diff --git a/.clang-format b/.clang-format index 9ba433b17362424973626470d930356c2173dd84..aff93435f58c522f5ed1090aef2005f76e91cf31 100644 --- a/.clang-format +++ b/.clang-format @@ -25,4 +25,3 @@ AllowAllParametersOfDeclarationOnNextLine: true BinPackParameters: false BinPackArguments: false ... - diff --git a/.gitignore b/.gitignore index 020d3f0c303f7d850f4ec9c0efe58ab2d57dce2e..ac56a3320ec85769d2c87c072512f5217eca0c24 100644 --- a/.gitignore +++ b/.gitignore @@ -28,3 +28,4 @@ cmake_install.cmake paddle/.timestamp python/paddlepaddle.egg-info/ paddle/pybind/pybind.h +python/paddle/version.py diff --git a/.travis.yml b/.travis.yml index c51e02eb79a9e53a2b8d1d663e8f0c3e0d8c3a61..e2d49daa1981396628efa5d16459eb70e9e76884 100644 --- a/.travis.yml +++ b/.travis.yml @@ -42,7 +42,7 @@ before_install: script: - | timeout 2580 paddle/scripts/travis/${JOB}.sh # 43min timeout - RESULT=$?; if [ $RESULT -eq 0 ] || [ $RESULT -eq 142 ]; then true; else false; fi; + RESULT=$?; if [ $RESULT -eq 0 ] || [ $RESULT -eq 142 ]; then true ;else exit 1; fi; - | if [[ "$JOB" != "build_doc" ]]; then exit 0; fi; if [[ "$TRAVIS_PULL_REQUEST" != "false" ]]; then exit 0; fi; diff --git a/CMakeLists.txt b/CMakeLists.txt index 65164b8472b902be8b0b9d5fb99807d012b8a666..5df83499d5dde29b205ee17fba81a63c9a643235 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -20,6 +20,8 @@ set(PADDLE_BINARY_DIR ${CMAKE_CURRENT_BINARY_DIR}) include(system) project(paddle CXX C Go) +message(STATUS "CXX compiler: " ${CMAKE_CXX_COMPILER} ", version: " ${CMAKE_CXX_COMPILER_VERSION}) +message(STATUS "C compiler: " ${CMAKE_C_COMPILER} ", version: " ${CMAKE_C_COMPILER_VERSION}) find_package(Sphinx) if(NOT CMAKE_CROSSCOMPILING) @@ -54,7 +56,9 @@ option(WITH_C_API "Compile PaddlePaddle with C-API(Prediction)" OFF) option(WITH_GOLANG "Compile PaddlePaddle with GOLANG" OFF) option(GLIDE_INSTALL "Download and install go dependencies " ON) option(USE_NNPACK "Compile PaddlePaddle with NNPACK library" OFF) +option(WITH_DISTRIBUTE "Compile with grpc distributed support" OFF) option(USE_EIGEN_FOR_BLAS "Use matrix multiplication in Eigen" OFF) +option(WITH_ARM_FP16 "Use half precision support on armv8.2-a cpu" OFF) # CMAKE_BUILD_TYPE if(NOT CMAKE_BUILD_TYPE) @@ -67,9 +71,6 @@ if(ANDROID OR IOS) if(ANDROID) if(${CMAKE_SYSTEM_VERSION} VERSION_LESS "16") message(FATAL_ERROR "Unsupport standalone toolchains with Android API level lower than 16") - elseif(${CMAKE_SYSTEM_VERSION} VERSION_LESS "21") - # TODO: support glog for Android api 16 ~ 19 in the future - message(WARNING "Using the unofficial git repository instead") endif() endif() @@ -83,6 +84,8 @@ if(ANDROID OR IOS) "Disable RDMA when cross-compiling for Android and iOS" FORCE) set(WITH_MKL OFF CACHE STRING "Disable MKL when cross-compiling for Android and iOS" FORCE) + set(WITH_GOLANG OFF CACHE STRING + "Disable golang when cross-compiling for Android and iOS" FORCE) # Compile PaddlePaddle mobile inference library if (NOT WITH_C_API) @@ -133,6 +136,8 @@ include(external/any) # download libn::any include(external/eigen) # download eigen3 include(external/pybind11) # download pybind11 include(external/nccl) +include(external/cares) +include(external/grpc) include(cudnn) # set cudnn libraries, must before configure include(configure) # add paddle env configuration @@ -194,6 +199,10 @@ if(WITH_GOLANG) endif(WITH_GOLANG) set(PADDLE_PYTHON_BUILD_DIR "${CMAKE_CURRENT_BINARY_DIR}/python/build") + +SET(CMAKE_CXX_FLAGS_RELWITHDEBINFO "-O3 -g -DNDEBUG") +SET(CMAKE_C_FLAGS_RELWITHDEBINFO "-O3 -g -DNDEBUG") + add_subdirectory(paddle) if(WITH_PYTHON) add_subdirectory(python) diff --git a/Dockerfile b/Dockerfile index 150344a8116e2be9b5bab8e5fdcc9c37f4025020..857d3f3e5f64791146741ffb29feabfcb2ecbb84 100644 --- a/Dockerfile +++ b/Dockerfile @@ -29,7 +29,7 @@ RUN apt-get update && \ automake locales clang-format swig doxygen cmake \ liblapack-dev liblapacke-dev libboost-dev \ clang-3.8 llvm-3.8 libclang-3.8-dev \ - net-tools && \ + net-tools libtool && \ apt-get clean -y # Install Go and glide diff --git a/README.md b/README.md index db0fbd88b250cdc2a3cc77521cc1c2cea77c6e87..577528e7aaf45ce002467590ec66b19afb145920 100644 --- a/README.md +++ b/README.md @@ -2,8 +2,8 @@ [![Build Status](https://travis-ci.org/PaddlePaddle/Paddle.svg?branch=develop)](https://travis-ci.org/PaddlePaddle/Paddle) -[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://doc.paddlepaddle.org/develop/doc/) -[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://doc.paddlepaddle.org/develop/doc_cn/) +[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/index_en.html) +[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://www.paddlepaddle.org/docs/develop/documentation/zh/getstarted/index_cn.html) [![Coverage Status](https://coveralls.io/repos/github/PaddlePaddle/Paddle/badge.svg?branch=develop)](https://coveralls.io/github/PaddlePaddle/Paddle?branch=develop) [![Release](https://img.shields.io/github/release/PaddlePaddle/Paddle.svg)](https://github.com/PaddlePaddle/Paddle/releases) [![License](https://img.shields.io/badge/license-Apache%202-blue.svg)](LICENSE) @@ -36,7 +36,7 @@ Please refer to our [release announcement](https://github.com/PaddlePaddle/Paddl examples: - Optimized math operations through SSE/AVX intrinsics, BLAS libraries - (e.g. MKL, ATLAS, cuBLAS) or customized CPU/GPU kernels. + (e.g. MKL, OpenBLAS, cuBLAS) or customized CPU/GPU kernels. - Highly optimized recurrent networks which can handle **variable-length** sequence without padding. - Optimized local and distributed training for models with high dimensional @@ -61,32 +61,32 @@ Please refer to our [release announcement](https://github.com/PaddlePaddle/Paddl ## Installation It is recommended to check out the -[Docker installation guide](http://doc.paddlepaddle.org/develop/doc/getstarted/build_and_install/docker_install_en.html) +[Docker installation guide](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/build_and_install/docker_install_en.html) before looking into the -[build from source guide](http://doc.paddlepaddle.org/develop/doc/getstarted/build_and_install/build_from_source_en.html). +[build from source guide](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/build_and_install/build_from_source_en.html). ## Documentation -We provide [English](http://doc.paddlepaddle.org/develop/doc/) and -[Chinese](http://doc.paddlepaddle.org/doc_cn/) documentation. +We provide [English](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/index_en.html) and +[Chinese](http://www.paddlepaddle.org/docs/develop/documentation/zh/getstarted/index_cn.html) documentation. -- [Deep Learning 101](http://book.paddlepaddle.org/index.html) +- [Deep Learning 101](http://www.paddlepaddle.org/docs/develop/book/01.fit_a_line/index.html) You might want to start from this online interactive book that can run in a Jupyter Notebook. -- [Distributed Training](http://doc.paddlepaddle.org/develop/doc/howto/usage/cluster/cluster_train_en.html) +- [Distributed Training](http://www.paddlepaddle.org/docs/develop/documentation/en/howto/usage/cluster/cluster_train_en.html) You can run distributed training jobs on MPI clusters. -- [Distributed Training on Kubernetes](http://doc.paddlepaddle.org/develop/doc/howto/usage/k8s/k8s_en.html) +- [Distributed Training on Kubernetes](http://www.paddlepaddle.org/docs/develop/documentation/en/howto/usage/cluster/k8s_en.html) You can also run distributed training jobs on Kubernetes clusters. -- [Python API](http://doc.paddlepaddle.org/develop/doc/api/index_en.html) +- [Python API](http://www.paddlepaddle.org/docs/develop/documentation/en/api/index_en.html) Our new API enables much shorter programs. -- [How to Contribute](http://doc.paddlepaddle.org/develop/doc/howto/dev/contribute_to_paddle_en.html) +- [How to Contribute](http://www.paddlepaddle.org/docs/develop/documentation/en/howto/dev/contribute_to_paddle_en.html) We appreciate your contributions! diff --git a/RELEASE.cn.md b/RELEASE.cn.md index 5deaf230a8f5dd3089993f0fc79b9460fd049750..494c59730dd3c2830514e8924aa3d59a34ac412e 100644 --- a/RELEASE.cn.md +++ b/RELEASE.cn.md @@ -1,3 +1,62 @@ +# v0.11.0版本 + +## PaddlePaddle Fluid + +- PaddlePaddle发布版本v0.11.0包含一个新的特性*PaddlePaddle Fluid*. Fluid 是设计用来让用户像Pytorch和Tensorflow Eager Execution一样执行程序。在这些系统中，不再有*模型*这个概念，应用也不再包含一个用于描述Operator图或者一系列层的符号描述，而是像通用程序那样描述训练或者预测的过程。而Fluid与PyTorch或Eager Execution的区别在于Fluid不依赖Python提供的控制流，例如 if-else-then或者for，而是提供了基于C++实现的控制流并暴露了对应的用with语法实现的Python接口。例如： + + https://github.com/PaddlePaddle/Paddle/blob/3df78ed2a98d37f7ae6725894cc7514effd5664b/python/paddle/v2/fluid/tests/test_while_op.py#L36-L44 + +- 在v0.11.0版本中，我们提供了一个C++类`Executor`用于运行一个Fluid程序。Executor类似一个解释器。在未来的版本中，我们将提升和优化Executor成为一个调试器，就像GDB。并可能提供一些编译器，这个编译器会读取一个上文所描述的应用然后编译成一个等价的 +源代码，这个源代码可以被nvcc编译成可以使用CUDA的二进制，或者被icc编译成可以充分利用Intel CPU的二进制。 + + +## 新特点 + +* 发布 `PaddlePaddle Fluid`。 +* 增加了用于模型预测的C-API。 +* 用Fluid API实现了一个简单的GAN的例子。 +* 增加了关于性能调优的文档。 +* 为`paddle.v2.dataset`下载数据集提供了重试机制. +* C++中使用protobuf-lite替换protobuf减少了二进制的大小。 +* 发布了新特性 [Elastic Deep Learning (EDL)](https://github.com/PaddlePaddle/cloud/tree/develop/doc/autoscale/experiment). +* 基于Bazel API利用cmake实现了一个的新的构建系统函数库。 +* 当使用编译选项`WITH_MKL=ON`时自动下载和编译Intel® [MKLML](https://github.com/01org/mkl-dnn/releases/download/v0.11/mklml_lnx_2018.0.1.20171007.tgz) 函数库. +* [Intel® MKL-DNN on PaddlePaddle](https://github.com/PaddlePaddle/Paddle/tree/develop/doc/design/mkldnn): + - 完成了 11个 MKL-DNN 层: Convolution, Fully connectivity, Pooling, ReLU, Tanh, ELU, Softmax, BatchNorm, AddTo, Concat, LRN。 + - 完成了 3个 MKL-DNN 网络: VGG-19, ResNet-50, GoogleNet + - 基于Intel Skylake 6148 CPU的[性能测试](https://github.com/PaddlePaddle/Paddle/blob/develop/benchmark/IntelOptimizedPaddle.md) : 相对于MKLML有2~3倍的训练加速。 +* 增加 [softsign activation](http://www.paddlepaddle.org/docs/develop/documentation/zh/api/v2/config/activation.html#softsign) +* 增加 [dot product layer](http://www.paddlepaddle.org/docs/develop/documentation/zh/api/v2/config/layer.html#dot-prod) +* 增加 [L2 distance layer](http://www.paddlepaddle.org/docs/develop/documentation/zh/api/v2/config/layer.html#l2-distance) +* 增加 [sub-nested sequence layer](http://www.paddlepaddle.org/docs/develop/documentation/zh/api/v2/config/layer.html#sub-nested-seq) +* 增加 [kmax sequence score layer](http://www.paddlepaddle.org/docs/develop/documentation/zh/api/v2/config/layer.html#kmax-sequence-score) +* 增加 [sequence slice layer](http://www.paddlepaddle.org/docs/develop/documentation/zh/api/v2/config/layer.html#seq-slice) +* 增加 [row convolution layer](http://www.paddlepaddle.org/docs/develop/documentation/zh/api/v2/config/layer.html#row-conv) +* 增加移动端友好的网页 + +## 改进 + +* 使用一个Python`whl`包即可安装. +* [V2 API可以实现用户定制化评估](https://github.com/PaddlePaddle/models/tree/develop/ltr#训练过程中输出自定义评估指标)。 +* 将 `PADDLE_ONLY_CPU` 改为 `PADDLE_WITH_GPU`, 因为我们会支持多种设备。 +* 删除了有一些bug的BarrierStat。 +* 清理和删除了paddle::Parameter中未使用的函数。 +* 删除了ProtoDataProvider。 +* Huber loss同时支持回归和分类。 +* 为sequence pooling 层增加`stride`参数。 +* v2 API自动使用cudnn batch normalization。 +* 可以使用一个固定的参数名共享BN层的参数。 +* 2D convolution operation支持variable-dimension input特性。 +* 重构cmake中关于CUDA的部分并实现自动检测GPU架构的功能。 +* 优化网页导航。 + +## 错误修复 + +* 修复ROI pooling的Bug. cc9a761 +* 修复当label是dense vector是AUC变成0的问题. #5274 +* 修复WarpCTC 层的Bug. + + # v0.10.0版本 我们非常高兴发布了PaddlePaddle V0.10.0版，并开发了新的[Python API](http://research.baidu.com/paddlepaddles-new-api-simplifies-deep-learning-programs/)。 diff --git a/RELEASE.md b/RELEASE.md index 146f7afa7dfbc152500b82fde28445ae3155c16c..5a62c955131007c9f3329d162c20d1b462550019 100644 --- a/RELEASE.md +++ b/RELEASE.md @@ -1,3 +1,75 @@ +# Release v0.11.0 + +## PaddlePaddle Fluid + +- Release 0.11.0 includes a new feature *PaddlePaddle Fluid*. Fluid is + designed to allow users to program like PyTorch and TensorFlow Eager Execution. + In these systems, there is no longer the concept *model* and applications + do not include a symbolic description of a graph of operators nor a sequence + of layers. Instead, applications look exactly like a usual program that + describes a process of training or inference. The difference between + Fluid and PyTorch or Eager Execution is that Fluid doesn't rely on Python's + control-flow, `if-then-else` nor `for`. Instead, Fluid provides its + C++ implementations and their Python binding using the `with` statement. For an example + + https://github.com/PaddlePaddle/Paddle/blob/3df78ed2a98d37f7ae6725894cc7514effd5664b/python/paddle/v2/fluid/tests/test_while_op.py#L36-L44 + +- In 0.11.0, we provides a C++ class `Executor` to run a Fluid program. +Executor works like an interpreter. In future version, we will improve +`Executor` into a debugger like GDB, and we might provide some compilers, +which, for example, takes an application like the above one, and outputs +an equivalent C++ source program, which can be compiled using +[`nvcc`](http://docs.nvidia.com/cuda/cuda-compiler-driver-nvcc/index.html) +to generate binaries that use CUDA, or using +[`icc`](https://software.intel.com/en-us/c-compilers) to generate binaries +that make full use of Intel CPUs. + +## New Features + +* Release `PaddlePaddle Fluid`. +* Add C-API for model inference +* Use fluid API to create a simple GAN demo. +* Add develop guide about performance tunning. +* Add retry when download `paddle.v2.dataset`. +* Linking protobuf-lite not protobuf in C++. Reduce the binary size. +* Feature [Elastic Deep Learning (EDL)](https://github.com/PaddlePaddle/cloud/tree/develop/doc/autoscale/experiment) released. +* A new style cmake functions for Paddle. It is based on Bazel API. +* Automatically download and compile with Intel® [MKLML](https://github.com/01org/mkl-dnn/releases/download/v0.11/mklml_lnx_2018.0.1.20171007.tgz) library as CBLAS when build `WITH_MKL=ON`. +* [Intel® MKL-DNN on PaddlePaddle](https://github.com/PaddlePaddle/Paddle/tree/develop/doc/design/mkldnn): + - Complete 11 MKL-DNN layers: Convolution, Fully connectivity, Pooling, ReLU, Tanh, ELU, Softmax, BatchNorm, AddTo, Concat, LRN. + - Complete 3 MKL-DNN networks: VGG-19, ResNet-50, GoogleNet + - [Benchmark](https://github.com/PaddlePaddle/Paddle/blob/develop/benchmark/IntelOptimizedPaddle.md) on Intel Skylake 6148 CPU: 2~3x training speedup compared with MKLML. +* Add the [`softsign` activation](http://www.paddlepaddle.org/docs/develop/documentation/zh/api/v2/config/activation.html#softsign). +* Add the [dot product layer](http://www.paddlepaddle.org/docs/develop/documentation/zh/api/v2/config/layer.html#dot-prod). +* Add the [L2 distance layer](http://www.paddlepaddle.org/docs/develop/documentation/zh/api/v2/config/layer.html#l2-distance). +* Add the [sub-nested sequence layer](http://www.paddlepaddle.org/docs/develop/documentation/zh/api/v2/config/layer.html#sub-nested-seq). +* Add the [kmax sequence score layer](http://www.paddlepaddle.org/docs/develop/documentation/zh/api/v2/config/layer.html#kmax-sequence-score). +* Add the [sequence slice layer](http://www.paddlepaddle.org/docs/develop/documentation/zh/api/v2/config/layer.html#seq-slice). +* Add the [row convolution layer](http://www.paddlepaddle.org/docs/develop/documentation/zh/api/v2/config/layer.html#row-conv) +* Add mobile friendly webpages. + +## Improvements + +* Build and install using a single `whl` package. +* [Custom evaluating in V2 API](https://github.com/PaddlePaddle/models/tree/develop/ltr#训练过程中输出自定义评估指标). +* Change `PADDLE_ONLY_CPU` to `PADDLE_WITH_GPU`, since we will support many kinds of devices. +* Remove buggy BarrierStat. +* Clean and remove unused functions in paddle::Parameter. +* Remove ProtoDataProvider. +* Huber loss supports both regression and classification. +* Add the `stride` parameter for sequence pooling layers. +* Enable v2 API use cudnn batch normalization automatically. +* The BN layer's parameter can be shared by a fixed the parameter name. +* Support variable-dimension input feature for 2D convolution operation. +* Refine cmake about CUDA to automatically detect GPU architecture. +* Improved website navigation. + +## Bug Fixes + +* Fix bug in ROI pooling. cc9a761 +* Fix AUC is zero when label is dense vector. #5274 +* Fix bug in WarpCTC layer. + # Release v0.10.0 We are glad to release version 0.10.0. In this version, we are happy to release the new diff --git a/benchmark/IntelOptimizedPaddle.md b/benchmark/IntelOptimizedPaddle.md index 040f5ffa41968cbf93a817faa1db86c18956341e..084d3237d9cfe9ca4837f77cf5f70a2449cfcc03 100644 --- a/benchmark/IntelOptimizedPaddle.md +++ b/benchmark/IntelOptimizedPaddle.md @@ -2,28 +2,27 @@ Machine: -- Server - - Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz, 2 Sockets, 20 Cores per socket -- Laptop - - DELL XPS15-9560-R1745: i7-7700HQ 8G 256GSSD - - i5 MacBook Pro (Retina, 13-inch, Early 2015) -- Desktop - - i7-6700k +- Server: Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz, 2 Sockets, 20 Cores per socket +- Laptop: TBD System: CentOS release 6.3 (Final), Docker 1.12.1. -PaddlePaddle: paddlepaddle/paddle:latest (TODO: will rerun after 0.11.0) - -- MKL-DNN tag v0.10 -- MKLML 2018.0.20170720 -- OpenBLAS v0.2.20 +PaddlePaddle: (TODO: will rerun after 0.11.0) +- paddlepaddle/paddle:latest (for MKLML and MKL-DNN) + - MKL-DNN tag v0.11 + - MKLML 2018.0.1.20171007 +- paddlepaddle/paddle:latest-openblas (for OpenBLAS) + - OpenBLAS v0.2.20 On each machine, we will test and compare the performance of training on single node using MKL-DNN / MKLML / OpenBLAS respectively. ## Benchmark Model ### Server + +#### Training Test on batch size 64, 128, 256 on Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz +Pay attetion that the speed below includes forward, backward and parameter update time. So we can not directly compare the data with the benchmark of caffe `time` [command](https://github.com/PaddlePaddle/Paddle/blob/develop/benchmark/caffe/image/run.sh#L9), which only contain forward and backward. The updating time of parameter would become very heavy when the weight size are large, especially on alexnet. Input image size - 3 * 224 * 224, Time: images/second @@ -31,18 +30,78 @@ Input image size - 3 * 224 * 224, Time: images/second | BatchSize | 64 | 128 | 256 | |--------------|-------| -----| --------| -| OpenBLAS | 7.82 | 8.62 | 10.34 | -| MKLML | 11.02 | 12.86 | 15.33 | -| MKL-DNN | 27.69 | 28.8 | 29.27 | +| OpenBLAS | 7.80 | 9.00 | 10.80 | +| MKLML | 12.12 | 13.70 | 16.18 | +| MKL-DNN | 28.46 | 29.83 | 30.44 | + -chart on batch size 128 -TBD + - ResNet-50 + +| BatchSize | 64 | 128 | 256 | +|--------------|-------| ------| -------| +| OpenBLAS | 25.22 | 25.68 | 27.12 | +| MKLML | 32.52 | 31.89 | 33.12 | +| MKL-DNN | 81.69 | 82.35 | 84.08 | + + - - ResNet - GoogLeNet +| BatchSize | 64 | 128 | 256 | +|--------------|-------| ------| -------| +| OpenBLAS | 89.52 | 96.97 | 108.25 | +| MKLML | 128.46| 137.89| 158.63 | +| MKL-DNN     | 250.46| 264.83| 269.50 | + + + +- Alexnet + +| BatchSize | 64 | 128 | 256 | +|--------------|--------| ------ | -------| +| OpenBLAS | 2.13 | 2.45 | 2.68 | +| MKLML | 66.37 | 105.60 | 144.04 | +| MKL-DNN | 399.00 | 498.94 | 626.53 | + +chart TBD + +#### Inference +Test on batch size 1, 2, 4, 8, 16 on Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz +- VGG-19 + +| BatchSize | 1 | 2 | 4 | 8 | 16 | +|-----------|-------|-------|-------|-------|-------| +| OpenBLAS | 1.07 | 1.08 | 1.06 | 0.88 | 0.65 | +| MKLML | 5.58 | 9.80 | 15.15 | 21.21 | 28.67 | +| MKL-DNN | 75.07 | 88.64 | 82.58 | 92.29 | 96.75 | + +- ResNet-50 + +| BatchSize | 1 | 2 | 4 | 8 | 16 | +|-----------|-------|--------|--------|--------|--------| +| OpenBLAS | 3.35 | 3.19 | 3.09 | 2.55 | 1.96 | +| MKLML | 6.33 | 12.02 | 22.88 | 40.53 | 63.09 | +| MKL-DNN | 107.83| 148.84 | 177.78 | 189.35 | 217.69 | + + +- GoogLeNet + +| BatchSize | 1 | 2 | 4 | 8 | 16 | +|-----------|--------|--------|--------|--------|--------| +| OpenBLAS | 12.04 | 11.31 | 10.00 | 9.07 | 4.34 | +| MKLML | 22.74 | 41.56 | 81.22 | 133.47 | 210.53 | +| MKL-DNN | 175.10 | 272.92 | 450.70 | 512.00 | 600.94 | + +- Alexnet + +| BatchSize | 1 | 2 | 4 | 8 | 16 | +|-----------|--------|--------|--------|--------|--------| +| OpenBLAS | | | | | | +| MKLML | 21.32 | 36.55 | 73.06 | 131.15 | 192.77 | +| MKL-DNN | 442.91 | 656.41 | 719.10 | 847.68 | 850.51 | + +chart TBD + ### Laptop TBD -### Desktop -TBD diff --git a/benchmark/figs/googlenet-cpu-train.png b/benchmark/figs/googlenet-cpu-train.png new file mode 100644 index 0000000000000000000000000000000000000000..c3f67faf096fe9b45dd815f294b41679dc7c9e54 Binary files /dev/null and b/benchmark/figs/googlenet-cpu-train.png differ diff --git a/benchmark/figs/resnet-cpu-train.png b/benchmark/figs/resnet-cpu-train.png new file mode 100644 index 0000000000000000000000000000000000000000..b96ecd5ff940c0d000613b1ed1f11fb16796cf47 Binary files /dev/null and b/benchmark/figs/resnet-cpu-train.png differ diff --git a/benchmark/figs/vgg-cpu-train.png b/benchmark/figs/vgg-cpu-train.png new file mode 100644 index 0000000000000000000000000000000000000000..f830ca6a87d10b72a5113636dd5686ab25a2e864 Binary files /dev/null and b/benchmark/figs/vgg-cpu-train.png differ diff --git a/benchmark/paddle/image/alexnet.py b/benchmark/paddle/image/alexnet.py index 3358d43a4b08c6a9b89d59e1a8be53ee1f12bbe0..cad6051f1413a5bb95f87a940f3aa81e49e5d282 100644 --- a/benchmark/paddle/image/alexnet.py +++ b/benchmark/paddle/image/alexnet.py @@ -6,10 +6,24 @@ height = 227 width = 227 num_class = 1000 batch_size = get_config_arg('batch_size', int, 128) +gp = get_config_arg('layer_num', int, 1) +is_infer = get_config_arg("is_infer", bool, False) +num_samples = get_config_arg('num_samples', int, 2560) -args = {'height': height, 'width': width, 'color': True, 'num_class': num_class} +args = { + 'height': height, + 'width': width, + 'color': True, + 'num_class': num_class, + 'is_infer': is_infer, + 'num_samples': num_samples +} define_py_data_sources2( - "train.list", None, module="provider", obj="process", args=args) + "train.list" if not is_infer else None, + "test.list" if is_infer else None, + module="provider", + obj="process", + args=args) settings( batch_size=batch_size, @@ -31,7 +45,7 @@ net = img_pool_layer(input=net, pool_size=3, stride=2) # conv2 net = img_conv_layer( - input=net, filter_size=5, num_filters=256, stride=1, padding=2, groups=1) + input=net, filter_size=5, num_filters=256, stride=1, padding=2, groups=gp) net = img_cmrnorm_layer(input=net, size=5, scale=0.0001, power=0.75) net = img_pool_layer(input=net, pool_size=3, stride=2) @@ -40,11 +54,11 @@ net = img_conv_layer( input=net, filter_size=3, num_filters=384, stride=1, padding=1) # conv4 net = img_conv_layer( - input=net, filter_size=3, num_filters=384, stride=1, padding=1, groups=1) + input=net, filter_size=3, num_filters=384, stride=1, padding=1, groups=gp) # conv5 net = img_conv_layer( - input=net, filter_size=3, num_filters=256, stride=1, padding=1, groups=1) + input=net, filter_size=3, num_filters=256, stride=1, padding=1, groups=gp) net = img_pool_layer(input=net, pool_size=3, stride=2) net = fc_layer( @@ -59,6 +73,9 @@ net = fc_layer( layer_attr=ExtraAttr(drop_rate=0.5)) net = fc_layer(input=net, size=1000, act=SoftmaxActivation()) -lab = data_layer('label', num_class) -loss = cross_entropy(input=net, label=lab) -outputs(loss) +if is_infer: + outputs(net) +else: + lab = data_layer('label', num_class) + loss = cross_entropy(input=net, label=lab) + outputs(loss) diff --git a/benchmark/paddle/image/googlenet.py b/benchmark/paddle/image/googlenet.py index bc893bab98c4d2e07c62fbd012d51a0939db4766..2a850ccb7f2c75b467554181fc5f4aa8f2b97a09 100644 --- a/benchmark/paddle/image/googlenet.py +++ b/benchmark/paddle/image/googlenet.py @@ -5,10 +5,24 @@ height = 224 width = 224 num_class = 1000 batch_size = get_config_arg('batch_size', int, 128) - -args = {'height': height, 'width': width, 'color': True, 'num_class': num_class} +use_gpu = get_config_arg('use_gpu', bool, True) +is_infer = get_config_arg("is_infer", bool, False) +num_samples = get_config_arg('num_samples', int, 2560) + +args = { + 'height': height, + 'width': width, + 'color': True, + 'num_class': num_class, + 'is_infer': is_infer, + 'num_samples': num_samples +} define_py_data_sources2( - "train.list", None, module="provider", obj="process", args=args) + "train.list" if not is_infer else None, + "test.list" if is_infer else None, + module="provider", + obj="process", + args=args) settings( batch_size=batch_size, @@ -16,6 +30,8 @@ settings( learning_method=MomentumOptimizer(0.9), regularization=L2Regularization(0.0005 * batch_size)) +conv_projection = conv_projection if use_gpu else img_conv_layer + def inception2(name, input, channels, \ filter1, filter3R, filter3, @@ -138,12 +154,11 @@ def inception(name, input, channels, \ cat = concat_layer( name=name, input=[cov1, cov3, cov5, covprj], - bias_attr=True, + bias_attr=True if use_gpu else False, act=ReluActivation()) return cat -lab = data_layer(name="label", size=1000) data = data_layer(name="input", size=3 * height * width) # stage 1 @@ -221,6 +236,10 @@ pool5 = img_pool_layer( dropout = dropout_layer(name="dropout", input=pool5, dropout_rate=0.4) out3 = fc_layer( name="output3", input=dropout, size=1000, act=SoftmaxActivation()) -loss3 = cross_entropy(name='loss3', input=out3, label=lab) -outputs(loss3) +if is_infer: + outputs(out3) +else: + lab = data_layer(name="label", size=num_class) + loss3 = cross_entropy(name='loss3', input=out3, label=lab) + outputs(loss3) diff --git a/benchmark/paddle/image/provider.py b/benchmark/paddle/image/provider.py index 4703944c8722552d56ba80a8e0663de5fb4df53d..1018ec9ce1e529f618ddd7b7afa72a84c5e876a1 100644 --- a/benchmark/paddle/image/provider.py +++ b/benchmark/paddle/image/provider.py @@ -13,14 +13,21 @@ def initHook(settings, height, width, color, num_class, **kwargs): settings.data_size = settings.height * settings.width * 3 else: settings.data_size = settings.height * settings.width - - settings.slots = [dense_vector(settings.data_size), integer_value(1)] + settings.is_infer = kwargs.get('is_infer', False) + settings.num_samples = kwargs.get('num_samples', 2560) + if settings.is_infer: + settings.slots = [dense_vector(settings.data_size)] + else: + settings.slots = [dense_vector(settings.data_size), integer_value(1)] @provider( init_hook=initHook, min_pool_size=-1, cache=CacheType.CACHE_PASS_IN_MEM) def process(settings, file_list): - for i in xrange(1024): + for i in xrange(settings.num_samples): img = np.random.rand(1, settings.data_size).reshape(-1, 1).flatten() - lab = random.randint(0, settings.num_class - 1) - yield img.astype('float32'), int(lab) + if settings.is_infer: + yield img.astype('float32') + else: + lab = random.randint(0, settings.num_class - 1) + yield img.astype('float32'), int(lab) diff --git a/benchmark/paddle/image/resnet.py b/benchmark/paddle/image/resnet.py index 6ae1857642e8df4b3859eec68a3a5227d1c4fcb3..2846e4763f1cda4602f03af5ec649d57ee6cf0d8 100644 --- a/benchmark/paddle/image/resnet.py +++ b/benchmark/paddle/image/resnet.py @@ -6,11 +6,23 @@ width = 224 num_class = 1000 batch_size = get_config_arg('batch_size', int, 64) layer_num = get_config_arg("layer_num", int, 50) -is_test = get_config_arg("is_test", bool, False) - -args = {'height': height, 'width': width, 'color': True, 'num_class': num_class} +is_infer = get_config_arg("is_infer", bool, False) +num_samples = get_config_arg('num_samples', int, 2560) + +args = { + 'height': height, + 'width': width, + 'color': True, + 'num_class': num_class, + 'is_infer': is_infer, + 'num_samples': num_samples +} define_py_data_sources2( - "train.list", None, module="provider", obj="process", args=args) + "train.list" if not is_infer else None, + "test.list" if is_infer else None, + module="provider", + obj="process", + args=args) settings( batch_size=batch_size, @@ -45,7 +57,10 @@ def conv_bn_layer(name, act=LinearActivation(), bias_attr=False) return batch_norm_layer( - name=name + "_bn", input=tmp, act=active_type, use_global_stats=is_test) + name=name + "_bn", + input=tmp, + act=active_type, + use_global_stats=is_infer) def bottleneck_block(name, input, num_filters1, num_filters2): @@ -207,7 +222,9 @@ elif layer_num == 152: else: print("Wrong layer number.") -lbl = data_layer(name="label", size=num_class) -loss = cross_entropy(name='loss', input=resnet, label=lbl) -inputs(img, lbl) -outputs(loss) +if is_infer: + outputs(resnet) +else: + lbl = data_layer(name="label", size=num_class) + loss = cross_entropy(name='loss', input=resnet, label=lbl) + outputs(loss) diff --git a/benchmark/paddle/image/run_mkl_infer.sh b/benchmark/paddle/image/run_mkl_infer.sh new file mode 100755 index 0000000000000000000000000000000000000000..62c9bf6efd3810f506fd4592b2ba3a21b1b7f0e7 --- /dev/null +++ b/benchmark/paddle/image/run_mkl_infer.sh @@ -0,0 +1,87 @@ +set -e + +function clock_to_seconds() { + hours=`echo $1 | awk -F ':' '{print $1}'` + mins=`echo $1 | awk -F ':' '{print $2}'` + secs=`echo $1 | awk -F ':' '{print $3}'` + echo `awk 'BEGIN{printf "%.2f",('$secs' + '$mins' * 60 + '$hours' * 3600)}'` +} + +function infer() { + unset OMP_NUM_THREADS MKL_NUM_THREADS OMP_DYNAMIC KMP_AFFINITY + topology=$1 + layer_num=$2 + bs=$3 + use_mkldnn=$4 + if [ $4 == "True" ]; then + thread=1 + log="logs/infer-${topology}-${layer_num}-mkldnn-${bs}.log" + elif [ $4 == "False" ]; then + thread=`nproc` + if [ $thread -gt $bs ]; then + thread=$bs + fi + log="logs/infer-${topology}-${layer_num}-${thread}mklml-${bs}.log" + else + echo "Wrong input $4, use True or False." + exit 0 + fi + + models_in="models/${topology}-${layer_num}/pass-00000/" + if [ ! -d $models_in ]; then + echo "Training model ${topology}_${layer_num}" + paddle train --job=train \ + --config="${topology}.py" \ + --use_mkldnn=True \ + --use_gpu=False \ + --trainer_count=1 \ + --num_passes=1 \ + --save_dir="models/${topology}-${layer_num}" \ + --config_args="batch_size=128,layer_num=${layer_num},num_samples=256" \ + > /dev/null 2>&1 + echo "Done" + fi + log_period=$((256 / bs)) + paddle train --job=test \ + --config="${topology}.py" \ + --use_mkldnn=$use_mkldnn \ + --use_gpu=False \ + --trainer_count=$thread \ + --log_period=$log_period \ + --config_args="batch_size=${bs},layer_num=${layer_num},is_infer=True" \ + --init_model_path=$models_in \ + 2>&1 | tee ${log} + + # calculate the last 5 logs period time of 1280 samples, + # the time before are burning time. + start=`tail ${log} -n 7 | head -n 1 | awk -F ' ' '{print $2}' | xargs` + end=`tail ${log} -n 2 | head -n 1 | awk -F ' ' '{print $2}' | xargs` + start_sec=`clock_to_seconds $start` + end_sec=`clock_to_seconds $end` + fps=`awk 'BEGIN{printf "%.2f",(1280 / ('$end_sec' - '$start_sec'))}'` + echo "Last 1280 samples start: ${start}(${start_sec} sec), end: ${end}(${end_sec} sec;" >> ${log} + echo "FPS: $fps images/sec" 2>&1 | tee -a ${log} +} + +if [ ! -f "train.list" ]; then + echo " " > train.list +fi +if [ ! -f "test.list" ]; then + echo " " > test.list +fi +if [ ! -d "logs" ]; then + mkdir logs +fi +if [ ! -d "models" ]; then + mkdir -p models +fi + +# inference benchmark +for use_mkldnn in True False; do + for batchsize in 1 2 4 8 16; do + infer vgg 19 $batchsize $use_mkldnn + infer resnet 50 $batchsize $use_mkldnn + infer googlenet v1 $batchsize $use_mkldnn + infer alexnet 2 $batchsize $use_mkldnn + done +done diff --git a/benchmark/paddle/image/run_mkldnn.sh b/benchmark/paddle/image/run_mkl_train.sh similarity index 59% rename from benchmark/paddle/image/run_mkldnn.sh rename to benchmark/paddle/image/run_mkl_train.sh index 3cc779b48d082985f75ab1c053fbe262bc6d58aa..03d2d378fb72e36f765d89af788f6ee96fe21d4e 100755 --- a/benchmark/paddle/image/run_mkldnn.sh +++ b/benchmark/paddle/image/run_mkl_train.sh @@ -8,13 +8,13 @@ function train() { use_mkldnn=$4 if [ $4 == "True" ]; then thread=1 - log="logs/${topology}-${layer_num}-mkldnn-${bs}.log" + log="logs/train-${topology}-${layer_num}-mkldnn-${bs}.log" elif [ $4 == "False" ]; then thread=`nproc` # each trainer_count use only 1 core to avoid conflict - log="logs/${topology}-${layer_num}-${thread}mklml-${bs}.log" + log="logs/train-${topology}-${layer_num}-${thread}mklml-${bs}.log" else - echo "Wrong input $3, use True or False." + echo "Wrong input $4, use True or False." exit 0 fi args="batch_size=${bs},layer_num=${layer_num}" @@ -28,18 +28,25 @@ function train() { --test_period=100 \ --config_args=$args \ 2>&1 | tee ${log} + + avg_time=`tail ${log} -n 1 | awk -F ' ' '{print $8}' | sed 's/avg=//'` + fps=`awk 'BEGIN{printf "%.2f",('$bs' / '$avg_time' * 1000)}'` + echo "FPS: $fps images/sec" 2>&1 | tee -a ${log} } -if [ ! -d "train.list" ]; then +if [ ! -f "train.list" ]; then echo " " > train.list fi if [ ! -d "logs" ]; then mkdir logs fi +# training benchmark for use_mkldnn in True False; do for batchsize in 64 128 256; do train vgg 19 $batchsize $use_mkldnn - train resnet 50 $batchsize $use_mkldnn + train resnet 50 $batchsize $use_mkldnn + train googlenet v1 $batchsize $use_mkldnn + train alexnet 2 $batchsize $use_mkldnn done done diff --git a/benchmark/paddle/image/run_openblas_infer.sh b/benchmark/paddle/image/run_openblas_infer.sh new file mode 100755 index 0000000000000000000000000000000000000000..71a49231a5527ebee9f45d5f4650ce2a4f6a1c31 --- /dev/null +++ b/benchmark/paddle/image/run_openblas_infer.sh @@ -0,0 +1,68 @@ +set -e + +function clock_to_seconds() { + hours=`echo $1 | awk -F ':' '{print $1}'` + mins=`echo $1 | awk -F ':' '{print $2}'` + secs=`echo $1 | awk -F ':' '{print $3}'` + echo `awk 'BEGIN{printf "%.2f",('$secs' + '$mins' * 60 + '$hours' * 3600)}'` +} + +function infer() { + topology=$1 + layer_num=$2 + bs=$3 + trainers=`nproc` + if [ $trainers -gt $bs ]; then + trainers=$bs + fi + log="logs/infer-${topology}-${layer_num}-${trainers}openblas-${bs}.log" + threads=$((`nproc` / trainers)) + if [ $threads -eq 0 ]; then + threads=1 + fi + export OPENBLAS_NUM_THREADS=$threads + + models_in="models/${topology}-${layer_num}/pass-00000/" + if [ ! -d $models_in ]; then + echo "./run_mkl_infer.sh to save the model first" + exit 0 + fi + log_period=$((32 / bs)) + paddle train --job=test \ + --config="${topology}.py" \ + --use_mkldnn=False \ + --use_gpu=False \ + --trainer_count=$trainers \ + --log_period=$log_period \ + --config_args="batch_size=${bs},layer_num=${layer_num},is_infer=True,num_samples=256" \ + --init_model_path=$models_in \ + 2>&1 | tee ${log} + + # calculate the last 5 logs period time of 160(=32*5) samples, + # the time before are burning time. + start=`tail ${log} -n 7 | head -n 1 | awk -F ' ' '{print $2}' | xargs` + end=`tail ${log} -n 2 | head -n 1 | awk -F ' ' '{print $2}' | xargs` + start_sec=`clock_to_seconds $start` + end_sec=`clock_to_seconds $end` + fps=`awk 'BEGIN{printf "%.2f",(160 / ('$end_sec' - '$start_sec'))}'` + echo "Last 160 samples start: ${start}(${start_sec} sec), end: ${end}(${end_sec} sec;" >> ${log} + echo "FPS: $fps images/sec" 2>&1 | tee -a ${log} +} + +if [ ! -f "train.list" ]; then + echo " " > train.list +fi +if [ ! -f "test.list" ]; then + echo " " > test.list +fi +if [ ! -d "logs" ]; then + mkdir logs +fi + +# inference benchmark +for batchsize in 1 2 4 8 16; do + infer vgg 19 $batchsize + infer resnet 50 $batchsize + infer googlenet v1 $batchsize + infer alexnet 2 $batchsize +done diff --git a/benchmark/paddle/image/run_openblas_train.sh b/benchmark/paddle/image/run_openblas_train.sh new file mode 100755 index 0000000000000000000000000000000000000000..935cff6f2c97d25d6de556cfee25e27dbe49b5b6 --- /dev/null +++ b/benchmark/paddle/image/run_openblas_train.sh @@ -0,0 +1,41 @@ +set -e + +function train() { + export OPENBLAS_NUM_THREADS=1 + topology=$1 + layer_num=$2 + bs=$3 + thread=`nproc` + # each trainer_count use only 1 core to avoid conflict + log="logs/train-${topology}-${layer_num}-${thread}openblas-${bs}.log" + args="batch_size=${bs},layer_num=${layer_num}" + config="${topology}.py" + paddle train --job=time \ + --config=$config \ + --use_mkldnn=False \ + --use_gpu=False \ + --trainer_count=$thread \ + --log_period=3 \ + --test_period=30 \ + --config_args=$args \ + 2>&1 | tee ${log} + + avg_time=`tail ${log} -n 1 | awk -F ' ' '{print $8}' | sed 's/avg=//'` + fps=`awk 'BEGIN{printf "%.2f",('$bs' / '$avg_time' * 1000)}'` + echo "FPS: $fps images/sec" 2>&1 | tee -a ${log} +} + +if [ ! -f "train.list" ]; then + echo " " > train.list +fi +if [ ! -d "logs" ]; then + mkdir logs +fi + +# training benchmark +for batchsize in 64 128 256; do + train vgg 19 $batchsize + train resnet 50 $batchsize + train googlenet v1 $batchsize + train alexnet 2 $batchsize +done diff --git a/benchmark/paddle/image/vgg.py b/benchmark/paddle/image/vgg.py index 420884ed8e1ae36a3f1772bfbe8323f3d0ea71e6..ca0a6798fb8c35b68cf84d263855955eb93ba0b0 100644 --- a/benchmark/paddle/image/vgg.py +++ b/benchmark/paddle/image/vgg.py @@ -6,10 +6,23 @@ width = 224 num_class = 1000 batch_size = get_config_arg('batch_size', int, 64) layer_num = get_config_arg('layer_num', int, 19) +is_infer = get_config_arg("is_infer", bool, False) +num_samples = get_config_arg('num_samples', int, 2560) -args = {'height': height, 'width': width, 'color': True, 'num_class': num_class} +args = { + 'height': height, + 'width': width, + 'color': True, + 'num_class': num_class, + 'is_infer': is_infer, + 'num_samples': num_samples +} define_py_data_sources2( - "train.list", None, module="provider", obj="process", args=args) + "train.list" if not is_infer else None, + "test.list" if is_infer else None, + module="provider", + obj="process", + args=args) settings( batch_size=batch_size, @@ -98,6 +111,9 @@ elif layer_num == 19: else: print("Wrong layer number.") -lab = data_layer('label', num_class) -loss = cross_entropy(input=vgg, label=lab) -outputs(loss) +if is_infer: + outputs(vgg) +else: + lab = data_layer('label', num_class) + loss = cross_entropy(input=vgg, label=lab) + outputs(loss) diff --git a/cmake/cblas.cmake b/cmake/cblas.cmake index b21fc43904d9aafe9f7d019dfbe5b1c0d3f9e2d6..6320b17520a687f88993b6f464d9115838b0f96b 100644 --- a/cmake/cblas.cmake +++ b/cmake/cblas.cmake @@ -3,7 +3,7 @@ # It will search MKLML, atlas, OpenBlas, reference-cblas in order. # # If any cblas implementation found, the following variable will be set. -# CBLAS_PROVIDER # one of MKLML, ATLAS, OPENBLAS, REFERENCE +# CBLAS_PROVIDER # one of MKLML, OPENBLAS, REFERENCE # CBLAS_INC_DIR # the include directory for cblas. # CBLAS_LIBS # a list of libraries should be linked by paddle. # # Each library should be full path to object file. @@ -17,7 +17,7 @@ if(WITH_MKLML AND MKLML_INC_DIR AND MKLML_LIB) set(CBLAS_INC_DIR ${MKLML_INC_DIR}) set(CBLAS_LIBRARIES ${MKLML_LIB}) - add_definitions(-DPADDLE_USE_MKLML) + add_definitions(-DPADDLE_WITH_MKLML) add_definitions(-DLAPACK_FOUND) message(STATUS "Found cblas and lapack in MKLML " @@ -25,42 +25,6 @@ if(WITH_MKLML AND MKLML_INC_DIR AND MKLML_LIB) return() endif() -## Then find atlas. -set(ATLAS_ROOT $ENV{ATLAS_ROOT} CACHE PATH "Folder contains Atlas") -set(ATLAS_INCLUDE_SEARCH_PATHS - ${ATLAS_ROOT}/include - /usr/include - /usr/include/atlas) -set(ATLAS_LIB_SEARCH_PATHS - ${ATLAS_ROOT}/lib - /usr/lib - /usr/lib/blas/atlas - /usr/lib/atlas - /usr/lib/atlas-base # special for ubuntu 14.04. - ) -find_path(ATLAS_INC_DIR NAMES cblas.h - PATHS ${ATLAS_INCLUDE_SEARCH_PATHS}) -find_path(ATLAS_CLAPACK_INC_DIR NAMES clapack.h - PATHS ${ATLAS_INCLUDE_SEARCH_PATHS}) -find_library(ATLAS_CBLAS_LIB NAMES cblas libcblas.so.3 - PATHS ${ATLAS_LIB_SEARCH_PATHS}) -find_library(ATLAS_CLAPACK_LIB NAMES lapack_atlas liblapack_atlas.so.3 - PATHS ${ATLAS_LIB_SEARCH_PATHS}) - -if(ATLAS_CLAPACK_INC_DIR AND ATLAS_INC_DIR AND ATLAS_CBLAS_LIB AND ATLAS_CLAPACK_LIB) - set(CBLAS_FOUND ON) - set(CBLAS_PROVIDER ATLAS) - set(CBLAS_INC_DIR ${ATLAS_INC_DIR} ${ATLAS_CLAPACK_INC_DIR}) - set(CBLAS_LIBRARIES ${ATLAS_CLAPACK_LIB} ${ATLAS_CBLAS_LIB}) - - add_definitions(-DPADDLE_USE_ATLAS) - add_definitions(-DLAPACK_FOUND) - - message(STATUS "Found ATLAS (include: ${ATLAS_INC_DIR}, library: ${CBLAS_LIBRARIES})") - message(STATUS "Found lapack in ATLAS (include: ${ATLAS_CLAPACK_INC_DIR})") - return() -endif() - ## Then find openblas. set(OPENBLAS_ROOT $ENV{OPENBLAS_ROOT} CACHE PATH "Folder contains Openblas") set(OPENBLAS_INCLUDE_SEARCH_PATHS diff --git a/cmake/configure.cmake b/cmake/configure.cmake index e550ec285668ea25757eeee9e7c5dc48fc9d339d..5c6bcfde76a1201f792d04766d698db8cd395a49 100644 --- a/cmake/configure.cmake +++ b/cmake/configure.cmake @@ -24,6 +24,11 @@ if(WITH_DOUBLE) add_definitions(-DPADDLE_TYPE_DOUBLE) endif(WITH_DOUBLE) +if(WITH_ARM_FP16) + add_definitions(-DPADDLE_ARM_FP16) + add_definitions("-march=armv8.2-a+fp16+simd") +endif(WITH_ARM_FP16) + if(WITH_TESTING) add_definitions(-DPADDLE_WITH_TESTING) endif(WITH_TESTING) diff --git a/cmake/external/cares.cmake b/cmake/external/cares.cmake new file mode 100644 index 0000000000000000000000000000000000000000..aec51410b33669f8a549f2eca193cc6aa2d07a13 --- /dev/null +++ b/cmake/external/cares.cmake @@ -0,0 +1,45 @@ +# 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. +# + +IF(MOBILE_INFERENCE OR NOT WITH_DISTRIBUTE) + return() +ENDIF() + +include (ExternalProject) + +# NOTE: c-ares is needed when linking with grpc. + +SET(CARES_SOURCES_DIR ${THIRD_PARTY_PATH}/cares) +SET(CARES_INSTALL_DIR ${THIRD_PARTY_PATH}/install/cares) +SET(CARES_INCLUDE_DIR "${CARES_INSTALL_DIR}/include/" CACHE PATH "cares include directory." FORCE) + +ExternalProject_Add( + extern_cares + GIT_REPOSITORY "https://github.com/c-ares/c-ares.git" + GIT_TAG "cares-1_13_0" + PREFIX ${CARES_SOURCES_DIR} + UPDATE_COMMAND "" + CONFIGURE_COMMAND ./buildconf && ./configure --disable-shared --prefix=${CARES_INSTALL_DIR} + BUILD_IN_SOURCE 1 + BUILD_COMMAND make -j8 + INSTALL_COMMAND make install +) + +ADD_LIBRARY(cares STATIC IMPORTED GLOBAL) +SET_PROPERTY(TARGET cares PROPERTY IMPORTED_LOCATION + "${CARES_INSTALL_DIR}/lib/libcares.a") + +include_directories(${CARES_INCLUDE_DIR}) +ADD_DEPENDENCIES(cares extern_cares) diff --git a/cmake/external/gflags.cmake b/cmake/external/gflags.cmake index c819eb4d70898e48eab499c666168d78262d4240..d4f252bb9f64c8db82b841fedf0817f5d8596501 100644 --- a/cmake/external/gflags.cmake +++ b/cmake/external/gflags.cmake @@ -28,15 +28,8 @@ INCLUDE_DIRECTORIES(${GFLAGS_INCLUDE_DIR}) ExternalProject_Add( extern_gflags ${EXTERNAL_PROJECT_LOG_ARGS} - # TODO(yiwang): The annoying warnings mentioned in - # https://github.com/PaddlePaddle/Paddle/issues/3277 are caused by - # gflags. I fired a PR https://github.com/gflags/gflags/pull/230 - # to fix it. Before it gets accepted by the gflags team, we use - # my personal fork, which contains above fix, temporarily. Let's - # change this back to the official Github repo once my PR is - # merged. - GIT_REPOSITORY "https://github.com/wangkuiyi/gflags.git" - GIT_TAG 986964c07427ecb9cdb5bd73f73ebbd40e54dadb + GIT_REPOSITORY "https://github.com/gflags/gflags.git" + GIT_TAG 77592648e3f3be87d6c7123eb81cbad75f9aef5a PREFIX ${GFLAGS_SOURCES_DIR} UPDATE_COMMAND "" CMAKE_ARGS -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER} diff --git a/cmake/external/glog.cmake b/cmake/external/glog.cmake index 08bdc1e1623b0d917061c7368e9b2a8f7e9517fd..0c6b3aafcb4e990b9d4549820137474e5968a7aa 100644 --- a/cmake/external/glog.cmake +++ b/cmake/external/glog.cmake @@ -26,12 +26,21 @@ ENDIF(WIN32) INCLUDE_DIRECTORIES(${GLOG_INCLUDE_DIR}) +IF(ANDROID AND ${CMAKE_SYSTEM_VERSION} VERSION_LESS "21") + # Using the unofficial glog for Android API < 21 + SET(GLOG_REPOSITORY "https://github.com/Xreki/glog.git") + SET(GLOG_TAG "8a547150548b284382ccb6582408e9140ff2bea8") +ELSE() + SET(GLOG_REPOSITORY "https://github.com/google/glog.git") + SET(GLOG_TAG "v0.3.5") +ENDIF() + ExternalProject_Add( extern_glog ${EXTERNAL_PROJECT_LOG_ARGS} DEPENDS gflags - GIT_REPOSITORY "https://github.com/google/glog.git" - GIT_TAG v0.3.5 + GIT_REPOSITORY ${GLOG_REPOSITORY} + GIT_TAG ${GLOG_TAG} PREFIX ${GLOG_SOURCES_DIR} UPDATE_COMMAND "" CMAKE_ARGS -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER} diff --git a/cmake/external/grpc.cmake b/cmake/external/grpc.cmake new file mode 100644 index 0000000000000000000000000000000000000000..abee6698e30b7e76ca42825ed225876bf2ba5ec0 --- /dev/null +++ b/cmake/external/grpc.cmake @@ -0,0 +1,66 @@ +# 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. +# + +IF(MOBILE_INFERENCE OR NOT WITH_DISTRIBUTE) + return() +ENDIF() + +include (ExternalProject) + +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) +IF(APPLE) + SET(BUILD_CMD make -n HAS_SYSTEM_PROTOBUF=false -s -j8 static grpc_cpp_plugin | sed "s/-Werror//g" | sh) +ELSE() + SET(BUILD_CMD make HAS_SYSTEM_PROTOBUF=false -s -j8 static grpc_cpp_plugin) +ENDIF() + +ExternalProject_Add( + extern_grpc + DEPENDS protobuf zlib + GIT_REPOSITORY "https://github.com/grpc/grpc.git" + GIT_TAG "v1.7.x" + PREFIX ${GRPC_SOURCES_DIR} + UPDATE_COMMAND "" + CONFIGURE_COMMAND "" + BUILD_IN_SOURCE 1 + # NOTE(yuyang18): + # Disable -Werror, otherwise the compile will fail in MacOS. + # It seems that we cannot configure that by make command. + # Just dry run make command and remove `-Werror`, then use a shell to run make commands + BUILD_COMMAND ${BUILD_CMD} + INSTALL_COMMAND make prefix=${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 + "${GRPC_INSTALL_DIR}/lib/libgrpc++_unsecure.a") + +ADD_LIBRARY(grpc++ STATIC IMPORTED GLOBAL) +SET_PROPERTY(TARGET grpc++ PROPERTY IMPORTED_LOCATION + "${GRPC_INSTALL_DIR}/lib/libgrpc++.a") +ADD_LIBRARY(gpr STATIC IMPORTED GLOBAL) +SET_PROPERTY(TARGET gpr PROPERTY IMPORTED_LOCATION + "${GRPC_INSTALL_DIR}/lib/libgpr.a") + +ADD_LIBRARY(grpc_unsecure STATIC IMPORTED GLOBAL) +SET_PROPERTY(TARGET grpc_unsecure PROPERTY IMPORTED_LOCATION + "${GRPC_INSTALL_DIR}/lib/libgrpc_unsecure.a") + +include_directories(${GRPC_INCLUDE_DIR}) +ADD_DEPENDENCIES(grpc++_unsecure extern_grpc) diff --git a/cmake/external/mkldnn.cmake b/cmake/external/mkldnn.cmake index fc52d339d7a336b44c97f2e0a9fc8d6604854365..5d24caebdcc5a28823164d718fb1628be5c4179d 100644 --- a/cmake/external/mkldnn.cmake +++ b/cmake/external/mkldnn.cmake @@ -67,5 +67,5 @@ ADD_LIBRARY(mkldnn SHARED IMPORTED GLOBAL) SET_PROPERTY(TARGET mkldnn PROPERTY IMPORTED_LOCATION ${MKLDNN_LIB}) ADD_DEPENDENCIES(mkldnn ${MKLDNN_PROJECT}) MESSAGE(STATUS "MKLDNN library: ${MKLDNN_LIB}") -add_definitions(-DPADDLE_USE_MKLDNN) +add_definitions(-DPADDLE_WITH_MKLDNN) LIST(APPEND external_project_dependencies mkldnn) diff --git a/cmake/external/openblas.cmake b/cmake/external/openblas.cmake index 4c4f59656dae68739f2f07f3febd510e727fe2dd..97857a686b38d935b19f510ecdcb66bcca91fe03 100644 --- a/cmake/external/openblas.cmake +++ b/cmake/external/openblas.cmake @@ -114,11 +114,7 @@ INCLUDE_DIRECTORIES(${CBLAS_INC_DIR}) # linear algebra libraries for cc_library(xxx SRCS xxx.c DEPS cblas) SET(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/cblas_dummy.c) FILE(WRITE ${dummyfile} "const char * dummy = \"${dummyfile}\";") -IF("${CBLAS_PROVIDER}" STREQUAL "MKLML") - ADD_LIBRARY(cblas SHARED ${dummyfile}) -ELSE() - ADD_LIBRARY(cblas STATIC ${dummyfile}) -ENDIF() +ADD_LIBRARY(cblas STATIC ${dummyfile}) TARGET_LINK_LIBRARIES(cblas ${CBLAS_LIBRARIES}) IF(NOT ${CBLAS_FOUND}) diff --git a/cmake/external/protobuf.cmake b/cmake/external/protobuf.cmake index be7f6a9465970711170bd15dcecaadeaa8a55f86..ff5855052dabaa0b63099cd219f3f04e22f1aa85 100644 --- a/cmake/external/protobuf.cmake +++ b/cmake/external/protobuf.cmake @@ -15,7 +15,18 @@ INCLUDE(ExternalProject) # Always invoke `FIND_PACKAGE(Protobuf)` for importing function protobuf_generate_cpp FIND_PACKAGE(Protobuf QUIET) -SET(PROTOBUF_FOUND "OFF") +macro(UNSET_VAR VAR_NAME) + UNSET(${VAR_NAME} CACHE) + UNSET(${VAR_NAME}) +endmacro() +UNSET_VAR(PROTOBUF_INCLUDE_DIR) +UNSET_VAR(PROTOBUF_FOUND) +UNSET_VAR(PROTOBUF_PROTOC_EXECUTABLE) +UNSET_VAR(PROTOBUF_PROTOC_LIBRARY) +UNSET_VAR(PROTOBUF_LITE_LIBRARY) +UNSET_VAR(PROTOBUF_LIBRARY) +UNSET_VAR(PROTOBUF_INCLUDE_DIR) +UNSET_VAR(Protobuf_PROTOC_EXECUTABLE) if(NOT COMMAND protobuf_generate_python) # before cmake 3.4, protobuf_genrerate_python is not defined. function(protobuf_generate_python SRCS) @@ -110,7 +121,6 @@ macro(PROMPT_PROTOBUF_LIB) # FIND_Protobuf.cmake uses `Protobuf_PROTOC_EXECUTABLE`. # make `protobuf_generate_cpp` happy. SET(Protobuf_PROTOC_EXECUTABLE ${PROTOBUF_PROTOC_EXECUTABLE}) - FOREACH(dep ${protobuf_DEPS}) ADD_DEPENDENCIES(protobuf ${dep}) ADD_DEPENDENCIES(protobuf_lite ${dep}) @@ -128,11 +138,11 @@ endmacro() set(PROTOBUF_ROOT "" CACHE PATH "Folder contains protobuf") if (NOT "${PROTOBUF_ROOT}" STREQUAL "") - find_path(PROTOBUF_INCLUDE_DIR google/protobuf/message.h PATHS ${PROTOBUF_ROOT}/include) - find_library(PROTOBUF_LIBRARY protobuf PATHS ${PROTOBUF_ROOT}/lib) - find_library(PROTOBUF_LITE_LIBRARY protobuf-lite PATHS ${PROTOBUF_ROOT}/lib) - find_library(PROTOBUF_PROTOC_LIBRARY protoc PATHS ${PROTOBUF_ROOT}/lib) - find_program(PROTOBUF_PROTOC_EXECUTABLE protoc PATHS ${PROTOBUF_ROOT}/bin) + find_path(PROTOBUF_INCLUDE_DIR google/protobuf/message.h PATHS ${PROTOBUF_ROOT}/include NO_DEFAULT_PATH) + find_library(PROTOBUF_LIBRARY protobuf PATHS ${PROTOBUF_ROOT}/lib NO_DEFAULT_PATH) + find_library(PROTOBUF_LITE_LIBRARY protobuf-lite PATHS ${PROTOBUF_ROOT}/lib NO_DEFAULT_PATH) + find_library(PROTOBUF_PROTOC_LIBRARY protoc PATHS ${PROTOBUF_ROOT}/lib NO_DEFAULT_PATH) + find_program(PROTOBUF_PROTOC_EXECUTABLE protoc PATHS ${PROTOBUF_ROOT}/bin NO_DEFAULT_PATH) if (PROTOBUF_INCLUDE_DIR AND PROTOBUF_LIBRARY AND PROTOBUF_LITE_LIBRARY AND PROTOBUF_PROTOC_LIBRARY AND PROTOBUF_PROTOC_EXECUTABLE) message(STATUS "Using custom protobuf library in ${PROTOBUF_ROOT}.") SET_PROTOBUF_VERSION() @@ -178,14 +188,26 @@ FUNCTION(build_protobuf TARGET_NAME BUILD_FOR_HOST) SET(OPTIONAL_CACHE_ARGS "-DZLIB_ROOT:STRING=${ZLIB_ROOT}") ENDIF() + SET(PROTOBUF_REPO "https://github.com/google/protobuf.git") + SET(PROTOBUF_TAG "9f75c5aa851cd877fb0d93ccc31b8567a6706546") + IF(MOBILE_INFERENCE) + # The reason why the official version is not used is described in + # https://github.com/PaddlePaddle/Paddle/issues/6114 + SET(PROTOBUF_REPO "https://github.com/qingqing01/protobuf.git") + SET(PROTOBUF_TAG "v3.2.0") + IF(NOT BUILD_FOR_HOST) + SET(OPTIONAL_ARGS ${OPTIONAL_ARGS} "-Dprotobuf_BUILD_PROTOC_BINARIES=OFF") + ENDIF() + ENDIF() + ExternalProject_Add( ${TARGET_NAME} ${EXTERNAL_PROJECT_LOG_ARGS} PREFIX ${PROTOBUF_SOURCES_DIR} UPDATE_COMMAND "" DEPENDS zlib - GIT_REPOSITORY "https://github.com/google/protobuf.git" - GIT_TAG "9f75c5aa851cd877fb0d93ccc31b8567a6706546" + GIT_REPOSITORY ${PROTOBUF_REPO} + GIT_TAG ${PROTOBUF_TAG} CONFIGURE_COMMAND ${CMAKE_COMMAND} ${PROTOBUF_SOURCES_DIR}/src/${TARGET_NAME}/cmake ${OPTIONAL_ARGS} @@ -203,7 +225,11 @@ FUNCTION(build_protobuf TARGET_NAME BUILD_FOR_HOST) ) ENDFUNCTION() -SET(PROTOBUF_VERSION 3.1) +IF(NOT MOBILE_INFERENCE) + SET(PROTOBUF_VERSION 3.1) +ELSE() + SET(PROTOBUF_VERSION 3.2) +ENDIF() IF(CMAKE_CROSSCOMPILING) build_protobuf(protobuf_host TRUE) LIST(APPEND external_project_dependencies protobuf_host) @@ -227,9 +253,9 @@ IF(NOT PROTOBUF_FOUND) IF(WITH_C_API) INSTALL(DIRECTORY ${PROTOBUF_INCLUDE_DIR} DESTINATION third_party/protobuf) IF(ANDROID) - INSTALL(FILES ${PROTOBUF_LIBRARY} DESTINATION third_party/protobuf/lib/${ANDROID_ABI}) + INSTALL(FILES ${PROTOBUF_LITE_LIBRARY} DESTINATION third_party/protobuf/lib/${ANDROID_ABI}) ELSE() - INSTALL(FILES ${PROTOBUF_LIBRARY} DESTINATION third_party/protobuf/lib) + INSTALL(FILES ${PROTOBUF_LITE_LIBRARY} DESTINATION third_party/protobuf/lib) ENDIF() ENDIF() diff --git a/cmake/external/zlib.cmake b/cmake/external/zlib.cmake index a98e069b7cd1654ddd5868560d0905eab6d9c692..1638cd8fdfc34575132462859e056a1907f0b2f1 100644 --- a/cmake/external/zlib.cmake +++ b/cmake/external/zlib.cmake @@ -50,6 +50,8 @@ ExternalProject_Add( ) LIST(APPEND external_project_dependencies zlib) +ADD_LIBRARY(zlib_target STATIC IMPORTED GLOBAL) +SET_PROPERTY(TARGET zlib_target PROPERTY IMPORTED_LOCATION ${ZLIB_LIBRARIES}) IF(WITH_C_API) INSTALL(DIRECTORY ${ZLIB_INCLUDE_DIR} DESTINATION third_party/zlib) diff --git a/cmake/flags.cmake b/cmake/flags.cmake index 2b125cef6aa8d1021afe8a7a0d232d84d36be4bc..1120677a37e0d44163816b66600121c8f0d545af 100644 --- a/cmake/flags.cmake +++ b/cmake/flags.cmake @@ -111,6 +111,8 @@ set(COMMON_FLAGS -Wno-error=sign-compare -Wno-error=unused-local-typedefs -Wno-error=parentheses-equality # Warnings in pybind11 + -Wno-error=ignored-attributes # Warnings in Eigen, gcc 6.3 + -Wno-error=terminate # Warning in PADDLE_ENFORCE ) set(GPU_COMMON_FLAGS diff --git a/cmake/generic.cmake b/cmake/generic.cmake index b9c1dde97bc444d793d67ff622fd6b13c6435a9a..66c8e3ad7ef7c80c1f388c25983425a0db5c0220 100644 --- a/cmake/generic.cmake +++ b/cmake/generic.cmake @@ -227,8 +227,8 @@ function(cc_test TARGET_NAME) set(multiValueArgs SRCS DEPS) cmake_parse_arguments(cc_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) add_executable(${TARGET_NAME} ${cc_test_SRCS}) - target_link_libraries(${TARGET_NAME} ${cc_test_DEPS} gtest gtest_main) - add_dependencies(${TARGET_NAME} ${cc_test_DEPS} gtest gtest_main) + target_link_libraries(${TARGET_NAME} ${cc_test_DEPS} paddle_gtest_main paddle_memory gtest gflags) + add_dependencies(${TARGET_NAME} ${cc_test_DEPS} paddle_gtest_main paddle_memory gtest gflags) add_test(NAME ${TARGET_NAME} COMMAND ${TARGET_NAME} WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}) endif() endfunction(cc_test) @@ -288,8 +288,8 @@ function(nv_test TARGET_NAME) set(multiValueArgs SRCS DEPS) cmake_parse_arguments(nv_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) cuda_add_executable(${TARGET_NAME} ${nv_test_SRCS}) - target_link_libraries(${TARGET_NAME} ${nv_test_DEPS} gtest gtest_main) - add_dependencies(${TARGET_NAME} ${nv_test_DEPS} gtest gtest_main) + target_link_libraries(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main paddle_memory gtest gflags) + add_dependencies(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main paddle_memory gtest gflags) add_test(${TARGET_NAME} ${TARGET_NAME}) endif() endfunction(nv_test) @@ -459,11 +459,58 @@ function(py_test TARGET_NAME) if(WITH_TESTING) set(options STATIC static SHARED shared) set(oneValueArgs "") - set(multiValueArgs SRCS DEPS) - cmake_parse_arguments(py_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) + set(multiValueArgs SRCS DEPS ARGS) + cmake_parse_arguments(py_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) add_test(NAME ${TARGET_NAME} COMMAND env PYTHONPATH=${PADDLE_PYTHON_BUILD_DIR}/lib-python - python2 ${py_test_SRCS} + ${PYTHON_EXECUTABLE} -u ${py_test_SRCS} ${py_test_ARGS} WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}) endif() endfunction() + +# grpc_library generate grpc code using grpc_cpp_plugin and protoc +# then build the generated protobuf code and grpc code with your +# implementation source codes together. Use SRCS argument for your +# implementation source files and PROTO argument for your .proto +# files. +# +# Usage: grpc_library(my_target SRCS my_client.cc PROTO my_target.proto DEPS my_dep) + +function(grpc_library TARGET_NAME) + set(oneValueArgs PROTO) + set(multiValueArgs SRCS DEPS) + set(options "") + cmake_parse_arguments(grpc_library "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) + + message(STATUS "generating grpc ${grpc_library_PROTO}") + + get_filename_component(ABS_PROTO ${grpc_library_PROTO} ABSOLUTE) + get_filename_component(PROTO_WE ${grpc_library_PROTO} NAME_WE) + get_filename_component(PROTO_PATH ${ABS_PROTO} PATH) + + protobuf_generate_cpp(grpc_proto_srcs grpc_proto_hdrs "${ABS_PROTO}") + set(grpc_grpc_srcs "${CMAKE_CURRENT_BINARY_DIR}/${PROTO_WE}.grpc.pb.cc") + set(grpc_grpc_hdrs "${CMAKE_CURRENT_BINARY_DIR}/${PROTO_WE}.grpc.pb.h") + cc_library("${TARGET_NAME}_proto" SRCS "${grpc_proto_srcs}") + + add_custom_command( + OUTPUT "${grpc_grpc_srcs}" "${grpc_grpc_hdrs}" + COMMAND ${PROTOBUF_PROTOC_EXECUTABLE} + ARGS --grpc_out "${CMAKE_CURRENT_BINARY_DIR}" -I "${PROTO_PATH}" + --plugin=protoc-gen-grpc="${GRPC_CPP_PLUGIN}" "${ABS_PROTO}" + DEPENDS "${ABS_PROTO}" ${PROTOBUF_PROTOC_EXECUTABLE} extern_grpc) + + # FIXME(typhoonzero): grpc generated code do not generate virtual-dtor, mark it + # as compiler warnings instead of error. Should try remove the warnings also. + set_source_files_properties( + ${grpc_grpc_srcs} + PROPERTIES + COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor") + cc_library("${TARGET_NAME}_grpc" SRCS "${grpc_grpc_srcs}") + + set_source_files_properties( + ${grpc_library_SRCS} + PROPERTIES + COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor") + cc_library("${TARGET_NAME}" SRCS "${grpc_library_SRCS}" DEPS "${TARGET_NAME}_grpc" "${TARGET_NAME}_proto" "${grpc_library_DEPS}") +endfunction() diff --git a/cmake/util.cmake b/cmake/util.cmake index ad905ab55ba3537054fa5b30b5fca4d83c406702..0dc33ce385175d1e2dc454d41db467d4b9d9cf9a 100644 --- a/cmake/util.cmake +++ b/cmake/util.cmake @@ -168,17 +168,3 @@ function(create_resources res_file output_file) COMMAND python ARGS ${PADDLE_SOURCE_DIR}/cmake/make_resource.py ${res_file} ${output_file} DEPENDS ${res_file} ${PADDLE_SOURCE_DIR}/cmake/make_resource.py) endfunction() - - -# Create a python unittest using run_python_tests.sh, -# which takes care of making correct running environment -function(add_python_test TEST_NAME) - foreach(arg ${ARGN}) - get_filename_component(py_fn ${arg} NAME_WE) - set(TRG_NAME ${TEST_NAME}_${py_fn}) - add_test(NAME ${TRG_NAME} - COMMAND env PYTHONPATH=${PADDLE_PYTHON_PACKAGE_DIR} - python2 ${arg} - WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}) - endforeach() -endfunction() diff --git a/doc/api/index_cn.rst b/doc/api/index_cn.rst index 9be0b370ee5e301aee4a6e31b1cfa905754968e8..84f9097a6cdc2da269bd6a0685796e14e26da37e 100644 --- a/doc/api/index_cn.rst +++ b/doc/api/index_cn.rst @@ -7,3 +7,4 @@ API 模型配置 数据访问 训练与应用 + v2/fluid.rst diff --git a/doc/api/index_en.rst b/doc/api/index_en.rst index 25c1dd00b9cbb3ab647e04cdc2b4c27c552a2332..e6f632e1a5b9c4b50b7c6aa96a120030bd6ce338 100644 --- a/doc/api/index_en.rst +++ b/doc/api/index_en.rst @@ -7,3 +7,4 @@ API v2/model_configs.rst v2/data.rst v2/run_logic.rst + v2/fluid.rst diff --git a/doc/api/v2/config/activation.rst b/doc/api/v2/config/activation.rst index eca3ce03bcdc599edca802d8dfca48d4f28275a2..5317e66b64bbd85c61f19700a9d2c1d239dee573 100644 --- a/doc/api/v2/config/activation.rst +++ b/doc/api/v2/config/activation.rst @@ -99,3 +99,10 @@ STanh .. automodule:: paddle.v2.activation :members: STanh :noindex: + +SoftSign +======== + +.. automodule:: paddle.v2.activation + :members: SoftSign + :noindex: diff --git a/doc/api/v2/config/layer.rst b/doc/api/v2/config/layer.rst index d4d182f6692e09b3e40f3620b77d9a0f20ec5af3..ddf0b055a92d80295b24255a5462d477e0d9c796 100644 --- a/doc/api/v2/config/layer.rst +++ b/doc/api/v2/config/layer.rst @@ -54,7 +54,7 @@ img_conv .. _api_v2.layer_context_projection: -context_projection +context_projection ------------------ .. autoclass:: paddle.v2.layer.context_projection :noindex: @@ -70,7 +70,7 @@ Image Pooling Layer img_pool -------- .. autoclass:: paddle.v2.layer.img_pool - :noindex: + :noindex: spp --- @@ -104,7 +104,7 @@ sum_to_one_norm --------------- .. autoclass:: paddle.v2.layer.sum_to_one_norm :noindex: - + cross_channel_norm ------------------ .. autoclass:: paddle.v2.layer.cross_channel_norm @@ -114,7 +114,7 @@ row_l2_norm ----------- .. autoclass:: paddle.v2.layer.row_l2_norm :noindex: - + Recurrent Layers ================ @@ -252,6 +252,11 @@ first_seq .. autoclass:: paddle.v2.layer.first_seq :noindex: +sub_seq +--------- +.. autoclass:: paddle.v2.layer.sub_seq + :noindex: + concat ------ .. autoclass:: paddle.v2.layer.concat @@ -415,6 +420,13 @@ multiplex .. autoclass:: paddle.v2.layer.multiplex :noindex: +Factorization Machine Layer +============================ + +factorization_machine +--------------------- +.. autoclass:: paddle.v2.layer.factorization_machine + :noindex: Slicing and Joining Layers ========================== @@ -460,7 +472,7 @@ lambda_cost :noindex: square_error_cost --------- +----------------- .. autoclass:: paddle.v2.layer.square_error_cost :noindex: @@ -526,7 +538,7 @@ Miscs ===== dropout --------------- +-------- .. autoclass:: paddle.v2.layer.dropout :noindex: diff --git a/doc/api/v2/fluid.rst b/doc/api/v2/fluid.rst new file mode 100644 index 0000000000000000000000000000000000000000..43fc19dc492bbc119f2356034b81c65e443db2fa --- /dev/null +++ b/doc/api/v2/fluid.rst @@ -0,0 +1,18 @@ +====================== +Fluid +====================== + +.. toctree:: + :maxdepth: 1 + + fluid/layers.rst + fluid/data_feeder.rst + fluid/executor.rst + fluid/initializer.rst + fluid/evaluator.rst + fluid/nets.rst + fluid/optimizer.rst + fluid/param_attr.rst + fluid/profiler.rst + fluid/regularizer.rst + diff --git a/doc/api/v2/fluid/data_feeder.rst b/doc/api/v2/fluid/data_feeder.rst new file mode 100644 index 0000000000000000000000000000000000000000..0fa78f7dfb04c13be7eb83b7fd35cb03f2f4a7fa --- /dev/null +++ b/doc/api/v2/fluid/data_feeder.rst @@ -0,0 +1,9 @@ +=========== +DataFeeder +=========== + +DataFeeder +----------- +.. automodule:: paddle.v2.fluid.data_feeder + :members: DataFeeder + :noindex: diff --git a/doc/api/v2/fluid/evaluator.rst b/doc/api/v2/fluid/evaluator.rst new file mode 100644 index 0000000000000000000000000000000000000000..a23f3301d0331e0ea3733f06444515eb4680cd31 --- /dev/null +++ b/doc/api/v2/fluid/evaluator.rst @@ -0,0 +1,9 @@ +=========== +Evaluator +=========== + +Evaluator +----------- +.. automodule:: paddle.v2.fluid.evaluator + :members: Evaluator + :noindex: diff --git a/doc/api/v2/fluid/executor.rst b/doc/api/v2/fluid/executor.rst new file mode 100644 index 0000000000000000000000000000000000000000..3a283538c120cfa1ef646c390bb71c6251c23675 --- /dev/null +++ b/doc/api/v2/fluid/executor.rst @@ -0,0 +1,9 @@ +=========== +Executor +=========== + +Executor +----------- +.. automodule:: paddle.v2.fluid.executor + :members: Executor + :noindex: diff --git a/doc/api/v2/fluid/initializer.rst b/doc/api/v2/fluid/initializer.rst new file mode 100644 index 0000000000000000000000000000000000000000..8f587837e9873370722062404f511654a9460587 --- /dev/null +++ b/doc/api/v2/fluid/initializer.rst @@ -0,0 +1,50 @@ +=========== +Initializer +=========== + + + +Initializer +----------- +.. automodule:: paddle.v2.fluid.initializer + :members: Initializer + :noindex: + + + +ConstantInitializer +------------------- +.. automodule:: paddle.v2.fluid.initializer + :members: ConstantInitializer + :noindex: + + + +UniformInitializer +------------------ +.. automodule:: paddle.v2.fluid.initializer + :members: UniformInitializer + :noindex: + + + +NormalInitializer +----------------- +.. automodule:: paddle.v2.fluid.initializer + :members: NormalInitializer + :noindex: + + +XavierInitializer +----------------- +.. automodule:: paddle.v2.fluid.initializer + :members: XavierInitializer + :noindex: + + +MSRAInitializer +--------------- +.. automodule:: paddle.v2.fluid.initializer + :members: MSRAInitializer + :noindex: + diff --git a/doc/api/v2/fluid/layers.rst b/doc/api/v2/fluid/layers.rst new file mode 100644 index 0000000000000000000000000000000000000000..004ee2d8c85ce7661886179570e693d7d61bc6d8 --- /dev/null +++ b/doc/api/v2/fluid/layers.rst @@ -0,0 +1,344 @@ +========== +Layers +========== + + +fc +--- +.. autofunction:: paddle.v2.fluid.layers.fc + :noindex: + +embedding +--------- +.. autofunction:: paddle.v2.fluid.layers.embedding + :noindex: + +dynamic_lstm +------------ +.. autofunction:: paddle.v2.fluid.layers.dynamic_lstm + :noindex: + +data +---- +.. autofunction:: paddle.v2.fluid.layers.data + :noindex: + +mean +---- +.. autofunction:: paddle.v2.fluid.layers.mean + :noindex: + +mul +--- +.. autofunction:: paddle.v2.fluid.layers.mul + :noindex: + +elementwise_add +--------------- +.. autofunction:: paddle.v2.fluid.layers.elementwise_add + :noindex: + +elementwise_div +--------------- +.. autofunction:: paddle.v2.fluid.layers.elementwise_div + :noindex: + + +dropout +------- +.. autofunction:: paddle.v2.fluid.layers.dropout + :noindex: + + +reshape +-------- +.. autofunction:: paddle.v2.fluid.layers.reshape + :noindex: + + +sigmoid +--------- +.. autofunction:: paddle.v2.fluid.layers.sigmoid + :noindex: + + +scale +--------- +.. autofunction:: paddle.v2.fluid.layers.scale + :noindex: + + +transpose +--------- +.. autofunction:: paddle.v2.fluid.layers.transpose + :noindex: + + +sigmoid_cross_entropy_with_logits +--------------------------------- +.. autofunction:: paddle.v2.fluid.layers.esigmoid_cross_entropy_with_logits + :noindex: + + +cast +---- +.. autofunction:: paddle.v2.fluid.layers.cast + :noindex: + + +concat +------- +.. autofunction:: paddle.v2.fluid.layers.concat + :noindex: + + +sums +---- +.. autofunction:: paddle.v2.fluid.layers.sums + :noindex: + + +linear_chain_crf +---------------- +.. autofunction:: paddle.v2.fluid.layers.linear_chain_crf + :noindex: + + +assign +------- +.. autofunction:: paddle.v2.fluid.layers.embedding + :noindex: + + +split_lod_tensor +---------------- +.. autofunction:: paddle.v2.fluid.layers.split_lod_tensor + :noindex: + + +merge_lod_tensor +---------------- +.. autofunction:: paddle.v2.fluid.layers.merge_lod_tensor + :noindex: + +cos_sim +-------- +.. autofunction:: paddle.v2.fluid.layers.cos_sim + :noindex: + + +cross_entropy +------------- +.. autofunction:: paddle.v2.fluid.layers.cross_entropy + :noindex: + + + +square_error_cost +----------------- +.. autofunction:: paddle.v2.fluid.layers.square_error_cost + :noindex: + + +accuracy +--------- +.. autofunction:: paddle.v2.fluid.layers.accuracy + :noindex: + + +sequence_conv +------------- +.. autofunction:: paddle.v2.fluid.layers.sequence_conv + :noindex: + + +conv2d +------ +.. autofunction:: paddle.v2.fluid.layers.conv2d + :noindex: + + +sequence_pool +------------- +.. autofunction:: paddle.v2.fluid.layers.sequence_pool + :noindex: + + +sequence_first_step +------------------- +.. autofunction:: paddle.v2.fluid.layers.sequence_first_step + :noindex: + + +sequence_last_step +------------------ +.. autofunction:: paddle.v2.fluid.layers.sequence_last_step + :noindex: + + +pool2d +------ +.. autofunction:: paddle.v2.fluid.layers.pool2d + :noindex: + + +batch_norm +---------- +.. autofunction:: paddle.v2.fluid.layers.batch_norm + :noindex: + + +beam_search_decode +------------------ +.. autofunction:: paddle.v2.fluid.layers.beam_search_decode + :noindex: + + +lod_rank_table +-------------- +.. autofunction:: paddle.v2.fluid.layers.lod_rank_table + :noindex: + + +max_sequence_len +---------------- +.. autofunction:: paddle.v2.fluid.layers.max_sequence_len + :noindex: + + +topk +----- +.. autofunction:: paddle.v2.fluid.layers.topk + :noindex: + + +lod_tensor_to_array +------------------- +.. autofunction:: paddle.v2.fluid.layers.lod_tensor_to_array + :noindex: + + + +array_to_lod_tensor +------------------- +.. autofunction:: paddle.v2.fluid.layers.array_to_lod_tensor + :noindex: + + + + +fill_constant +------------- +.. autofunction:: paddle.v2.fluid.layers.fill_constant + :noindex: + + + +fill_constant_batch_size_like +----------------------------- +.. autofunction:: paddle.v2.fluid.layers.fill_constant_batch_size_like + :noindex: + + +ones +---- +.. autofunction:: paddle.v2.fluid.layers.ones + :noindex: + + +zeros +----- +.. autofunction:: paddle.v2.fluid.layers.zeros + :noindex: + + +increment +--------- +.. autofunction:: paddle.v2.fluid.layers.increment + :noindex: + + +array_write +----------- +.. autofunction:: paddle.v2.fluid.layers.array_write + :noindex: + + + +create_array +------------ +.. autofunction:: paddle.v2.fluid.layers.create_array + :noindex: + + +less_than +--------- +.. autofunction:: paddle.v2.fluid.layers.less_than + :noindex: + + +array_read +---------- +.. autofunction:: paddle.v2.fluid.layers.array_read + :noindex: + + +shrink_memory +-------------- +.. autofunction:: paddle.v2.fluid.layers.shrink_memory + :noindex: + + +array_length +------------- +.. autofunction:: paddle.v2.fluid.layers.array_length + :noindex: + + +conv2d_transpose +---------------- +.. autofunction:: paddle.v2.fluid.layers.conv2d_transpose + :noindex: + + +sequence_expand +--------------- +.. autofunction:: paddle.v2.fluid.layers.sequence_expand + :noindex: + + +lstm_unit +--------- +.. autofunction:: paddle.v2.fluid.layers.lstm_unit + :noindex: + + +sequence_softmax +---------------- +.. autofunction:: paddle.v2.fluid.layers.sequence_softmax + :noindex: + + +reduce_sum +---------- +.. autofunction:: paddle.v2.fluid.layers.reduce_sum + :noindex: + + +reduce_mean +----------- +.. autofunction:: paddle.v2.fluid.layers.reduce_mean + :noindex: + + +reduce_max +---------- +.. autofunction:: paddle.v2.fluid.layers.reduce_max + :noindex: + + +reduce_min +---------- +.. autofunction:: paddle.v2.fluid.layers.reduce_min + :noindex: + diff --git a/doc/api/v2/fluid/nets.rst b/doc/api/v2/fluid/nets.rst new file mode 100644 index 0000000000000000000000000000000000000000..b792efb71f85ae643df655568da69c82414e9d5d --- /dev/null +++ b/doc/api/v2/fluid/nets.rst @@ -0,0 +1,22 @@ +=========== +Nets +=========== + +simple_img_conv_pool +-------------------- +.. autofunction:: paddle.v2.fluid.nets.simple_img_conv_pool + :noindex: + + +img_conv_group +--------------- +.. autofunction:: paddle.v2.fluid.nets.img_conv_group + :noindex: + + +sequence_conv_pool +------------------ +.. autofunction:: paddle.v2.fluid.nets.sequence_conv_pool + :noindex: + + diff --git a/doc/api/v2/fluid/optimizer.rst b/doc/api/v2/fluid/optimizer.rst new file mode 100644 index 0000000000000000000000000000000000000000..19b4940f08de3e2f7dc177f2961e538946d10a78 --- /dev/null +++ b/doc/api/v2/fluid/optimizer.rst @@ -0,0 +1,54 @@ +=========== +Optimizer +=========== + +Optimizer +----------- +.. automodule:: paddle.v2.fluid.optimizer + :members: Optimizer + :noindex: + + +SGDOptimizer +----------- +.. automodule:: paddle.v2.fluid.optimizer + :members: SGDOptimizer + :noindex: + + + +MomentumOptimizer +----------------- +.. automodule:: paddle.v2.fluid.optimizer + :members: MomentumOptimizer + :noindex: + + + +AdagradOptimizer +---------------- +.. automodule:: paddle.v2.fluid.optimizer + :members: AdagradOptimizer + :noindex: + + +AdamOptimizer +------------- +.. automodule:: paddle.v2.fluid.optimizer + :members: AdamOptimizer + :noindex: + + +AdamaxOptimizer +----------- +.. automodule:: paddle.v2.fluid.optimizer + :members: AdamaxOptimizer + :noindex: + + +DecayedAdagradOptimizer +----------------------- +.. automodule:: paddle.v2.fluid.optimizer + :members: DecayedAdagradOptimizer + :noindex: + diff --git a/doc/api/v2/fluid/param_attr.rst b/doc/api/v2/fluid/param_attr.rst new file mode 100644 index 0000000000000000000000000000000000000000..ca0c8af9e8c4f2271de7a131ad0d27c0e8635f50 --- /dev/null +++ b/doc/api/v2/fluid/param_attr.rst @@ -0,0 +1,11 @@ +=========== +ParamAttr +=========== + + + +ParamAttr +----------- +.. automodule:: paddle.v2.fluid.param_attr + :members: ParamAttr + :noindex: diff --git a/doc/api/v2/fluid/profiler.rst b/doc/api/v2/fluid/profiler.rst new file mode 100644 index 0000000000000000000000000000000000000000..7d4042d1f41c12c4a551ba6576559d612116872a --- /dev/null +++ b/doc/api/v2/fluid/profiler.rst @@ -0,0 +1,10 @@ +=========== +Profiler +=========== + + + +Profiler +----------- +.. autofunction:: paddle.v2.fluid.profiler.cuda_profiler + :noindex: diff --git a/doc/api/v2/fluid/regularizer.rst b/doc/api/v2/fluid/regularizer.rst new file mode 100644 index 0000000000000000000000000000000000000000..868e225ed3d59e79aeb217fb88081ea25f80fa2c --- /dev/null +++ b/doc/api/v2/fluid/regularizer.rst @@ -0,0 +1,25 @@ +=========== +Regularizer +=========== + +WeightDecayRegularizer +---------------------- +.. automodule:: paddle.v2.fluid.regularizer + :members: WeightDecayRegularizer + :noindex: + + +L2DecayRegularizer +------------------ +.. automodule:: paddle.v2.fluid.regularizer + :members: L2DecayRegularizer + :noindex: + + + +L1DecayRegularizer +------------------- +.. automodule:: paddle.v2.fluid.regularizer + :members: L1DecayRegularizer + + diff --git a/doc/design/backward.md b/doc/design/backward.md new file mode 100644 index 0000000000000000000000000000000000000000..35f03692bb052e0a04db18d28f6f8d901215a553 --- /dev/null +++ b/doc/design/backward.md @@ -0,0 +1,156 @@ +# Backward Building + +## Motivation + +In Neural Network, most models are solved by the backpropagation algorithm(known as **BP**) at present. Technically, BP calculates the gradient of the loss function, then propagates it back through the networks following the chain rule. However, when configuring the model structure, users do not need to define the backward part. So a mechanism is required by the framework which can complete the model's backward part automatically according to the given forward part. + +When implementing a specific `op`, the developer is also asked to implement its backward version, called `grad_op`. A `grad_op` takes gradients of its corresponding `op`'s outputs, and calculate gradients of the `op`'s inputs. During the building of a model's backward part, the framework creates each forward `op`'s `grad_op`, and then string them together in reverse order of forwarding part. In this way, gradients spread from the end to the beginning of the model, in another word, from the loss to parameters. + +## Challenges + +The motivation of backward building is apparent. However, implementation it correctly is not so easy. In the **Fluid** design, a deep learning model is described by `Program`, `Block`, `Op` and `Variable`. The `Block` itself can be nested. It means that the `op`s and `variable`s are scattered across different blocks rather than all be gathered in a single graph. Our backward building algorithm shall visit blocks in recursive order and be able to insert `grad_op`s and new created `variable`s into the right place. + +## Usage + +Although the whole algorithm is comprised of many functions, only one is exposed as API: + +```python +def append_backward(loss, parameter_list=None, no_grad_set=None): + """ + Append backward part to main_program + + Args: + loss(Variable): The variable generated by the cost function. + parameter_list(list): Parameters that need to be updated by optimizers. + If None, it means all parameters need to be updated. + + no_grad_set(set): Variables that have no gradients in Block 0. + If None, the set will be generated inside the function and + contains all variables with `step_gradient=True` from all blocks. + + Return: + (list[Variable]): list of (parameters, gradients) pair. + """ +``` + +By invoking this API, the framework appends backward part of the program where the `loss` is. It takes three arguments. `loss` means the final loss value. It must be a scalar and is usually the output of the loss layer. It is also where the gradient generated and backpropagation starts. `parameter_list` marks all parameters needs updating. If it's `None`, all parameter will be updated by optimizers. `no_grad_set` marks variables without gradient. if all outputs of some `grad_op` are in `no_grad_set`, the `grad_op` will not be run. + +This API will be invoked automatically before optimizer building. +As a result, in most cases, users do not need to invoke the API by themselves to append backward part. + +## Implementation + +The implementation of backward building algorithm is in `backward.py` file. The whole algorithm can be divided into two independent parts: creating `grad_op`s and creating new variables. + +### Creating `grad_op`s + +The creating of `grad_op`s is implemented by: + +```python +def _append_backward_ops_(target, + block, + target_block, + no_grad_dict, + grad_to_var): + """ + Create all grad ops, and insert them into given block + + Args: + target(Variable): the target variable of forward pass + block(Block): the block where forward ops are + target_block(Block): the block which is going to hold new generated grad ops + no_grad_dict(dict): + key(int) block index + val(set) a set of varibale names. These varibales have no gradient + grad_to_var(dict)(output argument): + key(str): grad variable name + val(str): corresponding forward variable name + """ +``` + +Given a `block`, the function will traverses all `op`s in this block in reverse order, gets corresponding `grad_op` from the C++ core via `core.get_grad_op_desc()`, then append it to `target_block`. + +However, some specific `op`(e.g. `while_op`, `if_else_op`) can hold its own sub-block. For these sub-blocks contains `op`s as well, the `grad_op` creating should be recursive. + +During the reverse traversal, we check each `op` whether it has an attribute named `sub_block`. If so, it means there is a sub-block and we need to deal with it first. After creating a new block whose father is the one in `op`'s attribute, we invoke `_append_backward_ops_()` recursively, assigning the new block to parameter `target_block` and the one in `op`'s attribute to `block`. The *pseudo-code* shows this process: + +``` +******* pseudo-code ******** +for op in reversed(block.ops): + if op has an attribute named 'sub_block': + Get the sub-block(`s_block`) from op's attribute. + Create a new block(`grad_s_block`), whose father is `s_block`. + Invoke _append_backward_ops_(), with `block=s_block` and `target_block=grad_s_block` + + Invoke `core.get_grad_op_desc()` to get op's grad_op. + Insert name correspondings between variables and their gradients of the grad_op to grad_to_var + Assign grad_s_block to grad_op as it's 'sub_block' attribute. + Append grad_op to current target_block. +``` + +The first invoking of `_append_backward_ops_()` is initiated by `append_backward()`, in which parameters `block` and `target_block` are all assigned with root block(the block with index 0). + +### Corner Cases of `grad_op` Creating + +In the previous section, we show the regular process of `grad_op` creating. However, in some corner cases, the conventional algorithm is not enough to get the correct result and appending handling is required. These additional processes run after the algorithm mentioned above and do some special adjusts on its output `grad_op`s. + +#### Shared Variables + +If a variable is read by more than one `op` in the forward pass, its gradient is likely to be written by more than one `grad_op`s in the next backward pass. To make the gradient result being the sum of all `grad_op`s' outputs instead of the last running one, we assign each output with a temporary variable and then add a `sum_op` to add them up. + +For the debug convenience, if the final gradient name is `w@GRAD`, it's corresponding temporary variables will be named as `w@GRAD@RENAME@0`, `w@GRAD@RENAME@1`... + +See function `_addup_repetitive_outputs_` in `backward.py` for implementation details. + +#### No Gradient Variables + +In our framework, variables can be marked as *no_gradient*, it means that the gradient of this variable is unnecessary and can be considered as zero in model training. Apparently, when all the outputs of some `grad_op` are marked as *no_gradient*, the `grad_op` itself can be skipped in backward pass. + +But these unnecessary gradients still need to be creating and initialized by something, otherwise following `grad_op`s who take these gradients as inputs take the risk of using uninitialized memory. In our code, we employ `fill_zeros_like_op` to initialize them as all zeros. + +This features are implemented in function `_remove_no_grad_branch_`. It checks new created `grad_op`s one-by-one, removes whose outputs are all in `no_grad_set` or inserts `fill_zeros_like_op` when its necessary. We can get the `no_grad_set` from the `_append_backward_ops_` argument `no_grad_dict` or generate it on the fly by scanning all variables' `no_gradient` attribute(True or False). + +### Creating Backward Variables + +Up to now, we have completed all creating and adjusting jobs of `grad_op`s. However, backward variables have not been created. Now they are only represented by `grad_op`'s input and output arguments. The backward variable creating job will be done by: + +```python +def _append_backward_vars_(block, + start_op_idx, + grad_to_var, + grad_info_map): + """ + Create new variables required by backward pass. + + Args: + block(Block): the block where new variables will be created + start_op_idx(int): Only variables required by ops in block.ops[start_op_idx : ] will be created + grad_to_var(dict): + key(str): grad variable name + val(str): corresponding forward variable name + In most cases, this dict is generated by _append_backward_ops_() + grad_info_map(dict)(output argument): + key(str): forward variable name + val(tuple): a tuple of (str, int), str is the corresponding grad name, int is the block index + """ +``` + +Given a `block`, this function traverses all the `grad_op`s in it(The argument `start_op_idx` indicates where the grad_op sequence starts.) and creates all the uncreated outputs. The *pseudo-code* shows this process: + +``` +for op in block.ops[start_op_idx : ]: + + if op has an attribute named 'sub_block': + Get the sub-block(`s_block`) from op's attribute. + Invoke _append_backward_vars_(), with `block=s_block` + + for var_name in op.all_output_names(): + if block.has_var_recursive(var_name) or var_name is the name of empty variable: + continue + create a new variable named 'var_name' in block + if grad_to_var.has_key(var_name): + set grad_info_map[grad_to_var[var_name]] as a tuple of (var_name. block) + + do op's var type inference + do op's shape inference +``` diff --git a/doc/design/block.md b/doc/design/block.md index 4066122c0e8dfa33776796c3d205ba5aec9e0f52..fab7f2dc481ae51aa982164dc5048d90fcdc2b0b 100644 --- a/doc/design/block.md +++ b/doc/design/block.md @@ -291,10 +291,10 @@ public: } void Run(const framework::Scope& scope, - const platform::DeviceContext& dev_ctx) const override { + const platform::Place& place) const override { PADDLE_ENFORCE(symbols_ready_, "operators and variables should be created first."); for (auto& op : runtime_table_.ops()) { - op->Run(scope, dev_ctx); + op->Run(scope, place); } } diff --git a/doc/design/evaluator.md b/doc/design/evaluator.md index a62d75ffef14962aec8c7587e172d78dfe0cb4be..11cc129d56905a9ee666da92fbe6f8559c6d325a 100644 --- a/doc/design/evaluator.md +++ b/doc/design/evaluator.md @@ -1,22 +1,22 @@ ## Evaluator Design -### The Problem +### Problem Statement -During training or serving, we provide the evaluation function to measure the model performance, e.g., accuracy, precision. In the operator based framework design, the data go through the network pipeline batch by batch. As a result, inside the operator, we only can calculate one minibatch metrics. We need to provide a mechanism to calculate the metrics for each N pass/batch the user wanted. +During training or inference, we provide an evaluation function to measure the model performance, for example, accuracy, precision, etc. In the operator based framework design, the data passes through the network pipeline batch by batch. As a result, inside the operator, we only calculate the metrics for one minibatch. Thus, we need to provide a mechanism to calculate the metrics for each N pass/batch the user wants. ### Evaluator Design -Currently, every operation is expressed in the graph. we divide the evaluator process into three steps. +Currently, every operation is expressed in the graph. We divide the evaluator process into three steps. 1. Initialize the metric state and add it into the block. -2. Calculate the statistic of the metric state in every mini-batch. The single operator is only responsible for calculating necessary statistics for one mini-batch. For example, accuracy operator only calculate a minibatch data if run once. +2. Calculate the concerned metrics for every mini-batch. The single evaluator operator is only responsible for calculating the necessary statistics for one mini-batch. For example, the accuracy operator only calculates the accuracy for a minibatch data if run once. 3. Merge the mini-batch statistics to form the evaluation result for multiple mini-batches. When it comes to distributed training/Multi-GPU training, aggregate the value from different devices. ### Implementation -This design is shown in python API. -Each metric operator need to caculate the metric statistic and return the batch aware states, Python side responsible for accumulate the states for each pass. +This design is shown in the Python API. +Each metric operator needs to caculate the metric statistic and return the batch-aware states. Python side is responsible for accumulating the states for each pass. ```python diff --git a/doc/design/executor.md b/doc/design/executor.md index b5fb6c5c3c1da3c112ce63878322083dd5c42b70..2d4b371cc56db82ce5747da6db07f05aa7f7e6c1 100644 --- a/doc/design/executor.md +++ b/doc/design/executor.md @@ -1,23 +1,29 @@ # Executor Design Doc ## Motivation +In [fluid](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/fluid.md), we encourage the user to use deep learning programming paradigms to describe the training process. When the user-written Python program is executed, it will first create a protobuf message +[`ProgramDesc`](https://github.com/PaddlePaddle/Paddle/blob/a91efdde6910ce92a78e3aa7157412c4c88d9ee8/paddle/framework/framework.proto#L145) that describes the process and is conceptually like an [abstract syntax tree](https://en.wikipedia.org/wiki/Abstract_syntax_tree). -We use executor to do the runtime evaluation of a `ProgramDesc`. +The executor runs the `ProgramDesc` like an interpreter. `ProgramDesc` contains the intrinsics (operators in this case) and variables which will be used, executor explicitly executes the stored precompiled code. ## Overview -An executor takes a `ProgramDesc`, a `block_id` and a `Scope`. The `ProgramDesc` is a list of blocks and each block contains the protobuf definition of all the parameters and operators. The `block_id` specifies the entrance block. And the `Scope` is the container of all the variable instance, which is persistent throughout different runs. +An executor takes a `ProgramDesc`, a `block_id` and a `Scope`. The `ProgramDesc` is a list of blocks and each block contains the protobuf definition of all the parameters and operators in the block. The `block_id` specifies the entrance block. And the `Scope` is the container of all the variable instances, which is persistent throughout different runs. -### What does executor do? +## Executor -It evaluates all the operators in the `block_id`th block of a `ProgramDesc`. +The `Executor` explicitly executes all the intrinsics (operators here) in the `block_id`th block of a `ProgramDesc`. Essentially, it instantiates Variables and Operators, then runs all the operators in sequence one-by-one. +It is very similar to how a push stack frame works when entering a block, following which it cleans up all the temporary variables when a mini-batch is finished. It does not however, have the stack frame pop process. -### What does executor NOT do? +### The interface +```c++ + Executor(places); +``` +A executor does not own any computing resources, a user can only construct an executor using the specified places. -It does not do runtime optimization, meaning intelligently parse the dependency of each op a choose which one to be run and in which order they should be run. +### Running an Executor -It does not do graph partitioning, meaning dividing the `ProgramDesc` into several small pieces and executing them on different devices. - -## Implementation - -`Executor` evaluates a `ProgramDesc`. Essentially, it instantiates Variables and Operators, then run all the operators in sequence. [[code]](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/executor.cc) +``` + void Run(ProgramDesc, Scope, block_id, create_local_scope); +``` +An `Executor` only provides a unified way to execute `ProgramDesc`. `ProgramDesc` is the target that will be executed, the `Scope` specifies the variable container, the `block_id` indicates the entrance block and `create_local_scope` is a boolean that states whether it will destroy the temporary variables after the execution is finished. diff --git a/doc/design/float16.md b/doc/design/float16.md index 078801ba2ed969d26dd31d5ec4ed268686cf7016..1ea95ed6b5d6792171569b6ff76d09be92fcb13e 100644 --- a/doc/design/float16.md +++ b/doc/design/float16.md @@ -28,6 +28,51 @@ The goal of float16 is to serve as a key for the executor to find and run the co - [Eigen](https://github.com/RLovelett/eigen) >= 3.3 supports float16 calculation on both GPU and CPU using the `Eigen::half` class. It is mostly useful for Nvidia GPUs because of the overloaded arithmetic operators using cuda intrinsics. It falls back to using software emulation on CPU for calculation and there is no special treatment to ARM processors. - [ARM compute library](https://github.com/ARM-software/ComputeLibrary) >= 17.02.01 supports NEON FP16 kernels (requires ARMv8.2-A CPU). +### CUDA version issue +There are currently three versions of CUDA that supports `__half` data type, namely, CUDA 7.5, 8.0, and 9.0. +CUDA 7.5 and 8.0 define `__half` as a simple struct that has a `uint16_t` data (see [`cuda_fp16.h`](https://github.com/ptillet/isaac/blob/9212ab5a3ddbe48f30ef373f9c1fb546804c7a8c/include/isaac/external/CUDA/cuda_fp16.h)) as follows: +``` +typedef struct __align__(2) { + unsigned short x; +} __half; + +typedef __half half; +``` +This struct does not define any overloaded arithmetic operators. So you have to directly use `__hadd` instead of `+` to correctly add two half types: +``` +__global__ void Add() { + half a, b, c; + c = __hadd(a, b); // correct + c = a + b; // compiler error: no operator "+" matches these operands +} +``` +CUDA 9.0 provides a major update to the half data type. The related code can be found in the updated [`cuda_fp16.h`](https://github.com/ptillet/isaac/blob/master/include/isaac/external/CUDA/cuda_fp16.h) and the newly added [`cuda_fp16.hpp`](https://github.com/ptillet/isaac/blob/master/include/isaac/external/CUDA/cuda_fp16.hpp). + +Essentially, CUDA 9.0 renames the original `__half` type in 7.5 and 8.0 as `__half_raw`, and defines a new `__half` class type that has constructors, conversion operators, and also provides overloaded arithmetic operators such as follows: +``` +typedef struct __CUDA_ALIGN__(2) { + unsigned short x; +} __half_raw; + + +struct __CUDA_ALIGN__(2) __half { +protected: + unsigned short __x; +public: + // constructors and conversion operators from/to + // __half_raw and other built-in data types +} + +typedef __half half; + +__device__ __forceinline__ +__half operator+(const __half &lh, const __half &rh) { + return __hadd(lh, rh); +} + +// Other overloaded operators +``` +This new design makes `c = a + b` work correctly for CUDA half data type. ## Implementation The float16 class holds a 16-bit `uint16_t` data internally. diff --git a/doc/design/fluid-compiler.graffle b/doc/design/fluid-compiler.graffle new file mode 100644 index 0000000000000000000000000000000000000000..c933df2cb855462c52b2d25f7f9a99b95652961d Binary files /dev/null and b/doc/design/fluid-compiler.graffle differ diff --git a/doc/design/fluid-compiler.png b/doc/design/fluid-compiler.png new file mode 100644 index 0000000000000000000000000000000000000000..1b0ffed2039c91a3a00bbb719da08c91c3acf7bb Binary files /dev/null and b/doc/design/fluid-compiler.png differ diff --git a/doc/design/fluid.md b/doc/design/fluid.md new file mode 100644 index 0000000000000000000000000000000000000000..585dc8ef39c0cfb30f470d79f7b27a59ceb5e940 --- /dev/null +++ b/doc/design/fluid.md @@ -0,0 +1,122 @@ +# Design Doc: PaddlePaddle Fluid + +## Why Fluid + +When Baidu developed PaddlePaddle in 2013, the only well-known open source deep learning system at the time was Caffe. However, when PaddlePaddle was open-sourced in 2016, many other choices were available. There was a challenge -- what is the need for open sourcing yet another deep learning framework? + +Fluid is the answer. Fluid is similar to PyTorch and TensorFlow Eager Execution, which describes the "process" of training or inference using the concept of a model. In fact in PyTorch, TensorFlow Eager Execution and Fluid, there is no concept of a model at all. The details are covered in the sections below. Fluid is currently more extreme in the above mentioned idea than PyTorch and Eager Execution, and we are trying to push Fluid towards the directions of a compiler and a new programming language for deep learning. + +## The Evolution of Deep Learning Systems + +Deep learning infrastructure is one of the fastest evolving technologies. Within four years, there have already been three generations of technologies invented. + +| Existed since | model as sequence of layers | model as graph of operators | No model | +|--|--|--|--| +| 2013 | Caffe, Theano, Torch, PaddlePaddle | | | +| 2015 | | TensorFlow, MxNet, Caffe2, ONNX, n-graph | | +| 2016 | | | PyTorch, TensorFlow Eager Execution, PaddlePaddle Fluid | + +From the above table, we see that the deep learning technology is evolving towards getting rid of the concept of a model. To understand the reasons behind this direction, a comparison of the *programming paradigms* or the ways to program deep learning applications using these systems, would be helpful. The following section goes over these. + +## Deep Learning Programming Paradigms + +With the systems listed as the first or second generation, e.g., Caffe or TensorFlow, an AI application training program looks like the following: + +```python +x = layer.data("image") +l = layer.data("label") +f = layer.fc(x, W) +s = layer.softmax(f) +c = layer.mse(l, s) + +for i in xrange(1000): # train for 1000 iterations + m = read_minibatch() + forward({input=x, data=m}, minimize=c) + backward(...) + +print W # print the trained model parameters. +``` + +The above program includes two parts: + +1. The first part describes the model, and +2. The second part describes the training process (or inference process) for the model. + +This paradigm has a well-known problem that limits the productivity of programmers. If the programmer made a mistake in configuring the model, the error messages wouldn't show up until the second part is executed and `forward` and `backward` propagations are performed. This makes it difficult for the programmer to debug and locate a mistake that is located blocks away from the actual error prompt. + +This problem of being hard to debug and re-iterate fast on a program is the primary reason that programmers, in general, prefer PyTorch over the older systems. Using PyTorch, we would write the above program as following: + +```python +W = tensor(...) + +for i in xrange(1000): # train for 1000 iterations + m = read_minibatch() + x = m["image"] + l = m["label"] + f = layer.fc(x, W) + s = layer.softmax(f) + c = layer.mse(l, s) + backward() + +print W # print the trained model parameters. +``` + +We can see that the main difference is the moving the model configuration part (the first step) into the training loop. This change would allow the mistakes in model configuration to be reported where they actually appear in the programming block. This change also represents the model better, or its forward pass, by keeping the configuration process in the training loop. + +## Describe Arbitrary Models for the Future + +Describing the process instead of the model also brings Fluid, the flexibility to define different non-standard models that haven't been invented yet. + +As we write out the program for the process, we can write an RNN as a loop, instead of an RNN as a layer or as an operator. A PyTorch example would look like the following: + +```python +for i in xrange(1000): + m = read_minibatch() + x = m["sentence"] + for t in xrange x.len(): + h[t] = the_step(x[t]) +``` + +With Fluid, the training loop and the RNN in the above program are not really Python loops, but just a "loop structure" provided by Fluid and implemented in C++ as the following: + +```python +train_loop = layers.While(cond) +with train_loop.block(): + m = read_minibatch() + x = m["sentence"] + rnn = layers.While(...) + with rnn.block(): + h[t] = the_step(input[t]) +``` + +An actual Fluid example is described [here](https://github.com/PaddlePaddle/Paddle/blob/a91efdde6910ce92a78e3aa7157412c4c88d9ee8/python/paddle/v2/fluid/tests/test_while_op.py#L36-L44). + +From the example, the Fluid programs look very similar to their PyTorch equivalent programs, except that Fluid's loop structure, wrapped with Python's `with` statement, could run much faster than just a Python loop. + +We have more examples of the [`if-then-else`](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/if_else_op.md) structure of Fluid. + +## Turing Completeness + +In computability theory, a system of data-manipulation rules, such as a programming language, is said to be Turing complete if it can be used to simulate any Turing machine. For a programming language, if it provides if-then-else and loop, it is Turing complete. From the above examples, Fluid seems to be Turing complete; however, it is noteworthy to notice that there is a slight difference between the `if-then-else` of Fluid and that of a programming language. The difference being that the former runs both of its branches and splits the input mini-batch into two -- one for the True condition and another for the False condition. This hasn't been researched in depth if this is equivalent to the `if-then-else` in programming languages that makes them Turing-complete. Based on a conversation with [Yuang Yu](https://research.google.com/pubs/104812.html), it seems to be the case but this needs to be looked into in-depth. + +## The Execution of a Fluid Program + +There are two ways to execute a Fluid program. When a program is executed, it creates a protobuf message [`ProgramDesc`](https://github.com/PaddlePaddle/Paddle/blob/a91efdde6910ce92a78e3aa7157412c4c88d9ee8/paddle/framework/framework.proto#L145) that describes the process and is conceptually like an [abstract syntax tree](https://en.wikipedia.org/wiki/Abstract_syntax_tree). + +There is a C++ class [`Executor`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/executor.h), which runs a `ProgramDesc`, similar to how an interpreter runs a Python program. + +Fluid is moving towards the direction of a compiler, which is explain in more detail later in this article. + +## Backward Compatibility of Fluid + +Given all the advantages from the removal of the concept of a *model*, hardware manufacturers might still prefer the existence of the concept of a model, so it would be easier for them to support multiple frameworks all at once and could run a trained model during inference. For example, Nervana, a startup company acquired by Intel, has been working on an XPU that reads the models in the format known as [n-graph](https://github.com/NervanaSystems/ngraph). Similarly, [Movidius](https://www.movidius.com/) is producing a mobile deep learning chip that reads and runs graphs of operators. The well-known [ONNX](https://github.com/onnx/onnx) is also a file format of graphs of operators. + +For Fluid, we can write a converter that extracts the parts in the `ProgramDesc` protobuf message, converts them into a graph of operators, and exports the graph into the ONNX or n-graph format. + +## Towards a Deep Learning Language and the Compiler + +We can change the `if-then-else` and loop structure a little bit in the above Fluid example programs, to make it into a new programming language, different than Python. + +Even if we do not invent a new language, as long as we get the `ProgramDesc` message filled in, we can write a transpiler, which translates each invocation to an operator, into a C++ call to a kernel function of that operator. For example, a transpiler that weaves the CUDA kernels outputs an NVIDIA-friendly C++ program, which can be built using `nvcc`. Another transpiler could generate MKL-friendly code that should be built using `icc` from Intel. More interestingly, we can translate a Fluid program into its distributed version of two `ProgramDesc` messages, one for running on the trainer process, and the other one for the parameter server. For more details of the last example, the [concurrent programming design](concurrent_programming.md) document would be a good pointer. The following figure explains the proposed two-stage process: + +![](fluid-compiler.png) diff --git a/paddle/framework/images/duplicate_op.graffle b/doc/design/images/duplicate_op.graffle similarity index 100% rename from paddle/framework/images/duplicate_op.graffle rename to doc/design/images/duplicate_op.graffle diff --git a/paddle/framework/images/duplicate_op.png b/doc/design/images/duplicate_op.png similarity index 100% rename from paddle/framework/images/duplicate_op.png rename to doc/design/images/duplicate_op.png diff --git a/paddle/framework/images/duplicate_op2.graffle b/doc/design/images/duplicate_op2.graffle similarity index 100% rename from paddle/framework/images/duplicate_op2.graffle rename to doc/design/images/duplicate_op2.graffle diff --git a/paddle/framework/images/duplicate_op2.png b/doc/design/images/duplicate_op2.png similarity index 100% rename from paddle/framework/images/duplicate_op2.png rename to doc/design/images/duplicate_op2.png diff --git a/doc/design/images/multigpu_allreduce.graffle b/doc/design/images/multigpu_allreduce.graffle new file mode 100644 index 0000000000000000000000000000000000000000..cb5bc420ceafe8ba4c87694d44ee4e5e4ad06779 Binary files /dev/null and b/doc/design/images/multigpu_allreduce.graffle differ diff --git a/doc/design/images/multigpu_allreduce.png b/doc/design/images/multigpu_allreduce.png new file mode 100644 index 0000000000000000000000000000000000000000..87a1b3e8f6dd4a713ec9df9f0037d1da04e9178a Binary files /dev/null and b/doc/design/images/multigpu_allreduce.png differ diff --git a/doc/design/images/multigpu_before_convert.graffle b/doc/design/images/multigpu_before_convert.graffle new file mode 100644 index 0000000000000000000000000000000000000000..6c35ab1b21fb76ceae82d3693ed0d085b5bc0855 Binary files /dev/null and b/doc/design/images/multigpu_before_convert.graffle differ diff --git a/doc/design/images/multigpu_before_convert.png b/doc/design/images/multigpu_before_convert.png new file mode 100644 index 0000000000000000000000000000000000000000..9c8f7711165d80a2fa3911280fdee91855a401b1 Binary files /dev/null and b/doc/design/images/multigpu_before_convert.png differ diff --git a/doc/design/kernel_hint_design.md b/doc/design/kernel_hint_design.md new file mode 100644 index 0000000000000000000000000000000000000000..a54b7da045e1a362626ef066f9ebb56af2c3181a --- /dev/null +++ b/doc/design/kernel_hint_design.md @@ -0,0 +1,57 @@ +## Problem +In PaddlePaddle's [Design](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/switch_kernel.md), one Operator may have multiple kernels. Users may have some personal preference to choose a certain type of kernel for an operator, such as `force_cpu` to choose a CPU kernel, `use_cudnn` to choose a CUDNN kernel, we need to provide a way for users to do this. + +In the current design, we use KernelType to describe one kernel. + +```cpp +struct KernelType { + Place place_; + DataType data_type_; + LayoutType layout_; +}; +``` + `place_` `data_type_` and `layout_` can be got from the input tensors of the operator, `GetActualKernelType(inputs)` use inputs to infer the proper kernel key that fit the incoming data, but users can not directly configure it. + +The [design](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/switch_kernel.md) also provides a virtual method `GetExpectedKernelType` that user can overload and use to choose the KernelType they want to use. + +So we should send the information user defined in proto to `GetExpectedKernelType` for choosing a kernel. + +The problem is, how should we define and send the information for `GetExpectedKernelType` to use? + +## Solution + +### Potential choice +1. Do nothing, let the user add the information they want to operator‘s attribute and get them inside `GetExpectedKernelType`, this can work properly. But there is a little problem that users may define many kinds of hints for the same purpose, such as `force_cpu`, `use_cpu`, `cpu_kernel` to choose CPU kernel, and `use_cudnn`, `force_cudnn`, `cudnn_kernel` to choose CUDNN kernel. + +2. Pre-define all the needed option and use a single attr key such as `kernel_hint` for the user, this is not so flexible if the user wants to define some more kind of hint. + +### Final choice +To provide enough flexibility while avoiding confusion definition, we can define some global constants for these attribute names, such as `force_cpu`, `use_cudnn`, `use_mkldnn` for a user to choose. + +In C++ + +```cpp +const std::string kForceCPU = "force_cpu"; +const std::string kUseCUDNN = "use_cudnn"; +const std::string kUseMKLDNN = "use_mkldnn"; + +KernelType GetExpectedKernelType() { + if (Attr(kForceCPU)) { + return KernelType(CPUPlace, ...) + } else { + ... + } +} +``` + +In Python code + +```python +FORCE_CPU = core.kForceCPU() + +def xx_layer(..., force_cpu=false): + layer_helper = LayerHelper(...) + layer_helper.append_op( + type="xx", + attr={FORCE_CPU: force_cpu}) +``` diff --git a/doc/design/mkl/image/engine.png b/doc/design/mkl/image/engine.png new file mode 100644 index 0000000000000000000000000000000000000000..1f5f65c2cc765a514a3ba9e7b7f468e1dc4b0c3b Binary files /dev/null and b/doc/design/mkl/image/engine.png differ diff --git a/doc/design/mkl/image/gradients.png b/doc/design/mkl/image/gradients.png new file mode 100644 index 0000000000000000000000000000000000000000..f031bcf8e4cec14e63075b8b9d2c7bbd9f1b1a3c Binary files /dev/null and b/doc/design/mkl/image/gradients.png differ diff --git a/doc/design/mkl/image/layers.png b/doc/design/mkl/image/layers.png new file mode 100644 index 0000000000000000000000000000000000000000..306f79b7a844610915eb8944128f57d2b7a3065a Binary files /dev/null and b/doc/design/mkl/image/layers.png differ diff --git a/doc/design/mkl/image/matrix.png b/doc/design/mkl/image/matrix.png new file mode 100644 index 0000000000000000000000000000000000000000..c33ce9cf0335e47cc8c1253304d0fe179186e6f2 Binary files /dev/null and b/doc/design/mkl/image/matrix.png differ diff --git a/doc/design/mkl/image/overview.png b/doc/design/mkl/image/overview.png new file mode 100644 index 0000000000000000000000000000000000000000..8fb7bbb9dd654bf363d701d0c8cd4a557043d188 Binary files /dev/null and b/doc/design/mkl/image/overview.png differ diff --git a/doc/design/mkl/mkl_packed.md b/doc/design/mkl/mkl_packed.md new file mode 100644 index 0000000000000000000000000000000000000000..0123315ad4368e68b377f66119949bfd6c1c7860 --- /dev/null +++ b/doc/design/mkl/mkl_packed.md @@ -0,0 +1,108 @@ +# Intel® MKL Packed on PaddlePaddle: Design Doc + + +## Contents + +- [Overview](#overview) +- [Key Points](#key-points) + - [Background](#background) + - [Solution](#solution) +- [Actions](#actions) + - [CMake](#cmake) + - [Layers](#layers) + - [Unit Tests](#unit-tests) + - [Python API](#python-api) + - [Benchmarking](#benchmarking) + + +## Overview +我们计划将 Intel® MKL 中引入的 GEMM Packed APIs\[[1](#references)\] 集成到 PaddlePaddle 中，充分发挥英特尔平台的优势，有效提升PaddlePaddle在英特尔架构上的性能。 +现阶段的优化主要针对 Recurrent Neural Network（以下简称RNN）相关层（包括`RecurrentLayer`, `GatedRecurrentLayer`和`LstmLayer`）， 以及 PaddlePaddle V1 API。 + +## Key Points + +### Background +目前PaddlePaddle采用了 Intel® MKL库的[cblas_?gemm](https://software.intel.com/en-us/mkl-developer-reference-c-cblas-gemm)函数，这个函数本身会在计算前将原数据转换为更适合英特尔平台的内部格式。 + +1. 转换耗时 \ +这一数据格式的转换操作（Packing），在问题本身的计算量比较小的时候，显得相对来说较为耗时。例如在DeepSpeech2 \[[2](#references)\] 的Vanilla RNN部分中，矩阵大小是`batch_size * 2048`。 +2. 转换冗余 \ +由于在现有的某些情况下（例如RNN），多次调用 cblas_?gemm 会使用相同的原数据，因此，每次调用时对原数据的重复Packing便成为了冗余。 + +为了最大程度减少多次调用 cblas_?gemm 在Packing上的耗时，Intel® MKL 引入了以下四个API: + * [cblas_?gemm_alloc](https://software.intel.com/en-us/mkl-developer-reference-c-cblas-gemm-alloc) + * [cblas_?gemm_pack](https://software.intel.com/en-us/mkl-developer-reference-c-cblas-gemm-pack) + * [cblas_?gemm_compute](https://software.intel.com/en-us/mkl-developer-reference-c-cblas-gemm-compute) + * [cblas_?gemm_free](https://software.intel.com/en-us/mkl-developer-reference-c-cblas-gemm-free) + +通过使用这些API，我们可以先完成对原数据的Packing操作，再把已转换为Packed格式的数据传递给那些复用同一数据的gemm_compute函数，从而避免了Packing冗余。 + +### Solution +在RNN的情况下，同一次前向、后向（forward/backward）过程中所有时间步（time step）共享同一个权重（weight）。当只做推断（inference）时，各次前向之间也都使用了相同的权重，没有必要在每次前向中每个时间步的计算时对权重进行重复的Packing操作。 + +我们通过使用新引入的GEMM Packed APIs，在层初始化的时候，先完成对权重的Packing操作，然后在前向，后向时复用已经转换过的权重，并在每次权重更新后，对新的权重进行转换用于下次迭代。 + +* 优化前，对于序列长度（sequence length）为`T`的网络模型（model）, `N`次迭代执行的转换次数为： + - `inference`： `N * T` + - `training`： `2 * N * T` +* 优化后，对于同样设置的网络模型，其转换次数减少至： + - `inference`： `1` + - `training`： `2 * N` + +## Actions + +添加的相关文件和目录结构如下： + +```txt +PaddlePaddle/Paddle +├── ... +└── paddle/ + ├── ... + └── gserver/ + ├── ... + ├── layers/ + │ ├── ... + │ ├── MKLPackedRecurrentLayer.* + | ├── MKLPackedGatedRecurrentLayer.* + | ├── MKLPackedLstmLayer.* + | └── MKLPackedGemm.h + └── tests/ + ├── ... + └── test_MKLPacked.cpp +``` + +### CMake +在对应的`CMakeLists.txt`中根据`WITH_MKL`是否打开，来决定是否开启MKL Packed相关功能。 + +### Layers +所有的`MKLPacked*Layer`都继承于PaddlePaddle的基类`Layer`, 并添加头文件 `MKLPackedGemm.h`，该文件对相关GEMM Packed APIs做了封装。 + +### Unit Tests +我们会添加`test_MKLPacked.cpp`用于MKL Packed优化后layer的测试。 +对于每一个新加的RNN layer，我们会对比如下2个方面： +1. 对比优化后layer自身，sequence mode（`rnn_use_batch=false`）与batch mode(`rnn_use_batch=true`)的结果。 +2. 对比优化后layer与相对应的PaddlePaddle原有layer, 在batch mode下的结果。 + +### Python API +计划在`paddle/utils.Flags`中添加`use_mkl_packed`的flag，用于选择是否使用相关功能，并且当编译时`WITH_MKL=ON`的情况下，默认设置为`true`。 + +同时，在`python/paddle/trainer/config_parser.py`中对应的layer处，添加`use_mkl_packed`这个选择，方便用户在Python端选择是否启用这个功能。 + +具体实现方式比如： + +```python +use_mkl_packed = bool(int(g_command_config_args.get("use_mkl_packed", 0))) +if use_mkl_packed: + self.layer_type = mkl_packed_* +``` + +所有相关的`layer_type`会以*mkl_packed_*开头，这些会在`MKLPacked*Layer`注册layer的时候保证，以示区分。 + + +### Benchmarking +会添加相应的脚本用于测试和对比在使用MKL Packed recurrent layers 前后的网络性能。 + +## References +1. [Introducing the new Packed APIs for GEMM](https://software.intel.com/en-us/articles/introducing-the-new-packed-apis-for-gemm) +2. [DeepSpeech2 on PaddlePaddle](https://github.com/PaddlePaddle/DeepSpeech#deepspeech2-on-paddlepaddle) + diff --git a/doc/design/mkl/mkldnn.md b/doc/design/mkl/mkldnn.md new file mode 100644 index 0000000000000000000000000000000000000000..e2fe1e6b26ffa73fda81863abfadf697c0acbfcf --- /dev/null +++ b/doc/design/mkl/mkldnn.md @@ -0,0 +1,210 @@ +# Intel® MKL-DNN on PaddlePaddle: Design Doc + +我们计划将英特尔深度神经网络数学库[Intel MKL-DNN](https://github.com/01org/mkl-dnn) +(Intel Math Kernel Library for Deep Neural Networks)集成到PaddlePaddle， +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 +
+ +近期目标 + +- 完成常用Layer的MKL-DNN实现。 +- 完成常见深度神经网络VGG，GoogLeNet 和 ResNet的MKL-DNN实现。 + +目前的优化，主要针对PaddlePaddle在重构之前的代码框架以及V1的API。 +具体的完成状态可以参见[这里](https://github.com/PaddlePaddle/Paddle/projects/21)。 + +## Contents + +- [Overview](#overview) +- [Actions](#actions) + - [CMake](#cmake) + - [Matrix](#matrix) + - [Layers](#layers) + - [Activations](#activations) + - [Parameters](#parameters) + - [Gradients](#gradients) + - [Unit Tests](#unit-tests) + - [Python API](#python-api) + - [Benchmarking](#benchmarking) + - [Others](#others) +- [Design Concerns](#design-concerns) + +## Overview + +我们会把MKL-DNN会作为第三方库集成进PaddlePaddle，与其他第三方库一样，会在编译PaddlePaddle的时候下载并编译MKL-DNN。 + +同时，为了进一步提升PaddlePaddle在基本数学运算的计算速度，我们也将MKLML即(MKL small library\[[1](#references)\]) +作为另一个第三方库集成进PaddlePaddle，它只会包括生成好的动态库和头文件。 + +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 | + +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 +
+ +## Actions + +添加的相关文件和目录结构如下： + +```txt +PaddlePaddle/Paddle +├── ... +├── cmake/ +│ ├── external/ +│ │ ├── ... +│ │ ├── mkldnn.cmake +│ │ └── mklml.cmake +└── paddle/ + ├── ... + ├── math/ + │ ├── ... + │ └── MKLDNNMatrix.* + └── gserver/ + ├── ... + ├── layers/ + │ ├── ... + │ └── MKLDNN*Layer.* + ├── activations/ + │ ├── ... + │ └── MKLDNNActivations.* + └── tests/ + ├── ... + ├── MKLDNNTester.* + └── test_MKLDNN.cpp +``` + +### CMake +在`CMakeLists.txt`中提供一个与MKL有关的总开关：`WITH_MKL`，它负责决定编译时是否使用MKLML和MKL-DNN + +- `WITH_MKLML` 控制是否使用MKLML库。 +当打开`WITH_MKL`时，会自动使用MKLML库作为PaddlePaddle的CBLAS和LAPACK库，同时会开启Intel OpenMP用于提高MKLML的性能。 +编译时会把对应的头文件和库放在`build/third_party/install/mklml/*`目录下对应的地方。 +MKLML的库目前都是动态库，主要包括`libiomp5.so`和`libmklml_intel.so`。 +- `WITH_MKLDNN` 控制是否使用MKL-DNN。 +当开启`WITH_MKL`时，会自动根据硬件配置[[2](#references)]选择是否编译MKL-DNN。 +编译时会把对应的头文件和库放在`build/third_party/install/mkldnn/*`目录下对应的地方。 +MKL-DNN的库目前只有动态库`libmkldnn.so`。 + +### Matrix +目前在PaddlePaddle中数据都是以`NCHW`的格式存储，但是在MKL-DNN中的排列方式不止这一种。 +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 +
+ +### Layers +所有MKL-DNN的Layers都会继承于`MKLDNNLayer`，该类继承于PaddlePaddle的基类`Layer`。 +在`MKLDNNLayer`中会提供一些必要的接口和函数，并且会写好`forward`和`backward`的基本逻辑， +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 +
+ +每个MKLDNNLayer都包含用于内部存储和外部存储的一系列MKLDNNMatrix： + +- 内部存储（internel memory）：`inVal_`,`inGrad_`,`outVal_`和`outGrad_`，分别代表输入数据，输入梯度，输出数据和输出梯度。 +- 外部存储（external memory）：都是以ext开头，比如`extInVal_`和`extInGrad_`，它们主要是用于， +当数据格式与PaddlePaddle默认的`NCHW`格式不匹配时，转换内存的工作。 +需要注意的是，PaddlePaddle的activation会直接使用`output_.value`和`output_.grad`， +所以`extOutVal_`和`extOutGrad_`必须分别与`output_.value`和`output_.grad`共享内存， +如果不需要外部存储用于转换，那么对应的内部存储也会与它们共享内存。 +- 转换函数（resetXXX）： 包括`resetInValue`，`resetInGrad`，`resetOutValue`和`resetOutGrad`， +表示对输入数据，输入梯度，输出数据和输出梯度的转换。 +这些函数会根据输入参数重新设置内部和外部存储，当然这两者也可以相等，即表示不需要转换。 + +注意：每个`MKLDNNlayer`的子类只需要使用内部存储就可以了，所有外部的转换工作都会在reset系列函数中都准备好。 + +### Activations +在重构前的PaddlePaddle中，激活函数是独立于`Layer`的概念，并且输入输出都是共用一块内存， +所以添加了对应的`MKLDNNActivation`来实现，方式类似于`MKLDNNLayer`。 + +### Parameters +对于有参数的层，我们会保证`MKLDNNLayer`使用的参数与PaddlePaddle申请的buffer共用一块内存。 +如果存在数据排列格式不一样的情况时，我们会在网络训练之前把格式转换为MKL-DNN希望的格式， +在训练结束的时候再保存为PaddlePaddle的格式，但是整个训练过程中不需要任何转换。 +这样既使得最终保存的参数格式与PaddlePaddle一致，又可以避免不必要的转换。 + +### Gradients +由于MKL-DNN的操作都是直接覆盖的形式，也就是说输出的结果不会在原来的数据上累加， +这样带来的好处就是不需要一直清空memory，节省了不必要的操作。 +但是注意的是，当网络出现分支且在`backward`的时候，需要累加不同Layer传过来的梯度。 +所以在`MKLDNNlayer`中实现了一个merge的方法，此时每个小分支的`Input Gradient` +会先临时保存在`MKLDNNMatrix`中，由分支处的Layer负责求和，并把结果放到当前层的`output_.grad`中。 +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 +
+ +### Unit Tests +我们会添加`test_MKLDNN.cpp`和`MKLDNNTester.*`用于MKL-DNN的测试。 +测试分为每个Layer（或Activation）的单元测试和简单网络的整体测试。 +每个测试会对比PaddlePaddle中CPU算出的结果与MKL-DNN的结果，小于某个比较小的阈值认为通过。 + +### Python API +目前只考虑**v1 API**。 + +计划在`python/paddle/trainer/config_parser.py`里面添加`use_mkldnn`这个选择，方便用户选择使用MKL-DNN的layers。 + +具体实现方式比如： + +```python +use_mkldnn = bool(int(g_command_config_args.get("use_mkldnn", 0))) +if use_mkldnn + self.layer_type = mkldnn_* +``` + +所有MKL-DNN的`layer_type`会以*mkldnn_*开头，这些会在`MKLDNN*Layer`注册layer的时候保证，以示区分。 + +同时,会在`paddle/utils.Flags`中添加一个`use_mkldnn`的flag，用于选择是否使用MKL-DNN的相关功能。 + +### Benchmarking +会添加相应的脚本在[这里](https://github.com/PaddlePaddle/Paddle/tree/develop/benchmark/paddle/image)，用于测试和对比在使用MKL-DNN前后的CNN网络性能。 +测试的性能对比结果会在[IntelOptimizedPaddle.md](https://github.com/PaddlePaddle/Paddle/blob/develop/benchmark/IntelOptimizedPaddle.md) + +### Others +1. 如果在使用MKL-DNN的情况下，会把CPU的Buffer对齐为4096，具体可以参考MKL-DNN中的[memory](https://github.com/01org/mkl-dnn/blob/master/include/mkldnn.hpp#L673)。 +2. 深入PaddlePaddle，寻找有没有其他可以优化的可能，进一步优化。比如可能会用OpenMP改进SGD的更新性能。 + +## Design Concerns + +为了更好的符合PaddlePaddle的代码风格\[[3](#references)\]，同时又尽可能少的牺牲MKL-DNN的性能\[[4](#references)\]。 + +我们总结出一些特别需要注意的点： + +1. 使用**deviceId_**。为了尽可能少的在父类Layer中添加变量或者函数， +我们决定使用已有的`deviceId_`变量来区分layer的属性，定义`-2`为`MKLDNNLayer`特有的设备ID。 +2. 重写父类Layer的**init**函数，修改`deviceId_`为`-2`，代表这个layer是用于跑在MKL-DNN的环境下。 +3. 创建`MKLDNNBase`，定义一些除了layer和memory相关的类和函数。 +包括MKL-DNN会用到`MKLDNNStream`和`CPUEngine`，和未来可能还会用到`FPGAEngine`等。 +4. 如果MKL-DNN layer的后面接有cpu device，那么就会使`output_.value`与`extOutVal_`共享内存， +同时数据格式就是`NCHW`，这样下一个cpu device就能拿到正确的数据。 +在有普通的CPU layer时， `extOutVal_`和`extOutGrad_`的格式始终是`NCHW`或者`NC`。 + +## References +1. [MKL small library](https://github.com/01org/mkl-dnn#linking-your-application)是[Intel MKL](https://software.intel.com/en-us/mkl)的一个子集。 +主要包括了深度学习相关的数学原语与操作，一般由MKL-DNN在发布[新版本](https://github.com/01org/mkl-dnn/releases)时一起更新。 +2. [MKL-DNN System Requirements](https://github.com/01org/mkl-dnn#system-requirements)。 +目前在PaddlePaddle中，仅会在支持AVX2指令集及以上的机器才使用MKL-DNN。 +3. [原来的方案](https://github.com/PaddlePaddle/Paddle/pull/3096)会引入**nextLayer**的信息。 +但是在PaddlePaddle中，无论是重构前的layer还是重构后的op，都不会想要知道next layer/op的信息。 +4. MKL-DNN的高性能格式与PaddlePaddle原有的`NCHW`不同(PaddlePaddle中的cuDNN部分使用的也是`NCHW`，所以不存在这个问题)。 +所以需要引入一个转换方法，并且只需要在必要的时候转换这种格式，才能更好的发挥MKL-DNN的性能。 diff --git a/doc/design/mkl/mkldnn_fluid.md b/doc/design/mkl/mkldnn_fluid.md new file mode 100644 index 0000000000000000000000000000000000000000..bef126f3f0577b69f646dfe5d10539b372c6a8a5 --- /dev/null +++ b/doc/design/mkl/mkldnn_fluid.md @@ -0,0 +1,149 @@ +# Design Doc: Add MKLDNN Kernel in Fluid Operator + +## Principles + +First of all, we should follow some basical principles like: +1. [How to write a new operator](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/dev/new_op_en.md). We are trying to add a new kind of kernel into operators, so basically we should follow this doc. +2. [Supporting new Device/Library](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/support_new_device.md). Since MKLDNN is a new library to fluid, we should add `MKLDNNDeviceContext` and maybe `mkldnn_helper.h`, just like [cudnn_helper.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/cudnn_helper.h). +3. [Switch Kernel](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/switch_kernel.md). Another important point is that we should ensure the data synchronization between different kernel types, which is this [topic](https://github.com/PaddlePaddle/Paddle/issues/6549). So basically we should override `GetExpectedKernelType` and `trans` functions to support switching kernels. +4. [The Keys of Operator Kernel Type](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/operator_kernel_type.md). Kernel Type is a pivotal conception which can record the `Place`, `Library`, `DataType` and `Layout`. + +## Sulution + +In general, there are four parts we should follow to run a MKL-DNN primitive. +- Create a primitive descriptor that describe this operator +- Create a primitive itself by primitive descriptor and the engine +- Create all memory buffers that primitive needed +- Launch a stream to execute the primitive created +More details can refer to [here](http://01org.github.io/mkl-dnn). + +It's better to avoid reinitialization of primitives and memory handles in the first three stages in every iteration. \ +So we plan to create a map to record all the `primitive` and `memory`, which should not take too much memories as discussed [here](https://github.com/PaddlePaddle/Paddle/issues/6822). + +It's assumed that following three conditions should be satisfied. +1. there is a unique key for each operator instance. May be the actual name of `Output Tensor`. +2. the `Input Tensor` inside `Compute` function is the one after converted. +3. we can get the phase(eg. `is_test`) inside `Compute` function, otherwise we need to expose this attribue to user. + +### Compute +The algorithm of `Compute` would be described as follow, let's take conv like an example. + +```c++ + + PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()), "It must use CPUPlace."); + PADDLE_ENFORCE(platform::is_mkldnn_library(ctx.GetLibrary()), "It must use MKLDNN Library."); + + auto& dev_ctx = ctx.template device_context(); + + // find primitive by unique key from mkldnn context + // the op_key should be a unique name of this op instance + auto& p = dev_ctx.findPrimitive(op_key + "_fwd"); + + // assuming the input tensor inside this compute function is the one after converted + // this point should be guarantee by another mechanism + auto& i = dev_ctx.findMemory(op_key + "_input"); + + if (p == nullptr || i == nullptr || inputSizeChanged(p, i)) { + auto fwd_primitive_desc = createPrimitiveDesc(ctx); + auto* input = ctx.Input("Input"); + auto* filter = ctx.Input("Filter"); + auto* output = ctx.Output("Output"); + shared_ptr in(new mkldnn::memory(fwd_primitive_desc->src_primitive_desc(), input->data())); + shared_ptr wgt(new mkldnn::memory(fwd_primitive_desc->weights_primitive_desc(), filter->data())); + shared_ptr out(new mkldnn::memory(fwd_primitive_desc->dst_primitive_desc(), output->mutable_data(ctx.GetPlace()))); + shared_ptr fwd_primitive(new mkldnn::conv_fwd(*fwd_primitive_desc, *in, *wgt, *out)); + + dev_ctx.addMemory(op_key+"_input", in); + dev_ctx.addMemory(op_key+"_output", out); + dev_ctx.addMemory(op_key+"_filer", wgt); + dev_ctx.addPrimitive(op_key+"_fwd", fwd_primitive); + dev_ctx.addPrimitiveDesc(op_key+"_fwd_PD", fwd_primitive_desc); + } + + p = dev_ctx.findPrimitive(op_key + "_fwd"); + + PADDLE_ENFORCE(p, "Should have forward Primitive"); + PADDLE_ENFORCE(dev_ctx.findMemory(op_unique_key+"_input"), "Should have input memory"); + PADDLE_ENFORCE(dev_ctx.findMemory(op_unique_key+"_output"), "Should have output memory"); + PADDLE_ENFORCE(dev_ctx.findMemory(op_unique_key+"_filter"), "Should have filter memory"); + PADDLE_ENFORCE(dev_ctx.findPrimitiveDesc(op_unique_key+"_fwd_PD"), "Should have forward PrimitiveDesc"); + dev_ctx.submit(p); + dev_ctx.execute(); // the convert primitive should have already contained. + +``` + +The `createPrimitiveDesc` returns the primitive descripotor of this operator, would be like this: +```c++ + auto* input = ctx.Input("Input"); + auto* filter = ctx.Input("Filter"); + auto* output = ctx.Output("Output"); + std::vector strides = ctx.Attr>("strides"); + std::vector paddings = ctx.Attr>("paddings"); + std::vector dilations = ctx.Attr>("dilations"); + int groups = ctx.Attr("groups"); + algorithm algo = static_cast(ctx.Attr("convolution_algorithm_option")); + prop_kind pk = ctx.Attr("is_test") ? prop_kind::forward_inference : prop_kind::forward_training; + + auto fwd_desc = mkldnn::conv_fwd::desc(/* all the setting above*/); + shared_ptr fwd_primitive_desc(new mkldnn::conv_fwd::primitive_desc(fwd_desc, ctx.getEngine())); + + return fwd_primitive_desc; + } +``` + +### MKLDNNDeviceContext +`MKLDNNDeviceContext`, which is very straightforward, should contain some base information like: `stream`, `engine` and the map needed. + + +### mkldnn_helper +Some functions would be put in `paddle/platform/mkldnn_helper.h`. +- create MKLDNN memories +- create MKLDNN primitives +- error check function +- etc + + +### Kernel Switch +We should `reorder` the different Layout from other device or to other device. `GetExpectedKernelType` and `trans` functions can help us to implement it. + +`GetExpectedKernelType` should get the context, and this operator can return the best `KernelType`. +`trans` would be like this: + +```c++ +void trans(inputs, ctx) override { + if (NoNeedTrans()) { + return; + } + // find reorder primitive by op_key from context + auto& dev_ctx = ctx.template device_context(); + auto& p = dev_ctx.findPrimitive(op_key + "_reorder_input"); + auto& i = dev_ctx.findMemory(op_key + "_src_input"); + + if (p == nullptr || i == nullptr || changeSized(i, input)) { + auto prim = createPrimitiveDesc(ctx); + auto src = createMemory(memoryDesc(input->dims(), actual_layout), input->data); + auto newbuffer = paddle::memory::Alloc(ctx.GetPlace(), input->size_in_bytes()); + auto dst = createMemory(p->expected_desc(), newbuffer->data); + auto reorder_primitive(new mkldnn::reorder(src, dst)); + + dev_ctx.addMemory(op_key+"_src_input", src); + dev_ctx.addMemory(op_key+"_input", dst); + dev_ctx.addPrimitive(op_key+"_reorder_input", reorder_primitive); + } + + p = dev_ctx.findPrimitive(op_key + "_reorder_input"); + PADDLE_ENFORCE(p, "Should have Reorder Primitive"); + dev_ctx.submit(p); + if (! this->isMKLDNNKernel()) { + // execute immediately only if this is not mkldnn kernel function. + // otherwise, it can be executed with the operator primitive in Compute + dev_ctx.stream(); + } + // after submit, the input tensor in ExecutionContext should be changed as the converted one + // there should be another mechanism to ensure this +} +``` + +### Unit Test +All the functions should be tested corresponding. +TBD diff --git a/doc/design/mkldnn/README.MD b/doc/design/mkldnn/README.MD deleted file mode 100644 index ec6d4681836e189f46dbb9b915a237dc15cda7cf..0000000000000000000000000000000000000000 --- a/doc/design/mkldnn/README.MD +++ /dev/null @@ -1,120 +0,0 @@ -# Intel® MKL-DNN on PaddlePaddle: Design Doc - -我们计划将Intel深度神经网络数学库(**MKL-DNN**\[[1](#references)\])集成到PaddlePaddle，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。 - -我们短期内的基本目标是： - -- 完成常用layer的MKL-DNN实现。 -- 完成常见深度神经网络VGG，GoogLeNet 和 ResNet的MKL-DNN实现。 - - -## Contents - -- [Overview](#overview) -- [Actions](#actions) - - [CMake](#cmake) - - [Layers](#layers) - - [Activations](#activations) - - [Weights](#weights) - - [Unit Tests](#unit-tests) - - [Protobuf Messages](#protobuf-messages) - - [Python API](#python-api) - - [Demos](#demos) - - [Benchmarking](#benchmarking) - - [Others](#others) -- [Design Concerns](#design-concerns) - -## Overview - -我们会把MKL-DNN作为第三方库集成进PaddlePaddle，整体框架图 -
-
-Figure 1. PaddlePaddle on IA. -
