502c3f2f0c8728f5c..d57d2189a45dc8cbcea7a8a5f25c5ec7ac71cca3 100644 --- a/paddle/parameter/LearningRateScheduler.cpp +++ b/paddle/parameter/LearningRateScheduler.cpp @@ -28,20 +28,20 @@ LearningRateScheduler* LearningRateScheduler::create( // LRS stands for LearningRateScheduler class BaseLRS : public LearningRateScheduler { -public: + public: explicit BaseLRS(const OptimizationConfig& config) : learningRate_(config.learning_rate()), a_(config.learning_rate_decay_a()), b_(config.learning_rate_decay_b()) {} -protected: + protected: real learningRate_; real a_; real b_; }; class ConstLRS : public BaseLRS { -public: + public: explicit ConstLRS(const OptimizationConfig& config) : BaseLRS(config) {} virtual real calcLearningRate(int64_t numSamplesProcessed, int64_t pass) { return learningRate_; @@ -50,7 +50,7 @@ public: REGISTER_LEARNING_RATE_SCHEDULER(constant, ConstLRS); class PolyLRS : public BaseLRS { -public: + public: explicit PolyLRS(const OptimizationConfig& config) : BaseLRS(config) {} virtual real calcLearningRate(int64_t numSamplesProcessed, int64_t pass) { return learningRate_ * pow(1.0 + a_ * numSamplesProcessed, -b_); @@ -59,7 +59,7 @@ public: REGISTER_LEARNING_RATE_SCHEDULER(poly, PolyLRS); class CaffePolyLRS : public BaseLRS { -public: + public: explicit CaffePolyLRS(const OptimizationConfig& config) : BaseLRS(config) {} virtual real calcLearningRate(int64_t numSamplesProcessed, int64_t pass) { if (numSamplesProcessed > a_) { @@ -78,7 +78,7 @@ public: REGISTER_LEARNING_RATE_SCHEDULER(caffe_poly, CaffePolyLRS); class ExpLRS : public BaseLRS { -public: + public: explicit ExpLRS(const OptimizationConfig& config) : BaseLRS(config) {} virtual real calcLearningRate(int64_t numSamplesProcessed, int64_t pass) { double decayRatio = (double)numSamplesProcessed / b_; @@ -88,7 +88,7 @@ public: REGISTER_LEARNING_RATE_SCHEDULER(exp, ExpLRS); class DiscreteExpLRS : public BaseLRS { -public: + public: explicit DiscreteExpLRS(const OptimizationConfig& config) : BaseLRS(config) {} virtual real calcLearningRate(int64_t numSamplesProcessed, int64_t pass) { int numDecays = floor(numSamplesProcessed / b_); @@ -98,7 +98,7 @@ public: REGISTER_LEARNING_RATE_SCHEDULER(discexp, DiscreteExpLRS); class LinearLRS : public BaseLRS { -public: + public: explicit LinearLRS(const OptimizationConfig& config) : BaseLRS(config) {} virtual real calcLearningRate(int64_t numSamplesProcessed, int64_t pass) { return std::max(learningRate_ - a_ * numSamplesProcessed, b_); @@ -113,7 +113,7 @@ REGISTER_LEARNING_RATE_SCHEDULER(linear, LinearLRS); then learning_rate = learning_rate_base * rate_i */ class ManualLRS : public BaseLRS { -public: + public: explicit ManualLRS(const OptimizationConfig& config) : BaseLRS(config), currentSegment_(0), lastNum_(0) { std::vector pieces; @@ -151,7 +151,7 @@ public: return learningRate_ * rates_.back(); } -protected: + protected: std::vector rates_; std::vector segments_; size_t currentSegment_; @@ -161,7 +161,7 @@ protected: REGISTER_LEARNING_RATE_SCHEDULER(manual, ManualLRS); class PassManualLRS : public ManualLRS { -public: + public: explicit PassManualLRS(const OptimizationConfig& config) : ManualLRS(config) {} virtual real calcLearningRate(int64_t numSamplesProcessed, int64_t pass) { diff --git a/paddle/parameter/LearningRateScheduler.h b/paddle/parameter/LearningRateScheduler.h index aea99a1c204b46e937135cbde22360a12d087ae2..3fad97040248dcf8a22988c38153df31f267ed37 100644 --- a/paddle/parameter/LearningRateScheduler.h +++ b/paddle/parameter/LearningRateScheduler.h @@ -26,7 +26,7 @@ namespace paddle { }) class LearningRateScheduler { -public: + public: static LearningRateScheduler* create(const OptimizationConfig& config); virtual ~LearningRateScheduler() {} virtual real calcLearningRate(int64_t numSamplesProcessed, int64_t pass) = 0; diff --git a/paddle/parameter/OptimizerWithRegularizer.h b/paddle/parameter/OptimizerWithRegularizer.h index 7219d96d924dfa26d3ab52b8c6a2ce1249e4f45c..bd29b3966324b2e206cfe56cc15678539d1e870e 100644 --- a/paddle/parameter/OptimizerWithRegularizer.h +++ b/paddle/parameter/OptimizerWithRegularizer.h @@ -20,7 +20,7 @@ namespace paddle { // add regularizer for objective function to do optimization class OptimizerWithRegularizer : public ParameterOptimizer { -public: + public: static ParameterOptimizer* create(const OptimizationConfig& optConfig, const ParameterConfig& paraConfig, bool isParameterSparse, @@ -67,7 +67,7 @@ public: regularizer_->update(vecs, config, optimizer_->getLearningRate(), 0, 1); } -protected: + protected: std::unique_ptr optimizer_; Regularizer* regularizer_; @@ -84,7 +84,7 @@ protected: // Regularized Loss function for every num of batches class OptimizerWithRegularizerEveryNumBatches : public OptimizerWithRegularizer { -public: + public: OptimizerWithRegularizerEveryNumBatches(const OptimizationConfig& optConfig, ParameterOptimizer* optimizer, Regularizer* regularizer) @@ -112,7 +112,7 @@ public: virtual TraverseCallback startCatchUpWith() const; virtual void finishCatchUpWith() { baseTimer_ = timer_; } -protected: + protected: bool isRegularizationBatch(const ParameterConfig& config) const { return ((timer_ + 1) % config.num_batches_regularization() == 0); } @@ -125,7 +125,7 @@ protected: // Regularized Loss function with Sparse support class OptimizerWithRegularizerSparse : public OptimizerWithRegularizer { -public: + public: OptimizerWithRegularizerSparse(const OptimizationConfig& optConfig, ParameterOptimizer* optimizer, Regularizer* regularizer) @@ -145,7 +145,7 @@ public: t0Vec_.assign(t0Vec_.size(), 0); } -protected: + protected: /** * t0Vec_ are last occur time of i rows * if one block is update by multi threads, diff --git a/paddle/parameter/Parameter.h b/paddle/parameter/Parameter.h index 24ac10f3fe5977553332a9a8402d6795577b5ad8..ef519bf35a4f051b4477eb04b5eb2c5f0b5e29e8 100644 --- a/paddle/parameter/Parameter.h +++ b/paddle/parameter/Parameter.h @@ -58,7 +58,7 @@ class Parameter; typedef std::shared_ptr ParameterPtr; class Parameter { -public: + public: Parameter(const ParameterConfig& config, bool useGpu, bool doInit = true); const std::string& getName() const { return config_.name(); } @@ -311,7 +311,7 @@ public: } } -protected: + protected: /** * @brief create matrix to matType. * @@ -326,7 +326,7 @@ protected: void clearUpdate() { updateCounter_ = 0; } -protected: + protected: ParameterConfig config_; bool useGpu_; @@ -363,7 +363,7 @@ protected: std::vector> updaterHooks_; -public: + public: void setSharedCount(int cnt) { sharedCount_ = cnt; } int getSharedCount() { return sharedCount_; } diff --git a/paddle/parameter/ParameterOptimizer.h b/paddle/parameter/ParameterOptimizer.h index a8d0ca72f21d04e0e65a9dd6a07e8f53b23e4223..019afa1358ae255fd096e84e5eb1d7b0b9d6859f 100644 --- a/paddle/parameter/ParameterOptimizer.h +++ b/paddle/parameter/ParameterOptimizer.h @@ -30,12 +30,12 @@ namespace paddle { * may be called many times, should be no state change between calls. */ class ParameterOptimizer { -public: + public: typedef std::function TraverseCallback; -public: + public: explicit ParameterOptimizer(const OptimizationConfig& optConfig) : applyDecay_(true), optConfig_(optConfig), @@ -175,7 +175,7 @@ public: static ParameterOptimizer* create(const OptimizationConfig& optConfig, bool inPserver = false); -protected: + protected: typedef std::vector TraverseCallbackVec; static TraverseCallback composeCallbacks( diff --git a/paddle/parameter/ParameterUpdaterBase.h b/paddle/parameter/ParameterUpdaterBase.h index 717e1c6721b6e4d3ff81172eb06213677c3bff98..493512886cad3ea9b74026d6dfcc4fc90f6aadb9 100644 --- a/paddle/parameter/ParameterUpdaterBase.h +++ b/paddle/parameter/ParameterUpdaterBase.h @@ -21,7 +21,7 @@ namespace paddle { class ParameterOptimizer; class ParameterUpdater { -public: + public: ParameterUpdater() : parameterTypes_{PARAMETER_VALUE, PARAMETER_GRADIENT} {} virtual ~ParameterUpdater() {} @@ -89,7 +89,7 @@ public: virtual void setForwardbackwardTime(uint64_t delta) {} #endif -protected: + protected: virtual void updateImpl(Parameter* para) = 0; std::vector parameterTypes_; @@ -101,7 +101,7 @@ protected: // part of all Parameters. It's useful when we need different // update strategy for different Parameter. class ParameterUpdaterComposite : public ParameterUpdater { -public: + public: ParameterUpdaterComposite() {} virtual ~ParameterUpdaterComposite() {} @@ -173,7 +173,7 @@ public: [&](int tid, size_t numThreads) { updaters_[tid]->restore(); }); } -protected: + protected: virtual void updateImpl(Parameter* para) {} std::vector> updaters_; std::unique_ptr syncThreadPool_; diff --git a/paddle/parameter/ParameterUpdaterHook.cpp b/paddle/parameter/ParameterUpdaterHook.cpp index e6aec3c34820764b3515f47f13a432961de1a673..989185b66a5b7785bb0572fba59a72adeef9797b 100644 --- a/paddle/parameter/ParameterUpdaterHook.cpp +++ b/paddle/parameter/ParameterUpdaterHook.cpp @@ -37,7 +37,7 @@ namespace paddle { */ class StaticPruningHook : public IParameterUpdaterHook { -public: + public: explicit StaticPruningHook(const ParameterUpdaterHookConfig &hookConfig) : initCount_(0) { sparsityRatio_ = hookConfig.sparsity_ratio(); @@ -96,7 +96,7 @@ public: paraVec->dotMul(*maskVec_); } -private: + private: SameThreadChecker updateThreadChecker_; std::atomic initCount_; VectorPtr maskVec_; @@ -116,12 +116,12 @@ IParameterUpdaterHook::~IParameterUpdaterHook() {} * May be extracted to Util.h to unify the hasher. */ class StringIntPairHasher { -public: + public: size_t operator()(const std::pair &k) const { return intHasher_(strHasher_(k.first) + k.second); } -private: + private: std::hash strHasher_; std::hash intHasher_; }; diff --git a/paddle/parameter/ParameterUpdaterHook.h b/paddle/parameter/ParameterUpdaterHook.h index d30530ec393c097bf77e5e376e3c4dc84b321ed8..cb96e4cf007572e9688c11719017a9d2771ecd51 100644 --- a/paddle/parameter/ParameterUpdaterHook.h +++ b/paddle/parameter/ParameterUpdaterHook.h @@ -29,7 +29,7 @@ class Parameter; * parameter optimization. */ class IParameterUpdaterHook { -public: + public: virtual ~IParameterUpdaterHook(); /** @@ -53,7 +53,7 @@ public: */ virtual void init(Parameter* para) = 0; -protected: + protected: /** * Ctor. */ diff --git a/paddle/parameter/Regularizer.h b/paddle/parameter/Regularizer.h index 6bed7b0ddfe7b72c697af60f5243f9037999d54a..fa5384e23251b918cc914df36c16ad790a5c59c5 100644 --- a/paddle/parameter/Regularizer.h +++ b/paddle/parameter/Regularizer.h @@ -20,7 +20,7 @@ namespace paddle { // Regularizer function for parameter, e.g. L1/L2 class Regularizer { -public: + public: virtual void update(const VectorPtr vecs[], const ParameterConfig& paraConfig, real learningRate, // learningrate from optimizer diff --git a/paddle/parameter/Weight.h b/paddle/parameter/Weight.h index 7314c29d0db92db06d5b921c09de39d3b0029ef3..113dd6530c82fe1e831ad4a35e9cbcb9880b9243 100644 --- a/paddle/parameter/Weight.h +++ b/paddle/parameter/Weight.h @@ -23,12 +23,12 @@ limitations under the License. */ namespace paddle { class Weight { -private: + private: MatrixPtr weight_; MatrixPtr weightGrad_; ParameterPtr parameter_; -public: + public: Weight(size_t height, size_t width, ParameterPtr parameter); Weight(size_t height, size_t width, ParameterPtr parameter, size_t offset); diff --git a/paddle/parameter/tests/test_common.cpp b/paddle/parameter/tests/test_common.cpp index 6e10becabbbbb8861095fed5aab9ac1e05bcac91..89dcc6c751eb2ec07bfe8297c93d56c824086211 100644 --- a/paddle/parameter/tests/test_common.cpp +++ b/paddle/parameter/tests/test_common.cpp @@ -24,7 +24,7 @@ limitations under the License. */ using namespace paddle; // NOLINT class CommonTest : public ::testing::Test { -protected: + protected: CommonTest() : testStat_("test") {} virtual ~CommonTest() {} virtual void SetUp() { @@ -51,7 +51,7 @@ protected: virtual void TreaDown() { LOG(INFO) << "All Test Finished."; } -protected: + protected: std::vector> valueUint_; std::vector sizeVec_; real learningRate_; diff --git a/paddle/pserver/BaseClient.h b/paddle/pserver/BaseClient.h index a932d34712f56de1cbbf84a9db4476f862febca0..d50230e73a3a7d128cbfd1d70517fddd228fb1bb 100644 --- a/paddle/pserver/BaseClient.h +++ b/paddle/pserver/BaseClient.h @@ -32,7 +32,7 @@ namespace paddle { * connections. */ class BaseClient { -protected: + protected: typedef std::unique_ptr ThreadPtr; typedef std::vector> InputIovs; typedef std::vector SendRequest; @@ -49,7 +49,7 @@ protected: SendDataRequestVec parallelDataRequests; }; -public: + public: explicit BaseClient(bool separate = false, int numPorts = FLAGS_ports_num); virtual ~BaseClient(); @@ -141,7 +141,7 @@ public: return dataType; } -protected: + protected: /// for a > 0, b > 0: /// return the smallest x s.t. b*x >= a static int divup(int a, int b) { return (a + b - 1) / b; } @@ -264,7 +264,7 @@ protected: */ virtual void recv(int threadId) = 0; -protected: + protected: bool stopping_; /// nodes * ports that means the number of real pservers int serviceNum_; diff --git a/paddle/pserver/CMakeLists.txt b/paddle/pserver/CMakeLists.txt index f75475a88f7224ee3889827795088c8aa920b63b..0ae9c6ef6afc6ec5a99a685b08883def0db51cf1 100644 --- a/paddle/pserver/CMakeLists.txt +++ b/paddle/pserver/CMakeLists.txt @@ -14,9 +14,6 @@ set(NETWORK_HEADERS add_library(paddle_network STATIC ${NETWORK_SOURCES}) -add_style_check_target(paddle_network ${NETWORK_SOURCES}) -add_style_check_target(paddle_network ${NETWORK_HEADERS}) - add_dependencies(paddle_network paddle_proto ${external_project_dependencies}) ################### paddle_pserver ###################### @@ -37,9 +34,6 @@ set(PSERVER_HEADERS add_library(paddle_pserver STATIC ${PSERVER_SOURCES}) -add_style_check_target(paddle_pserver ${PSERVER_SOURCES}) -add_style_check_target(paddle_pserver ${PSERVER_HEADERS}) - add_dependencies(paddle_pserver paddle_proto ${external_project_dependencies}) set(PSERVER_MAIN_SOURCES diff --git a/paddle/pserver/LightNetwork.h b/paddle/pserver/LightNetwork.h index 2aaa26a5c708f9c01f006136619f599bcfe0db71..bcfc9655e989e80e08e9dce9b8734c0643cbf661 100644 --- a/paddle/pserver/LightNetwork.h +++ b/paddle/pserver/LightNetwork.h @@ -41,7 +41,7 @@ class SocketServer : public Thread { // rdmaCpu controls the cpu affinity of RDMA server daemon, // which could benifit performance. rdmaCpu = -1 means TCP // is used instead of RDMA transport. -public: + public: SocketServer(const std::string& addr, int port, int rdmaCpu); ~SocketServer(); @@ -50,7 +50,7 @@ public: typedef std::function& outputIovs)> ResponseCallback; -protected: + protected: // // The derived class needs to implement this function // to handle the request received by SocketWorker @@ -70,13 +70,13 @@ protected: friend class SocketWorker; -private: + private: void rdmaServer(); void tcpServer(); void detach() {} // detach accept thread is forbidden -protected: + protected: enum ChannelType tcpRdma_; // for rdma int rdmaCpu_; @@ -96,7 +96,7 @@ protected: * @note all parameter processing will run in the context of this worker */ class SocketWorker : public Thread { -public: + public: SocketWorker(std::unique_ptr&& channel, SocketServer* server) : channel_(std::move(channel)), server_(server) {} @@ -104,7 +104,7 @@ public: virtual void run(); -protected: + protected: std::unique_ptr channel_; SocketServer* server_; enum ChannelType tcpRdma_; @@ -118,12 +118,12 @@ protected: * single cpu core for better load balance performance */ class RdmaClientDaemons { -private: + private: RdmaClientDaemons(); static std::unique_ptr daemons_; -public: + public: static RdmaClientDaemons* get() { std::call_once(RdmaClientDaemons::initDataFlag_, &RdmaClientDaemons::getInstance); @@ -141,10 +141,10 @@ public: ~RdmaClientDaemons(); -public: + public: friend class SocketClient; -private: + private: static std::once_flag initDataFlag_; static void getInstance() { if (!daemons_.get()) daemons_.reset(new RdmaClientDaemons()); @@ -162,19 +162,19 @@ private: * read data */ class SocketClient { -public: + public: SocketClient(const std::string& serverAddr, int serverPort, enum ChannelType channelType); SocketChannel* getChannel() { return channel_.get(); } -protected: + protected: std::unique_ptr channel_; struct sxi_socket* socketDaemon_; enum ChannelType tcpRdma_; -private: + private: void RdmaClient(const std::string& serverAddr, int serverPort); void TcpClient(const std::string& serverAddr, int serverPort); }; diff --git a/paddle/pserver/ParameterClient2.h b/paddle/pserver/ParameterClient2.h index d63273ccbc8ed30d9df50d9f8b1a4d1e4fba6720..c96bb787151a525556c8217629109de201762cff 100644 --- a/paddle/pserver/ParameterClient2.h +++ b/paddle/pserver/ParameterClient2.h @@ -50,11 +50,11 @@ struct PServerVector { * @brief A class to help to prepare server-side operations. */ class PreparedOperations { -protected: + protected: class ResultsAdder; struct LocalOperationResult; -public: + public: /** * Offers an easy way to prepare operations that will be performed on * server-side. @@ -93,7 +93,7 @@ public: return ResultsAdder(&localResults_.back()); } -protected: + protected: void addOperationHelper(Operation* op) {} /** @@ -151,7 +151,7 @@ protected: * @brief ResultsAdder offers easy ways to quickly store operation results. */ class ResultsAdder { - public: + public: explicit ResultsAdder(LocalOperationResult* localResult) : localResult_(localResult) {} template @@ -172,11 +172,11 @@ protected: addResult(args...); } - protected: + protected: LocalOperationResult* localResult_; }; -protected: + protected: DoOperationRequest request_; std::vector inputIovs_; struct LocalOperationResult { @@ -214,7 +214,7 @@ struct ParameterSegments { * waiting until all parameters are received to CPU host end. */ class ParameterClient2 : public BaseClient { -public: + public: /** Constructor. * @param separate True if sending and recieving activities are separated * into 2 threads, otherwise false. @@ -232,7 +232,7 @@ public: static int calcParameterBlockSize(const std::vector& parameters, size_t serviceNum); -public: + public: bool init(const std::vector& parameters); /// service functions @@ -514,7 +514,7 @@ public: void setForwardbackwardTime(uint64_t delta) { forwardbackwordTime_ = delta; } #endif -protected: + protected: template void multiCall(const char* funcName, const ProtoIn& request, @@ -529,7 +529,7 @@ protected: } } -private: + private: void destroy(); /** @@ -573,7 +573,7 @@ private: /// start necessary threads for threadPool void initThreads(); -protected: + protected: /// start port number of pserver /// it deduce all ports for dense and sparse with some rules int port_; diff --git a/paddle/pserver/ParameterServer2.h b/paddle/pserver/ParameterServer2.h index 3ed06b6b045802bcfd48bcff6bd0c1b34e9bbb86..0b8ef5c170c01ec8a5d53f01db9888f82ca68eec 100644 --- a/paddle/pserver/ParameterServer2.h +++ b/paddle/pserver/ParameterServer2.h @@ -71,7 +71,7 @@ namespace paddle { * to prevent from being polluted. */ class ParameterServer2 : public ProtoServer { -protected: + protected: /// parameter_ mutex. RWLock parameterMutex_; @@ -169,7 +169,7 @@ protected: template class ReadWriteBuffer : public std::vector> { - public: + public: static_assert(sizeof(T) % AlignBytes == 0 || AlignBytes % sizeof(T) == 0, "Type T must be able to aligned."); @@ -229,7 +229,7 @@ protected: return r; } - private: + private: size_t curOffset_; }; @@ -298,17 +298,17 @@ protected: /// barrier performance tuning sync-sgd required std::atomic batchId_; -public: + public: struct Buffer { real* base; size_t size; }; -protected: + protected: /// async gradient commit control bool asyncGrdientCommitCheckAndStat(const SendParameterRequest& request); -public: + public: /// disable default parameter for overloading /// @rdmaCpu:the id of cpu core hosting RDMA server(0-N) /// -1 means using TCP transport instead of RDMA @@ -437,7 +437,7 @@ public: void saveValueVector(const SaveValueRequest& request, ProtoResponseCallback callback); -public: + public: /** * @brief initialize parameter server */ @@ -512,7 +512,7 @@ public: SendParameterResponse* response, std::vector* outputBuffers); -protected: + protected: void mergeSegments(BlockSegments* segments); /// set the unused segments to zero @@ -641,7 +641,7 @@ protected: const VectorPtr vecs[], const ParameterOptimizer::TraverseCallback& callback); -public: + public: typedef void (ParameterServer2::*OperatorFunction)(const Operation& operation, OperationResult* result); diff --git a/paddle/pserver/ParameterServerController.h b/paddle/pserver/ParameterServerController.h index 3a9bc74edf240a12fe1f7bd266f0311555349311..1308d62fb1787f19123fe37d49f8e14039c5a39a 100644 --- a/paddle/pserver/ParameterServerController.h +++ b/paddle/pserver/ParameterServerController.h @@ -28,7 +28,7 @@ namespace paddle { * by gflags or proto. */ class ParameterServerController final { -public: + public: DISABLE_COPY(ParameterServerController); /** @@ -67,7 +67,7 @@ public: */ void wait(); -private: + private: std::vector> parameterServers_; }; diff --git a/paddle/pserver/ProtoServer.h b/paddle/pserver/ProtoServer.h index 3f78799dbfe1d4b80249e8cb27f269e6358903dd..2943867de5885ab1af1aa0f69e93a931092b28e3 100644 --- a/paddle/pserver/ProtoServer.h +++ b/paddle/pserver/ProtoServer.h @@ -34,7 +34,7 @@ namespace paddle { * for single NIC hardward with --port=N(N>1) for small cluster job. */ class ProtoServer : public SocketServer { -public: + public: /// rdmaCpu controls the cpu affinity of RDMA server daemon, /// which could benifit performance. rdmaCpu = -1 means TCP /// is used instead of RDMA transport. @@ -87,7 +87,7 @@ public: std::unique_ptr msgReader, ProtoResponseCallbackEx callback)> func); -protected: + protected: /** * @brief handle rpc request * @param[in] msgReader Message reader for reading data from connection @@ -111,7 +111,7 @@ protected: void registerServiceFunctionImp(const std::string& funcName, ServiceFunction func); -protected: + protected: /// Tuning bare network overhead: the beginning of receiving request ThreadLocal handleRequestBegin_; @@ -120,7 +120,7 @@ protected: }; class ProtoClient : public SocketClient { -public: + public: ProtoClient(const std::string& serverAddr, int serverPort, enum ChannelType channelType = F_TCP) diff --git a/paddle/pserver/SocketChannel.h b/paddle/pserver/SocketChannel.h index c0f30d0db760045a8c0cb001fcadaae8f0c03f9d..8b45ac56090ef82e77514566e7df6b366958655e 100644 --- a/paddle/pserver/SocketChannel.h +++ b/paddle/pserver/SocketChannel.h @@ -33,7 +33,7 @@ enum ChannelType { /// reading a set of blocks of data from SocketChannel. class MsgReader { -public: + public: MsgReader(SocketChannel* channel, size_t numIovs); ~MsgReader() { /// ensure all data blocks have been processed @@ -75,7 +75,7 @@ public: void readBlocks(const std::vector& bufs); void readNextBlock(void* buf); -protected: + protected: SocketChannel* channel_; std::vector blockLengths_; size_t currentBlockIndex_; @@ -84,7 +84,7 @@ protected: /// APIs for reading and writing byte stream data or naive iov data /// from the APIs both RDMA and TCP exhibits byte stream style class SocketChannel { -public: + public: SocketChannel(int socket, const std::string& peerName) : tcpSocket_(socket), peerName_(peerName) { tcpRdma_ = F_TCP; @@ -137,7 +137,7 @@ public: /// return null to indicate socket is closed std::unique_ptr readMessage(); -protected: + protected: struct MessageHeader { int64_t totalLength; /// include the header int64_t numIovs; diff --git a/paddle/pserver/SparseParameterDistribution.h b/paddle/pserver/SparseParameterDistribution.h index 13f199548d56262e77e91e45052f3e435dea407c..e168f36c75e9452fff547f139a67a553cc6b796a 100644 --- a/paddle/pserver/SparseParameterDistribution.h +++ b/paddle/pserver/SparseParameterDistribution.h @@ -31,7 +31,7 @@ namespace paddle { * if unbalanced distribution exhibts by default. */ class SparseParameterDistribution { -public: + public: /// serviceNum means the number of ParameterServers explicit SparseParameterDistribution(size_t serviceNum); ~SparseParameterDistribution() {} @@ -39,7 +39,7 @@ public: void probeDistribution(int serverId, size_t data); void checkAndResetDistribution(); -private: + private: std::vector data_; std::atomic totBytes_; diff --git a/paddle/pserver/test/SocketTest.cpp b/paddle/pserver/test/SocketTest.cpp index 6019dccaadf7fab5a1db7183c07cbbd9562dab2e..206cd17c379f529579c103893cfb492524bc6f8d 100644 --- a/paddle/pserver/test/SocketTest.cpp +++ b/paddle/pserver/test/SocketTest.cpp @@ -30,12 +30,12 @@ struct MessageHeader { }; class Thread { -public: + public: void start(); virtual void run() = 0; virtual ~Thread() {} -protected: + protected: std::unique_ptr thread_; }; @@ -44,13 +44,13 @@ void Thread::start() { } class SocketChannel { -public: + public: explicit SocketChannel(int socket) : socket_(socket) {} int getSocketFd() const { return socket_; } uint64_t readAll(void* buf, size_t size); uint64_t writeAll(const void* buf, size_t size); -protected: + protected: int socket_; }; @@ -79,7 +79,7 @@ uint64_t SocketChannel::writeAll(const void* buf, size_t size) { } class SocketWorker : public Thread { -public: + public: explicit SocketWorker(int socket) : channel_(socket) {} virtual void run(); @@ -88,19 +88,19 @@ public: // write n bytes -protected: + protected: SocketChannel channel_; std::string buffer_; }; class SocketServer : public Thread { -public: + public: explicit SocketServer(int port) : port_(port), socket_(0), maxPendingConnections_(100) {} virtual void run(); -protected: + protected: int port_; int socket_; int maxPendingConnections_; @@ -161,11 +161,11 @@ void SocketWorker::run() { } class SocketClient { -public: + public: SocketClient(const std::string& serverAddr, int serverPort); SocketChannel* getChannel() const { return channel_.get(); } -protected: + protected: std::unique_ptr channel_; }; diff --git a/paddle/pserver/test/test_ParameterServer2.cpp b/paddle/pserver/test/test_ParameterServer2.cpp index e742cd0871da865e02a60a125a936eea8f15e575..01d179258dffaf996a57022801ee3bd60a268f77 100644 --- a/paddle/pserver/test/test_ParameterServer2.cpp +++ b/paddle/pserver/test/test_ParameterServer2.cpp @@ -26,7 +26,7 @@ DEFINE_string(server_addr, "127.0.0.1", "assign server address"); DEFINE_int32(server_cpu, 0, "assign server cpu"); class ParameterServer2Tester : public ParameterServer2 { -public: + public: ParameterServer2Tester(std::string serverAddr, int port, int rdmaCpu = -1, @@ -88,7 +88,7 @@ public: void waitPassFinishTest(); void synchronizeTest(); -protected: + protected: ParameterClient2 client_; vector clientConfigs_; vector parameters_; diff --git a/paddle/pserver/test/test_ProtoServer.cpp b/paddle/pserver/test/test_ProtoServer.cpp index d68a8d2180cc3081346106132799498f6dc3fa20..a66b14a1cc58d11988e4936a9c35d98b8bf5edc1 100644 --- a/paddle/pserver/test/test_ProtoServer.cpp +++ b/paddle/pserver/test/test_ProtoServer.cpp @@ -28,7 +28,7 @@ DEFINE_bool(benchmark, false, "Do benchmark. Skip some tests"); using namespace paddle; // NOLINT class MyServer : public ProtoServer { -public: + public: explicit MyServer(int port, int rdmaCpu = -1) : ProtoServer(FLAGS_server_addr, port, rdmaCpu), status_(PSERVER_STATUS_NOT_SET) { @@ -62,7 +62,7 @@ public: callback(response); } -protected: + protected: PServerStatus status_; std::string buffer_; }; diff --git a/paddle/scripts/docker/build.sh b/paddle/scripts/docker/build.sh deleted file mode 100755 index 94628270228b9e7fd32405bdcb5e11c163ba4791..0000000000000000000000000000000000000000 --- a/paddle/scripts/docker/build.sh +++ /dev/null @@ -1,259 +0,0 @@ -#!/bin/bash - -function cmake_gen() { - mkdir -p /paddle/build - cd /paddle/build - - # build script will not fail if *.deb does not exist - rm *.deb 2>/dev/null || true - # delete previous built whl packages - rm -rf /paddle/paddle/dist 2>/dev/null || true - - # Support build for all python versions, currently - # including cp27-cp27m and cp27-cp27mu. - PYTHON_FLAGS="" - if [ "$1" != "" ]; then - echo "using python abi: $1" - if [ "$1" == "cp27-cp27m" ]; then - export LD_LIBRARY_PATH=/opt/_internal/cpython-2.7.11-ucs2/lib:${LD_LIBRARY_PATH#/opt/_internal/cpython-2.7.11-ucs4/lib:} - export PATH=/opt/python/cp27-cp27m/bin/:${PATH} - PYTHON_FLAGS="-DPYTHON_EXECUTABLE:FILEPATH=/opt/python/cp27-cp27m/bin/python - -DPYTHON_INCLUDE_DIR:PATH=/opt/python/cp27-cp27m/include/python2.7 - -DPYTHON_LIBRARIES:FILEPATH=/opt/_internal/cpython-2.7.11-ucs2/lib/libpython2.7.so" - elif [ "$1" == "cp27-cp27mu" ]; then - export LD_LIBRARY_PATH=/opt/_internal/cpython-2.7.11-ucs4/lib:${LD_LIBRARY_PATH#/opt/_internal/cpython-2.7.11-ucs2/lib:} - export PATH=/opt/python/cp27-cp27mu/bin/:${PATH} - PYTHON_FLAGS="-DPYTHON_EXECUTABLE:FILEPATH=/opt/python/cp27-cp27mu/bin/python - -DPYTHON_INCLUDE_DIR:PATH=/opt/python/cp27-cp27mu/include/python2.7 - -DPYTHON_LIBRARIES:FILEPATH=/opt/_internal/cpython-2.7.11-ucs4/lib/libpython2.7.so" - fi - fi - - cat < /paddle/build/Dockerfile < - ENV HOME /root -EOF - - if [[ ${WITH_GPU} == "ON" ]]; then - NCCL_DEPS="apt-get install -y libnccl2=2.1.2-1+cuda8.0 libnccl-dev=2.1.2-1+cuda8.0 &&" - else - NCCL_DEPS="" - fi - - if [[ ${WITH_FLUID_ONLY:-OFF} == "OFF" ]]; then - PADDLE_VERSION="paddle version" - CMD='"paddle", "version"' - else - PADDLE_VERSION="true" - CMD='"true"' - fi - - cat >> /paddle/build/Dockerfile <> /paddle/build/Dockerfile <> /paddle/build/Dockerfile <= 21." - ANDROID_API=21 - fi -else # armeabi, armeabi-v7a - ANDROID_ARCH=arm -fi - -ANDROID_STANDALONE_TOOLCHAIN=$ANDROID_TOOLCHAINS_DIR/$ANDROID_ARCH-android-$ANDROID_API - -cat <> /paddle/build/Dockerfile <> ${PADDLE_ROOT}/build/Dockerfile <> ${PADDLE_ROOT}/build/Dockerfile <> ${PADDLE_ROOT}/build/Dockerfile < /dev/null - return $? -} - function start_build_docker() { docker pull $IMG - if container_running "${CONTAINER_ID}"; then - docker stop "${CONTAINER_ID}" 1>/dev/null - docker rm -f "${CONTAINER_ID}" 1>/dev/null - fi - + apt_mirror='s#http://archive.ubuntu.com/ubuntu#mirror://mirrors.ubuntu.com/mirrors.txt#g' DOCKER_ENV=$(cat <&2 - echo "Please use pre-commit to check what is wrong." 1>&2 - exit 1 -} - -trap 'abort' 0 -set -e - -# install glide -curl https://glide.sh/get | bash -eval "$(GIMME_GO_VERSION=1.8.3 gimme)" - -# set up go environment for running gometalinter -mkdir -p $GOPATH/src/github.com/PaddlePaddle/ -ln -sf $TRAVIS_BUILD_DIR $GOPATH/src/github.com/PaddlePaddle/Paddle -cd $GOPATH/src/github.com/PaddlePaddle/Paddle/go; glide install; cd - - -go get github.com/alecthomas/gometalinter -gometalinter --install - -cd $TRAVIS_BUILD_DIR -export PATH=/usr/bin:$PATH -pre-commit install -clang-format --version - - - -if ! pre-commit run -a ; then - git diff - exit 1 -fi - -trap : 0 diff --git a/paddle/scripts/travis/deploy_key.enc b/paddle/scripts/travis/deploy_key.enc deleted file mode 100644 index b0aa45c5ac626c735735fd8541a43bf8b099d0a0..0000000000000000000000000000000000000000 Binary files a/paddle/scripts/travis/deploy_key.enc and /dev/null differ diff --git a/paddle/trainer/CMakeLists.txt b/paddle/trainer/CMakeLists.txt index 72911695bd4959d73d783897b0c5e674454c30bc..6192de4388c8c3f5165fb88b443d372748f7a17e 100644 --- a/paddle/trainer/CMakeLists.txt +++ b/paddle/trainer/CMakeLists.txt @@ -36,17 +36,12 @@ endif() add_library(paddle_trainer_lib STATIC ${TRAINER_SOURCES}) -add_style_check_target(paddle_trainer_lib - ${TRAINER_SOURCES}) -add_style_check_target(paddle_trainer_lib - ${TRAINER_HEADERS}) add_dependencies(paddle_trainer_lib paddle_proto ${external_project_dependencies}) macro(add_paddle_exe TARGET_NAME) add_executable(${TARGET_NAME} ${ARGN}) - add_style_check_target(${TARGET_NAME} ${ARGN}) link_paddle_exe(${TARGET_NAME}) endmacro() diff --git a/paddle/trainer/NewRemoteParameterUpdater.h b/paddle/trainer/NewRemoteParameterUpdater.h index 6223ba427c9b94494c2bee8f0847442f1b0574c9..02693c675e6f5cb574e52e9681963a5904676028 100644 --- a/paddle/trainer/NewRemoteParameterUpdater.h +++ b/paddle/trainer/NewRemoteParameterUpdater.h @@ -29,7 +29,7 @@ namespace paddle { * New remote parameter updater for dense parameters that use cclient of go. */ class NewRemoteParameterUpdater : public ParameterUpdater { -public: + public: NewRemoteParameterUpdater(const OptimizationConfig& config, const std::string pserverSpec); NewRemoteParameterUpdater(const OptimizationConfig& config, @@ -61,13 +61,13 @@ public: virtual void startPass(); virtual bool finishPass(); -protected: + protected: /** * work need to do after finishBatch */ virtual void updateImpl(Parameter* para); -private: + private: int parameterSize() { return (int)parameters_.size(); } /** @@ -104,7 +104,7 @@ private: } } -protected: + protected: const OptimizationConfig& trainerConfig_; /// internal parameter client object for exchanging data with pserver paddle_pserver_client parameterClient_; diff --git a/paddle/trainer/ParamUtil.h b/paddle/trainer/ParamUtil.h index 2e05595848760c9abd7d916003656c8103151abf..10746b4d58e3a82c081987a6aaad9e0b42272a03 100644 --- a/paddle/trainer/ParamUtil.h +++ b/paddle/trainer/ParamUtil.h @@ -56,7 +56,7 @@ struct ParameterUtilConfig { * Utility class for loading and saving parameters */ class ParameterUtil { -public: + public: /** * Ctor. * @@ -115,7 +115,7 @@ public: } } -private: + private: std::shared_ptr config_; std::unique_ptr intConfig_; GradientMachinePtr gserver_; diff --git a/paddle/trainer/ParameterUpdater.h b/paddle/trainer/ParameterUpdater.h index 9e9e948b8856d2712f8894b3d14db9c795d5f694..ef7ab92eca77bab2a8481561713f8034d2b8505d 100644 --- a/paddle/trainer/ParameterUpdater.h +++ b/paddle/trainer/ParameterUpdater.h @@ -36,7 +36,7 @@ namespace paddle { * @brief Parameter Updater for SGD, and local(not cluster) run. */ class SgdLocalUpdater : public ParameterUpdater { -public: + public: /** * @brief Ctor. Initialize optimizer locally by optConfig. * @param optConfig optimization config. @@ -131,7 +131,7 @@ public: } } -protected: + protected: /** * @brief update method. Update value from gradient. * @param para parameter that will be updated. @@ -159,7 +159,7 @@ protected: * @deprecated */ class SgdCpuUpdater : public SgdLocalUpdater, public Deprecated { -public: + public: explicit SgdCpuUpdater(const OptimizationConfig& optConfig) : SgdLocalUpdater(optConfig), Deprecated( @@ -178,7 +178,7 @@ public: optimizer_->finishBatch(); } -protected: + protected: /** * @brief do nothing. * @param para @@ -192,7 +192,7 @@ protected: * It will do model average in cpu to reduce gpu memory comsuption. */ class SgdUpdaterWithCpuAverager : public SgdLocalUpdater { -public: + public: /** * @brief Ctor. * @@ -233,12 +233,12 @@ public: */ virtual void restore(); -protected: + protected: virtual void updateImpl(Parameter* para); void updateFunc(Parameter* para); -protected: + protected: std::unique_ptr averager_; /** diff --git a/paddle/trainer/RemoteParameterUpdater.h b/paddle/trainer/RemoteParameterUpdater.h index 5e82c944751629632ea8d16992bd8f4178a2fbd5..3a40a46354efd6b92278884c8f5b72504a3ff283 100644 --- a/paddle/trainer/RemoteParameterUpdater.h +++ b/paddle/trainer/RemoteParameterUpdater.h @@ -53,7 +53,7 @@ namespace paddle { * backward and communication is not supported. */ class RemoteParameterUpdater : public ParameterUpdater { -public: + public: RemoteParameterUpdater( const OptimizationConfig& config, int expectedPassCount, @@ -101,7 +101,7 @@ public: virtual void apply(); virtual void restore(); -protected: + protected: /** * control all pservers with all trainers for sync-sgd */ @@ -128,7 +128,7 @@ protected: */ void copyParametersFromDevice(ParameterType parameterType); -protected: + protected: /// Optimization config used to guide initialization and finishBatch OptimizationConfig config_; /// internal parameter client object for exchanging data with pserver @@ -178,7 +178,7 @@ protected: * It contains separate send and recv thread for pipeline usage. */ class ConcurrentRemoteParameterUpdater : public RemoteParameterUpdater { -public: + public: ConcurrentRemoteParameterUpdater( OptimizationConfig config, int expectedPassCount, @@ -194,7 +194,7 @@ public: */ virtual void finishBatch(real cost); -protected: + protected: virtual void updateImpl(Parameter* para); /// internal thread called in send thread void send(Parameter* para); // para == NULL indicate end of a minibatch @@ -221,7 +221,7 @@ protected: return (numBatches_ + 1) % config_.num_batches_per_send_parameter() == 0; } -private: + private: /// send thread used for overlapping std::unique_ptr sendThread_; /// recv thread used for overlapping @@ -263,7 +263,7 @@ private: * to encapsulate sparse specified message for all pservers. */ class SparseRemoteParameterUpdater : public ParameterUpdater { -public: + public: SparseRemoteParameterUpdater(const OptimizationConfig& config, int expectedPassCount, bool testing); @@ -303,7 +303,7 @@ public: } #endif -protected: + protected: /// update implimentation, not implemented virtual void updateImpl(Parameter* para) {} @@ -313,7 +313,7 @@ protected: /// start controller thread void startController(); -protected: + protected: /// optimization config OptimizationConfig config_; /// internal parameter client @@ -335,7 +335,7 @@ protected: * it directly call internal dense and sparse udpater individually. */ class SparseRemoteParameterUpdaterComposite : public ParameterUpdaterComposite { -public: + public: enum { UPDATER_SPARSE_REMOTE = 0, // execute in sync thread pool(tid:0) UPDATER_NORMAL = 1, // execute in Owner thread(tid:1) @@ -364,7 +364,7 @@ public: }; class ParameterUpdaterCreators { -public: + public: /** * @brief add a creator to create custom ParameterUpdater while training. * The creator is a function with type (alogrithm, optConfig, isLocal, @@ -407,7 +407,7 @@ public: return nullptr; } -private: + private: static std::vector> constructors_; diff --git a/paddle/trainer/Tester.h b/paddle/trainer/Tester.h index e892744db278586f2fd5b3cb527aa7c17752c477..801c77e3116369732bf4b03107adce6a71dc2184 100644 --- a/paddle/trainer/Tester.h +++ b/paddle/trainer/Tester.h @@ -38,7 +38,7 @@ namespace paddle { * It is a private class for Trainer. */ class Tester { -public: + public: /** * Ctor * @param config Trainer Config. @@ -87,7 +87,7 @@ public: */ void test(); -protected: + protected: std::shared_ptr testParameterClient_; std::shared_ptr config_; std::unique_ptr intconfig_; @@ -107,7 +107,7 @@ protected: real cost; } testContext_; -private: + private: /** * Test one batch by batchId. It is only used for testOnePass. * diff --git a/paddle/trainer/ThreadParameterUpdater.h b/paddle/trainer/ThreadParameterUpdater.h index bc08a9e9f0eda1cab7776ba76c67e88add1028a9..b5e6a7ce3c8457364b10c921bca3386fbb6f6cbf 100644 --- a/paddle/trainer/ThreadParameterUpdater.h +++ b/paddle/trainer/ThreadParameterUpdater.h @@ -39,7 +39,7 @@ namespace paddle { class. */ class SgdThreadUpdater : public ParameterUpdater { -public: + public: explicit SgdThreadUpdater(const OptimizationConfig& optConfig); virtual ~SgdThreadUpdater() {} @@ -57,7 +57,7 @@ public: virtual void apply(); virtual void restore(); -protected: + protected: // This is the function that will be eventualy called by the GradientMachine. // used only for GPU update. virtual void updateImpl(Parameter* para); diff --git a/paddle/trainer/Trainer.h b/paddle/trainer/Trainer.h index fac589d1d711affcd008f90edf87d865c8362f69..78127b7be5cef34f51a4b540852c139625b571dd 100644 --- a/paddle/trainer/Trainer.h +++ b/paddle/trainer/Trainer.h @@ -41,7 +41,7 @@ namespace paddle { * train/test a NeuralNetwork. */ class Trainer { -public: + public: /** * Ctor. * @return @@ -138,7 +138,7 @@ public: */ ParameterUtil* getParameterUtilPtr(); -protected: + protected: /** * Train one pass of data. * @@ -159,10 +159,10 @@ protected: void createTester(); -private: + private: std::unique_ptr createTesterConfig(); -protected: + protected: std::shared_ptr config_; std::shared_ptr stats_; diff --git a/paddle/trainer/TrainerConfigHelper.h b/paddle/trainer/TrainerConfigHelper.h index f1366cc041b0d983e65a1bf5b02ec2128324c5a8..b21dda964e70fce6e5e9672cc131595ad5af3bbc 100644 --- a/paddle/trainer/TrainerConfigHelper.h +++ b/paddle/trainer/TrainerConfigHelper.h @@ -37,7 +37,7 @@ class DataConfig; * Define a macro to unify 'final' keyword */ class TrainerConfigHelper /*final*/ { -public: + public: DISABLE_COPY(TrainerConfigHelper); /** @@ -193,7 +193,7 @@ public: */ static std::shared_ptr createFromFlagConfig(); -private: + private: static std::string getConfigNameFromPassId(int passId, const std::string& modelPath); diff --git a/paddle/trainer/TrainerInternal.h b/paddle/trainer/TrainerInternal.h index 7018faab24744f7a087a53130acc56ec6314101e..48ee53a5e60f950bfc3cc299c754b0e72601c818 100644 --- a/paddle/trainer/TrainerInternal.h +++ b/paddle/trainer/TrainerInternal.h @@ -34,7 +34,7 @@ namespace paddle { * the core training class for driving training logic */ class TrainerInternal { -public: + public: struct ParaStat { real maxAbsGrad; real avgAbsGrad; @@ -126,7 +126,7 @@ public: UpdateCallback updateCallback, bool doPipelineUpdate); -protected: + protected: std::shared_ptr parameterUpdater_; GradientMachinePtr gradientMachine_; std::shared_ptr config_; diff --git a/paddle/trainer/TrainerInternalConfig.h b/paddle/trainer/TrainerInternalConfig.h index b47692720efc2ed4f2db84f61ca81fcb52d234c0..43aae381029784278ad58c9398f64af24dffa1df 100644 --- a/paddle/trainer/TrainerInternalConfig.h +++ b/paddle/trainer/TrainerInternalConfig.h @@ -37,7 +37,7 @@ namespace paddle { * through one mini-batch. */ class TrainerStats { -public: + public: /** * @brief reset all stats. * @@ -147,7 +147,7 @@ public: return os.str(); } -private: + private: int64_t numProcessed_; real totalCost_; real currentCost_; diff --git a/paddle/trainer/tests/picojson.h b/paddle/trainer/tests/picojson.h index eaa8b9baf6e4e753a441ab77811f494cbdab80cf..75349537b1c7f10d23bae788e8414a753c7ccab0 100644 --- a/paddle/trainer/tests/picojson.h +++ b/paddle/trainer/tests/picojson.h @@ -125,7 +125,7 @@ enum { INDENT_WIDTH = 2 }; struct null {}; class value { -public: + public: typedef std::vector array; typedef std::map object; union _storage { @@ -139,11 +139,11 @@ public: object* object_; }; -protected: + protected: int type_; _storage u_; -public: + public: value(); value(int type, bool); explicit value(bool b); @@ -179,7 +179,7 @@ public: void serialize(Iter os, bool prettify = false) const; std::string serialize(bool prettify = false) const; -private: + private: template value(const T*); // intentionally defined to block implicit conversion of // pointer to bool @@ -588,13 +588,13 @@ inline std::string value::_serialize(int indent) const { template class input { -protected: + protected: Iter cur_, end_; int last_ch_; bool ungot_; int line_; -public: + public: input(const Iter& first, const Iter& last) : cur_(first), end_(last), last_ch_(-1), ungot_(false), line_(1) {} int getc() { @@ -873,7 +873,7 @@ inline bool _parse(Context& ctx, input& in) { } class deny_parse_context { -public: + public: bool set_null() { return false; } bool set_bool(bool) { return false; } #ifdef PICOJSON_USE_INT64 @@ -898,10 +898,10 @@ public: }; class default_parse_context { -protected: + protected: value* out_; -public: + public: default_parse_context(value* out) : out_(out) {} bool set_null() { *out_ = value(); @@ -949,18 +949,18 @@ public: return _parse(ctx, in); } -private: + private: default_parse_context(const default_parse_context&); default_parse_context& operator=(const default_parse_context&); }; class null_parse_context { -public: + public: struct dummy_str { void push_back(int) {} }; -public: + public: null_parse_context() {} bool set_null() { return true; } bool set_bool(bool) { return true; } @@ -985,7 +985,7 @@ public: return _parse(*this, in); } -private: + private: null_parse_context(const null_parse_context&); null_parse_context& operator=(const null_parse_context&); }; diff --git a/paddle/trainer/tests/test_TrainerOnePass.cpp b/paddle/trainer/tests/test_TrainerOnePass.cpp index b2a93d4d5eea37ad716b59427f2aa4409d2f537d..de12c4d649c6041f497c0eeac0904ebfc0d5bf97 100644 --- a/paddle/trainer/tests/test_TrainerOnePass.cpp +++ b/paddle/trainer/tests/test_TrainerOnePass.cpp @@ -38,7 +38,7 @@ DECLARE_int32(num_passes); DECLARE_int32(saving_period); class TrainerForTest : public paddle::Trainer { -public: + public: inline const std::shared_ptr& getParameterUpdaterForTest() { return this->trainerInternal_.getParameterUpdater(); } diff --git a/paddle/utils/CMakeLists.txt b/paddle/utils/CMakeLists.txt index 6292e7fa52cd86c71724d9fe84ea622e98ff1e08..b42b2bae968a10c581c594054f853347eb5d5445 100644 --- a/paddle/utils/CMakeLists.txt +++ b/paddle/utils/CMakeLists.txt @@ -14,9 +14,6 @@ add_library(paddle_utils STATIC ${UTIL_SOURCES} ${UTIL_ARCH_SOURCES} ${UTIL_RES}) -add_style_check_target(paddle_utils ${UTIL_HEADERS}) -add_style_check_target(paddle_utils ${UTIL_SOURCES} - ${UTIL_ARCH_SOURCES}) add_dependencies(paddle_utils paddle_proto ${external_project_dependencies}) if(WITH_TESTING) add_subdirectory(tests) diff --git a/paddle/utils/ClassRegistrar.h b/paddle/utils/ClassRegistrar.h index 1ac27bafabd1945d1d01e3bead22b0dd200d8688..5f40a0b25e92c7adcfe3f8c4be96016be801da3b 100644 --- a/paddle/utils/ClassRegistrar.h +++ b/paddle/utils/ClassRegistrar.h @@ -41,7 +41,7 @@ namespace paddle { */ template class ClassRegistrar { -public: + public: typedef std::function ClassCreator; // Register a class using a creation function. @@ -74,7 +74,7 @@ public: } } -protected: + protected: std::map creatorMap_; }; diff --git a/paddle/utils/CpuId.h b/paddle/utils/CpuId.h index 869be5be541dafd699a87a8e8893aadadf59b711..ed58211d13ac1e0f80d6728950f0b88dc0ae625f 100644 --- a/paddle/utils/CpuId.h +++ b/paddle/utils/CpuId.h @@ -35,7 +35,7 @@ enum simd_t { // clang-format on class SIMDFlags final { -public: + public: DISABLE_COPY(SIMDFlags); SIMDFlags(); @@ -46,7 +46,7 @@ public: return !((simd_flags_ & flags) ^ flags); } -private: + private: int simd_flags_ = SIMD_NONE; }; diff --git a/paddle/utils/CustomStackTrace.h b/paddle/utils/CustomStackTrace.h index 52a6df94979fd3d8d7d540ed0e3898bb3375d975..b60077ea2d946366910780eeb773635972211e04 100644 --- a/paddle/utils/CustomStackTrace.h +++ b/paddle/utils/CustomStackTrace.h @@ -49,7 +49,7 @@ namespace paddle { */ template class CustomStackTrace { -public: + public: /** * @brief Pop out an item from the top of the stack if item == top. * Else, just set status to popping. @@ -136,7 +136,7 @@ public: p.push(item); } -private: + private: /** * Get thread local attribute, and save them into a map (threadId => TYPE*) * @@ -174,7 +174,7 @@ private: return this->getThreadLocal(this->isPushing_, this->pushingBuffers_); } -private: + private: mutable std::mutex mtx_; std::unordered_map*> stackBuffers_; diff --git a/paddle/utils/Error.h b/paddle/utils/Error.h index 7cde98306026ca1de76089749aaea265d151da33..1fc8482e3a1bef869d4df147bbd3cab6e62ccf49 100644 --- a/paddle/utils/Error.h +++ b/paddle/utils/Error.h @@ -95,7 +95,7 @@ namespace paddle { * log(FATAL) and CHECK in Paddle, 'check' method will be removed. */ class Error { -public: + public: /** * Construct a no-error value. */ @@ -138,7 +138,7 @@ public: */ bool isOK() const { return msg_ == nullptr; } -private: + private: std::shared_ptr msg_; }; diff --git a/paddle/utils/GlobalConstants.h b/paddle/utils/GlobalConstants.h index 0ec1c28dfbb2a7db9fa84c9eb2bc4dad806b78e9..3f45e82268435e4c22d1879e909b0c90838d6693 100644 --- a/paddle/utils/GlobalConstants.h +++ b/paddle/utils/GlobalConstants.h @@ -78,7 +78,7 @@ enum ParameterType { using namespace enumeration_wrapper; // NOLINT class TrainAlgorithm { -public: + public: static const std::string SGD; static const std::string AsyncSGD; static const std::string OWLQN; diff --git a/paddle/utils/Locks.h b/paddle/utils/Locks.h index e87abb9139f1c3f250f8b8fe1afdd8883f682647..65f983685f5e178345a6a875a79a6573ce1ccca1 100644 --- a/paddle/utils/Locks.h +++ b/paddle/utils/Locks.h @@ -42,7 +42,7 @@ namespace paddle { * Use unlock() to unlock the lock. */ class RWLock { -public: + public: RWLock() { pthread_rwlock_init(&rwlock_, NULL); } ~RWLock() { pthread_rwlock_destroy(&rwlock_); } RWLock(const RWLock&) = delete; @@ -62,7 +62,7 @@ public: void lock_shared() { pthread_rwlock_rdlock(&rwlock_); } void unlock() { pthread_rwlock_unlock(&rwlock_); } -protected: + protected: pthread_rwlock_t rwlock_; }; @@ -71,7 +71,7 @@ protected: * using RAII management mechanism. */ class ReadLockGuard { -public: + public: /** * @brief Construct Function. Lock on rwlock in read mode. */ @@ -86,7 +86,7 @@ public: */ ~ReadLockGuard() { rwlock_->unlock(); } -protected: + protected: RWLock* rwlock_; }; @@ -98,7 +98,7 @@ protected: */ class SpinLockPrivate; class SpinLock { -public: + public: DISABLE_COPY(SpinLock); SpinLock(); ~SpinLock(); @@ -107,7 +107,7 @@ public: void lock(); void unlock(); -private: + private: SpinLockPrivate* m; }; @@ -116,7 +116,7 @@ private: */ class SemaphorePrivate; class Semaphore { -public: + public: //! Disable copy & assign Semaphore(const Semaphore& other) = delete; Semaphore& operator=(const Semaphore&& other) = delete; @@ -124,7 +124,7 @@ public: //! Enable move. Semaphore(Semaphore&& other) : m(std::move(other.m)) {} -public: + public: /** * @brief Construct Function. * @param[in] initValue the initial value of the @@ -156,7 +156,7 @@ public: */ void post(); -private: + private: SemaphorePrivate* m; }; @@ -166,7 +166,7 @@ private: */ class ThreadBarrierPrivate; class ThreadBarrier { -public: + public: DISABLE_COPY(ThreadBarrier); /** @@ -184,7 +184,7 @@ public: */ void wait(); -private: + private: ThreadBarrierPrivate* m; }; @@ -192,7 +192,7 @@ private: * A wrapper for condition variable with mutex. */ class LockedCondition : public std::condition_variable { -public: + public: /** * @brief execute op and notify one thread which was blocked. * @param[in] op a thread can do something in op before notify. @@ -235,7 +235,7 @@ public: */ std::mutex* mutex() { return &mutex_; } -protected: + protected: std::mutex mutex_; }; diff --git a/paddle/utils/PythonUtil.h b/paddle/utils/PythonUtil.h index daebaffc855518425ae43942c22ec150d2e327f0..6f8d7e09309503e47aca7ae2d20774c748703b21 100644 --- a/paddle/utils/PythonUtil.h +++ b/paddle/utils/PythonUtil.h @@ -55,12 +55,12 @@ std::string callPythonFunc(const std::string& moduleName, * NOTE: the lock of this guard is reentrant or recursive. */ class PyGuard { -public: + public: PyGuard(); PyGuard(const PyGuard& other) = delete; PyGuard& operator=(const PyGuard& other) = delete; -private: + private: std::lock_guard guard_; }; @@ -133,7 +133,7 @@ std::string getPyCallStack(); * Implements getAttr method for object. */ class ObjectHelper { -public: + public: explicit ObjectHelper(const PyObjectPtr& obj) : obj_(obj) {} /** @@ -192,7 +192,7 @@ public: return PyObject_IsTrue(tmp.get()); } -private: + private: const PyObjectPtr& obj_; }; @@ -202,7 +202,7 @@ private: * The python sequence means list or tuple. */ class SequenceHelper { -public: + public: explicit SequenceHelper(const PyObjectPtr& seq) : seq_(seq.get()) { CHECK(PySequence_Check(seq_)); } @@ -248,12 +248,12 @@ public: } } -private: + private: PyObject* seq_; }; class DictHelper { -public: + public: explicit DictHelper(PyObject* d) : dict_(d) {} explicit DictHelper(const PyObjectPtr& d) : dict_(d.get()) {} @@ -275,7 +275,7 @@ public: this->set(key, list); } -private: + private: inline void checkDict() { CHECK(PyDict_Check(this->dict_)); } PyObject* dict_; @@ -289,7 +289,7 @@ inline static bool isCallable(const PyObjectPtr& obj) { * Wrap a callable object. */ class CallableHelper { -public: + public: explicit CallableHelper(const PyObjectPtr& obj) : obj_(obj) { CHECK(py::isCallable(obj_)); } @@ -315,7 +315,7 @@ public: return PyObject_Call(obj_.get(), args.get(), kwargs.get()); } -private: + private: const PyObjectPtr& obj_; PyObjectPtr args; PyObjectPtr kwargs; diff --git a/paddle/utils/Queue.h b/paddle/utils/Queue.h index f054738f87c02d2d749eec8d6c7bb55b506a6d91..189e1a14f7b2d133408a50418d96431164248f0e 100644 --- a/paddle/utils/Queue.h +++ b/paddle/utils/Queue.h @@ -56,7 +56,7 @@ namespace paddle { */ template class Queue { -public: + public: /** * @brief Construct Function. Default capacity of Queue is zero. */ @@ -147,7 +147,7 @@ public: }); } -private: + private: std::deque elements_; int numElements_; std::mutex queueLock_; @@ -185,7 +185,7 @@ private: */ template class BlockingQueue { -public: + public: /** * @brief Construct Function. * @param[in] capacity the max numer of elements the queue can have. @@ -244,7 +244,7 @@ public: return queue_.empty(); } -private: + private: std::mutex mutex_; std::condition_variable notEmpty_; std::condition_variable notFull_; diff --git a/paddle/utils/Stat.h b/paddle/utils/Stat.h index 79fd3b8cf043e62922dfd046754ee8ac261990c5..100e9eba909466fcca57f755405ab63b638a8ebd 100644 --- a/paddle/utils/Stat.h +++ b/paddle/utils/Stat.h @@ -33,7 +33,7 @@ namespace paddle { class Stat; class StatInfo { -public: + public: explicit StatInfo(Stat* stat = nullptr) : stat_(stat) { total_ = 0; max_ = 0; @@ -61,7 +61,7 @@ class Stat; typedef std::shared_ptr StatPtr; class StatSet { -public: + public: explicit StatSet(const std::string& name) : name_(name) {} ~StatSet() {} @@ -102,7 +102,7 @@ public: // pserver code logic, -_- ). void reset(bool clearRawData = true); -private: + private: std::unordered_map statSet_; const std::string name_; RWLock lock_; @@ -112,7 +112,7 @@ extern StatSet globalStat; /*@brief : a simple stat*/ class Stat { -public: + public: explicit Stat(const std::string& statName) : destructStat_(nullptr), name_(statName), openThreadInfo_(false) {} ~Stat() {} @@ -137,7 +137,7 @@ public: friend class StatInfo; -private: + private: void mergeThreadStat(StatInfo& allThreadStat); std::mutex lock_; @@ -164,7 +164,7 @@ inline uint64_t nowInMicroSec() { * A simple help class to measure time interval */ class Timer { -public: + public: explicit Timer(bool autoStart = true) : total_(0), startStamp_(0) { if (autoStart) { start(); @@ -181,13 +181,13 @@ public: void reset() { total_ = 0; } -protected: + protected: uint64_t total_; uint64_t startStamp_; }; class TimerOnce { -public: + public: TimerOnce(Stat* stat, const char* info = "", uint64_t threshold = -1, @@ -208,7 +208,7 @@ public: stat_->addSample(span); } -private: + private: Stat* stat_; const char* info_; Timer timer_; @@ -280,11 +280,11 @@ inline StatSet& registerTimerArg2(uint64_t threshold = -1, #endif // DISABLE_TIMER class GpuProfiler final { -public: + public: GpuProfiler(std::string statName, std::string info); ~GpuProfiler(); -private: + private: std::lock_guard guard_; }; diff --git a/paddle/utils/Thread.h b/paddle/utils/Thread.h index ef36a8c5b2b0e95d759da8a781d781b71d067b7a..2ee6eba1a68202282537788160a77f7689a2ffdb 100644 --- a/paddle/utils/Thread.h +++ b/paddle/utils/Thread.h @@ -29,7 +29,7 @@ namespace paddle { */ class Thread { -public: + public: /** * @brief Construct Function. Default thread pointer is null. */ @@ -62,7 +62,7 @@ public: */ virtual void run() = 0; -protected: + protected: std::unique_ptr thread_; }; @@ -73,7 +73,7 @@ protected: * Use addJob() to add a new job to the job queue. */ class ThreadWorker : protected Thread { -public: + public: typedef std::function JobFunc; /** @@ -116,7 +116,7 @@ public: finishCV_.wait([this] { return empty_; }); } -protected: + protected: /** * @brief Execute jobs in the job queue sequentianlly, * @note If finish all the jobs in the job queue, @@ -150,7 +150,7 @@ protected: * JobFunc can use tid to divide input data. */ class SyncThreadPool { -public: + public: typedef std::function JobFunc; /** @@ -236,7 +236,7 @@ public: } } -protected: + protected: /** * @brief Start all the workers in the pool, call their run() function. */ @@ -285,7 +285,7 @@ protected: } } -protected: + protected: pid_t ownerThreadId_; bool stopping_; ThreadBarrier jobStartBarrier_; @@ -323,7 +323,7 @@ protected: */ template class MultiThreadWorker { -public: + public: typedef T ResultType; typedef std::shared_ptr ResultPtrType; typedef std::function JobFunc; @@ -424,7 +424,7 @@ public: */ bool testResult() { return results_.empty(); } -protected: + protected: /** * @brief Do the jobs in the job queue sequentianlly * and enqueue the result into the result queue. @@ -476,7 +476,7 @@ protected: * thread pool. */ class AsyncThreadPool { -public: + public: typedef std::function JobFunc; AsyncThreadPool() { LOG(FATAL) << "Not implemented"; } @@ -594,7 +594,7 @@ public: } } -protected: + protected: /** * @brief Execute the jobs in the job queue. */ @@ -606,7 +606,7 @@ protected: } } -private: + private: std::vector> workers_; Queue jobs_; bool stopping_; diff --git a/paddle/utils/ThreadLocal.h b/paddle/utils/ThreadLocal.h index 0a27b8b97b83a9066af23039a317c437ea56777a..c5b07506d36875ead65887ea2e221e762be0d621 100644 --- a/paddle/utils/ThreadLocal.h +++ b/paddle/utils/ThreadLocal.h @@ -49,7 +49,7 @@ namespace paddle { */ template class ThreadLocal { -public: + public: ThreadLocal() { CHECK_EQ(pthread_key_create(&threadSpecificKey_, dataDestructor), 0); } @@ -92,7 +92,7 @@ public: */ operator T*() { return get(); } -private: + private: static void dataDestructor(void* p) { delete (T*)p; } pthread_key_t threadSpecificKey_; @@ -111,7 +111,7 @@ private: */ template class ThreadLocalD { -public: + public: ThreadLocalD() { CHECK_EQ(pthread_key_create(&threadSpecificKey_, NULL), 0); } ~ThreadLocalD() { pthread_key_delete(threadSpecificKey_); @@ -150,7 +150,7 @@ public: */ T& operator*() { return *get(); } -private: + private: static void dataDestructor(void* p) { delete (T*)p; } void updateMap(T* p) { @@ -172,7 +172,7 @@ private: * @brief Thread-safe C-style random API. */ class ThreadLocalRand { -public: + public: /** * initSeed just like srand, * called by main thread, @@ -205,7 +205,7 @@ public: */ static int getDefaultSeed() { return defaultSeed_; } -protected: + protected: static unsigned int defaultSeed_; static ThreadLocal seed_; }; @@ -214,7 +214,7 @@ protected: * @brief Thread-safe C++ style random engine. */ class ThreadLocalRandomEngine { -public: + public: /** * get random_engine for each thread. * @@ -222,7 +222,7 @@ public: */ static std::default_random_engine& get(); -protected: + protected: static ThreadLocal engine_; }; diff --git a/paddle/utils/Util.h b/paddle/utils/Util.h index 9579881ea3b92abab0189631184bab515afb67a3..e6f05e30d308b8b94935897e947350934a5971ee 100644 --- a/paddle/utils/Util.h +++ b/paddle/utils/Util.h @@ -179,7 +179,7 @@ void loadFileList(const std::string& fileListFileName, */ void registerInitFunction(std::function func, int priority = 0); class InitFunction { -public: + public: explicit InitFunction(std::function func, int priority = 0) { registerInitFunction(func, priority); } @@ -191,7 +191,7 @@ public: * When the SetDevice object is destructed, it will restore device environment. */ class SetDevice { -public: + public: explicit SetDevice(int deviceId) { isSet_ = deviceId >= 0; devId_ = 0; @@ -206,7 +206,7 @@ public: } } -protected: + protected: bool isSet_; int devId_; }; @@ -240,7 +240,7 @@ inline void enablePeerAccess(int d1, int d2) { * } */ class AsyncGpuBlock { -public: + public: AsyncGpuBlock() : syncFlag_(hl_get_sync_flag()) { hl_set_sync_flag(false); } ~AsyncGpuBlock() { if (syncFlag_) { @@ -249,7 +249,7 @@ public: } } -private: + private: bool syncFlag_; }; @@ -378,7 +378,7 @@ std::string join(const std::string& part1, * A Checker for each invoke of method in same thread. */ class SameThreadChecker { -public: + public: SameThreadChecker() {} /** @@ -400,7 +400,7 @@ public: << invokeThreadId_ << " current invoked in " << curThreadId; } -private: + private: std::once_flag onceFlag_; std::thread::id invokeThreadId_; }; @@ -421,7 +421,7 @@ private: */ template class WeakKVCache { -public: + public: WeakKVCache() {} std::shared_ptr get(const KType& key, @@ -442,7 +442,7 @@ public: return retVal; } -private: + private: std::mutex lock_; std::unordered_map, Hash> storage_; }; @@ -453,7 +453,7 @@ private: */ template class ScopedCallbacks { -public: + public: ScopedCallbacks(CallbackType enter, CallbackType exit, Args&... args) : exit_(std::bind(exit, args...)) { enter(args...); @@ -464,7 +464,7 @@ public: ~ScopedCallbacks() { exit_(); } -private: + private: std::function exit_; }; @@ -475,7 +475,7 @@ private: */ template class AlignedAllocator { -public: + public: /// std campatible typedefs. typedef T* pointer; typedef const T* const_pointer; @@ -552,12 +552,12 @@ public: return this->allocate(n); } -private: + private: AlignedAllocator& operator=(const AlignedAllocator&); // disable }; class Deprecated { -public: + public: explicit Deprecated(const std::string& msg = "") { if (msg.empty()) { LOG(WARNING) << "This class is deprecated, please do not use this class."; diff --git a/paddle/utils/arch/linux/Locks.cpp b/paddle/utils/arch/linux/Locks.cpp index a4e6c8f7b8397adc262588612c250bac5ef5eaa6..409af8bce3621c51bfd7a69c6b4ec1f9cc6be8e4 100644 --- a/paddle/utils/arch/linux/Locks.cpp +++ b/paddle/utils/arch/linux/Locks.cpp @@ -19,7 +19,7 @@ limitations under the License. */ namespace paddle { class SemaphorePrivate { -public: + public: sem_t sem; }; @@ -45,7 +45,7 @@ void Semaphore::post() { sem_post(&m->sem); } #ifdef PADDLE_USE_PTHREAD_SPINLOCK class SpinLockPrivate { -public: + public: inline SpinLockPrivate() { pthread_spin_init(&lock_, 0); } inline ~SpinLockPrivate() { pthread_spin_destroy(&lock_); } @@ -63,7 +63,7 @@ public: // clang-format on class SpinLockPrivate { -public: + public: inline void lock() { while (lock_.test_and_set(std::memory_order_acquire)) { } @@ -86,7 +86,7 @@ void SpinLock::unlock() { m->unlock(); } #ifdef PADDLE_USE_PTHREAD_BARRIER class ThreadBarrierPrivate { -public: + public: pthread_barrier_t barrier_; inline explicit ThreadBarrierPrivate(int count) { @@ -101,7 +101,7 @@ public: #else class ThreadBarrierPrivate { -public: + public: pthread_mutex_t mutex_; pthread_cond_t cond_; int count_; diff --git a/paddle/utils/arch/osx/Locks.cpp b/paddle/utils/arch/osx/Locks.cpp index e03992363fd6051a1970664d63406b2e7a47fce3..f3905091bd024ab02c3f5d39cfed6dbc38fabbbc 100644 --- a/paddle/utils/arch/osx/Locks.cpp +++ b/paddle/utils/arch/osx/Locks.cpp @@ -21,7 +21,7 @@ limitations under the License. */ namespace paddle { class SemaphorePrivate { -public: + public: ~SemaphorePrivate() { dispatch_release(sem); } dispatch_semaphore_t sem; @@ -45,7 +45,7 @@ void Semaphore::wait() { void Semaphore::post() { dispatch_semaphore_signal(m->sem); } class SpinLockPrivate { -public: + public: std::atomic_flag lock_ = ATOMIC_FLAG_INIT; char padding_[64 - sizeof(lock_)]; // Padding to cache line size }; @@ -61,7 +61,7 @@ void SpinLock::lock() { void SpinLock::unlock() { m->lock_.clear(std::memory_order_release); } class ThreadBarrierPrivate { -public: + public: pthread_mutex_t mutex_; pthread_cond_t cond_; int count_; diff --git a/paddle/utils/tests/test_CustomStackTrace.cpp b/paddle/utils/tests/test_CustomStackTrace.cpp index c320074fbadab3e211ed72ce715d595c90673d6d..4d5540b24cb9d52482cfa5a77dfa956b8bf4ef38 100644 --- a/paddle/utils/tests/test_CustomStackTrace.cpp +++ b/paddle/utils/tests/test_CustomStackTrace.cpp @@ -12,10 +12,8 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include - -#include -#include +#include // NOLINT +#include // NOLINT #include "paddle/utils/CustomStackTrace.h" #include "paddle/utils/Locks.h" @@ -39,14 +37,10 @@ void testNormalImpl( threads.reserve(FLAGS_test_thread_num); for (int32_t i = 0; i < FLAGS_test_thread_num; ++i) { - threads.emplace_back(new std::thread([&tracer, - &countDown, - &layerSize, - &startBarrier, - &doneBarrier, - &callback] { - callback(tracer, countDown, layerSize, startBarrier, doneBarrier); - })); + threads.emplace_back( + new std::thread([&tracer, &startBarrier, &doneBarrier, &callback] { + callback(tracer, countDown, layerSize, startBarrier, doneBarrier); + })); } size_t cntDown = countDown; while (cntDown-- > 0) { diff --git a/python/paddle/dataset/wmt16.py b/python/paddle/dataset/wmt16.py index ad23338a96df6856c7e15cb5e3bb713021a55bf0..540d43b692e0f65460f558dd74a52715ff4db68d 100644 --- a/python/paddle/dataset/wmt16.py +++ b/python/paddle/dataset/wmt16.py @@ -96,7 +96,7 @@ def __get_dict_size(src_dict_size, trg_dict_size, src_lang): src_dict_size = min(src_dict_size, (TOTAL_EN_WORDS if src_lang == "en" else TOTAL_DE_WORDS)) trg_dict_size = min(trg_dict_size, (TOTAL_DE_WORDS if src_lang == "en" else - TOTAL_ENG_WORDS)) + TOTAL_EN_WORDS)) return src_dict_size, trg_dict_size diff --git a/python/paddle/fluid/__init__.py b/python/paddle/fluid/__init__.py index e2502990d5b78eb0db7bdfd0c8ef9fb6688016df..859605d005328c030980a49a349742772de1cb6d 100644 --- a/python/paddle/fluid/__init__.py +++ b/python/paddle/fluid/__init__.py @@ -20,6 +20,16 @@ from framework import * import executor from executor import * +import trainer +from trainer import Trainer +from trainer import BeginEpochEvent +from trainer import EndEpochEvent +from trainer import BeginStepEvent +from trainer import EndStepEvent + +import inferencer +from inferencer import Inferencer + import io import evaluator import initializer @@ -30,51 +40,49 @@ import backward import regularizer import average import metrics +import transpiler from param_attr import ParamAttr, WeightNormParamAttr from data_feeder import DataFeeder from core import LoDTensor, CPUPlace, CUDAPlace, CUDAPinnedPlace -from distribute_transpiler import DistributeTranspiler -from distribute_transpiler_simple import SimpleDistributeTranspiler +from transpiler import DistributeTranspiler, SimpleDistributeTranspiler, \ + InferenceTranspiler, memory_optimize, release_memory from concurrency import (Go, make_channel, channel_send, channel_recv, channel_close, Select) -from inference_transpiler import InferenceTranspiler +from lod_tensor import create_lod_tensor, create_random_int_lodtensor import clip -from memory_optimization_transpiler import memory_optimize, release_memory import profiler import unique_name import recordio_writer -from parallel_executor import ParallelExecutor +import parallel_executor +from parallel_executor import * Tensor = LoDTensor -__all__ = framework.