- -## Actions -我们把集成方案大致分为了如下几个方面。 - -### CMake -我们会在`CMakeLists.txt`中会给用户添加一个`WITH_MKL`的开关，他是负责`WITH_MKLML`和`WITH_MKLDNN`的总开关。 - -当打开`WITH_MKL`时，会开启MKLML的功能，作为PaddlePaddle的CBLAS和LAPACK库，同时会开启Intel OpenMP用于提高MKLML的性能。 如果系统支持AVX2指令集及以上，同时会开启MKL-DNN功能。 - -当关闭`WITH_MKL`时，MKLML和MKL-DNN功能会同时关闭。 - -所以，我们会在`cmake/external`目录新建`mkldnn.cmake`和`mklml.cmake`文件，它们会在编译PaddlePaddle的时候下载对应的软件包，并放到PaddlePaddle的third party目录中。 - -### Layers -所有MKL-DNN相关的C++ layers，都会按照PaddlePaddle的目录结构存放在 -`paddle/gserver/layers`中，并且文件名都会一以*MKLDNN*开头。 - -所有MKL-DNN的layers都会继承于一个叫做`MKLDNNLayer`的父类，该父类继承于PaddlePaddle的基类`Layer`。 - -在`MKLDNNLayer`中会提供一些必要的接口和函数，并且会写好`forward`和`backward`的基本逻辑。部分函数定义为纯虚函数，子类只需要实现这些函数即可。 - -### Activations -由于在PaddlePaddle中，激活函数是独立于layer概念的，所以会在`paddle/gserver/activations`目录下添加`MKLDNNActivation.h`和`MKLDNNActivation.cpp`文件用于定义和使用MKL-DNN的接口。 - -### Weights -由于有些layer是含有参数的，我们会尽量让MKL-DNN的参数与PaddlePaddle中`parameter`共享一块内存。 -同时，由于MKL-DNN在训练时使用的参数layout可能与PaddlePaddle默认的`nchw`不一致，我们会在网络训练的开始和结束时分别转换这个layout，使得最终保存的参数格式与PaddlePaddle一致。 - -### Unit Tests -会在`paddle/gserver/test`目录下添加`test_MKLDNN.cpp`和`MKLDNNTester.*`用于MKL-DNN的测试。 -测试分为每个layer(或activation)的单元测试和简单网络的整体测试。 -每个测试会对比PaddlePaddle中CPU算出的结果与MKL-DNN的结果，小于某个比较小的阈值认为通过。 - -### Protobuf Messages -根据具体layer的需求可能会在`proto/ModelConfig.proto`里面添加必要的选项。 - -### Python API -目前只考虑**v1 API**。 - -计划在`python/paddle/trainer/config_parser.py`里面添加`use_mkldnn`这个选择，方便用户选择使用MKL-DNN的layers。 - -具体实现方式比如： - -```python -use_mkldnn = bool(int(g_command_config_args.get("use_mkldnn", 0))) -if use_mkldnn - self.layer_type = mkldnn_* -``` - -所有MKL-DNN的layer type会以*mkldnn_*开头，以示区分。 - -并且可能在`python/paddle/trainer_config_helper`目录下的`activations.py `和`layers.py`里面添加必要的MKL-DNN的接口。 - -### Demos - -会在`v1_api_demo`目录下添加一个`mkldnn`的文件夹，里面放入一些用于MKL-DNN测试的demo脚本。 - -### Benchmarking -会添加`benchmark/paddle/image/run_mkldnn.sh`，用于测试使用MKL-DNN之后的性能。 - -### Others -1. 如果在使用MKL-DNN的情况下，会把CPU的Buffer对齐为64。 -2. 深入PaddlePaddle，寻找有没有其他可以优化的可能，进一步优化。比如可能会用OpenMP改进SGD的更新性能。 - -## Design Concerns - -为了更好的符合PaddlePaddle的代码风格\[[2](#references)\]，同时又尽可能少的牺牲MKL-DNN的性能\[[3](#references)\]。 - -我们总结出一些特别需要注意的点： - -1. 使用**deviceId_**。为了尽可能少的在父类Layer中添加变量或者函数，我们决定使用已有的`deviceId_`变量来区分layer的属性，定义`-2`为`MKLDNNLayer`特有的设备ID。 -2. 重写父类Layer的**init**函数，修改`deviceId_`为`-2`，代表这个layer是用于跑在MKL-DNN的环境下。 -3. 创建`MKLDNNMatrix`，同时继承`CpuMatrix`和`mkldnn::memory`。用于管理MKL-DNN会用到的相关memory函数、接口以及会用的到格式信息。 -4. 创建`MKLDNNBase`，定义一些除了layer和memory相关的类和函数。包括MKL-DNN会用到`MKLDNNStream`和`CPUEngine`，和未来可能还会用到`FPGAEngine`等。 -5. 每个`MKLDNNlayer`都会有`inVal_`,`inGrad_`,`outVal_`和`outGrad_`，分别代表input value， input gradient，output value和output gradient。他们会存放MKL-DNN用到的internal memory。同时还会定义以*ext*开头的`MKLDNNMatrix`(表示external的memory)，主要是在格式与PaddlePaddle默认的`nchw`格式不匹配时，用于转换内存的工作。必要的转换函数也会在`MKLDNNLayer`中提前定义好，每个子类只需要调用定义好的reset buffer函数即可。 -6. 每个`MKLDNNlayer`的resetbuffer相关的函数（包括reset input、output的Value和grad），他们会根据输入参数reset internal和external的memory，当然这两者也可以相等，即表示不需要转换。只需要把握一个原则，每个`MKLDNNlayer`的子类，只需要使用internal的memory就可以了，所有external的转换工作在父类的reset函数中都提前准备好了。 -7. 一般来说，external的memory会尽量与PaddlePaddle中的`value`和`grad`共享内存。同时每个`MKLDNNLayer`中的external output value和gradient(也就是`extOutVal_`和`extOutGrad_`)必须分别与`output_.value`和`output_.grad`共享内存，因为PaddlePaddle的activation会直接使用`output_.value`和`output_.grad`。如果不需要external的buffer用于转换，那么internal的buffer也会与他们共享内存。 -8. 如果MKL-DNN layer的后面接有cpu device，那么就会使`output_.value`与`extOutVal_`共享内存，同时数据格式就是`nchw`，这样下一个cpu device就能拿到正确的数据。在有cpu device的时候，external的memory的格式始终是`nchw`或者`nc`。 -9. 由于MKL-DNN的输出操作都是覆盖data的，不是在原来的数据上累加，所以当网络出现分支时，在`backward`时会需要merge不同layer的梯度。`MKLDNNlayer`中会实现merge的方法，此时每个小分支的input gradient会先临时保存在一个`MKLDNNMatrix`中，由分支处的layer负责求和，并把结果放到这个layer的`output_.grad`中。所以整体上，每个子类并不会需要关心分支的事情，也是在父类都实现好了。 -10. 在原来的`FLAGS`中添加一个`use_mkldnn`的flag，用于选择是否使用MKL-DNN的相关功能。 - -## References - -1. [Intel Math Kernel Library for Deep Neural Networks (Intel MKL-DNN)](https://github.com/01org/mkl-dnn "Intel MKL-DNN") -2. [原来的方案](https://github.com/PaddlePaddle/Paddle/pull/3096)会引入**nextLayer**的信息。但是在PaddlePaddle中，无论是重构前的layer还是重构后的op，都不会想要知道next layer/op的信息。 -3. MKL-DNN的高性能格式与PaddlePaddle原有的`NCHW`不同(PaddlePaddle中的CUDNN部分使用的也是`NCHW`，所以不存在这个问题)，所以需要引入一个转换方法，并且只需要在必要的时候转换这种格式，才能更好的发挥MKL-DNN的性能。 - diff --git a/doc/design/mkldnn/image/overview.png b/doc/design/mkldnn/image/overview.png deleted file mode 100644 index 84b455c28230703599a2529f014cfbb222138fef..0000000000000000000000000000000000000000 Binary files a/doc/design/mkldnn/image/overview.png and /dev/null differ diff --git a/doc/design/operator_kernel_type.md b/doc/design/operator_kernel_type.md new file mode 100644 index 0000000000000000000000000000000000000000..aa82e96bf79319f1a57e2ad58aa9826e57be6470 --- /dev/null +++ b/doc/design/operator_kernel_type.md @@ -0,0 +1,91 @@ +# Design Doc: The Keys of Operator Kernel Type +## Problem +An operator can have different kernel implementations, and each operator will have a map to store the related kernels. Fluid uses `OpKernelType` as a key to identify a unique Kernel. Before an operator runs, an certain kernel must be chosen by a key of `OpKernelType`. Currently, `OpKernelType` is defined as follows: + +```cpp +struct OpKernelType { + platform::Place place_; + proto::DataType data_type_; +}; +``` +For more details, please refer to [codes](https://github.com/PaddlePaddle/Paddle/blob/2d5ec16bc8a09fb8e0f62c89b116b0cd1d333907/paddle/framework/operator.h#L348-L374) in github. + +It contains two keys, `Place` and `DataType`. And these two keys will be hashed to a unique key to represent a certain type of kernel. However, these two keys are not enough. We need a more complete representation of `OpKernelType`. + +We often implement a kernel of an operator with some computing library in certain device(place). Please remind that computing library and device are not one-to-one corresponding. A device can have a lot of computing libraries and a computing library can also support several devices. + +For example, Eigen library can support Nvidia GPU/AMD GPU/CPU. And MKLDNN library can support Intel CPU/Intel FPGA. Both `Place` and `Library` should be a key of `OpKernelType`. + +It's obvious that different DataTypes, like fp64/fp32/int8 will have different kernels. But the data layout of a Tensor will also lead to different implementation. Please refer to the batch norm operator [kernels](https://github.com/PaddlePaddle/Paddle/blob/a948fac4d0ad7e0412d373b8aabeb711c2899563/paddle/operators/batch_norm_op.cc#L180-L209). Data Layout should also be taken into consideration. + +## Solution + +There are four keys to determine a kernel type of an operator: `Place`/`Library`/`DataType`/`Layout`. + +```cpp +struct OpKernelType { + platform::Place place_; + platform::Library library_; + proto::DataType data_type_; + framework::Layout layout_; +}; +``` + +Following is the details: + +### Place + +`Place` is defined as follows: + +```cpp +typedef boost::variant Place; +``` + +`Place` is to represent the device memory where data is locating. + + +### Library + +One operator kernel is usually implemented based on one library. `Library` is defined as a enum variable: + +```cpp +enum Library { Plain, MKLDNN, CUDNN }; +``` + +We use `Plain` enumerator to represent default library. Since most operators in Fluid are implemented based on `Eigen` library, we take `Eigen` library as the `Plain` enumerator. +A library usually has a corresponding `DeviceContext` which contains some handles needed by computation. Fluid now have two default DeviceContexts in CPU and CUDA, `CPUDeviceContext` and `CUDADeviceContext`. `CPUDeviceContext` contains a Eigen library handle and `CDUADeviceContext` contains a Eigen library handle and cuBLAS handle. + +If we want to support new Library, a new enumerator need to be added to `Library` and a new corresponding `LibraryDeviceContext` will be created. + + +### DataType + + +`DataType` is defined in [framework.proto](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/framework.proto). Currently, int32/int64/fp32/fp64 are supported. + +### Layout + +Actually, a Tensor is a view of a block of memory. Besides a pointer to the memory, we also have to get some other descriptions of this block of memory, such as shape(ddim), stride, and layout. + +Different layout leads to different implementation of operator kernel. There are mainly 4 principles we have to follow to support layout in our fluid framework. + +- We take layout as a data member of Tensor. Layout is actually a enum variable. If fluid is built with MKLDNN, then, the memory format in MKLDNN will be added into this enum variable too. + +- Users have to set layout for input data. And some operators like fill_constant/random, also have to set layout of generating data. Of course, we can have some default layout, like NCHW. + +- The inference of Layout is at run-time, not compile-time. + +- Every operator have to implement different kernels for different layouts. Let's take MKLDNN as an example, if we want to implement a MKLDNN convolution operator, we have to realize all the kernels for different layout, list at [here](http://01org.github.io/mkl-dnn/structmkldnn_1_1memory.html). And we will have a special macro to do registering kernels for MKLDNN operators. + +`Layout` is also defined as a enum variable: + +```cpp +enum Layout { + kNCHW, + kNHWC, +#ifdef PADDLE_WITH_MKLDNN + knChw8c + ... +#endif +}; +``` diff --git a/doc/design/optimizer.md b/doc/design/optimizer.md index 202b4b65103c0b7c536a9cb466c4120ce134d8c3..691081c268b848811bf5ee6d6a41edfe0f47eec0 100644 --- a/doc/design/optimizer.md +++ b/doc/design/optimizer.md @@ -79,7 +79,7 @@ class Optimizer(object): def minimize(self, loss, parameter_list): """Add operations to minimize `loss` by updating `parameter_list`. - This method combines interface `append_backward_ops()` and + This method combines interface `append_backward()` and `create_optimization_pass()` into one. """ params_grads = self.create_backward_pass(loss, parameter_list) diff --git a/doc/design/paddle_nccl.md b/doc/design/paddle_nccl.md new file mode 100644 index 0000000000000000000000000000000000000000..c7dac70998a6cfec3a6d2fc72b698ff9722e6805 --- /dev/null +++ b/doc/design/paddle_nccl.md @@ -0,0 +1,65 @@ +# Design Doc: NCCL support in Paddle Fluid + +## Abstract + +This Design Doc refers to the NCCL feature in paddle. We propose an approach to support NCCL library both on a single machine and multiple machines. We wrapper the NCCL primitives `Broadcast`, `Allreduce`, `Reduce` as operators to utilize Multi-GPU powers in one script. + + +## Motivation + +[NCCL](https://developer.nvidia.com/nccl) is a NVIDIA library support Multi-GPU communicating and optimized for NVIDIA GPUs, it provides routines such as all-gather, all-reduce, broadcast, reduce, reduce-scatter, that can achieve high bandwidth over PCIe and NVLink high-speed interconnect. With NCCL library, we can easily accelerate the training in parallel. + +- Pros +1. easily plug-in with [NCCL2](https://developer.nvidia.com/nccl) library. +1. high performance in NVIDIA GPUs. +1. MPI like primitives, which have low learning cost for users. + +- Cons +1. Only design for NVIDIA GPUs, not a general multi-device solution. +1. Although NCCL1 is opensourced under BSD license, but NCCL2 is not opensourced anymore. + +At the beginning of training, the framework needs to distribute the same parameters to every GPU, and merge the gradients at any time user interests. + +As a result, during training, we need the operations of peer to peer copy between different GPUs, aggregating gradients/parameters from GPUs, and broadcasting parameters to GPUs. Every GPU only need to run the operator with correct place information. + +Besides, it needs interfaces to synchronize model update with each different GPU Cards. + +## Implementation + +As mentioned above, we wrap the NCCL routines as several kinds of operators. Need to note that NCCL need to create Communicator between gpu at the beginning, so there is a NCCLInit operator created. + +### Transpiler + +To be compatible with [parameter server design doc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/ops/dist_train.md), the transpiler compiles the user defined operation graph into sub-graphs to be executed on different devices. + +1. The user-defined model will be a single device program + +2. Broadcast/Reduce operators between GPUs will be inserted into the program, even for the multi-node, may insert the `Send`, `Recv` operator. + + *Broadcast, AllReduce in a single machine. And Broadcast, AllReduce, [Send, Recv](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/ops/dist_train.md#graph-converter) in multiple machines* + + + +After compiling, the graph as shows + + + +Operators are added to the sub-graphs. Every GPU assigned a role of `rank0`, `rank1` etc. + +- **Broadcast**. Broadcast operator distribute initialized parameter to all the GPUs from the GPU who owns it. e.g. from`rank0` GPU. +- **AllReduce**. AllReduce operator synchronizes parameters/gradients between GPUs. AllReduce implemented in the Ring-Based communicating method, avoid of the bottle neck in a single GPU. + +Need to notice that AllReduce operator force GPUs synchronized at that point. The whole training process in asynchronous or synchronous mode depends on the AllReduce point in the graph. + +As it shown in the picture, when each GPU compute the gradient of `W`, followed with a `AllReduce` operator, accumulate the `dW` to full batch of data, then run the optimize process individually and apply the gradient to its `W`. + +- **AllReduce** + Need to note that our AllReduce operator is a ring-base AllReduce implementation. If we use the NCCL2 AllReduce primitive, every GPU optimized full batch of data, wasted (n-1) GPU compute resources. In addition, NCCL2 built-in AllReduce will only utilize the communicating resource during synchronization, then update the gradient will be a subsequent phase. In fact, we can amortize the update gradient time cost into the communicating phase. The process is +1. Every parameter has its root card. That card will responsible for aggregating the gradients from GPUs. +2. The whole model's parameter will be hashed to different root card, ensure the load balance between GPUs. +3. Logically neighberhood card will start send parameter to the next one. After one round, the parameter main card will aggregate the full gradients. +4. Then the root card will optimize the parameter. +5. This parameter card will send its optimized result to its neighberhood, then the neighberhood will send parameter to its next one. +6. Finish the sychronization round. + +The total time cost will be 2 * (n-1) * per-parameter-send-time, we reach the goal of amortize the upgrade time into communicating phase. diff --git a/doc/design/reader/README.md b/doc/design/reader/README.md index 320dccec3ddc7bfe6042f4e65b2518ea7b1ad24a..2cd4b6225b61cf374458e40afabad7745f61ba71 100644 --- a/doc/design/reader/README.md +++ b/doc/design/reader/README.md @@ -1,25 +1,25 @@ # Python Data Reader Design Doc -At training and testing time, PaddlePaddle programs need to read data. To ease the users' work to write data reading code, we define that +During the training and testing phases, PaddlePaddle programs need to read data. To help the users write code that performs reading input data, we define the following: -- A *reader* is a function that reads data (from file, network, random number generator, etc) and yields data items. -- A *reader creator* is a function that returns a reader function. -- A *reader decorator* is a function, which accepts one or more readers, and returns a reader. -- A *batch reader* is a function that reads data (from *reader*, file, network, random number generator, etc) and yields a batch of data items. +- A *reader*: A function that reads data (from file, network, random number generator, etc) and yields the data items. +- A *reader creator*: A function that returns a reader function. +- A *reader decorator*: A function, which takes in one or more readers, and returns a reader. +- A *batch reader*: A function that reads data (from *reader*, file, network, random number generator, etc) and yields a batch of data items. -and provide function which converts reader to batch reader, frequently used reader creators and reader decorators. +and also provide a function which can convert a reader to a batch reader, frequently used reader creators and reader decorators. ## Data Reader Interface -Indeed, *data reader* doesn't have to be a function that reads and yields data items. It can be any function with no parameter that creates a iterable (anything can be used in `for x in iterable`): +*Data reader* doesn't have to be a function that reads and yields data items. It can just be any function without any parameters that creates an iterable (anything can be used in `for x in iterable`) as follows: ``` iterable = data_reader() ``` -Element produced from the iterable should be a **single** entry of data, **not** a mini batch. That entry of data could be a single item, or a tuple of items. Item should be of [supported type](http://www.paddlepaddle.org/doc/ui/data_provider/pydataprovider2.html?highlight=dense_vector#input-types) (e.g., numpy 1d array of float32, int, list of int) +The item produced from the iterable should be a **single** entry of data and **not** a mini batch. The entry of data could be a single item or a tuple of items. Item should be of one of the [supported types](http://www.paddlepaddle.org/doc/ui/data_provider/pydataprovider2.html?highlight=dense_vector#input-types) (e.g., numpy 1d array of float32, int, list of int etc.) -An example implementation for single item data reader creator: +An example implementation for single item data reader creator is as follows: ```python def reader_creator_random_image(width, height): @@ -29,7 +29,7 @@ def reader_creator_random_image(width, height): return reader ``` -An example implementation for multiple item data reader creator: +An example implementation for multiple item data reader creator is as follows: ```python def reader_creator_random_image_and_label(width, height, label): def reader(): @@ -40,9 +40,10 @@ def reader_creator_random_image_and_label(width, height, label): ## Batch Reader Interface -*batch reader* can be any function with no parameter that creates a iterable (anything can be used in `for x in iterable`). The output of the iterable should be a batch (list) of data items. Each item inside the list must be a tuple. +*Batch reader* can be any function without any parameters that creates an iterable (anything can be used in `for x in iterable`). The output of the iterable should be a batch (list) of data items. Each item inside the list should be a tuple. + +Here are some valid outputs: -Here are valid outputs: ```python # a mini batch of three data items. Each data item consist three columns of data, each of which is 1. [(1, 1, 1), @@ -58,20 +59,22 @@ Here are valid outputs: Please note that each item inside the list must be a tuple, below is an invalid output: ```python # wrong, [1,1,1] needs to be inside a tuple: ([1,1,1],). - # Otherwise it's ambiguous whether [1,1,1] means a single column of data [1, 1, 1], - # or three column of datas, each of which is 1. + # Otherwise it is ambiguous whether [1,1,1] means a single column of data [1, 1, 1], + # or three columns of data, each of which is 1. [[1,1,1], [2,2,2], [3,3,3]] ``` -It's easy to convert from reader to batch reader: +It is easy to convert from a reader to a batch reader: + ```python mnist_train = paddle.dataset.mnist.train() mnist_train_batch_reader = paddle.batch(mnist_train, 128) ``` -Also easy to create custom batch reader: +It is also straight forward to create a custom batch reader: + ```python def custom_batch_reader(): while True: @@ -85,7 +88,8 @@ mnist_random_image_batch_reader = custom_batch_reader ## Usage -batch reader, mapping from item(s) read to data layer, batch size and number of total pass will be passed into `paddle.train`: +Following is how we can use the reader with PaddlePaddle: +The batch reader, a mapping from item(s) to data layer, the batch size and the number of total passes will be passed into `paddle.train` as follows: ```python # two data layer is created: @@ -99,13 +103,13 @@ paddle.train(batch_reader, {"image":0, "label":1}, 128, 10, ...) ## Data Reader Decorator -*Data reader decorator* takes a single or multiple data reader, returns a new data reader. It is similar to a [python decorator](https://wiki.python.org/moin/PythonDecorators), but it does not use `@` syntax. +The *Data reader decorator* takes in a single reader or multiple data readers and returns a new data reader. It is similar to a [python decorator](https://wiki.python.org/moin/PythonDecorators), but it does not use `@` in the syntax. -Since we have a strict interface for data readers (no parameter, return a single data item). Data reader can be used flexiable via data reader decorators. Following are a few examples: +Since we have a strict interface for data readers (no parameters and return a single data item), a data reader can be used in a flexible way using data reader decorators. Following are a few examples: ### Prefetch Data -Since reading data may take time and training can not proceed without data. It is generally a good idea to prefetch data. +Since reading data may take some time and training can not proceed without data, it is generally a good idea to prefetch the data. Use `paddle.reader.buffered` to prefetch data: @@ -117,9 +121,9 @@ buffered_reader = paddle.reader.buffered(paddle.dataset.mnist.train(), 100) ### Compose Multiple Data Readers -For example, we want to use a source of real images (reusing mnist dataset), and a source of random images as input for [Generative Adversarial Networks](https://arxiv.org/abs/1406.2661). +For example, if we want to use a source of real images (say reusing mnist dataset), and a source of random images as input for [Generative Adversarial Networks](https://arxiv.org/abs/1406.2661). -We can do: +We can do the following : ```python def reader_creator_random_image(width, height): @@ -139,13 +143,13 @@ false_reader = reader_creator_bool(False) reader = paddle.reader.compose(paddle.dataset.mnist.train(), data_reader_creator_random_image(20, 20), true_reader, false_reader) # Skipped 1 because paddle.dataset.mnist.train() produces two items per data entry. -# And we don't care second item at this time. +# And we don't care about the second item at this time. paddle.train(paddle.batch(reader, 128), {"true_image":0, "fake_image": 2, "true_label": 3, "false_label": 4}, ...) ``` ### Shuffle -Given shuffle buffer size `n`, `paddle.reader.shuffle` will return a data reader that buffers `n` data entries and shuffle them before a data entry is read. +Given the shuffle buffer size `n`, `paddle.reader.shuffle` returns a data reader that buffers `n` data entries and shuffles them before a data entry is read. Example: ```python @@ -154,21 +158,21 @@ reader = paddle.reader.shuffle(paddle.dataset.mnist.train(), 512) ## Q & A -### Why reader return only a single entry, but not a mini batch? +### Why does a reader return only a single entry, and not a mini batch? -Always returning a single entry make reusing existing data readers much easier (e.g., if existing reader return not a single entry but 3 entries, training code will be more complex because it need to handle cases like batch size 2). +Returning a single entry makes reusing existing data readers much easier (for example, if an existing reader returns 3 entries instead if a single entry, the training code will be more complicated because it need to handle cases like a batch size 2). -We provide function `paddle.batch` to turn (single entry) reader into batch reader. +We provide a function: `paddle.batch` to turn (a single entry) reader into a batch reader. -### Why do we need batch reader, isn't train take reader and batch_size as arguments sufficient? +### Why do we need a batch reader, isn't is sufficient to give the reader and batch_size as arguments during training ? -In most of the case, train taking reader and batch_size as arguments would be sufficent. However sometimes user want to customize order of data entries inside a mini batch. Or even change batch size dynamically. +In most of the cases, it would be sufficient to give the reader and batch_size as arguments to the train method. However sometimes the user wants to customize the order of data entries inside a mini batch, or even change the batch size dynamically. For these cases using a batch reader is very efficient and helpful. -### Why use a dictionary but not a list to provide mapping? +### Why use a dictionary instead of a list to provide mapping? -We decided to use dictionary (`{"image":0, "label":1}`) instead of list (`["image", "label"]`) is because that user can easily resue item (e.g., using `{"image_a":0, "image_b":0, "label":1}`) or skip item (e.g., using `{"image_a":0, "label":2}`). +Using a dictionary (`{"image":0, "label":1}`) instead of a list (`["image", "label"]`) gives the advantage that the user can easily reuse the items (e.g., using `{"image_a":0, "image_b":0, "label":1}`) or even skip an item (e.g., using `{"image_a":0, "label":2}`). -### How to create custom data reader creator +### How to create a custom data reader creator ? ```python def image_reader_creator(image_path, label_path, n): @@ -192,7 +196,7 @@ paddle.train(paddle.batch(reader, 128), {"image":0, "label":1}, ...) ### How is `paddle.train` implemented -An example implementation of paddle.train could be: +An example implementation of paddle.train is: ```python def train(batch_reader, mapping, batch_size, total_pass): diff --git a/doc/design/refactor/distributed_architecture.md b/doc/design/refactor/distributed_architecture.md index 601772bff5671c915d9567a1dbc40db249ad1bd4..f228d481e7e949e1badd7086737d58197c2ddf42 100644 --- a/doc/design/refactor/distributed_architecture.md +++ b/doc/design/refactor/distributed_architecture.md @@ -2,148 +2,131 @@ ## Abstract -PaddlePaddle v0.10.0 uses the "trainer-parameter server" -architecture. We run multiple replicated instances of trainers (runs -the same code written by the user) and parameter servers for -distributed training. This architecture served us well, but has some -limitations: +PaddlePaddle version 0.10.0 uses the "trainer-parameter server" architecture. We run multiple instances of trainers (where each trainer runs the same model) and parameter servers for distributed training. This architecture serves well, but has few limitations: -1. Need to write special code to handle tasks which should only be run - by a single trainer. E.g., initializing model and saving model. +1. There is a need to write special code that handles tasks which should only be run on a single trainer. E.g., initializing the model, saving the model etc. -2. Model parallelism is hard: need to write if-else branches conditioned - on the trainer ID to partition model onto each trainer, and manually - write the inter-model-shard communication code. +2. Model parallelism is hard: It would need all the if-else branches conditioned on the trainer ID to partition the model onto the trainers, and eventually manually writing out the inter-model-shard communication code to communicate between different trainers. -3. The user can not directly specify the parameter update rule: need - to modify the parameter server C++ code and compile a new binary. - This adds complication for researchers: A lot of extra effort is - required. Besides, the training job submission program - may not allow running arbitrary binaries. +3. The user can not directly specify the parameter update rule: This would need to modify the parameter server code and compile a new binary. This makes things more complicated for researchers: A lot of extra effort is required to make this work. Besides, the training job submission program may not allow running arbitrary binaries. -This design doc discusses PaddlePaddle's new distributed training -architecture that addresses the above limitations. +This design doc discusses PaddlePaddle's new distributed training architecture that addresses the above mentioned limitations. ## Analysis -We will assume the user writes the trainer program by Python, the same -analysis holds if the trainer program is written in C++. +The assumption is that the user writes the trainer program in either Python or C++. ### Limitation 1 -If we look at the Python code that the user writes, there are two -kinds of functionalities: +There are two basic functionalities in the trainer program: -- The training logic such as load / save model and print log. -- The neural network definition such as the definition of the data - layer, the fully connected layer, the cost function and the +1. The training logic such as loading / saving the model and printing out the logs. +2. The neural network definition such as the definition of the data layer, the fully connected layer, the cost function and the optimizer. -When we training with PaddlePaddle v0.10.0 distributedly, multiple -replicated Python instances are running on different nodes: both the -training logic and the neural network computation is replicated. +When we train using PaddlePaddle v0.10.0 in a distributed fashion, multiple instances of the same Python code are run on different nodes, hence both: the +training logic as well as the neural network computation logic, is replicated. -The tasks that should only run once all belong to the training logic, -if we only replicate the neural network computation but do **not** -replicate the training logic, the limitation could be solved. +The tasks that only need to be run once belong to the training logic. Hence if we only replicate the neural network computation part, and do **not** +replicate the training logic, the limitation mentioned above can be avoided. ### Limitation 2 -Model parallelism means running a single model on multiple nodes by -partitioning the model onto different nodes and managing the -inter-model-shard communications. +Model parallelism means that a single model is partitioned into different components and each node runs one of the component separately. This comes at the extra cost of managing the +inter-model-shard communication between nodes. -PaddlePaddle should be able to modify the neural network computation -definition to support model parallelism automatically. However, the -computation is only specified in Python code, and PaddlePaddle cannot -modify Python code. +PaddlePaddle should ideally be able to modify the neural network computation and figure out the support for model parallelism automatically. However, the +computation is only specified in Python code which sits outside of PaddlePaddle, hence PaddlePaddle can not support the feature in this setup. -Just like compiler uses an intermediate representation (IR) so that -the programmer does not need to manually optimize their code in most of -the cases - the compiler will optimize the IR: +Similar to how a compiler uses an intermediate representation (IR) so that the programmer does not need to manually optimize their code for most of the cases, we can have an intermediate representation in PaddlePaddle as well. The compiler optimizes the IR as follows: -We can have our own IR which is called [Program](../program.md). -PaddlePaddle can support model parallel by -converting the IR so the user no longer need to manually do it in -Python: +PaddlePaddle can support model parallelism by converting the IR so that the user no longer needs to manually perform the computation and operations in the Python component: +The IR for PaddlePaddle after refactoring is called a `Block`, it specifies the computation dependency graph and the variables used in the computation. ### Limitation 3 -The user can not directly specify the parameter update rule for the -parameter server because the previous implementation hard coded that -parameter server only do vector's optimization algorithm by -configuration. The user can not specify the parameter server's -computation layer by layer. +The user can not directly specify the parameter update rule for the parameter server in the Python module, since the parameter server does not use the same computation definition as the trainer. Instead, the update rule is baked inside the parameter server. The user can not specify the update rule explicitly. -This could be fixed by making the parameter server run a separated -IR according to the trainer's variable (tensors, selectedrows) -definition. - -the same -computation definition of the trainer. For a detailed explanation, -please -see -[Design Doc: Operation Graph-Based Parameter Server](./parameter_server.md) +This could be fixed by making the parameter server run the same computation definition as the trainer (the user's Python module). For a detailed explanation, refer to this document - +[Design Doc: Operation Graph Based Parameter Server](./parameter_server.md) ## Distributed Training Architecture -The new distributed training architecture can address the above -limitations. Below is the illustration: +The revamped distributed training architecture can address the above discussed limitations. Below is the illustration of how it does so: -The architecture includes major components: *PaddlePaddle Python*, -*PaddlePaddle converter* and *PaddlePaddle runtime*: +The major components are: *Python API*, *Distribute Transpiler* and *Remote Executor*. -### PaddlePaddle Python +### Python API -PaddlePaddle Python is the Python library that user's Python trainer -invoke to build the neural network topology, start training, etc. +Python API is the Python library that user's Python code invokes, to read the data, build the neural network topology, and start training, etc. ```Python -paddle.init() -input = paddle.op.recordIO("/home/data/mnist.recordio") # file stored on the cluster -img, label = input[0], input[1] -hidden = paddle.layer.fc(input=img, size=200, act=paddle.activation.Tanh()) -prediction = paddle.layer.fc(input=img, size=10, act=paddle.activation.Softmax()) -cost = paddle.layer.classification_cost(input=prediction, label=label) -optimizer = paddle.optimizer.SGD(learning_rate=0.01) -opts = optimizer.minimize(cost) -exe = RemoteExecutor(num_trainer=3, num_ps=2, GPU_per_trainer=2, sync_batches=1) -# this will init variable data on both server and trainer -exe.run(framework.default_startup_program()) -exe.sync() - -for i in range(1000): - # feed data - ... - cost, acc = exe.run(framework.default_main_program(), - fetch_list=[avg_cost, acc_out]) - print cost, acc +images = fluid.layers.data(name='pixel', shape=[1, 28, 28], dtype='float32') +label = fluid.layers.data(name='label', shape=[1], dtype='int64') +... +predict = fluid.layers.fc(input=conv_pool_2, size=10, act="softmax") +cost = fluid.layers.cross_entropy(input=predict, label=label) +avg_cost = fluid.layers.mean(x=cost) +optimizer = fluid.optimizer.Adam(learning_rate=0.01) +optimizer.minimize(avg_cost) + +train_reader = paddle.batch( + paddle.reader.shuffle( + paddle.dataset.mnist.train(), buf_size=500), + batch_size=BATCH_SIZE) + +place = fluid.CPUPlace() +exe = fluid.Executor(place) + +for pass_id in range(10): + for data in train_reader(): + loss, acc = exe.run(trainer_prog, + feed=feeder.feed(data), + fetch_list=[avg_cost]) ``` -The code above is a typical Python trainer code, the neural network -topology is built using helper functions such as -`paddle.layer.fc`. The training is done by calling `Executor.run` +The code above is a typical local training program, the "Training Program" is built using helper functions such as +`fluid.layer.fc`. The training is done by calling `Executor.run` iteratively. -#### RemoteExecutor +For more details, the implementation of IR is [Program](../program.md), and `ProgramDesc` is the protobuf type. + +[Executor](../executor.md) simply runs the `ProgramDesc`. For local training you generally use +`Executor` to run the program locally. For any kind of distributed training, you can use +`RemoteExecutor` to specify desired distributed training method with some optional arguments. + +### Distributed Transpiler + +The Distributed Transpiler automatically converts the IR (in protobuf format) to partitioned IRs. Then +the Remote Executor dispatches the new IRs to Remote Executors across the cluster. +Below are the steps that are followed : + +1. User only need to change `Executor` to `RemoteExecutor` to change local program to distributed program. +1. `RemoteExecutor` calls `Distributed Transpiler` to "transpile" user's program to several IRs representing a + distributed training program: + 1. Parse configurations from `RemoteExecutor`. + 1. Determine the type of distributed program, can be DataParallelism, ModelParallelism or Streaming. + 1. Partition the `ProgramDesc` according to type and add `send` / `recv` OP pair on the boundaries. For + DataParallelism type for example, it removes the optimization operators and add a `send` OP to the + "trainer" role, then add the optimization operators to the parameter server role within the `recv` OP. +1. Dispatch the partitioned graph to different `RemoteExecutor` in the cluster. +1. `RemoteExecutor` on each node run the received `ProgramDesc` utill the end. -As shown in the graph, `RemoteExecutor.run` sends the IR to the -PaddlePaddle cluster for Execution. You can also use parameter -`fetch_list` to interactively fetch variable back to local for + +### RemoteExecutor + +As shown in the graph, `RemoteExecutor.run` sends the IR to the cluster for Execution. +You can also use parameter `fetch_list` to interactively fetch variable back to local for log printing. The Python `RemoteExecutor` is derived from `Executor` class. -For more information about `RemoteExecutor`, please -see [Design Doc: RemoteExecutor](./remote_executor.md). - -The `RemoteExecutor.run` interface defination is: ```python run(self, @@ -167,87 +150,24 @@ run(self, `JobDesc` object describe the distributed job resource specification to run on Cluster environment. -By default, `Executor.run` starts a PaddlePaddle Cloud -[TrainingJob](https://github.com/PaddlePaddle/cloud/blob/develop/doc/autoscale/README.md#training-job-resource), -or you can run each component in the -executor by your own method: - -- Data Parallelism - ```python - if os.getenv('PLACE_PSERVER'): - exe.run_pserver() - elif os.getenv('PLACE_TRAINER'): - exe.run_trainer() - ``` -- Model Parrallelism - ```python - for part in exe.get_parralle_parts(): - exe.run_part(part) - ``` - -#### Program and Executor - -As mentioned above, the implementation of IR is [Program](../program.md). - -[Executor](../executor.md) converts and parses the IR to a preferred -graph for final execution. For local training you generally use -`Executor` to run the graph locally. For any kind of distributed -training, you can use `RemoteExecutor` to specify desired distributed -training method with some optional arguments. - -### PaddlePaddle Converter - -PaddlePaddle converter automatically converts the IR in the request -(IR and evaluation inputs/targets) from PaddlePaddle Python to new -partitioned IRs and dispatch the new IRs and evaluation inputs/targets -to different PaddlePaddle runtimes. Below are the steps: - -1. Add `feed` OP that feeds the eval inputs, and `fetch` OP that - fetches the eval targets to the IR. - -1. Extract a new computation (sub)graph with `feed` and `fetch` OP as - the boundary. The runtime does not need to run the OP that is not - dependent on the `fetch` OP. - -1. Optimizes the computation graph. - -1. Place the OPs in the graph onto different devices on different - PaddlePaddle runtime according to a placement algorithm and device - constraint specified by the user. - -1. Partition the graph according to runtime boundaries and add `send` / - `recv` OP pair on the runtime boundaries. - -1. Dispatch the partitioned graph to different PaddlePaddle runtimes. + -1. PaddlePaddle runtimes with the `fetch` OP reports evaluation - results back to the converter, the convert reports the evaluation - results back to the PaddlePaddle Python. - -The output IRs will be cached to optimize the conversion latency. +`RemoteExecutor.run` sends the `ProgramDesc` and +[TrainingJob](https://github.com/PaddlePaddle/cloud/blob/develop/doc/autoscale/README.md#training-job-resource) +to a server in the cluster which executes `RemoteExecutor.listen`. This server is responsible +to start the final Kubernetes Jobs to run the different role of `ProgramDesc`. -#### Placement Algorithm +### Placement Algorithm -Our first implementation will only support "trainer-parameter server" -placement: the parameters, initializers, and optimizers are placed on -the PaddlePaddle runtimes with the parameter server role. And -everything else will be placed on the PaddlePaddle runtimes with the -trainer role. This has the same functionality of our -"trainer-parameter server" architecture of PaddlePaddle v0.10.0, but -is more general and flexible. +Our first implementation will only support "trainer-parameter server" placement: the parameters, initializers, and optimizers are all placed on the PaddlePaddle runtimes with the parameter server role. Everything else will be placed on the PaddlePaddle runtimes with the trainer role. This has the same functionality as the "trainer-parameter server" architecture of PaddlePaddle v0.10.0, but is more generic and flexible. -In the future, we will implement the general placement algorithm, -which makes placements according to the input IR, and a model of -device computation time and device communication time. Model -parallelism requires the general placement algorithm. +In the future, a more general placement algorithm should be implemented, which makes placements according to the input IR, and a model of device computation time and device communication time. Model parallelism requires the generic placement algorithm. ### Local Training Architecture -The local training architecture will be the same as the distributed -training architecture, the differences are everything runs locally, -and there is just one PaddlePaddle runtime: +The local training architecture will be the same as the distributed training architecture, the difference is that everything runs locally, and there is just one PaddlePaddle runtime: diff --git a/doc/design/refactor/multi_cpu.md b/doc/design/refactor/multi_cpu.md new file mode 100644 index 0000000000000000000000000000000000000000..a8d8ee0422acc84835170a44eb83f9b5f0c6bb40 --- /dev/null +++ b/doc/design/refactor/multi_cpu.md @@ -0,0 +1,43 @@ +# Design Doc: Execute the Program with Multi CPU + +## Abstract + +This Design Doc propose an approach to make the user-defined Op graph +running with multi-CPU, we will use an auto transpiler to convert the user-defined +Op graph to a multi-CPU Op graph, and run `ParallelDo` Op to run the graph. + +## Transpiler + + + +After converted: + + + +## Implement + +- `Multi-CPU Transpiler` will convert the graph to a multi-CPU graph + which would be executed with multi-threads. +- `BlockingCounter` will `Init/Decrement` an atomic counter, and Blocking `Wait` + for the atomic counter become `0`: + ```cpp + BlockingCounter bc(thread_count); + for (int i = 0; i < thread_count; ++i) { + thread_pool->Start([&bc] {bc.DecrementCount(); }) + } + bc.Wait(); + ``` +- `ParallelDo` Operator + - Initialize a thread pool which is a Singleton. + - Use a block id as the input, and create run the specify Block on independent scope + with multi-threads. + - Initialize a `BlockingCounter` instance and wait until all threads are done. +- `Split` Operator will split the Input Tensor into a TensorArray. +- `Merge` merge all the gradients which calculated in different threads + with `mean/sum/max/min...` method, and then run the Optimizer Op to optimize `W`. + +## TODO + +- Improve the optimizer stage with multi-threads, since we could + assign the parameters to the different threads and execute + optimizer with multi-threads. diff --git a/doc/design/refactor/src/multi-threads.graffle b/doc/design/refactor/src/multi-threads.graffle new file mode 100644 index 0000000000000000000000000000000000000000..e71173715fff92a0a933d0c7d83599ba948552c6 Binary files /dev/null and b/doc/design/refactor/src/multi-threads.graffle differ diff --git a/doc/design/refactor/src/multi-threads/multi-threads@3x.png b/doc/design/refactor/src/multi-threads/multi-threads@3x.png new file mode 100644 index 0000000000000000000000000000000000000000..e40a869987dbbf5019d4cb03c1dab55b74d6c9f9 Binary files /dev/null and b/doc/design/refactor/src/multi-threads/multi-threads@3x.png differ diff --git a/doc/design/refactor/src/multi-threads/single-thread@3x.png b/doc/design/refactor/src/multi-threads/single-thread@3x.png new file mode 100644 index 0000000000000000000000000000000000000000..4083aebfdd45af5fbac25fa2c4176bc08c3cb44a Binary files /dev/null and b/doc/design/refactor/src/multi-threads/single-thread@3x.png differ diff --git a/doc/design/releasing_process.md b/doc/design/releasing_process.md index 62ff8f3229bbbb5bc82e4da29259baffc30c2c87..14c081ea84282e52a2e36475c3c0ea755122d154 100644 --- a/doc/design/releasing_process.md +++ b/doc/design/releasing_process.md @@ -5,8 +5,9 @@ PaddlePaddle使用git-flow branching model做分支管理，使用[Semantic Vers PaddlePaddle每次发新的版本，遵循以下流程: 1. 从`develop`分支派生出新的分支，分支名为`release/版本号`。例如，`release/0.10.0` -2. 将新分支的版本打上tag，tag为`版本号rc.Patch号`。第一个tag为`0.10.0rc1`，第二个为`0.10.0rc2`，依次类推。 -3. 对这个版本的提交，做如下几个操作: +1. 将新分支的版本打上tag，tag为`版本号rc.Patch号`。第一个tag为`0.10.0rc1`，第二个为`0.10.0rc2`，依次类推。 +1. 对这个版本的提交，做如下几个操作: + * 修改`python/setup.py.in`中的版本信息,并将`istaged`字段设为`True`。 * 编译这个版本的Docker发行镜像，发布到dockerhub。如果失败，修复Docker编译镜像问题，Patch号加一，返回第二步 * 编译这个版本的Ubuntu Deb包。如果失败，修复Ubuntu Deb包编译问题，Patch号加一，返回第二步。 * 使用Regression Test List作为检查列表，测试Docker镜像/ubuntu安装包的功能正确性 @@ -20,9 +21,9 @@ PaddlePaddle每次发新的版本，遵循以下流程: pip install twine twine upload dist/[package to upload] ``` -4. 第三步完成后，将`release/版本号`分支合入master分支，并删除`release/版本号`分支。将master分支的合入commit打上tag，tag为`版本号`。同时再将`master`分支合入`develop`分支。最后删除`release/版本号`分支。 -5. 编译master分支的Docker发行镜像，发布到dockerhub。编译ubuntu的deb包，发布到github release页面 -6. 协同完成Release Note的书写 +1. 第三步完成后，将`release/版本号`分支合入master分支，并删除`release/版本号`分支。将master分支的合入commit打上tag，tag为`版本号`。同时再将`master`分支合入`develop`分支。最后删除`release/版本号`分支。 +1. 编译master分支的Docker发行镜像，发布到dockerhub。编译ubuntu的deb包，发布到github release页面 +1. 协同完成Release Note的书写 需要注意的是: @@ -30,7 +31,7 @@ PaddlePaddle每次发新的版本，遵循以下流程: * `release/版本号`分支一旦建立，一般不允许再从`develop`分支合入`release/版本号`。这样保证`release/版本号`分支功能的封闭，方便测试人员测试PaddlePaddle的行为。 * 在`release/版本号`分支存在的时候，如果有bugfix的行为，需要将bugfix的分支同时merge到`master`, `develop`和`release/版本号`这三个分支。 -# PaddlePaddle 分支规范 +## PaddlePaddle 分支规范 PaddlePaddle开发过程使用[git-flow](http://nvie.com/posts/a-successful-git-branching-model/)分支规范，并适应github的特性做了一些区别。 @@ -47,11 +48,11 @@ PaddlePaddle开发过程使用[git-flow](http://nvie.com/posts/a-successful-git- * BugFix分支也是在开发者自己的fork版本库维护，与功能分支不同的是，BugFix分支需要分别给主版本库的`master`、`develop`与可能有的`release/版本号`分支，同时提起`Pull Request`。 -# PaddlePaddle回归测试列表 +## PaddlePaddle回归测试列表 本列表说明PaddlePaddle发版之前需要测试的功能点。 -## PaddlePaddle Book中所有章节 +### PaddlePaddle Book中所有章节 PaddlePaddle每次发版本首先要保证PaddlePaddle Book中所有章节功能的正确性。功能的正确性包括验证PaddlePaddle目前的`paddle_trainer`训练和纯使用`Python`训练模型正确性。 diff --git a/doc/design/support_new_device.md b/doc/design/support_new_device.md new file mode 100644 index 0000000000000000000000000000000000000000..f54b2b3694cc2a8f1d892792fd4d39a0484dc750 --- /dev/null +++ b/doc/design/support_new_device.md @@ -0,0 +1,250 @@ +# Design Doc: Supporting new Device/Library + +## Background + +Deep learning has a high demand for computing resources. New high-performance devices and computing libraries are appearing very frequently. Deep learning frameworks have to integrate these high-performance devices and computing libraries flexibly and efficiently. + +On one hand, hardware and computing libraries usually do not have a one-to-one correspondence. For example,Intel CPUs support Eigen and MKL computing libraries while Nvidia GPUs support Eigen and cuDNN computing libraries. We have to implement operator specific kernels for each computing library. + +On the other hand, users usually do not want to care about the low-level hardware and computing libraries when writing a neural network configuration. In Fluid, `Layer` is exposed in `Python`, and `Operator` is exposed in `C++`. Both `Layer` and `Operator` are hardware independent. + +So, how to support a new Device/Library in Fluid becomes a challenge. + + +## Basic: Integrate A New Device/Library + +For a general overview of fluid, please refer to the [overview doc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/read_source.md). + +There are mainly three parts that we have to consider while integrating a new device/library: + +- Place and DeviceContext: indicates the device id and manages hardware resources + +- Memory and Tensor: malloc/free data on certain device + +- Math Functor and OpKernel: implement computing unit on certain devices/libraries + +### Place and DeviceContext + +Please remind that device and computing library are not one-to-one corresponding. A device can have a lot of computing libraries and a computing library can also support several devices. + +#### Place +Fluid uses class [Place](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/place.h#L55) to represent the device memory where data is located. If we add another device, we have to add corresponding `DevicePlace`. + +``` + | CPUPlace +Place --| CUDAPlace + | FPGAPlace +``` + +And `Place` is defined as follows: + +``` +typedef boost::variant Place; +``` + +#### DeviceContext + +Fluid uses class [DeviceContext](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/device_context.h#L30) to manage the resources in different libraries, such as CUDA stream in `CDUADeviceContext`. There are also inheritance relationships between different kinds of `DeviceContext`. + + +``` + /-> CPUDeviceContext --> MKLDeviceContext +DeviceContext ----> CUDADeviceContext --> CUDNNDeviceContext + \-> FPGADeviceContext +``` + +An example of Nvidia GPU is as follows: + +- DeviceContext + + +``` +class DeviceContext { + virtual Place GetPlace() const = 0; +}; +``` + + +- CUDADeviceContext + + +``` +class CUDADeviceContext : public DeviceContext { + Place GetPlace() const override { return place_; } +private: + CUDAPlace place_; + cudaStream_t stream_; + cublasHandle_t cublas_handle_; + std::unique_ptr eigen_device_; // binds with stream_ +}; +``` + +- CUDNNDeviceContext + +``` +class CUDNNDeviceContext : public CUDADeviceContext { + private: + cudnnHandle_t cudnn_handle_; +}; +``` + + +### Memory and Tensor + + +#### memory module + +Fluid provides the following [memory interfaces](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/memory/memory.h#L36): + +``` +template +void* Alloc(Place place, size_t size); + +template +void Free(Place place, void* ptr); + +template +size_t Used(Place place); +``` + +To implement these interfaces, we have to implement MemoryAllocator for different Devices. + + +#### Tensor + +[Tensor](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/tensor.h#L36) holds data with some shape in a specific Place. + +```cpp +class Tensor { + public: + /*! Return a pointer to mutable memory block. */ + template + inline T* data(); + + /** + * @brief Return a pointer to mutable memory block. + * @note If not exist, then allocation. + */ + template + inline T* mutable_data(platform::Place place); + + /** + * @brief Return a pointer to mutable memory block. + * + * @param[in] dims The dimensions of the memory block. + * @param[in] place The place of the memory block. + * + * @note If not exist, then allocation. + */ + template + inline T* mutable_data(DDim dims, platform::Place place); + + /*! Resize the dimensions of the memory block. */ + inline Tensor& Resize(const DDim& dims); + + /*! Return the dimensions of the memory block. */ + inline const DDim& dims() const; + + private: + /*! holds the memory block if allocated. */ + std::shared_ptr holder_; + + /*! points to dimensions of memory block. */ + DDim dim_; +}; +``` + +`Placeholder` is used to delay memory allocation; that is, we can first define a tensor, using `Resize` to configure its shape, and then call `mutuable_data` to allocate the actual memory. + +```cpp +paddle::framework::Tensor t; +paddle::platform::CPUPlace place; +// set size first +t.Resize({2, 3}); +// allocate memory on CPU later +t.mutable_data(place); +``` + + + +### Math Functor and OpKernel + +Fluid implements computing units based on different DeviceContexts. Some computing units are shared between operators. This common part will be put in operators/math directory as basic Functors. + +Let's take [MaxOutFunctor](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/math/maxouting.h#L27) as an example: + +The interface is defined in header file. + +``` +template +class MaxOutFunctor { + public: + void operator()(const DeviceContext& context, const framework::Tensor& input, + framework::Tensor* output, int groups); +}; +``` + +CPU implemention is in .cc file + +``` +template +class MaxOutFunctor { + public: + void operator()(const platform::CPUDeviceContext& context, + const framework::Tensor& input, framework::Tensor* output, + int groups) { + ... + } +}; +``` + +CUDA implemention is in .cu file + +``` +template +class MaxOutFunctor { + public: + void operator()(const platform::CUDADeviceContext& context, + const framework::Tensor& input, framework::Tensor* output, + int groups) { + ... + } +}; +``` + + +We get computing handle from a concrete DeviceContext, and make compution on tensors. + +The implemention of `OpKernel` is similar to math functors, the extra thing we need to do is to register the OpKernel in a global map. + +Fluid provides different register interfaces in op_registry.h + + +Let's take [Crop](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/crop_op.cc#L134) operator as an example: + +In .cc file: + +``` +REGISTER_OP_CPU_KERNEL(crop, ops::CropKernel); +REGISTER_OP_CPU_KERNEL( + crop_grad, ops::CropGradKernel); +``` + +In .cu file: + +``` +REGISTER_OP_CUDA_KERNEL(crop, ops::CropKernel); +REGISTER_OP_CUDA_KERNEL( + crop_grad, ops::CropGradKernel); +``` + + +## Advanced topics: How to switch between different Device/Library + +Generally, we will impelement OpKernel for all Device/Library of an Operator. We can easily train a Convolutional Neural Network in GPU. However, some OpKernel is not sutibale on a specific Device. For example, crf operator can only run on CPU, whereas most other operators can run at GPU. To achieve high performance in such circumstance, we have to switch between different Device/Library. + + +For more details, please refer to following docs: + +- operator kernel type [doc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/operator_kernel_type.md) +- switch kernel [doc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/switch_kernel.md) diff --git a/doc/design/switch_kernel.md b/doc/design/switch_kernel.md new file mode 100644 index 0000000000000000000000000000000000000000..1846e5d9f99dd433b44ac6b5ae52893ec8f0d451 --- /dev/null +++ b/doc/design/switch_kernel.md @@ -0,0 +1,66 @@ +## Background +Every operator has many kernels because there are multiple data types, places, data layout that Fluid supports. We use the `KernelType` to describe kernel types that operators can hold. + +The `KernelType` is as follows. + +``` +struct KernelType { + Place place_; + DataType data_type_; + LayoutType layout_; +}; +``` + +The `place_` is a descriptor of the device and the computational library, e.g., `MKLDNNPlace`, `CUDAPlace`. + +The `data_type_` is the data type that this kernel performs on, e.g., `FP32`, `INT64`. Note that one kernel may have inputs with different data types. However, it will be a major `data_type`. For example, the `cross_entropy` takes `int64` as it label, and `double`/`float` as its input logit and output cost. The major `data_type` of `cross_entropy` is `float`/`double`. + +The `layout` is useful for some computational library. One example is that MKLDNN uses many kinds of layout, such as `nChw8c`. Each kind of layout will invoke the different kernel. + +## Problem + +We register a kernel for every operator and every kernel type ideally. However, it is impracticable for the following situations. + +1. Some operators, like CRF, are complicated and inefficient to be implemented on GPU. The CRF operator will only have a CPU kernel. +2. Some operators will take too many memory. It is better to force them into CPU. However, the rest of operators in this neural network will be performed on GPU, i.e., model parallel problem. +3. Some layout and place are particular. One example is that MKLDNN uses `nChw8` and there is no other library uses `nChw8c`. + +Problems under these situations are similar. We can formalise this problem as follow. + +We register kernels with types $KT = \{kt_1, kt_2, kt_3, ...\}$ for one operator. The inputs of this operator should be run on kernel type $kt_{?}$, which the $kt_{?} \notin KT$. How to cast the input of this operator from $kt_{?}$ to any of kernel type in $KT$. + +## Solution + +It is clearly that transforming inputs of an operator toadapt another kernel type is not related to the particular operator. So we should register these transformation methods as global methods. + +We can infer a kernel type from the inputs of an operators. We let this kernel type as `actual kernel type`, which means this kernel type is the actually kernel type that operator should be performed. + +We can get a kernel type by 1) The configuration of operator description. (Users may want to force use `MKL` for `conv` operator). 2) The place of the current executor. (Executor is running on GPU). This kernel type is what we expect the operator will be performed on. We let this kernel type as `expect kernel type`. + +We transform the input data from `actual` to `expect` if the expect kernel type is not as same as actual kernel type. + +The algorithm is described as follow + +```cpp +using DataTransformationFN = std::function; +using KernelTypePair = std::pair; + +map g_data_transformation_; + +void OpWithKernel::Run() { + vec inputs = ... + auto actual_kernel_type = GetActualKernelType(inputs); + + // The expected kernel type is related to actual kernel type. + // For the most operators, the expected kernel type is as same as + // actual kernel type. + // + // So we pass `actual_kernel_type` as a parameter of + // GetExpectedKernelType + auto expect_kernel_type = GetExpectedKernelType(actual_kernel_type); + + auto trans = g_data_transformation_[{actual_kernel_type, expect_kernel_type}]; + + kernel.run(trans(inputs)); +} +``` diff --git a/doc/faq/build_and_install/index_cn.rst b/doc/faq/build_and_install/index_cn.rst index f1677e216f31d79b53ac29a0afbf6fbb886a0dcd..ed8a0c7e87da133138ecfc7ba6a8217d58b8f71d 100644 --- a/doc/faq/build_and_install/index_cn.rst +++ b/doc/faq/build_and_install/index_cn.rst @@ -14,7 +14,7 @@ $ export CUDA_SO="$(\ls usr/lib64/libcuda* | xargs -I{} echo '-v {}:{}') $(\ls /usr/lib64/libnvidia* | xargs -I{} echo '-v {}:{}')" $ export DEVICES=$(\ls /dev/nvidia* | xargs -I{} echo '--device {}:{}') - $ docker run ${CUDA_SO} ${DEVICES} -it paddledev/paddlepaddle:latest-gpu + $ docker run ${CUDA_SO} ${DEVICES} -it paddlepaddle/paddle:latest-gpu 更多关于Docker的安装与使用, 请参考 `PaddlePaddle Docker 文档 `_ 。 @@ -109,3 +109,31 @@ PaddlePaddle使用avx SIMD指令提高cpu执行效率，因此错误的使用二 解决办法是： * 卸载PaddlePaddle包 :code:`pip uninstall paddle`, 清理掉老旧的PaddlePaddle安装包，使得单元测试有一个干净的环境。如果PaddlePaddle包已经在python的site-packages里面，单元测试会引用site-packages里面的python包，而不是源码目录里 :code:`/python` 目录下的python包。同时，即便设置 :code:`PYTHONPATH` 到 :code:`/python` 也没用，因为python的搜索路径是优先已经安装的python包。 + +8. 下载MKLML库失败 +------------------ + +.. code-block:: bash + + make[2]: *** [third_party/mklml/src/extern_mklml-stamp/extern_mklml-download] 错误 4 + make[1]: *** [CMakeFiles/extern_mklml.dir/all] 错误 2 + make[1]: *** 正在等待未完成的任务.... + +原因：网速或SSL链接原因，导致MKLML库下载不成功。 + +解决办法是：手动下载并安装，具体步骤如下。 + +.. code-block:: bash + + // 1. 进入对应的目录 + cd build/third_party/mklml/src/extern_mklml + + // 2. 查看包的大小， 正常情况下是75M，如果小于75M，即下载失败： + du -sh mklml_lnx_2018.0.1.20171007.tgz + + // 3. 手动下载且解压缩，并手动生成download成功标签： + wget --no-check-certificate https://github.com/01org/mkl-dnn/releases/download/v0.11/mklml_lnx_2018.0.1.20171007.tgz -c -O mklml_lnx_2018.0.1.20171007.tgz + tar zxf mklml_lnx_2018.0.1.20171007.tgz + touch ../extern_mklml-stamp/extern_mklml-download + + // 4. 接着编译即可 diff --git a/doc/getstarted/basic_usage/index_cn.rst b/doc/getstarted/basic_usage/index_cn.rst deleted file mode 100644 index b473944fc7fb89d3e0a0b330933f2226734bb5bd..0000000000000000000000000000000000000000 --- a/doc/getstarted/basic_usage/index_cn.rst +++ /dev/null @@ -1,108 +0,0 @@ -经典的线性回归任务 -================== - -PaddlePaddle是源于百度的一个深度学习平台。这份简短的介绍将向你展示如何利用PaddlePaddle来解决一个经典的线性回归问题。 - -任务简介 --------- - -我们展示如何用PaddlePaddle解决 `单变量的线性回归 `_ 问题。线性回归的输入是一批点 `(x, y)` ，其中 `y = wx + b + ε`， 而 ε 是一个符合高斯分布的随机变量。线性回归的输出是从这批点估计出来的参数 `w` 和 `b` 。 - -一个例子是房产估值。我们假设房产的价格（y）是其大小（x）的一个线性函数，那么我们可以通过收集市场上房子的大小和价格，用来估计线性函数的参数w 和 b。 - -准备数据 ------------ - -假设变量 `x` 和 `y` 的真实关系为： `y = 2x + 0.3 + ε`，这里展示如何使用观测数据来拟合这一线性关系。首先，Python代码将随机产生2000个观测点，作为线性回归的输入。下面脚本符合PaddlePaddle期待的读取数据的Python程序的模式。 - -.. code-block:: python - - # dataprovider.py - from paddle.trainer.PyDataProvider2 import * - import random - - # 定义输入数据的类型: 2个浮点数 - @provider(input_types=[dense_vector(1), dense_vector(1)],use_seq=False) - def process(settings, input_file): - for i in xrange(2000): - x = random.random() - yield [x], [2*x+0.3] - -训练模型 ------------ - -为了还原 `y = 2x + 0.3`，我们先从一条随机的直线 `y' = wx + b` 开始，然后利用观测数据调整 `w` 和 `b` 使得 `y'` 和 `y` 的差距不断减小，最终趋于接近。这个过程就是模型的训练过程，而 `w` 和 `b` 就是模型的参数，即我们的训练目标。 - -在PaddlePaddle里，该模型的网络配置如下。 - -.. code-block:: python - - # trainer_config.py - from paddle.trainer_config_helpers import * - - # 1. 定义数据来源，调用上面的process函数获得观测数据 - data_file = 'empty.list' - with open(data_file, 'w') as f: f.writelines(' ') - define_py_data_sources2(train_list=data_file, test_list=None, - module='dataprovider', obj='process',args={}) - - # 2. 学习算法。控制如何改变模型参数 w 和 b - settings(batch_size=12, learning_rate=1e-3, learning_method=MomentumOptimizer()) - - # 3. 神经网络配置 - x = data_layer(name='x', size=1) - y = data_layer(name='y', size=1) - # 线性计算网络层: ȳ = wx + b - ȳ = fc_layer(input=x, param_attr=ParamAttr(name='w'), size=1, act=LinearActivation(), bias_attr=ParamAttr(name='b')) - # 计算误差函数，即 ȳ 和真实 y 之间的距离 - cost = square_error_cost(input= ȳ, label=y) - outputs(cost) - - -这段简短的配置展示了PaddlePaddle的基本用法： - -- 第一部分定义了数据输入。一般情况下，PaddlePaddle先从一个文件列表里获得数据文件地址，然后交给用户自定义的函数（例如上面的 `process`函数）进行读入和预处理从而得到真实输入。本文中由于输入数据是随机生成的不需要读输入文件，所以放一个空列表（`empty.list`）即可。 - -- 第二部分主要是选择学习算法，它定义了模型参数改变的规则。PaddlePaddle提供了很多优秀的学习算法，这里使用一个基于momentum的随机梯度下降(SGD)算法，该算法每批量(batch)读取12个采样数据进行随机梯度计算来更新更新。 - -- 最后一部分是神经网络的配置。由于PaddlePaddle已经实现了丰富的网络层，所以很多时候你需要做的只是定义正确的网络层并把它们连接起来。这里使用了三种网络单元： - - - **数据层**：数据层 `data_layer` 是神经网络的入口，它读入数据并将它们传输到接下来的网络层。这里数据层有两个，分别对应于变量 `x` 和 `y`。 - - **全连接层**：全连接层 `fc_layer` 是基础的计算单元，这里利用它建模变量之间的线性关系。计算单元是神经网络的核心，PaddlePaddle支持大量的计算单元和任意深度的网络连接，从而可以拟合任意的函数来学习复杂的数据关系。 - - **回归误差代价层**：回归误差代价层 `square_error_cost` 是众多误差代价函数层的一种，它们在训练过程作为网络的出口，用来计算模型的误差，是模型参数优化的目标函数。 - -定义了网络结构并保存为 `trainer_config.py` 之后，运行以下训练命令： - -.. code-block:: bash - - paddle train --config=trainer_config.py --save_dir=./output --num_passes=30 - -PaddlePaddle将在观测数据集上迭代训练30轮，并将每轮的模型结果存放在 `./output` 路径下。从输出日志可以看到，随着轮数增加误差代价函数的输出在不断的减小，这意味着模型在训练数据上不断的改进，直到逼近真实解：` y = 2x + 0.3 ` - -模型检验 ------------ - -训练完成后，我们希望能够检验模型的好坏。一种常用的做法是用学习的模型对另外一组测试数据进行预测，评价预测的效果。在这个例子中，由于已经知道了真实答案，我们可以直接观察模型的参数是否符合预期来进行检验。 - -PaddlePaddle将每个模型参数作为一个numpy数组单独存为一个文件，所以可以利用如下方法读取模型的参数。 - -.. code-block:: python - - import numpy as np - import os - - def load(file_name): - with open(file_name, 'rb') as f: - f.read(16) # skip header for float type. - return np.fromfile(f, dtype=np.float32) - - print 'w=%.6f, b=%.6f' % (load('output/pass-00029/w'), load('output/pass-00029/b')) - # w=1.999743, b=0.300137 - -.. image:: ./parameters.png - :align: center - :scale: 80 % - -从图中可以看到，虽然 `w` 和 `b` 都使用随机值初始化，但在起初的几轮训练中它们都在快速逼近真实值，并且后续仍在不断改进，使得最终得到的模型几乎与真实模型一致。 - -这样，我们用PaddlePaddle解决了单变量线性回归问题， 包括数据输入、模型训练和最后的结果验证。 diff --git a/doc/getstarted/basic_usage/index_en.rst b/doc/getstarted/basic_usage/index_en.rst deleted file mode 100644 index 2cc438ebbe0f97345d25354b93b4ebbd43502415..0000000000000000000000000000000000000000 --- a/doc/getstarted/basic_usage/index_en.rst +++ /dev/null @@ -1,101 +0,0 @@ -Simple Linear Regression -======================== - -PaddlePaddle is a deep learning platform open-sourced by Baidu. With PaddlePaddle, you can easily train a classic neural network within a couple lines of configuration, or you can build sophisticated models that provide state-of-the-art performance on difficult learning tasks like sentiment analysis, machine translation, image caption and so on. - -Problem Background ------------------- - -Now, to give you a hint of what using PaddlePaddle looks like, let's start with a fundamental learning problem - `simple linear regression `_: you have observed a set of two-dimensional data points of ``X`` and ``Y``, where ``X`` is an explanatory variable and ``Y`` is corresponding dependent variable, and you want to recover the underlying correlation between ``X`` and ``Y``. Linear regression can be used in many practical scenarios. For example, ``X`` can be a variable about house size, and ``Y`` a variable about house price. You can build a model that captures relationship between them by observing real estate markets. - -Prepare the Data ------------------ - -Suppose the true relationship can be characterized as ``Y = 2X + 0.3``, let's see how to recover this pattern only from observed data. Here is a piece of python code that feeds synthetic data to PaddlePaddle. The code is pretty self-explanatory, the only extra thing you need to add for PaddlePaddle is a definition of input data types. - - .. code-block:: python - - # dataprovider.py - from paddle.trainer.PyDataProvider2 import * - import random - - # define data types of input: 2 real numbers - @provider(input_types=[dense_vector(1), dense_vector(1)],use_seq=False) - def process(settings, input_file): - for i in xrange(2000): - x = random.random() - yield [x], [2*x+0.3] - -Train a NeuralNetwork ----------------------- - -To recover this relationship between ``X`` and ``Y``, we use a neural network with one layer of linear activation units and a square error cost layer. Don't worry if you are not familiar with these terminologies, it's just saying that we are starting from a random line ``Y' = wX + b`` , then we gradually adapt ``w`` and ``b`` to minimize the difference between ``Y'`` and ``Y``. Here is what it looks like in PaddlePaddle: - - .. code-block:: python - - # trainer_config.py - from paddle.trainer_config_helpers import * - - # 1. read data. Suppose you saved above python code as dataprovider.py - data_file = 'empty.list' - with open(data_file, 'w') as f: f.writelines(' ') - define_py_data_sources2(train_list=data_file, test_list=None, - module='dataprovider', obj='process',args={}) - - # 2. learning algorithm - settings(batch_size=12, learning_rate=1e-3, learning_method=MomentumOptimizer()) - - # 3. Network configuration - x = data_layer(name='x', size=1) - y = data_layer(name='y', size=1) - y_predict = fc_layer(input=x, param_attr=ParamAttr(name='w'), size=1, act=LinearActivation(), bias_attr=ParamAttr(name='b')) - cost = square_error_cost(input=y_predict, label=y) - outputs(cost) - -Some of the most fundamental usages of PaddlePaddle are demonstrated: - -- The first part shows how to feed data into PaddlePaddle. In general cases, PaddlePaddle reads raw data from a list of files, and then do some user-defined process to get real input. In this case, we only need to create a placeholder file since we are generating synthetic data on the fly. - -- The second part describes learning algorithm. It defines in what ways adjustments are made to model parameters. PaddlePaddle provides a rich set of optimizers, but a simple momentum based optimizer will suffice here, and it processes 12 data points each time. - -- Finally, the network configuration. It usually is as simple as "stacking" layers. Three kinds of layers are used in this configuration: - - **Data Layer**: a network always starts with one or more data layers. They provide input data to the rest of the network. In this problem, two data layers are used respectively for ``X`` and ``Y``. - - **FC Layer**: FC layer is short for Fully Connected Layer, which connects all the input units to current layer and does the actual computation specified as activation function. Computation layers like this are the fundamental building blocks of a deeper model. - - **Cost Layer**: in training phase, cost layers are usually the last layers of the network. They measure the performance of current model, and provide guidence to adjust parameters. - -Now that everything is ready, you can train the network with a simple command line call: - - .. code-block:: bash - - paddle train --config=trainer_config.py --save_dir=./output --num_passes=30 - - -This means that PaddlePaddle will train this network on the synthectic dataset for 30 passes, and save all the models under path ``./output``. You will see from the messages printed out during training phase that the model cost is decreasing as time goes by, which indicates we are getting a closer guess. - - -Evaluate the Model -------------------- - -Usually, a different dataset that left out during training phase should be used to evalute the models. However, we are lucky enough to know the real answer: ``w=2, b=0.3``, thus a better option is to check out model parameters directly. - -In PaddlePaddle, training is just to get a collection of model parameters, which are ``w`` and ``b`` in this case. Each parameter is saved in an individual file in the popular ``numpy`` array format. Here is the code that reads parameters from last pass. - - .. code-block:: python - - import numpy as np - import os - - def load(file_name): - with open(file_name, 'rb') as f: - f.read(16) # skip header for float type. - return np.fromfile(f, dtype=np.float32) - - print 'w=%.6f, b=%.6f' % (load('output/pass-00029/w'), load('output/pass-00029/b')) - # w=1.999743, b=0.300137 - - .. image:: parameters.png - :align: center - -Although starts from a random guess, you can see that value of ``w`` changes quickly towards 2 and ``b`` changes quickly towards 0.3. In the end, the predicted line is almost identical with real answer. - -There, you have recovered the underlying pattern between ``X`` and ``Y`` only from observed data. diff --git a/doc/getstarted/basic_usage/parameters.png b/doc/getstarted/basic_usage/parameters.png deleted file mode 100644 index 2ec67480951e21f0400bce1c34b3108dcd65c18c..0000000000000000000000000000000000000000 Binary files a/doc/getstarted/basic_usage/parameters.png and /dev/null differ diff --git a/doc/getstarted/build_and_install/build_from_source_cn.rst b/doc/getstarted/build_and_install/build_from_source_cn.rst new file mode 100644 index 0000000000000000000000000000000000000000..41ac07ca5674d2c121baba77c58226ad328cd681 --- /dev/null +++ b/doc/getstarted/build_and_install/build_from_source_cn.rst @@ -0,0 +1,141 @@ +从源码编译 +====================== + +.. _build_step: + +编译方法 +---------------- + +PaddlePaddle主要使用 `CMake `_ 以及GCC, G++作为编译工具。 +我们推荐您使用PaddlePaddle Docker编译环境镜像完成编译，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。 + +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。 + +编译PaddlePaddle，需要执行： + +.. code-block:: bash + + git clone https://github.com/PaddlePaddle/Paddle.git + cd Paddle + # 如果使用Docker编译环境，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 + # 如果不使用Docker编译环境，执行下面的命令 + mkdir build + cd build + cmake -DWITH_GPU=OFF -DWITH_TESTING=OFF .. + make + +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： + +.. code-block:: bash + + pip install build/python/dist/*.whl + + +.. _run_test: + +执行单元测试 +---------------- + +如果您期望在编译完成后立即执行所有的单元测试，可以按照下面的方法： + +使用Docker的情况下，设置 :code:`RUN_TEST=ON` 和 :code:`WITH_TESTING=ON` 就会在完成编译之后，立即执行单元测试。 +开启 :code:`WITH_GPU=ON` 可以指定同时执行GPU上的单元测试。 + +.. 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 + +如果不使用Docker，可以执行ctest命令即可： + +.. code-block:: bash + + mkdir build + cd build + cmake -DWITH_GPU=OFF -DWITH_TESTING=OFF .. + make + ctest + # 指定执行其中一个单元测试 test_mul_op + ctest -R test_mul_op + +.. _compile_deps: + +编译依赖 +---------------- + +PaddlePaddle编译需要使用到下面的依赖（包含但不限于），其他的依赖软件，会自动在编译时下载。 + +.. csv-table:: PaddlePaddle编译依赖 + :header: "依赖", "版本", "说明" + :widths: 10, 15, 30 + + "CMake", ">=3.2", "" + "GCC", "4.8.2", "推荐使用CentOS的devtools2" + "Python", "2.7.x", "依赖libpython2.7.so" + "pip", ">=9.0", "" + "numpy", "", "" + "SWIG", ">=2.0", "" + "Go", ">=1.8", "可选" + + +.. _build_options: + +编译选项 +---------------- + +PaddlePaddle的编译选项，包括生成CPU/GPU二进制文件、链接何种BLAS库等。 +用户可在调用cmake的时候设置它们，详细的cmake使用方法可以参考 +`官方文档 `_ 。 + +在cmake的命令行中，通过使用 ``-D`` 命令设置该类编译选项，例如： + +.. code-block:: bash + + cmake .. -DWITH_GPU=OFF + +.. csv-table:: 编译选项说明 + :header: "选项", "说明", "默认值" + :widths: 1, 7, 2 + + "WITH_GPU", "是否支持GPU", "ON" + "WITH_C_API", "是否仅编译CAPI", "OFF" + "WITH_DOUBLE", "是否使用双精度浮点数", "OFF" + "WITH_DSO", "是否运行时动态加载CUDA动态库，而非静态加载CUDA动态库。", "ON" + "WITH_AVX", "是否编译含有AVX指令集的PaddlePaddle二进制文件", "ON" + "WITH_PYTHON", "是否内嵌PYTHON解释器", "ON" + "WITH_STYLE_CHECK", "是否编译时进行代码风格检查", "ON" + "WITH_TESTING", "是否开启单元测试", "ON" + "WITH_DOC", "是否编译中英文文档", "OFF" + "WITH_SWIG_PY", "是否编译PYTHON的SWIG接口，该接口可用于预测和定制化训练", "Auto" + "WITH_GOLANG", "是否编译go语言的可容错parameter server", "ON" + "WITH_MKL", "是否使用MKL数学库，如果为否则是用OpenBLAS", "ON" + +BLAS ++++++ + +PaddlePaddle支持 `MKL `_ 和 +`OpenBlAS `_ 两种BLAS库。默认使用MKL。如果使用MKL并且机器含有AVX2指令集， +还会下载MKL-DNN数学库，详细参考 `这里 `_ 。 + +如果关闭MKL，则会使用OpenBLAS作为BLAS库。 + +CUDA/cuDNN ++++++++++++ + +PaddlePaddle在编译时/运行时会自动找到系统中安装的CUDA和cuDNN库进行编译和执行。 +使用参数 :code:`-DCUDA_ARCH_NAME=Auto` 可以指定开启自动检测SM架构，加速编译。 + +PaddlePaddle可以使用cuDNN v5.1之后的任何一个版本来编译运行，但尽量请保持编译和运行使用的cuDNN是同一个版本。 +我们推荐使用最新版本的cuDNN。 + +编译选项的设置 +++++++++++++++ + +PaddePaddle通过编译时指定路径来实现引用各种BLAS/CUDA/cuDNN库。cmake编译时，首先在系统路径（ :code:`/usr/lib:/usr/local/lib` ）中搜索这几个库，同时也会读取相关路径变量来进行搜索。 通过使用 ``-D`` 命令可以设置，例如 + +.. code-block:: bash + + cmake .. -DWITH_GPU=ON -DWITH_TESTING=OFF -DCUDNN_ROOT=/opt/cudnnv5 + +**注意：这几个编译选项的设置，只在第一次cmake的时候有效。如果之后想要重新设置，推荐清理整个编译目录（** :code:`rm -rf` ）**后，再指定。** diff --git a/doc/getstarted/build_and_install/build_from_source_en.md b/doc/getstarted/build_and_install/build_from_source_en.md deleted file mode 100644 index 2f1461489495618718d5abaeab9cbeda9b93700f..0000000000000000000000000000000000000000 --- a/doc/getstarted/build_and_install/build_from_source_en.md +++ /dev/null @@ -1,236 +0,0 @@ -Installing from Sources -========================== - -* [1. Download and Setup](#download) -* [2. Requirements](#requirements) -* [3. Build on Ubuntu](#ubuntu) -* [4. Build on Centos](#centos) - - -## Download and Setup -You can download PaddlePaddle from the [github source](https://github.com/PaddlePaddle/Paddle). - -```bash -git clone https://github.com/PaddlePaddle/Paddle paddle -cd paddle -``` -## Requirements - -To compile the source code, your computer must be equipped with the following dependencies. - -- **Compiler**: GCC >= 4.8 or Clang >= 3.3 (AppleClang >= 5.1) and gfortran compiler -- **CMake**: CMake >= 3.0 (at least CMake 3.4 on Mac OS X) -- **BLAS**: MKL, OpenBlas or ATLAS -- **Python**: only support Python 2.7 -- **Go** - -**Note:** For CUDA 7.0 and CUDA 7.5, GCC 5.0 and up are not supported! -For CUDA 8.0, GCC versions later than 5.3 are not supported! - -### Options - -PaddlePaddle supports some build options. - - - - - - - - - - - - - - - - - - - - - - - - - - -
OptionalDescription
WITH_GPUCompile PaddlePaddle with NVIDIA GPU
WITH_AVXCompile PaddlePaddle with AVX intrinsics
WITH_DSOCompile PaddlePaddle with dynamic linked CUDA
WITH_TESTINGCompile PaddlePaddle with unit testing
WITH_SWIG_PYCompile PaddlePaddle with inference api
WITH_STYLE_CHECKCompile PaddlePaddle with style check
WITH_PYTHONCompile PaddlePaddle with python interpreter
WITH_DOUBLECompile PaddlePaddle with double precision
WITH_RDMACompile PaddlePaddle with RDMA support
WITH_TIMERCompile PaddlePaddle with stats timer
WITH_PROFILERCompile PaddlePaddle with GPU profiler
WITH_DOCCompile PaddlePaddle with documentation
WITH_COVERAGECompile PaddlePaddle with code coverage
COVERALLS_UPLOADPackage code coverage data to coveralls
ON_TRAVISExclude special unit test on Travis CI
- - -**Note:** - - The GPU version works best with Cuda Toolkit 8.0 and cuDNN v5. - - Other versions like Cuda Toolkit 7.0, 7.5 and cuDNN v3, v4 are also supported. - - **To utilize cuDNN v5, Cuda Toolkit 7.5 is prerequisite and vice versa.** - -As a simple example, consider the following: - -1. **BLAS Dependencies(optional)** - - CMake will search BLAS libraries from the system. If not found, OpenBLAS will be downloaded, built and installed automatically. - To utilize preinstalled BLAS， you can simply specify MKL, OpenBLAS or ATLAS via `MKL_ROOT`, `OPENBLAS_ROOT` or `ATLAS_ROOT`. - - ```bash - # specify MKL - cmake .. -DMKL_ROOT= - # or specify OpenBLAS - cmake .. -DOPENBLAS_ROOT= - ``` - -2. **Doc Dependencies(optional)** - - To generate PaddlePaddle's documentation, install dependencies and set `-DWITH_DOC=ON` as follows: - - ```bash - pip install 'sphinx>=1.4.0' - pip install sphinx_rtd_theme recommonmark - - # install doxygen on Ubuntu - sudo apt-get install doxygen - # install doxygen on Mac OS X - brew install doxygen - - # active docs in cmake - cmake .. -DWITH_DOC=ON` - ``` - -## Build on Ubuntu 14.04 - -### Install Dependencies - -- **Paddle Dependencies** - - ```bash - # necessary - sudo apt-get update - sudo apt-get install -y git curl gcc g++ gfortran make build-essential automake - sudo apt-get install -y python python-pip python-numpy libpython-dev bison - sudo pip install 'protobuf==3.1.0.post1' - - # Install Go - # You can follow https://golang.org/doc/install for a detailed explanation. - wget -O go.tgz https://storage.googleapis.com/golang/go1.8.1.linux-amd64.tar.gz && \ - tar -C $HOME -xzf go.tgz && \ - mkdir $HOME/gopath && \ - rm go.tgz - - # Setup environment variables - export GOROOT=$HOME/go - export GOPATH=$HOME/gopath - export PATH=$PATH:$GOROOT/bin - - # install cmake 3.4 - curl -sSL https://cmake.org/files/v3.4/cmake-3.4.1.tar.gz | tar -xz && \ - cd cmake-3.4.1 && ./bootstrap && make -j4 && sudo make install && \ - cd .. && rm -rf cmake-3.4.1 - ``` - -- **GPU Dependencies (optional)** - - To build GPU version, you will need the following installed: - - 1. a CUDA-capable GPU - 2. A supported version of Linux with a GCC compiler and toolchain - 3. NVIDIA CUDA Toolkit (available at http://developer.nvidia.com/cuda-downloads) - 4. NVIDIA cuDNN Library (available at https://developer.nvidia.com/cudnn) - - The CUDA development environment relies on tight integration with the host development environment, - including the host compiler and C runtime libraries, and is therefore only supported on - distribution versions that have been qualified for this CUDA Toolkit release. - - After downloading cuDNN library, issue the following commands: - - ```bash - sudo tar -xzf cudnn-7.5-linux-x64-v5.1.tgz -C /usr/local - sudo chmod a+r /usr/local/cuda/include/cudnn.h /usr/local/cuda/lib64/libcudnn* - ``` - Then you need to set LD\_LIBRARY\_PATH, PATH environment variables in ~/.bashrc. - - ```bash - export LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH - export PATH=/usr/local/cuda/bin:$PATH - ``` - -### Build and Install - -As usual, the best option is to create build folder under paddle project directory. - -```bash -mkdir build && cd build -``` - -Finally, you can build and install PaddlePaddle: - -```bash -# you can add build option here, such as: -cmake .. -DCMAKE_INSTALL_PREFIX= -# please use sudo make install, if you want to install PaddlePaddle into the system -make -j `nproc` && make install -# set PaddlePaddle installation path in ~/.bashrc -export PATH=/bin:$PATH -# install PaddlePaddle Python modules. -sudo pip install /opt/paddle/share/wheels/*.whl -``` - -## Build on Centos 7 - -### Install Dependencies - -- **CPU Dependencies** - - ```bash - # necessary - sudo yum update - sudo yum install -y epel-release - sudo yum install -y make cmake3 python-devel python-pip gcc-gfortran swig git - sudo pip install wheel numpy - sudo pip install 'protobuf>=3.0.0' - ``` - -- **GPU Dependencies (optional)** - - To build GPU version, you will need the following installed: - - 1. a CUDA-capable GPU - 2. A supported version of Linux with a GCC compiler and toolchain - 3. NVIDIA CUDA Toolkit (available at http://developer.nvidia.com/cuda-downloads) - 4. NVIDIA cuDNN Library (available at https://developer.nvidia.com/cudnn) - - The CUDA development environment relies on tight integration with the host development environment, - including the host compiler and C runtime libraries, and is therefore only supported on - distribution versions that have been qualified for this CUDA Toolkit release. - - After downloading cuDNN library, issue the following commands: - - ```bash - sudo tar -xzf cudnn-7.5-linux-x64-v5.1.tgz -C /usr/local - sudo chmod a+r /usr/local/cuda/include/cudnn.h /usr/local/cuda/lib64/libcudnn* - ``` - Then you need to set LD\_LIBRARY\_PATH, PATH environment variables in ~/.bashrc. - - ```bash - export LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH - export PATH=/usr/local/cuda/bin:$PATH - ``` - -### Build and Install - -As usual, the best option is to create build folder under paddle project directory. - -```bash -mkdir build && cd build -``` - -Finally, you can build and install PaddlePaddle: - -```bash -# you can add build option here, such as: -cmake3 .. -DCMAKE_INSTALL_PREFIX= -# please use sudo make install, if you want to install PaddlePaddle into the system -make -j `nproc` && make install -# set PaddlePaddle installation path in ~/.bashrc -export PATH=/bin:$PATH -# install PaddlePaddle Python modules. -sudo pip install /opt/paddle/share/wheels/*.whl -``` diff --git a/doc/getstarted/build_and_install/build_from_source_en.rst b/doc/getstarted/build_and_install/build_from_source_en.rst new file mode 100644 index 0000000000000000000000000000000000000000..92211aee8c3bc0ae6e1a38311d40ddf92117cac7 --- /dev/null +++ b/doc/getstarted/build_and_install/build_from_source_en.rst @@ -0,0 +1,159 @@ +Build from Sources +========================== + +.. _build_step: + +How To Build +---------------- + +PaddlePaddle mainly uses `CMake `_ and GCC, G++ as compile +tools. We recommend you to use our pre-built Docker image to run the build +to avoid installing dependencies by yourself. We have several build environment +Docker images `here `_ . + +If you choose not to use Docker image for your build, you need to install the +below `Compile Dependencies`_ before run the build. + +Then run: + +.. code-block:: bash + + git clone https://github.com/PaddlePaddle/Paddle.git + cd Paddle + # run the following command to build a CPU-Only binaries if you are using docker + 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 + # else run these commands + mkdir build + cd build + cmake -DWITH_GPU=OFF -DWITH_TESTING=OFF .. + make + +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 +machine or copy it to the target machine. + +.. code-block:: bash + + pip install build/python/dist/*.whl + + +.. _run_test: + +Run Tests +---------------- + +If you wish to run the tests, you may follow the below steps: + +When using Docker, set :code:`RUN_TEST=ON` and :code:`WITH_TESTING=ON` will run test immediately after the build. +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 + +If you don't use Docker, just run ctest will start the tests: + +.. code-block:: bash + + mkdir build + cd build + cmake -DWITH_GPU=OFF -DWITH_TESTING=ON .. + make + ctest + # run a single test like test_mul_op + ctest -R test_mul_op + + +.. _compile_deps: + +Compile Dependencies +---------------- + +PaddlePaddle need the following dependencies when compiling, other dependencies +will be downloaded automatically. + +.. csv-table:: PaddlePaddle Compile Dependencies + :header: "Dependency", "Version", "Description" + :widths: 10, 15, 30 + + "CMake", ">=3.2", "" + "GCC", "4.8.2", "Recommend devtools2 for CentOS" + "Python", "2.7.x", "Need libpython2.7.so" + "pip", ">=9.0", "" + "numpy", "", "" + "SWIG", ">=2.0", "" + "Go", ">=1.8", "Optional" + + +.. _build_options: + +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 `_ 。 + +.. _build_options_bool: + +Bool Type Options +---------------- + +You can add :code:`-D` argument to pass such options, like: + +.. code-block:: bash + + cmake .. -DWITH_GPU=OFF + +.. csv-table:: Bool Type Options + :header: "Option", "Description", "Default" + :widths: 1, 7, 2 + + "WITH_GPU", "Build with GPU support", "ON" + "WITH_C_API", "Build only CAPI", "OFF" + "WITH_DOUBLE", "Build with double precision", "OFF" + "WITH_DSO", "Dynamically load CUDA libraries", "ON" + "WITH_AVX", "Build with AVX support", "ON" + "WITH_PYTHON", "Build with integrated Python interpreter", "ON" + "WITH_STYLE_CHECK", "Check code style when building", "ON" + "WITH_TESTING", "Build unit tests", "ON" + "WITH_DOC", "Build documentations", "OFF" + "WITH_SWIG_PY", "Build Python SWIG interface for V2 API", "Auto" + "WITH_GOLANG", "Build fault-tolerant parameter server written in go", "ON" + "WITH_MKL", "Use MKL as BLAS library, else use OpenBLAS", "ON" + + +BLAS ++++++ + +PaddlePaddle supports `MKL `_ and +`OpenBlAS `_ as BLAS library。By default it uses MKL. +If you are using MKL and your machine supports AVX2, MKL-DNN will also be downloaded +and used, for more `details `_ . + +If you choose not to use MKL, then OpenBlAS will be used. + +CUDA/cuDNN ++++++++++++ + +PaddlePaddle will automatically find CUDA and cuDNN when compiling and running. +parameter :code:`-DCUDA_ARCH_NAME=Auto` can be used to detect SM architecture +automatically in order to speed up the build. + +PaddlePaddle can build with any version later than cuDNN v5.1, and we intend to +keep on with latest cuDNN versions. Be sure to run with the same version of cuDNN +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 +:code:`/usr/lib:/usr/local/lib` and then search paths that you +passed to cmake, i.e. + +.. code-block:: bash + + cmake .. -DWITH_GPU=ON -DWITH_TESTING=OFF -DCUDNN_ROOT=/opt/cudnnv5 + +**NOTE: These options only take effect when running cmake for the first time, you need to clean the cmake cache or clean the build directory (** :code:`rm -rf` **) if you want to change it.** diff --git a/doc/getstarted/build_and_install/cmake.png b/doc/getstarted/build_and_install/cmake.png deleted file mode 100644 index a58cd09ad99cf27cc1ca5785fe54d726b83a82f6..0000000000000000000000000000000000000000 Binary files a/doc/getstarted/build_and_install/cmake.png and /dev/null differ diff --git a/doc/getstarted/build_and_install/cmake/build_from_source_cn.rst b/doc/getstarted/build_and_install/cmake/build_from_source_cn.rst deleted file mode 100644 index be0c1ffa451b2901ec06621dd4d886f800b4562e..0000000000000000000000000000000000000000 --- a/doc/getstarted/build_and_install/cmake/build_from_source_cn.rst +++ /dev/null @@ -1,43 +0,0 @@ -PaddlePaddle的编译选项 -====================== - -PaddlePaddle的编译选项，包括生成CPU/GPU二进制文件、链接何种BLAS库等。用户可在调用cmake的时候设置它们，详细的cmake使用方法可以参考 `官方文档 `_ 。 - -Bool型的编译选项 ----------------- -用户可在cmake的命令行中，通过使用 ``-D`` 命令设置该类编译选项，例如 - -.. code-block:: bash - - cmake .. -DWITH_GPU=OFF - -.. csv-table:: Bool型的编译选项 - :widths: 1, 7, 2 - :file: compile_options.csv - -BLAS/CUDA/Cudnn的编译选项 --------------------------- -BLAS -+++++ - -PaddlePaddle支持以下任意一种BLAS库：`MKL `_ ，`ATLAS `_ ，`OpenBlAS `_ 和 `REFERENCE BLAS `_ 。 - -.. csv-table:: BLAS路径相关的编译选项 - :widths: 1, 2, 7 - :file: cblas_settings.csv - -CUDA/Cudnn -+++++++++++ - -PaddlePaddle可以使用cudnn v2之后的任何一个版本来编译运行，但尽量请保持编译和运行使用的cudnn是同一个版本。 我们推荐使用最新版本的cudnn v5.1。 - -编译选项的设置 -++++++++++++++ - -PaddePaddle通过编译时指定路径来实现引用各种BLAS/CUDA/Cudnn库。cmake编译时，首先在系统路径(/usr/lib\:/usr/local/lib)中搜索这几个库，同时也会读取相关路径变量来进行搜索。 通过使用 ``-D`` 命令可以设置，例如 - -.. code-block:: bash - - cmake .. -DMKL_ROOT=/opt/mkl/ -DCUDNN_ROOT=/opt/cudnnv5 - -注意：这几个编译选项的设置，只在第一次cmake的时候有效。如果之后想要重新设置，推荐清理整个编译目录（``rm -rf``）后，再指定。 diff --git a/doc/getstarted/build_and_install/cmake/cblas_settings.csv b/doc/getstarted/build_and_install/cmake/cblas_settings.csv deleted file mode 100644 index a6356baf16a0d3d2499e39d2055d8ee878dcaef2..0000000000000000000000000000000000000000 --- a/doc/getstarted/build_and_install/cmake/cblas_settings.csv +++ /dev/null @@ -1,5 +0,0 @@ -编译选项,描述,注意 -MKL_ROOT,MKL的路径,${MKL_ROOT}/include下需要包含mkl.h，${MKL_ROOT}/lib目录下需要包含mkl_core，mkl_sequential和mkl_intel_lp64三个库。 -ATLAS_ROOT,ATLAS的路径,${ATLAS_ROOT}/include下需要包含cblas.h，${ATLAS_ROOT}/lib下需要包含cblas和atlas两个库。 -OPENBLAS_ROOT,OpenBLAS的路径,${OPENBLAS_ROOT}/include下需要包含cblas.h，${OPENBLAS_ROOT}/lib下需要包含openblas库。 -REFERENCE_CBLAS_ROOT,REFERENCE BLAS的路径,${REFERENCE_CBLAS_ROOT}/include下需要包含cblas.h，${REFERENCE_CBLAS_ROOT}/lib下需要包含cblas库。 \ No newline at end of file diff --git a/doc/getstarted/build_and_install/cmake/compile_options.csv b/doc/getstarted/build_and_install/cmake/compile_options.csv deleted file mode 100644 index 463b825470579d0c3736a408b1e82dd33e6f8d42..0000000000000000000000000000000000000000 --- a/doc/getstarted/build_and_install/cmake/compile_options.csv +++ /dev/null @@ -1,12 +0,0 @@ -选项,说明,默认值 -WITH_GPU,是否支持GPU。,取决于是否寻找到CUDA工具链 -WITH_DOUBLE,是否使用双精度浮点数。,否 -WITH_DSO,是否运行时动态加载CUDA动态库，而非静态加载CUDA动态库。,是 -WITH_AVX,是否编译含有AVX指令集的PaddlePaddle二进制文件,是 -WITH_PYTHON,是否内嵌PYTHON解释器。方便今后的嵌入式移植工作。,是 -WITH_STYLE_CHECK,是否编译时进行代码风格检查,是 -WITH_RDMA,是否开启RDMA,否 -WITH_TIMER,是否开启计时功能。如果开启会导致运行略慢，打印的日志变多，但是方便调试和测Benchmark,否 -WITH_TESTING,是否开启单元测试,取决于是否寻找到GTEST -WITH_DOC,是否编译中英文文档,否 -WITH_SWIG_PY,是否编译PYTHON的SWIG接口，该接口可用于预测和定制化训练,取决于是否寻找到SWIG \ No newline at end of file diff --git a/doc/getstarted/build_and_install/docker_install_cn.rst b/doc/getstarted/build_and_install/docker_install_cn.rst index 0d34dec8e908c5e61001500725187a2233797f46..bae42593ddc6f7a7eb47d603752ad6efa9820b45 100644 --- a/doc/getstarted/build_and_install/docker_install_cn.rst +++ b/doc/getstarted/build_and_install/docker_install_cn.rst @@ -1,222 +1,139 @@ -PaddlePaddle的Docker容器使用方式 +使用Docker安装运行 ================================ -PaddlePaddle目前唯一官方支持的运行的方式是Docker容器。因为Docker能在所有主要操作系统（包括Linux，Mac OS X和Windows）上运行。 请注意，您需要更改 `Dockers设置 `_ 才能充分利用Mac OS X和Windows上的硬件资源。 +使用Docker安装和运行PaddlePaddle可以无需考虑依赖环境即可运行。并且也可以在Windows的docker中运行。 +您可以在 `Docker官网 `_ 获得基本的Docker安装和使用方法。 -Docker使用入门 ------------------------------- - -几个基础的概念帮助理解和使用Docker： +如果您在使用Windows，可以参考 +`这篇 `_ +教程，完成在Windows上安装和使用Docker。 -- *镜像*：一个Docker镜像是一个打包好的软件。它包含了这个软件本身和它所依赖的运行环境。PaddlePaddle的Docker镜像就包含了PaddlePaddle的Python库以及其依赖的多个Python库。这样我们可以直接在Docker中运行需要的程序而不需要安装后在执行。可以执行： +在了解Docker的基本使用方法之后，即可开始下面的步骤： - .. code-block:: bash +.. _docker_pull: - docker images +获取PaddlePaddle的Docker镜像 +------------------------------ - 来列出当前系统中的所有镜像，同样可以执行： +执行下面的命令获取最新的PaddlePaddle Docker镜像，版本为cpu_avx_mkl： .. code-block:: bash - - docker pull paddlepaddle/paddle:0.10.0 - 来下载Docker镜像，paddlepaddle/paddle是从官方镜像源Dockerhub.com下载的，推荐国内用户使用docker.paddlepaddle.org/paddle下载。 + docker pull paddlepaddle/paddle -- *容器*： 如果说一个Docker镜像就是一个程序，那容器就是这个程序运行时产生的“进程”。 - 实际上，一个容器就是一个操作系统的进程，但是是运行在独立的进程空间，文件系统以及网络之上。 - 可以执行： +对于国内用户，我们提供了加速访问的镜像源： .. code-block:: bash - docker run paddlepaddle/paddle:0.10.0 + docker pull docker.paddlepaddle.org/paddle - 来使用一个镜像启动一个容器。 - -- 默认情况下，Docker容器会运行在独立的文件系统空间之上，我们无法在Docker容器中 - 访问到主机上的文件。可以通过*挂载Volume*的方式，将主机上的文件或目录挂载到 - Docker容器中。下面的命令把当前目录挂载到了容器中的 /data 目录下，容器使用 - debian镜像，并且启动后执行 :code:`ls /data`。 +下载GPU版本（cuda8.0_cudnn5_avx_mkl）的Docker镜像： .. code-block:: bash - docker run --rm -v $(pwd):/data debian ls /data - -PaddlePaddle发布的Docker镜像使用说明 ------------------------------- - -我们把PaddlePaddle的编译环境打包成一个镜像，称为开发镜像，里面涵盖了 -PaddlePaddle需要的所有编译工具。把编译出来的PaddlePaddle也打包成一个镜 -像，称为生产镜像，里面涵盖了PaddlePaddle运行所需的所有环境。每次 -PaddlePaddle发布新版本的时候都会发布对应版本的生产镜像以及开发镜像。运 -行镜像包括纯CPU版本和GPU版本以及其对应的非AVX版本。我们会在 -`dockerhub.com `_ -和国内镜像`docker.paddlepaddle.org` 提供最新 -的Docker镜像，可以在"tags"标签下找到最新的Paddle镜像版本。 - -**注意：为了方便在国内的开发者下载Docker镜像，我们提供了国内的镜像服务器供大家使用。如果您在国内，请把文档里命令中的paddlepaddle/paddle替换成docker.paddlepaddle.org/paddle。** - -1. 开发镜像：:code:`paddlepaddle/paddle:0.10.0-dev` - - 这个镜像包含了Paddle相关的开发工具以及编译和运行环境。用户可以使用开发镜像代替配置本地环境，完成开发，编译，发布， - 文档编写等工作。由于不同的Paddle的版本可能需要不同的依赖和工具，所以如果需要自行配置开发环境需要考虑版本的因素。 - 开发镜像包含了以下工具： - - - gcc/clang - - nvcc - - Python - - sphinx - - woboq - - sshd - 很多开发者会使用远程的安装有GPU的服务器工作，用户可以使用ssh登录到这台服务器上并执行 :code:`docker exec`进入开发镜像并开始工作， - 也可以在开发镜像中启动一个SSHD服务，方便开发者直接登录到镜像中进行开发: - - 以交互容器方式运行开发镜像： - - .. code-block:: bash - - docker run -it --rm -v $(pwd):/paddle paddlepaddle/paddle:0.10.0-dev /bin/bash - - 或者，可以以后台进程方式运行容器： - - .. code-block:: bash - - docker run -d -p 2202:22 -p 8888:8888 -v $(pwd):/paddle paddlepaddle/paddle:0.10.0-dev /usr/sbin/sshd -D - - 然后用密码 :code:`root` SSH进入容器： - - .. code-block:: bash - - ssh -p 2202 root@localhost - - SSH方式的一个优点是我们可以从多个终端进入容器。比如，一个终端运行vi，另一个终端运行Python。另一个好处是我们可以把PaddlePaddle容器运行在远程服务器上，并在笔记本上通过SSH与其连接。 - -2. 生产镜像：根据CPU、GPU和非AVX区分了如下4个镜像： - - - GPU/AVX：:code:`paddlepaddle/paddle:-gpu` - - GPU/no-AVX：:code:`paddlepaddle/paddle:-gpu-noavx` - - CPU/AVX：:code:`paddlepaddle/paddle:` - - CPU/no-AVX：:code:`paddlepaddle/paddle:-noavx` - - 纯CPU镜像以及GPU镜像都会用到AVX指令集，但是2008年之前生产的旧电脑不支持AVX。以下指令能检查Linux电脑是否支持AVX： - - .. code-block:: bash - - if cat /proc/cpuinfo | grep -i avx; then echo Yes; else echo No; fi - - 如果输出是No，就需要选择使用no-AVX的镜像 + docker pull paddlepaddle/paddle:latest-gpu + docker pull docker.paddlepaddle.org/paddle:latest-gpu - **注：在0.10.0之后的版本，PaddlePaddle都可以自动判断硬件是否支持AVX，所以无需判断AVX即可使用** +选择下载使用不同的BLAS库的Docker镜像： - 以上方法在GPU镜像里也能用，只是请不要忘记提前在物理机上安装GPU最新驱动。 - 为了保证GPU驱动能够在镜像里面正常运行，我们推荐使用[nvidia-docker](https://github.com/NVIDIA/nvidia-docker)来运行镜像。 - - .. code-block:: bash + .. code-block:: bash - nvidia-docker run -it --rm paddledev/paddle:0.10.0-gpu /bin/bash + # 默认是使用MKL的镜像 + docker pull paddlepaddle/paddle + # 使用OpenBLAS的镜像 + docker pull paddlepaddle/paddle:latest-openblas - 注意: 如果使用nvidia-docker存在问题，你也许可以尝试更老的方法，具体如下，但是我们并不推荐这种方法。： +下载指定版本的Docker镜像，可以从 `DockerHub网站 `_ 获取可选的tag，并执行下面的命令： - .. code-block:: bash - - export CUDA_SO="$(\ls /usr/lib64/libcuda* | xargs -I{} echo '-v {}:{}') $(\ls /usr/lib64/libnvidia* | xargs -I{} echo '-v {}:{}')" - export DEVICES=$(\ls /dev/nvidia* | xargs -I{} echo '--device {}:{}') - docker run ${CUDA_SO} ${DEVICES} -it paddledev/paddle:0.10.0-gpu + .. code-block:: bash -3. 运行以及发布您的AI程序 + docker pull paddlepaddle/paddle:[tag] + # 比如： + docker pull docker.paddlepaddle.org/paddle:0.10.0-gpu - 假设您已经完成了一个AI训练的python程序 :code:`a.py`，这个程序是您在开发机上使用开发镜像完成开发。此时您可以运行这个命令在开发机上进行测试运行： +.. _docker_run: - .. code-block:: bash +在Docker中执行PaddlePaddle训练程序 +---------------------------------- - docker run -it -v $PWD:/work paddle /work/a.py +假设您已经在当前目录（比如在/home/work）编写了一个PaddlePaddle的程序 :code:`train.py` （可以参考 +`PaddlePaddleBook `_ +编写），就可以使用下面的命令开始执行训练： - 如果要使用GPU，请运行： + .. code-block:: bash - .. code-block:: bash + cd /home/work + docker run -it -v $PWD:/work paddlepaddle/paddle /work/train.py + +上述命令中， :code:`-it` 参数说明容器已交互式运行； :code:`-v $PWD:/work` +指定将当前路径（Linux中$PWD变量会展开为当前路径的绝对路径）挂载到容器内部的 :code:`/work` +目录； :code:`paddlepaddle/paddle` 指定需要使用的容器； 最后 :code:`/work/train.py` +为容器内执行的命令，即运行训练程序。 - nvidia-docker run -it -v $PWD:/work paddle /work/a.py +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 - 这里`a.py`包含的所有依赖假设都可以在Paddle的运行容器中。如果需要包含更多的依赖、或者需要发布您的应用的镜像，可以编写`Dockerfile`使用`FROM paddledev/paddle:0.10.0` - 创建和发布自己的AI程序镜像。 +**注：PaddlePaddle Docker镜像为了减小体积，默认没有安装vim，您可以在容器中执行** :code:`apt-get install -y vim` **安装后，在容器中编辑代码。** -运行PaddlePaddle Book ---------------------- +.. _docker_run_book: -Jupyter Notebook是一个开源的web程序，大家可以通过它制作和分享带有代码、公式、图表、文字的交互式文档。用户可以通过网页浏览文档。 +使用Docker启动PaddlePaddle Book教程 +----------------------------------- +使用Docker可以快速在本地启动一个包含了PaddlePaddle官方Book教程的Jupyter Notebook，可以通过网页浏览。 PaddlePaddle Book是为用户和开发者制作的一个交互式的Jupyter Notebook。 如果您想要更深入了解deep learning，PaddlePaddle Book一定是您最好的选择。 +大家可以通过它阅读教程，或者制作和分享带有代码、公式、图表、文字的交互式文档。 我们提供可以直接运行PaddlePaddle Book的Docker镜像，直接运行： -.. code-block:: bash + .. code-block:: bash - docker run -p 8888:8888 paddlepaddle/book + docker run -p 8888:8888 paddlepaddle/book 然后在浏览器中输入以下网址： -.. code-block:: text + .. code-block:: text - http://localhost:8888/ + http://localhost:8888/ 就这么简单，享受您的旅程！ -通过Docker容器开发PaddlePaddle ------------------------------- - -开发人员可以在Docker开发镜像中开发PaddlePaddle。这样开发人员可以以一致的方式在不同的平台上工作 - Linux，Mac OS X和Windows。 - -1. 制作PaddlePaddle开发镜像 +.. _docker_run_gpu: - PaddlePaddle每次发布新版本都会发布对应的开发镜像供开发者直接使用。这里介绍如生成造这个开发镜像。 - 生成Docker镜像的方式有两个，一个是直接把一个容器转换成镜像，另一个是创建Dockerfile并运行docker build指令按照Dockerfile生成镜像。第一个方法的好处是简单快捷，适合自己实验，可以快速迭代。第二个方法的好处是Dockerfile可以把整个生成流程描述很清楚，其他人很容易看懂镜像生成过程，持续集成系统也可以简单地复现这个过程。我们采用第二个方法。Dockerfile位于PaddlePaddle repo的根目录。生成生产镜像只需要运行： +使用Docker执行GPU训练 +------------------------------ - .. code-block:: bash - - git clone https://github.com/PaddlePaddle/Paddle.git - cd Paddle - docker build -t paddle:dev . +为了保证GPU驱动能够在镜像里面正常运行，我们推荐使用 +`nvidia-docker `_ 来运行镜像。 +请不要忘记提前在物理机上安装GPU最新驱动。 - docker build这个命令的-t指定了生成的镜像的名字，这里我们用paddle:dev。到此，PaddlePaddle开发镜像就被构建完毕了。 + .. code-block:: bash -2. 制作PaddlePaddle生产镜像 + nvidia-docker run -it -v $PWD:/work paddlepaddle/paddle:latest-gpu /bin/bash - 生产镜像的生成分为两步，第一步是运行： +**注: 如果没有安装nvidia-docker，可以尝试以下的方法，将CUDA库和Linux设备挂载到Docker容器内：** - .. code-block:: bash - - docker run -v $(pwd):/paddle -e "WITH_GPU=OFF" -e "WITH_AVX=OFF" -e "WITH_TEST=ON" paddle:dev - - 以上命令会编译PaddlePaddle，生成运行程序，以及生成创建生产镜像的Dockerfile。所有生成的的文件都在build目录下。“WITH_GPU”控制生成的生产镜像是否支持GPU，“WITH_AVX”控制生成的生产镜像是否支持AVX，”WITH_TEST“控制是否生成单元测试。 + .. code-block:: bash - 第二步是运行： + export CUDA_SO="$(\ls /usr/lib64/libcuda* | xargs -I{} echo '-v {}:{}') $(\ls /usr/lib64/libnvidia* | xargs -I{} echo '-v {}:{}')" + export DEVICES=$(\ls /dev/nvidia* | xargs -I{} echo '--device {}:{}') + docker run ${CUDA_SO} ${DEVICES} -it paddlepaddle/paddle:latest-gpu - .. code-block:: bash - - docker build -t paddle:prod -f build/Dockerfile ./build +**关于AVX：** - 以上命令会按照生成的Dockerfile把生成的程序拷贝到生产镜像中并做相应的配置，最终生成名为paddle:prod的生产镜像。 +AVX是一种CPU指令集，可以加速PaddlePaddle的计算。最新的PaddlePaddle Docker镜像默认 +是开启AVX编译的，所以，如果您的电脑不支持AVX，需要单独 +`编译 <./build_from_source_cn.html>`_ PaddlePaddle为no-avx版本。 -3. 运行单元测试 - - 运行以下指令： +以下指令能检查Linux电脑是否支持AVX： .. code-block:: bash - - docker run -it -v $(pwd):/paddle paddle:dev bash -c "cd /paddle/build && ctest" - -文档 ----- - -Paddle的Docker开发镜像带有一个通过 `woboq code browser -`_ 生成的HTML版本的C++源代码，便于用户浏览C++源码。 -只要在Docker里启动PaddlePaddle的时候给它一个名字，就可以再运行另一个Nginx Docker镜像来服务HTML代码： - -.. code-block:: bash - - docker run -d --name paddle-cpu-doc paddle:0.10.0-dev - docker run -d --volumes-from paddle-cpu-doc -p 8088:80 nginx + if cat /proc/cpuinfo | grep -i avx; then echo Yes; else echo No; fi -接着我们就能够打开浏览器在 http://localhost:8088/paddle/ 浏览代码。 +如果输出是No，就需要选择使用no-AVX的镜像 diff --git a/doc/getstarted/build_and_install/docker_install_en.rst b/doc/getstarted/build_and_install/docker_install_en.rst index 94860240f6a4a9bed8a865684a8a79960489280e..56a7c68e4d39c45249fa55a964dc48b7081596a6 100644 --- a/doc/getstarted/build_and_install/docker_install_en.rst +++ b/doc/getstarted/build_and_install/docker_install_en.rst @@ -1,270 +1,146 @@ -PaddlePaddle in Docker Containers +Run in Docker Containers ================================= -Docker container is currently the only officially-supported way to -running PaddlePaddle. This is reasonable as Docker now runs on all -major operating systems including Linux, Mac OS X, and Windows. -Please be aware that you will need to change `Dockers settings -`_ to make full use -of your hardware resource on Mac OS X and Windows. +Run PaddlePaddle in Docker container so that you don't need to care about +runtime dependencies, also you can run under Windows system. You can get +tutorials at `here `_ . -Working With Docker -------------------- +If you are using Windows, please refer to +`this `_ +tutorial to start running docker under windows. -Docker is simple as long as we understand a few basic concepts: +After you've read above tutorials you may proceed the following steps. -- *image*: A Docker image is a pack of software. It could contain one or more programs and all their dependencies. For example, the PaddlePaddle's Docker image includes pre-built PaddlePaddle and Python and many Python packages. We can run a Docker image directly, other than installing all these software. We can type +.. _docker_pull: - .. code-block:: bash - - docker images +Pull PaddlePaddle Docker Image +------------------------------ - to list all images in the system. We can also run +Run the following command to download the latest Docker images, the version is cpu_avx_mkl: .. code-block:: bash - - docker pull paddlepaddle/paddle:0.10.0 - to download a Docker image, paddlepaddle/paddle in this example, - from Dockerhub.com. + docker pull paddlepaddle/paddle -- *container*: considering a Docker image a program, a container is a - "process" that runs the image. Indeed, a container is exactly an - operating system process, but with a virtualized filesystem, network - port space, and other virtualized environment. We can type +For users in China, we provide a faster mirror: .. code-block:: bash - docker run paddlepaddle/paddle:0.10.0 + docker pull docker.paddlepaddle.org/paddle - to start a container to run a Docker image, paddlepaddle/paddle in this example. - -- By default docker container have an isolated file system namespace, - we can not see the files in the host file system. By using *volume*, - mounted files in host will be visible inside docker container. - Following command will mount current dirctory into /data inside - docker container, run docker container from debian image with - command :code:`ls /data`. +Download GPU version (cuda8.0_cudnn5_avx_mkl) images: .. code-block:: bash - docker run --rm -v $(pwd):/data debian ls /data - -Usage of CPU-only and GPU Images ----------------------------------- - -We package PaddlePaddle's compile environment into a Docker image, -called the develop image, it contains all compiling tools that -PaddlePaddle needs. We package compiled PaddlePaddle program into a -Docker image as well, called the production image, it contains all -runtime environment that running PaddlePaddle needs. For each version -of PaddlePaddle, we release both of them. Production image includes -CPU-only version and a CUDA GPU version and their no-AVX versions. - -We put the docker images on `dockerhub.com -`_. You can find the -latest versions under "tags" tab at dockerhub.com. - -** NOTE: If you are in China, you can use our Docker image registry mirror to speed up the download process. To use it, please replace all paddlepaddle/paddle in the commands to docker.paddlepaddle.org/paddle.** - - -1. development image :code:`paddlepaddle/paddle:-dev` - - This image has packed related develop tools and runtime - environment. Users and developers can use this image instead of - their own local computer to accomplish development, build, - releasing, document writing etc. While different version of paddle - may depends on different version of libraries and tools, if you - want to setup a local environment, you must pay attention to the - versions. The development image contains: - - - gcc/clang - - nvcc - - Python - - sphinx - - woboq - - sshd - - Many developers use servers with GPUs, they can use ssh to login to - the server and run :code:`docker exec` to enter the docker - container and start their work. Also they can start a development - docker image with SSHD service, so they can login to the container - and start work. - -2. Production images, this image might have multiple variants: - - - GPU/AVX：:code:`paddlepaddle/paddle:-gpu` - - GPU/no-AVX：:code:`paddlepaddle/paddle:-gpu-noavx` - - CPU/AVX：:code:`paddlepaddle/paddle:` - - CPU/no-AVX：:code:`paddlepaddle/paddle:-noavx` - - Please be aware that the CPU-only and the GPU images both use the - AVX instruction set, but old computers produced before 2008 do not - support AVX. The following command checks if your Linux computer - supports AVX: - - .. code-block:: bash - - if cat /proc/cpuinfo | grep -i avx; then echo Yes; else echo No; fi - - **NOTE：versions after 0.10.0 will automatically detect system AVX support, so manual detect is not needed in this case.** - To run the CPU-only image as an interactive container: - - .. code-block:: bash - - docker run -it --rm paddlepaddle/paddle:0.10.0 /bin/bash - - Above method work with the GPU image too -- the recommended way is - using `nvidia-docker `_. - - Please install nvidia-docker first following this `tutorial - `_. - - Now you can run a GPU image: - - .. code-block:: bash - - nvidia-docker run -it --rm paddlepaddle/paddle:0.10.0-gpu /bin/bash - - -Train Model Using Python API ----------------------------- - -Our official docker image provides a runtime for PaddlePaddle -programs. The typical workflow will be as follows: - -Create a directory as workspace: - -.. code-block:: bash - - mkdir ~/workspace - -Edit a PaddlePaddle python program using your favourite editor - -.. code-block:: bash - - emacs ~/workspace/example.py - -Run the program using docker: - -.. code-block:: bash - - docker run --rm -v ~/workspace:/workspace paddlepaddle/paddle:0.10.0 python /workspace/example.py - -Or if you are using GPU for training: + docker pull paddlepaddle/paddle:latest-gpu + docker pull docker.paddlepaddle.org/paddle:latest-gpu -.. code-block:: bash +Choose between different BLAS version: - nvidia-docker run --rm -v ~/workspace:/workspace paddlepaddle/paddle:0.10.0-gpu python /workspace/example.py - -Above commands will start a docker container by running :code:`python -/workspace/example.py`. It will stop once :code:`python -/workspace/example.py` finishes. - -Another way is to tell docker to start a :code:`/bin/bash` session and -run PaddlePaddle program interactively: - -.. code-block:: bash - - docker run -it -v ~/workspace:/workspace paddlepaddle/paddle:0.10.0 /bin/bash - # now we are inside docker container - cd /workspace - python example.py - -Running with GPU is identical: - -.. code-block:: bash - - nvidia-docker run -it -v ~/workspace:/workspace paddlepaddle/paddle:0.10.0-gpu /bin/bash - # now we are inside docker container - cd /workspace - python example.py - - -Develop PaddlePaddle or Train Model Using C++ API ---------------------------------------------------- - -We will be using PaddlePaddle development image since it contains all -compiling tools and dependencies. + .. code-block:: bash -1. Build PaddlePaddle develop image + # image using MKL by default + docker pull paddlepaddle/paddle + # image using OpenBLAS + docker pull paddlepaddle/paddle:latest-openblas - Use following command to build PaddlePaddle develop image: - .. code-block:: bash +If you want to use legacy versions, choose a tag from +`DockerHub `_ +and run: - git clone https://github.com/PaddlePaddle/Paddle.git && cd Paddle - docker build -t paddle:dev . - -2. Build PaddlePaddle production image + .. code-block:: bash - There are two steps for building production image, the first step is to run: + docker pull paddlepaddle/paddle:[tag] + # i.e. + docker pull docker.paddlepaddle.org/paddle:0.10.0-gpu - .. code-block:: bash +.. _docker_run: - docker run -v $(pwd):/paddle -e "WITH_GPU=OFF" -e "WITH_AVX=OFF" -e "WITH_TEST=ON" paddle:dev +Launch your training program in Docker +-------------------------------------- - The above command will compile PaddlePaddle and create a Dockerfile for building production image. All the generated files are in the build directory. "WITH_GPU" controls if the generated production image supports GPU. "WITH_AVX" controls if the generated production image supports AVX. "WITH_TEST" controls if the unit test will be generated. +Assume that you have already written a PaddlePaddle program +named :code:`train.py` under directory :code:`/home/work` (refer to +`PaddlePaddleBook `_ +for more samples), then run the following command: - The second step is to run: + .. code-block:: bash - .. code-block:: bash + cd /home/work + docker run -it -v $PWD:/work paddlepaddle/paddle /work/train.py - docker build -t paddle:prod -f build/Dockerfile ./build +In the above command, :code:`-it` means run the container interactively; +:code:`-v $PWD:/work` means mount the current directory ($PWD will expand +to current absolute path in Linux) under :code:`/work` in the container. +:code:`paddlepaddle/paddle` to specify image to use; finnally +:code:`/work/train.py` is the command to run inside docker. - The above command will generate the production image by copying the compiled PaddlePaddle program into the image. +Also, you can go into the container shell, run or debug your code +interactively: -3. Run unit test + .. code-block:: bash + docker run -it -v $PWD:/work paddlepaddle/paddle /bin/bash + cd /work + python train.py - Following command will run unit test: +**NOTE: We did not install vim in the default docker image to reduce the image size, you can run** :code:`apt-get install -y vim` **to install it if you need to edit python files.** - .. code-block:: bash - - docker run -it -v $(pwd):/paddle paddle:dev bash -c "cd /paddle/build && ctest" +.. _docker_run_book: PaddlePaddle Book ------------------ -The Jupyter Notebook is an open-source web application that allows -you to create and share documents that contain live code, equations, -visualizations and explanatory text in a single browser. - -PaddlePaddle Book is an interactive Jupyter Notebook for users and developers. -We already exposed port 8888 for this book. If you want to +You can create a container serving PaddlePaddle Book using Jupyter Notebook in +one minute using Docker. PaddlePaddle Book is an interactive Jupyter Notebook +for users and developers.If you want to dig deeper into deep learning, PaddlePaddle Book definitely is your best choice. We provide a packaged book image, simply issue the command: -.. code-block:: bash + .. code-block:: bash - docker run -p 8888:8888 paddlepaddle/book + docker run -p 8888:8888 paddlepaddle/book Then, you would back and paste the address into the local browser: -.. code-block:: text + .. code-block:: text - http://localhost:8888/ + http://localhost:8888/ That's all. Enjoy your journey! +.. _docker_run_gpu: -Documentation -------------- +Train with Docker with GPU +------------------------------ -Paddle Docker images include an HTML version of C++ source code -generated using `woboq code browser -`_. This makes it easy -for users to browse and understand the C++ source code. +We recommend using +`nvidia-docker `_ +to run GPU training jobs. Please ensure you have latest +GPU driver installed before move on. -As long as we give the Paddle Docker container a name, we can run an -additional Nginx Docker container to serve the volume from the Paddle -container: + .. code-block:: bash -.. code-block:: bash + nvidia-docker run -it -v $PWD:/work paddlepaddle/paddle:latest-gpu /bin/bash - docker run -d --name paddle-cpu-doc paddle: - docker run -d --volumes-from paddle-cpu-doc -p 8088:80 nginx +**NOTE: If you don't have nvidia-docker installed, try the following method to mount CUDA libs and devices into the container.** + .. code-block:: bash -Then we can direct our Web browser to the HTML version of source code -at http://localhost:8088/paddle/ + export CUDA_SO="$(\ls /usr/lib64/libcuda* | xargs -I{} echo '-v {}:{}') $(\ls /usr/lib64/libnvidia* | xargs -I{} echo '-v {}:{}')" + export DEVICES=$(\ls /dev/nvidia* | xargs -I{} echo '--device {}:{}') + docker run ${CUDA_SO} ${DEVICES} -it paddlepaddle/paddle:latest-gpu + +**About AVX:** + +AVX is a kind of CPU instruction can accelerate PaddlePaddle's calculations. +The latest PaddlePaddle Docker image turns AVX on by default, so, if your +computer doesn't support AVX, you'll probably need to +`build <./build_from_source_en.html>`_ with :code:`WITH_AVX=OFF`. + +The following command will tell you whether your computer supports AVX. + + .. code-block:: bash + + if cat /proc/cpuinfo | grep -i avx; then echo Yes; else echo No; fi diff --git a/doc/getstarted/build_and_install/index_cn.rst b/doc/getstarted/build_and_install/index_cn.rst index dd9923697ab85825557aa89a08870bece7c76673..c9ba84c842b530162c92713046e64fdf82bd441b 100644 --- a/doc/getstarted/build_and_install/index_cn.rst +++ b/doc/getstarted/build_and_install/index_cn.rst @@ -6,22 +6,28 @@ 安装流程 ++++++++ -PaddlePaddle提供Docker镜像来部署环境。 +PaddlePaddle提供pip和Docker的安装方式： .. toctree:: :maxdepth: 1 - - docker_install_cn.rst + pip_install_cn.rst + docker_install_cn.rst + ../../howto/dev/build_cn.md 编译流程 ++++++++ .. warning:: - 编译流程主要推荐高级用户查看，普通用户请走安装流程。 + 建议直接使用上述安装流程，方便快速安装。只有在遇到需要独立定制的二进制时才需要编译。 .. toctree:: :maxdepth: 1 - cmake/build_from_source_cn.rst + build_from_source_cn.rst + +常见问题解答 +++++++++++ + +`常见问题解答 `_ diff --git a/doc/getstarted/build_and_install/index_en.rst b/doc/getstarted/build_and_install/index_en.rst index 8a53588e0439df8f4d5fd529b7a20262c67d4e58..32d66d63dd5b2a30d5de4a088dc80b680830cb84 100644 --- a/doc/getstarted/build_and_install/index_en.rst +++ b/doc/getstarted/build_and_install/index_en.rst @@ -1,22 +1,34 @@ Install and Build ================= -Install PaddlePaddle ----------------------- +.. _install_steps: -.. toctree:: - :maxdepth: 1 +Install Steps +++++++++ + +You can choose either pip or Docker to complete your install: + +.. toctree:: + :maxdepth: 1 + + pip_install_en.rst + docker_install_en.rst + ../../howto/dev/build_en.md - docker_install_en.rst Build from Source ----------------- .. warning:: - Please use :code:`docker` image to install paddle. The building guide is used for hacking or contributing PaddlePaddle source code. + We recommend to directly install via above installation steps, you'll only need to build PaddlePaddle from source when you need a modifed binary. .. toctree:: :maxdepth: 1 build_from_source_en.md + +FAQ +++++++++++ + +`FAQ `_ diff --git a/doc/getstarted/build_and_install/paddleci.png b/doc/getstarted/build_and_install/paddleci.png new file mode 100644 index 0000000000000000000000000000000000000000..16087ce059aa3c07ce8c927d983eb86351915825 Binary files /dev/null and b/doc/getstarted/build_and_install/paddleci.png differ diff --git a/doc/getstarted/build_and_install/pip_install_cn.rst b/doc/getstarted/build_and_install/pip_install_cn.rst new file mode 100644 index 0000000000000000000000000000000000000000..0c741e936b46eda5e7165e4ee54b545b14a28a19 --- /dev/null +++ b/doc/getstarted/build_and_install/pip_install_cn.rst @@ -0,0 +1,86 @@ +使用pip安装 +================================ + +PaddlePaddle可以使用常用的Python包管理工具 +`pip `_ +完成安装，并可以在大多数主流的Linux操作系统以及MacOS上执行。 + +.. _pip_install: + +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。 + + .. code-block:: bash + + pip install paddlepaddle + + +如果需要安装支持GPU的版本（cuda7.5_cudnn5_avx_openblas），需要执行： + + .. code-block:: bash + + pip install paddlepaddle-gpu + +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： + +如果在点击下面链接时出现如下登陆界面，点击“Log in as guest”即可开始下载： + +.. image:: paddleci.png + :scale: 50 % + :align: center + +.. csv-table:: 各个版本最新的whl包 + :header: "版本说明", "cp27-cp27mu", "cp27-cp27m", "C-API" + :widths: 1, 3, 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 `_", "`paddle.tgz `_" + "cpu_avx_openblas", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl `_", "暂无" + "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 `_", "`paddle.tgz `_" + "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 `_", "`paddle.tgz `_" + "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 `_", "`paddle.tgz `_" + +.. _pip_dependency: + +运行环境依赖 +------------------------------ + +PaddlePaddle安装包由于不仅仅包含.py程序，而且包含了C++编写的部分，所以我们确保发布的二进制包可以支持主流的Linux操作系统，比如CentOS 6以上，Ubuntu 14.04以上，MacOS 10.12以上。 + +PaddlePaddle发布的安装包会尽量对齐 `manylinux1 `_ 标准，通常使用CentOS 5作为编译环境。但由于CUDA库通常需要CentOS 6以上，而且CentOS 5即将停止维护，所以我们默认使用CentOS 6作为标准编译环境。 + +.. csv-table:: PaddlePaddle环境依赖 + :header: "依赖", "版本", "说明" + :widths: 10, 15, 30 + + "操作系统", "Linux, MacOS", "CentOS 6以上，Ubuntu 14.04以上，MacOS 10.12以上" + "Python", "2.7.x", "暂时不支持Python3" + "libc.so", "GLIBC_2.7", "glibc至少包含GLIBC_2.7以上的符号" + "libstdc++.so", "GLIBCXX_3.4.11, CXXABI_1.3.3", "至少包含GLIBCXX_3.4.11, CXXABI_1.3.3以上的符号" + "libgcc_s.so", "GCC_3.3", "至少包含GCC_3.3以上的符号" + +.. _pip_faq: + +安装常见问题和解决方法 +------------------------------ + +- paddlepaddle*.whl is not a supported wheel on this platform. + + 出现这个问题的主要原因是，没有找到和当前系统匹配的paddlepaddle安装包。请检查Python版本是否为2.7系列。另外最新的pip官方源中的安装包默认是manylinux1标准，需要使用最新的pip (>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 + + pip install --upgrade pip + + 如果仍然存在问题，可以执行： + + .. code-block:: bash + + python -c "import pip; print(pip.pep425tags.get_supported())" + + 获取当前系统支持的安装包格式，并检查和需安装的包是否匹配。pypi安装包可以在 `这个 `_ 链接中找到。 + + 如果系统支持的是 linux_x86_64 而安装包是 manylinux1_x86_64 ，需要升级pip版本到最新； 如果系统支持 manylinux1_x86_64 而安装包（本地）是 linux_x86_64 ，可以重命名这个whl包为 manylinux1_x86_64 再安装。 diff --git a/doc/getstarted/build_and_install/pip_install_en.rst b/doc/getstarted/build_and_install/pip_install_en.rst new file mode 100644 index 0000000000000000000000000000000000000000..285ed09805b09790beaef014f6813c227aff33ac --- /dev/null +++ b/doc/getstarted/build_and_install/pip_install_en.rst @@ -0,0 +1,104 @@ +Install Using pip +================================ + +You can use current widely used Python package management +tool `pip `_ +to install PaddlePaddle. This method can be used in +most of current Linux systems or MacOS. + +.. _pip_install: + +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. + + .. code-block:: bash + + pip install paddlepaddle + + +If you wish to install GPU version (cuda7.5_cudnn5_avx_openblas), just run: + + .. code-block:: bash + + pip install paddlepaddle-gpu + +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. + +If the links below shows up the login form, just click "Log in as guest" to start the download: + +.. image:: paddleci.png + :scale: 50 % + :align: center + +.. csv-table:: whl package of each version + :header: "version", "cp27-cp27mu", "cp27-cp27m", "C-API" + :widths: 1, 3, 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 `_", "`paddle.tgz `_" + "cpu_avx_openblas", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl `_", "Not Available" + "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 `_", "`paddle.tgz `_" + "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 `_", "`paddle.tgz `_" + "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 `_", "`paddle.tgz `_" + +.. _pip_dependency: + +Runtime Dependency +------------------------------ + +PaddlePaddle installation packages (whl) does not only contain .py files, +but also binaries built from C++ code. We ensure that PaddlePaddle can +run on current mainline Linux distributions, like CentOS 6, Ubuntu 14.04 +and MacOS 10.12. + +PaddlePaddle whl packages are trying to satisfy +`manylinux1 `_ +standard, which uses CentOS 5 as default build environment. But CUDA libraries +seems only run on CentOS 6 at least, also, CentOS 5 is about to end its lifetime, +so we use CentOS 6 as default build environment. + +.. csv-table:: PaddlePaddle Runtime Deps + :header: "Dependency", "version", "description" + :widths: 10, 15, 30 + + "OS", "Linux, MacOS", "CentOS 6 or later，Ubuntu 14.04 or later，MacOS 10.12 or later" + "Python", "2.7.x", "Currently Python3 is not supported" + "libc.so", "GLIBC_2.7", "glibc at least include GLIBC_2.7 symbols" + "libstdc++.so", "GLIBCXX_3.4.11, CXXABI_1.3.3", "At least include GLIBCXX_3.4.11, CXXABI_1.3.3 symbols" + "libgcc_s.so", "GCC_3.3", "At least include GCC_3.3 symbols" + +.. _pip_faq: + +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 + upgrade your pip to >9.0.0. Then run the below command: + + .. code-block:: bash + + pip install --upgrade pip + + If the problem still exists, run the following command: + + .. code-block:: bash + + python -c "import pip; print(pip.pep425tags.get_supported())" + + Then you'll get supported package suffixes, then check if it matches + the file name of the whl package. You can find default whl package at + `here `_ + + If your system supports linux_x86_64 but the whl package is manylinux1_x86_64, + you'll need to update pip to the latest version; If your system supports + manylinux1_x86_64 but the whl package is linux_x86_64 you can rename the + file to manylinux1_x86_64 suffix and then install. diff --git a/doc/getstarted/concepts/src/infer.py b/doc/getstarted/concepts/src/infer.py new file mode 100644 index 0000000000000000000000000000000000000000..4cc58dfee0bd6dade0340b4fd0ee1adb49ffebf6 --- /dev/null +++ b/doc/getstarted/concepts/src/infer.py @@ -0,0 +1,18 @@ +import paddle.v2 as paddle +import numpy as np + +paddle.init(use_gpu=False) +x = paddle.layer.data(name='x', type=paddle.data_type.dense_vector(2)) +y_predict = paddle.layer.fc(input=x, size=1, act=paddle.activation.Linear()) + +# loading the model which generated by training +with open('params_pass_90.tar', 'r') as f: + parameters = paddle.parameters.Parameters.from_tar(f) + +# Input multiple sets of data，Output the infer result in a array. +i = [[[1, 2]], [[3, 4]], [[5, 6]]] +print paddle.infer(output_layer=y_predict, parameters=parameters, input=i) +# Will print: +# [[ -3.24491572] +# [ -6.94668722] +# [-10.64845848]] diff --git a/doc/getstarted/concepts/src/train.py b/doc/getstarted/concepts/src/train.py index 8aceb23406a476f08639cc6223cdf730b728a705..4bccbfca3c70c12aec564e2cae3b8ca174b68777 100644 --- a/doc/getstarted/concepts/src/train.py +++ b/doc/getstarted/concepts/src/train.py @@ -26,6 +26,11 @@ def event_handler(event): if event.batch_id % 1 == 0: print "Pass %d, Batch %d, Cost %f" % (event.pass_id, event.batch_id, event.cost) + # product model every 10 pass + if isinstance(event, paddle.event.EndPass): + if event.pass_id % 10 == 0: + with open('params_pass_%d.tar' % event.pass_id, 'w') as f: + trainer.save_parameter_to_tar(f) # define training dataset reader diff --git a/doc/getstarted/concepts/use_concepts_cn.rst b/doc/getstarted/concepts/use_concepts_cn.rst index c243083794bb3c4659242de99b3b2715af9d7c24..e695ff283e2e806377a51c559b37e8068360a4ff 100644 --- a/doc/getstarted/concepts/use_concepts_cn.rst +++ b/doc/getstarted/concepts/use_concepts_cn.rst @@ -147,4 +147,9 @@ PaddlePaddle支持不同类型的输入数据，主要包括四种类型，和 .. literalinclude:: src/train.py :linenos: +使用以上训练好的模型进行预测，取其中一个模型params_pass_90.tar，输入需要预测的向量组，然后打印输出： + +.. literalinclude:: src/infer.py + :linenos: + 有关线性回归的实际应用，可以参考PaddlePaddle book的 `第一章节 `_。 diff --git a/doc/getstarted/index_cn.rst b/doc/getstarted/index_cn.rst index aa418c657a4ba16cce61c030066f4d3e14e891cc..9f6ee25987d51dcca3a37cf0f62a70a5a5a2d89a 100644 --- a/doc/getstarted/index_cn.rst +++ b/doc/getstarted/index_cn.rst @@ -1,10 +1,61 @@ 新手入门 ============ +.. _quick_install: + +快速安装 +++++++++ + +PaddlePaddle支持使用pip快速安装，目前支持CentOS 6以上, Ubuntu 14.04以及MacOS 10.12，并安装有Python2.7。 +执行下面的命令完成快速安装，版本为cpu_avx_openblas： + + .. code-block:: bash + + pip install paddlepaddle + +如果需要安装支持GPU的版本（cuda7.5_cudnn5_avx_openblas），需要执行： + + .. code-block:: bash + + pip install paddlepaddle-gpu + +更详细的安装和编译方法参考： + .. toctree:: :maxdepth: 1 build_and_install/index_cn.rst - concepts/use_concepts_cn.rst -- `深度学习入门课程 `_ +.. _quick_start: + +快速开始 +++++++++ + +创建一个 housing.py 并粘贴此Python代码： + + .. code-block:: python + + import paddle.v2 as paddle + + # Initialize PaddlePaddle. + paddle.init(use_gpu=False, trainer_count=1) + + # Configure the neural network. + x = paddle.layer.data(name='x', type=paddle.data_type.dense_vector(13)) + y_predict = paddle.layer.fc(input=x, size=1, act=paddle.activation.Linear()) + + # Infer using provided test data. + probs = paddle.infer( + output_layer=y_predict, + parameters=paddle.dataset.uci_housing.model(), + input=[item for item in paddle.dataset.uci_housing.test()()]) + + for i in xrange(len(probs)): + print 'Predicted price: ${:,.2f}'.format(probs[i][0] * 1000) + +执行 :code:`python housing.py` 瞧！ 它应该打印出预测住房数据的清单。 + +.. toctree:: + :maxdepth: 1 + + concepts/use_concepts_cn.rst diff --git a/doc/getstarted/index_en.rst b/doc/getstarted/index_en.rst index be3253e3d41b99a2b696e2c5ef6463ed49680d69..063d9d880c82550f7f5d47d3d0b1fff59865bca7 100644 --- a/doc/getstarted/index_en.rst +++ b/doc/getstarted/index_en.rst @@ -1,9 +1,61 @@ GET STARTED ============ +.. _quick_install: + +Quick Install +---------------------- + +You can use pip to install PaddlePaddle with a single command, supports +CentOS 6 above, Ubuntu 14.04 above or MacOS 10.12, with Python 2.7 installed. +Simply run the following command to install, the version is cpu_avx_openblas: + + .. code-block:: bash + + pip install paddlepaddle + +If you need to install GPU version (cuda7.5_cudnn5_avx_openblas), run: + + .. code-block:: bash + + pip install paddlepaddle-gpu + +For more details about installation and build: + .. toctree:: :maxdepth: 1 build_and_install/index_en.rst -- `Deep Learning 101 `_ + +.. _quick_start: + +Quick Start +++++++++ + +Create a new file called housing.py, and paste this Python +code: + + + .. code-block:: python + + import paddle.v2 as paddle + + # Initialize PaddlePaddle. + paddle.init(use_gpu=False, trainer_count=1) + + # Configure the neural network. + x = paddle.layer.data(name='x', type=paddle.data_type.dense_vector(13)) + y_predict = paddle.layer.fc(input=x, size=1, act=paddle.activation.Linear()) + + # Infer using provided test data. + probs = paddle.infer( + output_layer=y_predict, + parameters=paddle.dataset.uci_housing.model(), + input=[item for item in paddle.dataset.uci_housing.test()()]) + + for i in xrange(len(probs)): + print 'Predicted price: ${:,.2f}'.format(probs[i][0] * 1000) + +Run :code:`python housing.py` and voila! It should print out a list of predictions +for the test housing data. diff --git a/doc/howto/dev/build_cn.md b/doc/howto/dev/build_cn.md index 0b911f7b7509da4a147c65954acb7e7c38f489da..4a80a5245102fb992f513a749f6a02e1130188af 100644 --- a/doc/howto/dev/build_cn.md +++ b/doc/howto/dev/build_cn.md @@ -1,4 +1,4 @@ -# 编译PaddlePaddle和运行单元测试 +# 用Docker编译和测试PaddlePaddle ## 需要的软硬件 diff --git a/doc/howto/dev/build_en.md b/doc/howto/dev/build_en.md index d0048e3714a5861a503736879d6c0870e5906c95..91c41ef8ce3abdec5d69a9cbcebbc49b17d8f663 100644 --- a/doc/howto/dev/build_en.md +++ b/doc/howto/dev/build_en.md @@ -1,4 +1,4 @@ -# Build PaddlePaddle from Source Code and Run Unit Test +# Build using Docker ## What Developers Need diff --git a/doc/howto/dev/contribute_to_paddle_cn.md b/doc/howto/dev/contribute_to_paddle_cn.md index 699390145226ec2b65fdf5122db187e1d30d669e..3e0bf7b3973079a2063d33b6be4fe8a9dc5c07bb 100644 --- a/doc/howto/dev/contribute_to_paddle_cn.md +++ b/doc/howto/dev/contribute_to_paddle_cn.md @@ -76,18 +76,18 @@ no changes added to commit (use "git add" and/or "git commit -a") ## 构建和测试 -编译 PaddlePaddle 的源码以及生成文档需要多种开发工具。为了方便大家，我们的标准开发流程是把这些工具都装进一个Docker image，称为*开发镜像*，通常名字是 `paddle:dev`。然后所有用 `cmake && make` 的地方（比如IDE配置里）都用 `docker run paddle:dev`来代替。 +编译 PaddlePaddle 的源码以及生成文档需要多种开发工具。为了方便大家，我们的标准开发流程是把这些工具都装进一个Docker image，称为*开发镜像*，通常名字是 `paddle:latest-dev` 或者 `paddle:[version tag]-dev` 如 `paddle:0.11.0-dev`。然后所有用 `cmake && make` 的地方（比如IDE配置里）都用 `docker run paddle:latest-dev`来代替。 如要build这个开发镜像，在源码目录树的根目录中运行： ```bash -➜ docker build -t paddle:dev . +➜ docker build -t paddle:latest-dev . ``` 随后可以用这个开发镜像开始build PaddlePaddle的源码。比如如果要build一个不依赖GPU，但是支持AVX指令集，并且包括unit tests的PaddlePaddle，可以： ```bash -➜ docker run -v $(pwd):/paddle -e "WITH_GPU=OFF" -e "WITH_AVX=ON" -e "WITH_TEST=ON" paddle:dev +➜ docker run -v $(pwd):/paddle -e "WITH_GPU=OFF" -e "WITH_AVX=ON" -e "WITH_TESTING=ON" paddle:latest-dev ``` 这个过程除了编译PaddlePaddle为 `./build/libpaddle.so`，并且输出一个 `./build/paddle.deb`文件之外，还会输出一个 `build/Dockerfile`。我们只需要运行下面命令把编译好的PaddlePaddle打包成一个*生产镜像*（`paddle:prod`）： @@ -99,7 +99,7 @@ no changes added to commit (use "git add" and/or "git commit -a") 如果要运行所有的单元测试，可以用如下命令： ```bash -➜ docker run -it -v $(pwd):/paddle paddle:dev bash -c "cd /paddle/build && ctest" +➜ docker run -it -v $(pwd):/paddle paddle:latest-dev bash -c "cd /paddle/build && ctest" ``` 关于构建和测试的更多信息，请参见[这篇文档](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/getstarted/build_and_install/docker_install_cn.rst)。 diff --git a/doc/howto/dev/new_op_cn.md b/doc/howto/dev/new_op_cn.md index 6cfc9536f20e88571a9845a50be0341fe4d9f78b..3109d72001f13a38a93b9ca39d3f8525c8cea9f1 100644 --- a/doc/howto/dev/new_op_cn.md +++ b/doc/howto/dev/new_op_cn.md @@ -1,17 +1,18 @@ # 如何写新的Operator - [概念简介](#概念简介) - - [实现C++类](#实现C++类) - - [定义ProtoMaker类](#定义ProtoMaker类) - - [定义Operator类](#定义Operator类) - - [定义OpKernel类](#定义OpKernel类) - - [注册Operator](#注册Operator) + - [实现C++类](#实现c类) + - [定义ProtoMaker类](#定义protomaker类) + - [定义Operator类](#定义operator类) + - [定义OpKernel类](#定义opkernel类) + - [注册Operator](#注册operator) - [编译](#编译) - - [绑定Python](#绑定Python) + - [绑定Python](#绑定python) - [实现单元测试](#实现单元测试) - - [前向Operator单测](#前向Operator单测) - - [反向Operator单测](#反向Operator单测) + - [前向Operator单测](#前向operator单测) + - [反向Operator单测](#反向operator单测) - [编译和执行](#编译和执行) + - [注意事项](#注意事项) ## 概念简介 @@ -30,8 +31,8 @@ -------------- | :---------------------- OpProtoMake定义 | `.cc`文件，Backward Op不需要定义OpProtoMake Op定义 | `.cc`文件 -Kernel实现 | CPU、GPU共享Kernel实现在`.h`文件中，否则，CPU 实现在`.cc`文件中，GPU 实现在`.cu`文件中。 -注册Op | Op注册实现在`.cc`文件；Kernel注册CPU实现在`.cc`文件中，GPU实现在`.cu`文件中 +Kernel实现 | CPU、CUDA共享Kernel实现在`.h`文件中，否则，CPU 实现在`.cc`文件中，CUDA 实现在`.cu`文件中。 +注册Op | Op注册实现在`.cc`文件；Kernel注册CPU实现在`.cc`文件中，CUDA实现在`.cu`文件中 实现新的op都添加至目录[paddle/operators](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators)下，文件命名以`*_op.h`（如有） 、 `*_op.cc` 、`*_op.cu`（如有）结尾。**系统会根据文件名自动构建op和其对应的Python扩展。** @@ -43,7 +44,7 @@ Kernel实现 | CPU、GPU共享Kernel实现在`.h`文件中，否则，CPU ## 实现C++类 -### 1. 定义ProtoMaker类 +### 定义ProtoMaker类 矩阵乘法的公式：$Out = X * Y$, 可见该计算由两个输入，一个输出组成。 @@ -52,7 +53,7 @@ Kernel实现 | CPU、GPU共享Kernel实现在`.h`文件中，否则，CPU ```cpp class MulOpMaker : public framework::OpProtoAndCheckerMaker { public: - MulOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + MulOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "(Tensor), 2D tensor of size (M x K)"); AddInput("Y", "(Tensor), 2D tensor of size (K x N)"); @@ -81,7 +82,7 @@ The equation is: Out = X * Y template class ScaleOpMaker : public framework::OpProtoAndCheckerMaker { public: - ScaleOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + ScaleOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "The input tensor of scale operator.").NotInGradient(); AddOutput("Out", "The output tensor of scale operator.").NotInGradient(); @@ -100,7 +101,7 @@ The equation is: Out = scale*X - `AddAttr("scale", "...").SetDefault(1.0);` : 增加`scale`系数，作为参数属性，并且设置默认值为1.0。 -### 2. 定义Operator类 +### 定义Operator类 下面的点实现了MulOp的定义： @@ -149,11 +150,11 @@ MulOp(const std::string &type, const framework::VariableNameMap &inputs, 通常`OpProtoMaker`和`Op`类的定义写在`.cc`文件中，和下面将要介绍的注册函数一起放在`.cc`中 -### 3. 定义OpKernel类 +### 定义OpKernel类 `MulKernel`继承自`framework::OpKernel`，带有下面两个模板参数: -- `typename Place`: 表示设备类型，不同设备(CPU、GPU)共享同一个Kernel时，需加该模板参数，不共享则不加，一个不共享的例子是[`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43)。 +- `typename DeviceContext`: 表示设备类型，不同设备(CPU、CUDA)共享同一个Kernel时，需加该模板参数，不共享则不加，一个不共享的例子是[`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43)。 - `typename T` : 表示数据类型，如`float`, `double`等。 @@ -165,7 +166,7 @@ MulOp(const std::string &type, const framework::VariableNameMap &inputs, 下面是 `MulKernel` `Compute`的实现： ```cpp - template + template class MulKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { @@ -173,33 +174,32 @@ MulOp(const std::string &type, const framework::VariableNameMap &inputs, auto* Y = context.Input("Y"); auto* Z = context.Output("Out"); Z->mutable_data(context.GetPlace()); - auto* device_context = - const_cast(context.device_context_); - math::matmul(*X, false, *Y, false, 1, Z, 0, device_context); + auto& device_context = context.template device_context(); + math::matmul(*X, false, *Y, false, 1, Z, 0, device_context); } }; ``` -需要注意：**不同设备(CPU、GPU)共享一个Op定义，是否则共享同一个`OpKernel`，取决于`Compute`调用的函数是否支持不同设备。** +需要注意：**不同设备(CPU、CUDA)共享一个Op定义，是否则共享同一个`OpKernel`，取决于`Compute`调用的函数是否支持不同设备。** -`MulOp`的CPU、GPU实现共享同一个`Kernel`。`OpKernel`不共享的例子可以参考：[`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43)。 +`MulOp`的CPU、CUDA实现共享同一个`Kernel`。`OpKernel`不共享的例子可以参考：[`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43)。 -为了使`OpKernel`的计算过程书写更加简单，并且CPU、GPU的代码可以复用，我们通常借助 Eigen unsupported Tensor模块来实现`Compute`接口。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/howto/dev/use_eigen_cn.md)。 +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，我们通常借助 Eigen unsupported Tensor模块来实现`Compute`接口。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/howto/dev/use_eigen_cn.md)。 到此，前向Op实现完成。接下来，需要在`.cc`文件中注册该op和kernel。 反向Op类的定义，反向OpKernel的定义与前向Op类似，这里不再赘述。**但需注意反向Op没有`ProtoMaker`**。 -### 4. 注册Operator +### 注册Operator - 在`.cc`文件中注册前向、反向Op类，注册CPU Kernel。 ```cpp namespace ops = paddle::operators; REGISTER_OP(mul, ops::MulOp, ops::MulOpMaker, mul_grad, ops::MulOpGrad); - REGISTER_OP_CPU_KERNEL(mul, ops::MulKernel); + REGISTER_OP_CPU_KERNEL(mul, ops::MulKernel); REGISTER_OP_CPU_KERNEL(mul_grad, - ops::MulGradKernel); + ops::MulGradKernel); ``` 在上面的代码中： @@ -209,20 +209,20 @@ MulOp(const std::string &type, const framework::VariableNameMap &inputs, - `REGISTER_OP_CPU_KERNEL` ：注册`ops::MulKernel`类，并特化模板参数为`paddle::platform::CPUPlace`和`float`类型，同理，注册`ops::MulGradKernel`类。 -- 在 `.cu`文件中注册GPU Kernel。 - - 请注意，如果GPU Kernel的实现基于Eigen unsupported模块，那么在 `.cu`的开始请加上宏定义 `#define EIGEN_USE_GPU`，代码示例如下： +- 在 `.cu`文件中注册CUDA Kernel。 + - 请注意，如果CUDA Kernel的实现基于Eigen unsupported模块，那么在 `.cu`的开始请加上宏定义 `#define EIGEN_USE_GPU`，代码示例如下： ```cpp // if use Eigen unsupported module before include head files - // #define EIGEN_USE_GPU + #define EIGEN_USE_GPU namespace ops = paddle::operators; - REGISTER_OP_GPU_KERNEL(mul, ops::MulKernel); - REGISTER_OP_GPU_KERNEL(mul_grad, - ops::MulGradKernel); + REGISTER_OP_CUDA_KERNEL(mul, ops::MulKernel); + REGISTER_OP_CUDA_KERNEL(mul_grad, + ops::MulGradKernel); ``` -### 5. 编译 +### 编译 运行下面命令可以进行编译： @@ -236,71 +236,57 @@ make mul_op ## 实现单元测试 -单测包括对比前向Op不同设备(CPU、GPU)的实现、对比反向OP不同设备(CPU、GPU)的实现、反向Op的梯度测试。下面介绍介绍[`MulOp`的单元测试](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/framework/tests/test_mul_op.py)。 +单测包括对比前向Op不同设备(CPU、CUDA)的实现、对比反向OP不同设备(CPU、CUDA)的实现、反向Op的梯度测试。下面介绍介绍[`MulOp`的单元测试](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/framework/tests/test_mul_op.py)。 -### 前向Operator单元测试 +### 前向Operator单测 -前向Op单元测试继承自`unittest.TestCase`，并定义元类`__metaclass__ = OpTestMeta`。各项更加具体的单元测试在`OpTestMeta`里完成。测试前向Operator，需要： +Op单元测试继承自`OpTest`。各项更加具体的单元测试在`TestMulOp`里完成。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。 2. 生成随机的输入数据。 3. 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。 +4. 反向计算已经自动集成进测试框架，直接调用相应接口即可。 ```python import unittest import numpy as np - from gradient_checker import GradientChecker, create_op - from op_test_util import OpTestMeta + from op_test import OpTest - class TestMulOp(unittest.TestCase): - __metaclass__ = OpTestMeta + class TestMulOp(OpTest): def setUp(self): - self.type = "mul" + self.op_type = "mul" self.inputs = { 'X': np.random.random((32, 84)).astype("float32"), 'Y': np.random.random((84, 100)).astype("float32") } self.outputs = {'Out': np.dot(self.inputs['X'], self.inputs['Y'])} - ``` - -上面的代码首先导入依赖的包，下面是对`setUp`函数中操作的重要变量的详细解释： - -- `self.type = "mul" ` : 定义类型，与operator注册时注册的类型一致。 -- `self.inputs` : 定义输入，类型为`numpy.array`，并初始化。 -- `self.outputs` : 定义输出，并在Python脚本中完成与operator同样的计算逻辑，返回Python端的计算结果。 + def test_check_output(self): + self.check_output() -### 反向Operator单元测试 + def test_check_grad_normal(self): + self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.5) -反向Op单元测试继承自`GradientChecker`，而`GradientChecker`继承自`unittest.TestCase`，因此，**反向单元测试函数需要以`test_`开头**。 + def test_check_grad_ingore_x(self): + self.check_grad( + ['Y'], 'Out', max_relative_error=0.5, no_grad_set=set("X")) -```python -class TestMulGradOp(GradientChecker): - def setUp(self): - self.op = create_op("mul") - self.inputs = { - 'X': np.random.random((32, 84)).astype("float32"), - 'Y': np.random.random((84, 100)).astype("float32") - } - - def test_check_grad_normal(self): - # mul op will enlarge the relative error - self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.5) + def test_check_grad_ingore_y(self): + self.check_grad( + ['X'], 'Out', max_relative_error=0.5, no_grad_set=set('Y')) + ``` - def test_check_grad_ingore_x(self): - self.check_grad( - ['Y'], 'Out', max_relative_error=0.5, no_grad_set=set("X")) +上面的代码首先导入依赖的包，下面是对`setUp`函数中操作的重要变量的详细解释： - def test_check_grad_ingore_y(self): - self.check_grad( - ['X'], 'Out', max_relative_error=0.5, no_grad_set=set('Y')) -``` +- `self.op_type = "mul" ` : 定义类型，与operator注册时注册的类型一致。 +- `self.inputs` : 定义输入，类型为`numpy.array`，并初始化。 +- `self.outputs` : 定义输出，并在Python脚本中完成与operator同样的计算逻辑，返回Python端的计算结果。 -下面解释代码中一些关键的地方: +### 反向operator单测 -- 调用`create_op("mul")`创建反向Op对应的前向Op。 +而反向测试中： - `test_check_grad_normal`中调用`check_grad`使用数值法检测梯度正确性和稳定性。 - 第一个参数`["X", "Y"]` : 指定对输入变量`X`、`Y`做梯度检测。 - 第二个参数`"Out"` : 指定前向网络最终的输出目标变量`Out`。 @@ -308,7 +294,7 @@ class TestMulGradOp(GradientChecker): - `test_check_grad_ingore_x`和`test_check_grad_ingore_y`分支用来测试只需要计算一个输入梯度的情况。 -### 编译和执行单元测试 +### 编译和执行 `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。 @@ -328,5 +314,5 @@ ctest -R test_mul_op - 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，这将会导致编译出错。 - 注册Op时的类型名，需要和该Op的名字一样。即不允许在`A_op.cc`里面，注册`REGISTER_OP(B, ...)`等，这将会导致单元测试出错。 -- 如果Op没有实现GPU Kernel，请不要创建空的`*_op.cu`，这将会导致单元测试出错。 +- 如果Op没有实现CUDA Kernel，请不要创建空的`*_op.cu`，这将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。 diff --git a/doc/howto/dev/new_op_en.md b/doc/howto/dev/new_op_en.md index 1e88e1f5b4df710f1b69f0305d8d8a2921c4249a..7175d8370d6ce08c6d502eb42b8e53252db89bbb 100644 --- a/doc/howto/dev/new_op_en.md +++ b/doc/howto/dev/new_op_en.md @@ -1,8 +1,8 @@ # How to write a new operator - [Background](#background) - - [Implementing C++ Types](#implementing-c++-types) - - [Defining ProtoMaker](#defining-protoMaker) + - [Implementing C++ Types](#implementing-c-types) + - [Defining ProtoMaker](#defining-protomaker) - [Defining Operator](#defining-operator) - [Registering Operator](#registering-operator) - [Compilation](#compilation) @@ -28,8 +28,8 @@ An operator can be differentiated by whether in has kernel methods. An operator -------------- | :---------------------- OpProtoMake definition | `.cc`files, Backward Op does not need an OpProtoMake interface. Op definition | `.cc` files -Kernel implementation | The kernel methods shared between CPU and GPU are defined in `.h` files. CPU-specific kernels live in `.cc` files, while GPU-specific kernels are implemented in `.cu`files. -Registering the Op | Ops are registered in `.cc` files; For Kernel registration, `.cc` files contain the CPU implementation, while `.cu` files contain the GPU implementation. +Kernel implementation | The kernel methods shared between CPU and CUDA are defined in `.h` files. CPU-specific kernels live in `.cc` files, while CUDA-specific kernels are implemented in `.cu`files. +Registering the Op | Ops are registered in `.cc` files; For Kernel registration, `.cc` files contain the CPU implementation, while `.cu` files contain the CUDA implementation. New Operator implementations are added to the list [paddle/operators](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators), with file names in the format `*_op.h` (if applicable), `*_op.cc`, `*_op.cu` (if applicable).** The system will use the naming scheme to automatically build operators and their corresponding Python extensions. ** @@ -41,7 +41,7 @@ Let's take matrix multiplication operator, [MulOp](https://github.com/PaddlePadd ## Implementing C++ Types -### 1. Defining Class ProtoMaker +### Defining ProtoMaker Matrix Multiplication can be written as $Out = X * Y$, meaning that the operation consists of two inputs and pne output. @@ -50,7 +50,7 @@ First, define `ProtoMaker` to describe the Operator's input, output, and additio ```cpp class MulOpMaker : public framework::OpProtoAndCheckerMaker { public: - MulOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + MulOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "(Tensor), 2D tensor of size (M x K)"); AddInput("Y", "(Tensor), 2D tensor of size (K x N)"); @@ -79,7 +79,7 @@ An additional example [`ScaleOp`](https://github.com/PaddlePaddle/Paddle/blob/de template class ScaleOpMaker : public framework::OpProtoAndCheckerMaker { public: - ScaleOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + ScaleOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "The input tensor of scale operator.").NotInGradient(); AddOutput("Out", "The output tensor of scale operator.").NotInGradient(); @@ -98,7 +98,7 @@ There are two changes in this example: - `AddAttr("scale", "...").SetDefault(1.0);` adds `scale`constant as an attribute, and sets the default value to 1.0. -### 2. Defining Operator +### Defining Operator The following code defines the interface for MulOp: @@ -147,11 +147,11 @@ MulOp(const std::string &type, const framework::VariableNameMap &inputs, Usually `OpProtoMaker` and `Op`'s type definitions are written in `.cc` files, which also include the registration methods introduced later. -### 3. Defining OpKernel +### Defining OpKernel `MulKernel` inherits `framework::OpKernel`, which includes the following templates: -- `typename Place` denotes device type. When different devices, namely the CPU and the GPU, share the same kernel, this template needs to be added. If they don't share kernels, this must not be added. An example of a non-sharing kernel is [`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43). +- `typename DeviceContext` denotes device context type. When different devices, namely the CPUDeviceContext and the CUDADeviceContext, share the same kernel, this template needs to be added. If they don't share kernels, this must not be added. An example of a non-sharing kernel is [`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43). - `typename T` denotes data type, such as `float` or `double`. @@ -163,7 +163,7 @@ Usually `OpProtoMaker` and `Op`'s type definitions are written in `.cc` files, w `MulKernel`'s implementation of `Compute` is as follows: ```cpp - template + template class MulKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { @@ -171,16 +171,15 @@ Usually `OpProtoMaker` and `Op`'s type definitions are written in `.cc` files, w auto* Y = context.Input("Y"); auto* Z = context.Output("Out"); Z->mutable_data(context.GetPlace()); - auto* device_context = - const_cast(context.device_context_); - math::matmul(*X, false, *Y, false, 1, Z, 0, device_context); + auto& device_context = context.template device_context(); + math::matmul(*X, false, *Y, false, 1, Z, 0, device_context); } }; ``` -Note that **different devices (CPU, GPU)share an Op definition; whether or not they share the same `OpKernel` depends on whether `Compute` calls functions that support both devices.** +Note that **different devices (CPU, CUDA)share an Op definition; whether or not they share the same `OpKernel` depends on whether `Compute` calls functions that support both devices.** -`MulOp`'s CPU and GPU share the same `Kernel`. A non-sharing `OpKernel` example can be seen in [`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43). +`MulOp`'s CPU and CUDA share the same `Kernel`. A non-sharing `OpKernel` example can be seen in [`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43). To ease the writing of `OpKernel` compute, and for reusing code cross-device, [`Eigen-unsupported Tensor`](https://bitbucket.org/eigen/eigen/src/default/unsupported/Eigen/CXX11/src/Tensor/README.md?fileviewer=file-view-default) module is used to implement `Compute` interface. To learn about how the Eigen library is used in PaddlePaddle, please see [usage document](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/dev/use_eigen_cn.md). @@ -189,16 +188,16 @@ This concludes the forward implementation of an operator. Next its operation and The definition of its corresponding backward operator, if applicable, is similar to that of an forward operator. **Note that a backward operator does not include a `ProtoMaker`**. -### 4. Registering Operator +### Registering Operator - In `.cc` files, register forward and backward operator classes and the CPU kernel. ```cpp namespace ops = paddle::operators; REGISTER_OP(mul, ops::MulOp, ops::MulOpMaker, mul_grad, ops::MulOpGrad); - REGISTER_OP_CPU_KERNEL(mul, ops::MulKernel); + REGISTER_OP_CPU_KERNEL(mul, ops::MulKernel); REGISTER_OP_CPU_KERNEL(mul_grad, - ops::MulGradKernel); + ops::MulGradKernel); ``` In that code block, @@ -208,20 +207,20 @@ The definition of its corresponding backward operator, if applicable, is similar - `REGISTER_OP_CPU_KERNEL` registers `ops::MulKernel` class and specialized template types `paddle::platform::CPUPlace` and `float`, which also registers `ops::MulGradKernel`. -- Registering GPU Kernel in `.cu` files - - Note that if GPU Kernel is implemented using the `Eigen unsupported` module, then on top of `.cu`, a macro definition `#define EIGEN_USE_GPU` is needed, such as +- Registering CUDA Kernel in `.cu` files + - Note that if CUDA Kernel is implemented using the `Eigen unsupported` module, then on top of `.cu`, a macro definition `#define EIGEN_USE_GPU` is needed, such as ```cpp // if use Eigen unsupported module before include head files #define EIGEN_USE_GPU namespace ops = paddle::operators; - REGISTER_OP_GPU_KERNEL(mul, ops::MulKernel); - REGISTER_OP_GPU_KERNEL(mul_grad, - ops::MulGradKernel); + REGISTER_OP_CUDA_KERNEL(mul, ops::MulKernel); + REGISTER_OP_CUDA_KERNEL(mul_grad, + ops::MulGradKernel); ``` -### 5. Compilation +### Compilation Run the following commands to compile. @@ -253,62 +252,51 @@ A forward operator unit test inherits `unittest.TestCase` and defines metaclass 2. Generating random input data. -3. Implementing the same computation logic in a Python script: +3. Implementing the same computation logic in a Python script. + +4. Call check gradient function to check the backward operator. ```python import unittest import numpy as np - from gradient_checker import GradientChecker, create_op - from op_test_util import OpTestMeta + from op_test import OpTest - class TestMulOp(unittest.TestCase): - __metaclass__ = OpTestMeta + class TestMulOp(OpTest): def setUp(self): - self.type = "mul" + self.op_type = "mul" self.inputs = { 'X': np.random.random((32, 84)).astype("float32"), 'Y': np.random.random((84, 100)).astype("float32") } self.outputs = {'Out': np.dot(self.inputs['X'], self.inputs['Y'])} + + def test_check_output(self): + self.check_output() + + def test_check_grad_normal(self): + self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.5) + + def test_check_grad_ingore_x(self): + self.check_grad( + ['Y'], 'Out', max_relative_error=0.5, no_grad_set=set("X")) + + def test_check_grad_ingore_y(self): + self.check_grad( + ['X'], 'Out', max_relative_error=0.5, no_grad_set=set('Y')) ``` Get its output, and compare it with the forward operator's own output. The code above first loads required packages. In addition, we have -- `self.type = "mul" ` defines the type that is identical to what the operator's registered type. +- `self.op_type = "mul" ` defines the type that is identical to what the operator's registered type. - `self.inputs` defines input, with type `numpy.array` and initializes it. - `self.outputs` defines output and completes the same operator computation in the Python script, and returns its result from the Python script. ### Testing Backward Operators -A backward operator unit test inherits `GradientChecker`, which inherits `unittest.TestCase`. As a result, **a backward operator unit test needs to be have the prefix `test_`**. - -```python -class TestMulGradOp(GradientChecker): - def setUp(self): - self.op = create_op("mul") - self.inputs = { - 'X': np.random.random((32, 84)).astype("float32"), - 'Y': np.random.random((84, 100)).astype("float32") - } - - def test_check_grad_normal(self): - # mul op will enlarge the relative error - self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.5) - - def test_check_grad_ingore_x(self): - self.check_grad( - ['Y'], 'Out', max_relative_error=0.5, no_grad_set=set("X")) - - def test_check_grad_ingore_y(self): - self.check_grad( - ['X'], 'Out', max_relative_error=0.5, no_grad_set=set('Y')) -``` - -Some key points in the code above include: +Some key points in checking gradient above include: -- `create_op("mul")` creates the backward operator's corresponding forward operator. - `test_normal` calls `check_grad` to validate scaling tests' correctness and stability through numeric methods. - The first variable `["X", "Y"]` appoints `X` and `Y` to be scale tested. - The second variable `"Out"` points to the network's final output target `Out`. @@ -338,5 +326,5 @@ ctest -R test_mul_op - Every `*_op.h` (if applicable), `*_op.cc`, and `*_op.cu` (if applicable) must be created for a unique Op. Compiling will fail if multiple operators are included per file. - The type with which an operator is registered needs to be identical to the Op's name. Registering `REGISTER_OP(B, ...)` in `A_op.cc` will cause unit testing failures. -- If the operator does not implement a GPU kernel, please refrain from creating an empty `*_op.cu` file, or else unit tests will fail. +- If the operator does not implement a CUDA kernel, please refrain from creating an empty `*_op.cu` file, or else unit tests will fail. - If multiple operators rely on some shared methods, a file NOT named `*_op.*` can be created to store them, such as `gather.h`. diff --git a/doc/howto/dev/write_docs_cn.rst b/doc/howto/dev/write_docs_cn.rst index 61f3a223547b352cf7929615cf3682b29b9a738f..1bc947c260d7adb75ee5a2bb10e6b91bc0be2d4c 100644 --- a/doc/howto/dev/write_docs_cn.rst +++ b/doc/howto/dev/write_docs_cn.rst @@ -3,12 +3,64 @@ ################## PaddlePaddle的文档包括英文文档 ``doc`` 和中文文档 ``doc_cn`` 两个部分。文档都是通过 `cmake`_ 驱动 `sphinx`_ 编译生成，生成后的文档分别存储在编译目录的 ``doc`` 和 ``doc_cn`` 两个子目录下。 - +也可以利用PaddlePaddle 工具来编译文档，这个情况下所有的文件会存在整理过的的文件目录 .ppo_workspace/content 下 如何构建文档 ============ -PaddlePaddle的文档构建有两种方式。 +PaddlePaddle的文档构建有三种方式。 + + +使用PaddlePaddle.org工具 +-------------- +这个是目前推荐的使用方法。除了可以自动编译文档，也可以直接在网页预览文档。 + +文件工具是使用Docker，需要在系统里先安装好Docker工具包。Docker安装请参考Docker的官网。安装好Docker之后及可用以下命令启动工具 + +.. code-block:: bash + + mkdir paddlepaddle # Create paddlepaddle working directory + cd paddlepaddle + + # Clone the content repositories + git clone https://github.com/PaddlePaddle/Paddle.git + git clone https://github.com/PaddlePaddle/book.git + git clone https://github.com/PaddlePaddle/models.git + git clone https://github.com/PaddlePaddle/Mobile.git + + # Please specify the working directory through -v + docker run -it -p 8000:8000 -v `pwd`:/var/content paddlepaddle/paddlepaddle.org:latest + +注意: PaddlePaddle.org 会在 -v (volume) 指定的内容存储库运行命令 +之后再用网页连到http://localhost:8000就可以在网页上生成需要的文档 +编译后的文件将被存储在工作目录 /.ppo_workspace/content。 + +如果不想使用 Docker，你还可以通过运行Django框架直接激活工具的服务器。使用下面的命令来运行它。 + +.. code-block:: bash + + mkdir paddlepaddle # Create paddlepaddle working directory + cd paddlepaddle + + # Clone the content repositories and PaddlePaddle.org + git clone https://github.com/PaddlePaddle/Paddle.git + git clone https://github.com/PaddlePaddle/book.git + git clone https://github.com/PaddlePaddle/models.git + git clone https://github.com/PaddlePaddle/Mobile.git + git clone https://github.com/PaddlePaddle/PaddlePaddle.org.git + + # Please specify the PaddlePaddle working directory. In the current setting, it should be pwd + export CONTENT_DIR= + export ENV='' + cd PaddlePaddle.org/portal/ + pip install -r requirements.txt + python manage.py runserver + +工具服务器将读取环境变量 CONTENT_DIR 搜索代码库。请指定的PaddlePaddle工作目录给环境变量 CONTENT_DIR。 +之后再用网页连到http://localhost:8000就可以在网页上生成需要的文档。 +编译后的文件将被存储在工作目录 /.ppo_workspace/content。 + +想了解更多PaddlePaddle.org工具的详细信息，可以 `点击这里 `_ 。 使用Docker构建 -------------- @@ -47,17 +99,12 @@ PaddlePaddle的文档构建有两种方式。 PaddlePaddle文档使用 `sphinx`_ 自动生成，用户可以参考sphinx教程进行书写。 -如何更新文档主题 -================ - -PaddlePaddle文档主题在 `TO_YOUR_PADDLE_CLONE_PATH/doc_theme` 文件夹下，包含所有和前端网页设计相关的文件。 - -如何更新doc.paddlepaddle.org +如何更新www.paddlepaddle.org ============================ -更新的文档以PR的形式提交到github中，提交方式参见 `贡献文档 `_ 。 -目前PaddlePaddle的develop分支的文档是自动触发更新的，用户可以分别查看最新的 `中文文档 `_ 和 -`英文文档 `_ 。 +更新的文档以PR的形式提交到github中，提交方式参见 `贡献文档 `_ 。 +目前PaddlePaddle的develop分支的文档是自动触发更新的，用户可以分别查看最新的 `中文文档 `_ 和 +`英文文档 `_ 。 .. _cmake: https://cmake.org/ diff --git a/doc/howto/dev/write_docs_en.rst b/doc/howto/dev/write_docs_en.rst new file mode 100644 index 0000000000000000000000000000000000000000..b3ef07eb1d0012827df8e6a4f27c5fa643649492 --- /dev/null +++ b/doc/howto/dev/write_docs_en.rst @@ -0,0 +1,80 @@ +################## +Contribute Documentation +################## + +PaddlePaddle supports English documentation ``doc`` and Chinese documentation ``doc_cn``. +Both are compiled by `cmake`_ and `sphinx`_ , the compiled documentations will be stored under ``doc`` and ``doc_cn`` directories. +When using the PaddlePaddle.org to compile documentations, the compiled documentations will be stored under a consolidated directory: .ppo_workspace/content + +How to Build Documentations +============ + +We recommend using PaddlePaddle.org tool to build documentation + + +Use PaddlePaddle.org tool +-------------- +This is the recommended method to build documentation. It can compile documentation and preview the documentation in a web browser. + +The tool uses Docker, please install it on your system. Please check Docker official website on how to install Docker. You may use the following commands to activate the tool + +.. code-block:: bash + + mkdir paddlepaddle # Create paddlepaddle working directory + cd paddlepaddle + + # Clone the content repositories. You may only clone the contents you need + git clone https://github.com/PaddlePaddle/Paddle.git + git clone https://github.com/PaddlePaddle/book.git + git clone https://github.com/PaddlePaddle/models.git + git clone https://github.com/PaddlePaddle/Mobile.git + + # Please specify the working directory through -v + docker run -it -p 8000:8000 -v `pwd`:/var/content paddlepaddle/paddlepaddle.org:latest + +Note: PaddlePaddle.org will read the content repos specified in the -v (volume) flag of the docker run command +Use a web browser and navigate to http://localhost:8000, click the buttons to compile the documentation +The compiled documentations will be stored in /.ppo_workspace/content + + +If you don't wish to use Docker, you can also activate the tool through Django. Use the following the commands to set up + +.. code-block:: bash + + mkdir paddlepaddle # Create paddlepaddle working directory + cd paddlepaddle + + # Clone the content repositories and PaddlePaddle.org + git clone https://github.com/PaddlePaddle/Paddle.git + git clone https://github.com/PaddlePaddle/book.git + git clone https://github.com/PaddlePaddle/models.git + git clone https://github.com/PaddlePaddle/Mobile.git + git clone https://github.com/PaddlePaddle/PaddlePaddle.org.git + + # Please specify the PaddlePaddle working directory. In the current setting, it should be pwd + export CONTENT_DIR= + export ENV='' + cd PaddlePaddle.org/portal/ + pip install -r requirements.txt + python manage.py runserver + +Use a web browser and navigate to http://localhost:8000, click the buttons to compile the documentation +The compiled documentations will be stored in /.ppo_workspace/content + +If you want to learn more on the PaddlePaddle.org, please `click here `_ 。 + +How to write Documentations +============ + +PaddlePaddle uses `sphinx`_ to compile documentations，Please check sphinx official website for more detail. + + +How to update www.paddlepaddle.org +============================ + +Please create PRs and submit them to github, please check `Contribute Code `_ 。 +PaddlePaddle develop branch will update the documentation once the PR is merged. User may check latest `Chinese Docs `_ and +`English Docs `_ 。 + +.. _cmake: https://cmake.org/ +.. _sphinx: http://www.sphinx-doc.org/en/1.4.8/ diff --git a/doc/howto/index_cn.rst b/doc/howto/index_cn.rst index 76d3e0a0092f89005605a23e14e712530112a5ac..ccd909770253bb85dbc8a5a2560594076c2f68b0 100644 --- a/doc/howto/index_cn.rst +++ b/doc/howto/index_cn.rst @@ -9,9 +9,6 @@ usage/cmd_parameter/index_cn.rst usage/cluster/cluster_train_cn.md - usage/k8s/k8s_basis_cn.md - usage/k8s/k8s_cn.md - usage/k8s/k8s_distributed_cn.md 开发标准 -------- @@ -19,7 +16,7 @@ .. toctree:: :maxdepth: 1 - dev/build_cn.rst + dev/contribute_to_paddle_cn.md dev/write_docs_cn.rst 模型配置 diff --git a/doc/howto/index_en.rst b/doc/howto/index_en.rst index 1b6034be4edffd2cbc822018b733b9a3836ea84a..6d1bf7dfc003da6de31410ee0a7959233adfaf76 100644 --- a/doc/howto/index_en.rst +++ b/doc/howto/index_en.rst @@ -9,8 +9,6 @@ Usage usage/cmd_parameter/index_en.rst usage/cluster/cluster_train_en.md - usage/k8s/k8s_en.md - usage/k8s/k8s_aws_en.md Development ------------ @@ -18,9 +16,9 @@ Development .. toctree:: :maxdepth: 1 - dev/build_en.rst dev/new_layer_en.rst dev/contribute_to_paddle_en.md + dev/write_docs_en.rst Configuration ------------- diff --git a/doc/howto/optimization/cpu_profiling.md b/doc/howto/optimization/cpu_profiling.md new file mode 100644 index 0000000000000000000000000000000000000000..1775374cf6e518586c28bbd8e04946c74df7e4c5 --- /dev/null +++ b/doc/howto/optimization/cpu_profiling.md @@ -0,0 +1,197 @@ +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`. + +Profiling is the process that reveals performance bottlenecks, +which could be very different from what's in the developers' mind. +Performance tuning is done to fix these bottlenecks. Performance optimization +repeats the steps of profiling and tuning alternatively. + +PaddlePaddle users program AI applications by calling the Python API, which calls +into `libpaddle.so.` written in C++. In this tutorial, we focus on +the profiling and tuning of + +1. the Python code and +1. the mixture of Python and C++ code. + +## Profiling the Python Code + +### Generate the Performance Profiling File + +We can use Python standard +package, [`cProfile`](https://docs.python.org/2/library/profile.html), +to generate Python profiling file. For example: + +```bash +python -m cProfile -o profile.out main.py +``` + +where `main.py` is the program we are going to profile, `-o` specifies +the output file. Without `-o`, `cProfile` would outputs to standard +output. + +### Look into the Profiling File + +`cProfile` generates `profile.out` after `main.py` completes. We can +use [`cprofilev`](https://github.com/ymichael/cprofilev) to look into +the details: + +```bash +cprofilev -a 0.0.0.0 -p 3214 -f profile.out main.py +``` + +where `-a` specifies the HTTP IP, `-p` specifies the port, `-f` +specifies the profiling file, and `main.py` is the source file. + +Open the Web browser and points to the local IP and the specifies +port, we will see the output like the following: + +``` + ncalls tottime percall cumtime percall filename:lineno(function) + 1 0.284 0.284 29.514 29.514 main.py:1() + 4696 0.128 0.000 15.748 0.003 /home/yuyang/perf_test/.env/lib/python2.7/site-packages/paddle/v2/fluid/executor.py:20(run) + 4696 12.040 0.003 12.040 0.003 {built-in method run} + 1 0.144 0.144 6.534 6.534 /home/yuyang/perf_test/.env/lib/python2.7/site-packages/paddle/v2/__init__.py:14() +``` + +where each line corresponds to Python function, and the meaning of +each column is as follows: + +| column | meaning | +| --- | --- | +| ncalls | the number of calls into a function | +| tottime | the total execution time of the function, not including the + execution time of other functions called by the function | +| percall | tottime divided by ncalls | +| cumtime | the total execution time of the function, including the execution time of other functions being called | +| percall | cumtime divided by ncalls | +| filename:lineno(function) | where the function is defined | + +### Identify Performance Bottlenecks + +Usually, `tottime` and the related `percall` time is what we want to +focus on. We can sort above profiling file by tottime: + +```text + 4696 12.040 0.003 12.040 0.003 {built-in method run} + 300005 0.874 0.000 1.681 0.000 /home/yuyang/perf_test/.env/lib/python2.7/site-packages/paddle/v2/dataset/mnist.py:38(reader) + 107991 0.676 0.000 1.519 0.000 /home/yuyang/perf_test/.env/lib/python2.7/site-packages/paddle/v2/fluid/framework.py:219(__init__) + 4697 0.626 0.000 2.291 0.000 /home/yuyang/perf_test/.env/lib/python2.7/site-packages/paddle/v2/fluid/framework.py:428(sync_with_cpp) + 1 0.618 0.618 0.618 0.618 /home/yuyang/perf_test/.env/lib/python2.7/site-packages/paddle/v2/fluid/__init__.py:1() +``` + +We can see that the most time-consuming function is the `built-in +method run`, which is a C++ function in `libpaddle.so`. We will +explain how to profile C++ code in the next section. At this +moment, let's look into the third function `sync_with_cpp`, which is a +Python function. We can click it to understand more about it: + +``` +Called By: + + Ordered by: internal time + List reduced from 4497 to 2 due to restriction <'sync_with_cpp'> + +Function was called by... + ncalls tottime cumtime +/home/yuyang/perf_test/.env/lib/python2.7/site-packages/paddle/v2/fluid/framework.py:428(sync_with_cpp) <- 4697 0.626 2.291 /home/yuyang/perf_test/.env/lib/python2.7/site-packages/paddle/v2/fluid/framework.py:562(sync_with_cpp) +/home/yuyang/perf_test/.env/lib/python2.7/site-packages/paddle/v2/fluid/framework.py:562(sync_with_cpp) <- 4696 0.019 2.316 /home/yuyang/perf_test/.env/lib/python2.7/site-packages/paddle/v2/fluid/framework.py:487(clone) + 1 0.000 0.001 /home/yuyang/perf_test/.env/lib/python2.7/site-packages/paddle/v2/fluid/framework.py:534(append_backward) + + +Called: + + Ordered by: internal time + List reduced from 4497 to 2 due to restriction <'sync_with_cpp'> +``` + +The lists of the callers of `sync_with_cpp` might help us understand +how to improve the function definition. + +## Profiling Python and C++ Code + +### Generate the Profiling File + +To profile a mixture of Python and C++ code, we can use a Python +package, `yep`, that can work with Google's `perftools`, which is a +commonly-used profiler for C/C++ code. + +In Ubuntu systems, we can install `yep` and `perftools` by running the +following commands: + +```bash +apt update +apt install libgoogle-perftools-dev +pip install yep +``` + +Then we can run the following command + +```bash +python -m yep -v main.py +``` + +to generate the profiling file. The default filename is +`main.py.prof`. + +Please be aware of the `-v` command line option, which prints the +analysis results after generating the profiling file. By examining the + the print result, we'd know that if we stripped debug +information from `libpaddle.so` at build time. The following hints +help make sure that the analysis results are readable: + +1. Use GCC command line option `-g` when building `libpaddle.so` so to + include the debug information. The standard building system of + PaddlePaddle is CMake, so you might want to set + `CMAKE_BUILD_TYPE=RelWithDebInfo`. + +1. Use GCC command line option `-O2` or `-O3` to generate optimized + binary code. It doesn't make sense to profile `libpaddle.so` + without optimization, because it would anyway run slowly. + +1. Profiling the single-threaded binary file before the + multi-threading version, because the latter often generates tangled + profiling analysis result. You might want to set environment + variable `OMP_NUM_THREADS=1` to prevents OpenMP from automatically + starting multiple threads. + +### Examining the Profiling File + +The tool we used to examine the profiling file generated by +`perftools` is [`pprof`](https://github.com/google/pprof), which +provides a Web-based GUI like `cprofilev`. + +We can rely on the standard Go toolchain to retrieve the source code +of `pprof` and build it: + +```bash +go get github.com/google/pprof +``` + +Then we can use it to profile `main.py.prof` generated in the previous +section: + +```bash +pprof -http=0.0.0.0:3213 `which python` ./main.py.prof +``` + +Where `-http` specifies the IP and port of the HTTP service. +Directing our Web browser to the service, we would see something like +the following: + +![result](./pprof_1.png) + +### Identifying the Performance Bottlenecks + +Similar to how we work with `cprofilev`, we'd focus on `tottime` and +`cumtime`. + +![kernel_perf](./pprof_2.png) + +We can see that the execution time of multiplication and the computing +of the gradient of multiplication takes 2% to 4% of the total running +time, and `MomentumOp` takes about 17%. Obviously, we'd want to +optimize `MomentumOp`. + +`pprof` would mark performance critical parts of the program in +red. It's a good idea to follow the hints. diff --git a/doc/howto/optimization/cpu_profiling_cn.md b/doc/howto/optimization/cpu_profiling_cn.md new file mode 100644 index 0000000000000000000000000000000000000000..14eba0e2f34b115f5cd24920b5b1af07ec953d00 --- /dev/null +++ b/doc/howto/optimization/cpu_profiling_cn.md @@ -0,0 +1,155 @@ +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。 + +PaddlePaddle 用户一般通过调用 Python API 编写深度学习程序。大部分 Python API 调用用 C++ 写的 libpaddle.so。所以 PaddlePaddle 的性能分析与调优分为两个部分: + +* Python 代码的性能分析 +* Python 与 C++ 混合代码的性能分析 + + +## Python代码的性能分析 + +### 生成性能分析文件 + +Python标准库中提供了性能分析的工具包，[cProfile](https://docs.python.org/2/library/profile.html)。生成Python性能分析的命令如下: + +```bash +python -m cProfile -o profile.out main.py +``` + +其中 `main.py` 是我们要分析的程序，`-o`标识了一个输出的文件名，用来存储本次性能分析的结果。如果不指定这个文件，`cProfile`会打印到标准输出。 + +### 查看性能分析文件 + +`cProfile` 在main.py 运行完毕后输出`profile.out`。我们可以使用[`cprofilev`](https://github.com/ymichael/cprofilev)来查看性能分析结果。`cprofilev`是一个Python的第三方库。使用它会开启一个HTTP服务，将性能分析结果以网页的形式展示出来： + +```bash +cprofilev -a 0.0.0.0 -p 3214 -f profile.out main.py +``` + +其中`-a`标识HTTP服务绑定的IP。使用`0.0.0.0`允许外网访问这个HTTP服务。`-p`标识HTTP服务的端口。`-f`标识性能分析的结果文件。`main.py`标识被性能分析的源文件。 + +用Web浏览器访问对应网址，即可显示性能分析的结果： + +``` + ncalls tottime percall cumtime percall filename:lineno(function) + 1 0.284 0.284 29.514 29.514 main.py:1() + 4696 0.128 0.000 15.748 0.003 /home/yuyang/perf_test/.env/lib/python2.7/site-packages/paddle/v2/fluid/executor.py:20(run) + 4696 12.040 0.003 12.040 0.003 {built-in method run} + 1 0.144 0.144 6.534 6.534 /home/yuyang/perf_test/.env/lib/python2.7/site-packages/paddle/v2/__init__.py:14() +``` + +每一列的含义是: + +| 列名 | 含义 | +| --- | --- | +| ncalls | 函数的调用次数 | +| tottime | 函数实际使用的总时间。该时间去除掉本函数调用其他函数的时间 | +| percall | tottime的每次调用平均时间 | +| cumtime | 函数总时间。包含这个函数调用其他函数的时间 | +| percall | cumtime的每次调用平均时间 | +| filename:lineno(function) | 文件名, 行号，函数名 | + + +### 寻找性能瓶颈 + +通常`tottime`和`cumtime`是寻找瓶颈的关键指标。这两个指标代表了某一个函数真实的运行时间。 + +将性能分析结果按照tottime排序，效果如下: + +```text + 4696 12.040 0.003 12.040 0.003 {built-in method run} + 300005 0.874 0.000 1.681 0.000 /home/yuyang/perf_test/.env/lib/python2.7/site-packages/paddle/v2/dataset/mnist.py:38(reader) + 107991 0.676 0.000 1.519 0.000 /home/yuyang/perf_test/.env/lib/python2.7/site-packages/paddle/v2/fluid/framework.py:219(__init__) + 4697 0.626 0.000 2.291 0.000 /home/yuyang/perf_test/.env/lib/python2.7/site-packages/paddle/v2/fluid/framework.py:428(sync_with_cpp) + 1 0.618 0.618 0.618 0.618 /home/yuyang/perf_test/.env/lib/python2.7/site-packages/paddle/v2/fluid/__init__.py:1() +``` + +可以看到最耗时的函数是C++端的`run`函数。这需要联合我们第二节`Python`与`C++`混合代码的性能分析来进行调优。而`sync_with_cpp`函数的总共耗时很长，每次调用的耗时也很长。于是我们可以点击`sync_with_cpp`的详细信息，了解其调用关系。 + +```text +Called By: + + Ordered by: internal time + List reduced from 4497 to 2 due to restriction <'sync_with_cpp'> + +Function was called by... + ncalls tottime cumtime +/home/yuyang/perf_test/.env/lib/python2.7/site-packages/paddle/v2/fluid/framework.py:428(sync_with_cpp) <- 4697 0.626 2.291 /home/yuyang/perf_test/.env/lib/python2.7/site-packages/paddle/v2/fluid/framework.py:562(sync_with_cpp) +/home/yuyang/perf_test/.env/lib/python2.7/site-packages/paddle/v2/fluid/framework.py:562(sync_with_cpp) <- 4696 0.019 2.316 /home/yuyang/perf_test/.env/lib/python2.7/site-packages/paddle/v2/fluid/framework.py:487(clone) + 1 0.000 0.001 /home/yuyang/perf_test/.env/lib/python2.7/site-packages/paddle/v2/fluid/framework.py:534(append_backward) + + +Called: + + Ordered by: internal time + List reduced from 4497 to 2 due to restriction <'sync_with_cpp'> +``` + +通常观察热点函数间的调用关系，和对应行的代码，就可以了解到问题代码在哪里。当我们做出性能修正后，再次进行性能分析(profiling)即可检查我们调优后的修正是否能够改善程序的性能。 + + + +## Python与C++混合代码的性能分析 + +### 生成性能分析文件 + +C++的性能分析工具非常多。常见的包括`gprof`, `valgrind`, `google-perftools`。但是调试Python中使用的动态链接库与直接调试原始二进制相比增加了很多复杂度。幸而Python的一个第三方库`yep`提供了方便的和`google-perftools`交互的方法。于是这里使用`yep`进行Python与C++混合代码的性能分析 + +使用`yep`前需要安装`google-perftools`与`yep`包。ubuntu下安装命令为 + +```bash +apt update +apt install libgoogle-perftools-dev +pip install yep +``` + +安装完毕后，我们可以通过 + +```bash +python -m yep -v main.py +``` + +生成性能分析文件。生成的性能分析文件为`main.py.prof`。 + +命令行中的`-v`指定在生成性能分析文件之后，在命令行显示分析结果。我们可以在命令行中简单的看一下生成效果。因为C++与Python不同，编译时可能会去掉调试信息，运行时也可能因为多线程产生混乱不可读的性能分析结果。为了生成更可读的性能分析结果，可以采取下面几点措施: + +1. 编译时指定`-g`生成调试信息。使用cmake的话，可以将CMAKE_BUILD_TYPE指定为`RelWithDebInfo`。 +2. 编译时一定要开启优化。单纯的`Debug`编译性能会和`-O2`或者`-O3`有非常大的差别。`Debug`模式下的性能测试是没有意义的。 +3. 运行性能分析的时候，先从单线程开始，再开启多线程，进而多机。毕竟单线程调试更容易。可以设置`OMP_NUM_THREADS=1`这个环境变量关闭openmp优化。 + +### 查看性能分析文件 + +在运行完性能分析后，会生成性能分析结果文件。我们可以使用[`pprof`](https://github.com/google/pprof)来显示性能分析结果。注意，这里使用了用`Go`语言重构后的`pprof`，因为这个工具具有web服务界面，且展示效果更好。 + +安装`pprof`的命令和一般的`Go`程序是一样的，其命令如下: + +```bash +go get github.com/google/pprof +``` + +进而我们可以使用如下命令开启一个HTTP服务: + +```bash +pprof -http=0.0.0.0:3213 `which python` ./main.py.prof +``` + +这行命令中，`-http`指开启HTTP服务。`which python`会产生当前Python二进制的完整路径，进而指定了Python可执行文件的路径。`./main.py.prof`输入了性能分析结果。 + +访问对应的网址，我们可以查看性能分析的结果。结果如下图所示: + +![result](./pprof_1.png) + + +### 寻找性能瓶颈 + +与寻找Python代码的性能瓶颈类似，寻找Python与C++混合代码的性能瓶颈也是要看`tottime`和`cumtime`。而`pprof`展示的调用图也可以帮助我们发现性能中的问题。 + +例如下图中， + +![kernel_perf](./pprof_2.png) + +在一次训练中，乘法和乘法梯度的计算占用2%-4%左右的计算时间。而`MomentumOp`占用了17%左右的计算时间。显然，`MomentumOp`的性能有问题。 + +在`pprof`中，对于性能的关键路径都做出了红色标记。先检查关键路径的性能问题，再检查其他部分的性能问题，可以更有次序的完成性能的优化。 diff --git a/doc/howto/optimization/pprof_1.png b/doc/howto/optimization/pprof_1.png new file mode 100644 index 0000000000000000000000000000000000000000..8e9edbf377672d0ef40f2fc7bd39e746923550cb Binary files /dev/null and b/doc/howto/optimization/pprof_1.png differ diff --git a/doc/howto/optimization/pprof_2.png b/doc/howto/optimization/pprof_2.png new file mode 100644 index 0000000000000000000000000000000000000000..172ba20399ba974d27f4c072425277b69b02520b Binary files /dev/null and b/doc/howto/optimization/pprof_2.png differ diff --git a/doc/howto/read_source.md b/doc/howto/read_source.md new file mode 100644 index 0000000000000000000000000000000000000000..e4211abb3be9cace80bc14dbe3db3e0a31221dd0 --- /dev/null +++ b/doc/howto/read_source.md @@ -0,0 +1,67 @@ +# PaddlePaddle Fluid Source Code Overview + +Examples: https://github.com/PaddlePaddle/Paddle/tree/develop/python/paddle/v2/fluid/tests/book + +Core: https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/framework + +Operator: https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators + +Memory: https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/memory + +Platform: https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/platform + +# Compile Time + +The following **defines** the NN. The definition goes into this [protocol buffer](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/framework.proto). + +```python +x = fluid.layers.data(name='x', shape=[13], dtype='float32') +y = fluid.layers.data(name='y', shape=[1], dtype='float32') + +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(x=cost) + +sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.001) +sgd_optimizer.minimize(avg_cost) +``` + +- Variables: `x`, `y`, `y_predict`, `cost` and `avg_cost`. [Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/fluid/framework.py#L93) +- Layers: `fluid.layers.data`, `fluid.layers.fc` and `fluid.layers.mean` are layers. [Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/fluid/layers.py) + - Every Layer has one or more operators and variables/parameters + - All the operators are defined at [`paddle/operators/`](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators). Other worth-looking files: + - Base class: [`paddle/framework/operator.h`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/operator.h) + - Operator Registration: [`paddle/framework/op_registry.h`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/op_registry.h) + - Operator Lookup: [`paddle/framework/op_info.h`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/op_info.h) +- Optimizer: `fluid.optimizer.SGD`. It does the following + - Add backward operators. [[Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/fluid/backward.py), [C++](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/backward.cc)] + - Add optimizer operators. [[Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/fluid/optimizer.py), [C++](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/optimizer)] + +# Run Time + +The following **evaluates** the NN. Instantiates all the variables, operators. + +```python +place = fluid.CPUPlace() +feeder = fluid.DataFeeder(place=place, feed_list=[x, y]) +exe = fluid.Executor(place) + +# Allocate memory. Initialize Parameter. +exe.run(fluid.default_startup_program()) + +# Allocate memory. Do computation. +exe.run(fluid.default_main_program(), + feed=feeder.feed(data), + fetch_list=[avg_cost]) +``` + +- Place: `place`. one of CPU, GPU or FPGA. [C++](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/place.h) + - The device handle are at [paddle/platform/device_context.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/device_context.h) +- Executor: `fluid.Executor(place)`. [[Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/fluid/executor.py), [C++](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/executor.cc)] + - Feeds the data: `feed=feeder.feed(data)` + - Evaluates all the operators + - Fetches the result: `fetch_list=[avg_cost]` +- Other worth looking files: + - Scope: [paddle/framework/scope.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/scope.h). Where all the variables live + - Variable: [paddle/framework/variable.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/variable.h). Where all the data (most likely tensors) live + - Tensor: [paddle/framework/tensor.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/tensor.h). Where we allocate memory through [`paddle/memory/`](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/memory) diff --git a/doc/howto/usage/cluster/cluster_train_cn.md b/doc/howto/usage/cluster/cluster_train_cn.md index 2e98b3de3fe2284375f87e883ff4bac19255dbeb..659bae9c0ceaf2fb2df8446b9d406a822a9df0ea 100644 --- a/doc/howto/usage/cluster/cluster_train_cn.md +++ b/doc/howto/usage/cluster/cluster_train_cn.md @@ -1,25 +1,8 @@ -# PaddlePaddle分布式训练 - -* [概述](#概述) -* [环境准备](#环境准备) -* [启动参数说明](#启动参数说明) - * [启动参数服务器](#启动参数服务器) - * [启动计算节点](#启动计算节点) - * [准备数据集](#准备数据集) - * [准备训练程序](#准备训练程序) -* [使用分布式计算平台或工具](#使用分布式计算平台或工具) - * [使用Fabric启动集群作业](#使用fabric启动集群作业) - * [准备一个Linux集群](#准备一个linux集群) - * [启动集群作业](#启动集群作业) - * [终止集群作业](#终止集群作业) - * [检查集群训练结果](#检查集群训练结果) - * [检查模型输出](#检查模型输出) - * [在OpenMPI集群中提交训练作业](#在openmpi集群中提交训练作业) - * [准备OpenMPI集群](#准备OpenMPI集群) - * [启动集群作业](#启动集群作业-1) - * [在Kubernetes集群中提交训练作业](#在kubernetes集群中提交训练作业) +# 分布式训练 + ## 概述 + 本文将介绍如何使用PaddlePaddle在不同的集群框架下完成分布式训练。分布式训练架构如下图所示： @@ -32,10 +15,11 @@ 在使用同步SGD训练神经网络时，PaddlePaddle使用同步屏障（barrier），使梯度的提交和参数的更新按照顺序方式执行。在异步SGD中，则并不会等待所有trainer提交梯度才更新参数，这样极大地提高了计算的并行性：参数服务器之间不相互依赖，并行地接收梯度和更新参数，参数服务器也不会等待计算节点全部都提交梯度之后才开始下一步，计算节点之间也不会相互依赖，并行地执行模型的训练。可以看出，虽然异步SGD方式会提高参数更新并行度, 但是并不能保证参数同步更新，在任意时间某一台参数服务器上保存的参数可能比另一台要更新，与同步SGD相比，梯度会有噪声。 + ## 环境准备 1. 准备您的计算集群。计算集群通常由一组（几台到几千台规模）的Linux服务器组成。服务器之间可以通过局域网（LAN）联通，每台服务器具有集群中唯一的IP地址（或者可被DNS解析的主机名）。集群中的每台计算机通常被成为一个“节点”。 -1. 我们需要在集群的所有节点上安装 PaddlePaddle。 如果要启用GPU，还需要在节点上安装对应的GPU驱动以及CUDA。PaddlePaddle的安装可以参考[build_and_install](https://github.com/PaddlePaddle/Paddle/tree/develop/doc/getstarted/build_and_install)的多种安装方式。我们推荐使用[Docker](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/getstarted/build_and_install/docker_install_cn.rst)安装方式来快速安装PaddlePaddle。 +1. 我们需要在集群的所有节点上安装 PaddlePaddle。 如果要启用GPU，还需要在节点上安装对应的GPU驱动以及CUDA。PaddlePaddle的安装可以参考[build_and_install](http://www.paddlepaddle.org/docs/develop/documentation/zh/getstarted/build_and_install/index_cn.html)的多种安装方式。我们推荐使用[Docker](http://www.paddlepaddle.org/docs/develop/documentation/zh/getstarted/build_and_install/docker_install_cn.html)安装方式来快速安装PaddlePaddle。 安装完成之后，执行下面的命令可以查看已经安装的版本（docker安装方式可以进入docker容器执行：`docker run -it paddlepaddle/paddle:[tag] /bin/bash`）： ```bash @@ -63,12 +47,12 @@ $ paddle pserver --port=7164 --ports_num=1 --ports_num_for_sparse=1 --num_gradie $ stdbuf -oL /usr/bin/nohup paddle pserver --port=7164 --ports_num=1 --ports_num_for_sparse=1 --num_gradient_servers=1 &> pserver.log ``` -| 参数 | 是否必选 | 默认值 | 说明 | -| ------------- | ------------- | ------------- | ------------- | -| port | 必选 | 7164 | pserver监听的起始端口，根据ports_num决定
总端口个数，从起始端口监听多个端口用于通信 | -| ports_num | 必选 | 1 | 监听的端口个数 | -| ports_num_for_sparse | 必选 | 1 | 用于稀疏类型参数通信的端口个数 | -| num_gradient_servers | 必选 | 1 | 当前训练任务pserver总数 | +参数说明 + +- port：**必选，默认7164**，pserver监听的起始端口，根据ports_num决定总端口个数，从起始端口监听多个端口用于通信 +- ports_num：**必选，默认1**，监听的端口个数 +- ports_num_for_sparse：**必选，默认1**，用于稀疏类型参数通信的端口个数 +- num_gradient_servers：**必选，默认1**，当前训练任务pserver总数 ### 启动计算节点 执行以下命令启动使用python编写的trainer程序（文件名为任意文件名，如train.py） @@ -105,16 +89,16 @@ paddle.init( pservers="127.0.0.1") ``` -| 参数 | 是否必选 | 默认 | 说明 | -| ------------- | ------------- | ------------- | ------------- | -| use_gpu | 可选 | False | 是否启用GPU训练 | -| trainer_count | 必选 | 1 | 当前训练任务trainer总个数 | -| port | 必选 | 7164 | 连接到pserver的端口 | -| ports_num | 必选 | 1 | 连接到pserver的端口个数 | -| ports_num_for_sparse | 必选 | 1 | 和pserver之间用于稀疏类型参数通信的端口个数 | -| num_gradient_servers | 必选 | 1 | 当前训练任务pserver总数 | -| trainer_id | 必选 | 0 | 每个trainer的唯一ID，从0开始的整数 | -| pservers | 必选 | 127.0.0.1 | 当前训练任务启动的pserver的IP列表，多个IP使用“,”隔开 | +参数说明 + +- use_gpu： **可选，默认False**，是否启用GPU训练 +- trainer_count：**必选，默认1**，当前训练任务trainer总个数 +- port：**必选，默认7164**，连接到pserver的端口 +- ports_num：**必选，默认1**，连接到pserver的端口个数 +- ports_num_for_sparse：**必选，默认1**，和pserver之间用于稀疏类型参数通信的端口个数 +- num_gradient_servers：**必选，默认1**，当前训练任务pserver总数 +- trainer_id：**必选，默认0**，每个trainer的唯一ID，从0开始的整数 +- pservers：**必选，默认127.0.0.1**，当前训练任务启动的pserver的IP列表，多个IP使用“,”隔开 ### 准备数据集 @@ -171,7 +155,7 @@ test.txt-00002 - `my_lib.py`：会被`train.py`调用的一些用户定义的库函数，比如PIL库等。 - `word_dict.pickle`：在`train.py`中会使用到的字典数据文件。 -- `train.py`：训练程序，代码参考[api_train_v2_cluster.py](https://github.com/PaddlePaddle/Paddle/tree/develop/doc/howto/usage/cluster/src/word2vec/prepare.py)。***注意：*** 对于本样例代码，在使用不同的分布式计算平台时，您可能需要修改`train.py`开头的部分（如下），以便获得训练数据的位置和获取环境变量配置： +- `train.py`：训练程序，代码参考[api_train_v2_cluster.py](https://github.com/PaddlePaddle/Paddle/tree/develop/doc/howto/usage/cluster/src/word2vec/api_train_v2_cluster.py)。***注意：*** 对于本样例代码，在使用不同的分布式计算平台时，您可能需要修改`train.py`开头的部分（如下），以便获得训练数据的位置和获取环境变量配置： ```python cluster_train_file = "./train_data_dir/train/train.txt" @@ -195,91 +179,10 @@ PaddlePaddle可以使用多种分布式计算平台构建分布式计算任务 在使用分布式计算平台进行训练时，任务被调度在集群中时，分布式计算平台通常会通过API或者环境变量提供任务运行需要的参数，比如节点的ID、IP和任务节点个数等。 -### 使用Fabric启动集群作业 - -#### 准备一个Linux集群 -可以在`paddle/scripts/cluster_train_v2/fabric/docker_cluster`目录下，执行`kubectl -f ssh_servers.yaml`启动一个测试集群，并使用`kubectl get po -o wide`获得这些节点的IP地址。 - -#### 启动集群作业 - -`paddle.py` 提供了自动化脚本来启动不同节点中的所有 PaddlePaddle 集群进程。默认情况下，所有命令行选项可以设置为 `paddle.py` 命令选项并且 `paddle.py` 将透明、自动地将这些选项应用到 PaddlePaddle 底层进程。 - -`paddle.py` 为方便作业启动提供了两个独特的命令选项。 - -- `job_dispatch_package` 设为本地 `workspace` 目录，它将被分发到 `conf.py` 中设置的所有节点。它有助于帮助频繁修改和访问工作区文件的用户减少负担，否则频繁的多节点工作空间部署可能会很麻烦。 -- `job_workspace` 设为已部署的工作空间目录，`paddle.py` 将跳过分发阶段直接启动所有节点的集群作业。它可以帮助减少分发延迟。 - -`cluster_train/run.sh` 提供了命令样例来运行 `doc/howto/usage/cluster/src/word2vec` 集群任务，只需用您定义的目录修改 `job_dispatch_package` 和 `job_workspace`，然后： -``` -sh run.sh -``` - -集群作业将会在几秒后启动。 - -#### 终止集群作业 -`paddle.py`能获取`Ctrl + C` SIGINT 信号来自动终止它启动的所有进程。只需中断 `paddle.py` 任务来终止集群作业。如果程序崩溃你也可以手动终止。 - -#### 检查集群训练结果 -详细信息请检查 $workspace/log 里的日志，每一个节点都有相同的日志结构。 - -`paddle_trainer.INFO` -提供几乎所有训练的内部输出日志，与本地训练相同。这里检验运行时间模型的收敛。 - -`paddle_pserver2.INFO` -提供 pserver 运行日志，有助于诊断分布式错误。 - -`server.log` -提供 parameter server 进程的 stderr 和 stdout。训练失败时可以检查错误日志。 - -`train.log` -提供训练过程的 stderr 和 stdout。训练失败时可以检查错误日志。 - -#### 检查模型输出 -运行完成后，模型文件将被写入节点 0 的 `output` 目录中。 -工作空间中的 `nodefile` 表示当前集群作业的节点 ID。 - -### 在OpenMPI集群中提交训练作业 - -#### 准备OpenMPI集群 - -执行下面的命令以启动3个节点的OpenMPI集群和一个"head"节点： - -```bash -paddle/scripts/cluster_train_v2/openmpi/docker_cluster -kubectl create -f head.yaml -kubectl create -f mpi-nodes.yaml -``` - -然后可以从head节点ssh无密码登录到OpenMPI的每个节点上。 - -#### 启动集群作业 - -您可以按照下面的步骤在OpenMPI集群中提交paddle训练任务： - -```bash -# 获得head和node节点的IP地址 -kubectl get po -o wide -# 将node节点的IP地址保存到machines文件中 -kubectl get po -o wide | grep nodes | awk '{print $6}' > machines -# 拷贝必要的文件到head节点 -scp -i ssh/id_rsa.mpi.pub machines prepare.py train.py start_mpi_train.sh tutorial@[headIP]:~ -# ssh 登录到head节点 -ssh -i ssh/id_rsa.mpi.pub tutorial@[headIP] -# --------------- 以下操作均在head节点中执行 --------------- -# 准备训练数据 -python prepare.py -# 拷贝训练程序和字典文件到每台MPI节点 -cat machines | xargs -i scp word_dict.pickle train.py start_mpi_train.sh machines {}:/home/tutorial -# 创建日志目录 -mpirun -hostfile machines -n 3 mkdir /home/tutorial/logs -# 拷贝训练数据到各自的节点 -scp train.txt-00000 test.txt-00000 [node1IP]:/home/tutorial -scp train.txt-00001 test.txt-00001 [node2IP]:/home/tutorial -scp train.txt-00002 test.txt-00002 [node3IP]:/home/tutorial -# 启动训练任务 -mpirun -hostfile machines -n 3 /home/tutorial/start_mpi_train.sh -``` - -### 在Kubernetes集群中提交训练作业 +## 在不同集群中运行 -此部分的使用方法可以参考[here](../k8s/k8s_distributed_cn.md)。 + - [fabric集群](fabric_cn.md) + - [openmpi集群](openmpi_cn.md) + - [kubernetes单机](k8s_cn.md) + - [kubernetes distributed分布式](k8s_distributed_cn.md) + - [AWS上运行kubernetes集群训练](k8s_aws_cn.md) diff --git a/doc/howto/usage/cluster/cluster_train_en.md b/doc/howto/usage/cluster/cluster_train_en.md index baa97c0c02ae490fff8587071bd2d4adfb5325e3..915405ca5b446981515e301ca4b7ee065a82a9ff 100644 --- a/doc/howto/usage/cluster/cluster_train_en.md +++ b/doc/howto/usage/cluster/cluster_train_en.md @@ -1,23 +1,4 @@ -# PaddlePaddle Distributed Training - -* [Introduction](#introduction) -* [Preparations](#preparations) -* [Command-line arguments](#command-line-arguments) - * [Starting parameter server](#starting-parameter-server) - * [Starting trainer](#starting-trainer) - * [Prepare Training Dataset](#prepare-training-dataset) - * [Prepare Training program](#prepare-training-program) -* [Use cluster platforms or cluster management tools](#use-cluster-platforms-or-cluster-management-tools) - * [Cluster Training Using Fabric](#cluster-training-using-fabric) - * [Prepare a Linux cluster](#prepare-a-linux-cluster) - * [Launching Cluster Job](#launching-cluster-job) - * [Kill Cluster Job](#kill-cluster-job) - * [Check Cluster Training Result](#check-cluster-training-result) - * [Check Model Output](#check-model-output) - * [Cluster Training Using OpenMPI](#cluster-training-using-openmpi) - * [Prepare an OpenMPI cluster](#prepare-an-openmpi-cluster) - * [Launching Cluster Job](#launching-cluster-job-1) - * [Cluster Training Using Kubernetes](#cluster-training-using-kubernetes) +# Distributed Training ## Introduction @@ -35,7 +16,7 @@ When training with synchronize SGD, PaddlePaddle uses an internal "synchronize b ## Preparations 1. Prepare your computer cluster. It's normally a bunch of Linux servers connected by LAN. Each server will be assigned a unique IP address. The computers in the cluster can be called "nodes". -2. Install PaddlePaddle on every node. If you are going to take advantage of GPU cards, you'll also need to install proper driver and CUDA libraries. To install PaddlePaddle please read [this build and install](https://github.com/PaddlePaddle/Paddle/tree/develop/doc/getstarted/build_and_install) document. We strongly recommend using [Docker installation](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/getstarted/build_and_install/docker_install_en.rst). +2. Install PaddlePaddle on every node. If you are going to take advantage of GPU cards, you'll also need to install proper driver and CUDA libraries. To install PaddlePaddle please read [this build and install](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/build_and_install/index_en.html) document. We strongly recommend using [Docker installation](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/build_and_install/docker_install_en.html). After installation, you can check the version by typing the below command (run a docker container if using docker: `docker run -it paddlepaddle/paddle:[tag] /bin/bash`): @@ -67,12 +48,12 @@ If you wish to run parameter servers in background, and save a log file, you can $ stdbuf -oL /usr/bin/nohup paddle pserver --port=7164 --ports_num=1 --ports_num_for_sparse=1 --num_gradient_servers=1 &> pserver.log ``` -| param | required | default | description | -| ------------- | ------------- | ------------- | ------------- | -| port | required | 7164 | port which parameter server will listen on. If ports_num greater than 1, parameter server will listen on multiple ports for more network throughput | -| ports_num | required | 1 | total number of ports will listen on | -| ports_num_for_sparse | required | 1 | number of ports which serves sparse parameter update | -| num_gradient_servers | required | 1 | total number of gradient servers | +Parameter Description + +- port: **required, default 7164**, port which parameter server will listen on. If ports_num greater than 1, parameter server will listen on multiple ports for more network throughput. +- ports_num: **required, default 1**, total number of ports will listen on. +- ports_num_for_sparse: **required, default 1**, number of ports which serves sparse parameter update. +- num_gradient_servers: **required, default 1**, total number of gradient servers. ### Starting trainer Type the command below to start the trainer(name the file whatever you want, like "train.py") @@ -111,16 +92,16 @@ paddle.init( pservers="127.0.0.1") ``` -| param | required | default | description | -| ------------- | ------------- | ------------- | ------------- | -| use_gpu | optional | False | set to "True" to enable GPU training | -| trainer_count | required | 1 | total count of trainers in the training job | -| port | required | 7164 | port to connect to parameter server | -| ports_num | required | 1 | number of ports for communication | -| ports_num_for_sparse | required | 1 | number of ports for sparse type caculation | -| num_gradient_servers | required | 1 | total number of gradient server | -| trainer_id | required | 0 | ID for every trainer, start from 0 | -| pservers | required | 127.0.0.1 | list of IPs of parameter servers, separated by "," | +Parameter Description + +- use_gpu: **optional, default False**, set to "True" to enable GPU training. +- trainer_count: **required, default 1**, total count of trainers in the training job. +- port: **required, default 7164**, port to connect to parameter server. +- ports_num: **required, default 1**, number of ports for communication. +- ports_num_for_sparse: **required, default 1**, number of ports for sparse type caculation. +- num_gradient_servers: **required, default 1**, total number of gradient server. +- trainer_id: **required, default 0**, ID for every trainer, start from 0. +- pservers: **required, default 127.0.0.1**, list of IPs of parameter servers, separated by ",". ### Prepare Training Dataset @@ -178,7 +159,7 @@ Your workspace may looks like: - `my_lib.py`: user defined libraries, like PIL libs. This is optional. - `word_dict.pickle`: dict file for training word embeding. -- `train.py`: training program. Sample code: [api_train_v2_cluster.py](https://github.com/PaddlePaddle/Paddle/tree/develop/doc/howto/usage/cluster/src/word2vec/prepare.py). ***NOTE:*** You may need to modify the head part of `train.py` when using different cluster platform to retrive configuration environment variables: +- `train.py`: training program. Sample code: [api_train_v2_cluster.py](https://github.com/PaddlePaddle/Paddle/tree/develop/doc/howto/usage/cluster/src/word2vec/api_train_v2_cluster.py). ***NOTE:*** You may need to modify the head part of `train.py` when using different cluster platform to retrive configuration environment variables: ```python cluster_train_file = "./train_data_dir/train/train.txt" @@ -202,92 +183,9 @@ We'll introduce cluster job management on these platforms. The examples can be f These cluster platforms provide API or environment variables for training processes, when the job is dispatched to different nodes. Like node ID, IP or total number of nodes etc. -### Cluster Training Using Fabric - -#### Prepare a Linux cluster - -Run `kubectl -f ssh_servers.yaml` under the directory: `paddle/scripts/cluster_train_v2/fabric/docker_cluster` will launch a demo cluster. Run `kubectl get po -o wide` to get IP addresses of these nodes. - -#### Launching Cluster Job -`paddle.py` provides automatical scripts to start all PaddlePaddle cluster processes in different nodes. By default, all command line options can be set as `paddle.py` command options and `paddle.py` will transparently and automatically set these options to PaddlePaddle lower level processes. - -`paddle.py`provides two distinguished command option for easy job launching. - -- `job_dispatch_package` set it with local `workspace` directory, it will be dispatched to all nodes which is set in `conf.py`. It could be helpful for frequently manipulating workspace files. otherwise, frequent multi-nodes workspace deployment is very annoying. -- `job_workspace` set it with already deployed workspace directory, `paddle.py` will skip dispatch stage to directly launch cluster job with all nodes. It could help to reduce heavy -dispatch latency. - -`cluster_train/run.sh` provides command line sample to run `demo/recommendation` cluster job, just modify `job_dispatch_package` and `job_workspace` with your defined directory, then: -``` -sh run.sh -``` - -The cluster Job will start in several seconds. - -#### Kill Cluster Job -`paddle.py` can capture `Ctrl + C` SIGINT signal to automatically kill all processes launched by it. So just stop `paddle.py` to kill cluster job. You should manually kill the job if the program crashed. - -#### Check Cluster Training Result -Check log in $workspace/log for details, each node owns same log structure. - -`paddle_trainer.INFO` -It provides almost all internal output log for training, same as local training. Check runtime model convergence here. - -`paddle_pserver2.INFO` -It provides parameter server running log, which could help to diagnose distributed error. - -`server.log` -It provides stderr and stdout of parameter server process. Check error log if training crashes. - -`train.log` -It provides stderr and stdout of trainer process. Check error log if training crashes. - -#### Check Model Output -After one pass finished, model files will be written in `output` directory in node 0. -`nodefile` in workspace indicates the node id of current cluster job. - -### Cluster Training Using OpenMPI - -#### Prepare an OpenMPI cluster - -Run the following command to start a 3-node MPI cluster and one "head" node. - -```bash -cd paddle/scripts/cluster_train_v2/openmpi/docker_cluster -kubectl create -f head.yaml -kubectl create -f mpi-nodes.yaml -``` - -Then you can log in to every OpenMPI node using ssh without input any passwords. - -#### Launching Cluster Job - -Follow the steps to launch a PaddlePaddle training job in OpenMPI cluster:\ - -```bash -# find out node IP addresses -kubectl get po -o wide -# generate a "machines" file containing node IP addresses -kubectl get po -o wide | grep nodes | awk '{print $6}' > machines -# copy necessary files onto "head" node -scp -i ssh/id_rsa.mpi.pub machines prepare.py train.py start_mpi_train.sh tutorial@[headIP]:~ -# login to head node using ssh -ssh -i ssh/id_rsa.mpi.pub tutorial@[headIP] -# --------------- in head node --------------- -# prepare training data -python prepare.py -# copy training data and dict file to MPI nodes -cat machines | xargs -i scp word_dict.pickle train.py start_mpi_train.sh machines {}:/home/tutorial -# creat a directory for storing log files -mpirun -hostfile machines -n 3 mkdir /home/tutorial/logs -# copy training data to every node -scp train.txt-00000 test.txt-00000 [node1IP]:/home/tutorial -scp train.txt-00001 test.txt-00001 [node2IP]:/home/tutorial -scp train.txt-00002 test.txt-00002 [node3IP]:/home/tutorial -# start the job -mpirun -hostfile machines -n 3 /home/tutorial/start_mpi_train.sh -``` - -### Cluster Training Using Kubernetes +## Use different clusters -The details can be found [here](../k8s/k8s_cn.md) + - [fabric](fabric_en.md) + - [openmpi](openmpi_en.md) + - [kubernetes](k8s_en.md) + - [kubernetes on AWS](k8s_aws_en.md) diff --git a/doc/howto/usage/cluster/fabric_cn.md b/doc/howto/usage/cluster/fabric_cn.md new file mode 100644 index 0000000000000000000000000000000000000000..0385e401b399a51fad112e604dc56cb2f84c0a4b --- /dev/null +++ b/doc/howto/usage/cluster/fabric_cn.md @@ -0,0 +1,42 @@ +# 使用fabric启动集群训练 + +## 准备一个Linux集群 +可以在`paddle/scripts/cluster_train_v2/fabric/docker_cluster`目录下，执行`kubectl -f ssh_servers.yaml`启动一个测试集群，并使用`kubectl get po -o wide`获得这些节点的IP地址。 + +## 启动集群作业 + +`paddle.py` 提供了自动化脚本来启动不同节点中的所有 PaddlePaddle 集群进程。默认情况下，所有命令行选项可以设置为 `paddle.py` 命令选项并且 `paddle.py` 将透明、自动地将这些选项应用到 PaddlePaddle 底层进程。 + +`paddle.py` 为方便作业启动提供了两个独特的命令选项。 + +- `job_dispatch_package` 设为本地 `workspace` 目录，它将被分发到 `conf.py` 中设置的所有节点。它有助于帮助频繁修改和访问工作区文件的用户减少负担，否则频繁的多节点工作空间部署可能会很麻烦。 +- `job_workspace` 设为已部署的工作空间目录，`paddle.py` 将跳过分发阶段直接启动所有节点的集群作业。它可以帮助减少分发延迟。 + +`cluster_train/run.sh` 提供了命令样例来运行 `doc/howto/usage/cluster/src/word2vec` 集群任务，只需用您定义的目录修改 `job_dispatch_package` 和 `job_workspace`，然后： +``` +sh run.sh +``` + +集群作业将会在几秒后启动。 + +## 终止集群作业 +`paddle.py`能获取`Ctrl + C` SIGINT 信号来自动终止它启动的所有进程。只需中断 `paddle.py` 任务来终止集群作业。如果程序崩溃你也可以手动终止。 + +## 检查集群训练结果 +详细信息请检查 $workspace/log 里的日志，每一个节点都有相同的日志结构。 + +`paddle_trainer.INFO` +提供几乎所有训练的内部输出日志，与本地训练相同。这里检验运行时间模型的收敛。 + +`paddle_pserver2.INFO` +提供 pserver 运行日志，有助于诊断分布式错误。 + +`server.log` +提供 parameter server 进程的 stderr 和 stdout。训练失败时可以检查错误日志。 + +`train.log` +提供训练过程的 stderr 和 stdout。训练失败时可以检查错误日志。 + +## 检查模型输出 +运行完成后，模型文件将被写入节点 0 的 `output` 目录中。 +工作空间中的 `nodefile` 表示当前集群作业的节点 ID。 diff --git a/doc/howto/usage/cluster/fabric_en.md b/doc/howto/usage/cluster/fabric_en.md new file mode 100644 index 0000000000000000000000000000000000000000..bf270d89ab8514801ca4629cf412f73257429df9 --- /dev/null +++ b/doc/howto/usage/cluster/fabric_en.md @@ -0,0 +1,43 @@ +# Cluster Training Using Fabric + +## Prepare a Linux cluster + +Run `kubectl -f ssh_servers.yaml` under the directory: `paddle/scripts/cluster_train_v2/fabric/docker_cluster` will launch a demo cluster. Run `kubectl get po -o wide` to get IP addresses of these nodes. + +## Launching Cluster Job +`paddle.py` provides automatical scripts to start all PaddlePaddle cluster processes in different nodes. By default, all command line options can be set as `paddle.py` command options and `paddle.py` will transparently and automatically set these options to PaddlePaddle lower level processes. + +`paddle.py`provides two distinguished command option for easy job launching. + +- `job_dispatch_package` set it with local `workspace` directory, it will be dispatched to all nodes which is set in `conf.py`. It could be helpful for frequently manipulating workspace files. otherwise, frequent multi-nodes workspace deployment is very annoying. +- `job_workspace` set it with already deployed workspace directory, `paddle.py` will skip dispatch stage to directly launch cluster job with all nodes. It could help to reduce heavy +dispatch latency. + +`cluster_train/run.sh` provides command line sample to run `demo/recommendation` cluster job, just modify `job_dispatch_package` and `job_workspace` with your defined directory, then: +``` +sh run.sh +``` + +The cluster Job will start in several seconds. + +## Kill Cluster Job +`paddle.py` can capture `Ctrl + C` SIGINT signal to automatically kill all processes launched by it. So just stop `paddle.py` to kill cluster job. You should manually kill the job if the program crashed. + +## Check Cluster Training Result +Check log in $workspace/log for details, each node owns same log structure. + +`paddle_trainer.INFO` +It provides almost all internal output log for training, same as local training. Check runtime model convergence here. + +`paddle_pserver2.INFO` +It provides parameter server running log, which could help to diagnose distributed error. + +`server.log` +It provides stderr and stdout of parameter server process. Check error log if training crashes. + +`train.log` +It provides stderr and stdout of trainer process. Check error log if training crashes. + +## Check Model Output +After one pass finished, model files will be written in `output` directory in node 0. +`nodefile` in workspace indicates the node id of current cluster job. diff --git a/doc/howto/usage/cluster/k8s_aws_cn.md b/doc/howto/usage/cluster/k8s_aws_cn.md new file mode 120000 index 0000000000000000000000000000000000000000..c44cd9a731bed7067cdf19aa2f714abdce6c736a --- /dev/null +++ b/doc/howto/usage/cluster/k8s_aws_cn.md @@ -0,0 +1 @@ +k8s_aws_en.md \ No newline at end of file diff --git a/doc/howto/usage/k8s/k8s_aws_en.md b/doc/howto/usage/cluster/k8s_aws_en.md similarity index 98% rename from doc/howto/usage/k8s/k8s_aws_en.md rename to doc/howto/usage/cluster/k8s_aws_en.md index ce72b0803818d5bf0c18753c421848cf2fc1b668..0dfa8237a3fa2c9c3ee11e873c9fbbed3cd6018f 100644 --- a/doc/howto/usage/k8s/k8s_aws_en.md +++ b/doc/howto/usage/cluster/k8s_aws_en.md @@ -493,7 +493,7 @@ spec: spec: containers: - name: paddle-data - image: paddledev/paddle-tutorial:k8s_data + image: paddlepaddle/paddle-tutorial:k8s_data imagePullPolicy: Always volumeMounts: - mountPath: "/efs" @@ -522,7 +522,7 @@ NAME DESIRED SUCCESSFUL AGE paddle-data 1 1 6m ``` -Data preparation is done by docker image `paddledev/paddle-tutorial:k8s_data`, see [here](src/k8s_data/README.md) for how to build this docker image and source code. +Data preparation is done by docker image `paddlepaddle/paddle-tutorial:k8s_data`, see [here](src/k8s_data/README.md) for how to build this docker image and source code. #### Start Training @@ -545,7 +545,7 @@ spec: claimName: efsvol containers: - name: trainer - image: paddledev/paddle-tutorial:k8s_train + image: paddlepaddle/paddle-tutorial:k8s_train command: ["bin/bash", "-c", "/root/start.sh"] env: - name: JOB_NAME @@ -617,7 +617,7 @@ kubectl --kubeconfig=kubeconfig log -f POD_NAME Run `kubectl --kubeconfig=kubeconfig describe job paddle-cluster-job` to check training job status. It will complete in around 20 minutes. -The details for start `pserver` and `trainer` are hidden inside docker image `paddledev/paddle-tutorial:k8s_train`, see [here](src/k8s_train/README.md) for how to build the docker image and source code. +The details for start `pserver` and `trainer` are hidden inside docker image `paddlepaddle/paddle-tutorial:k8s_train`, see [here](src/k8s_train/README.md) for how to build the docker image and source code. #### Inspect Training Output diff --git a/doc/howto/usage/k8s/k8s_cn.md b/doc/howto/usage/cluster/k8s_cn.md similarity index 83% rename from doc/howto/usage/k8s/k8s_cn.md rename to doc/howto/usage/cluster/k8s_cn.md index ab07cb9cd5b135ddea82b3360720537f1dc5a801..c1a11f7165a2f9da9dd044641274447e7943a597 100644 --- a/doc/howto/usage/k8s/k8s_cn.md +++ b/doc/howto/usage/cluster/k8s_cn.md @@ -1,21 +1,22 @@ # Kubernetes单机训练 -在这篇文档里，我们介绍如何在 Kubernetes 集群上启动一个单机使用CPU的Paddle训练作业。在下一篇中，我们将介绍如何启动分布式训练作业。 +在这篇文档里，我们介绍如何在 Kubernetes 集群上启动一个单机使用CPU的PaddlePaddle训练作业。在下一篇中，我们将介绍如何启动分布式训练作业。 ## 制作Docker镜像 -在一个功能齐全的Kubernetes机群里，通常我们会安装Ceph等分布式文件系统来存储训练数据。这样的话，一个分布式Paddle训练任务中的每个进程都可以从Ceph读取数据。在这个例子里，我们只演示一个单机作业，所以可以简化对环境的要求，把训练数据直接放在 -Paddle的Docker image里。为此，我们需要制作一个包含训练数据的Paddle镜像。 +在一个功能齐全的Kubernetes机群里，通常我们会安装Ceph等分布式文件系统来存储训练数据。这样的话，一个分布式PaddlePaddle训练任务中 +的每个进程都可以从Ceph读取数据。在这个例子里，我们只演示一个单机作业，所以可以简化对环境的要求，把训练数据直接放在 +PaddlePaddle的Docker Image里。为此，我们需要制作一个包含训练数据的PaddlePaddle镜像。 + +PaddlePaddle的 `paddlepaddle/paddle:cpu-demo-latest` 镜像里有PaddlePaddle的源码与demo， +（请注意，默认的PaddlePaddle生产环境镜像 `paddlepaddle/paddle:latest` 是不包括源码的，PaddlePaddle的各版本镜像可以参考 +[Docker Installation Guide](http://paddlepaddle.org/docs/develop/documentation/zh/getstarted/build_and_install/docker_install_cn.html)）， +下面我们使用这个镜像来下载数据到Docker Container中，并把这个包含了训练数据的Container保存为一个新的镜像。 -Paddle 的 [Quick Start Tutorial](http://www.paddlepaddle.org/doc/demo/quick_start/index_en.html) -里介绍了用Paddle源码中的脚本下载训练数据的过程。 -而 `paddledev/paddle:cpu-demo-latest` 镜像里有 Paddle 源码与demo，（ 请注意，默认的 -Paddle镜像 `paddledev/paddle:cpu-latest` 是不包括源码的, Paddle的各版本镜像可以参考 [Docker installation guide](http://www.paddlepaddle.org/doc/build/docker_install.html) ），所以我们使用这个镜像来下载训练数据到Docker container中，然后把这个包含了训练数据的container保存为一个新的镜像。 - ### 运行容器 ``` -$ docker run --name quick_start_data -it paddledev/paddle:cpu-demo-latest +$ docker run --name quick_start_data -it paddlepaddle/paddle:cpu-demo-latest ``` ### 下载数据 @@ -103,7 +104,7 @@ spec: restartPolicy: Never ``` -### 创建Paddle Job +### 创建PaddlePaddle Job 使用上文创建的yaml文件创建Kubernetes Job，命令为： diff --git a/doc/howto/usage/k8s/k8s_distributed_cn.md b/doc/howto/usage/cluster/k8s_distributed_cn.md similarity index 88% rename from doc/howto/usage/k8s/k8s_distributed_cn.md rename to doc/howto/usage/cluster/k8s_distributed_cn.md index a9bebf09558b06993119803458977abedbbfbdd0..167089b8074b33e3b094fa3ec8e377630cec42ac 100644 --- a/doc/howto/usage/k8s/k8s_distributed_cn.md +++ b/doc/howto/usage/cluster/k8s_distributed_cn.md @@ -1,8 +1,6 @@ # Kubernetes分布式训练 -前一篇文章介绍了如何在Kubernetes集群上启动一个单机PaddlePaddle训练作业 (Job)。在这篇文章里，我们介绍如何在Kubernetes集群上进行分布式PaddlePaddle训练作业。关于PaddlePaddle的分布式训练，文章 [Cluster Training](https://github.com/baidu/Paddle/blob/develop/doc/cluster/opensource/cluster_train.md)介绍了一种通过SSH远程分发任务，进行分布式训练的方法，与此不同的是，本文将介绍在Kubernetes容器管理平台上快速构建PaddlePaddle容器集群，进行分布式训练的方案。 - -有关Kubernetes相关概念以及如何搭建和配置Kubernetes集群，可以参考[k8s_basis](./k8s_basis_cn.md)。 +前一篇文章介绍了如何在Kubernetes集群上启动一个单机PaddlePaddle训练作业 (Job)。在这篇文章里，我们介绍如何在Kubernetes集群上进行分布式PaddlePaddle训练作业。关于PaddlePaddle的分布式训练，文章 [Cluster Training](http://www.paddlepaddle.org/docs/develop/documentation/zh/howto/usage/cluster/cluster_train_cn.html)介绍了一种通过SSH远程分发任务，进行分布式训练的方法，与此不同的是，本文将介绍在Kubernetes容器管理平台上快速构建PaddlePaddle容器集群，进行分布式训练的方案。 ## 整体方案 @@ -28,7 +26,7 @@ PaddlePaddle镜像需要提供`paddle pserver`与`paddle train`进程的运行 - 拷贝训练文件到容器内 - 生成`paddle pserver`与`paddle train`进程的启动参数，并且启动训练 -因为官方镜像 `paddledev/paddle:cpu-latest` 内已经包含PaddlePaddle的执行程序但是还没上述功能，所以我们可以在这个基础上，添加启动脚本，制作新镜像来完成以上的工作。参考镜像的[*Dockerfile*](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/usage/cluster/k8s/src/k8s_train/Dockerfile)。 +因为官方镜像 `paddlepaddle/paddle:latest` 内已经包含PaddlePaddle的执行程序但是还没上述功能，所以我们可以在这个基础上，添加启动脚本，制作新镜像来完成以上的工作。参考镜像的[*Dockerfile*](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/usage/cluster/src/k8s_train/Dockerfile)。 ```bash $ cd doc/howto/usage/k8s/src/k8s_train @@ -62,7 +60,7 @@ spec: hostNetwork: true containers: - name: paddle-data - image: paddledev/paddle-tutorial:k8s_data + image: paddlepaddle/paddle-tutorial:k8s_data imagePullPolicy: Always volumeMounts: - mountPath: "/mnt" @@ -149,20 +147,19 @@ spec: 文件中，`metadata`下的`name`表示这个job的名字。`parallelism，completions`字段表示这个job会同时开启3个PaddlePaddle节点，成功训练且退出的pod数目为3时，这个job才算成功结束。然后申明一个存储卷`jobpath`，代表宿主机目录`/home/work/mfs`，在对容器的描述`containers`字段中，将此目录挂载为容器的`/home/jobpath`目录，这样容器的`/home/jobpath`目录就成为了共享存储，放在这个目录里的文件其实是保存到了MFS上。 -`env`字段表示容器的环境变量，我们将`paddle`运行的一些参数通过这种方式传递到容器内。 +`env`字段表示容器的环境变量，我们将`paddle`运行的一些参数通过这种方式传递到容器内： + -环境变量 | 说明 ---- | --- -JOB_PATH | 共享存储挂在的路径 -JOB_NAME | Job的名字 -TRAIN_CONFIG_DIR | 本次训练文件所在目录，与JOB_PATH,JOB_NAME组合可以找到本次训练需要的文件路径 -CONF_PADDLE_NIC | `paddle pserver`进程需要的`--nics`参数，即网卡名 -CONF_PADDLE_PORT | `paddle paserver`的`--port`参数 -CONF_PADDLE_PORTS_NUM | 稠密更新的端口数量，即`--ports_num`参数 -CONF_PADDLE_PORTS_NUM_SPARSE | 稀疏更新的端口数量，即`--ports_num_for_sparse`参数 -CONF_PADDLE_GRADIENT_NUM | 训练节点数量，即`--num_gradient_servers参数` +- JOB_PATH：共享存储挂在的路径 +- JOB_NAME：Job的名字 +- TRAIN_CONFIG_DIR：本次训练文件所在目录，与JOB_PATH,JOB_NAME组合可以找到本次训练需要的文件路径 +- CONF_PADDLE_NIC：`paddle pserver`进程需要的`--nics`参数，即网卡名 +- CONF_PADDLE_PORT：`paddle paserver`的`--port`参数 +- CONF_PADDLE_PORTS_NUM：稠密更新的端口数量，即`--ports_num`参数 +- CONF_PADDLE_PORTS_NUM_SPARSE：稀疏更新的端口数量，即`--ports_num_for_sparse`参数 +- CONF_PADDLE_GRADIENT_NUM：训练节点数量，即`--num_gradient_servers参数` -这些参数的具体描述，读者可以查看[这里](http://www.paddlepaddle.org/doc/ui/cmd_argument/detail_introduction.html#parameter-server-and-distributed-communication)。 +这些参数的具体描述，读者可以查看[这里](http://www.paddlepaddle.org/docs/develop/documentation/zh/howto/usage/cmd_parameter/detail_introduction_cn.html)。 编写完YAML文件后，可以使用Kubernetes的命令行工具创建job。 diff --git a/doc/howto/usage/k8s/k8s_en.md b/doc/howto/usage/cluster/k8s_en.md similarity index 79% rename from doc/howto/usage/k8s/k8s_en.md rename to doc/howto/usage/cluster/k8s_en.md index 0c3ab05b708e7a924577c26496b8c55126e76c62..c374f00a495d705ceddf8d3d930768ceeb93282b 100644 --- a/doc/howto/usage/k8s/k8s_en.md +++ b/doc/howto/usage/cluster/k8s_en.md @@ -1,18 +1,27 @@ -# Paddle On Kubernetes +# PaddlePaddle On Kubernetes ->In this article, we will introduce how to run Paddle training job on single CPU machine using Kubernetes. In next article, we will introduce how to run Paddle training job on distributed cluster. +In this article, we will introduce how to run PaddlePaddle training job on single CPU machine using Kubernetes. In next article, we will introduce how to run PaddlePaddle training job on distributed cluster. ## Build Docker Image -In distributed Kubernetes cluster, we will use Ceph or other shared storage system for storing training related data so that all processes in Paddle training can retrieve data from Ceph. In this example, we will only demo training job on single machine. In order to simplify the requirement of the environment, we will directly put training data into Paddle's Docker Image, so we need to create a Paddle Docker image that already includes the training data. +In distributed Kubernetes cluster, we will use Ceph or other distributed +storage system for storing training related data so that all processes in +PaddlePaddle training can retrieve data from Ceph. In this example, we will +only demo training job on single machine. In order to simplify the requirement +of the environment, we will directly put training data into the PaddlePaddle Docker Image, +so we need to create a PaddlePaddle Docker image that includes the training data. + +The production Docker Image `paddlepaddle/paddle:cpu-demo-latest` has the PaddlePaddle +source code and demo. (Caution: Default PaddlePaddle Docker Image `paddlepaddle/paddle:latest` doesn't include +the source code, PaddlePaddle's different versions of Docker Image can be referred here: +[Docker Installation Guide](http://paddlepaddle.org/docs/develop/documentation/zh/getstarted/build_and_install/docker_install_en.html)), +so we run this Docker Image and download the training data, and then commit the whole +Container to be a new Docker Image. -Paddle's [Quick Start Tutorial](http://www.paddlepaddle.org/doc/demo/quick_start/index_en.html) introduces how to download and train data by using script from Paddle's source code. -And `paddledev/paddle:cpu-demo-latest` image has the Paddle source code and demo. (Caution: Default Paddle image `paddledev/paddle:cpu-latest` doesn't include the source code, Paddle's different versions of image can be referred here: [Docker installation guide](http://www.paddlepaddle.org/doc/build/docker_install.html)), so we run this container and download the training data, and then commit the whole container to be a new Docker image. - ### Run Docker Container ``` -$ docker run --name quick_start_data -it paddledev/paddle:cpu-demo-latest +$ docker run --name quick_start_data -it paddlepaddle/paddle:cpu-demo-latest ``` ### Download Training Data @@ -67,7 +76,7 @@ $ docker commit quick_start_data mypaddle/paddle:quickstart ## Use Kubernetes For Training ->We will use Kubernetes job for training process, following steps shows how to do the training with Kubernetes. +We will use Kubernetes job for training process, following steps shows how to do the training with Kubernetes. ### Create Yaml Files @@ -99,7 +108,7 @@ spec: restartPolicy: Never ``` -### Start Paddle Job +### Start PaddlePaddle Job Using the above yaml file to start the Kubernetes job. diff --git a/doc/howto/usage/cluster/openmpi_cn.md b/doc/howto/usage/cluster/openmpi_cn.md new file mode 100644 index 0000000000000000000000000000000000000000..831cafdc03c6a908f31769d0467de022df42dab5 --- /dev/null +++ b/doc/howto/usage/cluster/openmpi_cn.md @@ -0,0 +1,41 @@ +# 在OpenMPI集群中提交训练作业 + +## 准备OpenMPI集群 + +执行下面的命令以启动3个节点的OpenMPI集群和一个"head"节点： + +```bash +paddle/scripts/cluster_train_v2/openmpi/docker_cluster +kubectl create -f head.yaml +kubectl create -f mpi-nodes.yaml +``` + +然后可以从head节点ssh无密码登录到OpenMPI的每个节点上。 + +## 启动集群作业 + +您可以按照下面的步骤在OpenMPI集群中提交paddle训练任务： + +```bash +# 获得head和node节点的IP地址 +kubectl get po -o wide +# 将node节点的IP地址保存到machines文件中 +kubectl get po -o wide | grep nodes | awk '{print $6}' > machines +# 拷贝必要的文件到head节点 +scp -i ssh/id_rsa.mpi.pub machines prepare.py train.py start_mpi_train.sh tutorial@[headIP]:~ +# ssh 登录到head节点 +ssh -i ssh/id_rsa.mpi.pub tutorial@[headIP] +# --------------- 以下操作均在head节点中执行 --------------- +# 准备训练数据 +python prepare.py +# 拷贝训练程序和字典文件到每台MPI节点 +cat machines | xargs -i scp word_dict.pickle train.py start_mpi_train.sh machines {}:/home/tutorial +# 创建日志目录 +mpirun -hostfile machines -n 3 mkdir /home/tutorial/logs +# 拷贝训练数据到各自的节点 +scp train.txt-00000 test.txt-00000 [node1IP]:/home/tutorial +scp train.txt-00001 test.txt-00001 [node2IP]:/home/tutorial +scp train.txt-00002 test.txt-00002 [node3IP]:/home/tutorial +# 启动训练任务 +mpirun -hostfile machines -n 3 /home/tutorial/start_mpi_train.sh +``` diff --git a/doc/howto/usage/cluster/openmpi_en.md b/doc/howto/usage/cluster/openmpi_en.md new file mode 100644 index 0000000000000000000000000000000000000000..09af46e25ebe1f843dc7c7be0997dc706413b65c --- /dev/null +++ b/doc/howto/usage/cluster/openmpi_en.md @@ -0,0 +1,41 @@ +# Cluster Training Using OpenMPI + +## Prepare an OpenMPI cluster + +Run the following command to start a 3-node MPI cluster and one "head" node. + +```bash +cd paddle/scripts/cluster_train_v2/openmpi/docker_cluster +kubectl create -f head.yaml +kubectl create -f mpi-nodes.yaml +``` + +Then you can log in to every OpenMPI node using ssh without input any passwords. + +## Launching Cluster Job + +Follow the steps to launch a PaddlePaddle training job in OpenMPI cluster:\ + +```bash +# find out node IP addresses +kubectl get po -o wide +# generate a "machines" file containing node IP addresses +kubectl get po -o wide | grep nodes | awk '{print $6}' > machines +# copy necessary files onto "head" node +scp -i ssh/id_rsa.mpi.pub machines prepare.py train.py start_mpi_train.sh tutorial@[headIP]:~ +# login to head node using ssh +ssh -i ssh/id_rsa.mpi.pub tutorial@[headIP] +# --------------- in head node --------------- +# prepare training data +python prepare.py +# copy training data and dict file to MPI nodes +cat machines | xargs -i scp word_dict.pickle train.py start_mpi_train.sh machines {}:/home/tutorial +# creat a directory for storing log files +mpirun -hostfile machines -n 3 mkdir /home/tutorial/logs +# copy training data to every node +scp train.txt-00000 test.txt-00000 [node1IP]:/home/tutorial +scp train.txt-00001 test.txt-00001 [node2IP]:/home/tutorial +scp train.txt-00002 test.txt-00002 [node3IP]:/home/tutorial +# start the job +mpirun -hostfile machines -n 3 /home/tutorial/start_mpi_train.sh +``` diff --git a/doc/howto/usage/k8s/src/Dockerfile b/doc/howto/usage/cluster/src/Dockerfile similarity index 54% rename from doc/howto/usage/k8s/src/Dockerfile rename to doc/howto/usage/cluster/src/Dockerfile index 3a73606c61432329b4cc2d2f8daadc5af8735c96..e178bf4da0f32fca9586b5b69a2c7419de5d9cb1 100644 --- a/doc/howto/usage/k8s/src/Dockerfile +++ b/doc/howto/usage/cluster/src/Dockerfile @@ -1,4 +1,4 @@ -FROM paddledev/paddle:cpu-latest +FROM paddlepaddle/paddle:latest MAINTAINER zjsxzong89@gmail.com diff --git a/doc/howto/usage/k8s/src/add_security_group.png b/doc/howto/usage/cluster/src/add_security_group.png similarity index 100% rename from doc/howto/usage/k8s/src/add_security_group.png rename to doc/howto/usage/cluster/src/add_security_group.png diff --git a/doc/howto/usage/k8s/src/create_efs.png b/doc/howto/usage/cluster/src/create_efs.png similarity index 100% rename from doc/howto/usage/k8s/src/create_efs.png rename to doc/howto/usage/cluster/src/create_efs.png diff --git a/doc/howto/usage/k8s/src/efs_mount.png b/doc/howto/usage/cluster/src/efs_mount.png similarity index 100% rename from doc/howto/usage/k8s/src/efs_mount.png rename to doc/howto/usage/cluster/src/efs_mount.png diff --git a/doc/howto/usage/cluster/src/k8s-paddle-arch.png b/doc/howto/usage/cluster/src/k8s-paddle-arch.png new file mode 100644 index 0000000000000000000000000000000000000000..b3800c4fe81302d35e49f7dbacb9221c4dfa5cde Binary files /dev/null and b/doc/howto/usage/cluster/src/k8s-paddle-arch.png differ diff --git a/doc/howto/usage/k8s/src/k8s_data/Dockerfile b/doc/howto/usage/cluster/src/k8s_data/Dockerfile similarity index 100% rename from doc/howto/usage/k8s/src/k8s_data/Dockerfile rename to doc/howto/usage/cluster/src/k8s_data/Dockerfile diff --git a/doc/howto/usage/k8s/src/k8s_data/README.md b/doc/howto/usage/cluster/src/k8s_data/README.md similarity index 100% rename from doc/howto/usage/k8s/src/k8s_data/README.md rename to doc/howto/usage/cluster/src/k8s_data/README.md diff --git a/doc/howto/usage/k8s/src/k8s_data/get_data.sh b/doc/howto/usage/cluster/src/k8s_data/get_data.sh similarity index 100% rename from doc/howto/usage/k8s/src/k8s_data/get_data.sh rename to doc/howto/usage/cluster/src/k8s_data/get_data.sh diff --git a/doc/howto/usage/k8s/src/k8s_train/Dockerfile b/doc/howto/usage/cluster/src/k8s_train/Dockerfile similarity index 77% rename from doc/howto/usage/k8s/src/k8s_train/Dockerfile rename to doc/howto/usage/cluster/src/k8s_train/Dockerfile index c0fca1f9a945921e6e8899fee2db8845e66136a1..77f021a89a70d934bf70424eaa3c6dc3f7c93a28 100644 --- a/doc/howto/usage/k8s/src/k8s_train/Dockerfile +++ b/doc/howto/usage/cluster/src/k8s_train/Dockerfile @@ -1,4 +1,4 @@ -FROM paddledev/paddle:cpu-latest +FROM paddlepaddle/paddle:latest COPY start.sh /root/ COPY start_paddle.py /root/ diff --git a/doc/howto/usage/k8s/src/k8s_train/README.md b/doc/howto/usage/cluster/src/k8s_train/README.md similarity index 100% rename from doc/howto/usage/k8s/src/k8s_train/README.md rename to doc/howto/usage/cluster/src/k8s_train/README.md diff --git a/doc/howto/usage/k8s/src/k8s_train/start.sh b/doc/howto/usage/cluster/src/k8s_train/start.sh similarity index 100% rename from doc/howto/usage/k8s/src/k8s_train/start.sh rename to doc/howto/usage/cluster/src/k8s_train/start.sh diff --git a/doc/howto/usage/k8s/src/k8s_train/start_paddle.py b/doc/howto/usage/cluster/src/k8s_train/start_paddle.py similarity index 100% rename from doc/howto/usage/k8s/src/k8s_train/start_paddle.py rename to doc/howto/usage/cluster/src/k8s_train/start_paddle.py diff --git a/doc/howto/usage/k8s/src/managed_policy.png b/doc/howto/usage/cluster/src/managed_policy.png similarity index 100% rename from doc/howto/usage/k8s/src/managed_policy.png rename to doc/howto/usage/cluster/src/managed_policy.png diff --git a/doc/howto/usage/k8s/src/pserver_and_trainer.png b/doc/howto/usage/cluster/src/pserver_and_trainer.png similarity index 100% rename from doc/howto/usage/k8s/src/pserver_and_trainer.png rename to doc/howto/usage/cluster/src/pserver_and_trainer.png diff --git a/doc/howto/usage/k8s/src/route53_create_recordset.png b/doc/howto/usage/cluster/src/route53_create_recordset.png similarity index 100% rename from doc/howto/usage/k8s/src/route53_create_recordset.png rename to doc/howto/usage/cluster/src/route53_create_recordset.png diff --git a/doc/howto/usage/k8s/src/route53_create_zone.png b/doc/howto/usage/cluster/src/route53_create_zone.png similarity index 100% rename from doc/howto/usage/k8s/src/route53_create_zone.png rename to doc/howto/usage/cluster/src/route53_create_zone.png diff --git a/doc/howto/usage/k8s/src/worker_security_group.png b/doc/howto/usage/cluster/src/worker_security_group.png similarity index 100% rename from doc/howto/usage/k8s/src/worker_security_group.png rename to doc/howto/usage/cluster/src/worker_security_group.png diff --git a/doc/howto/usage/k8s/k8s_basis_cn.md b/doc/howto/usage/k8s/k8s_basis_cn.md deleted file mode 100644 index 4c3dc81ed38f239c1f4a83d22b49cf57b5d16a8b..0000000000000000000000000000000000000000 --- a/doc/howto/usage/k8s/k8s_basis_cn.md +++ /dev/null @@ -1,75 +0,0 @@ -# Kubernetes 简介 - -[*Kubernetes*](http://kubernetes.io/)是Google开源的容器集群管理系统，其提供应用部署、维护、扩展机制等功能，利用Kubernetes能方便地管理跨机器运行容器化的应用。Kubernetes可以在物理机或虚拟机上运行，且支持部署到[AWS](http://kubernetes.io/docs/getting-started-guides/aws)，[Azure](http://kubernetes.io/docs/getting-started-guides/azure/)，[GCE](http://kubernetes.io/docs/getting-started-guides/gce)等多种公有云环境。介绍分布式训练之前，需要对[Kubernetes](http://kubernetes.io/)有一个基本的认识，下面先简要介绍一下本文用到的几个Kubernetes概念。 - -- [*Node*](http://kubernetes.io/docs/admin/node/) 表示一个Kubernetes集群中的一个工作节点，这个节点可以是物理机或者虚拟机，Kubernetes集群就是由node节点与master节点组成的。 - -- [*Pod*](http://kubernetes.io/docs/user-guide/pods/) 是一组(一个或多个)容器，pod是Kubernetes的最小调度单元，一个pod中的所有容器会被调度到同一个node上。Pod中的容器共享NET，PID，IPC，UTS等Linux namespace。由于容器之间共享NET namespace，所以它们使用同一个IP地址，可以通过*localhost*互相通信。不同pod之间可以通过IP地址访问。 - -- [*Job*](http://kubernetes.io/docs/user-guide/jobs/) 描述Kubernetes上运行的作业，一次作业称为一个job，通常每个job包括一个或者多个pods，job启动后会创建这些pod并开始执行一个程序，等待这个程序执行成功并返回0则成功退出，如果执行失败，也可以配置不同的重试机制。 - -- [*Volume*](http://kubernetes.io/docs/user-guide/volumes/) 存储卷，是pod内的容器都可以访问的共享目录，也是容器与node之间共享文件的方式，因为容器内的文件都是暂时存在的，当容器因为各种原因被销毁时，其内部的文件也会随之消失。通过volume，就可以将这些文件持久化存储。Kubernetes支持多种volume，例如hostPath(宿主机目录)，gcePersistentDisk，awsElasticBlockStore等。 - -- [*Namespaces*](https://kubernetes.io/docs/user-guide/namespaces/) 命名空间，在kubernetes中创建的所有资源对象(例如上文的pod，job)等都属于一个命名空间，在同一个命名空间中，资源对象的名字是唯一的，不同空间的资源名可以重复，命名空间主要为了对象进行逻辑上的分组便于管理。本文只使用了默认命名空间。 - -- [*PersistentVolume*](https://kubernetes.io/docs/user-guide/persistent-volumes/): 和[*PersistentVolumeClaim*](https://kubernetes.io/docs/user-guide/persistent-volumes/#persistentvolumeclaims)结合，将外部的存储服务在Kubernetes中描述成为统一的资源形式，便于存储资源管理和Pod引用。 - -## 部署Kubernetes集群 - -Kubernetes提供了多种集群部署的方案，本文档内不重复介绍。这里给出集中常见的部署方法： - -- [*minikube*](https://kubernetes.io/docs/getting-started-guides/minikube/): 快速在本地启动一个单机的kubernetes服务器，便于本地验证和测试。 -- [*kubeadm*](http://kubernetes.io/docs/getting-started-guides/kubeadm/): 在不同操作系统，不同主机(Bare-Metal, AWS, GCE)条件下，快速部署集群。 -- [*AWS EC2*](https://kubernetes.io/docs/getting-started-guides/aws/): 在aws上快速部署集群。 -- [*Bare-Metal*](https://kubernetes.io/docs/getting-started-guides/centos/centos_manual_config/): 在物理机上手动部署。 - -可以参考[这个表格](https://kubernetes.io/docs/getting-started-guides/#table-of-solutions)选择适合您的场景的合适方案。 - -## 选择存储方案 - -容器不会保留在运行时生成的数据，job或者应用程序在容器中运行时生成的数据会在容器销毁时消失。为了完成分布式机器学习训练任务，需要有一个外部的存储服务来保存训练所需数据和训练输出。 -常见的可选存储服务包括： - -- [*NFS*](https://github.com/kubernetes/kubernetes/tree/master/examples/volumes/nfs): 可以将磁盘上某个目录共享给网络中其他机器访问。部署和配置比较简单，可以用于小量数据的验证。不提供分布式存储，高可用，冗余等功能。NFS的部署方法可以参考[这里](http://www.tecmint.com/how-to-setup-nfs-server-in-linux/)。 -- [*GlusterFS*](http://gluster.readthedocs.io/en/latest/Quick-Start-Guide/Quickstart/): 网络分布式文件系统，可以在Kubernetes中按照[这个](https://github.com/kubernetes/kubernetes/tree/master/examples/volumes/glusterfs)例子使用。 -- [*Ceph*](http://docs.ceph.com/docs/master/): 分布式文件系统，支持rbd，POSIX API接口(ceph fs)和对象存储API，参考[这里](https://kubernetes.io/docs/user-guide/volumes/#rbd)。 -- [*MooseFS*](https://moosefs.com/documentation.html): 一个分布式的存储系统。需要先挂载到服务器Node上再通过kubernetes hostPath Volume挂载到容器中。 - -## 配置kubectl - -### 安装kubectl -``` -# OS X -curl -LO https://storage.googleapis.com/kubernetes-release/release/$(curl -s https://storage.googleapis.com/kubernetes-release/release/stable.txt)/bin/darwin/amd64/kubectl - -# Linux -curl -LO https://storage.googleapis.com/kubernetes-release/release/$(curl -s https://storage.googleapis.com/kubernetes-release/release/stable.txt)/bin/linux/amd64/kubectl - -# Windows -curl -LO https://storage.googleapis.com/kubernetes-release/release/$(curl -s https://storage.googleapis.com/kubernetes-release/release/stable.txt)/bin/windows/amd64/kubectl.exe -``` - -### 配置kubectl访问你的kubernetes集群 - -编辑`~/.kube/config`这个配置文件，修改`Master-IP`的地址。如果使用SSL认证，则需要配置`certificate-authority`和`users`中的用户证书。如果是使用非SSL方式访问（比如通过8080端口），也可以去掉这些证书的配置。 -``` -apiVersion: v1 -clusters: -- cluster: - certificate-authority: /path/to/ca.crt - server: https://[Master-IP]:443 - name: minikube -contexts: -- context: - cluster: minikube - user: minikube - name: minikube -current-context: minikube -kind: Config -preferences: {} -users: -- name: minikube - user: - client-certificate: /path/to/apiserver.crt - client-key: /Users/wuyi/.minikube/apiserver.key -``` diff --git a/doc/howto/usage/k8s/src/k8s-paddle-arch.png b/doc/howto/usage/k8s/src/k8s-paddle-arch.png deleted file mode 100644 index 2183a232ad402b76f82a67234a5c93e13ce97ac3..0000000000000000000000000000000000000000 Binary files a/doc/howto/usage/k8s/src/k8s-paddle-arch.png and /dev/null differ diff --git a/doc/mobile/cross_compiling_for_ios_cn.md b/doc/mobile/cross_compiling_for_ios_cn.md index 9da48e7f2119ce901fbb3abab73400df27be16d2..d5196d9a4c93c7692d2a624ec7d0650e32806338 100644 --- a/doc/mobile/cross_compiling_for_ios_cn.md +++ b/doc/mobile/cross_compiling_for_ios_cn.md @@ -18,11 +18,11 @@ PaddlePaddle为交叉编译提供了工具链配置文档[cmake/cross_compiling/ - `CMAKE_SYSTEM_NAME`，CMake编译的目标平台，必须设置为`iOS`。在设置`CMAKE_SYSTEM_NAME=iOS`后，PaddlePaddle的CMake系统会自动编译所有的第三方依赖库，并且强制设置一些PaddlePaddle参数的值（`WITH_C_API=ON`、`WITH_GPU=OFF`、`WITH_AVX=OFF`、`WITH_PYTHON=OFF`、`WITH_RDMA=OFF`）。 - `WITH_C_API`，是否编译C-API预测库，必须设置为ON。在iOS平台上只支持使用C-API来预测。 -- `WITH_SWIG_PY`，必须设置为ON。在iOS平台上不支持通过swig调用来训练或者预测。 +- `WITH_SWIG_PY`，必须设置为`OFF`。在iOS平台上不支持通过swig调用来训练或者预测。 iOS平台可选配置参数： -- `IOS_PLATFORM`，可设置为`OS/SIMULATOR`，默认值为`OS`。 +- `IOS_PLATFORM`，可设置为`OS`（默认值）或`SIMULATOR`。 - `OS`，构建目标为`arm`架构的iPhone或者iPad等物理设备。 - `SIMULATOR`，构建目标为`x86`架构的模拟器平台。 - `IOS_ARCH`，目标架构。针对不同的`IOS_PLATFORM`，可设置的目标架构如下表所示，默认编译所有架构： diff --git a/doc/mobile/cross_compiling_for_ios_en.md b/doc/mobile/cross_compiling_for_ios_en.md new file mode 100644 index 0000000000000000000000000000000000000000..aa390cd61f3fbd75e5a3b342f3559e76da35a918 --- /dev/null +++ b/doc/mobile/cross_compiling_for_ios_en.md @@ -0,0 +1,120 @@ +# PaddlePaddle Compiling Guide for iOS + +This tutorial will walk you through cross compiling the PaddlePaddle library for iOS from the source in MacOS. + +## Preparation + +Apple provides Xcode for cross-compiling and IDE for iOS development. Download from App store or [here](https://developer.apple.com/cn/xcode/). To verify your installation, run command as follows + +```bash +$ xcodebuild -version +Xcode 9.0 +Build version 9A235 +``` + +## Cross-compiling configurations + +PaddlePaddle provides cross-compiling toolchain configuration documentation [cmake/cross_compiling/ios.cmake](https://github.com/PaddlePaddle/Paddle/blob/develop/cmake/cross_compiling/ios.cmake), which has some default settings for frequently used compilers. + +There are some mandatory environment variables need to be set before cross compiling PaddlePaddle for iOS: + +- `CMAKE_SYSTEM_NAME`, CMake compiling target platform name, has to be `iOS`. PaddlePaddle CMake will compile all the third party dependencies and enforce some parameters (`WITH_C_API=ON`, `WITH_GPU=OFF`, `WITH_AVX=OFF`, `WITH_PYTHON=OFF`,`WITH_RDMA=OFF`) when this variable is set with value `iOS`. + +- `WITH_C_API`, Whether to compile inference C-API library, has to be `ON`, since C-API is the only supported interface for inferencing in iOS. +- `WITH_SWIG_PY`, has to be `OFF`. It's not supported to inference or train via swig in iOS. + +Optional environment variables for iOS are: + +- `IOS_PLATFORM`, either `OS` (default) or `SIMULATOR`. + - `OS`, build targets ARM-based physical devices like iPhone or iPad. + - `SIMULATOR`, build targets x86 architecture simulators. +- `IOS_ARCH`, target architecture. By default, all architecture types will be compiled. If you need to specify the architecture to compile for, please find valid values for different `IOS_PLATFORM` settings from the table below: + + + + + + + + + + + + + + + + + + + + + + +
IOS_PLATFORMIOS_ARCH
OSarmv7, armv7s, arm64
SIMULATORi386, x86_64
+ +- `IOS_DEPLOYMENT_TARGET`, minimum iOS version to deployment, `7.0` by default. +- `IOS_ENABLE_BITCODE`, whether to enable [Bitcode](https://developer.apple.com/library/content/documentation/IDEs/Conceptual/AppDistributionGuide/AppThinning/AppThinning.html#//apple_ref/doc/uid/TP40012582-CH35-SW3), values can be `ON/OFF`, `ON` by default. +- `IOS_USE_VECLIB_FOR_BLAS`, whether to use [vecLib](https://developer.apple.com/documentation/accelerate/veclib) framework for BLAS computing. values can be `ON/OFF`, `OFF` by default. +- `IOS_DEVELOPMENT_ROOT`, the path to `Developer` directory, can be explicitly set with your `/path/to/platform/Developer`. If left blank, PaddlePaddle will automatically pick the Xcode corresponding `platform`'s `Developer` directory based on your `IOS_PLATFORM` value. +- `IOS_SDK_ROOT`, the path to `SDK` root, can be explicitly set with your `/path/to/platform/Developer/SDKs/SDK`. if left black, PaddlePaddle will pick the latest SDK in the directory of `IOS_DEVELOPMENT_ROOT`. + +other settings： + +- `USE_EIGEN_FOR_BLAS`, whether to use Eigen for matrix computing. effective when `IOS_USE_VECLIB_FOR_BLAS=OFF`. Values can be `ON/OFF`, `OFF` by default. +- `HOST_C/CXX_COMPILER`, host C/C++ compiler. Uses value from environment variable `CC/CXX` by default or `cc/c++` if `CC/CXX` doesn't exist. + +some typical cmake configurations: + +```bash +cmake -DCMAKE_SYSTEM_NAME=iOS \ + -DIOS_PLATFORM=OS \ + -DIOS_ARCH="armv7;arm64" \ + -DIOS_ENABLE_BITCODE=ON \ + -DIOS_USE_VECLIB_FOR_BLAS=ON \ + -DCMAKE_INSTALL_PREFIX=your/path/to/install \ + -DWITH_C_API=ON \ + -DWITH_TESTING=OFF \ + -DWITH_SWIG_PY=OFF \ + .. +``` + +```bash +cmake -DCMAKE_SYSTEM_NAME=iOS \ + -DIOS_PLATFORM=SIMULATOR \ + -DIOS_ARCH="x86_64" \ + -DIOS_USE_VECLIB_FOR_BLAS=ON \ + -DCMAKE_INSTALL_PREFIX=your/path/to/install \ + -DWITH_C_API=ON \ + -DWITH_TESTING=OFF \ + -DWITH_SWIG_PY=OFF \ + .. +``` + +You can set other compiling parameters for your own need. I.E. if you are trying to minimize the library size, set `CMAKE_BUILD_TYPE` with `MinSizeRel`; or if the performance is your concern, set `CMAKE_BUILD_TYPE` with `Release`. You can even manipulate the PaddlePaddle compiling procedure by manually set `CMAKE_C/CXX_FLAGS` values. + +**TIPS for a better performance**: + +- set `CMAKE_BUILD_TYPE` with `Release` +- set `IOS_USE_VECLIB_FOR_BLAS` with `ON` + +## Compile and install + +After CMake, run following commands, PaddlePaddle will download the compile 3rd party dependencies, compile and install PaddlePaddle inference library. + +``` +$ make +$ make install +``` + +Please Note: if you compiled PaddlePaddle in the source directory for other platforms, do remove `third_party` and `build` directory within the source with `rm -rf` to ensure that all the 3rd party libraries dependencies and PaddlePaddle is newly compiled with current CMake configuration. + +`your/path/to/install` directory will have following directories after `compile` and `install`: + +- `include`, contains all the C-API header files. +- `lib`, contains PaddlePaddle C-API static library. +- `third_party` contains all the 3rd party libraries. + +Please note: if PaddlePaddle library need to support both physical devices and simulators, you will need to compile correspondingly, then merge fat library with `lipo`. + +Now you will have PaddlePaddle library compiled and installed, the fat library can be used in deep learning related iOS APPs. Please refer to C-API documentation for usage guides. diff --git a/doc/mobile/index_en.rst b/doc/mobile/index_en.rst index 3c08d736717cfe8d5fdf449dc58015086befbe60..ef421dacad458828cadf8cf505375d6c4bfd9dde 100644 --- a/doc/mobile/index_en.rst +++ b/doc/mobile/index_en.rst @@ -5,4 +5,5 @@ MOBILE :maxdepth: 1 cross_compiling_for_android_en.md + cross_compiling_for_ios_en.md cross_compiling_for_raspberry_en.md diff --git a/go/pserver/client/c/test/test_cclient.c b/go/pserver/client/c/test/test_cclient.c index 89c4d7f00aae2a92ae30ba7b4305550d150dd985..05ec421fff6e1c57b0bace080668d3793f85480f 100644 --- a/go/pserver/client/c/test/test_cclient.c +++ b/go/pserver/client/c/test/test_cclient.c @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 diff --git a/paddle/api/CMakeLists.txt b/paddle/api/CMakeLists.txt index d6b8464100d4497876aa3f6f7cbc666aafae4bfc..cf84568ecdf1227b0d0ed3606a4a9a6e5186af72 100644 --- a/paddle/api/CMakeLists.txt +++ b/paddle/api/CMakeLists.txt @@ -25,8 +25,18 @@ FILE(GLOB PY_PADDLE_PYTHON_FILES ${PADDLE_SOURCE_DIR}/paddle/py_paddle/*.py) SET_SOURCE_FILES_PROPERTIES(Paddle.i PROPERTIES CPLUSPLUS ON) +SET(SWIG_NEED_FLAGS + -ftls-model=global-dynamic + -Wno-parentheses-equality + -Wno-self-assign + -Wno-maybe-uninitialized + -Wno-missing-field-initializers) + FOREACH(flag ${SWIG_NEED_FLAGS}) + safe_set_cxxflag(SWIG_CXX_FLAGS ${flag}) +ENDFOREACH() + SET(CMAKE_SWIG_OUTDIR ${CMAKE_CURRENT_BINARY_DIR}) -SET(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-parentheses-equality -Wno-missing-field-initializers -Wno-self-assign -ftls-model=global-dynamic") +SET(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${SWIG_CXX_FLAGS}") SET(SWIG_MODULE_swig_paddle_EXTRA_DEPS paddle_parameter diff --git a/paddle/capi/CMakeLists.txt b/paddle/capi/CMakeLists.txt index d267b14657be2a773d1dacfd9ac3767cddc47415..ebb083c5a477d5be91ef14be74dd9de349d07931 100644 --- a/paddle/capi/CMakeLists.txt +++ b/paddle/capi/CMakeLists.txt @@ -4,6 +4,16 @@ else () set(PADDLE_FLOAT_TYPE float) endif() +execute_process( + COMMAND ${GIT_EXECUTABLE} log --pretty=format:%H -1 + WORKING_DIRECTORY ${PADDLE_SOURCE_DIR} + OUTPUT_VARIABLE PADDLE_GIT_COMMIT + RESULT_VARIABLE PADDLE_GIT_COMMIT_RESULT + ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE) +if(NOT PADDLE_GIT_COMMIT) + set(PADDLE_GIT_COMMIT "no commit information") +endif() + # config.h used for C-API. It will store Paddle building configuration as a # header. Make user just include PaddleCAPI.h then can get building # configuration without explicitly set -DPADDLE_WITH_DOUBLE when building their diff --git a/paddle/capi/Main.cpp b/paddle/capi/Main.cpp index bb8249a5511c089ec2f2263ff4cc290f0a5a8fce..c038789340033fcf6dcc07a41b033a50e980c965 100644 --- a/paddle/capi/Main.cpp +++ b/paddle/capi/Main.cpp @@ -43,4 +43,11 @@ paddle_error paddle_init(int argc, char** argv) { isInit = true; return kPD_NO_ERROR; } + +paddle_error paddle_init_thread() { + if (FLAGS_use_gpu) { + hl_init(FLAGS_gpu_id); + } + return kPD_NO_ERROR; +} } diff --git a/paddle/capi/Matrix.cpp b/paddle/capi/Matrix.cpp index d5b55e1c95f248f551e6a0a3b39123169dd7784f..cbacd1fb71c14f490ff548db714e728772292b4b 100644 --- a/paddle/capi/Matrix.cpp +++ b/paddle/capi/Matrix.cpp @@ -40,7 +40,7 @@ paddle_error paddle_matrix_destroy(paddle_matrix mat) { paddle_error paddle_matrix_set_row(paddle_matrix mat, uint64_t rowID, paddle_real* rowArray) { - if (mat == nullptr) return kPD_NULLPTR; + if (mat == nullptr || rowArray == nullptr) return kPD_NULLPTR; auto ptr = cast(mat); if (ptr->mat == nullptr) return kPD_NULLPTR; if (rowID >= ptr->mat->getHeight()) return kPD_OUT_OF_RANGE; @@ -55,7 +55,7 @@ paddle_error paddle_matrix_set_row(paddle_matrix mat, } PD_API paddle_error paddle_matrix_set_value(paddle_matrix mat, - paddle_real* value) { + paddle_real* value) { if (mat == nullptr || value == nullptr) return kPD_NULLPTR; auto ptr = cast(mat); if (ptr->mat == nullptr) return kPD_NULLPTR; @@ -75,7 +75,7 @@ PD_API paddle_error paddle_matrix_set_value(paddle_matrix mat, } PD_API paddle_error paddle_matrix_get_value(paddle_matrix mat, - paddle_real* result) { + paddle_real* result) { if (mat == nullptr || result == nullptr) return kPD_NULLPTR; auto ptr = cast(mat); if (ptr->mat == nullptr) return kPD_NULLPTR; diff --git a/paddle/capi/config.h.in b/paddle/capi/config.h.in index d205307588eb60b2e11accb9f825391f7c1453f2..0ddbd8c753c55ab95a89e1781c64b9416f7344e7 100644 --- a/paddle/capi/config.h.in +++ b/paddle/capi/config.h.in @@ -3,6 +3,9 @@ typedef @PADDLE_FLOAT_TYPE@ paddle_real; +#define __PADDLE_VERSION__ "@PADDLE_VERSION@" +#define __PADDLE_COMMIT__ "@PADDLE_GIT_COMMIT@" + // Since we only support linux and macos in compile, always use clang or // gcc 4.8+. DLL_IMPORT/DLL_EXPORT is as simple as below. #define PD_API __attribute__((visibility("default"))) diff --git a/paddle/capi/error.cpp b/paddle/capi/error.cpp new file mode 100644 index 0000000000000000000000000000000000000000..96ce31b45fc3f83237146443cbe4289af7bfa239 --- /dev/null +++ b/paddle/capi/error.cpp @@ -0,0 +1,32 @@ +/* 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. */ + +#include "error.h" + +extern "C" const char* paddle_error_string(paddle_error err) { + switch (err) { + case kPD_NULLPTR: + return "nullptr error"; + case kPD_OUT_OF_RANGE: + return "out of range error"; + case kPD_PROTOBUF_ERROR: + return "protobuf error"; + case kPD_NOT_SUPPORTED: + return "not supported error"; + case kPD_UNDEFINED_ERROR: + return "undefined error"; + default: + return ""; + } +} diff --git a/paddle/capi/error.h b/paddle/capi/error.h index 44d8c2040d1aad698398089baeee6f13c3deeb55..2da9e0a3ef604fbcd53bac271c72ef33b3105152 100644 --- a/paddle/capi/error.h +++ b/paddle/capi/error.h @@ -15,6 +15,8 @@ limitations under the License. */ #ifndef __PADDLE_CAPI_ERROR_H__ #define __PADDLE_CAPI_ERROR_H__ +#include "config.h" + /** * Error Type for Paddle API. */ @@ -27,4 +29,17 @@ typedef enum { kPD_UNDEFINED_ERROR = -1, } paddle_error; +#ifdef __cplusplus +extern "C" { +#endif + +/** + * Error string for Paddle API. + */ +PD_API const char* paddle_error_string(paddle_error err); + +#ifdef __cplusplus +} +#endif + #endif diff --git a/paddle/capi/examples/model_inference/dense/main.c b/paddle/capi/examples/model_inference/dense/main.c index 876af2aa7615c098d225b56ce2ea0b1529a6e3c6..5eeaf7e31fac7c9ed0b9269e74a7e467bde155ef 100644 --- a/paddle/capi/examples/model_inference/dense/main.c +++ b/paddle/capi/examples/model_inference/dense/main.c @@ -1,5 +1,6 @@ #include #include + #include "../common/common.h" #define CONFIG_BIN "./trainer_config.bin" @@ -27,20 +28,19 @@ int main() { CHECK(paddle_arguments_resize(in_args, 1)); // Create input matrix. - paddle_matrix mat = paddle_matrix_create(/* sample_num */ 10, + paddle_matrix mat = paddle_matrix_create(/* sample_num */ 1, /* size */ 784, /* useGPU */ false); srand(time(0)); - std::vector input; - input.resize(784 * 10); + paddle_real* array; + + // Get First row. + CHECK(paddle_matrix_get_row(mat, 0, &array)); - for (int i = 0; i < input.size(); ++i) { - input[i] = rand() / ((float)RAND_MAX); + for (int i = 0; i < 784; ++i) { + array[i] = rand() / ((float)RAND_MAX); } - - // Set value for the input matrix - CHECK(paddle_matrix_set_value(mat, input.data())); CHECK(paddle_arguments_set_value(in_args, 0, mat)); @@ -53,17 +53,18 @@ int main() { CHECK(paddle_arguments_get_value(out_args, 0, prob)); - std::std::vector result; - int height; - int width; + uint64_t height; + uint64_t width; - CHECK(paddle_matrix_get_shape(prob, &height, &width); - result.resize(height * width); - CHECK(paddle_matrix_get_value(prob, result.data())); + CHECK(paddle_matrix_get_shape(prob, &height, &width)); + CHECK(paddle_matrix_get_row(prob, 0, &array)); - printf("Prob: "); + printf("Prob: \n"); for (int i = 0; i < height * width; ++i) { - printf("%.2f ", result[i]); + printf("%.4f ", array[i]); + if ((i + 1) % width == 0) { + printf("\n"); + } } printf("\n"); diff --git a/paddle/capi/examples/model_inference/multi_thread/CMakeLists.txt b/paddle/capi/examples/model_inference/multi_thread/CMakeLists.txt index 98e411ddc02a46034e8f6ceb00657622d998c9f3..2fc8debddedeab6ae982b0df49ec2b73bc0f85f5 100644 --- a/paddle/capi/examples/model_inference/multi_thread/CMakeLists.txt +++ b/paddle/capi/examples/model_inference/multi_thread/CMakeLists.txt @@ -1,8 +1,29 @@ project(multi_thread) cmake_minimum_required(VERSION 2.8) -aux_source_directory(. SRC_LIST) -add_executable(${PROJECT_NAME} ${SRC_LIST}) + find_package (Threads) + +if(NOT PADDLE_ROOT) + set(PADDLE_ROOT $ENV{PADDLE_ROOT} CACHE PATH "Paddle Path") +endif() +if(PADDLE_ROOT) + include_directories(${PADDLE_ROOT}/include) + link_directories(${PADDLE_ROOT}/lib) +endif() + +set(CPU_SRCS main.c) +add_executable(${PROJECT_NAME} ${CPU_SRCS}) set_property(TARGET ${PROJECT_NAME} PROPERTY C_STANDARD 99) -target_link_libraries(${PROJECT_NAME} -lpaddle_capi_shared - ${CMAKE_THREAD_LIBS_INIT}) +target_link_libraries(${PROJECT_NAME} + -lpaddle_capi_shared + ${CMAKE_THREAD_LIBS_INIT}) + +find_package(CUDA QUIET) +if(CUDA_FOUND) + set(GPU_SRCS main_gpu.c) + cuda_add_executable(${PROJECT_NAME}_gpu ${GPU_SRCS}) + set_property(TARGET ${PROJECT_NAME}_gpu PROPERTY C_STANDARD 99) + target_link_libraries(${PROJECT_NAME}_gpu + -lpaddle_capi_shared + ${CMAKE_THREAD_LIBS_INIT}) +endif(CUDA_FOUND) diff --git a/paddle/capi/examples/model_inference/multi_thread/main_gpu.c b/paddle/capi/examples/model_inference/multi_thread/main_gpu.c new file mode 100644 index 0000000000000000000000000000000000000000..6fd376e0d1a2fee4f9a0f676b53c6f2891795cab --- /dev/null +++ b/paddle/capi/examples/model_inference/multi_thread/main_gpu.c @@ -0,0 +1,113 @@ +#include +#include +#include +#include "../common/common.h" + +#define CONFIG_BIN "./trainer_config.bin" +#define NUM_THREAD 4 +#define NUM_ITER 1000 + +pthread_mutex_t mutex; + +/* + * @brief It is an simple inference example that runs multi-threads on a GPU. + * Each thread holds it own local gradient_machine but shares the same + * parameters. + * If you want to run on different GPUs, you need to launch + * multi-processes or set trainer_count > 1. + */ +void* thread_main(void* gm_ptr) { + // Initialize the thread environment of Paddle. + CHECK(paddle_init_thread()); + + paddle_gradient_machine machine = (paddle_gradient_machine)(gm_ptr); + // Create input arguments. + paddle_arguments in_args = paddle_arguments_create_none(); + // Create input matrix. + paddle_matrix mat = paddle_matrix_create(/* sample_num */ 1, + /* size */ 784, + /* useGPU */ true); + // Create output arguments. + paddle_arguments out_args = paddle_arguments_create_none(); + // Create output matrix. + paddle_matrix prob = paddle_matrix_create_none(); + + // CPU buffer to cache the input and output. + paddle_real* cpu_input = (paddle_real*)malloc(784 * sizeof(paddle_real)); + paddle_real* cpu_output = (paddle_real*)malloc(10 * sizeof(paddle_real)); + for (int iter = 0; iter < NUM_ITER; ++iter) { + // There is only one input layer of this network. + CHECK(paddle_arguments_resize(in_args, 1)); + CHECK(paddle_arguments_set_value(in_args, 0, mat)); + + for (int i = 0; i < 784; ++i) { + cpu_input[i] = rand() / ((float)RAND_MAX); + } + CHECK(paddle_matrix_set_value(mat, cpu_input)); + + CHECK(paddle_gradient_machine_forward(machine, + in_args, + out_args, + /* isTrain */ false)); + + CHECK(paddle_arguments_get_value(out_args, 0, prob)); + CHECK(paddle_matrix_get_value(prob, cpu_output)); + + pthread_mutex_lock(&mutex); + printf("Prob: "); + for (int i = 0; i < 10; ++i) { + printf("%.2f ", cpu_output[i]); + } + printf("\n"); + pthread_mutex_unlock(&mutex); + } + + CHECK(paddle_matrix_destroy(prob)); + CHECK(paddle_arguments_destroy(out_args)); + CHECK(paddle_matrix_destroy(mat)); + CHECK(paddle_arguments_destroy(in_args)); + CHECK(paddle_gradient_machine_destroy(machine)); + + free(cpu_input); + free(cpu_output); + + return NULL; +} + +int main() { + // Initalize Paddle + char* argv[] = {"--use_gpu=True"}; + CHECK(paddle_init(1, (char**)argv)); + + // Reading config binary file. It is generated by `convert_protobin.sh` + long size; + void* buf = read_config(CONFIG_BIN, &size); + + // Create a gradient machine for inference. + paddle_gradient_machine machine; + CHECK(paddle_gradient_machine_create_for_inference(&machine, buf, (int)size)); + CHECK(paddle_gradient_machine_randomize_param(machine)); + + // Loading parameter. Uncomment the following line and change the directory. + // CHECK(paddle_gradient_machine_load_parameter_from_disk(machine, + // "./some_where_to_params")); + srand(time(0)); + pthread_mutex_init(&mutex, NULL); + + pthread_t threads[NUM_THREAD]; + + for (int i = 0; i < NUM_THREAD; ++i) { + paddle_gradient_machine thread_local_machine; + CHECK(paddle_gradient_machine_create_shared_param( + machine, buf, size, &thread_local_machine)); + pthread_create(&threads[i], NULL, thread_main, thread_local_machine); + } + + for (int i = 0; i < NUM_THREAD; ++i) { + pthread_join(threads[i], NULL); + } + + pthread_mutex_destroy(&mutex); + + return 0; +} diff --git a/paddle/capi/main.h b/paddle/capi/main.h index 893ebcbd58dd24cf835fb2005865c94c9ba2a810..99c4e8428dbaa14d36dc2d36b2a4f16c9ec3e0d1 100644 --- a/paddle/capi/main.h +++ b/paddle/capi/main.h @@ -26,6 +26,13 @@ extern "C" { */ PD_API paddle_error paddle_init(int argc, char** argv); +/** + * Initialize the thread environment of Paddle. + * @note it is requisite for GPU runs but optional for CPU runs. + * For GPU runs, all threads will run on the same GPU devices. + */ +PD_API paddle_error paddle_init_thread(); + #ifdef __cplusplus } #endif diff --git a/paddle/capi/matrix.h b/paddle/capi/matrix.h index 01b8bad2ee9f528f8622346f43b9ff82225a7e73..8cc3e0034e058daefc63c69efe0b1f575c586897 100644 --- a/paddle/capi/matrix.h +++ b/paddle/capi/matrix.h @@ -79,7 +79,7 @@ PD_API paddle_error paddle_matrix_set_row(paddle_matrix mat, * @note value should contain enough element of data to init the mat */ PD_API paddle_error paddle_matrix_set_value(paddle_matrix mat, - paddle_real* value); + paddle_real* value); /** * @brief PDMatGetRow Get raw row buffer from matrix @@ -93,14 +93,14 @@ PD_API paddle_error paddle_matrix_get_row(paddle_matrix mat, paddle_real** rawRowBuffer); /** - * @brief copy data from the matrix + * @brief copy data from the matrix * @param [in] mat Target matrix - * @param [out] result pointer to store the matrix data + * @param [out] result pointer to store the matrix data * @return paddle_error * @note the space of the result should allocated before invoke this API */ PD_API paddle_error paddle_matrix_get_value(paddle_matrix mat, - paddle_real* result); + paddle_real* result); /** * @brief PDMatCreateNone Create None Matrix * @return diff --git a/paddle/cuda/include/hl_cnn.h b/paddle/cuda/include/hl_cnn.h index 89c1f48edacbe0a4432957fe066481412db7e6e1..88418062927cd0f7714e992cc2495109da45d32f 100644 --- a/paddle/cuda/include/hl_cnn.h +++ b/paddle/cuda/include/hl_cnn.h @@ -116,6 +116,7 @@ extern void hl_maxpool_backward(const int frameCnt, * @param[in] paddingW padding width. * @param[out] tgtData output data. * @param[in] tgtStride stride between output data samples. + * @param[in] excludeMode whether to consider paddings for size. * */ extern void hl_avgpool_forward(const int frameCnt, @@ -132,7 +133,8 @@ extern void hl_avgpool_forward(const int frameCnt, const int paddingH, const int paddingW, real* tgtData, - const int tgtStride); + const int tgtStride, + bool excludeMode); /** * @brief Maximum pool backward. @@ -154,6 +156,7 @@ extern void hl_avgpool_forward(const int frameCnt, * @param[in] scaleB scale. * @param[out] backGrad output grad. * @param[in] outStride stride between output data samples. + * @param[in] excludeMode whether to consider paddings for size. * */ extern void hl_avgpool_backward(const int frameCnt, @@ -172,7 +175,8 @@ extern void hl_avgpool_backward(const int frameCnt, real scaleA, real scaleB, real* backGrad, - const int outStride); + const int outStride, + bool excludeMode); extern void hl_maxpool3D_forward(const int frameCnt, const real* inputData, diff --git a/paddle/cuda/include/stub/hl_cnn_stub.h b/paddle/cuda/include/stub/hl_cnn_stub.h index 968ed4840ffb0623b57bd6e6d839973e109394de..706cc59a8e394b109d2b290425f4b5f51d987f28 100644 --- a/paddle/cuda/include/stub/hl_cnn_stub.h +++ b/paddle/cuda/include/stub/hl_cnn_stub.h @@ -68,7 +68,8 @@ inline void hl_avgpool_forward(const int frameCnt, const int paddingH, const int paddingW, real* tgtData, - const int tgtStride) {} + const int tgtStride, + const bool excludeMode) {} inline void hl_avgpool_backward(const int frameCnt, const real* outGrad, @@ -86,7 +87,8 @@ inline void hl_avgpool_backward(const int frameCnt, real scaleA, real scaleB, real* backGrad, - const int outStride) {} + const int outStride, + const bool excludeMode) {} inline void hl_maxpool3D_forward(const int frameCnt, const real* inputData, diff --git a/paddle/cuda/src/hl_cuda_cnn.cu b/paddle/cuda/src/hl_cuda_cnn.cu index 3699b1e8ae9d8f813439eaeaa760c4a9f6e100a0..2d1bc4f6d55fac4b74f4e58d40fe56aa61d19cf9 100644 --- a/paddle/cuda/src/hl_cuda_cnn.cu +++ b/paddle/cuda/src/hl_cuda_cnn.cu @@ -210,7 +210,8 @@ __global__ void KeAvgPoolForward(const int nthreads, const int padH, const int padW, real* tgtData, - const int tgtStride) { + const int tgtStride, + const bool excludeMode) { int index = blockIdx.x * blockDim.x + threadIdx.x; if (index < nthreads) { int pw = index % pooledW; @@ -224,7 +225,8 @@ __global__ void KeAvgPoolForward(const int nthreads, int wend = min(wstart + sizeX, width); hstart = max(hstart, 0); wstart = max(wstart, 0); - int pool_size = (hend - hstart) * (wend - wstart); + int poolSize = + excludeMode ? (hend - hstart) * (wend - wstart) : sizeY * sizeX; real aveval = 0; inputData += (frameNum * channels + c) * height * width; @@ -235,7 +237,7 @@ __global__ void KeAvgPoolForward(const int nthreads, } int tgtIndex = index % (pooledW * pooledH * channels) + frameNum * tgtStride; - tgtData[tgtIndex] = aveval / pool_size; + tgtData[tgtIndex] = aveval / poolSize; } } @@ -253,7 +255,8 @@ void hl_avgpool_forward(const int frameCnt, const int paddingH, const int paddingW, real* tgtData, - const int tgtStride) { + const int tgtStride, + const bool excludeMode) { int num_kernels = pooledH * pooledW * channels * frameCnt; int blocks = (num_kernels + 1024 - 1) / 1024; KeAvgPoolForward<<>>(num_kernels, @@ -270,7 +273,8 @@ void hl_avgpool_forward(const int frameCnt, paddingH, paddingW, tgtData, - tgtStride); + tgtStride, + excludeMode); CHECK_SYNC("hl_avgpool_forward failed"); } @@ -290,7 +294,8 @@ __global__ void KeAvgPoolBackward(const int nthreads, real scaleA, real scaleB, real* tgtGrad, - const int outStride) { + const int outStride, + const bool excludeMode) { int index = blockIdx.x * blockDim.x + threadIdx.x; if (index < nthreads) { int offsetW = index % width + padW; @@ -314,8 +319,9 @@ __global__ void KeAvgPoolBackward(const int nthreads, int wstart = pw * strideW - padW; int wend = min(wstart + sizeX, width); wstart = max(wstart, 0); - int poolsize = (hend - hstart) * (wend - wstart); - gradient += outGrad[ph * pooledW + pw] / poolsize; + int poolSize = + excludeMode ? (hend - hstart) * (wend - wstart) : sizeY * sizeX; + gradient += outGrad[ph * pooledW + pw] / poolSize; } } tgtGrad[index] = scaleB * tgtGrad[index] + scaleA * gradient; @@ -338,7 +344,8 @@ void hl_avgpool_backward(const int frameCnt, real scaleA, real scaleB, real* backGrad, - const int outStride) { + const int outStride, + const bool excludeMode) { int num_kernels = height * width * channels * frameCnt; int blocks = (num_kernels + 1024 - 1) / 1024; @@ -358,7 +365,8 @@ void hl_avgpool_backward(const int frameCnt, scaleA, scaleB, backGrad, - outStride); + outStride, + excludeMode); CHECK_SYNC("hl_avgpool_backward failed"); } diff --git a/paddle/framework/CMakeLists.txt b/paddle/framework/CMakeLists.txt index c08e844847737b1172f6453767cc7f5e7b1a2bda..b4458eb9551724021636b628c5bf8c96f6e659aa 100644 --- a/paddle/framework/CMakeLists.txt +++ b/paddle/framework/CMakeLists.txt @@ -5,8 +5,19 @@ cc_library(ddim SRCS ddim.cc DEPS eigen3) cc_test(ddim_test SRCS ddim_test.cc DEPS ddim) nv_test(dim_test SRCS dim_test.cu DEPS ddim) -cc_library(tensor SRCS tensor.cc DEPS ddim place paddle_memory device_context) +if (WITH_GPU) + nv_library(tensor SRCS tensor.cc tensor_util.cu DEPS ddim place paddle_memory device_context framework_proto) +else() + cc_library(tensor SRCS tensor.cc tensor_util.cc DEPS ddim place paddle_memory device_context framework_proto) +endif () + cc_test(tensor_test SRCS tensor_test.cc DEPS tensor) +if (WITH_GPU) + nv_test(tensor_util_test SRCS tensor_util_test.cc tensor_util_test.cu DEPS tensor) +else() + cc_test(tensor_util_test SRCS tensor_util_test.cc DEPS tensor) +endif() + cc_test(eigen_test SRCS eigen_test.cc DEPS tensor) cc_library(lod_tensor SRCS lod_tensor.cc DEPS ddim place tensor framework_proto) @@ -18,6 +29,8 @@ cc_test(variable_test SRCS variable_test.cc) cc_library(scope SRCS scope.cc DEPS glog) cc_test(scope_test SRCS scope_test.cc DEPS scope) +cc_library(data_transform SRCS data_transform.cc DEPS math_function tensor framework_proto) +cc_test(data_transform_test SRCS data_transform_test.cc DEPS data_transform device_context) cc_library(attribute SRCS attribute.cc DEPS framework_proto) cc_test(program_desc_test SRCS program_desc_test.cc DEPS proto_desc @@ -26,12 +39,13 @@ cc_library(op_proto_maker SRCS op_proto_maker.cc DEPS framework_proto attribute) 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) -cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope glog shape_inference) -cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry) +cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope glog + shape_inference data_transform) +cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry init) cc_library(proto_desc SRCS var_desc.cc op_desc.cc block_desc.cc program_desc.cc DEPS shape_inference op_info operator glog) cc_library(op_registry SRCS op_registry.cc DEPS op_proto_maker op_info operator glog proto_desc) -cc_test(op_registry_test SRCS op_registry_test.cc DEPS op_registry) +nv_test(op_registry_test SRCS op_registry_test.cc DEPS op_registry) py_proto_compile(framework_py_proto SRCS framework.proto) # Generate an empty __init__.py to make framework_py_proto as a valid python module. @@ -51,11 +65,14 @@ cc_library(executor SRCS executor.cc DEPS op_registry device_context scope frame cc_library(prune SRCS prune.cc DEPS framework_proto) cc_test(prune_test SRCS prune_test.cc DEPS op_info prune recurrent_op device_context) - -cc_library(tensor_array SRCS tensor_array.cc DEPS lod_tensor) -cc_test(tensor_array_test SRCS tensor_array_test.cc DEPS tensor_array place) - cc_test(var_type_inference_test SRCS var_type_inference_test.cc DEPS op_registry proto_desc) cc_library(selected_rows SRCS selected_rows.cc DEPS tensor) cc_test(selected_rows_test SRCS selected_rows_test.cc DEPS selected_rows) + +cc_library(threadpool SRCS threadpool.cc) +cc_test(threadpool_test SRCS threadpool_test.cc DEPS threadpool) +cc_library(init SRCS init.cc DEPS gflags device_context place stringpiece) +cc_test(init_test SRCS init_test.cc DEPS init) + +cc_test(op_kernel_type_test SRCS op_kernel_type_test.cc DEPS place device_context framework_proto) diff --git a/paddle/framework/attribute.cc b/paddle/framework/attribute.cc index b1e17936417e4ce09bace1d1a5d346d1c9cfa710..b0fd4d2750eb2529706d871947332d39494505cd 100644 --- a/paddle/framework/attribute.cc +++ b/paddle/framework/attribute.cc @@ -19,42 +19,42 @@ limitations under the License. */ namespace paddle { namespace framework { -Attribute GetAttrValue(const OpDesc::Attr& attr_desc) { +Attribute GetAttrValue(const proto::OpDesc::Attr& attr_desc) { switch (attr_desc.type()) { - case framework::AttrType::BOOLEAN: { + case proto::AttrType::BOOLEAN: { return attr_desc.b(); } - case framework::AttrType::INT: { + case proto::AttrType::INT: { return attr_desc.i(); } - case framework::AttrType::FLOAT: { + case proto::AttrType::FLOAT: { return attr_desc.f(); } - case framework::AttrType::STRING: { + case proto::AttrType::STRING: { return attr_desc.s(); } - case framework::AttrType::BOOLEANS: { + case proto::AttrType::BOOLEANS: { std::vector val(attr_desc.bools_size()); for (int i = 0; i < attr_desc.bools_size(); ++i) { val[i] = attr_desc.bools(i); } return val; } - case framework::AttrType::INTS: { + case proto::AttrType::INTS: { std::vector val(attr_desc.ints_size()); for (int i = 0; i < attr_desc.ints_size(); ++i) { val[i] = attr_desc.ints(i); } return val; } - case framework::AttrType::FLOATS: { + case proto::AttrType::FLOATS: { std::vector val(attr_desc.floats_size()); for (int i = 0; i < attr_desc.floats_size(); ++i) { val[i] = attr_desc.floats(i); } return val; } - case framework::AttrType::STRINGS: { + case proto::AttrType::STRINGS: { std::vector val(attr_desc.strings_size()); for (int i = 0; i < attr_desc.strings_size(); ++i) { val[i] = attr_desc.strings(i); diff --git a/paddle/framework/attribute.h b/paddle/framework/attribute.h index 0641907d6ff7546df1601d3b0263ff42f4186968..c1c63d9cb13acb195b3bc3b30088f5fa7daf2a3d 100644 --- a/paddle/framework/attribute.h +++ b/paddle/framework/attribute.h @@ -27,12 +27,12 @@ limitations under the License. */ namespace paddle { namespace framework { template -inline AttrType AttrTypeID() { +inline proto::AttrType AttrTypeID() { Attribute tmp = T(); - return static_cast(tmp.which() - 1); + return static_cast(tmp.which() - 1); } -Attribute GetAttrValue(const OpDesc::Attr& attr_desc); +Attribute GetAttrValue(const proto::OpDesc::Attr& attr_desc); class AttrReader { public: diff --git a/paddle/framework/backward.cc b/paddle/framework/backward.cc index b9018ecdba8303fd6b37c87edd99e192aa604228..eaf13ddcefcd8dc5a6b0438f765d8d325925aa30 100644 --- a/paddle/framework/backward.cc +++ b/paddle/framework/backward.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/framework/backward.h" #include "paddle/operators/net_op.h" @@ -22,16 +22,27 @@ #include "paddle/framework/block_desc.h" #include "paddle/framework/op_registry.h" -#include "paddle/operators/dynamic_recurrent_op.h" #include "paddle/operators/net_op.h" namespace paddle { namespace framework { +static std::unordered_set* g_ctrl_flow_ops_ = nullptr; +// Control Flow operators's backward is significantly different from +// computational operators. Hack Code here. +// We should design a better way to backward CtrlFlowOps. +static std::unordered_set& CtrlFlowOps() { + if (g_ctrl_flow_ops_ == nullptr) { + g_ctrl_flow_ops_ = new std::unordered_set{ + "increment", "lod_rank_table", "less_than"}; + } + return *g_ctrl_flow_ops_; +} + static inline std::unique_ptr CreateGradOp( const OperatorBase& op, const std::unordered_set& no_grad_set, std::unordered_map* grad_to_var) { - OpDescBind op_desc; + OpDesc op_desc; op_desc.SetInputMap(op.Inputs()); op_desc.SetOutputMap(op.Outputs()); op_desc.SetType(op.Type()); @@ -42,7 +53,7 @@ static inline std::unique_ptr CreateGradOp( grad_ops.reserve(grad_descs.size()); std::transform(grad_descs.begin(), grad_descs.end(), std::back_inserter(grad_ops), - [](const std::unique_ptr& grad_desc) { + [](const std::unique_ptr& grad_desc) { return OpRegistry::CreateOp(*grad_desc); }); PADDLE_ENFORCE(!grad_ops.empty()); @@ -179,8 +190,9 @@ static std::unique_ptr BackwardRecursive( // collect all the offset for each alias, // insert a sum operator to add all aliases to output insert_position.push_back( - {dup_op.back(), OpRegistry::CreateOp("sum", {{"X", dup_outputs}}, - {{"Out", {name}}}, {})}); + {dup_op.back(), + OpRegistry::CreateOp("sum", {{"X", dup_outputs}}, {{"Out", {name}}}, + AttributeMap{})}); } // make sure the inserted `sum` ops follow the BFS order. @@ -205,7 +217,8 @@ static std::unique_ptr BackwardRecursive( // If part of input gradient of that operator is not calculated, fill // zero variables to that input gradient. net->AppendOp(OpRegistry::CreateOp("fill_zeros_like", {{"X", {prefix}}}, - {{"Y", {grad_input}}}, {})); + {{"Out", {grad_input}}}, + AttributeMap{})); } return false; }); @@ -218,21 +231,6 @@ static std::unique_ptr BackwardRecursive( return false; }); - // process recurrent gradient op as a special operator. - if (forwardOp.Type() == "dynamic_recurrent") { - // NOTE clean up cycle call somewhere (RNN's stepnet constains itself), - // or this will result in infinite loop. - const auto& rnnop = - *static_cast(&forwardOp); - auto rnn_grad_op = - static_cast(grad_op.get()); - const auto& stepnet_op = - *static_cast(&rnnop.rnn.GetStepUnit()); - // create stepnet's gradient op - rnn_grad_op->rnn.SetStepUnit( - BackwardRecursive(stepnet_op, no_grad_names, grad_to_var, uniq_id)); - } - if (net->ops_.empty()) { // Current no aux op is added to network return grad_op; } @@ -298,18 +296,30 @@ static std::string FwdName(const std::string& grad_name) { static void CreateGradVarInBlock( size_t grad_op_start_index, const std::unordered_map& param_name_map, - BlockDescBind* block_desc, + BlockDesc* block_desc, std::unordered_map* grad_var_record) { auto ops = block_desc->AllOps(); for (size_t op_index = grad_op_start_index; op_index < ops.size(); ++op_index) { std::unordered_set new_vars; + auto& ctrl_flow_ops = CtrlFlowOps(); ForEachVarName(ops[op_index]->Outputs(), [&](const std::string& grad_var_name) { - if (block_desc->HasVar(grad_var_name)) { + if (ctrl_flow_ops.find(ops[op_index]->Type()) != + ctrl_flow_ops.end()) { + if (block_desc->HasVarRecursive(grad_var_name)) { + return false; + } + } else { + if (block_desc->HasVar(grad_var_name)) { + return false; + } + } + if (grad_var_name == framework::kEmptyVarName) { return false; } auto var = block_desc->Var(grad_var_name); + VLOG(10) << "Creating Variable " << grad_var_name; new_vars.insert(var->Name()); auto it = param_name_map.find(grad_var_name); if (it == param_name_map.end()) { @@ -331,7 +341,7 @@ static void CreateGradVarInBlock( auto* param = block_desc->FindVarRecursive(pname); auto* grad = block_desc->FindVar(arg); if (param == nullptr) { - grad->SetDataType(DataType::FP32); + grad->SetDataType(proto::DataType::FP32); } else { grad->SetDataType(param->GetDataType()); } @@ -340,23 +350,33 @@ static void CreateGradVarInBlock( } } -std::vector> MakeOpGrad( - const OpDescBind* op_desc, std::unordered_set* no_grad_vars, +std::vector> MakeOpGrad( + const OpDesc* op_desc, std::unordered_set* no_grad_vars, std::unordered_map* grad_to_var, - const std::vector& grad_block = - std::vector()) { - std::vector> grad_op_descs; + const std::vector& grad_block = std::vector()) { + std::vector> grad_op_descs; // All input gradients of forwarding operator do not need to calculate. const std::vector& inputs = op_desc->InputArgumentNames(); if (AllGradInSet(inputs, *no_grad_vars)) { + VLOG(10) << "Drop operator " << op_desc->Type(); return grad_op_descs; // empty vector } + // All output gradients of forwarding operator do not need to calculate. const std::vector& outputs = op_desc->OutputArgumentNames(); + if (AllGradInSet(outputs, *no_grad_vars)) { - for (const std::string& name : inputs) { - no_grad_vars->insert(GradVarName(name)); + VLOG(10) << "Drop operator " << op_desc->Type(); + // FIXME: Hack code here + auto& ctrl_flow_ops = CtrlFlowOps(); + if (ctrl_flow_ops.find(op_desc->Type()) == ctrl_flow_ops.end()) { + // Only computational op need drop input's gradient. + for (const std::string& name : inputs) { + no_grad_vars->insert(GradVarName(name)); + VLOG(10) << " Also drop " << GradVarName(name); + } } + return grad_op_descs; // empty vector } @@ -365,7 +385,7 @@ std::vector> MakeOpGrad( .Get(op_desc->Type()) .GradOpMaker()(*op_desc, *no_grad_vars, grad_to_var, grad_block); - std::list> pending_fill_zeros_ops; + std::list> pending_fill_zeros_ops; for (auto& desc : grad_op_descs) { for (const std::string& in_name : desc->InputArgumentNames()) { if (no_grad_vars->count(in_name)) { @@ -373,8 +393,9 @@ std::vector> MakeOpGrad( 0, in_name.size() - sizeof(kGradVarSuffix) / sizeof(char) + 1); std::string new_name = prefix + kZeroVarSuffix; desc->Rename(in_name, new_name); - std::unique_ptr fill_zeros_op(new OpDescBind( - "fill_zeros_like", {{"X", {prefix}}}, {{"Y", {new_name}}}, {})); + std::unique_ptr fill_zeros_op( + new OpDesc("fill_zeros_like", {{"X", {prefix}}}, + {{"Out", {new_name}}}, AttributeMap{})); pending_fill_zeros_ops.push_back(std::move(fill_zeros_op)); } } @@ -386,36 +407,35 @@ std::vector> MakeOpGrad( return grad_op_descs; } -static BlockDescBind* CreateStepBlock( - ProgramDescBind& program_desc, - std::unordered_set* no_grad_vars, +static BlockDesc* CreateStepBlock( + ProgramDesc& program_desc, std::unordered_set* no_grad_vars, std::unordered_map* grad_to_var, int step_block_idx); -std::vector> MakeBlockBackward( - ProgramDescBind& program_desc, int block_idx, +std::vector> MakeBlockBackward( + ProgramDesc& program_desc, int block_idx, std::unordered_set* no_grad_vars, std::unordered_map* grad_to_var) { VLOG(5) << "MakeBlockBackward"; - BlockDescBind* cur_block = program_desc.MutableBlock(block_idx); - std::vector op_descs = cur_block->AllOps(); + BlockDesc* cur_block = program_desc.MutableBlock(block_idx); + std::vector op_descs = cur_block->AllOps(); std::unordered_map> dup_out_ops; size_t grad_desc_idx = 0; - std::vector> backward_descs; + std::vector> backward_descs; for (auto it = op_descs.rbegin(); it != op_descs.rend(); ++it) { VLOG(5) << "Making backward " << (*it)->Type() << " op"; - std::vector> op_grads; + std::vector> op_grads; if ((*it)->Type() == "recurrent" || (*it)->Type() == "while") { - int step_block_idx = (*it)->GetBlockAttr("step_block"); - BlockDescBind* backward_block = CreateStepBlock( - program_desc, no_grad_vars, grad_to_var, step_block_idx); + int step_block_idx = (*it)->GetBlockAttr("sub_block"); + BlockDesc* backward_block = CreateStepBlock(program_desc, no_grad_vars, + grad_to_var, step_block_idx); op_grads = MakeOpGrad(*it, no_grad_vars, grad_to_var, {backward_block}); } else if ((*it)->Type() == "conditional_block") { - BlockDescBind* backward_block = + BlockDesc* backward_block = CreateStepBlock(program_desc, no_grad_vars, grad_to_var, - (*it)->GetBlockAttr("block")); + (*it)->GetBlockAttr("sub_block")); op_grads = MakeOpGrad(*it, no_grad_vars, grad_to_var, {backward_block}); } else { op_grads = MakeOpGrad(*it, no_grad_vars, grad_to_var); @@ -441,14 +461,14 @@ std::vector> MakeBlockBackward( } ++grad_desc_idx; } - std::transform( - op_grads.begin(), op_grads.end(), std::back_inserter(backward_descs), - [](std::unique_ptr& ptr) { return std::move(ptr); }); + std::transform(op_grads.begin(), op_grads.end(), + std::back_inserter(backward_descs), + [](std::unique_ptr& ptr) { return std::move(ptr); }); } VLOG(5) << "Appending Sums"; // Check whether some variables are written more than once - std::list>> pending_sum_ops; + std::list>> pending_sum_ops; for (const auto& dup : dup_out_ops) { const std::string& out_name = dup.first; const std::vector dup_op = dup.second; @@ -464,17 +484,17 @@ std::vector> MakeBlockBackward( sum_op_inputs.emplace_back(new_name); next_g_name = sum_op_inputs.back(); } - std::unique_ptr sum_op(new OpDescBind( - "sum", {{"X", sum_op_inputs}}, {{"Out", {out_name}}}, {})); + std::unique_ptr sum_op(new OpDesc("sum", {{"X", sum_op_inputs}}, + {{"Out", {out_name}}}, + AttributeMap{})); pending_sum_ops.push_back({dup_op.back(), std::move(sum_op)}); } } - pending_sum_ops.sort( - [](const std::pair>& a, - const std::pair>& b) { - return a.first > b.first; - }); + pending_sum_ops.sort([](const std::pair>& a, + const std::pair>& b) { + return a.first > b.first; + }); for (auto& p : pending_sum_ops) { backward_descs.insert(backward_descs.begin() + p.first + 1, std::move(p.second)); @@ -485,14 +505,13 @@ std::vector> MakeBlockBackward( return backward_descs; } -static BlockDescBind* CreateStepBlock( - ProgramDescBind& program_desc, - std::unordered_set* no_grad_vars, +static BlockDesc* CreateStepBlock( + ProgramDesc& program_desc, std::unordered_set* no_grad_vars, std::unordered_map* grad_to_var, int step_block_idx) { auto backward_block_op_descs = MakeBlockBackward(program_desc, step_block_idx, no_grad_vars, grad_to_var); - BlockDescBind* backward_block = + BlockDesc* backward_block = program_desc.AppendBlock(*program_desc.MutableBlock(step_block_idx)); for (auto& ptr : backward_block_op_descs) { backward_block->AppendAllocatedOp(move(ptr)); @@ -501,7 +520,7 @@ static BlockDescBind* CreateStepBlock( } ParamGradInfoMap AppendBackward( - ProgramDescBind& program_desc, const VarDescBind& target, + ProgramDesc& program_desc, const VarDesc& target, const std::unordered_set& no_grad_vars) { std::unordered_set no_grad_var_names; no_grad_var_names.reserve(no_grad_vars.size() + 1); @@ -518,11 +537,11 @@ ParamGradInfoMap AppendBackward( PADDLE_ENFORCE(is_scalar, "target should be scalar"); VLOG(3) << "backward from loss=" << target.Name() << " data_type=" << target.GetDataType(); - std::unique_ptr fill_one_op( - new OpDescBind("fill_constant", {}, {{"Out", {fill_one_op_out}}}, - {{"shape", std::vector{1}}, - {"value", static_cast(1.0)}, - {"data_type", target.GetDataType()}})); + std::unique_ptr fill_one_op( + new OpDesc("fill_constant", {}, {{"Out", {fill_one_op_out}}}, + {{"shape", std::vector{1}}, + {"value", static_cast(1.0)}, + {"dtype", target.GetDataType()}})); // infer var type of fill_one_op fill_one_op->InferVarType(root_block); diff --git a/paddle/framework/backward.h b/paddle/framework/backward.h index 96154fa82cb7a486aa4762ae633982ed6735220b..69ee3802369c16a8b21c0710d2008ef3c085cc5c 100644 --- a/paddle/framework/backward.h +++ b/paddle/framework/backward.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 @@ -49,7 +49,7 @@ using ParamGradInfoMap = std::unordered_map; ParamGradInfoMap AppendBackward( - ProgramDescBind& program_desc, const VarDescBind& target, + ProgramDesc& program_desc, const VarDesc& target, const std::unordered_set& no_grad_vars); } // namespace framework diff --git a/paddle/framework/backward.md b/paddle/framework/backward.md deleted file mode 100644 index ac60be572419b62f4beb644ff192d413c35e19bb..0000000000000000000000000000000000000000 --- a/paddle/framework/backward.md +++ /dev/null @@ -1,100 +0,0 @@ -# Operator/expression 's Backward - -## Motivation - -In Neural Network, most models are solved by the backpropagation algorithm(known as **BP**) at present. Technically, BP calculates the gradient of the loss function, then propagates it back through the networks following the chain rule. Hence we need a module that chains the gradient operators/expressions together to construct the backward pass. Every forward network needs a backward network to construct the full computation graph. The operator/expression's backward pass will be generated with respect to the forward pass. - -## Implementation - -In this design doc, we exported only one API for generating the backward pass. - -```c++ -std::unique_ptr Backward(const OperatorBase& forwardOp, - const std::unordered_set& no_grad_vars); -``` - -The implementation behind it can be divided into two parts, **Backward Operator Creating** and **Backward Operator Building**. - -### Backward Operator Registry - -A backward network is built up with several backward operators. Backward operators take forward operators' inputs, outputs, and output gradients and then calculate its input gradients. - -| | forward operator | backward operator -| ---------------------- | ---------------- |------------------------- | -| **Operator::inputs_** | Inputs | Inputs, Outputs, OutputGradients | -| **Operator::outputs_** | Outputs | InputGradients | - - In most cases, there is a one-to-one relation between the forward and backward operators. These relations are recorded by a global hash map(`OpInfoMap`). To follow the philosophy of minimum core and to make operators pluggable, the registry mechanism is introduced. - -For example, we have `mul_op`, and we can register its information and corresponding backward operator by the following macro: - -```cpp -REGISTER_OP(mul, MulOp, MulOpMaker, mul_grad, MulOpGrad); -``` - -`mul` is the operator's type. `MulOp` and `MulOpMaker` are the operator class and the operator maker class respectively. - -`mul_grad` is the type of backward operator, and `MulOpGrad` is its class name. - -### Backward Opeartor Creating - -Given a certain forward operator, we can get its corresponding backward operator by calling: - -```cpp -OperatorBase* bwd_op = BuildGradOp(const OperatorBase* fwd_op); -``` - -The function `BuildGradOp` will sequentially execute following processes: - -1. Get the `type_` of given forward operator, and then get the corresponding backward operator's type by looking up the `OpInfoMap`. - -2. Build two maps named `inputs` and `outputs` to temporarily store backward operator's inputs and outputs. Copy forward operator's `inputs_` and `outputs_` to map `inputs`, except these, are not necessary for gradient computing. - -3. Add forward inputs' gradient variables into map `output`, adding forward outputs' gradient variables into map `input`. - -4. Building backward operator with `inputs`, `outputs` and forward operator's attributes. - -### Backward Network Building - -A backward network is a series of backward operators. The main idea of building a backward network is creating backward operators in the inverted sequence and appending them together one by one. There are some corner cases that need special processing. - -1. Op - - When the input forward network is an Op, return its gradient Operator immediately. If all of its outputs are in no gradient set, then return a special `NOP`. - -2. NetOp - - In our design, the network itself is also a kind of operator(**NetOp**). So the operators contained by a big network may be some small network. When the input forward network is a NetOp, it needs to call the sub NetOp/Operators backward function recursively. During the process, we need to collect the `OutputGradients` name according to the forward NetOp. - -3. RnnOp - - RnnOp is a nested stepnet operator. Backward module needs to recusively call `Backward` for every stepnet. - -4. Sharing Variables - - As illustrated in the figure 1 and figure 2, two operators share the same variable name **W@GRAD**, which will overwrite their shared input variable. - -