__all__ + executor.__all__ + concurrency.__all__ + [ - 'io', - 'initializer', - 'layers', - 'nets', - 'optimizer', - 'learning_rate_decay', - 'backward', - 'regularizer', - 'LoDTensor', - 'CPUPlace', - 'CUDAPlace', - 'CUDAPinnedPlace', - 'Tensor', - 'ParamAttr', - 'WeightNormParamAttr', - 'DataFeeder', - 'clip', - 'SimpleDistributeTranspiler', - 'DistributeTranspiler', - 'InferenceTranspiler', - 'memory_optimize', - 'release_memory', - 'profiler', - 'unique_name', - 'recordio_writer', - 'ParallelExecutor', -] +__all__ = framework.__all__ + executor.__all__ + concurrency.__all__ + \ + trainer.__all__ + inferencer.__all__ + transpiler.__all__ + \ + parallel_executor.__all__ + lod_tensor.__all__ + [ + 'io', + 'initializer', + 'layers', + 'transpiler' + 'nets', + 'optimizer', + 'learning_rate_decay', + 'backward', + 'regularizer', + 'LoDTensor', + 'CPUPlace', + 'CUDAPlace', + 'CUDAPinnedPlace', + 'Tensor', + 'ParamAttr', + 'WeightNormParamAttr', + 'DataFeeder', + 'clip', + 'profiler', + 'unique_name', + 'recordio_writer', + ] def __bootstrap__(): @@ -111,7 +119,9 @@ def __bootstrap__(): 'eager_delete_scope' ] if core.is_compiled_with_cuda(): - read_env_flags += ['fraction_of_gpu_memory_to_use'] + read_env_flags += [ + 'fraction_of_gpu_memory_to_use', 'cudnn_algo_use_autotune' + ] core.init_gflags([sys.argv[0]] + ["--tryfromenv=" + ",".join(read_env_flags)]) core.init_glog(sys.argv[0]) diff --git a/python/paddle/fluid/backward.py b/python/paddle/fluid/backward.py index 7af6ed1463ab737e871da487f2a687301652ef2d..4f9622d04dc98f41b503ceb780802d2a4e4c58a0 100644 --- a/python/paddle/fluid/backward.py +++ b/python/paddle/fluid/backward.py @@ -51,6 +51,12 @@ def _create_op_desc_(op_type, inputs, outputs, attrs): op_desc.set_input(para, args) for para, args in outputs.iteritems(): op_desc.set_output(para, args) + + op_role_attr_name = core.op_proto_and_checker_maker.kOpRoleAttrName() + + if op_role_attr_name not in attrs: + attrs[ + op_role_attr_name] = core.op_proto_and_checker_maker.OpRole.Backward for name, val in attrs.iteritems(): if isinstance(val, framework.Block): op_desc.set_block_attr(name, val.desc) @@ -141,7 +147,7 @@ def _addup_repetitive_outputs_(op_descs): else: if len(renamed_vars[var_name]) == 1: new_name = var_name + "@RENAME@" + \ - str(var_rename_count[var_name]) + str(var_rename_count[var_name]) var_rename_count[var_name] += 1 # rename original var_name renamed_vars[var_name][0] = new_name @@ -149,7 +155,7 @@ def _addup_repetitive_outputs_(op_descs): _rename_arg_(pending_sum_ops, var_name, new_name) new_name = var_name + "@RENAME@" + \ - str(var_rename_count[var_name]) + str(var_rename_count[var_name]) var_rename_count[var_name] += 1 op_desc.rename_output(var_name, new_name) renamed_vars[var_name].append(new_name) @@ -335,9 +341,12 @@ def _append_backward_ops_(block, no_grad_dict[block.idx]) # append op_desc in grad_op_descs to target_block + op_role_attr_name = core.op_proto_and_checker_maker.kOpRoleAttrName() + backward = core.op_proto_and_checker_maker.OpRole.Backward for op_desc in grad_op_descs: new_op_desc = target_block.desc.append_op() new_op_desc.copy_from(op_desc) + new_op_desc.set_attr(op_role_attr_name, backward) grad_to_var["__current_op_desc__"] = new_op_desc if callbacks is not None: assert (isinstance(callbacks, list)) @@ -439,6 +448,22 @@ def append_backward(loss, parameter_list=None, no_grad_set=None, (list[(Variable,Variable)]): list of (parameter, gradient) pair. """ assert isinstance(loss, framework.Variable) + + if loss.op is None: + # the loss is from a cloned program. Find loss op manually. + for op in reversed(loss.block.ops): + assert isinstance(op, framework.Operator) + if len(op.output_arg_names) == 1 and op.output_arg_names[ + 0] == loss.name: + loss.op = op + break + if loss.op is None: + raise ValueError("loss.op is None. Should not happend") + + loss.op.set_attr(core.op_proto_and_checker_maker.kOpRoleAttrName(), + int(core.op_proto_and_checker_maker.OpRole.Forward) | + int(core.op_proto_and_checker_maker.OpRole.Loss)) + if callbacks is not None: isinstance(callbacks, list) @@ -456,12 +481,16 @@ def append_backward(loss, parameter_list=None, no_grad_set=None, current_block_idx = program.current_block_idx grad_to_var = dict() - op_desc = _create_op_desc_("fill_constant", {}, { - "Out": [_append_grad_suffix_(loss.name)] - }, {"shape": [1], - "value": 1.0, - "dtype": loss.dtype, - "force_cpu": False}) + op_desc = _create_op_desc_( + "fill_constant", {}, {"Out": [_append_grad_suffix_(loss.name)]}, { + "shape": [1], + "value": 1.0, + "dtype": loss.dtype, + "force_cpu": False, + core.op_proto_and_checker_maker.kOpRoleAttrName(): + int(core.op_proto_and_checker_maker.OpRole.Backward) | + int(core.op_proto_and_checker_maker.OpRole.Loss), + }) root_block.desc.append_op().copy_from(op_desc) block_no_grad_set = set(map(_strip_grad_suffix_, no_grad_dict[0])) @@ -480,6 +509,8 @@ def append_backward(loss, parameter_list=None, no_grad_set=None, program.current_block_idx = current_block_idx program.sync_with_cpp() + # FIXME(zcd): prevent loss.grad optimized by mem_opt. + loss.block.var(_append_grad_suffix_(loss.name)).persistable = True if parameter_list is not None: parameters = parameter_list @@ -503,6 +534,24 @@ def append_backward(loss, parameter_list=None, no_grad_set=None, params_and_grads.append((param_var, grad_var)) else: params_and_grads.append((param_var, None)) + + op_role_var_attr_name = core.op_proto_and_checker_maker.kOpRoleVarAttrName() + for p, g in params_and_grads: + if g is None: + continue + for op in reversed(program.global_block().ops): + assert isinstance(op, framework.Operator) + if g.name in op.output_arg_names: + g.op = op + break + + if g.op is None: + raise ValueError("Unexpected branch") + attr_val = [p.name, g.name] + if g.op.has_attr(op_role_var_attr_name): + attr_val.extend(g.op.attr(op_role_var_attr_name)) + g.op.set_attr(op_role_var_attr_name, attr_val) + return params_and_grads diff --git a/python/paddle/fluid/clip.py b/python/paddle/fluid/clip.py index 12add9e686910c3936cf17fe87a5d0b78443b270..66c3fc6b66d61bc9578f84594409ad0f24c99910 100644 --- a/python/paddle/fluid/clip.py +++ b/python/paddle/fluid/clip.py @@ -214,21 +214,24 @@ def set_gradient_clip(clip, param_list=None, program=None): def append_gradient_clip_ops(param_grad): context = dict() - create_op_callbacks = [] for p, g in param_grad: - clip_attr = getattr(p, 'gradient_clip_attr', NullGradientClipAttr()) - if clip_attr is None: - clip_attr = NullGradientClipAttr() - if not isinstance(clip_attr, BaseGradientClipAttr): - raise TypeError( - "clip attribute should be an instance of BaseGradientClipAttr") + with p.block.program.optimized_guard(p): + clip_attr = getattr(p, 'gradient_clip_attr', NullGradientClipAttr()) + if clip_attr is None: + clip_attr = NullGradientClipAttr() + if not isinstance(clip_attr, BaseGradientClipAttr): + raise TypeError( + "clip attribute should be an instance of BaseGradientClipAttr" + ) - clip_attr.process_context(context=context, param=p, grad=g) - create_op_callbacks.append( - functools.partial( - clip_attr.create_operators, param=p, grad=g)) + clip_attr.process_context(context=context, param=p, grad=g) + + res = [] + for p, g in param_grad: + with p.block.program.optimized_guard(p): + res.append(clip_attr.create_operators(param=p, grad=g)) - return [each_callback() for each_callback in create_op_callbacks] + return res ClipByValue = GradientClipByValue diff --git a/python/paddle/fluid/data_feeder.py b/python/paddle/fluid/data_feeder.py index ac02401c79b787716b2e5f43e0d1c5686cf2bd13..7940dabcfb03cc9eb46f678365685a6e99bcceec 100644 --- a/python/paddle/fluid/data_feeder.py +++ b/python/paddle/fluid/data_feeder.py @@ -16,6 +16,7 @@ from __future__ import print_function import core import numpy import six.moves as six +import multiprocessing from framework import Variable, default_main_program @@ -35,9 +36,11 @@ class DataToLoDTensorConverter(object): self.dtype = 'float64' elif dtype == core.VarDesc.VarType.INT32: self.dtype = 'int32' + elif dtype == core.VarDesc.VarType.UINT8: + self.dtype = 'uint8' else: raise ValueError("dtype must be any of [int32, float32, int64, " - "float64]") + "float64, uint8]") self.data = [] self.lod = [] @@ -53,9 +56,9 @@ class DataToLoDTensorConverter(object): self.data.append(data) else: cur_lod_len = len(data) - lod[-1].append(lod[-1][-1] + cur_lod_len) + lod[0].append(lod[0][-1] + cur_lod_len) for each_data in data: - self._feed_impl_(each_data, lod[:-1], lod_level - 1) + self._feed_impl_(each_data, lod[1:], lod_level - 1) def done(self): arr = numpy.array(self.data, dtype=self.dtype).reshape(self.shape) @@ -116,3 +119,60 @@ class DataFeeder(object): for each_name, each_converter in six.zip(self.feed_names, converter): ret_dict[each_name] = each_converter.done() return ret_dict + + def feed_parallel(self, iterable, num_places=None): + if isinstance(self.place, core.CUDAPlace): + places = [ + core.CUDAPlace(i) + for i in six.xrange(self._get_number_of_places_(num_places)) + ] + else: + places = [ + core.CPUPlace() + for _ in six.xrange(self._get_number_of_places_(num_places)) + ] + + if len(iterable) != len(places): + raise ValueError("feed_parallel takes multiple mini-batches. Each " + "mini-batch will be feed on each device. The " + "number of devices and number of mini-batches " + "must be same.") + + place = self.place + for p, batch in six.zip(places, iterable): + self.place = p + yield self.feed(batch) + self.place = place + + def _get_number_of_places_(self, num_places): + if num_places is not None: + return int(num_places) + elif isinstance(self.place, core.CUDAPlace): + return core.get_cuda_device_count() + else: + return multiprocessing.cpu_count() + + def decorate_reader(self, + reader, + multi_devices, + num_places=None, + drop_last=True): + def __reader_creator__(): + if not multi_devices: + for item in reader(): + yield self.feed(item) + else: + num = self._get_number_of_places_(num_places) + item = [] + for batch in reader(): + item.append(batch) + if len(item) == num: + yield list(self.feed_parallel(item, num)) + item = [] + if not drop_last and len(item) != 0: + raise ValueError( + "The data batch which cannot fit for devices will be " + "dropped is not implementation. Other strategies are " + "not implemented") + + return __reader_creator__ diff --git a/python/paddle/fluid/debuger.py b/python/paddle/fluid/debugger.py similarity index 100% rename from python/paddle/fluid/debuger.py rename to python/paddle/fluid/debugger.py diff --git a/python/paddle/fluid/distributed_splitter.py b/python/paddle/fluid/distributed_splitter.py deleted file mode 100644 index 060c1df8ad2badc5132f45ff0f44d136d828faa1..0000000000000000000000000000000000000000 --- a/python/paddle/fluid/distributed_splitter.py +++ /dev/null @@ -1,57 +0,0 @@ -# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - - -def hash_name(varlist, pserver_endpoints): - """ - hash variable names to several endpoints. - - Args: - varlist(list): a list of Variables - - Returns(dict): a map of pserver endpoint -> varname - """ - - def _hash_block(block_str, total): - return hash(block_str) % total - - eplist = [] - for var in varlist: - server_id = _hash_block(var.name(), len(pserver_endpoints)) - server_for_param = pserver_endpoints[server_id] - eplist.append(server_for_param) - return eplist - - -def round_robin(varlist, pserver_endpoints): - """ - Distribute variables to several endpoints. - Args: - varlist(list): a list of variables - pserver_endpoints(list): a list of pserver endpoints - - Returns(list[int]): the endpoint for each variable - """ - assert (len(varlist) >= len(pserver_endpoints)) - - eplist = [] - pserver_idx = 0 - for var in varlist: - server_for_param = pserver_endpoints[pserver_idx] - eplist.append(server_for_param) - - pserver_idx += 1 - if pserver_idx >= len(pserver_endpoints): - pserver_idx = 0 - return eplist diff --git a/python/paddle/fluid/evaluator.py b/python/paddle/fluid/evaluator.py index 1ee1d3727174c079d2c217dede27ff1a0316c01c..7c6ad6f27dcfd7040f79c72c01413c8cc84a28ba 100644 --- a/python/paddle/fluid/evaluator.py +++ b/python/paddle/fluid/evaluator.py @@ -273,10 +273,11 @@ class DetectionMAP(Evaluator): [M, 6]. The layout is [label, confidence, xmin, ymin, xmax, ymax]. gt_label (Variable): The ground truth label index, which is a LoDTensor with shape [N, 1]. - gt_difficult (Variable): Whether this ground truth is a difficult - bounding box (bbox), which is a LoDTensor [N, 1]. gt_box (Variable): The ground truth bounding box (bbox), which is a LoDTensor with shape [N, 6]. The layout is [xmin, ymin, xmax, ymax]. + gt_difficult (Variable|None): Whether this ground truth is a difficult + bounding bbox, which can be a LoDTensor [N, 1] or not set. If None, + it means all the ground truth labels are not difficult bbox. class_num (int): The class number. background_label (int): The index of background label, the background label will be ignored. If set to -1, then all categories will be @@ -284,7 +285,8 @@ class DetectionMAP(Evaluator): overlap_threshold (float): The threshold for deciding true/false positive, 0.5 by defalut. evaluate_difficult (bool): Whether to consider difficult ground truth - for evaluation, True by defalut. + for evaluation, True by defalut. This argument does not work when + gt_difficult is None. ap_version (string): The average precision calculation ways, it must be 'integral' or '11point'. Please check https://sanchom.wordpress.com/tag/average-precision/ for details. @@ -295,7 +297,7 @@ class DetectionMAP(Evaluator): exe = fluid.executor(place) map_evaluator = fluid.Evaluator.DetectionMAP(input, - gt_label, gt_difficult, gt_box) + gt_label, gt_box, gt_difficult) cur_map, accum_map = map_evaluator.get_map_var() fetch = [cost, cur_map, accum_map] for epoch in PASS_NUM: @@ -313,8 +315,8 @@ class DetectionMAP(Evaluator): input, gt_label, gt_box, - gt_difficult, - class_num, + gt_difficult=None, + class_num=None, background_label=0, overlap_threshold=0.5, evaluate_difficult=True, @@ -322,8 +324,11 @@ class DetectionMAP(Evaluator): super(DetectionMAP, self).__init__("map_eval") gt_label = layers.cast(x=gt_label, dtype=gt_box.dtype) - gt_difficult = layers.cast(x=gt_difficult, dtype=gt_box.dtype) - label = layers.concat([gt_label, gt_difficult, gt_box], axis=1) + if gt_difficult: + gt_difficult = layers.cast(x=gt_difficult, dtype=gt_box.dtype) + label = layers.concat([gt_label, gt_difficult, gt_box], axis=1) + else: + label = layers.concat([gt_label, gt_box], axis=1) # calculate mean average precision (mAP) of current mini-batch map = layers.detection_map( diff --git a/python/paddle/fluid/executor.py b/python/paddle/fluid/executor.py index 7ad028714d3b47d93328dbf7c3297d55a2db1bd0..93aa5f908ec929a33089a62caa2186ba9e57fffe 100644 --- a/python/paddle/fluid/executor.py +++ b/python/paddle/fluid/executor.py @@ -299,14 +299,18 @@ class Executor(object): if feed is None: feed = {} if not isinstance(feed, dict): - raise TypeError("feed should be a map") + raise TypeError( + "feed requires dict as its Parameter. But you passed in %s" % + (type(feed))) if fetch_list is None: fetch_list = [] if program is None: program = default_main_program() if not isinstance(program, Program): - raise TypeError() + raise TypeError( + "Executor requires Program as its Parameter. But you passed in %s" + % (type(program))) if scope is None: scope = global_scope() diff --git a/python/paddle/fluid/framework.py b/python/paddle/fluid/framework.py index 340882ea9e7b0e2a0c52749c771308c6b860ed07..33b5caa0eab0ec192eb4a3b63cf82a672c58d2cb 100644 --- a/python/paddle/fluid/framework.py +++ b/python/paddle/fluid/framework.py @@ -72,6 +72,8 @@ def convert_np_dtype_to_dtype_(np_dtype): return core.VarDesc.VarType.INT64 elif dtype == np.bool: return core.VarDesc.VarType.BOOL + elif dtype == np.uint8: + return core.VarDesc.VarType.UINT8 else: raise ValueError("Not supported numpy dtype " + str(dtype)) @@ -160,6 +162,7 @@ class Variable(object): persistable=None, error_clip=None, stop_gradient=False, + is_data=False, **kwargs): self.block = block self.error_clip = error_clip @@ -238,6 +241,7 @@ class Variable(object): self.block.vars[name] = self self.op = None self.stop_gradient = stop_gradient + self.is_data = is_data def __str__(self): return self.to_string(True) @@ -400,6 +404,23 @@ class Operator(object): self.block = block self.desc = desc self.attrs = attrs + if self.attrs is None: + self.attrs = dict() + del attrs + + op_maker = core.op_proto_and_checker_maker + + if op_maker.kOpRoleAttrName() not in self.attrs: + self.attrs[op_maker.kOpRoleAttrName()] = self.block.program.op_role + + role_var_name = op_maker.kOpRoleVarAttrName() + if len(self.block.program. + op_role_var) != 0 and role_var_name not in self.attrs: + self.attrs[role_var_name] = self.block.program.op_role_var + + if role_var_name in self.attrs and len(self.attrs[role_var_name]) == 0: + del self.attrs[role_var_name] + if len(self.desc.type()) != 0: return if type is None: @@ -465,29 +486,30 @@ class Operator(object): arg.op = self self.desc.set_output(out_proto.name, out_arg_names) - if attrs is not None: - if not isinstance(attrs, dict): + if self.attrs is not None: + if not isinstance(self.attrs, dict): raise TypeError("'attrs' should be a dict.") for attr in proto.attrs: attr_name = attr.name - if (attr_name not in attrs) or (attrs[attr_name] is None): + if (attr_name not in self.attrs) or ( + self.attrs[attr_name] is None): continue - if isinstance(attrs[attr_name], Block): - self.desc.set_block_attr(attr_name, attrs[attr_name].desc) - elif isinstance(attrs[attr_name], core.BlockDesc) or \ - isinstance(attrs[attr_name], core.ProgramDesc): + if isinstance(self.attrs[attr_name], Block): + self.desc.set_block_attr(attr_name, + self.attrs[attr_name].desc) + elif isinstance(self.attrs[attr_name], core.BlockDesc) or \ + isinstance(self.attrs[attr_name], core.ProgramDesc): self.desc.set_serialized_attr( - attr_name, attrs[attr_name].serialize_to_string()) + attr_name, self.attrs[attr_name].serialize_to_string()) else: - self.desc.set_attr(attr_name, attrs[attr_name]) - + self.desc.set_attr(attr_name, self.attrs[attr_name]) self.desc.check_attrs() no_kernel_op_set = { 'feed', 'fetch', 'save', 'load', 'recurrent', 'go', 'rnn_memory_helper_grad', 'conditional_block', 'while', 'send', 'recv', 'listen_and_serv', 'parallel_do', 'save_combine', 'load_combine', 'ncclInit', 'channel_create', 'channel_close', - 'channel_send', 'channel_recv', 'select' + 'channel_send', 'channel_recv', 'select', 'gen_nccl_id' } if type not in no_kernel_op_set: self.desc.infer_var_type(self.block.desc) @@ -608,6 +630,10 @@ class Operator(object): """ return self.desc.attr_type(name) + def set_attr(self, name, val): + self.attrs[name] = val + self.desc.set_attr(name, val) + @property def attr_names(self): """ @@ -658,10 +684,10 @@ class Operator(object): class Block(object): def __init__(self, program, idx): self.desc = program.desc.block(idx) - self.vars = dict() # var_name --> var + self.vars = collections.OrderedDict() # var_name --> var self.ops = list() # operator list self.program = program - self.removed_vars = dict() + self.removed_vars = collections.OrderedDict() def __str__(self): return self.to_string(True) @@ -771,7 +797,7 @@ class Block(object): Rename variable in vars and ops' inputs and outputs """ if not self.has_var(name): - raise ValueError("var %s is not in current" % name) + raise ValueError("var %s is not in current block" % name) v = self.var(name) if type(v) == Parameter: var_type = "Parameter" @@ -817,6 +843,7 @@ class Block(object): self.vars[new_name] = var del self.vars[name] self.sync_with_cpp() + return var def remove_var(self, name): self.sync_with_cpp() @@ -848,17 +875,6 @@ class Block(object): self.desc.remove_op(index, index + 1) del self.ops[index] - def delete_ops(self, ops): - # remove from cpp - # FIXME(typhoonzero): remove only the first occurrence. - try: - start = list(self.ops).index(ops[0]) - end = list(self.ops).index(ops[-1]) - except Exception, e: - raise e - - self.desc.remove_op(start, end + 1) - def slice_ops(self, start, end): return self.ops[start:end] @@ -989,7 +1005,8 @@ class Block(object): shape=var.shape, dtype=var.dtype, type=var.type, - persistable=True) + persistable=True, + is_data=var.is_data) else: ret_var = self.create_var( name=var.name, @@ -997,7 +1014,8 @@ class Block(object): dtype=var.dtype, type=var.type, lod_level=var.lod_level, - persistable=True) + persistable=True, + is_data=var.is_data) return ret_var @@ -1007,6 +1025,33 @@ class Program(object): self.blocks = [Block(self, 0)] self.current_block_idx = 0 self._seed = 0 + self._current_role = core.op_proto_and_checker_maker.OpRole.Forward + self._op_role_var = [] + + @property + def op_role(self): + return self._current_role + + @op_role.setter + def set_op_role(self, role): + self._current_role = role + + @property + def op_role_var(self): + return self._op_role_var + + @op_role_var.setter + def set_op_role_var(self, var_name): + self._op_role_var = [var_name] + + @contextlib.contextmanager + def optimized_guard(self, var): + OpRole = core.op_proto_and_checker_maker.OpRole + self._current_role = OpRole.Optimize + self._op_role_var = [var.name if isinstance(var, Variable) else var] + yield + self._op_role_var = [] + self._current_role = OpRole.Forward def __str__(self): return self.to_string(True) @@ -1053,14 +1098,16 @@ class Program(object): Returns(Program): The cloned Program object. """ - p = Program() if for_test: - p.desc = core.inference_optimize(self.desc) + p = self.inference_optimize() else: + p = Program() p.desc = core.ProgramDesc(self.desc) - p.blocks = [Block(p, i) for i in xrange(self.desc.num_blocks())] - p.sync_with_cpp() + p.blocks = [Block(p, i) for i in xrange(self.desc.num_blocks())] + p.sync_with_cpp() + p.copy_param_info_from(self) + p.copy_data_info_from(self) return p def prune(self, targets): @@ -1070,16 +1117,25 @@ class Program(object): for t in targets: if not isinstance(t, Operator): if isinstance(t, Variable): - if t.op is None: - global_block = self.global_block() - for op in global_block.ops: - if t.name in op.output_arg_names: - t.op = op - break + # After transpiler processing, the op that output this + # variable maybe has been changed, so t.op is not reliable + # and we need to find the current op that generate this + # variable here. + t.op = None + global_block = self.global_block() + for idx, op in enumerate(global_block.ops): + if t.name in op.output_arg_names: + t.op = op + break + t = t.op + if t is None: + raise ValueError( + "The target variable must have an " + "associated operator that generates it.") else: - raise ValueError(("All targets of prune() can only be " - "Variable or Operator.")) + raise ValueError("All targets of prune() can only be " + "Variable or Operator.") targets_idx.append([t.block.idx, t.idx]) res = Program() @@ -1089,8 +1145,16 @@ class Program(object): return res def inference_optimize(self): + # this is an alternative implement before + # core.inference_optimize being fixed. res = Program() - res.desc = core.inference_optimize(self.desc) + res.desc = core.ProgramDesc(self.desc) + for i in xrange(res.desc.num_blocks()): + block = res.desc.block(i) + for j in xrange(block.op_size()): + op = block.op(j) + if op.has_attr('is_test'): + op.set_attr('is_test', True) res.blocks = [Block(res, i) for i in xrange(res.desc.num_blocks())] res.sync_with_cpp() return res @@ -1165,6 +1229,26 @@ class Program(object): "program, with represent the same topology") self.global_block().copy_param_info_from(other.global_block()) + def copy_data_info_from(self, other): + """ + Copy the information of data variables from other program. + Args: + other(Program): Other program + + Returns: + None + """ + if not isinstance(other, Program): + raise TypeError("copy_param_info_from should be invoked with " + "Program") + + if len(self.blocks) != len(other.blocks): + raise ValueError("copy_param_info_from should be invoked with two " + "program, with represent the same topology") + for var in other.global_block().vars.itervalues(): + if var.is_data: + self.global_block().var(var.name).is_data = True + def list_vars(self): for each_block in self.blocks: for each_var in each_block.vars.itervalues(): diff --git a/python/paddle/fluid/inferencer.py b/python/paddle/fluid/inferencer.py new file mode 100644 index 0000000000000000000000000000000000000000..9f242cf29a56573349f192307a68e135a409a4be --- /dev/null +++ b/python/paddle/fluid/inferencer.py @@ -0,0 +1,81 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import contextlib + +import core + +import executor +import framework +import io +import parallel_executor +import unique_name +from trainer import check_and_get_place + +__all__ = ['Inferencer', ] + + +class Inferencer(object): + def __init__(self, infer_func, param_path, place=None, parallel=False): + """ + :param infer_func: a function that will return predict Variable + :param param_path: the path where the inference model is saved by fluid.io.save_params + :param place: place to do the inference + :param parallel: use parallel_executor to run the inference, it will use multi CPU/GPU. + """ + self.param_path = param_path + self.scope = core.Scope() + self.parallel = parallel + self.place = check_and_get_place(place) + + self.inference_program = framework.Program() + with framework.program_guard(self.inference_program): + with unique_name.guard(): + self.predict_var = infer_func() + + with self._prog_and_scope_guard(): + # load params from param_path into scope + io.load_params(executor.Executor(self.place), param_path) + + if parallel: + with self._prog_and_scope_guard(): + self.exe = parallel_executor.ParallelExecutor( + use_cuda=isinstance(self.place, core.CUDAPlace), + loss_name=self.predict_var.name) + else: + self.exe = executor.Executor(self.place) + + def infer(self, inputs, return_numpy=True): + """ + :param inputs: a map of {"input_name": input_var} that will be feed into the inference program + to get the predict value + :return: the predict value of the inference model + """ + if not isinstance(inputs, dict): + raise ValueError( + "inputs should be a map of {'input_name': input_var}") + + with executor.scope_guard(self.scope): + results = self.exe.run(self.inference_program, + feed=inputs, + fetch_list=[self.predict_var], + return_numpy=return_numpy) + + return results + + @contextlib.contextmanager + def _prog_and_scope_guard(self): + with framework.program_guard(main_program=self.inference_program): + with executor.scope_guard(self.scope): + yield diff --git a/python/paddle/fluid/io.py b/python/paddle/fluid/io.py index f7f1ca2598a3e679b24fa8d62c52e4f4de788fe2..8e58e5eb794e1bb507ab05394a1f7b57a1d2ed42 100644 --- a/python/paddle/fluid/io.py +++ b/python/paddle/fluid/io.py @@ -13,21 +13,18 @@ # limitations under the License. import os +import time +import shutil from paddle.fluid.evaluator import Evaluator from paddle.fluid.framework import Program, Parameter, default_main_program, Variable from . import core __all__ = [ - 'save_vars', - 'save_params', - 'save_persistables', - 'load_vars', - 'load_params', - 'load_persistables', - 'save_inference_model', - 'load_inference_model', - 'get_inference_program', + 'save_vars', 'save_params', 'save_persistables', 'load_vars', 'load_params', + 'load_persistables', 'save_inference_model', 'load_inference_model', + 'get_inference_program', 'save_checkpoint', 'load_checkpoint', + 'clean_checkpoint' ] @@ -195,6 +192,8 @@ def load_vars(executor, load_var_map = {} for each_var in vars: assert isinstance(each_var, Variable) + if each_var.type == core.VarDesc.VarType.RAW: + continue new_var = _clone_var_in_block_(load_block, each_var) if filename is None: load_block.append_op( @@ -263,6 +262,9 @@ def get_inference_program(target_vars, main_program=None): def prepend_feed_ops(inference_program, feed_target_names, feed_holder_name='feed'): + if len(feed_target_names) == 0: + return + global_block = inference_program.global_block() feed_var = global_block.create_var( name=feed_holder_name, @@ -323,9 +325,10 @@ def save_inference_model(dirname, if isinstance(feeded_var_names, basestring): feeded_var_names = [feeded_var_names] else: - if not (bool(feeded_var_names) and all( - isinstance(name, basestring) for name in feeded_var_names)): - raise ValueError("'feed_var_names' should be a list of str.") + if len(feeded_var_names) > 0: + if not (bool(feeded_var_names) and all( + isinstance(name, basestring) for name in feeded_var_names)): + raise ValueError("'feed_var_names' should be a list of str.") if isinstance(target_vars, Variable): target_vars = [target_vars] @@ -336,7 +339,7 @@ def save_inference_model(dirname, if main_program is None: main_program = default_main_program() - copy_program = main_program + copy_program = main_program.clone() if not os.path.isdir(dirname): os.makedirs(dirname) @@ -450,3 +453,192 @@ def get_parameter_value_by_name(name, executor, program=None): program = default_main_program() var = program.global_block().var(name) return get_parameter_value(var, executor) + + +SUCCESS_MARK_FILENAME = "_SUCCESS" +CHECKPOINT_PREFIX = "checkpoint" +CHECKPOINT_SEPARATOR = "_" + + +def save_checkpoint(executor, + checkpoint_dir=None, + max_num_checkpoints=3, + save_interval_secs=600, + main_program=None): + """ + Save Checkpoint will save persistable LodTensor variables from main_program in checkpoint directory, + the directory named by serial number from 0 to (n -1), save_checkpoint use LRU strategy + to keep numbers of checkpoint directory, the numbers of checkpoint directory are max_num_checkpoints at most, + The interval between two saved checkpoints must greater than save_interval_secs. + + :param executor + :param checkpoint_dir + :param max_num_checkpoints + :param save_interval_secs + :param main_program + """ + if checkpoint_dir is None: + checkpoint_dir = os.getcwd() + + if not os.path.isdir(checkpoint_dir): + os.makedirs(checkpoint_dir) + + serial = _get_lastest_checkpoint_dir(checkpoint_dir) + if serial >= 0 and not _interval_secs_exceed( + _get_serial_dir(serial, checkpoint_dir), save_interval_secs): + return + + serial += 1 + cur_dir = _get_serial_dir(serial, checkpoint_dir) + + save_vars( + executor, + dirname=cur_dir, + main_program=main_program, + vars=None, + predicate=_is_checkpoint_var, + filename=None) + _write_success(cur_dir) + _lru_delete(checkpoint_dir, max_num_checkpoints) + + +def load_checkpoint(executor, checkpoint_dir=None, main_program=None): + """ + Load checkpoint from a directory by executor, + it will find the most recent saved checkpoint file and load it auto. + + :param executor + :param checkpoint_dir + :param main_program + """ + + if checkpoint_dir is None: + checkpoint_dir = os.getcwd() + + serial = _get_lastest_checkpoint_dir(checkpoint_dir) + + if serial < 0: + return + + cur_dir = _get_serial_dir(serial, checkpoint_dir) + + load_vars( + executor, + dirname=cur_dir, + main_program=main_program, + predicate=_is_checkpoint_var, + filename=None) + + +def clean_checkpoint(checkpoint_dir, delete_dir=False): + """ + clean the checkpoint dir, when the train exits normally, the trainer will call clean_checkpoint to delete checkpoint directory saved before. + delete_dir only works when the directory is empty, otherwise, OSError is raised. + """ + if checkpoint_dir is None: + checkpoint_dir = os.getcwd() + _lru_delete(checkpoint_dir, max_num_checkpoints=0) + + if delete_dir and not os.listdir(checkpoint_dir): + os.rmdir(checkpoint_dir) + + +def _get_serial_dir(serial, checkpoint_dir): + serial_folder = CHECKPOINT_PREFIX + CHECKPOINT_SEPARATOR + str(serial) + return os.path.join(checkpoint_dir, serial_folder) + + +def _is_checkpoint_var(var): + """ + the checkpoint will not save or load all the variables. + var type is FEED_MINIBATCH/FETCH_LIST/RAW or var name ends with @GRAD are discarded. + + :param var + """ + if var.desc.type() == core.VarDesc.VarType.FEED_MINIBATCH or \ + var.desc.type() == core.VarDesc.VarType.FETCH_LIST or \ + var.desc.type() == core.VarDesc.VarType.RAW: + return False + + if var.name.endswith("@GRAD"): + return False + + return var.persistable + + +def _interval_secs_exceed(dirname, save_interval_secs): + dir_time = os.path.getmtime(dirname) + if save_interval_secs > (time.time() - dir_time): + return False + return True + + +def _lru_delete(dirname, max_num_checkpoints=3): + dirs = os.listdir(dirname) + serials = [] + for serial in dirs: + try: + serials.append(int(serial)) + except ValueError: + continue + + if len(serials) <= max_num_checkpoints: + return + + serials.sort(reverse=True) + serials = serials[max_num_checkpoints:] + for serial in serials: + cur_dir = os.path.join(dirname, str(serial)) + shutil.rmtree(cur_dir) + + +def _write_success(dirname): + """ + write an empty file named "_SUCCESS" in checkpoint dir, indicate this checkpoint is correct. + + :param dirname + """ + success_file = os.path.join(dirname, SUCCESS_MARK_FILENAME) + with open(success_file, 'a') as f: + now = time.ctime() + f.write(now) + + +def _get_lastest_checkpoint_dir(checkpoint_dir): + """ + get the latest file in checkpoint directory, the _SUCCESS file must exist in the directory + + :param checkpoint_dir + """ + if not checkpoint_dir.strip(): + return -1 + + def has_success(checkpoint_dir, cur_dir): + """ + is _SUCCESS in this dir + """ + _, serial = cur_dir.split(CHECKPOINT_SEPARATOR) + + try: + int(serial) + except ValueError: + return -1 + + if not os.path.isdir(os.path.join(checkpoint_dir, cur_dir)): + return -1 + + success_path = os.path.join( + _get_serial_dir(serial, checkpoint_dir), SUCCESS_MARK_FILENAME) + if os.path.isfile(success_path): + return int(serial) + + if not os.path.isdir(checkpoint_dir): + return -1 + + current_dir = -1 + dirs = os.listdir(checkpoint_dir) + for cur_dir in dirs: + success_num = has_success(checkpoint_dir, cur_dir) + if success_num > current_dir: + current_dir = success_num + return current_dir diff --git a/python/paddle/fluid/layer_helper.py b/python/paddle/fluid/layer_helper.py index 62933b512529bd04fab8c9ded12e636ecfae685c..86efd1ff51cf29485ee28b4d60ffb1439af1aad9 100644 --- a/python/paddle/fluid/layer_helper.py +++ b/python/paddle/fluid/layer_helper.py @@ -400,11 +400,11 @@ class LayerHelper(object): if isinstance(act, basestring): act = {'type': act} + if 'use_cudnn' in self.kwargs and self.kwargs.get('use_cudnn'): + act['use_cudnn'] = self.kwargs.get('use_cudnn') if 'use_mkldnn' in self.kwargs: act['use_mkldnn'] = self.kwargs.get('use_mkldnn') act_type = act.pop('type') - if 'use_mkldnn' in self.kwargs: - act['use_mkldnn'] = self.kwargs.get('use_mkldnn') tmp = input_var # NOTE(dzhwinter): some activation support inplace compution. if not core.IsInplace(act_type): diff --git a/python/paddle/fluid/layers/control_flow.py b/python/paddle/fluid/layers/control_flow.py index 4b707973e27391a6bdcba138934f62a255e04bb2..d1ea9f148566d20988a43f4c9d421c4452697ef1 100644 --- a/python/paddle/fluid/layers/control_flow.py +++ b/python/paddle/fluid/layers/control_flow.py @@ -49,6 +49,7 @@ __all__ = [ 'reorder_lod_tensor_by_rank', 'ParallelDo', 'Print', + 'is_empty', ] @@ -1097,7 +1098,7 @@ class ConditionalBlock(object): input_set = set([ipt.name for ipt in self.inputs]) param_list = [ - parent_block.var(each_name) for each_name in params + parent_block.var_recursive(each_name) for each_name in params if each_name not in input_set ] @@ -1562,3 +1563,40 @@ def reorder_lod_tensor_by_rank(x, rank_table): 'RankTable': [rank_table]}, outputs={'Out': [out]}) return out + + +def is_empty(x, cond=None, **ignored): + """ + **Is Empty** + + This layer returns the truth value of whether the variable is empty. + + Args: + x(Variable): Operand of *is_empty* + cond(Variable|None): Optional output variable to store the result + of *is_empty* + + Returns: + Variable: The tensor variable storing the output of *is_empty*. + + Raises: + TypeError: If input cond is not a variable, or cond's dtype is + not bool + + Examples: + .. code-block:: python + + less = fluid.layers.is_empty(x=input) + """ + helper = LayerHelper("is_empty", **locals()) + if cond is None: + cond = helper.create_tmp_variable(dtype='bool') + cond.stop_gradient = True + elif not isinstance(cond, Variable): + raise TypeError("cond takes a variable") + elif cond.dtype != 'bool': + raise TypeError("The data type of cond must be bool") + + helper.append_op( + type='is_empty', inputs={'X': [x]}, outputs={'Out': [cond]}) + return cond diff --git a/python/paddle/fluid/layers/detection.py b/python/paddle/fluid/layers/detection.py index a5938fe494265778ef7032c56a8d6d35acd729c5..3a83db12fd13651578deeac6b562bac2f1e4e4b6 100644 --- a/python/paddle/fluid/layers/detection.py +++ b/python/paddle/fluid/layers/detection.py @@ -23,6 +23,7 @@ import nn import math __all__ = [ + 'prior_box', 'multi_box_head', 'bipartite_match', 'target_assign', @@ -564,6 +565,115 @@ def ssd_loss(location, return loss +def prior_box(input, + image, + min_sizes, + max_sizes=None, + aspect_ratios=[1.], + variance=[0.1, 0.1, 0.2, 0.2], + flip=False, + clip=False, + steps=[0.0, 0.0], + offset=0.5, + name=None): + """ + **Prior box operator** + + Generate prior boxes for SSD(Single Shot MultiBox Detector) algorithm. + Each position of the input produce N prior boxes, N is determined by + the count of min_sizes, max_sizes and aspect_ratios, The size of the + box is in range(min_size, max_size) interval, which is generated in + sequence according to the aspect_ratios. + + Args: + input(Variable): The Input Variables, the format is NCHW. + image(Variable): The input image data of PriorBoxOp, + the layout is NCHW. + min_sizes(list|tuple|float value): min sizes of generated prior boxes. + max_sizes(list|tuple|None): max sizes of generated prior boxes. + Default: None. + aspect_ratios(list|tuple|float value): the aspect ratios of generated + prior boxes. Default: [1.]. + variance(list|tuple): the variances to be encoded in prior boxes. + Default:[0.1, 0.1, 0.2, 0.2]. + flip(bool): Whether to flip aspect ratios. Default:False. + clip(bool): Whether to clip out-of-boundary boxes. Default: False. + step(list|turple): Prior boxes step across width and height, If + step[0] == 0.0/step[1] == 0.0, the prior boxes step across + height/weight of the input will be automatically calculated. + Default: [0., 0.] + offset(float): Prior boxes center offset. Default: 0.5 + name(str): Name of the prior box op. Default: None. + + Returns: + boxes(Variable): the output prior boxes of PriorBox. + The layout is [H, W, num_priors, 4]. + H is the height of input, W is the width of input, + num_priors is the total + box count of each position of input. + Variances(Variable): the expanded variances of PriorBox. + The layout is [H, W, num_priors, 4]. + H is the height of input, W is the width of input + num_priors is the total + box count of each position of input + + + Examples: + .. code-block:: python + box, var = prior_box( + input=conv1, + image=images, + min_sizes=[100.], + flip=True, + clip=True) + """ + helper = LayerHelper("prior_box", **locals()) + dtype = helper.input_dtype() + + def _is_list_or_tuple_(data): + return (isinstance(data, list) or isinstance(data, tuple)) + + if not _is_list_or_tuple_(min_sizes): + min_sizes = [min_sizes] + if not _is_list_or_tuple_(aspect_ratios): + aspect_ratios = [aspect_ratios] + if not (_is_list_or_tuple_(steps) and len(steps) == 2): + raise ValueError('steps should be a list or tuple ', + 'with length 2, (step_width, step_height).') + + min_sizes = list(map(float, min_sizes)) + aspect_ratios = list(map(float, aspect_ratios)) + steps = list(map(float, steps)) + + attrs = { + 'min_sizes': min_sizes, + 'aspect_ratios': aspect_ratios, + 'variances': variance, + 'flip': flip, + 'clip': clip, + 'step_w': steps[0], + 'step_h': steps[1], + 'offset': offset + } + if max_sizes is not None and len(max_sizes) > 0 and max_sizes[0] > 0: + if not _is_list_or_tuple_(max_sizes): + max_sizes = [max_sizes] + attrs['max_sizes'] = max_sizes + + box = helper.create_tmp_variable(dtype) + var = helper.create_tmp_variable(dtype) + helper.append_op( + type="prior_box", + inputs={"Input": input, + "Image": image}, + outputs={"Boxes": box, + "Variances": var}, + attrs=attrs, ) + box.stop_gradient = True + var.stop_gradient = True + return box, var + + def multi_box_head(inputs, image, base_size, @@ -660,47 +770,6 @@ def multi_box_head(inputs, clip=True) """ - def _prior_box_(input, - image, - min_sizes, - max_sizes, - aspect_ratios, - variance, - flip=False, - clip=False, - step_w=0.0, - step_h=0.0, - offset=0.5, - name=None): - helper = LayerHelper("prior_box", **locals()) - dtype = helper.input_dtype() - - attrs = { - 'min_sizes': min_sizes, - 'aspect_ratios': aspect_ratios, - 'variances': variance, - 'flip': flip, - 'clip': clip, - 'step_w': step_w, - 'step_h': step_h, - 'offset': offset - } - if len(max_sizes) > 0 and max_sizes[0] > 0: - attrs['max_sizes'] = max_sizes - - box = helper.create_tmp_variable(dtype) - var = helper.create_tmp_variable(dtype) - helper.append_op( - type="prior_box", - inputs={"Input": input, - "Image": image}, - outputs={"Boxes": box, - "Variances": var}, - attrs=attrs, ) - box.stop_gradient = True - var.stop_gradient = True - return box, var - def _reshape_with_axis_(input, axis=1): if not (axis > 0 and axis < len(input.shape)): raise ValueError("The axis should be smaller than " @@ -777,11 +846,10 @@ def multi_box_head(inputs, aspect_ratio = aspect_ratios[i] if not _is_list_or_tuple_(aspect_ratio): aspect_ratio = [aspect_ratio] + step = [step_w[i] if step_w else 0.0, step_h[i] if step_w else 0.0] - box, var = _prior_box_(input, image, min_size, max_size, aspect_ratio, - variance, flip, clip, step_w[i] - if step_w else 0.0, step_h[i] - if step_w else 0.0, offset) + box, var = prior_box(input, image, min_size, max_size, aspect_ratio, + variance, flip, clip, step, offset) box_results.append(box) var_results.append(var) diff --git a/python/paddle/fluid/layers/io.py b/python/paddle/fluid/layers/io.py index cc71c2136a6756ff094f6e06b8e200c6a68db06a..8758ac9f94ab91b5be5fc70917c64db38997d1c1 100644 --- a/python/paddle/fluid/layers/io.py +++ b/python/paddle/fluid/layers/io.py @@ -11,6 +11,7 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. +import contextlib from .. import core from ..framework import convert_np_dtype_to_dtype_, default_main_program, default_startup_program, Program @@ -21,7 +22,8 @@ from ..executor import global_scope __all__ = [ 'data', 'BlockGuardServ', 'ListenAndServ', 'Send', 'open_recordio_file', - 'open_files', 'read_file', 'shuffle', 'batch', 'double_buffer' + 'open_files', 'read_file', 'shuffle', 'batch', 'double_buffer', + 'random_data_generator', 'Preprocessor' ] @@ -50,8 +52,6 @@ def data(name, dtype(int|float): The type of data : float32, float_16, int etc type(VarType): The output type. By default it is LOD_TENSOR. lod_level(int): The LoD Level. 0 means the input data is not a sequence. - main_program(Program): Name of the main program that calls this - startup_program(Program): Name of the startup program stop_gradient(bool): A boolean that mentions whether gradient should flow. Returns: @@ -74,13 +74,15 @@ def data(name, if append_batch_size: shape = [-1] + shape # append batch size as -1 - return helper.create_global_variable( + data_var = helper.create_global_variable( name=name, shape=shape, dtype=dtype, type=type, stop_gradient=stop_gradient, - lod_level=lod_level) + lod_level=lod_level, + is_data=True) + return data_var class BlockGuardServ(BlockGuard): @@ -168,7 +170,9 @@ class ListenAndServ(object): 'endpoint': self.endpoint, 'Fanin': self.fan_in, 'OptimizeBlock': current_block, - 'PrefetchBlock': empty_block + 'PrefetchBlock': empty_block, + 'sync_mode': True, # did not support async now in layers + 'grad_to_block_id': [""] }) @@ -191,21 +195,23 @@ def Send(endpoints, send_vars, get_vars=None): endpoints = list(set(epmap)) helper = LayerHelper("Send", **locals()) - rpc_client_var = default_main_program().global_block().create_var( - name="RPC_CLIENT_VAR", persistable=True, type=core.VarDesc.VarType.RAW) if not get_vars: get_vars = [] for s in send_vars: v = helper.create_tmp_variable(dtype=s.dtype, stop_gradient=True) get_vars.append(v) + rpc_op_role_name = core.op_proto_and_checker_maker.kOpRoleAttrName() helper.append_op( type="send", inputs={"X": send_vars}, - outputs={"Out": get_vars, - "RPCClient": rpc_client_var}, - attrs={"endpoints": endpoints, - "epmap": epmap}) + outputs={"Out": get_vars}, + attrs={ + "endpoints": endpoints, + "epmap": epmap, + rpc_op_role_name: core.op_proto_and_checker_maker.OpRole.RPC + }) + return get_vars @@ -319,7 +325,7 @@ def open_recordio_file(filename, dtypes=['float32', 'int64']) # Via the reader, we can use 'read_file' layer to get data: - image, label = fluid.layers.read_file(reader) + image, label = fluid.layers.io.read_file(reader) """ dtypes = [convert_np_dtype_to_dtype_(dt) for dt in dtypes] shape_concat = [] @@ -357,6 +363,73 @@ def open_recordio_file(filename, return monkey_patch_reader_methods(main_prog_var) +def random_data_generator(low, high, shapes, lod_levels, for_parallel=True): + """ + Create a uniform random data generator + + This layer returns a Reader Variable. + Instead of opening a file and reading data from it, this + Reader Variable generates float uniform random data by itself. + It can be used as a dummy reader to test a network without + opening a real file. + + Args: + low(float): The lower bound of data's uniform distribution. + high(float): The upper bound of data's uniform distribution. + shapes(list): List of tuples which declaring data shapes. + lod_levels(list): List of ints which declaring data lod_level. + for_parallel(Bool): Set it as True if you are going to run + subsequent operators in parallel. + + Returns: + Variable: A Reader Variable from which we can get random data. + + Examples: + .. code-block:: python + + reader = fluid.layers.io.random_data_generator( + low=0.0, + high=1.0, + shapes=[(3,224,224), (1)], + lod_levels=[0, 0]) + + # Via the reader, we can use 'read_file' layer to get data: + image, label = fluid.layers.io.read_file(reader) + """ + dtypes = [core.VarDesc.VarType.FP32] * len(shapes) + shape_concat = [] + ranks = [] + + for shape in shapes: + shape_concat.extend(shape) + ranks.append(len(shape)) + + var_name = unique_name('random_data_generator') + + startup_blk = default_startup_program().current_block() + startup_var = startup_blk.create_var(name=var_name) + startup_blk.append_op( + type='create_random_data_generator', + outputs={'Out': [startup_var]}, + attrs={ + 'low': low, + 'high': high, + 'shape_concat': shape_concat, + 'lod_levels': lod_levels, + 'ranks': ranks + }) + + startup_var.desc.set_dtypes(dtypes) + startup_var.persistable = True + main_prog_var = _copy_reader_var_(default_main_program().current_block(), + startup_var) + + if for_parallel: + main_prog_var = parallel(reader=main_prog_var) + + return monkey_patch_reader_methods(main_prog_var) + + def open_files(filenames, shapes, lod_levels, @@ -466,8 +539,6 @@ def __create_unshared_decorated_reader__(op_type, reader, attrs, name=None): inputs={'UnderlyingReader': reader}, outputs={'Out': [new_reader]}, attrs=attrs) - new_reader.persistable = True - new_reader.stop_gradient = True return monkey_patch_reader_methods(new_reader) @@ -512,3 +583,82 @@ def read_file(file_obj): return out[0] else: return out + + +class Preprocessor(object): + BEFORE_SUB_BLOCK = 0 + IN_SUB_BLOCK = 1 + AFTER_SUB_BLOCK = 2 + + def __init__(self, reader, name=None): + self.underlying_reader = reader + new_reader_name = name if name is not None else unique_name( + "create_custom_reader") + self.main_prog = default_main_program() + self.reader = self.main_prog.current_block().create_var( + name=new_reader_name) + self.sub_block = None + self.source_var_names = None + self.sink_var_names = None + self.status = Preprocessor.BEFORE_SUB_BLOCK + + def is_completed(self): + return self.sub_block and self.source_var_names and self.sink_var_names + + @contextlib.contextmanager + def block(self): + self.status = Preprocessor.IN_SUB_BLOCK + self.sub_block = self.main_prog.create_block() + yield + self.main_prog.rollback() + self.status = Preprocessor.AFTER_SUB_BLOCK + if not self.is_completed(): + raise RuntimeError( + "The definition of preprocessor is incompleted! " + "Please make sure that you have set input and output " + "variables by invoking 'inputs' and 'outputs' in " + "Preprocessor's sub-block.") + + def inputs(self): + if self.status != Preprocessor.IN_SUB_BLOCK: + raise RuntimeError( + "Preprocessor.inputs() can only be invoked inside the sub-block." + ) + + source_shapes = self.underlying_reader.desc.shapes() + source_dtypes = self.underlying_reader.desc.dtypes() + source_lod_levels = self.underlying_reader.desc.lod_levels() + self.source_var_names = [ + unique_name("preprocessor_source") + for _ in xrange(len(source_shapes)) + ] + source_vars = [] + for var_name, shape, dtype, lod_level in zip( + self.source_var_names, source_shapes, source_dtypes, + source_lod_levels): + source_vars.append(self.main_prog.current_block().create_var( + name=var_name, shape=shape, dtype=dtype, lod_level=lod_level)) + return source_vars + + def outputs(self, *outs): + if self.status != Preprocessor.IN_SUB_BLOCK: + raise RuntimeError( + "Preprocessor.outputs() can only be invoked inside the sub-block." + ) + self.sink_var_names = [var.name for var in outs] + + def __call__(self, *args, **kwargs): + if self.status != Preprocessor.AFTER_SUB_BLOCK: + raise RuntimeError( + "Preprocessor output can only be retrieved after rnn block.") + + self.main_prog.current_block().append_op( + type="create_custom_reader", + inputs={'UnderlyingReader': self.underlying_reader}, + outputs={'Out': [self.reader]}, + attrs={ + "sub_block": self.sub_block, + "source_var_names": self.source_var_names, + "sink_var_names": self.sink_var_names + }) + return monkey_patch_reader_methods(self.reader) diff --git a/python/paddle/fluid/layers/layer_function_generator.py b/python/paddle/fluid/layers/layer_function_generator.py index 35b01a79914b3427836d4abd51aa2e2eb471d517..295d1b7190ec39bcc6efdf72aebede14a99807aa 100644 --- a/python/paddle/fluid/layers/layer_function_generator.py +++ b/python/paddle/fluid/layers/layer_function_generator.py @@ -113,7 +113,7 @@ def generate_layer_fn(op_type): if len(not_intermediate_outputs) != 1: raise ValueError("Only one non intermediate output operator can be", - "automatically generated.") + "automatically generated. {0}".format(op_type)) if not_intermediate_outputs[0].duplicable: raise ValueError( diff --git a/python/paddle/fluid/layers/math_op_patch.py b/python/paddle/fluid/layers/math_op_patch.py index 08a0184c2c2ad5f3c3792fd0a12f0ab0c746849b..1754061c4ba6f5b97bced3548bc412dfb1b7932c 100644 --- a/python/paddle/fluid/layers/math_op_patch.py +++ b/python/paddle/fluid/layers/math_op_patch.py @@ -169,7 +169,9 @@ def monkey_patch_variable(): # a*b == b*a. Do not need to reverse explicitly ("__rmul__", "elementwise_mul", False), ("__div__", "elementwise_div", False), + ("__truediv__", "elementwise_div", False), ("__rdiv__", "elementwise_div", True), + ("__rtruediv__", "elementwise_div", True), ("__pow__", "elementwise_pow", False), ("__rpow__", "elementwise_pow", True), # for logical compare diff --git a/python/paddle/fluid/layers/nn.py b/python/paddle/fluid/layers/nn.py index 9a0c328033cdfdae39da050fc482abba17032dd9..63ec83151477770ea64070cae4f5e4fcc497f7af 100644 --- a/python/paddle/fluid/layers/nn.py +++ b/python/paddle/fluid/layers/nn.py @@ -79,6 +79,10 @@ __all__ = [ 'lrn', 'pad', 'label_smooth', + 'roi_pool', + 'dice_loss', + 'upsampling_bilinear2d', + 'random_crop', ] @@ -87,6 +91,7 @@ def fc(input, num_flatten_dims=1, param_attr=None, bias_attr=None, + use_cudnn=False, use_mkldnn=False, act=None, is_test=False, @@ -150,7 +155,8 @@ def fc(input, Examples: .. code-block:: python - data = fluid.layers.data(name="data", shape=[32, 32], dtype="float32") + data = fluid.layers.data( + name="data", shape=[32, 32], dtype="float32") fc = fluid.layers.fc(input=data, size=1000, act="tanh") """ @@ -173,11 +179,8 @@ def fc(input, inputs={"X": input_var, "Y": w}, outputs={"Out": tmp}, - attrs={ - "x_num_col_dims": num_flatten_dims, - "y_num_col_dims": 1, - "use_mkldnn": use_mkldnn - }) + attrs={"x_num_col_dims": num_flatten_dims, + "y_num_col_dims": 1}) mul_results.append(tmp) if len(mul_results) == 1: @@ -345,7 +348,8 @@ def dynamic_lstm(input, cell_activation(str): The activation for cell output. Choices = ["sigmoid", "tanh", "relu", "identity"], default "tanh". candidate_activation(str): The activation for candidate hidden state. - Choices = ["sigmoid", "tanh", "relu", "identity"], + Choices = ["sigmoid", "tanh", + "relu", "identity"], default "tanh". dtype(str): Data type. Choices = ["float32", "float64"], default "float32". name(str|None): A name for this layer(optional). If set None, the layer @@ -512,10 +516,12 @@ def dynamic_lstmp(input, cell_activation(str): The activation for cell output. Choices = ["sigmoid", "tanh", "relu", "identity"], default "tanh". candidate_activation(str): The activation for candidate hidden state. - Choices = ["sigmoid", "tanh", "relu", "identity"], + Choices = ["sigmoid", "tanh", + "relu", "identity"], default "tanh". proj_activation(str): The activation for projection output. - Choices = ["sigmoid", "tanh", "relu", "identity"], + Choices = ["sigmoid", "tanh", + "relu", "identity"], default "tanh". dtype(str): Data type. Choices = ["float32", "float64"], default "float32". name(str|None): A name for this layer(optional). If set None, the layer @@ -697,8 +703,8 @@ def dynamic_gru(input, def gru_unit(input, hidden, size, - weight=None, - bias=None, + param_attr=None, + bias_attr=None, activation='tanh', gate_activation='sigmoid'): """ @@ -729,8 +735,8 @@ def gru_unit(input, input (Variable): The fc transformed input value of current step. hidden (Variable): The hidden value of lstm unit from previous step. size (integer): The input dimension value. - weight (ParamAttr): The weight parameters for gru unit. Default: None - bias (ParamAttr): The bias parameters for gru unit. Default: None + param_attr (ParamAttr): The weight parameters for gru unit. Default: None + bias_attr (ParamAttr): The bias parameters for gru unit. Default: None activation (string): The activation type for cell (actNode). Default: 'tanh' gate_activation (string): The activation type for gates (actGate). @@ -762,34 +768,31 @@ def gru_unit(input, size = size / 3 # create weight - if weight is None: - weight = helper.create_parameter( - attr=helper.param_attr, shape=[size, 3 * size], dtype=dtype) + weight = helper.create_parameter( + attr=helper.param_attr, shape=[size, 3 * size], dtype=dtype) + gate = helper.create_tmp_variable(dtype) + reset_hidden_pre = helper.create_tmp_variable(dtype) + updated_hidden = helper.create_tmp_variable(dtype) + inputs = {'Input': input, 'HiddenPrev': hidden, 'Weight': weight} # create bias - - if bias is None: + if helper.bias_attr: bias_size = [1, 3 * size] bias = helper.create_parameter( attr=helper.bias_attr, shape=bias_size, dtype=dtype, is_bias=True) - - gate = helper.create_tmp_variable(dtype) - reset_hidden_pre = helper.create_tmp_variable(dtype) - updated_hidden = helper.create_tmp_variable(dtype) + inputs['Bias'] = bias helper.append_op( type='gru_unit', - inputs={'Input': input, - 'HiddenPrev': hidden, - 'Weight': weight}, + inputs=inputs, outputs={ 'Gate': gate, 'ResetHiddenPrev': reset_hidden_pre, 'Hidden': updated_hidden, }, attrs={ - 'activation': 0, - 'gate_activation': 1, + 'activation': 2, # tanh + 'gate_activation': 1, # sigmoid }) return updated_hidden, reset_hidden_pre, gate @@ -854,7 +857,7 @@ def cos_sim(X, Y): return out -def dropout(x, dropout_prob, is_test=False, seed=None): +def dropout(x, dropout_prob, is_test=False, seed=None, name=None): """ Computes dropout. @@ -872,6 +875,8 @@ def dropout(x, dropout_prob, is_test=False, seed=None): parameter is set to None, a random seed is used. NOTE: If an integer seed is given, always the same output units will be dropped. DO NOT use a fixed seed in training. + name(str|None): A name for this layer(optional). If set None, the layer + will be named automatically. Returns: Variable: A tensor variable. @@ -1116,7 +1121,7 @@ def sequence_softmax(input, param_attr=None, bias_attr=None, use_cudnn=True): return softmax_out -def softmax(input, param_attr=None, bias_attr=None, use_cudnn=True): +def softmax(input, param_attr=None, bias_attr=None, use_cudnn=True, name=None): helper = LayerHelper('softmax', **locals()) dtype = helper.input_dtype() softmax_out = helper.create_tmp_variable(dtype) @@ -1327,6 +1332,8 @@ def sequence_pool(input, pool_type): sqrt : out.data = [2.82, 6.93, 4.24], where 2.82=(1+3)/sqrt(2), 6.93=(2+4+6)/sqrt(3), 4.24=(5+1)/sqrt(2) max : out.data = [3, 6, 5], where 3=max(1,3), 6=max(2,4,6), 5=max(5,1) + last : out.data = [3, 6, 1], where 3=last(1,3), 6=last(2,4,6), 1=last(5,1) + first : out.data = [1, 2, 5], where 1=first(1,3), 2=first(2,4,6), 5=first(5,1) Args: input(variable): The input variable which is a LoDTensor. @@ -1346,6 +1353,8 @@ def sequence_pool(input, pool_type): sum_x = fluid.layers.sequence_pool(input=x, pool_type='sum') sqrt_x = fluid.layers.sequence_pool(input=x, pool_type='sqrt') max_x = fluid.layers.sequence_pool(input=x, pool_type='max') + last_x = fluid.layers.sequence_pool(input=x, pool_type='last') + first_x = fluid.layers.sequence_pool(input=x, pool_type='first') """ helper = LayerHelper('sequence_pool', **locals()) dtype = helper.input_dtype() @@ -1495,6 +1504,7 @@ def batch_norm(input, bias_attr=None, data_layout='NCHW', in_place=False, + use_mkldnn=False, name=None, moving_mean_name=None, moving_variance_name=None, @@ -1573,9 +1583,12 @@ def batch_norm(input, "SavedMean": saved_mean, "SavedVariance": saved_variance }, - attrs={"momentum": momentum, - "epsilon": epsilon, - "is_test": is_test}) + attrs={ + "momentum": momentum, + "epsilon": epsilon, + "is_test": is_test, + "use_mkldnn": use_mkldnn + }) return helper.append_activation(batch_norm_out) @@ -1700,6 +1713,7 @@ def conv2d_transpose(input, padding=0, stride=1, dilation=1, + groups=None, param_attr=None, bias_attr=None, use_cudnn=True, @@ -1770,6 +1784,12 @@ def conv2d_transpose(input, dilation(int|tuple): The dilation size. If dilation is a tuple, it must contain two integers, (dilation_H, dilation_W). Otherwise, the dilation_H = dilation_W = dilation. Default: dilation = 1. + groups(int): The groups number of the Conv2d transpose layer. Inspired by + grouped convolution in Alex Krizhevsky's Deep CNN paper, in which + when group=2, the first half of the filters is only connected to the + first half of the input channels, while the second half of the + filters is only connected to the second half of the input channels. + Default: groups=1 param_attr(ParamAttr): The parameters to the Conv2d_transpose Layer. Default: None bias_attr(ParamAttr): Bias parameter for the Conv2d layer. Default: None @@ -1824,7 +1844,8 @@ def conv2d_transpose(input, filter_size = utils.convert_to_list(filter_size, 2, 'conv2d_transpose.filter_size') - filter_shape = [input_channel, num_filters] + filter_size + groups = 1 if groups is None else groups + filter_shape = [input_channel, num_filters / groups] + filter_size img_filter = helper.create_parameter( dtype=input.dtype, shape=filter_shape, attr=helper.param_attr) @@ -1838,6 +1859,7 @@ def conv2d_transpose(input, 'strides': stride, 'paddings': padding, 'dilations': dilation, + 'groups': groups, 'use_cudnn': use_cudnn }) @@ -2074,11 +2096,11 @@ def reduce_sum(input, dim=None, keep_dim=False, name=None): Args: input (Variable): The input variable which is a Tensor or LoDTensor. - dim (int|None): The dimension along which the sum is performed. If + dim (list|int|None): The dimensions along which the sum is performed. If :attr:`None`, sum all elements of :attr:`input` and return a Tensor variable with a single element, otherwise must be in the - range :math:`[-rank(input), rank(input))`. If :math:`dim < 0`, - the dimension to reduce is :math:`rank + dim`. + range :math:`[-rank(input), rank(input))`. If :math:`dim[i] < 0`, + the dimension to reduce is :math:`rank + dim[i]`. keep_dim (bool|False): Whether to reserve the reduced dimension in the output Tensor. The result tensor will have one fewer dimension than the :attr:`input` unless :attr:`keep_dim` is true. @@ -2099,15 +2121,25 @@ def reduce_sum(input, dim=None, keep_dim=False, name=None): fluid.layers.reduce_sum(x, dim=0) # [0.3, 0.5, 1.1, 1.6] fluid.layers.reduce_sum(x, dim=-1) # [1.9, 1.6] fluid.layers.reduce_sum(x, dim=1, keep_dim=True) # [[1.9], [1.6]] + + # x is a Tensor variable with shape [2, 2, 2] and elements as below: + # [[[1, 2], [3, 4]], + # [[5, 6], [7, 8]]] + # Each example is followed by the correspending output tensor. + fluid.layers.reduce_sum(x, dim=[1, 2]) # [10, 26] + fluid.layers.reduce_sum(x, dim=[0, 1]) # [16, 20] + """ helper = LayerHelper('reduce_sum', **locals()) out = helper.create_tmp_variable(dtype=helper.input_dtype()) + if dim is not None and not isinstance(dim, list): + dim = [dim] helper.append_op( type='reduce_sum', inputs={'X': input}, outputs={'Out': out}, attrs={ - 'dim': dim if dim != None else 0, + 'dim': dim if dim != None else [0], 'keep_dim': keep_dim, 'reduce_all': True if dim == None else False }) @@ -2120,11 +2152,11 @@ def reduce_mean(input, dim=None, keep_dim=False, name=None): Args: input (Variable): The input variable which is a Tensor or LoDTensor. - dim (int|None): The dimension along which the mean is computed. If + dim (list|int|None): The dimensions along which the mean is computed. If :attr:`None`, compute the mean over all elements of :attr:`input` and return a Tensor variable with a single element, otherwise must be in the range :math:`[-rank(input), rank(input))`. If - :math:`dim < 0`, the dimension to reduce is :math:`rank + dim`. + :math:`dim[i] < 0`, the dimension to reduce is :math:`rank + dim[i]`. keep_dim (bool): Whether to reserve the reduced dimension in the output Tensor. The result tensor will have one fewer dimension than the :attr:`input` unless :attr:`keep_dim` is true. @@ -2144,16 +2176,26 @@ def reduce_mean(input, dim=None, keep_dim=False, name=None): fluid.layers.reduce_mean(x) # [0.4375] fluid.layers.reduce_mean(x, dim=0) # [0.15, 0.25, 0.55, 0.8] fluid.layers.reduce_mean(x, dim=-1) # [0.475, 0.4] - fluid.layers.reduce_mean(x, dim=1, keep_dim=True) # [[0.475], [0.4]] + fluid.layers.reduce_mean( + x, dim=1, keep_dim=True) # [[0.475], [0.4]] + + # x is a Tensor variable with shape [2, 2, 2] and elements as below: + # [[[1.0, 2.0], [3.0, 4.0]], + # [[5.0, 6.0], [7.0, 8.0]]] + # Each example is followed by the correspending output tensor. + fluid.layers.reduce_mean(x, dim=[1, 2]) # [2.5, 6.5] + fluid.layers.reduce_mean(x, dim=[0, 1]) # [4.0, 5.0] """ helper = LayerHelper('reduce_mean', **locals()) out = helper.create_tmp_variable(dtype=helper.input_dtype()) + if dim is not None and not isinstance(dim, list): + dim = [dim] helper.append_op( type='reduce_mean', inputs={'X': input}, outputs={'Out': out}, attrs={ - 'dim': dim if dim != None else 0, + 'dim': dim if dim != None else [0], 'keep_dim': keep_dim, 'reduce_all': True if dim == None else False }) @@ -2166,11 +2208,11 @@ def reduce_max(input, dim=None, keep_dim=False, name=None): Args: input (Variable): The input variable which is a Tensor or LoDTensor. - dim (int|None): The dimension along which the maximum is computed. + dim (list|int|None): The dimension along which the maximum is computed. If :attr:`None`, compute the maximum over all elements of :attr:`input` and return a Tensor variable with a single element, otherwise must be in the range :math:`[-rank(input), rank(input))`. - If :math:`dim < 0`, the dimension to reduce is :math:`rank + dim`. + If :math:`dim[i] < 0`, the dimension to reduce is :math:`rank + dim[i]`. keep_dim (bool): Whether to reserve the reduced dimension in the output Tensor. The result tensor will have one fewer dimension than the :attr:`input` unless :attr:`keep_dim` is true. @@ -2191,15 +2233,24 @@ def reduce_max(input, dim=None, keep_dim=False, name=None): fluid.layers.reduce_max(x, dim=0) # [0.2, 0.3, 0.6, 0.9] fluid.layers.reduce_max(x, dim=-1) # [0.9, 0.7] fluid.layers.reduce_max(x, dim=1, keep_dim=True) # [[0.9], [0.7]] + + # x is a Tensor variable with shape [2, 2, 2] and elements as below: + # [[[1.0, 2.0], [3.0, 4.0]], + # [[5.0, 6.0], [7.0, 8.0]]] + # Each example is followed by the correspending output tensor. + fluid.layers.reduce_max(x, dim=[1, 2]) # [4.0, 8.0] + fluid.layers.reduce_max(x, dim=[0, 1]) # [7.0, 8.0] """ helper = LayerHelper('reduce_max', **locals()) out = helper.create_tmp_variable(dtype=helper.input_dtype()) + if dim is not None and not isinstance(dim, list): + dim = [dim] helper.append_op( type='reduce_max', inputs={'X': input}, outputs={'Out': out}, attrs={ - 'dim': dim if dim != None else 0, + 'dim': dim if dim != None else [0], 'keep_dim': keep_dim, 'reduce_all': True if dim == None else False }) @@ -2212,11 +2263,11 @@ def reduce_min(input, dim=None, keep_dim=False, name=None): Args: input (Variable): The input variable which is a Tensor or LoDTensor. - dim (int|None): The dimension along which the minimum is computed. + dim (list|int|None): The dimensions along which the minimum is computed. If :attr:`None`, compute the minimum over all elements of :attr:`input` and return a Tensor variable with a single element, otherwise must be in the range :math:`[-rank(input), rank(input))`. - If :math:`dim < 0`, the dimension to reduce is :math:`rank + dim`. + If :math:`dim[i] < 0`, the dimension to reduce is :math:`rank + dim[i]`. keep_dim (bool): Whether to reserve the reduced dimension in the output Tensor. The result tensor will have one fewer dimension than the :attr:`input` unless :attr:`keep_dim` is true. @@ -2237,15 +2288,24 @@ def reduce_min(input, dim=None, keep_dim=False, name=None): fluid.layers.reduce_min(x, dim=0) # [0.1, 0.2, 0.5, 0.7] fluid.layers.reduce_min(x, dim=-1) # [0.2, 0.1] fluid.layers.reduce_min(x, dim=1, keep_dim=True) # [[0.2], [0.1]] + + # x is a Tensor variable with shape [2, 2, 2] and elements as below: + # [[[1.0, 2.0], [3.0, 4.0]], + # [[5.0, 6.0], [7.0, 8.0]]] + # Each example is followed by the correspending output tensor. + fluid.layers.reduce_min(x, dim=[1, 2]) # [1.0, 5.0] + fluid.layers.reduce_min(x, dim=[0, 1]) # [1.0, 2.0] """ helper = LayerHelper('reduce_min', **locals()) out = helper.create_tmp_variable(dtype=helper.input_dtype()) + if dim is not None and not isinstance(dim, list): + dim = [dim] helper.append_op( type='reduce_min', inputs={'X': input}, outputs={'Out': out}, attrs={ - 'dim': dim if dim != None else 0, + 'dim': dim if dim != None else [0], 'keep_dim': keep_dim, 'reduce_all': True if dim == None else False }) @@ -2258,11 +2318,11 @@ def reduce_prod(input, dim=None, keep_dim=False, name=None): Args: input (Variable): The input variable which is a Tensor or LoDTensor. - dim (int|None): The dimension along which the product is performed. If + dim (list|int|None): The dimensions along which the product is performed. If :attr:`None`, multipy all elements of :attr:`input` and return a Tensor variable with a single element, otherwise must be in the - range :math:`[-rank(input), rank(input))`. If :math:`dim < 0`, - the dimension to reduce is :math:`rank + dim`. + range :math:`[-rank(input), rank(input))`. If :math:`dim[i] < 0`, + the dimension to reduce is :math:`rank + dim[i]`. keep_dim (bool|False): Whether to reserve the reduced dimension in the output Tensor. The result tensor will have one fewer dimension than the :attr:`input` unless :attr:`keep_dim` is true. @@ -2284,15 +2344,24 @@ def reduce_prod(input, dim=None, keep_dim=False, name=None): fluid.layers.reduce_prod(x, dim=-1) # [0.027, 0.0084] fluid.layers.reduce_prod(x, dim=1, keep_dim=True) # [[0.027], [0.0084]] + + # x is a Tensor variable with shape [2, 2, 2] and elements as below: + # [[[1.0, 2.0], [3.0, 4.0]], + # [[5.0, 6.0], [7.0, 8.0]]] + # Each example is followed by the correspending output tensor. + fluid.layers.reduce_prod(x, dim=[1, 2]) # [24.0, 1680.0] + fluid.layers.reduce_prod(x, dim=[0, 1]) # [105.0, 384.0] """ helper = LayerHelper('reduce_prod', **locals()) out = helper.create_tmp_variable(dtype=helper.input_dtype()) + if dim is not None and not isinstance(dim, list): + dim = [dim] helper.append_op( type='reduce_prod', inputs={'X': input}, outputs={'Out': out}, attrs={ - 'dim': dim if dim != None else 0, + 'dim': dim if dim != None else [0], 'keep_dim': keep_dim, 'reduce_all': True if dim == None else False }) @@ -2327,7 +2396,8 @@ def split(input, num_or_sections, dim=-1, name=None): x0.shape # [3, 3, 5] x1.shape # [3, 3, 5] x2.shape # [3, 3, 5] - x0, x1, x2 = fluid.layers.split(x, num_or_sections=[2, 3, 4], dim=1) + x0, x1, x2 = fluid.layers.split( + x, num_or_sections=[2, 3, 4], dim=1) x0.shape # [3, 2, 5] x1.shape # [3, 3, 5] x2.shape # [3, 4, 5] @@ -2395,7 +2465,6 @@ def l2_normalize(x, axis, epsilon=1e-12, name=None): if len(x.shape) == 1: axis = 0 - helper = LayerHelper("l2_normalize", **locals()) square = helper.create_tmp_variable(dtype=x.dtype) @@ -2407,7 +2476,7 @@ def l2_normalize(x, axis, epsilon=1e-12, name=None): inputs={"X": square}, outputs={"Out": reduced_sum}, attrs={ - "dim": 1 if axis is None else axis, + "dim": [1] if axis is None else [axis], "keep_dim": True, "reduce_all": False }) @@ -2547,7 +2616,7 @@ def matmul(x, y, transpose_x=False, transpose_y=False, name=None): return out -def topk(input, k): +def topk(input, k, name=None): """ This operator is used to find values and indices of the k largest entries for the last dimension. @@ -2563,6 +2632,8 @@ def topk(input, k): input(Variable): The input variable which can be a vector or Tensor with higher rank. k(int): An integer value to specify the top k largest elements. + name(str|None): A name for this layer(optional). If set None, the layer + will be named automatically. Returns: values(Variable): The k largest elements along each last dimensional @@ -3238,7 +3309,8 @@ def softmax_with_cross_entropy(logits, label, soft_label=False): data = fluid.layers.data(name='data', shape=[128], dtype='float32') label = fluid.layers.data(name='label', shape=[1], dtype='int64') fc = fluid.layers.fc(input=data, size=100) - out = fluid.layers.softmax_with_cross_entropy(logits=fc, label=label) + out = fluid.layers.softmax_with_cross_entropy( + logits=fc, label=label) """ helper = LayerHelper('softmax_with_cross_entropy', **locals()) softmax = helper.create_tmp_variable(dtype=logits.dtype) @@ -3257,35 +3329,36 @@ def smooth_l1(x, y, inside_weight=None, outside_weight=None, sigma=None): """ **Smooth L1 Loss Operator. ** - This operator computes the smooth l1 loss for X and Y. + This operator computes the smooth L1 loss for X and Y. The operator takes the first dimension of X and Y as batch size. - For each instance, it computes the smooth l1 loss element by element first + For each instance, it computes the smooth L1 loss element by element first and then sums all the losses. So the shape of Out is [batch_size, 1]. Args: x (Variable): A tensor with rank at least 2. The input value of smooth - l1 loss op with shape [batch_size, dim1, ..., dimN]. + L1 loss op with shape [batch_size, dim1, ..., dimN]. y (Variable): A tensor with rank at least 2. The target value of smooth - l1 loss op with same shape as x. + L1 loss op with same shape as x. inside_weight (Variable|None): A tensor with rank at least 2. This input is optional and should have same shape with x. If provided, the result of (x - y) will be multiplied by this tensor element by element. outside_weight (Variable|None): A tensor with rank at least 2. This input is optional and should have same shape with x. If provided, - the out smooth l1 loss will be multiplied by this tensor element + the out smooth L1 loss will be multiplied by this tensor element by element. - sigma (float|None): Hyper parameter of smooth l1 loss op. A float scalar + sigma (float|None): Hyper parameter of smooth L1 loss op. A float scalar with default value 1.0. Returns: - Variable: A tensor with rank be 2. The output smooth l1 loss with + Variable: A tensor with rank be 2. The output smooth L1 loss with shape [batch_size, 1]. Examples: .. code-block:: python data = fluid.layers.data(name='data', shape=[128], dtype='float32') - label = fluid.layers.data(name='label', shape=[100], dtype='int64') + label = fluid.layers.data( + name='label', shape=[100], dtype='float32') fc = fluid.layers.fc(input=data, size=100) out = fluid.layers.smooth_l1(x=fc, y=label) """ @@ -3607,7 +3680,8 @@ def lrn(input, n=5, k=1.0, alpha=1e-4, beta=0.75, name=None): Examples: .. code-block:: python - data = fluid.layers.data(name="data", shape=[3, 112, 112], dtype="float32") + data = fluid.layers.data( + name="data", shape=[3, 112, 112], dtype="float32") lrn = fluid.layers.lrn(input=data) """ helper = LayerHelper('lrn', **locals()) @@ -3759,3 +3833,190 @@ def label_smooth(label, outputs={"Out": smooth_label}, attrs={"epsilon": float(epsilon)}) return smooth_label + + +def roi_pool(input, rois, pooled_height=1, pooled_width=1, spatial_scale=1.0): + """ + Region of interest pooling (also known as RoI pooling) is to perform + is to perform max pooling on inputs of nonuniform sizes to obtain + fixed-size feature maps (e.g. 7*7). + The operator has three steps: + 1. Dividing each region proposal into equal-sized sections with + the pooled_width and pooled_height + 2. Finding the largest value in each section + 3. Copying these max values to the output buffer + + Args: + input (Variable): The input for ROI pooling. + rois (Variable): ROIs (Regions of Interest) to pool over. It should + be a 2-D one level LoTensor of shape [num_rois, 4]. + The layout is [x1, y1, x2, y2], where (x1, y1) + is the top left coordinates, and (x2, y2) is the + bottom right coordinates. The num_rois is the + total number of ROIs in this batch data. + pooled_height (integer): The pooled output height. Default: 1 + pooled_width (integer): The pooled output width. Default: 1 + spatial_scale (float): Multiplicative spatial scale factor. To + translate ROI coords from their input scale + to the scale used when pooling. Default: 1.0 + + Returns: + pool_out (Variable): The output is a 4-D tensor of the shape + (num_rois, channels, pooled_h, pooled_w). + + Examples: + .. code-block:: python + + pool_out = fluid.layers.roi_pool(input=x, rois=rois, 7, 7, 1.0) + """ + helper = LayerHelper('roi_pool', **locals()) + dtype = helper.input_dtype() + pool_out = helper.create_tmp_variable(dtype) + argmaxes = helper.create_tmp_variable(dtype='int32') + helper.append_op( + type="roi_pool", + inputs={"X": input, + "ROIs": rois}, + outputs={"Out": pool_out, + "Argmax": argmaxes}, + attrs={ + "pooled_height": pooled_height, + "pooled_width": pooled_width, + "spatial_scale": spatial_scale + }) + return pool_out + + +def dice_loss(input, label, epsilon=0.00001): + """ + **Dice loss Layer** + Dice loss for comparing the similarity of two batch of data, + usually is used for binary image segmentation i.e. labels are binary. + The dice loss can be defined as below equation: + + .. math:: + + dice\_loss &= 1 - \\frac{2 * intersection\_area}{total\_area} \\\\ + &= \\frac{(total\_area - intersection\_area) - intersection\_area}{total\_area} \\\\ + &= \\frac{(union\_area - intersection\_area)}{total\_area} + + + Args: + input (Variable): The predictions with rank>=2. The first dimension is batch size, + and the last dimension is class number. + label (Variable): The groud truth with the same rank with input. The first dimension + is batch size, and the last dimension is 1. + epsilon (float): The epsilon will be added to the numerator and denominator. + If both input and label are empty, it makes sure dice is 1. + Default: 0.00001 + + Returns: + dice_loss (Variable): The dice loss with shape [1]. + + Examples: + .. code-block:: python + + predictions = fluid.layers.softmax(x) + loss = fluid.layers.dice_loss(input=predictions, label=label, 2) + """ + label = one_hot(label, depth=input.shape[-1]) + reduce_dim = range(1, len(input.shape)) + inse = reduce_sum(input * label, dim=reduce_dim) + dice_denominator = reduce_sum( + input, dim=reduce_dim) + reduce_sum( + label, dim=reduce_dim) + dice_score = 1 - inse * 2 / (dice_denominator + epsilon) + return reduce_mean(dice_score) + + +def upsampling_bilinear2d(input, out_shape=None, scale=None, name=None): + """ + The mathematical meaning of upsampling_bilinear2d is also called + Bilinear interpolation. + Bilinear interpolation is an extension of linear interpolation for + interpolating functions of two variables (e.g. H-direction and + W-direction in this layer) on a rectilinear 2D grid. + + For details, please refer to Wikipedia: + https://en.wikipedia.org/wiki/Bilinear_interpolation + + Args: + input (Variable): The input tensor of bilinear interpolation, + This is a 4-D tensor of the shape + (num_batches, channels, in_h, in_w). + out_shape(list|tuple|None): Output shape of bilinear interpolation + layer, the shape is (out_h, out_w). + Default: None + scale(int|None): The multiplier for the input height or width. + At least one of out_shape or scale must be set. + And out_shape has a higher priority than scale. + Default: None + name(str|None): A name for this layer(optional). If set None, the layer + will be named automatically. + + Returns: + out (Variable): The output is a 4-D tensor of the shape + (num_batches, channls, out_h, out_w). + + Examples: + .. code-block:: python + + out = fluid.layers.bilinear_interp(input, out_shape=[12, 12]) + """ + if out_shape is None and scale is None: + raise ValueError("One of out_shape and scale must not be None") + helper = LayerHelper('bilinear_interp', **locals()) + dtype = helper.input_dtype() + + def _is_list_or_turple_(data): + return (isinstance(data, list) or isinstance(data, tuple)) + + if out_shape is not None: + if not (_is_list_or_turple_(out_shape) and len(out_shape) == 2): + raise ValueError('out_shape should be a list or tuple ', + 'with length 2, (out_h, out_w).') + out_shape = list(map(int, out_shape)) + out_h = out_shape[0] + out_w = out_shape[1] + else: + out_h = int(input.shape[2] * scale) + out_w = int(input.shape[3] * scale) + + out = helper.create_tmp_variable(dtype) + helper.append_op( + type="bilinear_interp", + inputs={"X": input}, + outputs={"Out": out}, + attrs={"out_h": out_h, + "out_w": out_w}) + return out + + +def random_crop(input, shape, seed=1): + helper = LayerHelper("random_crop", **locals()) + dtype = helper.input_dtype() + out = helper.create_tmp_variable(dtype) + if isinstance(seed, int): + seed_value = seed + seed = helper.create_tmp_variable(dtype="int64") + helper.append_op( + type="fill_constant", + inputs={}, + outputs={"Out": seed}, + attrs={ + "dtype": seed.dtype, + "shape": [1], + "value": float(seed_value), + "force_cpu": True + }) + elif not isinstance(seed, Variable): + raise ValueError("'seed' must be a Variable or an int.") + seed_out = helper.create_tmp_variable(dtype="int64") + helper.append_op( + type="random_crop", + inputs={"X": input, + "Seed": seed}, + outputs={"Out": out, + "SeedOut": seed_out}, + attrs={"shape": shape}) + return out diff --git a/python/paddle/fluid/layers/tensor.py b/python/paddle/fluid/layers/tensor.py index da066c34bdeba1f1b76f8d1cafd9244b2f7708fa..be34cc81a5d5ca0e781e5984b6c3eeaa4e25eb90 100644 --- a/python/paddle/fluid/layers/tensor.py +++ b/python/paddle/fluid/layers/tensor.py @@ -112,7 +112,7 @@ def cast(x, dtype): return out -def concat(input, axis=0): +def concat(input, axis=0, name=None): """ **Concat** @@ -122,6 +122,8 @@ def concat(input, axis=0): Args: input(list): List of tensors to be concatenated axis(int): Integer axis along which the tensors will be concatenated + name(str|None): A name for this layer(optional). If set None, the layer + will be named automatically. Returns: Variable: Output variable of the concatenation @@ -193,10 +195,7 @@ def assign(input, output): helper = LayerHelper('assign', **locals()) if isinstance(input, Variable): helper.append_op( - type='scale', - inputs={'X': [input]}, - outputs={'Out': [output]}, - attrs={'scale': 1.0}) + type='assign', inputs={'X': [input]}, outputs={'Out': [output]}) elif isinstance(input, numpy.ndarray): dtype = convert_np_dtype_to_dtype_(input.dtype) if dtype == VarDesc.VarType.FP32: diff --git a/python/paddle/fluid/lod_tensor.py b/python/paddle/fluid/lod_tensor.py new file mode 100644 index 0000000000000000000000000000000000000000..9946d0a4ff33b2f5040f6d2e31aa20fcf9c609a7 --- /dev/null +++ b/python/paddle/fluid/lod_tensor.py @@ -0,0 +1,189 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import core +import numpy as np + +__all__ = ['create_lod_tensor', 'create_random_int_lodtensor'] + + +def _validate_lod(lod, tensor_height=-1): + """Check whether the input length-based lod info is valid. + + There are several things to check: + 1. lod should be a list of lists. Empty list is fine. + 2. The length of each sublist (a lod level) should be at least one. + 3. Each element in each lod level should be an integer greater than 0. + 4. The sum of one lod level should be equal to the length of the next lod level. + 5. The sum of the last lod level should be equal to the tensor height. + Bypass this check if user does not provide tensor_height as input. + + Args: + lod: the length-based lod info, e.g., [[2, 3], [2, 1, 2, 3, 4]]. + tensor_height: the outermost dimension of the tensor with which the input + lod is associated with. + + Returns: + A boolean indicating whether the input lod is valid or not. + """ + assert isinstance(lod, list), "lod should be a list" + # Empty lod is fine + if len(lod) == 0: + return True + + lod_sum = [] + for level in lod: + assert isinstance(level, list), "each item in lod should be a list" + # Each level of lod should have at least one length info + if len(level) < 1: + return False + level_sum = 0 + for lod_len in level: + # Each length in a level should be > 0 + if lod_len <= 0: + return False + level_sum += lod_len + lod_sum.append(level_sum) + + for idx, val in enumerate(lod_sum[:-1]): + # Each level's sum should be equal to + # the number of items in the next level + if val != len(lod[idx + 1]): + return False + + if tensor_height == -1: + return True + else: + # Last level's sum should be equal to the tensor height + return lod_sum[-1] == tensor_height + + +def _convert_lod(lod): + """Convert a length-based lod to a offset-based lod. + + If the length-based lod is [[2, 3], [2, 1, 2, 3, 4]], + then the offset-based lod is [[0, 2, 5], [0, 2, 3, 5, 8, 12]]. + + Args: + lod: a length-based lod info. + + Returns: + A list of lists as the offset-based lod converted to from the input lod. + """ + new_lod = [] + for level in lod: + cur_len = 0 + new_level = [cur_len] + for lod_len in level: + cur_len += lod_len + new_level.append(cur_len) + new_lod.append(new_level) + return new_lod + + +def create_lod_tensor(data, lod, place): + """Create a lod tensor from a numpy array, a list, or an existing lod tensor. + + Create a lod tensor by doing the following: + 1. Check that the length-based input lod is valid. + 2. Convert the length-based lod to a offset-based LoD. + 3. Copy the data from a numpy array, a list or a existing lod tensor to + CPU or GPU device (based on input place). + 4. Set the level of detail (LoD) using the offset-based LoD. + + Use example: + Suppose we want LoDTensor to hold data for sequences of word, where each word is + represented by an integer. If we want to create a LoDTensor to represent two + sentences, one of 2 words, and one of 3 words. + + Then 'data' can be a numpy array of integers with shape (5, 1). + 'lod' will be [[2, 3]], indicating the length(# of words) in each sentence. + This length-based input lod [[2, 3]] will be converted to offset-based lod [[0, 2, 5]] + inside the function call. + + Please refer to + github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/lod_tensor.md + for more details regarding LoD. + + Args: + data: a numpy array or a LoDTensor or a list holding the data to be copied. + lod: a list of lists indicating the length-based LoD info specified by the user. + place: CPU or GPU place indicating where the data in the new LoDTensor will be stored. + + Returns: + A fluid LoDTensor object with tensor data and lod info. + """ + if isinstance(data, core.LoDTensor): + return create_lod_tensor(np.array(data), lod, place) + elif isinstance(data, list): + # When input data is a list, it only deal with the case where the base element + # is an index of shape [1] and dtype int64 (e.g., word id). Hence, the generated + # LoDTensor will be of shape [n, 1] and dtype int64, where `n` is the total number + # of words or other indexes in the sequence. + new_lod = [] + for seq in data: + new_lod.append(len(seq)) + assert [new_lod] == lod, "data and lod do not match" + flattened_data = np.concatenate(data, axis=0).astype("int64") + flattened_data = flattened_data.reshape([len(flattened_data), 1]) + return create_lod_tensor(flattened_data, lod, place) + elif isinstance(data, np.ndarray): + assert _validate_lod(lod, + data.shape[0]), "the provided lod info is invalid" + tensor = core.LoDTensor() + tensor.set(data, place) + tensor.set_lod(_convert_lod(lod)) + return tensor + else: + raise TypeError( + "data should be either a LoDTensor, a Numpy array or a list") + + +def create_random_int_lodtensor(lod, base_shape, place, low, high): + """Create a LoDTensor containing random integers. + + This function is frequently used in the book examples. So we revised it based on + the new create_lod_tensor API and put it here in the lod_tensor module to simplify + the code. + + The function does the following: + 1. Calculate the overall shape of the LoDTensor based on the length-based 'lod' input + and the shape of the basic element in 'base_shape'. + 2. Create a numpy array of this shape. + 3. Create the LoDTensor using create_lod_tensor API. + + Suppose we want LoDTensor to hold data for sequences of word, where each word is + represented by an integer. If we want to create a LoDTensor to represent two + sentences, one of 2 words, and one of 3 words. Then 'base_shape' is [1], input + length-based 'lod' is [[2, 3]]. Then the overall shape of the LoDTensor would be + [5, 1], holding 5 words for two sentences. + + Args: + data: a numpy array or a LoDTensor holding the data to be copied. + lod: a list of lists indicating the length-based LoD info specified by the user. + base_shape: the shape of the basic element to be held by the LoDTensor. + place: CPU or GPU place indicating where the data in the new LoDTensor will be stored. + low: the lower bound of the random integers. + high: the upper bound of the random integers. + + Returns: + A fluid LoDTensor object with tensor data and lod info. + """ + assert isinstance(base_shape, list), "base_shape should be a list" + converted_lod = _convert_lod(lod) + # append the total number of basic elements to the front of its shape + overall_shape = [converted_lod[-1][-1]] + base_shape + # the range of integer data elements is [low, high] + data = np.random.random_integers(low, high, overall_shape).astype("int64") + return create_lod_tensor(data, lod, place) diff --git a/python/paddle/fluid/metrics.py b/python/paddle/fluid/metrics.py index c618b02a768f2ca3e2b2914d8ee0134836d5c0d2..bb9c6fdc60089fc2b43573a6421a6f9781d2d4a8 100644 --- a/python/paddle/fluid/metrics.py +++ b/python/paddle/fluid/metrics.py @@ -251,7 +251,7 @@ class EditDistance(MetricBase): self.instance_error += seq_num - seq_right_count self.total_distance += total_distance - def eval(): + def eval(self): if self.seq_num == 0: raise ValueError( "There is no data in EditDistance Metric. Please check layers.edit_distance output has been added to EditDistance." @@ -280,6 +280,7 @@ class DetectionMAP(MetricBase): super(DetectionMAP, self).__init__(name) # the current map value self.value = .0 + self.weight = .0 def update(self, value, weight): if not _is_number_or_matrix_(value): @@ -340,8 +341,8 @@ class Auc(MetricBase): raise ValueError("The 'predictions' must be a numpy ndarray.") kepsilon = 1e-7 # to account for floating point imprecisions - thresholds = [(i + 1) * 1.0 / (num_thresholds - 1) - for i in range(num_thresholds - 2)] + thresholds = [(i + 1) * 1.0 / (self._num_thresholds - 1) + for i in range(self._num_thresholds - 2)] thresholds = [0.0 - kepsilon] + thresholds + [1.0 + kepsilon] # caculate TP, FN, TN, FP count @@ -358,19 +359,20 @@ class Auc(MetricBase): fp += 1 else: tn += 1 - tp_list[idx_thresh] += tp - fn_list[idx_thresh] += fn - tn_list[idx_thresh] += tn - fp_list[idx_thresh] += fp + self.tp_list[idx_thresh] += tp + self.fn_list[idx_thresh] += fn + self.tn_list[idx_thresh] += tn + self.fp_list[idx_thresh] += fp def eval(self): epsilon = self._epsilon num_thresholds = self._num_thresholds - tpr = (tp_list.astype("float32") + epsilon) / ( - tp_list + fn_list + epsilon) - fpr = fp_list.astype("float32") / (fp_list + tn_list + epsilon) - rec = (tp_list.astype("float32") + epsilon) / ( - tp_list + fp_list + epsilon) + tpr = (self.tp_list.astype("float32") + epsilon) / ( + self.tp_list + self.fn_list + epsilon) + fpr = self.fp_list.astype("float32") / ( + self.fp_list + self.tn_list + epsilon) + rec = (self.tp_list.astype("float32") + epsilon) / ( + self.tp_list + self.fp_list + epsilon) x = fpr[:num_thresholds - 1] - fpr[1:] y = (tpr[:num_thresholds - 1] + tpr[1:]) / 2.0 diff --git a/python/paddle/fluid/optimizer.py b/python/paddle/fluid/optimizer.py index 9ae43b3e93e4b7d337097a25379720c18dfd331c..115362c6bf33018342699a442c688e7356f3c206 100644 --- a/python/paddle/fluid/optimizer.py +++ b/python/paddle/fluid/optimizer.py @@ -28,7 +28,8 @@ from contextlib import contextmanager __all__ = [ 'SGD', 'Momentum', 'Adagrad', 'Adam', 'Adamax', 'DecayedAdagrad', 'SGDOptimizer', 'MomentumOptimizer', 'AdagradOptimizer', 'AdamOptimizer', - 'AdamaxOptimizer', 'DecayedAdagradOptimizer', 'Adadelta', 'ModelAverage' + 'AdamaxOptimizer', 'DecayedAdagradOptimizer', 'RMSPropOptimizer', + 'Adadelta', 'ModelAverage', 'Optimizer' ] @@ -46,6 +47,8 @@ class Optimizer(object): raise TypeError("learning rate should be float or Variable") self.regularization = regularization self._learning_rate = learning_rate + # the learning rate type should be inferenced from loss + self._dtype = None # each program should have a independent learning rate # program -> Variable(learning_rate) self._learning_rate_map = dict() @@ -76,7 +79,7 @@ class Optimizer(object): name=unique_name.generate("learning_rate"), shape=[1], value=float(self._learning_rate), - dtype='float32', + dtype='float32' if self._dtype == None else self._dtype, persistable=True) def global_learning_rate(self, program=None): @@ -199,6 +202,7 @@ class Optimizer(object): # Create any accumulators program = loss.block.program + self._dtype = loss.dtype with program_guard(program, startup_program): global_block = framework.default_main_program().global_block() start = len(global_block.ops) @@ -209,11 +213,13 @@ class Optimizer(object): optimize_ops = [] for param_and_grad in parameters_and_grads: - if param_and_grad[0].trainable is True and param_and_grad[ - 1] is not None: - optimize_op = self._append_optimize_op(loss.block, - param_and_grad) - optimize_ops.append(optimize_op) + with param_and_grad[0].block.program.optimized_guard( + param_and_grad[0]): + if param_and_grad[0].trainable is True and param_and_grad[ + 1] is not None: + optimize_op = self._append_optimize_op(loss.block, + param_and_grad) + optimize_ops.append(optimize_op) # Get custom finish ops for subclasses # FIXME: Need to fix this once we figure out how to handle dependencies @@ -390,7 +396,7 @@ class AdamOptimizer(Optimizer): beta_shape = [1] self._beta1_pow_acc = self.helper.create_global_variable( name=unique_name.generate('beta1_pow_acc'), - dtype='float32', + dtype='float32' if self._dtype == None else self._dtype, shape=beta_shape, lod_level=0, persistable=True) @@ -399,7 +405,7 @@ class AdamOptimizer(Optimizer): self._beta2_pow_acc = self.helper.create_global_variable( name=unique_name.generate('beta2_pow_acc'), - dtype='float32', + dtype='float32' if self._dtype == None else self._dtype, shape=beta_shape, lod_level=0, persistable=True) @@ -492,7 +498,7 @@ class AdamaxOptimizer(Optimizer): beta_shape = [1] self._beta1_pow_acc = self.helper.create_global_variable( name=unique_name.generate('beta1_pow_acc'), - dtype='float32', + dtype='float32' if self._dtype == None else self._dtype, shape=beta_shape, lod_level=0, persistable=True) @@ -899,8 +905,10 @@ class ModelAverage(Optimizer): # param = (sum_1 + sum_2 + sum_3) / (num_accumulates + old_num_accumulates) tmp = layers.sum(x=[num_accumulates, old_num_accumulates]) sum = layers.sum(x=[sum_1, sum_2, sum_3]) - tmp = layers.cast(x=tmp, dtype='float32') - sum = layers.cast(x=sum, dtype='float32') + tmp = layers.cast( + x=tmp, dtype='float32' if self._dtype == None else self._dtype) + sum = layers.cast( + x=sum, dtype='float32' if self._dtype == None else self._dtype) layers.elementwise_div(x=sum, y=tmp, out=param) def _add_average_restore_op(self, block, param_grad): diff --git a/python/paddle/fluid/parallel_executor.py b/python/paddle/fluid/parallel_executor.py index 4adbb2ea99b58c78c5c08c7ac8a556ca1de1615e..3117dfe00c7a3df1035c439dc31b81e67781d0cc 100644 --- a/python/paddle/fluid/parallel_executor.py +++ b/python/paddle/fluid/parallel_executor.py @@ -19,7 +19,10 @@ import executor import warnings import sys -__all__ = ['ParallelExecutor'] +__all__ = ['ParallelExecutor', 'ExecutionStrategy', 'BuildStrategy'] + +ExecutionStrategy = core.ParallelExecutor.ExecutionStrategy +BuildStrategy = core.ParallelExecutor.BuildStrategy class ParallelExecutor(object): @@ -27,10 +30,12 @@ class ParallelExecutor(object): use_cuda, loss_name=None, main_program=None, - num_threads=None, - allow_op_delay=False, share_vars_from=None, - customize_loss_grad=False): + exec_strategy=None, + build_strategy=None, + num_trainers=1, + trainer_id=0, + **kwargs): """ ParallelExecutor can run program in parallel. @@ -39,13 +44,13 @@ class ParallelExecutor(object): loss_name(str, default None): The loss name must set in training. main_program(Program, default None): The program that need to run, if not provided, then default_main_program will be used. - num_threads(int, default None): How many threads are used for - training. - allow_op_delay(bool, default False): Whether to delay and buffer - some operators together for scheduling or not, which may - improve performance in some cases, defalut False. share_vars_from(ParallelExecutor, default None): If provied, it will share variables from the specified ParallelExecutor. + num_trainers(int, default 1): If greater than 1, NCCL will be + initialized with multpile rank of nodes, each node should have + same number of GPUs. Distributed training will be enabled then. + trainer_id(int, default 0): Must use together with num_trainers. + trainer_id is the "rank" of current node starts from 0. Returns: A ParallelExecutor object. @@ -67,6 +72,25 @@ class ParallelExecutor(object): train_loss, = train_exe.run([loss.name], feed=feed_dict) test_loss, = test_exe.run([loss.name], feed=feed_dict) """ + if len(kwargs) != 0: + err_msg = "" + for key in kwargs: + if key in dir(ExecutionStrategy): + err_msg += \ + "Setting {0} by constructor is deprecated. Use " \ + "strategy=ExecutionStrategy(); strategy.{0}=xxx; " \ + "pe=ParallelExecutor(exec_strategy=strategy) " \ + "instead.\n ".format(key) + elif key in dir(BuildStrategy): + err_msg += \ + "Setting {0} by constructor is deprecated. Use " \ + "strategy=BuildStrategy(); See help(" \ + "paddle.fluid.ParallelExecutor.BuildStrategy) \n".format( + key) + else: + err_msg += "Setting {0} by constructor is deprecated. Use strategy.\n".format( + key) + raise ValueError(err_msg) self._places = [] self._act_places = [] @@ -84,15 +108,25 @@ class ParallelExecutor(object): self._places.append(p) assert self._places, "no place for execution" - if num_threads is None: + if exec_strategy is None: + exec_strategy = ExecutionStrategy() + if use_cuda: + exec_strategy.use_event = True + else: + exec_strategy.use_event = False + + if exec_strategy.num_threads == 0: if use_cuda: # Experiments on se-resnext shows that too many threads hurt # performance. Worth tunning for other models in the future. - num_threads = len(self._places) + exec_strategy.num_threads = len(self._places) * 2 else: - num_threads = min( + exec_strategy.num_threads = min( len(self._places) * 2, multiprocessing.cpu_count()) + if build_strategy is None: + build_strategy = BuildStrategy() + main = main_program main = main if main else framework.default_main_program() scope = executor.global_scope() @@ -111,20 +145,14 @@ class ParallelExecutor(object): ] self.executor = core.ParallelExecutor( - num_threads, - True if use_cuda else False, # use_event self._places, set([ p.name for p in main.global_block().iter_parameters() if not p.stop_gradient ]), - set(self.persistable_vars), - main.desc, - loss_name if loss_name else '', - scope, - local_scopes, - allow_op_delay, - customize_loss_grad) + set(self.persistable_vars), main.desc, loss_name + if loss_name else '', scope, local_scopes, exec_strategy, + build_strategy, num_trainers, trainer_id) self.scope = scope def run(self, fetch_list, feed=None, feed_dict=None): diff --git a/python/paddle/fluid/profiler.py b/python/paddle/fluid/profiler.py index 04fd05cc33cff3d720be75923d4af3767942669f..e2bd1d4c9a1ea5ddc0dfd19c769dcb40bfd6d04c 100644 --- a/python/paddle/fluid/profiler.py +++ b/python/paddle/fluid/profiler.py @@ -16,7 +16,10 @@ import core from contextlib import contextmanager import os -__all__ = ['cuda_profiler', 'reset_profiler', 'profiler'] +__all__ = [ + 'cuda_profiler', 'reset_profiler', 'profiler', 'start_profiler', + 'stop_profiler' +] NVPROF_CONFIG = [ "gpustarttimestamp", @@ -72,20 +75,31 @@ def reset_profiler(): core.reset_profiler() -@contextmanager -def profiler(state, sorted_key=None, profile_path='/tmp/profile'): - """The profiler interface. - Different from cuda_profiler, this profiler can be used to profile both CPU - and GPU program. By defalut, it records the CPU and GPU operator kernels, - if you want to profile other program, you can refer the profiling tutorial - to add more records. +def start_profiler(state): + """Enable the profiler. + + Args: + state (string) : The profiling state, which should be 'CPU', 'GPU' + or 'All'. 'CPU' means only profile CPU. 'GPU' means profiling + GPU as well. 'All' also generates timeline. + """ + if core.is_profiler_enabled(): + return + if state not in ['CPU', 'GPU', "All"]: + raise ValueError("The state must be 'CPU' or 'GPU' or 'All'.") + if state == "GPU": + prof_state = core.ProfilerState.kCUDA + elif state == "CPU": + prof_state = core.ProfilerState.kCPU + else: + prof_state = core.ProfilerState.kAll + core.enable_profiler(prof_state) + + +def stop_profiler(sorted_key=None, profile_path='/tmp/profile'): + """Stop the profiler. Args: - state (string) : The profiling state, which should be 'CPU' or 'GPU', - telling the profiler to use CPU timer or GPU timer for profiling. - Although users may have already specified the execution place - (CPUPlace/CUDAPlace) in the begining, for flexibility the profiler - would not inherit this place. sorted_key (string) : If None, the profiling results will be printed in the order of first end time of events. Otherwise, the profiling results will be sorted by the this flag. This flag should be one @@ -98,17 +112,8 @@ def profiler(state, sorted_key=None, profile_path='/tmp/profile'): profile_path (string) : If state == 'All', it will write a profile proto output file. """ - if state not in ['CPU', 'GPU', "All"]: - raise ValueError("The state must be 'CPU' or 'GPU' or 'All'.") - if state == "GPU": - prof_state = core.ProfilerState.kCUDA - elif state == "CPU": - prof_state = core.ProfilerState.kCPU - else: - prof_state = core.ProfilerState.kAll - core.enable_profiler(prof_state) - yield - + if not core.is_profiler_enabled(): + return sorted_key = 'default' if sorted_key is None else sorted_key if sorted_key not in ['default', 'calls', 'total', 'max', 'min', 'ave']: raise ValueError("The sorted_key must be None or in 'calls', 'total', " @@ -124,3 +129,34 @@ def profiler(state, sorted_key=None, profile_path='/tmp/profile'): # TODO(qingqing) : redirect C++ ostream to Python stream. # with core.ostream_redirect(stdout=True, stderr=True): core.disable_profiler(key_map[sorted_key], profile_path) + + +@contextmanager +def profiler(state, sorted_key=None, profile_path='/tmp/profile'): + """The profiler interface. + Different from cuda_profiler, this profiler can be used to profile both CPU + and GPU program. By defalut, it records the CPU and GPU operator kernels, + if you want to profile other program, you can refer the profiling tutorial + to add more records. + + Args: + state (string) : The profiling state, which should be 'CPU' or 'GPU', + telling the profiler to use CPU timer or GPU timer for profiling. + Although users may have already specified the execution place + (CPUPlace/CUDAPlace) in the begining, for flexibility the profiler + would not inherit this place. + sorted_key (string) : If None, the profiling results will be printed + in the order of first end time of events. Otherwise, the profiling + results will be sorted by the this flag. This flag should be one + of 'calls', 'total', 'max', 'min' or 'ave'. + The `calls` means sorting by the number of calls. + The `total` means sorting by the total execution time. + The `max` means sorting by the maximum execution time. + The `min` means sorting by the minimum execution time. + The `ave` means sorting by the average execution time. + profile_path (string) : If state == 'All', it will write a profile + proto output file. + """ + start_profiler(state) + yield + stop_profiler(sorted_key, profile_path) diff --git a/python/paddle/fluid/regularizer.py b/python/paddle/fluid/regularizer.py index c006bd9a66ddb422b7d80d2ca87aa7f56a6485db..c4d6829599616cb3ea7791a189e7070974de6ae3 100644 --- a/python/paddle/fluid/regularizer.py +++ b/python/paddle/fluid/regularizer.py @@ -43,31 +43,32 @@ def append_regularization_ops(parameters_and_grads, regularization=None): """ params_and_grads = [] for param, grad in parameters_and_grads: - # If no gradient then we don't need to do anything - if grad is None: + with param.block.program.optimized_guard(param): + # If no gradient then we don't need to do anything + if grad is None: + params_and_grads.append((param, grad)) + continue + + regularization_term = None + if param.regularizer is not None: + # Add variable for regularization term in grad block + regularization_term = param.regularizer(param, grad, grad.block) + elif regularization is not None: + regularization_term = regularization(param, grad, grad.block) + + # If no regularization specified, then we don't need to do anything + if regularization_term is None: + params_and_grads.append((param, grad)) + continue + + assert grad.shape == regularization_term.shape + + grad.block.append_op( + type='elementwise_add', + inputs={"X": grad, + "Y": regularization_term}, + outputs={"Out": grad}) params_and_grads.append((param, grad)) - continue - - regularization_term = None - if param.regularizer is not None: - # Add variable for regularization term in grad block - regularization_term = param.regularizer(param, grad, grad.block) - elif regularization is not None: - regularization_term = regularization(param, grad, grad.block) - - # If no regularization specified, then we don't need to do anything - if regularization_term is None: - params_and_grads.append((param, grad)) - continue - - assert grad.shape == regularization_term.shape - - grad.block.append_op( - type='elementwise_add', - inputs={"X": grad, - "Y": regularization_term}, - outputs={"Out": grad}) - params_and_grads.append((param, grad)) return params_and_grads diff --git a/python/paddle/fluid/tests/book/CMakeLists.txt b/python/paddle/fluid/tests/book/CMakeLists.txt index 673c965b662a022739f8d489c331f4de9455a926..ee734f3c782adb5196a03aca5718377009a5b4e7 100644 --- a/python/paddle/fluid/tests/book/CMakeLists.txt +++ b/python/paddle/fluid/tests/book/CMakeLists.txt @@ -5,3 +5,5 @@ string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}") foreach(src ${TEST_OPS}) py_test(${src} SRCS ${src}.py) endforeach() + +add_subdirectory(high-level-api) diff --git a/python/paddle/fluid/tests/book/high-level-api/CMakeLists.txt b/python/paddle/fluid/tests/book/high-level-api/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..efa5ee2d06af3d31e7d84122dd7eea37d6dcf3a3 --- /dev/null +++ b/python/paddle/fluid/tests/book/high-level-api/CMakeLists.txt @@ -0,0 +1,16 @@ +file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py") +string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}") + +# default test +foreach(src ${TEST_OPS}) + py_test(${src} SRCS ${src}.py) +endforeach() + +add_subdirectory(fit_a_line) +add_subdirectory(recognize_digits) +add_subdirectory(image_classification) +add_subdirectory(understand_sentiment) +add_subdirectory(label_semantic_roles) +add_subdirectory(word2vec) +add_subdirectory(recommender_system) +add_subdirectory(machine_translation) diff --git a/python/paddle/fluid/tests/book/high-level-api/fit_a_line/CMakeLists.txt b/python/paddle/fluid/tests/book/high-level-api/fit_a_line/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..673c965b662a022739f8d489c331f4de9455a926 --- /dev/null +++ b/python/paddle/fluid/tests/book/high-level-api/fit_a_line/CMakeLists.txt @@ -0,0 +1,7 @@ +file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py") +string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}") + +# default test +foreach(src ${TEST_OPS}) + py_test(${src} SRCS ${src}.py) +endforeach() diff --git a/python/paddle/fluid/tests/book/high-level-api/fit_a_line/test_fit_a_line.py b/python/paddle/fluid/tests/book/high-level-api/fit_a_line/test_fit_a_line.py new file mode 100644 index 0000000000000000000000000000000000000000..de3906fc6a005181b0ab04a846eb2e7ce14004c2 --- /dev/null +++ b/python/paddle/fluid/tests/book/high-level-api/fit_a_line/test_fit_a_line.py @@ -0,0 +1,131 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import paddle +import paddle.fluid as fluid +import contextlib +import numpy +import unittest + +# train reader +BATCH_SIZE = 20 + +train_reader = paddle.batch( + paddle.reader.shuffle( + paddle.dataset.uci_housing.train(), buf_size=500), + batch_size=BATCH_SIZE) + +test_reader = paddle.batch( + paddle.reader.shuffle( + paddle.dataset.uci_housing.test(), buf_size=500), + batch_size=BATCH_SIZE) + + +def inference_program(): + x = fluid.layers.data(name='x', shape=[13], dtype='float32') + y_predict = fluid.layers.fc(input=x, size=1, act=None) + return y_predict + + +def linear(): + y = fluid.layers.data(name='y', shape=[1], dtype='float32') + y_predict = inference_program() + + loss = fluid.layers.square_error_cost(input=y_predict, label=y) + avg_loss = fluid.layers.mean(loss) + + return avg_loss + + +def train(use_cuda, train_program, params_dirname): + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + + trainer = fluid.Trainer( + train_func=train_program, + place=place, + optimizer=fluid.optimizer.SGD(learning_rate=0.001)) + + def event_handler(event): + if isinstance(event, fluid.EndStepEvent): + if event.step == 10: + test_metrics = trainer.test( + reader=test_reader, feed_order=['x', 'y']) + print test_metrics + ''' + ... + ['25.768919467926025'] + ['15.343549569447836'] + ... + ''' + if params_dirname is not None: + trainer.save_params(params_dirname) + trainer.stop() + + trainer.train( + reader=train_reader, + num_epochs=100, + event_handler=event_handler, + feed_order=['x', 'y']) + + +# infer +def infer(use_cuda, inference_program, params_dirname=None): + if params_dirname is None: + return + + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + inferencer = fluid.Inferencer( + infer_func=inference_program, param_path=params_dirname, place=place) + + batch_size = 10 + tensor_x = numpy.random.uniform(0, 10, [batch_size, 13]).astype("float32") + + results = inferencer.infer({'x': tensor_x}) + print("infer results: ", results[0]) + + +def main(use_cuda): + if use_cuda and not fluid.core.is_compiled_with_cuda(): + return + + # Directory for saving the trained model + params_dirname = "fit_a_line.inference.model" + + train(use_cuda, linear, params_dirname) + infer(use_cuda, inference_program, params_dirname) + + +class TestFitALine(unittest.TestCase): + def test_cpu(self): + with self.program_scope_guard(): + with fluid.unique_name.guard(): + main(use_cuda=False) + + def test_cuda(self): + with self.program_scope_guard(): + with fluid.unique_name.guard(): + main(use_cuda=True) + + @contextlib.contextmanager + def program_scope_guard(self): + prog = fluid.Program() + startup_prog = fluid.Program() + scope = fluid.core.Scope() + with fluid.scope_guard(scope): + with fluid.program_guard(prog, startup_prog): + yield + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/book/high-level-api/image_classification/CMakeLists.txt b/python/paddle/fluid/tests/book/high-level-api/image_classification/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..673c965b662a022739f8d489c331f4de9455a926 --- /dev/null +++ b/python/paddle/fluid/tests/book/high-level-api/image_classification/CMakeLists.txt @@ -0,0 +1,7 @@ +file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py") +string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}") + +# default test +foreach(src ${TEST_OPS}) + py_test(${src} SRCS ${src}.py) +endforeach() diff --git a/python/paddle/fluid/tests/book/high-level-api/image_classification/cifar10_small_test_set.py b/python/paddle/fluid/tests/book/high-level-api/image_classification/cifar10_small_test_set.py new file mode 100644 index 0000000000000000000000000000000000000000..7fed6d914f75b690e34411aa154359c93b6ca989 --- /dev/null +++ b/python/paddle/fluid/tests/book/high-level-api/image_classification/cifar10_small_test_set.py @@ -0,0 +1,82 @@ +# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +CIFAR dataset. + +This module will download dataset from +https://www.cs.toronto.edu/~kriz/cifar.html and parse train/test set into +paddle reader creators. + +The CIFAR-10 dataset consists of 60000 32x32 colour images in 10 classes, +with 6000 images per class. There are 50000 training images and 10000 test +images. + +The CIFAR-100 dataset is just like the CIFAR-10, except it has 100 classes +containing 600 images each. There are 500 training images and 100 testing +images per class. + +""" + +import cPickle +import itertools +import numpy +import paddle.v2.dataset.common +import tarfile + +__all__ = ['train10'] + +URL_PREFIX = 'https://www.cs.toronto.edu/~kriz/' +CIFAR10_URL = URL_PREFIX + 'cifar-10-python.tar.gz' +CIFAR10_MD5 = 'c58f30108f718f92721af3b95e74349a' + + +def reader_creator(filename, sub_name, batch_size=None): + def read_batch(batch): + data = batch['data'] + labels = batch.get('labels', batch.get('fine_labels', None)) + assert labels is not None + for sample, label in itertools.izip(data, labels): + yield (sample / 255.0).astype(numpy.float32), int(label) + + def reader(): + with tarfile.open(filename, mode='r') as f: + names = (each_item.name for each_item in f + if sub_name in each_item.name) + + batch_count = 0 + for name in names: + batch = cPickle.load(f.extractfile(name)) + for item in read_batch(batch): + if isinstance(batch_size, int) and batch_count > batch_size: + break + batch_count += 1 + yield item + + return reader + + +def train10(batch_size=None): + """ + CIFAR-10 training set creator. + + It returns a reader creator, each sample in the reader is image pixels in + [0, 1] and label in [0, 9]. + + :return: Training reader creator + :rtype: callable + """ + return reader_creator( + paddle.v2.dataset.common.download(CIFAR10_URL, 'cifar', CIFAR10_MD5), + 'data_batch', + batch_size=batch_size) diff --git a/python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_resnet.py b/python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_resnet.py new file mode 100644 index 0000000000000000000000000000000000000000..63dc1b6ce30974ede22a3f7772b76bf207bbae39 --- /dev/null +++ b/python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_resnet.py @@ -0,0 +1,157 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import print_function + +import paddle +import paddle.fluid as fluid +import numpy +import cifar10_small_test_set + + +def resnet_cifar10(input, depth=32): + def conv_bn_layer(input, + ch_out, + filter_size, + stride, + padding, + act='relu', + bias_attr=False): + tmp = fluid.layers.conv2d( + input=input, + filter_size=filter_size, + num_filters=ch_out, + stride=stride, + padding=padding, + act=None, + bias_attr=bias_attr) + return fluid.layers.batch_norm(input=tmp, act=act) + + def shortcut(input, ch_in, ch_out, stride): + if ch_in != ch_out: + return conv_bn_layer(input, ch_out, 1, stride, 0, None) + else: + return input + + def basicblock(input, ch_in, ch_out, stride): + tmp = conv_bn_layer(input, ch_out, 3, stride, 1) + tmp = conv_bn_layer(tmp, ch_out, 3, 1, 1, act=None, bias_attr=True) + short = shortcut(input, ch_in, ch_out, stride) + return fluid.layers.elementwise_add(x=tmp, y=short, act='relu') + + def layer_warp(block_func, input, ch_in, ch_out, count, stride): + tmp = block_func(input, ch_in, ch_out, stride) + for i in range(1, count): + tmp = block_func(tmp, ch_out, ch_out, 1) + return tmp + + assert (depth - 2) % 6 == 0 + n = (depth - 2) / 6 + conv1 = conv_bn_layer( + input=input, ch_out=16, filter_size=3, stride=1, padding=1) + res1 = layer_warp(basicblock, conv1, 16, 16, n, 1) + res2 = layer_warp(basicblock, res1, 16, 32, n, 2) + res3 = layer_warp(basicblock, res2, 32, 64, n, 2) + pool = fluid.layers.pool2d( + input=res3, pool_size=8, pool_type='avg', pool_stride=1) + predict = fluid.layers.fc(input=pool, size=10, act='softmax') + return predict + + +def inference_network(): + data_shape = [3, 32, 32] + images = fluid.layers.data(name='pixel', shape=data_shape, dtype='float32') + predict = resnet_cifar10(images, 32) + return predict + + +def train_network(): + predict = inference_network() + label = fluid.layers.data(name='label', shape=[1], dtype='int64') + cost = fluid.layers.cross_entropy(input=predict, label=label) + avg_cost = fluid.layers.mean(cost) + accuracy = fluid.layers.accuracy(input=predict, label=label) + return [avg_cost, accuracy] + + +def train(use_cuda, train_program, params_dirname): + BATCH_SIZE = 128 + EPOCH_NUM = 1 + + train_reader = paddle.batch( + paddle.reader.shuffle( + cifar10_small_test_set.train10(batch_size=10), buf_size=128 * 10), + batch_size=BATCH_SIZE) + + test_reader = paddle.batch( + paddle.dataset.cifar.test10(), batch_size=BATCH_SIZE) + + def event_handler(event): + if isinstance(event, fluid.EndStepEvent): + avg_cost, accuracy = trainer.test( + reader=test_reader, feed_order=['pixel', 'label']) + + print('Loss {0:2.2}, Acc {1:2.2}'.format(avg_cost, accuracy)) + + if accuracy > 0.01: # Low threshold for speeding up CI + if params_dirname is not None: + trainer.save_params(params_dirname) + return + + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + trainer = fluid.Trainer( + train_func=train_program, + optimizer=fluid.optimizer.Adam(learning_rate=0.001), + place=place) + + trainer.train( + reader=train_reader, + num_epochs=EPOCH_NUM, + event_handler=event_handler, + feed_order=['pixel', 'label']) + + +def infer(use_cuda, inference_program, params_dirname=None): + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + inferencer = fluid.Inferencer( + infer_func=inference_program, param_path=params_dirname, place=place) + + # The input's dimension of conv should be 4-D or 5-D. + # Use normilized image pixels as input data, which should be in the range + # [0, 1.0]. + tensor_img = numpy.random.rand(1, 3, 32, 32).astype("float32") + results = inferencer.infer({'pixel': tensor_img}) + + print("infer results: ", results) + + +def main(use_cuda): + if use_cuda and not fluid.core.is_compiled_with_cuda(): + return + save_path = "image_classification_resnet.inference.model" + + train( + use_cuda=use_cuda, + train_program=train_network, + params_dirname=save_path) + + infer( + use_cuda=use_cuda, + inference_program=inference_network, + params_dirname=save_path) + + +if __name__ == '__main__': + for use_cuda in (False, True): + main(use_cuda=use_cuda) diff --git a/python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_vgg.py b/python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_vgg.py new file mode 100644 index 0000000000000000000000000000000000000000..0bf8f265a1c1b11364ecfa11061af183ce20d51e --- /dev/null +++ b/python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_vgg.py @@ -0,0 +1,134 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import print_function + +import paddle +import paddle.fluid as fluid +import numpy +import cifar10_small_test_set + + +def vgg16_bn_drop(input): + def conv_block(input, num_filter, groups, dropouts): + return fluid.nets.img_conv_group( + input=input, + pool_size=2, + pool_stride=2, + conv_num_filter=[num_filter] * groups, + conv_filter_size=3, + conv_act='relu', + conv_with_batchnorm=True, + conv_batchnorm_drop_rate=dropouts, + pool_type='max') + + conv1 = conv_block(input, 64, 2, [0.3, 0]) + conv2 = conv_block(conv1, 128, 2, [0.4, 0]) + conv3 = conv_block(conv2, 256, 3, [0.4, 0.4, 0]) + conv4 = conv_block(conv3, 512, 3, [0.4, 0.4, 0]) + conv5 = conv_block(conv4, 512, 3, [0.4, 0.4, 0]) + + drop = fluid.layers.dropout(x=conv5, dropout_prob=0.5) + fc1 = fluid.layers.fc(input=drop, size=4096, act=None) + bn = fluid.layers.batch_norm(input=fc1, act='relu') + drop2 = fluid.layers.dropout(x=bn, dropout_prob=0.5) + fc2 = fluid.layers.fc(input=drop2, size=4096, act=None) + predict = fluid.layers.fc(input=fc2, size=10, act='softmax') + return predict + + +def inference_network(): + data_shape = [3, 32, 32] + images = fluid.layers.data(name='pixel', shape=data_shape, dtype='float32') + predict = vgg16_bn_drop(images) + return predict + + +def train_network(): + predict = inference_network() + label = fluid.layers.data(name='label', shape=[1], dtype='int64') + cost = fluid.layers.cross_entropy(input=predict, label=label) + avg_cost = fluid.layers.mean(cost) + accuracy = fluid.layers.accuracy(input=predict, label=label) + return [avg_cost, accuracy] + + +def train(use_cuda, train_program, params_dirname): + BATCH_SIZE = 128 + train_reader = paddle.batch( + paddle.reader.shuffle( + cifar10_small_test_set.train10(batch_size=10), buf_size=128 * 10), + batch_size=BATCH_SIZE) + + test_reader = paddle.batch( + paddle.dataset.cifar.test10(), batch_size=BATCH_SIZE) + + def event_handler(event): + if isinstance(event, fluid.EndStepEvent): + avg_cost, accuracy = trainer.test( + reader=test_reader, feed_order=['pixel', 'label']) + + print('Loss {0:2.2}, Acc {1:2.2}'.format(avg_cost, accuracy)) + + if accuracy > 0.01: # Low threshold for speeding up CI + if params_dirname is not None: + trainer.save_params(params_dirname) + return + + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + trainer = fluid.Trainer( + train_func=train_program, + place=place, + optimizer=fluid.optimizer.Adam(learning_rate=0.001)) + + trainer.train( + reader=train_reader, + num_epochs=1, + event_handler=event_handler, + feed_order=['pixel', 'label']) + + +def infer(use_cuda, inference_program, params_dirname=None): + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + inferencer = fluid.Inferencer( + infer_func=inference_program, param_path=params_dirname, place=place) + + # The