-
- -​ Figure 1. Sharing variables in operators. - -

- -​ Sharing variable between operators or same input variable used in multiple operators can lead to duplicate gradient variables. As illustrated in figure 2, we need to rename the gradient names recursively and add a generic add operator to prevent overwriting. - -

-
- -​ Figure 2. Replace sharing variable's gradient with `Add` operator. - -

- -​ Because the framework finds variables according to their names, we need to rename the output links. We add an integer suffix to represent its position in the clockwise direction. - -5. Part of the Gradient is Zero. - - In the whole graph, there is some case of that one operator's gradient is not needed, but its input's gradient is a dependency link of other operator, we need to fill a same shape gradient matrix in the position. In our implementation, we insert a special `fillZeroLike` operator. - - -Follow these rules above, then collect the sub graph `OutputGradients`/`InputGradients` as the NetOp's and return it. diff --git a/paddle/framework/backward_test.cc b/paddle/framework/backward_test.cc index 2b858f5ea0874d7bf1a9cf38529f5d0d70cca7f2..692406b1c37d0c02714eafb5cf9a28329ed873bc 100644 --- a/paddle/framework/backward_test.cc +++ b/paddle/framework/backward_test.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/framework/backward.h" @@ -58,13 +58,13 @@ class RowWiseAddGradMaker : public SingleGradOpDescMaker { using SingleGradOpDescMaker::SingleGradOpDescMaker; protected: - std::unique_ptr Apply() const override { - auto grad_op = new OpDescBind(); + std::unique_ptr Apply() const override { + auto grad_op = new OpDesc(); grad_op->SetInput(GradVarName("Out"), OutputGrad("Out")); grad_op->SetOutput(GradVarName("X"), InputGrad("X")); grad_op->SetOutput(GradVarName("b"), InputGrad("b")); grad_op->SetType("rowwise_add_grad"); - return std::unique_ptr(grad_op); + return std::unique_ptr(grad_op); } }; @@ -106,15 +106,15 @@ class FcOp : public operators::NetOp { FcOp(const std::string &type, const VariableNameMap &inputs, const VariableNameMap &outputs, const AttributeMap &attrs) : NetOp(type, inputs, outputs, attrs) { - AppendOp(OpRegistry::CreateOp("mul", - {{"X", {Input("X")}}, {"Y", {Input("W")}}}, - {{"Out", {Output("mul_result")}}}, {})); + AppendOp(OpRegistry::CreateOp( + "mul", {{"X", {Input("X")}}, {"Y", {Input("W")}}}, + {{"Out", {Output("mul_result")}}}, AttributeMap{})); auto input_b = Inputs("b"); std::string before_act = "mul_result"; if (input_b.size() != 0) { AppendOp(OpRegistry::CreateOp( "rowwise_add", {{"X", {Output("mul_result")}}, {"b", {input_b[0]}}}, - {{"Out", {Output("add_result")}}}, {})); + {{"Out", {Output("add_result")}}}, AttributeMap{})); before_act = "add_result"; } else { auto out_varname = Output("add_result"); @@ -124,7 +124,7 @@ class FcOp : public operators::NetOp { } AppendOp(OpRegistry::CreateOp("sigmoid", {{"X", {Output(before_act)}}}, - {{"Out", {Output("Out")}}}, {})); + {{"Out", {Output("Out")}}}, AttributeMap{})); CompleteAddOp(false); } }; @@ -159,14 +159,14 @@ class FillZeroOpMaker : public OpProtoAndCheckerMaker { FillZeroOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "x"); - AddOutput("Y", "out"); + AddOutput("Out", "out"); AddComment(""); } }; class SumOpMaker : public framework::OpProtoAndCheckerMaker { public: - SumOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + SumOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "the input tensors of sum operator.").AsDuplicable(); AddOutput("Out", "the output tensor of sum operator."); @@ -190,11 +190,11 @@ class MinusGradOpDescMaker : public GradOpDescMakerBase { public: using GradOpDescMakerBase::GradOpDescMakerBase; - std::vector> operator()() const override { - std::vector> retv; + std::vector> operator()() const override { + std::vector> retv; auto x_g = InputGrad("X"); if (!x_g.empty()) { - auto *op_desc = new OpDescBind(); + auto *op_desc = new OpDesc(); op_desc->SetType("scale"); op_desc->SetInput("X", OutputGrad("Out")); op_desc->SetOutput("Out", x_g); @@ -204,7 +204,7 @@ class MinusGradOpDescMaker : public GradOpDescMakerBase { auto y_g = InputGrad("Y"); if (!y_g.empty()) { - auto *op_desc = new OpDescBind(); + auto *op_desc = new OpDesc(); op_desc->SetType("scale"); op_desc->SetInput("X", OutputGrad("Out")); op_desc->SetOutput("Out", y_g); @@ -278,8 +278,9 @@ REGISTER_OPERATOR(scale, f::NoneOp); REGISTER_OP_CPU_KERNEL(scale, f::NoneKernel); TEST(Backward, simple_op_not_need_grad) { - auto fwd = f::OpRegistry::CreateOp( - "rowwise_add", {{"X", {"x"}}, {"b", {"b"}}}, {{"Out", {"out"}}}, {}); + auto fwd = + f::OpRegistry::CreateOp("rowwise_add", {{"X", {"x"}}, {"b", {"b"}}}, + {{"Out", {"out"}}}, f::AttributeMap{}); ASSERT_NE(fwd, nullptr); auto gop = f::Backward(*fwd, {"x"}); ASSERT_EQ(gop->Output(f::GradVarName("X")), f::kEmptyVarName); @@ -296,9 +297,10 @@ TEST(Backward, net_fc_backward_normal) { {{"mul_result", {"mul_res"}}, {"add_result", {"add_re"}}, {"Out", {"out"}}}, - {}); + f::AttributeMap{}); ASSERT_NE(fwd, nullptr); - std::shared_ptr gop = f::Backward(*fwd, {}); + std::shared_ptr gop = + f::Backward(*fwd, std::unordered_set{}); ASSERT_TRUE(gop->IsNetOp()); auto net = static_cast(gop.get()); @@ -322,9 +324,10 @@ TEST(Backward, net_fc_backward_not_have_b) { {{"mul_result", {"mul_res"}}, {"add_result", {"add_res"}}, {"Out", {"tmp"}}}, - {}); + f::AttributeMap{}); ASSERT_NE(fwd, nullptr); - std::shared_ptr gop = f::Backward(*fwd, {}); + std::shared_ptr gop = + f::Backward(*fwd, std::unordered_set{}); ASSERT_TRUE(gop->IsNetOp()); auto net = static_cast(gop.get()); @@ -346,13 +349,13 @@ TEST(Backward, net_input_of_network_not_need_grad) { {{"mul_result", {"mul_tmp_0"}}, {"add_result", {"add_tmp_0"}}, {"Out", {"hidden0"}}}, - {})); + f::AttributeMap{})); net.AppendOp(f::OpRegistry::CreateOp( "fc", {{"X", {"hidden0"}}, {"W", {"W2"}}, {"b", {"b2"}}}, {{"mul_result", {"mul_tmp_1"}}, {"add_result", {"add_tmp_1"}}, {"Out", {"hidden1"}}}, - {})); + f::AttributeMap{})); net.CompleteAddOp(); auto bwd = Backward(net, {"x"}); // x@GRAD is not need. ASSERT_TRUE(bwd->IsNetOp()); @@ -381,12 +384,13 @@ TEST(Backward, net_input_of_network_not_need_grad) { TEST(Backward, net_shared_weight) { ops::NetOp net; net.AppendOp(f::OpRegistry::CreateOp("mul", {{"X", {"x"}}, {"Y", {"w"}}}, - {{"Out", {"out"}}}, {})); + {{"Out", {"out"}}}, f::AttributeMap{})); net.AppendOp(f::OpRegistry::CreateOp("mul", {{"X", {"out"}}, {"Y", {"w"}}}, - {{"Out", {"FinalOut"}}}, {})); + {{"Out", {"FinalOut"}}}, + f::AttributeMap{})); net.CompleteAddOp(); - auto bwd = f::Backward(net, {}); + auto bwd = f::Backward(net, std::unordered_set{}); ASSERT_TRUE(bwd->IsNetOp()); auto bwd_net = static_cast(bwd.get()); ASSERT_EQ(3UL, bwd_net->ops_.size()); @@ -394,8 +398,9 @@ TEST(Backward, net_shared_weight) { } TEST(Backward, op_all_input_are_not_need) { - auto fwd = f::OpRegistry::CreateOp( - "rowwise_add", {{"X", {"x"}}, {"b", {"b"}}}, {{"Out", {"out"}}}, {}); + auto fwd = + f::OpRegistry::CreateOp("rowwise_add", {{"X", {"x"}}, {"b", {"b"}}}, + {{"Out", {"out"}}}, f::AttributeMap{}); auto backward = f::Backward(*fwd, {"x", "b"}); ASSERT_TRUE(backward->IsNetOp()); auto net = static_cast(backward.get()); @@ -403,8 +408,9 @@ TEST(Backward, op_all_input_are_not_need) { } TEST(Backward, op_all_output_are_not_need) { - auto fwd = f::OpRegistry::CreateOp( - "rowwise_add", {{"X", {"x"}}, {"b", {"b"}}}, {{"Out", {"out"}}}, {}); + auto fwd = + f::OpRegistry::CreateOp("rowwise_add", {{"X", {"x"}}, {"b", {"b"}}}, + {{"Out", {"out"}}}, f::AttributeMap{}); auto backward = f::Backward(*fwd, {"out"}); ASSERT_TRUE(backward->IsNetOp()); auto net = static_cast(backward.get()); @@ -412,8 +418,9 @@ TEST(Backward, op_all_output_are_not_need) { } TEST(Backward, op_part_of_output_are_not_need) { - auto fwd = f::OpRegistry::CreateOp("many_output_op", {{"x", {"X"}}}, - {{"y", {"Y"}}, {"z", {"Z"}}}, {}); + auto fwd = + f::OpRegistry::CreateOp("many_output_op", {{"x", {"X"}}}, + {{"y", {"Y"}}, {"z", {"Z"}}}, f::AttributeMap{}); auto backward = f::Backward(*fwd, {"Z"}); ASSERT_TRUE(backward->IsNetOp()); auto net = static_cast(backward.get()); @@ -423,8 +430,8 @@ TEST(Backward, op_part_of_output_are_not_need) { ASSERT_EQ("fill_zeros_like", fill_zero.Type()); ASSERT_EQ(1UL, fill_zero.Inputs("X").size()); ASSERT_EQ("Z", fill_zero.Input("X")); - ASSERT_EQ(1UL, fill_zero.Outputs("Y").size()); - ASSERT_EQ(std::string("Z") + f::kZeroVarSuffix, fill_zero.Output("Y")); + ASSERT_EQ(1UL, fill_zero.Outputs("Out").size()); + ASSERT_EQ(std::string("Z") + f::kZeroVarSuffix, fill_zero.Output("Out")); auto &d_many_out = *net->ops_[1]; ASSERT_EQ("many_output_op_grad", d_many_out.Type()); @@ -437,7 +444,7 @@ TEST(Backward, op_part_of_output_are_not_need) { TEST(Backward, op_part_of_input_are_not_need) { auto fwd = f::OpRegistry::CreateOp("mul", {{"X", {"a"}}, {"Y", {"b"}}}, - {{"Out", {"out"}}}, {}); + {{"Out", {"out"}}}, f::AttributeMap{}); auto backward = f::Backward(*fwd, {"a"}); auto &grad_mul = *backward; ASSERT_EQ(grad_mul.Type(), "mul_grad"); @@ -458,19 +465,19 @@ TEST(Backward, linear_net_intermediate_variable_has_no_grad) { {{"mul_result", {"mul_out1"}}, {"add_result", {"add_out1"}}, {"Out", {"out1"}}}, - {})); + f::AttributeMap{})); net.AppendOp(f::OpRegistry::CreateOp( "fc", {{"X", {"out1"}}, {"W", {"w2"}}, {"b", {"b2"}}}, {{"mul_result", {"mul_out2"}}, {"add_result", {"tmp_out2"}}, {"Out", {"out2"}}}, - {})); + f::AttributeMap{})); net.AppendOp(f::OpRegistry::CreateOp( "fc", {{"X", {"out2"}}, {"W", {"w3"}}, {"b", {"b3"}}}, {{"mul_result", {"mul_out3"}}, {"add_result", {"tmp_out3"}}, {"Out", {"out3"}}}, - {})); + f::AttributeMap{})); net.CompleteAddOp(); auto backward = f::Backward(net, {"mul_out2", "tmp_out2", "out2"}); @@ -498,24 +505,25 @@ TEST(Backward, linear_net_intermediate_variable_has_no_grad) { } TEST(Backward, simple_single_op) { - f::ProgramDescBind program; - f::BlockDescBind *block = program.MutableBlock(0); + f::ProgramDesc program; + f::BlockDesc *block = program.MutableBlock(0); - f::OpDescBind *op = block->AppendOp(); + f::OpDesc *op = block->AppendOp(); op->SetType("rowwise_add"); op->SetInput("X", {"x"}); op->SetInput("b", {"b"}); op->SetOutput("Out", {"out"}); - auto target = f::VarDescBind("out"); + auto target = f::VarDesc("out"); target.SetShape({1}); - auto var_to_grad = AppendBackward(program, target, {}); + auto var_to_grad = + AppendBackward(program, target, std::unordered_set{}); ASSERT_EQ(block->AllOps().size(), 3UL); - f::OpDescBind *fill_op = block->AllOps()[1]; + f::OpDesc *fill_op = block->AllOps()[1]; EXPECT_EQ(fill_op->Type(), "fill_constant"); - f::OpDescBind *grad_op = block->AllOps()[2]; + f::OpDesc *grad_op = block->AllOps()[2]; EXPECT_EQ(grad_op->Type(), "rowwise_add_grad"); ASSERT_EQ(grad_op->InputNames().size(), 1UL); ASSERT_EQ(grad_op->OutputNames().size(), 2UL); @@ -535,63 +543,64 @@ TEST(Backward, simple_single_op) { } TEST(Backward, default_attribute) { - f::ProgramDescBind program; - f::BlockDescBind *block = program.MutableBlock(0); - f::OpDescBind *op = block->AppendOp(); + f::ProgramDesc program; + f::BlockDesc *block = program.MutableBlock(0); + f::OpDesc *op = block->AppendOp(); op->SetType("mul"); op->SetInput("X", {"x"}); op->SetInput("Y", {"y"}); op->SetOutput("Out", {"out"}); op->CheckAttrs(); - auto target = f::VarDescBind("out"); + auto target = f::VarDesc("out"); target.SetShape({1}); - AppendBackward(program, target, {}); + AppendBackward(program, target, std::unordered_set{}); ASSERT_EQ(block->AllOps().size(), 3UL); EXPECT_EQ(boost::get(op->GetAttr("x_num_col_dims")), 1); EXPECT_EQ(boost::get(op->GetAttr("y_num_col_dims")), 1); - f::OpDescBind *fill_op = block->AllOps()[1]; + f::OpDesc *fill_op = block->AllOps()[1]; EXPECT_EQ(fill_op->Type(), "fill_constant"); - f::OpDescBind *grad_op = block->AllOps()[2]; + f::OpDesc *grad_op = block->AllOps()[2]; ASSERT_EQ(grad_op->Type(), "mul_grad"); EXPECT_EQ(boost::get(grad_op->GetAttr("x_num_col_dims")), 1); EXPECT_EQ(boost::get(grad_op->GetAttr("y_num_col_dims")), 1); } TEST(Backward, simple_mult_op) { - f::ProgramDescBind program; - f::BlockDescBind *block = program.MutableBlock(0); - f::OpDescBind *op1 = block->AppendOp(); + f::ProgramDesc program; + f::BlockDesc *block = program.MutableBlock(0); + f::OpDesc *op1 = block->AppendOp(); op1->SetType("rowwise_add"); op1->SetInput("X", {"x1"}); op1->SetInput("b", {"b1"}); op1->SetOutput("Out", {"out1"}); - f::OpDescBind *op2 = block->AppendOp(); + f::OpDesc *op2 = block->AppendOp(); op2->SetType("mul"); op2->SetInput("X", {"out1"}); op2->SetInput("Y", {"y2"}); op2->SetOutput("Out", {"out2"}); - f::OpDescBind *op3 = block->AppendOp(); + f::OpDesc *op3 = block->AppendOp(); op3->SetType("rowwise_add"); op3->SetInput("X", {"out2"}); op3->SetInput("b", {"b3"}); op3->SetOutput("Out", {"out3"}); - auto target = f::VarDescBind("out3"); + auto target = f::VarDesc("out3"); target.SetShape({1}); size_t forward_len = block->AllOps().size(); - auto var_to_grad = AppendBackward(program, target, {}); + auto var_to_grad = + AppendBackward(program, target, std::unordered_set{}); ASSERT_EQ(block->AllOps().size(), 6UL + 1); - f::OpDescBind *fill_op = block->AllOps()[forward_len]; + f::OpDesc *fill_op = block->AllOps()[forward_len]; EXPECT_EQ(fill_op->Type(), "fill_constant"); - f::OpDescBind *grad_op1 = block->AllOps()[6]; + f::OpDesc *grad_op1 = block->AllOps()[6]; EXPECT_EQ(grad_op1->Type(), "rowwise_add_grad"); ASSERT_EQ(grad_op1->InputNames().size(), 1UL); ASSERT_EQ(grad_op1->OutputNames().size(), 2UL); @@ -602,7 +611,7 @@ TEST(Backward, simple_mult_op) { EXPECT_EQ(grad_op1->Output(f::GradVarName("b")), std::vector({f::GradVarName("b1")})); - f::OpDescBind *grad_op2 = block->AllOps()[5]; + f::OpDesc *grad_op2 = block->AllOps()[5]; EXPECT_EQ(grad_op2->Type(), "mul_grad"); ASSERT_EQ(grad_op2->InputNames().size(), 4UL); ASSERT_EQ(grad_op2->OutputNames().size(), 2UL); @@ -616,7 +625,7 @@ TEST(Backward, simple_mult_op) { EXPECT_EQ(grad_op2->Output(f::GradVarName("Y")), std::vector({f::GradVarName("y2")})); - f::OpDescBind *grad_op3 = block->AllOps()[4]; + f::OpDesc *grad_op3 = block->AllOps()[4]; EXPECT_EQ(grad_op3->Type(), "rowwise_add_grad"); ASSERT_EQ(grad_op3->InputNames().size(), 1UL); ASSERT_EQ(grad_op3->OutputNames().size(), 2UL); @@ -646,42 +655,42 @@ TEST(Backward, simple_mult_op) { } TEST(Backward, intermedia_var_no_grad) { - f::ProgramDescBind program; - f::BlockDescBind *block = program.MutableBlock(0); - f::OpDescBind *op1 = block->AppendOp(); + f::ProgramDesc program; + f::BlockDesc *block = program.MutableBlock(0); + f::OpDesc *op1 = block->AppendOp(); op1->SetType("rowwise_add"); op1->SetInput("X", {"x1"}); op1->SetInput("b", {"b1"}); op1->SetOutput("Out", {"out1"}); - f::OpDescBind *op2 = block->AppendOp(); + f::OpDesc *op2 = block->AppendOp(); op2->SetType("mul"); op2->SetInput("X", {"x2"}); op2->SetInput("Y", {"y2"}); op2->SetOutput("Out", {"out2"}); - f::OpDescBind *op3 = block->AppendOp(); + f::OpDesc *op3 = block->AppendOp(); op3->SetType("rowwise_add"); op3->SetInput("X", {"out2"}); op3->SetInput("b", {"b3"}); op3->SetOutput("Out", {"out3"}); - f::OpDescBind *op4 = block->AppendOp(); + f::OpDesc *op4 = block->AppendOp(); op4->SetType("mul"); op4->SetInput("X", {"out1"}); op4->SetInput("Y", {"out3"}); op4->SetOutput("Out", {"out4"}); - auto target = f::VarDescBind("out4"); + auto target = f::VarDesc("out4"); target.SetShape({1}); size_t forward_len = block->AllOps().size(); auto var_to_grad = AppendBackward(program, target, {"out3"}); ASSERT_EQ(block->AllOps().size(), 7UL); - f::OpDescBind *fill_op = block->AllOps()[forward_len]; + f::OpDesc *fill_op = block->AllOps()[forward_len]; EXPECT_EQ(fill_op->Type(), "fill_constant"); - f::OpDescBind *grad_op1 = block->AllOps()[6]; + f::OpDesc *grad_op1 = block->AllOps()[6]; EXPECT_EQ(grad_op1->Type(), "rowwise_add_grad"); ASSERT_EQ(grad_op1->InputNames().size(), 1UL); ASSERT_EQ(grad_op1->OutputNames().size(), 2UL); @@ -692,7 +701,7 @@ TEST(Backward, intermedia_var_no_grad) { EXPECT_EQ(grad_op1->Output(f::GradVarName("b")), std::vector({f::GradVarName("b1")})); - f::OpDescBind *grad_op4 = block->AllOps()[5]; + f::OpDesc *grad_op4 = block->AllOps()[5]; EXPECT_EQ(grad_op4->Type(), "mul_grad"); ASSERT_EQ(grad_op4->InputNames().size(), 4UL); ASSERT_EQ(grad_op4->OutputNames().size(), 2UL); @@ -717,32 +726,32 @@ TEST(Backward, intermedia_var_no_grad) { } TEST(Backward, var_no_grad) { - f::ProgramDescBind program; - f::BlockDescBind *block = program.MutableBlock(0); - f::OpDescBind *op1 = block->AppendOp(); + f::ProgramDesc program; + f::BlockDesc *block = program.MutableBlock(0); + f::OpDesc *op1 = block->AppendOp(); op1->SetType("mult_in_out"); op1->SetInput("X", {"x1"}); op1->SetInput("H", {"h1"}); op1->SetOutput("Y", {"y1"}); op1->SetOutput("Z", {"z1"}); - f::OpDescBind *op2 = block->AppendOp(); + f::OpDesc *op2 = block->AppendOp(); op2->SetType("mult_in_out"); op2->SetInput("X", {"y1"}); op2->SetInput("H", {"z1"}); op2->SetOutput("Y", {"y2"}); op2->SetOutput("Z", {"z2"}); - auto target = f::VarDescBind("z2"); + auto target = f::VarDesc("z2"); target.SetShape({1}); size_t forward_len = block->AllOps().size(); auto var_to_grad = AppendBackward(program, target, {"z1"}); ASSERT_EQ(block->AllOps().size(), 6UL); - f::OpDescBind *fill_op = block->AllOps()[forward_len]; + f::OpDesc *fill_op = block->AllOps()[forward_len]; EXPECT_EQ(fill_op->Type(), "fill_constant"); - f::OpDescBind *grad_op2 = block->AllOps()[3]; + f::OpDesc *grad_op2 = block->AllOps()[3]; ASSERT_EQ(grad_op2->Type(), "mult_in_out_grad"); ASSERT_EQ(grad_op2->InputNames().size(), 6UL); ASSERT_EQ(grad_op2->OutputNames().size(), 2UL); @@ -758,15 +767,15 @@ TEST(Backward, var_no_grad) { std::vector({f::GradVarName("y1")})); EXPECT_EQ(grad_op2->Output(f::GradVarName("H")), std::vector()); - f::OpDescBind *fill_zero_op = block->AllOps()[4]; + f::OpDesc *fill_zero_op = block->AllOps()[4]; ASSERT_EQ(fill_zero_op->Type(), "fill_zeros_like"); ASSERT_EQ(fill_zero_op->InputNames().size(), 1UL); ASSERT_EQ(fill_zero_op->OutputNames().size(), 1UL); EXPECT_EQ(fill_zero_op->Input("X"), std::vector({"z1"})); - EXPECT_EQ(fill_zero_op->Output("Y"), + EXPECT_EQ(fill_zero_op->Output("Out"), std::vector({std::string("z1") + f::kZeroVarSuffix})); - f::OpDescBind *grad_op1 = block->AllOps()[5]; + f::OpDesc *grad_op1 = block->AllOps()[5]; ASSERT_EQ(grad_op1->Type(), "mult_in_out_grad"); ASSERT_EQ(grad_op1->InputNames().size(), 6UL); ASSERT_EQ(grad_op1->OutputNames().size(), 2UL); @@ -794,36 +803,37 @@ TEST(Backward, var_no_grad) { } TEST(Backward, shared_var) { - f::ProgramDescBind program; - f::BlockDescBind *block = program.MutableBlock(0); - f::OpDescBind *op1 = block->AppendOp(); + f::ProgramDesc program; + f::BlockDesc *block = program.MutableBlock(0); + f::OpDesc *op1 = block->AppendOp(); op1->SetType("rowwise_add"); op1->SetInput("X", {"x1"}); op1->SetInput("b", {"b1"}); op1->SetOutput("Out", {"out1"}); - f::OpDescBind *op2 = block->AppendOp(); + f::OpDesc *op2 = block->AppendOp(); op2->SetType("mul"); op2->SetInput("X", {"out1"}); op2->SetInput("Y", {"y2"}); op2->SetOutput("Out", {"out2"}); - f::OpDescBind *op3 = block->AppendOp(); + f::OpDesc *op3 = block->AppendOp(); op3->SetType("rowwise_add"); op3->SetInput("X", {"out1"}); op3->SetInput("b", {"b3"}); op3->SetOutput("Out", {"out3"}); - auto target = f::VarDescBind("out3"); + auto target = f::VarDesc("out3"); target.SetShape({1}); size_t forward_len = block->AllOps().size(); - auto var_to_grad = AppendBackward(program, target, {}); + auto var_to_grad = + AppendBackward(program, target, std::unordered_set{}); ASSERT_EQ(block->AllOps().size(), 8UL); - f::OpDescBind *fill_op = block->AllOps()[forward_len]; + f::OpDesc *fill_op = block->AllOps()[forward_len]; EXPECT_EQ(fill_op->Type(), "fill_constant"); - f::OpDescBind *grad_op3 = block->AllOps()[4]; + f::OpDesc *grad_op3 = block->AllOps()[4]; ASSERT_EQ(grad_op3->Type(), "rowwise_add_grad"); ASSERT_EQ(grad_op3->InputNames().size(), 1UL); ASSERT_EQ(grad_op3->OutputNames().size(), 2UL); @@ -834,7 +844,7 @@ TEST(Backward, shared_var) { EXPECT_EQ(grad_op3->Output(f::GradVarName("b")), std::vector({f::GradVarName("b3")})); - f::OpDescBind *grad_op4 = block->AllOps()[5]; + f::OpDesc *grad_op4 = block->AllOps()[5]; ASSERT_EQ(grad_op4->Type(), "mul_grad"); ASSERT_EQ(grad_op4->InputNames().size(), 4UL); ASSERT_EQ(grad_op4->OutputNames().size(), 2UL); @@ -848,7 +858,7 @@ TEST(Backward, shared_var) { EXPECT_EQ(grad_op4->Output(f::GradVarName("Y")), std::vector({f::GradVarName("y2")})); - f::OpDescBind *sum_op = block->AllOps()[6]; + f::OpDesc *sum_op = block->AllOps()[6]; ASSERT_EQ(sum_op->Type(), "sum"); ASSERT_EQ(sum_op->InputNames().size(), 1UL); ASSERT_EQ(sum_op->OutputNames().size(), 1UL); @@ -858,7 +868,7 @@ TEST(Backward, shared_var) { EXPECT_EQ(sum_op->Output("Out"), std::vector({f::GradVarName("out1")})); - f::OpDescBind *grad_op1 = block->AllOps()[7]; + f::OpDesc *grad_op1 = block->AllOps()[7]; ASSERT_EQ(grad_op1->Type(), "rowwise_add_grad"); ASSERT_EQ(grad_op1->InputNames().size(), 1UL); ASSERT_EQ(grad_op1->OutputNames().size(), 2UL); @@ -885,19 +895,19 @@ TEST(Backward, shared_var) { } TEST(Backward, half_backward) { - f::ProgramDescBind program; - f::BlockDescBind *block = program.MutableBlock(0); + f::ProgramDesc program; + f::BlockDesc *block = program.MutableBlock(0); auto *op1 = block->AppendOp(); op1->SetType("minus"); op1->SetInput("X", {"a"}); op1->SetInput("Y", {"b"}); op1->SetOutput("Out", {"out"}); - auto target = f::VarDescBind("out"); + auto target = f::VarDesc("out"); target.SetShape({1}); size_t forward_len = block->AllOps().size(); auto var_to_grad = AppendBackward(program, target, {"b"}); - f::OpDescBind *fill_op = block->AllOps()[forward_len]; + f::OpDesc *fill_op = block->AllOps()[forward_len]; EXPECT_EQ(fill_op->Type(), "fill_constant"); auto ops = block->AllOps(); ASSERT_EQ(3UL, ops.size()); diff --git a/paddle/framework/block_desc.cc b/paddle/framework/block_desc.cc index 11764810e1d40e5e6eb3cd0d8e9b4b63a79855b4..0668b08ff7ab3c8ca4f1e989fc7af45a8ec5f63c 100644 --- a/paddle/framework/block_desc.cc +++ b/paddle/framework/block_desc.cc @@ -13,23 +13,24 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/framework/block_desc.h" +#include "paddle/framework/operator.h" #include "paddle/framework/program_desc.h" namespace paddle { namespace framework { -VarDescBind *BlockDescBind::Var(const std::string &name) { +VarDesc *BlockDesc::Var(const std::string &name) { auto it = vars_.find(name); if (it != vars_.end()) { return it->second.get(); } need_update_ = true; - auto *var = new VarDescBind(name); + auto *var = new VarDesc(name); vars_[name].reset(var); return var; } -VarDescBind *BlockDescBind::FindVar(const std::string &name) const { +VarDesc *BlockDesc::FindVar(const std::string &name) const { auto it = vars_.find(name); if (it == vars_.end()) { return nullptr; @@ -37,11 +38,13 @@ VarDescBind *BlockDescBind::FindVar(const std::string &name) const { return it->second.get(); } -bool BlockDescBind::HasVar(const std::string &name) const { +bool BlockDesc::HasVar(const std::string &name) const { return vars_.find(name) != vars_.end(); } -VarDescBind *BlockDescBind::FindVarRecursive(const std::string &name) const { +VarDesc *BlockDesc::FindVarRecursive(const std::string &name) const { + if (name == kEmptyVarName) return nullptr; + auto it = vars_.find(name); if (it == vars_.end()) { return Parent() == kNoneBlockIndex ? nullptr @@ -50,53 +53,67 @@ VarDescBind *BlockDescBind::FindVarRecursive(const std::string &name) const { return it->second.get(); } -VarDescBind *BlockDescBind::FindRecursiveOrCreateVar( - const std::string &name_bytes) { - VarDescBind *res = FindVarRecursive(name_bytes); +VarDesc *BlockDesc::FindRecursiveOrCreateVar(const std::string &name_bytes) { + VarDesc *res = FindVarRecursive(name_bytes); if (res == nullptr) { res = Var(name_bytes); } return res; } -bool BlockDescBind::HasVarRecursive(const std::string &name) const { +bool BlockDesc::HasVarRecursive(const std::string &name) const { return FindVarRecursive(name) != nullptr; } -std::vector BlockDescBind::AllVars() const { - std::vector res; +std::vector BlockDesc::AllVars() const { + std::vector res; for (const auto &p : vars_) { res.push_back(p.second.get()); } return res; } -OpDescBind *BlockDescBind::AppendOp() { +OpDesc *BlockDesc::AppendOp() { need_update_ = true; - ops_.emplace_back(new OpDescBind()); + ops_.emplace_back(new OpDesc()); return ops_.back().get(); } -void BlockDescBind::AppendAllocatedOp(std::unique_ptr &&op_desc) { +void BlockDesc::AppendAllocatedOp(std::unique_ptr &&op_desc) { need_update_ = true; ops_.emplace_back(std::move(op_desc)); } -OpDescBind *BlockDescBind::PrependOp() { +OpDesc *BlockDesc::PrependOp() { need_update_ = true; - ops_.emplace_front(new OpDescBind()); + ops_.emplace_front(new OpDesc()); return ops_.front().get(); } -std::vector BlockDescBind::AllOps() const { - std::vector res; +void BlockDesc::RemoveOp(size_t s, size_t e) { + if (ops_.begin() + s == ops_.end() || ops_.begin() + e == ops_.end()) { + return; + } + need_update_ = true; + for (auto it = ops_.begin() + s; it != ops_.begin() + e; it++) { + auto names = (*it)->InputArgumentNames(); + for (auto n : names) { + // TODO(typhoonzero): delete vars if no other op use it. + VLOG(3) << "deleting var " << n; + } + } + ops_.erase(ops_.begin() + s, ops_.begin() + e); +} + +std::vector BlockDesc::AllOps() const { + std::vector res; for (const auto &op : ops_) { res.push_back(op.get()); } return res; } -void BlockDescBind::Flush() { +void BlockDesc::Flush() { for (auto &op_desc : ops_) { op_desc->Flush(); } @@ -118,43 +135,43 @@ void BlockDescBind::Flush() { } } -BlockDescBind *BlockDescBind::ParentBlock() const { +BlockDesc *BlockDesc::ParentBlock() const { if (this->desc_->parent_idx() == kNoneBlockIndex) { return nullptr; } return prog_->MutableBlock(static_cast(this->desc_->parent_idx())); } -BlockDesc *BlockDescBind::Proto() { +proto::BlockDesc *BlockDesc::Proto() { Flush(); return desc_; } -BlockDescBind::BlockDescBind(ProgramDescBind *prog, BlockDesc *desc) +BlockDesc::BlockDesc(ProgramDesc *prog, proto::BlockDesc *desc) : prog_(prog), desc_(desc), need_update_(false) { - for (const VarDesc &var_desc : desc_->vars()) { - vars_[var_desc.name()].reset(new VarDescBind(var_desc)); + for (const proto::VarDesc &var_desc : desc_->vars()) { + vars_[var_desc.name()].reset(new VarDesc(var_desc)); } - for (const OpDesc &op_desc : desc_->ops()) { - ops_.emplace_back(new OpDescBind(op_desc, prog)); + for (const proto::OpDesc &op_desc : desc_->ops()) { + ops_.emplace_back(new OpDesc(op_desc, prog)); } } -BlockDescBind::BlockDescBind(const BlockDescBind &other, BlockDesc *desc, - ProgramDescBind *prog) +BlockDesc::BlockDesc(const BlockDesc &other, proto::BlockDesc *desc, + ProgramDesc *prog) : prog_(prog), desc_(desc) { need_update_ = true; for (auto &op : other.ops_) { - ops_.emplace_back(new OpDescBind(*op)); + ops_.emplace_back(new OpDesc(*op)); } for (auto &it : other.vars_) { - auto *var = new VarDescBind(*it.second); + auto *var = new VarDesc(*it.second); vars_[it.first].reset(var); } } -void BlockDescBind::ClearPBOps() { +void BlockDesc::ClearPBOps() { auto ops = this->desc_->mutable_ops(); while (!ops->empty()) { // we do not own the OpDesc, so release the ownership. @@ -162,7 +179,7 @@ void BlockDescBind::ClearPBOps() { } } -void BlockDescBind::ClearPBVars() { +void BlockDesc::ClearPBVars() { auto vars = this->desc_->mutable_vars(); while (!vars->empty()) { // we do not own the VarDesc, so release the ownership. diff --git a/paddle/framework/block_desc.h b/paddle/framework/block_desc.h index 8e967e5378eb47a7869efb59cc96a271f1cbb9a1..6c8c81b332d99e52db41018e117aa837be6745bc 100644 --- a/paddle/framework/block_desc.h +++ b/paddle/framework/block_desc.h @@ -28,20 +28,19 @@ limitations under the License. */ namespace paddle { namespace framework { -class ProgramDescBind; +class ProgramDesc; // Each Protobuf Message, we provide a XXXBind class. In that class, we optimize // read/write speed. Only when we want the protobuf message, the local changes // will be synchronized (by `Sync` method). -class BlockDescBind { +class BlockDesc { public: - BlockDescBind(ProgramDescBind *prog, BlockDesc *desc); + BlockDesc(ProgramDesc *prog, proto::BlockDesc *desc); - BlockDescBind(const BlockDescBind &other, BlockDesc *desc, - ProgramDescBind *prog); + BlockDesc(const BlockDesc &other, proto::BlockDesc *desc, ProgramDesc *prog); - ~BlockDescBind() { + ~BlockDesc() { this->ClearPBVars(); this->ClearPBOps(); } @@ -50,15 +49,15 @@ class BlockDescBind { int32_t Parent() const { return desc_->parent_idx(); } - VarDescBind *Var(const std::string &name_bytes); + VarDesc *Var(const std::string &name_bytes); - VarDescBind *FindVar(const std::string &name_bytes) const; + VarDesc *FindVar(const std::string &name_bytes) const; bool HasVar(const std::string &var_name) const; - VarDescBind *FindVarRecursive(const std::string &name_bytes) const; + VarDesc *FindVarRecursive(const std::string &name_bytes) const; - VarDescBind *FindRecursiveOrCreateVar(const std::string &name_bytes); + VarDesc *FindRecursiveOrCreateVar(const std::string &name_bytes); bool HasVarRecursive(const std::string &var_name) const; @@ -70,41 +69,43 @@ class BlockDescBind { return var_names; } - std::vector AllVars() const; + std::vector AllVars() const; - BlockDescBind *ParentBlock() const; + BlockDesc *ParentBlock() const; - OpDescBind *AppendOp(); + OpDesc *AppendOp(); - void AppendAllocatedOp(std::unique_ptr &&op_desc); + void AppendAllocatedOp(std::unique_ptr &&op_desc); - OpDescBind *PrependOp(); + OpDesc *PrependOp(); - std::vector AllOps() const; + void RemoveOp(size_t s, size_t e); + + std::vector AllOps() const; size_t OpSize() const { return ops_.size(); } - OpDescBind *Op(int idx) { return ops_.at(idx).get(); } + OpDesc *Op(int idx) { return ops_.at(idx).get(); } void Flush(); - BlockDesc *Proto(); + proto::BlockDesc *Proto(); - ProgramDescBind *Program() { return this->prog_; } + ProgramDesc *Program() { return this->prog_; } private: void ClearPBOps(); void ClearPBVars(); private: - ProgramDescBind *prog_; // not_own - BlockDesc *desc_; // not_own + ProgramDesc *prog_; // not_own + proto::BlockDesc *desc_; // not_own bool need_update_; - std::deque> ops_; - std::unordered_map> vars_; + std::deque> ops_; + std::unordered_map> vars_; - DISABLE_COPY_AND_ASSIGN(BlockDescBind); + DISABLE_COPY_AND_ASSIGN(BlockDesc); }; } // namespace framework } // namespace paddle diff --git a/paddle/framework/data_layout.h b/paddle/framework/data_layout.h new file mode 100644 index 0000000000000000000000000000000000000000..4a8669c3a41fceaad26878a79eabfd0affce86fd --- /dev/null +++ b/paddle/framework/data_layout.h @@ -0,0 +1,59 @@ +/* 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 +#include "paddle/platform/enforce.h" + +#include +#include "paddle/platform/enforce.h" + +namespace paddle { +namespace framework { + +enum class DataLayout { + kNHWC = 0, + kNCHW = 1, + kAnyLayout = 2, +}; + +inline DataLayout StringToDataLayout(const std::string& str) { + if (str == "NHWC" || str == "nhwc") { + return DataLayout::kNHWC; + } else if (str == "NCHW" || str == "nchw") { + return DataLayout::kNCHW; + } else { + PADDLE_THROW("Unknown storage order string: %s", str); + } +} + +inline std::string DataLayoutToString(const DataLayout& data_layout) { + switch (data_layout) { + case DataLayout::kNHWC: + return "NHWC"; + case DataLayout::kNCHW: + return "NCHW"; + case DataLayout::kAnyLayout: + return "ANY_LAYOUT"; + default: + PADDLE_THROW("unknown DataLayou %d", data_layout); + } +} + +inline std::ostream& operator<<(std::ostream& out, DataLayout l) { + out << DataLayoutToString(l); + return out; +} + +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/data_transform.cc b/paddle/framework/data_transform.cc new file mode 100644 index 0000000000000000000000000000000000000000..58780e386353845e380590f20540314dcd021649 --- /dev/null +++ b/paddle/framework/data_transform.cc @@ -0,0 +1,106 @@ +/* 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. */ + +#include "paddle/framework/data_transform.h" +#include "paddle/framework/lod_tensor.h" +#include "paddle/platform/device_context.h" + +namespace paddle { +namespace framework { + +DataTransformFnMap& DataTransformFnMap::Instance() { + static DataTransformFnMap data_transform_map; + return data_transform_map; +} + +auto KernelFP32 = OpKernelType(proto::DataType::FP32, platform::CPUPlace(), + DataLayout::kNHWC, LibraryType::kPlain); + +auto KernelFP64 = OpKernelType(proto::DataType::FP64, platform::CPUPlace(), + DataLayout::kNHWC, LibraryType::kPlain); + +auto KernelNHWC = OpKernelType(proto::DataType::FP64, platform::CPUPlace(), + DataLayout::kNHWC, LibraryType::kPlain); + +auto KernelNCHW = OpKernelType(proto::DataType::FP64, platform::CPUPlace(), + DataLayout::kNCHW, LibraryType::kPlain); + +void TransDataType(const platform::DeviceContext* ctx, + const KernelTypePair& kernel_pair, const Variable& in, + Variable* out) { + PADDLE_ENFORCE(in.IsType(), "Only Support Tensor transform!."); + PADDLE_ENFORCE( + platform::places_are_same_class(kernel_pair.first.place_, + kernel_pair.second.place_), + "TransDataType Only Support DataType transform on same place!"); + + auto src = in.Get(); + auto* dst = out->GetMutable(); + + auto dims = src.dims(); + dst->Resize(dims); + auto dst_type = kernel_pair.second.data_type_; + auto src_type = kernel_pair.first.data_type_; + + switch (src_type) { + case proto::DataType::FP32: + framework::VisitDataType(dst_type, CastDataType(src, dst, ctx)); + break; + case proto::DataType::FP64: + framework::VisitDataType(dst_type, CastDataType(src, dst, ctx)); + break; + case proto::DataType::INT32: + framework::VisitDataType(dst_type, CastDataType(src, dst, ctx)); + break; + case proto::DataType::INT64: + framework::VisitDataType(dst_type, CastDataType(src, dst, ctx)); + break; + case proto::DataType::BOOL: + framework::VisitDataType(dst_type, CastDataType(src, dst, ctx)); + break; + default: + PADDLE_THROW("Not support type %d", src_type); + } +} + +void TransDataLayout(const platform::DeviceContext* ctx, + const KernelTypePair& kernel_pair, const Variable& in, + Variable* out) { + PADDLE_ENFORCE(in.IsType(), "Only Support Tensor transform!."); + PADDLE_ENFORCE( + platform::places_are_same_class(kernel_pair.first.place_, + kernel_pair.second.place_), + "TransDataType Only Support DataType transform on same place!"); + + auto src = in.Get(); + auto* dst = out->GetMutable(); + PADDLE_ENFORCE(arity(src.dims()) == 4, "Input Arity Only Suppport 4!"); + + dst->Resize(src.dims()); + auto place = kernel_pair.second.place_; + CopyFrom(src, place, *ctx, dst); + const std::vector axis = {0, 2, 3, 1}; + + auto src_type = kernel_pair.first.data_type_; + framework::VisitDataType(src_type, CastDataLayout(src, dst, ctx, axis)); + + dst->set_layout(kernel_pair.second.data_layout_); +} + +} // namespace framework +} // namespace paddle + +namespace f = paddle::framework; +REGISTER_DATA_TRANSFORM_FN(f::KernelFP32, f::KernelFP64, f::TransDataType); +REGISTER_DATA_TRANSFORM_FN(f::KernelNHWC, f::KernelNCHW, f::TransDataLayout); diff --git a/paddle/framework/data_transform.h b/paddle/framework/data_transform.h new file mode 100644 index 0000000000000000000000000000000000000000..9abb3c99bf30fcf9deab59dc7ee9c02e7c7c775b --- /dev/null +++ b/paddle/framework/data_transform.h @@ -0,0 +1,171 @@ +/* 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 + +#include +#include +#include + +#include "paddle/framework/op_kernel_type.h" +#include "paddle/framework/tensor.h" +#include "paddle/framework/variable.h" +#include "paddle/operators/math/math_function.h" +#include "paddle/platform/device_context.h" +#include "paddle/platform/macros.h" +#include "paddle/platform/transform.h" + +namespace paddle { +namespace framework { + +using KernelTypePair = std::pair; + +using DataTransformFn = + std::function; + +struct KernelTypePairHash { + static void HashCombine(const OpKernelType& t, std::size_t* seed) { + OpKernelType::Hash kernel_type_hasher; + (*seed) ^= kernel_type_hasher(t) + 0x9e3779b9 + (*seed << 6) + (*seed >> 2); + } + + size_t operator()(const KernelTypePair& kernel_pair) const { + std::size_t seed = 0; + HashCombine(kernel_pair.first, &seed); + HashCombine(kernel_pair.second, &seed); + return seed; + } +}; + +template +struct CastDataTypeFunctor { + HOSTDEVICE inline OutType operator()(InType in) const { + return static_cast(in); + } +}; + +template +struct CastDataType { + CastDataType(const framework::Tensor& in, framework::Tensor* out, + const platform::DeviceContext* ctx) + : in_(in), out_(out), ctx_(ctx) {} + const framework::Tensor in_; + framework::Tensor* out_; + const platform::DeviceContext* ctx_; + + template + void operator()() { + auto place = ctx_->GetPlace(); + + auto* in_begin = in_.data(); + auto numel = in_.numel(); + auto* in_end = in_begin + numel; + auto* out_begin = out_->mutable_data(place); + if (platform::is_cpu_place(place)) { + platform::Transform trans; + auto* context = static_cast(ctx_); + trans(*context, in_begin, in_end, out_begin, + CastDataTypeFunctor()); + } else { + // TODO(dzhwinter): enhance CopyFrom CPU<->GPU with different data type? + PADDLE_THROW("Unsupport CPU <-> GPU!"); + } + } +}; + +struct CastDataLayout { + CastDataLayout(const framework::Tensor& in, framework::Tensor* out, + const platform::DeviceContext* ctx, + const std::vector& axis) + : in_(in), out_(out), ctx_(ctx), axis_(axis) {} + const framework::Tensor in_; + framework::Tensor* out_; + const platform::DeviceContext* ctx_; + const std::vector axis_; + + template + void operator()() { + auto place = ctx_->GetPlace(); + if (platform::is_cpu_place(place)) { + operators::math::Transpose trans4; + auto* context = static_cast(ctx_); + trans4(*context, in_, out_, axis_); + } else { + PADDLE_THROW("Unsupport CPU <-> GPU!"); + } + } +}; + +using DataTransformMap = + std::unordered_map; + +class DataTransformFnMap { + public: + static DataTransformFnMap& Instance(); + + bool Has(const KernelTypePair& key_pair) const { + return map_.find(key_pair) != map_.end(); + } + + void Insert(const OpKernelType& left, const OpKernelType& right, + const DataTransformFn& data_tranform_fn) { + Insert(std::make_pair(left, right), data_tranform_fn); + } + + void Insert(const KernelTypePair& kernel_type_pair, + const DataTransformFn& data_tranform_fn) { + PADDLE_ENFORCE(!Has(kernel_type_pair), + "KernelTypePair %s has been registered", ""); + map_.insert({kernel_type_pair, data_tranform_fn}); + } + + const DataTransformFn& Get(const KernelTypePair& key_pair) const { + auto data_transformer = GetNullable(key_pair); + PADDLE_ENFORCE_NOT_NULL(data_transformer, + "DataTransformFn should not be NULL"); + return *data_transformer; + } + + const DataTransformFn* GetNullable(const KernelTypePair& key_pair) const { + auto it = map_.find(key_pair); + if (it == map_.end()) { + return nullptr; + } else { + return &(it->second); + } + } + + const DataTransformMap& Map() const { return map_; } + + private: + DataTransformFnMap() = default; + DataTransformMap map_; + DISABLE_COPY_AND_ASSIGN(DataTransformFnMap); +}; + +// generate unique name with __LINE__ +// refs https://stackoverflow.com/questions/1597007 +#define TOKENPASTE(x, y) x##y +#define TOKENPASTE2(x, y) TOKENPASTE(x, y) +#define REGISTER_DATA_TRANSFORM_FN(from, to, fn) \ + static int TOKENPASTE2(fn_, __LINE__)() { \ + ::paddle::framework::DataTransformFnMap::Instance().Insert(from, to, fn); \ + return 0; \ + } \ + static int TOKENPASTE2(var_, __LINE__) __attribute__((unused)) = \ + TOKENPASTE2(fn_, __LINE__)() + +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/data_transform_test.cc b/paddle/framework/data_transform_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..5b01c8434b155fcb15f4bfb7d598c40bbac87d7b --- /dev/null +++ b/paddle/framework/data_transform_test.cc @@ -0,0 +1,156 @@ +/* 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. */ +#include +#include + +#include + +#include "paddle/framework/data_transform.h" +#include "paddle/platform/device_context.h" + +namespace paddle { +namespace framework { +using namespace platform; + +/** + * @brief cross validation of different kernel type transform + * We use four bit map represent different combination. + * If the field has multiple possible value, only choose two of them. + * For DataType, only test the FP32(float), FP64(double). + * e.g. 0000 -> FP32, CPUPlace, kNHWC, kPlain + * 1111 -> FP64, GPUPlace, kNCHW, kMKLDNN + */ + +std::array kDataType = {proto::DataType::FP32, + proto::DataType::FP64}; + +std::array kPlace = {CPUPlace(), CUDAPlace(0)}; + +std::array kDataLayout = { + DataLayout::kNHWC, DataLayout::kNCHW, +}; + +std::array kLibraryType = { + LibraryType::kPlain, LibraryType::kMKLDNN, +}; + +OpKernelType GenFromBit(const std::vector bits) { + return OpKernelType(kDataType[bits[0]], kPlace[bits[1]], kDataLayout[bits[2]], + kLibraryType[bits[3]]); +} + +int test_value = 0; + +auto kernel0 = GenFromBit({0, 0, 0, 0}); +auto kernel1 = GenFromBit({0, 0, 0, 1}); +auto kernel2 = GenFromBit({0, 0, 1, 0}); +auto kernel3 = GenFromBit({0, 0, 1, 1}); + +void TransDataType_t(const platform::DeviceContext* ctx, + const KernelTypePair& p, const Variable& in, + Variable* out) { + test_value++; +} + +void TransDataLayout_t(const platform::DeviceContext* ctx, + const KernelTypePair& p, const Variable& in, + Variable* out) { + test_value--; +} + +void TransLibraryType_t(const platform::DeviceContext* ctx, + const KernelTypePair& p, const Variable& in, + Variable* out) { + test_value += 2; +} + +} // namespace framework +} // namespace paddle + +namespace frw = paddle::framework; + +REGISTER_DATA_TRANSFORM_FN(frw::kernel0, frw::kernel1, frw::TransDataType_t); +REGISTER_DATA_TRANSFORM_FN(frw::kernel1, frw::kernel2, frw::TransDataLayout_t); +REGISTER_DATA_TRANSFORM_FN(frw::kernel0, frw::kernel2, frw::TransLibraryType_t); + +TEST(DataTransform, Register) { + using namespace paddle::framework; + using namespace paddle::platform; + + auto& instance = DataTransformFnMap::Instance(); + paddle::framework::Variable in; + paddle::framework::Variable out; + + DeviceContext* ctx = new CPUDeviceContext(); + auto pair0 = std::make_pair(frw::kernel0, frw::kernel1); + instance.Get(pair0)(ctx, pair0, in, &out); + ASSERT_EQ(test_value, 1); + + auto pair1 = std::make_pair(frw::kernel1, frw::kernel2); + instance.Get(pair1)(ctx, pair1, in, &out); + ASSERT_EQ(test_value, 0); + + auto pair3 = std::make_pair(frw::kernel0, frw::kernel2); + instance.Get(pair3)(ctx, pair3, in, &out); + ASSERT_EQ(test_value, 2); +} + +TEST(DataTransform, Layout) { + using namespace paddle::framework; + using namespace paddle::platform; + + auto& instance = DataTransformFnMap::Instance(); + Variable in; + Variable out; + Tensor* src = in.GetMutable(); + src->mutable_data(make_ddim({2, 3, 1, 2}), CPUPlace()); + src->set_layout(DataLayout::kNHWC); + + DeviceContext* ctx = new CPUDeviceContext(); + + { + auto kernel1 = GenFromBit({1, 0, 0, 0}); + auto kernel2 = GenFromBit({1, 0, 1, 0}); + auto pair0 = std::make_pair(kernel1, kernel2); + instance.Get(pair0)(ctx, pair0, in, &out); + } + + Tensor dst = out.Get(); + EXPECT_TRUE(dst.layout() != src->layout()); +} + +TEST(DataTransform, DataType) { + using namespace paddle::framework; + using namespace paddle::platform; + + auto& instance = DataTransformFnMap::Instance(); + DeviceContext* ctx = new CPUDeviceContext(); + + Variable in; + Variable out; + Tensor* src = in.GetMutable(); + float* ptr = src->mutable_data(make_ddim({2, 3}), CPUPlace()); + for (int i = 0; i < 6; ++i) { + ptr[i] = i / 3; + } + + { + auto kernel1 = GenFromBit({0, 0, 0, 0}); + auto kernel2 = GenFromBit({1, 0, 0, 0}); + auto pair0 = std::make_pair(kernel1, kernel2); + instance.Get(pair0)(ctx, pair0, in, &out); + } + Tensor dst = out.Get(); + EXPECT_TRUE(dst.data() != nullptr); +} diff --git a/paddle/framework/data_type.h b/paddle/framework/data_type.h index c54d2d4ddf09c445fb25c1fbe8a7498f233d8212..6a372ac32e48131eed28e2d42125feb5b92a11c7 100644 --- a/paddle/framework/data_type.h +++ b/paddle/framework/data_type.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 @@ -20,7 +20,8 @@ namespace paddle { namespace framework { -inline DataType ToDataType(std::type_index type) { +inline proto::DataType ToDataType(std::type_index type) { + using namespace paddle::framework::proto; if (typeid(float).hash_code() == type.hash_code()) { return DataType::FP32; } else if (typeid(double).hash_code() == type.hash_code()) { @@ -36,7 +37,8 @@ inline DataType ToDataType(std::type_index type) { } } -inline std::type_index ToTypeIndex(DataType type) { +inline std::type_index ToTypeIndex(proto::DataType type) { + using namespace paddle::framework::proto; switch (type) { case DataType::FP32: return typeid(float); @@ -54,7 +56,8 @@ inline std::type_index ToTypeIndex(DataType type) { } template -inline void VisitDataType(DataType type, Visitor visitor) { +inline void VisitDataType(proto::DataType type, Visitor visitor) { + using namespace paddle::framework::proto; switch (type) { case DataType::FP32: visitor.template operator()(); diff --git a/paddle/framework/ddim_test.cc b/paddle/framework/ddim_test.cc index 756232b1b56a49d2c91cc2cac950ca508c54fb3f..bc259d1f603fb34ac8546c388669d8c5c1250bd1 100644 --- a/paddle/framework/ddim_test.cc +++ b/paddle/framework/ddim_test.cc @@ -1,3 +1,16 @@ +/* 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. */ #include #include diff --git a/paddle/framework/details/op_registry.h b/paddle/framework/details/op_registry.h index f91e0e03410c95f84a65f02beed38b7bbfdcaa86..6d50e820b2b625f932768d2ca671d999071f1ca6 100644 --- a/paddle/framework/details/op_registry.h +++ b/paddle/framework/details/op_registry.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 @@ -90,7 +90,7 @@ struct OpInfoFiller { template struct OpInfoFiller { void operator()(const char* op_type, OpInfo* info) const { - info->proto_ = new OpProto; + info->proto_ = new proto::OpProto; info->checker_ = new OpAttrChecker(); auto maker = T(info->proto_, info->checker_); maker.Validate(); @@ -106,10 +106,10 @@ template struct OpInfoFiller { void operator()(const char* op_type, OpInfo* info) const { info->grad_op_maker_ = []( - const OpDescBind& fwd_op, + const OpDesc& fwd_op, const std::unordered_set& no_grad_set, std::unordered_map* grad_to_var, - const std::vector& grad_block) { + const std::vector& grad_block) { T maker(fwd_op, no_grad_set, grad_to_var, grad_block); return maker(); }; @@ -119,7 +119,7 @@ struct OpInfoFiller { template struct OpInfoFiller { void operator()(const char* op_type, OpInfo* info) const { - info->infer_var_type_ = [](const OpDescBind& fwd_op, BlockDescBind* block) { + info->infer_var_type_ = [](const OpDesc& fwd_op, BlockDesc* block) { T inference; inference(fwd_op, block); }; diff --git a/paddle/framework/executor.cc b/paddle/framework/executor.cc index adedd8cb0e8504fd6fc924e62a2ede3c1c7ce698..bf1f0471ccbfccf13cb6f74c8088da7acd68ec0b 100644 --- a/paddle/framework/executor.cc +++ b/paddle/framework/executor.cc @@ -14,18 +14,17 @@ limitations under the License. */ #include "paddle/framework/executor.h" -#include -#include -#include #include -#include +#include "gflags/gflags.h" #include "paddle/framework/feed_fetch_type.h" #include "paddle/framework/lod_rank_table.h" -#include "paddle/framework/lod_tensor.h" #include "paddle/framework/lod_tensor_array.h" #include "paddle/framework/op_registry.h" -#include "paddle/framework/scope.h" + +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 { @@ -33,48 +32,22 @@ namespace framework { const std::string kFeedOpType = "feed"; const std::string kFetchOpType = "fetch"; -Executor::Executor(const std::vector& places) : own_(true) { - PADDLE_ENFORCE_GT(places.size(), 0); - device_contexts_.resize(places.size()); - for (size_t i = 0; i < places.size(); i++) { - if (platform::is_cpu_place(places[i])) { - device_contexts_[i] = new platform::CPUDeviceContext( - boost::get(places[i])); - } else if (platform::is_gpu_place(places[i])) { -#ifdef PADDLE_WITH_CUDA - device_contexts_[i] = new platform::CUDADeviceContext( - boost::get(places[i])); -#else - PADDLE_THROW( - "'GPUPlace' is not supported, Please re-compile with WITH_GPU " - "option"); -#endif - } - } -} +Executor::Executor(const platform::Place& place) : place_(place) {} -Executor::~Executor() { - if (own_) { - for (auto& device_context : device_contexts_) { - delete device_context; - } - } -} - -static void CreateTensor(Variable* var, VarDesc::VarType var_type) { - if (var_type == VarDesc::LOD_TENSOR) { +static void CreateTensor(Variable* var, proto::VarDesc::VarType var_type) { + if (var_type == proto::VarDesc::LOD_TENSOR) { var->GetMutable(); - } else if (var_type == VarDesc::SELECTED_ROWS) { + } else if (var_type == proto::VarDesc::SELECTED_ROWS) { var->GetMutable(); - } else if (var_type == VarDesc::FEED_MINIBATCH) { + } else if (var_type == proto::VarDesc::FEED_MINIBATCH) { var->GetMutable(); - } else if (var_type == VarDesc::FETCH_LIST) { + } else if (var_type == proto::VarDesc::FETCH_LIST) { var->GetMutable(); - } else if (var_type == VarDesc::STEP_SCOPES) { + } else if (var_type == proto::VarDesc::STEP_SCOPES) { var->GetMutable>(); - } else if (var_type == VarDesc::LOD_RANK_TABLE) { + } else if (var_type == proto::VarDesc::LOD_RANK_TABLE) { var->GetMutable(); - } else if (var_type == VarDesc::LOD_TENSOR_ARRAY) { + } else if (var_type == proto::VarDesc::LOD_TENSOR_ARRAY) { var->GetMutable(); } else { PADDLE_THROW( @@ -84,52 +57,76 @@ static void CreateTensor(Variable* var, VarDesc::VarType var_type) { } } -void Executor::Run(const ProgramDescBind& pdesc, Scope* scope, int block_id, - bool create_local_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() && + tensor.type().hash_code() != typeid(double).hash_code()) { + return; + } + PADDLE_ENFORCE(!framework::HasInf(tensor), "Tensor %s has Inf", name); + PADDLE_ENFORCE(!framework::HasNAN(tensor), "Tensor %s has NAN", name); +} + +void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id, + bool create_local_scope, bool create_vars) { // TODO(tonyyang-svail): // - only runs on the first device (i.e. no interdevice communication) // - will change to use multiple blocks for RNN op and Cond Op PADDLE_ENFORCE_LT(static_cast(block_id), pdesc.Size()); auto& block = pdesc.Block(block_id); - auto& device = device_contexts_[0]; Scope* local_scope = scope; - if (create_local_scope) { - local_scope = &scope->NewScope(); - for (auto& var : block.AllVars()) { - if (var->Persistable()) { - auto* ptr = scope->Var(var->Name()); - CreateTensor(ptr, var->GetType()); - VLOG(3) << "Create Variable " << var->Name() - << " global, which pointer is " << ptr; - } else { + if (create_vars) { + if (create_local_scope) { + local_scope = &scope->NewScope(); + for (auto& var : block.AllVars()) { + if (var->Name() == framework::kEmptyVarName) { + continue; + } + + if (var->Persistable()) { + auto* ptr = scope->Var(var->Name()); + CreateTensor(ptr, var->GetType()); + VLOG(3) << "Create Variable " << var->Name() + << " global, which pointer is " << ptr; + } else { + auto* ptr = local_scope->Var(var->Name()); + CreateTensor(ptr, var->GetType()); + VLOG(3) << "Create Variable " << var->Name() + << " locally, which pointer is " << ptr; + } + } + } else { + for (auto& var : block.AllVars()) { auto* ptr = local_scope->Var(var->Name()); CreateTensor(ptr, var->GetType()); - VLOG(3) << "Create Variable " << var->Name() - << " locally, which pointer is " << ptr; + VLOG(3) << "Create variable " << var->Name() << ", which pointer is " + << ptr; } - } - } else { - for (auto& var : block.AllVars()) { - auto* ptr = local_scope->Var(var->Name()); - CreateTensor(ptr, var->GetType()); - VLOG(3) << "Create variable " << var->Name() << ", which pointer is " - << ptr; - } - } + } // if (create_local_scope) + } // if (create_vars) for (auto& op_desc : block.AllOps()) { auto op = paddle::framework::OpRegistry::CreateOp(*op_desc); - VLOG(10) << op->DebugString(); - op->Run(*local_scope, *device); + VLOG(3) << op->DebugString(); + op->Run(*local_scope, 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()); + } + } + } } - if (create_local_scope) { + if (create_vars && create_local_scope) { scope->DeleteScope(local_scope); } } -Executor::Executor(const platform::DeviceContext& device) - : device_contexts_({&device}), own_(false) {} - } // namespace framework } // namespace paddle diff --git a/paddle/framework/executor.h b/paddle/framework/executor.h index b745f4f6474ef688774f4c833a3958942e9aa8cb..d869e18901b82959a40cc296aa0844c20ea63ac1 100644 --- a/paddle/framework/executor.h +++ b/paddle/framework/executor.h @@ -18,15 +18,18 @@ limitations under the License. */ #include "paddle/framework/program_desc.h" #include "paddle/framework/scope.h" #include "paddle/framework/tensor.h" +#include "paddle/platform/device_context.h" namespace paddle { namespace framework { class Executor { public: - explicit Executor(const std::vector& places); - explicit Executor(const platform::DeviceContext& devices); - ~Executor(); + // TODO(dzhwinter) : Do not rely on this function, it will be removed + explicit Executor(const platform::DeviceContext& device) + : Executor(device.GetPlace()) {} + + explicit Executor(const platform::Place& place); /* @Brief * Runtime evaluation of the given ProgramDesc under certain Scope @@ -35,11 +38,11 @@ class Executor { * ProgramDesc * Scope */ - void Run(const ProgramDescBind&, Scope*, int, bool create_local_scope = true); + void Run(const ProgramDesc&, Scope*, int, bool create_local_scope = true, + bool create_vars = true); private: - std::vector device_contexts_; - bool own_; + const platform::Place place_; }; } // namespace framework diff --git a/paddle/framework/feed_fetch_type.h b/paddle/framework/feed_fetch_type.h index bc4ae440fc708f696c18bb9d5ab3ba7dd59e21ab..9bc4a90c44828ecb7458d524f59609f01848cc5c 100644 --- a/paddle/framework/feed_fetch_type.h +++ b/paddle/framework/feed_fetch_type.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 diff --git a/paddle/framework/framework.proto b/paddle/framework/framework.proto index f1fc4529e15502927560eefd74110f6ca7eab4a9..4f2746e4b86ee5fe095897ff6ef9d3f6473e8a14 100644 --- a/paddle/framework/framework.proto +++ b/paddle/framework/framework.proto @@ -14,7 +14,7 @@ limitations under the License. */ syntax = "proto2"; option optimize_for = LITE_RUNTIME; -package paddle.framework; +package paddle.framework.proto; enum AttrType { INT = 0; diff --git a/paddle/framework/grad_op_desc_maker.h b/paddle/framework/grad_op_desc_maker.h index 998186e33915a11f2864eb5387d19ed1bfbab51c..2de5242831835b47893a5825e5532500ad5ec3f9 100644 --- a/paddle/framework/grad_op_desc_maker.h +++ b/paddle/framework/grad_op_desc_maker.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 @@ -22,21 +22,27 @@ namespace paddle { namespace framework { +/* + This functor class is responsible for creating the gradient ops for the given + operator fwd_op. After it is called (through operator()), the pairs of + (gradient variable, corresponding input variable of fwd_op) will be added to + grad_to_var. If an input variable of fwd_op is contained in no_grad_set, its + gradient varialbe will be ignored or kEmptyVarName depending on the template + argument DropEmptyIG in the derived classes. + */ class GradOpDescMakerBase { public: explicit GradOpDescMakerBase( - const OpDescBind& fwd_op, - const std::unordered_set& no_grad_set, + const OpDesc& fwd_op, const std::unordered_set& no_grad_set, std::unordered_map* grad_to_var, - const std::vector& grad_block = - std::vector()) + const std::vector& grad_block = std::vector()) : fwd_op_(fwd_op), no_grad_set_(no_grad_set), grad_to_var_(grad_to_var), grad_block_(grad_block) {} virtual ~GradOpDescMakerBase() = default; - virtual std::vector> operator()() const = 0; + virtual std::vector> operator()() const = 0; protected: std::vector InputGrad(const std::string& name, @@ -58,6 +64,16 @@ class GradOpDescMakerBase { if (!drop_empty_grad) { return ret_val; } + PADDLE_ENFORCE_LE(var_names.size(), 1UL, + "BUG from operator developer:" + " for input argument with a list of variables, " + " drop_empty_grad is not allowed because it makes" + " the correspondence bewteen a variable and its gradient" + " ambiguous. Use REGISTER_OP_EX to register the op" + " or call InputGrad(?,false) in GradOpDescMaker." + " Op type %s", + fwd_op_.Type()); + std::vector dropped_ret_val; dropped_ret_val.reserve(ret_val.size()); std::copy_if(ret_val.begin(), ret_val.end(), @@ -105,26 +121,26 @@ class GradOpDescMakerBase { std::string ForwardOpType() const { return this->fwd_op_.Type(); } private: - const OpDescBind& fwd_op_; + const OpDesc& fwd_op_; const std::unordered_set& no_grad_set_; std::unordered_map* grad_to_var_; protected: - std::vector grad_block_; + std::vector grad_block_; }; class SingleGradOpDescMaker : public GradOpDescMakerBase { public: using GradOpDescMakerBase::GradOpDescMakerBase; - std::vector> operator()() const { - std::vector> retv; + std::vector> operator()() const { + std::vector> retv; retv.emplace_back(this->Apply()); return retv; } protected: - virtual std::unique_ptr Apply() const = 0; + virtual std::unique_ptr Apply() const = 0; }; template @@ -133,8 +149,8 @@ class DefaultGradOpDescMaker : public SingleGradOpDescMaker { using SingleGradOpDescMaker::SingleGradOpDescMaker; protected: - virtual std::unique_ptr Apply() const { - auto* grad = new OpDescBind(); + virtual std::unique_ptr Apply() const { + auto* grad = new OpDesc(); grad->SetType(this->GradOpType()); for (auto& input_param : this->InputNames()) { @@ -150,7 +166,7 @@ class DefaultGradOpDescMaker : public SingleGradOpDescMaker { grad->SetAttrMap(this->Attrs()); - return std::unique_ptr(grad); + return std::unique_ptr(grad); } virtual std::string GradOpType() const { @@ -161,7 +177,7 @@ class DefaultGradOpDescMaker : public SingleGradOpDescMaker { class EmptyGradOpMaker : public GradOpDescMakerBase { public: using GradOpDescMakerBase::GradOpDescMakerBase; - std::vector> operator()() const override { + std::vector> operator()() const override { return {}; } }; diff --git a/paddle/framework/init.cc b/paddle/framework/init.cc new file mode 100644 index 0000000000000000000000000000000000000000..682cff168d4d31e0565fc987604f97a671566fbd --- /dev/null +++ b/paddle/framework/init.cc @@ -0,0 +1,79 @@ +/* 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. */ +#include +#include + +#include "paddle/framework/init.h" +#include "paddle/platform/device_context.h" +#include "paddle/platform/place.h" +#include "paddle/string/piece.h" + +namespace paddle { +namespace framework { + +std::once_flag gflags_init_flag; + +// TODO(qijun) move init gflags to init.cc +void InitGflags(std::vector &argv) { + std::call_once(gflags_init_flag, [&]() { + int argc = argv.size(); + char **arr = new char *[argv.size()]; + std::string line; + for (size_t i = 0; i < argv.size(); i++) { + arr[i] = &argv[i][0]; + line += argv[i]; + line += ' '; + } + google::ParseCommandLineFlags(&argc, &arr, true); + VLOG(1) << "Init commandline: " << line; + }); +} + +bool InitDevices(const std::vector &devices) { + // device format + // CPU + // GPU:1 + // TODO(dzhwinter) : add device format annotation for users. + std::vector places; + for (auto &device : devices) { + auto p = string::Piece(device); + if (string::HasPrefix(p, "CPU")) { + places.emplace_back(platform::CPUPlace()); + } else if (string::HasPrefix(p, "GPU")) { +#ifdef PADDLE_WITH_CUDA + auto pos = string::RFind(p, ':', string::Piece::npos); + auto number = device.substr(pos + 1); + places.emplace_back(platform::CUDAPlace(std::stoi(number))); +#else + LOG(WARNING) + << "'GPU' is not supported, Please re-compile with WITH_GPU option"; +#endif + } else { + return false; + } + } + + if (std::find_if(places.begin(), places.end(), + [&](const platform::Place &place) { + return platform::is_cpu_place(place); + }) == places.end()) { + places.emplace_back(platform::CPUPlace()); + LOG(WARNING) << "Not specified CPU device, create CPU by Default."; + } + platform::DeviceContextPool::Init(places); + return true; +} + +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/init.h b/paddle/framework/init.h new file mode 100644 index 0000000000000000000000000000000000000000..33907f9eb00fb3469b53dcf8151557cc7a2d3791 --- /dev/null +++ b/paddle/framework/init.h @@ -0,0 +1,28 @@ +/* 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 +#include + +#include "gflags/gflags.h" +#include "glog/logging.h" + +namespace paddle { +namespace framework { + +void InitGflags(std::vector &argv); + +bool InitDevices(const std::vector &devices); + +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/init_test.cc b/paddle/framework/init_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..f0788051d4855a175d2d7ea1f1a0805c776c462b --- /dev/null +++ b/paddle/framework/init_test.cc @@ -0,0 +1,31 @@ +/* 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. */ +#include "gtest/gtest.h" + +#include "paddle/framework/init.h" + +TEST(Init, InitDevices) { + using paddle::framework::InitDevices; + std::vector ds1 = {"CPU"}; + ASSERT_EQ(InitDevices(ds1), true); + +#ifdef PADDLE_WITH_CUDA + std::vector ds2 = {"CPU", "GPU:0", "GPU:1"}; + ASSERT_EQ(InitDevices(ds2), true); + + // test re-init + std::vector ds3 = {"GPU:0", "GPU:1"}; + ASSERT_EQ(InitDevices(ds3), true); +#endif +} diff --git a/paddle/framework/library_type.h b/paddle/framework/library_type.h new file mode 100644 index 0000000000000000000000000000000000000000..7707799cae8c4edc304cd81725270a85f01fd28d --- /dev/null +++ b/paddle/framework/library_type.h @@ -0,0 +1,67 @@ +/* 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 + +namespace paddle { +namespace framework { + +// For more details about the design of LibraryType, Please refer to +// https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/operator_kernel_type.md#library + +enum class LibraryType { + kPlain = 0, + kMKLDNN = 1, + kCUDNN = 2, +}; + +inline std::string LibraryTypeToString(const LibraryType& library_type) { + switch (library_type) { + case LibraryType::kPlain: + return "PLAIN"; + case LibraryType::kMKLDNN: + return "MKLDNN"; + case LibraryType::kCUDNN: + return "CUDNN"; + default: + PADDLE_THROW("unknown LibraryType %d", static_cast(library_type)); + } +} + +inline LibraryType StringToLibraryType(const char* ctype) { + std::string s(ctype); + if (s == std::string("PLAIN")) { + return LibraryType::kPlain; + } else if (s == std::string("MKLDNN")) { + return LibraryType::kMKLDNN; + } else if (s == std::string("CUDNN")) { + return LibraryType::kCUDNN; + // To be compatible with register macro. + // CPU, CUDA, PLAIN are same library type. + } else if (s == std::string("CPU")) { + return LibraryType::kPlain; + } else if (s == std::string("CUDA")) { + return LibraryType::kPlain; + } else { + PADDLE_THROW("Unknown LibraryType %s", s.c_str()); + } +} + +inline std::ostream& operator<<(std::ostream& out, LibraryType l) { + out << LibraryTypeToString(l); + return out; +} + +} // namespace +} // framework diff --git a/paddle/framework/lod_rank_table.cc b/paddle/framework/lod_rank_table.cc index 1c2fba70c8ab0827ba6d1563f08cd0820650822e..704bce2a0eb60b974efd41a4edda0af2933da825 100644 --- a/paddle/framework/lod_rank_table.cc +++ b/paddle/framework/lod_rank_table.cc @@ -1,16 +1,16 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +/* 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 +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 + 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. */ +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/framework/lod_rank_table.h" @@ -46,4 +46,13 @@ void LoDRankTable::Reset(const LoD& lod, size_t level) { } } // namespace framework + +std::ostream& operator<<(std::ostream& out, + const framework::LoDRankTable& table) { + out << "NumOfSequence " << table.items().size() << "\n"; + for (auto& each_item : table.items()) { + out << "\tSeq #" << each_item.index << ", Len=" << each_item.length << "\n"; + } + return out; +} } // namespace paddle diff --git a/paddle/framework/lod_rank_table.h b/paddle/framework/lod_rank_table.h index 9faa3a4d7bdc55ab7b24e31f5e5434dacc0a4b36..df188709e91871ded0258fa5703ee16a5664f057 100644 --- a/paddle/framework/lod_rank_table.h +++ b/paddle/framework/lod_rank_table.h @@ -1,18 +1,19 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +/* 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 +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 + 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. */ +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/framework/lod_tensor.h" namespace paddle { @@ -52,4 +53,8 @@ class LoDRankTable { }; } // namespace framework + +std::ostream& operator<<(std::ostream& out, + const framework::LoDRankTable& table); + } // namespace paddle diff --git a/paddle/framework/lod_tensor.cc b/paddle/framework/lod_tensor.cc index a0f2906c749054c1ff9f624e47df432ec2bd6ac8..7b6dc09bdb5535488c8c4dbc71c9cd6a7998bd0b 100644 --- a/paddle/framework/lod_tensor.cc +++ b/paddle/framework/lod_tensor.cc @@ -1,18 +1,20 @@ /* 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 +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 + 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. */ +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/framework/lod_tensor.h" +#include "paddle/framework/data_type.h" +#include "paddle/framework/framework.pb.h" #include "paddle/memory/memcpy.h" #include "paddle/memory/memory.h" @@ -27,11 +29,11 @@ namespace paddle { namespace framework { -std::ostream& operator<<(std::ostream& os, const LoD& lod) { +std::ostream &operator<<(std::ostream &os, const LoD &lod) { os << "{"; - for (auto& v : lod) { + for (auto &v : lod) { os << "{"; - for (auto& i : v) { + for (auto &i : v) { os << i << ","; } os << "}"; @@ -41,7 +43,7 @@ std::ostream& operator<<(std::ostream& os, const LoD& lod) { return os; } -LoD SliceLevels(const LoD& in, size_t level_begin, size_t level_end) { +LoD SliceLevels(const LoD &in, size_t level_begin, size_t level_end) { LoD new_lod; new_lod.reserve(level_end - level_begin); for (size_t i = level_begin; i < level_end; i++) { @@ -53,7 +55,7 @@ LoD SliceLevels(const LoD& in, size_t level_begin, size_t level_end) { return new_lod; } -LoD SliceInLevel(const LoD& in, size_t level, size_t elem_begin, +LoD SliceInLevel(const LoD &in, size_t level, size_t elem_begin, size_t elem_end) { PADDLE_ENFORCE_LT(level, in.size()); PADDLE_ENFORCE_LT(elem_end, in[level].size()); @@ -64,9 +66,9 @@ LoD SliceInLevel(const LoD& in, size_t level, size_t elem_begin, res[0].assign(in[level].begin() + elem_begin, in[level].begin() + elem_end + 1); for (size_t lvl = 1; lvl < res.size(); lvl++) { - const auto& in_level = in[level + lvl]; - const auto& above_level = res[lvl - 1]; - auto& out_level = res[lvl]; + const auto &in_level = in[level + lvl]; + const auto &above_level = res[lvl - 1]; + auto &out_level = res[lvl]; out_level.assign(in_level.begin() + above_level.front(), in_level.begin() + above_level.back() + 1); } @@ -74,33 +76,33 @@ LoD SliceInLevel(const LoD& in, size_t level, size_t elem_begin, // to make the first offset equals 0, all the elements minus the first // element size_t front = res[lvl].front(); - for (auto& ele : res[lvl]) { + for (auto &ele : res[lvl]) { ele -= front; } } return res; } -LoD ToAbsOffset(const LoD& in) { +LoD ToAbsOffset(const LoD &in) { // the lowest level stores relative offsets if (in.empty() || in.size() == 1) return in; LoD result = in; for (int level = result.size() - 2; level >= 0; level--) { - for (auto& ele : result[level]) { + for (auto &ele : result[level]) { ele = result[level + 1][ele]; } } return result; } -bool operator==(const LoD& a, const LoD& b) { +bool operator==(const LoD &a, const LoD &b) { if (a.size() != b.size()) { return false; } for (size_t i = 0; i < a.size(); i++) { - const auto& a_level = a[i]; - const auto& b_level = b[i]; + const auto &a_level = a[i]; + const auto &b_level = b[i]; if (a_level.size() != b_level.size()) { return false; } @@ -151,7 +153,7 @@ void LoDTensor::ShrinkInLevel(size_t level, size_t elem_begin, } using LoDAndOffset = std::pair>; -LoDAndOffset GetSubLoDAndAbsoluteOffset(const LoD& lod, size_t start_idx, +LoDAndOffset GetSubLoDAndAbsoluteOffset(const LoD &lod, size_t start_idx, size_t end_idx, size_t start_level) { LoD sub_lod; @@ -170,7 +172,7 @@ LoDAndOffset GetSubLoDAndAbsoluteOffset(const LoD& lod, size_t start_idx, return LoDAndOffset{sub_lod, {start_idx, end_idx}}; } -void AppendLoD(LoD* lod, const LoD& lod_length) { +void AppendLoD(LoD *lod, const LoD &lod_length) { PADDLE_ENFORCE( lod->empty() || lod->size() == lod_length.size(), "The lod_length should has the same size with the appended lod."); @@ -178,12 +180,65 @@ void AppendLoD(LoD* lod, const LoD& lod_length) { *lod = LoD(lod_length.size(), std::vector({0})); } for (size_t i = 0; i < lod->size(); ++i) { - auto& level = (*lod)[i]; + auto &level = (*lod)[i]; for (size_t len : lod_length[i]) { level.push_back(level.back() + len); } } } +void SerializeToStream(std::ostream &os, const LoDTensor &tensor, + const platform::DeviceContext &dev_ctx) { + { // the 1st field, uint32_t version for LoDTensor + constexpr uint32_t version = 0; + os.write(reinterpret_cast(&version), sizeof(version)); + } + { + // the 2st field, LoD information + // uint64_t lod_level + // uint64_t lod_level_1 size in byte. + // int* lod_level_1 data + // ... + auto lod = tensor.lod(); + uint64_t size = lod.size(); + os.write(reinterpret_cast(&size), sizeof(size)); + + for (auto &each : lod) { + size = each.size() * sizeof(framework::LoD::value_type::value_type); + os.write(reinterpret_cast(&size), sizeof(size)); + os.write(reinterpret_cast(each.data()), + static_cast(size)); + } + } + // the 3st field, Tensor + SerializeToStream(os, static_cast(tensor), dev_ctx); +} + +void DeserializeFromStream(std::istream &is, LoDTensor *tensor) { + { + // the 1st field, unit32_t version for SelectedRows + uint32_t version; + is.read(reinterpret_cast(&version), sizeof(version)); + PADDLE_ENFORCE_EQ(version, 0U, "Only version 0 is supported"); + } + { + // the 2st field, LoD information + uint64_t lod_level; + is.read(reinterpret_cast(&lod_level), sizeof(lod_level)); + auto &lod = *tensor->mutable_lod(); + lod.resize(lod_level); + for (uint64_t i = 0; i < lod_level; ++i) { + uint64_t size; + is.read(reinterpret_cast(&size), sizeof(size)); + std::vector tmp(size / sizeof(size_t)); + is.read(reinterpret_cast(tmp.data()), + static_cast(size)); + lod[i] = tmp; + } + } + // the 3st filed, Tensor + DeserializeFromStream(is, static_cast(tensor)); +} + } // namespace framework } // namespace paddle diff --git a/paddle/framework/lod_tensor.h b/paddle/framework/lod_tensor.h index 7f8a51cc581e759bc707e506ac7cdeb3680f40ac..147db3ab0877662d9e47ae7ee6df05638b5fcbd1 100644 --- a/paddle/framework/lod_tensor.h +++ b/paddle/framework/lod_tensor.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 @@ -24,6 +24,7 @@ #include #include "paddle/framework/ddim.h" #include "paddle/framework/tensor.h" +#include "paddle/framework/tensor_util.h" #include "paddle/platform/enforce.h" #include "paddle/platform/place.h" @@ -175,18 +176,39 @@ LoDTensor LodExpand(const LoDTensor& source, const LoD& lod, size_t level, PADDLE_ENFORCE_EQ(num_instances, lod_level.size() - 1); for (size_t ins = 0; ins < num_instances; ins++) { for (size_t elem = lod_level[ins]; elem < lod_level[ins + 1]; elem++) { - tensor.Slice(elem, elem + 1) - .CopyFrom(source.Slice(ins, ins + 1), platform::CPUPlace(), - platform::CPUDeviceContext()); + auto slice = tensor.Slice(elem, elem + 1); + CopyFrom(source.Slice(ins, ins + 1), platform::CPUPlace(), + platform::CPUDeviceContext(), &slice); } } return tensor; } +// Get the absolute offset of a lod[start_level][start_idx:end_idx] and +// relative length of details for every levels(i.e., [start_level: ]). +// +// For example, +// lod = [[0, 3, 4, 8], [0, 9, 10, 11, 13, 17, 19, 22, 24]] +// start_level = 0 +// start_idx = 1 +// end_idx = 3 +// +// Returns: +// LoD = [[1, 4], [2, 4, 2, 3, 2]] +// pair = {11, 24} std::pair> GetSubLoDAndAbsoluteOffset( const LoD& lod, size_t start_idx, size_t end_idx, size_t start_level); void AppendLoD(LoD* lod, const LoD& lod_length); +/* + * Serialize/Desiralize LoDTensor to std::ostream + * You can pass ofstream or ostringstream to serilize to file + * or to a in memory string. GPU tensor will be copied to CPU. + */ +void SerializeToStream(std::ostream& os, const LoDTensor& tensor, + const platform::DeviceContext& dev_ctx); +void DeserializeFromStream(std::istream& is, LoDTensor* tensor); + } // namespace framework } // namespace paddle diff --git a/paddle/framework/lod_tensor_array.h b/paddle/framework/lod_tensor_array.h index 13f0608d24be97d8bba149b74f1a4deb57deeb48..4a8e7f4fa540b1c2f19a6e3ec236a0dd5c0daf0b 100644 --- a/paddle/framework/lod_tensor_array.h +++ b/paddle/framework/lod_tensor_array.h @@ -1,16 +1,16 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +/* 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 +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 + 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. */ +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 diff --git a/paddle/framework/lod_tensor_test.cc b/paddle/framework/lod_tensor_test.cc index 02d84b68233f2fdfc66e1df2fc7ce20307cadd94..0747c8db531d6ae443d76591b945cce0c9bbea2b 100644 --- a/paddle/framework/lod_tensor_test.cc +++ b/paddle/framework/lod_tensor_test.cc @@ -126,6 +126,20 @@ TEST_F(LoDTensorTester, ShrinkInLevel) { EXPECT_NE(t1.data(), lod_tensor_.data()); } +TEST_F(LoDTensorTester, SerializeAndDeserialize) { + LoDTensor dst_tensor; + platform::CPUDeviceContext cpu_ctx((platform::CPUPlace())); + std::ostringstream oss; + SerializeToStream(oss, lod_tensor_, cpu_ctx); + std::istringstream iss(oss.str()); + DeserializeFromStream(iss, &dst_tensor); + float* dst_ptr = dst_tensor.mutable_data(platform::CPUPlace()); + for (int i = 0; i < kLodTensorSize; ++i) { + EXPECT_EQ(dst_ptr[i], i); + } + EXPECT_EQ(dst_tensor.lod(), lod_tensor_.lod()); +} + TEST(LodExpand, test) { LoD lod{{0, 2}}; LoDTensor tensor; diff --git a/paddle/framework/lod_tensor_test.cu b/paddle/framework/lod_tensor_test.cu index 5b90fbfca7f6bec4f2c862d0ff18dfd7cf39e181..e8508ad2658ae850e4c98aa798b5db6d007e67d0 100644 --- a/paddle/framework/lod_tensor_test.cu +++ b/paddle/framework/lod_tensor_test.cu @@ -27,7 +27,7 @@ __global__ void test(size_t* a, int size) { TEST(LoDTensor, LoDInGPU) { paddle::framework::LoDTensor lod_tensor; - paddle::platform::GPUPlace place(0); + paddle::platform::CUDAPlace place(0); paddle::framework::LoD src_lod; src_lod.push_back(std::vector{0, 2, 4, 6, 8, 10, 12, 14}); diff --git a/paddle/framework/op_desc.cc b/paddle/framework/op_desc.cc index 48cd131550dea5ad3f368b25c31d753efbe0dff9..781bbb4c19f1c610df485c3061ca8b510e727019 100644 --- a/paddle/framework/op_desc.cc +++ b/paddle/framework/op_desc.cc @@ -25,12 +25,11 @@ limitations under the License. */ namespace paddle { namespace framework { -class OpDescBind; -class BlockDescBind; +class OpDesc; +class BlockDesc; class CompileTimeInferShapeContext : public InferShapeContext { public: - CompileTimeInferShapeContext(const OpDescBind &op, - const BlockDescBind &block); + CompileTimeInferShapeContext(const OpDesc &op, const BlockDesc &block); bool HasInput(const std::string &name) const override; @@ -58,31 +57,30 @@ class CompileTimeInferShapeContext : public InferShapeContext { PADDLE_ENFORCE_LT(j, Outputs(out).size()); auto *in_var = block_.FindVarRecursive(Inputs(in)[i]); auto *out_var = block_.FindVarRecursive(Outputs(out)[j]); - if (in_var->GetType() != VarDesc::LOD_TENSOR) { - VLOG(3) << "input " << in << "is not LodTensor"; + if (in_var->GetType() != proto::VarDesc::LOD_TENSOR) { + VLOG(3) << "input " << in << " is not LodTensor"; return; } - PADDLE_ENFORCE_EQ(in_var->GetType(), VarDesc::LOD_TENSOR, + PADDLE_ENFORCE_EQ(in_var->GetType(), proto::VarDesc::LOD_TENSOR, "The %d-th output of Output(%s) must be LoDTensor.", j, out); - in_var->SetLoDLevel(out_var->GetLodLevel()); + out_var->SetLoDLevel(in_var->GetLodLevel()); } bool IsRuntime() const override; protected: - VarDesc::VarType GetVarType(const std::string &name) const override; + proto::VarDesc::VarType GetVarType(const std::string &name) const override; DDim GetDim(const std::string &name) const override; void SetDim(const std::string &name, const DDim &dim) override; - const OpDescBind &op_; - const BlockDescBind &block_; + const OpDesc &op_; + const BlockDesc &block_; }; -OpDescBind::OpDescBind(const std::string &type, const VariableNameMap &inputs, - const VariableNameMap &outputs, - const AttributeMap &attrs) { +OpDesc::OpDesc(const std::string &type, const VariableNameMap &inputs, + const VariableNameMap &outputs, const AttributeMap &attrs) { desc_.set_type(type); inputs_ = inputs; outputs_ = outputs; @@ -90,12 +88,20 @@ OpDescBind::OpDescBind(const std::string &type, const VariableNameMap &inputs, need_update_ = true; } -OpDescBind::OpDescBind(const OpDesc &desc, ProgramDescBind *prog) +void OpDesc::CopyFrom(const OpDesc &op_desc) { + desc_.set_type(op_desc.Type()); + inputs_ = op_desc.inputs_; + outputs_ = op_desc.outputs_; + attrs_ = op_desc.attrs_; + need_update_ = true; +} + +OpDesc::OpDesc(const proto::OpDesc &desc, ProgramDesc *prog) : desc_(desc), need_update_(false) { // restore inputs_ int input_size = desc_.inputs_size(); for (int i = 0; i < input_size; ++i) { - const OpDesc::Var &var = desc_.inputs(i); + const proto::OpDesc::Var &var = desc_.inputs(i); std::vector &args = inputs_[var.parameter()]; int argu_size = var.arguments_size(); args.reserve(argu_size); @@ -106,7 +112,7 @@ OpDescBind::OpDescBind(const OpDesc &desc, ProgramDescBind *prog) // restore outputs_ int output_size = desc_.outputs_size(); for (int i = 0; i < output_size; ++i) { - const OpDesc::Var &var = desc_.outputs(i); + const proto::OpDesc::Var &var = desc_.outputs(i); std::vector &args = outputs_[var.parameter()]; int argu_size = var.arguments_size(); args.reserve(argu_size); @@ -115,9 +121,9 @@ OpDescBind::OpDescBind(const OpDesc &desc, ProgramDescBind *prog) } } // restore attrs_ - for (const OpDesc::Attr &attr : desc_.attrs()) { + for (const proto::OpDesc::Attr &attr : desc_.attrs()) { std::string attr_name = attr.name(); - if (attr.type() != AttrType::BLOCK) { + if (attr.type() != proto::AttrType::BLOCK) { attrs_[attr_name] = GetAttrValue(attr); } else { auto bid = attr.block_idx(); @@ -126,20 +132,19 @@ OpDescBind::OpDescBind(const OpDesc &desc, ProgramDescBind *prog) } } -OpDesc *OpDescBind::Proto() { +proto::OpDesc *OpDesc::Proto() { Flush(); return &desc_; } -const std::vector &OpDescBind::Input( - const std::string &name) const { +const std::vector &OpDesc::Input(const std::string &name) const { auto it = inputs_.find(name); PADDLE_ENFORCE(it != inputs_.end(), "Input %s cannot be found in Op %s", name, Type()); return it->second; } -std::vector OpDescBind::InputArgumentNames() const { +std::vector OpDesc::InputArgumentNames() const { std::vector retv; for (auto &ipt : this->inputs_) { retv.insert(retv.end(), ipt.second.begin(), ipt.second.end()); @@ -147,21 +152,20 @@ std::vector OpDescBind::InputArgumentNames() const { return retv; } -void OpDescBind::SetInput(const std::string ¶m_name, - const std::vector &args) { +void OpDesc::SetInput(const std::string ¶m_name, + const std::vector &args) { need_update_ = true; inputs_[param_name] = args; } -const std::vector &OpDescBind::Output( - const std::string &name) const { +const std::vector &OpDesc::Output(const std::string &name) const { auto it = outputs_.find(name); PADDLE_ENFORCE(it != outputs_.end(), "Output %s cannot be found in Op %s", name, Type()); return it->second; } -std::vector OpDescBind::OutputArgumentNames() const { +std::vector OpDesc::OutputArgumentNames() const { std::vector retv; for (auto &ipt : this->outputs_) { retv.insert(retv.end(), ipt.second.begin(), ipt.second.end()); @@ -169,19 +173,19 @@ std::vector OpDescBind::OutputArgumentNames() const { return retv; } -void OpDescBind::SetOutput(const std::string ¶m_name, - const std::vector &args) { +void OpDesc::SetOutput(const std::string ¶m_name, + const std::vector &args) { need_update_ = true; this->outputs_[param_name] = args; } -AttrType OpDescBind::GetAttrType(const std::string &name) const { +proto::AttrType OpDesc::GetAttrType(const std::string &name) const { auto it = attrs_.find(name); PADDLE_ENFORCE(it != attrs_.end(), "Attribute %s is not found", name); - return static_cast(it->second.which() - 1); + return static_cast(it->second.which() - 1); } -std::vector OpDescBind::AttrNames() const { +std::vector OpDesc::AttrNames() const { std::vector retv; retv.reserve(attrs_.size()); for (auto &attr : attrs_) { @@ -190,41 +194,39 @@ std::vector OpDescBind::AttrNames() const { return retv; } -void OpDescBind::SetAttr(const std::string &name, const Attribute &v) { +void OpDesc::SetAttr(const std::string &name, const Attribute &v) { this->attrs_[name] = v; need_update_ = true; } -void OpDescBind::SetBlockAttr(const std::string &name, BlockDescBind &block) { +void OpDesc::SetBlockAttr(const std::string &name, BlockDesc &block) { this->attrs_[name] = █ need_update_ = true; } -void OpDescBind::SetAttrMap( +void OpDesc::SetAttrMap( const std::unordered_map &attr_map) { attrs_ = attr_map; need_update_ = true; } -Attribute OpDescBind::GetAttr(const std::string &name) const { +Attribute OpDesc::GetAttr(const std::string &name) const { auto it = attrs_.find(name); PADDLE_ENFORCE(it != attrs_.end(), "Attribute %s is not found", name); return it->second; } -int OpDescBind::GetBlockAttr(const std::string &name) const { +int OpDesc::GetBlockAttr(const std::string &name) const { auto it = attrs_.find(name); PADDLE_ENFORCE(it != attrs_.end(), "Attribute %s is not found", name); - return boost::get(it->second)->ID(); + return boost::get(it->second)->ID(); } -const std::unordered_map &OpDescBind::GetAttrMap() - const { +const std::unordered_map &OpDesc::GetAttrMap() const { return attrs_; } -void OpDescBind::Rename(const std::string &old_name, - const std::string &new_name) { +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); } @@ -235,8 +237,8 @@ void OpDescBind::Rename(const std::string &old_name, need_update_ = true; } -void OpDescBind::RenameOutput(const std::string &old_name, - const std::string &new_name) { +void OpDesc::RenameOutput(const std::string &old_name, + const std::string &new_name) { for (auto &output : outputs_) { std::replace(output.second.begin(), output.second.end(), old_name, new_name); @@ -244,8 +246,8 @@ void OpDescBind::RenameOutput(const std::string &old_name, need_update_ = true; } -void OpDescBind::RenameInput(const std::string &old_name, - const std::string &new_name) { +void OpDesc::RenameInput(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); } @@ -253,8 +255,8 @@ void OpDescBind::RenameInput(const std::string &old_name, } struct SetAttrDescVisitor : public boost::static_visitor { - explicit SetAttrDescVisitor(OpDesc::Attr *attr) : attr_(attr) {} - mutable OpDesc::Attr *attr_; + explicit SetAttrDescVisitor(proto::OpDesc::Attr *attr) : attr_(attr) {} + mutable proto::OpDesc::Attr *attr_; void operator()(int v) const { attr_->set_i(v); } void operator()(float v) const { attr_->set_f(v); } void operator()(const std::string &v) const { attr_->set_s(v); } @@ -272,11 +274,13 @@ struct SetAttrDescVisitor : public boost::static_visitor { void operator()(const std::vector &v) const { VectorToRepeated(v, attr_->mutable_bools()); } - void operator()(BlockDesc *desc) const { attr_->set_block_idx(desc->idx()); } + void operator()(proto::BlockDesc *desc) const { + attr_->set_block_idx(desc->idx()); + } void operator()(boost::blank) const { PADDLE_THROW("Unexpected branch"); } }; -void OpDescBind::Flush() { +void OpDesc::Flush() { if (need_update_) { this->desc_.mutable_inputs()->Clear(); for (auto &ipt : inputs_) { @@ -297,7 +301,7 @@ void OpDescBind::Flush() { auto *attr_desc = desc_.add_attrs(); attr_desc->set_name(attr.first); attr_desc->set_type( - static_cast(attr.second.which() - 1)); + static_cast(attr.second.which() - 1)); SetAttrDescVisitor visitor(attr_desc); boost::apply_visitor(visitor, attr.second); } @@ -316,8 +320,8 @@ static void InitInferShapeFuncs() { for (auto &kern_pair : OperatorWithKernel::AllOpKernels()) { auto op_type = kern_pair.first; auto &op_info = info_map.at(op_type); - auto op = - static_cast(op_info.Creator()("", {}, {}, {})); + auto op = static_cast(op_info.Creator()( + "", VariableNameMap{}, VariableNameMap{}, AttributeMap{})); if (op_info.infer_shape_) { // infer_shape has been registered. continue; } @@ -328,7 +332,7 @@ static void InitInferShapeFuncs() { }); } -void OpDescBind::CheckAttrs() { +void OpDesc::CheckAttrs() { PADDLE_ENFORCE(!Type().empty(), "CheckAttr() can not be called before type is setted."); auto *checker = OpInfoMap::Instance().Get(Type()).Checker(); @@ -340,7 +344,7 @@ void OpDescBind::CheckAttrs() { checker->Check(attrs_); } -void OpDescBind::InferShape(const BlockDescBind &block) const { +void OpDesc::InferShape(const BlockDesc &block) const { VLOG(3) << "CompileTime infer shape on " << Type(); InitInferShapeFuncs(); auto &infer_shape = OpInfoMap::Instance().Get(this->Type()).infer_shape_; @@ -363,7 +367,7 @@ void OpDescBind::InferShape(const BlockDescBind &block) const { infer_shape(&ctx); } -void OpDescBind::InferVarType(BlockDescBind *block) const { +void OpDesc::InferVarType(BlockDesc *block) const { auto &info = OpInfoMap::Instance().Get(this->Type()); if (info.infer_var_type_) { info.infer_var_type_(*this, block); @@ -375,14 +379,14 @@ void OpDescBind::InferVarType(BlockDescBind *block) const { for (auto &out_pair : this->outputs_) { for (auto &out_var_name : out_pair.second) { block->FindRecursiveOrCreateVar(out_var_name) - ->SetType(VarDesc::LOD_TENSOR); + ->SetType(proto::VarDesc::LOD_TENSOR); } } } } CompileTimeInferShapeContext::CompileTimeInferShapeContext( - const OpDescBind &op, const BlockDescBind &block) + const OpDesc &op, const BlockDesc &block) : op_(op), block_(block) {} bool CompileTimeInferShapeContext::HasInput(const std::string &name) const { @@ -466,7 +470,12 @@ DDim CompileTimeInferShapeContext::GetDim(const std::string &name) const { auto var = block_.FindVarRecursive(name); PADDLE_ENFORCE(var != nullptr, "Cannot find variable %s", name); try { - return framework::make_ddim(var->Shape()); + auto shape = var->Shape(); + if (shape.empty()) { + return framework::make_ddim({0UL}); + } else { + return framework::make_ddim(var->Shape()); + } } catch (...) { VLOG(5) << "GetDim of variable " << name << " error"; std::rethrow_exception(std::current_exception()); @@ -479,7 +488,7 @@ void CompileTimeInferShapeContext::SetDim(const std::string &name, } bool CompileTimeInferShapeContext::IsRuntime() const { return false; } -VarDesc::VarType CompileTimeInferShapeContext::GetVarType( +proto::VarDesc::VarType CompileTimeInferShapeContext::GetVarType( const std::string &name) const { return block_.FindVarRecursive(name)->GetType(); } diff --git a/paddle/framework/op_desc.h b/paddle/framework/op_desc.h index da032319afa775571d3942bf6ae415db7d233735..4cf784a0d0d319d09caa27b4e2b589bd7ac4f324 100644 --- a/paddle/framework/op_desc.h +++ b/paddle/framework/op_desc.h @@ -23,19 +23,21 @@ limitations under the License. */ namespace paddle { namespace framework { -class BlockDescBind; -class ProgramDescBind; +class BlockDesc; +class ProgramDesc; -class OpDescBind { +class OpDesc { public: - OpDescBind() {} + OpDesc() {} - OpDescBind(const std::string &type, const VariableNameMap &inputs, - const VariableNameMap &outputs, const AttributeMap &attrs); + OpDesc(const std::string &type, const VariableNameMap &inputs, + const VariableNameMap &outputs, const AttributeMap &attrs); - OpDescBind(const OpDesc &desc, ProgramDescBind *prog); + OpDesc(const proto::OpDesc &desc, ProgramDesc *prog); - OpDesc *Proto(); + void CopyFrom(const OpDesc &op_desc); + + proto::OpDesc *Proto(); std::string Type() const { return desc_.type(); } @@ -59,13 +61,13 @@ class OpDescBind { return attrs_.find(name) != attrs_.end(); } - AttrType GetAttrType(const std::string &name) const; + proto::AttrType GetAttrType(const std::string &name) const; std::vector AttrNames() const; void SetAttr(const std::string &name, const Attribute &v); - void SetBlockAttr(const std::string &name, BlockDescBind &block); + void SetBlockAttr(const std::string &name, BlockDesc &block); Attribute GetAttr(const std::string &name) const; @@ -107,9 +109,9 @@ class OpDescBind { void CheckAttrs(); - void InferShape(const BlockDescBind &block) const; + void InferShape(const BlockDesc &block) const; - void InferVarType(BlockDescBind *block) const; + void InferVarType(BlockDesc *block) const; void MarkAsTarget() { desc_.set_is_target(true); } @@ -126,8 +128,10 @@ class OpDescBind { return ret_val; } - OpDesc desc_; + proto::OpDesc desc_; + // input arg name => output variable names VariableNameMap inputs_; + // output arg name => output variable names VariableNameMap outputs_; AttributeMap attrs_; diff --git a/paddle/framework/op_info.cc b/paddle/framework/op_info.cc index 81ba29797c5f478e5d6a91236f3e8de1e6b43e49..b520108109bb2f72b80f83559fa065a5ca58e9e1 100644 --- a/paddle/framework/op_info.cc +++ b/paddle/framework/op_info.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/framework/op_info.h" diff --git a/paddle/framework/op_info.h b/paddle/framework/op_info.h index d3b1a3b5fa2cf8f6a9571e92a319f3757666657e..d9b89f9cac9611fcecb18bef87940632df1e2234 100644 --- a/paddle/framework/op_info.h +++ b/paddle/framework/op_info.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 @@ -34,7 +34,7 @@ class InferShapeBase { struct OpInfo { OpCreator creator_; GradOpMakerFN grad_op_maker_; - OpProto* proto_{nullptr}; + proto::OpProto* proto_{nullptr}; OpAttrChecker* checker_{nullptr}; InferVarTypeFN infer_var_type_; InferShapeFN infer_shape_; @@ -43,7 +43,7 @@ struct OpInfo { return proto_ != nullptr && checker_ != nullptr; } - const OpProto& Proto() const { + const proto::OpProto& Proto() const { PADDLE_ENFORCE_NOT_NULL(proto_, "Operator Proto has not been registered"); PADDLE_ENFORCE(proto_->IsInitialized(), "Operator Proto must be initialized in op info"); diff --git a/paddle/framework/op_kernel_type.h b/paddle/framework/op_kernel_type.h new file mode 100644 index 0000000000000000000000000000000000000000..b06002096fb109da806809f7b908d9768cf095ba --- /dev/null +++ b/paddle/framework/op_kernel_type.h @@ -0,0 +1,90 @@ +/* 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 + +#include "paddle/framework/data_layout.h" +#include "paddle/framework/data_type.h" +#include "paddle/framework/library_type.h" +#include "paddle/platform/device_context.h" +#include "paddle/platform/place.h" + +namespace paddle { +namespace framework { + +struct OpKernelType { + struct Hash { + size_t operator()(const OpKernelType& key) const { + int place = key.place_.which() + (1 << LEFT_SHIFT); + int data_type = + static_cast(key.data_type_) + (1 << (LEFT_SHIFT + 1)); + int data_layout = + static_cast(key.data_layout_) + (1 << (LEFT_SHIFT + 2)); + int library_type = + static_cast(key.library_type_) + (1 << (LEFT_SHIFT + 3)); + std::hash hasher; + return hasher(place + data_type + data_layout + library_type); + } + }; + + // place, data_type, library_type kinds less than 2^8 + constexpr static int LEFT_SHIFT = 8; + + proto::DataType data_type_; + DataLayout data_layout_; + platform::Place place_; + LibraryType library_type_; + + OpKernelType(proto::DataType data_type, platform::Place place, + DataLayout data_layout = DataLayout::kAnyLayout, + LibraryType library_type = LibraryType::kPlain) + : data_type_(data_type), + data_layout_(data_layout), + place_(place), + library_type_(library_type) {} + + OpKernelType(proto::DataType data_type, + const platform::DeviceContext& dev_ctx, + DataLayout data_layout = DataLayout::kAnyLayout, + LibraryType library_type = LibraryType::kPlain) + : data_type_(data_type), + data_layout_(data_layout), + place_(dev_ctx.GetPlace()), + library_type_(library_type) {} + + bool operator==(const OpKernelType& o) const { + return platform::places_are_same_class(place_, o.place_) && + data_type_ == o.data_type_ && data_layout_ == o.data_layout_ && + library_type_ == o.library_type_; + } + + bool operator!=(const OpKernelType& o) const { return !(*this == o); } +}; + +inline std::ostream& operator<<(std::ostream& os, + const OpKernelType& kernel_key) { + os << "data_type[" << kernel_key.data_type_ << "]:data_layout[" + << kernel_key.data_layout_ << "]:place[" << kernel_key.place_ + << "]:library_type[" << kernel_key.library_type_ << "]"; + return os; +} + +inline std::string KernelTypeToString(const OpKernelType& kernel_key) { + std::ostringstream stream; + stream << kernel_key; + return stream.str(); +} + +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/op_kernel_type_test.cc b/paddle/framework/op_kernel_type_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..649afeee8a846b0579545f2edff77e9dbe3b4dd8 --- /dev/null +++ b/paddle/framework/op_kernel_type_test.cc @@ -0,0 +1,49 @@ +/* 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. */ + +#include "paddle/framework/op_kernel_type.h" +#include +#include + +TEST(OpKernelType, ToString) { + using OpKernelType = paddle::framework::OpKernelType; + using DataType = paddle::framework::proto::DataType; + using CPUPlace = paddle::platform::CPUPlace; + using DataLayout = paddle::framework::DataLayout; + using LibraryType = paddle::framework::LibraryType; + + OpKernelType op_kernel_type(DataType::FP32, CPUPlace(), DataLayout::kNCHW, + LibraryType::kCUDNN); + + ASSERT_EQ( + paddle::framework::KernelTypeToString(op_kernel_type), + "data_type[5]:data_layout[NCHW]:place[CPUPlace]:library_type[CUDNN]"); +} + +TEST(OpKernelType, Hash) { + using OpKernelType = paddle::framework::OpKernelType; + using DataType = paddle::framework::proto::DataType; + using CPUPlace = paddle::platform::CPUPlace; + using CUDAPlace = paddle::platform::CUDAPlace; + using DataLayout = paddle::framework::DataLayout; + using LibraryType = paddle::framework::LibraryType; + + OpKernelType op_kernel_type_1(DataType::FP32, CPUPlace(), DataLayout::kNCHW, + LibraryType::kCUDNN); + OpKernelType op_kernel_type_2(DataType::FP32, CUDAPlace(0), DataLayout::kNCHW, + LibraryType::kCUDNN); + + OpKernelType::Hash hasher; + ASSERT_NE(hasher(op_kernel_type_1), hasher(op_kernel_type_2)); +} diff --git a/paddle/framework/op_proto_maker.h b/paddle/framework/op_proto_maker.h index 44e8ab16895cc604f85bb83e240eab55739f8ba0..efd3a5ca535403d8d46a73adc899d914623b53e4 100644 --- a/paddle/framework/op_proto_maker.h +++ b/paddle/framework/op_proto_maker.h @@ -22,6 +22,8 @@ namespace framework { // 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) {} @@ -80,7 +82,7 @@ class OpProtoAndCheckerMaker { class NOPMaker : public OpProtoAndCheckerMaker { public: - NOPMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker) + NOPMaker(OpProto* proto, framework::OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) {} }; diff --git a/paddle/framework/op_proto_maker_test.cc b/paddle/framework/op_proto_maker_test.cc index 988a14cf4de8fdf052ca7e8c41bff0c05ba2daaa..f16cb6fa3aa095a6d9737d84c7ce58f385a7072b 100644 --- a/paddle/framework/op_proto_maker_test.cc +++ b/paddle/framework/op_proto_maker_test.cc @@ -18,7 +18,7 @@ limitations under the License. */ class TestAttrProtoMaker : public paddle::framework::OpProtoAndCheckerMaker { public: - TestAttrProtoMaker(paddle::framework::OpProto* proto, + TestAttrProtoMaker(paddle::framework::proto::OpProto* proto, paddle::framework::OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddAttr("scale", "scale of test op"); @@ -27,7 +27,7 @@ class TestAttrProtoMaker : public paddle::framework::OpProtoAndCheckerMaker { }; TEST(ProtoMaker, DuplicatedAttr) { - paddle::framework::OpProto op_proto; + 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); @@ -35,7 +35,7 @@ TEST(ProtoMaker, DuplicatedAttr) { class TestInOutProtoMaker : public paddle::framework::OpProtoAndCheckerMaker { public: - TestInOutProtoMaker(paddle::framework::OpProto* proto, + TestInOutProtoMaker(paddle::framework::proto::OpProto* proto, paddle::framework::OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("input", "input of test op"); @@ -44,7 +44,7 @@ class TestInOutProtoMaker : public paddle::framework::OpProtoAndCheckerMaker { }; TEST(ProtoMaker, DuplicatedInOut) { - paddle::framework::OpProto op_proto; + 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); diff --git a/paddle/framework/op_registry.cc b/paddle/framework/op_registry.cc index 8dedd873aad648174b770b84e5232cd17b577e72..dfa151316daeccfe92e26818165a694b78b5df62 100644 --- a/paddle/framework/op_registry.cc +++ b/paddle/framework/op_registry.cc @@ -31,7 +31,8 @@ std::unique_ptr OpRegistry::CreateOp( } static VariableNameMap ConvertOpDescVarsToVarNameMap( - const google::protobuf::RepeatedPtrField& op_desc_vars) { + const google::protobuf::RepeatedPtrField& + op_desc_vars) { VariableNameMap ret_val; for (auto& var : op_desc_vars) { auto& var_names = ret_val[var.parameter()]; @@ -43,9 +44,10 @@ static VariableNameMap ConvertOpDescVarsToVarNameMap( return ret_val; } -std::unique_ptr OpRegistry::CreateOp(const OpDesc& op_desc) { +std::unique_ptr OpRegistry::CreateOp( + const proto::OpDesc& op_desc) { VLOG(1) << "CreateOp directly from OpDesc is deprecated. It should only be" - "used in unit tests. Use CreateOp(const OpDescBind& op_desc) " + "used in unit tests. Use CreateOp(const OpDesc& op_desc) " "instead."; VariableNameMap inputs = ConvertOpDescVarsToVarNameMap(op_desc.inputs()); VariableNameMap outputs = ConvertOpDescVarsToVarNameMap(op_desc.outputs()); @@ -57,7 +59,7 @@ std::unique_ptr OpRegistry::CreateOp(const OpDesc& op_desc) { return CreateOp(op_desc.type(), inputs, outputs, attrs); } -std::unique_ptr OpRegistry::CreateOp(const OpDescBind& op_desc) { +std::unique_ptr OpRegistry::CreateOp(const OpDesc& op_desc) { return CreateOp(op_desc.Type(), op_desc.Inputs(), op_desc.Outputs(), op_desc.GetAttrMap()); } diff --git a/paddle/framework/op_registry.h b/paddle/framework/op_registry.h index daade439e5232f06be72bc5bb1e2285124f2c3a4..bdaa25918155caca4b64b0ed60aa3f6be03eb12f 100644 --- a/paddle/framework/op_registry.h +++ b/paddle/framework/op_registry.h @@ -61,25 +61,14 @@ struct OperatorRegistrar : public Registrar { class OpRegistry { public: - template - static void RegisterOp(const std::string& op_type, - const std::string& grad_op_type) { - OperatorRegistrar reg(op_type.c_str()); - reg.info.grad_op_type_ = grad_op_type; - // register gradient op - if (!grad_op_type.empty()) { - OperatorRegistrar grad_reg(grad_op_type.c_str()); - } - } - static std::unique_ptr CreateOp(const std::string& type, const VariableNameMap& inputs, const VariableNameMap& outputs, AttributeMap attrs); - static std::unique_ptr CreateOp(const OpDesc& op_desc); + static std::unique_ptr CreateOp(const proto::OpDesc& op_desc); - static std::unique_ptr CreateOp(const OpDescBind& op_desc); + static std::unique_ptr CreateOp(const OpDesc& op_desc); }; template @@ -90,30 +79,31 @@ struct OpKernelRegistrarFunctor { using KERNEL_TYPE = typename std::tuple_element>::type; - void operator()(const char* op_type) const { + void operator()(const char* op_type, const char* library_type) const { using T = typename KERNEL_TYPE::ELEMENT_TYPE; - OpKernelType key(ToDataType(std::type_index(typeid(T))), PlaceType()); + OpKernelType key(ToDataType(std::type_index(typeid(T))), PlaceType(), + DataLayout::kAnyLayout, StringToLibraryType(library_type)); OperatorWithKernel::AllOpKernels()[op_type][key].reset(new KERNEL_TYPE); constexpr auto size = std::tuple_size>::value; OpKernelRegistrarFunctor func; - func(op_type); + func(op_type, library_type); } }; template struct OpKernelRegistrarFunctor { - void operator()(const char* op_type) const {} + void operator()(const char* op_type, const char* library_type) const {} }; // User can register many kernel in one place. The data type could be different. template class OpKernelRegistrar : public Registrar { public: - explicit OpKernelRegistrar(const char* op_type) { + explicit OpKernelRegistrar(const char* op_type, const char* library_type) { OpKernelRegistrarFunctor func; - func(op_type); + func(op_type, library_type); } }; @@ -126,6 +116,14 @@ class OpKernelRegistrar : public Registrar { __test_global_namespace_##uniq_name##__>::value, \ msg) +/* + The variadic arguments should be class types derived from one of the + following classes: + OpProtoAndCheckerMaker + GradOpDescMakerBase + VarTypeInference + InferShapeBase +*/ #define REGISTER_OPERATOR(op_type, op_class, ...) \ STATIC_ASSERT_GLOBAL_NAMESPACE( \ __reg_op__##op_type, \ @@ -144,20 +142,29 @@ class OpKernelRegistrar : public Registrar { } /** - * Macro to register Operator. + * Macro to register Operator. When the input is duplicable, you should + * use REGISTER_OP_EX with deop_empty_grad=false instead. */ -#define REGISTER_OP(op_type, op_class, op_maker_class, grad_op_type, \ - grad_op_class) \ - REGISTER_OPERATOR(grad_op_type, grad_op_class); \ - class _GradOpDescMaker_##grad_op_type##_ \ - : public ::paddle::framework::DefaultGradOpDescMaker { \ - using ::paddle::framework::DefaultGradOpDescMaker< \ - true>::DefaultGradOpDescMaker; \ - \ - protected: \ - virtual std::string GradOpType() const { return #grad_op_type; } \ - }; \ - REGISTER_OPERATOR(op_type, op_class, _GradOpDescMaker_##grad_op_type##_, \ +#define REGISTER_OP(op_type, op_class, op_maker_class, grad_op_type, \ + grad_op_class) \ + REGISTER_OP_EX(op_type, op_class, op_maker_class, grad_op_type, \ + grad_op_class, true) + +// When an argument is duplicable, we need to use this version. +// Perhaps we can omit DropEmptyIG template parameter and +// only have one version of REGISTER_OP. +#define REGISTER_OP_EX(op_type, op_class, op_maker_class, grad_op_type, \ + grad_op_class, drop_empty_grad) \ + REGISTER_OPERATOR(grad_op_type, grad_op_class); \ + class _GradOpDescMaker_##grad_op_type##_ \ + : public ::paddle::framework::DefaultGradOpDescMaker { \ + using ::paddle::framework::DefaultGradOpDescMaker< \ + drop_empty_grad>::DefaultGradOpDescMaker; \ + \ + protected: \ + virtual std::string GradOpType() const { return #grad_op_type; } \ + }; \ + REGISTER_OPERATOR(op_type, op_class, _GradOpDescMaker_##grad_op_type##_, \ op_maker_class); #define REGISTER_OP_WITH_KERNEL(op_type, ...) \ @@ -175,14 +182,15 @@ class OpKernelRegistrar : public Registrar { __reg_op_kernel_##op_type##_##DEVICE_TYPE##__, \ "REGISTER_OP_KERNEL must be called in global namespace"); \ static ::paddle::framework::OpKernelRegistrar \ - __op_kernel_registrar_##op_type##_##DEVICE_TYPE##__(#op_type); \ + __op_kernel_registrar_##op_type##_##DEVICE_TYPE##__(#op_type, \ + #DEVICE_TYPE); \ int TouchOpKernelRegistrar_##op_type##_##DEVICE_TYPE() { \ __op_kernel_registrar_##op_type##_##DEVICE_TYPE##__.Touch(); \ return 0; \ } -#define REGISTER_OP_GPU_KERNEL(op_type, ...) \ - REGISTER_OP_KERNEL(op_type, GPU, ::paddle::platform::GPUPlace, __VA_ARGS__) +#define REGISTER_OP_CUDA_KERNEL(op_type, ...) \ + REGISTER_OP_KERNEL(op_type, CUDA, ::paddle::platform::CUDAPlace, __VA_ARGS__) #define REGISTER_OP_CPU_KERNEL(op_type, ...) \ REGISTER_OP_KERNEL(op_type, CPU, ::paddle::platform::CPUPlace, __VA_ARGS__) @@ -217,7 +225,7 @@ class OpKernelRegistrar : public Registrar { #else #define USE_OP_KERNEL(op_type) \ USE_OP_DEVICE_KERNEL(op_type, CPU); \ - USE_OP_DEVICE_KERNEL(op_type, GPU) + USE_OP_DEVICE_KERNEL(op_type, CUDA) #endif #define USE_NO_KERNEL_OP(op_type) USE_OP_ITSELF(op_type); @@ -226,9 +234,9 @@ class OpKernelRegistrar : public Registrar { USE_OP_ITSELF(op_type); \ USE_OP_DEVICE_KERNEL(op_type, CPU); -#define USE_GPU_ONLY_OP(op_type) \ - USE_OP_ITSELF(op_type); \ - USE_OP_DEVICE_KERNEL(op_type, GPU) +#define USE_CUDA_ONLY_OP(op_type) \ + USE_OP_ITSELF(op_type); \ + USE_OP_DEVICE_KERNEL(op_type, CUDA) #define USE_OP(op_type) \ USE_OP_ITSELF(op_type); \ diff --git a/paddle/framework/op_registry_test.cc b/paddle/framework/op_registry_test.cc index b860fe6cac773d1e85adecc43f5dfec42b6c7661..cef530c6e639f6e2188869fa57d114ec6b885aa8 100644 --- a/paddle/framework/op_registry_test.cc +++ b/paddle/framework/op_registry_test.cc @@ -1,3 +1,17 @@ +/* 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. */ + #include "paddle/framework/op_registry.h" #include @@ -8,8 +22,7 @@ namespace framework { class CosineOp : public OperatorBase { public: using OperatorBase::OperatorBase; - void Run(const Scope& scope, - const platform::DeviceContext& dev_ctx) const override {} + void Run(const Scope& scope, const platform::Place& place) const override {} }; class CosineOpProtoAndCheckerMaker : public OpProtoAndCheckerMaker { @@ -28,8 +41,7 @@ class CosineOpProtoAndCheckerMaker : public OpProtoAndCheckerMaker { class MyTestOp : public OperatorBase { public: using OperatorBase::OperatorBase; - void Run(const Scope& scope, - const platform::DeviceContext& dev_ctx) const override {} + void Run(const Scope& scope, const platform::Place& place) const override {} }; class MyTestOpProtoAndCheckerMaker : public OpProtoAndCheckerMaker { @@ -51,7 +63,7 @@ class MyTestOpProtoAndCheckerMaker : public OpProtoAndCheckerMaker { static void BuildVar(const std::string& param_name, std::initializer_list arguments, - paddle::framework::OpDesc::Var* var) { + paddle::framework::proto::OpDesc::Var* var) { var->set_parameter(param_name); for (auto& arg_name : arguments) { var->add_arguments(arg_name); @@ -63,7 +75,7 @@ REGISTER_OP_WITHOUT_GRADIENT(my_test_op, paddle::framework::MyTestOp, paddle::framework::MyTestOpProtoAndCheckerMaker); TEST(OpRegistry, CreateOp) { - paddle::framework::OpDesc op_desc; + paddle::framework::proto::OpDesc op_desc; op_desc.set_type("cos_sim"); BuildVar("input", {"aa"}, op_desc.add_inputs()); BuildVar("output", {"bb"}, op_desc.add_outputs()); @@ -71,26 +83,26 @@ TEST(OpRegistry, CreateOp) { float scale = 3.3; auto attr = op_desc.mutable_attrs()->Add(); attr->set_name("scale"); - attr->set_type(paddle::framework::AttrType::FLOAT); + attr->set_type(paddle::framework::proto::AttrType::FLOAT); attr->set_f(scale); auto op = paddle::framework::OpRegistry::CreateOp(op_desc); paddle::framework::Scope scope; - paddle::platform::CPUDeviceContext dev_ctx; - op->Run(scope, dev_ctx); + paddle::platform::CPUPlace cpu_place; + op->Run(scope, cpu_place); float scale_get = op->Attr("scale"); ASSERT_EQ(scale_get, scale); } TEST(OpRegistry, IllegalAttr) { - paddle::framework::OpDesc op_desc; + paddle::framework::proto::OpDesc op_desc; op_desc.set_type("cos_sim"); BuildVar("input", {"aa"}, op_desc.add_inputs()); BuildVar("output", {"bb"}, op_desc.add_outputs()); auto attr = op_desc.mutable_attrs()->Add(); attr->set_name("scale"); - attr->set_type(paddle::framework::AttrType::FLOAT); + attr->set_type(paddle::framework::proto::AttrType::FLOAT); attr->set_f(-2.0); bool caught = false; @@ -108,7 +120,7 @@ TEST(OpRegistry, IllegalAttr) { } TEST(OpRegistry, DefaultValue) { - paddle::framework::OpDesc op_desc; + paddle::framework::proto::OpDesc op_desc; op_desc.set_type("cos_sim"); BuildVar("input", {"aa"}, op_desc.add_inputs()); BuildVar("output", {"bb"}, op_desc.add_outputs()); @@ -117,13 +129,13 @@ TEST(OpRegistry, DefaultValue) { auto op = paddle::framework::OpRegistry::CreateOp(op_desc); paddle::framework::Scope scope; - paddle::platform::CPUDeviceContext dev_ctx; - op->Run(scope, dev_ctx); + paddle::platform::CPUPlace cpu_place; + op->Run(scope, cpu_place); ASSERT_EQ(op->Attr("scale"), 1.0); } TEST(OpRegistry, CustomChecker) { - paddle::framework::OpDesc op_desc; + paddle::framework::proto::OpDesc op_desc; op_desc.set_type("my_test_op"); BuildVar("input", {"ii"}, op_desc.add_inputs()); BuildVar("output", {"oo"}, op_desc.add_outputs()); @@ -145,7 +157,7 @@ TEST(OpRegistry, CustomChecker) { // set 'test_attr' set to an illegal value auto attr = op_desc.mutable_attrs()->Add(); attr->set_name("test_attr"); - attr->set_type(paddle::framework::AttrType::INT); + attr->set_type(paddle::framework::proto::AttrType::INT); attr->set_i(3); caught = false; try { @@ -164,12 +176,12 @@ TEST(OpRegistry, CustomChecker) { op_desc.mutable_attrs()->Clear(); attr = op_desc.mutable_attrs()->Add(); attr->set_name("test_attr"); - attr->set_type(paddle::framework::AttrType::INT); + attr->set_type(paddle::framework::proto::AttrType::INT); attr->set_i(4); auto op = paddle::framework::OpRegistry::CreateOp(op_desc); - paddle::platform::CPUDeviceContext dev_ctx; + paddle::platform::CPUPlace cpu_place; paddle::framework::Scope scope; - op->Run(scope, dev_ctx); + op->Run(scope, cpu_place); int test_attr = op->Attr("test_attr"); ASSERT_EQ(test_attr, 4); } @@ -184,3 +196,71 @@ TEST(OperatorRegistrar, Test) { using namespace paddle::framework; OperatorRegistrar reg("cos"); } + +namespace paddle { +namespace framework { + +class OpKernelTestMaker : public OpProtoAndCheckerMaker { + public: + OpKernelTestMaker(OpProto* proto, OpAttrChecker* op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddComment("NoGradOp, same input output. no Grad"); + } +}; + +class OpWithKernelTest : public OperatorWithKernel { + public: + using OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(InferShapeContext* ctx) const override {} + + framework::OpKernelType GetActualKernelType( + const framework::ExecutionContext& ctx) const override { + return framework::OpKernelType(proto::DataType::FP32, ctx.device_context()); + } +}; + +template +class OpKernelTest : public paddle::framework::OpKernel { + public: + void Compute(const paddle::framework::ExecutionContext& ctx) const {} +}; + +} // namespace framework +} // namespace paddle + +REGISTER_OP_WITHOUT_GRADIENT(op_with_kernel, + paddle::framework::OpWithKernelTest, + paddle::framework::OpKernelTestMaker); +REGISTER_OP_CPU_KERNEL( + op_with_kernel, + paddle::framework::OpKernelTest); + +REGISTER_OP_CUDA_KERNEL(op_with_kernel, + paddle::framework::OpKernelTest< + paddle::platform::CUDADeviceContext, float>); + +TEST(OperatorRegistrar, CPU) { + paddle::framework::proto::OpDesc op_desc; + paddle::platform::CPUPlace cpu_place; + paddle::framework::Scope scope; + + op_desc.set_type("op_with_kernel"); + auto op = paddle::framework::OpRegistry::CreateOp(op_desc); + + op->Run(scope, cpu_place); +} + +#ifdef PADDLE_WITH_CUDA +TEST(OperatorRegistrar, CUDA) { + paddle::framework::proto::OpDesc op_desc; + paddle::platform::CUDAPlace cuda_place(0); + paddle::framework::Scope scope; + + op_desc.set_type("op_with_kernel"); + auto op = paddle::framework::OpRegistry::CreateOp(op_desc); + + op->Run(scope, cuda_place); +} +#endif diff --git a/paddle/framework/operator.cc b/paddle/framework/operator.cc index 93467ab8ac796277b47a861a427de2837fb2d3d4..fc7091f1c89f8b3f998f6d1b68f032b76bad2197 100644 --- a/paddle/framework/operator.cc +++ b/paddle/framework/operator.cc @@ -12,30 +12,19 @@ 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/framework/operator.h" #include #include + +#include "paddle/framework/data_transform.h" +#include "paddle/framework/executor.h" #include "paddle/framework/lod_tensor_array.h" +#include "paddle/framework/operator.h" #include "paddle/framework/shape_inference.h" #include "paddle/framework/var_type.h" namespace paddle { namespace framework { -template <> -Eigen::DefaultDevice& ExecutionContext::GetEigenDevice< - platform::CPUPlace, Eigen::DefaultDevice>() const { - return *device_context_.GetEigenDevice(); -} - -#ifdef PADDLE_WITH_CUDA -template <> -Eigen::GpuDevice& -ExecutionContext::GetEigenDevice() const { - return *device_context_.GetEigenDevice(); -} -#endif - std::string OperatorBase::Input(const std::string& name) const { auto& ins = Inputs(name); PADDLE_ENFORCE_LE(ins.size(), 1UL, @@ -254,12 +243,6 @@ std::vector ExecutionContext::MultiOutput( return res; } -std::ostream& operator<<(std::ostream& os, const OpKernelType& kernel_key) { - os << "place[" << kernel_key.place_ << "]:data_type[" << kernel_key.data_type_ - << "]"; - return os; -} - bool OpSupportGPU(const std::string& op_type) { auto& all_kernels = OperatorWithKernel::AllOpKernels(); auto it = all_kernels.find(op_type); @@ -391,7 +374,7 @@ class RuntimeInferShapeContext : public InferShapeContext { } } - VarDesc::VarType GetVarType(const std::string& name) const override { + proto::VarDesc::VarType GetVarType(const std::string& name) const override { auto* var = scope_.FindVar(name); return ToVarType(var->Type()); } @@ -401,12 +384,30 @@ class RuntimeInferShapeContext : public InferShapeContext { const Scope& scope_; }; +const platform::DeviceContext* GetDeviceContext( + framework::KernelTypePair& kernel_pair) { + auto& actual_kernel_key = kernel_pair.first; + auto& expected_kernel_key = kernel_pair.second; + platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); + + if (platform::is_gpu_place(actual_kernel_key.place_) && + platform::is_cpu_place(expected_kernel_key.place_)) { + return pool.Get(actual_kernel_key.place_); + } else if (platform::is_cpu_place(actual_kernel_key.place_) && + platform::is_gpu_place(expected_kernel_key.place_)) { + return pool.Get(expected_kernel_key.place_); + } else { + PADDLE_THROW( + "Currently, model parallelism is only supported between CPU and CUDA"); + } +} + void OperatorWithKernel::Run(const Scope& scope, - const platform::DeviceContext& dev_ctx) const { + const platform::Place& place) const { RuntimeInferShapeContext infer_shape_ctx(*this, scope); this->InferShape(&infer_shape_ctx); - - ExecutionContext ctx(*this, scope, dev_ctx); + platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); + auto dev_ctx = pool.Get(place); // check if op[type] has kernel registered. auto& all_op_kernels = AllOpKernels(); @@ -418,23 +419,72 @@ void OperatorWithKernel::Run(const Scope& scope, // check if op[type] have kernel for kernel_key OpKernelMap& kernels = kernels_iter->second; - auto kernel_key = GetKernelType(ctx); - auto kernel_iter = kernels.find(kernel_key); + + ExecutionContext ctx(*this, scope, *dev_ctx); + auto actual_kernel_key = GetActualKernelType(ctx); + auto expected_kernel_key = GetExpectedKernelType(actual_kernel_key); + auto kernel_iter = kernels.find(expected_kernel_key); if (kernel_iter == kernels.end()) { - PADDLE_THROW("The operator %s does not support %s", type_, kernel_key); + PADDLE_THROW("The operator %s does not support %s", type_, + expected_kernel_key); } - kernel_iter->second->Compute(ctx); + if (actual_kernel_key == expected_kernel_key) { + PADDLE_ENFORCE_EQ(actual_kernel_key.place_, expected_kernel_key.place_, + "Currently, model parallelism is only supported between " + "CPU and other devices. For example, multi-GPU model " + "parallelism will failed."); + } else { + auto kernel_pair = std::make_pair(actual_kernel_key, expected_kernel_key); + const DataTransformFn* trans_fun = + DataTransformFnMap::Instance().GetNullable(kernel_pair); + if (trans_fun) { + auto input_vars = this->InputVars(); + // TODO(qijun) filter the input vars that do not need to be transformed + + // filter vars that has been transformed + std::vector need_trans; + for (auto var_name : input_vars) { + auto var_name_trans = + var_name + framework::KernelTypeToString(expected_kernel_key); + if (!scope.FindVar(var_name_trans)) { + const_cast(scope).Var(var_name_trans); + need_trans.push_back(var_name); + } + } - // throws errors if have. - dev_ctx.Finish(); + if (!need_trans.empty()) { + auto trans_dev_ctx = GetDeviceContext(kernel_pair); + + // Wait for transform starting + dev_ctx->Wait(); + + for (auto var_name : need_trans) { + (*trans_fun)(trans_dev_ctx, kernel_pair, *(scope.FindVar(var_name)), + scope.FindVar(var_name + framework::KernelTypeToString( + expected_kernel_key))); + } + // Wait for data transform finishing + trans_dev_ctx->Wait(); + } + } + } + + kernel_iter->second->Compute(ctx); } -OpKernelType OperatorWithKernel::GetKernelType( + +OpKernelType OperatorWithKernel::GetActualKernelType( const ExecutionContext& ctx) const { - return OpKernelType(IndicateDataType(ctx), ctx.device_context()); + return OpKernelType(IndicateDataType(ctx), ctx.GetPlace()); +} + +OpKernelType OperatorWithKernel::GetExpectedKernelType( + const OpKernelType& actual_kernel_type) const { + return actual_kernel_type; } -DataType OperatorWithKernel::IndicateDataType( + +proto::DataType OperatorWithKernel::IndicateDataType( const ExecutionContext& ctx) const { auto& scope = ctx.scope(); int data_type = -1; @@ -460,7 +510,7 @@ DataType OperatorWithKernel::IndicateDataType( } } PADDLE_ENFORCE(data_type != -1, "DataType should be indicated by input"); - return static_cast(data_type); + return static_cast(data_type); } } // namespace framework diff --git a/paddle/framework/operator.h b/paddle/framework/operator.h index 60861d92933dd100f877bec8d43f9b924f951e60..d0a9b643d565d6651fd7ec0b515f088362852ba3 100644 --- a/paddle/framework/operator.h +++ b/paddle/framework/operator.h @@ -23,15 +23,14 @@ limitations under the License. */ #include "glog/logging.h" // For VLOG #include "paddle/framework/attribute.h" #include "paddle/framework/block_desc.h" -#include "paddle/framework/data_type.h" #include "paddle/framework/framework.pb.h" #include "paddle/framework/lod_tensor.h" #include "paddle/framework/op_info.h" +#include "paddle/framework/op_kernel_type.h" #include "paddle/framework/scope.h" #include "paddle/framework/selected_rows.h" #include "paddle/framework/tensor.h" #include "paddle/platform/device_context.h" -#include "paddle/platform/place.h" #include "paddle/platform/variant.h" #include "paddle/utils/Error.h" @@ -53,6 +52,11 @@ constexpr char kGradVarSuffix[] = "@GRAD"; /// Variables with this suffix are supposed to be filled up with zeros. constexpr char kZeroVarSuffix[] = "@ZERO"; +// define some kernel hint +const std::string kUseCPU = "use_cpu"; +const std::string kUseCUDNN = "use_cudnn"; +const std::string kUseMKLDNN = "use_mkldnn"; + inline std::string GradVarName(const std::string& var_name) { return var_name + kGradVarSuffix; } @@ -83,8 +87,10 @@ class OperatorBase { virtual std::string DebugString() const; /// Net will call this function to Run an op. - virtual void Run(const Scope& scope, - const platform::DeviceContext& dev_ctx) const = 0; + virtual void Run(const Scope& scope, const platform::Place& place) const = 0; + + // FIXME(typhoonzero): this is only used for recv_op to stop event_loop. + virtual void Stop() {} virtual bool IsNetOp() const { return false; } @@ -159,8 +165,7 @@ class OperatorBase { class NOP : public OperatorBase { public: using OperatorBase::OperatorBase; - void Run(const Scope& scope, - const platform::DeviceContext& dev_ctx) const override {} + void Run(const Scope& scope, const platform::Place& place) const override {} std::unique_ptr Clone() const override { return std::unique_ptr(new NOP(*this)); } @@ -276,17 +281,25 @@ class ExecutionContext { out_tensor->set_lod(in_tensor.lod()); } - template ::EigenDeviceType> - DeviceType& GetEigenDevice() const; - platform::Place GetPlace() const { return device_context_.GetPlace(); } + template + const DeviceContextType& device_context() const { + return *reinterpret_cast(&device_context_); + } + const platform::DeviceContext& device_context() const { return device_context_; } +#ifdef PADDLE_WITH_CUDA + const inline platform::CUDADeviceContext& cuda_device_context() const { + PADDLE_ENFORCE(platform::is_gpu_place(device_context_.GetPlace())); + return *reinterpret_cast( + &device_context_); + } +#endif + //! Get actual name vector for this input. const std::vector& Inputs(const std::string& name) const { return op_.Inputs(name); @@ -297,14 +310,6 @@ class ExecutionContext { return op_.Outputs(name); } -#ifdef PADDLE_WITH_CUDA - const inline platform::CUDADeviceContext& cuda_device_context() const { - PADDLE_ENFORCE(platform::is_gpu_place(device_context_.GetPlace())); - return *reinterpret_cast( - &device_context_); - } -#endif - private: const OperatorBase& op_; const Scope& scope_; @@ -345,33 +350,6 @@ class OpKernel : public OpKernelBase { using ELEMENT_TYPE = T; }; -struct OpKernelType { - struct Hash { - std::hash hash_; - size_t operator()(const OpKernelType& key) const { - int place = key.place_.which(); - int data_type = static_cast(key.data_type_); - int pre_hash = data_type << NUM_PLACE_TYPE_LIMIT_IN_BIT | - (place & ((1 << NUM_PLACE_TYPE_LIMIT_IN_BIT) - 1)); - return hash_(pre_hash); - } - }; - - platform::Place place_; - DataType data_type_; - - OpKernelType(DataType data_type, platform::Place place) - : place_(place), data_type_(data_type) {} - - OpKernelType(DataType data_type, const platform::DeviceContext& dev_ctx) - : place_(dev_ctx.GetPlace()), data_type_(data_type) {} - - bool operator==(const OpKernelType& o) const { - return platform::places_are_same_class(place_, o.place_) && - data_type_ == o.data_type_; - } -}; - class OperatorWithKernel : public OperatorBase { public: using OpKernelMap = @@ -382,8 +360,7 @@ class OperatorWithKernel : public OperatorBase { const VariableNameMap& outputs, const AttributeMap& attrs) : OperatorBase(type, inputs, outputs, attrs) {} - void Run(const Scope& scope, - const platform::DeviceContext& dev_ctx) const final; + void Run(const Scope& scope, const platform::Place& place) const final; static std::unordered_map& AllOpKernels() { @@ -404,16 +381,16 @@ class OperatorWithKernel : public OperatorBase { } protected: - virtual OpKernelType GetKernelType(const ExecutionContext& ctx) const; + virtual OpKernelType GetActualKernelType(const ExecutionContext& ctx) const; + virtual OpKernelType GetExpectedKernelType( + const OpKernelType& actual_kernel_type) const; private: // indicate kernel DataType by input data. Defaultly all input data must be // same. - DataType IndicateDataType(const ExecutionContext& ctx) const; + proto::DataType IndicateDataType(const ExecutionContext& ctx) const; }; -std::ostream& operator<<(std::ostream& os, const OpKernelType& kernel_key); - extern bool OpSupportGPU(const std::string& op_type); } // namespace framework diff --git a/paddle/framework/operator_test.cc b/paddle/framework/operator_test.cc index 1e19f82b341768142258ba4a5dfa246d87ba4c43..4d38a7ada91af834aa1a19b49e36d606ebe786ba 100644 --- a/paddle/framework/operator_test.cc +++ b/paddle/framework/operator_test.cc @@ -11,11 +11,12 @@ 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/framework/operator.h" #include "gtest/gtest.h" + +#include "paddle/framework/init.h" #include "paddle/framework/op_info.h" #include "paddle/framework/op_registry.h" +#include "paddle/framework/operator.h" namespace paddle { namespace framework { @@ -27,8 +28,7 @@ class OpWithoutKernelTest : public OperatorBase { OpWithoutKernelTest(const std::string& type, const VariableNameMap& inputs, const VariableNameMap& outputs, const AttributeMap& attrs) : OperatorBase(type, inputs, outputs, attrs), x(1) {} - void Run(const Scope& scope, - const platform::DeviceContext& dev_ctx) const override { + void Run(const Scope& scope, const platform::Place& place) const override { ++op_run_num; ASSERT_EQ(static_cast(inputs_.size()), 1); ASSERT_EQ(static_cast(outputs_.size()), 1); @@ -41,10 +41,9 @@ class OpWithoutKernelTest : public OperatorBase { int x{0}; }; -class OpeWithoutKernelTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker { +class OpWithoutKernelCheckerMaker : public OpProtoAndCheckerMaker { public: - OpeWithoutKernelTestProtoAndCheckerMaker(OpProto* proto, - OpAttrChecker* op_checker) + OpWithoutKernelCheckerMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("input", "input of test op"); AddOutput("output", "output of test op"); @@ -58,35 +57,36 @@ class OpeWithoutKernelTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker { static void BuildVar(const std::string& param_name, std::initializer_list arguments, - paddle::framework::OpDesc::Var* var) { + paddle::framework::proto::OpDesc::Var* var) { var->set_parameter(param_name); for (auto& arg_name : arguments) { *var->mutable_arguments()->Add() = arg_name; } } -REGISTER_OP_WITHOUT_GRADIENT( - test_operator, paddle::framework::OpWithoutKernelTest, - paddle::framework::OpeWithoutKernelTestProtoAndCheckerMaker); +REGISTER_OP_WITHOUT_GRADIENT(test_operator, + paddle::framework::OpWithoutKernelTest, + paddle::framework::OpWithoutKernelCheckerMaker); TEST(OperatorBase, all) { - paddle::framework::OpDesc op_desc; + paddle::framework::InitDevices({"CPU"}); + paddle::framework::proto::OpDesc op_desc; op_desc.set_type("test_operator"); BuildVar("input", {"IN1"}, op_desc.add_inputs()); BuildVar("output", {"OUT1"}, op_desc.add_outputs()); auto attr = op_desc.mutable_attrs()->Add(); attr->set_name("scale"); - attr->set_type(paddle::framework::AttrType::FLOAT); + attr->set_type(paddle::framework::proto::AttrType::FLOAT); attr->set_f(3.14); - paddle::platform::CPUDeviceContext device_context; + paddle::platform::CPUPlace cpu_place; paddle::framework::Scope scope; auto op = paddle::framework::OpRegistry::CreateOp(op_desc); scope.Var("OUT1"); ASSERT_EQ(paddle::framework::op_run_num, 0); - op->Run(scope, device_context); + op->Run(scope, cpu_place); ASSERT_EQ(paddle::framework::op_run_num, 1); } @@ -114,8 +114,8 @@ class OpWithKernelTest : public OperatorWithKernel { protected: void InferShape(framework::InferShapeContext* ctx) const override {} - OpKernelType GetKernelType(const ExecutionContext& ctx) const override { - return OpKernelType(DataType::FP32, ctx.device_context()); + OpKernelType GetActualKernelType(const ExecutionContext& ctx) const override { + return OpKernelType(proto::DataType::FP32, ctx.GetPlace()); } }; @@ -123,7 +123,6 @@ template class CPUKernelTest : public OpKernel { public: void Compute(const ExecutionContext& ctx) const { - std::cout << "this is cpu kernel" << std::endl; std::cout << ctx.op().DebugString() << std::endl; cpu_kernel_run_num++; ASSERT_EQ(ctx.op().Input("x"), "IN1"); @@ -195,22 +194,23 @@ REGISTER_OP_CPU_KERNEL(op_with_kernel, // test with single input TEST(OpKernel, all) { - paddle::framework::OpDesc op_desc; + paddle::framework::InitDevices({"CPU"}); + paddle::framework::proto::OpDesc op_desc; op_desc.set_type("op_with_kernel"); BuildVar("x", {"IN1"}, op_desc.add_inputs()); BuildVar("y", {"OUT1"}, op_desc.add_outputs()); auto attr = op_desc.mutable_attrs()->Add(); attr->set_name("scale"); - attr->set_type(paddle::framework::AttrType::FLOAT); + attr->set_type(paddle::framework::proto::AttrType::FLOAT); attr->set_f(3.14); - paddle::platform::CPUDeviceContext cpu_device_context; + paddle::platform::CPUPlace cpu_place; paddle::framework::Scope scope; auto op = paddle::framework::OpRegistry::CreateOp(op_desc); ASSERT_EQ(paddle::framework::cpu_kernel_run_num, 0); - op->Run(scope, cpu_device_context); + op->Run(scope, cpu_place); ASSERT_EQ(paddle::framework::cpu_kernel_run_num, 1); } @@ -224,7 +224,9 @@ REGISTER_OP_CPU_KERNEL(op_multi_inputs_with_kernel, TEST(OpKernel, multi_inputs) { using namespace paddle::framework; - OpDesc op_desc; + paddle::framework::InitDevices({"CPU"}); + proto::OpDesc op_desc; + op_desc.set_type("op_multi_inputs_with_kernel"); BuildVar("xs", {"x0", "x1", "x2"}, op_desc.add_inputs()); BuildVar("k", {"k0"}, op_desc.add_inputs()); @@ -232,10 +234,10 @@ TEST(OpKernel, multi_inputs) { auto attr = op_desc.mutable_attrs()->Add(); attr->set_name("scale"); - attr->set_type(paddle::framework::AttrType::FLOAT); + attr->set_type(paddle::framework::proto::AttrType::FLOAT); attr->set_f(3.14); - paddle::platform::CPUDeviceContext cpu_device_context; + paddle::platform::CPUPlace cpu_place; paddle::framework::Scope scope; scope.Var("x0")->GetMutable(); scope.Var("x1")->GetMutable(); @@ -245,7 +247,7 @@ TEST(OpKernel, multi_inputs) { scope.Var("y1")->GetMutable(); auto op = paddle::framework::OpRegistry::CreateOp(op_desc); - op->Run(scope, cpu_device_context); + op->Run(scope, cpu_place); } class OperatorClone : public paddle::framework::OperatorBase { @@ -257,11 +259,14 @@ class OperatorClone : public paddle::framework::OperatorBase { const paddle::framework::AttributeMap& attrs) : OperatorBase(type, inputs, outputs, attrs) {} void Run(const paddle::framework::Scope& scope, - const paddle::platform::DeviceContext& dev_ctx) const override {} + const paddle::platform::Place& place) const override {} }; TEST(Operator, Clone) { - OperatorClone a("ABC", {}, {}, {}); + paddle::framework::InitDevices({"CPU"}); + OperatorClone a("ABC", paddle::framework::VariableNameMap{}, + paddle::framework::VariableNameMap{}, + paddle::framework::AttributeMap{}); auto b = a.Clone(); ASSERT_EQ(a.Type(), b->Type()); } diff --git a/paddle/framework/program_desc.cc b/paddle/framework/program_desc.cc index 4af8d94563ad0ecf6fcc6fe0575b0f69006a9a2d..b5d9e5e385c1ba57169ef885824fc23b0f130692 100644 --- a/paddle/framework/program_desc.cc +++ b/paddle/framework/program_desc.cc @@ -18,49 +18,49 @@ limitations under the License. */ namespace paddle { namespace framework { -BlockDescBind *ProgramDescBind::AppendBlock(const BlockDescBind &parent) { +BlockDesc *ProgramDesc::AppendBlock(const BlockDesc &parent) { auto *b = desc_.add_blocks(); b->set_parent_idx(parent.ID()); b->set_idx(desc_.blocks_size() - 1); - blocks_.emplace_back(new BlockDescBind(this, b)); + blocks_.emplace_back(new BlockDesc(this, b)); return blocks_.back().get(); } -ProgramDesc *ProgramDescBind::Proto() { +proto::ProgramDesc *ProgramDesc::Proto() { for (auto &block : blocks_) { block->Flush(); } return &desc_; } -ProgramDescBind::ProgramDescBind() { +ProgramDesc::ProgramDesc() { auto *block = desc_.mutable_blocks()->Add(); block->set_idx(kRootBlockIndex); block->set_parent_idx(kNoneBlockIndex); - blocks_.emplace_back(new BlockDescBind(this, block)); + blocks_.emplace_back(new BlockDesc(this, block)); } -ProgramDescBind::ProgramDescBind(const ProgramDescBind &o) { +ProgramDesc::ProgramDesc(const ProgramDesc &o) { desc_ = o.desc_; for (int i = 0; i < desc_.blocks_size(); ++i) { auto *block = desc_.mutable_blocks(i); - blocks_.emplace_back(new BlockDescBind(*o.blocks_[i], block, this)); + blocks_.emplace_back(new BlockDesc(*o.blocks_[i], block, this)); } } -ProgramDescBind::ProgramDescBind(const ProgramDesc &desc) { +ProgramDesc::ProgramDesc(const proto::ProgramDesc &desc) { desc_ = desc; for (auto &block_desc : *desc_.mutable_blocks()) { - blocks_.emplace_back(new BlockDescBind(this, &block_desc)); + blocks_.emplace_back(new BlockDesc(this, &block_desc)); } } -ProgramDescBind::ProgramDescBind(const std::string &binary_str) { +ProgramDesc::ProgramDesc(const std::string &binary_str) { PADDLE_ENFORCE(desc_.ParseFromString(binary_str), "Fail to parse program_desc from binary string."); for (auto &block_desc : *desc_.mutable_blocks()) { - blocks_.emplace_back(new BlockDescBind(this, &block_desc)); + blocks_.emplace_back(new BlockDesc(this, &block_desc)); } } diff --git a/paddle/framework/program_desc.h b/paddle/framework/program_desc.h index b1cb086de4345902482d8254b8aeec041ecf81bc..15a962bb696d6172acd1a83cf9bb1ffd0846d449 100644 --- a/paddle/framework/program_desc.h +++ b/paddle/framework/program_desc.h @@ -23,32 +23,32 @@ limitations under the License. */ namespace paddle { namespace framework { -class BlockDescBind; +class BlockDesc; -class ProgramDescBind { +class ProgramDesc { public: - ProgramDescBind(); + ProgramDesc(); - explicit ProgramDescBind(const ProgramDesc &desc); + explicit ProgramDesc(const proto::ProgramDesc &desc); - ProgramDescBind(const ProgramDescBind &o); + ProgramDesc(const ProgramDesc &o); - explicit ProgramDescBind(const std::string &binary_str); + explicit ProgramDesc(const std::string &binary_str); - BlockDescBind *AppendBlock(const BlockDescBind &parent); + BlockDesc *AppendBlock(const BlockDesc &parent); - BlockDescBind *MutableBlock(size_t idx) { return blocks_[idx].get(); } + BlockDesc *MutableBlock(size_t idx) { return blocks_[idx].get(); } - const BlockDescBind &Block(size_t idx) const { return *blocks_[idx]; } + const BlockDesc &Block(size_t idx) const { return *blocks_[idx]; } size_t Size() const { return blocks_.size(); } - ProgramDesc *Proto(); + proto::ProgramDesc *Proto(); private: - ProgramDesc desc_; + proto::ProgramDesc desc_; - std::vector> blocks_; + std::vector> blocks_; }; } // namespace framework } // namespace paddle diff --git a/paddle/framework/program_desc_test.cc b/paddle/framework/program_desc_test.cc index 83e7286e0ec3639fa589b0958922543a3ba16a00..59947c9f2189348226b7ff6c2b9315196bbf55fa 100644 --- a/paddle/framework/program_desc_test.cc +++ b/paddle/framework/program_desc_test.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/framework/program_desc.h" #include "gtest/gtest.h" @@ -19,18 +19,18 @@ namespace paddle { namespace framework { TEST(ProgramDesc, copy_ctor) { - ProgramDescBind program; + ProgramDesc program; auto* global_block = program.MutableBlock(0); auto* x = global_block->Var("X"); - x->SetType(VarDesc_VarType_LOD_TENSOR); + x->SetType(proto::VarDesc_VarType_LOD_TENSOR); x->SetLoDLevel(0); - x->SetDataType(FP32); + x->SetDataType(proto::FP32); x->SetShape({1000, 784}); auto* y = global_block->Var("Y"); - y->SetType(VarDesc_VarType_LOD_TENSOR); + y->SetType(proto::VarDesc_VarType_LOD_TENSOR); y->SetLoDLevel(0); - y->SetDataType(FP32); + y->SetDataType(proto::FP32); y->SetShape({784, 100}); auto* op = global_block->AppendOp(); @@ -39,15 +39,15 @@ TEST(ProgramDesc, copy_ctor) { op->SetInput("Y", {y->Name()}); auto* out = global_block->Var("Out"); - out->SetType(VarDesc_VarType_LOD_TENSOR); + out->SetType(proto::VarDesc_VarType_LOD_TENSOR); op->SetOutput("Y", {out->Name()}); - ProgramDescBind program_copy(program); + ProgramDesc program_copy(program); auto* global_block_copy = program_copy.MutableBlock(0); ASSERT_NE(global_block, global_block_copy); - auto assert_same_var = [&](const std::string& name, VarDescBind* var_before) { + auto assert_same_var = [&](const std::string& name, VarDesc* var_before) { ASSERT_TRUE(global_block_copy->HasVar(name)); auto* copy = global_block_copy->Var(name); ASSERT_NE(copy, var_before); @@ -81,18 +81,18 @@ TEST(ProgramDesc, copy_ctor) { } TEST(ProgramDescBind, serialize_and_deserialize) { - ProgramDescBind program_origin; + ProgramDesc program_origin; auto* global_block = program_origin.MutableBlock(0); auto* x = global_block->Var("X"); - x->SetType(VarDesc_VarType_LOD_TENSOR); + x->SetType(proto::VarDesc_VarType_LOD_TENSOR); x->SetLoDLevel(0); - x->SetDataType(FP32); + x->SetDataType(proto::FP32); x->SetShape({1000, 784}); auto* y = global_block->Var("Y"); - y->SetType(VarDesc_VarType_LOD_TENSOR); + y->SetType(proto::VarDesc_VarType_LOD_TENSOR); y->SetLoDLevel(0); - y->SetDataType(FP32); + y->SetDataType(proto::FP32); y->SetShape({784, 100}); auto* op = global_block->AppendOp(); @@ -101,17 +101,17 @@ TEST(ProgramDescBind, serialize_and_deserialize) { op->SetInput("Y", {y->Name()}); auto* out = global_block->Var("Out"); - out->SetType(VarDesc_VarType_LOD_TENSOR); + out->SetType(proto::VarDesc_VarType_LOD_TENSOR); op->SetOutput("Y", {out->Name()}); std::string binary_str; program_origin.Proto()->SerializeToString(&binary_str); - ProgramDescBind program_restored(binary_str); + ProgramDesc program_restored(binary_str); auto* global_block_restored = program_restored.MutableBlock(0); ASSERT_NE(global_block, global_block_restored); - auto assert_same_var = [&](const std::string& name, VarDescBind* var_before) { + auto assert_same_var = [&](const std::string& name, VarDesc* var_before) { ASSERT_TRUE(global_block_restored->HasVar(name)); auto* restored = global_block_restored->Var(name); ASSERT_NE(restored, var_before); diff --git a/paddle/framework/prune.cc b/paddle/framework/prune.cc index bf3066983cdcf44ae84f236ac72486e5d4fd5b92..25eb813ffb96e9b1e13299421ead9f85c02da59f 100644 --- a/paddle/framework/prune.cc +++ b/paddle/framework/prune.cc @@ -26,8 +26,10 @@ namespace framework { const std::string kFeedOpType = "feed"; const std::string kFetchOpType = "fetch"; +const std::string kDropOutOpType = "dropout"; +const std::string kBatchNormOpType = "batch_norm"; -bool HasDependentVar(const OpDesc& op_desc, +bool HasDependentVar(const proto::OpDesc& op_desc, const std::set& dependent_vars) { for (auto& var : op_desc.outputs()) { for (auto& argu : var.arguments()) { @@ -39,14 +41,15 @@ bool HasDependentVar(const OpDesc& op_desc, return false; } -bool IsTarget(const OpDesc& op_desc) { +bool IsTarget(const proto::OpDesc& op_desc) { if (op_desc.has_is_target()) { return op_desc.is_target(); } return false; } -void prune_impl(const ProgramDesc& input, ProgramDesc* output, int block_id) { +void prune_impl(const proto::ProgramDesc& input, proto::ProgramDesc* output, + int block_id) { // TODO(tonyyang-svail): // - will change to use multiple blocks for RNN op and Cond Op @@ -102,9 +105,31 @@ void prune_impl(const ProgramDesc& input, ProgramDesc* output, int block_id) { } // TODO(fengjiayi): Prune() could be inplaced to avoid unnecessary copies -void Prune(const ProgramDesc& input, ProgramDesc* output) { +void Prune(const proto::ProgramDesc& input, proto::ProgramDesc* output) { prune_impl(input, output, 0); } +void inference_optimize_impl(const proto::ProgramDesc& input, + proto::ProgramDesc* output, int block_id) { + *output = input; + auto* op_field = output->mutable_blocks(block_id)->mutable_ops(); + for (auto& op_desc : *op_field) { + if (op_desc.type() == kDropOutOpType || + op_desc.type() == kBatchNormOpType) { + for (auto& attr : *op_desc.mutable_attrs()) { + if (attr.name() == "is_test") { + attr.set_b(true); + break; + } + } + } + } +} + +void InferenceOptimize(const proto::ProgramDesc& input, + proto::ProgramDesc* output) { + inference_optimize_impl(input, output, 0); +} + } // namespace framework } // namespace paddle diff --git a/paddle/framework/prune.h b/paddle/framework/prune.h index 8cfb16343aa44dcc8a3349b01adecce33f1c2b5b..593292523d0c14136791bb804a4721a0740b47ba 100644 --- a/paddle/framework/prune.h +++ b/paddle/framework/prune.h @@ -20,7 +20,10 @@ limitations under the License. */ namespace paddle { namespace framework { -void Prune(const ProgramDesc& input, ProgramDesc* output); +void Prune(const proto::ProgramDesc& input, proto::ProgramDesc* output); + +void InferenceOptimize(const proto::ProgramDesc& input, + proto::ProgramDesc* output); } // namespace framework } // namespace paddle diff --git a/paddle/framework/prune_test.cc b/paddle/framework/prune_test.cc index 5988874809f51c09b3d3d279be6c1e8d43d7a782..d76c5abca94cb87220ce73537a8657c3ec695f4d 100644 --- a/paddle/framework/prune_test.cc +++ b/paddle/framework/prune_test.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/framework/prune.h" @@ -29,12 +29,12 @@ namespace ops = paddle::operators; void AddOp(const std::string &type, const f::VariableNameMap &inputs, const f::VariableNameMap &outputs, f::AttributeMap attrs, - paddle::framework::BlockDescBind *block) { + paddle::framework::BlockDesc *block) { // insert output for (auto kv : outputs) { for (auto v : kv.second) { auto var = block->Var(v); - var->SetDataType(paddle::framework::DataType::FP32); + var->SetDataType(paddle::framework::proto::DataType::FP32); } } @@ -51,88 +51,102 @@ void AddOp(const std::string &type, const f::VariableNameMap &inputs, } TEST(Prune, one_operator) { - f::ProgramDescBind program; - f::BlockDescBind *block = program.MutableBlock(0); + f::ProgramDesc program; + f::BlockDesc *block = program.MutableBlock(0); - AddOp("one_one", {{"input", {"a"}}}, {{"output", {"b"}}}, {}, block); + AddOp("one_one", {{"input", {"a"}}}, {{"output", {"b"}}}, f::AttributeMap{}, + block); - f::ProgramDesc *pdesc = program.Proto(); - f::ProgramDesc pruned; + f::proto::ProgramDesc *pdesc = program.Proto(); + f::proto::ProgramDesc pruned; - Prune(*pdesc, &pruned); + f::Prune(*pdesc, &pruned); PADDLE_ENFORCE_EQ(pruned.blocks(0).ops_size(), 0); pdesc->mutable_blocks(0)->mutable_ops(0)->set_is_target(true); - Prune(*pdesc, &pruned); + f::Prune(*pdesc, &pruned); PADDLE_ENFORCE_EQ(pruned.blocks(0).ops_size(), 1); } TEST(Prune, forward) { - f::ProgramDescBind program; - f::BlockDescBind *block = program.MutableBlock(0); + f::ProgramDesc program; + f::BlockDesc *block = program.MutableBlock(0); - AddOp("one_one", {{"input", {"a"}}}, {{"output", {"b"}}}, {}, block); - AddOp("one_one", {{"input", {"b"}}}, {{"output", {"c"}}}, {}, block); - AddOp("one_one", {{"input", {"c"}}}, {{"output", {"d"}}}, {}, block); - AddOp("one_one", {{"input", {"d"}}}, {{"output", {"e"}}}, {}, block); + AddOp("one_one", {{"input", {"a"}}}, {{"output", {"b"}}}, f::AttributeMap{}, + block); + AddOp("one_one", {{"input", {"b"}}}, {{"output", {"c"}}}, f::AttributeMap{}, + block); + AddOp("one_one", {{"input", {"c"}}}, {{"output", {"d"}}}, f::AttributeMap{}, + block); + AddOp("one_one", {{"input", {"d"}}}, {{"output", {"e"}}}, f::AttributeMap{}, + block); - f::ProgramDesc *pdesc = program.Proto(); + f::proto::ProgramDesc *pdesc = program.Proto(); for (int i = 0; i < pdesc->blocks(0).ops_size(); ++i) { - f::ProgramDesc pruned; + f::proto::ProgramDesc pruned; pdesc->mutable_blocks(0)->mutable_ops(i)->set_is_target(true); - Prune(*pdesc, &pruned); + f::Prune(*pdesc, &pruned); PADDLE_ENFORCE_EQ(pruned.blocks(0).ops_size(), i + 1); } } TEST(Prune, multi_input_op) { - f::ProgramDescBind program; - f::BlockDescBind *block = program.MutableBlock(0); + f::ProgramDesc program; + f::BlockDesc *block = program.MutableBlock(0); - AddOp("one_one", {{"input", {"a0"}}}, {{"output", {"b0"}}}, {}, block); - AddOp("one_one", {{"input", {"a1"}}}, {{"output", {"b1"}}}, {}, block); - AddOp("one_one", {{"input", {"a2"}}}, {{"output", {"b2"}}}, {}, block); - AddOp("three_one", {{"input", {"b0", "b1", "b2"}}}, {{"output", {"c"}}}, {}, + AddOp("one_one", {{"input", {"a0"}}}, {{"output", {"b0"}}}, f::AttributeMap{}, + block); + AddOp("one_one", {{"input", {"a1"}}}, {{"output", {"b1"}}}, f::AttributeMap{}, block); + AddOp("one_one", {{"input", {"a2"}}}, {{"output", {"b2"}}}, f::AttributeMap{}, + block); + AddOp("three_one", {{"input", {"b0", "b1", "b2"}}}, {{"output", {"c"}}}, + f::AttributeMap{}, block); - f::ProgramDesc *pdesc = program.Proto(); + f::proto::ProgramDesc *pdesc = program.Proto(); pdesc->mutable_blocks(0)->mutable_ops(3)->set_is_target(true); - f::ProgramDesc pruned; - Prune(*pdesc, &pruned); + f::proto::ProgramDesc pruned; + f::Prune(*pdesc, &pruned); PADDLE_ENFORCE_EQ(pruned.blocks(0).ops_size(), 4); } TEST(Prune, multi_output_op) { - f::ProgramDescBind program; - f::BlockDescBind *block = program.MutableBlock(0); + f::ProgramDesc program; + f::BlockDesc *block = program.MutableBlock(0); - AddOp("one_two", {{"input", {"a"}}}, {{"output", {"b", "c"}}}, {}, block); - AddOp("one_one", {{"input", {"b"}}}, {{"output", {"b1"}}}, {}, block); - AddOp("one_one", {{"input", {"c"}}}, {{"output", {"c1"}}}, {}, block); + AddOp("one_two", {{"input", {"a"}}}, {{"output", {"b", "c"}}}, + f::AttributeMap{}, block); + AddOp("one_one", {{"input", {"b"}}}, {{"output", {"b1"}}}, f::AttributeMap{}, + block); + AddOp("one_one", {{"input", {"c"}}}, {{"output", {"c1"}}}, f::AttributeMap{}, + block); - f::ProgramDesc *pdesc = program.Proto(); + f::proto::ProgramDesc *pdesc = program.Proto(); pdesc->mutable_blocks(0)->mutable_ops(2)->set_is_target(true); - f::ProgramDesc pruned; - Prune(*pdesc, &pruned); + f::proto::ProgramDesc pruned; + f::Prune(*pdesc, &pruned); PADDLE_ENFORCE_EQ(pruned.blocks(0).ops_size(), 2); } TEST(Prune, multi_target) { - f::ProgramDescBind program; - f::BlockDescBind *block = program.MutableBlock(0); + f::ProgramDesc program; + f::BlockDesc *block = program.MutableBlock(0); - AddOp("one_two", {{"input", {"a"}}}, {{"output", {"b", "c"}}}, {}, block); - AddOp("one_one", {{"input", {"b"}}}, {{"output", {"b1"}}}, {}, block); - AddOp("one_one", {{"input", {"c"}}}, {{"output", {"c1"}}}, {}, block); + AddOp("one_two", {{"input", {"a"}}}, {{"output", {"b", "c"}}}, + f::AttributeMap{}, block); + AddOp("one_one", {{"input", {"b"}}}, {{"output", {"b1"}}}, f::AttributeMap{}, + block); + AddOp("one_one", {{"input", {"c"}}}, {{"output", {"c1"}}}, f::AttributeMap{}, + block); - f::ProgramDesc *pdesc = program.Proto(); + f::proto::ProgramDesc *pdesc = program.Proto(); pdesc->mutable_blocks(0)->mutable_ops(1)->set_is_target(true); pdesc->mutable_blocks(0)->mutable_ops(2)->set_is_target(true); - f::ProgramDesc pruned; - Prune(*pdesc, &pruned); + f::proto::ProgramDesc pruned; + f::Prune(*pdesc, &pruned); PADDLE_ENFORCE_EQ(pruned.blocks(0).ops_size(), 3); } diff --git a/paddle/framework/scope.cc b/paddle/framework/scope.cc index 9ad6272c99dd6a85520ae44c1331ac232bc6a9a2..0c01d605bcd95f5796fba1e5a3351a2640b2898a 100644 --- a/paddle/framework/scope.cc +++ b/paddle/framework/scope.cc @@ -36,12 +36,9 @@ Scope& Scope::NewScope() const { } Variable* Scope::Var(const std::string& name) { - auto iter = vars_.find(name); - if (iter != vars_.end()) { - VLOG(3) << "Get existing variable " << name; - return iter->second; - } - Variable* v = new Variable(); + auto* v = FindVarLocally(name); + if (v != nullptr) return v; + v = new Variable(); vars_[name] = v; VLOG(3) << "Create variable " << name; v->name_ = &(vars_.find(name)->first); @@ -57,8 +54,10 @@ Variable* Scope::Var(std::string* name) { } Variable* Scope::FindVar(const std::string& name) const { - auto it = vars_.find(name); - if (it != vars_.end()) return it->second; + auto var = FindVarLocally(name); + if (var != nullptr) { + return var; + } return (parent_ == nullptr) ? nullptr : parent_->FindVar(name); } @@ -75,17 +74,9 @@ void Scope::DropKids() { kids_.clear(); } -std::vector Scope::GetAllNames(bool recursive) const { - std::vector known_vars(vars_.size()); - - if (recursive) { - for (auto& kid : kids_) { - auto kid_vars = kid->GetAllNames(); - for (auto& p : kid_vars) { - known_vars.emplace_back(p); - } - } - } +std::vector Scope::LocalVarNames() const { + std::vector known_vars; + known_vars.reserve(this->vars_.size()); for (auto& p : vars_) { known_vars.emplace_back(p.first); } @@ -116,6 +107,11 @@ std::string Scope::Rename(const std::string& origin_name) const { Rename(origin_name, var_name); return var_name; } +Variable* Scope::FindVarLocally(const std::string& name) const { + auto it = vars_.find(name); + if (it != vars_.end()) return it->second; + return nullptr; +} } // namespace framework } // namespace paddle diff --git a/paddle/framework/scope.h b/paddle/framework/scope.h index c2aafb6ad825f9bd9ffef754923a15afdeaa8e5c..10143326dfa201894c777b3e5e226d5ca5015eda 100644 --- a/paddle/framework/scope.h +++ b/paddle/framework/scope.h @@ -66,7 +66,7 @@ class Scope { void DropKids(); // enumerate all the variables current contains. - std::vector GetAllNames(bool recursive = false) const; + std::vector LocalVarNames() const; // Rename variable to a new name void Rename(const std::string& origin_name, @@ -76,6 +76,8 @@ class Scope { std::string Rename(const std::string& origin_name) const; private: + Variable* FindVarLocally(const std::string& name) const; + // Call Scope::NewScope for a sub-scope. explicit Scope(Scope const* parent) : parent_(parent) {} diff --git a/paddle/framework/scope_test.cc b/paddle/framework/scope_test.cc index f738d5ba9ecda57ea25bb5f84057d1d0106eef66..0f5b86061dbdebde08badca7f984f4a2c8d7bc79 100644 --- a/paddle/framework/scope_test.cc +++ b/paddle/framework/scope_test.cc @@ -61,7 +61,7 @@ TEST(Scope, GetAllNames) { Variable* v = s.Var("a"); EXPECT_EQ(&s, s.FindScope(v)); - std::vector ans = s.GetAllNames(); + std::vector ans = s.LocalVarNames(); std::string str; for (auto& var : ans) { str += var; diff --git a/paddle/framework/selected_rows.cc b/paddle/framework/selected_rows.cc index c74459c9dd7006a24615b1d6df041583088fb25c..82adfa7123a3cf40d929021602c45fe7d2e34ffa 100644 --- a/paddle/framework/selected_rows.cc +++ b/paddle/framework/selected_rows.cc @@ -12,5 +12,58 @@ limitations under the License. */ #include "paddle/framework/selected_rows.h" namespace paddle { -namespace framework {} // namespace framework +namespace framework { +void SerializeToStream(std::ostream& os, const SelectedRows& selected_rows, + const platform::DeviceContext& dev_ctx) { + { // the 1st field, uint32_t version + constexpr uint32_t version = 0; + os.write(reinterpret_cast(&version), sizeof(version)); + } + { + // the 2st field, rows information + auto& rows = selected_rows.rows(); + uint64_t size = rows.size(); + os.write(reinterpret_cast(&size), sizeof(size)); + for (uint64_t i = 0; i < size; ++i) { + os.write(reinterpret_cast(&rows[i]), sizeof(rows[i])); + } + } + { + // the 3st field, the height of SelectedRows + int64_t height = selected_rows.height(); + os.write(reinterpret_cast(&height), sizeof(height)); + } + // the 4st field, Tensor data + SerializeToStream(os, selected_rows.value(), dev_ctx); +} + +void DeserializeFromStream(std::istream& is, SelectedRows* selected_rows) { + auto tensor = *selected_rows->mutable_value(); + { + // the 1st field, unit32_t version for SelectedRows + uint32_t version; + is.read(reinterpret_cast(&version), sizeof(version)); + PADDLE_ENFORCE_EQ(version, 0U, "Only version 0 is supported"); + } + { + // the 2st field, rows information + uint64_t size; + is.read(reinterpret_cast(&size), sizeof(size)); + auto& rows = *selected_rows->mutable_rows(); + rows.resize(size); + for (uint64_t i = 0; i < size; ++i) { + is.read(reinterpret_cast(&rows[i]), sizeof(int64_t)); + } + } + { + // the 3st field, the height of the SelectedRows + int64_t height; + is.read(reinterpret_cast(&height), sizeof(int64_t)); + selected_rows->set_height(height); + } + // the 4st field, tensor which contains the data + DeserializeFromStream(is, &tensor); +} + +} // namespace framework } // namespace paddle diff --git a/paddle/framework/selected_rows.h b/paddle/framework/selected_rows.h index 0332b91323e3a4b4b80e02302ad3dcafe0986cde..699e392688e9889f050592172f8bfc45f855d0b1 100644 --- a/paddle/framework/selected_rows.h +++ b/paddle/framework/selected_rows.h @@ -59,5 +59,14 @@ class SelectedRows { int64_t height_; }; +/* + * Serialize/Desiralize SelectedRows to std::ostream + * You can pass ofstream or ostringstream to serilize to file + * or to a in memory string. GPU tensor will be copied to CPU. + */ +void SerializeToStream(std::ostream& os, const SelectedRows& selected_rows, + const platform::DeviceContext& dev_ctx); +void DeserializeFromStream(std::istream& is, SelectedRows* selected_rows); + } // namespace framework } // namespace paddle diff --git a/paddle/framework/selected_rows_test.cc b/paddle/framework/selected_rows_test.cc index 4ee13a65d72e44693573397bb686b355effb2227..75487c4010391aa9e519d73058184fa936dabb84 100644 --- a/paddle/framework/selected_rows_test.cc +++ b/paddle/framework/selected_rows_test.cc @@ -43,5 +43,19 @@ TEST_F(SelectedRowsTester, complete_dims) { ASSERT_EQ(selected_rows_->GetCompleteDims(), make_ddim({10, 100})); } +TEST_F(SelectedRowsTester, SerializeAndDeseralize) { + SelectedRows dst_tensor; + platform::CPUDeviceContext cpu_ctx(place_); + std::ostringstream oss; + + SerializeToStream(oss, *selected_rows_, cpu_ctx); + + std::istringstream iss(oss.str()); + DeserializeFromStream(iss, &dst_tensor); + + ASSERT_EQ(selected_rows_->rows(), dst_tensor.rows()); + ASSERT_EQ(selected_rows_->height(), dst_tensor.height()); +} + } // namespace framework } // namespace paddle diff --git a/paddle/framework/shape_inference.cc b/paddle/framework/shape_inference.cc index 0af41b164f5894db17b2f86d4eba371cf05e3b41..e53cc0cdabc623ae358f1a3e21823a2f38ec3c62 100644 --- a/paddle/framework/shape_inference.cc +++ b/paddle/framework/shape_inference.cc @@ -1,17 +1,19 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +/* 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 +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 + 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. */ +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/framework/shape_inference.h" +#include "grad_op_desc_maker.h" +#include "paddle/framework/operator.h" namespace paddle { namespace framework { @@ -22,6 +24,12 @@ std::vector InferShapeContext::GetInputsDim( return GetDims(names); } +DDim InferShapeContext::GetInputsElementDim(const std::string &name, + int idx) const { + const std::vector &names = Inputs(name); + return this->GetDim(names[idx]); +} + void InferShapeContext::SetOutputsDim( const std::string &name, const std::vector &dims) { auto &names = Outputs(name); @@ -43,20 +51,23 @@ void InferShapeContext::SetDims(const std::vector &names, size_t length = names.size(); PADDLE_ENFORCE_EQ(length, dims.size()); for (size_t i = 0; i < length; ++i) { + if (names[i] == framework::kEmptyVarName) { + continue; + } SetDim(names[i], dims[i]); } } -std::vector InferShapeContext::GetInputsVarType( +std::vector InferShapeContext::GetInputsVarType( const std::string &name) const { return GetVarTypes(Inputs(name)); } -std::vector InferShapeContext::GetOutputsVarType( +std::vector InferShapeContext::GetOutputsVarType( const std::string &name) const { return GetVarTypes(Outputs(name)); } -std::vector InferShapeContext::GetVarTypes( +std::vector InferShapeContext::GetVarTypes( const std::vector &names) const { - std::vector retv; + std::vector retv; retv.resize(names.size()); std::transform(names.begin(), names.end(), retv.begin(), std::bind(std::mem_fn(&InferShapeContext::GetVarType), this, diff --git a/paddle/framework/shape_inference.h b/paddle/framework/shape_inference.h index 05dc47f06ac81f0acb6d0317cbecb3009c7dd7f0..f93319d8f2fd4c5d388bd57fd595a6a5edd51775 100644 --- a/paddle/framework/shape_inference.h +++ b/paddle/framework/shape_inference.h @@ -27,8 +27,9 @@ class InferShapeContext { virtual bool HasInput(const std::string &name) const = 0; virtual bool HasOutput(const std::string &name) const = 0; - std::vector GetInputsVarType(const std::string &name) const; - std::vector GetOutputsVarType( + std::vector GetInputsVarType( + const std::string &name) const; + std::vector GetOutputsVarType( const std::string &name) const; virtual bool HasInputs(const std::string &name) const = 0; @@ -37,6 +38,7 @@ class InferShapeContext { virtual framework::DDim GetInputDim(const std::string &name) const = 0; std::vector GetInputsDim(const std::string &name) const; + DDim GetInputsElementDim(const std::string &name, int idx) const; virtual void SetOutputDim(const std::string &name, const DDim &dim) = 0; void SetOutputsDim(const std::string &name, @@ -64,10 +66,10 @@ class InferShapeContext { std::vector GetDims( const std::vector &names) const; - std::vector GetVarTypes( + std::vector GetVarTypes( const std::vector &names) const; - virtual VarDesc::VarType GetVarType(const std::string &name) const = 0; + virtual proto::VarDesc::VarType GetVarType(const std::string &name) const = 0; }; } // namespace framework diff --git a/paddle/framework/tensor.cc b/paddle/framework/tensor.cc index ea7b2a1f7b17d9abc2c2e14de5ecd1cf4a1a5027..f922e606249849e621e679f71d6dbe0f007c8464 100644 --- a/paddle/framework/tensor.cc +++ b/paddle/framework/tensor.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/framework/tensor.h" diff --git a/paddle/framework/tensor.h b/paddle/framework/tensor.h index 28d0fcf94ec31c82476e093f93ccee222a0c9d9a..341a6949beeb2dfa64b23d2079bd8f48750a94f8 100644 --- a/paddle/framework/tensor.h +++ b/paddle/framework/tensor.h @@ -20,12 +20,12 @@ limitations under the License. */ #include #include +#include "paddle/framework/data_layout.h" #include "paddle/framework/ddim.h" #include "paddle/memory/memory.h" #include "paddle/platform/device_context.h" #include "paddle/platform/enforce.h" #include "paddle/platform/place.h" -#include "unsupported/Eigen/CXX11/Tensor" namespace paddle { @@ -89,34 +89,6 @@ class Tensor { /*! The internal of two tensors share the same memory block. */ inline Tensor& ShareDataWith(const Tensor& src); - /** - * @brief Copy the content of external tensor to a new place. - * - * @param[in] src The external tensor. - * @param[in] dst_place The dst place. - * @param[in] ctx The device context contains device resources. - * - * @note CopyFrom supports CPU <-> GPU, GPU <-> GPU. - */ - // TODO(qijun): https://github.com/PaddlePaddle/Paddle/issues/4647 - // Remove `CopyFrom` and `CopyFromVector` from Tensor interface - // and make them global functions - inline void CopyFrom(const Tensor& src, const platform::Place& dst_place, - const platform::DeviceContext& ctx); - - /** - * @brief Copy the content of an external vector to a tensor. - * - * @param[in] src The external tensor. - * @param[in] ctx The device context contains device resources. - * - * * @note CopyFromVector assumes that the tensor has been resized - * before invoking. - */ - template - inline void CopyFromVector(const std::vector& src, - const platform::DeviceContext& ctx); - /** * @brief Return a sub-tensor of the given tensor. * @@ -141,9 +113,12 @@ class Tensor { size_t memory_size() const; - private: inline void check_memory_size() const; + inline DataLayout layout() const { return layout_; } + + inline void set_layout(const DataLayout layout) { layout_ = layout; } + private: friend class LoDTensor; @@ -202,6 +177,19 @@ class Tensor { DDim dims_; + /** + * @brief the layout of memory block, default is NHWC. + * + * @note the memory allocation order, describe how weight/data is stored + * For example, in 4-D Tensor(rank=4), there are three commonly + * used layout. They are + * NCHW, NHWC, CHWN. + * N,C,H,W for respectively the batch size, the number of + * feature maps, the height. + */ + + DataLayout layout_ = DataLayout::kNHWC; + /** * @brief A PlaceHolder may be shared by more than one tensor. * diff --git a/paddle/framework/tensor.md b/paddle/framework/tensor.md index 7a80816d8e4ffa3a9462f3d9b87eff0f048466aa..0a27ac9bb6b03649d42e12100fda9e80a56e7f56 100644 --- a/paddle/framework/tensor.md +++ b/paddle/framework/tensor.md @@ -71,7 +71,7 @@ private: ``` ```c++ -typedef boost::variant Place; +typedef boost::variant Place; typedef boost::variant, Dim<2>, Dim<3>, Dim<4>, Dim<5>, Dim<6>, Dim<7>, Dim<8>, Dim<9>> DDimVar; typedef boost::variant< diff --git a/paddle/framework/tensor_array.cc b/paddle/framework/tensor_array.cc deleted file mode 100644 index 0947e33548130a923e998f8bad68db00097af909..0000000000000000000000000000000000000000 --- a/paddle/framework/tensor_array.cc +++ /dev/null @@ -1,444 +0,0 @@ -/* 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. */ - -#include "paddle/framework/tensor_array.h" - -#include -#include -#include - -#include "paddle/framework/eigen.h" - -namespace paddle { -namespace framework { - -namespace detail { - -/* - * Offer an iterator over the length-sorted lod-tensor's top level. The top - * level of a lod-tensor stores batch-size of sequences, each top-level sequence - * may contains several lower-level sequences, sort top-level lod by the numbers - * of lower-level sequences in descending order, so that during RNN's running, - * the batch-size will keep decreasing, the short sentences will end at the tail - * of each batch. - * - * Let's take a simple lod-tensor for example - * - * |(0) |(1) top-level has two instances - * ||| ||||| lower-level - * - * sort by lower-level's length - * - * |(1) |(0) - * ||||| ||| - * - * when RNN runs, it get 5 batches (equals the number of elements the longest - * sequence has) - * - * ||||| - * ||| - * - * the first three batches has two elements, the last two elements just has 1 - * element each. - */ -struct DynamicBatchUnpacker { - using value_type = float; - - DynamicBatchUnpacker(const LoDTensor& source, size_t level, - bool descend = true) - : source(&source), level(level) { - BuildLengthSortedMeta(descend); - } - - LoDTensor GetBatch(size_t index); - - std::vector meta; - - LoDTensor const* source; - size_t level; - - protected: - void BuildLengthSortedMeta(bool descend); -}; - -LoDTensor PackDynamicBatch(const std::vector& source, - const std::vector& meta, const LoD& lod, - size_t level); - -std::vector GenDyBatchIndice(const DySeqMetaBatch& meta, int batch_id) { - // collect indice need to copy to the batch - std::vector indice; - for (const auto& seq : meta) { - size_t id = seq.begin + batch_id; - if (id >= seq.end) break; - indice.push_back(id); - } - return indice; -} - -} // namespace detail - -const LoDTensor& TensorArray::Read(size_t index) const { - PADDLE_ENFORCE_LE(index, MAX_SIZE, "index[%d] too large", index); - if (index >= size()) { - values_.resize(index + 1); - } - return values_[index]; -} - -void TensorArray::Write(size_t index, const LoDTensor& value) { - PADDLE_ENFORCE_LE(index, MAX_SIZE, "index[%d] too large", index); - - if (index >= size()) { - values_.resize(index + 1); - } - - values_[index].set_lod(value.lod()); - values_[index].Resize(value.dims()); - values_[index].mutable_data(value.place()); - values_[index].CopyFrom(value, value.place(), platform::CPUDeviceContext()); -} - -void TensorArray::WriteShared(size_t index, const LoDTensor& value) { - PADDLE_ENFORCE_LE(index, MAX_SIZE, "index[%d] too large", index); - if (index >= size()) { - values_.resize(index + 1); - } - - values_[index].set_lod(value.lod()); - values_[index].ShareDataWith(value); -} - -LoDTensor TensorArray::Pack(size_t level, const std::vector& meta, - const LoD& lod) const { - return detail::PackDynamicBatch(values_, meta, lod, level); -} - -DySeqMetaBatch TensorArray::Unpack(const LoDTensor& source, int level, - bool length_desend) { - detail::DynamicBatchUnpacker unpacker(source, level, - length_desend /*descend*/); - - // find max length of all the sequences - size_t max_length = 0; - for (const auto& seq : unpacker.meta) { - max_length = std::max(max_length, seq.end - seq.begin); - } - - // write batches to values - for (size_t batch_id = 0; batch_id < max_length; batch_id++) { - Write(batch_id, unpacker.GetBatch(batch_id)); - } - - PADDLE_ENFORCE(!unpacker.meta.empty()); - return unpacker.meta; -} - -LoDTensor TensorArray::LodPack(size_t level) const { - PADDLE_ENFORCE_GT(size(), 0UL, "no time step exists"); - // the levels should be no less than 2 - LoDTensor merged; - const LoDTensor *pre, *cur; - pre = &Read(0); - - for (size_t step = 1; step < size(); step++) { - cur = &Read(step); - PADDLE_ENFORCE_GT(cur->NumLevels(), 0); - PADDLE_ENFORCE_GT(pre->NumLevels(), 0); - PADDLE_ENFORCE_EQ(pre->NumLevels(), cur->NumLevels()); - PADDLE_ENFORCE_EQ(pre->NumElements(level), cur->NumElements(level)); - - merged = LodPackTwo(*pre, *cur, level); - pre = &merged; - } - return merged; -} - -/* - * NOTE currently, only the lowest level supports packing. - * The lowest LoD will be changed, while the relative offsets in levels above - * stay unchanged. - * - * previous step : [0] [1] [3] - * current step: [0 1 2] [2 3] [] - * packed to - * [0 0] [0 1] [0 2] [1 2] [1 3] [3] - */ -LoDTensor TensorArray::LodPackTwo(const LoDTensor& pre, const LoDTensor& cur, - size_t level) const { - PADDLE_ENFORCE_EQ(pre.NumLevels(), cur.NumLevels()); - PADDLE_ENFORCE_EQ(pre.NumLevels(), level + 1, - "Only the lowest LoD level supports pack temporarily."); - // calculate the result tensor's shape first - size_t num_instances = 0; - for (size_t elem = 0; elem < pre.NumElements(level); elem++) { - size_t prefix_size = pre.NumElements(level, elem); - size_t num_candidates = cur.NumElements(level, elem); - if (num_candidates > 0) { - num_instances += num_candidates * (prefix_size + 1); - } else { - num_instances += prefix_size; - } - } - - auto res_dims = pre.dims(); - res_dims[0] = num_instances; - LoDTensor result; - result.Resize(res_dims); - result.mutable_data(cur.place()); - - Vector last_lod_level; - // copy data - size_t index = 0; - last_lod_level.push_back(index); - for (size_t elem = 0; elem < pre.NumElements(level); elem++) { - size_t prefix_size = pre.NumElements(level, elem); - size_t num_candidates = cur.NumElements(level, elem); - - // slice the prefix Tensor - LoDTensor prefix = pre; - prefix.ShrinkInLevel(level, elem, elem + 1); - LoDTensor candidate = cur; - if (num_candidates > 0) { - candidate.ShrinkInLevel(level, elem, elem + 1); - } else { // just push prefix - result.Slice(index, index + prefix_size) - .CopyFrom(prefix, result.place(), platform::CPUDeviceContext()); - index += prefix_size; - last_lod_level.push_back(index); - } - for (size_t candi = 0; candi < num_candidates; candi++) { - // TODO(superjom) support GPU - result.Slice(index, index + prefix_size) - .CopyFrom(prefix, result.place(), platform::CPUDeviceContext()); - index += prefix_size; - // copy candidate record - result.Slice(index, index + 1) - .CopyFrom(candidate.Slice(candi, candi + 1), result.place(), - platform::CPUDeviceContext()); - index++; - last_lod_level.push_back(index); - } - } - - // update lod - auto lod = cur.lod(); - lod.back() = last_lod_level; - result.set_lod(lod); - return result; -} - -/* - * source [0 1 2] [3 4] [5 6 7] will be transformd to a list of LoDTensors such - * as - * [0 3 5] [1 4 6] [2 7] with 1-level LoDs: - * - [0 1 2 3] - * - [0 1 2 3] - * - [0 1 1 2], the [1,1) here means the second sequence is empty - * - * NOTE Unpack a LoDTensor in this approach may result in a big LoD. - */ -void TensorArray::LodUnpack(const LoDTensor& source, size_t level) { - PADDLE_ENFORCE_EQ(level, source.NumLevels() - 1, - "only the lowest LoD level supports unpack."); - const size_t non_empty_instances = source.dims()[0]; - size_t index = 0; - Vector lowest_lod_level; - lowest_lod_level.push_back(index); - - for (size_t step = 0; step < non_empty_instances; step++) { - size_t num_instances = 0; - for (size_t id = 0; id < source.NumElements(level); id++) { - auto instance = source; - instance.ShrinkInLevel(level, id, id + 1); - if (static_cast(instance.dims()[0]) > step) { - num_instances++; - index++; - } - lowest_lod_level.push_back(index); - } - - // create tensor for this time step - LoDTensor tensor; - auto dims = source.dims(); - dims[0] = num_instances; - // set lod - auto lod = source.lod(); - lod.back() = lowest_lod_level; - tensor.set_lod(lod); - - index = 0; - for (size_t id = 0; id < source.NumElements(level); id++) { - auto instance = source; - instance.ShrinkInLevel(level, id, id + 1); - if (static_cast(instance.dims()[0]) > step) { - // copy this instance - tensor.Slice(index, index + 1) - .CopyFrom(instance.Slice(step, step + 1), tensor.place(), - platform::CPUDeviceContext()); - index++; - } - } - Write(step, tensor); - } -} - -LoDTensor TensorArray::Stack() const { - LoDTensor result; - if (size() == 0) return result; - - const auto& first_dims = values_.front().dims(); - // check all the values have the same shape - // TODO(superjom) check the same dtypes - for (size_t idx = 1; idx < size(); idx++) { - const auto& value_dims = values_[idx].dims(); - PADDLE_ENFORCE_EQ(first_dims, value_dims); - } - - // copy - auto result_dims = vectorize(first_dims); - result_dims.insert(result_dims.begin(), size()); - result.Resize(make_ddim(result_dims)); - result.mutable_data(platform::CPUPlace()); - - for (size_t idx = 0; idx < size(); idx++) { - result.Slice(idx, idx + 1) - .CopyFrom(Read(idx), platform::CPUPlace(), - platform::CPUDeviceContext()); - } - return result; -} - -void TensorArray::Unstack(const LoDTensor& source) const { - Unstack(source, false /*data_shared*/); -} - -void TensorArray::UnstackShared(const LoDTensor& source) const { - Unstack(source, true /*data_shared*/); -} - -void TensorArray::Unstack(const LoDTensor& source, bool data_shared) const { - size_t first_dim = source.dims()[0]; - DDim value_dims = slice_ddim(source.dims(), 1, source.dims().size()); - PADDLE_ENFORCE_GT(first_dim, 0, - "source should have some data to be unstacked"); - - values_.resize(first_dim); - - for (size_t elem = 0; elem < first_dim; elem++) { - // create a new value - auto& value = values_[elem]; - if (data_shared) { - // share memory - value.ShareDataWith(source.Slice(elem, elem + 1)); - } else { - // copy - value.Resize(value_dims); - value.CopyFrom(source.Slice(elem, elem + 1), platform::CPUPlace(), - platform::CPUDeviceContext()); - } - } -} - -size_t TensorArray::size() const { return values_.size(); } - -namespace detail { - -void DynamicBatchUnpacker::BuildLengthSortedMeta(bool descend) { - PADDLE_ENFORCE(meta.empty(), "duplicate build meta"); - // collect meta for each sequence in some level - auto lod = SliceLevels(source->lod(), level, level + 1)[0]; - - for (size_t seq_id = 0; seq_id < lod.size() - 1; seq_id++) { - DySeqMeta seq_meta({lod[seq_id], lod[seq_id + 1], seq_id}); - meta.push_back(seq_meta); - } - - PADDLE_ENFORCE_GT(meta.size(), 0, "meta is empty"); - - // sort by length - sort(meta.begin(), meta.end(), - [descend](const DySeqMeta& a, const DySeqMeta& b) { - bool a_ge_b = (a.end - a.begin) > (b.end - b.begin); - return descend ? a_ge_b : !a_ge_b; - }); -} - -LoDTensor DynamicBatchUnpacker::GetBatch(size_t index) { - PADDLE_ENFORCE(!meta.empty(), "should build meta first"); - LoDTensor result; - - auto indice = detail::GenDyBatchIndice(meta, index); - PADDLE_ENFORCE(!indice.empty(), "invalid batch at %d", index); - - // copy the indice of records in LoDTensor - auto record_dims = slice_ddim(source->dims(), 1, source->dims().size()); - auto record_dims_vec = vectorize(record_dims); - record_dims_vec.insert(record_dims_vec.begin(), indice.size()); - result.Resize(make_ddim(record_dims_vec)); - result.mutable_data(platform::CPUPlace()); - - for (size_t i = 0; i < indice.size(); i++) { - auto index = indice[i]; - auto target = result.Slice(i, i + 1); - auto slice = source->Slice(index, index + 1); - - target.CopyFrom(slice, platform::CPUPlace(), platform::CPUDeviceContext()); - } - - return result; -} - -// TODO(supejom) to cache lod if reasonable -LoDTensor PackDynamicBatch(const std::vector& source, - const std::vector& meta, const LoD& lod, - size_t level) { - PADDLE_ENFORCE(!source.empty()); - PADDLE_ENFORCE(!meta.empty()); - PADDLE_ENFORCE(!lod.empty()); - - LoDTensor result; - - // init result space - auto record_dims = slice_ddim(source[0].dims(), 1, source[0].dims().size()); - auto record_dims_vec = vectorize(record_dims); - auto height = lod[level].back(); - record_dims_vec.insert(record_dims_vec.begin(), height); - result.Resize(make_ddim(record_dims_vec)); - result.mutable_data(platform::CPUPlace()); - - for (size_t batch_id = 0; batch_id < source.size(); batch_id++) { - for (size_t seq_id = 0; seq_id < meta.size(); seq_id++) { - const auto& seq_meta = meta[seq_id]; - // source is source[batch_id][seq_id] - // target is result[index] - auto index = seq_meta.begin + batch_id; - if (index >= seq_meta.end) break; - auto source_ = source[batch_id].Slice(seq_id, seq_id + 1); - auto target = result.Slice(index, index + 1); - target.CopyFrom(source_, platform::CPUPlace(), - platform::CPUDeviceContext()); - } - } - - result.set_lod(lod); - return result; -} - -} // namespace detail - -} // namespace framework -} // namespace paddle diff --git a/paddle/framework/tensor_array.h b/paddle/framework/tensor_array.h deleted file mode 100644 index 78fad8cab7e27a7f07ca542c2a083460ee9e2b79..0000000000000000000000000000000000000000 --- a/paddle/framework/tensor_array.h +++ /dev/null @@ -1,132 +0,0 @@ -/* 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 -#include - -#include "paddle/framework/lod_tensor.h" - -namespace paddle { -namespace framework { - -/* - * DyBatchSeqPosition stores indices of the basic element in tensor. It is used - * after lod-tensor's re-assembling, its info can be used to recover the order - * in original lod-tensor. - */ -struct DySeqMeta { - DySeqMeta(size_t begin, size_t end, size_t ori_idx) - : begin(begin), end(end), ori_idx(ori_idx) {} - - size_t begin; - size_t end; // not included - size_t ori_idx; -}; - -using DySeqMetaBatch = std::vector; - -/* - * Extract the indices of instances. - */ -std::vector GenDyBatchIndice(const DySeqMetaBatch &metas, int batch_id); - -/* - * TensorArray is a C-array-like array of tensors, it is meant to be used with - * dynamic iteration primitives such as while_loop. It is used to segment inputs - * and store states in all time steps. - * - * By providing some methods similar to a C++ array, the difinition of some - * state-based dynamic models such as RNN cound be more natural and highly - * flexible. - */ -class TensorArray { - public: - using value_type = float; - - // max number of values allowed to store. - const size_t MAX_SIZE{100000}; - - /* - * Read the value at location `index` in the `TensorArray`. - */ - const LoDTensor &Read(size_t index) const; - - /* - * Write value into the index of the TensorArray. - */ - void Write(size_t index, const LoDTensor &value); - - /* - * Write value into the index of the TensorArray, with memory shared. - */ - void WriteShared(size_t index, const LoDTensor &value); - - /* - * Recover the original LoD-arranged LoDTensor with the `values`, `level` and - * `indice_map`. - */ - LoDTensor Pack(size_t level, const DySeqMetaBatch &meta, - const LoD &lod) const; - - /* - * Split LoDTensor in some `level` and write the generated batches to - * `values`, if set `desend`, will sort by length in descending order else in - * ascending order. - */ - DySeqMetaBatch Unpack(const LoDTensor &source, int level, bool length_desend); - - /* - * Pack an array of LoDTensors to a LoDTensor. - */ - LoDTensor LodPack(size_t level) const; - - /* - * Unpack a LoDTensor to an array of LoDTensors. - */ - void LodUnpack(const LoDTensor &source, size_t level); - - /* - * Pack the values into a tensor with rank one higher than each tensor in - * values. - */ - LoDTensor Stack() const; - - /* - * Unstacks the given division of a rank-`R` tensor into rank-`(R-1)` tensors. - */ - void Unstack(const LoDTensor &source) const; - - /* - * Unstacks the given division of a rank-`R` tensor into rank-`(R-1)` tensors, - * with memory of tensors shared. - */ - void UnstackShared(const LoDTensor &source) const; - - /* - * Return the number of values. - */ - size_t size() const; - - protected: - void Unstack(const LoDTensor &source, bool data_shared) const; - - LoDTensor LodPackTwo(const LoDTensor &pre, const LoDTensor &cur, - size_t level) const; - - private: - mutable std::vector values_; -}; // class TensorArray - -} // namespace framework -} // namespace paddle diff --git a/paddle/framework/tensor_array_test.cc b/paddle/framework/tensor_array_test.cc deleted file mode 100644 index 83b52b442daf9b2f1fc40f23e458fcb67c5040e8..0000000000000000000000000000000000000000 --- a/paddle/framework/tensor_array_test.cc +++ /dev/null @@ -1,182 +0,0 @@ -/* 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. */ - -#include "paddle/framework/tensor_array.h" - -#include - -namespace paddle { -namespace framework { - -class TensorArrayTester : public ::testing::Test { - protected: - void SetUp() override { - LoDTensor source; - source.Resize(make_ddim({batch_size, dim})); - int* data = source.mutable_data(platform::CPUPlace()); - for (int i = 0; i < 16 * 32; i++) { - data[i] = i; - } - ta.Unstack(source); - } - - TensorArray ta; - const int batch_size = 16; - const int dim = 32; -}; - -TEST_F(TensorArrayTester, Read) { - for (int i = 0; i < batch_size; i++) { - const auto& tensor = ta.Read(i); - ASSERT_EQ(tensor.dims()[0], 1); - ASSERT_EQ(tensor.dims()[1], dim); - } -} - -TEST_F(TensorArrayTester, Write) { - LoDTensor source; - source.Resize(make_ddim({1, dim})); - for (int i = 0; i < dim; i++) { - *(source.mutable_data(platform::CPUPlace()) + i) = i; - } - - ta.Write(2, source); - - const auto& tensor = ta.Read(2); - for (int i = 0; i < dim; i++) { - EXPECT_EQ(*(tensor.data() + i), *(source.data() + i)); - } -} - -TEST_F(TensorArrayTester, WriteShared) { - LoDTensor source; - source.Resize(make_ddim({1, dim})); - for (int i = 0; i < dim; i++) { - *(source.mutable_data(platform::CPUPlace()) + i) = i; - } - - ta.WriteShared(2, source); - - const auto& tensor = ta.Read(2); - for (int i = 0; i < dim; i++) { - EXPECT_EQ(*(tensor.data() + i), *(source.data() + i)); - } - - EXPECT_EQ(source.data(), tensor.data()); -} - -class TensorArrayPackTester : public ::testing::Test { - protected: - virtual void SetUp() override { - lod.push_back(std::vector{0, 2, 9, 13}); - - source.set_lod(lod); - source.Resize(make_ddim({13, 128})); - source.mutable_data(platform::CPUPlace()); - - // content of each setence: 0 1 2 3 4 - const auto& level = lod.front(); - for (size_t i = 0; i < level.size() - 1; i++) { - size_t begin = level[i]; - size_t end = level[i + 1]; - for (size_t j = begin; j < end; j++) { - auto record = source.Slice(j, j + 1); - for (int dim = 0; dim < 128; dim++) { - record.mutable_data(platform::CPUPlace())[dim] = j - begin; - } - } - } - - // unpack - meta = ta.Unpack(source, 0, true); - } - - LoD lod; - TensorArray ta; - LoDTensor source; - std::vector meta; -}; - -TEST_F(TensorArrayPackTester, Unpack) { - ASSERT_EQ(ta.size(), 7UL); - - const auto& t0 = ta.Read(0); - const auto& t1 = ta.Read(1); - - ASSERT_EQ(t0.data()[0], int(0)); - ASSERT_EQ(t1.data()[0], int(1)); -} - -TEST_F(TensorArrayPackTester, Pack) { - LoDTensor packed = ta.Pack(0, meta, lod); -} - -TEST_F(TensorArrayTester, size) { - ASSERT_EQ(ta.size(), static_cast(batch_size)); -} - -TEST(TensorArray, LodPack) { - // three time steps, each step stores a LoDTensors - // - [0] [1] - // - [2 3], [4 5] - // - [6 7] [] [8], [9, 10] - // try to get a LoDTensor with content: - // - [0 2 6] - // - [0 2 7] - // - [0 3] - // - [1 4 8] - // - [1 5 9] - // - [1 5 10] - std::array tensors; - tensors[0].Resize(make_ddim({2, 1})); - tensors[1].Resize(make_ddim({4, 1})); - tensors[2].Resize(make_ddim({5, 1})); - int index = 0; - for (auto& t : tensors) { - t.mutable_data(platform::CPUPlace()); - for (int i = 0; i < t.dims()[0]; i++) { - t.data()[i] = index; - index++; - } - } - - std::array lods; - std::vector> levels{ - {0, 1, 2}, {0, 2, 4}, {0, 2, 2, 3, 5}}; - for (int i = 0; i < 3; i++) { - lods[i].emplace_back(levels[i].begin(), levels[i].end()); - } - - TensorArray ta; - for (int i = 0; i < 3; i++) { - tensors[i].set_lod(lods[i]); - ta.Write(i, tensors[i]); - } - - auto merged = ta.LodPack(0); - - std::vector target_tensor_data{{0, 2, 6, // 0 - 0, 2, 7, // 1 - 0, 3, // 2 - 1, 4, 8, // 3 - 1, 5, 9, // 5 - 1, 5, 10}}; - EXPECT_EQ(merged.dims()[0], (int)target_tensor_data.size()); - for (size_t i = 0; i < target_tensor_data.size(); i++) { - EXPECT_EQ(target_tensor_data[i], merged.data()[i]); - } -} - -} // namespace framework -} // namespace paddle diff --git a/paddle/framework/tensor_impl.h b/paddle/framework/tensor_impl.h index 7e88e039611007d17156d10f852eb46f3ee8e7a3..6c6f298edc187a87677089e54c4c9046821282df 100644 --- a/paddle/framework/tensor_impl.h +++ b/paddle/framework/tensor_impl.h @@ -125,11 +125,11 @@ inline void* Tensor::mutable_data(platform::Place place, std::type_index type) { boost::get(place), size, type)); } else if (platform::is_gpu_place(place)) { #ifndef PADDLE_WITH_CUDA - PADDLE_THROW("'GPUPlace' is not supported in CPU only device."); + PADDLE_THROW("'CUDAPlace' is not supported in CPU only device."); } #else - holder_.reset(new PlaceholderImpl( - boost::get(place), size, type)); + holder_.reset(new PlaceholderImpl( + boost::get(place), size, type)); } #endif offset_ = 0; @@ -150,84 +150,6 @@ inline Tensor& Tensor::ShareDataWith(const Tensor& src) { return *this; } -inline void Tensor::CopyFrom(const Tensor& src, - const platform::Place& dst_place, - const platform::DeviceContext& ctx) { - src.check_memory_size(); - Resize(src.dims()); - - auto src_place = src.holder_->place(); - auto src_ptr = src.data(); - - auto dst_ptr = 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) && - platform::is_cpu_place(dst_place)) { - auto src_gpu_place = boost::get(src_place); - auto dst_cpu_place = boost::get(dst_place); - auto ctx_place = ctx.GetPlace(); - PADDLE_ENFORCE(platform::is_gpu_place(ctx_place)); - auto ctx_gpu_place = boost::get(ctx_place); - PADDLE_ENFORCE_EQ(src_gpu_place, ctx_gpu_place); - memory::Copy( - dst_cpu_place, dst_ptr, src_gpu_place, src_ptr, size, - reinterpret_cast(ctx).stream()); - } 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); - auto ctx_place = ctx.GetPlace(); - PADDLE_ENFORCE(platform::is_gpu_place(ctx_place)); - auto ctx_gpu_place = boost::get(ctx_place); - PADDLE_ENFORCE_EQ(dst_gpu_place, ctx_gpu_place); - memory::Copy( - dst_gpu_place, dst_ptr, src_cpu_place, src_ptr, size, - reinterpret_cast(ctx).stream()); - } 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); - auto ctx_place = ctx.GetPlace(); - PADDLE_ENFORCE(platform::is_gpu_place(ctx_place)); - auto ctx_gpu_place = boost::get(ctx_place); - PADDLE_ENFORCE_EQ(src_gpu_place, ctx_gpu_place); - memory::Copy( - dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size, - reinterpret_cast(ctx).stream()); - } -#endif -} - -template -inline void Tensor::CopyFromVector(const std::vector& src, - const platform::DeviceContext& ctx) { - auto dst_place = ctx.GetPlace(); - auto src_ptr = static_cast(src.data()); - platform::CPUPlace src_place; - auto dst_ptr = static_cast(mutable_data(dst_place)); - auto size = src.size() * sizeof(T); - - if (platform::is_cpu_place(dst_place)) { - memory::Copy(boost::get(dst_place), dst_ptr, src_place, - src_ptr, size); - } -#ifdef PADDLE_WITH_CUDA - else if (platform::is_gpu_place(dst_place)) { - memory::Copy( - boost::get(dst_place), dst_ptr, src_place, src_ptr, - size, - reinterpret_cast(ctx).stream()); - } -#endif -} - inline Tensor Tensor::Slice(int begin_idx, int end_idx) const { check_memory_size(); PADDLE_ENFORCE_GE(begin_idx, 0, @@ -243,6 +165,7 @@ inline Tensor Tensor::Slice(int begin_idx, int end_idx) const { size_t base = numel() / dims_[0]; Tensor dst; dst.holder_ = holder_; + dst.set_layout(layout_); DDim dst_dims = dims_; dst_dims[0] = end_idx - begin_idx; dst.Resize(dst_dims); diff --git a/paddle/framework/tensor_test.cc b/paddle/framework/tensor_test.cc index 1bb0fb71b079940d35a995b78e04a531c074a8b2..a1b4a03289eca4c8b9d8c23ede4221853cb31f79 100644 --- a/paddle/framework/tensor_test.cc +++ b/paddle/framework/tensor_test.cc @@ -15,12 +15,13 @@ #include #include +namespace framework = paddle::framework; +namespace platform = paddle::platform; + TEST(Tensor, Dims) { - using namespace paddle::framework; - using namespace paddle::platform; - Tensor tt; + framework::Tensor tt; tt.Resize({2, 3, 4}); - DDim dims = tt.dims(); + framework::DDim dims = tt.dims(); ASSERT_EQ(arity(dims), 3); for (int i = 0; i < 3; ++i) { EXPECT_EQ(i + 2, dims[i]); @@ -28,12 +29,12 @@ TEST(Tensor, Dims) { } TEST(Tensor, DataAssert) { - paddle::framework::Tensor src_tensor; + framework::Tensor src_tensor; bool caught = false; try { src_tensor.data(); - } catch (paddle::platform::EnforceNotMet err) { + } catch (platform::EnforceNotMet err) { caught = true; std::string msg = "holder_ should not be null\nTensor holds no memory. Call " @@ -50,61 +51,65 @@ TEST(Tensor, DataAssert) { because Memory::Alloc() and Memory::Free() have not been ready. */ TEST(Tensor, MutableData) { - using namespace paddle::framework; - using namespace paddle::platform; { - Tensor src_tensor; + framework::Tensor src_tensor; float* p1 = nullptr; float* p2 = nullptr; // initialization - p1 = src_tensor.mutable_data(make_ddim({1, 2, 3}), CPUPlace()); + p1 = src_tensor.mutable_data(framework::make_ddim({1, 2, 3}), + platform::CPUPlace()); EXPECT_NE(p1, nullptr); // set src_tensor a new dim with large size // momery is supposed to be re-allocated - p2 = src_tensor.mutable_data(make_ddim({3, 4}), CPUPlace()); + p2 = src_tensor.mutable_data(framework::make_ddim({3, 4}), + platform::CPUPlace()); EXPECT_NE(p2, nullptr); EXPECT_NE(p1, p2); // set src_tensor a new dim with same size // momery block is supposed to be unchanged - p1 = src_tensor.mutable_data(make_ddim({2, 2, 3}), CPUPlace()); + p1 = src_tensor.mutable_data(framework::make_ddim({2, 2, 3}), + platform::CPUPlace()); EXPECT_EQ(p1, p2); // set src_tensor a new dim with smaller size // momery block is supposed to be unchanged - p2 = src_tensor.mutable_data(make_ddim({2, 2}), CPUPlace()); + p2 = src_tensor.mutable_data(framework::make_ddim({2, 2}), + platform::CPUPlace()); EXPECT_EQ(p1, p2); } #ifdef PADDLE_WITH_CUDA { - Tensor src_tensor; + framework::Tensor src_tensor; float* p1 = nullptr; float* p2 = nullptr; // initialization - p1 = src_tensor.mutable_data(make_ddim({1, 2, 3}), GPUPlace()); + p1 = src_tensor.mutable_data(framework::make_ddim({1, 2, 3}), + platform::CUDAPlace()); EXPECT_NE(p1, nullptr); // set src_tensor a new dim with large size // momery is supposed to be re-allocated - p2 = src_tensor.mutable_data(make_ddim({3, 4}), GPUPlace()); + p2 = src_tensor.mutable_data(framework::make_ddim({3, 4}), + platform::CUDAPlace()); EXPECT_NE(p2, nullptr); EXPECT_NE(p1, p2); // set src_tensor a new dim with same size // momery block is supposed to be unchanged - p1 = src_tensor.mutable_data(make_ddim({2, 2, 3}), GPUPlace()); + p1 = src_tensor.mutable_data(framework::make_ddim({2, 2, 3}), + platform::CUDAPlace()); EXPECT_EQ(p1, p2); // set src_tensor a new dim with smaller size // momery block is supposed to be unchanged - p2 = src_tensor.mutable_data(make_ddim({2, 2}), GPUPlace()); + p2 = src_tensor.mutable_data(framework::make_ddim({2, 2}), + platform::CUDAPlace()); EXPECT_EQ(p1, p2); } #endif } TEST(Tensor, ShareDataWith) { - using namespace paddle::framework; - using namespace paddle::platform; { - Tensor src_tensor; - Tensor dst_tensor; + framework::Tensor src_tensor; + framework::Tensor dst_tensor; // Try to share data form uninitialized tensor bool caught = false; try { @@ -121,16 +126,18 @@ TEST(Tensor, ShareDataWith) { } ASSERT_TRUE(caught); - src_tensor.mutable_data(make_ddim({2, 3, 4}), CPUPlace()); + src_tensor.mutable_data(framework::make_ddim({2, 3, 4}), + platform::CPUPlace()); dst_tensor.ShareDataWith(src_tensor); ASSERT_EQ(src_tensor.data(), dst_tensor.data()); } #ifdef PADDLE_WITH_CUDA { - Tensor src_tensor; - Tensor dst_tensor; - src_tensor.mutable_data(make_ddim({2, 3, 4}), GPUPlace()); + framework::Tensor src_tensor; + framework::Tensor dst_tensor; + src_tensor.mutable_data(framework::make_ddim({2, 3, 4}), + platform::CUDAPlace()); dst_tensor.ShareDataWith(src_tensor); ASSERT_EQ(src_tensor.data(), dst_tensor.data()); } @@ -138,13 +145,12 @@ TEST(Tensor, ShareDataWith) { } TEST(Tensor, Slice) { - using namespace paddle::framework; - using namespace paddle::platform; { - Tensor src_tensor; - src_tensor.mutable_data(make_ddim({5, 3, 4}), CPUPlace()); - Tensor slice_tensor = src_tensor.Slice(1, 3); - DDim slice_dims = slice_tensor.dims(); + framework::Tensor src_tensor; + src_tensor.mutable_data(framework::make_ddim({5, 3, 4}), + platform::CPUPlace()); + framework::Tensor slice_tensor = src_tensor.Slice(1, 3); + framework::DDim slice_dims = slice_tensor.dims(); ASSERT_EQ(arity(slice_dims), 3); EXPECT_EQ(slice_dims[0], 2); EXPECT_EQ(slice_dims[1], 3); @@ -153,11 +159,12 @@ TEST(Tensor, Slice) { uintptr_t src_data_address = reinterpret_cast(src_tensor.data()); uintptr_t src_mutable_data_address = reinterpret_cast( - src_tensor.mutable_data(src_tensor.dims(), CPUPlace())); + src_tensor.mutable_data(src_tensor.dims(), platform::CPUPlace())); uintptr_t slice_data_address = reinterpret_cast(slice_tensor.data()); - uintptr_t slice_mutable_data_address = reinterpret_cast( - slice_tensor.mutable_data(slice_tensor.dims(), CPUPlace())); + uintptr_t slice_mutable_data_address = + reinterpret_cast(slice_tensor.mutable_data( + slice_tensor.dims(), platform::CPUPlace())); EXPECT_EQ(src_data_address, src_mutable_data_address); EXPECT_EQ(slice_data_address, slice_mutable_data_address); EXPECT_EQ(src_data_address + 3 * 4 * 1 * sizeof(int), slice_data_address); @@ -165,22 +172,25 @@ TEST(Tensor, Slice) { #ifdef PADDLE_WITH_CUDA { - Tensor src_tensor; - src_tensor.mutable_data(make_ddim({6, 9}), GPUPlace()); - Tensor slice_tensor = src_tensor.Slice(2, 6); - DDim slice_dims = slice_tensor.dims(); + framework::Tensor src_tensor; + src_tensor.mutable_data(framework::make_ddim({6, 9}), + platform::CUDAPlace()); + framework::Tensor slice_tensor = src_tensor.Slice(2, 6); + framework::DDim slice_dims = slice_tensor.dims(); ASSERT_EQ(arity(slice_dims), 2); EXPECT_EQ(slice_dims[0], 4); EXPECT_EQ(slice_dims[1], 9); uintptr_t src_data_address = reinterpret_cast(src_tensor.data()); - uintptr_t src_mutable_data_address = reinterpret_cast( - src_tensor.mutable_data(src_tensor.dims(), GPUPlace())); + uintptr_t src_mutable_data_address = + reinterpret_cast(src_tensor.mutable_data( + src_tensor.dims(), platform::CUDAPlace())); uintptr_t slice_data_address = reinterpret_cast(slice_tensor.data()); - uintptr_t slice_mutable_data_address = reinterpret_cast( - slice_tensor.mutable_data(slice_tensor.dims(), GPUPlace())); + uintptr_t slice_mutable_data_address = + reinterpret_cast(slice_tensor.mutable_data( + slice_tensor.dims(), platform::CUDAPlace())); EXPECT_EQ(src_data_address, src_mutable_data_address); EXPECT_EQ(slice_data_address, slice_mutable_data_address); EXPECT_EQ(src_data_address + 9 * 2 * sizeof(double), slice_data_address); @@ -188,187 +198,20 @@ TEST(Tensor, Slice) { #endif } -TEST(Tensor, CopyFrom) { - using namespace paddle::framework; - using namespace paddle::platform; - { - Tensor src_tensor; - Tensor dst_tensor; - CPUDeviceContext cpu_ctx((CPUPlace())); - - int* src_ptr = src_tensor.mutable_data(make_ddim({3, 3}), CPUPlace()); - - int arr[9] = {1, 2, 3, 4, 5, 6, 7, 8, 9}; - memcpy(src_ptr, arr, 9 * sizeof(int)); - - auto cpu_place = new paddle::platform::CPUPlace(); - dst_tensor.CopyFrom(src_tensor, *cpu_place, cpu_ctx); - - const int* dst_ptr = dst_tensor.data(); - ASSERT_NE(src_ptr, dst_ptr); - for (size_t i = 0; i < 9; ++i) { - EXPECT_EQ(src_ptr[i], dst_ptr[i]); - } - - Tensor slice_tensor = src_tensor.Slice(1, 2); - dst_tensor.CopyFrom(slice_tensor, *cpu_place, cpu_ctx); - const int* slice_ptr = slice_tensor.data(); - dst_ptr = dst_tensor.data(); - ASSERT_NE(dst_ptr, slice_ptr); - for (size_t i = 0; i < 3; ++i) { - EXPECT_EQ(dst_ptr[i], slice_ptr[i]); - } - } -#ifdef PADDLE_WITH_CUDA - { - Tensor src_tensor; - Tensor gpu_tensor; - Tensor dst_tensor; - - int* src_ptr = src_tensor.mutable_data(make_ddim({3, 3}), CPUPlace()); - - int arr[9] = {1, 2, 3, 4, 5, 6, 7, 8, 9}; - memcpy(src_ptr, arr, 9 * sizeof(int)); - - // CPU Tensor to GPU Tensor - auto gpu_place = new paddle::platform::GPUPlace(0); - CUDADeviceContext gpu_ctx(*gpu_place); - gpu_tensor.CopyFrom(src_tensor, *gpu_place, gpu_ctx); - - // GPU Tensor to CPU Tensor - auto cpu_place = new paddle::platform::CPUPlace(); - dst_tensor.CopyFrom(gpu_tensor, *cpu_place, gpu_ctx); - - // Sync before Compare Tensors - gpu_ctx.Wait(); - const int* dst_ptr = dst_tensor.data(); - ASSERT_NE(src_ptr, dst_ptr); - for (size_t i = 0; i < 9; ++i) { - EXPECT_EQ(src_ptr[i], dst_ptr[i]); - } - - Tensor slice_tensor = src_tensor.Slice(1, 2); - - // CPU Slice Tensor to GPU Tensor - gpu_tensor.CopyFrom(slice_tensor, *gpu_place, gpu_ctx); - - // GPU Tensor to CPU Tensor - dst_tensor.CopyFrom(gpu_tensor, *cpu_place, gpu_ctx); - - // Sync before Compare Slice Tensors - gpu_ctx.Wait(); - const int* slice_ptr = slice_tensor.data(); - dst_ptr = dst_tensor.data(); - ASSERT_NE(dst_ptr, slice_ptr); - for (size_t i = 0; i < 3; ++i) { - EXPECT_EQ(dst_ptr[i], slice_ptr[i]); - } - } -#endif -} - -TEST(Tensor, CopyFromVector) { - using namespace paddle::framework; - using namespace paddle::platform; - { - std::vector src_vec = {1, 2, 3, 4, 5, 6, 7, 8, 9}; - Tensor cpu_tensor; - - // Copy to CPU Tensor - cpu_tensor.Resize(make_ddim({3, 3})); - auto cpu_place = new paddle::platform::CPUPlace(); - CPUDeviceContext cpu_ctx(*cpu_place); - cpu_tensor.CopyFromVector(src_vec, cpu_ctx); - - // Compare Tensors - const int* cpu_ptr = cpu_tensor.data(); - const int* src_ptr = src_vec.data(); - ASSERT_NE(src_ptr, cpu_ptr); - for (size_t i = 0; i < 9; ++i) { - EXPECT_EQ(src_ptr[i], cpu_ptr[i]); - } - - src_vec.erase(src_vec.begin(), src_vec.begin() + 5); - cpu_tensor.Resize(make_ddim({2, 2})); - cpu_tensor.CopyFromVector(src_vec, cpu_ctx); - cpu_ptr = cpu_tensor.data(); - src_ptr = src_vec.data(); - ASSERT_NE(src_ptr, cpu_ptr); - for (size_t i = 0; i < 5; ++i) { - EXPECT_EQ(src_ptr[i], cpu_ptr[i]); - } - - delete cpu_place; - } - -#ifdef PADDLE_WITH_CUDA - { - std::vector src_vec = {1, 2, 3, 4, 5, 6, 7, 8, 9}; - Tensor cpu_tensor; - Tensor gpu_tensor; - Tensor dst_tensor; - - // Copy to CPU Tensor - cpu_tensor.Resize(make_ddim({3, 3})); - auto cpu_place = new paddle::platform::CPUPlace(); - CPUDeviceContext cpu_ctx(*cpu_place); - cpu_tensor.CopyFromVector(src_vec, cpu_ctx); - - // Copy to GPUTensor - gpu_tensor.Resize(make_ddim({3, 3})); - auto gpu_place = new paddle::platform::GPUPlace(); - CUDADeviceContext gpu_ctx(*gpu_place); - gpu_tensor.CopyFromVector(src_vec, gpu_ctx); - // Copy from GPU to CPU tensor for comparison - dst_tensor.CopyFrom(gpu_tensor, *cpu_place, gpu_ctx); - - // Sync before Compare Tensors - gpu_ctx.Wait(); - const int* src_ptr = src_vec.data(); - const int* cpu_ptr = cpu_tensor.data(); - const int* dst_ptr = dst_tensor.data(); - ASSERT_NE(src_ptr, cpu_ptr); - ASSERT_NE(src_ptr, dst_ptr); - for (size_t i = 0; i < 9; ++i) { - EXPECT_EQ(src_ptr[i], cpu_ptr[i]); - EXPECT_EQ(src_ptr[i], dst_ptr[i]); - } - - src_vec.erase(src_vec.begin(), src_vec.begin() + 5); - - cpu_tensor.Resize(make_ddim({2, 2})); - cpu_tensor.CopyFromVector(src_vec, cpu_ctx); - gpu_tensor.Resize(make_ddim({2, 2})); - gpu_tensor.CopyFromVector(src_vec, gpu_ctx); - dst_tensor.CopyFrom(gpu_tensor, *cpu_place, gpu_ctx); - - // Sync before Compare Tensors - gpu_ctx.Wait(); - src_ptr = src_vec.data(); - cpu_ptr = cpu_tensor.data(); - dst_ptr = dst_tensor.data(); - ASSERT_NE(src_ptr, cpu_ptr); - ASSERT_NE(src_ptr, dst_ptr); - for (size_t i = 0; i < 5; ++i) { - EXPECT_EQ(src_ptr[i], cpu_ptr[i]); - EXPECT_EQ(src_ptr[i], dst_ptr[i]); - } - - delete cpu_place; - delete gpu_place; - } -#endif -} - TEST(Tensor, ReshapeToMatrix) { - using namespace paddle::framework; - using namespace paddle::platform; - Tensor src; - int* src_ptr = src.mutable_data({2, 3, 4, 9}, CPUPlace()); + framework::Tensor src; + int* src_ptr = src.mutable_data({2, 3, 4, 9}, platform::CPUPlace()); for (int i = 0; i < 2 * 3 * 4 * 9; ++i) { src_ptr[i] = i; } - Tensor res = ReshapeToMatrix(src, 2); + framework::Tensor res = framework::ReshapeToMatrix(src, 2); ASSERT_EQ(res.dims()[0], 2 * 3); ASSERT_EQ(res.dims()[1], 4 * 9); } + +TEST(Tensor, Layout) { + framework::Tensor src; + ASSERT_EQ(src.layout(), framework::DataLayout::kNHWC); + src.set_layout(framework::DataLayout::kAnyLayout); + ASSERT_EQ(src.layout(), framework::DataLayout::kAnyLayout); +} diff --git a/paddle/framework/tensor_util.cc b/paddle/framework/tensor_util.cc new file mode 100644 index 0000000000000000000000000000000000000000..7efc649d0bcda67c663d148e83bcbb6789b0f371 --- /dev/null +++ b/paddle/framework/tensor_util.cc @@ -0,0 +1,119 @@ +/* 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/framework/tensor_util.h" + +namespace paddle { +namespace framework { +template +struct AnyDTypeVisitor { + Predicate predicate_; + const Tensor& tensor_; + const DevCtx& ctx_; + Tensor* out_; + + AnyDTypeVisitor(Predicate predicate, const Tensor& tensor, const DevCtx& ctx, + Tensor* out) + : predicate_(predicate), tensor_(tensor), ctx_(ctx), out_(out) {} + + template + void operator()() const { + auto t = EigenVector::Flatten(tensor_); + auto o = EigenScalar::From(*out_); + // return any of predicate_(t) is true. + o.device(*ctx_.eigen_device()) = predicate_(t).any(); + } +}; + +template +inline void AnyImpl(Predicate predicate, const framework::Tensor& tensor, + const DevCtx& ctx, framework::Tensor* out) { + VisitDataType(ToDataType(tensor.type()), AnyDTypeVisitor( + predicate, tensor, ctx, out)); +} + +template +struct AnyVisitor : public boost::static_visitor { + const framework::Tensor& tensor_; + Predicate predicate_; + + AnyVisitor(const framework::Tensor& tensor, Predicate predicate) + : tensor_(tensor), predicate_(std::move(predicate)) {} + + template + bool operator()(const Place& place) const { + framework::Tensor out; + out.Resize({1}); + out.mutable_data(place); + auto* ctx = platform::DeviceContextPool::Instance().GetByPlace(place); + AnyImpl(predicate_, tensor_, *ctx, &out); + return this->GetResult(out, place); + } + + bool GetResult(const framework::Tensor& out, + const platform::CUDAPlace& gpu) const { + platform::CPUPlace cpu; + framework::Tensor tmp; + tmp.Resize({1}); + tmp.mutable_data(cpu); + auto gpuctx = platform::DeviceContextPool::Instance().Get(gpu); + gpuctx->Wait(); + CopyFrom(out, cpu, *gpuctx, &tmp); + gpuctx->Wait(); + return GetResult(tmp, cpu); + } + + bool GetResult(const framework::Tensor& out, + const platform::CPUPlace& cpu) const { + return *out.data(); + } +}; + +template +inline bool Any(const framework::Tensor& tensor, Predicate predicate) { + AnyVisitor visitor(tensor, predicate); + auto place = tensor.place(); + return platform::VisitPlace(place, visitor); +} + +struct HasNANPredicate { + template + auto operator()(const T& eigen_vec) const + -> decltype(std::declval().isnan()) { + // Cast eigen_vector to vector of bool. true if is inf. + return eigen_vec.isnan(); + } +}; + +bool HasNAN(const framework::Tensor& tensor) { + HasNANPredicate predicate; + return Any(tensor, predicate); +} + +struct HasInfPredicate { + template + auto operator()(const T& eigen_vec) const + -> decltype(std::declval().isinf()) { + // Cast eigen_vector to vector of bool. true if is inf. + return eigen_vec.isinf(); + } +}; + +bool HasInf(const framework::Tensor& tensor) { + HasInfPredicate predicate; + return Any(tensor, predicate); +} + +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/tensor_util.cu b/paddle/framework/tensor_util.cu new file mode 120000 index 0000000000000000000000000000000000000000..b00e6e59d93328bf3142597ea4de0dc225501e56 --- /dev/null +++ b/paddle/framework/tensor_util.cu @@ -0,0 +1 @@ +./tensor_util.cc \ No newline at end of file diff --git a/paddle/framework/tensor_util.h b/paddle/framework/tensor_util.h new file mode 100644 index 0000000000000000000000000000000000000000..6a21f8db1e3966fd23eee0da2346b2d61f9321fb --- /dev/null +++ b/paddle/framework/tensor_util.h @@ -0,0 +1,317 @@ +/* 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 +#include "paddle/framework/data_type.h" +#include "paddle/framework/eigen.h" +#include "paddle/framework/framework.pb.h" +#include "paddle/framework/tensor.h" +#include "paddle/platform/device_context.h" + +namespace paddle { +namespace framework { + +/** + * @brief Copy the content of external tensor to a new place. + * + * @param[in] src The external tensor. + * @param[in] dst_place The dst place. + * @param[in] ctx The device context contains device resources. + * + * @note CopyFrom supports CPU <-> GPU, GPU <-> GPU. + */ + +inline void CopyFrom(const Tensor& src, const platform::Place& dst_place, + const platform::DeviceContext& ctx, Tensor* dst) { + 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); + auto ctx_place = ctx.GetPlace(); + PADDLE_ENFORCE(platform::is_gpu_place(ctx_place)); + auto ctx_gpu_place = boost::get(ctx_place); + PADDLE_ENFORCE_EQ(src_gpu_place, ctx_gpu_place); + memory::Copy( + dst_cpu_place, dst_ptr, src_gpu_place, src_ptr, size, + reinterpret_cast(ctx).stream()); + } 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); + auto ctx_place = ctx.GetPlace(); + PADDLE_ENFORCE(platform::is_gpu_place(ctx_place)); + auto ctx_gpu_place = boost::get(ctx_place); + PADDLE_ENFORCE_EQ(dst_gpu_place, ctx_gpu_place); + memory::Copy( + dst_gpu_place, dst_ptr, src_cpu_place, src_ptr, size, + reinterpret_cast(ctx).stream()); + } 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); + auto ctx_place = ctx.GetPlace(); + PADDLE_ENFORCE(platform::is_gpu_place(ctx_place)); + auto ctx_gpu_place = boost::get(ctx_place); + PADDLE_ENFORCE_EQ(src_gpu_place, ctx_gpu_place); + memory::Copy( + dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size, + reinterpret_cast(ctx).stream()); + } +#endif +} + +/** + * @brief CopyFrom support CPU <-> CPU + */ +inline void CopyFrom(const Tensor& src, const platform::Place& dst_place, + Tensor* dst) { + 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()); + + PADDLE_ENFORCE(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); +} + +/** + * @brief Copy the content of an external vector to a tensor. + * + * @param[in] src The external tensor. + * @param[in] ctx The device context contains device resources. + * + * * @note CopyFromVector assumes that the tensor has been resized + * before invoking. + */ +template +inline void CopyFromVector(const std::vector& src, + const platform::DeviceContext& ctx, Tensor* dst) { + auto dst_place = ctx.GetPlace(); + auto src_ptr = static_cast(src.data()); + platform::CPUPlace src_place; + dst->Resize({static_cast(src.size())}); + auto dst_ptr = static_cast(dst->mutable_data(dst_place)); + auto size = src.size() * sizeof(T); + + if (platform::is_cpu_place(dst_place)) { + memory::Copy(boost::get(dst_place), dst_ptr, src_place, + src_ptr, size); + } +#ifdef PADDLE_WITH_CUDA + else if (platform::is_gpu_place(dst_place)) { // NOLINT + memory::Copy( + boost::get(dst_place), dst_ptr, src_place, src_ptr, + size, + reinterpret_cast(ctx).stream()); + } +#endif +} + +/** + * @brief CopyFromVector CPU vector -> CPU Tensor + */ +template +inline void CopyFromVector(const std::vector& src, Tensor* dst) { + platform::CPUPlace dst_place = platform::CPUPlace(); + auto src_ptr = static_cast(src.data()); + platform::CPUPlace src_place; + dst->Resize({static_cast(src.size())}); + auto dst_ptr = static_cast(dst->mutable_data(dst_place)); + auto size = src.size() * sizeof(T); + + memory::Copy(dst_place, dst_ptr, src_place, src_ptr, size); +} + +/** + * @brief Copy the content of a tensor to a vector + * + * @param[in] src The external tensor. + * @param[in] ctx The device context contains device resources. + * + * * @note CopyFromVector assumes that the tensor has been resized + * before invoking. + */ +template +inline void CopyToVector(const Tensor& src, const platform::DeviceContext& ctx, + std::vector* dst) { + auto src_ptr = static_cast(src.data()); + auto size = src.numel() * sizeof(T); + + platform::CPUPlace dst_place; + dst->resize(src.numel()); + auto dst_ptr = static_cast(dst->data()); + + if (platform::is_cpu_place(src.place())) { + memory::Copy(dst_place, dst_ptr, + boost::get(src.place()), src_ptr, size); + } +#ifdef PADDLE_WITH_CUDA + else if (platform::is_gpu_place(src.place())) { // NOLINT + memory::Copy( + dst_place, dst_ptr, boost::get(src.place()), + src_ptr, size, + reinterpret_cast(ctx).stream()); + } +#endif +} + +/** + * @brief CopyToVector CPUTensor <-> CPU Vector + */ +template +inline void CopyToVector(const Tensor& src, std::vector* dst) { + auto src_ptr = static_cast(src.data()); + auto size = src.numel() * sizeof(T); + + platform::CPUPlace dst_place; + dst->resize(src.numel()); + auto dst_ptr = static_cast(dst->data()); + + PADDLE_ENFORCE(platform::is_cpu_place(src.place())); + + memory::Copy(dst_place, dst_ptr, boost::get(src.place()), + src_ptr, size); +} + +// Returns true if a tensor contains NAN, i.e., Not A Number. +bool HasNAN(const framework::Tensor& tensor); + +// Returns true if a tensor contains Inf, i.e., Infinity. +bool HasInf(const framework::Tensor& tensor); + +inline void SerializeToStream(std::ostream& os, const Tensor& tensor, + const platform::DeviceContext& dev_ctx) { + // TODO(typhoonzero): serialize to ostream + { // the 1st field, uint32_t version + constexpr uint32_t version = 0; + os.write(reinterpret_cast(&version), sizeof(version)); + } + { // the 2nd field, tensor description + // int32_t size + // void* protobuf message + proto::TensorDesc desc; + desc.set_data_type(framework::ToDataType(tensor.type())); + auto dims = framework::vectorize(tensor.dims()); + auto* pb_dims = desc.mutable_dims(); + pb_dims->Resize(static_cast(dims.size()), 0); + std::copy(dims.begin(), dims.end(), pb_dims->begin()); + int32_t size = desc.ByteSize(); + os.write(reinterpret_cast(&size), sizeof(size)); + auto out = desc.SerializeAsString(); + os.write(out.data(), size); + } + { // the 3rd field, tensor data + uint64_t size = tensor.memory_size(); + auto* data_ptr = tensor.data(); + PADDLE_ENFORCE(size < std::numeric_limits::max(), + "Index overflow when writing tensor"); + if (platform::is_gpu_place(tensor.place())) { +#ifdef PADDLE_WITH_CUDA + constexpr size_t kBufSize = 1024 * 1024 * 64; // 64MB + std::unique_ptr buf(new char[kBufSize]); + auto& gpu_dev_ctx = + static_cast(dev_ctx); + platform::CPUPlace cpu; + uintptr_t data = reinterpret_cast(data_ptr); + while (size != 0) { + size_t size_to_write = std::min(kBufSize, static_cast(size)); + memory::Copy(cpu, buf.get(), + boost::get(tensor.place()), + reinterpret_cast(data), size_to_write, + gpu_dev_ctx.stream()); + gpu_dev_ctx.Wait(); + os.write(buf.get(), size_to_write); + data += size_to_write; + size -= size_to_write; + } +#else + PADDLE_THROW("Unexpected branch"); +#endif + } else { + os.write(static_cast(data_ptr), + static_cast(size)); + } + } +} + +inline void DeserializeFromStream(std::istream& is, Tensor* tensor) { + uint32_t version; + is.read(reinterpret_cast(&version), sizeof(version)); + PADDLE_ENFORCE_EQ(version, 0U, "Only version 0 is supported"); + proto::TensorDesc desc; + { // int32_t size + // proto buffer + int32_t size; + is.read(reinterpret_cast(&size), sizeof(size)); + std::unique_ptr buf(new char[size]); + is.read(reinterpret_cast(buf.get()), size); + PADDLE_ENFORCE(desc.ParseFromArray(buf.get(), size), + "Cannot parse tensor desc"); + } + { // read tensor + std::vector dims; + dims.reserve(static_cast(desc.dims().size())); + std::copy(desc.dims().begin(), desc.dims().end(), std::back_inserter(dims)); + tensor->Resize(framework::make_ddim(dims)); + + void* buf; + platform::Place cpu = platform::CPUPlace(); + // TODO(Yancey1989): use VisiterDataType instead of DataType switch + switch (desc.data_type()) { + case proto::FP32: + buf = tensor->mutable_data(cpu); + break; + case proto::FP64: + buf = tensor->mutable_data(cpu); + break; + case proto::INT32: + buf = tensor->mutable_data(cpu); + break; + case proto::INT64: + buf = tensor->mutable_data(cpu); + break; + default: + PADDLE_THROW("DataType %d not supported", desc.data_type()); + } + is.read(static_cast(buf), tensor->memory_size()); + } +} + +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/tensor_util_test.cc b/paddle/framework/tensor_util_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..0dc5166fcabf77b48b8681ab1f050e2bc88f44ab --- /dev/null +++ b/paddle/framework/tensor_util_test.cc @@ -0,0 +1,308 @@ +/* + 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. +*/ + +#include "paddle/framework/tensor_util.h" +#include +#include +#include + +namespace paddle { +namespace framework { + +TEST(CopyFrom, Tensor) { + Tensor src_tensor; + Tensor dst_tensor; + platform::CPUDeviceContext cpu_ctx((platform::CPUPlace())); + + int* src_ptr = + src_tensor.mutable_data(make_ddim({3, 3}), platform::CPUPlace()); + + int arr[9] = {1, 2, 3, 4, 5, 6, 7, 8, 9}; + memcpy(src_ptr, arr, 9 * sizeof(int)); + src_tensor.set_layout(DataLayout::kAnyLayout); + + auto cpu_place = new platform::CPUPlace(); + CopyFrom(src_tensor, *cpu_place, &dst_tensor); + + const int* dst_ptr = dst_tensor.data(); + ASSERT_NE(src_ptr, dst_ptr); + for (size_t i = 0; i < 9; ++i) { + EXPECT_EQ(src_ptr[i], dst_ptr[i]); + } + + EXPECT_TRUE(dst_tensor.layout() == src_tensor.layout()); + + Tensor slice_tensor = src_tensor.Slice(1, 2); + CopyFrom(slice_tensor, *cpu_place, &dst_tensor); + const int* slice_ptr = slice_tensor.data(); + dst_ptr = dst_tensor.data(); + ASSERT_NE(dst_ptr, slice_ptr); + for (size_t i = 0; i < 3; ++i) { + EXPECT_EQ(dst_ptr[i], slice_ptr[i]); + } + EXPECT_TRUE(dst_tensor.layout() == src_tensor.layout()); + +#ifdef PADDLE_WITH_CUDA + { + Tensor src_tensor; + Tensor gpu_tensor; + Tensor dst_tensor; + + int* src_ptr = + src_tensor.mutable_data(make_ddim({3, 3}), platform::CPUPlace()); + + int arr[9] = {1, 2, 3, 4, 5, 6, 7, 8, 9}; + memcpy(src_ptr, arr, 9 * sizeof(int)); + + // CPU Tensor to GPU Tensor + auto gpu_place = new platform::CUDAPlace(0); + platform::CUDADeviceContext gpu_ctx(*gpu_place); + CopyFrom(src_tensor, *gpu_place, gpu_ctx, &gpu_tensor); + + // GPU Tensor to CPU Tensor + auto cpu_place = new platform::CPUPlace(); + CopyFrom(gpu_tensor, *cpu_place, gpu_ctx, &dst_tensor); + + // Sync before Compare Tensors + gpu_ctx.Wait(); + const int* dst_ptr = dst_tensor.data(); + ASSERT_NE(src_ptr, dst_ptr); + for (size_t i = 0; i < 9; ++i) { + EXPECT_EQ(src_ptr[i], dst_ptr[i]); + } + + Tensor slice_tensor = src_tensor.Slice(1, 2); + + // CPU Slice Tensor to GPU Tensor + CopyFrom(slice_tensor, *gpu_place, gpu_ctx, &gpu_tensor); + + // GPU Tensor to CPU Tensor + CopyFrom(gpu_tensor, *cpu_place, gpu_ctx, &dst_tensor); + + // Sync before Compare Slice Tensors + gpu_ctx.Wait(); + const int* slice_ptr = slice_tensor.data(); + dst_ptr = dst_tensor.data(); + ASSERT_NE(dst_ptr, slice_ptr); + for (size_t i = 0; i < 3; ++i) { + EXPECT_EQ(dst_ptr[i], slice_ptr[i]); + } + + EXPECT_TRUE(dst_tensor.layout() == src_tensor.layout()); + } +#endif +} + +TEST(CopyFromVector, Tensor) { + using namespace paddle::framework; + using namespace paddle::platform; + { + std::vector src_vec = {1, 2, 3, 4, 5, 6, 7, 8, 9}; + Tensor cpu_tensor; + + // Copy to CPU Tensor + cpu_tensor.Resize(make_ddim({3, 3})); + auto cpu_place = new paddle::platform::CPUPlace(); + CopyFromVector(src_vec, &cpu_tensor); + + // Compare Tensors + const int* cpu_ptr = cpu_tensor.data(); + const int* src_ptr = src_vec.data(); + ASSERT_NE(src_ptr, cpu_ptr); + for (size_t i = 0; i < 9; ++i) { + EXPECT_EQ(src_ptr[i], cpu_ptr[i]); + } + + src_vec.erase(src_vec.begin(), src_vec.begin() + 5); + cpu_tensor.Resize(make_ddim({2, 2})); + CopyFromVector(src_vec, &cpu_tensor); + cpu_ptr = cpu_tensor.data(); + src_ptr = src_vec.data(); + ASSERT_NE(src_ptr, cpu_ptr); + for (size_t i = 0; i < 5; ++i) { + EXPECT_EQ(src_ptr[i], cpu_ptr[i]); + } + + delete cpu_place; + } + +#ifdef PADDLE_WITH_CUDA + { + std::vector src_vec = {1, 2, 3, 4, 5, 6, 7, 8, 9}; + Tensor cpu_tensor; + Tensor gpu_tensor; + Tensor dst_tensor; + + // Copy to CPU Tensor + cpu_tensor.Resize(make_ddim({3, 3})); + auto cpu_place = new paddle::platform::CPUPlace(); + CPUDeviceContext cpu_ctx(*cpu_place); + CopyFromVector(src_vec, cpu_ctx, &cpu_tensor); + + // Copy to GPUTensor + gpu_tensor.Resize(make_ddim({3, 3})); + auto gpu_place = new paddle::platform::CUDAPlace(); + CUDADeviceContext gpu_ctx(*gpu_place); + CopyFromVector(src_vec, gpu_ctx, &gpu_tensor); + // Copy from GPU to CPU tensor for comparison + CopyFrom(gpu_tensor, *cpu_place, gpu_ctx, &dst_tensor); + + // Sync before Compare Tensors + gpu_ctx.Wait(); + const int* src_ptr = src_vec.data(); + const int* cpu_ptr = cpu_tensor.data(); + const int* dst_ptr = dst_tensor.data(); + ASSERT_NE(src_ptr, cpu_ptr); + ASSERT_NE(src_ptr, dst_ptr); + for (size_t i = 0; i < 9; ++i) { + EXPECT_EQ(src_ptr[i], cpu_ptr[i]); + EXPECT_EQ(src_ptr[i], dst_ptr[i]); + } + + src_vec.erase(src_vec.begin(), src_vec.begin() + 5); + + cpu_tensor.Resize(make_ddim({2, 2})); + CopyFromVector(src_vec, cpu_ctx, &cpu_tensor); + gpu_tensor.Resize(make_ddim({2, 2})); + CopyFromVector(src_vec, gpu_ctx, &gpu_tensor); + CopyFrom(gpu_tensor, *cpu_place, gpu_ctx, &dst_tensor); + + // Sync before Compare Tensors + gpu_ctx.Wait(); + src_ptr = src_vec.data(); + cpu_ptr = cpu_tensor.data(); + dst_ptr = dst_tensor.data(); + ASSERT_NE(src_ptr, cpu_ptr); + ASSERT_NE(src_ptr, dst_ptr); + for (size_t i = 0; i < 5; ++i) { + EXPECT_EQ(src_ptr[i], cpu_ptr[i]); + EXPECT_EQ(src_ptr[i], dst_ptr[i]); + } + + delete cpu_place; + delete gpu_place; + } +#endif +} + +TEST(CopyToVector, Tensor) { + using namespace paddle::framework; + using namespace paddle::platform; + { + Tensor src; + int* src_ptr = src.mutable_data({3, 3}, CPUPlace()); + for (int i = 0; i < 3 * 3; ++i) { + src_ptr[i] = i; + } + + CPUPlace place; + std::vector dst; + CopyToVector(src, &dst); + + for (int i = 0; i < 3 * 3; ++i) { + EXPECT_EQ(src_ptr[i], dst[i]); + } + } +#ifdef PADDLE_WITH_CUDA + { + std::vector src_vec = {1, 2, 3, 4, 5, 6, 7, 8, 9}; + Tensor gpu_tensor; + CUDAPlace place; + CUDADeviceContext gpu_ctx(place); + CopyFromVector(src_vec, gpu_ctx, &gpu_tensor); + + std::vector dst; + CopyToVector(gpu_tensor, gpu_ctx, &dst); + + for (int i = 0; i < 3 * 3; ++i) { + EXPECT_EQ(src_vec[i], dst[i]); + } + } +#endif +} + +TEST(HasNAN, CPU) { + using namespace paddle::framework; + using namespace paddle::platform; + Tensor src; + float* buf = src.mutable_data({3}, CPUPlace()); + buf[0] = 0.0; + buf[1] = NAN; + buf[2] = 0.0; + + ASSERT_TRUE(HasNAN(src)); +} + +TEST(HasInf, CPU) { + using namespace paddle::framework; + using namespace paddle::platform; + Tensor src; + double* buf = src.mutable_data({3}, CPUPlace()); + buf[0] = 1.0; + buf[1] = INFINITY; + buf[2] = 0.0; + ASSERT_TRUE(HasInf(src)); +} + +TEST(Tensor, SerializeAndDeserialize) { + framework::Tensor src_tensor; + int array[6] = {1, 2, 3, 4, 5, 6}; + src_tensor.Resize({2, 3}); + int* src_ptr = src_tensor.mutable_data(platform::CPUPlace()); + for (int i = 0; i < 6; ++i) { + src_ptr[i] = array[i]; + } + { + framework::Tensor dst_tensor; + auto place = new platform::CPUPlace(); + platform::CPUDeviceContext cpu_ctx(*place); + std::ostringstream oss; + SerializeToStream(oss, src_tensor, cpu_ctx); + + std::istringstream iss(oss.str()); + DeserializeFromStream(iss, &dst_tensor); + int* dst_ptr = dst_tensor.mutable_data(platform::CPUPlace()); + for (int i = 0; i < 5; ++i) { + ASSERT_EQ(dst_ptr[i], array[i]); + } + delete place; + } +#ifdef PADDLE_WITH_CUDA + { + Tensor gpu_tensor; + gpu_tensor.Resize({2, 3}); + Tensor dst_tensor; + + auto gpu_place = new platform::CUDAPlace(); + platform::CUDADeviceContext gpu_ctx(*gpu_place); + + CopyFrom(src_tensor, *gpu_place, gpu_ctx, &gpu_tensor); + + std::ostringstream oss; + SerializeToStream(oss, gpu_tensor, gpu_ctx); + + std::istringstream iss(oss.str()); + DeserializeFromStream(iss, &dst_tensor); + + int* dst_ptr = dst_tensor.mutable_data(platform::CPUPlace()); + for (int i = 0; i < 6; ++i) { + ASSERT_EQ(dst_ptr[i], array[i]); + } + + delete gpu_place; + } +#endif +} + +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/tensor_util_test.cu b/paddle/framework/tensor_util_test.cu new file mode 100644 index 0000000000000000000000000000000000000000..ebd35fdf6c2a1388fec23057070f723c8ef9da9c --- /dev/null +++ b/paddle/framework/tensor_util_test.cu @@ -0,0 +1,57 @@ +/* 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 "gtest/gtest.h" +#include "paddle/framework/tensor_util.h" +#include "paddle/platform/device_context.h" +#include "paddle/platform/place.h" + +namespace paddle { +namespace framework { + +static __global__ void FillNAN(float* buf) { + buf[0] = 0.0; + buf[1] = 0.1; + buf[2] = NAN; +} +static __global__ void FillInf(float* buf) { + buf[0] = 0.0; + buf[1] = INFINITY; + buf[2] = 0.5; +} + +TEST(HasNAN, GPU) { + Tensor tensor; + platform::CUDAPlace gpu(0); + auto& pool = platform::DeviceContextPool::Instance(); + auto* cuda_ctx = pool.GetByPlace(gpu); + float* buf = tensor.mutable_data({3}, gpu); + FillNAN<<<1, 1, 0, cuda_ctx->stream()>>>(buf); + cuda_ctx->Wait(); + ASSERT_TRUE(HasNAN(tensor)); +} + +TEST(HasInf, GPU) { + Tensor tensor; + platform::CUDAPlace gpu(0); + auto& pool = platform::DeviceContextPool::Instance(); + auto* cuda_ctx = pool.GetByPlace(gpu); + float* buf = tensor.mutable_data({3}, gpu); + FillInf<<<1, 1, 0, cuda_ctx->stream()>>>(buf); + cuda_ctx->Wait(); + ASSERT_TRUE(HasInf(tensor)); +} + +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/threadpool.cc b/paddle/framework/threadpool.cc new file mode 100644 index 0000000000000000000000000000000000000000..109a7e7dc440d91e8223f2c0924f489f54a06f64 --- /dev/null +++ b/paddle/framework/threadpool.cc @@ -0,0 +1,24 @@ +/* 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. */ + +#include "paddle/framework/threadpool.h" + +namespace paddle { +namespace framework { + +std::unique_ptr ThreadPool::threadpool(nullptr); +std::once_flag ThreadPool::init_flag; + +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/threadpool.h b/paddle/framework/threadpool.h new file mode 100644 index 0000000000000000000000000000000000000000..bcd8190755083ec30687675602a1c95a9c15c69e --- /dev/null +++ b/paddle/framework/threadpool.h @@ -0,0 +1,163 @@ +/* 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 + +#include +#include +#include +#include +#include +#include + +#include "paddle/platform/enforce.h" + +namespace paddle { +namespace framework { + +class ThreadPool { + public: + typedef std::packaged_task Task; + typedef std::function Fun; + + /** + * @brief Get a instance of threadpool, the thread number will + * be specified as the number of hardware thread contexts + */ + static ThreadPool* GetInstance() { + std::call_once(init_flag, &ThreadPool::Init); + return threadpool.get(); + } + + ~ThreadPool() { + { + // notify all threads to stop running + running_ = false; + scheduled_.notify_all(); + } + + for (auto& t : threads_) { + t->join(); + t.reset(nullptr); + } + } + + int GetNumThreads() const { return num_threads_; } + + int GetAvailable() { + std::unique_lock lock(mutex_); + return available_; + } + + /** + * @brief Push a function to the queue, and will be scheduled and + * executed if a thread is available. + * @param[in] Task, will be pushed to the task queue. + * @return std::future, we could wait for the task finished by + * f.wait(). + */ + std::future Run(const Fun& fn) { + std::unique_lock lock(mutex_); + Task task(std::bind(fn)); + std::future f = task.get_future(); + tasks_.push(std::move(task)); + lock.unlock(); + scheduled_.notify_one(); + return f; + } + + /** + * @brief Wait until all the tasks are completed. + */ + void Wait() { + std::unique_lock lock(mutex_); + completed_.wait(lock, [=] { return Done() == true; }); + } + + private: + DISABLE_COPY_AND_ASSIGN(ThreadPool); + + explicit ThreadPool(int num_threads) + : num_threads_(num_threads), available_(num_threads), running_(true) { + threads_.resize(num_threads); + for (auto& thread : threads_) { + // TODO(Yancey1989): binding the thread on the specify CPU number + thread.reset(new std::thread(std::bind(&ThreadPool::TaskLoop, this))); + } + } + + /** + * @brief If the task queue is empty and avaialbe + * is equal to the number of threads, means that + * all tasks are completed. + * + * Note: this function is not thread-safe. + * + * @return true if all tasks are completed. + */ + bool Done() { return tasks_.empty() && available_ == num_threads_; } + + void TaskLoop() { + while (running_) { + std::unique_lock lock(mutex_); + scheduled_.wait(lock, [=] { return !tasks_.empty() || !running_; }); + + if (!running_) { + break; + } + // pop a task from the task queue + auto task = std::move(tasks_.front()); + tasks_.pop(); + + --available_; + lock.unlock(); + + // run the task + task(); + + { + std::unique_lock lock(mutex_); + ++available_; + if (Done()) { + completed_.notify_all(); + } + } + } + } + + static void Init() { + if (threadpool.get() == nullptr) { + // TODO(Yancey1989): specify the max threads number + int num_threads = std::thread::hardware_concurrency(); + PADDLE_ENFORCE_GT(num_threads, 0); + threadpool.reset(new ThreadPool(num_threads)); + } + } + + private: + static std::unique_ptr threadpool; + static std::once_flag init_flag; + + int num_threads_; + int available_; + bool running_; + std::queue tasks_; + std::vector> threads_; + std::mutex mutex_; + std::condition_variable scheduled_; + std::condition_variable completed_; +}; + +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/threadpool_test.cc b/paddle/framework/threadpool_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..50b6238cd8786be9d8cf2d5f821daadea12bd208 --- /dev/null +++ b/paddle/framework/threadpool_test.cc @@ -0,0 +1,61 @@ +/* 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. */ + +#include +#include + +#include "threadpool.h" + +namespace framework = paddle::framework; + +void do_sum(framework::ThreadPool* pool, std::atomic& sum, int cnt) { + std::vector> fs; + for (int i = 0; i < cnt; ++i) { + auto f = pool->Run([&sum]() { sum.fetch_add(1); }); + fs.push_back(std::move(f)); + } + for (auto& f : fs) { + f.wait(); + } +} + +TEST(ThreadPool, ConcurrentInit) { + framework::ThreadPool* pool; + int n = 50; + std::vector threads; + for (int i = 0; i < n; ++i) { + std::thread t([&pool]() { pool = framework::ThreadPool::GetInstance(); }); + threads.push_back(std::move(t)); + } + for (auto& t : threads) { + t.join(); + } +} + +TEST(ThreadPool, ConcurrentRun) { + framework::ThreadPool* pool = framework::ThreadPool::GetInstance(); + std::atomic sum(0); + std::vector threads; + int n = 50; + // sum = (n * (n + 1)) / 2 + for (int i = 1; i <= n; ++i) { + std::thread t(do_sum, pool, std::ref(sum), i); + threads.push_back(std::move(t)); + } + for (auto& t : threads) { + t.join(); + } + pool->Wait(); + EXPECT_EQ(sum, ((n + 1) * n) / 2); +} diff --git a/paddle/framework/type_defs.h b/paddle/framework/type_defs.h index baeb98c9bd49ec65da5931bcbe33ab788f86f3e8..d834d343759fa279a1444c6337956ffce1b9061a 100644 --- a/paddle/framework/type_defs.h +++ b/paddle/framework/type_defs.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 @@ -25,11 +25,9 @@ namespace paddle { namespace framework { class OperatorBase; -class OpDescBind; -class BlockDescBind; -class BlockDesc; +class OpDesc; class InferShapeContext; -class BlockDescBind; +class BlockDesc; using VariableNameMap = std::map>; @@ -37,7 +35,7 @@ using VariableNameMap = std::map>; using Attribute = boost::variant, std::vector, std::vector, bool, - std::vector, BlockDescBind*>; + std::vector, BlockDesc*>; using AttributeMap = std::unordered_map; @@ -45,13 +43,13 @@ using OpCreator = std::function; -using GradOpMakerFN = std::function>( - const OpDescBind&, const std::unordered_set& /*no_grad_set*/, +using GradOpMakerFN = std::function>( + const OpDesc&, const std::unordered_set& /*no_grad_set*/, std::unordered_map* /*grad_to_var*/, - const std::vector& grad_block)>; + const std::vector& grad_block)>; -using InferVarTypeFN = std::function; +using InferVarTypeFN = + std::function; using InferShapeFN = std::function; diff --git a/paddle/framework/var_desc.cc b/paddle/framework/var_desc.cc index 0babec29f6f4412ed29deeafe24470e86b30a636..7d002b9ea0b597730685ee03b021c4982f787f49 100644 --- a/paddle/framework/var_desc.cc +++ b/paddle/framework/var_desc.cc @@ -18,30 +18,32 @@ limitations under the License. */ namespace paddle { namespace framework { -VarDesc::VarType VarDescBind::GetType() const { return desc_.type(); } +proto::VarDesc::VarType VarDesc::GetType() const { return desc_.type(); } -void VarDescBind::SetType(VarDesc::VarType type) { desc_.set_type(type); } +void VarDesc::SetType(proto::VarDesc::VarType type) { desc_.set_type(type); } -void VarDescBind::SetShape(const std::vector &dims) { +void VarDesc::SetShape(const std::vector &dims) { VectorToRepeated(dims, mutable_tensor_desc()->mutable_dims()); } -void VarDescBind::SetDataType(DataType data_type) { +void VarDesc::SetDataType(proto::DataType data_type) { mutable_tensor_desc()->set_data_type(data_type); } -std::vector VarDescBind::Shape() const { +std::vector VarDesc::Shape() const { return RepeatedToVector(tensor_desc().dims()); } -DataType VarDescBind::GetDataType() const { return tensor_desc().data_type(); } +proto::DataType VarDesc::GetDataType() const { + return tensor_desc().data_type(); +} -void VarDescBind::SetLoDLevel(int32_t lod_level) { +void VarDesc::SetLoDLevel(int32_t lod_level) { switch (desc_.type()) { - case VarDesc::LOD_TENSOR: + case proto::VarDesc::LOD_TENSOR: desc_.mutable_lod_tensor()->set_lod_level(lod_level); break; - case VarDesc::LOD_TENSOR_ARRAY: + case proto::VarDesc::LOD_TENSOR_ARRAY: desc_.mutable_tensor_array()->set_lod_level(lod_level); break; default: @@ -50,11 +52,11 @@ void VarDescBind::SetLoDLevel(int32_t lod_level) { } } -int32_t VarDescBind::GetLodLevel() const { +int32_t VarDesc::GetLodLevel() const { switch (desc_.type()) { - case VarDesc::LOD_TENSOR: + case proto::VarDesc::LOD_TENSOR: return desc_.lod_tensor().lod_level(); - case VarDesc::LOD_TENSOR_ARRAY: + case proto::VarDesc::LOD_TENSOR_ARRAY: return desc_.tensor_array().lod_level(); default: PADDLE_THROW("Tensor type=%d does not support LoDLevel", @@ -62,29 +64,29 @@ int32_t VarDescBind::GetLodLevel() const { } } -const TensorDesc &VarDescBind::tensor_desc() const { +const proto::TensorDesc &VarDesc::tensor_desc() const { PADDLE_ENFORCE(desc_.has_type(), "invoke TensorDesc must after set type"); switch (desc_.type()) { - case VarDesc::SELECTED_ROWS: + case proto::VarDesc::SELECTED_ROWS: return desc_.selected_rows(); - case VarDesc::LOD_TENSOR: + case proto::VarDesc::LOD_TENSOR: return desc_.lod_tensor().tensor(); - case VarDesc::LOD_TENSOR_ARRAY: + case proto::VarDesc::LOD_TENSOR_ARRAY: return desc_.tensor_array().tensor(); default: - PADDLE_THROW("Unexpected branch."); + PADDLE_THROW("The type of var '", this->Name(), "' is unsupported."); } } -TensorDesc *VarDescBind::mutable_tensor_desc() { +proto::TensorDesc *VarDesc::mutable_tensor_desc() { PADDLE_ENFORCE(desc_.has_type(), "invoke MutableTensorDesc must after set type"); switch (desc_.type()) { - case VarDesc::SELECTED_ROWS: + case proto::VarDesc::SELECTED_ROWS: return desc_.mutable_selected_rows(); - case VarDesc::LOD_TENSOR: + case proto::VarDesc::LOD_TENSOR: return desc_.mutable_lod_tensor()->mutable_tensor(); - case VarDesc::LOD_TENSOR_ARRAY: + case proto::VarDesc::LOD_TENSOR_ARRAY: return desc_.mutable_tensor_array()->mutable_tensor(); default: PADDLE_THROW("Unexpected branch."); diff --git a/paddle/framework/var_desc.h b/paddle/framework/var_desc.h index 5cf4608944c5011d798fbde060002a57be8f6102..4fd2abe7fb215c3ac454de3e30754685111eb570 100644 --- a/paddle/framework/var_desc.h +++ b/paddle/framework/var_desc.h @@ -53,44 +53,44 @@ inline void VectorToRepeated(const std::vector &vec, } } -class VarDescBind { +class VarDesc { public: - explicit VarDescBind(const std::string &name) { + explicit VarDesc(const std::string &name) { desc_.set_name(name); - desc_.set_type(VarDesc::LOD_TENSOR); + desc_.set_type(proto::VarDesc::LOD_TENSOR); } - explicit VarDescBind(const VarDesc &desc) : desc_(desc) {} + explicit VarDesc(const proto::VarDesc &desc) : desc_(desc) {} - VarDesc *Proto() { return &desc_; } + proto::VarDesc *Proto() { return &desc_; } std::string Name() const { return desc_.name(); } void SetShape(const std::vector &dims); - void SetDataType(DataType data_type); + void SetDataType(proto::DataType data_type); std::vector Shape() const; - DataType GetDataType() const; + proto::DataType GetDataType() const; void SetLoDLevel(int32_t lod_level); int32_t GetLodLevel() const; - VarDesc::VarType GetType() const; + proto::VarDesc::VarType GetType() const; - void SetType(VarDesc::VarType type); + void SetType(proto::VarDesc::VarType type); bool Persistable() const { return desc_.persistable(); } void SetPersistable(bool persistable) { desc_.set_persistable(persistable); } private: - const TensorDesc &tensor_desc() const; - TensorDesc *mutable_tensor_desc(); + const proto::TensorDesc &tensor_desc() const; + proto::TensorDesc *mutable_tensor_desc(); - VarDesc desc_; + proto::VarDesc desc_; }; } // namespace framework } // namespace paddle diff --git a/paddle/framework/var_type.h b/paddle/framework/var_type.h index 0f19870bec3e69d07278507cc556a86bbd25d12d..0e6ea8dc69fe9b7fdaa1163b8d63295624cd3abc 100644 --- a/paddle/framework/var_type.h +++ b/paddle/framework/var_type.h @@ -1,16 +1,16 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +/* 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 +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 + 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. */ +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/framework/framework.pb.h" @@ -20,15 +20,15 @@ namespace paddle { namespace framework { -inline VarDesc::VarType ToVarType(std::type_index type) { +inline proto::VarDesc::VarType ToVarType(std::type_index type) { if (type.hash_code() == typeid(LoDTensor).hash_code()) { - return VarDesc_VarType_LOD_TENSOR; + return proto::VarDesc_VarType_LOD_TENSOR; } else if (type.hash_code() == typeid(LoDRankTable).hash_code()) { - return VarDesc_VarType_LOD_RANK_TABLE; + return proto::VarDesc_VarType_LOD_RANK_TABLE; } else if (type.hash_code() == typeid(LoDTensorArray).hash_code()) { - return VarDesc_VarType_LOD_TENSOR_ARRAY; + return proto::VarDesc_VarType_LOD_TENSOR_ARRAY; } else if (type.hash_code() == typeid(SelectedRows).hash_code()) { - return VarDesc_VarType_SELECTED_ROWS; + return proto::VarDesc_VarType_SELECTED_ROWS; } else { PADDLE_THROW("ToVarType:Unsupported type %s", type.name()); } @@ -37,16 +37,16 @@ inline VarDesc::VarType ToVarType(std::type_index type) { template inline void VisitVarType(const Variable& var, Visitor visitor) { switch (ToVarType(var.Type())) { - case VarDesc_VarType_LOD_TENSOR: + case proto::VarDesc_VarType_LOD_TENSOR: visitor(var.Get()); return; - case VarDesc_VarType_LOD_RANK_TABLE: + case proto::VarDesc_VarType_LOD_RANK_TABLE: visitor(var.Get()); return; - case VarDesc_VarType_LOD_TENSOR_ARRAY: + case proto::VarDesc_VarType_LOD_TENSOR_ARRAY: visitor(var.Get()); return; - case VarDesc_VarType_SELECTED_ROWS: + case proto::VarDesc_VarType_SELECTED_ROWS: visitor(var.Get()); return; default: diff --git a/paddle/framework/var_type_inference.h b/paddle/framework/var_type_inference.h index 32abbeb33479444c5e7a9889f4211f59af07f98f..6c11f2fee7f554fb008f559bb33aeafea5c5a556 100644 --- a/paddle/framework/var_type_inference.h +++ b/paddle/framework/var_type_inference.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/framework/type_defs.h" @@ -21,8 +21,7 @@ namespace framework { class VarTypeInference { public: virtual ~VarTypeInference() {} - virtual void operator()(const OpDescBind& op_desc, - BlockDescBind* block) const = 0; + virtual void operator()(const OpDesc& op_desc, BlockDesc* block) const = 0; }; } // namespace framework diff --git a/paddle/framework/var_type_inference_test.cc b/paddle/framework/var_type_inference_test.cc index 9035e63fa48ffdf7c72061b0a4248538d7a357e4..fa6018b1c583abaccb0259b82c9bb61c0fc10820 100644 --- a/paddle/framework/var_type_inference_test.cc +++ b/paddle/framework/var_type_inference_test.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/framework/var_type_inference.h" #include "gtest/gtest.h" @@ -33,17 +33,16 @@ class SumOpMaker : public OpProtoAndCheckerMaker { class SumOpVarTypeInference : public VarTypeInference { public: - void operator()(const OpDescBind &op_desc, - BlockDescBind *block) const override { + void operator()(const OpDesc &op_desc, BlockDesc *block) const override { auto &inputs = op_desc.Input("X"); - auto default_var_type = VarDesc::SELECTED_ROWS; + auto default_var_type = proto::VarDesc::SELECTED_ROWS; bool any_input_is_lod_tensor = std::any_of( inputs.begin(), inputs.end(), [block](const std::string &name) { - return block->Var(name)->GetType() == VarDesc::LOD_TENSOR; + return block->Var(name)->GetType() == proto::VarDesc::LOD_TENSOR; }); if (any_input_is_lod_tensor) { - default_var_type = VarDesc::LOD_TENSOR; + default_var_type = proto::VarDesc::LOD_TENSOR; } auto out_var_name = op_desc.Output("Out").front(); @@ -62,43 +61,43 @@ namespace paddle { namespace framework { TEST(InferVarType, sum_op) { - ProgramDescBind prog; + ProgramDesc prog; auto *op = prog.MutableBlock(0)->AppendOp(); op->SetType("sum"); op->SetInput("X", {"test_a", "test_b", "test_c"}); op->SetOutput("Out", {"test_out"}); - prog.MutableBlock(0)->Var("test_a")->SetType(VarDesc::SELECTED_ROWS); - prog.MutableBlock(0)->Var("test_b")->SetType(VarDesc::SELECTED_ROWS); - prog.MutableBlock(0)->Var("test_c")->SetType(VarDesc::SELECTED_ROWS); + prog.MutableBlock(0)->Var("test_a")->SetType(proto::VarDesc::SELECTED_ROWS); + prog.MutableBlock(0)->Var("test_b")->SetType(proto::VarDesc::SELECTED_ROWS); + prog.MutableBlock(0)->Var("test_c")->SetType(proto::VarDesc::SELECTED_ROWS); prog.MutableBlock(0)->Var("test_out"); op->InferVarType(prog.MutableBlock(0)); - ASSERT_EQ(VarDesc::SELECTED_ROWS, + ASSERT_EQ(proto::VarDesc::SELECTED_ROWS, prog.MutableBlock(0)->Var("test_out")->GetType()); - prog.MutableBlock(0)->Var("test_b")->SetType(VarDesc::LOD_TENSOR); + prog.MutableBlock(0)->Var("test_b")->SetType(proto::VarDesc::LOD_TENSOR); op->InferVarType(prog.MutableBlock(0)); - ASSERT_EQ(VarDesc::LOD_TENSOR, + ASSERT_EQ(proto::VarDesc::LOD_TENSOR, prog.MutableBlock(0)->Var("test_out")->GetType()); } TEST(InferVarType, sum_op_without_infer_var_type) { - ProgramDescBind prog; + ProgramDesc prog; auto *op = prog.MutableBlock(0)->AppendOp(); op->SetType("sum_without_infer_var_type"); op->SetInput("X", {"test2_a", "test2_b", "test2_c"}); op->SetOutput("Out", {"test2_out"}); - prog.MutableBlock(0)->Var("test2_a")->SetType(VarDesc::SELECTED_ROWS); - prog.MutableBlock(0)->Var("test2_b")->SetType(VarDesc::SELECTED_ROWS); - prog.MutableBlock(0)->Var("test2_c")->SetType(VarDesc::SELECTED_ROWS); + prog.MutableBlock(0)->Var("test2_a")->SetType(proto::VarDesc::SELECTED_ROWS); + prog.MutableBlock(0)->Var("test2_b")->SetType(proto::VarDesc::SELECTED_ROWS); + prog.MutableBlock(0)->Var("test2_c")->SetType(proto::VarDesc::SELECTED_ROWS); prog.MutableBlock(0)->Var("test2_out"); op->InferVarType(prog.MutableBlock(0)); - ASSERT_EQ(VarDesc_VarType_LOD_TENSOR, + ASSERT_EQ(proto::VarDesc_VarType_LOD_TENSOR, prog.MutableBlock(0)->Var("test2_out")->GetType()); } diff --git a/paddle/function/EigenGemm.cpp b/paddle/function/EigenGemm.cpp index b3e666e860d29d89650d48a23cf44917035a02d7..644098a9e7873fb59b6343e805163e4892f060a8 100644 --- a/paddle/function/EigenGemm.cpp +++ b/paddle/function/EigenGemm.cpp @@ -21,7 +21,7 @@ template struct EigenBlasGemm { typedef Eigen::TensorMap, Eigen::Aligned> - Matrix; + EigenMatrix; static void compute(const bool transA, const bool transB, @@ -56,14 +56,13 @@ struct EigenBlasGemm { sizeB[1] = N; CHECK_EQ(N, ldb); } - Eigen::array sizeC; - sizeC[0] = M; - sizeC[1] = N; - CHECK_EQ(N, ldc); + Eigen::array sizeC = {{M, ldc}}; + Eigen::array offsetC = {{0, 0}}; + Eigen::array extentC = {{M, N}}; - const Matrix a(const_cast(A), sizeA); - const Matrix b(const_cast(B), sizeB); - Matrix c(C, sizeC); + const EigenMatrix a(const_cast(A), sizeA); + const EigenMatrix b(const_cast(B), sizeB); + EigenMatrix c(C, sizeC); typedef typename Eigen::Tensor::DimensionPair DimPair; Eigen::array dims; @@ -72,12 +71,23 @@ struct EigenBlasGemm { dims[0].second = transB ? 