input's dimension of conv should be 4-D or 5-D. + # Use normilized image pixels as input data, which should be in the range + # [0, 1.0]. + tensor_img = numpy.random.rand(1, 3, 32, 32).astype("float32") + results = inferencer.infer({'pixel': tensor_img}) + + print("infer results: ", results) + + +def main(use_cuda): + if use_cuda and not fluid.core.is_compiled_with_cuda(): + return + save_path = "image_classification_vgg.inference.model" + + train( + use_cuda=use_cuda, + train_program=train_network, + params_dirname=save_path) + + infer( + use_cuda=use_cuda, + inference_program=inference_network, + params_dirname=save_path) + + +if __name__ == '__main__': + for use_cuda in (False, True): + main(use_cuda=use_cuda) diff --git a/python/paddle/fluid/tests/book/high-level-api/label_semantic_roles/CMakeLists.txt b/python/paddle/fluid/tests/book/high-level-api/label_semantic_roles/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..673c965b662a022739f8d489c331f4de9455a926 --- /dev/null +++ b/python/paddle/fluid/tests/book/high-level-api/label_semantic_roles/CMakeLists.txt @@ -0,0 +1,7 @@ +file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py") +string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}") + +# default test +foreach(src ${TEST_OPS}) + py_test(${src} SRCS ${src}.py) +endforeach() diff --git a/python/paddle/fluid/tests/book/high-level-api/label_semantic_roles/test_label_semantic_roles_newapi.py b/python/paddle/fluid/tests/book/high-level-api/label_semantic_roles/test_label_semantic_roles_newapi.py new file mode 100755 index 0000000000000000000000000000000000000000..8cce398ff33695dc15ae6fb01a887194596af001 --- /dev/null +++ b/python/paddle/fluid/tests/book/high-level-api/label_semantic_roles/test_label_semantic_roles_newapi.py @@ -0,0 +1,261 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import print_function + +import paddle +import paddle.fluid as fluid +import numpy as np + +WORD_DICT, VERB_DICT, LABEL_DICT = paddle.dataset.conll05.get_dict() +WORD_DICT_LEN = len(WORD_DICT) +LABEL_DICT_LEN = len(LABEL_DICT) +PRED_DICT_LEN = len(VERB_DICT) +MARK_DICT_LEN = 2 +IS_SPARSE = True +BATCH_SIZE = 10 +EMBEDDING_NAME = 'emb' + + +def lstm_net(): + WORD_DIM = 32 + MARK_DIM = 5 + HIDDEN_DIM = 512 + DEPTH = 8 + + # Data definitions + word = fluid.layers.data( + name='word_data', shape=[1], dtype='int64', lod_level=1) + predicate = fluid.layers.data( + name='verb_data', shape=[1], dtype='int64', lod_level=1) + ctx_n2 = fluid.layers.data( + name='ctx_n2_data', shape=[1], dtype='int64', lod_level=1) + ctx_n1 = fluid.layers.data( + name='ctx_n1_data', shape=[1], dtype='int64', lod_level=1) + ctx_0 = fluid.layers.data( + name='ctx_0_data', shape=[1], dtype='int64', lod_level=1) + ctx_p1 = fluid.layers.data( + name='ctx_p1_data', shape=[1], dtype='int64', lod_level=1) + ctx_p2 = fluid.layers.data( + name='ctx_p2_data', shape=[1], dtype='int64', lod_level=1) + mark = fluid.layers.data( + name='mark_data', shape=[1], dtype='int64', lod_level=1) + + # 8 features + predicate_embedding = fluid.layers.embedding( + input=predicate, + size=[PRED_DICT_LEN, WORD_DIM], + dtype='float32', + is_sparse=IS_SPARSE, + param_attr='vemb') + + mark_embedding = fluid.layers.embedding( + input=mark, + size=[MARK_DICT_LEN, MARK_DIM], + dtype='float32', + is_sparse=IS_SPARSE) + + word_input = [word, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2] + emb_layers = [ + fluid.layers.embedding( + size=[WORD_DICT_LEN, WORD_DIM], + input=x, + param_attr=fluid.ParamAttr(name=EMBEDDING_NAME)) + for x in word_input + #name=EMBEDDING_NAME, trainable=False)) for x in word_input + ] + emb_layers.append(predicate_embedding) + emb_layers.append(mark_embedding) + + hidden_0_layers = [ + fluid.layers.fc(input=emb, size=HIDDEN_DIM, act='tanh') + for emb in emb_layers + ] + + hidden_0 = fluid.layers.sums(input=hidden_0_layers) + + lstm_0 = fluid.layers.dynamic_lstm( + input=hidden_0, + size=HIDDEN_DIM, + candidate_activation='relu', + gate_activation='sigmoid', + cell_activation='sigmoid') + + # stack L-LSTM and R-LSTM with direct edges + input_tmp = [hidden_0, lstm_0] + + for i in range(1, DEPTH): + mix_hidden = fluid.layers.sums(input=[ + fluid.layers.fc(input=input_tmp[0], size=HIDDEN_DIM, act='tanh'), + fluid.layers.fc(input=input_tmp[1], size=HIDDEN_DIM, act='tanh') + ]) + + lstm = fluid.layers.dynamic_lstm( + input=mix_hidden, + size=HIDDEN_DIM, + candidate_activation='relu', + gate_activation='sigmoid', + cell_activation='sigmoid', + is_reverse=((i % 2) == 1)) + + input_tmp = [mix_hidden, lstm] + + feature_out = fluid.layers.sums(input=[ + fluid.layers.fc(input=input_tmp[0], size=LABEL_DICT_LEN, act='tanh'), + fluid.layers.fc(input=input_tmp[1], size=LABEL_DICT_LEN, act='tanh') + ]) + + return feature_out + + +def inference_program(): + predict = lstm_net() + + return predict + + +def train_program(): + MIX_HIDDEN_LR = 1e-3 + + predict = lstm_net() + target = fluid.layers.data( + name='target', shape=[1], dtype='int64', lod_level=1) + crf_cost = fluid.layers.linear_chain_crf( + input=predict, + label=target, + param_attr=fluid.ParamAttr( + name='crfw', learning_rate=MIX_HIDDEN_LR)) + avg_cost = fluid.layers.mean(crf_cost) + + return [avg_cost] + + +def train(use_cuda, train_program, params_dirname): + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + optimizer = fluid.optimizer.SGD(learning_rate=0.01) + + trainer = fluid.Trainer( + train_func=train_program, place=place, optimizer=optimizer) + + feed_order = [ + 'word_data', 'ctx_n2_data', 'ctx_n1_data', 'ctx_0_data', 'ctx_p1_data', + 'ctx_p2_data', 'verb_data', 'mark_data', 'target' + ] + + #embedding_param = fluid.global_scope().find_var( + # EMBEDDING_NAME).get_tensor() + #embedding_param.set( + # load_parameter(conll05.get_embedding(), WORD_DICT_LEN, WORD_DIM), + # place) + + def event_handler(event): + if isinstance(event, fluid.EndEpochEvent): + test_reader = paddle.batch( + paddle.dataset.conll05.test(), batch_size=BATCH_SIZE) + avg_cost_set = trainer.test( + reader=test_reader, feed_order=feed_order) + + # get avg cost + avg_cost = np.array(avg_cost_set).mean() + + print("avg_cost: %s" % avg_cost) + + if float(avg_cost) < 100.0: # Large value to increase CI speed + trainer.save_params(params_dirname) + else: + print('BatchID {0}, Test Loss {1:0.2}'.format(event.epoch + 1, + float(avg_cost))) + if math.isnan(float(avg_cost)): + sys.exit("got NaN loss, training failed.") + + elif isinstance(event, fluid.EndStepEvent): + print("Step {0}, Epoch {1} Metrics {2}".format( + event.step, event.epoch, map(np.array, event.metrics))) + if event.step == 1: # Run 2 iterations to speed CI + trainer.save_params(params_dirname) + trainer.stop() + + train_reader = paddle.batch( + paddle.reader.shuffle( + paddle.dataset.conll05.test(), buf_size=8192), + batch_size=BATCH_SIZE) + trainer.train( + num_epochs=1, + event_handler=event_handler, + reader=train_reader, + feed_order=feed_order) + + +def infer(use_cuda, inference_program, params_dirname): + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + inferencer = fluid.Inferencer( + inference_program, param_path=params_dirname, place=place) + + # Setup inputs by creating LoDTensors to represent sequences of words. + # Here each word is the basic element of these LoDTensors and the shape of + # each word (base_shape) should be [1] since it is simply an index to + # look up for the corresponding word vector. + # Suppose the length_based level of detail (lod) info is set to [[3, 4, 2]], + # which has only one lod level. Then the created LoDTensors will have only + # one higher level structure (sequence of words, or sentence) than the basic + # element (word). Hence the LoDTensor will hold data for three sentences of + # length 3, 4 and 2, respectively. + # Note that lod info should be a list of lists. + lod = [[3, 4, 2]] + base_shape = [1] + # The range of random integers is [low, high] + word = fluid.create_random_int_lodtensor( + lod, base_shape, place, low=0, high=WORD_DICT_LEN - 1) + ctx_n2 = fluid.create_random_int_lodtensor( + lod, base_shape, place, low=0, high=WORD_DICT_LEN - 1) + ctx_n1 = fluid.create_random_int_lodtensor( + lod, base_shape, place, low=0, high=WORD_DICT_LEN - 1) + ctx_0 = fluid.create_random_int_lodtensor( + lod, base_shape, place, low=0, high=WORD_DICT_LEN - 1) + ctx_p1 = fluid.create_random_int_lodtensor( + lod, base_shape, place, low=0, high=WORD_DICT_LEN - 1) + ctx_p2 = fluid.create_random_int_lodtensor( + lod, base_shape, place, low=0, high=WORD_DICT_LEN - 1) + pred = fluid.create_random_int_lodtensor( + lod, base_shape, place, low=0, high=PRED_DICT_LEN - 1) + mark = fluid.create_random_int_lodtensor( + lod, base_shape, place, low=0, high=MARK_DICT_LEN - 1) + + results = inferencer.infer( + { + 'word_data': word, + 'ctx_n2_data': ctx_n2, + 'ctx_n1_data': ctx_n1, + 'ctx_0_data': ctx_0, + 'ctx_p1_data': ctx_p1, + 'ctx_p2_data': ctx_p2, + 'verb_data': pred, + 'mark_data': mark + }, + return_numpy=False) + + print("infer results: ", np.array(results[0])) + + +def main(use_cuda): + if use_cuda and not fluid.core.is_compiled_with_cuda(): + return + params_dirname = "label_semantic_roles.inference.model" + train(use_cuda, train_program, params_dirname) + infer(use_cuda, inference_program, params_dirname) + + +if __name__ == '__main__': + for use_cuda in (False, True): + main(use_cuda=use_cuda) diff --git a/python/paddle/fluid/tests/book/high-level-api/machine_translation/CMakeLists.txt b/python/paddle/fluid/tests/book/high-level-api/machine_translation/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..673c965b662a022739f8d489c331f4de9455a926 --- /dev/null +++ b/python/paddle/fluid/tests/book/high-level-api/machine_translation/CMakeLists.txt @@ -0,0 +1,7 @@ +file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py") +string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}") + +# default test +foreach(src ${TEST_OPS}) + py_test(${src} SRCS ${src}.py) +endforeach() diff --git a/python/paddle/fluid/tests/book/high-level-api/machine_translation/test_machine_translation.py b/python/paddle/fluid/tests/book/high-level-api/machine_translation/test_machine_translation.py new file mode 100644 index 0000000000000000000000000000000000000000..d4b723d3e6b619709ab3dc76a32ae87f1cdec274 --- /dev/null +++ b/python/paddle/fluid/tests/book/high-level-api/machine_translation/test_machine_translation.py @@ -0,0 +1,302 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +import contextlib + +import numpy as np +import paddle +import paddle.fluid as fluid +import paddle.fluid.framework as framework +import paddle.fluid.layers as pd +from paddle.fluid.executor import Executor +from functools import partial +import unittest +import os + +dict_size = 30000 +source_dict_dim = target_dict_dim = dict_size +hidden_dim = 32 +word_dim = 16 +batch_size = 2 +max_length = 8 +topk_size = 50 +trg_dic_size = 10000 +beam_size = 2 + +decoder_size = hidden_dim + + +def encoder(is_sparse): + # encoder + src_word_id = pd.data( + name="src_word_id", shape=[1], dtype='int64', lod_level=1) + src_embedding = pd.embedding( + input=src_word_id, + size=[dict_size, word_dim], + dtype='float32', + is_sparse=is_sparse, + param_attr=fluid.ParamAttr(name='vemb')) + + fc1 = pd.fc(input=src_embedding, size=hidden_dim * 4, act='tanh') + lstm_hidden0, lstm_0 = pd.dynamic_lstm(input=fc1, size=hidden_dim * 4) + encoder_out = pd.sequence_last_step(input=lstm_hidden0) + return encoder_out + + +def train_decoder(context, is_sparse): + # decoder + trg_language_word = pd.data( + name="target_language_word", shape=[1], dtype='int64', lod_level=1) + trg_embedding = pd.embedding( + input=trg_language_word, + size=[dict_size, word_dim], + dtype='float32', + is_sparse=is_sparse, + param_attr=fluid.ParamAttr(name='vemb')) + + rnn = pd.DynamicRNN() + with rnn.block(): + current_word = rnn.step_input(trg_embedding) + pre_state = rnn.memory(init=context) + current_state = pd.fc(input=[current_word, pre_state], + size=decoder_size, + act='tanh') + + current_score = pd.fc(input=current_state, + size=target_dict_dim, + act='softmax') + rnn.update_memory(pre_state, current_state) + rnn.output(current_score) + + return rnn() + + +def decode(context, is_sparse): + init_state = context + array_len = pd.fill_constant(shape=[1], dtype='int64', value=max_length) + counter = pd.zeros(shape=[1], dtype='int64', force_cpu=True) + + # fill the first element with init_state + state_array = pd.create_array('float32') + pd.array_write(init_state, array=state_array, i=counter) + + # ids, scores as memory + ids_array = pd.create_array('int64') + scores_array = pd.create_array('float32') + + init_ids = pd.data(name="init_ids", shape=[1], dtype="int64", lod_level=2) + init_scores = pd.data( + name="init_scores", shape=[1], dtype="float32", lod_level=2) + + pd.array_write(init_ids, array=ids_array, i=counter) + pd.array_write(init_scores, array=scores_array, i=counter) + + cond = pd.less_than(x=counter, y=array_len) + + while_op = pd.While(cond=cond) + with while_op.block(): + pre_ids = pd.array_read(array=ids_array, i=counter) + pre_state = pd.array_read(array=state_array, i=counter) + pre_score = pd.array_read(array=scores_array, i=counter) + + # expand the lod of pre_state to be the same with pre_score + pre_state_expanded = pd.sequence_expand(pre_state, pre_score) + + pre_ids_emb = pd.embedding( + input=pre_ids, + size=[dict_size, word_dim], + dtype='float32', + is_sparse=is_sparse) + + # use rnn unit to update rnn + current_state = pd.fc(input=[pre_state_expanded, pre_ids_emb], + size=decoder_size, + act='tanh') + current_state_with_lod = pd.lod_reset(x=current_state, y=pre_score) + # use score to do beam search + current_score = pd.fc(input=current_state_with_lod, + size=target_dict_dim, + act='softmax') + topk_scores, topk_indices = pd.topk(current_score, k=topk_size) + selected_ids, selected_scores = pd.beam_search( + pre_ids, topk_indices, topk_scores, beam_size, end_id=10, level=0) + + pd.increment(x=counter, value=1, in_place=True) + + # update the memories + pd.array_write(current_state, array=state_array, i=counter) + pd.array_write(selected_ids, array=ids_array, i=counter) + pd.array_write(selected_scores, array=scores_array, i=counter) + + pd.less_than(x=counter, y=array_len, cond=cond) + + translation_ids, translation_scores = pd.beam_search_decode( + ids=ids_array, scores=scores_array) + + # return init_ids, init_scores + + return translation_ids, translation_scores + + +def train_program(is_sparse): + context = encoder(is_sparse) + rnn_out = train_decoder(context, is_sparse) + label = pd.data( + name="target_language_next_word", shape=[1], dtype='int64', lod_level=1) + cost = pd.cross_entropy(input=rnn_out, label=label) + avg_cost = pd.mean(cost) + return avg_cost + + +def train(use_cuda, is_sparse, is_local=True): + EPOCH_NUM = 1 + + if use_cuda and not fluid.core.is_compiled_with_cuda(): + return + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + + train_reader = paddle.batch( + paddle.reader.shuffle( + paddle.dataset.wmt14.train(dict_size), buf_size=1000), + batch_size=batch_size) + + feed_order = [ + 'src_word_id', 'target_language_word', 'target_language_next_word' + ] + + def event_handler(event): + if isinstance(event, fluid.EndStepEvent): + print('pass_id=' + str(event.epoch) + ' batch=' + str(event.step)) + if event.step == 10: + trainer.stop() + + trainer = fluid.Trainer( + train_func=partial(train_program, is_sparse), + optimizer=fluid.optimizer.Adagrad( + learning_rate=1e-4, + regularization=fluid.regularizer.L2DecayRegularizer( + regularization_coeff=0.1)), + place=place) + + trainer.train( + reader=train_reader, + num_epochs=EPOCH_NUM, + event_handler=event_handler, + feed_order=feed_order) + + +def decode_main(use_cuda, is_sparse): + if use_cuda and not fluid.core.is_compiled_with_cuda(): + return + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + + context = encoder(is_sparse) + translation_ids, translation_scores = decode(context, is_sparse) + + exe = Executor(place) + exe.run(framework.default_startup_program()) + + init_ids_data = np.array([1 for _ in range(batch_size)], dtype='int64') + init_scores_data = np.array( + [1. for _ in range(batch_size)], dtype='float32') + init_ids_data = init_ids_data.reshape((batch_size, 1)) + init_scores_data = init_scores_data.reshape((batch_size, 1)) + init_lod = [1] * batch_size + init_lod = [init_lod, init_lod] + + init_ids = fluid.create_lod_tensor(init_ids_data, init_lod, place) + init_scores = fluid.create_lod_tensor(init_scores_data, init_lod, place) + + train_data = paddle.batch( + paddle.reader.shuffle( + paddle.dataset.wmt14.train(dict_size), buf_size=1000), + batch_size=batch_size) + + feed_order = ['src_word_id'] + feed_list = [ + framework.default_main_program().global_block().var(var_name) + for var_name in feed_order + ] + feeder = fluid.DataFeeder(feed_list, place) + + for data in train_data(): + feed_dict = feeder.feed(map(lambda x: [x[0]], data)) + feed_dict['init_ids'] = init_ids + feed_dict['init_scores'] = init_scores + + result_ids, result_scores = exe.run( + framework.default_main_program(), + feed=feed_dict, + fetch_list=[translation_ids, translation_scores], + return_numpy=False) + print result_ids.lod() + break + + +class TestMachineTranslation(unittest.TestCase): + pass + + +@contextlib.contextmanager +def scope_prog_guard(): + prog = fluid.Program() + startup_prog = fluid.Program() + scope = fluid.core.Scope() + with fluid.scope_guard(scope): + with fluid.program_guard(prog, startup_prog): + yield + + +def inject_test_train(use_cuda, is_sparse): + f_name = 'test_{0}_{1}_train'.format('cuda' if use_cuda else 'cpu', 'sparse' + if is_sparse else 'dense') + + def f(*args): + with scope_prog_guard(): + train(use_cuda, is_sparse) + + setattr(TestMachineTranslation, f_name, f) + + +def inject_test_decode(use_cuda, is_sparse, decorator=None): + f_name = 'test_{0}_{1}_decode'.format('cuda' + if use_cuda else 'cpu', 'sparse' + if is_sparse else 'dense') + + def f(*args): + with scope_prog_guard(): + decode_main(use_cuda, is_sparse) + + if decorator is not None: + f = decorator(f) + + setattr(TestMachineTranslation, f_name, f) + + +for _use_cuda_ in (False, True): + for _is_sparse_ in (False, True): + inject_test_train(_use_cuda_, _is_sparse_) + +for _use_cuda_ in (False, True): + for _is_sparse_ in (False, True): + + _decorator_ = None + if _use_cuda_: + _decorator_ = unittest.skip( + reason='Beam Search does not support CUDA!') + + inject_test_decode( + is_sparse=_is_sparse_, use_cuda=_use_cuda_, decorator=_decorator_) + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/book/high-level-api/recognize_digits/CMakeLists.txt b/python/paddle/fluid/tests/book/high-level-api/recognize_digits/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..673c965b662a022739f8d489c331f4de9455a926 --- /dev/null +++ b/python/paddle/fluid/tests/book/high-level-api/recognize_digits/CMakeLists.txt @@ -0,0 +1,7 @@ +file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py") +string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}") + +# default test +foreach(src ${TEST_OPS}) + py_test(${src} SRCS ${src}.py) +endforeach() diff --git a/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_conv.py b/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_conv.py new file mode 100644 index 0000000000000000000000000000000000000000..03439cbd37671b4727879bf3d0793f016f55247a --- /dev/null +++ b/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_conv.py @@ -0,0 +1,134 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from __future__ import print_function +import argparse +import paddle.fluid as fluid +import paddle +import sys +import numpy +import unittest +import math +import sys +import os + +BATCH_SIZE = 64 + + +def inference_program(): + img = fluid.layers.data(name='img', shape=[1, 28, 28], dtype='float32') + + conv_pool_1 = fluid.nets.simple_img_conv_pool( + input=img, + filter_size=5, + num_filters=20, + pool_size=2, + pool_stride=2, + act="relu") + conv_pool_1 = fluid.layers.batch_norm(conv_pool_1) + conv_pool_2 = fluid.nets.simple_img_conv_pool( + input=conv_pool_1, + filter_size=5, + num_filters=50, + pool_size=2, + pool_stride=2, + act="relu") + prediction = fluid.layers.fc(input=conv_pool_2, size=10, act='softmax') + return prediction + + +def train_program(): + label = fluid.layers.data(name='label', shape=[1], dtype='int64') + + predict = inference_program() + cost = fluid.layers.cross_entropy(input=predict, label=label) + avg_cost = fluid.layers.mean(cost) + acc = fluid.layers.accuracy(input=predict, label=label) + return [avg_cost, acc] + + +def train(use_cuda, train_program, params_dirname): + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + optimizer = fluid.optimizer.Adam(learning_rate=0.001) + + trainer = fluid.Trainer( + train_func=train_program, + place=place, + optimizer=optimizer, + parallel=True) + + def event_handler(event): + if isinstance(event, fluid.EndEpochEvent): + test_reader = paddle.batch( + paddle.dataset.mnist.test(), batch_size=BATCH_SIZE) + avg_cost, acc = trainer.test( + reader=test_reader, feed_order=['img', 'label']) + + print("avg_cost: %s" % avg_cost) + print("acc : %s" % acc) + + if acc > 0.2: # Smaller value to increase CI speed + trainer.save_params(params_dirname) + else: + print('BatchID {0}, Test Loss {1:0.2}, Acc {2:0.2}'.format( + event.epoch + 1, avg_cost, acc)) + if math.isnan(avg_cost): + sys.exit("got NaN loss, training failed.") + elif isinstance(event, fluid.EndStepEvent): + print("Step {0}, Epoch {1} Metrics {2}".format( + event.step, event.epoch, map(numpy.array, event.metrics))) + + train_reader = paddle.batch( + paddle.reader.shuffle( + paddle.dataset.mnist.train(), buf_size=500), + batch_size=BATCH_SIZE) + + trainer.train( + num_epochs=1, + event_handler=event_handler, + reader=train_reader, + feed_order=['img', 'label']) + + +def infer(use_cuda, inference_program, params_dirname=None): + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + + inferencer = fluid.Inferencer( + infer_func=inference_program, param_path=params_dirname, place=place) + + batch_size = 1 + tensor_img = numpy.random.uniform(-1.0, 1.0, + [batch_size, 1, 28, 28]).astype("float32") + + results = inferencer.infer({'img': tensor_img}) + + print("infer results: ", results[0]) + + +def main(use_cuda): + params_dirname = "recognize_digits_conv.inference.model" + + # call train() with is_local argument to run distributed train + train( + use_cuda=use_cuda, + train_program=train_program, + params_dirname=params_dirname) + infer( + use_cuda=use_cuda, + inference_program=inference_program, + params_dirname=params_dirname) + + +if __name__ == '__main__': + # for use_cuda in (False, True): + main(use_cuda=True) diff --git a/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_mlp.py b/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_mlp.py new file mode 100644 index 0000000000000000000000000000000000000000..89bbd21bea7d64a8dd6fc32829b6addb680da62e --- /dev/null +++ b/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_mlp.py @@ -0,0 +1,115 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from __future__ import print_function +import argparse +import paddle.fluid as fluid +import paddle +import sys +import numpy +import unittest +import math +import sys +import os + +BATCH_SIZE = 64 + + +def inference_program(): + img = fluid.layers.data(name='img', shape=[1, 28, 28], dtype='float32') + + hidden = fluid.layers.fc(input=img, size=200, act='tanh') + hidden = fluid.layers.fc(input=hidden, size=200, act='tanh') + prediction = fluid.layers.fc(input=hidden, size=10, act='softmax') + return prediction + + +def train_program(): + label = fluid.layers.data(name='label', shape=[1], dtype='int64') + + predict = inference_program() + cost = fluid.layers.cross_entropy(input=predict, label=label) + avg_cost = fluid.layers.mean(cost) + acc = fluid.layers.accuracy(input=predict, label=label) + return [avg_cost, acc] + + +def train(use_cuda, train_program, params_dirname): + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + optimizer = fluid.optimizer.Adam(learning_rate=0.001) + + trainer = fluid.Trainer( + train_func=train_program, place=place, optimizer=optimizer) + + def event_handler(event): + if isinstance(event, fluid.EndEpochEvent): + test_reader = paddle.batch( + paddle.dataset.mnist.test(), batch_size=BATCH_SIZE) + avg_cost, acc = trainer.test( + reader=test_reader, feed_order=['img', 'label']) + + print("avg_cost: %s" % avg_cost) + print("acc : %s" % acc) + + if acc > 0.2: # Smaller value to increase CI speed + trainer.save_params(params_dirname) + else: + print('BatchID {0}, Test Loss {1:0.2}, Acc {2:0.2}'.format( + event.epoch + 1, avg_cost, acc)) + if math.isnan(avg_cost): + sys.exit("got NaN loss, training failed.") + + train_reader = paddle.batch( + paddle.reader.shuffle( + paddle.dataset.mnist.train(), buf_size=500), + batch_size=BATCH_SIZE) + + trainer.train( + num_epochs=1, + event_handler=event_handler, + reader=train_reader, + feed_order=['img', 'label']) + + +def infer(use_cuda, inference_program, params_dirname=None): + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + + inferencer = fluid.Inferencer( + infer_func=inference_program, param_path=params_dirname, place=place) + + batch_size = 1 + tensor_img = numpy.random.uniform(-1.0, 1.0, + [batch_size, 1, 28, 28]).astype("float32") + + results = inferencer.infer({'img': tensor_img}) + + print("infer results: ", results[0]) + + +def main(use_cuda): + params_dirname = "recognize_digits_mlp.inference.model" + + # call train() with is_local argument to run distributed train + train( + use_cuda=use_cuda, + train_program=train_program, + params_dirname=params_dirname) + infer( + use_cuda=use_cuda, + inference_program=inference_program, + params_dirname=params_dirname) + + +if __name__ == '__main__': + # for use_cuda in (False, True): + main(use_cuda=False) diff --git a/python/paddle/fluid/tests/book/high-level-api/recommender_system/CMakeLists.txt b/python/paddle/fluid/tests/book/high-level-api/recommender_system/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..673c965b662a022739f8d489c331f4de9455a926 --- /dev/null +++ b/python/paddle/fluid/tests/book/high-level-api/recommender_system/CMakeLists.txt @@ -0,0 +1,7 @@ +file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py") +string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}") + +# default test +foreach(src ${TEST_OPS}) + py_test(${src} SRCS ${src}.py) +endforeach() diff --git a/python/paddle/fluid/tests/book/high-level-api/recommender_system/test_recommender_system_newapi.py b/python/paddle/fluid/tests/book/high-level-api/recommender_system/test_recommender_system_newapi.py new file mode 100644 index 0000000000000000000000000000000000000000..dfc7325acf23176c05fe42761b9997b98d23372a --- /dev/null +++ b/python/paddle/fluid/tests/book/high-level-api/recommender_system/test_recommender_system_newapi.py @@ -0,0 +1,255 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import math +import sys +import numpy as np +import paddle +import paddle.fluid as fluid +import paddle.fluid.layers as layers +import paddle.fluid.nets as nets + +IS_SPARSE = True +USE_GPU = False +BATCH_SIZE = 256 + + +def get_usr_combined_features(): + # FIXME(dzh) : old API integer_value(10) may have range check. + # currently we don't have user configurated check. + + USR_DICT_SIZE = paddle.dataset.movielens.max_user_id() + 1 + + uid = layers.data(name='user_id', shape=[1], dtype='int64') + + usr_emb = layers.embedding( + input=uid, + dtype='float32', + size=[USR_DICT_SIZE, 32], + param_attr='user_table', + is_sparse=IS_SPARSE) + + usr_fc = layers.fc(input=usr_emb, size=32) + + USR_GENDER_DICT_SIZE = 2 + + usr_gender_id = layers.data(name='gender_id', shape=[1], dtype='int64') + + usr_gender_emb = layers.embedding( + input=usr_gender_id, + size=[USR_GENDER_DICT_SIZE, 16], + param_attr='gender_table', + is_sparse=IS_SPARSE) + + usr_gender_fc = layers.fc(input=usr_gender_emb, size=16) + + USR_AGE_DICT_SIZE = len(paddle.dataset.movielens.age_table) + usr_age_id = layers.data(name='age_id', shape=[1], dtype="int64") + + usr_age_emb = layers.embedding( + input=usr_age_id, + size=[USR_AGE_DICT_SIZE, 16], + is_sparse=IS_SPARSE, + param_attr='age_table') + + usr_age_fc = layers.fc(input=usr_age_emb, size=16) + + USR_JOB_DICT_SIZE = paddle.dataset.movielens.max_job_id() + 1 + usr_job_id = layers.data(name='job_id', shape=[1], dtype="int64") + + usr_job_emb = layers.embedding( + input=usr_job_id, + size=[USR_JOB_DICT_SIZE, 16], + param_attr='job_table', + is_sparse=IS_SPARSE) + + usr_job_fc = layers.fc(input=usr_job_emb, size=16) + + concat_embed = layers.concat( + input=[usr_fc, usr_gender_fc, usr_age_fc, usr_job_fc], axis=1) + + usr_combined_features = layers.fc(input=concat_embed, size=200, act="tanh") + + return usr_combined_features + + +def get_mov_combined_features(): + + MOV_DICT_SIZE = paddle.dataset.movielens.max_movie_id() + 1 + + mov_id = layers.data(name='movie_id', shape=[1], dtype='int64') + + mov_emb = layers.embedding( + input=mov_id, + dtype='float32', + size=[MOV_DICT_SIZE, 32], + param_attr='movie_table', + is_sparse=IS_SPARSE) + + mov_fc = layers.fc(input=mov_emb, size=32) + + CATEGORY_DICT_SIZE = len(paddle.dataset.movielens.movie_categories()) + + category_id = layers.data( + name='category_id', shape=[1], dtype='int64', lod_level=1) + + mov_categories_emb = layers.embedding( + input=category_id, size=[CATEGORY_DICT_SIZE, 32], is_sparse=IS_SPARSE) + + mov_categories_hidden = layers.sequence_pool( + input=mov_categories_emb, pool_type="sum") + + MOV_TITLE_DICT_SIZE = len(paddle.dataset.movielens.get_movie_title_dict()) + + mov_title_id = layers.data( + name='movie_title', shape=[1], dtype='int64', lod_level=1) + + mov_title_emb = layers.embedding( + input=mov_title_id, size=[MOV_TITLE_DICT_SIZE, 32], is_sparse=IS_SPARSE) + + mov_title_conv = nets.sequence_conv_pool( + input=mov_title_emb, + num_filters=32, + filter_size=3, + act="tanh", + pool_type="sum") + + concat_embed = layers.concat( + input=[mov_fc, mov_categories_hidden, mov_title_conv], axis=1) + + # FIXME(dzh) : need tanh operator + mov_combined_features = layers.fc(input=concat_embed, size=200, act="tanh") + + return mov_combined_features + + +def inference_program(): + usr_combined_features = get_usr_combined_features() + mov_combined_features = get_mov_combined_features() + + inference = layers.cos_sim(X=usr_combined_features, Y=mov_combined_features) + scale_infer = layers.scale(x=inference, scale=5.0) + + return scale_infer + + +def train_program(): + + scale_infer = inference_program() + + label = layers.data(name='score', shape=[1], dtype='float32') + square_cost = layers.square_error_cost(input=scale_infer, label=label) + avg_cost = layers.mean(square_cost) + + return [avg_cost, scale_infer] + + +def train(use_cuda, train_program, params_dirname): + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + optimizer = fluid.optimizer.SGD(learning_rate=0.2) + + trainer = fluid.Trainer( + train_func=train_program, place=place, optimizer=optimizer) + + feed_order = [ + 'user_id', 'gender_id', 'age_id', 'job_id', 'movie_id', 'category_id', + 'movie_title', 'score' + ] + + def event_handler(event): + if isinstance(event, fluid.EndStepEvent): + test_reader = paddle.batch( + paddle.dataset.movielens.test(), batch_size=BATCH_SIZE) + avg_cost_set = trainer.test( + reader=test_reader, feed_order=feed_order) + + # get avg cost + avg_cost = np.array(avg_cost_set).mean() + + print("avg_cost: %s" % avg_cost) + + if float(avg_cost) < 4: # Smaller value to increase CI speed + trainer.save_params(params_dirname) + trainer.stop() + else: + print('BatchID {0}, Test Loss {1:0.2}'.format(event.epoch + 1, + float(avg_cost))) + if math.isnan(float(avg_cost)): + sys.exit("got NaN loss, training failed.") + + train_reader = paddle.batch( + paddle.reader.shuffle( + paddle.dataset.movielens.train(), buf_size=8192), + batch_size=BATCH_SIZE) + + trainer.train( + num_epochs=1, + event_handler=event_handler, + reader=train_reader, + feed_order=feed_order) + + +def infer(use_cuda, inference_program, params_dirname): + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + inferencer = fluid.Inferencer( + inference_program, param_path=params_dirname, place=place) + + # Use the first data from paddle.dataset.movielens.test() as input. + # Use create_lod_tensor(data, lod, place) API to generate LoD Tensor, + # where `data` is a list of sequences of index numbers, `lod` is + # the level of detail (lod) info associated with `data`. + # For example, data = [[10, 2, 3], [2, 3]] means that it contains + # two sequences of indexes, of length 3 and 2, respectively. + # Correspondingly, lod = [[3, 2]] contains one level of detail info, + # indicating that `data` consists of two sequences of length 3 and 2. + user_id = fluid.create_lod_tensor([[1]], [[1]], place) + gender_id = fluid.create_lod_tensor([[1]], [[1]], place) + age_id = fluid.create_lod_tensor([[0]], [[1]], place) + job_id = fluid.create_lod_tensor([[10]], [[1]], place) + movie_id = fluid.create_lod_tensor([[783]], [[1]], place) + category_id = fluid.create_lod_tensor([[10, 8, 9]], [[3]], place) + movie_title = fluid.create_lod_tensor([[1069, 4140, 2923, 710, 988]], [[5]], + place) + + results = inferencer.infer( + { + 'user_id': user_id, + 'gender_id': gender_id, + 'age_id': age_id, + 'job_id': job_id, + 'movie_id': movie_id, + 'category_id': category_id, + 'movie_title': movie_title + }, + return_numpy=False) + + print("infer results: ", np.array(results[0])) + + +def main(use_cuda): + if use_cuda and not fluid.core.is_compiled_with_cuda(): + return + params_dirname = "recommender_system.inference.model" + train( + use_cuda=use_cuda, + train_program=train_program, + params_dirname=params_dirname) + infer( + use_cuda=use_cuda, + inference_program=inference_program, + params_dirname=params_dirname) + + +if __name__ == '__main__': + main(USE_GPU) diff --git a/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/CMakeLists.txt b/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..d71147a85e77ea6dc5b6391aa169abd9b02a0aa1 --- /dev/null +++ b/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/CMakeLists.txt @@ -0,0 +1,12 @@ +file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py") +string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}") + +# This test is buggy +# py_test(test_understand_sentiment_dynamic_rnn SRCS +# test_understand_sentiment_dynamic_rnn.py SERIAL) +LIST(REMOVE_ITEM TEST_OPS test_understand_sentiment_dynamic_rnn) + +# default test +foreach(src ${TEST_OPS}) + py_test(${src} SRCS ${src}.py) +endforeach() diff --git a/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_conv.py b/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_conv.py new file mode 100644 index 0000000000000000000000000000000000000000..11e9fd1bec1450f6753dbe38c7014090d6e136b6 --- /dev/null +++ b/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_conv.py @@ -0,0 +1,153 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import print_function + +import paddle +import paddle.fluid as fluid +from functools import partial +import numpy as np + +CLASS_DIM = 2 +EMB_DIM = 128 +HID_DIM = 512 +BATCH_SIZE = 128 + + +def convolution_net(data, input_dim, class_dim, emb_dim, hid_dim): + emb = fluid.layers.embedding( + input=data, size=[input_dim, emb_dim], is_sparse=True) + conv_3 = fluid.nets.sequence_conv_pool( + input=emb, + num_filters=hid_dim, + filter_size=3, + act="tanh", + pool_type="sqrt") + conv_4 = fluid.nets.sequence_conv_pool( + input=emb, + num_filters=hid_dim, + filter_size=4, + act="tanh", + pool_type="sqrt") + prediction = fluid.layers.fc(input=[conv_3, conv_4], + size=class_dim, + act="softmax") + return prediction + + +def inference_program(word_dict): + data = fluid.layers.data( + name="words", shape=[1], dtype="int64", lod_level=1) + + dict_dim = len(word_dict) + net = convolution_net(data, dict_dim, CLASS_DIM, EMB_DIM, HID_DIM) + return net + + +def train_program(word_dict): + prediction = inference_program(word_dict) + label = fluid.layers.data(name="label", shape=[1], dtype="int64") + cost = fluid.layers.cross_entropy(input=prediction, label=label) + avg_cost = fluid.layers.mean(cost) + accuracy = fluid.layers.accuracy(input=prediction, label=label) + return [avg_cost, accuracy] + + +def train(use_cuda, train_program, params_dirname): + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + optimizer = fluid.optimizer.Adagrad(learning_rate=0.002) + + word_dict = paddle.dataset.imdb.word_dict() + trainer = fluid.Trainer( + train_func=partial(train_program, word_dict), + place=place, + optimizer=optimizer) + + def event_handler(event): + if isinstance(event, fluid.EndEpochEvent): + test_reader = paddle.batch( + paddle.dataset.imdb.test(word_dict), batch_size=BATCH_SIZE) + avg_cost, acc = trainer.test( + reader=test_reader, feed_order=['words', 'label']) + + print("avg_cost: %s" % avg_cost) + print("acc : %s" % acc) + + if acc > 0.2: # Smaller value to increase CI speed + trainer.save_params(params_dirname) + trainer.stop() + + else: + print('BatchID {0}, Test Loss {1:0.2}, Acc {2:0.2}'.format( + event.epoch + 1, avg_cost, acc)) + if math.isnan(avg_cost): + sys.exit("got NaN loss, training failed.") + elif isinstance(event, fluid.EndStepEvent): + print("Step {0}, Epoch {1} Metrics {2}".format( + event.step, event.epoch, map(np.array, event.metrics))) + if event.step == 1: # Run 2 iterations to speed CI + trainer.save_params(params_dirname) + trainer.stop() + + train_reader = paddle.batch( + paddle.reader.shuffle( + paddle.dataset.imdb.train(word_dict), buf_size=25000), + batch_size=BATCH_SIZE) + + trainer.train( + num_epochs=1, + event_handler=event_handler, + reader=train_reader, + feed_order=['words', 'label']) + + +def infer(use_cuda, inference_program, params_dirname=None): + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + word_dict = paddle.dataset.imdb.word_dict() + + inferencer = fluid.Inferencer( + infer_func=partial(inference_program, word_dict), + param_path=params_dirname, + place=place) + + # Setup input by creating LoDTensor to represent sequence of words. + # Here each word is the basic element of the LoDTensor and the shape of + # each word (base_shape) should be [1] since it is simply an index to + # look up for the corresponding word vector. + # Suppose the length_based level of detail (lod) info is set to [[3, 4, 2]], + # which has only one lod level. Then the created LoDTensor will have only + # one higher level structure (sequence of words, or sentence) than the basic + # element (word). Hence the LoDTensor will hold data for three sentences of + # length 3, 4 and 2, respectively. + # Note that lod info should be a list of lists. + lod = [[3, 4, 2]] + base_shape = [1] + # The range of random integers is [low, high] + tensor_words = fluid.create_random_int_lodtensor( + lod, base_shape, place, low=0, high=len(word_dict) - 1) + results = inferencer.infer({'words': tensor_words}) + print("infer results: ", results) + + +def main(use_cuda): + if use_cuda and not fluid.core.is_compiled_with_cuda(): + return + params_dirname = "understand_sentiment_conv.inference.model" + train(use_cuda, train_program, params_dirname) + infer(use_cuda, inference_program, params_dirname) + + +if __name__ == '__main__': + for use_cuda in (False, True): + main(use_cuda=use_cuda) diff --git a/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_dynamic_rnn.py b/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_dynamic_rnn.py new file mode 100644 index 0000000000000000000000000000000000000000..90757d54f99715163518ce5a094e6ba3a67efed3 --- /dev/null +++ b/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_dynamic_rnn.py @@ -0,0 +1,168 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import print_function + +import paddle +import paddle.fluid as fluid +from functools import partial +import numpy as np + +CLASS_DIM = 2 +EMB_DIM = 128 +BATCH_SIZE = 128 +LSTM_SIZE = 128 + + +def dynamic_rnn_lstm(data, input_dim, class_dim, emb_dim, lstm_size): + emb = fluid.layers.embedding( + input=data, size=[input_dim, emb_dim], is_sparse=True) + sentence = fluid.layers.fc(input=emb, size=lstm_size, act='tanh') + + rnn = fluid.layers.DynamicRNN() + with rnn.block(): + word = rnn.step_input(sentence) + prev_hidden = rnn.memory(value=0.0, shape=[lstm_size]) + prev_cell = rnn.memory(value=0.0, shape=[lstm_size]) + + def gate_common(ipt, hidden, size): + gate0 = fluid.layers.fc(input=ipt, size=size, bias_attr=True) + gate1 = fluid.layers.fc(input=hidden, size=size, bias_attr=False) + return gate0 + gate1 + + forget_gate = fluid.layers.sigmoid(x=gate_common(word, prev_hidden, + lstm_size)) + input_gate = fluid.layers.sigmoid(x=gate_common(word, prev_hidden, + lstm_size)) + output_gate = fluid.layers.sigmoid(x=gate_common(word, prev_hidden, + lstm_size)) + cell_gate = fluid.layers.sigmoid(x=gate_common(word, prev_hidden, + lstm_size)) + + cell = forget_gate * prev_cell + input_gate * cell_gate + hidden = output_gate * fluid.layers.tanh(x=cell) + rnn.update_memory(prev_cell, cell) + rnn.update_memory(prev_hidden, hidden) + rnn.output(hidden) + + last = fluid.layers.sequence_last_step(rnn()) + prediction = fluid.layers.fc(input=last, size=class_dim, act="softmax") + return prediction + + +def inference_program(word_dict): + data = fluid.layers.data( + name="words", shape=[1], dtype="int64", lod_level=1) + + dict_dim = len(word_dict) + pred = dynamic_rnn_lstm(data, dict_dim, CLASS_DIM, EMB_DIM, LSTM_SIZE) + return pred + + +def train_program(word_dict): + prediction = inference_program(word_dict) + label = fluid.layers.data(name="label", shape=[1], dtype="int64") + cost = fluid.layers.cross_entropy(input=prediction, label=label) + avg_cost = fluid.layers.mean(cost) + accuracy = fluid.layers.accuracy(input=prediction, label=label) + return [avg_cost, accuracy] + + +def train(use_cuda, train_program, params_dirname): + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + optimizer = fluid.optimizer.Adagrad(learning_rate=0.002) + + word_dict = paddle.dataset.imdb.word_dict() + trainer = fluid.Trainer( + train_func=partial(train_program, word_dict), + place=place, + optimizer=optimizer) + + def event_handler(event): + if isinstance(event, fluid.EndEpochEvent): + test_reader = paddle.batch( + paddle.dataset.imdb.test(word_dict), batch_size=BATCH_SIZE) + avg_cost, acc = trainer.test( + reader=test_reader, feed_order=['words', 'label']) + + print("avg_cost: %s" % avg_cost) + print("acc : %s" % acc) + + if acc > 0.2: # Smaller value to increase CI speed + trainer.save_params(params_dirname) + trainer.stop() + + else: + print('BatchID {0}, Test Loss {1:0.2}, Acc {2:0.2}'.format( + event.epoch + 1, avg_cost, acc)) + if math.isnan(avg_cost): + sys.exit("got NaN loss, training failed.") + elif isinstance(event, fluid.EndStepEvent): + print("Step {0}, Epoch {1} Metrics {2}".format( + event.step, event.epoch, map(np.array, event.metrics))) + if event.step == 1: # Run 2 iterations to speed CI + trainer.save_params(params_dirname) + trainer.stop() + + train_reader = paddle.batch( + paddle.reader.shuffle( + paddle.dataset.imdb.train(word_dict), buf_size=25000), + batch_size=BATCH_SIZE) + + trainer.train( + num_epochs=1, + event_handler=event_handler, + reader=train_reader, + feed_order=['words', 'label']) + + +def infer(use_cuda, inference_program, params_dirname=None): + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + word_dict = paddle.dataset.imdb.word_dict() + + inferencer = fluid.Inferencer( + infer_func=partial(inference_program, word_dict), + param_path=params_dirname, + place=place) + + # Setup input by creating LoDTensor to represent sequence of words. + # Here each word is the basic element of the LoDTensor and the shape of + # each word (base_shape) should be [1] since it is simply an index to + # look up for the corresponding word vector. + # Suppose the length_based level of detail (lod) info is set to [[3, 4, 2]], + # which has only one lod level. Then the created LoDTensor will have only + # one higher level structure (sequence of words, or sentence) than the basic + # element (word). Hence the LoDTensor will hold data for three sentences of + # length 3, 4 and 2, respectively. + # Note that lod info should be a list of lists. + lod = [[3, 4, 2]] + base_shape = [1] + # The range of random integers is [low, high] + tensor_words = fluid.create_random_int_lodtensor( + lod, base_shape, place, low=0, high=len(word_dict) - 1) + results = inferencer.infer({'words': tensor_words}) + print("infer results: ", results) + + +def main(use_cuda): + if use_cuda and not fluid.core.is_compiled_with_cuda(): + return + params_dirname = "understand_sentiment_conv.inference.model" + train(use_cuda, train_program, params_dirname) + infer(use_cuda, inference_program, params_dirname) + + +if __name__ == '__main__': + for use_cuda in (False, True): + main(use_cuda=use_cuda) diff --git a/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_stacked_lstm.py b/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_stacked_lstm.py new file mode 100644 index 0000000000000000000000000000000000000000..52b7d4a83779d01936afb3d9d1e4864b05d55b5a --- /dev/null +++ b/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_stacked_lstm.py @@ -0,0 +1,160 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import print_function + +import paddle +import paddle.fluid as fluid +from functools import partial +import numpy as np + +CLASS_DIM = 2 +EMB_DIM = 128 +HID_DIM = 512 +STACKED_NUM = 3 +BATCH_SIZE = 128 + + +def stacked_lstm_net(data, input_dim, class_dim, emb_dim, hid_dim, stacked_num): + assert stacked_num % 2 == 1 + + emb = fluid.layers.embedding( + input=data, size=[input_dim, emb_dim], is_sparse=True) + + fc1 = fluid.layers.fc(input=emb, size=hid_dim) + lstm1, cell1 = fluid.layers.dynamic_lstm(input=fc1, size=hid_dim) + + inputs = [fc1, lstm1] + + for i in range(2, stacked_num + 1): + fc = fluid.layers.fc(input=inputs, size=hid_dim) + lstm, cell = fluid.layers.dynamic_lstm( + input=fc, size=hid_dim, is_reverse=(i % 2) == 0) + inputs = [fc, lstm] + + fc_last = fluid.layers.sequence_pool(input=inputs[0], pool_type='max') + lstm_last = fluid.layers.sequence_pool(input=inputs[1], pool_type='max') + + prediction = fluid.layers.fc(input=[fc_last, lstm_last], + size=class_dim, + act='softmax') + return prediction + + +def inference_program(word_dict): + data = fluid.layers.data( + name="words", shape=[1], dtype="int64", lod_level=1) + + dict_dim = len(word_dict) + net = stacked_lstm_net(data, dict_dim, CLASS_DIM, EMB_DIM, HID_DIM, + STACKED_NUM) + return net + + +def train_program(word_dict): + prediction = inference_program(word_dict) + label = fluid.layers.data(name="label", shape=[1], dtype="int64") + cost = fluid.layers.cross_entropy(input=prediction, label=label) + avg_cost = fluid.layers.mean(cost) + accuracy = fluid.layers.accuracy(input=prediction, label=label) + return [avg_cost, accuracy] + + +def train(use_cuda, train_program, params_dirname): + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + optimizer = fluid.optimizer.Adagrad(learning_rate=0.002) + + word_dict = paddle.dataset.imdb.word_dict() + trainer = fluid.Trainer( + train_func=partial(train_program, word_dict), + place=place, + optimizer=optimizer) + + def event_handler(event): + if isinstance(event, fluid.EndEpochEvent): + test_reader = paddle.batch( + paddle.dataset.imdb.test(word_dict), batch_size=BATCH_SIZE) + avg_cost, acc = trainer.test( + reader=test_reader, feed_order=['words', 'label']) + + print("avg_cost: %s" % avg_cost) + print("acc : %s" % acc) + + if acc > 0.2: # Smaller value to increase CI speed + trainer.save_params(params_dirname) + trainer.stop() + + else: + print('BatchID {0}, Test Loss {1:0.2}, Acc {2:0.2}'.format( + event.epoch + 1, avg_cost, acc)) + if math.isnan(avg_cost): + sys.exit("got NaN loss, training failed.") + elif isinstance(event, fluid.EndStepEvent): + print("Step {0}, Epoch {1} Metrics {2}".format( + event.step, event.epoch, map(np.array, event.metrics))) + if event.step == 1: # Run 2 iterations to speed CI + trainer.save_params(params_dirname) + trainer.stop() + + train_reader = paddle.batch( + paddle.reader.shuffle( + paddle.dataset.imdb.train(word_dict), buf_size=25000), + batch_size=BATCH_SIZE) + + trainer.train( + num_epochs=1, + event_handler=event_handler, + reader=train_reader, + feed_order=['words', 'label']) + + +def infer(use_cuda, inference_program, params_dirname=None): + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + word_dict = paddle.dataset.imdb.word_dict() + + inferencer = fluid.Inferencer( + infer_func=partial(inference_program, word_dict), + param_path=params_dirname, + place=place) + + # Setup input by creating LoDTensor to represent sequence of words. + # Here each word is the basic element of the LoDTensor and the shape of + # each word (base_shape) should be [1] since it is simply an index to + # look up for the corresponding word vector. + # Suppose the length_based level of detail (lod) info is set to [[3, 4, 2]], + # which has only one lod level. Then the created LoDTensor will have only + # one higher level structure (sequence of words, or sentence) than the basic + # element (word). Hence the LoDTensor will hold data for three sentences of + # length 3, 4 and 2, respectively. + # Note that lod info should be a list of lists. + lod = [[3, 4, 2]] + base_shape = [1] + # The range of random integers is [low, high] + tensor_words = fluid.create_random_int_lodtensor( + lod, base_shape, place, low=0, high=len(word_dict) - 1) + results = inferencer.infer({'words': tensor_words}) + print("infer results: ", results) + + +def main(use_cuda): + if use_cuda and not fluid.core.is_compiled_with_cuda(): + return + params_dirname = "understand_sentiment_stacked_lstm.inference.model" + train(use_cuda, train_program, params_dirname) + infer(use_cuda, inference_program, params_dirname) + + +if __name__ == '__main__': + for use_cuda in (False, True): + main(use_cuda=use_cuda) diff --git a/python/paddle/fluid/tests/book/high-level-api/word2vec/CMakeLists.txt b/python/paddle/fluid/tests/book/high-level-api/word2vec/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..673c965b662a022739f8d489c331f4de9455a926 --- /dev/null +++ b/python/paddle/fluid/tests/book/high-level-api/word2vec/CMakeLists.txt @@ -0,0 +1,7 @@ +file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py") +string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}") + +# default test +foreach(src ${TEST_OPS}) + py_test(${src} SRCS ${src}.py) +endforeach() diff --git a/python/paddle/fluid/tests/book/high-level-api/word2vec/test_word2vec_new_api.py b/python/paddle/fluid/tests/book/high-level-api/word2vec/test_word2vec_new_api.py new file mode 100644 index 0000000000000000000000000000000000000000..eeb8e67087334ea96aab9cdb6272e34e2eb99939 --- /dev/null +++ b/python/paddle/fluid/tests/book/high-level-api/word2vec/test_word2vec_new_api.py @@ -0,0 +1,172 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import paddle +import paddle.fluid as fluid +import numpy as np +import math +import sys +from functools import partial + +PASS_NUM = 100 +EMBED_SIZE = 32 +HIDDEN_SIZE = 256 +N = 5 +BATCH_SIZE = 32 + +word_dict = paddle.dataset.imikolov.build_dict() +dict_size = len(word_dict) + + +def inference_program(is_sparse): + first_word = fluid.layers.data(name='firstw', shape=[1], dtype='int64') + second_word = fluid.layers.data(name='secondw', shape=[1], dtype='int64') + third_word = fluid.layers.data(name='thirdw', shape=[1], dtype='int64') + forth_word = fluid.layers.data(name='forthw', shape=[1], dtype='int64') + + embed_first = fluid.layers.embedding( + input=first_word, + size=[dict_size, EMBED_SIZE], + dtype='float32', + is_sparse=is_sparse, + param_attr='shared_w') + embed_second = fluid.layers.embedding( + input=second_word, + size=[dict_size, EMBED_SIZE], + dtype='float32', + is_sparse=is_sparse, + param_attr='shared_w') + embed_third = fluid.layers.embedding( + input=third_word, + size=[dict_size, EMBED_SIZE], + dtype='float32', + is_sparse=is_sparse, + param_attr='shared_w') + embed_forth = fluid.layers.embedding( + input=forth_word, + size=[dict_size, EMBED_SIZE], + dtype='float32', + is_sparse=is_sparse, + param_attr='shared_w') + + concat_embed = fluid.layers.concat( + input=[embed_first, embed_second, embed_third, embed_forth], axis=1) + hidden1 = fluid.layers.fc(input=concat_embed, + size=HIDDEN_SIZE, + act='sigmoid') + predict_word = fluid.layers.fc(input=hidden1, size=dict_size, act='softmax') + return predict_word + + +def train_program(is_sparse): + # The declaration of 'next_word' must be after the invoking of inference_program, + # or the data input order of train program would be [next_word, firstw, secondw, + # thirdw, forthw], which is not correct. + predict_word = inference_program(is_sparse) + next_word = fluid.layers.data(name='nextw', shape=[1], dtype='int64') + cost = fluid.layers.cross_entropy(input=predict_word, label=next_word) + avg_cost = fluid.layers.mean(cost) + return avg_cost + + +def train(use_cuda, train_program, params_dirname): + train_reader = paddle.batch( + paddle.dataset.imikolov.train(word_dict, N), BATCH_SIZE) + test_reader = paddle.batch( + paddle.dataset.imikolov.test(word_dict, N), BATCH_SIZE) + + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + + def event_handler(event): + if isinstance(event, fluid.EndStepEvent): + outs = trainer.test( + reader=test_reader, + feed_order=['firstw', 'secondw', 'thirdw', 'forthw', 'nextw']) + avg_cost = outs[0] + print("loss= ", avg_cost) + + if avg_cost < 10.0: + trainer.save_params(params_dirname) + trainer.stop() + + if math.isnan(avg_cost): + sys.exit("got NaN loss, training failed.") + + trainer = fluid.Trainer( + train_func=train_program, + optimizer=fluid.optimizer.SGD(learning_rate=0.001), + place=place) + + trainer.train( + reader=train_reader, + num_epochs=1, + event_handler=event_handler, + feed_order=['firstw', 'secondw', 'thirdw', 'forthw', 'nextw']) + + +def infer(use_cuda, inference_program, params_dirname=None): + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + inferencer = fluid.Inferencer( + infer_func=inference_program, param_path=params_dirname, place=place) + + # Setup inputs by creating 4 LoDTensors representing 4 words. Here each word + # is simply an index to look up for the corresponding word vector and hence + # the shape of word (base_shape) should be [1]. The length-based level of + # detail (lod) info of each LoDtensor should be [[1]] meaning there is only + # one lod_level and there is only one sequence of one word on this level. + # Note that lod info should be a list of lists. + lod = [[1]] + base_shape = [1] + # The range of random integers is [low, high] + first_word = fluid.create_random_int_lodtensor( + lod, base_shape, place, low=0, high=dict_size - 1) + second_word = fluid.create_random_int_lodtensor( + lod, base_shape, place, low=0, high=dict_size - 1) + third_word = fluid.create_random_int_lodtensor( + lod, base_shape, place, low=0, high=dict_size - 1) + fourth_word = fluid.create_random_int_lodtensor( + lod, base_shape, place, low=0, high=dict_size - 1) + + result = inferencer.infer( + { + 'firstw': first_word, + 'secondw': second_word, + 'thirdw': third_word, + 'forthw': fourth_word + }, + return_numpy=False) + print(np.array(result[0])) + + +def main(use_cuda, is_sparse): + if use_cuda and not fluid.core.is_compiled_with_cuda(): + return + + params_dirname = "word2vec.inference.model" + + train( + use_cuda=use_cuda, + train_program=partial(train_program, is_sparse), + params_dirname=params_dirname) + + infer( + use_cuda=use_cuda, + inference_program=partial(inference_program, is_sparse), + params_dirname=params_dirname) + + +if __name__ == '__main__': + for use_cuda in (False, True): + for is_sparse in (False, True): + main(use_cuda=use_cuda, is_sparse=is_sparse) diff --git a/python/paddle/fluid/tests/book/test_understand_sentiment.py b/python/paddle/fluid/tests/book/notest_understand_sentiment.py similarity index 92% rename from python/paddle/fluid/tests/book/test_understand_sentiment.py rename to python/paddle/fluid/tests/book/notest_understand_sentiment.py index dedd153778d7ad9caeb5fa7090a980bc7f177dea..c6687e8ad7fcc45c82d6dcb2256e9055a81cc61c 100644 --- a/python/paddle/fluid/tests/book/test_understand_sentiment.py +++ b/python/paddle/fluid/tests/book/notest_understand_sentiment.py @@ -125,14 +125,6 @@ def stacked_lstm_net(data, return avg_cost, accuracy, prediction -def create_random_lodtensor(lod, place, low, high): - data = np.random.random_integers(low, high, [lod[-1], 1]).astype("int64") - res = fluid.LoDTensor() - res.set(data, place) - res.set_lod([lod]) - return res - - def train(word_dict, net_method, use_cuda, @@ -170,7 +162,7 @@ def train(word_dict, assert save_dirname is None adagrad = fluid.optimizer.Adagrad(learning_rate=0.002) - optimize_ops, params_grads = adagrad.minimize(cost) + adagrad.minimize(cost) train_data = paddle.batch( paddle.reader.shuffle( @@ -213,12 +205,7 @@ def train(word_dict, trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID")) training_role = os.getenv("TRAINING_ROLE", "TRAINER") t = fluid.DistributeTranspiler() - t.transpile( - optimize_ops, - params_grads, - trainer_id, - pservers=pserver_endpoints, - trainers=trainers) + t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers) if training_role == "PSERVER": pserver_prog = t.get_pserver_program(current_endpoint) pserver_startup = t.get_startup_program(current_endpoint, @@ -247,9 +234,21 @@ def infer(word_dict, use_cuda, save_dirname=None): word_dict_len = len(word_dict) - lod = [0, 4, 10] - tensor_words = create_random_lodtensor( - lod, place, low=0, high=word_dict_len - 1) + # Setup input by creating LoDTensor to represent sequence of words. + # Here each word is the basic element of the LoDTensor and the shape of + # each word (base_shape) should be [1] since it is simply an index to + # look up for the corresponding word vector. + # Suppose the length_based level of detail (lod) info is set to [[3, 4, 2]], + # which has only one lod level. Then the created LoDTensor will have only + # one higher level structure (sequence of words, or sentence) than the basic + # element (word). Hence the LoDTensor will hold data for three sentences of + # length 3, 4 and 2, respectively. + # Note that lod info should be a list of lists. + lod = [[3, 4, 2]] + base_shape = [1] + # The range of random integers is [low, high] + tensor_words = fluid.create_random_int_lodtensor( + lod, base_shape, place, low=0, high=word_dict_len - 1) # Construct feed as a dictionary of {feed_target_name: feed_target_data} # and results will contain a list of data corresponding to fetch_targets. diff --git a/python/paddle/fluid/tests/book/test_fit_a_line.py b/python/paddle/fluid/tests/book/test_fit_a_line.py index 6dfc2997ae0328a41fe22d13dfa8fc51d4d021a6..b1a6b524d33cae97c8982ffb8f780b1b07761a09 100644 --- a/python/paddle/fluid/tests/book/test_fit_a_line.py +++ b/python/paddle/fluid/tests/book/test_fit_a_line.py @@ -33,7 +33,7 @@ def train(use_cuda, save_dirname, is_local): avg_cost = fluid.layers.mean(cost) sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.001) - optimize_ops, params_grads = sgd_optimizer.minimize(avg_cost) + sgd_optimizer.minimize(avg_cost) BATCH_SIZE = 20 @@ -80,12 +80,7 @@ def train(use_cuda, save_dirname, is_local): trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID")) training_role = os.getenv("TRAINING_ROLE", "TRAINER") t = fluid.DistributeTranspiler() - t.transpile( - optimize_ops, - params_grads, - trainer_id, - pservers=pserver_endpoints, - trainers=trainers) + t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers) if training_role == "PSERVER": pserver_prog = t.get_pserver_program(current_endpoint) pserver_startup = t.get_startup_program(current_endpoint, diff --git a/python/paddle/fluid/tests/book/test_image_classification.py b/python/paddle/fluid/tests/book/test_image_classification.py index db96c82ce2d8376b029e9dcc54ffab669f1def9a..0f3a4c9242a81a3c1fb90268245715a8e59a207a 100644 --- a/python/paddle/fluid/tests/book/test_image_classification.py +++ b/python/paddle/fluid/tests/book/test_image_classification.py @@ -125,7 +125,7 @@ def train(net_type, use_cuda, save_dirname, is_local): test_program = fluid.default_main_program().clone(for_test=True) optimizer = fluid.optimizer.Adam(learning_rate=0.001) - optimize_ops, params_grads = optimizer.minimize(avg_cost) + optimizer.minimize(avg_cost) BATCH_SIZE = 128 PASS_NUM = 1 @@ -189,12 +189,7 @@ def train(net_type, use_cuda, save_dirname, is_local): trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID")) training_role = os.getenv("TRAINING_ROLE", "TRAINER") t = fluid.DistributeTranspiler() - t.transpile( - optimize_ops, - params_grads, - trainer_id, - pservers=pserver_endpoints, - trainers=trainers) + t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers) if training_role == "PSERVER": pserver_prog = t.get_pserver_program(current_endpoint) pserver_startup = t.get_startup_program(current_endpoint, diff --git a/python/paddle/fluid/tests/book/test_label_semantic_roles.py b/python/paddle/fluid/tests/book/test_label_semantic_roles.py index d9cd76952e31f8185512ab45f9f3ab2ce7d9da48..bc8a1aafc82d62501cecfa71be0cc3851c75eae2 100644 --- a/python/paddle/fluid/tests/book/test_label_semantic_roles.py +++ b/python/paddle/fluid/tests/book/test_label_semantic_roles.py @@ -36,7 +36,7 @@ depth = 8 mix_hidden_lr = 1e-3 IS_SPARSE = True -PASS_NUM = 100 +PASS_NUM = 10 BATCH_SIZE = 10 embedding_name = 'emb' @@ -116,29 +116,6 @@ def db_lstm(word, predicate, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2, mark, return feature_out -def to_lodtensor(data, place): - seq_lens = [len(seq) for seq in data] - cur_len = 0 - lod = [cur_len] - for l in seq_lens: - cur_len += l - lod.append(cur_len) - flattened_data = np.concatenate(data, axis=0).astype("int64") - flattened_data = flattened_data.reshape([len(flattened_data), 1]) - res = fluid.LoDTensor() - res.set(flattened_data, place) - res.set_lod([lod]) - return res - - -def create_random_lodtensor(lod, place, low, high): - data = np.random.random_integers(low, high, [lod[-1], 1]).astype("int64") - res = fluid.LoDTensor() - res.set(data, place) - res.set_lod([lod]) - return res - - def train(use_cuda, save_dirname=None, is_local=True): # define network topology word = fluid.layers.data( @@ -175,19 +152,13 @@ def train(use_cuda, save_dirname=None, is_local=True): decay_steps=100000, decay_rate=0.5, staircase=True)) - optimize_ops, params_grads = sgd_optimizer.minimize(avg_cost) + sgd_optimizer.minimize(avg_cost) # TODO(qiao) # add dependency track and move this config before optimizer crf_decode = fluid.layers.crf_decoding( input=feature_out, param_attr=fluid.ParamAttr(name='crfw')) - chunk_evaluator = fluid.evaluator.ChunkEvaluator( - input=crf_decode, - label=target, - chunk_scheme="IOB", - num_chunk_types=int(math.ceil((label_dict_len - 1) / 2.0))) - train_data = paddle.batch( paddle.reader.shuffle( paddle.dataset.conll05.test(), buf_size=8192), @@ -203,7 +174,6 @@ def train(use_cuda, save_dirname=None, is_local=True): def train_loop(main_program): exe.run(fluid.default_startup_program()) - embedding_param = fluid.global_scope().find_var( embedding_name).get_tensor() embedding_param.set( @@ -213,27 +183,19 @@ def train(use_cuda, save_dirname=None, is_local=True): start_time = time.time() batch_id = 0 for pass_id in xrange(PASS_NUM): - chunk_evaluator.reset(exe) for data in train_data(): - cost, precision, recall, f1_score = exe.run( - main_program, - feed=feeder.feed(data), - fetch_list=[avg_cost] + chunk_evaluator.metrics) - pass_precision, pass_recall, pass_f1_score = chunk_evaluator.eval( - exe) + cost = exe.run(main_program, + feed=feeder.feed(data), + fetch_list=[avg_cost]) + cost = cost[0] if batch_id % 10 == 0: - print("avg_cost:" + str(cost) + " precision:" + str( - precision) + " recall:" + str(recall) + " f1_score:" + - str(f1_score) + " pass_precision:" + str( - pass_precision) + " pass_recall:" + str( - pass_recall) + " pass_f1_score:" + str( - pass_f1_score)) + print("avg_cost:" + str(cost)) if batch_id != 0: print("second per batch: " + str((time.time( ) - start_time) / batch_id)) # Set the threshold low to speed up the CI test - if float(pass_precision) > 0.01: + if float(cost) < 60.0: if save_dirname is not None: # TODO(liuyiqun): Change the target to crf_decode fluid.io.save_inference_model(save_dirname, [ @@ -259,12 +221,7 @@ def train(use_cuda, save_dirname=None, is_local=True): trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID")) training_role = os.getenv("TRAINING_ROLE", "TRAINER") t = fluid.DistributeTranspiler() - t.transpile( - optimize_ops, - params_grads, - trainer_id, - pservers=pserver_endpoints, - trainers=trainers) + t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers) if training_role == "PSERVER": pserver_prog = t.get_pserver_program(current_endpoint) pserver_startup = t.get_startup_program(current_endpoint, @@ -291,23 +248,35 @@ def infer(use_cuda, save_dirname=None): [inference_program, feed_target_names, fetch_targets] = fluid.io.load_inference_model(save_dirname, exe) - lod = [0, 4, 10] - word = create_random_lodtensor( - lod, place, low=0, high=word_dict_len - 1) - pred = create_random_lodtensor( - lod, place, low=0, high=pred_dict_len - 1) - ctx_n2 = create_random_lodtensor( - lod, place, low=0, high=word_dict_len - 1) - ctx_n1 = create_random_lodtensor( - lod, place, low=0, high=word_dict_len - 1) - ctx_0 = create_random_lodtensor( - lod, place, low=0, high=word_dict_len - 1) - ctx_p1 = create_random_lodtensor( - lod, place, low=0, high=word_dict_len - 1) - ctx_p2 = create_random_lodtensor( - lod, place, low=0, high=word_dict_len - 1) - mark = create_random_lodtensor( - lod, place, low=0, high=mark_dict_len - 1) + # Setup inputs by creating LoDTensors to represent sequences of words. + # Here each word is the basic element of these LoDTensors and the shape of + # each word (base_shape) should be [1] since it is simply an index to + # look up for the corresponding word vector. + # Suppose the length_based level of detail (lod) info is set to [[3, 4, 2]], + # which has only one lod level. Then the created LoDTensors will have only + # one higher level structure (sequence of words, or sentence) than the basic + # element (word). Hence the LoDTensor will hold data for three sentences of + # length 3, 4 and 2, respectively. + # Note that lod info should be a list of lists. + lod = [[3, 4, 2]] + base_shape = [1] + # The range of random integers is [low, high] + word = fluid.create_random_int_lodtensor( + lod, base_shape, place, low=0, high=word_dict_len - 1) + pred = fluid.create_random_int_lodtensor( + lod, base_shape, place, low=0, high=pred_dict_len - 1) + ctx_n2 = fluid.create_random_int_lodtensor( + lod, base_shape, place, low=0, high=word_dict_len - 1) + ctx_n1 = fluid.create_random_int_lodtensor( + lod, base_shape, place, low=0, high=word_dict_len - 1) + ctx_0 = fluid.create_random_int_lodtensor( + lod, base_shape, place, low=0, high=word_dict_len - 1) + ctx_p1 = fluid.create_random_int_lodtensor( + lod, base_shape, place, low=0, high=word_dict_len - 1) + ctx_p2 = fluid.create_random_int_lodtensor( + lod, base_shape, place, low=0, high=word_dict_len - 1) + mark = fluid.create_random_int_lodtensor( + lod, base_shape, place, low=0, high=mark_dict_len - 1) # Construct feed as a dictionary of {feed_target_name: feed_target_data} # and results will contain a list of data corresponding to fetch_targets. diff --git a/python/paddle/fluid/tests/book/test_machine_translation.py b/python/paddle/fluid/tests/book/test_machine_translation.py index 830d78df8b9e56b45f7e928562ef4b89e88f696d..23e5900f127a7a3253c551f8f7fbceba08382209 100644 --- a/python/paddle/fluid/tests/book/test_machine_translation.py +++ b/python/paddle/fluid/tests/book/test_machine_translation.py @@ -147,28 +147,6 @@ def decoder_decode(context, is_sparse): return translation_ids, translation_scores -def set_init_lod(data, lod, place): - res = fluid.LoDTensor() - res.set(data, place) - res.set_lod(lod) - return res - - -def to_lodtensor(data, place): - seq_lens = [len(seq) for seq in data] - cur_len = 0 - lod = [cur_len] - for l in seq_lens: - cur_len += l - lod.append(cur_len) - flattened_data = np.concatenate(data, axis=0).astype("int64") - flattened_data = flattened_data.reshape([len(flattened_data), 1]) - res = fluid.LoDTensor() - res.set(flattened_data, place) - res.set_lod([lod]) - return res - - def train_main(use_cuda, is_sparse, is_local=True): if use_cuda and not fluid.core.is_compiled_with_cuda(): return @@ -185,30 +163,32 @@ def train_main(use_cuda, is_sparse, is_local=True): learning_rate=1e-4, regularization=fluid.regularizer.L2DecayRegularizer( regularization_coeff=0.1)) - optimize_ops, params_grads = optimizer.minimize(avg_cost) + optimizer.minimize(avg_cost) train_data = paddle.batch( paddle.reader.shuffle( paddle.dataset.wmt14.train(dict_size), buf_size=1000), batch_size=batch_size) + feed_order = [ + 'src_word_id', 'target_language_word', 'target_language_next_word' + ] + exe = Executor(place) def train_loop(main_program): exe.run(framework.default_startup_program()) + feed_list = [ + main_program.global_block().var(var_name) for var_name in feed_order + ] + feeder = fluid.DataFeeder(feed_list, place) + batch_id = 0 for pass_id in xrange(1): for data in train_data(): - word_data = to_lodtensor(map(lambda x: x[0], data), place) - trg_word = to_lodtensor(map(lambda x: x[1], data), place) - trg_word_next = to_lodtensor(map(lambda x: x[2], data), place) outs = exe.run(main_program, - feed={ - 'src_word_id': word_data, - 'target_language_word': trg_word, - 'target_language_next_word': trg_word_next - }, + feed=feeder.feed(data), fetch_list=[avg_cost]) avg_cost_val = np.array(outs[0]) print('pass_id=' + str(pass_id) + ' batch=' + str(batch_id) + @@ -231,12 +211,7 @@ def train_main(use_cuda, is_sparse, is_local=True): trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID")) training_role = os.getenv("TRAINING_ROLE", "TRAINER") t = fluid.DistributeTranspiler() - t.transpile( - optimize_ops, - params_grads, - trainer_id, - pservers=pserver_endpoints, - trainers=trainers) + t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers) if training_role == "PSERVER": pserver_prog = t.get_pserver_program(current_endpoint) pserver_startup = t.get_startup_program(current_endpoint, @@ -263,26 +238,32 @@ def decode_main(use_cuda, is_sparse): [1. for _ in range(batch_size)], dtype='float32') init_ids_data = init_ids_data.reshape((batch_size, 1)) init_scores_data = init_scores_data.reshape((batch_size, 1)) - init_lod = [i for i in range(batch_size)] + [batch_size] + init_lod = [1] * batch_size init_lod = [init_lod, init_lod] + init_ids = fluid.create_lod_tensor(init_ids_data, init_lod, place) + init_scores = fluid.create_lod_tensor(init_scores_data, init_lod, place) + train_data = paddle.batch( paddle.reader.shuffle( paddle.dataset.wmt14.train(dict_size), buf_size=1000), batch_size=batch_size) - for _, data in enumerate(train_data()): - init_ids = set_init_lod(init_ids_data, init_lod, place) - init_scores = set_init_lod(init_scores_data, init_lod, place) - src_word_data = to_lodtensor(map(lambda x: x[0], data), place) + feed_order = ['src_word_id'] + feed_list = [ + framework.default_main_program().global_block().var(var_name) + for var_name in feed_order + ] + feeder = fluid.DataFeeder(feed_list, place) + + for data in train_data(): + feed_dict = feeder.feed(map(lambda x: [x[0]], data)) + feed_dict['init_ids'] = init_ids + feed_dict['init_scores'] = init_scores result_ids, result_scores = exe.run( framework.default_main_program(), - feed={ - 'src_word_id': src_word_data, - 'init_ids': init_ids, - 'init_scores': init_scores - }, + feed=feed_dict, fetch_list=[translation_ids, translation_scores], return_numpy=False) print result_ids.lod() diff --git a/python/paddle/fluid/tests/book/test_recognize_digits.py b/python/paddle/fluid/tests/book/test_recognize_digits.py index 5ec6890c1b0dabd2804a92071b63c9610299e67c..578b1162fbd7e3a1b1c0cc934406818f2e07e019 100644 --- a/python/paddle/fluid/tests/book/test_recognize_digits.py +++ b/python/paddle/fluid/tests/book/test_recognize_digits.py @@ -95,7 +95,7 @@ def train(nn_type, test_program = fluid.default_main_program().clone(for_test=True) optimizer = fluid.optimizer.Adam(learning_rate=0.001) - optimize_ops, params_grads = optimizer.minimize(avg_loss) + optimizer.minimize(avg_loss) place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() @@ -162,12 +162,7 @@ def train(nn_type, trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID")) training_role = os.getenv("TRAINING_ROLE", "TRAINER") t = fluid.DistributeTranspiler() - t.transpile( - optimize_ops, - params_grads, - trainer_id, - pservers=pserver_endpoints, - trainers=trainers) + t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers) if training_role == "PSERVER": pserver_prog = t.get_pserver_program(current_endpoint) pserver_startup = t.get_startup_program(current_endpoint, diff --git a/python/paddle/fluid/tests/book/test_recommender_system.py b/python/paddle/fluid/tests/book/test_recommender_system.py index 2172c275b8082689a6ff5f2c3c27a2ff4e92275a..65d6552acc9b3d31a97a45290e4613a633fffa3c 100644 --- a/python/paddle/fluid/tests/book/test_recommender_system.py +++ b/python/paddle/fluid/tests/book/test_recommender_system.py @@ -160,7 +160,7 @@ def train(use_cuda, save_dirname, is_local=True): test_program = fluid.default_main_program().clone(for_test=True) sgd_optimizer = SGDOptimizer(learning_rate=0.2) - optimize_ops, params_grads = sgd_optimizer.minimize(avg_cost) + sgd_optimizer.minimize(avg_cost) place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() @@ -173,63 +173,33 @@ def train(use_cuda, save_dirname, is_local=True): test_reader = paddle.batch( paddle.dataset.movielens.test(), batch_size=BATCH_SIZE) - feeding = { - 'user_id': 0, - 'gender_id': 1, - 'age_id': 2, - 'job_id': 3, - 'movie_id': 4, - 'category_id': 5, - 'movie_title': 6, - 'score': 7 - } - - def func_feed(feeding, data): - feed_tensors = {} - for (key, idx) in feeding.iteritems(): - tensor = fluid.LoDTensor() - if key != "category_id" and key != "movie_title": - if key == "score": - numpy_data = np.array(map(lambda x: x[idx], data)).astype( - "float32") - else: - numpy_data = np.array(map(lambda x: x[idx], data)).astype( - "int64") - else: - numpy_data = map(lambda x: np.array(x[idx]).astype("int64"), - data) - lod_info = [len(item) for item in numpy_data] - offset = 0 - lod = [offset] - for item in lod_info: - offset += item - lod.append(offset) - numpy_data = np.concatenate(numpy_data, axis=0) - tensor.set_lod([lod]) - - numpy_data = numpy_data.reshape([numpy_data.shape[0], 1]) - tensor.set(numpy_data, place) - feed_tensors[key] = tensor - return feed_tensors + feed_order = [ + 'user_id', 'gender_id', 'age_id', 'job_id', 'movie_id', 'category_id', + 'movie_title', 'score' + ] def train_loop(main_program): exe.run(framework.default_startup_program()) + feed_list = [ + main_program.global_block().var(var_name) for var_name in feed_order + ] + feeder = fluid.DataFeeder(feed_list, place) + PASS_NUM = 100 for pass_id in range(PASS_NUM): for batch_id, data in enumerate(train_reader()): # train a mini-batch outs = exe.run(program=main_program, - feed=func_feed(feeding, data), + feed=feeder.feed(data), fetch_list=[avg_cost]) out = np.array(outs[0]) if (batch_id + 1) % 10 == 0: avg_cost_set = [] for test_data in test_reader(): - avg_cost_np = exe.run( - program=test_program, - feed=func_feed(feeding, test_data), - fetch_list=[avg_cost]) + avg_cost_np = exe.run(program=test_program, + feed=feeder.feed(test_data), + fetch_list=[avg_cost]) avg_cost_set.append(avg_cost_np[0]) break # test only 1 segment for speeding up CI @@ -261,12 +231,7 @@ def train(use_cuda, save_dirname, is_local=True): trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID")) training_role = os.getenv("TRAINING_ROLE", "TRAINER") t = fluid.DistributeTranspiler() - t.transpile( - optimize_ops, - params_grads, - trainer_id, - pservers=pserver_endpoints, - trainers=trainers) + t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers) if training_role == "PSERVER": pserver_prog = t.get_pserver_program(current_endpoint) pserver_startup = t.get_startup_program(current_endpoint, @@ -284,23 +249,6 @@ def infer(use_cuda, save_dirname=None): place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() exe = fluid.Executor(place) - def create_lod_tensor(data, lod=None): - tensor = fluid.LoDTensor() - if lod is None: - # Tensor, the shape is [batch_size, 1] - index = 0 - lod_0 = [index] - for l in range(len(data)): - index += 1 - lod_0.append(index) - lod = [lod_0] - tensor.set_lod(lod) - - flattened_data = np.concatenate(data, axis=0).astype("int64") - flattened_data = flattened_data.reshape([len(flattened_data), 1]) - tensor.set(flattened_data, place) - return tensor - inference_scope = fluid.core.Scope() with fluid.scope_guard(inference_scope): # Use fluid.io.load_inference_model to obtain the inference program desc, @@ -312,26 +260,33 @@ def infer(use_cuda, save_dirname=None): # Use the first data from paddle.dataset.movielens.test() as input assert feed_target_names[0] == "user_id" - user_id = create_lod_tensor([[1]]) + # Use create_lod_tensor(data, lod, place) API to generate LoD Tensor + # where `data` is a list of sequences of index numbers, `lod` is + # the level of detail (lod) info associated with `data`. + # For example, data = [[10, 2, 3], [2, 3]] means that it contains + # two sequences of indexes, of length 3 and 2, respectively. + # Correspondingly, lod = [[3, 2]] contains one level of detail info, + # indicating that `data` consists of two sequences of length 3 and 2. + user_id = fluid.create_lod_tensor([[1]], [[1]], place) assert feed_target_names[1] == "gender_id" - gender_id = create_lod_tensor([[1]]) + gender_id = fluid.create_lod_tensor([[1]], [[1]], place) assert feed_target_names[2] == "age_id" - age_id = create_lod_tensor([[0]]) + age_id = fluid.create_lod_tensor([[0]], [[1]], place) assert feed_target_names[3] == "job_id" - job_id = create_lod_tensor([[10]]) + job_id = fluid.create_lod_tensor([[10]], [[1]], place) assert feed_target_names[4] == "movie_id" - movie_id = create_lod_tensor([[783]]) + movie_id = fluid.create_lod_tensor([[783]], [[1]], place) assert feed_target_names[5] == "category_id" - category_id = create_lod_tensor([[10], [8], [9]], [[0, 3]]) + category_id = fluid.create_lod_tensor([[10, 8, 9]], [[3]], place) assert feed_target_names[6] == "movie_title" - movie_title = create_lod_tensor([[1069], [4140], [2923], [710], [988]], - [[0, 5]]) + movie_title = fluid.create_lod_tensor([[1069, 4140, 2923, 710, 988]], + [[5]], place) # Construct feed as a dictionary of {feed_target_name: feed_target_data} # and results will contain a list of data corresponding to fetch_targets. diff --git a/python/paddle/fluid/tests/book/notest_rnn_encoder_decoer.py b/python/paddle/fluid/tests/book/test_rnn_encoder_decoder.py similarity index 87% rename from python/paddle/fluid/tests/book/notest_rnn_encoder_decoer.py rename to python/paddle/fluid/tests/book/test_rnn_encoder_decoder.py index ce640dece8a5067bd10f410a2bb58874b7cc0908..7ada57def6bfedb113ea1a56f9677116b80488ea 100644 --- a/python/paddle/fluid/tests/book/notest_rnn_encoder_decoer.py +++ b/python/paddle/fluid/tests/book/test_rnn_encoder_decoder.py @@ -152,29 +152,6 @@ def seq_to_seq_net(): return avg_cost, prediction -def to_lodtensor(data, place): - seq_lens = [len(seq) for seq in data] - cur_len = 0 - lod = [cur_len] - for l in seq_lens: - cur_len += l - lod.append(cur_len) - flattened_data = np.concatenate(data, axis=0).astype("int64") - flattened_data = flattened_data.reshape([len(flattened_data), 1]) - res = core.LoDTensor() - res.set(flattened_data, place) - res.set_lod([lod]) - return res - - -def create_random_lodtensor(lod, place, low, high): - data = np.random.random_integers(low, high, [lod[-1], 1]).astype("int64") - res = fluid.LoDTensor() - res.set(data, place) - res.set_lod([lod]) - return res - - def train(use_cuda, save_dirname=None): [avg_cost, prediction] = seq_to_seq_net() @@ -188,22 +165,20 @@ def train(use_cuda, save_dirname=None): place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() exe = Executor(place) - exe.run(framework.default_startup_program()) + feed_order = ['source_sequence', 'target_sequence', 'label_sequence'] + feed_list = [ + framework.default_main_program().global_block().var(var_name) + for var_name in feed_order + ] + feeder = fluid.DataFeeder(feed_list, place) + batch_id = 0 for pass_id in xrange(2): for data in train_data(): - word_data = to_lodtensor(map(lambda x: x[0], data), place) - trg_word = to_lodtensor(map(lambda x: x[1], data), place) - trg_word_next = to_lodtensor(map(lambda x: x[2], data), place) - outs = exe.run(framework.default_main_program(), - feed={ - 'source_sequence': word_data, - 'target_sequence': trg_word, - 'label_sequence': trg_word_next - }, + feed=feeder.feed(data), fetch_list=[avg_cost]) avg_cost_val = np.array(outs[0]) @@ -237,9 +212,23 @@ def infer(use_cuda, save_dirname=None): [inference_program, feed_target_names, fetch_targets] = fluid.io.load_inference_model(save_dirname, exe) - lod = [0, 4, 10] - word_data = create_random_lodtensor(lod, place, low=0, high=1) - trg_word = create_random_lodtensor(lod, place, low=0, high=1) + # Setup input by creating LoDTensor to represent sequence of words. + # Here each word is the basic element of the LoDTensor and the shape of + # each word (base_shape) should be [1] since it is simply an index to + # look up for the corresponding word vector. + # Suppose the length_based level of detail (lod) info is set to [[4, 6]], + # which has only one lod level. Then the created LoDTensor will have only + # one higher level structure (sequence of words, or sentence) than the basic + # element (word). Hence the LoDTensor will hold data for two sentences of + # length 4 and 6, respectively. + # Note that lod info should be a list of lists. + lod = [[4, 6]] + base_shape = [1] + # The range of random integers is [low, high] + word_data = fluid.create_random_int_lodtensor( + lod, base_shape, place, low=0, high=1) + trg_word = fluid.create_random_int_lodtensor( + lod, base_shape, place, low=0, high=1) # Construct feed as a dictionary of {feed_target_name: feed_target_data} # and results will contain a list of data corresponding to fetch_targets. diff --git a/python/paddle/fluid/tests/book/test_word2vec.py b/python/paddle/fluid/tests/book/test_word2vec.py index 8929779de9448d036e1528b64330b37463ab3988..3118d88701e5f64ae50f7ee774ea8174aa7758eb 100644 --- a/python/paddle/fluid/tests/book/test_word2vec.py +++ b/python/paddle/fluid/tests/book/test_word2vec.py @@ -21,15 +21,6 @@ import math import sys -def create_random_lodtensor(lod, place, low, high): - # The range of data elements is [low, high] - data = np.random.random_integers(low, high, [lod[-1], 1]).astype("int64") - res = fluid.LoDTensor() - res.set(data, place) - res.set_lod([lod]) - return res - - def train(use_cuda, is_sparse, is_parallel, save_dirname, is_local=True): PASS_NUM = 100 EMBED_SIZE = 32 @@ -101,7 +92,7 @@ def train(use_cuda, is_sparse, is_parallel, save_dirname, is_local=True): avg_cost = fluid.layers.mean(pd()) sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.001) - optimize_ops, params_grads = sgd_optimizer.minimize(avg_cost) + sgd_optimizer.minimize(avg_cost) train_reader = paddle.batch( paddle.dataset.imikolov.train(word_dict, N), BATCH_SIZE) @@ -145,12 +136,7 @@ def train(use_cuda, is_sparse, is_parallel, save_dirname, is_local=True): trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID")) training_role = os.getenv("TRAINING_ROLE", "TRAINER") t = fluid.DistributeTranspiler() - t.transpile( - optimize_ops, - params_grads, - trainer_id, - pservers=pserver_endpoints, - trainers=trainers) + t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers) if training_role == "PSERVER": pserver_prog = t.get_pserver_program(current_endpoint) pserver_startup = t.get_startup_program(current_endpoint, @@ -180,16 +166,23 @@ def infer(use_cuda, save_dirname=None): word_dict = paddle.dataset.imikolov.build_dict() dict_size = len(word_dict) - # Setup inputs, by creating 4 words, the lod of which should be [0, 1] - lod = [0, 1] - first_word = create_random_lodtensor( - lod, place, low=0, high=dict_size - 1) - second_word = create_random_lodtensor( - lod, place, low=0, high=dict_size - 1) - third_word = create_random_lodtensor( - lod, place, low=0, high=dict_size - 1) - fourth_word = create_random_lodtensor( - lod, place, low=0, high=dict_size - 1) + # Setup inputs by creating 4 LoDTensors representing 4 words. Here each word + # is simply an index to look up for the corresponding word vector and hence + # the shape of word (base_shape) should be [1]. The length-based level of + # detail (lod) info of each LoDtensor should be [[1]] meaning there is only + # one lod_level and there is only one sequence of one word on this level. + # Note that lod info should be a list of lists. + lod = [[1]] + base_shape = [1] + # The range of random integers is [low, high] + first_word = fluid.create_random_int_lodtensor( + lod, base_shape, place, low=0, high=dict_size - 1) + second_word = fluid.create_random_int_lodtensor( + lod, base_shape, place, low=0, high=dict_size - 1) + third_word = fluid.create_random_int_lodtensor( + lod, base_shape, place, low=0, high=dict_size - 1) + fourth_word = fluid.create_random_int_lodtensor( + lod, base_shape, place, low=0, high=dict_size - 1) assert feed_target_names[0] == 'firstw' assert feed_target_names[1] == 'secondw' diff --git a/python/paddle/fluid/tests/test_cpp_reader.py b/python/paddle/fluid/tests/test_cpp_reader.py index e54c73b2956dd99ee57804318130c261e133d21a..6cc291dfcffdd7083f498389834e37bd06ca4572 100644 --- a/python/paddle/fluid/tests/test_cpp_reader.py +++ b/python/paddle/fluid/tests/test_cpp_reader.py @@ -44,8 +44,8 @@ create_random_data_generator_op = startup_block.append_op( attrs={ "shape_concat": [1, 2, 1, 1], "ranks": [2, 2], - "min": 0.0, - "max": 1.0, + "low": 0.0, + "high": 1.0, 'lod_levels': [0, 0] }) diff --git a/python/paddle/fluid/tests/test_data_feeder.py b/python/paddle/fluid/tests/test_data_feeder.py index 861dd3174a21d59fe12e0b794ecb2a934946ac71..ce3ba3ebc50d7b015f379b5e80b179463a7b231a 100644 --- a/python/paddle/fluid/tests/test_data_feeder.py +++ b/python/paddle/fluid/tests/test_data_feeder.py @@ -13,15 +13,62 @@ # limitations under the License. import paddle.fluid as fluid +import unittest -def test_converter(): - img = fluid.layers.data(name='image', shape=[1, 28, 28]) - label = fluid.layers.data(name='label', shape=[1], dtype='int64') - feeder = fluid.DataFeeder([img, label], fluid.CPUPlace()) - result = feeder.feed([[[0] * 784, [9]], [[1] * 784, [1]]]) - print(result) +class TestDataFeeder(unittest.TestCase): + def test_lod_level_0_converter(self): + img = fluid.layers.data(name='image', shape=[1, 28, 28]) + label = fluid.layers.data(name='label', shape=[1], dtype='int64') + feeder = fluid.DataFeeder([img, label], fluid.CPUPlace()) + result = feeder.feed([([0] * 784, [9]), ([1] * 784, [1])]) + print(result) + + self.assertEqual(result['image'].shape(), [2, 1, 28, 28]) + self.assertEqual(result['label'].shape(), [2, 1]) + self.assertEqual(result['image'].lod(), []) + self.assertEqual(result['label'].lod(), []) + + def test_lod_level_1_converter(self): + # lod_level = 1 + # each sentence has a different number of words + sentences = fluid.layers.data( + name='sentences', shape=[1], dtype='int64', lod_level=1) + label = fluid.layers.data(name='label', shape=[1], dtype='int64') + feeder = fluid.DataFeeder([sentences, label], fluid.CPUPlace()) + + # lod = [[0, 3, 5, 9]] + # data = [[1, 2, 3], [4, 5], [6, 7, 8, 9]] + # label = [1] * len(data) + result = feeder.feed( + [([1, 2, 3], [1]), ([4, 5], [1]), ([6, 7, 8, 9], [1])]) + print(result) + + self.assertEqual(result['sentences'].shape(), [9, 1]) + self.assertEqual(result['label'].shape(), [3, 1]) + self.assertEqual(result['sentences'].lod(), [[0, 3, 5, 9]]) + self.assertEqual(result['label'].lod(), []) + + def test_lod_level_2_converter(self): + # lod_level = 2 + # paragraphs -> sentences -> words + paragraphs = fluid.layers.data( + name='paragraphs', shape=[1], dtype='int64', lod_level=2) + label = fluid.layers.data(name='label', shape=[1], dtype='int64') + feeder = fluid.DataFeeder([paragraphs, label], fluid.CPUPlace()) + + # lod = [[0, 2, 3], [0, 3, 5, 9]] + # data = [[[1, 2, 3], [4, 5]], [[6, 7, 8, 9]]] + # label = [1] * len(data) + result = feeder.feed( + [([[1, 2, 3], [4, 5]], [1]), ([[6, 7, 8, 9]], [1])]) + print(result) + + self.assertEqual(result['paragraphs'].shape(), [9, 1]) + self.assertEqual(result['label'].shape(), [2, 1]) + self.assertEqual(result['paragraphs'].lod(), [[0, 2, 3], [0, 3, 5, 9]]) + self.assertEqual(result['label'].lod(), []) if __name__ == '__main__': - test_converter() + unittest.main() diff --git a/python/paddle/fluid/tests/test_detection.py b/python/paddle/fluid/tests/test_detection.py index 921260ef3f4b1f9e4c65b3ffb440dc34cb0a9376..8569d838bdd414eb84c6c87674990a25a2fdcdf9 100644 --- a/python/paddle/fluid/tests/test_detection.py +++ b/python/paddle/fluid/tests/test_detection.py @@ -109,6 +109,24 @@ class TestDetection(unittest.TestCase): print(str(program)) +class TestPriorBox(unittest.TestCase): + def test_prior_box(self): + data_shape = [3, 224, 224] + images = fluid.layers.data( + name='pixel', shape=data_shape, dtype='float32') + conv1 = fluid.layers.conv2d(images, 3, 3, 2) + box, var = layers.prior_box( + input=conv1, + image=images, + min_sizes=[100.0], + aspect_ratios=[1.], + flip=True, + clip=True) + assert len(box.shape) == 4 + assert box.shape == var.shape + assert box.shape[3] == 4 + + class TestMultiBoxHead(unittest.TestCase): def test_multi_box_head(self): data_shape = [3, 224, 224] diff --git a/python/paddle/fluid/tests/test_lod_tensor.py b/python/paddle/fluid/tests/test_lod_tensor.py new file mode 100644 index 0000000000000000000000000000000000000000..013d72f418cf7ac11eb31fd221052039e896e203 --- /dev/null +++ b/python/paddle/fluid/tests/test_lod_tensor.py @@ -0,0 +1,91 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import paddle.fluid as fluid +from paddle.fluid.lod_tensor import create_lod_tensor, create_random_int_lodtensor, _validate_lod, _convert_lod +import numpy +import unittest + + +class TestLoDTensor(unittest.TestCase): + def test_validate_lod(self): + lod = (1, 2, 1) + self.assertRaises(AssertionError, _validate_lod, lod, -1) + lod = [[1, 2], (2, 3)] + self.assertRaises(AssertionError, _validate_lod, lod, -1) + lod = [1, 2, 3] + self.assertRaises(AssertionError, _validate_lod, lod, -1) + + lod = [] + self.assertTrue(_validate_lod(lod, -1)) + lod = [[], [1], [3]] + self.assertFalse(_validate_lod(lod, -1)) + lod = [[0], [-1], [3]] + self.assertFalse(_validate_lod(lod, -1)) + + # Each level's sum should be equal to the number of items in the next level + # Moreover, last level's sum should be equal to the tensor height + lod = [[2, 3], [1, 3, 1, 2, 1]] + self.assertTrue(_validate_lod(lod, tensor_height=8)) + lod = [[1, 3], [2, 1, 3]] + self.assertFalse(_validate_lod(lod, tensor_height=6)) + lod = [[1, 3], [2, 1, 3, 4]] + self.assertFalse(_validate_lod(lod, tensor_height=5)) + + def test_convert_lod(self): + lod = [[1, 2, 3]] + converted_lod = [[0, 1, 3, 6]] + self.assertEqual(_convert_lod(lod), converted_lod) + + lod = [[2, 3], [1, 3, 1, 2, 1]] + converted_lod = [[0, 2, 5], [0, 1, 4, 5, 7, 8]] + self.assertEqual(_convert_lod(lod), converted_lod) + + def test_create_lod_tensor(self): + # Create LoDTensor from a list + data = [[1, 2, 3], [3, 4]] + wrong_lod = [[2, 2]] + correct_lod = [[3, 2]] + self.assertRaises(AssertionError, create_lod_tensor, data, wrong_lod, + fluid.CPUPlace()) + tensor = create_lod_tensor(data, correct_lod, fluid.CPUPlace()) + self.assertEqual(tensor.lod(), [[0, 3, 5]]) + + # Create LoDTensor from numpy array + data = numpy.random.random([10, 1]) + lod = [[2, 1], [3, 3, 4]] + tensor = create_lod_tensor(data, lod, fluid.CPUPlace()) + self.assertEqual(tensor.lod(), [[0, 2, 3], [0, 3, 6, 10]]) + + # Create LoDTensor from another LoDTensor, they are differnt instances + new_lod = [[2, 2, 1], [1, 2, 2, 3, 2]] + new_tensor = create_lod_tensor(tensor, new_lod, fluid.CPUPlace()) + self.assertEqual(tensor.lod(), [[0, 2, 3], [0, 3, 6, 10]]) + self.assertEqual(new_tensor.lod(), [[0, 2, 4, 5], [0, 1, 3, 5, 8, 10]]) + + def test_create_random_int_lodtensor(self): + # The shape of a word, commonly used in speech and NLP problem, is [1] + shape = [1] + lod = [[2, 3, 5]] + dict_size = 10000 + low = 0 + high = dict_size - 1 + tensor = create_random_int_lodtensor(lod, shape, + fluid.CPUPlace(), low, high) + self.assertEqual(tensor.lod(), [[0, 2, 5, 10]]) + self.assertEqual(tensor.shape(), [10, 1]) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/CMakeLists.txt b/python/paddle/fluid/tests/unittests/CMakeLists.txt index d9190408e151283ece8460286dd67818dd39da3e..fead95ffdab25c7ea96b7ef223efc0abf7eea3e3 100644 --- a/python/paddle/fluid/tests/unittests/CMakeLists.txt +++ b/python/paddle/fluid/tests/unittests/CMakeLists.txt @@ -17,7 +17,7 @@ endif(NOT WITH_DISTRIBUTE) list(REMOVE_ITEM TEST_OPS test_seq_concat_op) # FIXME(helin): https://github.com/PaddlePaddle/Paddle/issues/8290 list(REMOVE_ITEM TEST_OPS test_modified_huber_loss_op) # FIXME(qijun) https://github.com/PaddlePaddle/Paddle/issues/5184 list(REMOVE_ITEM TEST_OPS test_lstm_unit_op) # # FIXME(qijun) https://github.com/PaddlePaddle/Paddle/issues/5185 -list(REMOVE_ITEM TEST_OPS test_nce) # IXME(qijun) https://github.com/PaddlePaddle/Paddle/issues/7778 +list(REMOVE_ITEM TEST_OPS test_nce) # FIXME(qijun) https://github.com/PaddlePaddle/Paddle/issues/7778 list(REMOVE_ITEM TEST_OPS test_recurrent_op) # FIXME(qijun) https://github.com/PaddlePaddle/Paddle/issues/6152 list(REMOVE_ITEM TEST_OPS test_cond_op) # FIXME(qijun): https://github.com/PaddlePaddle/Paddle/issues/5101#issuecomment-339814957 @@ -26,82 +26,27 @@ list(REMOVE_ITEM TEST_OPS decorators) # decorators is a helper python file, not function(py_test_modules TARGET_NAME) if(WITH_TESTING) - set(options "") + set(options SERIAL) set(oneValueArgs "") - set(multiValueArgs MODULES DEPS ARGS ENVS) + set(multiValueArgs MODULES DEPS ENVS) cmake_parse_arguments(py_test_modules "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) add_test(NAME ${TARGET_NAME} COMMAND env PYTHONPATH=${PADDLE_BINARY_DIR}/python ${py_test_modules_ENVS} - ${PYTHON_EXECUTABLE} -u -m unittest --verbose ${py_test_modules_MODULES} ${py_test_modules_ARGS} + ${PYTHON_EXECUTABLE} ${PADDLE_SOURCE_DIR}/tools/test_runner.py ${py_test_modules_MODULES} WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}) + if (py_test_modules_SERIAL) + set_property(TEST ${TARGET_NAME} PROPERTY SERIAL 1) + endif() endif() endfunction() - -list(REMOVE_ITEM TEST_OPS test_sequence_expand) - -# test time consuming OPs in a separate process for expliot parallism -list(REMOVE_ITEM TEST_OPS test_parallel_executor) list(REMOVE_ITEM TEST_OPS test_warpctc_op) -list(REMOVE_ITEM TEST_OPS test_dyn_rnn) -list(REMOVE_ITEM TEST_OPS test_mul_op) - -# tests that need to be run in separate process. -list(REMOVE_ITEM TEST_OPS test_multihead_attention) -list(REMOVE_ITEM TEST_OPS test_calc_gradient) -list(REMOVE_ITEM TEST_OPS test_while_op) -list(REMOVE_ITEM TEST_OPS test_lod_array_length_op) -list(REMOVE_ITEM TEST_OPS test_reorder_lod_tensor) -list(REMOVE_ITEM TEST_OPS test_profiler) -list(REMOVE_ITEM TEST_OPS test_nvprof) -list(REMOVE_ITEM TEST_OPS test_normalization_wrapper) -list(REMOVE_ITEM TEST_OPS test_executor_and_mul) -list(REMOVE_ITEM TEST_OPS test_assign_value_op) -list(REMOVE_ITEM TEST_OPS test_array_read_write_op) -list(REMOVE_ITEM TEST_OPS test_lod_rank_table) -list(REMOVE_ITEM TEST_OPS test_weight_normalization) -list(REMOVE_ITEM TEST_OPS test_conditional_block) -list(REMOVE_ITEM TEST_OPS test_parameter) -list(REMOVE_ITEM TEST_OPS test_registry) -list(REMOVE_ITEM TEST_OPS test_fetch_var) -list(REMOVE_ITEM TEST_OPS test_parallel_op) -list(REMOVE_ITEM TEST_OPS test_dynrnn_static_input) list(REMOVE_ITEM TEST_OPS test_dist_train) - -# tests that can be bundled together in one python process for speed. -if(WITH_FAST_BUNDLE_TEST) - py_test_modules("test_all_ops" MODULES ${TEST_OPS}) -else() - foreach(TEST_OP ${TEST_OPS}) - py_test_modules(${TEST_OP} MODULES ${TEST_OP}) - endforeach(TEST_OP) -endif(WITH_FAST_BUNDLE_TEST) - -# -py_test_modules(test_sequence_expand MODULES test_sequence_expand) -# tests with high overhead -py_test_modules(test_parallel_executor MODULES test_parallel_executor) -py_test_modules(test_warpctc_op MODULES test_warpctc_op ENVS FLAGS_warpctc_dir=${WARPCTC_LIB_DIR}) -py_test_modules(test_train_dyn_rnn MODULES test_dyn_rnn) -py_test_modules(test_mul_op MODULES test_mul_op) - -# tests that need to be run in separate process. -py_test_modules(test_multihead_attention MODULES test_multihead_attention) -py_test_modules(test_calc_gradient MODULES test_calc_gradient) -py_test_modules(test_while_op MODULES test_while_op) -py_test_modules(test_lod_array_length_op MODULES test_lod_array_length_op) -py_test_modules(test_reorder_lod_tensor MODULES test_reorder_lod_tensor) -py_test_modules(test_profiler MODULES test_profiler) -py_test_modules(test_nvprof MODULES test_nvprof) -py_test_modules(test_normalization_wrapper MODULES test_normalization_wrapper) -py_test_modules(test_executor_and_mul MODULES test_executor_and_mul) -py_test_modules(test_assign_value_op MODULES test_assign_value_op) -py_test_modules(test_array_read_write_op MODULES test_array_read_write_op) -py_test_modules(test_lod_rank_table MODULES test_lod_rank_table) -py_test_modules(test_weight_normalization MODULES test_weight_normalization) -py_test_modules(test_conditional_block MODULES test_conditional_block) -py_test_modules(test_parameter MODULES test_parameter) -py_test_modules(test_registry MODULES test_registry) -py_test_modules(test_fetch_var MODULES test_fetch_var) -py_test_modules(test_dynrnn_static_input MODULES test_dynrnn_static_input) -py_test_modules(test_parallel_op MODULES test_parallel_op) -py_test_modules(test_dist_train MODULES test_dist_train) +list(REMOVE_ITEM TEST_OPS test_parallel_executor_crf) +list(REMOVE_ITEM TEST_OPS test_parallel_executor_fetch_feed) +foreach(TEST_OP ${TEST_OPS}) + py_test_modules(${TEST_OP} MODULES ${TEST_OP}) +endforeach(TEST_OP) +py_test_modules(test_warpctc_op MODULES test_warpctc_op ENVS FLAGS_warpctc_dir=${WARPCTC_LIB_DIR} SERIAL) +py_test_modules(test_dist_train MODULES test_dist_train SERIAL) +# tests that need to be done in fixed timeout +set_tests_properties(test_listen_and_serv_op PROPERTIES TIMEOUT 20) diff --git a/python/paddle/fluid/tests/unittests/op_test.py b/python/paddle/fluid/tests/unittests/op_test.py index 299ab8e51f017e1980a8b40e3830fc42b1ff7ccc..b611470fa1ff326df960c349b71006f52d586d8e 100644 --- a/python/paddle/fluid/tests/unittests/op_test.py +++ b/python/paddle/fluid/tests/unittests/op_test.py @@ -36,6 +36,12 @@ def randomize_probability(batch_size, class_num, dtype='float32'): def create_op(scope, op_type, inputs, outputs, attrs): kwargs = dict() + op_maker = core.op_proto_and_checker_maker + op_role_attr_name = op_maker.kOpRoleAttrName() + + if op_role_attr_name not in attrs: + attrs[op_role_attr_name] = int(op_maker.OpRole.Forward) + def __create_var__(name, var_name): scope.var(var_name).get_tensor() kwargs[name].append(var_name) @@ -473,9 +479,9 @@ class OpTest(unittest.TestCase): def np_dtype_to_fluid_dtype(input): """Change the dtype of float16 numpy array - numpy float16 is binded to paddle::platform::float16 + numpy float16 is binded to paddle::platform::float16 in tensor_py.h via the help of uint16 data type since - the internal memory representation of float16 is + the internal memory representation of float16 is uint16_t in paddle and np.uint16 in numpy, which are themselves binded together by pybind. @@ -483,9 +489,9 @@ class OpTest(unittest.TestCase): input: input numpy array Returns: - input: The dtype of input will be changed to np.uint16 if + input: The dtype of input will be changed to np.uint16 if it is originally np.float16, such that the internal memory - of input will be reinterpreted as of dtype np.uint16. + of input will be reinterpreted as of dtype np.uint16. """ if input.dtype == np.float16: input.dtype = np.uint16 diff --git a/python/paddle/fluid/tests/unittests/parallel_executor_test_base.py b/python/paddle/fluid/tests/unittests/parallel_executor_test_base.py new file mode 100644 index 0000000000000000000000000000000000000000..c9c3c648717814c28c39a401487925824e885946 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/parallel_executor_test_base.py @@ -0,0 +1,96 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest +import paddle.fluid as fluid +import time +import numpy as np + +__all__ = ['TestParallelExecutorBase'] + + +class TestParallelExecutorBase(unittest.TestCase): + def check_network_convergence(self, + method, + memory_opt=True, + iter=50, + batch_size=None, + allow_op_delay=False, + feed_dict=None, + seed=None, + use_parallel_executor=True, + balance_parameter_opt_between_cards=False): + def run_executor(exe, feed, fetch_list, program=None): + if isinstance(exe, fluid.ParallelExecutor): + res = exe.run(fetch_list=fetch_list, feed=feed) + elif isinstance(exe, fluid.Executor): + if program is None: + program = fluid.default_main_program() + res = exe.run(program=program, feed=feed, fetch_list=fetch_list) + else: + raise ValueError('Unkown type exe') + return res + + main = fluid.Program() + startup = fluid.Program() + startup.random_seed = 1 # Fix random seed + with fluid.program_guard(main, startup): + if seed is not None: + startup.random_seed = seed + loss = method(use_feed=feed_dict is not None) + adam = fluid.optimizer.Adam() + adam.minimize(loss) + if memory_opt: + fluid.memory_optimize(main) + place = fluid.CUDAPlace(0) + startup_exe = fluid.Executor(place) + startup_exe.run(startup) + exec_strategy = fluid.ExecutionStrategy() + exec_strategy.allow_op_delay = allow_op_delay + + build_strategy = fluid.BuildStrategy() + build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce if balance_parameter_opt_between_cards else fluid.BuildStrategy.ReduceStrategy.AllReduce + + if use_parallel_executor: + exe = fluid.ParallelExecutor( + True, + loss_name=loss.name, + exec_strategy=exec_strategy, + build_strategy=build_strategy) + else: + exe = fluid.Executor(place=place) + + if batch_size is not None: + batch_size *= fluid.core.get_cuda_device_count() + begin = time.time() + first_loss, = run_executor( + exe=exe, feed=feed_dict, fetch_list=[loss.name]) + first_loss = np.array(first_loss) + + for i in xrange(iter): + run_executor(exe=exe, feed=feed_dict, fetch_list=[]) + + last_loss, = run_executor( + exe=exe, feed=feed_dict, fetch_list=[loss.name]) + end = time.time() + + if batch_size is not None: + print "%.4f Instance per second" % ( + (batch_size * iter + 2) / (end - begin)) + + last_loss = np.array(last_loss) + + print first_loss, last_loss + # self.assertGreater(first_loss[0], last_loss[0]) + return first_loss, last_loss diff --git a/python/paddle/fluid/tests/unittests/test_batch_norm_mkldnn_op.py b/python/paddle/fluid/tests/unittests/test_batch_norm_mkldnn_op.py new file mode 100644 index 0000000000000000000000000000000000000000..f6097d4b846e8da1c4ee3cc49b31f9873660056d --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_batch_norm_mkldnn_op.py @@ -0,0 +1,56 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest +import numpy as np +import paddle.fluid.core as core +from paddle.fluid.op import Operator +import paddle.fluid as fluid +from op_test import OpTest +from paddle.fluid.framework import grad_var_name +from test_batch_norm_op import TestBatchNormOpInference, TestBatchNormOpTraining, _reference_training, _reference_grad + + +class TestMKLDNNBatchNormOpTraining(TestBatchNormOpTraining): + def init_kernel_type(self): + self.use_mkldnn = True + self.data_formats = ["NCHW"] + + def ref_forward_backward(self, x, y_grad, scale, bias, mean, variance, + epsilon, momentum, shape, data_layout): + # run forward + y, saved_mean, saved_variance = _reference_training( + x, scale, bias, epsilon, data_layout) + mean_out = saved_mean * (1. - momentum) + momentum * mean + variance_out = saved_variance * (1. - momentum) + momentum * variance + # run backward + x_grad, scale_grad, bias_grad = _reference_grad( + x, y_grad, scale, saved_mean, saved_variance, epsilon, data_layout) + + return y, mean_out, variance_out, saved_mean, saved_variance, x_grad, scale_grad, bias_grad + + +class TestMKLDNNBatchNormOpInference(TestBatchNormOpInference): + def init_kernel_type(self): + self.use_mkldnn = True + + def test_check_output(self): + place = core.CPUPlace() + data_format = "NCHW" + + self.check_with_place(place, data_format, self.dtype, [2, 3, 4, 5]) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_batch_norm_op.py b/python/paddle/fluid/tests/unittests/test_batch_norm_op.py index 6afb6fa6e753d3d6478313c840b158c3895b3efb..4216d83653b27ec7f18034e576fbedbecc3f1cfe 100644 --- a/python/paddle/fluid/tests/unittests/test_batch_norm_op.py +++ b/python/paddle/fluid/tests/unittests/test_batch_norm_op.py @@ -158,6 +158,8 @@ def set_output_grad(scope, outputs, place, feed_dict=None): class TestBatchNormOpInference(unittest.TestCase): def setUp(self): self.dtype = np.float32 + self.use_mkldnn = False + self.init_kernel_type() def __assert_close(self, tensor, np_array, msg, atol=1e-4): self.assertTrue(np.allclose(np.array(tensor), np_array, atol=atol), msg) @@ -230,6 +232,7 @@ class TestBatchNormOpInference(unittest.TestCase): # attrs is_test=True, data_layout=data_layout, + use_mkldnn=self.use_mkldnn, epsilon=epsilon) batch_norm_op.run(scope, place) @@ -254,10 +257,15 @@ class TestBatchNormOpInference(unittest.TestCase): [2, 3, 4, 5]) self.check_with_place(place, data_format, self.dtype, [2, 3]) + def init_kernel_type(self): + pass + class TestFP16BatchNormOpInference(TestBatchNormOpInference): def setUp(self): self.dtype = np.float16 + self.use_mkldnn = False + self.init_kernel_type() def test_check_output(self): places = [] @@ -274,11 +282,27 @@ class TestFP16BatchNormOpInference(TestBatchNormOpInference): class TestBatchNormOpTraining(unittest.TestCase): + def setUp(self): + self.use_mkldnn = False + self.data_formats = ["NCHW", "NHWC"] + self.init_kernel_type() + def __assert_close(self, tensor, np_array, msg, atol=1e-4): - if not np.allclose(np.array(tensor), np_array, atol=atol): - import pdb - pdb.set_trace() - self.assertTrue(np.allclose(np.array(tensor), np_array, atol=atol), msg) + np.allclose(np.array(tensor), np_array, atol=atol) + + def ref_forward_backward(self, x, y_grad, scale, bias, mean, variance, + epsilon, momentum, shape, data_layout): + # run forward + y, saved_mean, var_ref = _reference_training(x, scale, bias, epsilon, + data_layout) + mean_out = saved_mean * (1. - momentum) + momentum * mean + variance_out = var_ref * (1. - momentum) + momentum * variance + saved_variance = 1. / np.sqrt(var_ref + epsilon) + # run backward + x_grad, scale_grad, bias_grad = _reference_grad( + x, y_grad, scale, saved_mean, var_ref, epsilon, data_layout) + + return y, mean_out, variance_out, saved_mean, saved_variance, x_grad, scale_grad, bias_grad def test_forward_backward(self): def test_with_place(place, data_layout, shape): @@ -298,16 +322,11 @@ class TestBatchNormOpTraining(unittest.TestCase): mean = np.zeros(scale_shape).astype(np.float32) variance = np.ones(scale_shape).astype(np.float32) - # run forward - y, saved_mean, var_ref = _reference_training(x, scale, bias, - epsilon, data_layout) - mean_out = saved_mean * (1. - momentum) + momentum * mean - variance_out = var_ref * (1. - momentum) + momentum * variance - saved_variance = 1. / np.sqrt(var_ref + epsilon) - # run backward y_grad = np.random.random_sample(shape).astype(np.float32) - x_grad, scale_grad, bias_grad = _reference_grad( - x, y_grad, scale, saved_mean, var_ref, epsilon, data_layout) + + y, mean_out, variance_out, saved_mean, saved_variance, x_grad, scale_grad, bias_grad = self.ref_forward_backward( + x, y_grad, scale, bias, mean, variance, epsilon, momentum, + shape, data_layout) var_dict = locals() var_dict['y@GRAD'] = y_grad @@ -347,7 +366,8 @@ class TestBatchNormOpTraining(unittest.TestCase): "momentum": momentum, "epsilon": epsilon, "is_test": False, - "data_layout": data_layout + "data_layout": data_layout, + "use_mkldnn": self.use_mkldnn }) block.create_var(name='y@GRAD', dtype='float32', shape=y.shape) @@ -390,13 +410,17 @@ class TestBatchNormOpTraining(unittest.TestCase): print "op test forward passed: ", str(place), data_layout places = [core.CPUPlace()] + if core.is_compiled_with_cuda() and core.op_support_gpu("batch_norm"): places.append(core.CUDAPlace(0)) for place in places: - for data_format in ["NCHW", "NHWC"]: + for data_format in self.data_formats: test_with_place(place, data_format, [2, 3, 4, 5]) + def init_kernel_type(self): + pass + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_beam_search_decode_op.py b/python/paddle/fluid/tests/unittests/test_beam_search_decode_op.py index 4ee00605e22ba45d9e46a8bba27712c3fd97872a..7976dd7c3f14390fb00bc8ab39121b6a686e3039 100644 --- a/python/paddle/fluid/tests/unittests/test_beam_search_decode_op.py +++ b/python/paddle/fluid/tests/unittests/test_beam_search_decode_op.py @@ -22,12 +22,12 @@ from paddle.fluid.op import Operator class TestBeamSearchDecodeOp(unittest.TestCase): def setUp(self): self.scope = core.Scope() - self.cpu_place = core.CPUPlace() + self.place = core.CPUPlace() def append_lod_tensor(self, tensor_array, lod, data): lod_tensor = core.LoDTensor() lod_tensor.set_lod(lod) - lod_tensor.set(data, self.cpu_place) + lod_tensor.set(data, self.place) tensor_array.append(lod_tensor) def test_get_set(self): @@ -71,7 +71,7 @@ class TestBeamSearchDecodeOp(unittest.TestCase): SentenceIds="sentence_ids", SentenceScores="sentence_scores") - beam_search_decode_op.run(self.scope, self.cpu_place) + beam_search_decode_op.run(self.scope, self.place) expected_lod = [[0, 4, 8], [0, 1, 3, 6, 9, 10, 13, 16, 19]] self.assertEqual(sentence_ids.lod(), expected_lod) @@ -84,5 +84,11 @@ class TestBeamSearchDecodeOp(unittest.TestCase): np.array_equal(np.array(sentence_scores), expected_data)) +class TestBeamSearchDecodeOpGPU(TestBeamSearchDecodeOp): + def setUp(self): + self.scope = core.Scope() + self.place = core.CUDAPlace(0) + + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_conv2d_transpose_op.py b/python/paddle/fluid/tests/unittests/test_conv2d_transpose_op.py index d864b9b348e961c585749d47d449d775b2dfebc9..ded2f130288a4a959a1c859b2cc8ccf0912efb12 100644 --- a/python/paddle/fluid/tests/unittests/test_conv2d_transpose_op.py +++ b/python/paddle/fluid/tests/unittests/test_conv2d_transpose_op.py @@ -21,8 +21,11 @@ from op_test import OpTest def conv2dtranspose_forward_naive(input_, filter_, attrs): in_n, in_c, in_h, in_w = input_.shape - f_c, out_c, f_h, f_w = filter_.shape + f_c, f_out_c, f_h, f_w = filter_.shape + groups = attrs['groups'] assert in_c == f_c + out_c = f_out_c * groups + sub_in_c = in_c / groups stride, pad, dilations = attrs['strides'], attrs['paddings'], attrs[ 'dilations'] @@ -36,15 +39,21 @@ def conv2dtranspose_forward_naive(input_, filter_, attrs): for n in range(in_n): for i in range(in_h): for j in range(in_w): - input_masked = input_[n, :, i, j] # (c) - input_masked = np.reshape(input_masked, (in_c, 1, 1)) - input_masked = np.tile(input_masked, (1, f_h, f_w)) - - for k in range(out_c): - tmp_out = np.sum(input_masked * filter_[:, k, :, :], axis=0) - i1, i2 = i * stride[0], i * stride[0] + d_bolck_h - j1, j2 = j * stride[0], j * stride[0] + d_bolck_h - out[n, k, i1:i2:dilations[0], j1:j2:dilations[1]] += tmp_out + for g in range(groups): + input_masked = input_[n, g * sub_in_c:(g + 1) * sub_in_c, i, + j] # (c) + input_masked = np.reshape(input_masked, (sub_in_c, 1, 1)) + input_masked = np.tile(input_masked, (1, f_h, f_w)) + + for k in range(f_out_c): + tmp_out = np.sum( + input_masked * + filter_[g * sub_in_c:(g + 1) * sub_in_c, k, :, :], + axis=0) + i1, i2 = i * stride[0], i * stride[0] + d_bolck_h + j1, j2 = j * stride[0], j * stride[0] + d_bolck_h + out[n, g * f_out_c + k, i1:i2:dilations[0], j1:j2: + dilations[1]] += tmp_out out = out[:, :, pad[0]:out_h - pad[0], pad[1]:out_w - pad[1]] return out @@ -64,6 +73,7 @@ class TestConv2dTransposeOp(OpTest): self.attrs = { 'strides': self.stride, 'paddings': self.pad, + 'groups': self.groups, 'dilations': self.dilations, 'use_cudnn': self.use_cudnn, 'data_format': 'AnyLayout' # TODO(dzhwinter) : should be fix latter @@ -127,6 +137,7 @@ class TestConv2dTransposeOp(OpTest): self.pad = [0, 0] self.stride = [1, 1] self.dilations = [1, 1] + self.groups = 1 self.input_size = [2, 3, 5, 5] # NCHW f_c = self.input_size[1] self.filter_size = [f_c, 6, 3, 3] @@ -140,16 +151,29 @@ class TestWithPad(TestConv2dTransposeOp): self.pad = [1, 1] self.stride = [1, 1] self.dilations = [1, 1] + self.groups = 1 self.input_size = [2, 3, 5, 5] # NCHW f_c = self.input_size[1] self.filter_size = [f_c, 6, 3, 3] +class TestWithGroups(TestConv2dTransposeOp): + def init_test_case(self): + self.pad = [1, 1] + self.stride = [1, 1] + self.dilations = [1, 1] + self.groups = 2 + self.input_size = [2, 4, 5, 5] # NCHW + f_c = self.input_size[1] + self.filter_size = [f_c, 3, 3, 3] + + class TestWithStride(TestConv2dTransposeOp): def init_test_case(self): self.pad = [1, 1] self.stride = [2, 2] self.dilations = [1, 1] + self.groups = 1 self.input_size = [2, 3, 5, 5] # NCHW f_c = self.input_size[1] self.filter_size = [f_c, 6, 3, 3] @@ -159,6 +183,7 @@ class TestWithDilation(TestConv2dTransposeOp): def init_test_case(self): self.pad = [1, 1] self.stride = [1, 1] + self.groups = 1 self.dilations = [2, 2] self.input_size = [2, 3, 5, 5] # NCHW f_c = self.input_size[1] @@ -176,6 +201,7 @@ class TestCUDNNWithPad(TestWithPad): def init_test_case(self): self.pad = [1, 1] self.stride = [1, 1] + self.groups = 1 self.dilations = [1, 1] self.input_size = [2, 3, 5, 5] # NCHW f_c = self.input_size[1] @@ -190,6 +216,7 @@ class TestCUDNNWithStride(TestWithStride): def init_test_case(self): self.pad = [1, 1] self.stride = [2, 2] + self.groups = 1 self.dilations = [1, 1] self.input_size = [2, 3, 5, 5] # NCHW f_c = self.input_size[1] @@ -200,6 +227,21 @@ class TestCUDNNWithStride(TestWithStride): self.op_type = "conv2d_transpose" +class TestCUDNNWithGroups(TestWithGroups): + def init_test_case(self): + self.pad = [1, 1] + self.stride = [1, 1] + self.dilations = [1, 1] + self.groups = 2 + self.input_size = [2, 4, 5, 5] # NCHW + f_c = self.input_size[1] + self.filter_size = [f_c, 3, 3, 3] + + def init_op_type(self): + self.use_cudnn = True + self.op_type = "conv2d_transpose" + + # Please Don't remove the following code. # Currently, CI use cudnn V5.0 which not support dilation conv. # class TestCUDNNWithDilation(TestWithDilation): diff --git a/python/paddle/fluid/tests/unittests/test_conv3d_op.py b/python/paddle/fluid/tests/unittests/test_conv3d_op.py index 7703dfe0135b402f830bcdeaf47c26e5e3f8ca58..dd4ef7cc94ea1e8de5fe4775408389907d47d0d6 100644 --- a/python/paddle/fluid/tests/unittests/test_conv3d_op.py +++ b/python/paddle/fluid/tests/unittests/test_conv3d_op.py @@ -70,9 +70,11 @@ def conv3d_forward_naive(input, filter, group, conv_param): class TestConv3dOp(OpTest): def setUp(self): + self.op_type = "conv3d" self.use_cudnn = False + self.dtype = np.float32 + self.init_kernel_type() self.init_group() - self.init_op_type() self.init_dilation() self.init_test_case() @@ -80,20 +82,24 @@ class TestConv3dOp(OpTest): 'stride': self.stride, 'pad': self.pad, 'dilations': self.dilations, - 'use_cudnn': self.use_cudnn, 'data_format': 'AnyLayout' # TODO(dzhwinter) : should be fix latter } - input = np.random.random(self.input_size).astype("float32") - filter = np.random.random(self.filter_size).astype("float32") + + input = np.random.random(self.input_size).astype(self.dtype) + filter = np.random.random(self.filter_size).astype(self.dtype) output = conv3d_forward_naive(input, filter, self.groups, - conv3d_param).astype("float32") + conv3d_param).astype(self.dtype) - self.inputs = {'Input': input, 'Filter': filter} + self.inputs = { + 'Input': OpTest.np_dtype_to_fluid_dtype(input), + 'Filter': OpTest.np_dtype_to_fluid_dtype(filter) + } self.attrs = { 'strides': self.stride, 'paddings': self.pad, 'groups': self.groups, - 'dilations': self.dilations + 'dilations': self.dilations, + 'use_cudnn': self.use_cudnn } self.outputs = {'Output': output} @@ -108,6 +114,8 @@ class TestConv3dOp(OpTest): self.check_output() def test_check_grad(self): + if self.dtype == np.float16: + return if self.testcudnn(): place = core.CUDAPlace(0) self.check_grad_with_place( @@ -120,6 +128,8 @@ class TestConv3dOp(OpTest): set(['Input', 'Filter']), 'Output', max_relative_error=0.03) def test_check_grad_no_filter(self): + if self.dtype == np.float16: + return if self.testcudnn(): place = core.CUDAPlace(0) self.check_grad_with_place( @@ -135,6 +145,8 @@ class TestConv3dOp(OpTest): no_grad_set=set(['Filter'])) def test_check_grad_no_input(self): + if self.dtype == np.float16: + return if self.testcudnn(): place = core.CUDAPlace(0) self.check_grad_with_place( @@ -163,8 +175,8 @@ class TestConv3dOp(OpTest): def init_group(self): self.groups = 1 - def init_op_type(self): - self.op_type = "conv3d" + def init_kernel_type(self): + pass class TestCase1(TestConv3dOp): @@ -235,34 +247,90 @@ class TestWithDilation(TestConv3dOp): self.groups = 3 +#----------------Conv3dCUDNN---------------- class TestCUDNN(TestConv3dOp): - def init_op_type(self): + def init_kernel_type(self): self.use_cudnn = True - self.op_type = "conv3d" + + +class TestFP16CUDNN(TestConv3dOp): + def init_kernel_type(self): + self.use_cudnn = True + self.dtype = np.float16 + + def test_check_output(self): + if core.is_compiled_with_cuda(): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_output_with_place(place, atol=2e-2) class TestWithGroup1CUDNN(TestWithGroup1): - def init_op_type(self): + def init_kernel_type(self): self.use_cudnn = True - self.op_type = "conv3d" + + +class TestFP16WithGroup1CUDNN(TestWithGroup1): + def init_kernel_type(self): + self.use_cudnn = True + self.dtype = np.float16 + + def test_check_output(self): + if core.is_compiled_with_cuda(): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_output_with_place(place, atol=2e-2) class TestWithGroup2CUDNN(TestWithGroup2): - def init_op_type(self): + def init_kernel_type(self): self.use_cudnn = True - self.op_type = "conv3d" + + +class TestFP16WithGroup2CUDNN(TestWithGroup2): + def init_kernel_type(self): + self.use_cudnn = True + self.dtype = np.float16 + + def test_check_output(self): + if core.is_compiled_with_cuda(): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_output_with_place(place, atol=2e-2) class TestWith1x1CUDNN(TestWith1x1): - def init_op_type(self): + def init_kernel_type(self): self.use_cudnn = True - self.op_type = "conv3d" + + +class TestFP16With1x1CUDNN(TestWith1x1): + def init_kernel_type(self): + self.use_cudnn = True + self.dtype = np.float16 + + def test_check_output(self): + if core.is_compiled_with_cuda(): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_output_with_place(place, atol=2e-2) class TestWithInput1x1Filter1x1CUDNN(TestWithInput1x1Filter1x1): - def init_op_type(self): + def init_kernel_type(self): self.use_cudnn = True - self.op_type = "conv3d" + + +class TestFP16WithInput1x1Filter1x1CUDNN(TestWithInput1x1Filter1x1): + def init_kernel_type(self): + self.use_cudnn = True + self.dtype = np.float16 + + def test_check_output(self): + if core.is_compiled_with_cuda(): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_output_with_place(place, atol=2e-2) # FIXME(typhoonzero): find a way to determine if diff --git a/python/paddle/fluid/tests/unittests/test_conv3d_transpose_op.py b/python/paddle/fluid/tests/unittests/test_conv3d_transpose_op.py index 55ba238710c56dd0daea388cd2dcdb79243bb71e..c9f26d10df8ff39d6bd77b1597336600f676d362 100644 --- a/python/paddle/fluid/tests/unittests/test_conv3d_transpose_op.py +++ b/python/paddle/fluid/tests/unittests/test_conv3d_transpose_op.py @@ -21,8 +21,11 @@ from op_test import OpTest def conv3dtranspose_forward_naive(input_, filter_, attrs): in_n, in_c, in_d, in_h, in_w = input_.shape - f_c, out_c, f_d, f_h, f_w = filter_.shape + f_c, f_out_c, f_d, f_h, f_w = filter_.shape + groups = attrs['groups'] assert in_c == f_c + out_c = f_out_c * groups + sub_in_c = in_c / groups stride, pad, dilations = attrs['strides'], attrs['paddings'], attrs[ 'dilations'] @@ -39,18 +42,23 @@ def conv3dtranspose_forward_naive(input_, filter_, attrs): for d in range(in_d): for i in range(in_h): for j in range(in_w): - input_masked = input_[n, :, d, i, j] # (c) - input_masked = np.reshape(input_masked, (in_c, 1, 1, 1)) - input_masked = np.tile(input_masked, (1, f_d, f_h, f_w)) - - for k in range(out_c): - tmp_out = np.sum(input_masked * filter_[:, k, :, :, :], - axis=0) - d1, d2 = d * stride[0], d * stride[0] + d_bolck_d - i1, i2 = i * stride[1], i * stride[1] + d_bolck_h - j1, j2 = j * stride[2], j * stride[2] + d_bolck_w - out[n, k, d1:d2:dilations[0], i1:i2:dilations[1], j1:j2: - dilations[2]] += tmp_out + for g in range(groups): + input_masked = input_[n, g * sub_in_c:(g + 1 + ) * sub_in_c, d, + i, j] # (c) + input_masked = np.reshape(input_masked, + (sub_in_c, 1, 1, 1)) + input_masked = np.tile(input_masked, (1, f_d, f_h, f_w)) + + for k in range(f_out_c): + tmp_out = np.sum(input_masked * filter_[ + g * sub_in_c:(g + 1) * sub_in_c, k, :, :, :], + axis=0) + d1, d2 = d * stride[0], d * stride[0] + d_bolck_d + i1, i2 = i * stride[1], i * stride[1] + d_bolck_h + j1, j2 = j * stride[2], j * stride[2] + d_bolck_w + out[n, g * f_out_c + k, d1:d2:dilations[0], i1:i2: + dilations[1], j1:j2:dilations[2]] += tmp_out out = out[:, :, pad[0]:out_d - pad[0], pad[1]:out_h - pad[1], pad[2]:out_w - pad[2]] @@ -72,6 +80,7 @@ class TestConv3dTransposeOp(OpTest): 'strides': self.stride, 'paddings': self.pad, 'dilations': self.dilations, + 'groups': self.groups, 'use_cudnn': self.use_cudnn, 'data_format': 'AnyLayout' # TODO(dzhwinter) : should be fix latter } @@ -134,6 +143,7 @@ class TestConv3dTransposeOp(OpTest): self.pad = [0, 0, 0] self.stride = [1, 1, 1] self.dilations = [1, 1, 1] + self.groups = 1 self.input_size = [2, 3, 5, 5, 5] # NCDHW f_c = self.input_size[1] self.filter_size = [f_c, 6, 3, 3, 3] @@ -147,16 +157,29 @@ class TestWithPad(TestConv3dTransposeOp): self.pad = [1, 1, 1] self.stride = [1, 1, 1] self.dilations = [1, 1, 1] + self.groups = 1 self.input_size = [2, 3, 5, 5, 5] # NCDHW f_c = self.input_size[1] self.filter_size = [f_c, 6, 3, 3, 3] +class TestWithGroups(TestConv3dTransposeOp): + def init_test_case(self): + self.pad = [1, 1, 1] + self.stride = [1, 1, 1] + self.dilations = [1, 1, 1] + self.groups = 2 + self.input_size = [2, 4, 5, 5, 5] # NCHW + f_c = self.input_size[1] + self.filter_size = [f_c, 3, 3, 3, 3] + + class TestWithStride(TestConv3dTransposeOp): def init_test_case(self): self.pad = [1, 1, 1] self.stride = [2, 2, 2] self.dilations = [1, 1, 1] + self.groups = 1 self.input_size = [2, 3, 5, 5, 5] # NCDHW f_c = self.input_size[1] self.filter_size = [f_c, 6, 3, 3, 3] @@ -167,6 +190,7 @@ class TestWithDilation(TestConv3dTransposeOp): self.pad = [1, 1, 1] self.stride = [1, 1, 1] self.dilations = [2, 2, 2] + self.groups = 1 self.input_size = [2, 3, 5, 5, 5] # NCDHW f_c = self.input_size[1] self.filter_size = [f_c, 6, 3, 3, 3] @@ -184,6 +208,7 @@ class TestCUDNNWithPad(TestWithPad): self.pad = [1, 1, 1] self.stride = [1, 1, 1] self.dilations = [1, 1, 1] + self.groups = 1 self.input_size = [2, 3, 5, 5, 5] # NCDHW f_c = self.input_size[1] self.filter_size = [f_c, 6, 3, 3, 3] @@ -198,6 +223,7 @@ class TestCUDNNWithStride(TestWithStride): self.pad = [1, 1, 1] self.stride = [2, 2, 2] self.dilations = [1, 1, 1] + self.groups = 1 self.input_size = [2, 3, 5, 5, 5] # NCDHW f_c = self.input_size[1] self.filter_size = [f_c, 6, 3, 3, 3] @@ -207,6 +233,21 @@ class TestCUDNNWithStride(TestWithStride): self.op_type = "conv3d_transpose" +class TestCUDNNWithGroups(TestWithGroups): + def init_test_case(self): + self.pad = [1, 1, 1] + self.stride = [1, 1, 1] + self.dilations = [1, 1, 1] + self.groups = 2 + self.input_size = [2, 4, 5, 5, 5] # NCHW + f_c = self.input_size[1] + self.filter_size = [f_c, 3, 3, 3, 3] + + def init_op_type(self): + self.use_cudnn = True + self.op_type = "conv3d_transpose" + + # Please Don't remove the following code. # Currently, CI use cudnn V5.0 which not support dilation conv. # class TestCUDNNWithDilation(TestWithDilation): diff --git a/python/paddle/fluid/tests/unittests/test_debugger.py b/python/paddle/fluid/tests/unittests/test_debugger.py index 67b03f635b6f8a3003efabe5425325080d47f61c..870952f2f916dcdec5991ac5c10d2da3a7ab18a8 100644 --- a/python/paddle/fluid/tests/unittests/test_debugger.py +++ b/python/paddle/fluid/tests/unittests/test_debugger.py @@ -15,7 +15,7 @@ import unittest import paddle.fluid as fluid import paddle.fluid.core as core -from paddle.fluid import debuger +from paddle.fluid import debugger from paddle.fluid.framework import Program @@ -51,9 +51,9 @@ class TestDebugger(unittest.TestCase): outputs={"Out": mul_out}, attrs={"x_num_col_dims": 1}) - print(debuger.pprint_program_codes(p)) + print(debugger.pprint_program_codes(p)) - debuger.draw_block_graphviz(p.block(0), path="./test.dot") + debugger.draw_block_graphviz(p.block(0), path="./test.dot") if __name__ == '__main__': diff --git a/python/paddle/fluid/tests/unittests/test_detection_map_op.py b/python/paddle/fluid/tests/unittests/test_detection_map_op.py index a905a854ad157ffa3d7816dfbd445f3e344a1249..f545ad155ccd28c2d34e424d307eed49b37f20fb 100644 --- a/python/paddle/fluid/tests/unittests/test_detection_map_op.py +++ b/python/paddle/fluid/tests/unittests/test_detection_map_op.py @@ -160,7 +160,9 @@ class TestDetectionMAPOp(OpTest): label_count, true_pos, false_pos = get_input_pos( self.class_pos_count, self.true_pos, self.true_pos_lod, self.false_pos, self.false_pos_lod) - for (label, difficult, xmin, ymin, xmax, ymax) in self.label: + for v in self.label: + label = v[0] + difficult = False if len(v) == 5 else v[1] if self.evaluate_difficult: label_count[label] += 1 elif not difficult: @@ -245,6 +247,15 @@ class TestDetectionMAPOpSkipDiff(TestDetectionMAPOp): [2, 0.8, 0, 1], [2, 0.1, 1, 0], [3, 0.2, 0, 1]] +class TestDetectionMAPOpWithoutDiff(TestDetectionMAPOp): + def init_test_case(self): + super(TestDetectionMAPOpWithoutDiff, self).init_test_case() + + # label xmin ymin xmax ymax + self.label = [[1, 0.1, 0.1, 0.3, 0.3], [1, 0.6, 0.6, 0.8, 0.8], + [2, 0.3, 0.3, 0.6, 0.5], [1, 0.7, 0.1, 0.9, 0.3]] + + class TestDetectionMAPOp11Point(TestDetectionMAPOp): def init_test_case(self): super(TestDetectionMAPOp11Point, self).init_test_case() diff --git a/python/paddle/fluid/tests/unittests/test_dist_train.py b/python/paddle/fluid/tests/unittests/test_dist_train.py index c7fdd06f105e3b5fd906d3524d41df8f84160e63..2314bb2ed8a4eeb34752fd5d040f8a8476798aa6 100644 --- a/python/paddle/fluid/tests/unittests/test_dist_train.py +++ b/python/paddle/fluid/tests/unittests/test_dist_train.py @@ -12,19 +12,21 @@ # See the License for the specific language governing permissions and # limitations under the License. +import os +import time import unittest +from multiprocessing import Process + +import numpy import paddle.fluid as fluid -import paddle.fluid.core as core import paddle.fluid.layers as layers -import numpy -from multiprocessing import Process -from threading import Thread -import os, sys -import time class TestSendOp(unittest.TestCase): + @unittest.skip( + "This test is buggy. We cannot use time.sleep to sync processes, the connection may fail in unittest." + ) def test_send(self): # Run init_serv in a thread place = fluid.CPUPlace() @@ -34,7 +36,7 @@ class TestSendOp(unittest.TestCase): p.start() time.sleep(10) - with open("/tmp/paddle.selected_port", "r") as fn: + with open("/tmp/paddle.