1 : 0; Eigen::DefaultDevice device; - if (alpha == T(1) && beta == T(0)) { - c.device(device) = a.contract(b, dims); - } else if (alpha == T(1) && beta == T(1)) { - c.device(device) += a.contract(b, dims); + if (N == ldc) { + if (alpha == T(1) && beta == T(0)) { + c.device(device) = a.contract(b, dims); + } else if (alpha == T(1) && beta == T(1)) { + c.device(device) += a.contract(b, dims); + } else { + c.device(device) = alpha * a.contract(b, dims) + beta * c; + } } else { - c.device(device) = alpha * a.contract(b, dims) + beta * c; + if (alpha == T(1) && beta == T(0)) { + 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); + } else { + c.slice(offsetC, extentC).device(device) = + alpha * a.contract(b, dims) + beta * c.slice(offsetC, extentC); + } } } }; diff --git a/paddle/function/GemmConvOp.cpp b/paddle/function/GemmConvOp.cpp index 8d34eee886a6202691e5dec2ab62e7c5b0ac7fb1..cbdbf5335d32d55a0221728758025c9d2cb3e7d1 100644 --- a/paddle/function/GemmConvOp.cpp +++ b/paddle/function/GemmConvOp.cpp @@ -129,6 +129,162 @@ public: } }; +#ifdef PADDLE_MOBILE_INFERENCE + +/* + * \brief Forward calculation of convolution, optimized for mobile. + */ +template +class GemmConvMobileFunction : public ConvFunctionBase { +public: + void init(const FuncConfig& config) override { + ConvFunctionBase::init(config); + } + + void check(const BufferArgs& inputs, const BufferArgs& outputs) override { + const TensorShape& input = inputs[0].shape(); + const TensorShape& filter = inputs[1].shape(); + const TensorShape& output = outputs[0].shape(); + checkShape(input, filter, output); + } + + void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { + CHECK_EQ(numInputs_, inputs.size()); + CHECK_EQ(numOutputs_, outputs.size()); + check(inputs, outputs); + // TODO(hedaoyuan): Need to define some index macros, + // to avoid useing 0 and 1. + const TensorShape& input = inputs[0].shape(); + const TensorShape& filter = inputs[1].shape(); + const TensorShape& output = outputs[0].shape(); + + real beta; + if (outputs[0].getArgType() == ADD_TO) { + beta = 1.0; + } else { + beta = 0.0; + } + + size_t batchSize = input[0]; + size_t inputChannels = input[1]; + size_t inputHeight = input[2]; + size_t inputWidth = input[3]; + size_t filterHeight = getFilterHeight(filter); + size_t filterWidth = getFilterWidth(filter); + size_t outputChannels = output[1]; + size_t outputHeight = output[2]; + size_t outputWidth = output[3]; + + real* inputData = inputs[0].data(); + real* filterData = inputs[1].data(); + real* outputData = outputs[0].data(); + bool needIm2col = isNeedIm2col(filter); + + TensorShape imShape = + TensorShape({inputChannels / groups_, inputHeight, inputWidth}); + + TensorShape colShape; + real* colData = NULL; + + size_t colHeight = inputChannels / groups_ * filterHeight * filterWidth; + size_t colWidth = outputHeight * outputWidth; + // Max col matrix height 256, Max col matrix width 1024 + size_t stepColHeight = std::min(colHeight, static_cast(256)); + size_t stepColWidth = std::min(colWidth, static_cast(2048)); + + if (needIm2col) { + colShape = TensorShape({inputChannels / groups_, + filterHeight, + filterWidth, + outputHeight, + outputWidth}); + + resizeBuffer(stepColHeight * stepColWidth * sizeof(real)); + colData = reinterpret_cast(memory_->getBuf()); + } + + Im2ColMobileFunctor im2col; + size_t inputOffset = imShape.getElements(); + size_t outputOffset = + (outputChannels / groups_) * outputHeight * outputWidth; + size_t filterOffset = filter.getElements() / groups_; + + int nStride = colWidth; + int kStride = colHeight; + for (size_t i = 0; i < batchSize; i++) { + for (size_t g = 0; g < groups_; g++) { + if (needIm2col) { + real beta_ = beta; + for (size_t colHeightStart = 0; colHeightStart < colHeight; + colHeightStart += stepColHeight) { + for (size_t colWidthStart = 0; colWidthStart < colWidth; + colWidthStart += stepColWidth) { + int N = std::min(colWidth - colWidthStart, stepColWidth); + int K = std::min(colHeight - colHeightStart, stepColHeight); + // im2col + im2col(inputData + g * inputOffset, + imShape, + colData, + colShape, + strideH(), + strideW(), + paddingH(), + paddingW(), + dilationH(), + dilationW(), + colHeightStart, + K, + colWidthStart, + N); + + // gemm + int M = outputChannels / groups_; + BlasGemm::compute( + false, + false, + M, + N, + K, + 1.0f, + filterData + g * filterOffset + colHeightStart, + kStride, + colData, + N, + beta_, + outputData + g * outputOffset + colWidthStart, + nStride); + } + beta_ = 1.0; + } + } else { + int M = outputChannels / groups_; + int N = outputHeight * outputWidth; + int K = inputChannels / groups_ * filterHeight * filterWidth; + BlasGemm::compute(false, + false, + M, + N, + K, + 1.0f, + filterData + g * filterOffset, + K, + inputData + g * inputOffset, + N, + beta, + outputData + g * outputOffset, + N); + } + } + inputData += inputChannels * inputHeight * inputWidth; + outputData += outputChannels * outputHeight * outputWidth; + } + + memory_.reset(); + } +}; + +#endif + /* * \brief Backward input calculation of convolution. */ @@ -343,7 +499,11 @@ public: } }; +#ifdef PADDLE_MOBILE_INFERENCE +REGISTER_TYPED_FUNC(GemmConv, CPU, GemmConvMobileFunction); +#else REGISTER_TYPED_FUNC(GemmConv, CPU, GemmConvFunction); +#endif REGISTER_TYPED_FUNC(GemmConvGradInput, CPU, GemmConvGradInputFunction); REGISTER_TYPED_FUNC(GemmConvGradFilter, CPU, GemmConvGradFilterFunction); #ifdef PADDLE_WITH_CUDA diff --git a/paddle/function/Im2Col.h b/paddle/function/Im2Col.h index 0c37fc972484bfbede01d23652e384071bf883af..36a9bcf84e4b14965c83627821b71d1c7c0da1b2 100644 --- a/paddle/function/Im2Col.h +++ b/paddle/function/Im2Col.h @@ -98,4 +98,54 @@ public: int dilationWidth = 1); }; +template +class Im2ColMobileFunctor { +public: + void operator()(const T* imData, + const TensorShape& imShape, + T* colData, + const TensorShape& colShape, + int strideHeight, + int strideWidth, + int paddingHeight, + int paddingWidth, + int dilationHeight, + int dilationWidth, + int colHeightStart, + int colHeightSize, + int colWidthStart, + int colWidthSize) { + int inputHeight = imShape[1]; + int inputWidth = imShape[2]; + int filterHeight = colShape[1]; + int filterWidth = colShape[2]; + int outputWidth = colShape[4]; + + for (int colh = 0; colh < colHeightSize; colh++) { + int wOffset = (colHeightStart + colh) % filterWidth; + int hOffset = ((colHeightStart + colh) / filterWidth) % filterHeight; + int c_im = (colHeightStart + colh) / filterWidth / filterHeight; + + for (int colw = 0; colw < colWidthSize; colw++) { + int h = (colWidthStart + colw) / outputWidth; + int w = (colWidthStart + colw) % outputWidth; + + int imRowIdx = h * strideHeight + hOffset * dilationHeight; + int imColIdx = w * strideWidth + wOffset * dilationWidth; + if ((imRowIdx - paddingHeight) < 0 || + (imRowIdx - paddingHeight) >= inputHeight || + (imColIdx - paddingWidth) < 0 || + (imColIdx - paddingWidth) >= inputWidth) { + colData[colh * colWidthSize + colw] = static_cast(0); + } else { + imRowIdx += c_im * inputHeight - paddingHeight; + imColIdx -= paddingWidth; + colData[colh * colWidthSize + colw] = + imData[imRowIdx * inputWidth + imColIdx]; + } + } + } + } +}; + } // namespace paddle diff --git a/paddle/function/Im2ColTest.cpp b/paddle/function/Im2ColTest.cpp index 1f085538d81904dbd5b5d6bcd014adaed22e37d7..3ba866dcdd845403d52f7a85adfef08cbb11c305 100644 --- a/paddle/function/Im2ColTest.cpp +++ b/paddle/function/Im2ColTest.cpp @@ -138,4 +138,86 @@ TEST(Im2ColFunctor, GPU) { TestIm2ColFunctor(); } #endif +template +void TestIm2ColMobileFunctor() { + for (size_t channels : {32}) { + for (size_t inputHeight : {33, 100}) { + for (size_t inputWidth : {32, 96}) { + for (size_t filterHeight : {5}) { + for (size_t filterWidth : {7}) { + for (size_t stride : {2}) { + for (size_t padding : {1}) { + for (size_t dilation : {1, 3}) { + size_t filterSizeH = (filterHeight - 1) * dilation + 1; + size_t filterSizeW = (filterWidth - 1) * dilation + 1; + if (inputHeight + 2 * padding < filterSizeH || + inputWidth + 2 * padding < filterSizeW) + break; + if (padding >= filterSizeH || padding >= filterSizeW) break; + size_t outputHeight = + (inputHeight - filterSizeH + 2 * padding) / stride + 1; + size_t outputWidth = + (inputWidth - filterSizeW + 2 * padding) / stride + 1; + + TensorShape imShape = + TensorShape({channels, inputHeight, inputWidth}); + TensorShape colShape1 = TensorShape({channels, + filterHeight, + filterWidth, + outputHeight, + outputWidth}); + + size_t height = channels * filterHeight * filterWidth; + size_t width = outputHeight * outputWidth; + VectorPtr input1 = + Vector::create(imShape.getElements(), false); + VectorPtr input2 = + Vector::create(imShape.getElements(), false); + MatrixPtr output1 = + Matrix::create(height, width, false, false); + MatrixPtr output2 = + Matrix::create(height, width, false, false); + input1->uniform(0.001, 1); + input2->copyFrom(*input1); + + Im2ColFunctor im2Col1; + Im2ColMobileFunctor im2Col2; + im2Col1(input1->getData(), + imShape, + output1->getData(), + colShape1, + stride, + stride, + padding, + padding, + dilation, + dilation); + im2Col2(input2->getData(), + imShape, + output2->getData(), + colShape1, + stride, + stride, + padding, + padding, + dilation, + dilation, + 0, + height, + 0, + width); + + autotest::TensorCheckEqual(*output1, *output2); + } + } + } + } + } + } + } + } +} + +TEST(Im2ColFunctor, Mobile) { TestIm2ColMobileFunctor(); } + } // namespace paddle diff --git a/paddle/gserver/activations/ActivationFunction.cpp b/paddle/gserver/activations/ActivationFunction.cpp index 8b7b2e9b65898950e036ebc023cd28990cef303f..57c890e4884da38e2087d89dc199e20af51495ea 100644 --- a/paddle/gserver/activations/ActivationFunction.cpp +++ b/paddle/gserver/activations/ActivationFunction.cpp @@ -24,7 +24,7 @@ limitations under the License. */ #include "paddle/utils/ClassRegistrar.h" #include "paddle/utils/Logging.h" -#ifdef PADDLE_USE_MKLDNN +#ifdef PADDLE_WITH_MKLDNN #include "MKLDNNActivation.h" #endif @@ -212,6 +212,37 @@ Error __must_check backward(Argument& act) { } END_DEFINE_ACTIVATION(sequence_softmax) +/* + * @brief SoftSign Activation. + * \f[ + * f(z) = \frac{z}{1 + |z|} + * \f] + */ +BEGIN_DEFINE_ACTIVATION(softsign) +private: +MatrixPtr denominator_; + +Error __must_check forward(Argument& act) { + size_t height = act.value->getHeight(); + size_t width = act.value->getWidth(); + Matrix::resizeOrCreate( + denominator_, height, width, false, useGpu(act.deviceId)); + denominator_->assign(*act.value); + denominator_->abs2(); + denominator_->add(1.); + + act.value->dotDiv(*act.value, *denominator_); + return Error(); +} + +Error __must_check backward(Argument& act) { + denominator_->square2(); + denominator_->scalarDiv(*denominator_, 1.); + act.grad->dotMul(*act.grad, *denominator_); + return Error(); +} +END_DEFINE_ACTIVATION(softsign) + /** * @brief Relu Activation. * forward. y = max(0, z) @@ -459,7 +490,7 @@ Error __must_check backward(Argument& act) { END_DEFINE_ACTIVATION(log) ActivationFunction* ActivationFunction::create(const std::string& type) { -#ifdef PADDLE_USE_MKLDNN +#ifdef PADDLE_WITH_MKLDNN if (!type.empty() && type.compare(0, 7, "mkldnn_") == 0) { return MKLDNNActivation::create(type); } diff --git a/paddle/gserver/gradientmachines/NeuralNetwork.cpp b/paddle/gserver/gradientmachines/NeuralNetwork.cpp index be112b41239cace3fa9b9ee97923f8c3c7a9a98f..68bf37d59db65ddc8096e2db3391be25c37b57e6 100644 --- a/paddle/gserver/gradientmachines/NeuralNetwork.cpp +++ b/paddle/gserver/gradientmachines/NeuralNetwork.cpp @@ -20,7 +20,7 @@ limitations under the License. */ #include "paddle/utils/Logging.h" #include "paddle/utils/Stat.h" -#ifdef PADDLE_USE_MKLDNN +#ifdef PADDLE_WITH_MKLDNN #include "paddle/gserver/layers/MKLDNNLayer.h" #endif @@ -307,7 +307,7 @@ void NeuralNetwork::backward(const UpdateCallback& callback) { } void NeuralNetwork::finish() { -#ifdef PADDLE_USE_MKLDNN +#ifdef PADDLE_WITH_MKLDNN FOR_EACH_R(layer, layers_) { MKLDNNLayerPtr dnnLayer = std::dynamic_pointer_cast(*layer); if (dnnLayer) { diff --git a/paddle/gserver/layers/BatchNormBaseLayer.cpp b/paddle/gserver/layers/BatchNormBaseLayer.cpp index bc7d1c83a48aefeb4bc6d3baa32b78aba712e58d..925af31289d0c8ca534a30a16b14bfd2df90b013 100644 --- a/paddle/gserver/layers/BatchNormBaseLayer.cpp +++ b/paddle/gserver/layers/BatchNormBaseLayer.cpp @@ -41,6 +41,7 @@ bool BatchNormBaseLayer::init(const LayerMap& layerMap, useGlobalStats_ = config_.use_global_stats(); } movingAvgFraction_ = config_.moving_average_fraction(); + epsilon_ = config_.epsilon(); weight_.reset(new Weight(1, channels_, parameters_[0])); movingMean_.reset(new Weight(1, channels_, parameters_[1])); diff --git a/paddle/gserver/layers/BatchNormBaseLayer.h b/paddle/gserver/layers/BatchNormBaseLayer.h index e721d2d267a31cae46407673b8b1281e87055608..2ac3cd9d670d0fcf9c40ad2f117d5a72479663a3 100644 --- a/paddle/gserver/layers/BatchNormBaseLayer.h +++ b/paddle/gserver/layers/BatchNormBaseLayer.h @@ -94,6 +94,8 @@ protected: bool useGlobalStats_; // use to compute moving mean and variance. real movingAvgFraction_; + // Epsilon is a small random noise used in batch normalization for stability. + real epsilon_; }; } // namespace paddle diff --git a/paddle/gserver/layers/BatchNormalizationLayer.cpp b/paddle/gserver/layers/BatchNormalizationLayer.cpp index dacff25e5927daf9c991577a71be86b160228317..25ab5cd927792d18f78bc1fa33eee4029b427cc7 100644 --- a/paddle/gserver/layers/BatchNormalizationLayer.cpp +++ b/paddle/gserver/layers/BatchNormalizationLayer.cpp @@ -22,8 +22,6 @@ namespace paddle { REGISTER_LAYER(batch_norm, BatchNormalizationLayer); -const real BatchNormalizationLayer::EPS = 1E-5; - bool BatchNormalizationLayer::init(const LayerMap& layerMap, const ParameterMap& parameterMap) { /* Initialize the basic parent class */ @@ -53,7 +51,7 @@ void BatchNormalizationLayer::calMeanAndStd(const MatrixPtr& mat) { calMovingMeanAndVar(); - savedInvVar_->subScalar(-EPS); + savedInvVar_->subScalar(-epsilon_); savedInvVar_->sqrt2(*savedInvVar_); } @@ -74,7 +72,7 @@ void BatchNormalizationLayer::setMeanAndStd() { savedInvVar_->copyFrom(*(movingVar_->getW())); savedInvVar_->downClip(real(0.0)); - savedInvVar_->subScalar(-EPS); + savedInvVar_->subScalar(-epsilon_); savedInvVar_->sqrt2(*savedInvVar_); } diff --git a/paddle/gserver/layers/BatchNormalizationLayer.h b/paddle/gserver/layers/BatchNormalizationLayer.h index f6115801fc6b341c0718f8851617de43bdeeec09..1fdb5e2070259a14ab6f70957c9cf03f0699f734 100644 --- a/paddle/gserver/layers/BatchNormalizationLayer.h +++ b/paddle/gserver/layers/BatchNormalizationLayer.h @@ -39,9 +39,6 @@ public: void backward(const UpdateCallback& callback = nullptr) override; protected: - /// Epsilon value used in the batch normalization formula. - static const real EPS; - /// Load pre-calculated mean and std. void setMeanAndStd(); diff --git a/paddle/gserver/layers/ConvTransProjection.cpp b/paddle/gserver/layers/ConvTransProjection.cpp index 48132a3ce4cc4b50fea6d755d84d7254d2055bec..e7f081c0232d185c223fc2f48ca79dc84c7f721d 100644 --- a/paddle/gserver/layers/ConvTransProjection.cpp +++ b/paddle/gserver/layers/ConvTransProjection.cpp @@ -24,13 +24,13 @@ size_t ConvTransProjection::calOutputSize() { if (outputH_ == 0) outputH_ = configOutH_; if (outputW_ == 0) outputW_ = configOutW_; imageH_ = imageSize(outputH_, - filterH_, + (filterH_ - 1) * dilationH_ + 1, paddingH_, strideH_, /* caffeMode */ true); imageW_ = imageSize(outputW_, - filterW_, + (filterW_ - 1) * dilationW_ + 1, paddingW_, strideW_, /* caffeMode */ true); diff --git a/paddle/gserver/layers/CudnnBatchNormLayer.cpp b/paddle/gserver/layers/CudnnBatchNormLayer.cpp index 49a9540c0b6e36b59ed786287ff5c4569b69a6a5..8390b55026c895b661cb514714ba92c05a7bf02e 100644 --- a/paddle/gserver/layers/CudnnBatchNormLayer.cpp +++ b/paddle/gserver/layers/CudnnBatchNormLayer.cpp @@ -21,8 +21,6 @@ namespace paddle { REGISTER_LAYER(cudnn_batch_norm, CudnnBatchNormLayer); -const double CudnnBatchNormLayer::EPS = 1E-5; - bool CudnnBatchNormLayer::init(const LayerMap& layerMap, const ParameterMap& parameterMap) { /* Initialize the basic parent class */ @@ -61,6 +59,9 @@ void CudnnBatchNormLayer::forward(PassType passType) { real* movingMean = movingMean_->getW()->getData(); real* movingVar = movingVar_->getW()->getData(); + // cuDNN does not allow an epsilon value less than CUDNN_BN_MIN_EPSILON. + eps_ = std::max(CUDNN_BN_MIN_EPSILON, static_cast(epsilon_)); + if (!useGlobalStats_) { REGISTER_TIMER_INFO("CudnnBatchFwTimer", getName().c_str()); real* savedMean = savedMean_->getData(); @@ -75,7 +76,7 @@ void CudnnBatchNormLayer::forward(PassType passType) { 1.0 - movingAvgFraction_, movingMean, movingVar, - EPS, + eps_, savedMean, savedInvVar); } else { @@ -90,7 +91,7 @@ void CudnnBatchNormLayer::forward(PassType passType) { beta, movingMean, movingVar, - EPS); + eps_); } else { // There is a limitation in cudnn library. // When the batch size is larger than 1024 in cuDNN v5.1, @@ -101,7 +102,7 @@ void CudnnBatchNormLayer::forward(PassType passType) { beta, movingMean, movingVar, - EPS, + eps_, batchSize, channels_, imageH_ * imageD_, @@ -128,6 +129,9 @@ void CudnnBatchNormLayer::backward(const UpdateCallback& callback) { real* savedMean = savedMean_->getData(); real* savedInvVar = savedInvVar_->getData(); + // cuDNN does not allow an epsilon value less than CUDNN_BN_MIN_EPSILON. + eps_ = std::max(CUDNN_BN_MIN_EPSILON, static_cast(epsilon_)); + auto create = [](MatrixPtr& m, size_t h, size_t w, real** p) { Matrix::resizeOrCreate(m, h, w, false, true); m->zeroMem(); @@ -157,7 +161,7 @@ void CudnnBatchNormLayer::backward(const UpdateCallback& callback) { gamma, gammaGrad, betaGrad, - EPS, + eps_, savedMean, savedInvVar); diff --git a/paddle/gserver/layers/CudnnBatchNormLayer.h b/paddle/gserver/layers/CudnnBatchNormLayer.h index 413efd4d3ecd734b343efbcf8328ac0592daddda..1a3f0c0cbf8a1540e77cef70c753c91298728484 100644 --- a/paddle/gserver/layers/CudnnBatchNormLayer.h +++ b/paddle/gserver/layers/CudnnBatchNormLayer.h @@ -14,6 +14,7 @@ limitations under the License. */ #pragma once +#include #include "BatchNormBaseLayer.h" #include "Layer.h" #include "paddle/utils/Stat.h" @@ -46,12 +47,9 @@ public: void backward(const UpdateCallback& callback = nullptr) override; protected: - /** - * Epsilon value used in the batch normalization formula. - * Minimum allowed value is CUDNN_BN_MIN_EPSILON defined in cudnn.h. - * Same epsilon value should be used in forward and backward functions. - */ - static const double EPS; + /// Epsilon value used in the batch normalization formula. + /// Same epsilon value should be used in forward and backward functions. + double eps_; /// Input/output tensor descriptor desc hl_tensor_descriptor ioDesc_; diff --git a/paddle/gserver/layers/FactorizationMachineLayer.cpp b/paddle/gserver/layers/FactorizationMachineLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..be26b9ba88c279036f73b0a0baaff164755fe067 --- /dev/null +++ b/paddle/gserver/layers/FactorizationMachineLayer.cpp @@ -0,0 +1,158 @@ +/* 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. */ + +#include "FactorizationMachineLayer.h" +#include +#include +#include "paddle/math/SparseMatrix.h" +#include "paddle/utils/Logging.h" +#include "paddle/utils/Stat.h" + +namespace paddle { + +REGISTER_LAYER(factorization_machine, FactorizationMachineLayer); + +bool FactorizationMachineLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + /* Initialize the basic parent class */ + Layer::init(layerMap, parameterMap); + + factorSize_ = config_.factor_size(); + + /* initialize the latentVectors_ */ + CHECK_EQ(inputLayers_.size(), 1UL); + size_t inputSize = inputLayers_[0]->getSize(); + CHECK_EQ(parameters_[0]->getSize(), inputSize * factorSize_); + latentVectors_ = std::unique_ptr( + new Weight(inputSize, factorSize_, parameters_[0])); + + return true; +} + +void FactorizationMachineLayer::forward(PassType passType) { + Layer::forward(passType); + + const MatrixPtr& inputV = getInputValue(0); + + size_t batchSize = inputV->getHeight(); + size_t outputSize = getSize(); + size_t inputSize = inputLayers_[0]->getSize(); + reserveOutput(batchSize, outputSize); + + MatrixPtr outV = getOutputValue(); + + Matrix::resizeOrCreate( + latentVectorsSquare_, inputSize, factorSize_, false, useGpu_); + Matrix::resizeOrCreate( + inputMulFactor_, batchSize, factorSize_, false, useGpu_); + Matrix::resizeOrCreate(tmpOut_, batchSize, factorSize_, false, useGpu_); + + REGISTER_TIMER_INFO("FmInputMulFactorTimer", getName().c_str()); + inputMulFactor_->mul(*inputV, *latentVectors_->getW()); + inputMulFactor_->square2(*tmpOut_); + outV->sumRows(*tmpOut_, 0.5, 0); + + if (dynamic_cast(inputV.get())) { + Matrix::resizeOrCreateSparseMatrix(inputSquare_, + inputV->getHeight(), + inputV->getWidth(), + inputV->getElementCnt(), + inputV->getValueType()); + inputSquare_->copyFrom(*inputV); + (dynamic_cast(inputSquare_.get()))->square2(); + } else { + Matrix::resizeOrCreate( + inputSquare_, inputV->getHeight(), inputV->getWidth(), false, useGpu_); + inputV->square2(*inputSquare_); + } + latentVectors_->getW()->square2(*latentVectorsSquare_); + tmpOut_->mul(*inputSquare_, *latentVectorsSquare_); + outV->sumRows(*tmpOut_, -0.5, 1.0); + + /* activation */ { + REGISTER_TIMER_INFO("FmFwAtvTimer", getName().c_str()); + forwardActivation(); + } +} + +void FactorizationMachineLayer::backward(const UpdateCallback& callback) { + /* Do derivation */ { backwardActivation(); } + + const MatrixPtr& inputV = getInputValue(0); + const MatrixPtr& oGrad = getOutputGrad(); + + Matrix::resizeOrCreate( + tmpSum_, 1, latentVectors_->getW()->getHeight(), false, useGpu_); + MatrixPtr tmpSumTrans = Matrix::create(tmpSum_->getRowBuf(0), + latentVectors_->getW()->getHeight(), + 1, + false, + useGpu_); + + /* Calculate the gradients of the latentVectors_ matrix */ + if (latentVectors_->getWGrad()) { + if (dynamic_cast(inputV.get())) { + Matrix::resizeOrCreateSparseMatrix(tmpInput_, + inputV->getHeight(), + inputV->getWidth(), + inputV->getElementCnt()); + + CpuSparseMatrix* sparseInputV = + dynamic_cast(inputV.get()); + CpuSparseMatrix* sparseInputSquare = + dynamic_cast(inputSquare_.get()); + CpuSparseMatrix* sparseTmpInput = + dynamic_cast(tmpInput_.get()); + sparseTmpInput->copyFrom(*sparseInputV); + + sparseTmpInput->rowScale(0, *sparseInputV, *oGrad); + latentVectors_->getWGrad()->mul( + *sparseTmpInput->getTranspose(), *inputMulFactor_, 1, 1); + sparseTmpInput->rowScale(0, *sparseInputSquare, *oGrad); + + Matrix::resizeOrCreate(negOnes_, 1, inputV->getHeight(), false, useGpu_); + negOnes_->zeroMem(); + negOnes_->add(-1); + tmpSum_->mul(*negOnes_, *sparseTmpInput, 1, 0); + } else { + Matrix::resizeOrCreate( + tmpInput_, inputV->getHeight(), inputV->getWidth(), false, useGpu_); + + tmpInput_->rowScale(0, *inputV, *oGrad); + latentVectors_->getWGrad()->mul( + *tmpInput_->getTranspose(), *inputMulFactor_, 1, 1); + tmpInput_->rowScale(0, *inputSquare_, *oGrad); + + tmpSum_->sumCols(*tmpInput_, -1, 0); + } + + latentVectors_->getWGrad()->addRowScale( + 0, *latentVectors_->getW(), *tmpSumTrans); + + /* Increasing the number of gradient */ + latentVectors_->getParameterPtr()->incUpdate(callback); + } + + /* Calculate the input layers gradient */ + MatrixPtr inGrad = getInputGrad(0); + if (inGrad != NULL) { + inGrad->mul( + *inputMulFactor_, *latentVectors_->getW()->getTranspose(), 1, 1); + tmpSumTrans->sumRows(*latentVectorsSquare_, -1, 0); + inGrad->addColScale(0, *inputV, *tmpSum_); + inGrad->rowScale(0, *inGrad, *oGrad); + } +} + +} // namespace paddle diff --git a/paddle/gserver/layers/FactorizationMachineLayer.h b/paddle/gserver/layers/FactorizationMachineLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..df20a49934d5dd444f127842c8fdb7c77f4ebeb1 --- /dev/null +++ b/paddle/gserver/layers/FactorizationMachineLayer.h @@ -0,0 +1,80 @@ +/* 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 + +#include "Layer.h" +#include "paddle/math/Matrix.h" +#include "paddle/utils/ThreadLocal.h" + +namespace paddle { +/** + * @brief The Factorization Machine models pairwise (order-2) feature + * interactions as inner product of the learned latent vectors corresponding + * to each input feature. + * + * The Factorization Machine can effectively capture feature interactions + * especially when the input is sparse. While in principle FM can model higher + * order feature interaction, in practice usually only order-2 feature + * interactions are considered. The Factorization Machine Layer here only + * computes the order-2 interations with the formula: + * + * \f[ + * y = \sum_{i=1}^{n-1}\sum_{j=i+1}^n\langle v_i, v_j \rangle x_i x_j + * \f] + * + * The detailed calculation for forward and backward can be found at this paper: + * + * Factorization machines. + * + * The config file api is factorization_machine. + */ + +class FactorizationMachineLayer : public Layer { +protected: + // The latent vectors, shape: (size, factorSize_) + // Each row of the latentVectors_ matrix is the latent vector + // corresponding to one input feature dimension + std::unique_ptr latentVectors_; + // The hyperparameter that defines the dimensionality of the factorization + size_t factorSize_; + +private: + // Store the square values of the letent vectors matrix + MatrixPtr latentVectorsSquare_; + // Store the square values of input matrix + MatrixPtr inputSquare_; + // The result of input matrix * latent vector matrix that will be used in + // both forward and backward step + MatrixPtr inputMulFactor_; + // Store temporary calculation result + MatrixPtr tmpOut_; + MatrixPtr tmpSum_; + MatrixPtr tmpInput_; + // Negative identity matrix + MatrixPtr negOnes_; + +public: + explicit FactorizationMachineLayer(const LayerConfig& config) + : Layer(config) {} + ~FactorizationMachineLayer() {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; +}; + +} // namespace paddle diff --git a/paddle/gserver/layers/HierarchicalSigmoidLayer.cpp b/paddle/gserver/layers/HierarchicalSigmoidLayer.cpp index d62a8d846e5b347aa44ce1951c043d5813a5b3ff..236f8096bdb6e024cf3c9c73eba422616a777a23 100644 --- a/paddle/gserver/layers/HierarchicalSigmoidLayer.cpp +++ b/paddle/gserver/layers/HierarchicalSigmoidLayer.cpp @@ -64,49 +64,111 @@ void HierarchicalSigmoidLayer::forward(PassType passType) { batchSize, codeLength_, /* trans */ false, - useGpu(deviceId_)); + false); Matrix::resizeOrCreate(preOutput_.grad, batchSize, codeLength_, /* trans */ false, - useGpu(deviceId_)); - + false); IVectorPtr label = getInput(*getLabelLayer()).ids; - preOutput_.value->zeroMem(); + if (useGpu_) { + Matrix::resizeOrCreate(cpuOutput_, + output_.value->getHeight(), + output_.value->getWidth(), + /* trans */ false, + false); + IVector::resizeOrCreate(cpuLabel_, label->getSize(), false); + cpuLabel_->copyFrom(*label); + cpuOutput_->copyFrom(*output_.value); + } else { + cpuOutput_ = output_.value; + cpuLabel_ = label; + } /* add the bias-vector */ if (biases_.get() != NULL) { - preOutput_.value->addByBitCode(numClasses_, *label, *biases_->getW()); + if (useGpu_) { + Matrix::resizeOrCreate(cpuBias_, + 1, + numClasses_ - 1, + /* trans */ false, + false); + cpuBias_->copyFrom(*biases_->getW()); + } else { + cpuBias_ = biases_->getW(); + } + preOutput_.value->addByBitCode(numClasses_, *cpuLabel_, *cpuBias_); } for (size_t i = 0; i < inputLayers_.size() - 1; ++i) { MatrixPtr input = getInputValue(i); + if (useGpu_) { + Matrix::resizeOrCreate(cpuInput_, + input->getHeight(), + input->getWidth(), + /* trans */ false, + false); + Matrix::resizeOrCreate(cpuWeight_, + weights_[i]->getW()->getHeight(), + weights_[i]->getW()->getWidth(), + /* trans */ false, + false); + cpuInput_->copyFrom(*input); + cpuWeight_->copyFrom(*weights_[i]->getW()); + } else { + cpuInput_ = input; + cpuWeight_ = weights_[i]->getW(); + } preOutput_.value->mulByBitCode( - numClasses_, *label, *weights_[i]->getW(), *input); + numClasses_, *cpuLabel_, *cpuWeight_, *cpuInput_); } // keep consistent with the clipping in the following softrelu preOutput_.value->clip(-40.0, 40.0); preOutput_.value->sumByBitCode(numClasses_, - *label, - *output_.value, + *cpuLabel_, + *cpuOutput_, -1); // scaleSum preOutput_.value->softrelu(*preOutput_.value); - MatrixPtr sum = - Matrix::create(batchSize, 1, /* trans= */ false, useGpu(deviceId_)); + MatrixPtr sum = Matrix::create(batchSize, 1, /* trans= */ false, false); preOutput_.value->rowSum(*sum); - output_.value->add(*sum); + cpuOutput_->add(*sum); + if (useGpu_) { + output_.value->copyFrom(*cpuOutput_); + } else { + output_.value = cpuOutput_; + } } void HierarchicalSigmoidLayer::backward(const UpdateCallback& callback) { IVectorPtr label = getInput(*getLabelLayer()).ids; + if (useGpu_) { + IVector::resizeOrCreate(cpuLabel_, label->getSize(), false); + cpuLabel_->copyFrom(*label); + } else { + cpuLabel_ = label; + } preOutput_.grad->one(); preOutput_.grad->softreluDerivative(*preOutput_.value); - preOutput_.grad->subByBitCode(numClasses_, *label); + preOutput_.grad->subByBitCode(numClasses_, *cpuLabel_); if (biases_ && biases_->getWGrad()) { - preOutput_.grad->addByBitCodeBackward( - numClasses_, *label, *biases_->getWGrad()); - + MatrixPtr biases_grad = biases_->getWGrad(); + if (useGpu_) { + Matrix::resizeOrCreate(cpuBias_, + 1, + numClasses_ - 1, + /* trans */ false, + false); + cpuBias_->copyFrom(*biases_grad); + } else { + cpuBias_ = biases_grad; + } + preOutput_.grad->addByBitCodeBackward(numClasses_, *cpuLabel_, *cpuBias_); + if (useGpu_) { + biases_grad->copyFrom(*cpuBias_); + } else { + biases_grad = cpuBias_; + } /* Increasing the number of gradient */ biases_->getParameterPtr()->incUpdate(callback); } @@ -115,9 +177,31 @@ void HierarchicalSigmoidLayer::backward(const UpdateCallback& callback) { /* Calculate the W-gradient for the current layer */ MatrixPtr input = getInputValue(i); if (weights_[i]->getWGrad()) { + MatrixPtr weights_grad = weights_[i]->getWGrad(); + if (useGpu_) { + Matrix::resizeOrCreate(cpuInput_, + input->getHeight(), + input->getWidth(), + /* trans */ false, + false); + Matrix::resizeOrCreate(cpuWeightGrad_, + weights_grad->getHeight(), + weights_grad->getWidth(), + /* trans */ false, + false); + cpuInput_->copyFrom(*input); + cpuWeightGrad_->copyFrom(*weights_grad); + } else { + cpuInput_ = input; + cpuWeightGrad_ = weights_grad; + } preOutput_.grad->mulByBitCodeBackwardWeight( - numClasses_, *label, *weights_[i]->getWGrad(), *input); - + numClasses_, *cpuLabel_, *cpuWeightGrad_, *cpuInput_); + if (useGpu_) { + weights_grad->copyFrom(*cpuWeightGrad_); + } else { + weights_grad = cpuWeightGrad_; + } /* Increasing the number of gradient */ weights_[i]->getParameterPtr()->incUpdate(callback); } @@ -125,8 +209,30 @@ void HierarchicalSigmoidLayer::backward(const UpdateCallback& callback) { /* Calculate the input layers error */ MatrixPtr inputGrad = getInputGrad(i); if (inputGrad) { + if (useGpu_) { + Matrix::resizeOrCreate(cpuInputGrad_, + inputGrad->getHeight(), + inputGrad->getWidth(), + /* trans */ false, + false); + Matrix::resizeOrCreate(cpuWeight_, + weights_[i]->getW()->getHeight(), + weights_[i]->getW()->getWidth(), + /* trans */ false, + false); + cpuInputGrad_->copyFrom(*inputGrad); + cpuWeight_->copyFrom(*weights_[i]->getW()); + } else { + cpuInputGrad_ = inputGrad; + cpuWeight_ = weights_[i]->getW(); + } preOutput_.grad->mulByBitCodeBackwardError( - numClasses_, *label, *weights_[i]->getW(), *inputGrad); + numClasses_, *cpuLabel_, *cpuWeight_, *cpuInputGrad_); + if (useGpu_) { + inputGrad->copyFrom(*cpuInputGrad_); + } else { + inputGrad = cpuInputGrad_; + } } } } diff --git a/paddle/gserver/layers/HierarchicalSigmoidLayer.h b/paddle/gserver/layers/HierarchicalSigmoidLayer.h index 9afd40b1674680da962d6e51caa56b46279b70de..7f896e61ca26e3e22b99b65b1285384a121f7f02 100644 --- a/paddle/gserver/layers/HierarchicalSigmoidLayer.h +++ b/paddle/gserver/layers/HierarchicalSigmoidLayer.h @@ -80,6 +80,15 @@ protected: int codeLength_; /// temporary result of output_ Argument preOutput_; + + /// The temporary variables in CPU memory. + MatrixPtr cpuWeight_; + MatrixPtr cpuWeightGrad_; + MatrixPtr cpuInput_; + MatrixPtr cpuInputGrad_; + MatrixPtr cpuBias_; + MatrixPtr cpuOutput_; + IVectorPtr cpuLabel_; }; } // namespace paddle diff --git a/paddle/gserver/layers/MKLDNNAddtoLayer.cpp b/paddle/gserver/layers/MKLDNNAddtoLayer.cpp index 0f2b67fd758ec1513f42c4cb1a36f2f3915f4740..39bffc26f7ddcd159130c492115b41080e32ce7f 100644 --- a/paddle/gserver/layers/MKLDNNAddtoLayer.cpp +++ b/paddle/gserver/layers/MKLDNNAddtoLayer.cpp @@ -38,12 +38,13 @@ bool MKLDNNAddtoLayer::init(const LayerMap& layerMap, } void MKLDNNAddtoLayer::reshape( - int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) { + int& bs, int& ic, int& ih, int& iw, int& oc, int& oh, int& ow) { CHECK_EQ(layerSize_, getSize()) << "this layer size can not be changed"; reshapeInput(bs, ih, iw); ic = inputLayers_[0]->getSize() / ih / iw; CHECK_EQ((size_t)ic * ih * iw, inputLayers_[0]->getSize()); - CHECK_EQ(inputElemenCnt_, (size_t)bs * ic * ih * iw); + CHECK_EQ(inputLayers_[0]->getOutputValue()->getElementCnt(), + (size_t)bs * ic * ih * iw); for (size_t i = 0; i < inputLayers_.size(); i++) { CHECK_EQ(int64_t(bs), inputLayers_[i]->getOutput().getBatchSize()); CHECK_EQ(layerSize_, inputLayers_[i]->getSize()); @@ -57,47 +58,43 @@ void MKLDNNAddtoLayer::reshape( } void MKLDNNAddtoLayer::resetFwd(std::vector& pipeline, - MKLDNNMatrixPtr& in, - MKLDNNMatrixPtr& wgt, - MKLDNNMatrixPtr& bias, + std::vector& inputs, MKLDNNMatrixPtr& out) { - resetFwdBuffers(inVals_, bias, out); - in = inVals_[0]; + resetFwdBuffers(inputs, biasVal_, out); std::shared_ptr fwdPD; std::shared_ptr biasPD; - resetFwdPD(fwdPD, biasPD, inVals_, bias, out); + resetFwdPD(fwdPD, biasPD, inputs, biasVal_, out); - resetFwdPipeline(pipeline, fwdPD, biasPD, inVals_, bias, out); + resetFwdPipeline(pipeline, fwdPD, biasPD, inputs, biasVal_, out); } void MKLDNNAddtoLayer::resetBwd(std::vector& pipeline, - MKLDNNMatrixPtr& in, - MKLDNNMatrixPtr& wgt, - MKLDNNMatrixPtr& bias, + std::vector& inputs, MKLDNNMatrixPtr& out) { - resetBwdBuffers(inGrads_, bias, out); - in = inGrads_[0]; + resetBwdBuffers(inputs, biasGrad_, out); // backward only need share output grad to input grad - for (size_t i = 0; i < inGrads_.size(); i++) { - if (inGrads_[i] != nullptr) { - inGrads_[i] = out; - inputLayers_[i]->getOutputGrad()->setData(inGrads_[i]->getData()); + for (size_t i = 0; i < inputs.size(); i++) { + if (inputs[i] != nullptr) { + inputs[i] = out; + inputLayers_[i]->getOutputGrad()->setData(inputs[i]->getData()); } } // backward bias bwdBias_ = nullptr; - if (bias) { + if (biasGrad_) { std::vector scales(bs_, 1.0); - std::vector srcPDs(bs_, bias->getPrimitiveDesc()); - auto biasPD = sum::primitive_desc(bias->getMemoryDesc(), scales, srcPDs); + std::vector srcPDs(bs_, + biasGrad_->getPrimitiveDesc()); + auto biasPD = + sum::primitive_desc(biasGrad_->getMemoryDesc(), scales, srcPDs); std::vector srcs; for (size_t i = 0; i < grads_.size(); ++i) { srcs.push_back(*(grads_[i])); } - bwdBias_.reset(new sum(biasPD, srcs, *bias)); + bwdBias_.reset(new sum(biasPD, srcs, *biasGrad_)); pipeline.push_back(*bwdBias_); } } @@ -208,7 +205,7 @@ void MKLDNNAddtoLayer::resetBwdBuffers(std::vector& inputs, inputs.resize(inputLayers_.size()); for (size_t i = 0; i < inputs.size(); i++) { - resetInGrad(inputs[i], inVal_->getPrimitiveDesc(), i); + resetInGrad(inputs[i], inVals_[i]->getPrimitiveDesc(), i); CHECK_PRIMITIVE_DESC_EQ(inputs[i], out->getPrimitiveDesc()); } diff --git a/paddle/gserver/layers/MKLDNNAddtoLayer.h b/paddle/gserver/layers/MKLDNNAddtoLayer.h index 24504b7b4f50726e2b2757ca3029461cdc27b411..0ea3e208e5fab8cbed8b53390a9381e6f2bb5733 100644 --- a/paddle/gserver/layers/MKLDNNAddtoLayer.h +++ b/paddle/gserver/layers/MKLDNNAddtoLayer.h @@ -26,9 +26,6 @@ namespace paddle { */ class MKLDNNAddtoLayer : public MKLDNNLayer { protected: - std::vector inVals_; - std::vector inGrads_; - // layer size == ic * ih * iw == oc * oh *ow, and can not be changed size_t layerSize_; @@ -50,52 +47,19 @@ public: const ParameterMap& parameterMap) override; void reshape( - int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) override; + int& bs, int& ic, int& ih, int& iw, int& oc, int& oh, int& ow) override; void resetFwd(std::vector& pipeline, - MKLDNNMatrixPtr& in, - MKLDNNMatrixPtr& wgt, - MKLDNNMatrixPtr& bias, + std::vector& inputs, MKLDNNMatrixPtr& out) override; void resetBwd(std::vector& pipeline, - MKLDNNMatrixPtr& in, - MKLDNNMatrixPtr& wgt, - MKLDNNMatrixPtr& bias, + std::vector& inputs, MKLDNNMatrixPtr& out) override; void updateWeights(const UpdateCallback& callback) override; - void printValueFormat() override { - for (size_t i = 0; i < inVals_.size(); ++i) { - VLOG(MKLDNN_FMTS) << i << " input: " << inVals_[i]->getFormat() << " >>>"; - } - if (outVal_) { - VLOG(MKLDNN_FMTS) << outVal_->getFormat() << " >>> "; - } - if (extOutVal_) { - VLOG(MKLDNN_FMTS) << extOutVal_->getFormat(); - } - } - - void printGradFormat() override { - if (extOutGrad_) { - VLOG(MKLDNN_FMTS) << extOutGrad_->getFormat(); - } - if (outGrad_) { - VLOG(MKLDNN_FMTS) << outGrad_->getFormat() << " <<< "; - } - for (size_t i = 0; i < inGrads_.size(); ++i) { - VLOG(MKLDNN_FMTS) << i << " input: " << inGrads_[i]->getFormat() << "<<<"; - } - } - protected: - /** - * Forward functions: reset buffers(inputs, output, bias), - * reset primitive descriptor, - * reset pipeline. - */ void resetFwdBuffers(std::vector& inputs, MKLDNNMatrixPtr& bias, MKLDNNMatrixPtr& out); @@ -110,17 +74,10 @@ protected: std::vector& inputs, MKLDNNMatrixPtr& bias, MKLDNNMatrixPtr& out); - - /** - * Backward functions: reset buffers(inputs, output, bias) - */ void resetBwdBuffers(std::vector& inputs, MKLDNNMatrixPtr& bias, MKLDNNMatrixPtr& out); - /** - * prepare for bias - */ void prepareBias(MKLDNNMatrixPtr& bias, const MatrixPtr& biasMat, const MKLDNNMatrixPtr& out, diff --git a/paddle/gserver/layers/MKLDNNBatchNormLayer.cpp b/paddle/gserver/layers/MKLDNNBatchNormLayer.cpp index 071bdf54d5dc9538d5ced580a73b9c0fbcea41fb..7faca0f8b7f54fa0a09e8fdab11064c8c26df375 100644 --- a/paddle/gserver/layers/MKLDNNBatchNormLayer.cpp +++ b/paddle/gserver/layers/MKLDNNBatchNormLayer.cpp @@ -21,8 +21,6 @@ namespace paddle { REGISTER_LAYER(mkldnn_batch_norm, MKLDNNBatchNormLayer); -const real MKLDNNBatchNormLayer::EPS = 1E-5; - bool MKLDNNBatchNormLayer::init(const LayerMap& layerMap, const ParameterMap& parameterMap) { if (!MKLDNNLayer::init(layerMap, parameterMap)) { @@ -50,6 +48,8 @@ bool MKLDNNBatchNormLayer::init(const LayerMap& layerMap, useGlobalStats_ = config_.use_global_stats(); } movingAvgFraction_ = config_.moving_average_fraction(); + epsilon_ = config_.epsilon(); + VLOG(MKLDNN_BASE) << "--- " << (useGlobalStats_ ? "use" : "do not use") << " --- global stats"; VLOG(MKLDNN_BASE) << "Moving average fraction: " << movingAvgFraction_; @@ -116,21 +116,20 @@ void MKLDNNBatchNormLayer::calMovingMeanAndVar() { } void MKLDNNBatchNormLayer::reshape( - int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) { + int& bs, int& ic, int& ih, int& iw, int& oc, int& oh, int& ow) { reshapeInput(bs, ih, iw); oh = ih; ow = iw; // ic_ and oc can not be changed - CHECK_EQ(inputElemenCnt_ / bs / ih / iw, (size_t)ic) + CHECK_EQ((size_t)ic, + inputLayers_[0]->getOutputValue()->getElementCnt() / bs / ih / iw) << "Input channel can not be changed"; reshapeOutput(oh, ow); resizeOutput(bs, oc * oh * ow); } void MKLDNNBatchNormLayer::resetFwd(std::vector& pipeline, - MKLDNNMatrixPtr& in, - MKLDNNMatrixPtr& wgt, - MKLDNNMatrixPtr& bias, + std::vector& inputs, MKLDNNMatrixPtr& out) { // In training phase, it will always calculate mean and var, // so useGlobalStats must be false. @@ -140,25 +139,23 @@ void MKLDNNBatchNormLayer::resetFwd(std::vector& pipeline, useGlobalStats_ = false; } - resetFwdBuffers(in, wgt, out); + resetFwdBuffers(inputs[0], wgtVal_, out); - resetFwdPD(fwdPD_, in, wgt, out); + resetFwdPD(fwdPD_, inputs[0], wgtVal_, out); - resetFwdPipeline(pipeline, fwdPD_, in, wgt, out); + resetFwdPipeline(pipeline, fwdPD_, inputs[0], wgtVal_, out); } void MKLDNNBatchNormLayer::resetBwd(std::vector& pipeline, - MKLDNNMatrixPtr& in, - MKLDNNMatrixPtr& wgt, - MKLDNNMatrixPtr& bias, + std::vector& inputs, MKLDNNMatrixPtr& out) { std::shared_ptr pd; - resetBwdBuffers(in, wgt, out); + resetBwdBuffers(inputs[0], wgtGrad_, out); - resetBwdPD(pd, in, wgt, out); + resetBwdPD(pd, inputs[0], wgtGrad_, out); - resetBwdPipeline(pipeline, pd, in, wgt, out); + resetBwdPipeline(pipeline, pd, inputs[0], wgtGrad_, out); } void MKLDNNBatchNormLayer::forward(PassType passType) { @@ -213,7 +210,7 @@ void MKLDNNBatchNormLayer::resetFwdPD( if (wgt) { flags_ = (flags_ | batch_normalization_flag::use_scale_shift); } - auto fwdDesc = bn_fwd::desc(pk, in->getMemoryDesc(), EPS, flags_); + auto fwdDesc = bn_fwd::desc(pk, in->getMemoryDesc(), epsilon_, flags_); pd.reset(new bn_fwd::primitive_desc(fwdDesc, engine_)); CHECK_PRIMITIVE_DESC_EQ(out, pd->dst_primitive_desc()); if (wgt) { @@ -260,9 +257,9 @@ void MKLDNNBatchNormLayer::resetFwdPipeline( void MKLDNNBatchNormLayer::resetBwdBuffers(MKLDNNMatrixPtr& in, MKLDNNMatrixPtr& wgt, MKLDNNMatrixPtr& out) { - CHECK(inVal_ && outVal_); + CHECK(inVals_[0] && outVal_); resetOutGrad(out, outVal_->getPrimitiveDesc()); - resetInGrad(in, inVal_->getPrimitiveDesc()); + resetInGrad(in, inVals_[0]->getPrimitiveDesc()); if (gradScaleShift_) { CHECK(wgtVal_); resetWithMatrix(wgt, gradScaleShift_, wgtVal_->getPrimitiveDesc()); @@ -280,7 +277,7 @@ void MKLDNNBatchNormLayer::resetBwdPD( } CHECK_PRIMITIVE_DESC_EQ(out, in->getPrimitiveDesc()); auto md = in->getMemoryDesc(); - auto bwdDesc = bn_bwd::desc(prop_kind::backward, md, md, EPS, flags_); + auto bwdDesc = bn_bwd::desc(prop_kind::backward, md, md, epsilon_, flags_); pd.reset(new bn_bwd::primitive_desc(bwdDesc, engine_, *fwdPD_)); CHECK(pd->weights_primitive_desc() == fwdPD_->weights_primitive_desc()); CHECK_PRIMITIVE_DESC_EQ(wgt, pd->diff_weights_primitive_desc()); @@ -297,11 +294,12 @@ void MKLDNNBatchNormLayer::resetBwdPipeline( if (pd == nullptr) { return; } - CHECK(inVal_); + CHECK(inVals_[0]); bwdData_.reset( wgt && wgtVal_ - ? new bn_bwd(*pd, *inVal_, *mean_, *var_, *out, *wgtVal_, *in, *wgt) - : new bn_bwd(*pd, *inVal_, *mean_, *var_, *out, *in)); + ? new bn_bwd( + *pd, *inVals_[0], *mean_, *var_, *out, *wgtVal_, *in, *wgt) + : new bn_bwd(*pd, *inVals_[0], *mean_, *var_, *out, *in)); pipeline.push_back(*bwdData_); } diff --git a/paddle/gserver/layers/MKLDNNBatchNormLayer.h b/paddle/gserver/layers/MKLDNNBatchNormLayer.h index 456c0424ecb8dde17f98a900c5d77268cc672e34..1cf33cb34fa9cd7c9b8487a0a4a0011fb129e311 100644 --- a/paddle/gserver/layers/MKLDNNBatchNormLayer.h +++ b/paddle/gserver/layers/MKLDNNBatchNormLayer.h @@ -32,7 +32,8 @@ protected: std::shared_ptr fwdPD_; // Epsilon value used in the batch normalization formula. - static const real EPS; + real epsilon_; + // weight and bias in paddle std::unique_ptr weight_; std::unique_ptr biases_; @@ -73,18 +74,14 @@ public: void forward(PassType passType) override; void reshape( - int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) override; + int& bs, int& ic, int& ih, int& iw, int& oc, int& oh, int& ow) override; void resetFwd(std::vector& pipeline, - MKLDNNMatrixPtr& in, - MKLDNNMatrixPtr& wgt, - MKLDNNMatrixPtr& bias, + std::vector& inputs, MKLDNNMatrixPtr& out) override; void resetBwd(std::vector& pipeline, - MKLDNNMatrixPtr& in, - MKLDNNMatrixPtr& wgt, - MKLDNNMatrixPtr& bias, + std::vector& inputs, MKLDNNMatrixPtr& out) override; void updateWeights(const UpdateCallback& callback) override; @@ -98,11 +95,7 @@ protected: * moving = moving * AvgFraction + local * (1 - AvgFraction) */ void calMovingMeanAndVar(); - /** - * Forward functions: reset buffers(input, weight, output), - * reset primitive descriptor, - * reset pipeline. - */ + void resetFwdBuffers(MKLDNNMatrixPtr& in, MKLDNNMatrixPtr& wgt, MKLDNNMatrixPtr& out); @@ -115,12 +108,6 @@ protected: MKLDNNMatrixPtr& in, MKLDNNMatrixPtr& wgt, MKLDNNMatrixPtr& out); - - /** - * Backward functions: reset buffers(input, weight, output), - * reset primitive descriptor, - * reset pipeline. - */ void resetBwdBuffers(MKLDNNMatrixPtr& in, MKLDNNMatrixPtr& wgt, MKLDNNMatrixPtr& out); diff --git a/paddle/gserver/layers/MKLDNNConcatLayer.cpp b/paddle/gserver/layers/MKLDNNConcatLayer.cpp index c9099297cc5c741fbae0b42f21b988e6c561ef11..44bb0883b89c712d70e2d4fdfe16bdfde86f81b7 100644 --- a/paddle/gserver/layers/MKLDNNConcatLayer.cpp +++ b/paddle/gserver/layers/MKLDNNConcatLayer.cpp @@ -32,17 +32,16 @@ bool MKLDNNConcatLayer::init(const LayerMap& layerMap, } void MKLDNNConcatLayer::reshape( - int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) { + int& bs, int& ic, int& ih, int& iw, int& oc, int& oh, int& ow) { reshapeInput(bs, ih, iw); ic = inputLayers_[0]->getSize() / ih / iw; CHECK_EQ((size_t)ic * ih * iw, inputLayers_[0]->getSize()); - CHECK_EQ(inputElemenCnt_, (size_t)bs * ic * ih * iw); + CHECK_EQ(inputLayers_[0]->getOutputValue()->getElementCnt(), + (size_t)bs * ic * ih * iw); CHECK_GT(inputLayers_.size(), 1UL); channels_.resize(inputLayers_.size()); channels_[0] = ic; - // need change the output channel, so use oc_ instead - // TODO(TJ): change API, use &oc - oc_ = ic; + oc = ic; for (size_t i = 1; i < inputLayers_.size(); i++) { int batchsize, height, witdh; reshapeInput(batchsize, height, witdh, i); @@ -52,37 +51,31 @@ void MKLDNNConcatLayer::reshape( channels_[i] = inputLayers_[i]->getSize() / height / witdh; CHECK_EQ((size_t)channels_[i] * height * witdh, inputLayers_[i]->getSize()); - oc_ += channels_[i]; + oc += channels_[i]; } oh = ih; ow = iw; reshapeOutput(oh, ow); - resizeOutput(bs, oc_ * oh * ow); + resizeOutput(bs, oc * oh * ow); } void MKLDNNConcatLayer::resetFwd(std::vector& pipeline, - MKLDNNMatrixPtr& in, - MKLDNNMatrixPtr& wgt, - MKLDNNMatrixPtr& bias, + std::vector& inputs, MKLDNNMatrixPtr& out) { - resetFwdBuffers(inVals_, out); - in = inVals_[0]; + resetFwdBuffers(inputs, out); std::shared_ptr fwdPD; - resetFwdPD(fwdPD, inVals_, out); + resetFwdPD(fwdPD, inputs, out); - resetFwdPipeline(pipeline, fwdPD, inVals_, out); + resetFwdPipeline(pipeline, fwdPD, inputs, out); } void MKLDNNConcatLayer::resetBwd(std::vector& pipeline, - MKLDNNMatrixPtr& in, - MKLDNNMatrixPtr& wgt, - MKLDNNMatrixPtr& bias, + std::vector& inputs, MKLDNNMatrixPtr& out) { - resetBwdBuffers(inGrads_, out); - in = inGrads_[0]; + resetBwdBuffers(inputs, out); - resetBwdPipeline(pipeline, bwds_, inGrads_, out); + resetBwdPipeline(pipeline, bwds_, inputs, out); } void MKLDNNConcatLayer::resetFwdBuffers(std::vector& inputs, @@ -90,10 +83,7 @@ void MKLDNNConcatLayer::resetFwdBuffers(std::vector& inputs, inputs.resize(inputLayers_.size()); bool has8c = false, has16c = false, hasnc = false; for (size_t i = 0; i < inputs.size(); i++) { - // resetInValue will use ic_ so temporary change as current input's channel - // TODO(TJ): change ic_ as vector then can remove channels_ - ic_ = channels_[i]; - resetInValue(inputs[i], nullptr, i); + resetInValue(inputs[i], nullptr, i, channels_[i]); CHECK(inputs[i]); auto dm = inputs[i]->getDims(); // inputs format can be different, but ndims must equal @@ -114,8 +104,6 @@ void MKLDNNConcatLayer::resetFwdBuffers(std::vector& inputs, has16c = true; } } - // change back, ic_ always save the input 0 size - ic_ = channels_[0]; format outFmt; if (has16c && oc_ % 16 == 0) { @@ -168,14 +156,9 @@ void MKLDNNConcatLayer::resetBwdBuffers(std::vector& inputs, inputs.resize(inputLayers_.size()); for (size_t i = 0; i < inputs.size(); i++) { CHECK(inVals_[i]); - // resetInGrad will use inVal_ - // TODO(TJ): change move inVals_ to MKLDNNLayer ans remove inVal_ - inVal_ = inVals_[i]; resetInGrad(inputs[i], inVals_[i]->getPrimitiveDesc(), i); CHECK_PRIMITIVE_DESC_EQ(inputs[i], inVals_[i]->getPrimitiveDesc()); } - // change back, inVal_ always save the input 0 - inVal_ = inVals_[0]; } void MKLDNNConcatLayer::resetBwdPipeline( diff --git a/paddle/gserver/layers/MKLDNNConcatLayer.h b/paddle/gserver/layers/MKLDNNConcatLayer.h index d5749d327e4259b81541a234f48a4538ab035fe4..37f3a26c5ed5db10cdba507368874c9557fb75ef 100644 --- a/paddle/gserver/layers/MKLDNNConcatLayer.h +++ b/paddle/gserver/layers/MKLDNNConcatLayer.h @@ -26,8 +26,6 @@ namespace paddle { */ class MKLDNNConcatLayer : public MKLDNNLayer { protected: - std::vector inVals_; - std::vector inGrads_; std::vector> bwds_; // input channel numbers std::vector channels_; @@ -47,18 +45,14 @@ public: const ParameterMap& parameterMap) override; void reshape( - int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) override; + int& bs, int& ic, int& ih, int& iw, int& oc, int& oh, int& ow) override; void resetFwd(std::vector& pipeline, - MKLDNNMatrixPtr& in, - MKLDNNMatrixPtr& wgt, - MKLDNNMatrixPtr& bias, + std::vector& inputs, MKLDNNMatrixPtr& out) override; void resetBwd(std::vector& pipeline, - MKLDNNMatrixPtr& in, - MKLDNNMatrixPtr& wgt, - MKLDNNMatrixPtr& bias, + std::vector& inputs, MKLDNNMatrixPtr& out) override; void printSizeInfo() override { @@ -72,38 +66,16 @@ public: << ", " << ow_; } - void printValueFormat() override { - for (size_t i = 0; i < inVals_.size(); ++i) { - VLOG(MKLDNN_FMTS) << "Input " << i << ", " << inputLayers_[i]->getName() - << ": " << inVals_[i]->getFormat() << " >>>"; - } - if (outVal_) { - VLOG(MKLDNN_FMTS) << outVal_->getFormat() << " >>> "; - } - if (extOutVal_) { - VLOG(MKLDNN_FMTS) << extOutVal_->getFormat(); - } - } - - void printGradFormat() override { - if (extOutGrad_) { - VLOG(MKLDNN_FMTS) << extOutGrad_->getFormat(); - } - if (outGrad_) { - VLOG(MKLDNN_FMTS) << outGrad_->getFormat() << " <<< "; - } - for (size_t i = 0; i < inGrads_.size(); ++i) { - VLOG(MKLDNN_FMTS) << "Input " << i << ", " << inputLayers_[i]->getName() - << ": " << inGrads_[i]->getFormat() << "<<<"; + size_t keepCondition() { + // reset when the total element size of all inputs changed + size_t totalSize = inputLayers_[0]->getOutputValue()->getElementCnt(); + for (size_t i = 1; i < inputLayers_.size(); ++i) { + totalSize += inputLayers_[i]->getOutputValue()->getElementCnt(); } + return totalSize; } protected: - /** - * Forward functions: reset buffers(inputs, output, bias), - * reset primitive descriptor, - * reset pipeline. - */ void resetFwdBuffers(std::vector& inputs, MKLDNNMatrixPtr& out); void resetFwdPD(std::shared_ptr& pd, @@ -113,11 +85,6 @@ protected: std::shared_ptr& pd, std::vector& inputs, MKLDNNMatrixPtr& out); - - /** - * Backward functions: reset buffers(inputs, output, bias) - * reset primitives and pipeline - */ void resetBwdBuffers(std::vector& inputs, MKLDNNMatrixPtr& out); void resetBwdPipeline(std::vector& pipeline, diff --git a/paddle/gserver/layers/MKLDNNConvLayer.cpp b/paddle/gserver/layers/MKLDNNConvLayer.cpp index 8aa54e0a9efa7adb766cbb6009f6a29410c6ae7d..ab1d0f7b049a349c00c6e23deb37d789382de64f 100644 --- a/paddle/gserver/layers/MKLDNNConvLayer.cpp +++ b/paddle/gserver/layers/MKLDNNConvLayer.cpp @@ -90,7 +90,7 @@ void MKLDNNConvLayer::convertWeightsToPaddle() { } void MKLDNNConvLayer::reshape( - int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) { + int& bs, int& ic, int& ih, int& iw, int& oc, int& oh, int& ow) { reshapeInput(bs, ih, iw); // cal output sizes @@ -105,21 +105,17 @@ void MKLDNNConvLayer::reshape( } void MKLDNNConvLayer::resetFwd(std::vector& pipeline, - MKLDNNMatrixPtr& in, - MKLDNNMatrixPtr& wgt, - MKLDNNMatrixPtr& bias, + std::vector& inputs, MKLDNNMatrixPtr& out) { resetFwdPD(fwdPD_); - resetFwdBuffers(fwdPD_, in, wgt, bias, out); + resetFwdBuffers(fwdPD_, inputs[0], wgtVal_, biasVal_, out); - resetFwdPipeline(pipeline, fwdPD_, in, wgt, bias, out); + resetFwdPipeline(pipeline, fwdPD_, inputs[0], wgtVal_, biasVal_, out); } void MKLDNNConvLayer::resetBwd(std::vector& pipeline, - MKLDNNMatrixPtr& in, - MKLDNNMatrixPtr& wgt, - MKLDNNMatrixPtr& bias, + std::vector& inputs, MKLDNNMatrixPtr& out) { std::shared_ptr bwdWgtPD; std::shared_ptr bwdDataPD; @@ -128,9 +124,10 @@ void MKLDNNConvLayer::resetBwd(std::vector& pipeline, resetBwdDataPD(bwdDataPD); - resetBwdBuffers(bwdWgtPD, bwdDataPD, in, wgt, bias, out); + resetBwdBuffers(bwdWgtPD, bwdDataPD, inputs[0], wgtGrad_, biasGrad_, out); - resetBwdPipeline(pipeline, bwdWgtPD, bwdDataPD, in, wgt, bias, out); + resetBwdPipeline( + pipeline, bwdWgtPD, bwdDataPD, inputs[0], wgtGrad_, biasGrad_, out); } void MKLDNNConvLayer::updateWeights(const UpdateCallback& callback) { @@ -236,14 +233,14 @@ void MKLDNNConvLayer::resetBwdWgtPD( loadConvSettings(wgtDims, biasDims, strides, dilations, padL, padR); // create backward weight using input, output and weight value memory desc - CHECK(inVal_) << "Should have internal input value"; + CHECK(inVals_[0]) << "Should have internal input value"; CHECK(outVal_) << "Should have internal output value"; CHECK(wgtVal_) << "Should have weight value"; algorithm algo = algorithm::convolution_direct; padding_kind padKind = padding_kind::zero; auto bwdWgtDesc = biasVal_ != nullptr ? conv_bwdWgt::desc(algo, - inVal_->getMemoryDesc(), + inVals_[0]->getMemoryDesc(), wgtVal_->getMemoryDesc(), biasVal_->getMemoryDesc(), outVal_->getMemoryDesc(), @@ -252,7 +249,7 @@ void MKLDNNConvLayer::resetBwdWgtPD( padR, padKind) : conv_bwdWgt::desc(algo, - inVal_->getMemoryDesc(), + inVals_[0]->getMemoryDesc(), wgtVal_->getMemoryDesc(), outVal_->getMemoryDesc(), strides, @@ -260,7 +257,7 @@ void MKLDNNConvLayer::resetBwdWgtPD( padR, padKind); pd.reset(new conv_bwdWgt::primitive_desc(bwdWgtDesc, engine_, *fwdPD_)); - CHECK_PRIMITIVE_DESC_EQ(inVal_, pd->src_primitive_desc()); + CHECK_PRIMITIVE_DESC_EQ(inVals_[0], pd->src_primitive_desc()); CHECK_PRIMITIVE_DESC_EQ( outVal_, pd->diff_dst_primitive_desc(), @@ -280,12 +277,12 @@ void MKLDNNConvLayer::resetBwdDataPD( memory::dims wgtDims, biasDims, strides, dilations, padL, padR; loadConvSettings(wgtDims, biasDims, strides, dilations, padL, padR); - CHECK(inVal_) << "Should have internal input value"; + CHECK(inVals_[0]) << "Should have internal input value"; CHECK(outVal_) << "Should have internal output value"; // create backward data using input and output value memory desc // but using weight memory desc with any format auto bwdDataDesc = conv_bwdData::desc(algorithm::convolution_direct, - inVal_->getMemoryDesc(), + inVals_[0]->getMemoryDesc(), MKLDNNMatrix::createMemoryDesc(wgtDims), outVal_->getMemoryDesc(), strides, @@ -294,7 +291,7 @@ void MKLDNNConvLayer::resetBwdDataPD( padding_kind::zero); pd.reset(new conv_bwdData::primitive_desc(bwdDataDesc, engine_, *fwdPD_)); CHECK_PRIMITIVE_DESC_EQ( - inVal_, + inVals_[0], pd->diff_src_primitive_desc(), "primitive desc of in value and grad should be equal"); CHECK_PRIMITIVE_DESC_EQ( @@ -346,12 +343,12 @@ void MKLDNNConvLayer::resetBwdPipeline( MKLDNNMatrixPtr& wgt, MKLDNNMatrixPtr& bias, MKLDNNMatrixPtr& out) { - CHECK(inVal_); + CHECK(inVals_[0]); // add bwdWgt handle if (bias) { - bwdWgt_.reset(new conv_bwdWgt(*wgtPD, *inVal_, *out, *wgt, *bias)); + bwdWgt_.reset(new conv_bwdWgt(*wgtPD, *inVals_[0], *out, *wgt, *bias)); } else { - bwdWgt_.reset(new conv_bwdWgt(*wgtPD, *inVal_, *out, *wgt)); + bwdWgt_.reset(new conv_bwdWgt(*wgtPD, *inVals_[0], *out, *wgt)); } pipeline.push_back(*bwdWgt_); diff --git a/paddle/gserver/layers/MKLDNNConvLayer.h b/paddle/gserver/layers/MKLDNNConvLayer.h index 9c69136684e5f9005860b476ec6ed1bbc9ceff6c..3e754a0e65771879e836c13d63d5a5c8be3a699a 100644 --- a/paddle/gserver/layers/MKLDNNConvLayer.h +++ b/paddle/gserver/layers/MKLDNNConvLayer.h @@ -69,18 +69,14 @@ public: const ParameterMap& parameterMap) override; void reshape( - int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) override; + int& bs, int& ic, int& ih, int& iw, int& oc, int& oh, int& ow) override; void resetFwd(std::vector& pipeline, - MKLDNNMatrixPtr& in, - MKLDNNMatrixPtr& wgt, - MKLDNNMatrixPtr& bias, + std::vector& inputs, MKLDNNMatrixPtr& out) override; void resetBwd(std::vector& pipeline, - MKLDNNMatrixPtr& in, - MKLDNNMatrixPtr& wgt, - MKLDNNMatrixPtr& bias, + std::vector& inputs, MKLDNNMatrixPtr& out) override; void updateWeights(const UpdateCallback& callback) override; @@ -107,48 +103,26 @@ protected: mkldnn::memory::dims& padL, mkldnn::memory::dims& padR); - /** - * reset the forward primitive descriptor. - */ void resetFwdPD(std::shared_ptr& pd); - /** - * reset the MKLDNNMatrix buffers used in forward. - */ void resetFwdBuffers(std::shared_ptr& pd, MKLDNNMatrixPtr& in, MKLDNNMatrixPtr& wgt, MKLDNNMatrixPtr& bias, MKLDNNMatrixPtr& out); - /** - * reset the forward pipeline. - */ void resetFwdPipeline(std::vector& pipeline, std::shared_ptr& pd, MKLDNNMatrixPtr& in, MKLDNNMatrixPtr& wgt, MKLDNNMatrixPtr& bias, MKLDNNMatrixPtr& out); - - /** - * reset the backward weight primitive descriptor. - */ void resetBwdWgtPD(std::shared_ptr& pd); - /** - * reset the backward data primitive descriptor. - */ void resetBwdDataPD(std::shared_ptr& pd); - /** - * reset the MKLDNNMatrix buffers used in backward. - */ void resetBwdBuffers(std::shared_ptr& wgtPD, std::shared_ptr& dataPD, MKLDNNMatrixPtr& in, MKLDNNMatrixPtr& wgt, MKLDNNMatrixPtr& bias, MKLDNNMatrixPtr& out); - /** - * reset the backward pipeline. - */ void resetBwdPipeline(std::vector& pipeline, std::shared_ptr& wgtPD, std::shared_ptr& dataPD, diff --git a/paddle/gserver/layers/MKLDNNFcLayer.cpp b/paddle/gserver/layers/MKLDNNFcLayer.cpp index 350ec65fffbc73c3a6e4245f763f4c6aa868f574..c8778bdd077c4b6d170140be92bdcdd7e8e81bb2 100644 --- a/paddle/gserver/layers/MKLDNNFcLayer.cpp +++ b/paddle/gserver/layers/MKLDNNFcLayer.cpp @@ -74,7 +74,7 @@ void MKLDNNFcLayer::convertWeightsToPaddle() { } void MKLDNNFcLayer::reshape( - int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) { + int& bs, int& ic, int& ih, int& iw, int& oc, int& oh, int& ow) { reshapeInput(bs, ih, iw); CHECK_EQ(iLayerSize_, inputLayers_[0]->getSize()); @@ -87,32 +87,29 @@ void MKLDNNFcLayer::reshape( } void MKLDNNFcLayer::resetFwd(std::vector& pipeline, - MKLDNNMatrixPtr& in, - MKLDNNMatrixPtr& wgt, - MKLDNNMatrixPtr& bias, + std::vector& inputs, MKLDNNMatrixPtr& out) { - resetFwdBuffers(in, wgt, bias, out); + resetFwdBuffers(inputs[0], wgtVal_, biasVal_, out); - resetFwdPD(fwdPD_, in, wgt, bias, out); + resetFwdPD(fwdPD_, inputs[0], wgtVal_, biasVal_, out); - resetFwdPipeline(pipeline, fwdPD_, in, wgt, bias, out); + resetFwdPipeline(pipeline, fwdPD_, inputs[0], wgtVal_, biasVal_, out); } void MKLDNNFcLayer::resetBwd(std::vector& pipeline, - MKLDNNMatrixPtr& in, - MKLDNNMatrixPtr& wgt, - MKLDNNMatrixPtr& bias, + std::vector& inputs, MKLDNNMatrixPtr& out) { std::shared_ptr bwdWgtPD; std::shared_ptr bwdDataPD; - resetBwdBuffers(in, wgt, bias, out); + resetBwdBuffers(inputs[0], wgtGrad_, biasGrad_, out); - resetBwdWgtPD(bwdWgtPD, wgt, bias, out); + resetBwdWgtPD(bwdWgtPD, wgtGrad_, biasGrad_, out); - resetBwdDataPD(bwdDataPD, in, out); + resetBwdDataPD(bwdDataPD, inputs[0], out); - resetBwdPipeline(pipeline, bwdWgtPD, bwdDataPD, in, wgt, bias, out); + resetBwdPipeline( + pipeline, bwdWgtPD, bwdDataPD, inputs[0], wgtGrad_, biasGrad_, out); } void MKLDNNFcLayer::updateWeights(const UpdateCallback& callback) { @@ -193,9 +190,9 @@ void MKLDNNFcLayer::resetBwdBuffers(MKLDNNMatrixPtr& in, MKLDNNMatrixPtr& wgt, MKLDNNMatrixPtr& bias, MKLDNNMatrixPtr& out) { - CHECK(inVal_ && outVal_); + CHECK(inVals_[0] && outVal_); resetOutGrad(out, outVal_->getPrimitiveDesc()); - resetInGrad(in, inVal_->getPrimitiveDesc()); + resetInGrad(in, inVals_[0]->getPrimitiveDesc()); CHECK(wgtVal_); resetWithMatrix(wgt, weight_->getWGrad(), wgtVal_->getPrimitiveDesc()); @@ -212,14 +209,15 @@ void MKLDNNFcLayer::resetBwdWgtPD( MKLDNNMatrixPtr& wgt, MKLDNNMatrixPtr& bias, MKLDNNMatrixPtr& out) { - CHECK(inVal_); - fc_bwdWgt::desc bwdWgtDesc = bias ? fc_bwdWgt::desc(inVal_->getMemoryDesc(), - wgt->getMemoryDesc(), - bias->getMemoryDesc(), - out->getMemoryDesc()) - : fc_bwdWgt::desc(inVal_->getMemoryDesc(), - wgt->getMemoryDesc(), - out->getMemoryDesc()); + CHECK(inVals_[0]); + fc_bwdWgt::desc bwdWgtDesc = + bias ? fc_bwdWgt::desc(inVals_[0]->getMemoryDesc(), + wgt->getMemoryDesc(), + bias->getMemoryDesc(), + out->getMemoryDesc()) + : fc_bwdWgt::desc(inVals_[0]->getMemoryDesc(), + wgt->getMemoryDesc(), + out->getMemoryDesc()); pd.reset(new fc_bwdWgt::primitive_desc(bwdWgtDesc, engine_, *fwdPD_)); } @@ -245,11 +243,11 @@ void MKLDNNFcLayer::resetBwdPipeline( MKLDNNMatrixPtr& wgt, MKLDNNMatrixPtr& bias, MKLDNNMatrixPtr& out) { - CHECK(inVal_); + CHECK(inVals_[0]); if (bias) { - bwdWgt_.reset(new fc_bwdWgt(*bwdWgtPD, *inVal_, *out, *wgt, *bias)); + bwdWgt_.reset(new fc_bwdWgt(*bwdWgtPD, *inVals_[0], *out, *wgt, *bias)); } else { - bwdWgt_.reset(new fc_bwdWgt(*bwdWgtPD, *inVal_, *out, *wgt)); + bwdWgt_.reset(new fc_bwdWgt(*bwdWgtPD, *inVals_[0], *out, *wgt)); } pipeline.push_back(*bwdWgt_); diff --git a/paddle/gserver/layers/MKLDNNFcLayer.h b/paddle/gserver/layers/MKLDNNFcLayer.h index ee861763ff3dc10ddb4c119358b80dbe1614aecb..283dc9b540531f6009ae6e2485b7c12d4e5cf2e3 100644 --- a/paddle/gserver/layers/MKLDNNFcLayer.h +++ b/paddle/gserver/layers/MKLDNNFcLayer.h @@ -52,18 +52,14 @@ public: const ParameterMap& parameterMap) override; void reshape( - int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) override; + int& bs, int& ic, int& ih, int& iw, int& oc, int& oh, int& ow) override; void resetFwd(std::vector& pipeline, - MKLDNNMatrixPtr& in, - MKLDNNMatrixPtr& wgt, - MKLDNNMatrixPtr& bias, + std::vector& inputs, MKLDNNMatrixPtr& out) override; void resetBwd(std::vector& pipeline, - MKLDNNMatrixPtr& in, - MKLDNNMatrixPtr& wgt, - MKLDNNMatrixPtr& bias, + std::vector& inputs, MKLDNNMatrixPtr& out) override; void updateWeights(const UpdateCallback& callback) override; @@ -73,11 +69,6 @@ public: void convertWeightsToPaddle() override; protected: - /** - * Forward functions: reset buffers(input, output, weight and bias), - * reset primitive descriptor, - * reset pipeline. - */ void resetFwdBuffers(MKLDNNMatrixPtr& in, MKLDNNMatrixPtr& wgt, MKLDNNMatrixPtr& bias, @@ -93,13 +84,6 @@ protected: MKLDNNMatrixPtr& wgt, MKLDNNMatrixPtr& bias, MKLDNNMatrixPtr& out); - - /** - * Backward functions: reset buffers(input, output, weight and bias), - * reset primitive descriptor for backward weight, - * reset primitive descriptor for backward data, - * reset pipeline. - */ void resetBwdBuffers(MKLDNNMatrixPtr& in, MKLDNNMatrixPtr& wgt, MKLDNNMatrixPtr& bias, diff --git a/paddle/gserver/layers/MKLDNNLRNLayer.cpp b/paddle/gserver/layers/MKLDNNLRNLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..ac217f1363dbd0360645bbe07cd71a17cc931a79 --- /dev/null +++ b/paddle/gserver/layers/MKLDNNLRNLayer.cpp @@ -0,0 +1,163 @@ +/* Copyright (c) 2017 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. */ + +#include "MKLDNNLRNLayer.h" +#include "paddle/utils/Logging.h" + +using namespace mkldnn; // NOLINT +typedef memory::format format; + +namespace paddle { + +REGISTER_LAYER(mkldnn_lrn, MKLDNNLRNLayer); + +bool MKLDNNLRNLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + if (!MKLDNNLayer::init(layerMap, parameterMap)) { + return false; + } + + /* the size of inputs for norm-layer is 1 */ + CHECK_EQ(config_.inputs_size(), 1); + const NormConfig& conf = config_.inputs(0).norm_conf(); + localSize_ = conf.size(); + alpha_ = conf.scale(); + beta_ = conf.pow(); + + ic_ = conf.channels(); + oc_ = ic_; + iw_ = conf.img_size(); + ow_ = conf.output_x(); + ih_ = conf.has_img_size_y() ? conf.img_size_y() : conf.img_size(); + oh_ = conf.has_output_y() ? conf.output_y() : conf.output_x(); + CHECK_EQ(iw_, ow_); + CHECK_EQ(ih_, oh_); + return true; +} + +void MKLDNNLRNLayer::reshape( + int& bs, int& ic, int& ih, int& iw, int& oc, int& oh, int& ow) { + CHECK_EQ(inputLayers_.size(), 1UL); + reshapeInput(bs, ih, iw); + // ic_ and oc can not be changed + CHECK_EQ((size_t)ic, + inputLayers_[0]->getOutputValue()->getElementCnt() / bs / ih / iw) + << "Input channel can not be changed"; + oh = ih; + ow = iw; + reshapeOutput(oh, ow); + resizeOutput(bs, oc * oh * ow); +} + +void MKLDNNLRNLayer::resetFwd(std::vector& pipeline, + std::vector& inputs, + MKLDNNMatrixPtr& out) { + resetFwdBuffers(inputs[0], out); + + resetFwdPD(fwdPD_, inputs[0], out); + + resetFwdPipeline(pipeline, fwdPD_, inputs[0], out); +} + +void MKLDNNLRNLayer::resetBwd(std::vector& pipeline, + std::vector& inputs, + MKLDNNMatrixPtr& out) { + std::shared_ptr pd; + + resetBwdBuffers(inputs[0], out); + + resetBwdPD(pd, inputs[0], out); + + resetBwdPipeline(pipeline, pd, inputs[0], out); +} + +void MKLDNNLRNLayer::resetFwdBuffers(MKLDNNMatrixPtr& in, + MKLDNNMatrixPtr& out) { + resetInValue(in); + CHECK(in); + resetOutValue(out, in->getPrimitiveDesc()); +} + +void MKLDNNLRNLayer::resetFwdPD(std::shared_ptr& pd, + MKLDNNMatrixPtr in, + MKLDNNMatrixPtr out) { + prop_kind pk = passType_ == PASS_TEST ? prop_kind::forward_scoring + : prop_kind::forward_training; + auto fwdDesc = lrn_fwd::desc(pk, + algorithm::lrn_across_channels, + in->getMemoryDesc(), + localSize_, + alpha_, + beta_, + 1.0f); + pd.reset(new lrn_fwd::primitive_desc(fwdDesc, engine_)); + // prepare workspace if necessary + workspace_ = + passType_ != PASS_TEST + ? std::make_shared(memory(pd->workspace_primitive_desc())) + : nullptr; +} + +void MKLDNNLRNLayer::resetFwdPipeline( + std::vector& pipeline, + std::shared_ptr& pd, + MKLDNNMatrixPtr& in, + MKLDNNMatrixPtr& out) { + fwd_ = workspace_ + ? std::make_shared(lrn_fwd(*pd, *in, *workspace_, *out)) + : std::make_shared(lrn_fwd(*pd, *in, *out)); + pipeline.push_back(*fwd_); +} + +void MKLDNNLRNLayer::resetBwdBuffers(MKLDNNMatrixPtr& in, + MKLDNNMatrixPtr& out) { + CHECK(inVals_[0] && outVal_); + resetOutGrad(out, outVal_->getPrimitiveDesc()); + resetInGrad(in, inVals_[0]->getPrimitiveDesc()); +} + +void MKLDNNLRNLayer::resetBwdPD(std::shared_ptr& pd, + MKLDNNMatrixPtr& in, + MKLDNNMatrixPtr& out) { + pd = nullptr; + if (in == nullptr) { + return; + } + CHECK(out); + auto bwdDesc = lrn_bwd::desc(algorithm::lrn_across_channels, + in->getMemoryDesc(), + out->getMemoryDesc(), + localSize_, + alpha_, + beta_, + 1.0f); + pd.reset(new lrn_bwd::primitive_desc(bwdDesc, engine_, *fwdPD_)); +} + +void MKLDNNLRNLayer::resetBwdPipeline( + std::vector& pipeline, + std::shared_ptr& pd, + MKLDNNMatrixPtr& in, + MKLDNNMatrixPtr& out) { + if (pd == nullptr) { + return; + } + CHECK(inVals_[0]); + CHECK(workspace_); + bwdData_ = std::make_shared( + lrn_bwd(*pd, *inVals_[0], *out, *workspace_, *in)); + pipeline.push_back(*bwdData_); +} + +} // namespace paddle diff --git a/paddle/gserver/layers/MKLDNNLRNLayer.h b/paddle/gserver/layers/MKLDNNLRNLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..cfe5621252c71a1de9a0a42a2a88e221e3e56972 --- /dev/null +++ b/paddle/gserver/layers/MKLDNNLRNLayer.h @@ -0,0 +1,78 @@ +/* Copyright (c) 2017 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 + +#include "MKLDNNLayer.h" +#include "mkldnn.hpp" + +namespace paddle { +typedef mkldnn::lrn_forward lrn_fwd; +typedef mkldnn::lrn_backward lrn_bwd; + +/** + * @brief A subclass of MKLDNNLayer LRN(Local Response Norm) layer. + * + * The config file api is mkldnn_lrn + */ +class MKLDNNLRNLayer : public MKLDNNLayer { +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/ + // test_lrn_backward.cpp, lrn need workspace for backward + std::shared_ptr workspace_; + + int localSize_; + float alpha_, beta_; // scale and pow in paddle + +public: + explicit MKLDNNLRNLayer(const LayerConfig& config) : MKLDNNLayer(config) {} + + ~MKLDNNLRNLayer() {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void reshape( + int& bs, int& ic, int& ih, int& iw, int& oc, int& oh, int& ow) override; + + void resetFwd(std::vector& pipeline, + std::vector& inputs, + MKLDNNMatrixPtr& out) override; + + void resetBwd(std::vector& pipeline, + std::vector& inputs, + MKLDNNMatrixPtr& out) override; + +protected: + void resetFwdBuffers(MKLDNNMatrixPtr& in, MKLDNNMatrixPtr& out); + void resetFwdPD(std::shared_ptr& pd, + MKLDNNMatrixPtr in, + MKLDNNMatrixPtr out); + void resetFwdPipeline(std::vector& pipeline, + std::shared_ptr& pd, + MKLDNNMatrixPtr& in, + MKLDNNMatrixPtr& out); + void resetBwdBuffers(MKLDNNMatrixPtr& in, MKLDNNMatrixPtr& out); + void resetBwdPD(std::shared_ptr& pd, + MKLDNNMatrixPtr& in, + MKLDNNMatrixPtr& out); + void resetBwdPipeline(std::vector& pipeline, + std::shared_ptr& pd, + MKLDNNMatrixPtr& in, + MKLDNNMatrixPtr& out); +}; + +} // namespace paddle diff --git a/paddle/gserver/layers/MKLDNNLayer.cpp b/paddle/gserver/layers/MKLDNNLayer.cpp index cf42da0735282d667d6b87061c8c59bf2f96e0be..6fbf3c7fdec2f537769adb660c67c5a597beb609 100644 --- a/paddle/gserver/layers/MKLDNNLayer.cpp +++ b/paddle/gserver/layers/MKLDNNLayer.cpp @@ -48,31 +48,20 @@ void MKLDNNLayer::forward(PassType passType) { REGISTER_TIMER_INFO("mkldnn_FwdTimer", getName().c_str()); CHECK(!inputLayers_.empty()); copySeqInfoToOutputs(); - size_t elemenCnt = inputLayers_[0]->getOutputValue()->getElementCnt(); - if (inputElemenCnt_ != elemenCnt) { + if (condition_ != keepCondition()) { VLOG(MKLDNN_BASE) << getName() << " reset mkldnn forward"; - // reset when input total sizes changed, not only the batchsize - inputElemenCnt_ = elemenCnt; - pipelineFwd_.clear(); + condition_ = keepCondition(); reshape(bs_, ic_, ih_, iw_, oc_, oh_, ow_); - // all cpu device output grad or value share output's + printSizeInfo(); + // the output_.value and output_.grad are shared with CPU device shareCPUDevice(); - resetFwd(pipelineFwd_, inVal_, wgtVal_, biasVal_, outVal_); - // MKLDNNLayer output value should be MKLDNNMatrix - // so external output value is necessary. - // Then external input value is not necessary, - // since input may be mkldnn internal buffer. - CHECK(extOutVal_) << "external output value is necessary"; - output_.value = std::dynamic_pointer_cast(extOutVal_); - CHECK(inVal_ && outVal_) << "internal memories are necessary"; - if (cvtInVal_) { - pipelineFwd_.insert(pipelineFwd_.begin(), *cvtInVal_); - } - if (cvtOutVal_) { - pipelineFwd_.push_back(*cvtOutVal_); - } + pipelineFwd_.clear(); + inVals_.resize(inputLayers_.size(), nullptr); + extInVals_.resize(inputLayers_.size(), nullptr); + cvtInVals_.resize(inputLayers_.size(), nullptr); + resetFwd(pipelineFwd_, inVals_, outVal_); + prepareValueConversions(pipelineFwd_); convertWeightsFromPaddle(); - printSizeInfo(); printValueFormat(); needResetBwd_ = true; } @@ -80,8 +69,8 @@ void MKLDNNLayer::forward(PassType passType) { if (inputLayers_[0]->getType() == "data" && inputLayers_.size() == 1) { // Update input value data when input layer is "data" type, // since the input value data address might be changed. - CHECK(extInVal_); - extInVal_->setData(getInputValue(0, CPU_DEVICE)->getData()); + CHECK(extInVals_[0]); + extInVals_[0]->setData(getInputValue(0, CPU_DEVICE)->getData()); } if (!outputOnlyMKLDNN_) { @@ -99,22 +88,13 @@ void MKLDNNLayer::backward(const UpdateCallback& callback) { if (needResetBwd_) { VLOG(MKLDNN_BASE) << getName() << " reset mkldnn backward"; pipelineBwd_.clear(); + inGrads_.resize(inputLayers_.size(), nullptr); + extInGrads_.resize(inputLayers_.size(), nullptr); + cvtInGrads_.resize(inputLayers_.size(), nullptr); pipelineMergeGrad_.clear(); mergeGrad_ = nullptr; - resetBwd(pipelineBwd_, inGrad_, wgtGrad_, biasGrad_, outGrad_); - // external output grad is not necessary - // since output may be mkldnn internal buffer or merge them directly. - CHECK(outGrad_) << "internal output grad is necessary"; - if (extOutGrad_) { - CHECK_EQ(extOutGrad_->getData(), output_.grad->getData()) - << "the external buffer should share the same data with output_.grad"; - } - if (cvtOutGrad_) { - pipelineBwd_.insert(pipelineBwd_.begin(), *cvtOutGrad_); - } - if (cvtInGrad_) { - pipelineBwd_.push_back(*cvtInGrad_); - } + resetBwd(pipelineBwd_, inGrads_, outGrad_); + prepareGradConversions(pipelineBwd_); printGradFormat(); needResetBwd_ = false; } @@ -141,8 +121,8 @@ void MKLDNNLayer::backward(const UpdateCallback& callback) { void MKLDNNLayer::reshapeInput(int& batchsize, int& height, int& width, - size_t inputIdx) { - const Argument& input = inputLayers_[inputIdx]->getOutput(); + size_t idx) { + const Argument& input = inputLayers_[idx]->getOutput(); batchsize = input.getBatchSize(); int h = input.getFrameHeight(); int w = input.getFrameWidth(); @@ -176,27 +156,30 @@ void MKLDNNLayer::resetWithMatrix(MKLDNNMatrixPtr& dnn, void MKLDNNLayer::resetInValue( MKLDNNMatrixPtr& in, const std::shared_ptr& intPD, - size_t inputIdx) { - cvtInVal_ = nullptr; - extInVal_ = nullptr; + size_t idx, + int inputChannel) { + cvtInVals_[idx] = nullptr; + extInVals_[idx] = nullptr; in = nullptr; - CHECK_GT(bs_ * ic_ * ih_ * iw_, 0); + inputChannel = inputChannel == 0 ? ic_ : inputChannel; + CHECK_GT(bs_ * inputChannel * ih_ * iw_, 0); auto extPD = MKLDNNMatrix::createPrimitiveDesc( - {bs_, ic_, ih_, iw_}, format::nchw, engine_); - const MatrixPtr& inMat = inputLayers_[inputIdx]->getOutputValue(); - extInVal_ = std::dynamic_pointer_cast(inMat); - CHECK_EQ(inputIsOnlyMKLDNN(), extInVal_ != nullptr); - if (extInVal_ == nullptr || extInVal_->getFormat() == format::nc) { - extInVal_ = MKLDNNMatrix::create(extPD, inMat); + {bs_, inputChannel, ih_, iw_}, format::nchw, engine_); + const MatrixPtr& inMat = inputLayers_[idx]->getOutputValue(); + extInVals_[idx] = std::dynamic_pointer_cast(inMat); + CHECK_EQ(inputIsOnlyMKLDNN(), extInVals_[idx] != nullptr); + if (extInVals_[idx] == nullptr || + extInVals_[idx]->getFormat() == format::nc) { + extInVals_[idx] = MKLDNNMatrix::create(extPD, inMat); } - in = extInVal_; + in = extInVals_[idx]; if (nullptr == intPD || in->getPrimitiveDesc() == *intPD) { return; } // need create reorder in = MKLDNNMatrix::create(*intPD); - cvtInVal_ = MKLDNNMatrix::createReorder(extInVal_, in); - CHECK(cvtInVal_) << "should not be emptry"; + cvtInVals_[idx] = MKLDNNMatrix::createReorder(extInVals_[idx], in); + CHECK(cvtInVals_[idx]) << "should not be emptry"; } void MKLDNNLayer::resetOutValue(MKLDNNMatrixPtr& out, @@ -218,11 +201,11 @@ void MKLDNNLayer::resetOutValue(MKLDNNMatrixPtr& out, void MKLDNNLayer::resetInGrad(MKLDNNMatrixPtr& in, memory::primitive_desc intPD, - size_t inputIdx) { - cvtInGrad_ = nullptr; - extInGrad_ = nullptr; + size_t idx) { + cvtInGrads_[idx] = nullptr; + extInGrads_[idx] = nullptr; in = nullptr; - LayerPtr& input = inputLayers_[inputIdx]; + LayerPtr& input = inputLayers_[idx]; if (input->getOutputGrad() == nullptr) { // no need input grad return; @@ -237,23 +220,25 @@ void MKLDNNLayer::resetInGrad(MKLDNNMatrixPtr& in, in = MKLDNNMatrix::create(intPD, inMat); Argument& arg = input->getOutput(this->getName()); arg.grad = std::dynamic_pointer_cast(in); - CHECK_PRIMITIVE_DESC_EQ(inVal_, intPD); + CHECK_PRIMITIVE_DESC_EQ(inVals_[idx], intPD); if (inputIsOnlyMKLDNN()) { return; } - extInGrad_ = in; - if (isPaddleFormat(extInGrad_->getFormat())) { + extInGrads_[idx] = in; + if (isPaddleFormat(extInGrads_[idx]->getFormat())) { return; } // need create reorder - CHECK(extInVal_ != nullptr && isPaddleFormat(extInVal_->getFormat())) + CHECK(extInVals_[idx] != nullptr && + isPaddleFormat(extInVals_[idx]->getFormat())) << "should have external input value and the format must be nchw(nc)"; - extInGrad_ = MKLDNNMatrix::create(extInVal_->getPrimitiveDesc(), inMat); - CHECK_PRIMITIVE_DESC_EQ(inVal_, intPD); + extInGrads_[idx] = + MKLDNNMatrix::create(extInVals_[idx]->getPrimitiveDesc(), inMat); + CHECK_PRIMITIVE_DESC_EQ(inVals_[idx], intPD); in = MKLDNNMatrix::create(intPD); - cvtInGrad_ = MKLDNNMatrix::createReorder(in, extInGrad_); - CHECK(cvtInGrad_); + cvtInGrads_[idx] = MKLDNNMatrix::createReorder(in, extInGrads_[idx]); + CHECK(cvtInGrads_[idx]); } void MKLDNNLayer::resetOutGrad(MKLDNNMatrixPtr& out, @@ -309,22 +294,8 @@ void MKLDNNLayer::resetMergeGrad(MKLDNNMatrixPtr& out) { srcs.push_back(*src); } - // TODO(TJ): remove me when mkldnn sum support different formats - for (size_t i = 1; i < srcPDs.size(); ++i) { - CHECK(srcPDs[0] == srcPDs[i]); - } - tmpOutGrad_ = out; - tmpCvt_ = nullptr; - if (out->getPrimitiveDesc() != srcPDs[0]) { - tmpOutGrad_ = MKLDNNMatrix::create(srcPDs[0]); - tmpCvt_ = MKLDNNMatrix::createReorder(tmpOutGrad_, out); - CHECK(tmpCvt_); - pipelineMergeGrad_.push_back(*tmpCvt_); - } - - auto sumPD = - sum::primitive_desc(tmpOutGrad_->getMemoryDesc(), scales, srcPDs); - mergeGrad_.reset(new sum(sumPD, srcs, *tmpOutGrad_)); + auto sumPD = sum::primitive_desc(out->getMemoryDesc(), scales, srcPDs); + mergeGrad_.reset(new sum(sumPD, srcs, *out)); pipelineMergeGrad_.insert(pipelineMergeGrad_.begin(), *mergeGrad_); } diff --git a/paddle/gserver/layers/MKLDNNLayer.h b/paddle/gserver/layers/MKLDNNLayer.h index 4c42df1bee75fa7b28c2001c30797cc0df7c5554..e48b9b5a91f7f17cb3f31e9140f1428ba8954a20 100644 --- a/paddle/gserver/layers/MKLDNNLayer.h +++ b/paddle/gserver/layers/MKLDNNLayer.h @@ -34,15 +34,16 @@ typedef std::shared_ptr MKLDNNLayerPtr; */ class MKLDNNLayer : public Layer { protected: - // input value element count - size_t inputElemenCnt_; // batch size int bs_; + // their sizes are always from the first input layer // input image channel, height and width int ic_, ih_, iw_; // output image channel, height and width int oc_, oh_, ow_; + // the condition that forward need be reset + size_t condition_; // backward also need reset after reset forward handle bool needResetBwd_; @@ -67,18 +68,18 @@ protected: * When all layers are mkldnn layers, they could save internal data. */ // below MKLDNNMatrix buffers are all internal buffers - MKLDNNMatrixPtr inVal_; - MKLDNNMatrixPtr inGrad_; + std::vector inVals_; + std::vector inGrads_; MKLDNNMatrixPtr outVal_; MKLDNNMatrixPtr outGrad_; // below are external value and grad - MKLDNNMatrixPtr extInVal_; - MKLDNNMatrixPtr extInGrad_; + std::vector extInVals_; + std::vector extInGrads_; MKLDNNMatrixPtr extOutVal_; MKLDNNMatrixPtr extOutGrad_; // convert handle between external and internal buffers - std::shared_ptr cvtInVal_; - std::shared_ptr cvtInGrad_; + std::vector> cvtInVals_; + std::vector> cvtInGrads_; std::shared_ptr cvtOutVal_; std::shared_ptr cvtOutGrad_; @@ -93,23 +94,11 @@ protected: std::vector pipelineMergeGrad_; // tmp input argument to save input grad, only used to merge grad Argument tmpInArg_; - // since mkldnn sum do not support different formats: - // can refer to https://github.com/01org/mkl-dnn/issues/134 - // so need create reorder manually and save tmp MKLDNNMatrix - MKLDNNMatrixPtr tmpOutGrad_; - std::shared_ptr tmpCvt_; public: explicit MKLDNNLayer(const LayerConfig& config) : Layer(config), - inputElemenCnt_(0), - bs_(0), - ic_(0), - ih_(0), - iw_(0), - oc_(0), - oh_(0), - ow_(0), + condition_(0), needResetBwd_(true), outputOnlyMKLDNN_(false), engine_(mkldnn::engine::cpu, 0), @@ -125,31 +114,28 @@ public: virtual void backward(const UpdateCallback& callback); /** - * reshape the input image sizes - * and reset output image and buffer size - * output channel can not be changed + * reshape the input and output channels and image sizes + * and reset output buffer size */ virtual void reshape( - int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) = 0; + int& bs, int& ic, int& ih, int& iw, int& oc, int& oh, int& ow) = 0; /** * reset the mkldnn forward primitve and memories * only would be called when input size changes + * weight and bias buffers should be coverd by child class itself */ virtual void resetFwd(std::vector& pipeline, - MKLDNNMatrixPtr& in, - MKLDNNMatrixPtr& wgt, - MKLDNNMatrixPtr& bias, + std::vector& inputs, MKLDNNMatrixPtr& out) = 0; /** * reset the mkldnn backward primitve and memories * only would be called when needed + * weight and bias buffers should be coverd by child class itself */ virtual void resetBwd(std::vector& pipeline, - MKLDNNMatrixPtr& in, - MKLDNNMatrixPtr& wgt, - MKLDNNMatrixPtr& bias, + std::vector& inputs, MKLDNNMatrixPtr& out) = 0; /** @@ -175,13 +161,19 @@ public: void addOutputArgument(int deviceId) { Layer::addOutputArgument(deviceId); } protected: + /** + * Some layers may have different condition to reset the forward. + * The function returns the condition that do not need reset forward. + */ + inline virtual size_t keepCondition() { + // reset when the first input element size changed, not only the batchsize + return inputLayers_[0]->getOutputValue()->getElementCnt(); + } + /** * reshape the input image sizes and input batchsize */ - void reshapeInput(int& batchsize, - int& height, - int& width, - size_t inputIdx = 0); + void reshapeInput(int& batchsize, int& height, int& width, size_t idx = 0); /** * reshape output image sizes @@ -199,11 +191,13 @@ protected: /** * reset input value from input MKLDNNMatrix and internal primitive desc. * reset both internal and external buffer and create reorder if necessary. + * input channel may be different in concat. */ void resetInValue( MKLDNNMatrixPtr& in, const std::shared_ptr& intPD = nullptr, - size_t inputIdx = 0); + size_t idx = 0, + int inputChannel = 0); /** * reset output value from internal primitive desc. @@ -218,7 +212,7 @@ protected: */ void resetInGrad(MKLDNNMatrixPtr& in, mkldnn::memory::primitive_desc intPD, - size_t inputIdx = 0); + size_t idx = 0); /** * reset output grad from internal primitive desc. @@ -296,17 +290,19 @@ protected: * print the mkldnn memory format of value */ virtual void printValueFormat() { - if (extInVal_) { - VLOG(MKLDNN_FMTS) << extInVal_->getFormat() << " >>> "; - } - if (inVal_) { - VLOG(MKLDNN_FMTS) << inVal_->getFormat() << " >>>"; + for (size_t i = 0; i < inVals_.size(); ++i) { + if (!inVals_[i]) { + continue; + } + VLOG(MKLDNN_FMTS) << "Input " << i << ", " << inputLayers_[i]->getName() + << ": " << (extInVals_[i] ? extInVals_[i]->getFormat() + : inVals_[i]->getFormat()) + << " >>> " << inVals_[i]->getFormat() << " >>>"; } if (outVal_) { - VLOG(MKLDNN_FMTS) << outVal_->getFormat() << " >>> "; - } - if (extOutVal_) { - VLOG(MKLDNN_FMTS) << extOutVal_->getFormat(); + VLOG(MKLDNN_FMTS) << outVal_->getFormat() << " >>> " + << (extOutVal_ ? extOutVal_->getFormat() + : outVal_->getFormat()); } if (wgtVal_) { VLOG(MKLDNN_FMTS) << "Weight value format: " << wgtVal_->getFormat(); @@ -320,17 +316,19 @@ protected: * print the mkldnn memory format of grad */ virtual void printGradFormat() { - if (extOutGrad_) { - VLOG(MKLDNN_FMTS) << extOutGrad_->getFormat(); - } if (outGrad_) { - VLOG(MKLDNN_FMTS) << outGrad_->getFormat() << " <<< "; + VLOG(MKLDNN_FMTS) << outGrad_->getFormat() << " <<< " + << (extOutGrad_ ? extOutGrad_->getFormat() + : outGrad_->getFormat()); } - if (inGrad_) { - VLOG(MKLDNN_FMTS) << inGrad_->getFormat() << " <<<"; - } - if (extInGrad_) { - VLOG(MKLDNN_FMTS) << extInGrad_->getFormat() << " <<< "; + for (size_t i = 0; i < inGrads_.size(); ++i) { + if (!inGrads_[i]) { + continue; + } + VLOG(MKLDNN_FMTS) << "Input " << i << ", " << inputLayers_[i]->getName() + << ": " << (extInGrads_[i] ? extInGrads_[i]->getFormat() + : inGrads_[i]->getFormat()) + << " <<< " << inGrads_[i]->getFormat() << " <<<"; } if (wgtGrad_) { VLOG(MKLDNN_FMTS) << "Weight grad format: " << wgtGrad_->getFormat(); @@ -437,6 +435,41 @@ private: outputOtherDevice_[i].cpuSequenceDims = output_.cpuSequenceDims; } } + + void prepareValueConversions(std::vector& pipeline) { + // MKLDNNLayer output value should be MKLDNNMatrix + // so external output value is necessary. + // Then external input value is not necessary, + // since input may be mkldnn internal buffer. + CHECK(extOutVal_) << "external output value is necessary"; + output_.value = std::dynamic_pointer_cast(extOutVal_); + CHECK(inVals_[0] && outVal_) << "internal memories are necessary"; + for (size_t i = 0; i < cvtInVals_.size(); ++i) { + if (cvtInVals_[i]) { + pipeline.insert(pipeline.begin(), *cvtInVals_[i]); + } + } + if (cvtOutVal_) { + pipeline.push_back(*cvtOutVal_); + } + } + void prepareGradConversions(std::vector& pipeline) { + // external output grad is not necessary + // since output may be mkldnn internal buffer or merge them directly. + CHECK(outGrad_) << "internal output grad is necessary"; + if (extOutGrad_) { + CHECK_EQ(extOutGrad_->getData(), output_.grad->getData()) + << "the external buffer should share the same data with output_.grad"; + } + if (cvtOutGrad_) { + pipeline.insert(pipeline.begin(), *cvtOutGrad_); + } + for (size_t i = 0; i < cvtInGrads_.size(); ++i) { + if (cvtInGrads_[i]) { + pipeline.push_back(*cvtInGrads_[i]); + } + } + } }; } // namespace paddle diff --git a/paddle/gserver/layers/MKLDNNPoolLayer.cpp b/paddle/gserver/layers/MKLDNNPoolLayer.cpp index a18c455beab96ef25b5545281bae4d48cec98d9e..a8252593c8fbb8013ab909e74a057850ba54bcaa 100644 --- a/paddle/gserver/layers/MKLDNNPoolLayer.cpp +++ b/paddle/gserver/layers/MKLDNNPoolLayer.cpp @@ -58,10 +58,11 @@ bool MKLDNNPoolLayer::init(const LayerMap& layerMap, } void MKLDNNPoolLayer::reshape( - int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) { + int& bs, int& ic, int& ih, int& iw, int& oc, int& oh, int& ow) { reshapeInput(bs, ih, iw); // ic_ and oc can not be changed - CHECK_EQ(inputElemenCnt_ / bs / ih / iw, (size_t)ic) + CHECK_EQ((size_t)ic, + inputLayers_[0]->getOutputValue()->getElementCnt() / bs / ih / iw) << "Input channel can not be changed"; // cal output sizes @@ -74,29 +75,25 @@ void MKLDNNPoolLayer::reshape( } void MKLDNNPoolLayer::resetFwd(std::vector& pipeline, - MKLDNNMatrixPtr& in, - MKLDNNMatrixPtr& wgt, - MKLDNNMatrixPtr& bias, + std::vector& inputs, MKLDNNMatrixPtr& out) { - resetFwdBuffers(in, out); + resetFwdBuffers(inputs[0], out); - resetFwdPD(fwdPD_, in, out); + resetFwdPD(fwdPD_, inputs[0], out); - resetFwdPipeline(pipeline, fwdPD_, in, out); + resetFwdPipeline(pipeline, fwdPD_, inputs[0], out); } void MKLDNNPoolLayer::resetBwd(std::vector& pipeline, - MKLDNNMatrixPtr& in, - MKLDNNMatrixPtr& wgt, - MKLDNNMatrixPtr& bias, + std::vector& inputs, MKLDNNMatrixPtr& out) { std::shared_ptr pd; - resetBwdBuffers(in, out); + resetBwdBuffers(inputs[0], out); - resetBwdPD(pd, in, out); + resetBwdPD(pd, inputs[0], out); - resetBwdPipeline(pipeline, pd, in, out); + resetBwdPipeline(pipeline, pd, inputs[0], out); } void MKLDNNPoolLayer::resetFwdBuffers(MKLDNNMatrixPtr& in, @@ -151,9 +148,9 @@ void MKLDNNPoolLayer::resetFwdPipeline( void MKLDNNPoolLayer::resetBwdBuffers(MKLDNNMatrixPtr& in, MKLDNNMatrixPtr& out) { - CHECK(inVal_ && outVal_); + CHECK(inVals_[0] && outVal_); resetOutGrad(out, outVal_->getPrimitiveDesc()); - resetInGrad(in, inVal_->getPrimitiveDesc()); + resetInGrad(in, inVals_[0]->getPrimitiveDesc()); } void MKLDNNPoolLayer::resetBwdPD(std::shared_ptr& pd, diff --git a/paddle/gserver/layers/MKLDNNPoolLayer.h b/paddle/gserver/layers/MKLDNNPoolLayer.h index c5ec87828bfb28b4502b4ec6b47287089c514204..dad60156f0ef7caa059ff6c70d1040e7e34c938f 100644 --- a/paddle/gserver/layers/MKLDNNPoolLayer.h +++ b/paddle/gserver/layers/MKLDNNPoolLayer.h @@ -53,18 +53,14 @@ public: const ParameterMap& parameterMap) override; void reshape( - int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) override; + int& bs, int& ic, int& ih, int& iw, int& oc, int& oh, int& ow) override; void resetFwd(std::vector& pipeline, - MKLDNNMatrixPtr& in, - MKLDNNMatrixPtr& wgt, - MKLDNNMatrixPtr& bias, + std::vector& inputs, MKLDNNMatrixPtr& out) override; void resetBwd(std::vector& pipeline, - MKLDNNMatrixPtr& in, - MKLDNNMatrixPtr& wgt, - MKLDNNMatrixPtr& bias, + std::vector& inputs, MKLDNNMatrixPtr& out) override; void printSizeInfo() override { @@ -75,11 +71,6 @@ public: } protected: - /** - * Forward functions: reset buffers(input, output), - * reset primitive descriptor, - * reset pipeline. - */ void resetFwdBuffers(MKLDNNMatrixPtr& in, MKLDNNMatrixPtr& out); void resetFwdPD(std::shared_ptr& pd, MKLDNNMatrixPtr in, @@ -88,12 +79,6 @@ protected: std::shared_ptr& pd, MKLDNNMatrixPtr& in, MKLDNNMatrixPtr& out); - - /** - * Backward functions: reset buffers(input, output), - * reset primitive descriptor, - * reset pipeline. - */ void resetBwdBuffers(MKLDNNMatrixPtr& in, MKLDNNMatrixPtr& out); void resetBwdPD(std::shared_ptr& pd, MKLDNNMatrixPtr& in, diff --git a/paddle/gserver/layers/PoolLayer.cpp b/paddle/gserver/layers/PoolLayer.cpp index 87613a96c5b3c2da212f63e9e678bcd22308b08e..fceb389d06d8d2cb0357186bf83edda9957c6c19 100644 --- a/paddle/gserver/layers/PoolLayer.cpp +++ b/paddle/gserver/layers/PoolLayer.cpp @@ -45,6 +45,8 @@ bool PoolLayer::init(const LayerMap& layerMap, strideY_ = conf.has_stride_y() ? conf.stride_y() : conf.stride(); confPaddingY_ = conf.has_padding_y() ? conf.padding_y() : conf.padding(); outputY_ = conf.has_output_y() ? conf.output_y() : conf.output_x(); + + excludeMode_ = conf.has_exclude_mode() ? conf.exclude_mode() : true; return true; } diff --git a/paddle/gserver/layers/PoolLayer.h b/paddle/gserver/layers/PoolLayer.h index d43292ad2d4bbe1229ca59ca21bee92c9ec006a3..9df672a935868e9c61f4dd1fd47a9c309b214f12 100644 --- a/paddle/gserver/layers/PoolLayer.h +++ b/paddle/gserver/layers/PoolLayer.h @@ -38,6 +38,8 @@ protected: std::string poolType_; + bool excludeMode_; + public: explicit PoolLayer(const LayerConfig& config) : Layer(config) {} diff --git a/paddle/gserver/layers/PoolProjection.cpp b/paddle/gserver/layers/PoolProjection.cpp index d90b438448eb72e72e22e9a91a3cbcd84ac7e6cb..6a9de394cee3769784a38f5512b15f52b1ed6fa1 100644 --- a/paddle/gserver/layers/PoolProjection.cpp +++ b/paddle/gserver/layers/PoolProjection.cpp @@ -36,6 +36,8 @@ PoolProjection::PoolProjection(const ProjectionConfig& config, strideY_ = conf.has_stride_y() ? conf.stride_y() : conf.stride(); confPaddingY_ = conf.has_padding_y() ? conf.padding_y() : conf.padding(); outputY_ = conf.has_output_y() ? conf.output_y() : conf.output_x(); + + excludeMode_ = conf.has_exclude_mode() ? conf.exclude_mode() : true; } size_t PoolProjection::getSize() { @@ -141,7 +143,8 @@ void AvgPoolProjection::forward() { outputY_, outputX_, confPaddingY_, - confPadding_); + confPadding_, + excludeMode_); } void AvgPoolProjection::backward(const UpdateCallback& callback) { @@ -166,6 +169,7 @@ void AvgPoolProjection::backward(const UpdateCallback& callback) { 1, 1, confPaddingY_, - confPadding_); + confPadding_, + excludeMode_); } } // namespace paddle diff --git a/paddle/gserver/layers/PoolProjection.h b/paddle/gserver/layers/PoolProjection.h index 9a75f465f6fbb2f2a928b0e36fcfbe0e510d7b3a..a0412714bca7a273e999e4d6bd552e833d20d69c 100644 --- a/paddle/gserver/layers/PoolProjection.h +++ b/paddle/gserver/layers/PoolProjection.h @@ -28,6 +28,7 @@ protected: int confPaddingY_, confPadding_; size_t channels_; std::string poolType_; + bool excludeMode_; public: PoolProjection(const ProjectionConfig& config, diff --git a/paddle/gserver/layers/ROIPoolLayer.cpp b/paddle/gserver/layers/ROIPoolLayer.cpp index 02402894d3354a6af221948a3360ef830881bf39..7d7c30b4d89e2dd137e7fc7de3159c07bbab9fb4 100644 --- a/paddle/gserver/layers/ROIPoolLayer.cpp +++ b/paddle/gserver/layers/ROIPoolLayer.cpp @@ -13,6 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "ROIPoolLayer.h" +#include namespace paddle { @@ -83,12 +84,15 @@ void ROIPoolLayer::forward(PassType passType) { size_t poolChannelOffset = pooledHeight_ * pooledWidth_; real* outputData = outputValue->getData(); - Matrix::resizeOrCreate(maxIdxs_, - numROIs, - channels_ * pooledHeight_ * pooledWidth_, - false, - false); - real* argmaxData = maxIdxs_->getData(); + real* argmaxData = nullptr; + if (passType != PASS_TEST) { + Matrix::resizeOrCreate(maxIdxs_, + numROIs, + channels_ * pooledHeight_ * pooledWidth_, + false, + false); + argmaxData = maxIdxs_->getData(); + } for (size_t n = 0; n < numROIs; ++n) { // the first five elememts of each RoI should be: @@ -126,8 +130,8 @@ void ROIPoolLayer::forward(PassType passType) { bool isEmpty = (hend <= hstart) || (wend <= wstart); size_t poolIndex = ph * pooledWidth_ + pw; - if (isEmpty) { - outputData[poolIndex] = 0; + outputData[poolIndex] = isEmpty ? 0 : -FLT_MAX; + if (argmaxData) { argmaxData[poolIndex] = -1; } @@ -136,7 +140,9 @@ void ROIPoolLayer::forward(PassType passType) { size_t index = h * width_ + w; if (batchData[index] > outputData[poolIndex]) { outputData[poolIndex] = batchData[index]; - argmaxData[poolIndex] = index; + if (argmaxData) { + argmaxData[poolIndex] = index; + } } } } @@ -144,7 +150,9 @@ void ROIPoolLayer::forward(PassType passType) { } batchData += channelOffset; outputData += poolChannelOffset; - argmaxData += poolChannelOffset; + if (argmaxData) { + argmaxData += poolChannelOffset; + } } bottomROIs += roiOffset; } diff --git a/paddle/gserver/layers/SequenceToBatch.cpp b/paddle/gserver/layers/SequenceToBatch.cpp index 5fa7b6f4881b9582b540a5b1bfe849220cc2a4ea..6b769378d24838364701d0f128a7308c6195cc41 100644 --- a/paddle/gserver/layers/SequenceToBatch.cpp +++ b/paddle/gserver/layers/SequenceToBatch.cpp @@ -171,12 +171,31 @@ void SequenceToBatch::sequence2BatchCopy(Matrix &batch, hl_sequence2batch_copy( batchData, seqData, idxData, seqWidth, batchCount, seq2batch); } else { - for (int i = 0; i < batchCount; ++i) { - if (seq2batch) { + if (seq2batch) { +#ifdef PADDLE_USE_MKLML + const int blockMemSize = 8 * 1024; + const int blockSize = blockMemSize / sizeof(real); +#pragma omp parallel for collapse(2) + for (int i = 0; i < batchCount; ++i) { + for (int j = 0; j < seqWidth; j += blockSize) { + memcpy(batch.rowBuf(i) + j, + sequence.rowBuf(idxData[i]) + j, + (j + blockSize > seqWidth) ? (seqWidth - j) * sizeof(real) + : blockMemSize); + } + } +#else + for (int i = 0; i < batchCount; ++i) { memcpy(batch.rowBuf(i), sequence.rowBuf(idxData[i]), seqWidth * sizeof(real)); - } else { + } +#endif + } else { +#ifdef PADDLE_USE_MKLML +#pragma omp parallel for +#endif + for (int i = 0; i < batchCount; ++i) { memcpy(sequence.rowBuf(idxData[i]), batch.rowBuf(i), seqWidth * sizeof(real)); diff --git a/paddle/gserver/tests/CMakeLists.txt b/paddle/gserver/tests/CMakeLists.txt index c295ea19c9ccb3d05c509a41925d2c36efdba8ef..b578a906c2027a1169a0098b93f8d0742920f99d 100644 --- a/paddle/gserver/tests/CMakeLists.txt +++ b/paddle/gserver/tests/CMakeLists.txt @@ -1,5 +1,4 @@ # gserver pacakge unittests - add_simple_unittest(test_LinearChainCRF) add_simple_unittest(test_RecurrentLayer) @@ -29,6 +28,26 @@ gserver_test(test_KmaxSeqScore) gserver_test(test_Expand) gserver_test(test_MaxPoolingWithMaskOutput) +set(PYTHON_PATH + ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d + ${PADDLE_SOURCE_DIR}/python/:${PADDLE_SOURCE_DIR}/paddle/gserver/tests) +function(gserver_test_with_python TARGET) + add_unittest_without_exec(${TARGET} ${TARGET}.cpp) + add_test(NAME ${TARGET} + COMMAND ${PYTHON_PATH} ${CMAKE_CURRENT_BINARY_DIR}/${TARGET} + WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/) +endfunction() + +gserver_test_with_python(test_PyDataProvider2) +if(WITH_PYTHON) + gserver_test_with_python(test_PyDataProvider) +endif() +if(NOT MOBILE_INFERENCE) + gserver_test_with_python(test_CompareTwoNets) + # TODO(yuyang18): There is some bug in test_RecurrentGradientMachine, I will fix it. + gserver_test_with_python(test_RecurrentGradientMachine) +endif() + ########## test_MKLDNN layers and activations ########## if(WITH_MKLDNN) add_unittest_without_exec(test_MKLDNN @@ -36,87 +55,43 @@ if(WITH_MKLDNN) MKLDNNTester.cpp LayerGradUtil.cpp) add_test(NAME test_MKLDNN - COMMAND .set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python - ${CMAKE_CURRENT_BINARY_DIR}/test_MKLDNN + COMMAND ${PYTHON_PATH} ${CMAKE_CURRENT_BINARY_DIR}/test_MKLDNN WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle) endif() -############## test_PyDataProvider ######################## -if(WITH_PYTHON) - add_unittest_without_exec(test_PyDataProvider - test_PyDataProvider.cpp) - - add_test(NAME test_PyDataProvider - COMMAND .set_python_path.sh -d ./gserver/tests:${PADDLE_SOURCE_DIR}/python/ ${CMAKE_CURRENT_BINARY_DIR}/test_PyDataProvider - WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle) -endif() - ############### test_WarpCTCLayer ####################### if(NOT WITH_DOUBLE AND NOT MOBILE_INFERENCE) add_unittest_without_exec(test_WarpCTCLayer test_WarpCTCLayer.cpp) - add_test(NAME test_WarpCTCLayer COMMAND ${CMAKE_CURRENT_BINARY_DIR}/test_WarpCTCLayer --warpctc_dir=${WARPCTC_LIB_DIR} WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle) endif() if(NOT MOBILE_INFERENCE) -################## test_Evaluator ####################### + ################## test_Evaluator ############# add_unittest(test_Evaluator test_Evaluator.cpp) -############### test_RecurrentGradientMachine ############### - # TODO(yuyang18): There is some bug in test_RecurrentGradientMachine - # I will fix it. - add_unittest_without_exec(test_RecurrentGradientMachine - test_RecurrentGradientMachine.cpp) - add_test(NAME test_RecurrentGradientMachine - COMMAND .set_python_path.sh -d - ${PADDLE_SOURCE_DIR}/python:${PADDLE_SOURCE_DIR}/paddle/gserver/tests - ${CMAKE_CURRENT_BINARY_DIR}/test_RecurrentGradientMachine - WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle) - -############### test_NetworkCompare ############### + ########### test_NetworkCompare ############### add_unittest_without_exec(test_NetworkCompare test_NetworkCompare.cpp) if(WITH_GPU) - add_test(NAME test_NetworkCompare - COMMAND .set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python ${CMAKE_CURRENT_BINARY_DIR}/test_NetworkCompare --use_gpu=true - WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle) + set(use_gpu true) else() - add_test(NAME test_NetworkCompare - COMMAND .set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python ${CMAKE_CURRENT_BINARY_DIR}/test_NetworkCompare --use_gpu=false - WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle) + set(use_gpu false) endif() -endif() - - -add_unittest_without_exec(test_PyDataProvider2 - test_PyDataProvider2.cpp) - -add_test(NAME test_PyDataProvider2 - COMMAND .set_python_path.sh -d ${PADDLE_SOURCE_DIR}/paddle/gserver/tests:${PADDLE_SOURCE_DIR}/python ${CMAKE_CURRENT_BINARY_DIR}/test_PyDataProvider2 - WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle -) + add_test(NAME test_NetworkCompare + COMMAND ${PYTHON_PATH} ${CMAKE_CURRENT_BINARY_DIR}/test_NetworkCompare --use_gpu=${use_gpu} + WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle) -################# test_CompareSparse ################## -add_unittest_without_exec(test_CompareSparse - test_CompareSparse.cpp) -if(NOT ON_TRAVIS) - add_test(NAME test_CompareSparse - COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d - ${PADDLE_SOURCE_DIR}/python:${PADDLE_SOURCE_DIR}/paddle/gserver/tests - ./.set_port.sh -p port -n 6 - ${CMAKE_CURRENT_BINARY_DIR}/test_CompareSparse - WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/) + ############ test_CompareSparse ################ + add_unittest_without_exec(test_CompareSparse + test_CompareSparse.cpp) + if(NOT ON_TRAVIS) + add_test(NAME test_CompareSparse + COMMAND ${PYTHON_PATH} ./.set_port.sh -p port -n 6 + ${CMAKE_CURRENT_BINARY_DIR}/test_CompareSparse + WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/) + endif() endif() - -################ test_CompareTwoNets ###################### -add_unittest_without_exec(test_CompareTwoNets - test_CompareTwoNets.cpp) -add_test(NAME test_CompareTwoNets - COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d - ${PADDLE_SOURCE_DIR}/python:${PADDLE_SOURCE_DIR}/paddle/gserver/tests - ${CMAKE_CURRENT_BINARY_DIR}/test_CompareTwoNets - WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/) diff --git a/paddle/gserver/tests/mkldnn_simple_net.conf b/paddle/gserver/tests/mkldnn_simple_net.conf index 8bbe91e56d0ba6da06475ad16f3162ee1103ee02..0e9d6b31fa8776136b4eee29311383ae6bb21644 100644 --- a/paddle/gserver/tests/mkldnn_simple_net.conf +++ b/paddle/gserver/tests/mkldnn_simple_net.conf @@ -51,6 +51,8 @@ tmp = img_pool_layer(input=tmp, padding=1, pool_type=MaxPooling()) +tmp = img_cmrnorm_layer(input=tmp, size=5, scale=0.0001, power=0.75) + tmp = fc_layer(input=tmp, size=channels, bias_attr=False, diff --git a/paddle/gserver/tests/sequence_rnn_matched_inputs.py b/paddle/gserver/tests/sequence_rnn_matched_inputs.py index e2635b4400b13517bac716a5a0affeb16c218b09..59e8c91733c42b6f13f723321d21bca98ab78bb7 100644 --- a/paddle/gserver/tests/sequence_rnn_matched_inputs.py +++ b/paddle/gserver/tests/sequence_rnn_matched_inputs.py @@ -41,7 +41,7 @@ nonseq = embedding_layer(input=label, size=word_dim) # This hierarchical RNN is designed to be equivalent to the simple RNN in -# sequence_rnn_multi_unequalength_inputs.conf +# sequence_rnn_mixed_inputs.conf def outer_step(subseq, seq, nonseq, encoding): outer_mem = memory(name="outer_rnn_state", size=hidden_dim) diff --git a/paddle/gserver/tests/sequence_rnn_mixed_inputs.py b/paddle/gserver/tests/sequence_rnn_mixed_inputs.py index 84a66e294495c01e03dc83b38a531e482bed1292..6fe9dca6e2cb0e14fee346b8307f67b804328471 100644 --- a/paddle/gserver/tests/sequence_rnn_mixed_inputs.py +++ b/paddle/gserver/tests/sequence_rnn_mixed_inputs.py @@ -37,7 +37,7 @@ encoding = embedding_layer(input=data2, size=word_dim) # This hierarchical RNN is designed to be equivalent to the simple RNN in -# sequence_rnn_multi_unequalength_inputs.conf +# sequence_rnn_matched_inputs.conf def outer_step(subseq, seq, nonseq, encoding): outer_mem = memory(name="outer_rnn_state", size=hidden_dim) diff --git a/paddle/gserver/tests/test_LayerGrad.cpp b/paddle/gserver/tests/test_LayerGrad.cpp index cacf10692942f5eca2f6c498183f4acc00768460..a2f07937b8834e3f3fa7a6bf2ae10f29a8d84f29 100644 --- a/paddle/gserver/tests/test_LayerGrad.cpp +++ b/paddle/gserver/tests/test_LayerGrad.cpp @@ -238,9 +238,24 @@ void testProjectionConv(size_t groups, bool isDeconv) { /* caffeMode */ true); conv->set_output_x(output_x); conv->set_output_y(output_y); + LOG(INFO) << "DILATION:" << DILATION << "; output_x: " << output_x + << "; output_y: " << output_y; if (isDeconv) { + int deconv_image_x = imageSize(output_x, + (conv->filter_size() - 1) * DILATION + 1, + conv->padding(), + conv->stride(), + /* caffeMode */ true); + int deconv_image_y = imageSize(output_y, + (conv->filter_size_y() - 1) * DILATION + 1, + conv->padding_y(), + conv->stride_y(), + /* caffeMode */ true); + + LOG(INFO) << " deconv_image_x: " << deconv_image_x + << "; deconv_image_y: " << deconv_image_y; conf.set_input_size(output_x * output_y * CHANNELS); - conf.set_output_size(IMAGE_SIZE * IMAGE_SIZE * NUM_FILTERS); + conf.set_output_size(deconv_image_x * deconv_image_y * NUM_FILTERS); } else { conf.set_input_size(IMAGE_SIZE * IMAGE_SIZE * CHANNELS); conf.set_output_size(output_x * output_y * NUM_FILTERS); @@ -681,12 +696,13 @@ TEST(Layer, hsigmoidLayer) { config.layerConfig.add_inputs(); config.layerConfig.add_inputs(); - // Not support GPU now - testLayerGrad(config, - "hsigmoid", - 100, - /* trans */ false, /* useGpu */ - false); + for (auto useGpu : {false, true}) { + testLayerGrad(config, + "hsigmoid", + 100, + /* trans */ false, + /* useGpu */ useGpu); + } } TEST(Layer, multi_cross) { @@ -1210,7 +1226,10 @@ void setPoolConfig(TestConfig* config, pool->set_output_y(oh); } -void testPoolLayer(const string& poolType, bool trans, bool useGpu) { +void testPoolLayer(const string& poolType, + bool trans, + bool useGpu, + bool excludeMode = true) { TestConfig config; config.inputDefs.push_back({INPUT_DATA, "layer_0", 3136, 0}); LayerInputConfig* input = config.layerConfig.add_inputs(); @@ -1218,6 +1237,7 @@ void testPoolLayer(const string& poolType, bool trans, bool useGpu) { pool->set_img_size(14); pool->set_img_size_y(14); + pool->set_exclude_mode(excludeMode); setPoolConfig(&config, pool, poolType); config.layerConfig.set_size(pool->output_x() * pool->output_y() * pool->channels()); @@ -1249,16 +1269,26 @@ void testPoolLayer2(const string& poolType, bool trans, bool useGpu) { TEST(Layer, PoolLayer) { testPoolLayer("avg-projection", /* trans= */ false, /* useGpu= */ false); + testPoolLayer("avg-projection", + /* trans= */ false, + /* useGpu= */ false, + /* excludeMode= */ false); testPoolLayer("max-projection", /* trans= */ false, /* useGpu= */ false); testPoolLayer("max-pool-with-mask", /* trans= */ false, /* useGpu= */ false); #ifdef PADDLE_WITH_CUDA testPoolLayer("avg-projection", /* trans= */ false, /* useGpu= */ true); + testPoolLayer("avg-projection", + /* trans= */ false, + /* useGpu= */ true, + /* excludeMode= */ false); testPoolLayer("max-projection", /* trans= */ false, /* useGpu= */ true); testPoolLayer("cudnn-max-pool", /* trans= */ false, /* useGpu= */ true); testPoolLayer("cudnn-avg-pool", /* trans= */ false, /* useGpu= */ true); testPoolLayer2("cudnn-max-pool", /* trans= */ false, /* useGpu= */ true); testPoolLayer2("cudnn-avg-pool", /* trans= */ false, /* useGpu= */ true); + testPoolLayer2( + "cudnn-avg-incl-pad-pool", /* trans= */ false, /* useGpu= */ true); testPoolLayer("max-pool-with-mask", /* trans= */ false, /* useGpu= */ true); #endif } @@ -2464,6 +2494,25 @@ TEST(Layer, L2DistanceLayer) { } } +void testFactorizationMachineLayer(InputType type, bool useGpu) { + const int FACTOR_SIZE = 10; + TestConfig config; + config.layerConfig.set_type("factorization_machine"); + config.layerConfig.set_factor_size(FACTOR_SIZE); + config.layerConfig.set_size(1); + config.biasSize = 0; + config.inputDefs.push_back({type, "layer_0", 128, 1280}); + config.layerConfig.add_inputs(); + testLayerGrad(config, "factorization_machine", 16, false, useGpu, false); +} + +TEST(Layer, FactorizationMachineLayer) { + for (auto useGpu : {false, true}) { + testFactorizationMachineLayer(INPUT_DATA, useGpu); + } + testFactorizationMachineLayer(INPUT_SPARSE_FLOAT_VALUE_DATA, false); +} + int main(int argc, char** argv) { testing::InitGoogleTest(&argc, argv); initMain(argc, argv); diff --git a/paddle/gserver/tests/test_MKLDNN.cpp b/paddle/gserver/tests/test_MKLDNN.cpp index 42644e9601a82ea81c417adc6441edeb036998e2..ad1dbc3ee2bfd00a94de06f1e1b2ffe64f19b417 100644 --- a/paddle/gserver/tests/test_MKLDNN.cpp +++ b/paddle/gserver/tests/test_MKLDNN.cpp @@ -272,6 +272,51 @@ TEST(MKLDNNLayer, BatchNormLayer) { testBatchNormLayer({4, 16, 8, 10}); } +struct testLRNDesc { + int bs, ic, ih, iw; + float scale, pow; + int localSize; +}; + +void getMKLDNNLRNConfig(TestConfig& cfg, const testLRNDesc& pm) { + cfg.layerConfig.set_type("mkldnn_lrn"); + cfg.layerConfig.set_active_type("relu"); + size_t layerSize = pm.ic * pm.ih * pm.iw; + cfg.inputDefs.push_back({INPUT_DATA, "layer_0", layerSize, 0}); + LayerInputConfig* input = cfg.layerConfig.add_inputs(); + NormConfig* norm = input->mutable_norm_conf(); + norm->set_channels(pm.ic); + norm->set_size(pm.localSize); + norm->set_scale(pm.scale); + norm->set_pow(pm.pow); + norm->set_blocked(0); + norm->set_img_size(pm.iw); + norm->set_img_size_y(pm.ih); + norm->set_output_x(norm->img_size()); + norm->set_output_y(norm->img_size_y()); + cfg.layerConfig.set_size(layerSize); + cfg.biasSize = 0; +} + +void testLRNLayer(const testLRNDesc& pm) { + TestConfig dnnConfig; + getMKLDNNLRNConfig(dnnConfig, pm); + // mkldnn_lrn <==> norm with cmrnorm-projection type + TestConfig refConfig = dnnConfig; + refConfig.layerConfig.set_type("norm"); + LayerInputConfig* input = refConfig.layerConfig.mutable_inputs(0); + NormConfig* norm = input->mutable_norm_conf(); + norm->set_norm_type("cmrnorm-projection"); + norm->set_scale(norm->scale() / norm->size()); + RUN_MKLDNN_TEST(dnnConfig, refConfig, pm) +} + +TEST(MKLDNNLayer, LRNLayer) { + testLRNLayer({4, 10, 12, 12, 0.001f, 0.75f, 5}); + testLRNLayer({2, 32, 6, 6, 0.001f, 0.75f, 5}); + testLRNLayer({4, 16, 8, 10, 0.01f, 0.5f, 5}); +} + struct testImageDesc { int bs, ic, ih, iw; }; @@ -315,7 +360,7 @@ TEST(MKLDNNLayer, AddtoLayer) { static void getMKLDNNConcatConfig(TestConfig& cfg, const std::vector& inputs) { - CHECK_GE(inputs.size(), 2) << "at least two inputs"; + CHECK_GE(inputs.size(), 2UL) << "at least two inputs"; int oc = inputs[0].ic; for (size_t i = 1; i < inputs.size(); ++i) { CHECK_EQ(inputs[i].bs, inputs[0].bs); diff --git a/paddle/math/Allocator.h b/paddle/math/Allocator.h index 94ef561f066a127496e2849a419835e175c526d7..17563bf5e1649361b83b896bf864b922296a5487 100644 --- a/paddle/math/Allocator.h +++ b/paddle/math/Allocator.h @@ -48,7 +48,7 @@ public: */ virtual void* alloc(size_t size) { void* ptr; -#ifdef PADDLE_USE_MKLDNN +#ifdef PADDLE_WITH_MKLDNN // refer to https://github.com/01org/mkl-dnn/blob/master/include/mkldnn.hpp // memory alignment CHECK_EQ(posix_memalign(&ptr, 4096ul, size), 0); diff --git a/paddle/math/CMakeLists.txt b/paddle/math/CMakeLists.txt index 86bb270a4372841b3e6f4676e222d2190549c153..922fb5172273da24f9c48786961a6d850b1ed7c5 100644 --- a/paddle/math/CMakeLists.txt +++ b/paddle/math/CMakeLists.txt @@ -26,8 +26,6 @@ else() endif() if(MOBILE_INFERENCE) - list(REMOVE_ITEM MATH_SOURCES - ${CMAKE_CURRENT_SOURCE_DIR}/SIMDFunctions.cpp) # Remove sparse list(REMOVE_ITEM MATH_HEADERS ${CMAKE_CURRENT_SOURCE_DIR}/CpuSparseMatrix.h diff --git a/paddle/math/CpuSparseMatrix.cpp b/paddle/math/CpuSparseMatrix.cpp index bf62229c03bb1d6e2bdf86d8c56a8157938fb832..dc6979cf5a5229fb09866189f28217889d58c2d0 100644 --- a/paddle/math/CpuSparseMatrix.cpp +++ b/paddle/math/CpuSparseMatrix.cpp @@ -260,6 +260,35 @@ void CpuSparseMatrix::printOneRow(std::ostream& os, size_t idx) const { os << ";"; } +void CpuSparseMatrix::rowScale(size_t cCol, CpuSparseMatrix& b, Matrix& c) { + CHECK(getFormat() != SPARSE_CSC) << "Not supported"; + CHECK_EQ(height_, b.getHeight()); + CHECK_EQ(width_, b.getWidth()); + real* A = getValue(); + real* B = b.getValue(); + if (b.getValueType() == FLOAT_VALUE) { + for (size_t i = 0; i < height_; i++) { + size_t start = getRowStartIdx(i); + size_t end = getRowStartIdx(i + 1); + CHECK_EQ(start, b.getRowStartIdx(i)); + CHECK_EQ(end, b.getRowStartIdx(i + 1)); + for (size_t j = start; j < end; j++) { + A[j] = B[j] * c.getElement(i, cCol); + } + } + } else if (b.getValueType() == NO_VALUE) { + for (size_t i = 0; i < height_; i++) { + size_t start = getRowStartIdx(i); + size_t end = getRowStartIdx(i + 1); + CHECK_EQ(start, b.getRowStartIdx(i)); + CHECK_EQ(end, b.getRowStartIdx(i + 1)); + for (size_t j = start; j < end; j++) { + A[j] = c.getElement(i, cCol); + } + } + } +} + void CpuSparseMatrix::randomizeUniform() { CHECK_LE(elementCnt_, height_ * width_); if (valueType_ == FLOAT_VALUE) { diff --git a/paddle/math/CpuSparseMatrix.h b/paddle/math/CpuSparseMatrix.h index aad1348353d558abca72ed0fa5cf943237e3ac78..522b436a2a69179d3f4f17c919d5ba024102db7b 100644 --- a/paddle/math/CpuSparseMatrix.h +++ b/paddle/math/CpuSparseMatrix.h @@ -239,6 +239,15 @@ public: const unsigned int* cols, const real* values); + /** + * @brief this_row = b_row * c_row[cCol] + * + * @param[in] cCol the column of matrix c used to scale each row of b + * @param[in] b CpuSparseMatrix + * @param[in] c Matrix + */ + void rowScale(size_t cCol, CpuSparseMatrix& b, Matrix& c); + void randomizeUniform(); void copyFrom(const GpuSparseMatrix& src, hl_stream_t stream); diff --git a/paddle/math/MathFunctions.cpp b/paddle/math/MathFunctions.cpp index ba86eacbb5d53ee43a60d2cd1dd922333a5d48f0..28ab54b450c96b4bdefdf36813595766162b1434 100644 --- a/paddle/math/MathFunctions.cpp +++ b/paddle/math/MathFunctions.cpp @@ -206,7 +206,7 @@ double dotProduct(const int n, const double* x, const double* y) { } #endif -#if defined(PADDLE_USE_MKLML) +#if defined(PADDLE_WITH_MKLML) template <> void vExp(const int n, const float* a, float* r) { diff --git a/paddle/math/MathFunctions.h b/paddle/math/MathFunctions.h index f6e77029bdd75a602f88b688ca810f47ba4ee615..29fe36e3a4bd5e5d372480950a03142822262d41 100644 --- a/paddle/math/MathFunctions.h +++ b/paddle/math/MathFunctions.h @@ -15,7 +15,7 @@ limitations under the License. */ #ifndef MATHFUNCTIONS_H_ #define MATHFUNCTIONS_H_ -#ifdef PADDLE_USE_MKLML +#ifdef PADDLE_WITH_MKLML #include #include #include diff --git a/paddle/math/Matrix.cpp b/paddle/math/Matrix.cpp index 88e9180690606c92cf46c5b295d80f14e5d64567..1ec4336cabbc7d3073b7638b7484bf61e83a2dc5 100644 --- a/paddle/math/Matrix.cpp +++ b/paddle/math/Matrix.cpp @@ -28,6 +28,7 @@ limitations under the License. */ #include "hl_top_k.h" #include "paddle/utils/Logging.h" +#include "NEONFunctions.h" #include "paddle/function/GemmFunctor.h" #include "paddle/utils/ThreadLocal.h" @@ -1130,7 +1131,8 @@ void GpuMatrix::avgPoolForward(Matrix& inputMat, size_t outputH, size_t outputW, size_t paddingH, - size_t paddingW) { + size_t paddingW, + bool excludeMode) { CHECK(inputMat.useGpu_ == true) << "Matrix type are not equal"; real* inputData = inputMat.getData(); @@ -1153,7 +1155,8 @@ void GpuMatrix::avgPoolForward(Matrix& inputMat, paddingH, paddingW, data_, - getStride()); + getStride(), + excludeMode); } void GpuMatrix::avgPoolBackward(Matrix& outGrad, @@ -1168,7 +1171,8 @@ void GpuMatrix::avgPoolBackward(Matrix& outGrad, real scaleTargets, real scaleOutput, size_t paddingH, - size_t paddingW) { + size_t paddingW, + bool excludeMode) { CHECK(outGrad.useGpu_ == true) << "Matrix type are not equal"; real* outDiff = outGrad.getData(); @@ -1194,7 +1198,8 @@ void GpuMatrix::avgPoolBackward(Matrix& outGrad, scaleTargets, scaleOutput, data_, - outGrad.getStride()); + outGrad.getStride(), + excludeMode); } void GpuMatrix::maxPool3DForward(Matrix& inputMat, @@ -2136,7 +2141,8 @@ void CpuMatrix::avgPoolForward(Matrix& input, size_t outputH, size_t outputW, size_t paddingH, - size_t paddingW) { + size_t paddingW, + bool excludeMode) { // The main loop size_t num = input.getHeight(); size_t inLength = imgSizeH * imgSizeW; @@ -2165,7 +2171,8 @@ void CpuMatrix::avgPoolForward(Matrix& input, tgtData[ph * outputW + pw] += inData[h * imgSizeW + w]; } } - int poolSize = (hend - hstart) * (wend - wstart); + int poolSize = + excludeMode ? (hend - hstart) * (wend - wstart) : sizeY * sizeX; CHECK(poolSize); tgtData[ph * outputW + pw] /= poolSize; } @@ -2189,7 +2196,8 @@ void CpuMatrix::avgPoolBackward(Matrix& input, real scaleTargets, real scaleOutput, size_t paddingH, - size_t paddingW) { + size_t paddingW, + bool excludeMode) { size_t num = input.getHeight(); size_t channels = input.getWidth() / outputH / outputW; size_t inLength = imgSizeH * imgSizeW; @@ -2211,7 +2219,8 @@ void CpuMatrix::avgPoolBackward(Matrix& input, int wstart = pw * strideW - paddingW; int wend = std::min(wstart + sizeX, imgSizeW); wstart = std::max(wstart, 0); - int poolSize = (hend - hstart) * (wend - wstart); + int poolSize = + excludeMode ? (hend - hstart) * (wend - wstart) : sizeY * sizeX; CHECK(poolSize); for (int h = hstart; h < hend; ++h) { @@ -4157,16 +4166,36 @@ void CpuMatrix::print(std::ostream& os) const { void CpuMatrix::paramReluForward(Matrix& data, Matrix& W) { real* input = data.getData(); real* w = W.getData(); + real* output = data_; size_t numElements = data.getWidth(); size_t numSamples = data.getHeight(); size_t paraSize = W.getHeight() * W.getWidth(); CHECK(!(numElements % paraSize)); // this check from ParameterReluLayer::init + size_t partial_sum = numElements / paraSize; + if (paraSize == numElements) { + for (size_t n = 0; n < numSamples * numElements; ++n) { + output[n] = input[n] > 0 ? input[n] : input[n] * w[n % numElements]; + } + return; + } + +#if defined(__ARM_NEON__) || defined(__ARM_NEON) + for (size_t n = 0; n < numSamples; ++n) { + for (size_t i = 0; i < paraSize; i++) { + neon::prelu( + input + i * partial_sum, w[i], output + i * partial_sum, partial_sum); + } + input = input + numElements; + output = output + numElements; + } +#else for (size_t n = 0, k = 0; n < numSamples; ++n) { for (size_t i = 0; i < numElements; ++i, ++k) { - data_[k] = input[k] > 0 ? input[k] : input[k] * w[i / partial_sum]; + output[k] = input[k] > 0 ? input[k] : input[k] * w[i / partial_sum]; } } +#endif } void CpuMatrix::paramReluBackwardW(Matrix& oGrad, Matrix& data) { diff --git a/paddle/math/Matrix.h b/paddle/math/Matrix.h index e273f1123690e31984c97185c5a8bc5e7b92c38c..c8e690e6421668bdade4e50a61882c915b2ddc7c 100644 --- a/paddle/math/Matrix.h +++ b/paddle/math/Matrix.h @@ -911,7 +911,8 @@ public: size_t outputH, size_t outputW, size_t paddingH, - size_t paddingW) { + size_t paddingW, + bool excludeMode = true) { LOG(FATAL) << "Not implemeted"; } @@ -927,9 +928,11 @@ public: real scaleTargets, real scaleOutput, size_t paddingH, - size_t paddingW) { + size_t paddingW, + bool excludeMode = true) { LOG(FATAL) << "Not implemeted"; } + /** * Pooling 3D forward operation, pick out the largest element * in the sizeX of value @@ -1458,7 +1461,8 @@ public: size_t outputH, size_t outputW, size_t paddingH, - size_t paddingW); + size_t paddingW, + bool excludeMode = true); void avgPoolBackward(Matrix& input, size_t imgSizeH, @@ -1472,7 +1476,8 @@ public: real scaleTargets, real scaleOutput, size_t paddingH, - size_t paddingW); + size_t paddingW, + bool excludeMode = true); void maxPool3DForward(Matrix& inputMat, Matrix& maxPoolIdx, @@ -1730,7 +1735,8 @@ public: size_t outputH, size_t outputW, size_t paddingH, - size_t paddingW); + size_t paddingW, + bool excludeMode = true); void avgPoolBackward(Matrix& input, size_t imgSizeH, @@ -1744,7 +1750,8 @@ public: real scaleTargets, real scaleOutput, size_t paddingH, - size_t paddingW); + size_t paddingW, + bool excludeMode = true); void maxPool3DForward(Matrix& inputMat, Matrix& maxPoolIdx, diff --git a/paddle/math/NEONFunctions.cpp b/paddle/math/NEONFunctions.cpp index 3bf47901f1069ac228fa1b877e29848d8cc130e8..0f8314942290a71dd327437b8a6da2d64fe48444 100644 --- a/paddle/math/NEONFunctions.cpp +++ b/paddle/math/NEONFunctions.cpp @@ -49,6 +49,46 @@ void relu(const float* a, float* b, int len) { } } +// b[i] = a[i] > 0.0f ? a[i] : a[i] * w +void prelu(const float* a, float w, float* b, int len) { + int offset = len % 16; + float32x4_t ma0, ma1, ma2, ma3; + + float32x4_t zero = vdupq_n_f32(0.f); + float32x4_t vw = vdupq_n_f32(w); + + for (int k = 0; k < len / 16; k++, a += 16, b += 16) { + ma0 = vld1q_f32(a); + ma1 = vld1q_f32(a + 4); + ma2 = vld1q_f32(a + 8); + ma3 = vld1q_f32(a + 12); + + uint32x4_t flag0 = vcgtq_f32(ma0, zero); + uint32x4_t flag1 = vcgtq_f32(ma1, zero); + uint32x4_t flag2 = vcgtq_f32(ma2, zero); + uint32x4_t flag3 = vcgtq_f32(ma3, zero); + + float32x4_t mul0 = vmulq_f32(ma0, vw); + float32x4_t mul1 = vmulq_f32(ma1, vw); + float32x4_t mul2 = vmulq_f32(ma2, vw); + float32x4_t mul3 = vmulq_f32(ma3, vw); + + ma0 = vbslq_f32(flag0, ma0, mul0); + ma1 = vbslq_f32(flag1, ma1, mul1); + ma2 = vbslq_f32(flag2, ma2, mul2); + ma3 = vbslq_f32(flag3, ma3, mul3); + + vst1q_f32(b, ma0); + vst1q_f32(b + 4, ma1); + vst1q_f32(b + 8, ma2); + vst1q_f32(b + 12, ma3); + } + + for (int i = 0; i < offset; i++) { + b[i] = a[i] > 0.0f ? a[i] : a[i] * w; + } +} + } // namespace neon } // namespace paddle diff --git a/paddle/math/NEONFunctions.h b/paddle/math/NEONFunctions.h index 69085e333547a31a341fbfde247f1e30adb957ee..d67b2f47a85a963949d23415e4f6881658203bb7 100644 --- a/paddle/math/NEONFunctions.h +++ b/paddle/math/NEONFunctions.h @@ -18,6 +18,7 @@ namespace paddle { namespace neon { void relu(const float* a, float* b, int len); +void prelu(const float* a, float w, float* b, int len); } // namespace neon } // namespace paddle diff --git a/paddle/math/SIMDFunctions.h b/paddle/math/SIMDFunctions.h index 439f11b79d134d7054f45f2d0a70fc5a6fde6c13..76909720f6aef0eea7cdf0dfe618237403d52c99 100644 --- a/paddle/math/SIMDFunctions.h +++ b/paddle/math/SIMDFunctions.h @@ -116,9 +116,11 @@ inline bool vec_check(size_t len) { } namespace internal { +#ifdef __SSE3__ void addToImpl(float* a, const float* b, size_t len); void batchAddToImpl(float* a, const float* b[], int batch, size_t len); void colMaxImpl(float* result, const float* data, int dim, int numSamples); +#endif #ifdef __AVX__ void decayL1AvxImpl(float* dst, float* src, float lambda, size_t len); void decayL1AvxImpl( diff --git a/paddle/math/float16.h b/paddle/math/float16.h new file mode 100644 index 0000000000000000000000000000000000000000..efebbce50405018c6b7ce2049f8d55c33680469f --- /dev/null +++ b/paddle/math/float16.h @@ -0,0 +1,739 @@ +/* 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 + +#include + +#ifdef PADDLE_WITH_CUDA +#include +#endif // PADDLE_WITH_CUDA + +#include "unsupported/Eigen/CXX11/Tensor" + +#include "paddle/platform/hostdevice.h" + +#ifdef __GNUC__ +#define PADDLE_GNUC_VER (__GNUC__ * 10 + __GNUC_MINOR__) +#else +#define PADDLE_GNUC_VER 0 +#endif // __GNUC__ + +#ifdef __clang__ +#define PADDLE_CLANG_VER (__clang_major__ * 10 + __clang_minor__) +#else +#define PADDLE_CLANG_VER 0 +#endif // __clang__ + +#if defined(__CUDACC__) && CUDA_VERSION >= 7050 +#define PADDLE_CUDA_FP16 +#include +#endif + +#if defined(__arm__) || defined(__aarch64__) +#define PADDLE_ARM +#endif + +#if defined(__ARM_NEON) || defined(__ARM_NEON__) +#define PADDLE_NEON +#include +#endif + +#if defined(PADDLE_NEON) && defined(PADDLE_ARM_FP16) && \ + (PADDLE_GNUC_VER >= 62 || PADDLE_CLANG_VER >= 37) +#define PADDLE_WITH_NATIVE_FP16 +#endif + +#ifndef PADDLE_ARM +#include +#endif // PADDLE_ARM + +#define PADDLE_ALIGN(x) __attribute__((aligned(x))) + +namespace paddle { + +// Use PADDLE_ALIGNED(2) to ensure that each float16 will be allocated +// and aligned at least on a 2-byte boundary, which leads to efficient +// memory access of float16 struct and also makes float16 compatible +// with CUDA half, ARM float16_t, and Eigen::half data types. +struct PADDLE_ALIGN(2) float16 { +public: + uint16_t x; + + // Constructors + HOSTDEVICE inline float16() : x(0) {} + + HOSTDEVICE inline float16(const float16& h) : x(h.x) {} + +#ifdef PADDLE_CUDA_FP16 + HOSTDEVICE inline explicit float16(const half& h) { +#if CUDA_VERSION >= 9000 + x = reinterpret_cast<__half_raw*>(const_cast(&h))->x; +#else + x = h.x; +#endif // CUDA_VERSION >= 9000 + } +#endif // PADDLE_CUDA_FP16 + + HOSTDEVICE inline explicit float16(const Eigen::half& h) : x(h.x) {} + +#ifdef PADDLE_WITH_NATIVE_FP16 + // __fp16 is a native half precision data type for arm cpu, + // float16_t is an alias for __fp16 + HOSTDEVICE inline explicit float16(const float16_t& h) { + x = *reinterpret_cast(&h); + } +#endif + + HOSTDEVICE inline explicit float16(float val) { +#if defined(PADDLE_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300 + half tmp = __float2half(val); + x = *reinterpret_cast(&tmp); + +#elif defined(PADDLE_WITH_NATIVE_FP16) + float32x4_t tmp = vld1q_dup_f32(&val); + float16_t res = vget_lane_f16(vcvt_f16_f32(tmp), 0); + x = *reinterpret_cast(&res); + +#elif defined(__F16C__) + x = _cvtss_sh(val, 0); + +#else + // Conversion routine adapted from + // http://stackoverflow.com/questions/1659440/32-bit-to-16-bit-floating-point-conversion + Bits v, s; + v.f = val; + uint32_t sign = v.si & sigN; + v.si ^= sign; + sign >>= shiftSign; // logical shift + s.si = mulN; + s.si = s.f * v.f; // correct subnormals + v.si ^= (s.si ^ v.si) & -(minN > v.si); + v.si ^= (infN ^ v.si) & -((infN > v.si) & (v.si > maxN)); + v.si ^= (nanN ^ v.si) & -((nanN > v.si) & (v.si > infN)); + v.ui >>= shift; // logical shift + v.si ^= ((v.si - maxD) ^ v.si) & -(v.si > maxC); + v.si ^= ((v.si - minD) ^ v.si) & -(v.si > subC); + x = v.ui | sign; + +#endif + } + + HOSTDEVICE inline explicit float16(bool b) : x(b ? 0x3c00 : 0) {} + + template + HOSTDEVICE inline explicit float16(const T& val) + : x(float16(static_cast(val)).x) {} + + HOSTDEVICE inline float16& operator=(const float16& rhs) { + x = rhs.x; + return *this; + } + +// Assignment operators +#ifdef PADDLE_CUDA_FP16 + HOSTDEVICE inline float16& operator=(const half& rhs) { +#if CUDA_VERSION >= 9000 + x = reinterpret_cast<__half_raw*>(const_cast(&rhs))->x; +#else + x = rhs.x; +#endif + return *this; + } +#endif + + HOSTDEVICE inline float16& operator=(const Eigen::half& rhs) { + x = rhs.x; + return *this; + } + +#ifdef PADDLE_WITH_NATIVE_FP16 + HOSTDEVICE inline float16& operator=(const float16_t& rhs) { + x = *reinterpret_cast(&rhs); + return *this; + } +#endif + + HOSTDEVICE inline float16& operator=(bool b) { + x = b ? 0x3c00 : 0; + return *this; + } + + HOSTDEVICE inline float16& operator=(int8_t val) { + x = float16(val).x; + return *this; + } + + HOSTDEVICE inline float16& operator=(uint8_t val) { + x = float16(val).x; + return *this; + } + + HOSTDEVICE inline float16& operator=(int16_t val) { + x = float16(val).x; + return *this; + } + + HOSTDEVICE inline float16& operator=(uint16_t val) { + x = float16(val).x; + return *this; + } + + HOSTDEVICE inline float16& operator=(int32_t val) { + x = float16(val).x; + return *this; + } + + HOSTDEVICE inline float16& operator=(uint32_t val) { + x = float16(val).x; + return *this; + } + + HOSTDEVICE inline float16& operator=(int64_t val) { + x = float16(val).x; + return *this; + } + + HOSTDEVICE inline float16& operator=(uint64_t val) { + x = float16(val).x; + return *this; + } + + HOSTDEVICE inline float16& operator=(float val) { + x = float16(val).x; + return *this; + } + + HOSTDEVICE inline float16& operator=(double val) { + x = float16(val).x; + return *this; + } + +// Conversion opertors +#ifdef PADDLE_CUDA_FP16 + HOSTDEVICE inline explicit operator half() const { +#if CUDA_VERSION >= 9000 + __half_raw h; + h.x = x; + return half(h); +#else + half h; + h.x = x; + return h; +#endif // CUDA_VERSION >= 9000 + } +#endif // PADDLE_CUDA_FP16 + + HOSTDEVICE inline explicit operator Eigen::half() const { + Eigen::half h; + h.x = x; + return h; + } + +#ifdef PADDLE_WITH_NATIVE_FP16 + HOSTDEVICE inline explicit operator float16_t() const { + return *reinterpret_cast(this); + } +#endif + + HOSTDEVICE inline explicit operator float() const { +#if defined(PADDLE_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300 + half tmp = *reinterpret_cast(this); + return __half2float(tmp); + +#elif defined(PADDLE_WITH_NATIVE_FP16) + float16x4_t res = vld1_dup_f16(reinterpret_cast(this)); + return vgetq_lane_f32(vcvt_f32_f16(res), 0); + +#elif defined(__F16C__) + return _cvtsh_ss(this->x); + +#else + // Conversion routine adapted from + // http://stackoverflow.com/questions/1659440/32-bit-to-16-bit-floating-point-conversion + Bits v; + v.ui = this->x; + int32_t sign = v.si & sigC; + v.si ^= sign; + sign <<= shiftSign; + v.si ^= ((v.si + minD) ^ v.si) & -(v.si > subC); + v.si ^= ((v.si + maxD) ^ v.si) & -(v.si > maxC); + Bits s; + s.si = mulC; + s.f *= v.si; + int32_t mask = -(norC > v.si); + v.si <<= shift; + v.si ^= (s.si ^ v.si) & mask; + v.si |= sign; + return v.f; + +#endif + } + + HOSTDEVICE inline explicit operator bool() const { return (x & 0x7fff) != 0; } + + HOSTDEVICE inline explicit operator int8_t() const { + return static_cast(float(*this)); + } + + HOSTDEVICE inline explicit operator uint8_t() const { + return static_cast(float(*this)); + } + + HOSTDEVICE inline explicit operator int16_t() const { + return static_cast(float(*this)); + } + + HOSTDEVICE inline explicit operator uint16_t() const { + return static_cast(float(*this)); + } + + HOSTDEVICE inline explicit operator int32_t() const { + return static_cast(float(*this)); + } + + HOSTDEVICE inline explicit operator uint32_t() const { + return static_cast(float(*this)); + } + + HOSTDEVICE inline explicit operator int64_t() const { + return static_cast(float(*this)); + } + + HOSTDEVICE inline explicit operator uint64_t() const { + return static_cast(float(*this)); + } + + HOSTDEVICE inline explicit operator double() const { + return static_cast(float(*this)); + } + +private: + union Bits { + float f; + int32_t si; + uint32_t ui; + }; + + static const int shift = 13; + static const int shiftSign = 16; + + static const int32_t infN = 0x7F800000; + static const int32_t maxN = 0x477FE000; // max flt16 as flt32 + static const int32_t minN = 0x38800000; // min flt16 normal as flt32 + static const int32_t sigN = 0x80000000; // sign bit + + static constexpr int32_t infC = infN >> shift; + static constexpr int32_t nanN = (infC + 1) + << shift; // minimum flt16 nan as float32 + static constexpr int32_t maxC = maxN >> shift; + static constexpr int32_t minC = minN >> shift; + static constexpr int32_t sigC = sigN >> shiftSign; + + static const int32_t mulN = 0x52000000; // (1 << 23) / minN + static const int32_t mulC = 0x33800000; // minN / (1 << (23 - shift)) + static const int32_t subC = 0x003FF; // max flt32 subnormal downshifted + static const int32_t norC = 0x00400; // min flt32 normal downshifted + + static constexpr int32_t maxD = infC - maxC - 1; + static constexpr int32_t minD = minC - subC - 1; +}; + +// Arithmetic operators on GPU +// CUDA 9.0 provides built-in arithmetic operators for half while +// CUDA 7.5 and 8.0 do not. The arithmetic operators defined here are +// for users to write similar CUDA code in CUDA 7.5 and 8.0 as in +// CUDA 9.0 regarding the half data type. +#if defined(PADDLE_CUDA_FP16) && CUDA_VERSION < 9000 + +DEVICE inline half operator+(const half& a, const half& b) { +#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530 + return __hadd(a, b); +#else + float res = float(float16(a)) + float(float16(b)); + return half(float16(res)); +#endif +} + +DEVICE inline half operator-(const half& a, const half& b) { +#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530 + return __hsub(a, b); +#else + float res = float(float16(a)) - float(float16(b)); + return half(float16(res)); +#endif +} + +DEVICE inline half operator*(const half& a, const half& b) { +#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530 + return __hmul(a, b); +#else + float res = float(float16(a)) * float(float16(b)); + return half(float16(res)); +#endif +} + +DEVICE inline half operator/(const half& a, const half& b) { +#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300 + float num = __half2float(a); + float denom = __half2float(b); + return __float2half(num / denom); +#else + float res = float(float16(a)) / float(float16(b)); + return half(float16(res)); +#endif +} + +DEVICE inline half operator-(const half& a) { +#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530 + return __hneg(a); +#else + float res = -float(float16(a)); + return half(float16(res)); +#endif +} + +DEVICE inline half& operator+=(half& a, const half& b) { + a = a + b; + return a; +} + +DEVICE inline half& operator-=(half& a, const half& b) { + a = a - b; + return a; +} + +DEVICE inline half& operator*=(half& a, const half& b) { + a = a * b; + return a; +} + +DEVICE inline half& operator/=(half& a, const half& b) { + a = a / b; + return a; +} + +DEVICE inline bool operator==(const half& a, const half& b) { +#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530 + return __heq(a, b); +#else + return float(float16(a)) == float(float16(b)); +#endif +} + +DEVICE inline bool operator!=(const half& a, const half& b) { +#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530 + return __hne(a, b); +#else + return float(float16(a)) != float(float16(b)); +#endif +} + +DEVICE inline bool operator<(const half& a, const half& b) { +#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530 + return __hlt(a, b); +#else + return float(float16(a)) < float(float16(b)); +#endif +} + +DEVICE inline bool operator<=(const half& a, const half& b) { +#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530 + return __hle(a, b); +#else + return float(float16(a)) <= float(float16(b)); +#endif +} + +DEVICE inline bool operator>(const half& a, const half& b) { +#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530 + return __hgt(a, b); +#else + return float(float16(a)) > float(float16(b)); +#endif +} + +DEVICE inline bool operator>=(const half& a, const half& b) { +#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530 + return __hge(a, b); +#else + return float(float16(a)) >= float(float16(b)); +#endif +} + +#endif // PADDLE_CUDA_FP16 + +// Arithmetic operators on ARMv8.2-A CPU +#if defined(PADDLE_WITH_NATIVE_FP16) +HOST inline float16 operator+(const float16& a, const float16& b) { + float16 res; + asm volatile( + "ld1 {v0.h}[0], [%[a_ptr]]\n" + "ld1 {v1.h}[0], [%[b_ptr]]\n" + "fadd h0, h0, h1\n" + "st1 {v0.h}[0], [%[res_ptr]]\n" + : // outputs + : // inputs + [a_ptr] "r"(&(a.x)), + [b_ptr] "r"(&(b.x)), + [res_ptr] "r"(&(res.x)) + : // clobbers + "memory", "v0", "v1"); + return res; +} + +HOST inline float16 operator-(const float16& a, const float16& b) { + float16 res; + asm volatile( + "ld1 {v0.h}[0], [%[a_ptr]]\n" + "ld1 {v1.h}[0], [%[b_ptr]]\n" + "fsub h0, h0, h1\n" + "st1 {v0.h}[0], [%[res_ptr]]\n" + : // outputs + : // inputs + [a_ptr] "r"(&(a.x)), + [b_ptr] "r"(&(b.x)), + [res_ptr] "r"(&(res.x)) + : // clobbers + "memory", "v0", "v1"); + return res; +} + +HOST inline float16 operator*(const float16& a, const float16& b) { + float16 res; + asm volatile( + "ld1 {v0.h}[0], [%[a_ptr]]\n" + "ld1 {v1.h}[0], [%[b_ptr]]\n" + "fmul h0, h0, h1\n" + "st1 {v0.h}[0], [%[res_ptr]]\n" + : // outputs + : // inputs + [a_ptr] "r"(&(a.x)), + [b_ptr] "r"(&(b.x)), + [res_ptr] "r"(&(res.x)) + : // clobbers + "memory", "v0", "v1"); + return res; +} + +HOST inline float16 operator/(const float16& a, const float16& b) { + float16 res; + asm volatile( + "ld1 {v0.h}[0], [%[a_ptr]]\n" + "ld1 {v1.h}[0], [%[b_ptr]]\n" + "fdiv h0, h0, h1\n" + "st1 {v0.h}[0], [%[res_ptr]]\n" + : // outputs + : // inputs + [a_ptr] "r"(&(a.x)), + [b_ptr] "r"(&(b.x)), + [res_ptr] "r"(&(res.x)) + : // clobbers + "memory", "v0", "v1"); + return res; +} + +HOST inline float16 operator-(const float16& a) { + float16 res; + asm volatile( + "ld1 {v0.h}[0], [%[a_ptr]]\n" + "fneg h0, h0\n" + "st1 {v0.h}[0], [%[res_ptr]]\n" + : // outputs + : // inputs + [a_ptr] "r"(&(a.x)), + [res_ptr] "r"(&(res.x)) + : // clobbers + "memory", "v0"); + return res; +} + +HOST inline float16& operator+=(float16& a, const float16& b) { + a = a + b; + return a; +} + +HOST inline float16& operator-=(float16& a, const float16& b) { + a = a - b; + return a; +} + +HOST inline float16& operator*=(float16& a, const float16& b) { + a = a * b; + return a; +} + +HOST inline float16& operator/=(float16& a, const float16& b) { + a = a / b; + return a; +} + +HOST inline bool operator==(const float16& a, const float16& b) { + uint16_t res; + asm volatile( + "ld1 {v0.h}[0], [%[a_ptr]]\n" + "ld1 {v1.h}[0], [%[b_ptr]]\n" + "fcmeq h0, h0, h1\n" + "st1 {v0.h}[0], [%[res_ptr]]\n" + : // outputs + : // inputs + [a_ptr] "r"(&(a.x)), + [b_ptr] "r"(&(b.x)), + [res_ptr] "r"(&res) + : // clobbers + "memory", "v0", "v1"); + return (res & 0xffff) != 0; +} + +HOST inline bool operator!=(const float16& a, const float16& b) { + return !(a == b); +} + +HOST inline bool operator<(const float16& a, const float16& b) { + uint16_t res; + asm volatile( + "ld1 {v1.h}[0], [%[a_ptr]]\n" + "ld1 {v0.h}[0], [%[b_ptr]]\n" + "fcmgt h0, h0, h1\n" + "st1 {v0.h}[0], [%[res_ptr]]\n" + : // outputs + : // inputs + [a_ptr] "r"(&(a.x)), + [b_ptr] "r"(&(b.x)), + [res_ptr] "r"(&res) + : // clobbers + "memory", "v0", "v1"); + return (res & 0xffff) != 0; +} + +HOST inline bool operator<=(const float16& a, const float16& b) { + uint16_t res; + asm volatile( + "ld1 {v1.h}[0], [%[a_ptr]]\n" + "ld1 {v0.h}[0], [%[b_ptr]]\n" + "fcmge h0, h0, h1\n" + "st1 {v0.h}[0], [%[res_ptr]]\n" + : // outputs + : // inputs + [a_ptr] "r"(&(a.x)), + [b_ptr] "r"(&(b.x)), + [res_ptr] "r"(&res) + : // clobbers + "memory", "v0", "v1"); + return (res & 0xffff) != 0; +} + +HOST inline bool operator>(const float16& a, const float16& b) { + uint16_t res; + asm volatile( + "ld1 {v0.h}[0], [%[a_ptr]]\n" + "ld1 {v1.h}[0], [%[b_ptr]]\n" + "fcmgt h0, h0, h1\n" + "st1 {v0.h}[0], [%[res_ptr]]\n" + : // outputs + : // inputs + [a_ptr] "r"(&(a.x)), + [b_ptr] "r"(&(b.x)), + [res_ptr] "r"(&res) + : // clobbers + "memory", "v0", "v1"); + return (res & 0xffff) != 0; +} + +HOST inline bool operator>=(const float16& a, const float16& b) { + uint16_t res; + asm volatile( + "ld1 {v0.h}[0], [%[a_ptr]]\n" + "ld1 {v1.h}[0], [%[b_ptr]]\n" + "fcmge h0, h0, h1\n" + "st1 {v0.h}[0], [%[res_ptr]]\n" + : // outputs + : // inputs + [a_ptr] "r"(&(a.x)), + [b_ptr] "r"(&(b.x)), + [res_ptr] "r"(&res) + : // clobbers + "memory", "v0", "v1"); + return (res & 0xffff) != 0; +} + +// Arithmetic operators, software emulated on other CPU +#else +HOSTDEVICE inline float16 operator+(const float16& a, const float16& b) { + return float16(float(a) + float(b)); +} + +HOSTDEVICE inline float16 operator-(const float16& a, const float16& b) { + return float16(float(a) - float(b)); +} + +HOSTDEVICE inline float16 operator*(const float16& a, const float16& b) { + return float16(float(a) * float(b)); +} + +HOSTDEVICE inline float16 operator/(const float16& a, const float16& b) { + return float16(float(a) / float(b)); +} + +HOSTDEVICE inline float16 operator-(const float16& a) { + float16 res; + res.x = a.x ^ 0x8000; + return res; +} + +HOSTDEVICE inline float16& operator+=(float16& a, const float16& b) { + a = float16(float(a) + float(b)); + return a; +} + +HOSTDEVICE inline float16& operator-=(float16& a, const float16& b) { + a = float16(float(a) - float(b)); + return a; +} + +HOSTDEVICE inline float16& operator*=(float16& a, const float16& b) { + a = float16(float(a) * float(b)); + return a; +} + +HOSTDEVICE inline float16& operator/=(float16& a, const float16& b) { + a = float16(float(a) / float(b)); + return a; +} + +HOSTDEVICE inline bool operator==(const float16& a, const float16& b) { + return float(a) == float(b); +} + +HOSTDEVICE inline bool operator!=(const float16& a, const float16& b) { + return float(a) != float(b); +} + +HOSTDEVICE inline bool operator<(const float16& a, const float16& b) { + return float(a) < float(b); +} + +HOSTDEVICE inline bool operator<=(const float16& a, const float16& b) { + return float(a) <= float(b); +} + +HOSTDEVICE inline bool operator>(const float16& a, const float16& b) { + return float(a) > float(b); +} + +HOSTDEVICE inline bool operator>=(const float16& a, const float16& b) { + return float(a) >= float(b); +} +#endif +} // namespace paddle diff --git a/paddle/math/tests/CMakeLists.txt b/paddle/math/tests/CMakeLists.txt index d8b7f9e3fc74040189ade83049e4a1c3348e08de..dcd2a34583417993a4bf2976f7a3bc5a10d496ac 100644 --- a/paddle/math/tests/CMakeLists.txt +++ b/paddle/math/tests/CMakeLists.txt @@ -22,6 +22,7 @@ if(WITH_GPU) link_paddle_test(test_Tensor) CUDA_ADD_EXECUTABLE(test_lazyAssign test_lazyAssign.cu) link_paddle_test(test_lazyAssign) + nv_test(test_float16_gpu SRCS test_float16.cu) else() compile_cu_as_cpp(test_Tensor.cu) add_unittest(test_Tensor test_Tensor.cu) @@ -33,3 +34,4 @@ add_simple_unittest(test_FPException) add_simple_unittest(test_GpuProfiler) add_simple_unittest(test_BaseMatrix) add_simple_unittest(test_Matrix) +add_simple_unittest(test_float16) diff --git a/paddle/math/tests/test_float16.cpp b/paddle/math/tests/test_float16.cpp new file mode 100644 index 0000000000000000000000000000000000000000..74cc55aa3792f5e9f86b4f56f28dad97f35996a0 --- /dev/null +++ b/paddle/math/tests/test_float16.cpp @@ -0,0 +1,119 @@ +/* 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. */ + +#include "paddle/math/float16.h" + +#include + +namespace paddle { + +TEST(float16, conversion_cpu) { + // Explicit conversion from Eigen::half + EXPECT_EQ(float16(Eigen::half(1.0f)).x, 0x3c00); + EXPECT_EQ(float16(Eigen::half(0.5f)).x, 0x3800); + EXPECT_EQ(float16(Eigen::half(0.33333f)).x, 0x3555); + EXPECT_EQ(float16(Eigen::half(0.0f)).x, 0x0000); + EXPECT_EQ(float16(Eigen::half(-0.0f)).x, 0x8000); + EXPECT_EQ(float16(Eigen::half(65504.0f)).x, 0x7bff); + EXPECT_EQ(float16(Eigen::half(65536.0f)).x, 0x7c00); + + // Conversion from float + EXPECT_EQ(float16(1.0f).x, 0x3c00); + EXPECT_EQ(float16(0.5f).x, 0x3800); + EXPECT_EQ(float16(0.33333f).x, 0x3555); + EXPECT_EQ(float16(0.0f).x, 0x0000); + EXPECT_EQ(float16(-0.0f).x, 0x8000); + EXPECT_EQ(float16(65504.0f).x, 0x7bff); + EXPECT_EQ(float16(65536.0f).x, 0x7c00); + + // Conversion from double + EXPECT_EQ(float16(1.0).x, 0x3c00); + EXPECT_EQ(float16(0.5).x, 0x3800); + EXPECT_EQ(float16(0.33333).x, 0x3555); + EXPECT_EQ(float16(0.0).x, 0x0000); + EXPECT_EQ(float16(-0.0).x, 0x8000); + EXPECT_EQ(float16(65504.0).x, 0x7bff); + EXPECT_EQ(float16(65536.0).x, 0x7c00); + + // Conversion from int + EXPECT_EQ(float16(-1).x, 0xbc00); + EXPECT_EQ(float16(0).x, 0x0000); + EXPECT_EQ(float16(1).x, 0x3c00); + EXPECT_EQ(float16(2).x, 0x4000); + EXPECT_EQ(float16(3).x, 0x4200); + + // Conversion from bool + EXPECT_EQ(float16(true).x, 0x3c00); + EXPECT_EQ(float16(false).x, 0x0000); + + // Default constructor + float16 v_def; + EXPECT_EQ(v_def.x, 0x0000); + + // Assignment operator + float16 v_assign; + v_assign = v_def; + EXPECT_EQ(v_assign.x, 0x0000); + v_assign = Eigen::half(1.0f); + EXPECT_EQ(v_assign.x, 0x3c00); + v_assign = 0.5f; + EXPECT_EQ(v_assign.x, 0x3800); + v_assign = 0.33333; + EXPECT_EQ(v_assign.x, 0x3555); + v_assign = -1; + EXPECT_EQ(v_assign.x, 0xbc00); + v_assign = true; + EXPECT_EQ(v_assign.x, 0x3c00); + + // Conversion operator + EXPECT_EQ(Eigen::half(float16(1.0f)).x, 0x3c00); + EXPECT_EQ(float(float16(0.5f)), 0.5f); + EXPECT_NEAR(double(float16(0.33333)), 0.33333, 0.0001); + EXPECT_EQ(int(float16(-1)), -1); + EXPECT_EQ(bool(float16(true)), true); +} + +TEST(float16, arithmetic_cpu) { + EXPECT_EQ(float(float16(1) + float16(1)), 2); + EXPECT_EQ(float(float16(5) + float16(-5)), 0); + EXPECT_NEAR(float(float16(0.33333f) + float16(0.66667f)), 1.0f, 0.001); + EXPECT_EQ(float(float16(3) - float16(5)), -2); + EXPECT_NEAR(float(float16(0.66667f) - float16(0.33333f)), 0.33334f, 0.001); + EXPECT_NEAR(float(float16(3.3f) * float16(2.0f)), 6.6f, 0.01); + EXPECT_NEAR(float(float16(-2.1f) * float16(-3.0f)), 6.3f, 0.01); + EXPECT_NEAR(float(float16(2.0f) / float16(3.0f)), 0.66667f, 0.001); + EXPECT_EQ(float(float16(1.0f) / float16(2.0f)), 0.5f); + EXPECT_EQ(float(-float16(512.0f)), -512.0f); + EXPECT_EQ(float(-float16(-512.0f)), 512.0f); +} + +TEST(float16, comparison_cpu) { + EXPECT_TRUE(float16(1.0f) == float16(1.0f)); + EXPECT_FALSE(float16(-1.0f) == float16(-0.5f)); + EXPECT_TRUE(float16(1.0f) != float16(0.5f)); + EXPECT_FALSE(float16(-1.0f) != float16(-1.0f)); + EXPECT_TRUE(float16(1.0f) < float16(2.0f)); + EXPECT_FALSE(float16(-1.0f) < float16(-1.0f)); + EXPECT_TRUE(float16(1.0f) <= float16(1.0f)); + EXPECT_TRUE(float16(2.0f) > float16(1.0f)); + EXPECT_FALSE(float16(-2.0f) > float16(-2.0f)); + EXPECT_TRUE(float16(2.0f) >= float16(2.0f)); + + EXPECT_TRUE(float16(0.0f) == float16(-0.0f)); + EXPECT_TRUE(float16(0.0f) <= float16(-0.0f)); + EXPECT_TRUE(float16(0.0f) >= float16(-0.0f)); + EXPECT_FALSE(float16(0.0f) < float16(-0.0f)); + EXPECT_FALSE(float16(-0.0f) < float16(0.0f)); + EXPECT_FALSE(float16(0.0f) > float16(-0.0f)); + EXPECT_FALSE(float16(-0.0f) > float16(0.0f)); +} + +} // namespace paddle diff --git a/paddle/math/tests/test_float16.cu b/paddle/math/tests/test_float16.cu new file mode 100644 index 0000000000000000000000000000000000000000..4b520feaaf552302a969d8caee8aa28cc143304b --- /dev/null +++ b/paddle/math/tests/test_float16.cu @@ -0,0 +1,213 @@ +/* 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. */ + +#include "paddle/math/float16.h" + +#include + +#include "paddle/utils/Logging.h" + +#define ARITHMETIC_KERNEL(op_type, sign) \ + __global__ void op_type(const half* in1, const half* in2, half* out) { \ + out[0] = in1[0] sign in2[0]; \ + } + +#define COMPOUND_KERNEL(op_type, sign) \ + __global__ void op_type(half* in1, const half* in2) { in1[0] sign in2[0]; } + +#define COMPARISON_KERNEL(op_type, sign) \ + __global__ void op_type(const half* in1, const half* in2, bool* out) { \ + out[0] = in1[0] sign in2[0]; \ + } + +#define ARITHMETIC_KERNEL_LAUNCH(op_type) \ + void Test##op_type(float v_in1, float v_in2, float v_out) { \ + LOG(INFO) << "Test " << #op_type << " on GPU!"; \ + half *in1, *in2, *out; \ + half *d_in1, *d_in2, *d_out; \ + int size = sizeof(half); \ + cudaMalloc((void**)&d_in1, size); \ + cudaMalloc((void**)&d_in2, size); \ + cudaMalloc((void**)&d_out, size); \ + in1 = (half*)malloc(size); \ + in2 = (half*)malloc(size); \ + out = (half*)malloc(size); \ + in1[0] = half(float16(v_in1)); \ + in2[0] = half(float16(v_in2)); \ + cudaMemcpy(d_in1, in1, size, cudaMemcpyHostToDevice); \ + cudaMemcpy(d_in2, in2, size, cudaMemcpyHostToDevice); \ + op_type<<<1, 1>>>(d_in1, d_in2, d_out); \ + cudaMemcpy(out, d_out, size, cudaMemcpyDeviceToHost); \ + EXPECT_EQ(float(float16(out[0])), v_out); \ + free(in1); \ + free(in2); \ + free(out); \ + cudaFree(d_in1); \ + cudaFree(d_in2); \ + cudaFree(d_out); \ + } + +#define COMPOUND_KERNEL_LAUNCH(op_type) \ + void Test##op_type(float v_in1, float v_in2, float v_out) { \ + LOG(INFO) << "Test " << #op_type << " on GPU!"; \ + half *in1, *in2; \ + half *d_in1, *d_in2; \ + int size = sizeof(half); \ + cudaMalloc((void**)&d_in1, size); \ + cudaMalloc((void**)&d_in2, size); \ + in1 = (half*)malloc(size); \ + in2 = (half*)malloc(size); \ + in1[0] = half(float16(v_in1)); \ + in2[0] = half(float16(v_in2)); \ + cudaMemcpy(d_in1, in1, size, cudaMemcpyHostToDevice); \ + cudaMemcpy(d_in2, in2, size, cudaMemcpyHostToDevice); \ + op_type<<<1, 1>>>(d_in1, d_in2); \ + cudaMemcpy(in1, d_in1, size, cudaMemcpyDeviceToHost); \ + EXPECT_EQ(float(float16(in1[0])), v_out); \ + free(in1); \ + free(in2); \ + cudaFree(d_in1); \ + cudaFree(d_in2); \ + } + +#define COMPARISON_KERNEL_LAUNCH(op_type) \ + void Test##op_type(float v_in1, float v_in2, bool v_out) { \ + LOG(INFO) << "Test " << #op_type << " on GPU!"; \ + half *in1, *in2; \ + half *d_in1, *d_in2; \ + bool *out, *d_out; \ + int size = sizeof(half); \ + cudaMalloc((void**)&d_in1, size); \ + cudaMalloc((void**)&d_in2, size); \ + cudaMalloc((void**)&d_out, 1); \ + in1 = (half*)malloc(size); \ + in2 = (half*)malloc(size); \ + out = (bool*)malloc(1); \ + in1[0] = half(float16(v_in1)); \ + in2[0] = half(float16(v_in2)); \ + cudaMemcpy(d_in1, in1, size, cudaMemcpyHostToDevice); \ + cudaMemcpy(d_in2, in2, size, cudaMemcpyHostToDevice); \ + op_type<<<1, 1>>>(d_in1, d_in2, d_out); \ + cudaMemcpy(out, d_out, 1, cudaMemcpyDeviceToHost); \ + EXPECT_EQ(out[0], v_out); \ + free(in1); \ + free(in2); \ + free(out); \ + cudaFree(d_in1); \ + cudaFree(d_in2); \ + cudaFree(d_out); \ + } + +#ifdef PADDLE_CUDA_FP16 +namespace paddle { + +#if CUDA_VERSION < 9000 +ARITHMETIC_KERNEL(Add, +) +ARITHMETIC_KERNEL(Sub, -) +ARITHMETIC_KERNEL(Mul, *) +ARITHMETIC_KERNEL(Div, /) + +ARITHMETIC_KERNEL_LAUNCH(Add) +ARITHMETIC_KERNEL_LAUNCH(Sub) +ARITHMETIC_KERNEL_LAUNCH(Mul) +ARITHMETIC_KERNEL_LAUNCH(Div) + +// Negative sign kernel +__global__ void Neg(half* in) { in[0] = -in[0]; } + +void TestNeg(float v_in, float v_out) { + LOG(INFO) << "Test Neg on GPU!"; + half *in, *d_in; + int size = sizeof(half); + cudaMalloc((void**)&d_in, size); + in = (half*)malloc(size); + in[0] = half(float16(v_in)); + cudaMemcpy(d_in, in, size, cudaMemcpyHostToDevice); + Neg<<<1, 1>>>(d_in); + cudaMemcpy(in, d_in, size, cudaMemcpyDeviceToHost); + EXPECT_EQ(float(float16(in[0])), v_out); + free(in); + cudaFree(d_in); +} + +COMPOUND_KERNEL(AddAssign, +=) +COMPOUND_KERNEL(SubAssign, -=) +COMPOUND_KERNEL(MulAssign, *=) +COMPOUND_KERNEL(DivAssign, /=) + +COMPOUND_KERNEL_LAUNCH(AddAssign) +COMPOUND_KERNEL_LAUNCH(SubAssign) +COMPOUND_KERNEL_LAUNCH(MulAssign) +COMPOUND_KERNEL_LAUNCH(DivAssign) + +COMPARISON_KERNEL(Equal, ==) +COMPARISON_KERNEL(NotEqual, !=) +COMPARISON_KERNEL(Less, <) +COMPARISON_KERNEL(LessEqual, <=) +COMPARISON_KERNEL(Greater, >) +COMPARISON_KERNEL(GreaterEqual, >=) + +COMPARISON_KERNEL_LAUNCH(Equal) +COMPARISON_KERNEL_LAUNCH(NotEqual) +COMPARISON_KERNEL_LAUNCH(Less) +COMPARISON_KERNEL_LAUNCH(LessEqual) +COMPARISON_KERNEL_LAUNCH(Greater) +COMPARISON_KERNEL_LAUNCH(GreaterEqual) + +TEST(float16, arithmetic_on_gpu) { + TestAdd(1, 2, 3); + TestSub(2, 1, 1); + TestMul(2, 3, 6); + TestDiv(6, 2, 3); + TestNeg(1, -1); +} + +TEST(float16, compound_on_gpu) { + TestAddAssign(1, 2, 3); + TestSubAssign(2, 1, 1); + TestMulAssign(2, 3, 6); + TestDivAssign(6, 2, 3); +} + +TEST(float16, comparision_on_gpu) { + TestEqual(1, 1, true); + TestEqual(1, 2, false); + TestNotEqual(2, 3, true); + TestNotEqual(2, 2, false); + TestLess(3, 4, true); + TestLess(3, 3, false); + TestLessEqual(3, 3, true); + TestLessEqual(3, 2, false); + TestGreater(4, 3, true); + TestGreater(4, 4, false); + TestGreaterEqual(4, 4, true); + TestGreaterEqual(4, 5, false); +} +#endif // CUDA_VERSION + +TEST(float16, conversion_on_gpu) { + // Explicit conversion to and from cuda half + EXPECT_EQ(float16(half(float16(1.0f))).x, 0x3c00); + EXPECT_EQ(float16(half(float16(0.5f))).x, 0x3800); + EXPECT_EQ(float16(half(float16(0.33333f))).x, 0x3555); + EXPECT_EQ(float16(half(float16(0.0f))).x, 0x0000); + EXPECT_EQ(float16(half(float16(-0.0f))).x, 0x8000); + EXPECT_EQ(float16(half(float16(65504.0f))).x, 0x7bff); + EXPECT_EQ(float16(half(float16(65536.0f))).x, 0x7c00); + + // Assignment operator + float16 v_assign; + v_assign = half(float16(1.0f)); + EXPECT_EQ(v_assign.x, 0x3c00); +} + +} // namespace paddle +#endif // PADDLE_CUDA_FP16 diff --git a/paddle/math/tests/test_matrixCompare.cpp b/paddle/math/tests/test_matrixCompare.cpp index 7e5a1db44a5302e3b4e5d2768755824666e880ba..afb8d9d599b15a0b6d19b7ecca5e91b623695dea 100644 --- a/paddle/math/tests/test_matrixCompare.cpp +++ b/paddle/math/tests/test_matrixCompare.cpp @@ -244,7 +244,7 @@ TEST(Matrix, unary) { LOG(WARNING) << "This version of PaddlePaddle was not built with LAPACK" << "support so we cannot test matrix inverse. To test " << "matrix inverse, please install LAPACKE " - << "and MKL/Openblas/ATLAS, and re-build PaddlePaddle."; + << "and MKL/Openblas, and re-build PaddlePaddle."; #endif } } diff --git a/paddle/memory/CMakeLists.txt b/paddle/memory/CMakeLists.txt index aed5275dbf9be707cc6e19e729133ba8eab58195..8841c14ee083fccfd2271efd0c331805919a09d9 100644 --- a/paddle/memory/CMakeLists.txt +++ b/paddle/memory/CMakeLists.txt @@ -1,6 +1,6 @@ add_subdirectory(detail) -cc_library(memory SRCS memory.cc DEPS place) +cc_library(memory SRCS memory.cc DEPS place enforce) cc_library(memcpy SRCS memcpy.cc) cc_library(paddle_memory diff --git a/paddle/memory/README.md b/paddle/memory/README.md index 6cb003c50bc7d142d65b0591e7e5235431d2ea42..7cf61d089b39041b7a15184e0ea9211d14a66f5e 100644 --- a/paddle/memory/README.md +++ b/paddle/memory/README.md @@ -12,13 +12,13 @@ p = memory::Alloc(platform::CPUPlace(), 4*1024); To allocate 4KB memory on the 3rd GPU: ```cpp -p = memory::Alloc(platform::GPUPlace(2), 4*1024); +p = memory::Alloc(platform::CUDAPlace(2), 4*1024); ``` To free memory and check the so-far used amount of memory on a place: ```cpp -auto pl = platform::GPUPlace(0); +auto pl = platform::CUDAPlace(0); p = memory::Alloc(pl, 4*1024); cout << memory::Used(pl); memory::Free(pl, p); @@ -36,7 +36,7 @@ template size_t Used(Place); } // namespace memory ``` -These function templates have specializations on either `platform::CPUPlace` or `platform::GPUPlace`: +These function templates have specializations on either `platform::CPUPlace` or `platform::CUDAPlace`: ```cpp template<> @@ -49,7 +49,7 @@ and ```cpp template<> -void Alloc(GPUPlace p, size_t size) { +void Alloc(CUDAPlace p, size_t size) { return GetGPUBuddyAllocator(p.id)->Alloc(size); } ``` @@ -122,7 +122,7 @@ There are two implementations of `Context`: 1. [`CPUContext`](https://github.com/caffe2/caffe2/blob/v0.7.0/caffe2/core/context.h#L105), whose [`New` method](https://github.com/caffe2/caffe2/blob/v0.7.0/caffe2/core/context.h#L131) calls [`g_cpu_allocator.get()->New(size_t)`](https://github.com/caffe2/caffe2/blob/v0.7.0/caffe2/core/context.cc#L15) to allocate the memory. -1. [`CUDAContext`](https://github.com/caffe2/caffe2/blob/v0.7.0/caffe2/core/context_gpu.h#L99), which has a data member [`int gpu_id_`](https://github.com/caffe2/caffe2/blob/v0.7.0/caffe2/core/context_gpu.h#L202). This looks very similar to class `majel::GPUPlace`, who also has an `int id_` data member. `CUDAContext::New(size_t)` calls [`g_cub_allocator->DeviceAllocate(&ptr, nbytes)`](https://github.com/caffe2/caffe2/blob/v0.7.0/caffe2/core/context_gpu.cu#L355) to allocate the memory. +1. [`CUDAContext`](https://github.com/caffe2/caffe2/blob/v0.7.0/caffe2/core/context_gpu.h#L99), which has a data member [`int gpu_id_`](https://github.com/caffe2/caffe2/blob/v0.7.0/caffe2/core/context_gpu.h#L202). This looks very similar to class `majel::CUDAPlace`, who also has an `int id_` data member. `CUDAContext::New(size_t)` calls [`g_cub_allocator->DeviceAllocate(&ptr, nbytes)`](https://github.com/caffe2/caffe2/blob/v0.7.0/caffe2/core/context_gpu.cu#L355) to allocate the memory. ### Majel diff --git a/paddle/memory/detail/buddy_allocator.cc b/paddle/memory/detail/buddy_allocator.cc index 64ee53803891f192302bb915027f0499dfa36411..2bc2c06a15702b29c8bdf755978aebe0e6219b4a 100644 --- a/paddle/memory/detail/buddy_allocator.cc +++ b/paddle/memory/detail/buddy_allocator.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/memory/detail/buddy_allocator.h" #include "glog/logging.h" diff --git a/paddle/memory/detail/buddy_allocator.h b/paddle/memory/detail/buddy_allocator.h index 9c41378483993101a098fc4ad1068c1ef908e566..4e0135dd655d04b7f99722a2159795738c1b29c7 100644 --- a/paddle/memory/detail/buddy_allocator.h +++ b/paddle/memory/detail/buddy_allocator.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 diff --git a/paddle/memory/detail/memory_block.cc b/paddle/memory/detail/memory_block.cc index fc40993208323f1f5d18103165c8835b5f829613..f50eceba096477d7b2f50f7c406770c8e9595332 100644 --- a/paddle/memory/detail/memory_block.cc +++ b/paddle/memory/detail/memory_block.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/memory/detail/memory_block.h" #include "paddle/memory/detail/meta_cache.h" diff --git a/paddle/memory/detail/memory_block.h b/paddle/memory/detail/memory_block.h index a5168b519f3a3747f34ef2ea7b87d72dce70064d..a4ca51b31b0df269b4b48f824bfd2b74f2a059fb 100644 --- a/paddle/memory/detail/memory_block.h +++ b/paddle/memory/detail/memory_block.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 diff --git a/paddle/memory/detail/meta_cache.cc b/paddle/memory/detail/meta_cache.cc index 7e2f92b00ca5d787c1114176c5dc3304ca3ebe26..2bacca75108f9f80e7aa291fcb4fd66112201394 100644 --- a/paddle/memory/detail/meta_cache.cc +++ b/paddle/memory/detail/meta_cache.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/memory/detail/meta_cache.h" #include "glog/logging.h" diff --git a/paddle/memory/detail/meta_cache.h b/paddle/memory/detail/meta_cache.h index cf5815644284c23a1d2abc904f8c5053ce107a72..db8ffd49ae30cf72ca691894af2df08a7106d02f 100644 --- a/paddle/memory/detail/meta_cache.h +++ b/paddle/memory/detail/meta_cache.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 diff --git a/paddle/memory/detail/meta_data.cc b/paddle/memory/detail/meta_data.cc index 70c5c1f439e84ec33cf0507beae33f9cdfa51727..dc57d4d2376ab1cfe6fb49c92af4591b3972a53a 100644 --- a/paddle/memory/detail/meta_data.cc +++ b/paddle/memory/detail/meta_data.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/memory/detail/meta_data.h" diff --git a/paddle/memory/detail/meta_data.h b/paddle/memory/detail/meta_data.h index 628cf1f2e347e288d1bf34c14c7b2f13a28d3662..6b83c42eb851f0487bed6d625d848cf90db00929 100644 --- a/paddle/memory/detail/meta_data.h +++ b/paddle/memory/detail/meta_data.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 diff --git a/paddle/memory/detail/system_allocator.cc b/paddle/memory/detail/system_allocator.cc index 6b4e46f56a0c9c9836c5b353ec9c554454ab0491..509250debc2b2fd2e87078ab5f233ae2db6fd898 100644 --- a/paddle/memory/detail/system_allocator.cc +++ b/paddle/memory/detail/system_allocator.cc @@ -19,6 +19,7 @@ limitations under the License. */ #include // for malloc and free #include // for mlock and munlock +#include // for std::max #include "gflags/gflags.h" @@ -28,7 +29,7 @@ limitations under the License. */ // of memory available to the system for paging. So, by default, we // should set false to use_pinned_memory. DEFINE_bool(use_pinned_memory, true, "If set, allocate cpu pinned memory."); - +DECLARE_double(fraction_of_gpu_memory_to_use); namespace paddle { namespace memory { namespace detail { @@ -43,7 +44,7 @@ void* CPUAllocator::Alloc(size_t& index, size_t size) { void* p; -#ifdef PADDLE_USE_MKLDNN +#ifdef PADDLE_WITH_MKLDNN // refer to https://github.com/01org/mkl-dnn/blob/master/include/mkldnn.hpp // memory alignment PADDLE_ENFORCE_EQ(posix_memalign(&p, 4096ul, size), 0); @@ -77,45 +78,20 @@ void* GPUAllocator::Alloc(size_t& index, size_t size) { // CUDA documentation doesn't explain if cudaMalloc returns nullptr // if size is 0. We just make sure it does. if (size <= 0) return nullptr; - - size_t available = 0; - size_t capacity = 0; - paddle::platform::GpuMemoryUsage(available, capacity); - - // Reserve memory for page tables, etc. - size_t reserving = capacity - paddle::platform::GpuMaxAllocSize(); - size_t usable = available > reserving ? available - reserving : 0; - - // If remaining size no less than expected size, using general - // cudaMalloc to allocate GPU memory. - void* p = 0; - if (size <= usable) { - cudaError_t result = cudaMalloc(&p, size); - if (result == cudaSuccess) { - index = 0; - gpu_alloc_size_ += size; - return p; - } - } - - // If remaining size less than expected size or cudaMalloc failed, - // cudaMallocHost will be considered as a fallback allocator. - // - // NOTE: here, we use GpuMaxAllocSize() as the maximum memory size - // of host fallback allocation. Allocates too much would reduce - // the amount of memory available to the underlying system for paging. - usable = paddle::platform::GpuMaxAllocSize() - fallback_alloc_size_; - - if (size > usable) return nullptr; - - cudaError_t result = cudaMallocHost(&p, size); + void* p; + cudaError_t result = cudaMalloc(&p, size); if (result == cudaSuccess) { - index = 1; - fallback_alloc_size_ += size; + index = 0; + gpu_alloc_size_ += size; return p; + } else { + LOG(WARNING) + << "Cannot malloc " << size / 1024.0 / 1024.0 + << " MB GPU memory. Please shrink FLAGS_fraction_of_gpu_memory_to_use " + "environment variable to a lower value. Current value is " + << FLAGS_fraction_of_gpu_memory_to_use; + return nullptr; } - - return nullptr; } void GPUAllocator::Free(void* p, size_t size, size_t index) { diff --git a/paddle/memory/memcpy.cc b/paddle/memory/memcpy.cc index 1df88a6da9fb0c50d0d7ecd083c0533d8a886a67..b46141aafd7146bd3def12d86108c10f1f143d20 100644 --- a/paddle/memory/memcpy.cc +++ b/paddle/memory/memcpy.cc @@ -28,31 +28,25 @@ void Copy(platform::CPUPlace, void* dst, #ifdef PADDLE_WITH_CUDA template <> -void Copy(platform::CPUPlace dst_place, - void* dst, - platform::GPUPlace src_place, - const void* src, size_t num, - cudaStream_t stream) { +void Copy( + platform::CPUPlace dst_place, void* dst, platform::CUDAPlace src_place, + const void* src, size_t num, cudaStream_t stream) { platform::SetDeviceId(src_place.device); platform::GpuMemcpyAsync(dst, src, num, cudaMemcpyDeviceToHost, stream); } template <> -void Copy(platform::GPUPlace dst_place, - void* dst, - platform::CPUPlace src_place, - const void* src, size_t num, - cudaStream_t stream) { +void Copy( + platform::CUDAPlace dst_place, void* dst, platform::CPUPlace src_place, + const void* src, size_t num, cudaStream_t stream) { platform::SetDeviceId(dst_place.device); platform::GpuMemcpyAsync(dst, src, num, cudaMemcpyHostToDevice, stream); } template <> -void Copy(platform::GPUPlace dst_place, - void* dst, - platform::GPUPlace src_place, - const void* src, size_t num, - cudaStream_t stream) { +void Copy( + platform::CUDAPlace dst_place, void* dst, platform::CUDAPlace src_place, + const void* src, size_t num, cudaStream_t stream) { if (dst_place == src_place) { platform::SetDeviceId(src_place.device); platform::GpuMemcpyAsync(dst, src, num, cudaMemcpyDeviceToDevice, stream); @@ -62,33 +56,6 @@ void Copy(platform::GPUPlace dst_place, } } -template <> -void Copy(platform::CPUPlace dst_place, - void* dst, - platform::GPUPlace src_place, - const void* src, size_t num) { - platform::SetDeviceId(src_place.device); - platform::GpuMemcpySync(dst, src, num, cudaMemcpyDeviceToHost); -} - -template <> -void Copy(platform::GPUPlace dst_place, - void* dst, - platform::CPUPlace src_place, - const void* src, size_t num) { - platform::SetDeviceId(dst_place.device); - platform::GpuMemcpySync(dst, src, num, cudaMemcpyHostToDevice); -} - -template <> -void Copy(platform::GPUPlace dst_place, - void* dst, - platform::GPUPlace src_place, - const void* src, size_t num) { - platform::SetDeviceId(dst_place.device); - platform::GpuMemcpySync(dst, src, num, cudaMemcpyDeviceToDevice); -} - #endif } // namespace memory diff --git a/paddle/memory/memory.cc b/paddle/memory/memory.cc index 5eb1c44eb6fc45db31ef44bf79e74b79193e08aa..c4bb6baee7ebf2941cee5915ca2723c298689261 100644 --- a/paddle/memory/memory.cc +++ b/paddle/memory/memory.cc @@ -64,35 +64,52 @@ BuddyAllocator* GetGPUBuddyAllocator(int gpu_id) { int gpu_num = platform::GetCUDADeviceCount(); as = new BuddyAllocator*[gpu_num]; for (int gpu = 0; gpu < gpu_num; gpu++) { - platform::SetDeviceId(gpu); - as[gpu] = new BuddyAllocator(new detail::GPUAllocator, - platform::GpuMinChunkSize(), - platform::GpuMaxChunkSize()); + as[gpu] = nullptr; } + } + platform::SetDeviceId(gpu_id); + if (!as[gpu_id]) { + as[gpu_id] = new BuddyAllocator(new detail::GPUAllocator, + platform::GpuMinChunkSize(), + platform::GpuMaxChunkSize()); VLOG(10) << "\n\nNOTE: each GPU device use " << FLAGS_fraction_of_gpu_memory_to_use * 100 << "% of GPU memory.\n" - << "You can set environment variable '" - << platform::kEnvFractionGpuMemoryToUse + << "You can set GFlags environment variable '" + << "FLAGS_fraction_of_gpu_memory_to_use" << "' to change the fraction of GPU usage.\n\n"; } - platform::SetDeviceId(gpu_id); return as[gpu_id]; } template <> -void* Alloc(platform::GPUPlace place, size_t size) { - return GetGPUBuddyAllocator(place.device)->Alloc(size); +size_t Used(platform::CUDAPlace place) { + return GetGPUBuddyAllocator(place.device)->Used(); } template <> -void Free(platform::GPUPlace place, void* p) { - GetGPUBuddyAllocator(place.device)->Free(p); +void* Alloc(platform::CUDAPlace place, size_t size) { + auto* buddy_allocator = GetGPUBuddyAllocator(place.device); + auto* ptr = buddy_allocator->Alloc(size); + if (ptr == nullptr) { + int cur_dev = platform::GetCurrentDeviceId(); + platform::SetDeviceId(place.device); + size_t avail, total; + platform::GpuMemoryUsage(avail, total); + LOG(WARNING) << "Cannot allocate " << size << " bytes in GPU " + << place.device << ", available " << avail << " bytes"; + LOG(WARNING) << "total " << total; + LOG(WARNING) << "GpuMinChunkSize " << platform::GpuMinChunkSize(); + LOG(WARNING) << "GpuMaxChunkSize " << platform::GpuMaxChunkSize(); + LOG(WARNING) << "GPU memory used: " << Used(place); + platform::SetDeviceId(cur_dev); + } + return ptr; } template <> -size_t Used(platform::GPUPlace place) { - return GetGPUBuddyAllocator(place.device)->Used(); +void Free(platform::CUDAPlace place, void* p) { + GetGPUBuddyAllocator(place.device)->Free(p); } #endif diff --git a/paddle/memory/memory_test.cc b/paddle/memory/memory_test.cc index 2444931e26774ae80b916fbb7bd46ff93025d9ed..f476bf71264da59a5c546968f4689145e1d8801b 100644 --- a/paddle/memory/memory_test.cc +++ b/paddle/memory/memory_test.cc @@ -82,7 +82,7 @@ TEST(BuddyAllocator, CPUMultAlloc) { #ifdef PADDLE_WITH_CUDA -size_t align(size_t size, paddle::platform::GPUPlace place) { +size_t align(size_t size, paddle::platform::CUDAPlace place) { size += sizeof(paddle::memory::detail::Metadata); size_t alignment = paddle::platform::GpuMinChunkSize(); size_t remaining = size % alignment; @@ -94,7 +94,7 @@ TEST(BuddyAllocator, GPUAllocation) { EXPECT_EQ(p, nullptr); - paddle::platform::GPUPlace gpu(0); + paddle::platform::CUDAPlace gpu(0); p = paddle::memory::Alloc(gpu, 4096); EXPECT_NE(p, nullptr); @@ -103,7 +103,7 @@ TEST(BuddyAllocator, GPUAllocation) { } TEST(BuddyAllocator, GPUMultAlloc) { - paddle::platform::GPUPlace gpu; + paddle::platform::CUDAPlace gpu; std::unordered_map ps; diff --git a/paddle/operators/CMakeLists.txt b/paddle/operators/CMakeLists.txt index d0fe5b4635174fa0f74658509c4e8ca58a1d056a..bfcc70b31defd378fba0e505b1522471ca285ac4 100644 --- a/paddle/operators/CMakeLists.txt +++ b/paddle/operators/CMakeLists.txt @@ -1,5 +1,6 @@ file(GLOB GENERAL_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "*_op.cc") string(REPLACE ".cc" "" GENERAL_OPS "${GENERAL_OPS}") +set(DEPS_OPS "") set(pybind_file ${PADDLE_SOURCE_DIR}/paddle/pybind/pybind.h) file(WRITE ${pybind_file} "// Generated by the paddle/operator/CMakeLists.txt. DO NOT EDIT!\n\n") function(op_library TARGET) @@ -48,6 +49,10 @@ function(op_library TARGET) message(FATAL_ERROR "The op library ${TARGET} should contains at least one .cc file") endif() + list(LENGTH op_library_DEPS op_library_DEPS_len) + if (${op_library_DEPS_len} GREATER 0) + set(DEPS_OPS ${TARGET} ${DEPS_OPS} PARENT_SCOPE) + endif() if (WITH_GPU) nv_library(${TARGET} SRCS ${cc_srcs} ${cu_cc_srcs} ${cu_srcs} DEPS ${op_library_DEPS} ${op_common_deps}) @@ -73,6 +78,13 @@ function(op_library TARGET) file(APPEND ${pybind_file} "USE_OP(conv2d);\n") endif() + # conv_cudnn_op contains several operators + if ("${TARGET}" STREQUAL "conv_cudnn_op") + set(pybind_flag 1) + # It's enough to just adding one operator to pybind + file(APPEND ${pybind_file} "USE_OP(conv2d_cudnn);\n") + endif() + # pool_op contains several operators if ("${TARGET}" STREQUAL "pool_op") set(pybind_flag 1) @@ -131,7 +143,7 @@ function(op_library TARGET) if ("${TARGET}" STREQUAL "nccl_op") set(pybind_flag 1) # It's enough to just adding one operator to pybind - file(APPEND ${pybind_file} "USE_GPU_ONLY_OP(ncclAllReduce);\n") + file(APPEND ${pybind_file} "USE_CUDA_ONLY_OP(ncclAllReduce);\n") endif() # reduce_op contains several operators @@ -178,12 +190,13 @@ set(DEPS_OPS cond_op cross_entropy_op recurrent_op - dynamic_recurrent_op softmax_with_cross_entropy_op softmax_op sequence_softmax_op sum_op pool_op + maxout_op + unpool_op pool_with_index_op conv_op conv_transpose_op @@ -193,44 +206,63 @@ set(DEPS_OPS lod_rank_table_op lod_tensor_to_array_op array_to_lod_tensor_op + max_sequence_len_op lstm_op - tensor_array_read_write_op gru_op adagrad_op - sgd_op) + sgd_op + save_op + load_op + send_op + recv_op + detection_output_op) +if(WITH_GPU) + op_library(nccl_op DEPS nccl_common) +else() + set(DEPS_OPS ${DEPS_OPS} nccl_op) +endif() +if(WITH_DISTRIBUTE) + add_subdirectory(detail) + set(DISTRIBUTE_DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib_target 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}) + set_source_files_properties(send_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) + op_library(recv_op DEPS ${DISTRIBUTE_DEPS}) + set_source_files_properties(recv_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) + cc_test(test_send_recv SRCS send_recv_op_test.cc DEPS send_op recv_op sum_op executor) +else() + set(DEPS_OPS ${DEPS_OPS} send_op recv_op) +endif() -op_library(cond_op SRCS cond_op.cc DEPS framework_proto tensor operator net_op) +op_library(cond_op DEPS framework_proto tensor net_op) 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(detection_output_op DEPS softmax) op_library(sequence_softmax_op DEPS softmax) op_library(sum_op DEPS selected_rows_functor) op_library(sgd_op DEPS selected_rows_functor) op_library(adagrad_op DEPS selected_rows_functor) op_library(conv_op DEPS vol2col) op_library(pool_op DEPS pooling) +op_library(maxout_op DEPS maxouting) +op_library(unpool_op DEPS unpooling) op_library(pool_with_index_op DEPS pooling) -op_library(lod_rank_table_op SRCS lod_rank_table_op.cc DEPS lod_rank_table) -op_library(lod_tensor_to_array_op SRCS lod_tensor_to_array_op.cc DEPS lod_rank_table_op) -op_library(array_to_lod_tensor_op SRCS array_to_lod_tensor_op.cc DEPS lod_rank_table_op) -op_library(tensor_array_read_write_op SRCS tensor_array_read_write_op.cc) -if(WITH_GPU) -op_library(nccl_op DEPS nccl_common) -endif() +op_library(lod_rank_table_op DEPS lod_rank_table) +op_library(lod_tensor_to_array_op DEPS lod_rank_table_op) +op_library(array_to_lod_tensor_op DEPS lod_rank_table_op) +op_library(max_sequence_len_op DEPS lod_rank_table) op_library(sequence_conv_op DEPS context_project) op_library(sequence_pool_op DEPS sequence_pooling) op_library(lstm_op DEPS sequence2batch lstm_compute) op_library(conv_transpose_op DEPS vol2col) op_library(gru_op DEPS sequence2batch gru_compute) -if(WITH_TESTING) - op_library(dynamic_recurrent_op SRCS dynamic_recurrent_op.cc rnn/recurrent_op_utils.cc - DEPS net_op tensor_array gtest) -else() - op_library(dynamic_recurrent_op SRCS dynamic_recurrent_op.cc rnn/recurrent_op_utils.cc - DEPS net_op tensor_array) -endif() -op_library(recurrent_op SRCS recurrent_op.cc DEPS executor) +op_library(recurrent_op DEPS executor) +op_library(cos_sim_op DEPS cos_sim_functor) +# FIXME(typhoonzero): save/load depends lodtensor serialization functions +op_library(save_op DEPS lod_tensor) +op_library(load_op DEPS lod_tensor) list(REMOVE_ITEM GENERAL_OPS ${DEPS_OPS}) foreach(src ${GENERAL_OPS}) @@ -239,15 +271,13 @@ endforeach() set(GLOB_OP_LIB ${OP_LIBRARY} CACHE INTERNAL "Global OP library") + cc_test(gather_test SRCS gather_test.cc DEPS tensor) cc_test(net_op_test SRCS net_op_test.cc DEPS net_op) cc_test(scatter_test SRCS scatter_test.cc DEPS tensor) cc_test(beam_search_decode_op_test SRCS beam_search_decode_op_test.cc DEPS lod_tensor) cc_test(strided_memcpy_test SRCS strided_memcpy_test.cc DEPS tensor paddle_memory) -cc_test(dynamic_recurrent_op_test SRCS dynamic_recurrent_op_test.cc - rnn/recurrent_op_utils.cc - DEPS dynamic_recurrent_op) if(WITH_GPU) - cc_test(nccl_op_test SRCS nccl_op_test.cu.cc DEPS nccl_op gpu_info device_context) + cc_test(nccl_op_test SRCS nccl_op_test.cu.cc DEPS nccl_op gpu_info device_context) endif() cc_test(save_load_op_test SRCS save_load_op_test.cc DEPS save_op load_op) diff --git a/paddle/operators/accuracy_op.cc b/paddle/operators/accuracy_op.cc index 2785a8c6fb62527db4d203788be88ebead068a19..d7baa6e90538484b400f32587a052d394a8d10d5 100644 --- a/paddle/operators/accuracy_op.cc +++ b/paddle/operators/accuracy_op.cc @@ -53,18 +53,17 @@ class AccuracyOp : public framework::OperatorWithKernel { } protected: - framework::OpKernelType GetKernelType( + framework::OpKernelType GetActualKernelType( const framework::ExecutionContext &ctx) const override { return framework::OpKernelType( framework::ToDataType(ctx.Input("Out")->type()), - ctx.device_context()); + ctx.GetPlace()); } }; class AccuracyOpMaker : public framework::OpProtoAndCheckerMaker { public: - AccuracyOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + AccuracyOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { // TODO(typhoonzero): support both inference value and indices. AddInput("Out", "The network output of topk (inferences)"); diff --git a/paddle/operators/accuracy_op.cu b/paddle/operators/accuracy_op.cu index d2dcab4e548b99c6beecfaa570ac31804fd07d82..0aadd5af41531e54b357756441f92da668d4ec01 100644 --- a/paddle/operators/accuracy_op.cu +++ b/paddle/operators/accuracy_op.cu @@ -26,7 +26,7 @@ template __global__ void AccuracyCudaKernel(const int N, const int D, const int64_t* Xdata, const int64_t* labeldata, int* correct_data, - float* accuracy) { + float* accuracy, int* total_data) { int count = 0; __shared__ int total[BlockSize]; @@ -47,6 +47,7 @@ __global__ void AccuracyCudaKernel(const int N, const int D, if (threadIdx.x == 0) { *correct_data = result; *accuracy = static_cast(result) / static_cast(N); + *total_data = N; } } @@ -55,7 +56,7 @@ class AccuracyOpCUDAKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), - "It must use GPUPlace."); + "It must use CUDAPlace."); auto* inference = ctx.Input("Out"); auto* indices = ctx.Input("Indices"); auto* label = ctx.Input("Label"); @@ -80,22 +81,11 @@ class AccuracyOpCUDAKernel : public framework::OpKernel { if (num_samples == 0) { return; } - platform::GpuMemcpyAsync(total_data, &num_samples, sizeof(int), - cudaMemcpyHostToDevice, stream); AccuracyCudaKernel< PADDLE_CUDA_NUM_THREADS><<<1, PADDLE_CUDA_NUM_THREADS, 0, stream>>>( num_samples, infer_width, indices_data, label_data, correct_data, - accuracy_data); - - int d_num_samples, d_num_correct; - float d_accuracy; - platform::GpuMemcpyAsync(&d_num_correct, correct_data, sizeof(int), - cudaMemcpyDeviceToHost, stream); - platform::GpuMemcpyAsync(&d_num_samples, total_data, sizeof(int), - cudaMemcpyDeviceToHost, stream); - platform::GpuMemcpyAsync(&d_accuracy, accuracy_data, sizeof(float), - cudaMemcpyDeviceToHost, stream); + accuracy_data, total_data); } }; @@ -104,5 +94,6 @@ class AccuracyOpCUDAKernel : public framework::OpKernel { // FIXME(typhoonzero): types of T is for inference data. // label data is always int64 -REGISTER_OP_GPU_KERNEL(accuracy, paddle::operators::AccuracyOpCUDAKernel, - paddle::operators::AccuracyOpCUDAKernel); +REGISTER_OP_CUDA_KERNEL(accuracy, + paddle::operators::AccuracyOpCUDAKernel, + paddle::operators::AccuracyOpCUDAKernel); diff --git a/paddle/operators/accuracy_op.h b/paddle/operators/accuracy_op.h index d060e6edddb31ecc1a4d27836f80b8ac5fa7d36d..04104a695fac6a967ad94780e31ba3fdd2ca2eda 100644 --- a/paddle/operators/accuracy_op.h +++ b/paddle/operators/accuracy_op.h @@ -21,7 +21,7 @@ namespace operators { using Tensor = framework::Tensor; -template +template class AccuracyKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { diff --git a/paddle/operators/activation_op.cc b/paddle/operators/activation_op.cc index 83d35a450d0e8ebf5311cdfd948b066642ccec8c..4188858a90daf8b2c10eb6960393de977d467371 100644 --- a/paddle/operators/activation_op.cc +++ b/paddle/operators/activation_op.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/operators/activation_op.h" @@ -22,8 +22,8 @@ class ActivationOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; void InferShape(framework::InferShapeContext *ctx) const override { - ctx->SetOutputDim("Y", ctx->GetInputDim("X")); - ctx->ShareLoD("X", /*->*/ "Y"); + ctx->SetOutputDim("Out", ctx->GetInputDim("X")); + ctx->ShareLoD("X", /*->*/ "Out"); } }; @@ -32,21 +32,20 @@ class ActivationOpGrad : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; void InferShape(framework::InferShapeContext *ctx) const override { - ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("Y")); + ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("Out")); } }; class SigmoidOpMaker : public framework::OpProtoAndCheckerMaker { public: - SigmoidOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + SigmoidOpMaker(OpProto *proto, OpAttrChecker *op_checker) + : framework::OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "Input of Sigmoid operator"); - AddOutput("Y", "Output of Sigmoid operator"); + AddOutput("Out", "Output of Sigmoid operator"); AddComment(R"DOC( -Sigmoid Activation Operator. +Sigmoid Activation Operator -$y = 1 / (1 + e^{-x})$ +$$out = \frac{1}{1 + e^{-x}}$$ )DOC"); } @@ -54,15 +53,14 @@ $y = 1 / (1 + e^{-x})$ class LogSigmoidOpMaker : public framework::OpProtoAndCheckerMaker { public: - LogSigmoidOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + LogSigmoidOpMaker(OpProto *proto, OpAttrChecker *op_checker) + : framework::OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "Input of LogSigmoid operator"); - AddOutput("Y", "Output of LogSigmoid operator"); + AddOutput("Out", "Output of LogSigmoid operator"); AddComment(R"DOC( -Logsigmoid Activation Operator. +Logsigmoid Activation Operator -$y = \log(1 / (1 + e^{-x}))$ +$$out = \log \frac{1}{1 + e^{-x}}$$ )DOC"); } @@ -70,14 +68,14 @@ $y = \log(1 / (1 + e^{-x}))$ class ExpOpMaker : public framework::OpProtoAndCheckerMaker { public: - ExpOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + ExpOpMaker(OpProto *proto, OpAttrChecker *op_checker) + : framework::OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "Input of Exp operator"); - AddOutput("Y", "Output of Exp operator"); + AddOutput("Out", "Output of Exp operator"); AddComment(R"DOC( Exp Activation Operator. -$y = e^x$ +$out = e^x$ )DOC"); } @@ -85,53 +83,47 @@ $y = e^x$ class ReluOpMaker : public framework::OpProtoAndCheckerMaker { public: - ReluOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + ReluOpMaker(OpProto *proto, OpAttrChecker *op_checker) + : framework::OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "Input of Relu operator"); - AddOutput("Y", "Output of Relu operator"); + AddOutput("Out", "Output of Relu operator"); AddComment(R"DOC( Relu Activation Operator. -$y = \max(x, 0)$ +$out = \max(x, 0)$ )DOC"); } }; -template class LeakyReluOpMaker : public framework::OpProtoAndCheckerMaker { public: - LeakyReluOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + LeakyReluOpMaker(OpProto *proto, OpAttrChecker *op_checker) + : framework::OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "Input of LeakyRelu operator"); - AddOutput("Y", "Output of LeakyRelu operator"); - AddAttr("alpha", "The small negative slope") - .SetDefault(static_cast(0.02f)); + AddOutput("Out", "Output of LeakyRelu operator"); + AddAttr("alpha", "The small negative slope").SetDefault(0.02f); AddComment(R"DOC( LeakyRelu Activation Operator. -$y = \max(x, \alpha * x)$ +$out = \max(x, \alpha * x)$ )DOC"); } }; -template class SoftShrinkOpMaker : public framework::OpProtoAndCheckerMaker { public: - SoftShrinkOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + SoftShrinkOpMaker(OpProto *proto, OpAttrChecker *op_checker) + : framework::OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "Input of Softshrink operator"); - AddOutput("Y", "Output of Softshrink operator"); - AddAttr("lambda", "non-negative offset") - .SetDefault(static_cast(0.5f)); + AddOutput("Out", "Output of Softshrink operator"); + AddAttr("lambda", "non-negative offset").SetDefault(0.5f); AddComment(R"DOC( Softshrink Activation Operator. $$ -y = \begin{cases} +out = \begin{cases} x - \lambda, \text{if } x > \lambda \\ x + \lambda, \text{if } x < -\lambda \\ 0, \text{otherwise} @@ -144,14 +136,14 @@ $$ class TanhOpMaker : public framework::OpProtoAndCheckerMaker { public: - TanhOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + TanhOpMaker(OpProto *proto, OpAttrChecker *op_checker) + : framework::OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "Input of Tanh operator"); - AddOutput("Y", "Output of Tanh operator"); + AddOutput("Out", "Output of Tanh operator"); AddComment(R"DOC( Tanh Activation Operator. -$$y = \frac{e^{x} - e^{-x}}{e^{x} + e^{-x}}$$ +$$out = \frac{e^{x} - e^{-x}}{e^{x} + e^{-x}}$$ )DOC"); } @@ -159,35 +151,32 @@ $$y = \frac{e^{x} - e^{-x}}{e^{x} + e^{-x}}$$ class TanhShrinkOpMaker : public framework::OpProtoAndCheckerMaker { public: - TanhShrinkOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + TanhShrinkOpMaker(OpProto *proto, OpAttrChecker *op_checker) + : framework::OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "Input of TanhShrink operator"); - AddOutput("Y", "Output of TanhShrink operator"); + AddOutput("Out", "Output of TanhShrink operator"); AddComment(R"DOC( TanhShrink Activation Operator. -$$y = x - \frac{e^{x} - e^{-x}}{e^{x} + e^{-x}}$$ +$$out = x - \frac{e^{x} - e^{-x}}{e^{x} + e^{-x}}$$ )DOC"); } }; -template class HardShrinkOpMaker : public framework::OpProtoAndCheckerMaker { public: - HardShrinkOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + HardShrinkOpMaker(OpProto *proto, OpAttrChecker *op_checker) + : framework::OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "Input of HardShrink operator"); - AddOutput("Y", "Output of HardShrink operator"); - AddAttr("threshold", "The value of threshold for HardShrink") - .SetDefault(static_cast(0.5)); + AddOutput("Out", "Output of HardShrink operator"); + AddAttr("threshold", "The value of threshold for HardShrink") + .SetDefault(0.5f); AddComment(R"DOC( HardShrink Activation Operator. $$ -y = \begin{cases} +out = \begin{cases} x, \text{if } x > \lambda \\ x, \text{if } x < -\lambda \\ 0, \text{otherwise} @@ -200,14 +189,14 @@ $$ class SqrtOpMaker : public framework::OpProtoAndCheckerMaker { public: - SqrtOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + SqrtOpMaker(OpProto *proto, OpAttrChecker *op_checker) + : framework::OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "Input of Sqrt operator"); - AddOutput("Y", "Output of Sqrt operator"); + AddOutput("Out", "Output of Sqrt operator"); AddComment(R"DOC( Sqrt Activation Operator. -$y = \sqrt{x}$ +$out = \sqrt{x}$ )DOC"); } @@ -215,14 +204,59 @@ $y = \sqrt{x}$ class AbsOpMaker : public framework::OpProtoAndCheckerMaker { public: - AbsOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + AbsOpMaker(OpProto *proto, OpAttrChecker *op_checker) + : framework::OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "Input of Abs operator"); - AddOutput("Y", "Output of Abs operator"); + AddOutput("Out", "Output of Abs operator"); AddComment(R"DOC( Abs Activation Operator. -$y = |x|$ +$out = |x|$ + +)DOC"); + } +}; + +class CeilOpMaker : public framework::OpProtoAndCheckerMaker { + public: + CeilOpMaker(OpProto *proto, OpAttrChecker *op_checker) + : framework::OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "Input of Ceil operator"); + AddOutput("Out", "Output of Ceil operator"); + AddComment(R"DOC( +Ceil Activation Operator. + +$out = ceil(x)$ + +)DOC"); + } +}; + +class FloorOpMaker : public framework::OpProtoAndCheckerMaker { + public: + FloorOpMaker(OpProto *proto, OpAttrChecker *op_checker) + : framework::OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "Input of Floor operator"); + AddOutput("Out", "Output of Floor operator"); + AddComment(R"DOC( +Floor Activation Operator. + +$out = floor(x)$ + +)DOC"); + } +}; + +class RoundOpMaker : public framework::OpProtoAndCheckerMaker { + public: + RoundOpMaker(OpProto *proto, OpAttrChecker *op_checker) + : framework::OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "Input of Round operator"); + AddOutput("Out", "Output of Round operator"); + AddComment(R"DOC( +Round Activation Operator. + +$out = [x]$ )DOC"); } @@ -230,15 +264,14 @@ $y = |x|$ class ReciprocalOpMaker : public framework::OpProtoAndCheckerMaker { public: - ReciprocalOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + ReciprocalOpMaker(OpProto *proto, OpAttrChecker *op_checker) + : framework::OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "Input of Reciprocal operator"); - AddOutput("Y", "Output of Reciprocal operator"); + AddOutput("Out", "Output of Reciprocal operator"); AddComment(R"DOC( Reciprocal Activation Operator. -$$y = \frac{1}{x}$$ +$$out = \frac{1}{x}$$ )DOC"); } @@ -246,14 +279,14 @@ $$y = \frac{1}{x}$$ class LogOpMaker : public framework::OpProtoAndCheckerMaker { public: - LogOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + LogOpMaker(OpProto *proto, OpAttrChecker *op_checker) + : framework::OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "Input of Log operator"); - AddOutput("Y", "Output of Log operator"); + AddOutput("Out", "Output of Log operator"); AddComment(R"DOC( Log Activation Operator. -$y = \ln(x)$ +$out = \ln(x)$ Natural logarithm of x. @@ -263,14 +296,14 @@ Natural logarithm of x. class SquareOpMaker : public framework::OpProtoAndCheckerMaker { public: - SquareOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + SquareOpMaker(OpProto *proto, OpAttrChecker *op_checker) + : framework::OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "Input of Square operator"); - AddOutput("Y", "Output of Square operator"); + AddOutput("Out", "Output of Square operator"); AddComment(R"DOC( Square Activation Operator. -$y = x^2$ +$out = x^2$ )DOC"); } @@ -278,15 +311,14 @@ $y = x^2$ class SoftplusOpMaker : public framework::OpProtoAndCheckerMaker { public: - SoftplusOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + SoftplusOpMaker(OpProto *proto, OpAttrChecker *op_checker) + : framework::OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "Input of Softplus operator"); - AddOutput("Y", "Output of Softplus operator"); + AddOutput("Out", "Output of Softplus operator"); AddComment(R"DOC( Softplus Activation Operator. -$y = \ln(1 + e^{x})$ +$out = \ln(1 + e^{x})$ )DOC"); } @@ -294,151 +326,139 @@ $y = \ln(1 + e^{x})$ class SoftsignOpMaker : public framework::OpProtoAndCheckerMaker { public: - SoftsignOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + SoftsignOpMaker(OpProto *proto, OpAttrChecker *op_checker) + : framework::OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "Input of Softsign operator"); - AddOutput("Y", "Output of Softsign operator"); + AddOutput("Out", "Output of Softsign operator"); AddComment(R"DOC( Softsign Activation Operator. -$$y = \frac{x}{1 + |x|}$$ +$$out = \frac{x}{1 + |x|}$$ )DOC"); } }; -template class BReluOpMaker : public framework::OpProtoAndCheckerMaker { public: - BReluOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + BReluOpMaker(OpProto *proto, OpAttrChecker *op_checker) + : framework::OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "Input of BRelu operator"); - AddOutput("Y", "Output of BRelu operator"); - AddAttr("t_min", "The min marginal value of BRelu") - .SetDefault(static_cast(0)); - AddAttr("t_max", "The max marginal value of BRelu") - .SetDefault(static_cast(24)); + AddOutput("Out", "Output of BRelu operator"); + AddAttr("t_min", "The min marginal value of BRelu") + .SetDefault(static_cast(0)); + AddAttr("t_max", "The max marginal value of BRelu") + .SetDefault(static_cast(24)); AddComment(R"DOC( BRelu Activation Operator. -$y = \max(\min(x, t_{min}), t_{max})$ +$out = \max(\min(x, t_{min}), t_{max})$ )DOC"); } }; -template class SoftReluOpMaker : public framework::OpProtoAndCheckerMaker { public: - SoftReluOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + SoftReluOpMaker(OpProto *proto, OpAttrChecker *op_checker) + : framework::OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "Input of SoftRelu operator"); - AddOutput("Y", "Output of SoftRelu operator"); - AddAttr("threshold", "The threshold value of SoftRelu") - .SetDefault(static_cast(40)); + AddOutput("Out", "Output of SoftRelu operator"); + AddAttr("threshold", "The threshold value of SoftRelu") + .SetDefault(40.0f); AddComment(R"DOC( SoftRelu Activation Operator. -$y = \ln(1 + \exp(\max(\min(x, threshold), threshold))$ +$out = \ln(1 + \exp(\max(\min(x, threshold), threshold))$ )DOC"); } }; -template class ELUOpMaker : public framework::OpProtoAndCheckerMaker { public: - ELUOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + ELUOpMaker(OpProto *proto, OpAttrChecker *op_checker) + : framework::OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "Input of ELU operator"); - AddOutput("Y", "Output of ELU operator"); - AddAttr("alpha", "The alpha value of ELU") - .SetDefault(static_cast(1.0f)); + AddOutput("Out", "Output of ELU operator"); + AddAttr("alpha", "The alpha value of ELU").SetDefault(1.0f); AddComment(R"DOC( ELU Activation Operator. Applies the following element-wise computation on the input according to https://arxiv.org/abs/1511.07289. -$y = \max(0, x) + \min(0, \alpha * (e^x - 1))$ +$out = \max(0, x) + \min(0, \alpha * (e^x - 1))$ )DOC"); } }; -template class Relu6OpMaker : public framework::OpProtoAndCheckerMaker { public: - Relu6OpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + Relu6OpMaker(OpProto *proto, OpAttrChecker *op_checker) + : framework::OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "Input of Relu6 operator"); - AddOutput("Y", "Output of Relu6 operator"); - AddAttr("threshold", "The threshold value of Relu6") - .SetDefault(static_cast(6)); + AddOutput("Out", "Output of Relu6 operator"); + AddAttr("threshold", "The threshold value of Relu6") + .SetDefault(6.0f); AddComment(R"DOC( Relu6 Activation Operator. -$y = \min(\max(0, x), 6)$ +$out = \min(\max(0, x), 6)$ )DOC"); } }; -template class PowOpMaker : public framework::OpProtoAndCheckerMaker { public: - PowOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + PowOpMaker(OpProto *proto, OpAttrChecker *op_checker) + : framework::OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "Input of Pow operator"); - AddOutput("Y", "Output of Pow operator"); - AddAttr("factor", "The exponential factor of Pow") - .SetDefault(static_cast(1)); + AddOutput("Out", "Output of Pow operator"); + AddAttr("factor", "The exponential factor of Pow").SetDefault(1.0f); AddComment(R"DOC( Pow Activation Operator. -$y = x^{factor}$ +$out = x^{factor}$ )DOC"); } }; -template class STanhOpMaker : public framework::OpProtoAndCheckerMaker { public: - STanhOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + STanhOpMaker(OpProto *proto, OpAttrChecker *op_checker) + : framework::OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "Input of STanh operator"); - AddOutput("Y", "Output of STanh operator"); - AddAttr("scale_a", "The scale parameter of a for the input") - .SetDefault(static_cast(2 / 3)); - AddAttr("scale_b", "The scale parameter of b for the input") - .SetDefault(static_cast(1.7159)); + AddOutput("Out", "Output of STanh operator"); + AddAttr("scale_a", "The scale parameter of a for the input") + .SetDefault(2.0f / 3.0f); + AddAttr("scale_b", "The scale parameter of b for the input") + .SetDefault(1.7159f); AddComment(R"DOC( STanh Activation Operator. -$$y = b * \frac{e^{a * x} - e^{-a * x}}{e^{a * x} + e^{-a * x}}$$ +$$out = b * \frac{e^{a * x} - e^{-a * x}}{e^{a * x} + e^{-a * x}}$$ )DOC"); } }; -template class ThresholdedReluOpMaker : public framework::OpProtoAndCheckerMaker { public: - ThresholdedReluOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + ThresholdedReluOpMaker(OpProto *proto, OpAttrChecker *op_checker) + : framework::OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "Input of ThresholdedRelu operator"); - AddOutput("Y", "Output of ThresholdedRelu operator"); - AddAttr("threshold", "The threshold location of activation") - .SetDefault(static_cast(1.0)); + AddOutput("Out", "Output of ThresholdedRelu operator"); + AddAttr("threshold", "The threshold location of activation") + .SetDefault(1.0f); AddComment(R"DOC( ThresholdedRelu Activation Operator. $$ -y = \begin{cases} +out = \begin{cases} x, \text{if } x > threshold \\ 0, \text{otherwise} \end{cases} @@ -448,25 +468,23 @@ $$ } }; -template class HardSigmoidOpMaker : public framework::OpProtoAndCheckerMaker { public: - HardSigmoidOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + HardSigmoidOpMaker(OpProto *proto, OpAttrChecker *op_checker) + : framework::OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "Input of HardSigmoid operator"); - AddOutput("Y", "Output of HardSigmoid operator"); - AddAttr("slope", "Slope for linear approximation of sigmoid") - .SetDefault(static_cast(0.2)); - AddAttr("offset", "Offset for linear approximation of sigmoid") - .SetDefault(static_cast(0.5)); + AddOutput("Out", "Output of HardSigmoid operator"); + AddAttr("slope", "Slope for linear approximation of sigmoid") + .SetDefault(0.2f); + AddAttr("offset", "Offset for linear approximation of sigmoid") + .SetDefault(0.5f); AddComment(R"DOC( HardSigmoid Activation Operator. Segment-wise linear approximation of sigmoid(https://arxiv.org/abs/1603.00391), which is much faster than sigmoid. -$y = \max(0, \min(1, slope * x + shift))$ +$out = \max(0, \min(1, slope * x + shift))$ The slope should be positive. The offset can be either positive or negative. The default slope and shift are set according to the above reference. @@ -476,6 +494,22 @@ 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) { + AddInput("X", "Input of Swish operator"); + AddOutput("Out", "Output of Swish operator"); + AddAttr("beta", "Constant beta of swish operator").SetDefault(1.0f); + AddComment(R"DOC( +Swish Activation Operator. + +$$out = \frac{x}{1 + e^{- \beta x}}$$ + +)DOC"); + } +}; + } // namespace operators } // namespace paddle @@ -499,7 +533,7 @@ REGISTER_OP(tanh, ops::ActivationOp, ops::TanhOpMaker, tanh_grad, REGISTER_OP(tanh_shrink, ops::ActivationOp, ops::TanhShrinkOpMaker, tanh_shrink_grad, ops::ActivationOpGrad); -REGISTER_OP(softshrink, ops::ActivationOp, ops::SoftShrinkOpMaker, +REGISTER_OP(softshrink, ops::ActivationOp, ops::SoftShrinkOpMaker, softshrink_grad, ops::ActivationOpGrad); REGISTER_OP(sqrt, ops::ActivationOp, ops::SqrtOpMaker, sqrt_grad, @@ -508,6 +542,15 @@ REGISTER_OP(sqrt, ops::ActivationOp, ops::SqrtOpMaker, sqrt_grad, REGISTER_OP(abs, ops::ActivationOp, ops::AbsOpMaker, abs_grad, ops::ActivationOpGrad); +REGISTER_OP(ceil, ops::ActivationOp, ops::CeilOpMaker, ceil_grad, + ops::ActivationOpGrad); + +REGISTER_OP(floor, ops::ActivationOp, ops::FloorOpMaker, floor_grad, + ops::ActivationOpGrad); + +REGISTER_OP(round, ops::ActivationOp, ops::RoundOpMaker, round_grad, + ops::ActivationOpGrad); + REGISTER_OP(reciprocal, ops::ActivationOp, ops::ReciprocalOpMaker, reciprocal_grad, ops::ActivationOpGrad); @@ -523,47 +566,50 @@ REGISTER_OP(softplus, ops::ActivationOp, ops::SoftplusOpMaker, softplus_grad, REGISTER_OP(softsign, ops::ActivationOp, ops::SoftsignOpMaker, softsign_grad, ops::ActivationOpGrad); -REGISTER_OP(brelu, ops::ActivationOp, ops::BReluOpMaker, brelu_grad, +REGISTER_OP(brelu, ops::ActivationOp, ops::BReluOpMaker, brelu_grad, ops::ActivationOpGrad); -REGISTER_OP(leaky_relu, ops::ActivationOp, ops::LeakyReluOpMaker, +REGISTER_OP(leaky_relu, ops::ActivationOp, ops::LeakyReluOpMaker, leaky_relu_grad, ops::ActivationOpGrad); -REGISTER_OP(soft_relu, ops::ActivationOp, ops::SoftReluOpMaker, - soft_relu_grad, ops::ActivationOpGrad); +REGISTER_OP(soft_relu, ops::ActivationOp, ops::SoftReluOpMaker, soft_relu_grad, + ops::ActivationOpGrad); -REGISTER_OP(elu, ops::ActivationOp, ops::ELUOpMaker, elu_grad, +REGISTER_OP(elu, ops::ActivationOp, ops::ELUOpMaker, elu_grad, ops::ActivationOpGrad); -REGISTER_OP(relu6, ops::ActivationOp, ops::Relu6OpMaker, relu6_grad, +REGISTER_OP(relu6, ops::ActivationOp, ops::Relu6OpMaker, relu6_grad, ops::ActivationOpGrad); -REGISTER_OP(pow, ops::ActivationOp, ops::PowOpMaker, pow_grad, +REGISTER_OP(pow, ops::ActivationOp, ops::PowOpMaker, pow_grad, ops::ActivationOpGrad); -REGISTER_OP(stanh, ops::ActivationOp, ops::STanhOpMaker, stanh_grad, +REGISTER_OP(stanh, ops::ActivationOp, ops::STanhOpMaker, stanh_grad, ops::ActivationOpGrad); -REGISTER_OP(hard_shrink, ops::ActivationOp, ops::HardShrinkOpMaker, +REGISTER_OP(hard_shrink, ops::ActivationOp, ops::HardShrinkOpMaker, hard_shrink_grad, ops::ActivationOpGrad); -REGISTER_OP(thresholded_relu, ops::ActivationOp, - ops::ThresholdedReluOpMaker, thresholded_relu_grad, - ops::ActivationOpGrad); +REGISTER_OP(thresholded_relu, ops::ActivationOp, ops::ThresholdedReluOpMaker, + thresholded_relu_grad, ops::ActivationOpGrad); -REGISTER_OP(hard_sigmoid, ops::ActivationOp, ops::HardSigmoidOpMaker, +REGISTER_OP(hard_sigmoid, ops::ActivationOp, ops::HardSigmoidOpMaker, hard_sigmoid_grad, ops::ActivationOpGrad); -#define REGISTER_ACTIVATION_CPU_KERNEL(act_type, functor, grad_functor) \ - REGISTER_OP_CPU_KERNEL( \ - act_type, \ - ops::ActivationKernel>, \ - ops::ActivationKernel>); \ - REGISTER_OP_CPU_KERNEL( \ - act_type##_grad, ops::ActivationGradKernel>, \ - ops::ActivationGradKernel>, \ + ops::ActivationKernel>); \ + REGISTER_OP_CPU_KERNEL( \ + act_type##_grad, \ + ops::ActivationGradKernel>, \ + ops::ActivationGradKernel>); FOR_EACH_KERNEL_FUNCTOR(REGISTER_ACTIVATION_CPU_KERNEL); diff --git a/paddle/operators/activation_op.cu b/paddle/operators/activation_op.cu index 97737857ab25dfa92163b64a750fd7a7d9ea0ac3..b9ccdf639cf4a9ea80d530e550c16089e50c44e0 100644 --- a/paddle/operators/activation_op.cu +++ b/paddle/operators/activation_op.cu @@ -1,32 +1,33 @@ /* 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 +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 + 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. */ +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. */ #define EIGEN_USE_GPU #include "paddle/operators/activation_op.h" namespace ops = paddle::operators; -#define REGISTER_ACTIVATION_GPU_KERNEL(act_type, functor, grad_functor) \ - REGISTER_OP_GPU_KERNEL( \ - act_type, \ - ops::ActivationKernel>, \ - ops::ActivationKernel>); \ - REGISTER_OP_GPU_KERNEL( \ - act_type##_grad, ops::ActivationGradKernel>, \ - ops::ActivationGradKernel>, \ + ops::ActivationKernel>); \ + REGISTER_OP_CUDA_KERNEL( \ + act_type##_grad, \ + ops::ActivationGradKernel>, \ + ops::ActivationGradKernel>); -FOR_EACH_KERNEL_FUNCTOR(REGISTER_ACTIVATION_GPU_KERNEL); +FOR_EACH_KERNEL_FUNCTOR(REGISTER_ACTIVATION_CUDA_KERNEL); diff --git a/paddle/operators/activation_op.h b/paddle/operators/activation_op.h index ceb4b4e40b67473f42e67e3f02f8e012e1b1eb50..0885f7c570b9b52dc51597347295734fd689da8d 100644 --- a/paddle/operators/activation_op.h +++ b/paddle/operators/activation_op.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/framework/eigen.h" @@ -19,7 +19,7 @@ namespace paddle { namespace operators { -template +template class ActivationKernel : public framework::OpKernel { public: @@ -27,45 +27,48 @@ class ActivationKernel void Compute(const framework::ExecutionContext& context) const override { auto* X = context.Input("X"); - auto* Y = context.Output("Y"); - Y->mutable_data(context.GetPlace()); + auto* Out = context.Output("Out"); + Out->mutable_data(context.GetPlace()); auto x = framework::EigenVector::Flatten(*X); - auto y = framework::EigenVector::Flatten(*Y); - auto place = context.GetEigenDevice(); + auto out = framework::EigenVector::Flatten(*Out); + auto* place = + context.template device_context().eigen_device(); Functor functor; auto attrs = functor.GetAttrs(); for (auto& attr : attrs) { *attr.second = context.Attr(attr.first); } - functor(place, x, y); + functor(*place, x, out); } }; -template +template class ActivationGradKernel : public framework::OpKernel { public: using T = typename Functor::ELEMENT_TYPE; void Compute(const framework::ExecutionContext& context) const override { auto* X = context.Input("X"); - auto* Y = context.Input("Y"); - auto* dY = context.Input(framework::GradVarName("Y")); + auto* Out = context.Input("Out"); + auto* dOut = + context.Input(framework::GradVarName("Out")); auto* dX = context.Output(framework::GradVarName("X")); dX->mutable_data(context.GetPlace()); - auto dy = framework::EigenVector::Flatten(*dY); + auto dout = framework::EigenVector::Flatten(*dOut); auto x = framework::EigenVector::Flatten(*X); - auto y = framework::EigenVector::Flatten(*Y); + auto out = framework::EigenVector::Flatten(*Out); auto dx = framework::EigenVector::Flatten(*dX); - auto place = context.GetEigenDevice(); + auto* place = + context.template device_context().eigen_device(); Functor functor; auto attrs = functor.GetAttrs(); for (auto& attr : attrs) { *attr.second = context.Attr(attr.first); } - functor(place, x, y, dy, dx); + functor(*place, x, out, dout, dx); } }; @@ -81,17 +84,18 @@ struct BaseActivationFunctor { // sigmoid(x) = 1 / (1 + exp(-x)) template struct SigmoidFunctor : public BaseActivationFunctor { - template - void operator()(Device d, X x, Y y) const { - y.device(d) = static_cast(1) / (static_cast(1) + (-x).exp()); + template + void operator()(Device d, X x, Out out) const { + out.device(d) = static_cast(1) / (static_cast(1) + (-x).exp()); } }; template struct SigmoidGradFunctor : public BaseActivationFunctor { - template - void operator()(Device d, X x, Y y, dY dy, dX dx) const { - dx.device(d) = dy * y * (static_cast(1) - y); + template + void operator()(Device d, X x, Out out, dOut dout, dX dx) const { + dx.device(d) = dout * out * (static_cast(1) - out); } }; @@ -99,7 +103,7 @@ struct SigmoidGradFunctor : public BaseActivationFunctor { // For numerical stability, we can use the log-sum-exp trick: // https://hips.seas.harvard.edu/blog/2013/01/09/computing-log-sum-exp/ // We can rewrite the above equation as: -// y = -log( exp(0) + exp(-x)) [since exp(0) = 1] +// out = -log( exp(0) + exp(-x)) [since exp(0) = 1] // = -log( exp(max(-x, 0) - max(-x, 0)) + exp(-x + max(-x, 0) - max(-x, 0))) // = -log( exp(max(-x, 0)) * exp(-max(-x, 0)) - exp(max(-x, 0)) * exp(-x - // max(-x, 0))) @@ -110,10 +114,10 @@ struct SigmoidGradFunctor : public BaseActivationFunctor { // + exp(-x - max(-x, 0)))) template struct LogSigmoidFunctor : public BaseActivationFunctor { - template - void operator()(Device d, X x, Y y) const { + template + void operator()(Device d, X x, Out out) const { auto temp = (-x).cwiseMax(static_cast(0)); // temp = max(-x, 0) - y.device(d) = -temp - (((-temp).exp() + (-x - temp).exp()).log()); + out.device(d) = -temp - (((-temp).exp() + (-x - temp).exp()).log()); } }; @@ -122,62 +126,66 @@ struct LogSigmoidFunctor : public BaseActivationFunctor { // exp(-x - max(-x, 0))) template struct LogSigmoidGradFunctor : public BaseActivationFunctor { - template - void operator()(Device d, X x, Y y, dY dy, dX dx) const { + template + void operator()(Device d, X x, Out out, dOut dout, dX dx) const { auto temp = (-x).cwiseMax(static_cast(0)); // temp = max(-x, 0) dx.device(d) = - dy * ((-x - temp).exp() / ((-temp).exp() + (-x - temp).exp())); + dout * ((-x - temp).exp() / ((-temp).exp() + (-x - temp).exp())); } }; // exp(x) = e^x template struct ExpFunctor : public BaseActivationFunctor { - template - void operator()(Device d, X x, Y y) const { - y.device(d) = x.exp(); + template + void operator()(Device d, X x, Out out) const { + out.device(d) = x.exp(); } }; template struct ExpGradFunctor : public BaseActivationFunctor { - template - void operator()(Device d, X x, Y y, dY dy, dX dx) const { - dx.device(d) = dy * y; + template + void operator()(Device d, X x, Out out, dOut dout, dX dx) const { + dx.device(d) = dout * out; } }; // relu(x) = max(x, 0) template struct ReluFunctor : public BaseActivationFunctor { - template - void operator()(Device d, X x, Y y) const { - y.device(d) = x.cwiseMax(static_cast(0)); + template + void operator()(Device d, X x, Out out) const { + out.device(d) = x.cwiseMax(static_cast(0)); } }; template struct ReluGradFunctor : public BaseActivationFunctor { - template - void operator()(Device d, X x, Y y, dY dy, dX dx) const { - dx.device(d) = dy * (x > static_cast(0)).template cast(); + template + void operator()(Device d, X x, Out out, dOut dout, dX dx) const { + dx.device(d) = dout * (x > static_cast(0)).template cast(); } }; // tanh(x) = (exp(x) - exp(-x)) / (exp(x) + exp(-x)) template struct TanhFunctor : public BaseActivationFunctor { - template - void operator()(Device d, X x, Y y) const { - y.device(d) = x.tanh(); + template + void operator()(Device d, X x, Out out) const { + out.device(d) = x.tanh(); } }; template struct TanhGradFunctor : public BaseActivationFunctor { - template - void operator()(Device d, X x, Y y, dY dy, dX dx) const { - dx.device(d) = dy * (static_cast(1) - y * y); + template + void operator()(Device d, X x, Out out, dOut dout, dX dx) const { + dx.device(d) = dout * (static_cast(1) - out * out); } }; @@ -185,17 +193,18 @@ struct TanhGradFunctor : public BaseActivationFunctor { // where tanh(x) = (exp(x) - exp(-x)) / (exp(x) + exp(-x)) template struct TanhShrinkFunctor : public BaseActivationFunctor { - template - void operator()(Device d, X x, Y y) const { - y.device(d) = x - x.tanh(); + template + void operator()(Device d, X x, Out out) const { + out.device(d) = x - x.tanh(); } }; template struct TanhShrinkGradFunctor : public BaseActivationFunctor { - template - void operator()(Device d, X x, Y y, dY dy, dX dx) const { - dx.device(d) = dy * (x.tanh() * x.tanh()); + template + void operator()(Device d, X x, Out out, dOut dout, dX dx) const { + dx.device(d) = dout * (x.tanh() * x.tanh()); } }; @@ -208,11 +217,11 @@ struct HardShrinkFunctor : public BaseActivationFunctor { typename BaseActivationFunctor::AttrPair GetAttrs() { return {{"threshold", &threshold}}; } - template - void operator()(Device d, X x, Y y) const { + template + void operator()(Device d, X x, Out out) const { auto temp1 = (x < static_cast(threshold * -1)).template cast().eval(); auto temp2 = (x > static_cast(threshold)).template cast().eval(); - y.device(d) = x * (temp1 + temp2); + out.device(d) = x * (temp1 + temp2); } }; @@ -224,11 +233,12 @@ struct HardShrinkGradFunctor : public BaseActivationFunctor { return {{"threshold", &threshold}}; } - template - void operator()(Device d, X x, Y y, dY dy, dX dx) const { + template + void operator()(Device d, X x, Out out, dOut dout, dX dx) const { auto temp1 = (x < static_cast(threshold * -1)).template cast().eval(); auto temp2 = (x > static_cast(threshold)).template cast().eval(); - dx.device(d) = dy * (temp1 + temp2).template cast(); + dx.device(d) = dout * (temp1 + temp2).template cast(); } }; @@ -241,12 +251,12 @@ struct SoftShrinkFunctor : public BaseActivationFunctor { return {{"lambda", &lambda}}; } - template - void operator()(Device d, X x, Y y) const { + template + void operator()(Device d, X x, Out out) const { auto lambdaT = static_cast(lambda); auto temp1 = (x > lambdaT).template cast().eval(); auto temp2 = (x < -lambdaT).template cast().eval(); - y.device(d) = temp1 * (x - lambdaT) + temp2 * (x + lambdaT); + out.device(d) = temp1 * (x - lambdaT) + temp2 * (x + lambdaT); } }; @@ -256,98 +266,140 @@ struct SoftShrinkGradFunctor : public BaseActivationFunctor { typename BaseActivationFunctor::AttrPair GetAttrs() { return {{"lambda", &lambda}}; } - template - void operator()(Device d, X x, Y y, dY dy, dX dx) const { + template + void operator()(Device d, X x, Out out, dOut dout, dX dx) const { auto lambdaT = static_cast(lambda); auto temp1 = (x > lambdaT).template cast().eval(); auto temp2 = (x < -lambdaT).template cast().eval(); - dx.device(d) = dy * (temp1 + temp2).template cast(); + dx.device(d) = dout * (temp1 + temp2).template cast(); } }; // sqrt(x) = x^(1/2) template struct SqrtFunctor : public BaseActivationFunctor { - template - void operator()(Device d, X x, Y y) const { - y.device(d) = x.sqrt(); + template + void operator()(Device d, X x, Out out) const { + out.device(d) = x.sqrt(); } }; template struct SqrtGradFunctor : public BaseActivationFunctor { - template - void operator()(Device d, X x, Y y, dY dy, dX dx) const { - const Y y_conj = Eigen::numext::conj(y); - dx.device(d) = static_cast(0.5) * dy / y_conj; + template + void operator()(Device d, X x, Out out, dOut dout, dX dx) const { + const Out out_conj = Eigen::numext::conj(out); + dx.device(d) = static_cast(0.5) * dout / out_conj; + } +}; + +// ceil(x) = ceiling(x) +template +struct CeilFunctor : public BaseActivationFunctor { + template + void operator()(Device d, X x, Out out) const { + out.device(d) = x.ceil(); + } +}; + +template +struct ZeroGradFunctor : public BaseActivationFunctor { + template + void operator()(Device d, X x, Out out, dOut dout, dX dx) const { + dx.device(d) = static_cast(0) / x; + } +}; + +// floor(x) = flooring(x) +template +struct FloorFunctor : public BaseActivationFunctor { + template + void operator()(Device d, X x, Out out) const { + out.device(d) = x.ceil(); + } +}; + +// round(x) = [x] +template +struct RoundFunctor : public BaseActivationFunctor { + template + void operator()(Device d, X x, Out out) const { + out.device(d) = x.round(); } }; // abs(x) = |x| template struct AbsFunctor : public BaseActivationFunctor { - template - void operator()(Device d, X x, Y y) const { - y.device(d) = x.abs(); + template + void operator()(Device d, X x, Out out) const { + out.device(d) = x.abs(); } }; template struct AbsGradFunctor : public BaseActivationFunctor { - template - void operator()(Device d, X x, Y y, dY dy, dX dx) const { - dx.device(d) = dy * x.sign(); + template + void operator()(Device d, X x, Out out, dOut dout, dX dx) const { + dx.device(d) = dout * x.sign(); } }; // reciprocal(x) = 1 / x template struct ReciprocalFunctor : public BaseActivationFunctor { - template - void operator()(Device d, X x, Y y) const { - y.device(d) = static_cast(1) / x; + template + void operator()(Device d, X x, Out out) const { + out.device(d) = static_cast(1) / x; } }; template struct ReciprocalGradFunctor : public BaseActivationFunctor { - template - void operator()(Device d, X x, Y y, dY dy, dX dx) const { - dx.device(d) = dy * static_cast(-1) * y * y; + template + void operator()(Device d, X x, Out out, dOut dout, dX dx) const { + dx.device(d) = dout * static_cast(-1) * out * out; } }; // log(x) = natural logarithm of x template struct LogFunctor : public BaseActivationFunctor { - template - void operator()(Device d, X x, Y y) const { - y.device(d) = x.log(); + template + void operator()(Device d, X x, Out out) const { + out.device(d) = x.log(); } }; template struct LogGradFunctor : public BaseActivationFunctor { - template - void operator()(Device d, X x, Y y, dY dy, dX dx) const { - dx.device(d) = dy * (static_cast(1) / x); + template + void operator()(Device d, X x, Out out, dOut dout, dX dx) const { + dx.device(d) = dout * (static_cast(1) / x); } }; // square(x) = x^2 template struct SquareFunctor : public BaseActivationFunctor { - template - void operator()(Device d, X x, Y y) const { - y.device(d) = x.square(); + template + void operator()(Device d, X x, Out out) const { + out.device(d) = x.square(); } }; template struct SquareGradFunctor : public BaseActivationFunctor { - template - void operator()(Device d, X x, Y y, dY dy, dX dx) const { - dx.device(d) = dy * static_cast(2) * x; + template + void operator()(Device d, X x, Out out, dOut dout, dX dx) const { + dx.device(d) = dout * static_cast(2) * x; } }; @@ -362,9 +414,9 @@ struct BReluFunctor : public BaseActivationFunctor { return {{"t_min", &t_min}, {"t_max", &t_max}}; } - template - void operator()(Device d, X x, Y y) const { - y.device(d) = + template + void operator()(Device d, X x, Out out) const { + out.device(d) = x.cwiseMax(static_cast(t_min)).cwiseMin(static_cast(t_max)); } }; @@ -376,9 +428,10 @@ struct BReluGradFunctor : public BaseActivationFunctor { typename BaseActivationFunctor::AttrPair GetAttrs() { return {{"t_min", &t_min}, {"t_max", &t_max}}; } - template - void operator()(Device d, X x, Y y, dY dy, dX dx) const { - dx.device(d) = dy * + template + void operator()(Device d, X x, Out out, dOut dout, dX dx) const { + dx.device(d) = dout * ((x > static_cast(t_min)) * (x < static_cast(t_max))) .template cast(); } @@ -393,9 +446,9 @@ struct Relu6Functor : public BaseActivationFunctor { return {{"threshold", &threshold}}; } - template - void operator()(Device d, X x, Y y) const { - y.device(d) = + template + void operator()(Device d, X x, Out out) const { + out.device(d) = x.cwiseMax(static_cast(0)).cwiseMin(static_cast(threshold)); } }; @@ -406,9 +459,10 @@ struct Relu6GradFunctor : public BaseActivationFunctor { typename BaseActivationFunctor::AttrPair GetAttrs() { return {{"threshold", &threshold}}; } - template - void operator()(Device d, X x, Y y, dY dy, dX dx) const { - dx.device(d) = dy * + template + void operator()(Device d, X x, Out out, dOut dout, dX dx) const { + dx.device(d) = dout * ((x > static_cast(0)) * (x < static_cast(threshold))) .template cast(); } @@ -421,10 +475,10 @@ struct Relu6GradFunctor : public BaseActivationFunctor { // Then: softplus(x) = max(x, 0) + log(exp(-max(x, 0)) + exp(x - max(x, 0))) template struct SoftplusFunctor : public BaseActivationFunctor { - template - void operator()(Device d, X x, Y y) { + template + void operator()(Device d, X x, Out out) { auto temp = x.cwiseMax(static_cast(0)); // temp = max(x, 0) - y.device(d) = temp + (((-temp).exp() + (x - temp).exp()).log()); + out.device(d) = temp + (((-temp).exp() + (x - temp).exp()).log()); } }; @@ -434,19 +488,21 @@ struct SoftplusFunctor : public BaseActivationFunctor { // exp(x - max(x, 0))) template struct SoftplusGradFunctor : public BaseActivationFunctor { - template - void operator()(Device d, X x, Y y, dY dy, dX dx) { + template + void operator()(Device d, X x, Out out, dOut dout, dX dx) { auto temp = x.cwiseMax(static_cast(0)); // temp = max(x, 0) - dx.device(d) = dy * ((x - temp).exp() / ((-temp).exp() + (x - temp).exp())); + dx.device(d) = + dout * ((x - temp).exp() / ((-temp).exp() + (x - temp).exp())); } }; // softsign(x) = x / (1 + |x|) template struct SoftsignFunctor : public BaseActivationFunctor { - template - void operator()(Device d, X x, Y y) { - y.device(d) = x / (static_cast(1) + x.abs()); + template + void operator()(Device d, X x, Out out) { + out.device(d) = x / (static_cast(1) + x.abs()); } }; @@ -454,10 +510,11 @@ struct SoftsignFunctor : public BaseActivationFunctor { // Taken from https://en.wikipedia.org/wiki/Activation_function template struct SoftsignGradFunctor : public BaseActivationFunctor { - template - void operator()(Device d, X x, Y y, dY dy, dX dx) { + template + void operator()(Device d, X x, Out out, dOut dout, dX dx) { dx.device(d) = - dy * (static_cast(1) / (static_cast(1) + x.abs()).square()); + dout * (static_cast(1) / (static_cast(1) + x.abs()).square()); } }; @@ -468,11 +525,11 @@ struct SoftReluFunctor : public BaseActivationFunctor { return {{"threshold", &threshold}}; } - template - void operator()(Device d, X x, Y y) const { + template + void operator()(Device d, X x, Out out) const { auto tmp = static_cast(threshold); auto temp = x.cwiseMax(-tmp).cwiseMin(tmp); - y.device(d) = (static_cast(1) + temp.exp()).log(); + out.device(d) = (static_cast(1) + temp.exp()).log(); } }; @@ -482,11 +539,12 @@ struct SoftReluGradFunctor : public BaseActivationFunctor { typename BaseActivationFunctor::AttrPair GetAttrs() { return {{"threshold", &threshold}}; } - template - void operator()(Device d, X x, Y y, dY dy, dX dx) const { + template + void operator()(Device d, X x, Out out, dOut dout, dX dx) const { auto tmp = static_cast(threshold); auto temp = ((x > -tmp) * (x < tmp)).template cast().eval(); - dx.device(d) = dy * (static_cast(1) - (-y).exp()) * temp; + dx.device(d) = dout * (static_cast(1) - (-out).exp()) * temp; } }; @@ -497,9 +555,9 @@ struct LeakyReluFunctor : public BaseActivationFunctor { return {{"alpha", &alpha}}; } - template - void operator()(Device d, X x, Y y) const { - y.device(d) = x.cwiseMax(static_cast(alpha) * x); + template + void operator()(Device d, X x, Out out) const { + out.device(d) = x.cwiseMax(static_cast(alpha) * x); } }; @@ -509,12 +567,13 @@ struct LeakyReluGradFunctor : public BaseActivationFunctor { typename BaseActivationFunctor::AttrPair GetAttrs() { return {{"alpha", &alpha}}; } - template - void operator()(Device d, X x, Y y, dY dy, dX dx) const { + template + void operator()(Device d, X x, Out out, dOut dout, dX dx) const { auto temp1 = static_cast(alpha) * (x < static_cast(0)).template cast().eval(); auto temp2 = (x >= static_cast(0)).template cast().eval(); - dx.device(d) = dy * (temp1 + temp2).template cast(); + dx.device(d) = dout * (temp1 + temp2).template cast(); } }; @@ -525,11 +584,11 @@ struct ELUFunctor : public BaseActivationFunctor { return {{"alpha", &alpha}}; } - template - void operator()(Device d, X x, Y y) const { - y.device(d) = x.cwiseMax(static_cast(0)) + - (static_cast(alpha) * (x.exp() - static_cast(1))) - .cwiseMin(static_cast(0)); + template + void operator()(Device d, X x, Out out) const { + out.device(d) = x.cwiseMax(static_cast(0)) + + (static_cast(alpha) * (x.exp() - static_cast(1))) + .cwiseMin(static_cast(0)); } }; @@ -539,10 +598,11 @@ struct ELUGradFunctor : public BaseActivationFunctor { typename BaseActivationFunctor::AttrPair GetAttrs() { return {{"alpha", &alpha}}; } - template - void operator()(Device d, X x, Y y, dY dy, dX dx) const { - dx.device(d) = dy * (x > static_cast(0)).template cast() + - dy * (y + static_cast(alpha)) * + template + void operator()(Device d, X x, Out out, dOut dout, dX dx) const { + dx.device(d) = dout * (x > static_cast(0)).template cast() + + dout * (out + static_cast(alpha)) * (x < static_cast(0)).template cast(); } }; @@ -554,9 +614,9 @@ struct PowFunctor : public BaseActivationFunctor { typename BaseActivationFunctor::AttrPair GetAttrs() { return {{"factor", &factor}}; } - template - void operator()(Device d, X x, Y y) const { - y.device(d) = x.pow(static_cast(factor)); + template + void operator()(Device d, X x, Out out) const { + out.device(d) = x.pow(static_cast(factor)); } }; @@ -566,9 +626,10 @@ struct PowGradFunctor : public BaseActivationFunctor { typename BaseActivationFunctor::AttrPair GetAttrs() { return {{"factor", &factor}}; } - template - void operator()(Device d, X x, Y y, dY dy, dX dx) const { - dx.device(d) = dy * static_cast(factor) * + template + void operator()(Device d, X x, Out out, dOut dout, dX dx) const { + dx.device(d) = dout * static_cast(factor) * x.pow(static_cast(factor - static_cast(1))); } }; @@ -581,9 +642,9 @@ struct STanhFunctor : public BaseActivationFunctor { return {{"scale_a", &scale_a}, {"scale_b", &scale_b}}; } - template - void operator()(Device d, X x, Y y) const { - y.device(d) = + template + void operator()(Device d, X x, Out out) const { + out.device(d) = static_cast(scale_b) * (static_cast(scale_a) * x).tanh(); } }; @@ -596,12 +657,13 @@ struct STanhGradFunctor : public BaseActivationFunctor { return {{"scale_a", &scale_a}, {"scale_b", &scale_b}}; } - template - void operator()(Device d, X x, Y y, dY dy, dX dx) const { + template + void operator()(Device d, X x, Out out, dOut dout, dX dx) const { auto a = static_cast(scale_a); auto b = static_cast(scale_b); auto temp = (a * x).tanh() * (a * x).tanh(); - dx.device(d) = dy * a * b * (static_cast(1) - temp); + dx.device(d) = dout * a * b * (static_cast(1) - temp); } }; @@ -612,10 +674,10 @@ struct ThresholdedReluFunctor : public BaseActivationFunctor { return {{"threshold", &threshold}}; } - template - void operator()(Device d, X x, Y y) const { + template + void operator()(Device d, X x, Out out) const { auto th = static_cast(threshold); - y.device(d) = (x > th).template cast() * x; + out.device(d) = (x > th).template cast() * x; } }; @@ -626,10 +688,11 @@ struct ThresholdedReluGradFunctor : public BaseActivationFunctor { return {{"threshold", &threshold}}; } - template - void operator()(Device d, X x, Y y, dY dy, dX dx) const { + template + void operator()(Device d, X x, Out out, dOut dout, dX dx) const { auto th = static_cast(threshold); - dx.device(d) = dy * (x > th).template cast(); + dx.device(d) = dout * (x > th).template cast(); } }; @@ -641,10 +704,11 @@ struct HardSigmoidFunctor : public BaseActivationFunctor { return {{"slope", &slope}, {"offset", &offset}}; } - template - void operator()(Device d, X x, Y y) const { + template + void operator()(Device d, X x, Out out) const { auto temp = x * static_cast(slope) + static_cast(offset); - y.device(d) = temp.cwiseMax(static_cast(0)).cwiseMin(static_cast(1)); + out.device(d) = + temp.cwiseMax(static_cast(0)).cwiseMin(static_cast(1)); } }; @@ -656,12 +720,43 @@ struct HardSigmoidGradFunctor : public BaseActivationFunctor { return {{"slope", &slope}, {"offset", &offset}}; } - template - void operator()(Device d, X x, Y y, dY dy, dX dx) const { - dx.device(d) = - dy * - ((y > static_cast(0)) * (y < static_cast(1))).template cast() * - static_cast(slope); + template + void operator()(Device d, X x, Out out, dOut dout, dX dx) const { + dx.device(d) = dout * + ((out > static_cast(0)) * (out < static_cast(1))) + .template cast() * + static_cast(slope); + } +}; + +template +struct SwishFunctor : public BaseActivationFunctor { + float beta; + typename BaseActivationFunctor::AttrPair GetAttrs() { + return {{"beta", &beta}}; + } + + template + void operator()(Device d, X x, Out out) const { + out.device(d) = x / (static_cast(1) + (static_cast(-beta) * x).exp()); + } +}; + +template +struct SwishGradFunctor : public BaseActivationFunctor { + float beta; + typename BaseActivationFunctor::AttrPair GetAttrs() { + return {{"beta", &beta}}; + } + + template + void operator()(Device d, X x, Out out, dOut dout, dX dx) const { + auto temp1 = static_cast(1) / + (static_cast(1) + (static_cast(-beta) * x).exp()); + auto temp2 = temp1 * (static_cast(1) - (beta * out)); + dx.device(d) = dout * ((beta * out) + temp2); } }; @@ -677,6 +772,9 @@ struct HardSigmoidGradFunctor : public BaseActivationFunctor { __macro(softshrink, SoftShrinkFunctor, SoftShrinkGradFunctor); \ __macro(sqrt, SqrtFunctor, SqrtGradFunctor); \ __macro(abs, AbsFunctor, AbsGradFunctor); \ + __macro(ceil, CeilFunctor, ZeroGradFunctor); \ + __macro(floor, FloorFunctor, ZeroGradFunctor); \ + __macro(round, RoundFunctor, ZeroGradFunctor); \ __macro(reciprocal, ReciprocalFunctor, ReciprocalGradFunctor); \ __macro(log, LogFunctor, LogGradFunctor); \ __macro(square, SquareFunctor, SquareGradFunctor); \ @@ -692,4 +790,5 @@ struct HardSigmoidGradFunctor : public BaseActivationFunctor { __macro(elu, ELUFunctor, ELUGradFunctor); \ __macro(hard_shrink, HardShrinkFunctor, HardShrinkGradFunctor); \ __macro(hard_sigmoid, HardSigmoidFunctor, HardSigmoidGradFunctor); \ + __macro(swish, SwishFunctor, SwishGradFunctor); \ __macro(thresholded_relu, ThresholdedReluFunctor, ThresholdedReluGradFunctor); diff --git a/paddle/operators/adadelta_op.cc b/paddle/operators/adadelta_op.cc index b717e1647e4b89285b841420650dc69e8a1e0c58..d8a9491c8247ac463e01606dac248780d5284236 100644 --- a/paddle/operators/adadelta_op.cc +++ b/paddle/operators/adadelta_op.cc @@ -59,8 +59,7 @@ class AdadeltaOp : public framework::OperatorWithKernel { class AdadeltaOpMaker : public framework::OpProtoAndCheckerMaker { public: - AdadeltaOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + AdadeltaOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("Param", "(Tensor) Input parameter"); AddInput("Grad", "(Tensor) Input gradient"); @@ -92,12 +91,12 @@ for gradient descent. Adadelta updates are as follows: -$$avgSquaredGradOut = \rho * avgSquaredGrad + (1 - \rho) * grad * grad \break -paramUpdate = - $\sqrt{((avgSquaredUpdate + \epsilon) / - (avgSquaredGrad_out + \epsilon))}$ * grad \break -avgSquaredUpdateOut = \rho * avgSquaredUpdate + (1 - \rho) * - {(paramUpdate)}^2 \break -paramOut = param + paramUpdate$$ +$$ +avg\_squared\_grad\_out = \rho * avg\_squared\_grad + (1 - \rho) * grad * grad \\ +param\_update = - \sqrt{\frac{avg\_squared\_update + \epsilon}{avg\_squared\_grad\_out + \epsilon}} * grad \\ +avg\_squared\_update\_out = \rho * avg\_squared\_update + (1 - \rho) * {param\_update}^2 \\ +param\_out = param + param\_update +$$ )DOC"); } @@ -109,4 +108,5 @@ paramOut = param + paramUpdate$$ namespace ops = paddle::operators; REGISTER_OP_WITHOUT_GRADIENT(adadelta, ops::AdadeltaOp, ops::AdadeltaOpMaker); REGISTER_OP_CPU_KERNEL( - adadelta, ops::AdadeltaOpKernel); + adadelta, ops::AdadeltaOpKernel, + ops::AdadeltaOpKernel); diff --git a/paddle/operators/adadelta_op.cu b/paddle/operators/adadelta_op.cu index 3af1c8c8e9861138a33b3156818f704c3b20363f..91294a0d5d148a43bb95ab83ae8176b475fde9de 100644 --- a/paddle/operators/adadelta_op.cu +++ b/paddle/operators/adadelta_op.cu @@ -1,20 +1,21 @@ /* 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 +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 + 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. */ +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. */ #define EIGEN_USE_GPU #include "paddle/operators/adadelta_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL( - adadelta, ops::AdadeltaOpKernel); +REGISTER_OP_CUDA_KERNEL( + adadelta, ops::AdadeltaOpKernel, + ops::AdadeltaOpKernel); diff --git a/paddle/operators/adadelta_op.h b/paddle/operators/adadelta_op.h index d29e15c43583bd447fbacb548a326f303f7d1463..819d0845dbdafab95d993a455013300fa71495e2 100644 --- a/paddle/operators/adadelta_op.h +++ b/paddle/operators/adadelta_op.h @@ -19,7 +19,7 @@ limitations under the License. */ namespace paddle { namespace operators { -template +template class AdadeltaOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { @@ -33,8 +33,8 @@ class AdadeltaOpKernel : public framework::OpKernel { avg_squared_grad_out_tensor->mutable_data(ctx.GetPlace()); avg_squared_update_out_tensor->mutable_data(ctx.GetPlace()); - float rho = ctx.Attr("rho"); - float epsilon = ctx.Attr("epsilon"); + T rho = static_cast(ctx.Attr("rho")); + T epsilon = static_cast(ctx.Attr("epsilon")); auto param = framework::EigenVector::Flatten( *ctx.Input("Param")); @@ -51,7 +51,7 @@ class AdadeltaOpKernel : public framework::OpKernel { framework::EigenVector::Flatten(*avg_squared_grad_out_tensor); auto avg_squared_update_out = framework::EigenVector::Flatten(*avg_squared_update_out_tensor); - auto place = ctx.GetEigenDevice(); + auto& place = *ctx.template device_context().eigen_device(); avg_squared_grad_out.device(place) = rho * avg_squared_grad + (1 - rho) * grad.square(); diff --git a/paddle/operators/adagrad_op.cc b/paddle/operators/adagrad_op.cc index d6686e3ef3165976cf4c077a7a0f213082aa7716..c83318a272302a474c37ce86365201acf56b9cad 100644 --- a/paddle/operators/adagrad_op.cc +++ b/paddle/operators/adagrad_op.cc @@ -59,8 +59,7 @@ class AdagradOp : public framework::OperatorWithKernel { class AdagradOpMaker : public framework::OpProtoAndCheckerMaker { public: - AdagradOpMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker) + AdagradOpMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("Param", "(Tensor) Input parameter"); AddInput("Grad", "(Tensor) Input gradient"); @@ -80,8 +79,8 @@ Adaptive Gradient Algorithm (Adagrad). The update is done as follows: -$$momentOut = moment + grad * grad \break -paramOut = param - learningRate * grad / ($\sqrt{momentOut}$ + \epsilon) \break +$$moment\_out = moment + grad * grad \\ +param\_out = param - \frac{learning\_rate * grad}{\sqrt{moment\_out} + \epsilon} $$ The original paper(http://www.jmlr.org/papers/volume12/duchi11a/duchi11a.pdf) @@ -100,54 +99,24 @@ size_t FindPos(const std::vector& rows, int64_t value) { } // namespace template -struct SparseAdagradFunctor { - void operator()(const platform::DeviceContext& context, +struct SparseAdagradFunctor { + void operator()(const platform::CPUDeviceContext& context, const framework::SelectedRows& grad, const framework::Tensor& learning_rate, T epsilon, framework::Tensor* moment, framework::Tensor* param) { // 1. g_m.rows = set(g.rows) - auto grad_rows = grad.rows(); - std::set row_set(grad_rows.begin(), grad_rows.end()); - std::vector merge_rows(row_set.begin(), row_set.end()); - auto grad_width = grad.value().dims()[1]; - std::unique_ptr grad_merge{ - new framework::SelectedRows()}; - grad_merge->set_rows(merge_rows); - grad_merge->set_height(grad.height()); - grad_merge->mutable_value()->mutable_data( - framework::make_ddim( - {static_cast(merge_rows.size()), grad_width}), - context.GetPlace()); - - math::SetConstant constant_functor; - constant_functor(context, grad_merge->mutable_value(), 0.0); - - auto* grad_merge_data = grad_merge->mutable_value()->data(); - auto* grad_data = grad.value().data(); - - for (size_t i = 0; i < grad_rows.size(); i++) { - size_t grad_merge_i = FindPos(merge_rows, grad_rows[i]); - for (int64_t j = 0; j < grad_width; j++) { - grad_merge_data[grad_merge_i * grad_width + j] += - grad_data[i * grad_width + j]; - } - } + math::scatter::MergeAdd merge_func; + auto grad_merge = merge_func(context, grad); + auto& merge_rows = grad_merge.rows(); + auto* grad_merge_data = grad_merge.mutable_value()->template data(); // 2. m += g_m * g_m - std::unique_ptr grad_square{ - new framework::SelectedRows()}; - grad_square->set_rows(grad_merge->rows()); - grad_square->set_height(grad_merge->height()); - grad_square->mutable_value()->mutable_data(grad_merge->value().dims(), - context.GetPlace()); - auto gs = - framework::EigenVector::Flatten(*(grad_square->mutable_value())); - auto gm = framework::EigenVector::Flatten(grad_merge->value()); - gs.device(*context.GetEigenDevice()) = gm * gm; - - math::SelectedRowsAddToTensor functor; - functor(context, *grad_square, moment); + math::scatter::Mul sqare_func; + auto grad_square = sqare_func(context, grad_merge, grad_merge); + + math::SelectedRowsAddToTensor functor; + functor(context, grad_square, moment); // 3. update parameter auto* lr = learning_rate.data(); @@ -164,13 +133,13 @@ struct SparseAdagradFunctor { } }; -template struct SparseAdagradFunctor; -template struct SparseAdagradFunctor; +template struct SparseAdagradFunctor; +template struct SparseAdagradFunctor; } // namespace operators } // namespace paddle namespace ops = paddle::operators; REGISTER_OP_WITHOUT_GRADIENT(adagrad, ops::AdagradOp, ops::AdagradOpMaker); REGISTER_OP_CPU_KERNEL( - adagrad, ops::AdagradOpKernel, - ops::AdagradOpKernel); + adagrad, ops::AdagradOpKernel, + ops::AdagradOpKernel); diff --git a/paddle/operators/adagrad_op.cu b/paddle/operators/adagrad_op.cu index 5b869e6bc5f4604ba6055ffd62fa21e4a1f41b93..4e579387924a5b0499f29609bc6b1322030a3c0d 100644 --- a/paddle/operators/adagrad_op.cu +++ b/paddle/operators/adagrad_op.cu @@ -1,21 +1,21 @@ /* 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 +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 + 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. */ +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. */ #define EIGEN_USE_GPU #include "paddle/operators/adagrad_op.h" -#include "paddle/operators/math/selected_rows_functor.h" #include "paddle/operators/math/math_function.h" +#include "paddle/operators/math/selected_rows_functor.h" #include "paddle/platform/cuda_helper.h" namespace paddle { @@ -72,80 +72,48 @@ __global__ void SparseAdagradFunctorKernel(const T* grad, const int64_t* rows, } // namespace template -struct SparseAdagradFunctor { - void operator()(const platform::DeviceContext& context, +struct SparseAdagradFunctor { + void operator()(const platform::CUDADeviceContext& context, const framework::SelectedRows& grad, const framework::Tensor& learning_rate, T epsilon, framework::Tensor* moment, framework::Tensor* param) { // 1. g_m.rows = set(g.rows) - auto grad_rows = grad.rows(); - std::set row_set(grad_rows.begin(), grad_rows.end()); - std::vector merge_rows(row_set.begin(), row_set.end()); - auto grad_width = grad.value().dims()[1]; - std::unique_ptr grad_merge{ - new framework::SelectedRows()}; - grad_merge->set_rows(merge_rows); - grad_merge->set_height(grad.height()); - grad_merge->mutable_value()->mutable_data( - framework::make_ddim( - {static_cast(merge_rows.size()), grad_width}), - context.GetPlace()); - - math::SetConstant constant_functor; - constant_functor(context, grad_merge->mutable_value(), 0.0); - - auto* grad_merge_data = grad_merge->mutable_value()->data(); - auto* grad_data = grad.value().data(); - - const int block_size = 256; - dim3 threads(block_size, 1); - dim3 grid1(1, grad_rows.size()); - - MergeGradKernel< - T, 256><<(context) - .stream()>>>(grad_data, grad.rows().data(), - grad_merge_data, grad_merge->rows().data(), - grad_merge->rows().size(), grad_width); - + math::scatter::MergeAdd merge_func; + auto grad_merge = merge_func(context, grad); + auto* grad_merge_data = grad_merge.mutable_value()->template data(); + auto& merge_rows = grad_merge.rows(); // 2. m += g_m * g_m - std::unique_ptr grad_square{ - new framework::SelectedRows()}; - grad_square->set_rows(grad_merge->rows()); - grad_square->set_height(grad_merge->height()); - grad_square->mutable_value()->mutable_data(grad_merge->value().dims(), - context.GetPlace()); - auto gs = - framework::EigenVector::Flatten(*(grad_square->mutable_value())); - auto gm = framework::EigenVector::Flatten(grad_merge->value()); - gs.device(*context.GetEigenDevice()) = gm * gm; - - math::SelectedRowsAddToTensor functor; - functor(context, *grad_square, moment); + math::scatter::Mul sqare_func; + auto grad_square = sqare_func(context, grad_merge, grad_merge); + + math::SelectedRowsAddToTensor functor; + functor(context, grad_square, moment); // 3. update parameter auto* lr = learning_rate.data(); auto* param_data = param->data(); auto* moment_data = moment->data(); + const int block_size = 256; + dim3 threads(block_size, 1); dim3 grid2(1, merge_rows.size()); SparseAdagradFunctorKernel< T, 256><<(context) - .stream()>>>(grad_merge_data, grad_merge->rows().data(), - lr, param_data, - moment_data, grad_width, epsilon); + .stream()>>>(grad_merge_data, grad_merge.rows().data(), + lr, param_data, moment_data, grad_width, + epsilon); } }; -template struct SparseAdagradFunctor; -template struct SparseAdagradFunctor; +template struct SparseAdagradFunctor; +template struct SparseAdagradFunctor; } // namespace operators } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL( - adagrad, ops::AdagradOpKernel, - ops::AdagradOpKernel); +REGISTER_OP_CUDA_KERNEL( + adagrad, ops::AdagradOpKernel, + ops::AdagradOpKernel); diff --git a/paddle/operators/adagrad_op.h b/paddle/operators/adagrad_op.h index 4d4a6434c7c472d8ceb01edfc4050fbb009d6c9f..0d77dbcbacd4efb6c1900e57b5c4ea9e9b136771 100644 --- a/paddle/operators/adagrad_op.h +++ b/paddle/operators/adagrad_op.h @@ -19,15 +19,15 @@ limitations under the License. */ namespace paddle { namespace operators { -template +template struct SparseAdagradFunctor { - void operator()(const platform::DeviceContext& context, + void operator()(const DeviceContext& context, const framework::SelectedRows& grad, const framework::Tensor& learning_rate, T epsilon, framework::Tensor* moment, framework::Tensor* param); }; -template +template class AdagradOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { @@ -52,11 +52,11 @@ class AdagradOpKernel : public framework::OpKernel { auto param_out = framework::EigenVector::Flatten(*param_out_tensor); auto moment_out = framework::EigenVector::Flatten(*moment_out_tensor); - auto place = ctx.GetEigenDevice(); + auto* place = ctx.template device_context().eigen_device(); - moment_out.device(place) = moment + grad * grad; + moment_out.device(*place) = moment + grad * grad; Eigen::DSizes m_dsize(moment_out_tensor->numel()); - param_out.device(place) = + param_out.device(*place) = param - lr.broadcast(m_dsize) * grad / (moment_out.sqrt() + epsilon); } else if (grad_var->IsType()) { auto* param_tensor = ctx.Input("Param"); @@ -65,8 +65,9 @@ class AdagradOpKernel : public framework::OpKernel { auto* moment_tensor = ctx.Input("Moment"); PADDLE_ENFORCE_EQ(moment_tensor, moment_out_tensor); - SparseAdagradFunctor functor; - functor(ctx.device_context(), *ctx.Input("Grad"), + SparseAdagradFunctor functor; + functor(ctx.template device_context(), + *ctx.Input("Grad"), *ctx.Input("LearningRate"), epsilon, moment_out_tensor, param_out_tensor); } else { diff --git a/paddle/operators/adam_op.cc b/paddle/operators/adam_op.cc index 97a091ae766abfba5412bbd32c34a6f80701fbf7..03527de936bf736d572fb0140033bde4db990981 100644 --- a/paddle/operators/adam_op.cc +++ b/paddle/operators/adam_op.cc @@ -73,7 +73,7 @@ class AdamOp : public framework::OperatorWithKernel { class AdamOpMaker : public framework::OpProtoAndCheckerMaker { public: - AdamOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + AdamOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("Param", "(Tensor) Input parameter"); AddInput("Grad", "(Tensor) Input gradient"); @@ -112,11 +112,13 @@ adaptive estimates of lower-order moments. Adam updates: -$$moment_1_{out} = \beta_1 * moment_1 + (1 - \beta_1) * grad \break -moment_2_{out} = \beta_2 * moment_2 + (1 - \beta_2) * grad * grad \break -learningRate = learningRate * - $\sqrt{(1 - \beta_2_{pow})}$ / (1 - \beta_1_{pow}) \break -paramOut = param - learningRate * moment_1/ ($\sqrt{(moment_2)} + \epsilon)$$ +$$ +moment\_1\_out = \beta_1 * moment\_1 + (1 - \beta_1) * grad \\ +moment\_2_\out = \beta_2 * moment\_2 + (1 - \beta_2) * grad * grad \\ +learning\_rate = learning\_rate * + \frac{\sqrt{1 - \beta_{2\_pow}}}{1 - \beta_{1\_pow}} \\ +param\_out = param - learning\_rate * \frac{moment\_1}{\sqrt{moment\_2} + \epsilon} +$$ )DOC"); } @@ -126,5 +128,6 @@ paramOut = param - learningRate * moment_1/ ($\sqrt{(moment_2)} + \epsilon)$$ namespace ops = paddle::operators; REGISTER_OP_WITHOUT_GRADIENT(adam, ops::AdamOp, ops::AdamOpMaker); -REGISTER_OP_CPU_KERNEL(adam, - ops::AdamOpKernel); +REGISTER_OP_CPU_KERNEL( + adam, ops::AdamOpKernel, + ops::AdamOpKernel); diff --git a/paddle/operators/adam_op.cu b/paddle/operators/adam_op.cu index a3def912e540454275350209435eb01ae2151331..94f840c188942a900858429bc621c3a18d5900ad 100644 --- a/paddle/operators/adam_op.cu +++ b/paddle/operators/adam_op.cu @@ -1,20 +1,21 @@ /* 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 +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 + 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. */ +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. */ #define EIGEN_USE_GPU #include "paddle/operators/adam_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(adam, - ops::AdamOpKernel); +REGISTER_OP_CUDA_KERNEL( + adam, ops::AdamOpKernel, + ops::AdamOpKernel); diff --git a/paddle/operators/adam_op.h b/paddle/operators/adam_op.h index 45938006db1231a7a134964d729df6ca114d4dbe..9cc34bdded780e61e8700eb4fa4a295c84fb48bc 100644 --- a/paddle/operators/adam_op.h +++ b/paddle/operators/adam_op.h @@ -13,59 +13,210 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once -#include "paddle/framework/eigen.h" +#include // for sqrt in CPU and CUDA #include "paddle/framework/op_registry.h" +#include "paddle/operators/detail/safe_ref.h" +#include "paddle/operators/math/selected_rows_functor.h" +#include "paddle/platform/for_range.h" namespace paddle { namespace operators { -template +namespace scatter = paddle::operators::math::scatter; + +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) {} + + inline HOSTDEVICE void operator()(size_t i) const { + // Merge all memory access together. + T g = grad_[i]; + T mom1 = moment1_[i]; + T mom2 = moment2_[i]; + T lr = *lr_; + T beta1_pow = *beta1_pow_; + T beta2_pow = *beta2_pow_; + T p = param_[i]; + + // 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_)); + + // Write back to global memory + moment1_out_[i] = mom1; + moment2_out_[i] = mom2; + param_out_[i] = p; + } +}; + +template +struct SparseAdamFunctor { + 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_; + + const int64_t* rows_; + int64_t row_numel_; + + SparseAdamFunctor(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, const int64_t* rows, + int64_t row_numel) + : 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), + rows_(rows), + row_numel_(row_numel) {} + + inline HOSTDEVICE void operator()(size_t i) const { + T beta1_pow = *beta1_pow_; + T beta2_pow = *beta2_pow_; + for (int64_t j = 0; j < row_numel_; ++j) { + T g = grad_[i * row_numel_ + j]; + T mom1 = moment1_[rows_[i] * row_numel_ + j]; + T mom2 = moment2_[rows_[i] * row_numel_ + j]; + T lr = *lr_; + 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_)); + + moment1_out_[rows_[i] * row_numel_ + j] = mom1; + moment2_out_[rows_[i] * row_numel_ + j] = mom2; + param_out_[rows_[i] * row_numel_ + j] = p; + } // for col id + } +}; + +template class AdamOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { - auto param_out_tensor = ctx.Output("ParamOut"); - auto moment1_out_tensor = ctx.Output("Moment1Out"); - auto moment2_out_tensor = ctx.Output("Moment2Out"); - - param_out_tensor->mutable_data(ctx.GetPlace()); - moment1_out_tensor->mutable_data(ctx.GetPlace()); - moment2_out_tensor->mutable_data(ctx.GetPlace()); - - float beta1 = ctx.Attr("beta1"); - float beta2 = ctx.Attr("beta2"); - float epsilon = ctx.Attr("epsilon"); - - auto param = framework::EigenVector::Flatten( - *ctx.Input("Param")); - auto grad = framework::EigenVector::Flatten( - *ctx.Input("Grad")); - auto moment1 = framework::EigenVector::Flatten( - *ctx.Input("Moment1")); - auto moment2 = framework::EigenVector::Flatten( - *ctx.Input("Moment2")); - auto lr = framework::EigenVector::Flatten( - *ctx.Input("LearningRate")); - auto beta1_pow = framework::EigenVector::Flatten( - *ctx.Input("Beta1Pow")); - auto beta2_pow = framework::EigenVector::Flatten( - *ctx.Input("Beta2Pow")); - auto param_out = framework::EigenVector::Flatten(*param_out_tensor); - auto moment1_out = framework::EigenVector::Flatten(*moment1_out_tensor); - auto moment2_out = framework::EigenVector::Flatten(*moment2_out_tensor); - auto place = ctx.GetEigenDevice(); - - moment1_out.device(place) = beta1 * moment1 + (1 - beta1) * grad; - moment2_out.device(place) = beta2 * moment2 + (1 - beta2) * grad.square(); - - // All of these are tensors of 1 element - auto lr_t = lr * (1 - beta2_pow).sqrt() / (1 - beta1_pow); - // Eigen does not support automatic broadcast - // Get dimensions of moment vector to broadcast lr_t - Eigen::DSizes m_dsize(moment1_out_tensor->numel()); - param_out.device(place) = - param - - lr_t.broadcast(m_dsize) * - (moment1_out / (moment2_out.sqrt() + epsilon)); + using paddle::framework::LoDTensor; + using paddle::operators::detail::Ref; + + T beta1 = static_cast(ctx.Attr("beta1")); + T beta2 = static_cast(ctx.Attr("beta2")); + T epsilon = static_cast(ctx.Attr("epsilon")); + auto& param = Ref(ctx.Input("Param"), "Must set Param"); + // auto& grad = Ref(ctx.Input("Grad"), "Must set Grad"); + auto* grad_var = ctx.InputVar("Grad"); + auto& mom1 = Ref(ctx.Input("Moment1"), "Must set Moment1"); + auto& mom2 = Ref(ctx.Input("Moment2"), "Must set Moment2"); + auto& lr = + Ref(ctx.Input("LearningRate"), "Must set LearningRate"); + + auto& beta1_pow = + Ref(ctx.Input("Beta1Pow"), "Must set Beta1Pow"); + auto& beta2_pow = + Ref(ctx.Input("Beta2Pow"), "Must set Beta2Pow"); + + auto& param_out = + Ref(ctx.Output("ParamOut"), "Must set ParamOut"); + auto& mom1_out = + Ref(ctx.Output("Moment1Out"), "Must set Moment1Out"); + auto& mom2_out = + Ref(ctx.Output("Moment2Out"), "Must set Moment1Out"); + + 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); + } else if (grad_var->IsType()) { + auto& grad = + Ref(ctx.Input("Grad"), "Must set Grad"); + // merge duplicated rows if any. + scatter::MergeAdd merge_func; + auto grad_merge = + merge_func(ctx.template device_context(), grad); + auto& grad_tensor = grad_merge.value(); + const T* grad_data = grad_tensor.template data(); + auto* rows = grad_merge.rows().data(); + auto row_numel = grad_tensor.numel() / grad_merge.rows().size(); + + SparseAdamFunctor 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_data, param.template data(), + param_out.template mutable_data(ctx.GetPlace()), rows, row_numel); + platform::ForRange for_range( + static_cast(ctx.device_context()), + grad_merge.rows().size()); + for_range(functor); + } else { + PADDLE_THROW("Variable type not supported by adam_op"); + } } }; diff --git a/paddle/operators/adamax_op.cc b/paddle/operators/adamax_op.cc index 14cf3841b33a8153549e4c99ed2b75286e9c64db..3b0b71418477ea128dbb31a8d7cd44cf6bf023a1 100644 --- a/paddle/operators/adamax_op.cc +++ b/paddle/operators/adamax_op.cc @@ -67,7 +67,7 @@ class AdamaxOp : public framework::OperatorWithKernel { class AdamaxOpMaker : public framework::OpProtoAndCheckerMaker { public: - AdamaxOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + AdamaxOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("Param", "(Tensor) Input parameter"); AddInput("Grad", "(Tensor) Input gradient"); @@ -107,10 +107,12 @@ Adam algorithm based on the infinity norm. Adamax updates: -$$momentOut = \beta_1 * moment + (1 - \beta_1) * grad \break -infNormOut = max(\beta_2 * infNorm + \epsilon, |grad|) \break -learningRate = learningRate /(1 - \beta_1_{pow}) \break -paramOut = param - learningRate * momentPut / infNormOut$$ +$$ +moment\_out = \beta_1 * moment + (1 - \beta_1) * grad \\ +inf\_norm\_out = max(\beta_2 * inf\_norm + \epsilon, |grad|) \\ +learning\_rate = \frac{learning\_rate}{1 - \beta_{1\_pow}} \\ +param\_out = param - learning\_rate * \frac{moment\_out}{inf\_norm\_out} +$$ The original paper does not have an epsilon attribute. However, it is added here for numerical stability to prevent the @@ -125,5 +127,6 @@ division by 0 error. namespace ops = paddle::operators; REGISTER_OP_WITHOUT_GRADIENT(adamax, ops::AdamaxOp, ops::AdamaxOpMaker); -REGISTER_OP_CPU_KERNEL(adamax, - ops::AdamaxOpKernel); +REGISTER_OP_CPU_KERNEL( + adamax, ops::AdamaxOpKernel, + ops::AdamaxOpKernel); diff --git a/paddle/operators/adamax_op.cu b/paddle/operators/adamax_op.cu index fee3b6fc6b656917d79b84f48da8e63be7683890..8f87bb28671018a184f25a014f9bdb7615f3040c 100644 --- a/paddle/operators/adamax_op.cu +++ b/paddle/operators/adamax_op.cu @@ -1,20 +1,21 @@ /* 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 +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 + 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. */ +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. */ #define EIGEN_USE_GPU #include "paddle/operators/adamax_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(adamax, - ops::AdamaxOpKernel); +REGISTER_OP_CUDA_KERNEL( + adamax, ops::AdamaxOpKernel, + ops::AdamaxOpKernel); diff --git a/paddle/operators/adamax_op.h b/paddle/operators/adamax_op.h index 2c99832ec08e9c1d9b5458c467d5238f9b1b3c37..172c179c5fabf5ca106bf11479aff2d94a4e21d2 100644 --- a/paddle/operators/adamax_op.h +++ b/paddle/operators/adamax_op.h @@ -19,7 +19,7 @@ limitations under the License. */ namespace paddle { namespace operators { -template +template class AdamaxOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { @@ -31,9 +31,9 @@ class AdamaxOpKernel : public framework::OpKernel { moment_out_tensor->mutable_data(ctx.GetPlace()); inf_norm_out_tensor->mutable_data(ctx.GetPlace()); - float beta1 = ctx.Attr("beta1"); - float beta2 = ctx.Attr("beta2"); - float epsilon = ctx.Attr("epsilon"); + T beta1 = static_cast(ctx.Attr("beta1")); + T beta2 = static_cast(ctx.Attr("beta2")); + T epsilon = static_cast(ctx.Attr("epsilon")); auto param = framework::EigenVector::Flatten( *ctx.Input("Param")); @@ -51,14 +51,14 @@ class AdamaxOpKernel : public framework::OpKernel { auto moment_out = framework::EigenVector::Flatten(*moment_out_tensor); auto inf_norm_out = framework::EigenVector::Flatten(*inf_norm_out_tensor); - auto place = ctx.GetEigenDevice(); + auto* place = ctx.template device_context().eigen_device(); - moment_out.device(place) = beta1 * moment + (1 - beta1) * grad; - inf_norm_out.device(place) = + moment_out.device(*place) = beta1 * moment + (1 - beta1) * grad; + inf_norm_out.device(*place) = grad.abs().cwiseMax((beta2 * inf_norm) + epsilon); auto lr_t = lr / (1 - beta1_pow); Eigen::DSizes m_dsize(moment_out_tensor->numel()); - param_out.device(place) = + param_out.device(*place) = param - lr_t.broadcast(m_dsize) * (moment_out / inf_norm_out); } }; diff --git a/paddle/operators/array_operator.h b/paddle/operators/array_operator.h index 233a81198e336d3190565fb18556f96979cec0ce..e0eef5d9f93d70930ee82d663de9610cc0176e33 100644 --- a/paddle/operators/array_operator.h +++ b/paddle/operators/array_operator.h @@ -1,20 +1,21 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +/* 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 +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 + 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. */ +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/framework/lod_tensor_array.h" #include "paddle/framework/op_registry.h" +#include "paddle/platform/device_context.h" namespace paddle { namespace operators { @@ -27,16 +28,21 @@ class ArrayOp : public framework::OperatorBase { protected: size_t GetOffset(const framework::Scope &scope, - const platform::DeviceContext &dev_ctx) const { + const platform::Place &place) const { auto *i = scope.FindVar(Input("I")); PADDLE_ENFORCE(i != nullptr, "I must be set"); auto &i_tensor = i->Get(); PADDLE_ENFORCE_EQ(i_tensor.numel(), 1); + + // get device context from pool + platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance(); + auto &dev_ctx = *pool.Get(place); + size_t offset; if (platform::is_gpu_place(i_tensor.place())) { // FIXME: Avoid copy from GPU to CPU framework::Tensor t; - t.CopyFrom(i_tensor, platform::CPUPlace(), dev_ctx); + framework::CopyFrom(i_tensor, platform::CPUPlace(), dev_ctx, &t); dev_ctx.Wait(); offset = static_cast(*t.data()); } else { diff --git a/paddle/operators/array_to_lod_tensor_op.cc b/paddle/operators/array_to_lod_tensor_op.cc index c0903bb4e5ca7f160e19eefab99af7e3e4a8ed76..49366fee8df5a44a97b7b4e87cbf0b7c813a414a 100644 --- a/paddle/operators/array_to_lod_tensor_op.cc +++ b/paddle/operators/array_to_lod_tensor_op.cc @@ -1,21 +1,23 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +/* 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 +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 + 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. */ +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/framework/lod_rank_table.h" #include "paddle/framework/lod_tensor_array.h" #include "paddle/framework/op_registry.h" #include "paddle/memory/memcpy.h" +#include "paddle/platform/device_context.h" namespace paddle { namespace operators { @@ -30,7 +32,7 @@ class ArrayToLoDTensorOp : public framework::OperatorBase { const framework::AttributeMap &attrs) : OperatorBase(type, inputs, outputs, attrs) {} void Run(const framework::Scope &scope, - const platform::DeviceContext &dev_ctx) const override { + const platform::Place &dev_place) const override { auto &x = scope.FindVar(Input("X"))->Get(); auto &rank_table = scope.FindVar(Input("RankTable"))->Get(); @@ -102,8 +104,14 @@ class ArrayToLoDTensorOp : public framework::OperatorBase { if (len == 0) { continue; } - out->Slice(out_offset, out_offset + len) - .CopyFrom(x[x_idx].Slice(start_offset, end_offset), place, dev_ctx); + auto slice = out->Slice(out_offset, out_offset + len); + + platform::DeviceContextPool &pool = + platform::DeviceContextPool::Instance(); + auto &dev_ctx = *pool.Get(place); + + framework::CopyFrom(x[x_idx].Slice(start_offset, end_offset), place, + dev_ctx, &slice); out_offset += len; } } @@ -113,8 +121,7 @@ class ArrayToLoDTensorOp : public framework::OperatorBase { class ArrayToLoDTensorOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - ArrayToLoDTensorOpProtoMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + ArrayToLoDTensorOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "(std::vector) A vector of tensors that is going to " @@ -149,14 +156,14 @@ class ArrayToLoDTensorGradMaker : public framework::SingleGradOpDescMaker { using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; protected: - std::unique_ptr Apply() const override { - auto *grad_op = new framework::OpDescBind(); + std::unique_ptr Apply() const override { + auto *grad_op = new framework::OpDesc(); grad_op->SetType("lod_tensor_to_array"); grad_op->SetInput("X", OutputGrad("Out")); grad_op->SetInput("RankTable", Input("RankTable")); grad_op->SetOutput("Out", InputGrad("X")); grad_op->SetAttrMap(Attrs()); - return std::unique_ptr(grad_op); + return std::unique_ptr(grad_op); } }; diff --git a/paddle/operators/assign_op.cc b/paddle/operators/assign_op.cc index 609e915b932e2bc4d5abee1e5f868cc07a7619d3..7d77be3be1034bb38f6c92c181aa525214073eec 100644 --- a/paddle/operators/assign_op.cc +++ b/paddle/operators/assign_op.cc @@ -1,20 +1,21 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +/* 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 +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 + 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. */ +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/framework/data_type.h" #include "paddle/framework/op_registry.h" #include "paddle/framework/var_type.h" +#include "paddle/platform/device_context.h" namespace paddle { namespace operators { @@ -43,7 +44,8 @@ class AssignFunctor { out_rows.set_rows(rows.rows()); out_rows.set_height(rows.height()); auto &t = rows.value(); - out_rows.mutable_value()->CopyFrom(t, t.place(), dev_ctx_); + auto *m = out_rows.mutable_value(); + framework::CopyFrom(t, t.place(), dev_ctx_, m); } template @@ -55,7 +57,7 @@ class AssignFunctor { void copy_tensor(const framework::LoDTensor &lod_tensor, framework::LoDTensor *out) const { auto &out_tensor = *out; - out_tensor.CopyFrom(lod_tensor, lod_tensor.place(), dev_ctx_); + CopyFrom(lod_tensor, lod_tensor.place(), dev_ctx_, &out_tensor); out_tensor.set_lod(lod_tensor.lod()); } @@ -70,7 +72,7 @@ class AssignOp : public framework::OperatorBase { const framework::AttributeMap &attrs) : OperatorBase(type, inputs, outputs, attrs) {} void Run(const framework::Scope &scope, - const platform::DeviceContext &dev_ctx) const override { + const platform::Place &place) const override { auto *x = scope.FindVar(Input("X")); if (x == nullptr) { return; @@ -79,14 +81,17 @@ class AssignOp : public framework::OperatorBase { PADDLE_ENFORCE( out != nullptr, "The Output(Out) should not be null if the Input(X) is set."); + + platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance(); + auto &dev_ctx = *pool.Get(place); + framework::VisitVarType(*x, AssignFunctor(out, dev_ctx)); } }; class AssignOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - AssignOpProtoMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + AssignOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "(LoDTensor, SelectedRows or LoDTensorArray) The input variable " @@ -108,8 +113,8 @@ class AssignInferShape : public framework::InferShapeBase { void operator()(framework::InferShapeContext *context) const override { if (context->HasInput("X")) { auto type = context->GetInputsVarType("X")[0]; - if (type == framework::VarDesc_VarType_SELECTED_ROWS || - type == framework::VarDesc_VarType_LOD_TENSOR) { + if (type == framework::proto::VarDesc_VarType_SELECTED_ROWS || + type == framework::proto::VarDesc_VarType_LOD_TENSOR) { context->SetOutputDim("Out", context->GetInputDim("X")); } } @@ -121,12 +126,12 @@ class AssignGradMaker : public framework::SingleGradOpDescMaker { using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; protected: - std::unique_ptr Apply() const override { - auto *op = new framework::OpDescBind(); + std::unique_ptr Apply() const override { + auto *op = new framework::OpDesc(); op->SetType("assign"); op->SetInput("X", OutputGrad("Out")); op->SetOutput("Out", InputGrad("X")); - return std::unique_ptr(op); + return std::unique_ptr(op); } }; diff --git a/paddle/operators/auc_op.cc b/paddle/operators/auc_op.cc index 6c3f67ec32fb1b942241997e87a1e9c4752e707d..c16bc11931e6733d567107913521eafc34a30066 100644 --- a/paddle/operators/auc_op.cc +++ b/paddle/operators/auc_op.cc @@ -39,7 +39,7 @@ class AucOp : public framework::OperatorWithKernel { } protected: - framework::OpKernelType GetKernelType( + framework::OpKernelType GetActualKernelType( const framework::ExecutionContext &ctx) const override { return framework::OpKernelType( framework::ToDataType(ctx.Input("Out")->type()), @@ -49,7 +49,7 @@ class AucOp : public framework::OperatorWithKernel { class AucOpMaker : public framework::OpProtoAndCheckerMaker { public: - AucOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + AucOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("Out", "A floating point 2D tensor, values are in the range [0, 1]." diff --git a/paddle/operators/auc_op.h b/paddle/operators/auc_op.h index e5ac57b038ac32ed35bce35e477ede0cdb5da813..b80509e2a99a2a255dff2a98d950257588a21d29 100644 --- a/paddle/operators/auc_op.h +++ b/paddle/operators/auc_op.h @@ -25,7 +25,7 @@ template using EigenVector = framework::EigenVector; -template +template class AucKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { diff --git a/paddle/operators/batch_norm_op.cc b/paddle/operators/batch_norm_op.cc index f884e6efa917ce3f8554dce0e248f2b29273e3f3..98db28ddee7c6cdb37fe7732649d4fc38de7b873 100644 --- a/paddle/operators/batch_norm_op.cc +++ b/paddle/operators/batch_norm_op.cc @@ -13,12 +13,14 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/operators/batch_norm_op.h" +#include "paddle/framework/data_layout.h" namespace paddle { namespace operators { using Tensor = framework::Tensor; using LoDTensor = framework::LoDTensor; +using DataLayout = framework::DataLayout; template using EigenArrayMap = @@ -48,10 +50,6 @@ class BatchNormOp : public framework::OperatorWithKernel { PADDLE_ENFORCE(ctx->HasOutput("SavedMean"), ""); PADDLE_ENFORCE(ctx->HasOutput("SavedVariance"), ""); - const float epsilon = ctx->Attrs().Get("epsilon"); - PADDLE_ENFORCE_GE(epsilon, 0.0, "epsilon should be larger than 0"); - PADDLE_ENFORCE_LE(epsilon, 0.001, "epsilon should not be too large"); - // make sure Mean/MeanOut and Variance/VarianceOut share memory in Python PADDLE_ENFORCE_EQ(ctx->Inputs("Mean")[0], ctx->Outputs("MeanOut")[0], "Mean and MeanOut should share the same memory"); @@ -60,14 +58,15 @@ class BatchNormOp : public framework::OperatorWithKernel { "Variance and VarianceOut should share the same memory"); const auto x_dims = ctx->GetInputDim("X"); - const TensorFormat tensor_format = - StringToTensorFormat(ctx->Attrs().Get("tensor_format")); - const int C = - (tensor_format == TensorFormat::NCHW ? x_dims[1] - : x_dims[x_dims.size() - 1]); + const DataLayout data_layout = framework::StringToDataLayout( + ctx->Attrs().Get("data_layout")); + + PADDLE_ENFORCE(x_dims.size() >= 2 && x_dims.size() <= 5, + "Input X must have 2 to 5 dimensions."); - PADDLE_ENFORCE(x_dims.size() >= 3 && x_dims.size() <= 5, - "Input X must have 3 to 5 dimensions."); + const int C = + (data_layout == DataLayout::kNCHW ? x_dims[1] + : x_dims[x_dims.size() - 1]); PADDLE_ENFORCE_EQ(ctx->GetInputDim("Scale").size(), 1UL); PADDLE_ENFORCE_EQ(ctx->GetInputDim("Scale")[0], C); @@ -84,13 +83,17 @@ class BatchNormOp : public framework::OperatorWithKernel { class BatchNormOpMaker : public framework::OpProtoAndCheckerMaker { public: - BatchNormOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + BatchNormOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddAttr("is_test", "").SetDefault(false); AddAttr("momentum", "").SetDefault(0.9); - AddAttr("epsilon", "").SetDefault(1e-5); - AddAttr("tensor_format", "").SetDefault("NCHW"); + AddAttr("epsilon", "") + .SetDefault(1e-5) + .AddCustomChecker([](const float &epsilon) { + PADDLE_ENFORCE(epsilon >= 0.0f && epsilon <= 0.001f, + "'epsilon' should be between 0.0 and 0.001."); + }); + AddAttr("data_layout", "").SetDefault("NCHW"); AddInput("X", "The input tensor"); AddInput("Scale", "Scale is a 1-dimensional tensor of size C " @@ -134,24 +137,25 @@ The required data format for this layer is one of the following: }; template -class BatchNormKernel : public framework::OpKernel { +class BatchNormKernel + : public framework::OpKernel { public: void Compute(const framework::ExecutionContext &ctx) const override { const float epsilon = ctx.Attr("epsilon"); const float momentum = ctx.Attr("momentum"); const bool is_test = ctx.Attr("is_test"); - const std::string tensor_format_str = - ctx.Attr("tensor_format"); - const TensorFormat tensor_format = StringToTensorFormat(tensor_format_str); + const std::string data_layout_str = ctx.Attr("data_layout"); + const DataLayout data_layout = + framework::StringToDataLayout(data_layout_str); const auto *x = ctx.Input("X"); const auto &x_dims = x->dims(); - PADDLE_ENFORCE(x_dims.size() >= 3 && x_dims.size() <= 5, - "The Input dim size should be between 3 and 5"); + PADDLE_ENFORCE(x_dims.size() >= 2 && x_dims.size() <= 5, + "The Input dim size should be between 2 and 5"); const int N = x_dims[0]; const int C = - (tensor_format == TensorFormat::NCHW ? x_dims[1] - : x_dims[x_dims.size() - 1]); + (data_layout == DataLayout::kNCHW ? x_dims[1] + : x_dims[x_dims.size() - 1]); const int sample_size = x->numel() / N / C; auto *y = ctx.Output("Y"); @@ -176,8 +180,8 @@ class BatchNormKernel : public framework::OpKernel { saved_mean_e.setZero(); saved_variance_e.setZero(); - switch (tensor_format) { - case TensorFormat::NCHW: { + switch (data_layout) { + case DataLayout::kNCHW: { ConstEigenArrayMap x_arr(x->data(), sample_size, N * C); for (int nc = 0; nc < N * C; ++nc) { saved_mean_e(nc % C) += x_arr.col(nc).sum(); @@ -190,7 +194,7 @@ class BatchNormKernel : public framework::OpKernel { saved_variance_e /= N * sample_size; break; } - case TensorFormat::NHWC: { + case DataLayout::kNHWC: { ConstEigenArrayMap x_arr(x->data(), C, N * sample_size); for (int i = 0; i < N * sample_size; ++i) { saved_mean_e += x_arr.col(i); @@ -204,7 +208,7 @@ class BatchNormKernel : public framework::OpKernel { break; } default: - PADDLE_THROW("Unknown storage order: %s", tensor_format_str); + PADDLE_THROW("Unknown storage order: %s", data_layout_str); } EigenVectorArrayMap running_mean_arr( @@ -246,8 +250,8 @@ class BatchNormKernel : public framework::OpKernel { Eigen::Array new_bias = bias_arr - mean_arr * inv_std * scale_arr; - switch (tensor_format) { - case TensorFormat::NCHW: { + switch (data_layout) { + case DataLayout::kNCHW: { EigenArrayMap y_arr(y->mutable_data(ctx.GetPlace()), sample_size, N * C); ConstEigenArrayMap x_arr(x->data(), sample_size, N * C); @@ -256,7 +260,7 @@ class BatchNormKernel : public framework::OpKernel { } break; } - case TensorFormat::NHWC: { + case DataLayout::kNHWC: { EigenArrayMap(y->mutable_data(ctx.GetPlace()), C, N * sample_size) = (ConstEigenArrayMap(x->data(), C, N * sample_size).colwise() * @@ -266,7 +270,7 @@ class BatchNormKernel : public framework::OpKernel { break; } default: - PADDLE_THROW("Unknown storage order: %d", tensor_format); + PADDLE_THROW("Unknown storage order: %d", data_layout); } } }; @@ -289,11 +293,11 @@ class BatchNormGradOp : public framework::OperatorWithKernel { PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("Bias")), ""); const auto x_dims = ctx->GetInputDim("X"); - const TensorFormat tensor_format = - StringToTensorFormat(ctx->Attrs().Get("tensor_format")); + const DataLayout data_layout = framework::StringToDataLayout( + ctx->Attrs().Get("data_layout")); const int C = - (tensor_format == TensorFormat::NCHW ? x_dims[1] - : x_dims[x_dims.size() - 1]); + (data_layout == DataLayout::kNCHW ? x_dims[1] + : x_dims[x_dims.size() - 1]); ctx->SetOutputDim(framework::GradVarName("X"), x_dims); ctx->SetOutputDim(framework::GradVarName("Scale"), {C}); @@ -301,7 +305,7 @@ class BatchNormGradOp : public framework::OperatorWithKernel { } protected: - framework::OpKernelType GetKernelType( + framework::OpKernelType GetActualKernelType( const framework::ExecutionContext &ctx) const override { const auto *var = ctx.InputVar(framework::GradVarName("Y")); if (var == nullptr) { @@ -317,12 +321,12 @@ class BatchNormGradOp : public framework::OperatorWithKernel { PADDLE_THROW("can't find Y@GRAD"); } return framework::OpKernelType(framework::ToDataType(t->type()), - ctx.device_context()); + ctx.GetPlace()); } }; template -class BatchNormGradKernel +class BatchNormGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext &ctx) const override { @@ -332,19 +336,19 @@ class BatchNormGradKernel const auto *saved_mean = ctx.Input("SavedMean"); // SavedVariance have been reverted in forward operator const auto *saved_inv_variance = ctx.Input("SavedVariance"); - const std::string tensor_format_str = - ctx.Attr("tensor_format"); - const TensorFormat tensor_format = StringToTensorFormat(tensor_format_str); + const std::string data_layout_str = ctx.Attr("data_layout"); + const DataLayout data_layout = + framework::StringToDataLayout(data_layout_str); // Get the size for each dimension. // NCHW [batch_size, in_channels, in_height, in_width] const auto &x_dims = x->dims(); - PADDLE_ENFORCE(x_dims.size() >= 3 && x_dims.size() <= 5, - "The Input dim size should be between 3 and 5"); + PADDLE_ENFORCE(x_dims.size() >= 2 && x_dims.size() <= 5, + "The Input dim size should be between 2 and 5"); const int N = x_dims[0]; const int C = - (tensor_format == TensorFormat::NCHW ? x_dims[1] - : x_dims[x_dims.size() - 1]); + (data_layout == DataLayout::kNCHW ? x_dims[1] + : x_dims[x_dims.size() - 1]); const int sample_size = x->numel() / N / C; ConstEigenVectorArrayMap scale_arr(scale->data(), C); @@ -375,8 +379,8 @@ class BatchNormGradKernel const auto scale_inv_var_nhw = scale_arr * inv_var_arr / (N * sample_size); - switch (tensor_format) { - case TensorFormat::NCHW: { + switch (data_layout) { + case DataLayout::kNCHW: { ConstEigenArrayMap x_arr(x->data(), sample_size, N * C); ConstEigenArrayMap d_y_arr(d_y->data(), sample_size, N * C); EigenArrayMap d_x_arr(d_x->mutable_data(ctx.GetPlace()), @@ -399,7 +403,7 @@ class BatchNormGradKernel } break; } - case TensorFormat::NHWC: { + case DataLayout::kNHWC: { ConstEigenArrayMap x_arr(x->data(), C, N * sample_size); ConstEigenArrayMap d_y_arr(d_y->data(), C, N * sample_size); EigenArrayMap d_x_arr(d_x->mutable_data(ctx.GetPlace()), C, @@ -424,7 +428,7 @@ class BatchNormGradKernel break; } default: - PADDLE_THROW("Unknown storage order: %s", tensor_format_str); + PADDLE_THROW("Unknown storage order: %s", data_layout_str); } } }; @@ -435,8 +439,9 @@ class BatchNormGradKernel namespace ops = paddle::operators; REGISTER_OP(batch_norm, ops::BatchNormOp, ops::BatchNormOpMaker, batch_norm_grad, ops::BatchNormGradOp); -REGISTER_OP_CPU_KERNEL(batch_norm, - ops::BatchNormKernel); +REGISTER_OP_CPU_KERNEL( + batch_norm, + ops::BatchNormKernel); REGISTER_OP_CPU_KERNEL( batch_norm_grad, - ops::BatchNormGradKernel); + ops::BatchNormGradKernel); diff --git a/paddle/operators/batch_norm_op.cu.cc b/paddle/operators/batch_norm_op.cu.cc index 726d1ea1b8d7ced93f94bb0e5bb4df9e43b0ac7b..3d17725ab47682355b2093782848849857f9bf59 100644 --- a/paddle/operators/batch_norm_op.cu.cc +++ b/paddle/operators/batch_norm_op.cu.cc @@ -13,6 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/operators/batch_norm_op.h" +#include "paddle/framework/data_layout.h" #include #include "paddle/operators/math/math_function.h" @@ -22,44 +23,52 @@ namespace paddle { namespace operators { using Tensor = framework::Tensor; +using DataLayout = framework::DataLayout; template using CudnnDataType = platform::CudnnDataType; -void ExtractNCWHD(const framework::DDim &dims, - const TensorFormat &tensor_format, int *N, int *C, int *H, - int *W, int *D) { +void ExtractNCWHD(const framework::DDim &dims, const DataLayout &data_layout, + int *N, int *C, int *H, int *W, int *D) { *N = dims[0]; - *C = tensor_format == TensorFormat::NCHW ? dims[1] : dims[dims.size() - 1]; - *H = tensor_format == TensorFormat::NCHW ? dims[2] : dims[1]; - *W = dims.size() > 3 - ? (tensor_format == TensorFormat::NCHW ? dims[3] : dims[2]) - : 1; - *D = dims.size() > 4 - ? (tensor_format == TensorFormat::NCHW ? dims[4] : dims[3]) - : 1; + if (dims.size() == 2) { + *C = dims[1]; + *H = 1; + *W = 1; + *D = 1; + } else { + *C = data_layout == DataLayout::kNCHW ? dims[1] : dims[dims.size() - 1]; + *H = data_layout == DataLayout::kNCHW ? dims[2] : dims[1]; + *W = dims.size() > 3 + ? (data_layout == DataLayout::kNCHW ? dims[3] : dims[2]) + : 1; + *D = dims.size() > 4 + ? (data_layout == DataLayout::kNCHW ? dims[4] : dims[3]) + : 1; + } } template -class BatchNormKernel : public framework::OpKernel { +class BatchNormKernel + : public framework::OpKernel { public: void Compute(const framework::ExecutionContext &ctx) const override { PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), - "It must use GPUPlace."); + "It must use CUDAPlace."); double epsilon = static_cast(ctx.Attr("epsilon")); const float momentum = ctx.Attr("momentum"); const bool is_test = ctx.Attr("is_test"); - const std::string tensor_format_str = - ctx.Attr("tensor_format"); - const TensorFormat tensor_format = StringToTensorFormat(tensor_format_str); + const std::string data_layout_str = ctx.Attr("data_layout"); + const DataLayout data_layout = + framework::StringToDataLayout(data_layout_str); // Get the size for each dimension. // NCHW [batch_size, in_channels, in_height, in_width] const auto *x = ctx.Input("X"); const auto &x_dims = x->dims(); - PADDLE_ENFORCE(x_dims.size() >= 3 && x_dims.size() <= 5, - "The Input dim size should be between 3 and 5"); + PADDLE_ENFORCE(x_dims.size() >= 2 && x_dims.size() <= 5, + "The Input dim size should be between 2 and 5"); int N, C, H, W, D; - ExtractNCWHD(x_dims, tensor_format, &N, &C, &H, &W, &D); + ExtractNCWHD(x_dims, data_layout, &N, &C, &H, &W, &D); // ------------------- cudnn descriptors --------------------- cudnnTensorDescriptor_t data_desc_; @@ -85,7 +94,7 @@ class BatchNormKernel : public framework::OpKernel { VLOG(1) << "Setting descriptors."; std::vector dims; std::vector strides; - if (tensor_format == TensorFormat::NCHW) { + if (data_layout == DataLayout::kNCHW) { dims = {N, C, H, W, D}; strides = {C * H * W * D, H * W * D, W * D, D, 1}; } else { @@ -114,11 +123,12 @@ class BatchNormKernel : public framework::OpKernel { saved_mean->mutable_data(ctx.GetPlace()); saved_variance->mutable_data(ctx.GetPlace()); - math::SetConstant functor; - functor(ctx.device_context(), saved_mean, 0); - functor(ctx.device_context(), saved_variance, 0); + auto &dev_ctx = ctx.template device_context(); + math::SetConstant functor; + functor(dev_ctx, saved_mean, 0); + functor(dev_ctx, saved_variance, 0); - auto handle = ctx.cuda_device_context().cudnn_handle(); + auto handle = dev_ctx.cudnn_handle(); // Now, depending on whether we are running test or not, we have two paths. if (is_test) { @@ -164,26 +174,26 @@ class BatchNormKernel : public framework::OpKernel { }; template -class BatchNormGradKernel +class BatchNormGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext &ctx) const override { PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), - "It must use GPUPlace."); + "It must use CUDAPlace."); double epsilon = static_cast(ctx.Attr("epsilon")); - const std::string tensor_format_str = - ctx.Attr("tensor_format"); - const TensorFormat tensor_format = StringToTensorFormat(tensor_format_str); + const std::string data_layout_str = ctx.Attr("data_layout"); + const DataLayout data_layout = + framework::StringToDataLayout(data_layout_str); const auto *x = ctx.Input("X"); const auto *d_y = ctx.Input(framework::GradVarName("Y")); const auto *scale = ctx.Input("Scale"); const auto &x_dims = x->dims(); - PADDLE_ENFORCE(x_dims.size() >= 3 && x_dims.size() <= 5, - "The Input dim size should be between 3 and 5"); + PADDLE_ENFORCE(x_dims.size() >= 2 && x_dims.size() <= 5, + "The Input dim size should be between 2 and 5"); int N, C, H, W, D; - ExtractNCWHD(x_dims, tensor_format, &N, &C, &H, &W, &D); + ExtractNCWHD(x_dims, data_layout, &N, &C, &H, &W, &D); PADDLE_ENFORCE_EQ(scale->dims().size(), 1UL); PADDLE_ENFORCE_EQ(scale->dims()[0], C); @@ -210,7 +220,7 @@ class BatchNormGradKernel std::vector dims; std::vector strides; - if (tensor_format == TensorFormat::NCHW) { + if (data_layout == DataLayout::kNCHW) { dims = {N, C, H, W, D}; strides = {C * H * W * D, H * W * D, W * D, D, 1}; } else { @@ -237,11 +247,12 @@ class BatchNormGradKernel const void *saved_mean_data = saved_mean->template data(); const void *saved_var_data = saved_var->template data(); + auto &dev_ctx = ctx.template device_context(); CUDNN_ENFORCE(platform::dynload::cudnnBatchNormalizationBackward( - ctx.cuda_device_context().cudnn_handle(), mode_, - CudnnDataType::kOne(), CudnnDataType::kZero(), - CudnnDataType::kOne(), CudnnDataType::kZero(), data_desc_, - x->template data(), data_desc_, d_y->template data(), data_desc_, + dev_ctx.cudnn_handle(), mode_, CudnnDataType::kOne(), + CudnnDataType::kZero(), CudnnDataType::kOne(), + CudnnDataType::kZero(), data_desc_, x->template data(), + data_desc_, d_y->template data(), data_desc_, d_x->template mutable_data(ctx.GetPlace()), bn_param_desc_, scale->template data(), d_scale->template mutable_data(ctx.GetPlace()), @@ -259,8 +270,9 @@ class BatchNormGradKernel } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(batch_norm, - ops::BatchNormKernel); -REGISTER_OP_GPU_KERNEL( +REGISTER_OP_CUDA_KERNEL( + batch_norm, + ops::BatchNormKernel); +REGISTER_OP_CUDA_KERNEL( batch_norm_grad, - ops::BatchNormGradKernel); + ops::BatchNormGradKernel); diff --git a/paddle/operators/batch_norm_op.h b/paddle/operators/batch_norm_op.h index 4e80134a1acf3b4d66154453dd0ed709133d1c7c..a817ef41fc87da33ad87923c99a75ee7c3c7bbfe 100644 --- a/paddle/operators/batch_norm_op.h +++ b/paddle/operators/batch_norm_op.h @@ -19,28 +19,13 @@ limitations under the License. */ namespace paddle { namespace operators { -enum TensorFormat { - NHWC = 0, - NCHW = 1, -}; - -inline TensorFormat StringToTensorFormat(const std::string& str) { - if (str == "NHWC" || str == "nhwc") { - return TensorFormat::NHWC; - } else if (str == "NCHW" || str == "nchw") { - return TensorFormat::NCHW; - } else { - PADDLE_THROW("Unknown storage order string: %s", str); - } -} - -template +template class BatchNormKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override; }; -template +template class BatchNormGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override; diff --git a/paddle/operators/beam_search_decode_op.cc b/paddle/operators/beam_search_decode_op.cc index 3904a97d58166cfeeb2be7d2144700dbd8bc5721..72e05607b0b612807d552b4c45b58f9d9ce9c2af 100644 --- a/paddle/operators/beam_search_decode_op.cc +++ b/paddle/operators/beam_search_decode_op.cc @@ -13,10 +13,41 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/operators/beam_search_decode_op.h" +#include "paddle/platform/device_context.h" namespace paddle { namespace operators { +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), + id_tensor_(id_tensor), + score_tensor_(score_tensor) {} + + template + void operator()() const; + + const LoDTensorArray& step_ids_; + const LoDTensorArray& step_scores_; + LoDTensor* id_tensor_; + LoDTensor* score_tensor_; +}; + +template +void BeamSearchDecodeFunctor::operator()() const { + BeamSearchDecoder beam_search_decoder; + beam_search_decoder.PackAllSteps(step_ids_, step_scores_, id_tensor_, + score_tensor_); +} + +template <> +void BeamSearchDecodeFunctor::operator()() const { + PADDLE_THROW("beam search decode op does not support bool!"); +} + class BeamSearchDecodeOp : public framework::OperatorBase { public: BeamSearchDecodeOp(const std::string& type, @@ -25,7 +56,10 @@ class BeamSearchDecodeOp : public framework::OperatorBase { const framework::AttributeMap& attrs) : OperatorBase(type, inputs, outputs, attrs) {} void Run(const framework::Scope& scope, - const platform::DeviceContext& dev_ctx) const override { + const platform::Place& dev_place) const override { + platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); + auto& dev_ctx = *pool.Get(dev_place); + framework::ExecutionContext ctx(*this, scope, dev_ctx); const LoDTensorArray* ids = ctx.Input("Ids"); @@ -45,17 +79,16 @@ class BeamSearchDecodeOp : public framework::OperatorBase { LoDTensor* sentenceIds = ctx.Output("SentenceIds"); LoDTensor* sentenceScores = ctx.Output("SentenceScores"); - BeamSearchDecoder beam_search_decoder; - beam_search_decoder.PackAllSteps(*ids, *scores, sentenceIds, - sentenceScores); + framework::VisitDataType( + framework::ToDataType(scores->at(0).type()), + BeamSearchDecodeFunctor(*ids, *scores, sentenceIds, sentenceScores)); } }; class BeamSearchDecodeOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - BeamSearchDecodeOpProtoMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + BeamSearchDecodeOpProtoMaker(OpProto* proto, OpAttrChecker* op_checker) + : framework::OpProtoAndCheckerMaker(proto, op_checker) { AddInput("Ids", "(LodTensorArray)" "score of the candidate words in each step"); @@ -90,13 +123,13 @@ class BeamSearchDecodeInferShape : public framework::InferShapeBase { class BeamSearchDecodeInferVarType : public framework::VarTypeInference { public: - void operator()(const framework::OpDescBind& op_desc, - framework::BlockDescBind* block) const override { + void operator()(const framework::OpDesc& op_desc, + framework::BlockDesc* block) const override { for (auto& o : op_desc.Output("SentenceIds")) { - block->Var(o)->SetType(framework::VarDesc::LOD_TENSOR); + block->Var(o)->SetType(framework::proto::VarDesc::LOD_TENSOR); } for (auto& o : op_desc.Output("SentenceScores")) { - block->Var(o)->SetType(framework::VarDesc::LOD_TENSOR); + block->Var(o)->SetType(framework::proto::VarDesc::LOD_TENSOR); } } }; diff --git a/paddle/operators/beam_search_decode_op.h b/paddle/operators/beam_search_decode_op.h index 0f007ec22f9a66572971516a711317f348e1ec5a..3b1c6cd7a1045bfbb896725c79dc1ae2e22f43dc 100644 --- a/paddle/operators/beam_search_decode_op.h +++ b/paddle/operators/beam_search_decode_op.h @@ -232,12 +232,12 @@ void BeamSearchDecoder::ConvertSentenceVectorToLodTensor( id_tensor->set_lod(lod); id_tensor->Resize({static_cast(id_data.size())}); id_tensor->mutable_data(paddle::platform::CPUPlace()); - id_tensor->CopyFromVector(id_data, cpu_ctx); + framework::CopyFromVector(id_data, cpu_ctx, id_tensor); score_tensor->set_lod(lod); score_tensor->Resize({static_cast(score_data.size())}); score_tensor->mutable_data(paddle::platform::CPUPlace()); - score_tensor->CopyFromVector(score_data, cpu_ctx); + framework::CopyFromVector(score_data, cpu_ctx, score_tensor); } template diff --git a/paddle/operators/beam_search_op.cc b/paddle/operators/beam_search_op.cc index 17926a813d5b0b8ace6a1b20066cd0007703c696..2e0513b37a24b9737532b3a71f8f0724fbdd2c13 100644 --- a/paddle/operators/beam_search_op.cc +++ b/paddle/operators/beam_search_op.