%d.port" % p.pid, "r") as fn: selected_port = int(fn.readlines()[0]) self.init_client(place, selected_port) @@ -52,15 +54,18 @@ class TestSendOp(unittest.TestCase): serv = layers.ListenAndServ( "127.0.0.1:0", ["X"], optimizer_mode=False) with serv.do(): + out_var = main.global_block().create_var( + name="scale_0.tmp_0", + psersistable=True, + dtype="float32", + shape=[32, 32]) x = layers.data( shape=[32, 32], dtype='float32', name="X", append_batch_size=False) fluid.initializer.Constant(value=1.0)(x, main.global_block()) - o = layers.scale(x=x, scale=10.0) - main.global_block().create_var( - name=o.name, psersistable=False, dtype=o.dtype, shape=o.shape) + layers.scale(x=x, scale=10.0, out=out_var) self.server_exe = fluid.Executor(place) self.server_exe.run(main) diff --git a/python/paddle/fluid/tests/unittests/test_dist_transpiler.py b/python/paddle/fluid/tests/unittests/test_dist_transpiler.py new file mode 100644 index 0000000000000000000000000000000000000000..fa49bd41a5876847d046682dce5c3d3868a18500 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_dist_transpiler.py @@ -0,0 +1,116 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest + +import paddle.fluid as fluid +import paddle.fluid.core as core +import paddle.fluid.layers as layers +from paddle.fluid.transpiler.distribute_transpiler import delete_ops +import numpy + + +class TestDistTranspiler(unittest.TestCase): + def setUp(self): + self.trainer_id = 0 + self.trainers = 2 + self.pservers = 2 + self.pserver_eps = "127.0.0.1:6174,127.0.0.1:6175" + self.current_pserver_ep = "127.0.0.1:6174" + + def net_conf(self): + x = fluid.layers.data(name='x', shape=[1000], dtype='float32') + + y_predict = fluid.layers.fc(input=x, + size=1000, + act=None, + param_attr=fluid.ParamAttr(name='fc_w')) + + y = fluid.layers.data(name='y', shape=[1], dtype='float32') + + cost = fluid.layers.square_error_cost(input=y_predict, label=y) + avg_cost = fluid.layers.mean(cost) + sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.1) + + optimize_ops, params_grads = sgd_optimizer.minimize(avg_cost) + return optimize_ops, params_grads + + def test_transpiler(self): + trainer = self.get_trainer() + pserver, startup = self.get_pserver(self.current_pserver_ep) + self.assertEqual([op.type for op in trainer.global_block().ops], + self.get_expect_trainer_ops()) + + self.assertEqual(len(pserver.blocks), 3) + # block0: listen_and_serv + self.assertEqual([op.type for op in pserver.blocks[0].ops], + ["listen_and_serv"]) + # block2: optimize pass + self.assertEqual([op.type for op in pserver.blocks[1].ops], + ["sum", "scale", "sgd"]) + + # confirm startup program + + self.assertEqual([op.type for op in startup.global_block().ops], [ + "fill_constant", "fill_constant", "uniform_random", "uniform_random" + ]) + + # the variable #fc_w will be split into two blocks + fc_w_var = startup.global_block().var("fc_w.block1") + self.assertEqual(fc_w_var.shape, (500, 1000)) + + def get_main_program(self): + main = fluid.Program() + + with fluid.program_guard(main): + self.net_conf() + + return main + + def get_expect_trainer_ops(self): + trainer = fluid.Program() + + with fluid.program_guard(trainer): + optimize_ops, params_grads = self.net_conf() + + delete_ops(trainer.global_block(), optimize_ops) + ops = [op.type for op in trainer.global_block().ops] + [ + "split_byref", "send_vars", "send_barrier", "recv", "recv", + "fetch_barrier", "concat" + ] + ops.insert(ops.index("elementwise_add_grad") + 1, "send_vars") + return ops + + def get_trainer(self): + return self._transpiler_instance().get_trainer_program() + + def get_pserver(self, ep): + t = self._transpiler_instance() + pserver = t.get_pserver_program(ep) + startup = t.get_startup_program(ep, pserver) + return pserver, startup + + def _transpiler_instance(self): + main = self.get_main_program() + t = fluid.DistributeTranspiler() + t.transpile( + self.trainer_id, + program=main, + pservers=self.pserver_eps, + trainers=self.trainers) + return t + + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_fake_dequantize_op.py b/python/paddle/fluid/tests/unittests/test_fake_dequantize_op.py new file mode 100644 index 0000000000000000000000000000000000000000..281068e945e76a42635868d19573498f79fde1f3 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_fake_dequantize_op.py @@ -0,0 +1,60 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest +import numpy as np +import math +from op_test import OpTest + + +def quantize_max_abs(x, num_bits): + range = math.pow(2, num_bits) - 1 + scale = np.max(np.abs(x).flatten()) + y = np.round(x / scale * range) + return y, scale + + +def dequantize_max_abs(x, num_bits, scale): + range = math.pow(2, num_bits) - 1 + y = (scale / range) * x + return y + + +class TestFakeDequantizeMaxAbsOp(OpTest): + def set_args(self): + self.num_bits = 8 + + def setUp(self): + self.set_args() + self.op_type = "fake_dequantize_max_abs" + x = np.random.randn(31, 65).astype("float32") + yq, scale = quantize_max_abs(x, self.num_bits) + print 'scale ', scale + ydq = dequantize_max_abs(yq, self.num_bits, scale) + + self.inputs = {'X': yq} + self.attrs = {'num_bits': self.num_bits, 'scale': float(scale)} + self.outputs = {'Out': ydq} + + def test_check_output(self): + self.check_output() + + +class TestFakeDequantizeMaxAbsOp5Bits(OpTest): + def set_args(self): + self.num_bits = 5 + + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_fill_constant_batch_size_like_op.py b/python/paddle/fluid/tests/unittests/test_fill_constant_batch_size_like_op.py index 66e3e2d51d118756d4881955b4df8eb4d2bbc094..533d8ccfac82a2e298af16181ab16bf7aa3db282 100644 --- a/python/paddle/fluid/tests/unittests/test_fill_constant_batch_size_like_op.py +++ b/python/paddle/fluid/tests/unittests/test_fill_constant_batch_size_like_op.py @@ -50,5 +50,27 @@ class TestFillConstantBatchSizeLikeWhenSecondDimIsBatchSize(OpTest): self.check_output() +class TestFillConstantBatchSizeLikeWithLoDTensor(OpTest): + def setUp(self): + self.op_type = "fill_constant_batch_size_like" + self.inputs = { + 'Input': (np.random.random((31, 28)).astype("float32"), + [[0, 9, 23, 31]]) + } + self.attrs = { + 'value': 3.5, + 'shape': [-1, 16], + 'input_dim_idx': 0, + 'output_dim_idx': 0 + } + + out = np.random.random((3, 16)).astype("float32") + out.fill(3.5) + self.outputs = {'Out': out} + + def test_check_output(self): + self.check_output() + + if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_is_empty_op.py b/python/paddle/fluid/tests/unittests/test_is_empty_op.py index 4d11cf226be2ba4ffbe015198fed3191f1e02f72..11121d9b65351eab639b7618fac0e54714cf4680 100644 --- a/python/paddle/fluid/tests/unittests/test_is_empty_op.py +++ b/python/paddle/fluid/tests/unittests/test_is_empty_op.py @@ -14,42 +14,24 @@ import unittest import numpy as np -from paddle.fluid.op import Operator -import paddle.fluid.core as core +from op_test import OpTest -def create_tensor(scope, name, np_data): - tensor = scope.var(name).get_tensor() - tensor.set_dims(np_data.shape) - tensor.set(np_data, core.CPUPlace()) - return tensor - - -class TestIsEmptyOp(unittest.TestCase): +class TestEmpty(OpTest): def setUp(self): - self.scope = core.Scope() - # create input variables - np_data0 = np.array([0, 1, 2]) - create_tensor(self.scope, "X0", np_data0) - - np_data1 = np.array([1]) - t = create_tensor(self.scope, "X1", np_data1) - t.set_dims([0]) + self.op_type = "is_empty" + self.inputs = {'X': np.array([1, 2, 3])} + self.outputs = {'Out': np.array([False])} - # create output variables - self.scope.var("out") + def test_check_output(self): + self.check_output() - def test_no_empty(self): - self.one_case("X0", False) - def test_empty(self): - self.one_case("X1", True) - - def one_case(self, input, target): - op = Operator(type="is_empty", X=input, Out="out") - op.run(self.scope, core.CPUPlace()) - out = self.scope.var("out").get_tensor() - self.assertEqual(np.array(out)[0], target) +class TestNotEmpty(TestEmpty): + def setUp(self): + self.op_type = "is_empty" + self.inputs = {'X': np.array([])} + self.outputs = {'Out': np.array([True])} if __name__ == "__main__": diff --git a/python/paddle/fluid/tests/unittests/test_layers.py b/python/paddle/fluid/tests/unittests/test_layers.py index 17d6afdee161426e5da398ffa2ec148a027c905e..60dc1f83fc32e2551eb2a04ef35f1c8a0ffec769 100644 --- a/python/paddle/fluid/tests/unittests/test_layers.py +++ b/python/paddle/fluid/tests/unittests/test_layers.py @@ -359,6 +359,26 @@ class TestBook(unittest.TestCase): self.assertIsNotNone(indices) print(str(program)) + def test_roi_pool(self): + program = Program() + with program_guard(program): + x = layers.data(name="x", shape=[256, 30, 30], dtype="float32") + rois = layers.data( + name="rois", shape=[4], dtype="float32", lod_level=1) + output = layers.roi_pool(x, rois, 7, 7, 0.6) + self.assertIsNotNone(output) + print(str(program)) + + def test_upsampling_bilinear2d(self): + program = Program() + with program_guard(program): + x = layers.data(name='x', shape=[3, 9, 6], dtype="float32") + output = layers.upsampling_bilinear2d(x, out_shape=[12, 12]) + self.assertIsNotNone(output) + output = layers.upsampling_bilinear2d(x, scale=3) + self.assertIsNotNone(output) + print(str(program)) + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_listen_and_serv_op.py b/python/paddle/fluid/tests/unittests/test_listen_and_serv_op.py new file mode 100644 index 0000000000000000000000000000000000000000..cf89f9d0ebf6200933e539ef7fa8cbdc8f6db058 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_listen_and_serv_op.py @@ -0,0 +1,109 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import paddle +import paddle.fluid as fluid +import os +import signal +import subprocess +import time +import unittest +from multiprocessing import Process +from op_test import OpTest + + +def run_pserver(use_cuda, sync_mode, ip, port, trainer_count, trainer_id): + x = fluid.layers.data(name='x', shape=[1], dtype='float32') + y_predict = fluid.layers.fc(input=x, size=1, act=None) + y = fluid.layers.data(name='y', shape=[1], dtype='float32') + + # loss function + cost = fluid.layers.square_error_cost(input=y_predict, label=y) + avg_cost = fluid.layers.mean(cost) + + # optimizer + sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.001) + sgd_optimizer.minimize(avg_cost) + + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + exe = fluid.Executor(place) + + port = os.getenv("PADDLE_INIT_PORT", port) + pserver_ips = os.getenv("PADDLE_INIT_PSERVERS", ip) # ip,ip... + eplist = [] + for ip in pserver_ips.split(","): + eplist.append(':'.join([ip, port])) + pserver_endpoints = ",".join(eplist) # ip:port,ip:port... + trainers = int(os.getenv("TRAINERS", trainer_count)) + current_endpoint = os.getenv("POD_IP", ip) + ":" + port + trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID", trainer_id)) + t = fluid.DistributeTranspiler() + t.transpile( + trainer_id, + pservers=pserver_endpoints, + trainers=trainers, + sync_mode=sync_mode) + pserver_prog = t.get_pserver_program(current_endpoint) + pserver_startup = t.get_startup_program(current_endpoint, pserver_prog) + exe.run(pserver_startup) + exe.run(pserver_prog) + + +class TestListenAndServOp(OpTest): + def setUp(self): + self.sleep_time = 5 + self.ip = "127.0.0.1" + self.port = "6173" + self.trainer_count = 1 + self.trainer_id = 1 + + def _raise_signal(self, parent_pid, raised_signal): + time.sleep(self.sleep_time) + ps_command = subprocess.Popen( + "ps -o pid --ppid %d --noheaders" % parent_pid, + shell=True, + stdout=subprocess.PIPE) + ps_output = ps_command.stdout.read() + retcode = ps_command.wait() + assert retcode == 0, "ps command returned %d" % retcode + + for pid_str in ps_output.split("\n")[:-1]: + try: + os.kill(int(pid_str), raised_signal) + except Exception: + continue + + def _start_pserver(self, use_cuda, sync_mode): + p = Process( + target=run_pserver, + args=(use_cuda, sync_mode, self.ip, self.port, self.trainer_count, + self.trainer_id)) + p.start() + + def test_handle_signal_in_serv_op(self): + # run pserver on CPU in sync mode + self._start_pserver(False, True) + + # raise SIGINT to pserver + self._raise_signal(os.getpid(), signal.SIGINT) + + # run pserver on CPU in async mode + self._start_pserver(False, False) + + # raise SIGTERM to pserver + self._raise_signal(os.getpid(), signal.SIGTERM) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_lookup_sparse_table_op.py b/python/paddle/fluid/tests/unittests/test_lookup_sparse_table_op.py new file mode 100644 index 0000000000000000000000000000000000000000..aa9eae1e882f55ef51f38e158317a1a9aeed641c --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_lookup_sparse_table_op.py @@ -0,0 +1,86 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest +import numpy as np +from op_test import OpTest +import paddle.fluid.core as core +from paddle.fluid.op import Operator + + +def output_hist(out): + hist, _ = np.histogram(out, range=(-5, 10)) + hist = hist.astype("float32") + hist /= float(out.size) + prob = 0.1 * np.ones((10)) + return hist, prob + + +class TestLookupSpraseTable(OpTest): + def check_with_place(self, place): + scope = core.Scope() + + # create and initialize Id Variable + ids = scope.var("Ids").get_tensor() + ids_array = np.array([0, 2, 3, 5, 100]).astype("int64") + ids.set(ids_array, place) + + # create and initialize W Variable + rows = [0, 1, 2, 3, 4, 5, 6] + row_numel = 10000 + + w_selected_rows = scope.var('W').get_selected_rows() + w_selected_rows.set_height(len(rows)) + w_selected_rows.set_rows(rows) + w_array = np.ones((len(rows), row_numel)).astype("float32") + for i in range(len(rows)): + w_array[i] *= i + w_tensor = w_selected_rows.get_tensor() + w_tensor.set(w_array, place) + + # create Out Variable + out_tensor = scope.var('Out').get_tensor() + + # create and run lookup_table operator + lookup_table = Operator( + "lookup_sparse_table", + W='W', + Ids='Ids', + Out='Out', + min=-5.0, + max=10.0, + seed=10) + lookup_table.run(scope, place) + + # get result from Out + result_array = np.array(out_tensor) + # all(): return True if all elements of the iterable are true (or if the iterable is empty) + for idx, row in enumerate(ids_array[:-2]): + assert (row == result_array[idx]).all() + + # check the random value + hist, prob = output_hist(result_array[-1]) + self.assertTrue( + np.allclose( + hist, prob, rtol=0, atol=0.01), "hist: " + str(hist)) + + def test_w_is_selected_rows(self): + places = [core.CPUPlace()] + # currently only support CPU + for place in places: + self.check_with_place(place) + + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_matmul_op.py b/python/paddle/fluid/tests/unittests/test_matmul_op.py index 44ac4683891ffd3141a126740f4fddb47550e183..cae2c8fa87d9857de8f26cf4962d9370eca66243 100644 --- a/python/paddle/fluid/tests/unittests/test_matmul_op.py +++ b/python/paddle/fluid/tests/unittests/test_matmul_op.py @@ -111,21 +111,24 @@ class Generator(object): # Generate test cases for all possibilities -for dim_X in [1, 2, 3]: - for dim_Y in [1, 2, 3]: - for transpose_X in [False, True]: - for transpose_Y in [False, True]: - test_name = ( - 'TestMatMulOp_dimX_{}_dim_Y_{}_transX_{}_transY_{}'.format( - dim_X, dim_Y, transpose_X, transpose_Y)) - shape_X, shape_Y = generate_compatible_shapes( - dim_X, dim_Y, transpose_X, transpose_Y) - globals()[test_name] = type(test_name, (Generator, OpTest), { - 'shape_X': shape_X, - 'shape_Y': shape_Y, - 'transpose_X': transpose_X, - 'transpose_Y': transpose_Y, - }) +def inject_test(dim_x, dim_y, trans_x, trans_y): + test_name = ('TestMatMulOp_dimX_{}_dim_Y_{}_transX_{}_transY_{}'.format( + dim_x, dim_y, trans_x, trans_y)) + shape_x, shape_y = generate_compatible_shapes(dim_x, dim_y, trans_x, + trans_y) + globals()[test_name] = type(test_name, (Generator, OpTest), { + 'shape_X': shape_x, + 'shape_Y': shape_y, + 'transpose_X': trans_x, + 'transpose_Y': trans_y, + }) + + +for dim_X in (1, 2, 3): + for dim_Y in (1, 2, 3): + for transose_x in (False, True): + for transose_y in (False, True): + inject_test(dim_X, dim_Y, transose_x, transose_y) # Test case n-dim @@ -149,7 +152,7 @@ def generate_compatible_shapes(dim, transpose_X, transpose_Y): return shape_X, shape_Y -# Test case n-dim +# # Test case n-dim for dim in [4]: for transpose_X in [False, True]: for transpose_Y in [False, True]: diff --git a/python/paddle/fluid/tests/unittests/test_memory_optimization_transpiler.py b/python/paddle/fluid/tests/unittests/test_memory_optimization_transpiler.py index f3dcca6b0107a9c4a6efcb0c0fd50324aaf92648..cfd6e63e12258a92447e68b4afbc7ead91b68cc1 100644 --- a/python/paddle/fluid/tests/unittests/test_memory_optimization_transpiler.py +++ b/python/paddle/fluid/tests/unittests/test_memory_optimization_transpiler.py @@ -18,7 +18,7 @@ import unittest import paddle.fluid.layers as layers import paddle.fluid.optimizer as optimizer from paddle.fluid.framework import Program, program_guard -from paddle.fluid.memory_optimization_transpiler import memory_optimize +from paddle.fluid.transpiler import memory_optimize class TestControlFlowGraph(unittest.TestCase): diff --git a/python/paddle/fluid/tests/unittests/test_mul_mkldnn_op.py b/python/paddle/fluid/tests/unittests/test_mul_mkldnn_op.py deleted file mode 100644 index 42d68ef376dc4a664a96ff5a24545c1997ee924a..0000000000000000000000000000000000000000 --- a/python/paddle/fluid/tests/unittests/test_mul_mkldnn_op.py +++ /dev/null @@ -1,44 +0,0 @@ -# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -import unittest -from test_mul_op import TestMulOp, TestMulOp2, TestFP16MulOp1, TestFP16MulOp2 - - -class TestMKLDNNMulOp(TestMulOp): - def init_op_test(self): - super(TestMKLDNNMulOp, self).setUp() - self.attrs = {"use_mkldnn": True} - - -class TestMKLDNNMulOp2(TestMulOp2): - def init_op_test(self): - super(TestMKLDNNMulOp2, self).setUp() - self.attrs = {"use_mkldnn": True} - - -class TestMKLDNNFP16MulOp1(TestFP16MulOp1): - def init_op_test(self): - super(TestMKLDNNFP16MulOp1, self).setUp() - self.attrs = {"use_mkldnn": True} - - -class TestMKLDNNFP16MulOp2(TestFP16MulOp2): - def init_op_test(self): - super(TestMKLDNNFP16MulOp2, self).setUp() - self.attrs = {"use_mkldnn": True} - - -if __name__ == "__main__": - unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_mul_op.py b/python/paddle/fluid/tests/unittests/test_mul_op.py index d984393c89f44f5b9679a22bf7bb6182599233e3..862b7f8cb93620da4dd4673028776cfe565eeb0b 100644 --- a/python/paddle/fluid/tests/unittests/test_mul_op.py +++ b/python/paddle/fluid/tests/unittests/test_mul_op.py @@ -21,12 +21,10 @@ from op_test import OpTest class TestMulOp(OpTest): def setUp(self): self.op_type = "mul" - self.use_mkldnn = False self.inputs = { 'X': np.random.random((32, 84)).astype("float32"), 'Y': np.random.random((84, 100)).astype("float32") } - self.attrs = {'use_mkldnn': self.use_mkldnn} self.outputs = {'Out': np.dot(self.inputs['X'], self.inputs['Y'])} def test_check_output(self): @@ -47,16 +45,11 @@ class TestMulOp(OpTest): class TestMulOp2(OpTest): def setUp(self): self.op_type = "mul" - self.use_mkldnn = False self.inputs = { 'X': np.random.random((15, 4, 12, 10)).astype("float32"), 'Y': np.random.random((4, 30, 8, 2, 9)).astype("float32") } - self.attrs = { - 'x_num_col_dims': 2, - 'y_num_col_dims': 2, - 'use_mkldnn': self.use_mkldnn - } + self.attrs = {'x_num_col_dims': 2, 'y_num_col_dims': 2} result = np.dot(self.inputs['X'].reshape(15 * 4, 12 * 10), self.inputs['Y'].reshape(4 * 30, 8 * 2 * 9)) result = result.reshape(15, 4, 8, 2, 9) @@ -80,11 +73,9 @@ class TestMulOp2(OpTest): class TestFP16MulOp1(OpTest): def setUp(self): self.op_type = "mul" - self.use_mkldnn = False x = np.random.random((32, 84)).astype("float16") y = np.random.random((84, 100)).astype("float16") self.inputs = {'X': x.view(np.uint16), 'Y': y.view(np.uint16)} - self.attrs = {'use_mkldnn': self.use_mkldnn} self.outputs = {'Out': np.dot(x, y)} def test_check_output(self): @@ -97,15 +88,10 @@ class TestFP16MulOp1(OpTest): class TestFP16MulOp2(OpTest): def setUp(self): self.op_type = "mul" - self.use_mkldnn = False x = np.random.random((15, 4, 12, 10)).astype("float16") y = np.random.random((4, 30, 8, 2, 9)).astype("float16") self.inputs = {'X': x.view(np.uint16), 'Y': y.view(np.uint16)} - self.attrs = { - 'x_num_col_dims': 2, - 'y_num_col_dims': 2, - 'use_mkldnn': self.use_mkldnn - } + self.attrs = {'x_num_col_dims': 2, 'y_num_col_dims': 2} result = np.dot( x.reshape(15 * 4, 12 * 10), y.reshape(4 * 30, 8 * 2 * 9)) result = result.reshape(15, 4, 8, 2, 9) diff --git a/python/paddle/fluid/tests/unittests/test_network_with_dtype.py b/python/paddle/fluid/tests/unittests/test_network_with_dtype.py new file mode 100644 index 0000000000000000000000000000000000000000..d4835dd18405fc7a0d508a780a734922e0abd12c --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_network_with_dtype.py @@ -0,0 +1,74 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest + +import numpy as np +import paddle +import paddle.fluid as fluid +import paddle.fluid.core as core +from paddle.fluid.executor import Executor + +BATCH_SIZE = 20 + + +class TestNetWithDtype(unittest.TestCase): + def setUp(self): + self.dtype = "float64" + self.init_dtype() + + def run_net_on_place(self, place): + main = fluid.Program() + startup = fluid.Program() + with fluid.program_guard(main, startup): + x = fluid.layers.data(name='x', shape=[13], dtype=self.dtype) + y = fluid.layers.data(name='y', shape=[1], dtype=self.dtype) + y_predict = fluid.layers.fc(input=x, size=1, act=None) + cost = fluid.layers.square_error_cost(input=y_predict, label=y) + avg_cost = fluid.layers.mean(cost) + sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.001) + sgd_optimizer.minimize(avg_cost) + + fetch_list = [avg_cost] + train_reader = paddle.batch( + paddle.dataset.uci_housing.train(), batch_size=BATCH_SIZE) + feeder = fluid.DataFeeder(place=place, feed_list=[x, y]) + exe = fluid.Executor(place) + exe.run(startup) + for data in train_reader(): + exe.run(main, feed=feeder.feed(data), fetch_list=fetch_list) + # the main program is runable, the datatype is fully supported + break + + def init_dtype(self): + pass + + def test_cpu(self): + place = fluid.CPUPlace() + self.run_net_on_place(place) + + def test_gpu(self): + if not core.is_compiled_with_cuda(): + return + place = fluid.CUDAPlace(0) + self.run_net_on_place(place) + + +# TODO(dzhwinter): make sure the fp16 is runable +# class TestFloat16(TestNetWithDtype): +# def init_dtype(self): +# self.dtype = "float16" + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_operator_desc.py b/python/paddle/fluid/tests/unittests/test_operator_desc.py index 779ae388f04496a7be9a6d5aa4e39b8245022925..c098a5a0cb0364f9ec93c95c1ef50912e574b3d9 100644 --- a/python/paddle/fluid/tests/unittests/test_operator_desc.py +++ b/python/paddle/fluid/tests/unittests/test_operator_desc.py @@ -63,7 +63,7 @@ class TestOperator(unittest.TestCase): self.assertEqual(mul_op.output("Out"), ["mul.out"]) self.assertEqual( set(mul_op.attr_names), - set(["x_num_col_dims", "y_num_col_dims", "use_mkldnn"])) + set(["x_num_col_dims", "y_num_col_dims", "op_role", "op_role_var"])) self.assertEqual(mul_op.has_attr("x_num_col_dims"), True) self.assertEqual(mul_op.attr_type("x_num_col_dims"), core.AttrType.INT) self.assertEqual(mul_op.attr("x_num_col_dims"), 1) diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor.py b/python/paddle/fluid/tests/unittests/test_parallel_executor.py deleted file mode 100644 index c783a142467f3f6a9cd210425acfc526a32a6f71..0000000000000000000000000000000000000000 --- a/python/paddle/fluid/tests/unittests/test_parallel_executor.py +++ /dev/null @@ -1,696 +0,0 @@ -# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -import numpy -import unittest - -import paddle.fluid as fluid -import paddle -import paddle.dataset.mnist as mnist -import paddle.dataset.wmt16 as wmt16 - - -def simple_fc_net(use_feed): - if use_feed: - img = fluid.layers.data(name='image', shape=[784], dtype='float32') - label = fluid.layers.data(name='label', shape=[1], dtype='int64') - else: - reader = fluid.layers.open_files( - filenames=['./mnist.recordio'], - shapes=[[-1, 784], [-1, 1]], - lod_levels=[0, 0], - dtypes=['float32', 'int64'], - thread_num=1, - for_parallel=True) - reader = fluid.layers.io.double_buffer(reader) - img, label = fluid.layers.read_file(reader) - hidden = img - for _ in xrange(4): - hidden = fluid.layers.fc( - hidden, - size=200, - act='tanh', - bias_attr=fluid.ParamAttr( - initializer=fluid.initializer.Constant(value=1.0))) - prediction = fluid.layers.fc(hidden, size=10, act='softmax') - loss = fluid.layers.cross_entropy(input=prediction, label=label) - loss = fluid.layers.mean(loss) - return loss - - -def fc_with_batchnorm(use_feed): - if use_feed: - img = fluid.layers.data(name='image', shape=[784], dtype='float32') - label = fluid.layers.data(name='label', shape=[1], dtype='int64') - else: - reader = fluid.layers.open_files( - filenames=['mnist.recordio'], - shapes=[[-1, 784], [-1, 1]], - lod_levels=[0, 0], - dtypes=['float32', 'int64'], - thread_num=1, - for_parallel=True) - reader = fluid.layers.io.double_buffer(reader) - img, label = fluid.layers.read_file(reader) - - hidden = img - for _ in xrange(1): - hidden = fluid.layers.fc( - hidden, - size=200, - act='tanh', - bias_attr=fluid.ParamAttr( - initializer=fluid.initializer.Constant(value=1.0))) - - hidden = fluid.layers.batch_norm(input=hidden) - - prediction = fluid.layers.fc(hidden, size=10, act='softmax') - loss = fluid.layers.cross_entropy(input=prediction, label=label) - loss = fluid.layers.mean(loss) - return loss - - -def squeeze_excitation(input, num_channels, reduction_ratio): - # pool = fluid.layers.pool2d( - # input=input, pool_size=0, pool_type='avg', global_pooling=True) - conv = input - shape = conv.shape - reshape = fluid.layers.reshape( - x=conv, shape=[-1, shape[1], shape[2] * shape[3]]) - pool = fluid.layers.reduce_mean(input=reshape, dim=2) - - squeeze = fluid.layers.fc(input=pool, - size=num_channels / reduction_ratio, - act='relu') - excitation = fluid.layers.fc(input=squeeze, - size=num_channels, - act='sigmoid') - scale = fluid.layers.elementwise_mul(x=input, y=excitation, axis=0) - return scale - - -def conv_bn_layer(input, num_filters, filter_size, stride=1, groups=1, - act=None): - conv = fluid.layers.conv2d( - input=input, - num_filters=num_filters, - filter_size=filter_size, - stride=stride, - padding=(filter_size - 1) / 2, - groups=groups, - act=None, - bias_attr=False) - return fluid.layers.batch_norm(input=conv, act=act, momentum=0.1) - - -def shortcut(input, ch_out, stride): - ch_in = input.shape[1] - if ch_in != ch_out: - if stride == 1: - filter_size = 1 - else: - filter_size = 3 - return conv_bn_layer(input, ch_out, filter_size, stride) - else: - return input - - -def bottleneck_block(input, num_filters, stride, cardinality, reduction_ratio): - # The number of first 1x1 convolutional channels for each bottleneck build block - # was halved to reduce the compution cost. - conv0 = conv_bn_layer( - input=input, num_filters=num_filters, filter_size=1, act='relu') - conv1 = conv_bn_layer( - input=conv0, - num_filters=num_filters * 2, - filter_size=3, - stride=stride, - groups=cardinality, - act='relu') - conv2 = conv_bn_layer( - input=conv1, num_filters=num_filters * 2, filter_size=1, act=None) - scale = squeeze_excitation( - input=conv2, - num_channels=num_filters * 2, - reduction_ratio=reduction_ratio) - - short = shortcut(input, num_filters * 2, stride) - - return fluid.layers.elementwise_add(x=short, y=scale, act='relu') - - -def SE_ResNeXt50Small(batch_size=2, use_feed=False): - assert not use_feed, "SE_ResNeXt doesn't support feed yet" - - img = fluid.layers.fill_constant( - shape=[batch_size, 3, 224, 224], dtype='float32', value=0.0) - label = fluid.layers.fill_constant( - shape=[batch_size, 1], dtype='int64', value=0.0) - - conv = conv_bn_layer( - input=img, num_filters=16, filter_size=3, stride=2, act='relu') - conv = conv_bn_layer( - input=conv, num_filters=16, filter_size=3, stride=1, act='relu') - conv = conv_bn_layer( - input=conv, num_filters=16, filter_size=3, stride=1, act='relu') - conv = fluid.layers.pool2d( - input=conv, pool_size=3, pool_stride=2, pool_padding=1, pool_type='max') - - cardinality = 32 - reduction_ratio = 16 - depth = [3, 4, 6, 3] - num_filters = [128, 256, 512, 1024] - - for block in range(len(depth)): - for i in range(depth[block]): - conv = bottleneck_block( - input=conv, - num_filters=num_filters[block], - stride=2 if i == 0 and block != 0 else 1, - cardinality=cardinality, - reduction_ratio=reduction_ratio) - - shape = conv.shape - reshape = fluid.layers.reshape( - x=conv, shape=[-1, shape[1], shape[2] * shape[3]]) - pool = fluid.layers.reduce_mean(input=reshape, dim=2) - dropout = fluid.layers.dropout(x=pool, dropout_prob=0.2) - # Classifier layer: - prediction = fluid.layers.fc(input=dropout, size=1000, act='softmax') - loss = fluid.layers.cross_entropy(input=prediction, label=label) - loss = fluid.layers.mean(loss) - return loss - - -import time - - -class TestParallelExecutorBase(unittest.TestCase): - def check_network_convergence(self, - method, - memory_opt=True, - iter=50, - batch_size=None, - allow_op_delay=False, - feed_dict=None, - seed=None, - use_parallel_executor=True): - def run_executor(exe, feed, fetch_list, program=None): - if isinstance(exe, fluid.ParallelExecutor): - res = exe.run(fetch_list=fetch_list, feed=feed) - elif isinstance(exe, fluid.Executor): - if program is None: - program = fluid.default_main_program() - res = exe.run(program=program, feed=feed, fetch_list=fetch_list) - else: - raise ValueError('Unkown type exe') - return res - - main = fluid.Program() - startup = fluid.Program() - startup.random_seed = 1 # Fix random seed - with fluid.program_guard(main, startup): - if seed is not None: - startup.random_seed = seed - loss = method(use_feed=feed_dict is not None) - adam = fluid.optimizer.Adam() - adam.minimize(loss) - if memory_opt: - fluid.memory_optimize(main) - place = fluid.CUDAPlace(0) - startup_exe = fluid.Executor(place) - startup_exe.run(startup) - - if use_parallel_executor: - exe = fluid.ParallelExecutor( - True, loss_name=loss.name, allow_op_delay=allow_op_delay) - else: - exe = fluid.Executor(place=place) - - if batch_size is not None: - batch_size *= fluid.core.get_cuda_device_count() - begin = time.time() - first_loss, = run_executor( - exe=exe, feed=feed_dict, fetch_list=[loss.name]) - first_loss = numpy.array(first_loss) - - for i in xrange(iter): - run_executor(exe=exe, feed=feed_dict, fetch_list=[]) - - last_loss, = run_executor( - exe=exe, feed=feed_dict, fetch_list=[loss.name]) - end = time.time() - - if batch_size is not None: - print "%.4f Instance per second" % ( - (batch_size * iter + 2) / (end - begin)) - - last_loss = numpy.array(last_loss) - - print first_loss, last_loss - # self.assertGreater(first_loss[0], last_loss[0]) - return first_loss, last_loss - - -class TestMNIST(TestParallelExecutorBase): - @classmethod - def setUpClass(cls): - # Convert mnist to recordio file - with fluid.program_guard(fluid.Program(), fluid.Program()): - reader = paddle.batch(mnist.train(), batch_size=4) - feeder = fluid.DataFeeder( - feed_list=[ # order is image and label - fluid.layers.data( - name='image', shape=[784]), - fluid.layers.data( - name='label', shape=[1], dtype='int64'), - ], - place=fluid.CPUPlace()) - fluid.recordio_writer.convert_reader_to_recordio_file( - './mnist.recordio', reader, feeder) - - def test_simple_fc(self): - self.check_network_convergence(simple_fc_net) - self.check_network_convergence(simple_fc_net, allow_op_delay=True) - - img = numpy.zeros(shape=[32, 784], dtype='float32') - label = numpy.ones(shape=[32, 1], dtype='int64') - self.check_network_convergence( - simple_fc_net, feed_dict={"image": img, - "label": label}) - - def test_simple_fc_parallel_accuracy(self): - img = numpy.zeros(shape=[32, 784], dtype='float32') - label = numpy.ones(shape=[32, 1], dtype='int64') - single_first_loss, single_last_loss = self.check_network_convergence( - method=simple_fc_net, - seed=1000, - feed_dict={"image": img, - "label": label}, - use_parallel_executor=False) - parallel_first_loss, parallel_last_loss = self.check_network_convergence( - method=simple_fc_net, - seed=1000, - feed_dict={"image": img, - "label": label}, - use_parallel_executor=True) - - for p_f in parallel_first_loss: - self.assertAlmostEquals(p_f, single_first_loss[0], delta=1e-6) - for p_l in parallel_last_loss: - self.assertAlmostEquals(p_l, single_last_loss[0], delta=1e-6) - - def test_batchnorm_fc(self): - self.check_network_convergence(fc_with_batchnorm) - img = numpy.zeros(shape=[32, 784], dtype='float32') - label = numpy.ones(shape=[32, 1], dtype='int64') - self.check_network_convergence( - fc_with_batchnorm, feed_dict={"image": img, - "label": label}) - - -class TestResnet(TestParallelExecutorBase): - # @classmethod - # def setUpClass(cls): - # # import os - # # if os.path.exists('./flowers.recordio'): - # # return - # with fluid.program_guard(fluid.Program(), fluid.Program()): - # reader = paddle.batch(flowers.train(), batch_size=4) - # feeder = fluid.DataFeeder( - # feed_list=[ - # fluid.layers.data( - # name='image', shape=[3, 224, 224]), - # fluid.layers.data( - # name='label', shape=[1], dtype='int64'), - # ], - # place=fluid.CPUPlace()) - # fluid.recordio_writer.convert_reader_to_recordio_file( - # "./flowers.recordio", reader, feeder, compressor=fluid.core.RecordIOWriter.Compressor.NoCompress) - - def test_resnet(self): - import functools - batch_size = 2 - self.check_network_convergence( - functools.partial( - SE_ResNeXt50Small, batch_size=batch_size), - iter=20, - batch_size=batch_size) - - -class ModelHyperParams(object): - # Dictionary size for source and target language. This model directly uses - # paddle.dataset.wmt16 in which , and token has - # alreay been added, but the token is not added. Transformer requires - # sequences in a mini-batch are padded to have the same length. A token is - # added into the original dictionary in paddle.dateset.wmt16. - - # size of source word dictionary. - src_vocab_size = 10000 - # index for token in source language. - src_pad_idx = src_vocab_size - - # size of target word dictionay - trg_vocab_size = 10000 - # index for token in target language. - trg_pad_idx = trg_vocab_size - - # position value corresponding to the token. - pos_pad_idx = 0 - - # max length of sequences. It should plus 1 to include position - # padding token for position encoding. - max_length = 50 - - # the dimension for word embeddings, which is also the last dimension of - # the input and output of multi-head attention, position-wise feed-forward - # networks, encoder and decoder. - - d_model = 512 - # size of the hidden layer in position-wise feed-forward networks. - d_inner_hid = 1024 - # the dimension that keys are projected to for dot-product attention. - d_key = 64 - # the dimension that values are projected to for dot-product attention. - d_value = 64 - # number of head used in multi-head attention. - n_head = 8 - # number of sub-layers to be stacked in the encoder and decoder. - n_layer = 6 - # dropout rate used by all dropout layers. - dropout = 0.1 - - -import numpy as np - - -def prepare_batch_input(insts, src_pad_idx, trg_pad_idx, n_head): - """ - Pad the instances to the max sequence length in batch, and generate the - corresponding position data and attention bias. Then, convert the numpy - data to tensors and return a dict mapping names to tensors. - """ - - def __pad_batch_data(insts, - pad_idx, - is_target=False, - return_pos=True, - return_attn_bias=True, - return_max_len=True): - """ - Pad the instances to the max sequence length in batch, and generate the - corresponding position data and attention bias. - """ - return_list = [] - max_len = max(len(inst) for inst in insts) - inst_data = np.array( - [inst + [pad_idx] * (max_len - len(inst)) for inst in insts]) - return_list += [inst_data.astype("int64").reshape([-1, 1])] - if return_pos: - inst_pos = np.array([[ - pos_i + 1 if w_i != pad_idx else 0 - for pos_i, w_i in enumerate(inst) - ] for inst in inst_data]) - - return_list += [inst_pos.astype("int64").reshape([-1, 1])] - if return_attn_bias: - if is_target: - # This is used to avoid attention on paddings and subsequent - # words. - slf_attn_bias_data = np.ones((inst_data.shape[0], max_len, - max_len)) - slf_attn_bias_data = np.triu(slf_attn_bias_data, 1).reshape( - [-1, 1, max_len, max_len]) - slf_attn_bias_data = np.tile(slf_attn_bias_data, - [1, n_head, 1, 1]) * [-1e9] - else: - # This is used to avoid attention on paddings. - slf_attn_bias_data = np.array([[0] * len(inst) + [-1e9] * - (max_len - len(inst)) - for inst in insts]) - slf_attn_bias_data = np.tile( - slf_attn_bias_data.reshape([-1, 1, 1, max_len]), - [1, n_head, max_len, 1]) - return_list += [slf_attn_bias_data.astype("float32")] - if return_max_len: - return_list += [max_len] - return return_list if len(return_list) > 1 else return_list[0] - - def data_to_tensor(data_list, name_list, input_dict, place): - assert len(data_list) == len(name_list) - for i in range(len(name_list)): - tensor = fluid.LoDTensor() - tensor.set(data_list[i], place) - input_dict[name_list[i]] = tensor - - src_word, src_pos, src_slf_attn_bias, src_max_len = __pad_batch_data( - [inst[0] for inst in insts], src_pad_idx, is_target=False) - trg_word, trg_pos, trg_slf_attn_bias, trg_max_len = __pad_batch_data( - [inst[1] for inst in insts], trg_pad_idx, is_target=True) - trg_src_attn_bias = np.tile(src_slf_attn_bias[:, :, ::src_max_len, :], - [1, 1, trg_max_len, 1]).astype("float32") - lbl_word = __pad_batch_data([inst[2] for inst in insts], trg_pad_idx, False, - False, False, False) - lbl_weight = (lbl_word != trg_pad_idx).astype("float32").reshape([-1, 1]) - - return [ - src_word, src_pos, trg_word, trg_pos, src_slf_attn_bias, - trg_slf_attn_bias, trg_src_attn_bias, lbl_word, lbl_weight - ] - - -import transformer_model - - -def transformer(use_feed): - assert not use_feed, "transfomer doesn't support feed yet" - return transformer_model.transformer( - ModelHyperParams.src_vocab_size + 1, - ModelHyperParams.trg_vocab_size + 1, ModelHyperParams.max_length + 1, - ModelHyperParams.n_layer, ModelHyperParams.n_head, - ModelHyperParams.d_key, ModelHyperParams.d_value, - ModelHyperParams.d_model, ModelHyperParams.d_inner_hid, - ModelHyperParams.dropout, ModelHyperParams.src_pad_idx, - ModelHyperParams.trg_pad_idx, ModelHyperParams.pos_pad_idx) - - -class TestTransformer(TestParallelExecutorBase): - @classmethod - def setUpClass(cls): - reader = paddle.batch( - wmt16.train(ModelHyperParams.src_vocab_size, - ModelHyperParams.trg_vocab_size), - batch_size=transformer_model.batch_size) - - with fluid.recordio_writer.create_recordio_writer( - "./wmt16.recordio") as writer: - for batch in reader(): - for tensor in prepare_batch_input( - batch, ModelHyperParams.src_pad_idx, - ModelHyperParams.trg_pad_idx, ModelHyperParams.n_head): - t = fluid.LoDTensor() - t.set(tensor, fluid.CPUPlace()) - writer.append_tensor(t) - writer.complete_append_tensor() - - @unittest.skip("transformer is buggy in multi gpu") - def test_main(self): - self.check_network_convergence(transformer) - - -class ParallelExecutorTestingDuringTraining(unittest.TestCase): - def test_parallel_testing(self): - main = fluid.Program() - startup = fluid.Program() - with fluid.program_guard(main, startup): - loss = simple_fc_net(True) - test_program = main.clone(for_test=True) - - opt = fluid.optimizer.SGD(learning_rate=0.001) - opt.minimize(loss) - - batch_size = 32 - image = numpy.random.normal(size=(batch_size, - 784)).astype('float32') - label = numpy.random.randint(0, 10, (batch_size, 1), dtype="int64") - - place = fluid.CUDAPlace(0) - exe = fluid.Executor(place) - exe.run(startup) - feed_dict = {'image': image, 'label': label} - - train_exe = fluid.ParallelExecutor( - use_cuda=True, loss_name=loss.name, main_program=main) - - test_exe = fluid.ParallelExecutor( - use_cuda=True, - main_program=test_program, - share_vars_from=train_exe) - - for i in xrange(5): - test_loss, = test_exe.run([loss.name], feed=feed_dict) - test_loss = numpy.array(test_loss) - - train_loss, = train_exe.run([loss.name], feed=feed_dict) - train_loss = numpy.array(train_loss) - self.assertTrue( - numpy.allclose( - train_loss, test_loss, atol=1e-8), - "Train loss: " + str(train_loss) + "\n Test loss:" + - str(test_loss)) - - -import paddle.dataset.conll05 as conll05 -import paddle.fluid as fluid - -word_dict, verb_dict, label_dict = conll05.get_dict() -word_dict_len = len(word_dict) -label_dict_len = len(label_dict) -pred_dict_len = len(verb_dict) -mark_dict_len = 2 -word_dim = 32 -mark_dim = 5 -hidden_dim = 512 -depth = 8 -mix_hidden_lr = 1e-3 -embedding_name = 'emb' - - -def db_lstm(word, predicate, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2, mark, - **ignored): - # 8 features - predicate_embedding = fluid.layers.embedding( - input=predicate, - size=[pred_dict_len, word_dim], - dtype='float32', - param_attr='vemb') - - mark_embedding = fluid.layers.embedding( - input=mark, size=[mark_dict_len, mark_dim], dtype='float32') - - word_input = [word, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2] - emb_layers = [ - fluid.layers.embedding( - size=[word_dict_len, word_dim], - input=x, - param_attr=fluid.ParamAttr( - name=embedding_name, trainable=False)) for x in word_input - ] - emb_layers.append(predicate_embedding) - emb_layers.append(mark_embedding) - - hidden_0_layers = [ - fluid.layers.fc(input=emb, size=hidden_dim, act='tanh') - for emb in emb_layers - ] - - hidden_0 = fluid.layers.sums(input=hidden_0_layers) - - lstm_0 = fluid.layers.dynamic_lstm( - input=hidden_0, - size=hidden_dim, - candidate_activation='relu', - gate_activation='sigmoid', - cell_activation='sigmoid') - - # stack L-LSTM and R-LSTM with direct edges - input_tmp = [hidden_0, lstm_0] - - for i in range(1, depth): - mix_hidden = fluid.layers.sums(input=[ - fluid.layers.fc(input=input_tmp[0], size=hidden_dim, act='tanh'), - fluid.layers.fc(input=input_tmp[1], size=hidden_dim, act='tanh') - ]) - - lstm = fluid.layers.dynamic_lstm( - input=mix_hidden, - size=hidden_dim, - candidate_activation='relu', - gate_activation='sigmoid', - cell_activation='sigmoid', - is_reverse=((i % 2) == 1)) - - input_tmp = [mix_hidden, lstm] - - feature_out = fluid.layers.sums(input=[ - fluid.layers.fc(input=input_tmp[0], size=label_dict_len, act='tanh'), - fluid.layers.fc(input=input_tmp[1], size=label_dict_len, act='tanh') - ]) - - return feature_out - - -class TestCRFModel(unittest.TestCase): - def test_all(self): - main = fluid.Program() - startup = fluid.Program() - with fluid.program_guard(main, startup): - word = fluid.layers.data( - name='word_data', shape=[1], dtype='int64', lod_level=1) - predicate = fluid.layers.data( - name='verb_data', shape=[1], dtype='int64', lod_level=1) - ctx_n2 = fluid.layers.data( - name='ctx_n2_data', shape=[1], dtype='int64', lod_level=1) - ctx_n1 = fluid.layers.data( - name='ctx_n1_data', shape=[1], dtype='int64', lod_level=1) - ctx_0 = fluid.layers.data( - name='ctx_0_data', shape=[1], dtype='int64', lod_level=1) - ctx_p1 = fluid.layers.data( - name='ctx_p1_data', shape=[1], dtype='int64', lod_level=1) - ctx_p2 = fluid.layers.data( - name='ctx_p2_data', shape=[1], dtype='int64', lod_level=1) - mark = fluid.layers.data( - name='mark_data', shape=[1], dtype='int64', lod_level=1) - feature_out = db_lstm(**locals()) - target = fluid.layers.data( - name='target', shape=[1], dtype='int64', lod_level=1) - crf_cost = fluid.layers.linear_chain_crf( - input=feature_out, - label=target, - param_attr=fluid.ParamAttr( - name='crfw', learning_rate=1e-1)) - avg_cost = fluid.layers.mean(crf_cost) - - sgd_optimizer = fluid.optimizer.SGD( - learning_rate=fluid.layers.exponential_decay( - learning_rate=0.01, - decay_steps=100000, - decay_rate=0.5, - staircase=True)) - sgd_optimizer.minimize(avg_cost) - - train_data = paddle.batch( - paddle.reader.shuffle( - paddle.dataset.conll05.test(), buf_size=8192), - batch_size=16) - - place = fluid.CUDAPlace(0) - exe = fluid.Executor(place) - exe.run(startup) - - pe = fluid.ParallelExecutor(use_cuda=True, loss_name=avg_cost.name) - - feeder = fluid.DataFeeder( - feed_list=[ - word, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2, predicate, - mark, target - ], - place=fluid.CPUPlace()) - - data = train_data() - for i in xrange(10): - cur_batch = next(data) - print map(numpy.array, - pe.run(feed=feeder.feed(cur_batch), - fetch_list=[avg_cost.name]))[0] diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor_crf.py b/python/paddle/fluid/tests/unittests/test_parallel_executor_crf.py new file mode 100644 index 0000000000000000000000000000000000000000..66e138b03f3b170aca4fb2207438eb9af1783c33 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_parallel_executor_crf.py @@ -0,0 +1,197 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import paddle.dataset.conll05 as conll05 +import paddle.fluid as fluid +import unittest +import paddle +import numpy as np + +word_dict, verb_dict, label_dict = conll05.get_dict() +word_dict_len = len(word_dict) +label_dict_len = len(label_dict) +pred_dict_len = len(verb_dict) +mark_dict_len = 2 +word_dim = 32 +mark_dim = 5 +hidden_dim = 512 +depth = 8 +mix_hidden_lr = 1e-3 +embedding_name = 'emb' + + +def db_lstm(word, predicate, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2, mark, + is_sparse, **ignored): + # 8 features + predicate_embedding = fluid.layers.embedding( + input=predicate, + is_sparse=is_sparse, + size=[pred_dict_len, word_dim], + dtype='float32', + param_attr='vemb') + + mark_embedding = fluid.layers.embedding( + input=mark, + is_sparse=is_sparse, + size=[mark_dict_len, mark_dim], + dtype='float32') + + word_input = [word, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2] + emb_layers = [ + fluid.layers.embedding( + size=[word_dict_len, word_dim], + is_sparse=is_sparse, + input=x, + param_attr=fluid.ParamAttr( + name=embedding_name, trainable=False)) for x in word_input + ] + emb_layers.append(predicate_embedding) + emb_layers.append(mark_embedding) + + hidden_0_layers = [ + fluid.layers.fc(input=emb, size=hidden_dim, act='tanh') + for emb in emb_layers + ] + + hidden_0 = fluid.layers.sums(input=hidden_0_layers) + + lstm_0 = fluid.layers.dynamic_lstm( + input=hidden_0, + size=hidden_dim, + candidate_activation='relu', + gate_activation='sigmoid', + cell_activation='sigmoid') + + # stack L-LSTM and R-LSTM with direct edges + input_tmp = [hidden_0, lstm_0] + + for i in range(1, depth): + mix_hidden = fluid.layers.sums(input=[ + fluid.layers.fc(input=input_tmp[0], size=hidden_dim, act='tanh'), + fluid.layers.fc(input=input_tmp[1], size=hidden_dim, act='tanh') + ]) + + lstm = fluid.layers.dynamic_lstm( + input=mix_hidden, + size=hidden_dim, + candidate_activation='relu', + gate_activation='sigmoid', + cell_activation='sigmoid', + is_reverse=((i % 2) == 1)) + + input_tmp = [mix_hidden, lstm] + + feature_out = fluid.layers.sums(input=[ + fluid.layers.fc(input=input_tmp[0], size=label_dict_len, act='tanh'), + fluid.layers.fc(input=input_tmp[1], size=label_dict_len, act='tanh') + ]) + + return feature_out + + +class TestCRFModel(unittest.TestCase): + def check_network_convergence(self, is_sparse, build_strategy=None): + main = fluid.Program() + startup = fluid.Program() + with fluid.program_guard(main, startup): + word = fluid.layers.data( + name='word_data', shape=[1], dtype='int64', lod_level=1) + predicate = fluid.layers.data( + name='verb_data', shape=[1], dtype='int64', lod_level=1) + ctx_n2 = fluid.layers.data( + name='ctx_n2_data', shape=[1], dtype='int64', lod_level=1) + ctx_n1 = fluid.layers.data( + name='ctx_n1_data', shape=[1], dtype='int64', lod_level=1) + ctx_0 = fluid.layers.data( + name='ctx_0_data', shape=[1], dtype='int64', lod_level=1) + ctx_p1 = fluid.layers.data( + name='ctx_p1_data', shape=[1], dtype='int64', lod_level=1) + ctx_p2 = fluid.layers.data( + name='ctx_p2_data', shape=[1], dtype='int64', lod_level=1) + mark = fluid.layers.data( + name='mark_data', shape=[1], dtype='int64', lod_level=1) + + feature_out = db_lstm(**locals()) + target = fluid.layers.data( + name='target', shape=[1], dtype='int64', lod_level=1) + crf_cost = fluid.layers.linear_chain_crf( + input=feature_out, + label=target, + param_attr=fluid.ParamAttr( + name='crfw', learning_rate=1e-1)) + avg_cost = fluid.layers.mean(crf_cost) + + sgd_optimizer = fluid.optimizer.SGD( + learning_rate=fluid.layers.exponential_decay( + learning_rate=0.01, + decay_steps=100000, + decay_rate=0.5, + staircase=True)) + sgd_optimizer.minimize(avg_cost) + + train_data = paddle.batch( + paddle.reader.shuffle( + paddle.dataset.conll05.test(), buf_size=8192), + batch_size=16) + + place = fluid.CUDAPlace(0) + exe = fluid.Executor(place) + exe.run(startup) + + pe = fluid.ParallelExecutor( + use_cuda=True, + loss_name=avg_cost.name, + build_strategy=build_strategy) + + feeder = fluid.DataFeeder( + feed_list=[ + word, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2, predicate, + mark, target + ], + place=fluid.CPUPlace()) + + data = train_data() + for i in xrange(10): + cur_batch = next(data) + print map(np.array, + pe.run(feed=feeder.feed(cur_batch), + fetch_list=[avg_cost.name]))[0] + + def test_update_sparse_parameter_all_reduce(self): + build_strategy = fluid.BuildStrategy() + build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.AllReduce + self.check_network_convergence( + is_sparse=True, build_strategy=build_strategy) + + def test_update_dense_parameter_all_reduce(self): + build_strategy = fluid.BuildStrategy() + build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.AllReduce + self.check_network_convergence( + is_sparse=False, build_strategy=build_strategy) + + def test_update_sparse_parameter_reduce(self): + build_strategy = fluid.BuildStrategy() + build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce + self.check_network_convergence( + is_sparse=True, build_strategy=build_strategy) + + def test_update_dense_parameter_reduce(self): + build_strategy = fluid.BuildStrategy() + build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce + self.check_network_convergence( + is_sparse=False, build_strategy=build_strategy) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor_fetch_feed.py b/python/paddle/fluid/tests/unittests/test_parallel_executor_fetch_feed.py new file mode 100644 index 0000000000000000000000000000000000000000..24f8d28c0304a77a99213374b25d0db728eca265 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_parallel_executor_fetch_feed.py @@ -0,0 +1,132 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import paddle.dataset.flowers as flowers +import math +import paddle.fluid as fluid +import unittest +import numpy as np +import paddle + + +def Lenet(data, class_dim): + conv1 = fluid.layers.conv2d(data, 32, 5, 1, act=None) + bn1 = fluid.layers.batch_norm(conv1, act='relu') + pool1 = fluid.layers.pool2d(bn1, 2, 'max', 2) + conv2 = fluid.layers.conv2d(pool1, 50, 5, 1, act=None) + bn2 = fluid.layers.batch_norm(conv2, act='relu') + pool2 = fluid.layers.pool2d(bn2, 2, 'max', 2) + + fc1 = fluid.layers.fc(pool2, size=500, act='relu') + fc2 = fluid.layers.fc(fc1, size=class_dim, act='softmax') + + return fc2 + + +class TestFetchOp(unittest.TestCase): + def parallel_exe(self, train_inputs, seed): + main = fluid.Program() + startup = fluid.Program() + startup.random_seed = seed + with fluid.program_guard(main, startup): + data = fluid.layers.data( + name='image', shape=[3, 224, 224], dtype='float32') + label = fluid.layers.data(name='label', shape=[1], dtype='int64') + out = Lenet(data, class_dim=102) + loss = fluid.layers.cross_entropy(input=out, label=label) + loss = fluid.layers.mean(loss) + + opt = fluid.optimizer.Momentum( + learning_rate=0.1, + momentum=0.9, + regularization=fluid.regularizer.L2Decay(1e-4)) + + opt.minimize(loss) + + # TODO(zcd): I found that onece the memory optimizer is open, + # parallel_exe doesn't fetch some variable, such as conv2d_0.b_0@GRAD, + # conv2d_1.b_0@GRAD. Those variables should not be pruned. + # fluid.memory_optimize(main) + + place = fluid.CUDAPlace(0) + exe = fluid.Executor(place) + exe.run(startup) + + feeder = fluid.DataFeeder(place=place, feed_list=[data, label]) + pe = fluid.ParallelExecutor( + use_cuda=True, loss_name=loss.name, main_program=main) + + fetch_list = [] + all_vars = main.global_block().vars + for k, v in all_vars.iteritems(): + if 'tmp' not in k and k[0] is not '_' or v.persistable: + fetch_list.append(k) + + for data in train_inputs: + ret = pe.run(fetch_list, feed=feeder.feed(data)) + for i in range(len(fetch_list)): + assert not math.isnan(np.sum(ret[i])) and \ + not math.isinf(np.sum(ret[i])) + + def test_fetch_op(self): + tst_reader = paddle.batch(flowers.test(use_xmap=False), batch_size=16) + tst_reader_iter = tst_reader() + + iters = 3 + train_inputs = [] + for i in range(iters): + train_inputs.append(tst_reader_iter.next()) + + self.parallel_exe(train_inputs, seed=1) + + +class TestFeedParallel(unittest.TestCase): + def test_main(self): + main = fluid.Program() + startup = fluid.Program() + startup.random_seed = 1 + with fluid.scope_guard(fluid.core.Scope()): + with fluid.program_guard(main, startup): + data = fluid.layers.data( + name='image', shape=[3, 224, 224], dtype='float32') + label = fluid.layers.data( + name='label', shape=[1], dtype='int64') + out = Lenet(data, class_dim=102) + loss = fluid.layers.cross_entropy(input=out, label=label) + loss = fluid.layers.mean(loss) + opt = fluid.optimizer.Momentum( + learning_rate=0.1, + momentum=0.9, + regularization=fluid.regularizer.L2Decay(1e-4)) + + opt.minimize(loss) + place = fluid.CUDAPlace(0) + feeder = fluid.DataFeeder(place=place, feed_list=[data, label]) + reader = feeder.decorate_reader( + paddle.batch( + flowers.train(), batch_size=16), multi_devices=True) + exe = fluid.Executor(place) + exe.run(startup) + pe = fluid.ParallelExecutor( + use_cuda=True, loss_name=loss.name, main_program=main) + + for batch_id, data in enumerate(reader()): + loss_np = np.array(pe.run(feed=data, fetch_list=[loss.name])[0]) + print batch_id, loss_np + if batch_id == 2: + break + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor_mnist.py b/python/paddle/fluid/tests/unittests/test_parallel_executor_mnist.py new file mode 100644 index 0000000000000000000000000000000000000000..015703c3e25f4e11e64ab6a7de99da12bee608f6 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_parallel_executor_mnist.py @@ -0,0 +1,171 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from parallel_executor_test_base import TestParallelExecutorBase +import paddle.fluid as fluid +import numpy as np +import paddle +import paddle.dataset.mnist as mnist +import unittest + +MNIST_RECORDIO_FILE = "./mnist_test_pe.recordio" + + +def simple_fc_net(use_feed): + if use_feed: + img = fluid.layers.data(name='image', shape=[784], dtype='float32') + label = fluid.layers.data(name='label', shape=[1], dtype='int64') + else: + reader = fluid.layers.open_files( + filenames=[MNIST_RECORDIO_FILE], + shapes=[[-1, 784], [-1, 1]], + lod_levels=[0, 0], + dtypes=['float32', 'int64'], + thread_num=1, + for_parallel=True) + reader = fluid.layers.io.double_buffer(reader) + img, label = fluid.layers.read_file(reader) + hidden = img + for _ in xrange(4): + hidden = fluid.layers.fc( + hidden, + size=200, + act='tanh', + bias_attr=fluid.ParamAttr( + initializer=fluid.initializer.Constant(value=1.0))) + prediction = fluid.layers.fc(hidden, size=10, act='softmax') + loss = fluid.layers.cross_entropy(input=prediction, label=label) + loss = fluid.layers.mean(loss) + return loss + + +def fc_with_batchnorm(use_feed): + if use_feed: + img = fluid.layers.data(name='image', shape=[784], dtype='float32') + label = fluid.layers.data(name='label', shape=[1], dtype='int64') + else: + reader = fluid.layers.open_files( + filenames=[MNIST_RECORDIO_FILE], + shapes=[[-1, 784], [-1, 1]], + lod_levels=[0, 0], + dtypes=['float32', 'int64'], + thread_num=1, + for_parallel=True) + reader = fluid.layers.io.double_buffer(reader) + img, label = fluid.layers.read_file(reader) + + hidden = img + for _ in xrange(1): + hidden = fluid.layers.fc( + hidden, + size=200, + act='tanh', + bias_attr=fluid.ParamAttr( + initializer=fluid.initializer.Constant(value=1.0))) + + hidden = fluid.layers.batch_norm(input=hidden) + + prediction = fluid.layers.fc(hidden, size=10, act='softmax') + loss = fluid.layers.cross_entropy(input=prediction, label=label) + loss = fluid.layers.mean(loss) + return loss + + +class TestMNIST(TestParallelExecutorBase): + @classmethod + def setUpClass(cls): + # Convert mnist to recordio file + with fluid.program_guard(fluid.Program(), fluid.Program()): + reader = paddle.batch(mnist.train(), batch_size=4) + feeder = fluid.DataFeeder( + feed_list=[ # order is image and label + fluid.layers.data( + name='image', shape=[784]), + fluid.layers.data( + name='label', shape=[1], dtype='int64'), + ], + place=fluid.CPUPlace()) + fluid.recordio_writer.convert_reader_to_recordio_file( + MNIST_RECORDIO_FILE, reader, feeder) + + def check_simple_fc_convergence(self, balance_parameter_opt_between_cards): + self.check_network_convergence(simple_fc_net) + self.check_network_convergence(simple_fc_net, allow_op_delay=True) + + img = np.zeros(shape=[32, 784], dtype='float32') + label = np.ones(shape=[32, 1], dtype='int64') + self.check_network_convergence( + simple_fc_net, + feed_dict={"image": img, + "label": label}, + balance_parameter_opt_between_cards=balance_parameter_opt_between_cards + ) + + def test_simple_fc(self): + self.check_simple_fc_convergence(False) + + def test_simple_fc_with_new_strategy(self): + self.check_simple_fc_convergence(True) + + def check_simple_fc_parallel_accuracy(self, + balance_parameter_opt_between_cards): + img = np.zeros(shape=[32, 784], dtype='float32') + label = np.ones(shape=[32, 1], dtype='int64') + single_first_loss, single_last_loss = self.check_network_convergence( + method=simple_fc_net, + seed=1000, + feed_dict={"image": img, + "label": label}, + use_parallel_executor=False) + parallel_first_loss, parallel_last_loss = self.check_network_convergence( + method=simple_fc_net, + seed=1000, + feed_dict={"image": img, + "label": label}, + use_parallel_executor=True, + balance_parameter_opt_between_cards=balance_parameter_opt_between_cards + ) + + for p_f in parallel_first_loss: + self.assertAlmostEquals(p_f, single_first_loss[0], delta=1e-6) + for p_l in parallel_last_loss: + self.assertAlmostEquals(p_l, single_last_loss[0], delta=1e-6) + + def test_simple_fc_parallel_accuracy(self): + self.check_simple_fc_parallel_accuracy(False) + + def test_simple_fc_parallel_accuracy_with_new_strategy(self): + self.check_simple_fc_parallel_accuracy(True) + + def check_batchnorm_fc_convergence(self, + balance_parameter_opt_between_cards): + self.check_network_convergence(fc_with_batchnorm) + img = np.zeros(shape=[32, 784], dtype='float32') + label = np.ones(shape=[32, 1], dtype='int64') + self.check_network_convergence( + fc_with_batchnorm, + feed_dict={"image": img, + "label": label}, + balance_parameter_opt_between_cards=balance_parameter_opt_between_cards + ) + + def test_batchnorm_fc(self): + self.check_batchnorm_fc_convergence(False) + + def test_batchnorm_fc_with_new_strategy(self): + self.check_batchnorm_fc_convergence(True) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor_seresnext.py b/python/paddle/fluid/tests/unittests/test_parallel_executor_seresnext.py new file mode 100644 index 0000000000000000000000000000000000000000..a3fa140cbb7994a36d2cbee26d598165f1f771d2 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_parallel_executor_seresnext.py @@ -0,0 +1,152 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import paddle.fluid as fluid +from parallel_executor_test_base import TestParallelExecutorBase +import unittest + + +def squeeze_excitation(input, num_channels, reduction_ratio): + # pool = fluid.layers.pool2d( + # input=input, pool_size=0, pool_type='avg', global_pooling=True) + conv = input + shape = conv.shape + reshape = fluid.layers.reshape( + x=conv, shape=[-1, shape[1], shape[2] * shape[3]]) + pool = fluid.layers.reduce_mean(input=reshape, dim=2) + + squeeze = fluid.layers.fc(input=pool, + size=num_channels / reduction_ratio, + act='relu') + excitation = fluid.layers.fc(input=squeeze, + size=num_channels, + act='sigmoid') + scale = fluid.layers.elementwise_mul(x=input, y=excitation, axis=0) + return scale + + +def conv_bn_layer(input, num_filters, filter_size, stride=1, groups=1, + act=None): + conv = fluid.layers.conv2d( + input=input, + num_filters=num_filters, + filter_size=filter_size, + stride=stride, + padding=(filter_size - 1) / 2, + groups=groups, + act=None, + bias_attr=False) + return fluid.layers.batch_norm(input=conv, act=act, momentum=0.1) + + +def shortcut(input, ch_out, stride): + ch_in = input.shape[1] + if ch_in != ch_out: + if stride == 1: + filter_size = 1 + else: + filter_size = 3 + return conv_bn_layer(input, ch_out, filter_size, stride) + else: + return input + + +def bottleneck_block(input, num_filters, stride, cardinality, reduction_ratio): + # The number of first 1x1 convolutional channels for each bottleneck build block + # was halved to reduce the compution cost. + conv0 = conv_bn_layer( + input=input, num_filters=num_filters, filter_size=1, act='relu') + conv1 = conv_bn_layer( + input=conv0, + num_filters=num_filters * 2, + filter_size=3, + stride=stride, + groups=cardinality, + act='relu') + conv2 = conv_bn_layer( + input=conv1, num_filters=num_filters * 2, filter_size=1, act=None) + scale = squeeze_excitation( + input=conv2, + num_channels=num_filters * 2, + reduction_ratio=reduction_ratio) + + short = shortcut(input, num_filters * 2, stride) + + return fluid.layers.elementwise_add(x=short, y=scale, act='relu') + + +def SE_ResNeXt50Small(batch_size=2, use_feed=False): + assert not use_feed, "SE_ResNeXt doesn't support feed yet" + + img = fluid.layers.fill_constant( + shape=[batch_size, 3, 224, 224], dtype='float32', value=0.0) + label = fluid.layers.fill_constant( + shape=[batch_size, 1], dtype='int64', value=0.0) + + conv = conv_bn_layer( + input=img, num_filters=16, filter_size=3, stride=2, act='relu') + conv = conv_bn_layer( + input=conv, num_filters=16, filter_size=3, stride=1, act='relu') + conv = conv_bn_layer( + input=conv, num_filters=16, filter_size=3, stride=1, act='relu') + conv = fluid.layers.pool2d( + input=conv, pool_size=3, pool_stride=2, pool_padding=1, pool_type='max') + + cardinality = 32 + reduction_ratio = 16 + depth = [3, 4, 6, 3] + num_filters = [128, 256, 512, 1024] + + for block in range(len(depth)): + for i in range(depth[block]): + conv = bottleneck_block( + input=conv, + num_filters=num_filters[block], + stride=2 if i == 0 and block != 0 else 1, + cardinality=cardinality, + reduction_ratio=reduction_ratio) + + shape = conv.shape + reshape = fluid.layers.reshape( + x=conv, shape=[-1, shape[1], shape[2] * shape[3]]) + pool = fluid.layers.reduce_mean(input=reshape, dim=2) + dropout = fluid.layers.dropout(x=pool, dropout_prob=0.2) + # Classifier layer: + prediction = fluid.layers.fc(input=dropout, size=1000, act='softmax') + loss = fluid.layers.cross_entropy(input=prediction, label=label) + loss = fluid.layers.mean(loss) + return loss + + +class TestResnet(TestParallelExecutorBase): + def check_resnet_convergence(self, balance_parameter_opt_between_cards): + import functools + batch_size = 2 + self.check_network_convergence( + functools.partial( + SE_ResNeXt50Small, batch_size=batch_size), + iter=20, + batch_size=batch_size, + balance_parameter_opt_between_cards=balance_parameter_opt_between_cards + ) + + def test_resnet(self): + self.check_resnet_convergence(False) + + def test_resnet_with_new_strategy(self): + self.check_resnet_convergence(True) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor_test_while_train.py b/python/paddle/fluid/tests/unittests/test_parallel_executor_test_while_train.py new file mode 100644 index 0000000000000000000000000000000000000000..93a5f767867d68110cf7b8f441cc740ecd843cf9 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_parallel_executor_test_while_train.py @@ -0,0 +1,93 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import paddle.fluid as fluid +import numpy as np +import unittest + + +def simple_fc_net(): + img = fluid.layers.data(name='image', shape=[784], dtype='float32') + label = fluid.layers.data(name='label', shape=[1], dtype='int64') + hidden = img + for _ in xrange(4): + hidden = fluid.layers.fc( + hidden, + size=200, + act='tanh', + bias_attr=fluid.ParamAttr( + initializer=fluid.initializer.Constant(value=1.0))) + prediction = fluid.layers.fc(hidden, size=10, act='softmax') + loss = fluid.layers.cross_entropy(input=prediction, label=label) + loss = fluid.layers.mean(loss) + return loss + + +class ParallelExecutorTestingDuringTraining(unittest.TestCase): + def check_network_convergence(self, build_strategy=None): + main = fluid.Program() + startup = fluid.Program() + with fluid.program_guard(main, startup): + loss = simple_fc_net() + test_program = main.clone(for_test=True) + + opt = fluid.optimizer.SGD(learning_rate=0.001) + opt.minimize(loss) + + batch_size = 32 + image = np.random.normal(size=(batch_size, 784)).astype('float32') + label = np.random.randint(0, 10, (batch_size, 1), dtype="int64") + + place = fluid.CUDAPlace(0) + exe = fluid.Executor(place) + exe.run(startup) + feed_dict = {'image': image, 'label': label} + + train_exe = fluid.ParallelExecutor( + use_cuda=True, + loss_name=loss.name, + main_program=main, + build_strategy=build_strategy) + + test_exe = fluid.ParallelExecutor( + use_cuda=True, + main_program=test_program, + share_vars_from=train_exe, + build_strategy=build_strategy) + + for i in xrange(5): + test_loss, = test_exe.run([loss.name], feed=feed_dict) + test_loss = np.array(test_loss) + + train_loss, = train_exe.run([loss.name], feed=feed_dict) + train_loss = np.array(train_loss) + self.assertTrue( + np.allclose( + train_loss, test_loss, atol=1e-8), + "Train loss: " + str(train_loss) + "\n Test loss:" + + str(test_loss)) + + def test_parallel_testing(self): + build_strategy = fluid.BuildStrategy() + build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.AllReduce + self.check_network_convergence(build_strategy) + + def test_parallel_testing_with_new_strategy(self): + build_strategy = fluid.BuildStrategy() + build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce + self.check_network_convergence(build_strategy) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor_transformer.py b/python/paddle/fluid/tests/unittests/test_parallel_executor_transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..c81df66d987f3d3856af0e19fc935df7de2edacc --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_parallel_executor_transformer.py @@ -0,0 +1,174 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import paddle.fluid as fluid +import transformer_model +import numpy as np +from parallel_executor_test_base import TestParallelExecutorBase +import unittest +import paddle +import paddle.dataset.wmt16 as wmt16 + +WMT16_RECORDIO_FILE = "./wmt16_test_pe.recordio" + + +class ModelHyperParams(object): + # Dictionary size for source and target language. This model directly uses + # paddle.dataset.wmt16 in which , and token has + # alreay been added, but the token is not added. Transformer requires + # sequences in a mini-batch are padded to have the same length. A token is + # added into the original dictionary in paddle.dateset.wmt16. + + # size of source word dictionary. + src_vocab_size = 10000 + # index for token in source language. + src_pad_idx = src_vocab_size + + # size of target word dictionay + trg_vocab_size = 10000 + # index for token in target language. + trg_pad_idx = trg_vocab_size + + # position value corresponding to the token. + pos_pad_idx = 0 + + # max length of sequences. It should plus 1 to include position + # padding token for position encoding. + max_length = 50 + + # the dimension for word embeddings, which is also the last dimension of + # the input and output of multi-head attention, position-wise feed-forward + # networks, encoder and decoder. + + d_model = 512 + # size of the hidden layer in position-wise feed-forward networks. + d_inner_hid = 1024 + # the dimension that keys are projected to for dot-product attention. + d_key = 64 + # the dimension that values are projected to for dot-product attention. + d_value = 64 + # number of head used in multi-head attention. + n_head = 8 + # number of sub-layers to be stacked in the encoder and decoder. + n_layer = 6 + # dropout rate used by all dropout layers. + dropout = 0.1 + + +def prepare_batch_input(insts, src_pad_idx, trg_pad_idx, n_head): + """ + Pad the instances to the max sequence length in batch, and generate the + corresponding position data and attention bias. Then, convert the numpy + data to tensors and return a dict mapping names to tensors. + """ + + def __pad_batch_data(insts, + pad_idx, + is_target=False, + return_pos=True, + return_attn_bias=True, + return_max_len=True): + """ + Pad the instances to the max sequence length in batch, and generate the + corresponding position data and attention bias. + """ + return_list = [] + max_len = max(len(inst) for inst in insts) + inst_data = np.array( + [inst + [pad_idx] * (max_len - len(inst)) for inst in insts]) + return_list += [inst_data.astype("int64").reshape([-1, 1])] + if return_pos: + inst_pos = np.array([[ + pos_i + 1 if w_i != pad_idx else 0 + for pos_i, w_i in enumerate(inst) + ] for inst in inst_data]) + + return_list += [inst_pos.astype("int64").reshape([-1, 1])] + if return_attn_bias: + if is_target: + # This is used to avoid attention on paddings and subsequent + # words. + slf_attn_bias_data = np.ones((inst_data.shape[0], max_len, + max_len)) + slf_attn_bias_data = np.triu(slf_attn_bias_data, 1).reshape( + [-1, 1, max_len, max_len]) + slf_attn_bias_data = np.tile(slf_attn_bias_data, + [1, n_head, 1, 1]) * [-1e9] + else: + # This is used to avoid attention on paddings. + slf_attn_bias_data = np.array([[0] * len(inst) + [-1e9] * + (max_len - len(inst)) + for inst in insts]) + slf_attn_bias_data = np.tile( + slf_attn_bias_data.reshape([-1, 1, 1, max_len]), + [1, n_head, max_len, 1]) + return_list += [slf_attn_bias_data.astype("float32")] + if return_max_len: + return_list += [max_len] + return return_list if len(return_list) > 1 else return_list[0] + + src_word, src_pos, src_slf_attn_bias, src_max_len = __pad_batch_data( + [inst[0] for inst in insts], src_pad_idx, is_target=False) + trg_word, trg_pos, trg_slf_attn_bias, trg_max_len = __pad_batch_data( + [inst[1] for inst in insts], trg_pad_idx, is_target=True) + trg_src_attn_bias = np.tile(src_slf_attn_bias[:, :, ::src_max_len, :], + [1, 1, trg_max_len, 1]).astype("float32") + lbl_word = __pad_batch_data([inst[2] for inst in insts], trg_pad_idx, False, + False, False, False) + lbl_weight = (lbl_word != trg_pad_idx).astype("float32").reshape([-1, 1]) + + return [ + src_word, src_pos, trg_word, trg_pos, src_slf_attn_bias, + trg_slf_attn_bias, trg_src_attn_bias, lbl_word, lbl_weight + ] + + +def transformer(use_feed): + assert not use_feed, "transfomer doesn't support feed yet" + return transformer_model.transformer( + ModelHyperParams.src_vocab_size + 1, + ModelHyperParams.trg_vocab_size + 1, ModelHyperParams.max_length + 1, + ModelHyperParams.n_layer, ModelHyperParams.n_head, + ModelHyperParams.d_key, ModelHyperParams.d_value, + ModelHyperParams.d_model, ModelHyperParams.d_inner_hid, + ModelHyperParams.dropout, ModelHyperParams.src_pad_idx, + ModelHyperParams.trg_pad_idx, ModelHyperParams.pos_pad_idx) + + +class TestTransformer(TestParallelExecutorBase): + @classmethod + def setUpClass(cls): + reader = paddle.batch( + wmt16.train(ModelHyperParams.src_vocab_size, + ModelHyperParams.trg_vocab_size), + batch_size=transformer_model.batch_size) + + with fluid.recordio_writer.create_recordio_writer( + WMT16_RECORDIO_FILE) as writer: + for batch in reader(): + for tensor in prepare_batch_input( + batch, ModelHyperParams.src_pad_idx, + ModelHyperParams.trg_pad_idx, ModelHyperParams.n_head): + t = fluid.LoDTensor() + t.set(tensor, fluid.CPUPlace()) + writer.append_tensor(t) + writer.complete_append_tensor() + + @unittest.skip("transformer is buggy in multi gpu") + def test_main(self): + self.check_network_convergence(transformer) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_polygon_box_transform.py b/python/paddle/fluid/tests/unittests/test_polygon_box_transform.py new file mode 100644 index 0000000000000000000000000000000000000000..2105d320665367e3ec1bfd7b3a353a144c91244f --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_polygon_box_transform.py @@ -0,0 +1,68 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest +import numpy as np +from op_test import OpTest + + +def PolygonBoxRestore(input): + shape = input.shape + batch_size = shape[0] + geo_channels = shape[1] + h = shape[2] + w = shape[3] + h_indexes = np.array(range(h) * w).reshape( + [w, h]).transpose()[np.newaxis, :] # [1, h, w] + w_indexes = np.array(range(w) * h).reshape( + [h, w])[np.newaxis, :] # [1, h, w] + indexes = np.concatenate( + (w_indexes, h_indexes))[np.newaxis, :] # [1, 2, h, w] + indexes = indexes.repeat( + [geo_channels / 2], + axis=0)[np.newaxis, :] # [1, geo_channels/2, 2, h, w] + indexes = indexes.repeat( + [batch_size], axis=0) # [batch_size, geo_channels/2, 2, h, w] + return indexes.reshape( + input.shape) - input # [batch_size, geo_channels, h, w] + + +class TestPolygonBoxRestoreOp(OpTest): + def config(self): + self.input_shape = (1, 8, 2, 2) + + def setUp(self): + self.config() + self.op_type = "polygon_box_transform" + input = np.random.random(self.input_shape).astype("float32") + self.inputs = {'Input': input} + output = PolygonBoxRestore(input) + self.outputs = {'Output': output} + + def test_check_output(self): + self.check_output() + + +class TestCase1(TestPolygonBoxRestoreOp): + def config(self): + self.input_shape = (2, 10, 3, 2) + + +class TestCase2(TestPolygonBoxRestoreOp): + def config(self): + self.input_shape = (3, 12, 4, 5) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_pool3d_op.py b/python/paddle/fluid/tests/unittests/test_pool3d_op.py index aaa94842513691c836e04353aa4bc5ce5e66c5c3..142165f29beeaedfaa660f04424147e06710d192 100644 --- a/python/paddle/fluid/tests/unittests/test_pool3d_op.py +++ b/python/paddle/fluid/tests/unittests/test_pool3d_op.py @@ -90,20 +90,22 @@ def avg_pool3D_forward_naive(x, class TestPool3d_Op(OpTest): def setUp(self): + self.op_type = "pool3d" self.use_cudnn = False + self.dtype = np.float32 self.init_test_case() self.init_global_pool() - self.init_op_type() + self.init_kernel_type() self.init_pool_type() self.init_ceil_mode() if self.global_pool: self.paddings = [0 for _ in range(len(self.paddings))] - input = np.random.random(self.shape).astype("float32") + input = np.random.random(self.shape).astype(self.dtype) output = self.pool3D_forward_naive(input, self.ksize, self.strides, self.paddings, self.global_pool, - self.ceil_mode).astype("float32") - self.inputs = {'X': input} + self.ceil_mode).astype(self.dtype) + self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(input)} self.attrs = { 'strides': self.strides, @@ -116,7 +118,7 @@ class TestPool3d_Op(OpTest): 'data_format': 'AnyLayout' # TODO(dzhwinter) : should be fix latter } - self.outputs = {'Out': output.astype('float32')} + self.outputs = {'Out': output} def testcudnn(self): return core.is_compiled_with_cuda() and self.use_cudnn @@ -129,6 +131,8 @@ class TestPool3d_Op(OpTest): self.check_output() def test_check_grad(self): + if self.dtype == np.float16: + return if self.testcudnn() and self.pool_type != "max": place = core.CUDAPlace(0) self.check_grad_with_place( @@ -142,8 +146,8 @@ class TestPool3d_Op(OpTest): self.strides = [1, 1, 1] self.paddings = [0, 0, 0] - def init_op_type(self): - self.op_type = "pool3d" + def init_kernel_type(self): + pass def init_pool_type(self): self.pool_type = "avg" @@ -158,15 +162,11 @@ class TestPool3d_Op(OpTest): class TestCase1(TestPool3d_Op): def init_test_case(self): - self.op_type = "pool3d" self.shape = [2, 3, 7, 7, 7] self.ksize = [3, 3, 3] self.strides = [1, 1, 1] self.paddings = [0, 0, 0] - def init_op_type(self): - self.op_type = "pool3d" - def init_pool_type(self): self.pool_type = "avg" self.pool3D_forward_naive = avg_pool3D_forward_naive @@ -182,9 +182,6 @@ class TestCase2(TestPool3d_Op): self.strides = [1, 1, 1] self.paddings = [1, 1, 1] - def init_op_type(self): - self.op_type = "pool3d" - def init_pool_type(self): self.pool_type = "avg" self.pool3D_forward_naive = avg_pool3D_forward_naive @@ -194,27 +191,18 @@ class TestCase2(TestPool3d_Op): class TestCase3(TestPool3d_Op): - def init_op_type(self): - self.op_type = "pool3d" - def init_pool_type(self): self.pool_type = "max" self.pool3D_forward_naive = max_pool3D_forward_naive class TestCase4(TestCase1): - def init_op_type(self): - self.op_type = "pool3d" - def init_pool_type(self): self.pool_type = "max" self.pool3D_forward_naive = max_pool3D_forward_naive class TestCase5(TestCase2): - def init_op_type(self): - self.op_type = "pool3d" - def init_pool_type(self): self.pool_type = "max" self.pool3D_forward_naive = max_pool3D_forward_naive @@ -222,39 +210,105 @@ class TestCase5(TestCase2): #--------------------test pool3d-------------------- class TestCUDNNCase1(TestPool3d_Op): - def init_op_type(self): + def init_kernel_type(self): self.use_cudnn = True - self.op_type = "pool3d" + + +class TestFP16CUDNNCase1(TestPool3d_Op): + def init_kernel_type(self): + self.use_cudnn = True + self.dtype = np.float16 + + def test_check_output(self): + if core.is_compiled_with_cuda(): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_output_with_place(place, atol=1e-3) class TestCUDNNCase2(TestCase1): - def init_op_type(self): + def init_kernel_type(self): self.use_cudnn = True - self.op_type = "pool3d" + + +class TestFP16CUDNNCase2(TestCase1): + def init_kernel_type(self): + self.use_cudnn = True + self.dtype = np.float16 + + def test_check_output(self): + if core.is_compiled_with_cuda(): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_output_with_place(place, atol=1e-3) class TestCUDNNCase3(TestCase2): - def init_op_type(self): + def init_kernel_type(self): self.use_cudnn = True - self.op_type = "pool3d" + + +class TestFP16CUDNNCase3(TestCase2): + def init_kernel_type(self): + self.use_cudnn = True + self.dtype = np.float16 + + def test_check_output(self): + if core.is_compiled_with_cuda(): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_output_with_place(place, atol=1e-3) class TestCUDNNCase4(TestCase3): - def init_op_type(self): + def init_kernel_type(self): self.use_cudnn = True - self.op_type = "pool3d" + + +class TestFP16CUDNNCase4(TestCase3): + def init_kernel_type(self): + self.use_cudnn = True + self.dtype = np.float16 + + def test_check_output(self): + if core.is_compiled_with_cuda(): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_output_with_place(place, atol=1e-3) class TestCUDNNCase5(TestCase4): - def init_op_type(self): + def init_kernel_type(self): self.use_cudnn = True - self.op_type = "pool3d" + + +class TestFP16CUDNNCase5(TestCase4): + def init_kernel_type(self): + self.use_cudnn = True + self.dtype = np.float16 + + def test_check_output(self): + if core.is_compiled_with_cuda(): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_output_with_place(place, atol=1e-3) class TestCUDNNCase6(TestCase5): - def init_op_type(self): + def init_kernel_type(self): self.use_cudnn = True - self.op_type = "pool3d" + + +class TestFP16CUDNNCase6(TestCase5): + def init_kernel_type(self): + self.use_cudnn = True + self.dtype = np.float16 + + def test_check_output(self): + if core.is_compiled_with_cuda(): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_output_with_place(place, atol=1e-3) class TestCeilModeCase1(TestCUDNNCase1): diff --git a/python/paddle/fluid/tests/unittests/test_preprocessor.py b/python/paddle/fluid/tests/unittests/test_preprocessor.py new file mode 100644 index 0000000000000000000000000000000000000000..cbf1a7e0c50a87cd43507ffdb94109873cf4e5d9 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_preprocessor.py @@ -0,0 +1,93 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest +import numpy as np + +import paddle.fluid as fluid +import paddle.v2 as paddle +import paddle.v2.dataset.mnist as mnist + + +class TestPreprocessor(unittest.TestCase): + def setUp(self): + with fluid.program_guard(fluid.Program(), fluid.Program()): + reader = paddle.batch(mnist.train(), batch_size=32) + feeder = fluid.DataFeeder( + feed_list=[ # order is image and label + fluid.layers.data( + name='image', shape=[784]), + fluid.layers.data( + name='label', shape=[1], dtype='int64'), + ], + place=fluid.CPUPlace()) + self.num_batches = fluid.recordio_writer.convert_reader_to_recordio_file( + './mnist_for_preprocessor_test.recordio', reader, feeder) + + def test_main(self): + N = 10 + + img_expected_res = [] + lbl_expected_res = [] + with fluid.program_guard(fluid.Program(), fluid.Program()): + data_file = fluid.layers.io.open_recordio_file( + './mnist_for_preprocessor_test.recordio', + shapes=[[-1, 784], [-1, 1]], + lod_levels=[0, 0], + dtypes=['float32', 'int64']) + img, lbl = fluid.layers.io.read_file(data_file) + + if fluid.core.is_compiled_with_cuda(): + place = fluid.CUDAPlace(0) + else: + place = fluid.CPUPlace() + exe = fluid.Executor(place) + exe.run(fluid.default_startup_program()) + for _ in range(N): + img_v, lbl_v = exe.run(fetch_list=[img, lbl]) + img_expected_res.append(img_v / 2) + lbl_expected_res.append(lbl_v + 1) + + img_actual_res = [] + lbl_actual_res = [] + with fluid.program_guard(fluid.Program(), fluid.Program()): + data_file = fluid.layers.io.open_recordio_file( + './mnist_for_preprocessor_test.recordio', + shapes=[[-1, 784], [-1, 1]], + lod_levels=[0, 0], + dtypes=['float32', 'int64']) + preprocessor = fluid.layers.io.Preprocessor(reader=data_file) + with preprocessor.block(): + img, lbl = preprocessor.inputs() + img_out = img / 2 + lbl_out = lbl + 1 + preprocessor.outputs(img_out, lbl_out) + + data_file = fluid.layers.io.double_buffer(preprocessor()) + img, lbl = fluid.layers.io.read_file(data_file) + + if fluid.core.is_compiled_with_cuda(): + place = fluid.CUDAPlace(0) + else: + place = fluid.CPUPlace() + exe = fluid.Executor(place) + exe.run(fluid.default_startup_program()) + for _ in range(N): + img_v, lbl_v = exe.run(fetch_list=[img, lbl]) + img_actual_res.append(img_v) + lbl_actual_res.append(lbl_v) + + for idx in range(N): + np.allclose(img_expected_res[idx], img_actual_res[idx]) + np.allclose(lbl_expected_res[idx], lbl_actual_res[idx]) diff --git a/python/paddle/fluid/tests/unittests/test_random_crop_op.py b/python/paddle/fluid/tests/unittests/test_random_crop_op.py new file mode 100644 index 0000000000000000000000000000000000000000..1c708d0386da4028f1f3d177d0a3fd494c077c6e --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_random_crop_op.py @@ -0,0 +1,46 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest +import numpy as np +import paddle.fluid.core as core +from op_test import OpTest + + +class TestRandomCropOp(OpTest): + def setUp(self): + to_crop = np.array([[[1, 2, 3, 4], [5, 6, 7, 8], [9, 10, 11, 12]]] * + 5).astype("float32") + self.possible_res = [ + np.array([[1, 2, 3], [5, 6, 7]]), np.array([[2, 3, 4], [6, 7, 8]]), + np.array([[5, 6, 7], [9, 10, 11]]), + np.array([[6, 7, 8], [10, 11, 12]]) + ] + self.op_type = "random_crop" + self.inputs = {'X': to_crop, 'Seed': np.array([10])} + self.outputs = {'Out': np.array([]), 'SeedOut': np.array([])} + self.attrs = {'shape': [2, 3]} + + def test_check_output(self): + self.check_output_customized(self.verify_output) + + def verify_output(self, outs): + out = np.array(outs[1]) + for ins in out[:]: + is_equal = [(ins == res).all() for res in self.possible_res] + self.assertIn(True, is_equal) + + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_reduce_op.py b/python/paddle/fluid/tests/unittests/test_reduce_op.py index 9b0cc3534dc551e7fdf7ef8111cad1c172f8bfa4..865c2b7df085aa6a6cb0d6eb461c342ce08695cd 100644 --- a/python/paddle/fluid/tests/unittests/test_reduce_op.py +++ b/python/paddle/fluid/tests/unittests/test_reduce_op.py @@ -34,8 +34,10 @@ class TestMeanOp(OpTest): def setUp(self): self.op_type = "reduce_mean" self.inputs = {'X': np.random.random((5, 6, 2, 10)).astype("float64")} - self.attrs = {'dim': 1} - self.outputs = {'Out': self.inputs['X'].mean(axis=self.attrs['dim'])} + self.attrs = {'dim': [1]} + self.outputs = { + 'Out': self.inputs['X'].mean(axis=tuple(self.attrs['dim'])) + } def test_check_output(self): self.check_output() @@ -50,8 +52,10 @@ class TestMaxOp(OpTest): def setUp(self): self.op_type = "reduce_max" self.inputs = {'X': np.random.random((5, 6, 10)).astype("float64")} - self.attrs = {'dim': -1} - self.outputs = {'Out': self.inputs['X'].max(axis=self.attrs['dim'])} + self.attrs = {'dim': [-1]} + self.outputs = { + 'Out': self.inputs['X'].max(axis=tuple(self.attrs['dim'])) + } def test_check_output(self): self.check_output() @@ -63,8 +67,10 @@ class TestMinOp(OpTest): def setUp(self): self.op_type = "reduce_min" self.inputs = {'X': np.random.random((5, 6, 10)).astype("float64")} - self.attrs = {'dim': 2} - self.outputs = {'Out': self.inputs['X'].min(axis=self.attrs['dim'])} + self.attrs = {'dim': [2]} + self.outputs = { + 'Out': self.inputs['X'].min(axis=tuple(self.attrs['dim'])) + } def test_check_output(self): self.check_output() @@ -87,9 +93,10 @@ class TestKeepDimReduce(OpTest): def setUp(self): self.op_type = "reduce_sum" self.inputs = {'X': np.random.random((5, 6, 10)).astype("float64")} - self.attrs = {'dim': -2, 'keep_dim': True} + self.attrs = {'dim': [-2], 'keep_dim': True} self.outputs = { - 'Out': self.inputs['X'].sum(axis=self.attrs['dim'], keepdims=True) + 'Out': + self.inputs['X'].sum(axis=tuple(self.attrs['dim']), keepdims=True) } def test_check_output(self): @@ -126,5 +133,67 @@ class TestReduceAll(OpTest): self.check_grad(['X'], 'Out') +## reduction in multi dims +class TestReduceMeanOpMultiAxises(OpTest): + def setUp(self): + self.op_type = "reduce_mean" + self.inputs = {'X': np.random.random((5, 6, 2, 10)).astype("float64")} + self.attrs = {'dim': [1, 2]} + self.outputs = {'Out': self.inputs['X'].mean(axis=(1, 2))} + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad(['X'], 'Out') + + +class TestReduceMaxOpMultiAxises(OpTest): + """Remove Max with subgradient from gradient check to confirm the success of CI.""" + + def setUp(self): + self.op_type = "reduce_max" + self.inputs = {'X': np.random.random((5, 6, 10)).astype("float64")} + self.attrs = {'dim': [-2, -1]} + self.outputs = { + 'Out': self.inputs['X'].max(axis=tuple(self.attrs['dim'])) + } + + def test_check_output(self): + self.check_output() + + +class TestReduceMinOpMultiAxises(OpTest): + """Remove Min with subgradient from gradient check to confirm the success of CI.""" + + def setUp(self): + self.op_type = "reduce_min" + self.inputs = {'X': np.random.random((5, 6, 10)).astype("float64")} + self.attrs = {'dim': [1, 2]} + self.outputs = { + 'Out': self.inputs['X'].min(axis=tuple(self.attrs['dim'])) + } + + def test_check_output(self): + self.check_output() + + +class TestKeepDimReduceSumMultiAxises(OpTest): + def setUp(self): + self.op_type = "reduce_sum" + self.inputs = {'X': np.random.random((5, 6, 10)).astype("float64")} + self.attrs = {'dim': [-2, -1], 'keep_dim': True} + self.outputs = { + 'Out': + self.inputs['X'].sum(axis=tuple(self.attrs['dim']), keepdims=True) + } + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad(['X'], 'Out') + + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_roi_pool_op.py b/python/paddle/fluid/tests/unittests/test_roi_pool_op.py index e556d51b021217063e23190e40bc0e8f9fdc816c..3d754aff3a73e7168e2123483b26e5e3a3585a4e 100644 --- a/python/paddle/fluid/tests/unittests/test_roi_pool_op.py +++ b/python/paddle/fluid/tests/unittests/test_roi_pool_op.py @@ -25,7 +25,7 @@ class TestROIPoolOp(OpTest): self.make_rois() self.calc_roi_pool() - self.inputs = {'X': self.x, 'ROIs': self.rois} + self.inputs = {'X': self.x, 'ROIs': (self.rois[:, 1:5], self.rois_lod)} self.attrs = { 'spatial_scale': self.spatial_scale, @@ -36,7 +36,7 @@ class TestROIPoolOp(OpTest): self.outputs = {'Out': self.outs, 'Argmax': self.argmaxes} def init_test_case(self): - self.batch_size = 5 + self.batch_size = 3 self.channels = 3 self.height = 6 self.width = 4 @@ -47,7 +47,6 @@ class TestROIPoolOp(OpTest): self.spatial_scale = 1.0 / 4.0 self.pooled_height = 2 self.pooled_width = 2 - self.rois_num = 2 self.x = np.random.random(self.x_dim).astype('float32') @@ -106,20 +105,24 @@ class TestROIPoolOp(OpTest): def make_rois(self): rois = [] - batch_ids = np.random.randint(0, self.batch_size, size=self.rois_num) - for i in range(self.rois_num): - x1 = np.random.random_integers( - 0, self.width / self.spatial_scale - self.pooled_width) - y1 = np.random.random_integers( - 0, self.height / self.spatial_scale - self.pooled_height) - - x2 = np.random.random_integers(x1 + self.pooled_width, - self.width / self.spatial_scale) - y2 = np.random.random_integers(y1 + self.pooled_height, - self.height / self.spatial_scale) - - roi = [batch_ids[i], x1, y1, x2, y2] - rois.append(roi) + self.rois_lod = [[]] + for bno in range(self.batch_size): + self.rois_lod[0].append(len(rois)) + for i in range(bno + 1): + x1 = np.random.random_integers( + 0, self.width / self.spatial_scale - self.pooled_width) + y1 = np.random.random_integers( + 0, self.height / self.spatial_scale - self.pooled_height) + + x2 = np.random.random_integers(x1 + self.pooled_width, + self.width / self.spatial_scale) + y2 = np.random.random_integers(y1 + self.pooled_height, + self.height / self.spatial_scale) + + roi = [bno, x1, y1, x2, y2] + rois.append(roi) + self.rois_lod[0].append(len(rois)) + self.rois_num = len(rois) self.rois = np.array(rois).astype("int64") def setUp(self): diff --git a/python/paddle/fluid/tests/unittests/test_split_var.py b/python/paddle/fluid/tests/unittests/test_split_var.py index 104ceb4fe7beb70b9016f57cef0ef895a3eb8ba6..157def9b56e44092a86023035d1ab444c938aa07 100644 --- a/python/paddle/fluid/tests/unittests/test_split_var.py +++ b/python/paddle/fluid/tests/unittests/test_split_var.py @@ -14,22 +14,14 @@ import math import unittest -from paddle.fluid.distribute_transpiler import split_dense_variable +from paddle.fluid.transpiler.distribute_transpiler import split_variable import paddle.fluid as fluid import paddle.fluid.core as core import random class TestSplitVar(unittest.TestCase): - def test_check_output(self): - # split below shapes to 10 servers - shapes = [[3, 5], [1024], [28, 784], [8, 1020], [800, 10]] - expected_sizes = [ - [15], [1024], - [2352, 2352, 2352, 2352, 2352, 2352, 2352, 2352, 2352, 784], - [2040, 2040, 2040, 2040], - [1150, 1150, 1150, 1150, 1150, 1150, 1100] - ] + def check_split_output(self, shapes, expected_sizes, min_size): var_list = [] program = fluid.Program() for shape in shapes: @@ -39,7 +31,7 @@ class TestSplitVar(unittest.TestCase): # dtype=core.VarDesc.VarType.LOD_TENSOR, shape=shape) var_list.append(var) - blocks = split_dense_variable(var_list, 10) + blocks = split_variable(var_list, 10, min_size) all_sizes = [] for s in expected_sizes: for s2 in s: @@ -48,6 +40,25 @@ class TestSplitVar(unittest.TestCase): varname, block_id, size = block_str.split(":") self.assertEqual(int(size), all_sizes[i]) + def test_1k(self): + shapes = [[3, 5], [1024], [28, 784], [8, 1020], [800, 10]] + expected_sizes = [ + [15], [1024], + [2352, 2352, 2352, 2352, 2352, 2352, 2352, 2352, 2352, 784], + [2040, 2040, 2040, 2040], + [1150, 1150, 1150, 1150, 1150, 1150, 1100] + ] + + self.check_split_output(shapes, expected_sizes, 1024) + + def test_check_output_8k(self): + shapes = [[3, 5], [1024], [28, 784], [8, 1020], [800, 10], + [6, 33, 33, 33]] + expected_sizes = [[15], [1024], [10976, 10976], [8160], [8000], + [35937, 35937, 35937, 35937, 35937, 35937]] + + self.check_split_output(shapes, expected_sizes, 8192) + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/trainer.py b/python/paddle/fluid/trainer.py new file mode 100644 index 0000000000000000000000000000000000000000..7da123dd92ed9d111d68cd70efb8ce1493452609 --- /dev/null +++ b/python/paddle/fluid/trainer.py @@ -0,0 +1,372 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import contextlib +import os + +import core + +import data_feeder +import executor +import framework +import io +# optimizer is same as the parameter of Trainer.__init__. Rename it to opt_module +import optimizer as opt_module +import parallel_executor +from transpiler import distribute_transpiler + +__all__ = [ + 'Trainer', + 'BeginEpochEvent', + 'EndEpochEvent', + 'BeginStepEvent', + 'EndStepEvent', +] + + +class BeginEpochEvent(object): + def __init__(self, epoch_id): + self.epoch = epoch_id + + +class EndEpochEvent(object): + def __init__(self, epoch_id): + self.epoch = epoch_id + + +class BeginStepEvent(object): + def __init__(self, epoch_id, step_id): + self.epoch = epoch_id + self.step = step_id + self.fetch_metrics = True + + +class EndStepEvent(object): + def __init__(self, epoch_id, step_id, metrics): + self.epoch = epoch_id + self.step = step_id + self.metrics = metrics + + +def check_and_get_place(place): + """ + Check the type of place or get the default place + Args: + place(None|core.CUDAPlace|core.CPUPlace): the place that trainer will be executed on. + + Raises: + TypeError if the type mismatched. + + Returns: + the original place if it is not None. + if fluid is compiled with CUDA, returns CUDAPlace(0) by default. + Otherwise returns CPUPlace by default. + """ + if place is None: + if core.is_compiled_with_cuda(): + return core.CUDAPlace(0) + else: + return core.CPUPlace() + else: + if not isinstance(place, core.CUDAPlace) and not isinstance( + place, core.CPUPlace): + raise TypeError("Place should be either CUDAPlace or CPUPlace") + return place + + +class Trainer(object): + """ + + Args: + train_func(callable): A function which will return loss. The loss must be a scalar. + optimizer(optimizer.Optimizer): The optimizer should be an instance of Optimizer + place: The device place of this trainer. + """ + + def __init__(self, + train_func, + optimizer, + param_path=None, + place=None, + parallel=False): + self.__stop = False + self.parallel = parallel + # 1. we need to generate a framework.Program by calling + # program_func. Reference: fluid.program_guard in + # test_word2vec.py + if not isinstance(optimizer, opt_module.Optimizer): + raise TypeError("The optimizer should be an instance of Optimizer") + + self.scope = core.Scope() + + self.startup_program = framework.Program() + self.train_program = framework.Program() + + with framework.program_guard(self.train_program, self.startup_program): + program_func_outs = train_func() + self.train_func_outputs = program_func_outs if isinstance( + program_func_outs, list) else [program_func_outs] + self.test_program = self.train_program.clone() + if not isinstance(optimizer, opt_module.Optimizer): + raise TypeError( + "The optimizer should be an instance of Optimizer") + # The fisrt element of program_func_outs is loss. + loss = self.train_func_outputs[0] + optimize_ops, params_grads = optimizer.minimize(loss) + + self.place = check_and_get_place(place) + + self._dist_transpile_if_necessary(optimize_ops, params_grads) + + # 2. move the default_main_program to self.program and run the + # default_startup program on an empty core.Scope() + # Run startup program + with self._prog_and_scope_guard(): + exe = executor.Executor(place) + exe.run(self.startup_program) + + if param_path: + # load params from param_path into scope + io.load_persistables(exe, dirname=param_path) + + def _transpile_nccl2_dist(self): + # PADDLE_TRAINER_IPS + if "PADDLE_TRAINER_IPS" not in os.environ: + self.nccl_id_var = None + else: + self.trainer_id = int(os.getenv("PADDLE_TRAINER_ID")) + port = os.getenv("PADDLE_PSERVER_PORT") + worker_ips = os.getenv("PADDLE_TRAINER_IPS") + worker_endpoints = [] + for ip in worker_ips.split(","): + worker_endpoints.append(':'.join([ip, port])) + self.num_trainers = len(worker_endpoints) + current_endpoint = os.getenv("POD_IP") + ":" + port + worker_endpoints.remove(current_endpoint) + # TODO(wuyi): use self.nccl_id_var, self.num_trainers and self.trainer_id + # in ParallelExecutor to start + # distributed training using NCCL2 + self.nccl_id_var = self.startup_program.global_block().create_var( + name="NCCLID", persistable=True, type=core.VarDesc.VarType.RAW) + self.startup_program.global_block().append_op( + type="gen_nccl_id", + inputs={}, + outputs={"NCCLID": self.nccl_id_var}, + attrs={ + "endpoint": current_endpoint, + "endpoint_list": worker_endpoints, + "trainer_id": self.trainer_id + }) + + def _dist_transpile_if_necessary(self, optimize_ops, params_grads): + self._transpile_nccl2_dist() + if self.nccl_id_var != None: + return + + if "PADDLE_TRAINING_ROLE" not in os.environ: + return + + # the port of all pservers, needed by both trainer and pserver + port = os.getenv("PADDLE_PSERVER_PORT", "6174") + # comma separated ips of all pservers, needed by trainer and + # pserver + pserver_ips = os.getenv("PADDLE_PSERVER_IPS", "") + eplist = [] + for ip in pserver_ips.split(","): + eplist.append(':'.join([ip, port])) + pserver_endpoints = ",".join(eplist) + # total number of workers/trainers in the job, needed by + # trainer and pserver + trainers = int(os.getenv("PADDLE_TRAINERS")) + # the IP of the local machine, needed by pserver only + current_endpoint = os.getenv("PADDLE_CURRENT_IP", "") + ":" + port + # the unique trainer id, starting from 0, needed by trainer + # only + trainer_id = int(os.getenv("PADDLE_TRAINER_ID", "0")) + # the role, should be either PSERVER or TRAINER + training_role = os.getenv("PADDLE_TRAINING_ROLE") + with self._prog_and_scope_guard(): + t = distribute_transpiler.DistributeTranspiler() + t.transpile( + trainer_id, pservers=pserver_endpoints, trainers=trainers) + if training_role == "PSERVER": + self.train_program = t.get_pserver_program(current_endpoint) + self.startup_program = t.get_startup_program(current_endpoint, + self.train_program) + elif training_role == "TRAINER": + self.train_program = t.get_trainer_program() + else: + raise ValueError( + 'TRAINING_ROLE environment variable must be either TRAINER or PSERVER' + ) + + def stop(self): + """ + stop training + """ + self.__stop = True + + def train(self, num_epochs, event_handler, reader=None, feed_order=None): + """ + Train the model. + + Args: + num_epochs: The number of epoch. An epoch will process all data in reader + event_handler: The event handler. A function with type (ev:Event)->void + reader: + feed_order: Feeding order of reader. None will following the defining + order in program + + Returns: + + """ + training_role = os.getenv("PADDLE_TRAINING_ROLE", "") + if training_role == "PSERVER": + with self._prog_and_scope_guard(): + exe = executor.Executor(self.place) + exe.run() + return + if self.parallel: + self._train_by_parallel_executor(num_epochs, event_handler, reader, + feed_order) + else: + self._train_by_executor(num_epochs, event_handler, reader, + feed_order) + + def test(self, reader, feed_order): + """ + Test the model on given test data + + Args: + reader: The reader that yields test data. + feed_order: Feeding order of reader. None will following the defining + order in program + """ + + return self._test_by_executor(reader, feed_order, + self.train_func_outputs) + + def save_params(self, param_path): + # reference: save_persistables in io.py + with self._prog_and_scope_guard(): + exe = executor.Executor(self.place) + io.save_persistables(exe, dirname=param_path) + + @contextlib.contextmanager + def _prog_and_scope_guard(self): + with framework.program_guard( + main_program=self.train_program, + startup_program=self.startup_program): + with executor.scope_guard(self.scope): + yield + + def _train_by_executor(self, num_epochs, event_handler, reader, feed_order): + """ + Train by Executor and single device. + + Args: + num_epochs: + event_handler: + reader: + feed_order: + + Returns: + + """ + with self._prog_and_scope_guard(): + feed_var_list = build_feed_var_list(self.train_program, feed_order) + feeder = data_feeder.DataFeeder( + feed_list=feed_var_list, place=self.place) + exe = executor.Executor(self.place) + reader = feeder.decorate_reader(reader, multi_devices=False) + self._train_by_any_executor(event_handler, exe, num_epochs, reader) + + def _train_by_any_executor(self, event_handler, exe, num_epochs, reader): + for epoch_id in range(num_epochs): + event_handler(BeginEpochEvent(epoch_id)) + for step_id, data in enumerate(reader()): + if self.__stop: + return + begin_event = BeginStepEvent(epoch_id, step_id) + event_handler(begin_event) + if begin_event.fetch_metrics: + metrics = exe.run(feed=data, + fetch_list=[ + var.name + for var in self.train_func_outputs + ]) + else: + metrics = exe.run(feed=data, fetch_list=[]) + event_handler(EndStepEvent(epoch_id, step_id, metrics)) + event_handler(EndEpochEvent(epoch_id)) + + def _test_by_executor(self, reader, feed_order, fetch_list): + with executor.scope_guard(self.scope): + feed_var_list = build_feed_var_list(self.test_program, feed_order) + feeder = data_feeder.DataFeeder( + feed_list=feed_var_list, place=self.place) + exe = executor.Executor(self.place) + accumulated = len(fetch_list) * [0] + count = 0 + for data in reader(): + outs = exe.run(program=self.test_program, + feed=feeder.feed(data), + fetch_list=fetch_list) + accumulated = [x[0] + x[1][0] for x in zip(accumulated, outs)] + count += 1 + + return [x / count for x in accumulated] + + def _train_by_parallel_executor(self, num_epochs, event_handler, reader, + feed_order): + with self._prog_and_scope_guard(): + pe = self._get_or_create_parallel_executor() + feed_var_list = build_feed_var_list(self.train_program, feed_order) + feeder = data_feeder.DataFeeder( + feed_list=feed_var_list, place=self.place) + reader = feeder.decorate_reader(reader, multi_devices=True) + self._train_by_any_executor(event_handler, pe, num_epochs, reader) + + def _get_parallel_executor(self): + return getattr(self, 'parallel_executor', None) + + def _get_or_create_parallel_executor(self): + if self._get_parallel_executor() is None: + self.parallel_executor = parallel_executor.ParallelExecutor( + use_cuda=isinstance(self.place, core.CUDAPlace), + loss_name=self.train_func_outputs[0].name) + return self._get_parallel_executor() + + +def build_feed_var_list(program, feed_order): + if not isinstance(program, framework.Program): + raise TypeError("The 'program' should be an object of Program") + + if isinstance(feed_order, list): + feed_var_list = [ + program.global_block().var(var_name) for var_name in feed_order + ] + else: + if not isinstance(feed_order, dict): + raise TypeError( + "The 'feed_order' should be either None, list or dict.") + if not sorted(feed_order.values()) == range(len(feed_order)): + raise ValueError( + "The values of 'feed_order' should be a permutation of [0, len(feed_order))" + ) + sorted_pair_list = sorted(feed_order.items(), key=lambda item: item[1]) + feed_var_list = [ + program.global_block().var(pair[0]) for pair in sorted_pair_list + ] + return feed_var_list diff --git a/python/paddle/fluid/transpiler/__init__.py b/python/paddle/fluid/transpiler/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..045ca537b2e84c02298d6375a7ef5bdbb5517380 --- /dev/null +++ b/python/paddle/fluid/transpiler/__init__.py @@ -0,0 +1,24 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from distribute_transpiler import DistributeTranspiler +from inference_transpiler import InferenceTranspiler +from memory_optimization_transpiler import memory_optimize, release_memory +from distribute_transpiler_simple import SimpleDistributeTranspiler +from ps_dispatcher import HashName, RoundRobin + +__all__ = [ + "DistributeTranspiler", "InferenceTranspiler", "SimpleDistributeTranspiler", + "memory_optimize", "release_memory", "HashName", "RoundRobin" +] diff --git a/python/paddle/fluid/transpiler/details/__init__.py b/python/paddle/fluid/transpiler/details/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..dc597c33849dc06cc975b245099672f64c3539d3 --- /dev/null +++ b/python/paddle/fluid/transpiler/details/__init__.py @@ -0,0 +1,16 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from program_utils import * +from ufind import * diff --git a/python/paddle/fluid/transpiler/details/program_utils.py b/python/paddle/fluid/transpiler/details/program_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..f10b496306a002ee131d01798a0698b807d379ca --- /dev/null +++ b/python/paddle/fluid/transpiler/details/program_utils.py @@ -0,0 +1,37 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +def delete_ops(block, ops): + try: + start = list(block.ops).index(ops[0]) + end = list(block.ops).index(ops[-1]) + [block.remove_op(start) for _ in xrange(end - start + 1)] + except Exception, e: + raise e + block.program.sync_with_cpp() + + +def find_op_by_input_arg(block, arg_name): + for index, op in enumerate(block.ops): + if arg_name in op.input_arg_names: + return index + return -1 + + +def find_op_by_output_arg(block, arg_name): + for index, op in enumerate(block.ops): + if arg_name in op.output_arg_names: + return index + return -1 diff --git a/python/paddle/fluid/transpiler/details/ufind.py b/python/paddle/fluid/transpiler/details/ufind.py new file mode 100644 index 0000000000000000000000000000000000000000..0e30d0e3f9c5712c494daf17b2b4bcec86f69c23 --- /dev/null +++ b/python/paddle/fluid/transpiler/details/ufind.py @@ -0,0 +1,64 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +class UnionFind(object): + """ Union-find data structure. + + Union-find is a data structure that keeps track of a set of elements partitioned + into a number of disjoint (non-overlapping) subsets. + + Reference: + https://en.wikipedia.org/wiki/Disjoint-set_data_structure + + Args: + elements(list): The initialize element list. + """ + + def __init__(self, elementes=None): + self._parents = [] # index -> parent index + self._index = {} # element -> index + self._curr_idx = 0 + if not elementes: + elementes = [] + for ele in elementes: + self._parents.append(self._curr_idx) + self._index.update({ele: self._curr_idx}) + self._curr_idx += 1 + + def find(self, x): + # Find the root index of given element x, + # execute the path compress while findind the root index + if not x in self._index: + return -1 + idx = self._index[x] + while idx != self._parents[idx]: + t = self._parents[idx] + self._parents[idx] = self._parents[t] + idx = t + return idx + + def union(self, x, y): + # Union two given element + x_root = self.find(x) + y_root = self.find(y) + + if x_root == y_root: + return + self._parents[x_root] = y_root + + def is_connected(self, x, y): + # If two given elements have the same root index, + # then they are connected. + return self.find(x) == self.find(y) diff --git a/python/paddle/fluid/distribute_transpiler.py b/python/paddle/fluid/transpiler/distribute_transpiler.py similarity index 58% rename from python/paddle/fluid/distribute_transpiler.py rename to python/paddle/fluid/transpiler/distribute_transpiler.py index d17475cd28b3ae57032d3be811542fc89246e299..819073257c584692b2316f0703255d2ef7777915 100644 --- a/python/paddle/fluid/distribute_transpiler.py +++ b/python/paddle/fluid/transpiler/distribute_transpiler.py @@ -11,19 +11,48 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. +""" +Transpile the program to distributed data-parallelism programs. +The main_program will be transformed to use a remote parameter server +to do parameter optimization. And the optimization graph will be put +into a parameter server program. + +Use different methods to split trainable variables to different +parameter servers. + +Steps to transpile trainer: +1. split variable to multiple blocks, aligned by product(dim[1:]) (width). +2. rename splited grad variables to add trainer_id suffix ".trainer_%d". +3. modify trainer program add split_op to each grad variable. +4. append send_op to send splited variables to server and fetch + params(splited blocks or origin param) from server. +5. append concat_op to merge splited blocks to update local weights. + +Steps to transpile pserver: +1. create new program for parameter server. +2. create params and grad variables that assigned to current server instance. +3. create a sub-block in the server side program +4. append ops that should run on current server instance. +5. add listen_and_serv op +""" from __future__ import print_function import math -import distributed_splitter as splitter -import framework -from framework import Program, default_main_program, Variable, Parameter -from . import core +from ps_dispatcher import RoundRobin, HashName, PSDispatcher +from .. import core, framework +from ..framework import Program, default_main_program, \ + default_startup_program, \ + Variable, Parameter, grad_var_name +from details import * LOOKUP_TABLE_TYPE = "lookup_table" LOOKUP_TABLE_GRAD_TYPE = "lookup_table_grad" -RPC_CLIENT_VAR_NAME = "RPC_CLIENT_VAR" +OP_ROLE_VAR_ATTR_NAME = core.op_proto_and_checker_maker.kOpRoleVarAttrName() +RPC_OP_ROLE_ATTR_NAME = op_role_attr_name = core.op_proto_and_checker_maker.kOpRoleAttrName( +) +RPC_OP_ROLE_ATTR_VALUE = core.op_proto_and_checker_maker.OpRole.RPC class VarBlock: @@ -37,85 +66,37 @@ class VarBlock: return "%s:%d:%d" % (self.varname, self.offset, self.size) -class UnionFind(object): - """ Union-find data structure. - - Union-find is a data structure that keeps track of a set of elements partitioned - into a number of disjoint (non-overlapping) subsets. - - Reference: - https://en.wikipedia.org/wiki/Disjoint-set_data_structure - - Args: - elements(list): The initialize element list. - """ - - def __init__(self, elementes=None): - self._parents = [] # index -> parent index - self._index = {} # element -> index - self._curr_idx = 0 - if not elementes: - elementes = [] - for ele in elementes: - self._parents.append(self._curr_idx) - self._index.update({ele: self._curr_idx}) - self._curr_idx += 1 - - def find(self, x): - # Find the root index of given element x, - # execute the path compress while findind the root index - if not x in self._index: - return -1 - idx = self._index[x] - while idx != self._parents[idx]: - t = self._parents[idx] - self._parents[idx] = self._parents[t] - idx = t - return idx - - def union(self, x, y): - # Union two given element - x_root = self.find(x) - y_root = self.find(y) - - if x_root == y_root: - return - self._parents[x_root] = y_root - - def is_connected(self, x, y): - # If two given elements have the same root index, - # then they are connected. - return self.find(x) == self.find(y) - - def same_or_split_var(p_name, var_name): return p_name == var_name or p_name.startswith(var_name + ".block") -def split_dense_variable(var_list, - pserver_count, - min_block_size=1024, - max_block_size=1048576): +def split_variable(var_list, service_count, min_block_size=8192): """ - We may need to split dense tensor to one or more blocks and put - them equally onto parameter server. One block is a sub-tensor - aligned by dim[0] of the tensor. - - We need to have a minimal block size so that the calculations in - the parameter server side can gain better performance. By default - minimum block size is 1024. The max block size is used to prevent - very large blocks that may cause send error. - :return: A list of VarBlocks. Each VarBlock specifies a shard of - the var. + We may need to split dense tensor to one or more blocks and put + them equally onto parameter server. One block is a sub-tensor + aligned by dim[0] of the tensor. + + We need to have a minimal block size so that the calculations in + the parameter server side can gain better performance. By default + minimum block size 8K elements (maybe 16bit or 32bit or 64bit). + + Args: + var_list (list): List of variables. + service_count (int): Numel of pserver services. A pserver may have two + or more listening ports. + min_block_size (int): Minimum splitted block size. + Returns: + blocks (list[(varname, block_id, current_block_size)]): A list + of VarBlocks. Each VarBlock specifies a shard of the var. """ blocks = [] for var in var_list: - split_count = pserver_count + split_count = service_count var_numel = reduce(lambda x, y: x * y, var.shape) max_pserver_count = int(math.floor(var_numel / float(min_block_size))) if max_pserver_count == 0: max_pserver_count = 1 - if max_pserver_count < pserver_count: + if max_pserver_count < service_count: split_count = max_pserver_count block_size = int(math.ceil(var_numel / float(split_count))) @@ -136,80 +117,14 @@ def split_dense_variable(var_list, class DistributeTranspiler: - def transpile(self, - optimize_ops, - params_grads, - trainer_id, - program=None, - pservers="127.0.0.1:6174", - trainers=1, - split_method=splitter.round_robin, - sync_mode=True): - """ - Transpile the program to distributed data-parallelism programs. - The main_program will be transformed to use a remote parameter server - to do parameter optimization. And the optimization graph will be put - into a parameter server program. - - Use different methods to split trainable variables to different - parameter servers. - - Steps to transpile trainer: - 1. split variable to multiple blocks, aligned by product(dim[1:]) (width). - 2. rename splited grad variables to add trainer_id suffix ".trainer_%d". - 3. modify trainer program add split_op to each grad variable. - 4. append send_op to send splited variables to server and fetch - params(splited blocks or origin param) from server. - 5. append concat_op to merge splited blocks to update local weights. - - Steps to transpile pserver: - 1. create new program for parameter server. - 2. create params and grad variables that assigned to current server instance. - 3. create a sub-block in the server side program - 4. append ops that should run on current server instance. - 5. add listen_and_serv op - - :param optimize_ops: op list of optimization, should be the - return value of Optimizer.minimize - :type optimize_ops: list - :param params_grads: list of tuple(weight, gradient) - :type params_grads: list - :param trainer_id: one unique id for each trainer in a job. - :type trainer_id: int - :param program: program to transpile, default is default_main_program - :type program: Program - :param pservers: parameter server endpoints like "m1:6174,m2:6174" - :type pservers: string - :param trainers: total number of workers/trainers in the job - :type trainers: int - :param split_method: A function to determin how to split variables - to different servers equally. - :type split_method: function - :type sync_mode: boolean default True - :param sync_mode: if sync_mode is set True, it means that dist transpiler - will transpile the program into sync_mode pserver and trainer program. - """ - assert (callable(split_method)) - if program is None: - program = default_main_program() - self.origin_program = program - self.trainer_num = trainers - self.optimize_ops = optimize_ops - self.sync_mode = sync_mode - # TODO(typhoonzero): currently trainer_id is fetched from cluster system - # like Kubernetes, we should port this to use etcd later when developing - # fluid distributed training with fault-tolerance. - self.trainer_id = trainer_id - pserver_endpoints = pservers.split(",") - self.pserver_endpoints = pserver_endpoints - + def _has_distributed_lookup_table(self): # process lookup_table_op # 1. check all lookup_table_op is distributed # 2. check all lookup_table_op share the same table. distributed_lookup_table_ops = [] # support only one distributed_lookup_table now self.table_name = None - for op in program.global_block().ops: + for op in self.origin_program.global_block().ops: if op.type == LOOKUP_TABLE_TYPE: if op.attrs['is_distributed'] is True: if self.table_name is None: @@ -222,91 +137,216 @@ class DistributeTranspiler: if self.table_name is not None: assert op.input("W")[0] != self.table_name - self.has_distributed_lookup_table = len( - distributed_lookup_table_ops) > 0 - - # step1: For large parameters and gradients, split them into smaller - # blocks. - param_list = [] - grad_list = [] - for p, g in params_grads: - # skip parameter marked not trainable - if type(p) == Parameter and p.trainable == False: - continue - param_list.append(p) - grad_list.append(g) + return len(distributed_lookup_table_ops) > 0 + def _update_dist_lookup_table_vars(self, param_list, grad_list, + params_grads): + # TODO(wuyi): put find a way to put dist lookup table stuff all together. + # update self.table_param_grad and self.trainer_side_table_grad_list + program = self.origin_program if self.has_distributed_lookup_table: param_list = [ param for param in param_list if param.name != self.table_name ] grad_list = [ grad for grad in grad_list - if grad.name != framework.grad_var_name(self.table_name) + if grad.name != grad_var_name(self.table_name) ] self.table_param_grad = [ param_grad for param_grad in params_grads if param_grad[0].name == self.table_name ][0] table_grad_var = self.table_param_grad[1] - self.table_grad_list = [ - program.global_block().create_var( - name="%s.trainer_%d.pserver_%d" % - (table_grad_var.name, trainer_id, index), - type=table_grad_var.type, - shape=table_grad_var.shape, - dtype=table_grad_var.dtype) - for index in range(len(self.pserver_endpoints)) - ] + if self.sync_mode: + self.trainer_side_table_grad_list = [ + program.global_block().create_var( + name="%s.trainer_%d.pserver_%d" % + (table_grad_var.name, self.trainer_id, index), + type=table_grad_var.type, + shape=table_grad_var.shape, + dtype=table_grad_var.dtype) + for index in range(len(self.pserver_endpoints)) + ] + else: + self.trainer_side_table_grad_list = [ + program.global_block().create_var( + name="%s.pserver_%d" % (table_grad_var.name, index), + type=table_grad_var.type, + shape=table_grad_var.shape, + dtype=table_grad_var.dtype) + for index in range(len(self.pserver_endpoints)) + ] + + def _init_splited_vars(self, split_method): + # update these mappings for further transpile: + # 1. param_var_mapping: param var name -> [splited params vars] + # 2. grad_var_mapping: grad var name -> [splited grads vars] + # 3. grad_param_mapping: grad.blockx -> param.blockx + # 4. param_grad_ep_mapping: ep -> {"params": [], "grads": []} + + param_list = [] + grad_list = [] + param_grad_set = set() + for p, g in self.params_grads: + # skip parameter marked not trainable + if type(p) == Parameter and p.trainable == False: + continue + if p.name not in param_grad_set: + param_list.append(p) + param_grad_set.add(p.name) + if g.name not in param_grad_set: + grad_list.append(g) + param_grad_set.add(g.name) + + self._update_dist_lookup_table_vars(param_list, grad_list, + self.params_grads) + + grad_blocks = split_variable(grad_list, len(self.pserver_endpoints)) + param_blocks = split_variable(param_list, len(self.pserver_endpoints)) + assert (len(grad_blocks) == len(param_blocks)) + # origin_varname -> [splited_var] + self.param_var_mapping = self._create_vars_from_blocklist( + self.origin_program, param_blocks) + self.grad_var_mapping = self._create_vars_from_blocklist( + self.origin_program, + grad_blocks, + add_trainer_suffix=self.trainer_num > 1) + self.grad_param_mapping = dict() + for g, p in zip(grad_blocks, param_blocks): + g_name, g_bid, _ = g.split(":") + p_name, p_bid, _ = p.split(":") + self.grad_param_mapping[self.grad_var_mapping[g_name][int(g_bid)]] = \ + self.param_var_mapping[p_name][int(p_bid)] - grad_blocks = split_dense_variable(grad_list, len(pserver_endpoints)) - param_blocks = split_dense_variable(param_list, len(pserver_endpoints)) - # step2: Create new vars for the parameters and gradients blocks and - # add ops to do the split. - grad_var_mapping = self._append_split_op(program, grad_blocks) - param_var_mapping = self._create_vars_from_blocklist(program, - param_blocks) - # step3: Add gradients as send op inputs and parameters as send - # op outputs. - send_inputs = [] - send_outputs = [] - for b in grad_blocks: # append by order - varname, block_id, _ = b.split(":") - send_inputs.append(grad_var_mapping[varname][int(block_id)]) - for b in param_blocks: - varname, block_id, _ = b.split(":") - send_outputs.append(param_var_mapping[varname][int(block_id)]) - # let send_op know which endpoint to send which var to, eplist has the same - # order as send_inputs. - eplist = split_method(send_inputs, pserver_endpoints) # create mapping of endpoint -> split var to create pserver side program self.param_grad_ep_mapping = dict() + [ + self.param_grad_ep_mapping.update({ + ep: { + "params": [], + "grads": [] + } + }) for ep in self.pserver_endpoints + ] + + def transpile(self, + trainer_id, + program=None, + pservers="127.0.0.1:6174", + trainers=1, + split_method=RoundRobin, + sync_mode=True): + """ + :param trainer_id: one unique id for each trainer in a job. + :type trainer_id: int + :param program: program to transpile, default is default_main_program + :type program: Program + :param pservers: parameter server endpoints like "m1:6174,m2:6174" + :type pservers: string + :param trainers: total number of workers/trainers in the job + :type trainers: int + :param split_method: A function to determin how to split variables + to different servers equally. + :type split_method: function + :type sync_mode: boolean default True + :param sync_mode: if sync_mode is set True, it means that dist transpiler + will transpile the program into sync_mode pserver and trainer program. + """ + assert (split_method.__bases__[0] == PSDispatcher) + if program is None: + program = default_main_program() + self.origin_program = program + self.trainer_num = trainers + self.sync_mode = sync_mode + self.trainer_id = trainer_id + pserver_endpoints = pservers.split(",") + self.pserver_endpoints = pserver_endpoints + self.optimize_ops, self.params_grads = self._get_optimize_pass() + + ps_dispatcher = split_method(self.pserver_endpoints) + self.has_distributed_lookup_table = self._has_distributed_lookup_table() + + # split and create vars, then put splited vars in dicts for later use. + self._init_splited_vars(split_method) + + # step 3.1: insert send op to send gradient vars to parameter servers + ps_dispatcher.reset() + send_vars = [] + for orig_varname, splited_vars in self.grad_var_mapping.items(): + eplist = ps_dispatcher.dispatch(splited_vars) + if len(splited_vars) == 1: + orig_varname = splited_vars[0].name + index = find_op_by_output_arg(program.global_block(), + orig_varname) + elif len(splited_vars) > 1: + orig_var = program.global_block().vars[orig_varname] + index = find_op_by_output_arg(program.global_block(), + orig_varname) + self._insert_split_op(program, orig_var, index, splited_vars) + index += 1 + else: + AssertionError("Can not insert the send op by original " + "variable name :", orig_varname) + + program.global_block().insert_op( + index=index + 1, + type="send_vars", + inputs={"X": splited_vars}, + outputs={}, + attrs={ + "epmap": eplist, + RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE + }) + for _, var in enumerate(splited_vars): + send_vars.append(var) + + if self.sync_mode: + program.global_block().append_op( + type="send_barrier", + inputs={}, + outputs={}, + attrs={ + "endpoints": pserver_endpoints, + "sync_mode": self.sync_mode, + RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE + }) + + # step 3.2: insert recv op to receive parameters from parameter server + recv_vars = [] + for _, var in enumerate(send_vars): + recv_vars.append(self.grad_param_mapping[var]) + ps_dispatcher.reset() + eplist = ps_dispatcher.dispatch(recv_vars) + for i, ep in enumerate(eplist): - param = send_outputs[i] - grad = send_inputs[i] - if not self.param_grad_ep_mapping.has_key(ep): - self.param_grad_ep_mapping[ep] = {"params": [], "grads": []} - self.param_grad_ep_mapping[ep]["params"].append(param) - self.param_grad_ep_mapping[ep]["grads"].append(grad) - - rpc_client_var = program.global_block().create_var( - name=RPC_CLIENT_VAR_NAME, - persistable=True, - type=core.VarDesc.VarType.RAW) - - # create send_op + self.param_grad_ep_mapping[ep]["params"].append(recv_vars[i]) + self.param_grad_ep_mapping[ep]["grads"].append(send_vars[i]) + # step4: Concat the parameters splits together after recv. + for varname, splited_var in self.param_var_mapping.iteritems(): + eps = [] + for var in splited_var: + index = [v.name for v in recv_vars].index(var.name) + eps.append(eplist[index]) + + program.global_block().append_op( + type="recv", + inputs={}, + outputs={"Out": splited_var}, + attrs={ + "epmap": eps, + RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE + }) + program.global_block().append_op( - type="send", - inputs={"X": send_inputs}, - outputs={"Out": send_outputs, - "RPCClient": rpc_client_var}, + type="fetch_barrier", + inputs={}, + outputs={}, attrs={ "endpoints": pserver_endpoints, - "epmap": eplist, - "sync_mode": self.sync_mode + RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE }) - # step4: Concat the parameters splits together after recv. - for varname, splited_var in param_var_mapping.iteritems(): + + for varname, splited_var in self.param_var_mapping.iteritems(): if len(splited_var) <= 1: continue orig_param = program.global_block().vars[varname] @@ -317,15 +357,13 @@ class DistributeTranspiler: attrs={"axis": 0}) if self.has_distributed_lookup_table: - self._replace_lookup_table_op_with_prefetch(program, rpc_client_var, - eplist) - self._split_table_grad_and_add_send_vars(program, rpc_client_var, - pserver_endpoints) + self._replace_lookup_table_op_with_prefetch(program, + pserver_endpoints) + self._split_table_grad_and_add_send_vars(program, pserver_endpoints) def get_trainer_program(self): # remove optimize ops and add a send op to main_program - self.origin_program.global_block().delete_ops(self.optimize_ops) - self.origin_program.sync_with_cpp() + delete_ops(self.origin_program.global_block(), self.optimize_ops) # FIXME(typhoonzero): serialize once will fix error occurs when clone. self.origin_program.__str__() return self.origin_program @@ -349,7 +387,6 @@ class DistributeTranspiler: # we don't need to create them when grad arrives. # change client side var name to origin name by # removing ".trainer_%d" suffix - suff_idx = v.name.find(".trainer_") if suff_idx >= 0: orig_var_name = v.name[:suff_idx] @@ -386,40 +423,33 @@ class DistributeTranspiler: # located on current pserver opt_op_on_pserver = [] for _, op in enumerate(self.optimize_ops): - if self._is_opt_op(op) and self._is_opt_op_on_pserver(endpoint, op): + if self._is_optimizer_op(op) and self._is_opt_op_on_pserver( + endpoint, op): opt_op_on_pserver.append(op) # step 3.3 # Iterate through the ops, and if an op and the optimize ops # which located on current pserver are in one set, then # append it into the sub program. - # We try to put optimization program run parallelly, assume - # optimization program always looks like: - # - # prevop -> prevop -> opt op -> following op -> following op; -> - # prevop -> prevop -> opt op -> following op -> following op; -> - # global op -> global op - # - # we put operators that can run parallelly to many program blocks. - # in above example, we seperate ops by the ";". Global ops must run - # after all the optimize ops finished. - global_ops = [] # HACK: optimization global ops only used to scale beta1 and beta2 # replace it with dependency engine. for op in self.optimize_ops: - if op.type == "scale": - for in_name in op.input_arg_names: - if in_name.startswith("beta1_pow_acc") or \ - in_name.startswith("beta2_pow_acc"): - global_ops.append(op) - - def __append_optimize_op__(op, block, grad_to_block_id): - if self._is_opt_op(op): + if self._is_adam_connected_op(op): + global_ops.append(op) + + def __append_optimize_op__(op, block, grad_to_block_id, merged_var): + if self._is_optimizer_op(op): self._append_pserver_ops(block, op, endpoint, grad_to_block_id, - default_main_program()) + self.origin_program, merged_var) else: - self._append_pserver_non_opt_ops(block, op) + self._append_pserver_non_opt_ops(block, op, endpoint) + + def __op_have_grad_input__(op): + for varname in op.input_arg_names: + if varname.find("@GRAD") >= 0: + return varname + return "" # append lr decay ops to the child block if exists lr_ops = self._get_lr_ops() @@ -427,39 +457,41 @@ class DistributeTranspiler: lr_decay_block = pserver_program.create_block( pserver_program.num_blocks - 1) for _, op in enumerate(lr_ops): - self._append_pserver_non_opt_ops(lr_decay_block, op) + self._append_pserver_non_opt_ops(lr_decay_block, op, endpoint) # append op to the current block grad_to_block_id = [] pre_block_idx = pserver_program.num_blocks - 1 for idx, opt_op in enumerate(opt_op_on_pserver): per_opt_block = pserver_program.create_block(pre_block_idx) + # append grad merging ops before clip and weight decay + for _, op in enumerate(self.optimize_ops): + # find the origin @GRAD var before clipping + grad_varname_for_block = __op_have_grad_input__(op) + if ufind.is_connected(op, opt_op) and grad_varname_for_block: + merged_var = self._append_pserver_grad_merge_ops( + per_opt_block, grad_varname_for_block, endpoint, + grad_to_block_id, self.origin_program) for _, op in enumerate(self.optimize_ops): # optimizer is connected to itself if ufind.is_connected(op, opt_op) and op not in global_ops: - __append_optimize_op__(op, per_opt_block, grad_to_block_id) + __append_optimize_op__(op, per_opt_block, grad_to_block_id, + merged_var) # append global ops if global_ops: opt_state_block = pserver_program.create_block( pserver_program.num_blocks - 1) for glb_op in global_ops: - __append_optimize_op__(glb_op, opt_state_block) - - # NOT USED: single block version: - # - # for _, op in enumerate(self.optimize_ops): - # for _, opt_op in enumerate(opt_op_on_pserver): - # if ufind.is_connected(op, opt_op): - # __append_optimize_op__(glb_op, optimize_block) - # break + __append_optimize_op__(glb_op, opt_state_block, + grad_to_block_id, None) # process distributed lookup_table prefetch_block = None if self.has_distributed_lookup_table: pserver_index = self.pserver_endpoints.index(endpoint) table_opt_block = self._create_table_optimize_block( - pserver_index, pserver_program, pre_block_idx) + pserver_index, pserver_program, pre_block_idx, grad_to_block_id) prefetch_block = self._create_prefetch_block( pserver_index, pserver_program, table_opt_block) @@ -495,7 +527,7 @@ class DistributeTranspiler: were split to several blocks. """ s_prog = Program() - orig_s_prog = framework.default_startup_program() + orig_s_prog = default_startup_program() params = self.param_grad_ep_mapping[endpoint]["params"] def _get_splited_name_and_shape(varname): @@ -542,9 +574,11 @@ class DistributeTranspiler: attrs=op.attrs) return s_prog + # ====================== private transpiler functions ===================== + # transpiler function for dis lookup_table - def _replace_lookup_table_op_with_prefetch(self, program, rpc_client_var, - eplist): + def _replace_lookup_table_op_with_prefetch(self, program, + pserver_endpoints): # 1. replace lookup_table_op with split_ids_op -> prefetch_op -> sum_op self.prefetch_input_vars = None self.prefetch_output_vars = None @@ -591,11 +625,11 @@ class DistributeTranspiler: index=op_index + 1, type="prefetch", inputs={'X': self.prefetch_input_vars}, - outputs={ - "Out": self.prefetch_output_vars, - "RPCClient": rpc_client_var - }, - attrs={"epmap": eplist}) + outputs={"Out": self.prefetch_output_vars}, + attrs={ + "epmap": pserver_endpoints, + RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE + }) # insert concat_op program.global_block().insert_op( @@ -611,17 +645,15 @@ class DistributeTranspiler: attrs={"axis": 0}) # delete lookup_table_op - program.global_block().delete_ops([op]) - program.sync_with_cpp() + delete_ops(program.global_block(), [op]) # break for loop break - def _split_table_grad_and_add_send_vars(self, program, rpc_client_var, - pserver_endpoints): + def _split_table_grad_and_add_send_vars(self, program, pserver_endpoints): # 2. add split_ids_op and send_vars_op to send gradient to pservers # there should only be one table_name all_ops = program.global_block().ops - table_grad_name = framework.grad_var_name(self.table_name) + table_grad_name = grad_var_name(self.table_name) for op in all_ops: if table_grad_name in op.output_arg_names: op_index = list(all_ops).index(op) @@ -632,14 +664,17 @@ class DistributeTranspiler: inputs={ 'Ids': [program.global_block().vars[table_grad_name]] }, - outputs={"Out": self.table_grad_list}) + outputs={"Out": self.trainer_side_table_grad_list}) program.global_block().insert_op( index=op_index + 2, type="send_vars", - inputs={'X': self.table_grad_list}, - outputs={"RPCClient": rpc_client_var}, - attrs={"sync_send": True, - "epmap": pserver_endpoints}) + inputs={'X': self.trainer_side_table_grad_list}, + outputs={}, + attrs={ + "sync_send": True, + "epmap": pserver_endpoints, + RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE + }) break def _create_prefetch_block(self, pserver_index, pserver_program, @@ -660,7 +695,7 @@ class DistributeTranspiler: shape=trainer_out.shape, dtype=trainer_out.dtype) prefetch_block.append_op( - type=LOOKUP_TABLE_TYPE, + type="lookup_sparse_table", inputs={'Ids': pserver_ids, "W": table_var}, outputs={"Out": pserver_out}, @@ -672,26 +707,22 @@ class DistributeTranspiler: return prefetch_block def _create_table_optimize_block(self, pserver_index, pserver_program, - pre_block_idx): - def _clone_var(block, var, persistable=True): - assert isinstance(var, Variable) - return block.create_var( - name=var.name, - shape=var.shape, - dtype=var.dtype, - type=var.type, - persistable=persistable) - + pre_block_idx, grad_to_block_id): # STEP: create table optimize block # create table param and grad var in pserver program - param_var = _clone_var( - pserver_program.global_block(), - self.origin_program.global_block().vars[self.table_name]) - grad_var = _clone_var( - pserver_program.global_block(), - self.origin_program.global_block().vars[framework.grad_var_name( - self.table_name)], - persistable=False) + origin_param_var = self.origin_program.global_block().vars[ + self.table_name] + param_var = pserver_program.global_block().create_var( + name=origin_param_var.name, + shape=origin_param_var.shape, + dtype=origin_param_var.dtype, + type=core.VarDesc.VarType.SELECTED_ROWS, + persistable=True) + # parameter must be selected rows + param_var.desc.set_type(core.VarDesc.VarType.SELECTED_ROWS) + grad_var = pserver_program.global_block().clone_variable( + self.origin_program.global_block().vars[grad_var_name( + self.table_name)]) # create table optimize block in pserver program table_opt_op = [ @@ -705,7 +736,7 @@ class DistributeTranspiler: if self.sync_mode: # create grad vars in pserver program table_grad_var = self.table_param_grad[1] - table_grad_list = [ + pserver_side_table_grad_list = [ pserver_program.global_block().create_var( name="%s.trainer_%d.pserver_%d" % (table_grad_var.name, index, pserver_index), @@ -715,11 +746,21 @@ class DistributeTranspiler: for index in range(self.trainer_num) ] - # append sum op for table_grad_list + # append sum op for pserver_side_table_grad_list table_opt_block.append_op( type="sum", - inputs={"X": table_grad_list}, + inputs={"X": pserver_side_table_grad_list}, outputs={"Out": [grad_var]}) + else: + # in async_mode, for table gradient, it also need to be splited to each parameter server + origin_grad_name = grad_var.name + splited_grad_name = self.trainer_side_table_grad_list[ + pserver_index].name + if not splited_grad_name.startswith(origin_grad_name): + raise ValueError("origin_grad_var: " + splited_grad_name + + " grad_var:" + grad_var.name) + grad_var = pserver_program.global_block().rename_var( + origin_grad_name, splited_grad_name) lr_var = pserver_program.global_block().vars[table_opt_op.input( "LearningRate")[0]] @@ -735,9 +776,11 @@ class DistributeTranspiler: outputs=outputs, attrs=table_opt_op.attrs) + # add table parameter gradient and it's block id to grad_to_block_id + grad_to_block_id.append(grad_var.name + ":" + str(table_opt_block.idx)) + return table_opt_block - # ====================== private transpiler functions ===================== def _create_vars_from_blocklist(self, program, block_list, @@ -746,15 +789,27 @@ class DistributeTranspiler: Create vars for each split. NOTE: only grads need to be named for different trainers, use add_trainer_suffix to rename the grad vars. - :return: A dict mapping from original var name to each var split. + Args: + program (ProgramDesc): ProgramDesc which gradients blong. + block_list (list[(varname, block_id, block_size)]): List of gradient blocks. + add_trainer_suffix (Bool): Add trainer suffix to new variable's name if set True. + Returns: + var_mapping (dict(varname->[new_varname_variable])):A dict mapping + from original var name to each var split. """ + + # varname->[(block_id, current_block_size)] block_map = dict() + var_mapping = dict() for block_str in block_list: varname, offset, size = block_str.split(":") if not block_map.has_key(varname): block_map[varname] = [] block_map[varname].append((long(offset), long(size))) + # Do not remove this important debug message: + print("block map: %s" % block_map) + for varname, splited in block_map.iteritems(): orig_var = program.global_block().var(varname) if len(splited) == 1: @@ -818,41 +873,31 @@ class DistributeTranspiler: lod_level=var.lod_level, persistable=persistable) - def _append_split_op(self, program, gradblocks): - # Split variables that need to be split and append respective ops - add_suffix = False - if self.trainer_num > 1: - add_suffix = True - var_mapping = self._create_vars_from_blocklist( - program, gradblocks, add_trainer_suffix=add_suffix) - for varname, splited_vars in var_mapping.iteritems(): - # variable that don't need to split have empty splited_vars - if len(splited_vars) <= 1: - continue - orig_var = program.global_block().vars[varname] - if orig_var.type == core.VarDesc.VarType.SELECTED_ROWS: - height_sections = [] - for v in splited_vars: - height_sections.append(v.shape[0]) - program.global_block().append_op( - type="split_selected_rows", - inputs={"X": orig_var}, - outputs={"Out": splited_vars}, - attrs={"height_sections": height_sections}) - elif orig_var.type == core.VarDesc.VarType.LOD_TENSOR: - sections = [] - for v in splited_vars: - sections.append(v.shape[0]) - program.global_block().append_op( - type="split_byref", - inputs={"X": orig_var}, - outputs={"Out": splited_vars}, - attrs={"sections": sections} # assume split evenly - ) - else: - AssertionError("Variable type should be in set " - "[LOD_TENSOR, SELECTED_ROWS]") - return var_mapping + def _insert_split_op(self, program, orig_var, index, splited_vars): + if orig_var.type == core.VarDesc.VarType.SELECTED_ROWS: + height_sections = [] + for v in splited_vars: + height_sections.append(v.shape[0]) + program.global_block().insert_op( + index=index + 1, + type="split_selected_rows", + inputs={"X": orig_var}, + outputs={"Out": splited_vars}, + attrs={"height_sections": height_sections}) + elif orig_var.type == core.VarDesc.VarType.LOD_TENSOR: + sections = [] + for v in splited_vars: + sections.append(v.shape[0]) + program.global_block().insert_op( + index=index + 1, + type="split_byref", + inputs={"X": orig_var}, + outputs={"Out": splited_vars}, + attrs={"sections": sections} # assume split evenly + ) + else: + AssertionError("Variable type should be in set " + "[LOD_TENSOR, SELECTED_ROWS]") def _get_optimizer_input_shape(self, op_type, varkey, orig_shape, param_shape): @@ -881,17 +926,74 @@ class DistributeTranspiler: pass return orig_shape - def _orig_varname(self, varname): - suff_idx = varname.find(".trainer_") + def _get_varname_parts(self, varname): + # returns origin, blockid, trainerid orig_var_name = "" - if suff_idx >= 0: - orig_var_name = varname[:suff_idx] + trainer_part = "" + block_part = "" + trainer_idx = varname.find(".trainer_") + if trainer_idx >= 0: + trainer_part = varname[trainer_idx + 1:] + else: + trainer_idx = len(varname) + block_index = varname.find(".block") + if block_index >= 0: + block_part = varname[block_index + 1:trainer_idx] else: - orig_var_name = varname - return orig_var_name + block_index = len(varname) + orig_var_name = varname[0:min(block_index, trainer_idx)] + return orig_var_name, block_part, trainer_part + + def _orig_varname(self, varname): + orig, _, _ = self._get_varname_parts(varname) + return orig + + def _append_pserver_grad_merge_ops(self, optimize_block, + grad_varname_for_block, endpoint, + grad_to_block_id, origin_program): + program = optimize_block.program + pserver_block = program.global_block() + grad_block = None + for g in self.param_grad_ep_mapping[endpoint]["grads"]: + if self._orig_varname(g.name) == \ + self._orig_varname(grad_varname_for_block): + grad_block = g + break + if not grad_block: + # do not append this op if current endpoint + # is not dealing with this grad block + return + orig_varname, block_name, trainer_name = self._get_varname_parts( + grad_block.name) + if block_name: + merged_var_name = '.'.join([orig_varname, block_name]) + else: + merged_var_name = orig_varname + merged_var = \ + pserver_block.vars[merged_var_name] + grad_to_block_id.append(merged_var.name + ":" + str(optimize_block.idx)) + if self.sync_mode and self.trainer_num > 1: + vars2merge = [] + for i in xrange(self.trainer_num): + per_trainer_name = "%s.trainer_%d" % \ + (merged_var_name, i) + vars2merge.append(pserver_block.vars[per_trainer_name]) + + optimize_block.append_op( + type="sum", + inputs={"X": vars2merge}, + outputs={"Out": merged_var}) + # TODO(panyx0718): What if it's SELECTED_ROWS. + if not merged_var.type == core.VarDesc.VarType.SELECTED_ROWS: + optimize_block.append_op( + type="scale", + inputs={"X": merged_var}, + outputs={"Out": merged_var}, + attrs={"scale": 1.0 / float(self.trainer_num)}) + return merged_var def _append_pserver_ops(self, optimize_block, opt_op, endpoint, - grad_to_block_id, origin_program): + grad_to_block_id, origin_program, merged_var): program = optimize_block.program pserver_block = program.global_block() new_inputs = dict() @@ -899,40 +1001,6 @@ class DistributeTranspiler: # moment can use the updated shape for key in opt_op.input_names: if key == "Grad": - grad_block = None - for g in self.param_grad_ep_mapping[endpoint]["grads"]: - if same_or_split_var( - self._orig_varname(g.name), - self._orig_varname(opt_op.input(key)[0])): - grad_block = g - break - if not grad_block: - # do not append this op if current endpoint - # is not dealing with this grad block - return - merged_var = \ - pserver_block.vars[self._orig_varname(grad_block.name)] - grad_to_block_id.append(merged_var.name + ":" + str( - optimize_block.idx)) - if self.sync_mode and self.trainer_num > 1: - vars2merge = [] - for i in xrange(self.trainer_num): - per_trainer_name = "%s.trainer_%d" % \ - (self._orig_varname(grad_block.name), i) - vars2merge.append(pserver_block.vars[per_trainer_name]) - - optimize_block.append_op( - type="sum", - inputs={"X": vars2merge}, - outputs={"Out": merged_var}) - # TODO(panyx0718): What if it's SELECTED_ROWS. - if not merged_var.type == core.VarDesc.VarType.SELECTED_ROWS: - optimize_block.append_op( - type="scale", - inputs={"X": merged_var}, - outputs={"Out": merged_var}, - attrs={"scale": 1.0 / float(self.trainer_num)}) - new_inputs[key] = merged_var elif key == "Param": # param is already created on global program @@ -991,17 +1059,31 @@ class DistributeTranspiler: outputs=outputs, attrs=opt_op.attrs) - def _append_pserver_non_opt_ops(self, optimize_block, opt_op): + def _is_splited_grad_var(self, var, var_dict): + grad_block = None + for _, g in var_dict.iteritems(): + if self._orig_varname(g.name) == self._orig_varname(var.name): + if g.name.find(".trainer_") == -1: + grad_block = g + break + return grad_block + + def _append_pserver_non_opt_ops(self, optimize_block, opt_op, endpoint): program = optimize_block.program # Append the ops for parameters that do not need to be optimized/updated inputs = self._get_input_map_from_op( self.origin_program.global_block().vars, opt_op) - for varlist in inputs.itervalues(): + for key, varlist in inputs.iteritems(): if not isinstance(varlist, list): varlist = [varlist] - for var in varlist: - if not program.global_block().vars.has_key(var.name): + # for ops like clipping and weight decay, get the splited var + # for inputs/outputs + grad_block = self._is_splited_grad_var( + var, program.global_block().vars) + if grad_block: + inputs[key] = grad_block + elif not program.global_block().vars.has_key(var.name): program.global_block().create_var( name=var.name, persistable=var.persistable, @@ -1010,13 +1092,16 @@ class DistributeTranspiler: outputs = self._get_output_map_from_op( self.origin_program.global_block().vars, opt_op) - - for varlist in outputs.itervalues(): + for key, varlist in outputs.iteritems(): if not isinstance(varlist, list): varlist = [varlist] - for var in varlist: - program.global_block().clone_variable(var) + grad_block = self._is_splited_grad_var( + var, program.global_block().vars) + if grad_block: + outputs[key] = grad_block + elif not program.global_block().vars.has_key(var.name): + program.global_block().clone_variable(var) optimize_block.append_op( type=opt_op.type, @@ -1062,9 +1147,17 @@ class DistributeTranspiler: ufind.union(op1, op2) return ufind - def _is_opt_op(self, op): - # NOTE: It's a HACK implement. - # optimize op: SGDOptimize, MomentumOptimizer, AdamOptimizer and etc... + def _is_opt_role_op(self, op): + # NOTE: depend on oprole to find out whether this op is for + # optimize + op_maker = core.op_proto_and_checker_maker + optimize_role = core.op_proto_and_checker_maker.OpRole.Optimize + if op_maker.kOpRoleAttrName() in op.attrs and \ + int(op.attrs[op_maker.kOpRoleAttrName()]) == int(optimize_role): + return True + return False + + def _is_optimizer_op(self, op): if "Param" in op.input_names and \ "LearningRate" in op.input_names: return True @@ -1114,7 +1207,7 @@ class DistributeTranspiler: # find learning rate variables by optimize op lr_vars = set() for op in self.optimize_ops: - if self._is_opt_op(op): + if self._is_optimizer_op(op): lr_vars.add(op.input("LearningRate")[0]) find_ops = [] @@ -1131,7 +1224,7 @@ class DistributeTranspiler: # NOTE: we need to skip all optimize ops, since it is connected # with forward/backward ops and lr ops, we only need the lr ops. if op1 != op2 and self._is_op_connected(op1, op2) and \ - not self._is_opt_op(op1) and not self._is_opt_op(op2): + not self._is_optimizer_op(op1) and not self._is_optimizer_op(op2): ufind.union(op1, op2) # find all ops which is related with lr var for op1 in block.ops: @@ -1141,3 +1234,46 @@ class DistributeTranspiler: # we only need to append op for once break return lr_ops + + def _get_optimize_pass(self): + """ + Get optimizer operators, paramters and gradients from origin_program + Returns: + opt_ops (list): optimize operators. + params_grads (dict): paramter->gradient. + """ + block = self.origin_program.global_block() + opt_ops = [] + params_grads = [] + origin_var_dict = self.origin_program.global_block().vars + for op in block.ops: + if self._is_opt_role_op(op): + opt_ops.append(op) + # HACK(wuyi): if we find grad vars from input of optimize + # ops, we may get the output of clip op. Use syntax "@GRAD" + # and op_role_var to get the pair. + for input_name in op.input_arg_names: + if input_name.find("@GRAD") != -1 and \ + op.attrs[RPC_OP_ROLE_ATTR_NAME]: + param_name = op.attrs[OP_ROLE_VAR_ATTR_NAME][0] + params_grads.append([ + origin_var_dict[param_name], + origin_var_dict[input_name] + ]) + elif self._is_adam_connected_op(op): + opt_ops.append(op) + else: + pass + return opt_ops, params_grads + + def _is_adam_connected_op(self, op): + """ + A hack function to determinate whether the input operator + is connected to optimize operator. + """ + if op.type == "scale": + for in_name in op.input_arg_names: + if in_name.startswith("beta1_pow_acc") or \ + in_name.startswith("beta2_pow_acc"): + return True + return False diff --git a/python/paddle/fluid/distribute_transpiler_simple.py b/python/paddle/fluid/transpiler/distribute_transpiler_simple.py similarity index 98% rename from python/paddle/fluid/distribute_transpiler_simple.py rename to python/paddle/fluid/transpiler/distribute_transpiler_simple.py index e94bbb6c39f7a017e2d0b79d050e6ff8e4371a14..ea8c27cdca885dbbf90349b35df9691951264061 100644 --- a/python/paddle/fluid/distribute_transpiler_simple.py +++ b/python/paddle/fluid/transpiler/distribute_transpiler_simple.py @@ -12,10 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. -import framework -from framework import Program, default_main_program, Parameter, Variable -import optimizer -from layer_helper import LayerHelper +from ..framework import Program, default_main_program, Parameter, Variable +from ..layer_helper import LayerHelper def hash_name_to_server(params_grads, pserver_endpoints): diff --git a/python/paddle/fluid/inference_transpiler.py b/python/paddle/fluid/transpiler/inference_transpiler.py similarity index 99% rename from python/paddle/fluid/inference_transpiler.py rename to python/paddle/fluid/transpiler/inference_transpiler.py index 39b01610f96018e1775405a30147e77006cecc16..202aa76084432b4b2378470919b2e924301f2130 100644 --- a/python/paddle/fluid/inference_transpiler.py +++ b/python/paddle/fluid/transpiler/inference_transpiler.py @@ -13,9 +13,9 @@ # limitations under the License. import numpy as np -from framework import Program -from executor import global_scope -from . import core +from .. import core +from ..framework import Program +from ..executor import global_scope class InferenceTranspiler: diff --git a/python/paddle/fluid/memory_optimization_transpiler.py b/python/paddle/fluid/transpiler/memory_optimization_transpiler.py similarity index 98% rename from python/paddle/fluid/memory_optimization_transpiler.py rename to python/paddle/fluid/transpiler/memory_optimization_transpiler.py index 20ed19104207c1f0aa45db8f44570377011f3cde..9ff0ae6fca27d4681891b2033e2f8f95bd825942 100644 --- a/python/paddle/fluid/memory_optimization_transpiler.py +++ b/python/paddle/fluid/transpiler/memory_optimization_transpiler.py @@ -13,11 +13,9 @@ # limitations under the License. from collections import defaultdict -import framework -from framework import Program, default_main_program, Parameter, Variable -import backward -from backward import _rename_arg_ -from . import core +from .. import core +from ..framework import Program, default_main_program, Parameter, Variable +from ..backward import _rename_arg_ dtype_to_size = { core.VarDesc.VarType.FP16: 2, @@ -26,7 +24,8 @@ dtype_to_size = { core.VarDesc.VarType.INT16: 2, core.VarDesc.VarType.INT32: 4, core.VarDesc.VarType.INT64: 8, - core.VarDesc.VarType.BOOL: 1 + core.VarDesc.VarType.BOOL: 1, + core.VarDesc.VarType.UINT8: 1, } SUB_BLOCK_OPS = [ @@ -108,7 +107,7 @@ class ControlFlowGraph(object): # Repeatedly apply liveness updates until the algorithm stablize # on a complete set live input vars and live output vars. while True: - for i in range(self.op_size, 0, -1): + for i in reversed(range(self.op_size)): live_in[i] = set(self._live_in[i]) live_out[i] = set(self._live_out[i]) for s in self._successors[i]: diff --git a/python/paddle/fluid/transpiler/ps_dispatcher.py b/python/paddle/fluid/transpiler/ps_dispatcher.py new file mode 100644 index 0000000000000000000000000000000000000000..d6a68677527deb09ace0e3a23cbc093d6d7b4349 --- /dev/null +++ b/python/paddle/fluid/transpiler/ps_dispatcher.py @@ -0,0 +1,78 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +class PSDispatcher(object): + """ + PSDispatcher is the base class for dispatching vars + into different pserver instance. + You need to implement the `dispatch` inferface. + """ + + def __init__(self, pserver_endpoints): + self._eps = pserver_endpoints + self._step = 0 + + @property + def eps(self): + return self._eps + + def reset(self): + self._step = 0 + + def dispatch(self, varlist): + """ + :param varlist: a list of Variables + :return: a map of pserver endpoint -> varname + """ + AssertionError("Interface has not been implemented.") + + +class HashName(PSDispatcher): + """ + Hash variable names to several endpoints + """ + + def __init__(self, pserver_endpoints): + super(self.__class__, self).__init__(pserver_endpoints) + + def _hash_block(self, block_str, total): + return hash(block_str) % total + + def dispatch(self, varlist): + eplist = [] + for var in varlist: + server_id = self._hash_block(var.name(), len(self._eps)) + server_for_param = self._eps[server_id] + eplist.append(server_for_param) + return eplist + + +class RoundRobin(PSDispatcher): + """ + Distribute variables to serveral endpoints. + """ + + def __init__(self, pserver_endpoints): + super(self.__class__, self).__init__(pserver_endpoints) + + def dispatch(self, varlist): + eplist = [] + for var in varlist: + server_for_param = self._eps[self._step] + eplist.append(server_for_param) + self._step += 1 + if self._step >= len(self._eps): + self._step = 0 + return eplist diff --git a/python/requirements.txt b/python/requirements.txt index daf3f368b92408408897e33223118fe3647aa6de..ea827e9d5a0dcf8eb2ede1f6eaa88c777a138816 100644 --- a/python/requirements.txt +++ b/python/requirements.txt @@ -8,3 +8,4 @@ scipy>=0.19.0 Pillow nltk>=3.2.2 graphviz +LinkChecker diff --git a/python/setup.py.in b/python/setup.py.in index a811b509a90b8b0d84451f54462a0308c062d022..d1464e36150b1a599d81c8679b4f45dcb9feb4d7 100644 --- a/python/setup.py.in +++ b/python/setup.py.in @@ -5,10 +5,10 @@ class BinaryDistribution(Distribution): return True MAJOR = 0 -MINOR = 11 +MINOR = 13 PATCH = 0 RC = 0 -ISTAGED = False +ISTAGED = True @@ -68,7 +68,9 @@ packages=['paddle', 'paddle.fluid', 'paddle.fluid.proto', 'paddle.fluid.proto.profiler', - 'paddle.fluid.layers'] + 'paddle.fluid.layers', + 'paddle.fluid.transpiler', + 'paddle.fluid.transpiler.details'] if '${WITH_FLUID_ONLY}'== 'OFF': packages+=['paddle.proto', diff --git a/tools/aws_benchmarking/server/cluster_master.py b/tools/aws_benchmarking/server/cluster_master.py index 7952e61159ec31a4be5394b50f30cbc20f9b414e..a9b24846544d8aca5e4c7bd5709e70564c088431 100644 --- a/tools/aws_benchmarking/server/cluster_master.py +++ b/tools/aws_benchmarking/server/cluster_master.py @@ -20,6 +20,7 @@ import time import threading import logging import copy +import csv import netaddr import boto3 @@ -136,6 +137,12 @@ parser.add_argument( parser.add_argument( '--master_server_ip', type=str, default="", help="master server private ip") +parser.add_argument( + '--metric_data_identifier', + type=str, + default="**metrics_data: ", + help="key string to identify metrics data") + parser.add_argument( '--no_clean_up', type=str2bool, @@ -155,6 +162,11 @@ logging.basicConfig( log_files = ["master.log"] +metrics = {} + +metrics_csv_file_name = "metrics.csv" +is_metrics_file_created = False + def create_subnet(): # if no vpc id provided, list vpcs @@ -329,12 +341,42 @@ def create_pservers(): cleanup(args.task_name) +def save_metrics_data(str_msg): + #parse msg + logging.info("found metrics data, saving it to csv file") + global is_metrics_file_created + metrics_raw = str_msg.split(",") + with open(args.log_path + metrics_csv_file_name, 'a') as csvfile: + csv_fieldnames = [] + csv_write_data = {} + for metric in metrics_raw: + metric_data = metric.split("=") + metric_key = metric_data[0].strip() + metric_val = float(metric_data[1].strip()) + if not metric_key in metrics: + metrics[metric_key] = [] + metric_repo = metrics[metric_key] + metric_repo.append(metric_val) + csv_fieldnames.append(metric_key) + csv_write_data[metric_key] = metric_val + writer = csv.DictWriter(csvfile, fieldnames=csv_fieldnames) + if not is_metrics_file_created: + writer.writeheader() + is_metrics_file_created = True + writer.writerow(csv_write_data) + logging.info("csv file appended") + + def log_to_file(source, filename): if not filename in log_files: log_files.append(filename) with open(args.log_path + filename, "a") as log_file: for line in iter(source.readline, ""): log_file.write(line) + if (line.startswith(args.metric_data_identifier)): + #found key data, trying to add to csv + line = line.replace(args.metric_data_identifier, "") + save_metrics_data(line) def parse_command(command_raw, defaults={}): @@ -640,6 +682,7 @@ def start_server(args): elif request_path == "/cleanup": self._set_headers() logging.info("Received request to cleanup cluster") + args.no_clean_up = False cleanup(args.task_name) self.wfile.write("cleanup in progress") diff --git a/tools/codestyle/cpplint_pre_commit.hook b/tools/codestyle/cpplint_pre_commit.hook index 94d1e23ce716f7f1d723bad5f1f4c60030f19eb7..b194af76dc529fd52b0aedfab9c41d625fe64c0d 100755 --- a/tools/codestyle/cpplint_pre_commit.hook +++ b/tools/codestyle/cpplint_pre_commit.hook @@ -4,8 +4,12 @@ TOTAL_ERRORS=0 # The trick to remove deleted files: https://stackoverflow.com/a/2413151 for file in $(git diff --cached --name-status | awk '$1 != "D" {print $2}'); do - cpplint $file; - TOTAL_ERRORS=$(expr $TOTAL_ERRORS + $?); + if [[ $file =~ ^(paddle/api/.*|paddle/capi/.*|paddle/contrib/.*|paddle/cuda/.*|paddle/function/.*|paddle/gserver/.*|paddle/math/.*|paddle/optimizer/.*|paddle/parameter/.*|paddle/pserver/.*|paddle/trainer/.*|paddle/utils/.*) ]]; then + continue; + else + cpplint $file; + TOTAL_ERRORS=$(expr $TOTAL_ERRORS + $?); + fi done exit $TOTAL_ERRORS diff --git a/tools/codestyle/docstring_checker.py b/tools/codestyle/docstring_checker.py new file mode 100644 index 0000000000000000000000000000000000000000..48100e5bf989520043b5ca372b02883faea8a9fd --- /dev/null +++ b/tools/codestyle/docstring_checker.py @@ -0,0 +1,334 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +"""DocstringChecker is used to check python doc string's style.""" + +import six +import astroid + +from pylint.checkers import BaseChecker, utils +from pylint.interfaces import IAstroidChecker + +from collections import defaultdict +import re + + +def register(linter): + """Register checkers.""" + linter.register_checker(DocstringChecker(linter)) + + +class Docstring(object): + """Docstring class holds the parsed doc string elements. + """ + + def __init__(self): + self.d = defaultdict(list) #name->[] + self.clear() + + def clear(self): + self.d['Args'] = [] + self.d['Examples'] = [] + self.d['Returns'] = [] + self.d['Raises'] = [] + self.args = {} #arg_name->arg_type + + def get_level(self, string, indent=' '): + level = 0 + unit_size = len(indent) + while string[:unit_size] == indent: + string = string[unit_size:] + level += 1 + + return level + + def parse(self, doc): + """parse gets sections from doc + Such as Args, Returns, Raises, Examples s + Args: + doc (string): is the astroid node doc string. + Returns: + True if doc is parsed successfully. + """ + self.clear() + + lines = doc.splitlines() + state = ("others", -1) + for l in lines: + c = l.strip() + if len(c) <= 0: + continue + + level = self.get_level(l) + if c.startswith("Args:"): + state = ("Args", level) + elif c.startswith("Returns:"): + state = ("Returns", level) + elif c.startswith("Raises:"): + state = ("Raises", level) + elif c.startswith("Examples:"): + state = ("Examples", level) + else: + if level > state[1]: + self.d[state[0]].append(c) + continue + + state = ("others", -1) + self.d[state[0]].append(c) + + self._arg_with_type() + return True + + def get_returns(self): + return self.d['Returns'] + + def get_raises(self): + return self.d['Raises'] + + def get_examples(self): + return self.d['Examples'] + + def _arg_with_type(self): + + for t in self.d['Args']: + m = re.search('([A-Za-z0-9_-]+)\s{0,4}(\(.+\))\s{0,4}:', t) + if m: + self.args[m.group(1)] = m.group(2) + + return self.args + + +class DocstringChecker(BaseChecker): + """DosstringChecker is pylint checker to + check docstring style. + """ + __implements__ = (IAstroidChecker, ) + + POSITIONAL_MESSAGE_ID = 'str-used-on-positional-format-argument' + KEYWORD_MESSAGE_ID = 'str-used-on-keyword-format-argument' + + name = 'doc-string-checker' + symbol = "doc-string" + priority = -1 + msgs = { + 'W9001': ('One line doc string on > 1 lines', symbol + "-one-line", + 'Used when a short doc string is on multiple lines'), + 'W9002': + ('Doc string does not end with "." period', symbol + "-end-with", + 'Used when a doc string does not end with a period'), + 'W9003': ('All args with their types must be mentioned in doc string', + symbol + "-with-all-args", + 'Used when not all arguments are in the doc string '), + 'W9005': ('Missing docstring or docstring is too short', + symbol + "-missing", 'Add docstring longer >=10'), + 'W9006': ('Docstring indent error, use 4 space for indent', + symbol + "-indent-error", 'Use 4 space for indent'), + 'W9007': ('You should add `Returns` in comments', + symbol + "-with-returns", + 'There should be a `Returns` section in comments'), + 'W9008': ('You should add `Raises` section in comments', + symbol + "-with-raises", + 'There should be a `Raises` section in comments'), + } + options = () + + def visit_functiondef(self, node): + """visit_functiondef checks Function node docstring style. + Args: + node (astroid.node): The visiting node. + Returns: + True if successful other wise False. + """ + + self.check_doc_string(node) + + if node.tolineno - node.fromlineno <= 10: + return True + + if not node.doc: + return True + + doc = Docstring() + doc.parse(node.doc) + + self.all_args_in_doc(node, doc) + self.with_returns(node, doc) + self.with_raises(node, doc) + + def visit_module(self, node): + self.check_doc_string(node) + + def visit_classdef(self, node): + self.check_doc_string(node) + + def check_doc_string(self, node): + self.missing_doc_string(node) + self.one_line(node) + self.has_period(node) + self.indent_style(node) + + def missing_doc_string(self, node): + if node.tolineno - node.fromlineno <= 10: + return True + + if node.doc is None or len(node.doc) < 10: + self.add_message('W9005', node=node, line=node.fromlineno) + return False + + # FIXME(gongwb): give the docstring line-no + def indent_style(self, node, indent=4): + """indent_style checks docstring's indent style + Args: + node (astroid.node): The visiting node. + indent (int): The default indent of style + Returns: + True if successful other wise False. + """ + if node.doc is None: + return True + + doc = node.doc + lines = doc.splitlines() + + for l in lines: + cur_indent = len(l) - len(l.lstrip()) + if cur_indent % indent != 0: + self.add_message('W9006', node=node, line=node.fromlineno) + return False + + return True + + def one_line(self, node): + """one_line checks if docstring (len < 40) is on one line. + Args: + node (astroid.node): The node visiting. + Returns: + True if successful otherwise False. + """ + + doc = node.doc + if doc is None: + return True + + if len(doc) > 40: + return True + elif sum(doc.find(nl) for nl in ('\n', '\r', '\n\r')) == -3: + return True + else: + self.add_message('W9001', node=node, line=node.fromlineno) + return False + + return True + + def has_period(self, node): + """has_period checks if one line doc end-with '.' . + Args: + node (astroid.node): the node is visiting. + Returns: + True if successful otherwise False. + """ + if node.doc is None: + return True + + if len(node.doc.splitlines()) > 1: + return True + + if not node.doc.strip().endswith('.'): + self.add_message('W9002', node=node, line=node.fromlineno) + return False + + return True + + def with_raises(self, node, doc): + """with_raises checks if one line doc end-with '.' . + Args: + node (astroid.node): the node is visiting. + doc (Docstring): Docstring object. + Returns: + True if successful otherwise False. + """ + + find = False + for t in node.body: + if not isinstance(t, astroid.Raise): + continue + + find = True + break + + if not find: + return True + + if len(doc.get_raises()) == 0: + self.add_message('W9008', node=node, line=node.fromlineno) + return False + + return True + + def with_returns(self, node, doc): + """with_returns checks if docstring comments what are returned . + Args: + node (astroid.node): the node is visiting. + doc (Docstring): Docstring object. + Returns: + True if successful otherwise False. + """ + + find = False + for t in node.body: + if not isinstance(t, astroid.Return): + continue + + find = True + break + + if not find: + return True + + if len(doc.get_returns()) == 0: + self.add_message('W9007', node=node, line=node.fromlineno) + return False + + return True + + def all_args_in_doc(self, node, doc): + """all_args_in_doc checks if arguments are mentioned in doc + Args: + node (astroid.node): the node is visiting. + doc (Docstring): Docstring object + Returns: + True if successful otherwise False. + """ + args = [] + for arg in node.args.get_children(): + if (not isinstance(arg, astroid.AssignName)) \ + or arg.name == "self": + continue + args.append(arg.name) + + if len(args) <= 0: + return True + + parsed_args = doc.args + if len(args) > 0 and len(parsed_args) <= 0: + print "debug:parsed args: ", parsed_args + self.add_message('W9003', node=node, line=node.fromlineno) + return False + + for t in args: + if t not in parsed_args: + print t, " with (type) not in ", parsed_args + self.add_message('W9003', node=node, line=node.fromlineno) + return False + + return True diff --git a/tools/codestyle/docstring_checker.pyc b/tools/codestyle/docstring_checker.pyc new file mode 100644 index 0000000000000000000000000000000000000000..1ce612ca2318ccb9b9f28d51cb93ce8e5e1d0680 Binary files /dev/null and b/tools/codestyle/docstring_checker.pyc differ diff --git a/tools/codestyle/pylint_pre_commit.hook b/tools/codestyle/pylint_pre_commit.hook new file mode 100755 index 0000000000000000000000000000000000000000..e7c92ba671e0eb778b2ab5447bea7c4b14fe761b --- /dev/null +++ b/tools/codestyle/pylint_pre_commit.hook @@ -0,0 +1,19 @@ +#!/bin/bash + +TOTAL_ERRORS=0 + + +DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" && pwd )" +export PYTHONPATH=$DIR:$PYTHONPATH + +# The trick to remove deleted files: https://stackoverflow.com/a/2413151 +for file in $(git diff --cached --name-status | awk '$1 != "D" {print $2}'); do + pylint --disable=all --load-plugins=docstring_checker \ + --enable=doc-string-one-line,doc-string-end-with,doc-string-with-all-args,doc-string-triple-quotes,doc-string-missing,doc-string-indent-error,doc-string-with-returns,doc-string-with-raises $file; + TOTAL_ERRORS=$(expr $TOTAL_ERRORS + $?); +done + +#exit $TOTAL_ERRORS +#For now, just warning: +exit 0 + diff --git a/tools/codestyle/test_docstring_checker.py b/tools/codestyle/test_docstring_checker.py new file mode 100644 index 0000000000000000000000000000000000000000..0547f7d1610c64b0ca6efa9384e97d658c8276fe --- /dev/null +++ b/tools/codestyle/test_docstring_checker.py @@ -0,0 +1,232 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import docstring_checker +import pylint.testutils +import astroid +import pytest +import sys + + +class TestDocstring(pylint.testutils.CheckerTestCase): + CHECKER_CLASS = docstring_checker.DocstringChecker + + def test_one_line(self): + func_node = astroid.extract_node(''' + def test(): + """get + news. + """ + if True: + return 5 + return 5 + ''') + + self.checker.visit_functiondef(func_node) + got = self.linter.release_messages() + assert len(got) == 1 + assert 'W9001' == got[0][0] + + def test_one_line(self): + func_node = astroid.extract_node(''' + def test(): + """get news""" + if True: + return 5 + return 5 + ''') + + self.checker.visit_functiondef(func_node) + got = self.linter.release_messages() + assert len(got) == 1 + assert 'W9002' == got[0][0] + + def test_args(self): + func_node = astroid.extract_node(''' + def test(scale, mean): + """get news. + Args: + scale (int): scale is the number. + """ + mean=scale + mean=scale + mean=scale + mean=scale + mean=scale + mean=scale + mean=scale + ''') + + self.checker.visit_functiondef(func_node) + got = self.linter.release_messages() + assert len(got) == 1 + assert 'W9003' == got[0][0] + + def test_missing(self): + func_node = astroid.extract_node(''' + def test(): + mean=scale + mean=scale + mean=scale + mean=scale + mean=scale + mean=scale + mean=scale + mean=scale + mean=scale + mean=scale + mean=scale + ''') + + self.checker.visit_functiondef(func_node) + got = self.linter.release_messages() + assert len(got) == 1 + assert 'W9005' == got[0][0] + + def test_indent(self): + func_node = astroid.extract_node(''' + def test(): + """ get get get get get get get get + get get get get get get get get. + """ + pass + ''') + + self.checker.visit_functiondef(func_node) + got = self.linter.release_messages() + assert len(got) == 1 + assert 'W9006' == got[0][0] + + def test_with_resturns(self): + func_node = astroid.extract_node(''' + def test(): + """get news. + Args: + scale (int): scale is the number. + """ + mean=scale + mean=scale + mean=scale + mean=scale + mean=scale + mean=scale + mean=scale + mean=scale + mean=scale + mean=scale + mean=scale + return mean + ''') + + self.checker.visit_functiondef(func_node) + got = self.linter.release_messages() + assert len(got) == 1 + assert 'W9007' == got[0][0] + + def test_with_raises(self): + func_node = astroid.extract_node(''' + def test(): + """get news. + Args: + scale (int): scale is the number. + """ + mean=scale + mean=scale + mean=scale + mean=scale + mean=scale + mean=scale + mean=scale + mean=scale + mean=scale + mean=scale + mean=scale + raise ValueError('A very specific bad thing happened.') + ''') + + self.checker.visit_functiondef(func_node) + got = self.linter.release_messages() + assert len(got) == 1 + assert 'W9008' == got[0][0] + + def test_no_message(self): + p = ''' +def fc(input, + size, + num_flatten_dims=1, + param_attr=None, + bias_attr=None, + act=None, + name=None): + """ + **Fully Connected Layer** + The fully connected layer can take multiple tensors as its inputs. It + creates a variable called weights for each input tensor, which represents + a fully connected weight matrix from each input unit to each output unit. + The fully connected layer multiplies each input tensor with its coresponding + weight to produce an output Tensor. If multiple input tensors are given, + the results of multiple multiplications will be sumed up. If bias_attr is + not None, a bias variable will be created and added to the output. Finally, + if activation is not None, it will be applied to the output as well. + This process can be formulated as follows: + + Args: + input (Variable|list of Variable): The input tensor(s) of this layer, and the dimension of + the input tensor(s) is at least 2. + size(int): The number of output units in this layer. + num_flatten_dims (int, default 1): The fc layer can accept an input tensor with more than + two dimensions. If this happens, the multidimensional tensor will first be flattened + into a 2-dimensional matrix. The parameter `num_flatten_dims` determines how the input + tensor is flattened: the first `num_flatten_dims` (inclusive, index starts from 1) + dimensions will be flatten to form the first dimension of the final matrix (height of + the matrix), and the rest `rank(X) - num_flatten_dims` dimensions are flattened to + form the second dimension of the final matrix (width of the matrix). For example, suppose + `X` is a 6-dimensional tensor with a shape [2, 3, 4, 5, 6], and `num_flatten_dims` = 3. + Then, the flattened matrix will have a shape [2 x 3 x 4, 5 x 6] = [24, 30]. + param_attr (ParamAttr|list of ParamAttr, default None): The parameter attribute for learnable + parameters/weights of this layer. + bias_attr (ParamAttr|list of ParamAttr, default None): The parameter attribute for the bias + of this layer. If it is set to None, no bias will be added to the output units. + act (str, default None): Activation to be applied to the output of this layer. + name (str, default None): The name of this layer. + Returns: + A tensor variable storing the transformation result. + Raises: + ValueError: If rank of the input tensor is less than 2. + Examples: + .. code-block:: python + data = fluid.layers.data(name="data", shape=[32, 32], dtype="float32") + fc = fluid.layers.fc(input=data, size=1000, act="tanh") + """ + raise ValueError('A very specific bad thing happened.') + size = 1 + size = 1 + size = 1 + size = 1 + size = 1 + size = 1 + size = 1 + size = 1 + size = 1 + size = 1 + size = 1 + size = 1 + size = 1 + return size + ''' + + func_node = astroid.extract_node(p) + self.checker.visit_functiondef(func_node) + got = self.linter.release_messages() + assert len(got) == 0 diff --git a/tools/manylinux1/README.md b/tools/manylinux1/README.md index 898e00bd37c7b7bcbcb4a56476ff10c87381e47a..0e5905040175047f5b79939d97a3efcf38992944 100644 --- a/tools/manylinux1/README.md +++ b/tools/manylinux1/README.md @@ -28,3 +28,38 @@ git clone https://github.com/paddlepaddle/paddle cd paddle/tools/manylinux1 REPO=[yourrepo] ./build_all.sh ``` + +## Build PaddlePaddle for the different Python ABIs + +Choose one of the following Python ABI and set the correct environment variables. + +- cp27-cp27m + + ```bash + export LD_LIBRARY_PATH=/opt/_internal/cpython-2.7.11-ucs2/lib:${LD_LIBRARY_PATH#/opt/_internal/cpython-2.7.11-ucs4/lib:} + export PATH=/opt/python/cp27-cp27m/bin/:${PATH} + export PYTHON_FLAGS="-DPYTHON_EXECUTABLE:FILEPATH=/opt/python/cp27-cp27m/bin/python + -DPYTHON_INCLUDE_DIR:PATH=/opt/python/cp27-cp27m/include/python2.7 + -DPYTHON_LIBRARIES:FILEPATH=/opt/_internal/cpython-2.7.11-ucs2/lib/libpython2.7.so" + ``` + +- cp27-cp27mu + + ```bash + export LD_LIBRARY_PATH=/opt/_internal/cpython-2.7.11-ucs4/lib:${LD_LIBRARY_PATH#/opt/_internal/cpython-2.7.11-ucs2/lib:} + export PATH=/opt/python/cp27-cp27mu/bin/:${PATH} + export PYTHON_FLAGS="-DPYTHON_EXECUTABLE:FILEPATH=/opt/python/cp27-cp27mu/bin/python + -DPYTHON_INCLUDE_DIR:PATH=/opt/python/cp27-cp27mu/include/python2.7 + -DPYTHON_LIBRARIES:FILEPATH=/opt/_internal/cpython-2.7.11-ucs4/lib/libpython2.7.so" + ``` + +And then add the `PYTHON_FLAGS` as your cmake flags: + +```bash +cmake .. + ${PYTHON_FLAGS} \ + -DWITH_GPU=OFF \ + ... +``` + +You can find more details about cmake flags at [here](http://www.paddlepaddle.org/docs/develop/documentation/fluid/en/build_and_install/build_from_source_en.html#appendix-build-options) diff --git a/tools/manylinux1/build_scripts/install_nccl2.sh b/tools/manylinux1/build_scripts/install_nccl2.sh index 7efc1fe8651a02684c92fa4b924a4f4416503710..282c5c290da14bd3c04346ab01fdb48423c23f88 100644 --- a/tools/manylinux1/build_scripts/install_nccl2.sh +++ b/tools/manylinux1/build_scripts/install_nccl2.sh @@ -1,11 +1,18 @@ #!/bin/bash -DEB="nccl-repo-ubuntu1604-2.1.4-ga-cuda8.0_1-1_amd64.deb" +VERSION=$(nvcc --version | grep release | grep -oEi "release ([0-9]+)\.([0-9])"| sed "s/release //") +if [ "$VERSION" == "9.0" ]; then + DEB="nccl-repo-ubuntu1604-2.1.15-ga-cuda9.0_1-1_amd64.deb" + URL="http://nccl2-deb.gz.bcebos.com/nccl-repo-ubuntu1604-2.1.15-ga-cuda9.0_1-1_amd64.deb" +else + DEB="nccl-repo-ubuntu1604-2.1.15-ga-cuda8.0_1-1_amd64.deb" + URL="http://nccl2-deb.gz.bcebos.com/nccl-repo-ubuntu1604-2.1.15-ga-cuda8.0_1-1_amd64.deb" +fi + DIR="/nccl2" mkdir -p $DIR # we cached the nccl2 deb package in BOS, so we can download it with wget # install nccl2: http://docs.nvidia.com/deeplearning/sdk/nccl-install-guide/index.html#down -wget -O $DIR/$DEB \ - "http://nccl2-deb.gz.bcebos.com/nccl-repo-ubuntu1604-2.1.4-ga-cuda8.0_1-1_amd64.deb?responseContentDisposition=attachment" +wget -O $DIR/$DEB $URL cd $DIR && ar x $DEB && tar xf data.tar.xz DEBS=$(find ./var/ -name "*.deb") diff --git a/tools/test_runner.py b/tools/test_runner.py new file mode 100644 index 0000000000000000000000000000000000000000..9dc750b89058cd73355a2f7984d577252c03526d --- /dev/null +++ b/tools/test_runner.py @@ -0,0 +1,48 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest +import os +import sys +import paddle.fluid as fluid +import importlib +import cStringIO + + +def main(): + sys.path.append(os.getcwd()) + some_test_failed = False + for module_name in sys.argv[1:]: + buffer = cStringIO.StringIO() + main = fluid.Program() + startup = fluid.Program() + scope = fluid.core.Scope() + with fluid.program_guard(main, startup): + with fluid.scope_guard(scope): + with fluid.unique_name.guard(): + test_loader = unittest.TestLoader() + module = importlib.import_module(module_name) + tests = test_loader.loadTestsFromModule(module) + res = unittest.TextTestRunner(stream=buffer).run(tests) + if not res.wasSuccessful(): + some_test_failed = True + print >> sys.stderr, module_name, 'failed\n', buffer.getvalue( + ) + + if some_test_failed: + exit(1) + + +if __name__ == '__main__': + main() diff --git a/tools/timeline.py b/tools/timeline.py index f4083c824e7333a74661d096d4954609f767c83e..b413bb6fe0505df8fb09fa0759fefb6509b95bc9 100644 --- a/tools/timeline.py +++ b/tools/timeline.py @@ -22,7 +22,11 @@ import paddle.fluid.proto.profiler.profiler_pb2 as profiler_pb2 parser = argparse.ArgumentParser(description=__doc__) parser.add_argument( - '--profile_path', type=str, default='', help='Input profile file name.') + '--profile_path', + type=str, + default='', + help='Input profile file name. If there are multiple file, the format ' + 'should be trainer1=file1,trainer2=file2,ps=file3') parser.add_argument( '--timeline_path', type=str, default='', help='Output timeline file name.') args = parser.parse_args() @@ -108,8 +112,8 @@ class _ChromeTraceFormatter(object): class Timeline(object): - def __init__(self, profile_pb): - self._profile_pb = profile_pb + def __init__(self, profile_dict): + self._profile_dict = profile_dict self._pid = 0 self._devices = dict() self._chrome_trace = _ChromeTraceFormatter() @@ -120,35 +124,37 @@ class Timeline(object): return cur_pid def _allocate_pids(self): - for event in self._profile_pb.events: - if event.type == profiler_pb2.Event.CPU: - if (event.device_id, "CPU") not in self._devices: - pid = self._allocate_pid() - self._devices[(event.device_id, "CPU")] = pid - self._chrome_trace.emit_pid("cpu:block:%d" % - (event.device_id), pid) - elif event.type == profiler_pb2.Event.GPUKernel: - if (event.device_id, "GPUKernel") not in self._devices: - pid = self._allocate_pid() - self._devices[(event.device_id, "GPUKernel")] = pid - self._chrome_trace.emit_pid("gpu:%d" % (event.device_id), - pid) + for k, profile_pb in self._profile_dict.iteritems(): + for event in profile_pb.events: + if event.type == profiler_pb2.Event.CPU: + if (k, event.device_id, "CPU") not in self._devices: + pid = self._allocate_pid() + self._devices[(k, event.device_id, "CPU")] = pid + self._chrome_trace.emit_pid("%s:cpu:block:%d" % + (k, event.device_id), pid) + elif event.type == profiler_pb2.Event.GPUKernel: + if (k, event.device_id, "GPUKernel") not in self._devices: + pid = self._allocate_pid() + self._devices[(k, event.device_id, "GPUKernel")] = pid + self._chrome_trace.emit_pid("%s:gpu:%d" % + (k, event.device_id), pid) def _allocate_events(self): - for event in self._profile_pb.events: - if event.type == profiler_pb2.Event.CPU: - type = "CPU" - elif event.type == profiler_pb2.Event.GPUKernel: - type = "GPUKernel" - pid = self._devices[(event.device_id, type)] - args = {'name': event.name} - if event.memcopy.bytes > 0: - args = {'mem_bytes': event.memcopy.bytes} - # TODO(panyx0718): Chrome tracing only handles ms. However, some - # ops takes micro-seconds. Hence, we keep the ns here. - self._chrome_trace.emit_region( - event.start_ns, (event.end_ns - event.start_ns) / 1.0, pid, - event.sub_device_id, 'Op', event.name, args) + for k, profile_pb in self._profile_dict.iteritems(): + for event in profile_pb.events: + if event.type == profiler_pb2.Event.CPU: + type = "CPU" + elif event.type == profiler_pb2.Event.GPUKernel: + type = "GPUKernel" + pid = self._devices[(k, event.device_id, type)] + args = {'name': event.name} + if event.memcopy.bytes > 0: + args = {'mem_bytes': event.memcopy.bytes} + # TODO(panyx0718): Chrome tracing only handles ms. However, some + # ops takes micro-seconds. Hence, we keep the ns here. + self._chrome_trace.emit_region( + event.start_ns, (event.end_ns - event.start_ns) / 1.0, pid, + event.sub_device_id, 'Op', event.name, args) def generate_chrome_trace(self): self._allocate_pids() @@ -163,11 +169,23 @@ timeline_path = '/tmp/timeline' if args.timeline_path: timeline_path = args.timeline_path -with open(profile_path, 'r') as f: - profile_s = f.read() - profile_pb = profiler_pb2.Profile() - profile_pb.ParseFromString(profile_s) - -tl = Timeline(profile_pb) +profile_paths = profile_path.split(',') +profile_dict = dict() +if len(profile_paths) == 1: + with open(profile_path, 'r') as f: + profile_s = f.read() + profile_pb = profiler_pb2.Profile() + profile_pb.ParseFromString(profile_s) + profile_dict['trainer'] = profile_pb +else: + for profile_path in profile_paths: + k, v = profile_path.split('=') + with open(v, 'r') as f: + profile_s = f.read() + profile_pb = profiler_pb2.Profile() + profile_pb.ParseFromString(profile_s) + profile_dict[k] = profile_pb + +tl = Timeline(profile_dict) with open(timeline_path, 'w') as f: f.write(tl.generate_chrome_trace())