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/operators/beam_search_op.h" @@ -139,7 +139,7 @@ bool BeamSearch::NextItemSet(std::vector *items) { items->reserve(framework::product(ids.dims())); for (size_t offset = abs_lod[lod_level_][sent_offset_]; offset < abs_lod[lod_level_][sent_offset_ + 1]; offset++) { - for (int d = 0; d < instance_dim; d++) { + for (size_t d = 0; d < instance_dim; d++) { const size_t dim_offset = offset * instance_dim + d; items->emplace_back(offset, ids_data[dim_offset], scores_data[dim_offset]); @@ -153,8 +153,7 @@ bool BeamSearch::NextItemSet(std::vector *items) { class BeamSearchProtoAndCheckerMaker : public framework::OpProtoAndCheckerMaker { public: - BeamSearchProtoAndCheckerMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + BeamSearchProtoAndCheckerMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { // inputs and outputs stored in proto AddInput("pre_ids", "ids in previous step"); diff --git a/paddle/operators/beam_search_op.h b/paddle/operators/beam_search_op.h index cc556bfe42ab12d73c0eb503d033efc272b5dd68..08b551ef9bd63106ed222d3a956a912294f827ec 100644 --- a/paddle/operators/beam_search_op.h +++ b/paddle/operators/beam_search_op.h @@ -189,7 +189,7 @@ class BeamSearchOp : public framework::OperatorBase { } void Run(const framework::Scope& scope, - const platform::DeviceContext& dev_ctx) const override { + const platform::Place& dev_place) const override { LOG(INFO) << "run beam search op"; auto ids_var = scope.FindVar(Input("ids")); auto scores_var = scope.FindVar(Input("scores")); diff --git a/paddle/operators/bilinear_tensor_product_op.cc b/paddle/operators/bilinear_tensor_product_op.cc index c65ba7eb262f3aabe2c00837b79806c0b40b60fd..7640147a12d66a924f16eaf168227b6ce6a96040 100644 --- a/paddle/operators/bilinear_tensor_product_op.cc +++ b/paddle/operators/bilinear_tensor_product_op.cc @@ -65,8 +65,7 @@ class BilinearTensorProductOp : public framework::OperatorWithKernel { class BilinearTensorProductOpMaker : public framework::OpProtoAndCheckerMaker { public: - BilinearTensorProductOpMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker) + BilinearTensorProductOpMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "The first input of bilinear_tensor_product operator."); AddInput("Y", "The second input of bilinear_tensor_product operator."); @@ -77,11 +76,19 @@ class BilinearTensorProductOpMaker : public framework::OpProtoAndCheckerMaker { AddOutput("Out", "The output of bilinear_tensor_product operator."); AddComment(R"DOC( Bilinear Tensor Product operator. -Given input X and Y, a 3D tensor weight, and bias. Each column of the -output is computed by one slice i = 1, . . . , k of the tensor: +Given input X and Y, a 3D tensor Weight and a Bias. Each column of the +Output is computed by one slice $i = 1, . . . , k$ of the tensor: - M = (X W_i) \cdot Y - Out_i = \sum_i {M_i} + Bias_i +$$ +M = (X W_i) * Y \\ +Out_i = \sum_j {M_j} + Bias_i +$$ + +Where $W_i$ is the $i$-th slice of Input(Weight); + $M_j$ is the $j$-th column of $M$; + $Out_i$ is the $i$-th column of Output(Out); + $Bias_i$ is a column vector, each element of it is equal to + the $i$-th element of $Bias$; )DOC"); } @@ -151,9 +158,12 @@ REGISTER_OP(bilinear_tensor_product, ops::BilinearTensorProductOp, ops::BilinearTensorProductOpGrad); REGISTER_OP_CPU_KERNEL( bilinear_tensor_product, - ops::BilinearTensorProductKernel, - ops::BilinearTensorProductKernel); + ops::BilinearTensorProductKernel, + ops::BilinearTensorProductKernel); REGISTER_OP_CPU_KERNEL( bilinear_tensor_product_grad, - ops::BilinearTensorProductGradKernel, - ops::BilinearTensorProductGradKernel); + ops::BilinearTensorProductGradKernel, + ops::BilinearTensorProductGradKernel); diff --git a/paddle/operators/bilinear_tensor_product_op.cu b/paddle/operators/bilinear_tensor_product_op.cu index 858d2668d01379afe8082cd1eda32a2a5d09bd18..0f48010716f086a64c0b6a35b76e06a42430ab84 100644 --- a/paddle/operators/bilinear_tensor_product_op.cu +++ b/paddle/operators/bilinear_tensor_product_op.cu @@ -16,11 +16,15 @@ limitations under the License. */ #include "paddle/operators/bilinear_tensor_product_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL( +REGISTER_OP_CUDA_KERNEL( bilinear_tensor_product, - ops::BilinearTensorProductKernel, - ops::BilinearTensorProductKernel); -REGISTER_OP_GPU_KERNEL( + ops::BilinearTensorProductKernel, + ops::BilinearTensorProductKernel); +REGISTER_OP_CUDA_KERNEL( bilinear_tensor_product_grad, - ops::BilinearTensorProductGradKernel, - ops::BilinearTensorProductGradKernel); + ops::BilinearTensorProductGradKernel, + ops::BilinearTensorProductGradKernel); diff --git a/paddle/operators/bilinear_tensor_product_op.h b/paddle/operators/bilinear_tensor_product_op.h index 1113a4c6f357edb4f6b14b73c6eec9c6cca24ce5..ba9a2c5ce3c024a82e864a399ad90281d8dcdb20 100644 --- a/paddle/operators/bilinear_tensor_product_op.h +++ b/paddle/operators/bilinear_tensor_product_op.h @@ -27,7 +27,7 @@ template using EigenMatrix = framework::EigenMatrix; -template +template class BilinearTensorProductKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { @@ -46,7 +46,8 @@ class BilinearTensorProductKernel : public framework::OpKernel { int out_dim = weight_dims[0]; auto x_dim = weight_dims[1]; auto y_dim = weight_dims[2]; - auto place = ctx.GetEigenDevice(); + auto& place = *ctx.template device_context().eigen_device(); + auto& dev_ctx = ctx.template device_context(); // Create the intermediate variable to caculate the result of // Input(X) multiplied by Input(Weight_i), the formula is: @@ -60,9 +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(ctx.device_context(), CblasNoTrans, CblasNoTrans, - batch_size, y_dim, x_dim, 1, x->data(), - weight_mat.data(), 0, left_mul.data()); + math::gemm(dev_ctx, 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)); } @@ -74,7 +75,7 @@ class BilinearTensorProductKernel : public framework::OpKernel { } }; -template +template class BilinearTensorProductGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { @@ -96,8 +97,8 @@ class BilinearTensorProductGradKernel : public framework::OpKernel { auto x_mat = EigenMatrix::From(*x); auto y_mat = EigenMatrix::From(*y); auto d_out_mat = EigenMatrix::From(*d_out); - auto place = ctx.GetEigenDevice(); - + auto& place = *ctx.template device_context().eigen_device(); + auto& dev_ctx = ctx.template device_context(); // Create the intermediate variable to caculate the Output(Y@Grad). Tensor x_scale; x_scale.mutable_data(framework::make_ddim({batch_size, x_dim}), @@ -110,18 +111,18 @@ class BilinearTensorProductGradKernel : public framework::OpKernel { ctx.GetPlace()); auto y_scale_mat = EigenMatrix::From(y_scale); - math::SetConstant set_zero; + math::SetConstant set_zero; // Set Output(X@Grad) be zero. if (d_x) { d_x->mutable_data(ctx.GetPlace()); - set_zero(ctx.device_context(), d_x, static_cast(0)); + set_zero(dev_ctx, d_x, static_cast(0)); } // Set Output(Y@Grad) be zero. if (d_y) { d_y->mutable_data(ctx.GetPlace()); - set_zero(ctx.device_context(), d_y, static_cast(0)); + set_zero(dev_ctx, d_y, static_cast(0)); } // Caculate the Output(X@Grad) and Output(Y@Grad). @@ -137,18 +138,18 @@ class BilinearTensorProductGradKernel : public framework::OpKernel { output_vec.reshape(Eigen::DSizes(batch_size, 1)) .broadcast(bcast_for_x) * y_mat; - math::gemm(ctx.device_context(), CblasNoTrans, CblasTrans, - batch_size, x_dim, y_dim, 1, y_scale.data(), - weight_i.data(), 1, d_x->data()); + math::gemm( + dev_ctx, 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(ctx.device_context(), CblasNoTrans, CblasNoTrans, - batch_size, y_dim, x_dim, 1, x_scale.data(), - weight_i.data(), 1, d_y->data()); + math::gemm( + dev_ctx, CblasNoTrans, CblasNoTrans, batch_size, y_dim, x_dim, 1, + x_scale.data(), weight_i.data(), 1, d_y->data()); } } } @@ -165,9 +166,9 @@ class BilinearTensorProductGradKernel : public framework::OpKernel { output_vec.reshape(Eigen::DSizes(batch_size, 1)) .broadcast(bcast_for_weight) * x_mat; - math::gemm(ctx.device_context(), CblasTrans, CblasNoTrans, - x_dim, y_dim, batch_size, 1, x_scale.data(), - y->data(), 0, d_weight_i.data()); + math::gemm(dev_ctx, CblasTrans, CblasNoTrans, x_dim, + y_dim, batch_size, 1, x_scale.data(), + y->data(), 0, d_weight_i.data()); } } diff --git a/paddle/operators/cast_op.cc b/paddle/operators/cast_op.cc index 70ee7861bab3a982eae60dd85b10c2e41f5827d0..446976edafca56f3c56fe573c8b5ef76a333089f 100644 --- a/paddle/operators/cast_op.cc +++ b/paddle/operators/cast_op.cc @@ -1,16 +1,16 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +/* 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 +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 + 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. */ +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/operators/cast_op.h" #include "paddle/framework/op_registry.h" @@ -20,13 +20,12 @@ namespace operators { class CastOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - CastOpProtoMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + CastOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "The input tensor of cast op"); AddOutput("Out", "The output tensor of cast op"); - AddAttr("out_data_type", "output data type"); - AddAttr("in_data_type", "input data type"); + AddAttr("out_dtype", "output data type"); + AddAttr("in_dtype", "input data type"); AddComment(R"DOC( Cast Operator. @@ -53,14 +52,14 @@ class CastOpGradMaker : public framework::SingleGradOpDescMaker { using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; protected: - std::unique_ptr Apply() const override { - auto grad = new framework::OpDescBind(); + std::unique_ptr Apply() const override { + auto grad = new framework::OpDesc(); grad->SetType("cast"); grad->SetInput("X", OutputGrad("Out")); grad->SetOutput("Out", InputGrad("X")); - grad->SetAttr("out_data_type", GetAttr("in_data_type")); - grad->SetAttr("in_data_type", GetAttr("out_data_type")); - return std::unique_ptr(grad); + grad->SetAttr("out_dtype", GetAttr("in_dtype")); + grad->SetAttr("in_dtype", GetAttr("out_dtype")); + return std::unique_ptr(grad); } }; @@ -68,10 +67,11 @@ class CastOpGradMaker : public framework::SingleGradOpDescMaker { } // namespace paddle namespace ops = paddle::operators; -using CPU = paddle::platform::CPUPlace; +using CPU = paddle::platform::CPUDeviceContext; REGISTER_OP_WITH_KERNEL(cast, ops::CastOpGradMaker, ops::CastOpInferShape, ops::CastOpProtoMaker); REGISTER_OP_CPU_KERNEL(cast, ops::CastOpKernel, ops::CastOpKernel, ops::CastOpKernel, - ops::CastOpKernel); + ops::CastOpKernel, + ops::CastOpKernel); diff --git a/paddle/operators/cast_op.cu b/paddle/operators/cast_op.cu index fb75ddbabfefd8d00420d8c96f958abcb8fdce62..d68bbe6e39a2fbaa92787731145ae324288b981a 100644 --- a/paddle/operators/cast_op.cu +++ b/paddle/operators/cast_op.cu @@ -1,22 +1,23 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +/* 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 +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 + 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. */ +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/operators/cast_op.h" template using CastOpKernel = - paddle::operators::CastOpKernel; + paddle::operators::CastOpKernel; -REGISTER_OP_GPU_KERNEL(cast, CastOpKernel, CastOpKernel, - CastOpKernel, CastOpKernel); +REGISTER_OP_CUDA_KERNEL(cast, CastOpKernel, CastOpKernel, + CastOpKernel, CastOpKernel, + CastOpKernel); diff --git a/paddle/operators/cast_op.h b/paddle/operators/cast_op.h index ffdbff7030afedab2efc06479ac86ad70c185f48..9f39d91edd49d236d74019ca81b42002e4f35d36 100644 --- a/paddle/operators/cast_op.h +++ b/paddle/operators/cast_op.h @@ -1,16 +1,16 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +/* 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 +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 + 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. */ +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 @@ -27,13 +27,13 @@ struct CastOpTransformFunctor { HOSTDEVICE OutT operator()(InT in) const { return static_cast(in); } }; -template +template struct CastOpFunctor { const framework::Tensor* in_; framework::Tensor* out_; - const platform::DeviceContext& ctx_; + const DeviceContext& ctx_; CastOpFunctor(const framework::Tensor* in, framework::Tensor* out, - const platform::DeviceContext& ctx) + const DeviceContext& ctx) : in_(in), out_(out), ctx_(ctx) {} template @@ -42,21 +42,22 @@ struct CastOpFunctor { auto numel = in_->numel(); auto* in_end = in_begin + numel; auto* out_begin = out_->mutable_data(ctx_.GetPlace()); - platform::Transform trans; + platform::Transform trans; trans(ctx_, in_begin, in_end, out_begin, CastOpTransformFunctor()); } }; -template +template class CastOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto* in = context.Input("X"); auto* out = context.Output("Out"); framework::VisitDataType( - static_cast(context.Attr("out_data_type")), - CastOpFunctor(in, out, context.device_context())); + static_cast(context.Attr("out_dtype")), + CastOpFunctor( + in, out, context.template device_context())); } }; diff --git a/paddle/operators/chunk_eval_op.cc b/paddle/operators/chunk_eval_op.cc index 309660b01fe7052de2f9300acdf00779d0228221..a040404266c3cd44230b141cfed1aaede3f05187 100644 --- a/paddle/operators/chunk_eval_op.cc +++ b/paddle/operators/chunk_eval_op.cc @@ -32,6 +32,13 @@ class ChunkEvalOp : public framework::OperatorWithKernel { "Output(Recall) of ChunkEvalOp should not be null."); PADDLE_ENFORCE(ctx->HasOutput("F1-Score"), "Output(F1-Score) of ChunkEvalOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("NumInferChunks"), + "Output(NumInferChunks) of ChunkEvalOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("NumLabelChunks"), + "Output(NumLabelChunks) of ChunkEvalOp should not be null."); + PADDLE_ENFORCE( + ctx->HasOutput("NumCorrectChunks"), + "Output(NumCorrectChunks) of ChunkEvalOp should not be null."); auto inference_dim = ctx->GetInputDim("Inference"); auto label_dim = ctx->GetInputDim("Label"); @@ -42,25 +49,28 @@ class ChunkEvalOp : public framework::OperatorWithKernel { ctx->SetOutputDim("Precision", {1}); ctx->SetOutputDim("Recall", {1}); ctx->SetOutputDim("F1-Score", {1}); + ctx->SetOutputDim("NumInferChunks", {1}); + ctx->SetOutputDim("NumLabelChunks", {1}); + ctx->SetOutputDim("NumCorrectChunks", {1}); } protected: - framework::OpKernelType GetKernelType( + framework::OpKernelType GetActualKernelType( const framework::ExecutionContext &ctx) const override { - return framework::OpKernelType(framework::DataType::FP32, + return framework::OpKernelType(framework::proto::DataType::FP32, ctx.device_context()); } }; class ChunkEvalOpMaker : public framework::OpProtoAndCheckerMaker { public: - ChunkEvalOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + ChunkEvalOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("Inference", - "(Tensor, default: Tensor). Predictions from the network."); + "(Tensor, default: Tensor). " + "Predictions from the network."); AddInput("Label", - "(Tensor, default: Tensor). The true tag sequences."); + "(Tensor, default: Tensor). The true tag sequences."); AddOutput("Precision", "(float). The evaluated precision (called positive predictive " "value) of chunks on the given mini-batch."); @@ -69,6 +79,16 @@ class ChunkEvalOpMaker : public framework::OpProtoAndCheckerMaker { "sensitivity) of chunks on the given mini-batch."); AddOutput("F1-Score", "(float). The evaluated F1-Score on the given mini-batch."); + AddOutput("NumInferChunks", + "(int64_t). The number of chunks in Inference on the given " + "mini-batch."); + AddOutput( + "NumLabelChunks", + "(int64_t). The number of chunks in Label on the given mini-batch."); + AddOutput( + "NumCorrectChunks", + "(int64_t). The number of chunks both in Inference and Label on the " + "given mini-batch."); AddAttr("num_chunk_types", "(int). The number of chunk type. See below for details."); AddAttr( @@ -84,7 +104,7 @@ class ChunkEvalOpMaker : public framework::OpProtoAndCheckerMaker { .SetDefault(std::vector{}); AddComment(R"DOC( For some basics of chunking, please refer to -‘Chunking with Support Vector Mechines ’. +‘Chunking with Support Vector Machines ’. CheckEvalOp computes the precision, recall, and F1-score of chunk detection, @@ -97,7 +117,7 @@ Here is a NER example of labeling for these tagging schemes: IOE: I-PER E-PER O O I-ORG I-ORG I-ORG E-ORG O E-LOC IOBES: B-PER E-PER O O I-ORG I-ORG I-ORG E-ORG O S-LOC -There are three chunk types(named entity types) including PER(person), ORG(orgnazation) +There are three chunk types(named entity types) including PER(person), ORG(organization) and LOC(LOCATION), and we can see that the labels have the form -. Since the calculations actually use label ids rather than labels, extra attention diff --git a/paddle/operators/chunk_eval_op.h b/paddle/operators/chunk_eval_op.h index 81aa07817b673b2ff85a35a51cc43742b7ad7fed..74ab435c860b22b2ee3f485743540976a7a31b96 100644 --- a/paddle/operators/chunk_eval_op.h +++ b/paddle/operators/chunk_eval_op.h @@ -23,7 +23,7 @@ namespace operators { using Tensor = framework::Tensor; using LoDTensor = framework::LoDTensor; -template +template class ChunkEvalKernel : public framework::OpKernel { public: struct Segment { @@ -35,10 +35,10 @@ class ChunkEvalKernel : public framework::OpKernel { } }; - void GetSegments(const int* label, int length, std::vector& segments, - int num_chunk_types, int num_tag_types, int other_chunk_type, - int tag_begin, int tag_inside, int tag_end, - int tag_single) const { + void GetSegments(const int64_t* label, int length, + std::vector& segments, int num_chunk_types, + int num_tag_types, int other_chunk_type, int tag_begin, + int tag_inside, int tag_end, int tag_single) const { segments.clear(); segments.reserve(length); int chunk_start = 0; @@ -111,9 +111,7 @@ class ChunkEvalKernel : public framework::OpKernel { std::vector label_segments; std::vector output_segments; std::set excluded_chunk_types; - int64_t num_output_segments = 0; - int64_t num_label_segments = 0; - int64_t num_correct = 0; + if (context.Attr("chunk_scheme") == "IOB") { num_tag_types = 2; tag_begin = 0; @@ -151,12 +149,24 @@ class ChunkEvalKernel : public framework::OpKernel { auto* precision = context.Output("Precision"); auto* recall = context.Output("Recall"); auto* f1 = context.Output("F1-Score"); + auto* num_infer_chunks = context.Output("NumInferChunks"); + auto* num_label_chunks = context.Output("NumLabelChunks"); + auto* num_correct_chunks = context.Output("NumCorrectChunks"); - const int* inference_data = inference->data(); - const int* label_data = label->data(); + const int64_t* inference_data = inference->data(); + const int64_t* label_data = label->data(); T* precision_data = precision->mutable_data(context.GetPlace()); T* racall_data = recall->mutable_data(context.GetPlace()); T* f1_data = f1->mutable_data(context.GetPlace()); + int64_t* num_infer_chunks_data = + num_infer_chunks->mutable_data(context.GetPlace()); + int64_t* num_label_chunks_data = + num_label_chunks->mutable_data(context.GetPlace()); + int64_t* num_correct_chunks_data = + num_correct_chunks->mutable_data(context.GetPlace()); + *num_infer_chunks_data = 0; + *num_label_chunks_data = 0; + *num_correct_chunks_data = 0; auto lod = label->lod(); PADDLE_ENFORCE_EQ(lod.size(), 1UL, "Only support one level sequence now."); @@ -166,20 +176,26 @@ class ChunkEvalKernel : public framework::OpKernel { for (int i = 0; i < num_sequences; ++i) { int seq_length = lod[0][i + 1] - lod[0][i]; EvalOneSeq(inference_data + lod[0][i], label_data + lod[0][i], seq_length, - output_segments, label_segments, num_output_segments, - num_label_segments, num_correct, num_chunk_types, - num_tag_types, other_chunk_type, tag_begin, tag_inside, - tag_end, tag_single, excluded_chunk_types); + output_segments, label_segments, *num_infer_chunks_data, + *num_label_chunks_data, *num_correct_chunks_data, + num_chunk_types, num_tag_types, other_chunk_type, tag_begin, + tag_inside, tag_end, tag_single, excluded_chunk_types); } - *precision_data = !num_output_segments ? 0 : static_cast(num_correct) / - num_output_segments; - *racall_data = !num_label_segments ? 0 : static_cast(num_correct) / - num_label_segments; - *f1_data = !num_correct ? 0 : 2 * (*precision_data) * (*racall_data) / - ((*precision_data) + (*racall_data)); + *precision_data = !(*num_infer_chunks_data) + ? 0 + : static_cast(*num_correct_chunks_data) / + (*num_infer_chunks_data); + *racall_data = !(*num_label_chunks_data) + ? 0 + : static_cast(*num_correct_chunks_data) / + (*num_label_chunks_data); + *f1_data = !(*num_correct_chunks_data) + ? 0 + : 2 * (*precision_data) * (*racall_data) / + ((*precision_data) + (*racall_data)); } - void EvalOneSeq(const int* output, const int* label, int length, + void EvalOneSeq(const int64_t* output, const int64_t* label, int length, std::vector& output_segments, std::vector& label_segments, int64_t& num_output_segments, int64_t& num_label_segments, diff --git a/paddle/operators/clip_by_norm_op.cc b/paddle/operators/clip_by_norm_op.cc index d9fc532e39500fa397be80396b075e866bad9362..b90921d79baa920f0b6f92cde2f7e1ca9183d0d2 100644 --- a/paddle/operators/clip_by_norm_op.cc +++ b/paddle/operators/clip_by_norm_op.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/operators/clip_by_norm_op.h" @@ -37,8 +37,7 @@ class ClipByNormOp : public framework::OperatorWithKernel { class ClipByNormOpMaker : public framework::OpProtoAndCheckerMaker { public: - ClipByNormOpMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker) + ClipByNormOpMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "(Tensor) The input of clip_by_norm op." @@ -47,15 +46,19 @@ class ClipByNormOpMaker : public framework::OpProtoAndCheckerMaker { "(Tensor) The output of clip_by_norm op with shape as input(X)"); AddAttr("max_norm", "(float) The maximum norm value."); AddComment(R"DOC( -ClipByNorm operator limits the L2 norm of the input 'X' within 'max_norm'. -If the L2 norm of 'X' is less than or equal to 'max_norm', 'Out' will be -the same as 'X'. If the L2 norm of 'X' is greater than 'max_norm', 'X' will -be linearly scaled to make the L2 norm of 'Out' equal to 'max_norm', as -shown in the following formula： +ClipByNorm Operator. -'Out' = 'max_norm' * 'X' / norm('X'), +This operator limits the L2 norm of the input $X$ within $max\_norm$. +If the L2 norm of $X$ is less than or equal to $max\_norm$, $Out$ will be +the same as $X$. If the L2 norm of $X$ is greater than $max\_norm$, $X$ will +be linearly scaled to make the L2 norm of $Out$ equal to $max\_norm$, as +shown in the following formula: -where norm('X') represents the L2 norm of 'X'. +$$ +Out = \frac{max\_norm * X}{norm(X)}, +$$ + +where $norm(X)$ represents the L2 norm of $X$. )DOC"); } }; @@ -67,4 +70,5 @@ namespace ops = paddle::operators; REGISTER_OP_WITHOUT_GRADIENT(clip_by_norm, ops::ClipByNormOp, ops::ClipByNormOpMaker); REGISTER_OP_CPU_KERNEL( - clip_by_norm, ops::ClipByNormKernel); + clip_by_norm, + ops::ClipByNormKernel); diff --git a/paddle/operators/clip_by_norm_op.cu b/paddle/operators/clip_by_norm_op.cu index 2593a24ebbf56ecd286a726e527d2414247576e8..cbf8fa44133739f948fed13e18fc5cbaabd3abb0 100644 --- a/paddle/operators/clip_by_norm_op.cu +++ b/paddle/operators/clip_by_norm_op.cu @@ -1,19 +1,20 @@ /* 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 +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 + 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. */ +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/operators/clip_by_norm_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL( - clip_by_norm, ops::ClipByNormKernel); +REGISTER_OP_CUDA_KERNEL( + clip_by_norm, + ops::ClipByNormKernel); diff --git a/paddle/operators/clip_by_norm_op.h b/paddle/operators/clip_by_norm_op.h index b26476cae9b5b2fa290bc9186b9a64c48ba703d6..87956a707cf58afa2336602b8ab6acf73b0ff814 100644 --- a/paddle/operators/clip_by_norm_op.h +++ b/paddle/operators/clip_by_norm_op.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 @@ -26,7 +26,7 @@ template using EigenVector = framework::EigenVector; -template +template class ClipByNormKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { @@ -38,7 +38,8 @@ class ClipByNormKernel : public framework::OpKernel { auto x = EigenVector::Flatten(*input); auto out = EigenVector::Flatten(*output); auto x_norm = x.square().sum().sqrt(); - auto place = context.GetEigenDevice(); + auto& place = + *context.template device_context().eigen_device(); auto temp = (x_norm <= max_norm).template cast().eval(); auto scaling = temp + (static_cast(1) - temp) * max_norm / x_norm; diff --git a/paddle/operators/clip_op.cc b/paddle/operators/clip_op.cc index 3e9066ceb2a4a4dc19fdf5ef02bb7fadaab4bfff..573bb9c7dfdac2366c2458dd9f27a035a9f9b813 100644 --- a/paddle/operators/clip_op.cc +++ b/paddle/operators/clip_op.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/operators/clip_op.h" @@ -38,7 +38,7 @@ class ClipOp : public framework::OperatorWithKernel { template class ClipOpMaker : public framework::OpProtoAndCheckerMaker { public: - ClipOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker) + ClipOpMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "(Tensor)The input of clip op." @@ -52,7 +52,11 @@ class ClipOpMaker : public framework::OpProtoAndCheckerMaker { Clip Operator. The clip operator limits the value of given input within an interval. The interval is -specified with arguments 'min' and 'max'. +specified with arguments 'min' and 'max': + +$$ +Out = \min(\max(X, min), max) +$$ )DOC"); } @@ -79,7 +83,7 @@ class ClipOpGrad : public framework::OperatorWithKernel { 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); +REGISTER_OP_CPU_KERNEL( + clip, ops::ClipKernel); +REGISTER_OP_CPU_KERNEL( + clip_grad, ops::ClipGradKernel); diff --git a/paddle/operators/clip_op.cu b/paddle/operators/clip_op.cu index ca9701298fdae3fabe234925edaf9e4d775cc66e..5ccbc9643407c65e8734711744ceac9814f4c6a2 100644 --- a/paddle/operators/clip_op.cu +++ b/paddle/operators/clip_op.cu @@ -1,21 +1,21 @@ /* 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 +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 + 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. */ +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/operators/clip_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(clip, - ops::ClipKernel); -REGISTER_OP_GPU_KERNEL(clip_grad, - ops::ClipGradKernel); +REGISTER_OP_CUDA_KERNEL( + clip, ops::ClipKernel); +REGISTER_OP_CUDA_KERNEL( + clip_grad, ops::ClipGradKernel); diff --git a/paddle/operators/clip_op.h b/paddle/operators/clip_op.h index ac702e9935201ba5263a80ebeb1ab22fa0bd1340..51db185dffd80cc3b839d063acaf3f936d732817 100644 --- a/paddle/operators/clip_op.h +++ b/paddle/operators/clip_op.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 @@ -55,7 +55,7 @@ class ClipGradFunctor { T max_; }; -template +template class ClipKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { @@ -66,13 +66,13 @@ class ClipKernel : public framework::OpKernel { T* out_data = out->mutable_data(context.GetPlace()); const T* x_data = x->data(); int64_t numel = x->numel(); - Transform trans; - trans(context.device_context(), x_data, x_data + numel, out_data, - ClipFunctor(min, max)); + Transform trans; + trans(context.template device_context(), x_data, + x_data + numel, out_data, ClipFunctor(min, max)); } }; -template +template class ClipGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { @@ -86,9 +86,9 @@ class ClipGradKernel : public framework::OpKernel { auto* d_x_data = d_x->mutable_data(context.GetPlace()); const T* d_out_data = d_out->data(); const T* x_data = x->data(); - Transform trans; - trans(context.device_context(), d_out_data, d_out_data + numel, x_data, - d_x_data, ClipGradFunctor(min, max)); + Transform trans; + trans(context.template device_context(), d_out_data, + d_out_data + numel, x_data, d_x_data, ClipGradFunctor(min, max)); } } }; diff --git a/paddle/operators/compare_op.cc b/paddle/operators/compare_op.cc index bf7e88368157d29e627c3c06384f28b6e5e4ecc1..44665b7872acab8178552e5504916408cf566d13 100644 --- a/paddle/operators/compare_op.cc +++ b/paddle/operators/compare_op.cc @@ -1,16 +1,16 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +/* 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 +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 + 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. */ +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/operators/compare_op.h" #include "paddle/framework/op_registry.h" @@ -20,8 +20,7 @@ namespace operators { template class CompareOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - CompareOpProtoMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + CompareOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { OpComment comment; AddInput("X", @@ -67,9 +66,9 @@ class CompareOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - framework::OpKernelType GetKernelType( + framework::OpKernelType GetActualKernelType( const framework::ExecutionContext &ctx) const override { - framework::OpKernelType kt = OperatorWithKernel::GetKernelType(ctx); + framework::OpKernelType kt = OperatorWithKernel::GetActualKernelType(ctx); // CompareOp kernel's device type is decided by input tensor place kt.place_ = ctx.Input("X")->place(); return kt; diff --git a/paddle/operators/compare_op.cu b/paddle/operators/compare_op.cu index 6ac8c124b9b2e7c808808ecc8802a2e5aeaa5b5d..26049271befd1fe57001659d1a406e73de0004a7 100644 --- a/paddle/operators/compare_op.cu +++ b/paddle/operators/compare_op.cu @@ -1,23 +1,23 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +/* 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 +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 + 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. */ +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/operators/compare_op.h" -REGISTER_LOGICAL_KERNEL(less_than, GPU, paddle::operators::LessThanFunctor); -REGISTER_LOGICAL_KERNEL(less_equal, GPU, paddle::operators::LessEqualFunctor); -REGISTER_LOGICAL_KERNEL(greater_than, GPU, +REGISTER_LOGICAL_KERNEL(less_than, CUDA, paddle::operators::LessThanFunctor); +REGISTER_LOGICAL_KERNEL(less_equal, CUDA, paddle::operators::LessEqualFunctor); +REGISTER_LOGICAL_KERNEL(greater_than, CUDA, paddle::operators::GreaterThanFunctor); -REGISTER_LOGICAL_KERNEL(greater_equal, GPU, +REGISTER_LOGICAL_KERNEL(greater_equal, CUDA, paddle::operators::GreaterEqualFunctor); -REGISTER_LOGICAL_KERNEL(equal, GPU, paddle::operators::EqualFunctor); +REGISTER_LOGICAL_KERNEL(equal, CUDA, paddle::operators::EqualFunctor); diff --git a/paddle/operators/compare_op.h b/paddle/operators/compare_op.h index afdf3ab3e098b4e7f4c996471617d97ec49264b1..567e89c0a727ad0cdd2add8ec8b2a42c86a58007 100644 --- a/paddle/operators/compare_op.h +++ b/paddle/operators/compare_op.h @@ -1,16 +1,16 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +/* 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 +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 + 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. */ +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 @@ -59,7 +59,7 @@ struct EqualFunctor { } }; -template +template class CompareOpKernel : public framework::OpKernel { public: @@ -69,24 +69,23 @@ class CompareOpKernel auto* y = context.Input("Y"); auto* out = context.Output("Out"); Functor binary_func; - platform::Transform trans; - trans(context.device_context(), x->data(), x->data() + x->numel(), - y->data(), out->mutable_data(context.GetPlace()), - binary_func); + platform::Transform trans; + trans(context.template device_context(), x->data(), + x->data() + x->numel(), y->data(), + out->mutable_data(context.GetPlace()), binary_func); } }; } // namespace operators } // namespace paddle -#define REGISTER_LOGICAL_KERNEL(op_type, dev, functor) \ - REGISTER_OP_##dev##_KERNEL( \ - op_type, \ - ::paddle::operators::CompareOpKernel<::paddle::platform::dev##Place, \ - functor>, \ - ::paddle::operators::CompareOpKernel<::paddle::platform::dev##Place, \ - functor>, \ - ::paddle::operators::CompareOpKernel<::paddle::platform::dev##Place, \ - functor>, \ - ::paddle::operators::CompareOpKernel<::paddle::platform::dev##Place, \ - functor>); +#define REGISTER_LOGICAL_KERNEL(op_type, dev, functor) \ + REGISTER_OP_##dev##_KERNEL( \ + op_type, ::paddle::operators::CompareOpKernel< \ + ::paddle::platform::dev##DeviceContext, functor>, \ + ::paddle::operators::CompareOpKernel< \ + ::paddle::platform::dev##DeviceContext, functor>, \ + ::paddle::operators::CompareOpKernel< \ + ::paddle::platform::dev##DeviceContext, functor>, \ + ::paddle::operators::CompareOpKernel< \ + ::paddle::platform::dev##DeviceContext, functor>); diff --git a/paddle/operators/concat_op.cc b/paddle/operators/concat_op.cc index 5f052689251bc023df635d41c1e64a660a0aa488..32b61edfd0dd163e5ef8f3d1de133c55314458b5 100644 --- a/paddle/operators/concat_op.cc +++ b/paddle/operators/concat_op.cc @@ -25,7 +25,7 @@ class ConcatOp : public framework::OperatorWithKernel { void InferShape(framework::InferShapeContext *ctx) const override { PADDLE_ENFORCE_GE(ctx->Inputs("X").size(), 1UL, - "Inputs(X) of ConcatOp should be empty.") + "Inputs(X) of ConcatOp should be empty."); PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) of ConcatOp should not be null."); @@ -41,20 +41,24 @@ class ConcatOp : public framework::OperatorWithKernel { for (size_t j = 0; j < in_zero_dims_size; j++) { if (j == axis) { out_dims[axis] += ins[i][j]; - continue; + } else { + PADDLE_ENFORCE_EQ(out_dims[j], ins[i][j], + "Input tensors should have the same " + "elements except the specify axis."); } - PADDLE_ENFORCE_EQ(out_dims[j], ins[i][j], - "Input tensors should have the same " - "elements except the specify axis.") } } + if (out_dims[axis] < 0) { + out_dims[axis] = -1; + } ctx->SetOutputDim("Out", out_dims); + ctx->ShareLoD("X", /*->*/ "Out"); } }; class ConcatOpMaker : public framework::OpProtoAndCheckerMaker { public: - ConcatOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + ConcatOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "Input tensors of concat operator.").AsDuplicable(); AddOutput("Out", "Output tensor of concat operator."); @@ -94,8 +98,8 @@ class ConcatOpGrad : public framework::OperatorWithKernel { } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP(concat, ops::ConcatOp, ops::ConcatOpMaker, concat_grad, - ops::ConcatOpGrad) +REGISTER_OP_EX(concat, ops::ConcatOp, ops::ConcatOpMaker, concat_grad, + ops::ConcatOpGrad, false) REGISTER_OP_CPU_KERNEL(concat, ops::ConcatKernel) REGISTER_OP_CPU_KERNEL(concat_grad, diff --git a/paddle/operators/concat_op.cu.cc b/paddle/operators/concat_op.cu.cc index ede832ddcd486729db56bba016683b33875f8837..7b46452d3d5db58799923a3dc76bb9df3471d9e7 100644 --- a/paddle/operators/concat_op.cu.cc +++ b/paddle/operators/concat_op.cu.cc @@ -14,7 +14,8 @@ limitations under the License. */ #include "paddle/operators/concat_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(concat, - ops::ConcatKernel); -REGISTER_OP_GPU_KERNEL( - concat_grad, ops::ConcatGradKernel); +REGISTER_OP_CUDA_KERNEL( + concat, ops::ConcatKernel); +REGISTER_OP_CUDA_KERNEL( + concat_grad, + ops::ConcatGradKernel); diff --git a/paddle/operators/concat_op.h b/paddle/operators/concat_op.h index c113f19fb5cf806709bff845ee0f1078b34014bb..de4011585af81363368a096a5c361ff3f7aeecdb 100644 --- a/paddle/operators/concat_op.h +++ b/paddle/operators/concat_op.h @@ -21,7 +21,7 @@ limitations under the License. */ namespace paddle { namespace operators { -template +template class ConcatKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { @@ -43,7 +43,7 @@ class ConcatKernel : public framework::OpKernel { } }; -template +template class ConcatGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const { diff --git a/paddle/operators/cond_op.cc b/paddle/operators/cond_op.cc index b809bdc3a0fea727f2fb6ea0a55672ee9b0bbd04..e333002bfd1ab40c62882f09cd207a12a0939648 100644 --- a/paddle/operators/cond_op.cc +++ b/paddle/operators/cond_op.cc @@ -13,9 +13,9 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/operators/cond_op.h" - #include "paddle/operators/gather.h" #include "paddle/operators/scatter.h" +#include "paddle/platform/device_context.h" namespace paddle { namespace operators { @@ -193,20 +193,22 @@ void CondOp::MergeDataFromSubnet(const framework::Scope& scope, } } -void CondOp::Run(const Scope& scope, - const platform::DeviceContext& dev_ctx) const { +void CondOp::Run(const Scope& scope, const platform::Place& place) const { + // get device context from pool + platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); + auto& dev_ctx = *pool.Get(place); + PrepareDataForSubnet(scope, dev_ctx); std::vector& sub_scopes = GetSubScopes(scope); for (int i = 0; i < BRANCH_NUM; ++i) { - sub_net_op_[i]->Run(*sub_scopes[i], dev_ctx); + sub_net_op_[i]->Run(*sub_scopes[i], place); } MergeDataFromSubnet(scope, dev_ctx); } class CondOpProtoAndCheckerMaker : public framework::OpProtoAndCheckerMaker { public: - CondOpProtoAndCheckerMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker) + CondOpProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("Cond", "The condition, which is a bool vector"); AddInput("Xs", "Inputs of Subnets").AsDuplicable(); diff --git a/paddle/operators/cond_op.h b/paddle/operators/cond_op.h index 93121fb31be287794249b5a62386d5a8dd268a0c..7dcdc47e0b2ff216bea92d083fe5897009384d39 100644 --- a/paddle/operators/cond_op.h +++ b/paddle/operators/cond_op.h @@ -78,7 +78,7 @@ class CondOp : public framework::OperatorBase { } void Run(const framework::Scope& scope, - const platform::DeviceContext& dev_ctx) const override; + const platform::Place& place) const override; private: const int TRUE_BRANCH = 0; diff --git a/paddle/operators/conditional_block_op.cc b/paddle/operators/conditional_block_op.cc index d5b124682d755ffb39f32c9f001a3cf113a01a2c..3cae61a438431e72cb24d714c761676cc0c3a41f 100644 --- a/paddle/operators/conditional_block_op.cc +++ b/paddle/operators/conditional_block_op.cc @@ -1,16 +1,16 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +/* 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 +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 + 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. */ +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/framework/executor.h" #include "paddle/framework/op_registry.h" @@ -51,7 +51,7 @@ class ConditionalBlockOp : public ConditionalOp { const framework::AttributeMap &attrs) : ConditionalOp(type, inputs, outputs, attrs) {} void Run(const framework::Scope &scope, - const platform::DeviceContext &dev_ctx) const override { + const platform::Place &dev_place) const override { auto xs = InputTensors(scope); bool need_run = std::all_of( xs.begin(), xs.end(), @@ -65,8 +65,8 @@ class ConditionalBlockOp : public ConditionalOp { scopes->front() = &scope.NewScope(); auto &cur_scope = *scopes->front(); - auto *block = Attr("block"); - framework::Executor exec(dev_ctx); + framework::Executor exec(dev_place); + auto *block = Attr("sub_block"); exec.Run(*block->Program(), &cur_scope, block->ID(), false); } } @@ -74,8 +74,7 @@ class ConditionalBlockOp : public ConditionalOp { class ConditionalBlockOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - ConditionalBlockOpProtoMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + ConditionalBlockOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "The conditional variable of this operator. If X is empty, the " @@ -87,8 +86,8 @@ class ConditionalBlockOpProtoMaker : public framework::OpProtoAndCheckerMaker { "(std::vector) The step scope of conditional block. To " "unify the conditional block, rnn and while op, the type of " "scope is std::vector"); - AddAttr( - "block", "The step block of conditional block operator"); + AddAttr( + "sub_block", "The step block of conditional block operator"); AddComment(R"DOC(Conditional block operator Run the sub-block if X is not empty. Params is the other inputs and Out is the @@ -105,7 +104,7 @@ class ConditionalBlockGradOp : public ConditionalOp { const framework::AttributeMap &attrs) : ConditionalOp(type, inputs, outputs, attrs) {} void Run(const framework::Scope &scope, - const platform::DeviceContext &dev_ctx) const override { + const platform::Place &dev_place) const override { auto xs = this->InputTensors(scope); bool need_run = std::all_of( xs.begin(), xs.end(), @@ -117,21 +116,21 @@ class ConditionalBlockGradOp : public ConditionalOp { auto &scopes = scope_var->Get>(); framework::Scope &cur_scope = *scopes[0]; - auto *block = Attr("block"); - framework::Executor exec(dev_ctx); + framework::Executor exec(dev_place); + auto *block = Attr("sub_block"); exec.Run(*block->Program(), &cur_scope, block->ID(), false); - AssignLocalGradientToGlobal(dev_ctx, cur_scope, Inputs("Params"), + AssignLocalGradientToGlobal(dev_place, cur_scope, Inputs("Params"), Outputs(framework::GradVarName("Params"))); - AssignLocalGradientToGlobal(dev_ctx, cur_scope, Inputs("X"), + AssignLocalGradientToGlobal(dev_place, cur_scope, Inputs("X"), Outputs(framework::GradVarName("X"))); } } private: void AssignLocalGradientToGlobal( - const platform::DeviceContext &dev_ctx, const framework::Scope &cur_scope, + const platform::Place &place, const framework::Scope &cur_scope, const std::vector &p_names, const std::vector &pg_names) const { for (size_t i = 0; i < p_names.size(); ++i) { @@ -142,10 +141,10 @@ class ConditionalBlockGradOp : public ConditionalOp { continue; } auto new_in_grad_name = cur_scope.Rename(in_grad_name); - auto assign = - framework::OpRegistry::CreateOp("assign", {{"X", {new_in_grad_name}}}, - {{"Out", {out_grad_name}}}, {}); - assign->Run(cur_scope, dev_ctx); + auto assign = framework::OpRegistry::CreateOp( + "assign", {{"X", {new_in_grad_name}}}, {{"Out", {out_grad_name}}}, + framework::AttributeMap{}); + assign->Run(cur_scope, place); cur_scope.Rename(new_in_grad_name, in_grad_name); } } @@ -171,18 +170,19 @@ class ConditionalBlockGradMaker : public framework::SingleGradOpDescMaker { using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; protected: - std::unique_ptr Apply() const override { - auto grad_op = new framework::OpDescBind(); + std::unique_ptr Apply() const override { + auto grad_op = new framework::OpDesc(); grad_op->SetType("conditional_block_grad"); grad_op->SetInput("X", Input("X")); grad_op->SetInput("Params", Input("Params")); grad_op->SetInput("Out", Output("Out")); grad_op->SetInput(framework::GradVarName("Out"), OutputGrad("Out")); grad_op->SetInput("Scope", Output("Scope")); - grad_op->SetOutput(framework::GradVarName("X"), InputGrad("X")); - grad_op->SetOutput(framework::GradVarName("Params"), InputGrad("Params")); - grad_op->SetBlockAttr("block", *this->grad_block_[0]); - return std::unique_ptr(grad_op); + grad_op->SetOutput(framework::GradVarName("X"), InputGrad("X", false)); + grad_op->SetOutput(framework::GradVarName("Params"), + InputGrad("Params", false)); + grad_op->SetBlockAttr("sub_block", *this->grad_block_[0]); + return std::unique_ptr(grad_op); } }; diff --git a/paddle/operators/conv_cudnn_op.cc b/paddle/operators/conv_cudnn_op.cc index 4c65b60d2349d2989128f4b1da705ea18391b8a3..84d9ce1973a4cccadcb8f78feaecbcaa9e7af312 100644 --- a/paddle/operators/conv_cudnn_op.cc +++ b/paddle/operators/conv_cudnn_op.cc @@ -1,26 +1,25 @@ /* 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 +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 + 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. */ +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/operators/conv_op.h" namespace paddle { namespace operators { -class CudnnConvOpMaker : public Conv2DOpMaker { +class CudnnConv2DOpMaker : public Conv2DOpMaker { public: - CudnnConvOpMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker) + CudnnConv2DOpMaker(OpProto* proto, OpAttrChecker* op_checker) : Conv2DOpMaker(proto, op_checker) { AddAttr("workspace_size_MB", "workspace size for cudnn, in MB, " @@ -32,15 +31,44 @@ class CudnnConvOpMaker : public Conv2DOpMaker { } }; +class CudnnConv3DOpMaker : public Conv3DOpMaker { + public: + CudnnConv3DOpMaker(OpProto* proto, OpAttrChecker* op_checker) + : Conv3DOpMaker(proto, op_checker) { + AddAttr("workspace_size_MB", + "workspace size for cudnn, in MB, " + "workspace is a section of GPU memory which will be " + "allocated/freed each time the operator runs, larger " + "workspace size can increase performance but also requires " + "better hardware. This size should be chosen carefully.") + .SetDefault(4096); + } +}; + } // namespace operators } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP(conv_cudnn, ops::ConvOp, ops::CudnnConvOpMaker, conv_cudnn_grad, - ops::ConvOpGrad); +REGISTER_OP(conv2d_cudnn, ops::ConvOp, ops::CudnnConv2DOpMaker, + conv2d_cudnn_grad, ops::ConvOpGrad); + +REGISTER_OP(conv3d_cudnn, ops::ConvOp, ops::CudnnConv3DOpMaker, + conv3d_cudnn_grad, ops::ConvOpGrad); -REGISTER_OP_CPU_KERNEL(conv_cudnn, - ops::GemmConvKernel); REGISTER_OP_CPU_KERNEL( - conv_cudnn_grad, - ops::GemmConvGradKernel); + conv2d_cudnn, + ops::GemmConvKernel, + ops::GemmConvKernel); +REGISTER_OP_CPU_KERNEL( + conv2d_cudnn_grad, + ops::GemmConvGradKernel, + ops::GemmConvGradKernel); + +REGISTER_OP_CPU_KERNEL( + conv3d_cudnn, + ops::GemmConvKernel, + ops::GemmConvKernel); +REGISTER_OP_CPU_KERNEL( + conv3d_cudnn_grad, + ops::GemmConvGradKernel, + ops::GemmConvGradKernel); diff --git a/paddle/operators/conv_cudnn_op.cu.cc b/paddle/operators/conv_cudnn_op.cu.cc index 4900f7b086c869b496c492743c71ab7047c5f672..0aa7dd48cafc3e2387ac902882d84ce9029cfcd0 100644 --- a/paddle/operators/conv_cudnn_op.cu.cc +++ b/paddle/operators/conv_cudnn_op.cu.cc @@ -1,16 +1,16 @@ -/* Copyright (c) 2016 PaddlePaddle Authors All Rights Reserve. +/* 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 +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 + 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. */ +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/framework/eigen.h" #include "paddle/framework/op_registry.h" @@ -28,14 +28,15 @@ using ScopedFilterDescriptor = platform::ScopedFilterDescriptor; using ScopedConvolutionDescriptor = platform::ScopedConvolutionDescriptor; using DataLayout = platform::DataLayout; -static constexpr size_t kCONV_CUDNN_WORKSPACE_LIMIT_BYTES = 1024 * 1024 * 1024; +static constexpr size_t kCONV_CUDNN_WORKSPACE_LIMIT_BYTES = + static_cast(1024) * 1024 * 1024; template class CudnnConvOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), - "It must use GPUPlace."); + "It must use CUDAPlace."); auto* input = ctx.Input("Input"); auto* filter = ctx.Input("Filter"); auto* output = ctx.Output("Output"); @@ -44,7 +45,8 @@ class CudnnConvOpKernel : public framework::OpKernel { std::vector paddings = ctx.Attr>("paddings"); std::vector dilations = ctx.Attr>("dilations"); int groups = ctx.Attr("groups"); - int user_workspace_size = ctx.Attr("workspace_size_MB"); + int64_t user_workspace_size = + static_cast(ctx.Attr("workspace_size_MB")); const T* input_data = input->data(); const T* filter_data = filter->data(); @@ -56,6 +58,21 @@ class CudnnConvOpKernel : public framework::OpKernel { ScopedFilterDescriptor filter_desc; ScopedConvolutionDescriptor conv_desc; DataLayout layout = DataLayout::kNCHW; + if (input->dims().size() == 5) { + layout = DataLayout::kNCDHW; + } + + cudnnConvolutionDescriptor_t cudnn_conv_desc = + conv_desc.descriptor(paddings, strides, dilations); + +#if CUDNN_VERSION_MIN(7, 0, 1) + // cudnn 7 can support groups, no need to do it mannually + // FIXME(typhoonzero): find a better way to disable groups + // rather than setting it to 1. + PADDLE_ENFORCE(platform::dynload::cudnnSetConvolutionGroupCount( + cudnn_conv_desc, groups)); + groups = 1; +#endif cudnnTensorDescriptor_t cudnn_input_desc = input_desc.descriptor( layout, framework::vectorize2int(input->dims()), groups); @@ -63,19 +80,34 @@ class CudnnConvOpKernel : public framework::OpKernel { layout, framework::vectorize2int(output->dims()), groups); cudnnFilterDescriptor_t cudnn_filter_desc = filter_desc.descriptor( layout, framework::vectorize2int(filter->dims()), groups); - cudnnConvolutionDescriptor_t cudnn_conv_desc = - conv_desc.descriptor(paddings, strides, dilations); int input_channels = input->dims()[1]; - int input_height = input->dims()[2]; - int input_width = input->dims()[3]; - int output_channels = output->dims()[1]; - int output_height = output->dims()[2]; - int output_width = output->dims()[3]; + int input_height, input_width, input_depth; + if (input->dims().size() == 5) { + input_depth = input->dims()[2]; + input_height = input->dims()[3]; + input_width = input->dims()[4]; + } else { // dim size is enforced in InferShape + input_depth = 1; + input_height = input->dims()[2]; + input_width = input->dims()[3]; + } + int output_channels = filter->dims()[0]; + int output_height, output_width, output_depth; + if (output->dims().size() == 5) { + output_depth = output->dims()[2]; + output_height = output->dims()[3]; + output_width = output->dims()[4]; + } else { + output_depth = 1; + output_height = output->dims()[2]; + output_width = output->dims()[3]; + } - int group_offset_in = input_channels / groups * input_height * input_width; + int group_offset_in = + input_channels / groups * input_height * input_width * input_depth; int group_offset_out = - output_channels / groups * output_height * output_width; + output_channels / groups * output_height * output_width * output_depth; int group_offset_filter = filter->numel() / groups; // ------------------- cudnn conv workspace --------------------- void* cudnn_workspace = nullptr; @@ -86,7 +118,8 @@ class CudnnConvOpKernel : public framework::OpKernel { } // ------------------- cudnn conv algorithm --------------------- cudnnConvolutionFwdAlgo_t algo; - auto handle = ctx.cuda_device_context().cudnn_handle(); + auto& dev_ctx = ctx.template device_context(); + auto handle = dev_ctx.cudnn_handle(); PADDLE_ENFORCE(platform::dynload::cudnnGetConvolutionForwardAlgorithm( handle, cudnn_input_desc, cudnn_filter_desc, cudnn_conv_desc, @@ -97,7 +130,7 @@ class CudnnConvOpKernel : public framework::OpKernel { handle, cudnn_input_desc, cudnn_filter_desc, cudnn_conv_desc, cudnn_output_desc, algo, &workspace_size_in_bytes)); // Allocate on GPU memory - platform::GPUPlace gpu = boost::get(ctx.GetPlace()); + platform::CUDAPlace gpu = boost::get(ctx.GetPlace()); cudnn_workspace = paddle::memory::Alloc(gpu, workspace_size_in_bytes); // ------------------- cudnn conv forward --------------------- T alpha = 1.0f, beta = 0.0f; @@ -118,7 +151,7 @@ class CudnnConvGradOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), - "It must use GPUPlace."); + "It must use CUDAPlace."); auto input = ctx.Input("Input"); auto filter = ctx.Input("Filter"); auto output_grad = ctx.Input(framework::GradVarName("Output")); @@ -133,17 +166,32 @@ class CudnnConvGradOpKernel : public framework::OpKernel { std::vector paddings = ctx.Attr>("paddings"); std::vector dilations = ctx.Attr>("dilations"); int groups = ctx.Attr("groups"); - int user_workspace_size = ctx.Attr("workspace_size_MB"); + int64_t user_workspace_size = + static_cast(ctx.Attr("workspace_size_MB")); // ------------------- cudnn descriptors --------------------- ScopedTensorDescriptor input_desc; ScopedTensorDescriptor output_grad_desc; - ScopedTensorDescriptor input_grad_desc; ScopedFilterDescriptor filter_desc; ScopedFilterDescriptor filter_grad_desc; ScopedConvolutionDescriptor conv_desc; DataLayout layout = DataLayout::kNCHW; + if (input->dims().size() == 5) { + layout = DataLayout::kNCDHW; + } + + cudnnConvolutionDescriptor_t cudnn_conv_desc = + conv_desc.descriptor(paddings, strides, dilations); + +#if CUDNN_VERSION_MIN(7, 0, 1) + // cudnn 7 can support groups, no need to do it mannually + // FIXME(typhoonzero): find a better way to disable groups + // rather than setting it to 1. + PADDLE_ENFORCE(platform::dynload::cudnnSetConvolutionGroupCount( + cudnn_conv_desc, groups)); + groups = 1; +#endif cudnnTensorDescriptor_t cudnn_input_desc = input_desc.descriptor( layout, framework::vectorize2int(input->dims()), groups); @@ -152,22 +200,35 @@ class CudnnConvGradOpKernel : public framework::OpKernel { layout, framework::vectorize2int(output_grad->dims()), groups); cudnnFilterDescriptor_t cudnn_filter_desc = filter_desc.descriptor( layout, framework::vectorize2int(filter->dims()), groups); - cudnnTensorDescriptor_t cudnn_input_grad_desc = nullptr; - cudnnFilterDescriptor_t cudnn_filter_grad_desc = nullptr; - - cudnnConvolutionDescriptor_t cudnn_conv_desc = - conv_desc.descriptor(paddings, strides, dilations); int input_channels = input->dims()[1]; - int input_height = input->dims()[2]; - int input_width = input->dims()[3]; + int input_height, input_width, input_depth; + if (input->dims().size() == 5) { + input_depth = input->dims()[2]; + input_height = input->dims()[3]; + input_width = input->dims()[4]; + } else { // dim size is enforced in InferShape + input_depth = 1; + input_height = input->dims()[2]; + input_width = input->dims()[3]; + } + int output_grad_channels = filter->dims()[0]; - int output_grad_height = output_grad->dims()[2]; - int output_grad_width = output_grad->dims()[3]; + int output_grad_height, output_grad_width, output_grad_depth; + if (input->dims().size() == 5) { + output_grad_depth = output_grad->dims()[2]; + output_grad_height = output_grad->dims()[3]; + output_grad_width = output_grad->dims()[4]; + } else { + output_grad_depth = 1; + output_grad_height = output_grad->dims()[2]; + output_grad_width = output_grad->dims()[3]; + } - int group_offset_in = input_channels / groups * input_height * input_width; - int group_offset_out = - output_grad_channels / groups * output_grad_height * output_grad_width; + int group_offset_in = + input_channels / groups * input_height * input_width * input_depth; + int group_offset_out = output_grad_channels / groups * output_grad_height * + output_grad_width * output_grad_depth; int group_offset_filter = filter->numel() / groups; // ------------------- cudnn backward algorithm --------------------- cudnnConvolutionBwdDataAlgo_t data_algo; @@ -178,10 +239,9 @@ class CudnnConvGradOpKernel : public framework::OpKernel { workspace_size_limit = user_workspace_size * 1024 * 1024; } - auto handle = ctx.cuda_device_context().cudnn_handle(); + auto& dev_ctx = ctx.template device_context(); + auto handle = dev_ctx.cudnn_handle(); if (input_grad) { - cudnn_input_grad_desc = input_grad_desc.descriptor( - layout, framework::vectorize2int(input_grad->dims()), groups); PADDLE_ENFORCE( platform::dynload::cudnnGetConvolutionBackwardDataAlgorithm( handle, cudnn_filter_desc, @@ -190,19 +250,17 @@ class CudnnConvGradOpKernel : public framework::OpKernel { cudnn_output_grad_desc, cudnn_conv_desc, // dxDesc: Handle to the previously initialized output tensor // descriptor. - cudnn_input_grad_desc, + cudnn_input_desc, CUDNN_CONVOLUTION_BWD_DATA_SPECIFY_WORKSPACE_LIMIT, workspace_size_limit, &data_algo)); PADDLE_ENFORCE( platform::dynload::cudnnGetConvolutionBackwardDataWorkspaceSize( handle, cudnn_filter_desc, cudnn_output_grad_desc, - cudnn_conv_desc, cudnn_input_grad_desc, data_algo, &tmp_size)); + cudnn_conv_desc, cudnn_input_desc, data_algo, &tmp_size)); workspace_size_in_bytes = std::max(workspace_size_in_bytes, tmp_size); } if (filter_grad) { - cudnn_filter_grad_desc = filter_grad_desc.descriptor( - layout, framework::vectorize2int(filter_grad->dims()), groups); PADDLE_ENFORCE( platform::dynload::cudnnGetConvolutionBackwardFilterAlgorithm( handle, cudnn_input_desc, cudnn_output_grad_desc, cudnn_conv_desc, @@ -219,10 +277,9 @@ class CudnnConvGradOpKernel : public framework::OpKernel { // ------------------- cudnn conv workspace --------------------- // Already on GPU void* cudnn_workspace = nullptr; - platform::GPUPlace gpu = boost::get(ctx.GetPlace()); + platform::CUDAPlace gpu = boost::get(ctx.GetPlace()); 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. T alpha = 1.0f, beta = 0.0f; if (input_grad) { T* input_grad_data = input_grad->mutable_data(ctx.GetPlace()); @@ -233,21 +290,20 @@ class CudnnConvGradOpKernel : public framework::OpKernel { handle, &alpha, cudnn_filter_desc, filter_data + i * group_offset_filter, cudnn_output_grad_desc, output_grad_data + i * group_offset_out, cudnn_conv_desc, data_algo, - cudnn_workspace, workspace_size_in_bytes, &beta, - cudnn_input_grad_desc, input_grad_data + i * group_offset_in)); + cudnn_workspace, workspace_size_in_bytes, &beta, cudnn_input_desc, + input_grad_data + i * group_offset_in)); } } // ------------------- cudnn conv backward filter --------------------- if (filter_grad) { T* filter_grad_data = filter_grad->mutable_data(ctx.GetPlace()); // Because beta is zero, it is unnecessary to reset filter_grad. - for (int i = 0; i < groups; i++) { PADDLE_ENFORCE(platform::dynload::cudnnConvolutionBackwardFilter( handle, &alpha, cudnn_input_desc, input_data + i * group_offset_in, cudnn_output_grad_desc, output_grad_data + i * group_offset_out, cudnn_conv_desc, filter_algo, cudnn_workspace, - workspace_size_in_bytes, &beta, cudnn_filter_grad_desc, + workspace_size_in_bytes, &beta, cudnn_filter_desc, filter_grad_data + i * group_offset_filter)); } } @@ -259,6 +315,20 @@ class CudnnConvGradOpKernel : public framework::OpKernel { } // namespace operators } // namespace paddle -REGISTER_OP_GPU_KERNEL(conv_cudnn, paddle::operators::CudnnConvOpKernel); -REGISTER_OP_GPU_KERNEL(conv_cudnn_grad, - paddle::operators::CudnnConvGradOpKernel); +REGISTER_OP_KERNEL(conv2d, CUDNN, paddle::platform::CUDAPlace, + paddle::operators::CudnnConvOpKernel, + paddle::operators::CudnnConvOpKernel); + +REGISTER_OP_CUDA_KERNEL(conv2d_cudnn, + paddle::operators::CudnnConvOpKernel, + paddle::operators::CudnnConvOpKernel); +REGISTER_OP_CUDA_KERNEL(conv2d_cudnn_grad, + paddle::operators::CudnnConvGradOpKernel, + paddle::operators::CudnnConvGradOpKernel); + +REGISTER_OP_CUDA_KERNEL(conv3d_cudnn, + paddle::operators::CudnnConvOpKernel, + paddle::operators::CudnnConvOpKernel); +REGISTER_OP_CUDA_KERNEL(conv3d_cudnn_grad, + paddle::operators::CudnnConvGradOpKernel, + paddle::operators::CudnnConvGradOpKernel); diff --git a/paddle/operators/conv_op.cc b/paddle/operators/conv_op.cc index 7a36a9b21aa6a1b415ac5a232e65eda8051c87f8..e65a5dce52c3c51d3d6bee1684c1e97230203d38 100644 --- a/paddle/operators/conv_op.cc +++ b/paddle/operators/conv_op.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/operators/conv_op.h" @@ -31,8 +31,6 @@ void ConvOp::InferShape(framework::InferShapeContext* ctx) const { std::vector paddings = ctx->Attrs().Get>("paddings"); int groups = ctx->Attrs().Get("groups"); std::vector dilations = ctx->Attrs().Get>("dilations"); - int input_channels = in_dims[1]; - int output_channels = filter_dims[0]; PADDLE_ENFORCE(in_dims.size() == 4 || in_dims.size() == 5, "Conv intput should be 4-D or 5-D tensor."); @@ -45,9 +43,13 @@ void ConvOp::InferShape(framework::InferShapeContext* ctx) const { PADDLE_ENFORCE_EQ( paddings.size(), strides.size(), "Conv paddings dimension and Conv strides dimension should be the same."); + + int input_channels = in_dims[1]; PADDLE_ENFORCE_EQ(input_channels, filter_dims[1] * groups, "The number of input channels should be equal to filter " "channels * groups."); + + int output_channels = filter_dims[0]; PADDLE_ENFORCE_EQ( output_channels % groups, 0, "The number of output channels should be divided by groups."); @@ -66,8 +68,7 @@ void ConvOp::InferShape(framework::InferShapeContext* ctx) const { ctx->SetOutputDim("Output", framework::make_ddim(output_shape)); } -Conv2DOpMaker::Conv2DOpMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker) +Conv2DOpMaker::Conv2DOpMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput( "Input", @@ -97,7 +98,7 @@ Conv2DOpMaker::Conv2DOpMaker(framework::OpProto* proto, .SetDefault({0, 0}); AddAttr( "groups", - "(int default:1), the group size of convolution operator. " + "(int default:1), the groups number of the convolution operator. " "According to grouped convolution in Alex Krizhevsky's Deep CNN paper: " "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 " @@ -112,28 +113,33 @@ Conv2DOpMaker::Conv2DOpMaker(framework::OpProto* proto, Convolution Operator. The convolution operation calculates the output based on the input, filter -and strides, paddings, groups, dilations parameters. The size of each dimension of the +and strides, paddings, dilations, groups parameters. The size of each dimension of the parameters is checked in the infer-shape. -Input(Input, Filter) and output(Output) are in NCHW format. Where N is batch +Input(Input) and Output(Output) are in NCHW format. Where N is batch size, C is the number of channels, H is the height of the feature, and W is -the width of the feature. Parameters(ksize, strides, paddings, dilations) are two elements. -These two elements represent height and width, respectively. +the width of the feature. +Filters(Input) is MCHW format. Where M is the number of output image channels, C is +the number of input image channels, H is the height of the filter, and W +is the width of the filter. +Parameters(strides, paddings, dilations) are two elements. These two elements represent +height and width, respectively. The input(X) size and output(Out) size may be different. Example: Input: - Input shape: (N, C_in, H_in, W_in) - Filter shape: (C_out, C_in, H_f, W_f) + Input shape: $(N, C_{in}, H_{in}, W_{in})$ + Filter shape: $(C_{out}, C_{in}, H_f, W_f)$ Output: - Output shape: (N, C_out, H_out, W_out) - where - H_out = (H_in + 2 * paddings[0] - (dilations[0]*(filter_size[0] - 1) + 1)) / strides[0] + 1; - W_out = (W_in + 2 * paddings[1] - (dilations[1]*(filter_size[1] - 1) + 1)) / strides[1] + 1; + Output shape: $(N, C_{out}, H_{out}, W_{out})$ + Where +$$ + H_{out}= \frac{(H_{in} + 2 * paddings[0] - (dilations[0] * (H_f - 1) + 1))}{strides[0]}+ 1 \\ + W_{out}= \frac{(W_{in} + 2 * paddings[1] - (dilations[1] * (W_f - 1) + 1))}{strides[1]}+ 1 +$$ )DOC"); } -Conv3DOpMaker::Conv3DOpMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker) +Conv3DOpMaker::Conv3DOpMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput( "Input", @@ -165,7 +171,7 @@ Conv3DOpMaker::Conv3DOpMaker(framework::OpProto* proto, .SetDefault({0, 0, 0}); AddAttr( "groups", - "(int default:1), the group size of convolution operator. " + "(int default:1), the groups number of the convolution operator. " "According to grouped convolution in Alex Krizhevsky's Deep CNN paper: " "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 " @@ -174,32 +180,37 @@ Conv3DOpMaker::Conv3DOpMaker(framework::OpProto* proto, AddAttr>("dilations", "(vector default:{1, 1, 1}), the " "dilations(d_dilation, h_dilation, w_dilation) of " - "convolution operator. Currently, conv3d doesn't " - "support dilation.") + "convolution operator.") .SetDefault({1, 1, 1}); AddComment(R"DOC( Convolution3D Operator. The convolution operation calculates the output based on the input, filter -and strides, paddings, groups parameters. The size of each dimension of the +and strides, paddings, dilations, groups parameters. The size of each dimension of the parameters is checked in the infer-shape. -Input(Input, Filter) and output(Output) are in NCDHW format. Where N is batch +Input(Input) and output(Output) are in NCDHW format, where N is batch size, C is the number of channels,D is the depth of the feature, H is the height of -the feature, and W is the width of the feature. Parameters(ksize, strides, paddings) -are three elements. These three elements represent depth, height and width, respectively. +the feature, and W is the width of the feature. +Filters(Input) is MCDHW format, where M is the number of output image channels, +C is the number of input image channels, D is the depth of the filter, +H is the height of the filter, and W is the width of the filter. +Parameters(strides, paddings, dilations) are three elements. These three elements +represent depth, height and width, respectively. The input(X) size and output(Out) size may be different. Example: Input: - Input shape: (N, C_in, D_in, H_in, W_in) - Filter shape: (C_out, C_in, D_f, H_f, W_f) + Input shape: $(N, C_{in}, D_{in}, H_{in}, W_{in})$ + Filter shape: $(C_{out}, C_{in}, D_f, H_f, W_f)$ Output: - Output shape: (N, C_out, D_out, H_out, W_out) - where - D_out = (D_in - filter_size[0] + 2 * paddings[0]) / strides[0] + 1; - H_out = (H_in - filter_size[1] + 2 * paddings[1]) / strides[1] + 1; - W_out = (W_in - filter_size[2] + 2 * paddings[2]) / strides[2] + 1; + Output shape: $(N, C_{out}, D_{out}, H_{out}, W_{out})$ + Where + $$ + D_{out}= \frac{(D_{in} + 2 * paddings[0] - (dilations[0] * (D_f - 1) + 1))}{ strides[0]}+ 1 \\ + H_{out}= \frac{(H_{in} + 2 * paddings[1] - (dilations[1] * (H_f - 1) + 1))}{ strides[1]}+ 1 \\ + W_{out}= \frac{(W_{in} + 2 * paddings[2] - (dilations[2] * (W_f - 1) + 1))}{ strides[2]}+ 1 + $$ )DOC"); } @@ -224,16 +235,18 @@ namespace ops = paddle::operators; REGISTER_OP(conv3d, ops::ConvOp, ops::Conv3DOpMaker, conv3d_grad, ops::ConvOpGrad); -REGISTER_OP_CPU_KERNEL(conv2d, - ops::GemmConvKernel, - ops::GemmConvKernel); REGISTER_OP_CPU_KERNEL( - conv2d_grad, ops::GemmConvGradKernel, - ops::GemmConvGradKernel); + conv2d, ops::GemmConvKernel, + ops::GemmConvKernel); +REGISTER_OP_CPU_KERNEL( + conv2d_grad, + ops::GemmConvGradKernel, + ops::GemmConvGradKernel); -REGISTER_OP_CPU_KERNEL(conv3d, - ops::GemmConvKernel, - ops::GemmConvKernel); REGISTER_OP_CPU_KERNEL( - conv3d_grad, ops::GemmConvGradKernel, - ops::GemmConvGradKernel); + conv3d, ops::GemmConvKernel, + ops::GemmConvKernel); +REGISTER_OP_CPU_KERNEL( + conv3d_grad, + ops::GemmConvGradKernel, + ops::GemmConvGradKernel); diff --git a/paddle/operators/conv_op.cu.cc b/paddle/operators/conv_op.cu.cc index 546451234a1ed1a4d3119cb175c6d37ae3f0aac1..4f942444f3eb5584f07399b8d1b4d6a5087496d4 100644 --- a/paddle/operators/conv_op.cu.cc +++ b/paddle/operators/conv_op.cu.cc @@ -1,31 +1,33 @@ -/* Copyright (c) 2016 PaddlePaddle Authors All Rights Reserve. +/* 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 +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 + 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. */ +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/operators/conv_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(conv2d, - ops::GemmConvKernel, - ops::GemmConvKernel); -REGISTER_OP_GPU_KERNEL( - conv2d_grad, ops::GemmConvGradKernel, - ops::GemmConvGradKernel); +REGISTER_OP_CUDA_KERNEL( + conv2d, ops::GemmConvKernel, + ops::GemmConvKernel); +REGISTER_OP_CUDA_KERNEL( + conv2d_grad, + ops::GemmConvGradKernel, + ops::GemmConvGradKernel); -REGISTER_OP_GPU_KERNEL(conv3d, - ops::GemmConvKernel, - ops::GemmConvKernel); -REGISTER_OP_GPU_KERNEL( - conv3d_grad, ops::GemmConvGradKernel, - ops::GemmConvGradKernel); +REGISTER_OP_CUDA_KERNEL( + conv3d, ops::GemmConvKernel, + ops::GemmConvKernel); +REGISTER_OP_CUDA_KERNEL( + conv3d_grad, + ops::GemmConvGradKernel, + ops::GemmConvGradKernel); diff --git a/paddle/operators/conv_op.h b/paddle/operators/conv_op.h index fac5f1d0e25fe205f89fc7eeb9fadfd8431517d5..83786e2329e7ae3c2908fdfdaeb1f79d19a53f47 100644 --- a/paddle/operators/conv_op.h +++ b/paddle/operators/conv_op.h @@ -38,7 +38,7 @@ inline bool IsExpand(std::vector& filter_dim, std::vector& dilations) { bool filter_1 = true, strides_1 = true, padding_0 = true, dilation_1 = true; for (size_t j = 0; j < strides.size(); ++j) { - filter_1 = filter_1 && (static_cast(filter_dim[j]) == 1); + filter_1 = filter_1 && (static_cast(filter_dim[j + 2]) == 1); strides_1 = strides_1 && (strides[j] == 1); padding_0 = padding_0 && (paddings[j] == 0); dilation_1 = dilation_1 && (dilations[j] == 1); @@ -50,14 +50,12 @@ inline bool IsExpand(std::vector& filter_dim, // operator implementations can reuse the code. class Conv2DOpMaker : public framework::OpProtoAndCheckerMaker { public: - Conv2DOpMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker); + Conv2DOpMaker(OpProto* proto, OpAttrChecker* op_checker); }; class Conv3DOpMaker : public framework::OpProtoAndCheckerMaker { public: - Conv3DOpMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker); + Conv3DOpMaker(OpProto* proto, OpAttrChecker* op_checker); }; class ConvOp : public framework::OperatorWithKernel { @@ -72,7 +70,7 @@ class ConvOpGrad : public framework::OperatorWithKernel { void InferShape(framework::InferShapeContext* ctx) const override; }; -template +template class GemmConvKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { @@ -91,32 +89,28 @@ class GemmConvKernel : public framework::OpKernel { const int batch_size = static_cast(input->dims()[0]); - // filter_shape_vec: {k_h, k_w} or {k_d, k_h, k_w} + // filter_shape_vec: {k_o, k_i, k_h, k_w} or {k_o, k_i, k_d, k_h, k_w} std::vector filter_shape_vec(framework::vectorize(filter.dims())); - filter_shape_vec.erase(filter_shape_vec.begin(), - filter_shape_vec.begin() + 2); - - // output_shape_vec: {o_h, o_w} or {o_d, o_h, o_w} + // output_shape_vec: {o_n, o_c, o_h, o_w} or {o_n, o_c, o_d, o_h, o_w} std::vector output_shape_vec(framework::vectorize(output->dims())); - output_shape_vec.erase(output_shape_vec.begin(), - output_shape_vec.begin() + 2); // use col_shape in the im2col calculation // col_shape_vec: {i_c/g, k_h, k_w, o_h, o_w} or {i_c/g, k_d, k_h, k_w, o_d, // o_h, o_w} - std::vector col_shape_vec; - col_shape_vec.push_back(input->dims()[1] / groups); - col_shape_vec.insert(col_shape_vec.end(), filter_shape_vec.begin(), - filter_shape_vec.end()); - col_shape_vec.insert(col_shape_vec.end(), output_shape_vec.begin(), - output_shape_vec.end()); + size_t data_dim = filter_shape_vec.size() - 2; + std::vector col_shape_vec(1 + 2 * data_dim); + col_shape_vec[0] = input->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] = output_shape_vec[j + 2]; + } framework::DDim col_shape(framework::make_ddim(col_shape_vec)); // use col_matrix_shape in the gemm calculation // size: (i_c/g * k_h * k_w, o_h * o_w) or (i_c/g * k_d * k_h * k_w, o_d * // o_h * o_w) framework::DDim col_matrix_shape = - framework::flatten_to_2d(col_shape, filter_shape_vec.size() + 1); + framework::flatten_to_2d(col_shape, data_dim + 1); bool is_expand = IsExpand(filter_shape_vec, strides, paddings, dilations); Tensor col; @@ -145,9 +139,10 @@ class GemmConvKernel : public framework::OpKernel { int in_step = static_cast(input->dims()[1]) / groups; int out_step = static_cast(output->dims()[1]) / groups; - math::Vol2ColFunctor vol2col; - math::Im2ColFunctor im2col; + math::Vol2ColFunctor vol2col; + math::Im2ColFunctor im2col; + auto& dev_ctx = context.template device_context(); 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); @@ -159,29 +154,28 @@ class GemmConvKernel : public framework::OpKernel { col.ShareDataWith(in_slice); col_matrix.ShareDataWith(col); col_matrix.Resize(col_matrix_shape); - } else if (filter_shape_vec.size() == 2) { + } else if (data_dim == 2U) { // im2col - im2col(context.device_context(), in_slice, dilations, strides, + im2col(dev_ctx, in_slice, dilations, strides, std::vector{paddings[0], paddings[1], paddings[0], paddings[1]}, &col); - } else if (filter_shape_vec.size() == 3) { + } else if (data_dim == 3U) { // vol2col - vol2col(context.device_context(), in_slice, dilations, strides, - paddings, &col); + vol2col(dev_ctx, in_slice, dilations, strides, paddings, &col); } // 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(context.device_context(), filter_slice, false, - col_matrix, false, T(1.0), &out_slice, T(0.0)); + math::matmul(dev_ctx, filter_slice, false, col_matrix, + false, T(1.0), &out_slice, T(0.0)); } } } }; -template +template class GemmConvGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { @@ -206,26 +200,22 @@ class GemmConvGradKernel : public framework::OpKernel { const int batch_size = static_cast(input->dims()[0]); - // filter_shape_vec: {k_h, k_w} or {k_d, k_h, k_w} + // filter_shape_vec: {k_o, k_i, k_h, k_w} or {k_o, k_i, k_d, k_h, k_w} std::vector filter_shape_vec(framework::vectorize(filter.dims())); - filter_shape_vec.erase(filter_shape_vec.begin(), - filter_shape_vec.begin() + 2); - - // output_shape_vec: {o_h, o_w} or {o_d, o_h, o_w} + // output_shape_vec: {o_n, o_c, o_h, o_w} or {o_n, o_c, o_d, o_h, o_w} std::vector output_shape_vec( framework::vectorize(output_grad->dims())); - output_shape_vec.erase(output_shape_vec.begin(), - output_shape_vec.begin() + 2); // use col_shape in the im2col calculation // col_shape_vec: {i_c/g, k_h, k_w, o_h, o_w} or {i_c/g, k_d, k_h, k_w, o_d, // o_h, o_w} - std::vector col_shape_vec; - col_shape_vec.push_back(input->dims()[1] / groups); - col_shape_vec.insert(col_shape_vec.end(), filter_shape_vec.begin(), - filter_shape_vec.end()); - col_shape_vec.insert(col_shape_vec.end(), output_shape_vec.begin(), - output_shape_vec.end()); + size_t data_dim = filter_shape_vec.size() - 2; + std::vector col_shape_vec(1 + 2 * data_dim); + col_shape_vec[0] = input->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] = output_shape_vec[j + 2]; + } framework::DDim col_shape(framework::make_ddim(col_shape_vec)); // use col_matrix_shape in the gemm calculation @@ -233,7 +223,7 @@ class GemmConvGradKernel : public framework::OpKernel { // or // (i_c/g * k_d * k_h * k_w, o_d * o_h * o_w) framework::DDim col_matrix_shape = - framework::flatten_to_2d(col_shape, filter_shape_vec.size() + 1); + framework::flatten_to_2d(col_shape, data_dim + 1); framework::DDim input_shape = framework::slice_ddim( input->dims(), 1, static_cast(input->dims().size())); @@ -264,14 +254,19 @@ class GemmConvGradKernel : public framework::OpKernel { col_matrix.Resize(col_matrix_shape); } - math::SetConstant set_zero; + math::SetConstant set_zero; + auto& dev_ctx = context.template device_context(); if (input_grad) { input_grad->mutable_data(context.GetPlace()); - set_zero(context.device_context(), input_grad, static_cast(0)); - math::Col2VolFunctor col2vol; - math::Col2ImFunctor col2im; + // if is_expand is false, the operation of set_zero is unnecessary, + // because math::matmul will reset input_grad. + if (is_expand) { + set_zero(dev_ctx, input_grad, static_cast(0)); + } + math::Col2VolFunctor col2vol; + math::Col2ImFunctor col2im; for (int i = 0; i < batch_size; i++) { Tensor out_grad_batch = @@ -290,18 +285,17 @@ class GemmConvGradKernel : public framework::OpKernel { col_matrix.ShareDataWith(in_grad_slice); col_matrix.Resize(col_matrix_shape); } - math::matmul(context.device_context(), filter_slice, true, - out_grad_slice, false, T(1.0), &col_matrix, - T(0.0)); + math::matmul(dev_ctx, filter_slice, true, + out_grad_slice, false, T(1.0), + &col_matrix, T(0.0)); - if (is_expand && filter_shape_vec.size() == 2) { - col2im(context.device_context(), col, dilations, strides, + if (is_expand && data_dim == 2U) { + col2im(dev_ctx, col, dilations, strides, std::vector{paddings[0], paddings[1], paddings[0], paddings[1]}, &in_grad_slice); - } else if (is_expand && filter_shape_vec.size() == 3) { - col2vol(context.device_context(), col, dilations, strides, paddings, - &in_grad_slice); + } else if (is_expand && data_dim == 3U) { + col2vol(dev_ctx, col, dilations, strides, paddings, &in_grad_slice); } } } @@ -311,9 +305,9 @@ class GemmConvGradKernel : public framework::OpKernel { filter_grad->mutable_data(context.GetPlace()); Tensor filter_grad_ = *filter_grad; filter_grad_.Resize(filter_matrix_shape); - set_zero(context.device_context(), filter_grad, static_cast(0)); - math::Im2ColFunctor im2col; - math::Vol2ColFunctor vol2col; + set_zero(dev_ctx, filter_grad, static_cast(0)); + math::Im2ColFunctor im2col; + math::Vol2ColFunctor vol2col; for (int i = 0; i < batch_size; i++) { Tensor out_grad_batch = output_grad->Slice(i, i + 1).Resize(output_matrix_shape); @@ -328,22 +322,21 @@ class GemmConvGradKernel : public framework::OpKernel { col.ShareDataWith(in_slice); col_matrix.ShareDataWith(col); col_matrix.Resize(col_matrix_shape); - } else if (filter_shape_vec.size() == 2) { - im2col(context.device_context(), in_slice, dilations, strides, + } else if (data_dim == 2U) { + im2col(dev_ctx, in_slice, dilations, strides, std::vector{paddings[0], paddings[1], paddings[0], paddings[1]}, &col); - } else if (filter_shape_vec.size() == 3) { - vol2col(context.device_context(), in_slice, dilations, strides, - paddings, &col); + } else if (data_dim == 3U) { + vol2col(dev_ctx, in_slice, dilations, strides, paddings, &col); } // gemm Tensor filter_grad_slice = filter_grad_.Slice(g * out_step, (g + 1) * out_step); - math::matmul(context.device_context(), out_grad_slice, - false, col_matrix, true, T(1.0), - &filter_grad_slice, T(1.0)); + math::matmul(dev_ctx, out_grad_slice, false, + col_matrix, true, T(1.0), + &filter_grad_slice, T(1.0)); } } } diff --git a/paddle/operators/conv_shift_op.cc b/paddle/operators/conv_shift_op.cc index a4150a5664690e750d2501a1849767c23209186b..106b68a0a0e787a0c9da2de924f4646c77b42b41 100644 --- a/paddle/operators/conv_shift_op.cc +++ b/paddle/operators/conv_shift_op.cc @@ -1,16 +1,16 @@ -/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserve. +/* 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 +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 + 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. */ +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/operators/conv_shift_op.h" #include "paddle/framework/eigen.h" @@ -75,8 +75,7 @@ class ConvShiftGradOp : public framework::OperatorWithKernel { class ConvShiftOpMaker : public framework::OpProtoAndCheckerMaker { public: - ConvShiftOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + ConvShiftOpMaker(OpProto *proto, OpAttrChecker *op_checker) : framework::OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "(Tensor, default Tensor), a 2-D tensor with shape B x M, " diff --git a/paddle/operators/conv_shift_op.cu b/paddle/operators/conv_shift_op.cu index 95e13c38a8dd234f49393d2d4808607a447b0d4c..cf7abc196e1293ab1b998d1a8cb9c361a7c2d427 100644 --- a/paddle/operators/conv_shift_op.cu +++ b/paddle/operators/conv_shift_op.cu @@ -1,16 +1,16 @@ -/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserve. +/* 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 +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 + 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. */ +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/operators/conv_shift_op.h" #include "paddle/operators/math/math_function.h" @@ -111,7 +111,8 @@ __global__ void ConvShiftDy(const T *x, const T *dout, int x_width, int y_width, } // namespace template -class ConvShiftKernel : public framework::OpKernel { +class ConvShiftKernel + : public framework::OpKernel { public: void Compute(const framework::ExecutionContext &context) const override { const Tensor *X = context.Input("X"); @@ -132,7 +133,8 @@ class ConvShiftKernel : public framework::OpKernel { dim3 grid_dim(num_x_blocks, batch_size); - auto stream = context.cuda_device_context().stream(); + auto stream = + context.template device_context().stream(); ConvShiftForward<<>>( x_data, y_data, x_width, y_width, y_half_width, batch_size, out_data); @@ -140,7 +142,7 @@ class ConvShiftKernel : public framework::OpKernel { }; template -class ConvShiftGradKernel +class ConvShiftGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext &context) const override { @@ -159,8 +161,9 @@ class ConvShiftGradKernel int y_width = Y->dims()[1]; int y_half_width = (y_width - 1) / 2; - auto &device_ctx = context.cuda_device_context(); - math::SetConstant zero; + auto &device_ctx = + context.template device_context(); + math::SetConstant zero; const int x_per_block = 256; int num_x_blocks = DivUp(x_width, x_per_block); @@ -186,8 +189,9 @@ class ConvShiftGradKernel } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(conv_shift, - ops::ConvShiftKernel); -REGISTER_OP_GPU_KERNEL( +REGISTER_OP_CUDA_KERNEL( + conv_shift, + ops::ConvShiftKernel); +REGISTER_OP_CUDA_KERNEL( conv_shift_grad, - ops::ConvShiftGradKernel); + ops::ConvShiftGradKernel); diff --git a/paddle/operators/conv_shift_op.h b/paddle/operators/conv_shift_op.h index 5a160b0f1696c70868fc48d219b38cde2018e8a3..6781d87ef0d99a0b0fc4747245920b6a38a33804 100644 --- a/paddle/operators/conv_shift_op.h +++ b/paddle/operators/conv_shift_op.h @@ -1,16 +1,16 @@ -/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserve. +/* 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 +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 + 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. */ +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/framework/op_registry.h" @@ -18,13 +18,13 @@ namespace paddle { namespace operators { -template +template class ConvShiftKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext &context) const override; }; -template +template class ConvShiftGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext &context) const override; diff --git a/paddle/operators/conv_transpose_cudnn_op.cc b/paddle/operators/conv_transpose_cudnn_op.cc index dbd1bc3c3bc2d026f13ddcf62919db6cf7d87bc5..2e5333a265f2f59f31c651b8bb080599ec6e31a4 100644 --- a/paddle/operators/conv_transpose_cudnn_op.cc +++ b/paddle/operators/conv_transpose_cudnn_op.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/operators/conv_transpose_op.h" @@ -19,11 +19,8 @@ namespace operators { class CudnnConv2DTransposeOpMaker : public Conv2DTransposeOpMaker { public: - CudnnConv2DTransposeOpMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker) + CudnnConv2DTransposeOpMaker(OpProto* proto, OpAttrChecker* op_checker) : Conv2DTransposeOpMaker(proto, op_checker) { - AddAttr>("dilations", "dilations of convolution operator.") - .SetDefault({1, 1}); AddAttr("workspace_size_MB", "workspace size for cudnn, in MB, " "workspace is a section of GPU memory which will be " @@ -36,11 +33,8 @@ class CudnnConv2DTransposeOpMaker : public Conv2DTransposeOpMaker { class CudnnConv3DTransposeOpMaker : public Conv3DTransposeOpMaker { public: - CudnnConv3DTransposeOpMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker) + CudnnConv3DTransposeOpMaker(OpProto* proto, OpAttrChecker* op_checker) : Conv3DTransposeOpMaker(proto, op_checker) { - AddAttr>("dilations", "dilations of convolution operator.") - .SetDefault({1, 1, 1}); AddAttr("workspace_size_MB", "workspace size for cudnn, in MB, " "workspace is a section of GPU memory which will be " @@ -61,10 +55,13 @@ REGISTER_OP(conv2d_transpose_cudnn, ops::ConvTransposeOp, REGISTER_OP_CPU_KERNEL( conv2d_transpose_cudnn, - ops::GemmConvTransposeKernel); + ops::GemmConvTransposeKernel, + ops::GemmConvTransposeKernel); REGISTER_OP_CPU_KERNEL( conv2d_transpose_cudnn_grad, - ops::GemmConvTransposeGradKernel); + ops::GemmConvTransposeGradKernel, + ops::GemmConvTransposeGradKernel); REGISTER_OP(conv3d_transpose_cudnn, ops::ConvTransposeOp, ops::CudnnConv3DTransposeOpMaker, conv3d_transpose_cudnn_grad, @@ -72,7 +69,10 @@ REGISTER_OP(conv3d_transpose_cudnn, ops::ConvTransposeOp, REGISTER_OP_CPU_KERNEL( conv3d_transpose_cudnn, - ops::GemmConvTransposeKernel); + ops::GemmConvTransposeKernel, + ops::GemmConvTransposeKernel); REGISTER_OP_CPU_KERNEL( conv3d_transpose_cudnn_grad, - ops::GemmConvTransposeGradKernel); + ops::GemmConvTransposeGradKernel, + ops::GemmConvTransposeGradKernel); diff --git a/paddle/operators/conv_transpose_cudnn_op.cu.cc b/paddle/operators/conv_transpose_cudnn_op.cu.cc index e2ba77086e737a07471f14e483cbd32ab1d4ee12..fc37776ba1ed35aa6b2523eb593e9713cfcc54eb 100644 --- a/paddle/operators/conv_transpose_cudnn_op.cu.cc +++ b/paddle/operators/conv_transpose_cudnn_op.cu.cc @@ -1,16 +1,16 @@ -/* Copyright (c) 2016 PaddlePaddle Authors All Rights Reserve. +/* 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 +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 + 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. */ +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/framework/eigen.h" #include "paddle/framework/op_registry.h" @@ -35,7 +35,7 @@ class CudnnConvTransposeOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), - "It must use GPUPlace."); + "It must use CUDAPlace."); auto* input = ctx.Input("Input"); auto* filter = ctx.Input("Filter"); auto* output = ctx.Output("Output"); @@ -83,7 +83,8 @@ class CudnnConvTransposeOpKernel : public framework::OpKernel { } // ------------------- cudnn conv algorithm --------------------- cudnnConvolutionBwdDataAlgo_t algo; - auto handle = ctx.cuda_device_context().cudnn_handle(); + auto& dev_ctx = ctx.template device_context(); + auto handle = dev_ctx.cudnn_handle(); // Get the algorithm PADDLE_ENFORCE(platform::dynload::cudnnGetConvolutionBackwardDataAlgorithm( handle, cudnn_filter_desc, cudnn_input_desc, cudnn_conv_desc, @@ -99,7 +100,7 @@ class CudnnConvTransposeOpKernel : public framework::OpKernel { cudnn_output_desc, algo, &workspace_size_in_bytes)); // Allocate on GPU memory - platform::GPUPlace gpu = boost::get(ctx.GetPlace()); + platform::CUDAPlace gpu = boost::get(ctx.GetPlace()); cudnn_workspace = paddle::memory::Alloc(gpu, workspace_size_in_bytes); // ------------------- cudnn conv transpose forward --------------------- @@ -119,7 +120,7 @@ class CudnnConvTransposeGradOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), - "It must use GPUPlace."); + "It must use CUDAPlace."); auto input = ctx.Input("Input"); auto filter = ctx.Input("Filter"); auto output_grad = ctx.Input(framework::GradVarName("Output")); @@ -165,7 +166,8 @@ class CudnnConvTransposeGradOpKernel : public framework::OpKernel { workspace_size_limit = user_workspace_size * 1024 * 1024; } - auto handle = ctx.cuda_device_context().cudnn_handle(); + auto& dev_ctx = ctx.template device_context(); + auto handle = dev_ctx.cudnn_handle(); if (input_grad) { // choose backward algorithm for data PADDLE_ENFORCE(platform::dynload::cudnnGetConvolutionForwardAlgorithm( @@ -199,7 +201,7 @@ class CudnnConvTransposeGradOpKernel : public framework::OpKernel { // ------------------- cudnn conv workspace --------------------- // Already on GPU void* cudnn_workspace = nullptr; - platform::GPUPlace gpu = boost::get(ctx.GetPlace()); + platform::CUDAPlace gpu = boost::get(ctx.GetPlace()); 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. @@ -234,12 +236,16 @@ class CudnnConvTransposeGradOpKernel : public framework::OpKernel { namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(conv2d_transpose_cudnn, - ops::CudnnConvTransposeOpKernel); -REGISTER_OP_GPU_KERNEL(conv2d_transpose_cudnn_grad, - ops::CudnnConvTransposeGradOpKernel); - -REGISTER_OP_GPU_KERNEL(conv3d_transpose_cudnn, - ops::CudnnConvTransposeOpKernel); -REGISTER_OP_GPU_KERNEL(conv3d_transpose_cudnn_grad, - ops::CudnnConvTransposeGradOpKernel); +REGISTER_OP_CUDA_KERNEL(conv2d_transpose_cudnn, + ops::CudnnConvTransposeOpKernel, + ops::CudnnConvTransposeOpKernel); +REGISTER_OP_CUDA_KERNEL(conv2d_transpose_cudnn_grad, + ops::CudnnConvTransposeGradOpKernel, + ops::CudnnConvTransposeGradOpKernel); + +REGISTER_OP_CUDA_KERNEL(conv3d_transpose_cudnn, + ops::CudnnConvTransposeOpKernel, + ops::CudnnConvTransposeOpKernel); +REGISTER_OP_CUDA_KERNEL(conv3d_transpose_cudnn_grad, + ops::CudnnConvTransposeGradOpKernel, + ops::CudnnConvTransposeGradOpKernel); diff --git a/paddle/operators/conv_transpose_op.cc b/paddle/operators/conv_transpose_op.cc index 3e55ef036a7fb976117054574d1347fa943acd55..74636d138f1e40474a1cc5453609dafe14fcaaab 100644 --- a/paddle/operators/conv_transpose_op.cc +++ b/paddle/operators/conv_transpose_op.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/operators/conv_transpose_op.h" @@ -29,6 +29,7 @@ void ConvTransposeOp::InferShape(framework::InferShapeContext* ctx) const { auto filter_dims = ctx->GetInputDim("Filter"); std::vector strides = ctx->Attrs().Get>("strides"); std::vector paddings = ctx->Attrs().Get>("paddings"); + std::vector dilations = ctx->Attrs().Get>("dilations"); PADDLE_ENFORCE(in_dims.size() == 4 || in_dims.size() == 5, "ConvTransposeOp intput should be 4-D or 5-D tensor."); @@ -39,7 +40,10 @@ void ConvTransposeOp::InferShape(framework::InferShapeContext* ctx) const { "ConvTransposeOp input dimension and strides dimension should " "be consistent."); PADDLE_ENFORCE_EQ(paddings.size(), strides.size(), - "ConvTransposeOp paddings dimension and Conv strides " + "ConvTransposeOp paddings dimension and strides " + "dimension should be the same."); + PADDLE_ENFORCE_EQ(paddings.size(), dilations.size(), + "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 " @@ -47,14 +51,15 @@ void ConvTransposeOp::InferShape(framework::InferShapeContext* ctx) const { std::vector output_shape({in_dims[0], filter_dims[1]}); 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] + - filter_dims[i + 2]); + filter_extent); } ctx->SetOutputDim("Output", framework::make_ddim(output_shape)); } -Conv2DTransposeOpMaker::Conv2DTransposeOpMaker( - framework::OpProto* proto, framework::OpAttrChecker* op_checker) +Conv2DTransposeOpMaker::Conv2DTransposeOpMaker(OpProto* proto, + OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput( "Input", @@ -62,52 +67,64 @@ Conv2DTransposeOpMaker::Conv2DTransposeOpMaker( "The format of input tensor is NCHW. Where N is batch size, C is the " "number of input channels, H is the height of the feature, and " "W is the width of the feature."); - AddInput("Filter", - "(Tensor) The filter tensor of convolution transpose operator. " - "The format of the filter tensor is CMHW, where C is the number of " - "output image channels, M is the number of input image channels, " - "H is the height of the filter, and W is the width of the filter. " - "We enforce groups number == 1 and padding == 0 in " - "the convolution transpose scenario."); + AddInput( + "Filter", + "(Tensor) The filter tensor of convolution transpose operator. " + "The format of the filter tensor is MCHW, where M is the number of " + "input feature channels, C is the number of " + "output feature channels," + "H is the height of the filter, and W is the width of the filter. " + "We enforce groups number == 1 in the convolution transpose scenario."); AddOutput("Output", "(Tensor) The output tensor of convolution transpose operator. " "The format of output tensor is also NCHW."); + + AddAttr>("dilations", + "(vector default:{1, 1}), the " + "dilations(h_dilation, w_dilation) of convolution " + "transpose operator.") + .SetDefault({1, 1}); AddAttr>( "strides", - "(vector defalut:{1, 1}), the strides(h_stride, w_stride) of " + "(vector default:{1, 1}), the strides(h_stride, w_stride) of " "convolution transpose operator.") .SetDefault({1, 1}); AddAttr>( "paddings", - "(vector defalut:{0, 0}), the paddings(h_pad, w_pad) of convolution " + "(vector default:{0, 0}), the paddings(h_pad, w_pad) of convolution " "transpose operator.") .SetDefault({0, 0}); AddComment(R"DOC( Convolution2D Transpose Operator. The convolution transpose operation calculates the output based on the input, filter -and strides, paddings, groups parameters. The size of each dimension of the +and dilations, strides, paddings, groups parameters. The size of each dimension of the parameters is checked in the infer-shape. - -Input(Input, Filter) and output(Output) are in NCHW format. Where N is batch -size, C is the number of channels, H is the height of the feature, and -W is the width of the feature. Parameters(ksize, strides, paddings) are two elements. -These two elements represent height and width, respectively. +Input(Input) and output(Output) are in NCHW format. Where N is batchsize, C is the +number of channels, H is the height of the feature, and W is the width of the feature. +Filter(Input) is in MCHW format. Where M is the number of input feature channels, +C is the number of output feature channels, H is the height of the filter, +and W is the width of the filter. +Parameters(strides, paddings) are two elements. These two elements represent height +and width, respectively. The input(X) size and output(Out) size may be different. + Example: Input: - Input shape: (N, C_in, H_in, W_in) - Filter shape: (C_in, C_out, H_f, W_f) + Input shape: $(N, C_{in}, H_{in}, W_{in})$ + Filter shape: $(C_{in}, C_{out}, H_f, W_f)$ Output: - Output shape: (N, C_out, H_out, W_out) - where - H_out = (H_in - 1) * strides[0] - 2 * paddings[0] + filter_size[0]; - W_out = (W_in - 1) * strides[1] - 2 * paddings[1] + filter_size[1]; + Output shape: $(N, C_{out}, H_{out}, W_{out})$ + Where + $$ + H_{out} = (H_{in} - 1) * strides[0] - 2 * paddings[0] + H_f \\ + W_{out} = (W_{in} - 1) * strides[1] - 2 * paddings[1] + W_f + $$ )DOC"); } -Conv3DTransposeOpMaker::Conv3DTransposeOpMaker( - framework::OpProto* proto, framework::OpAttrChecker* op_checker) +Conv3DTransposeOpMaker::Conv3DTransposeOpMaker(OpProto* proto, + OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("Input", "(Tensor) The input tensor of convolution transpose operator." @@ -117,8 +134,9 @@ Conv3DTransposeOpMaker::Conv3DTransposeOpMaker( "W is the width of the feature."); AddInput("Filter", "(Tensor) The filter tensor of convolution transpose operator." - "The format of the filter tensor is CMDHW, where C is the number of " - "output image channels, M is the number of input image channels, D " + "The format of the filter tensor is MCDHW, where M is the number of " + "input feature channels, C is the number of " + "output feature channels, D " "is the depth of the filter, H is the height of the filter, and " "W is the width of the filter." "We enforce groups number == 1 and padding == 0 in " @@ -129,38 +147,50 @@ Conv3DTransposeOpMaker::Conv3DTransposeOpMaker( "Where N is batch size, C is " "the number of channels, D is the depth of the feature, H is the " "height of the feature, and W is the width of the feature."); + + AddAttr>( + "dilations", + "(vector default:{1, 1, 1}), the " + "dilations(d_dilation,h_dilation, w_dilation) of convolution " + "transpose operator.") + .SetDefault({1, 1, 1}); AddAttr>("strides", - "(vector defalut:{1, 1, 1}), the " + "(vector default:{1, 1, 1}), the " "strides{d_stride, h_stride, w_stride} of " "convolution transpose operator.") .SetDefault({1, 1, 1}); AddAttr>("paddings", - "(vector defalut:{0, 0, 0}), paddings(d_pad, " + "(vector default:{0, 0, 0}), paddings(d_pad, " "h_pad, w_pad) of convolution transpose operator.") .SetDefault({0, 0, 0}); AddComment(R"DOC( Convolution3D Transpose Operator. The convolution transpose operation calculates the output based on the input, filter -and strides, paddings, groups parameters. The size of each dimension of the +and dilations, strides, paddings, groups parameters. The size of each dimension of the parameters is checked in the infer-shape. - -Input(Input, Filter) and output(Output) are in NCDHW format. Where N is batch -size, C is the number of channels, D is the depth of the feature, -H is the height of the feature, and W is the width of the feature. -Parameters(ksize, strides, paddings) are three elements. -These three elements represent depth, height and width, respectively. +Input(Input) and output(Output) are in NCDHW format. Where N is batch size, C is the +number of channels, D is the depth of the feature, H is the height of the feature, +and W is the width of the feature. +Filter(Input) is in MCDHW format. Where M is the number of input feature channels, +C is the number of output feature channels, D is the depth of the filter,H is the +height of the filter, and W is the width of the filter. +Parameters(strides, paddings) are three elements. These three elements represent +depth, height and width, respectively. The input(X) size and output(Out) size may be different. -Example: + +Example: Input: - Input shape: (N, C_in, D_in, H_in, W_in) - Filter shape: (C_in, C_out, D_f, H_f, W_f) + Input shape: $(N, C_{in}, D_{in}, H_{in}, W_{in})$ + Filter shape: $(C_{in}, C_{out}, D_f, H_f, W_f)$ Output: - Output shape: (N, C_out, D_out, H_out, W_out) - where - D_out = (D_in - 1) * strides[0] - 2 * paddings[0] + filter_size[0]; - H_out = (H_in - 1) * strides[1] - 2 * paddings[1] + filter_size[1]; - W_out = (W_in - 1) * strides[2] - 2 * paddings[2] + filter_size[2]; + Output shape: $(N, C_{out}, D_{out}, H_{out}, W_{out})$ + Where + $$ + D_{out} = (D_{in} - 1) * strides[0] - 2 * paddings[0] + D_f \\ + H_{out} = (H_{in} - 1) * strides[1] - 2 * paddings[1] + H_f \\ + W_{out} = (W_{in} - 1) * strides[2] - 2 * paddings[2] + W_f + $$ )DOC"); } @@ -185,21 +215,23 @@ REGISTER_OP(conv2d_transpose, ops::ConvTransposeOp, ops::Conv2DTransposeOpMaker, REGISTER_OP_CPU_KERNEL( conv2d_transpose, - ops::GemmConvTransposeKernel, - ops::GemmConvTransposeKernel); + ops::GemmConvTransposeKernel, + ops::GemmConvTransposeKernel); REGISTER_OP_CPU_KERNEL( conv2d_transpose_grad, - ops::GemmConvTransposeGradKernel, - ops::GemmConvTransposeGradKernel); + ops::GemmConvTransposeGradKernel, + ops::GemmConvTransposeGradKernel); REGISTER_OP(conv3d_transpose, ops::ConvTransposeOp, ops::Conv3DTransposeOpMaker, conv3d_transpose_grad, ops::ConvTransposeOpGrad); REGISTER_OP_CPU_KERNEL( conv3d_transpose, - ops::GemmConvTransposeKernel, - ops::GemmConvTransposeKernel); + ops::GemmConvTransposeKernel, + ops::GemmConvTransposeKernel); REGISTER_OP_CPU_KERNEL( conv3d_transpose_grad, - ops::GemmConvTransposeGradKernel, - ops::GemmConvTransposeGradKernel); + ops::GemmConvTransposeGradKernel, + ops::GemmConvTransposeGradKernel); diff --git a/paddle/operators/conv_transpose_op.cu.cc b/paddle/operators/conv_transpose_op.cu.cc index 4165eb0c7b048b83bbd94c57b971530043b66545..f1d827c606283440debb9a0edb25168816a3a08c 100644 --- a/paddle/operators/conv_transpose_op.cu.cc +++ b/paddle/operators/conv_transpose_op.cu.cc @@ -1,35 +1,39 @@ -/* Copyright (c) 2016 PaddlePaddle Authors All Rights Reserve. +/* 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 +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 + 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. */ +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/operators/conv_transpose_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL( +REGISTER_OP_CUDA_KERNEL( conv2d_transpose, - ops::GemmConvTransposeKernel, - ops::GemmConvTransposeKernel); -REGISTER_OP_GPU_KERNEL( + ops::GemmConvTransposeKernel, + ops::GemmConvTransposeKernel); +REGISTER_OP_CUDA_KERNEL( conv2d_transpose_grad, - ops::GemmConvTransposeGradKernel, - ops::GemmConvTransposeGradKernel); + ops::GemmConvTransposeGradKernel, + ops::GemmConvTransposeGradKernel); -REGISTER_OP_GPU_KERNEL( +REGISTER_OP_CUDA_KERNEL( conv3d_transpose, - ops::GemmConvTransposeKernel, - ops::GemmConvTransposeKernel); -REGISTER_OP_GPU_KERNEL( + ops::GemmConvTransposeKernel, + ops::GemmConvTransposeKernel); +REGISTER_OP_CUDA_KERNEL( conv3d_transpose_grad, - ops::GemmConvTransposeGradKernel, - ops::GemmConvTransposeGradKernel); + ops::GemmConvTransposeGradKernel, + ops::GemmConvTransposeGradKernel); diff --git a/paddle/operators/conv_transpose_op.h b/paddle/operators/conv_transpose_op.h index ab336ad23ce1c180b68d04e4c85b299e301d5376..4c8f8a80672788e8b2919e500d3627adec1ad035 100644 --- a/paddle/operators/conv_transpose_op.h +++ b/paddle/operators/conv_transpose_op.h @@ -30,14 +30,12 @@ using DDim = framework::DDim; // operator implementations can reuse the code. class Conv2DTransposeOpMaker : public framework::OpProtoAndCheckerMaker { public: - Conv2DTransposeOpMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker); + Conv2DTransposeOpMaker(OpProto* proto, OpAttrChecker* op_checker); }; class Conv3DTransposeOpMaker : public framework::OpProtoAndCheckerMaker { public: - Conv3DTransposeOpMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker); + Conv3DTransposeOpMaker(OpProto* proto, OpAttrChecker* op_checker); }; class ConvTransposeOp : public framework::OperatorWithKernel { @@ -52,7 +50,7 @@ class ConvTransposeOpGrad : public framework::OperatorWithKernel { void InferShape(framework::InferShapeContext* ctx) const override; }; -template +template class GemmConvTransposeKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { @@ -63,35 +61,31 @@ class GemmConvTransposeKernel : public framework::OpKernel { std::vector strides = context.Attr>("strides"); std::vector paddings = context.Attr>("paddings"); - // TODO(Zhuoyuan): Paddings can be added in future. + std::vector dilations = context.Attr>("dilations"); // groups will alway be disabled in conv2dtranspose. const int batch_size = static_cast(input->dims()[0]); - // input_shape_vec: {h, w} or {d, h, w} + // input_shape_vec: {n, c, h, w} or {n, c, d, h, w} std::vector input_shape_vec = framework::vectorize(input->dims()); - input_shape_vec.erase(input_shape_vec.begin(), input_shape_vec.begin() + 2); - - // filter_shape_vec: {k_h, k_w} or {k_d, k_h, k_w} + // filter_shape_vec: {k_o, k_c, k_h, k_w} or {k_o, k_c, k_d, k_h, k_w} std::vector filter_shape_vec = framework::vectorize(filter.dims()); - filter_shape_vec.erase(filter_shape_vec.begin(), - filter_shape_vec.begin() + 2); // 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} - std::vector col_shape_vec; - col_shape_vec.push_back(output->dims()[1]); - col_shape_vec.insert(col_shape_vec.end(), filter_shape_vec.begin(), - filter_shape_vec.end()); - col_shape_vec.insert(col_shape_vec.end(), input_shape_vec.begin(), - input_shape_vec.end()); + 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]; + 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]; + } 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) - DDim col_matrix_shape = - framework::flatten_to_2d(col_shape, filter_shape_vec.size() + 1); + DDim col_matrix_shape = framework::flatten_to_2d(col_shape, data_dim + 1); Tensor col; col.mutable_data(col_shape, context.GetPlace()); @@ -114,12 +108,12 @@ class GemmConvTransposeKernel : public framework::OpKernel { filter.Resize(filter_matrix_shape); output->mutable_data(context.GetPlace()); - math::SetConstant set_zero; - set_zero(context.device_context(), output, static_cast(0)); + math::SetConstant set_zero; + auto& dev_ctx = context.template device_context(); + set_zero(dev_ctx, output, static_cast(0)); - math::Col2ImFunctor col2im; - math::Col2VolFunctor col2vol; - std::vector dilations({1, 1, 1}); + math::Col2ImFunctor col2im; + math::Col2VolFunctor col2vol; // convolution transpose: gemm + col2im or col2vol (similar to conv-backward // on input) @@ -132,29 +126,27 @@ class GemmConvTransposeKernel : public framework::OpKernel { // 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(context.device_context(), filter, true, - input_batch, false, static_cast(1.0), - &col_matrix, static_cast(0.0)); + math::matmul(dev_ctx, filter, true, input_batch, false, + static_cast(1.0), &col_matrix, + static_cast(0.0)); - if (filter_shape_vec.size() == 2) { + if (data_dim == 2U) { // col2im: col_matrix -> dy // from (c * k_h * k_w, h * w) to (c, o_h, o_w) - col2im(context.device_context(), col, - std::vector{dilations[0], dilations[1]}, strides, - std::vector{paddings[0], paddings[1], paddings[0], - paddings[1]}, + col2im(dev_ctx, col, std::vector{dilations[0], dilations[1]}, + strides, std::vector{paddings[0], paddings[1], paddings[0], + paddings[1]}, &output_batch); - } else if (filter_shape_vec.size() == 3) { + } 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(context.device_context(), col, dilations, strides, paddings, - &output_batch); + col2vol(dev_ctx, col, dilations, strides, paddings, &output_batch); } } } }; -template +template class GemmConvTransposeGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { @@ -173,33 +165,30 @@ class GemmConvTransposeGradKernel : public framework::OpKernel { std::vector strides = context.Attr>("strides"); std::vector paddings = context.Attr>("paddings"); + std::vector dilations = context.Attr>("dilations"); const int batch_size = static_cast(input->dims()[0]); - // input_shape_vec: {h, w} or {d, h, w} + // input_shape_vec: {n, c, h, w} or {n, c, d, h, w} std::vector input_shape_vec = framework::vectorize(input->dims()); - input_shape_vec.erase(input_shape_vec.begin(), input_shape_vec.begin() + 2); - - // filter_shape_vec: {k_h, k_w} or {k_d, k_h, k_w} + // filter_shape_vec: {k_o, k_c, k_h, k_w} or {k_o, k_c, k_d, k_h, k_w} std::vector filter_shape_vec = framework::vectorize(filter.dims()); - filter_shape_vec.erase(filter_shape_vec.begin(), - filter_shape_vec.begin() + 2); // 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} - std::vector col_shape_vec; - col_shape_vec.push_back(output_grad->dims()[1]); - col_shape_vec.insert(col_shape_vec.end(), filter_shape_vec.begin(), - filter_shape_vec.end()); - col_shape_vec.insert(col_shape_vec.end(), input_shape_vec.begin(), - input_shape_vec.end()); + size_t data_dim = filter_shape_vec.size() - 2; + std::vector col_shape_vec(1 + 2 * data_dim); + col_shape_vec[0] = output_grad->dims()[1]; + 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]; + } 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) - DDim col_matrix_shape = - framework::flatten_to_2d(col_shape, filter_shape_vec.size() + 1); + DDim col_matrix_shape = framework::flatten_to_2d(col_shape, data_dim + 1); // output size: (c, o_h, o_w) or (c, o_d, o_h, o_w) DDim output_shape = framework::slice_ddim(output_grad->dims(), 1, @@ -215,6 +204,7 @@ class GemmConvTransposeGradKernel : public framework::OpKernel { // convolution transpose grad on input: // im2col + gemm (similar to conv-forward) // input need to compute gradient + auto& dev_ctx = context.template device_context(); if (input_grad || filter_grad) { Tensor col; col.mutable_data(col_shape, context.GetPlace()); @@ -226,19 +216,17 @@ class GemmConvTransposeGradKernel : public framework::OpKernel { col_matrix.Resize(col_matrix_shape); Tensor filter_grad_; - math::SetConstant set_zero; + math::SetConstant set_zero; - math::Im2ColFunctor im2col; - math::Vol2ColFunctor vol2col; - std::vector dilations({1, 1, 1}); + math::Im2ColFunctor im2col; + math::Vol2ColFunctor vol2col; if (input_grad) { input_grad->mutable_data(context.GetPlace()); - set_zero(context.device_context(), input_grad, static_cast(0)); } if (filter_grad) { // filter size (m, c, k_h, k_w) filter_grad->mutable_data(context.GetPlace()); - set_zero(context.device_context(), filter_grad, static_cast(0)); + set_zero(dev_ctx, filter_grad, static_cast(0)); filter_grad_ = *filter_grad; filter_grad_.Resize(filter_matrix_shape); } @@ -248,19 +236,19 @@ class GemmConvTransposeGradKernel : public framework::OpKernel { Tensor output_grad_batch = output_grad->Slice(i, i + 1).Resize(output_shape); - if (filter_shape_vec.size() == 2) { + if (data_dim == 2U) { // im2col: dy -> col matrix // from (c, o_h, o_w) to (c * k_h * k_w, h * w) - im2col(context.device_context(), output_grad_batch, + im2col(dev_ctx, output_grad_batch, std::vector{dilations[0], dilations[1]}, strides, std::vector{paddings[0], paddings[1], paddings[0], paddings[1]}, &col); - } else if (filter_shape_vec.size() == 3) { + } else if (data_dim == 3U) { // vol2col: dy -> col_matrix // from (c, o_d, o_h, o_w) to (c * k_d * k_h * k_w, d * h * w) - vol2col(context.device_context(), output_grad_batch, dilations, - strides, paddings, &col); + vol2col(dev_ctx, output_grad_batch, dilations, strides, paddings, + &col); } if (input_grad) { @@ -272,9 +260,9 @@ 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(context.device_context(), filter, false, - col_matrix, false, static_cast(1.0), - &input_grad_batch, static_cast(0.0)); + math::matmul( + dev_ctx, filter, false, col_matrix, false, static_cast(1.0), + &input_grad_batch, static_cast(0.0)); } if (filter_grad) { // input batch @@ -284,9 +272,9 @@ 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(context.device_context(), in_batch, false, - col_matrix, true, static_cast(1.0), - &filter_grad_, static_cast(1.0)); + math::matmul(dev_ctx, in_batch, false, col_matrix, + true, static_cast(1.0), + &filter_grad_, static_cast(1.0)); } } } diff --git a/paddle/operators/cos_sim_op.cc b/paddle/operators/cos_sim_op.cc index 312264ccd48d1405a247a2c864d9f5897c897bea..9019a1edb379be4007e38d3c0dc71feae23ae4e8 100644 --- a/paddle/operators/cos_sim_op.cc +++ b/paddle/operators/cos_sim_op.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/operators/cos_sim_op.h" @@ -62,7 +62,7 @@ class CosSimOp : public framework::OperatorWithKernel { class CosSimOpMaker : public framework::OpProtoAndCheckerMaker { public: - CosSimOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker) + CosSimOpMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "The 1st input of cos_sim op."); AddInput("Y", "The 2nd input of cos_sim op."); @@ -155,7 +155,8 @@ class CosSimOpGrad : public framework::OperatorWithKernel { namespace ops = paddle::operators; REGISTER_OP(cos_sim, ops::CosSimOp, ops::CosSimOpMaker, cos_sim_grad, ops::CosSimOpGrad); -REGISTER_OP_CPU_KERNEL(cos_sim, - ops::CosSimKernel); REGISTER_OP_CPU_KERNEL( - cos_sim_grad, ops::CosSimGradKernel); + cos_sim, ops::CosSimKernel); +REGISTER_OP_CPU_KERNEL( + cos_sim_grad, + ops::CosSimGradKernel); diff --git a/paddle/operators/cos_sim_op.cu b/paddle/operators/cos_sim_op.cu index 0cb8fd26de47a4a464db98664263544e3e503d63..9e5d1b6e4f0b6e482edd96df93d535e05dba3bc6 100644 --- a/paddle/operators/cos_sim_op.cu +++ b/paddle/operators/cos_sim_op.cu @@ -1,22 +1,23 @@ /* 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 +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 + 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. */ +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. */ #define EIGEN_USE_GPU #include "paddle/operators/cos_sim_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(cos_sim, - ops::CosSimKernel); -REGISTER_OP_GPU_KERNEL( - cos_sim_grad, ops::CosSimGradKernel); +REGISTER_OP_CUDA_KERNEL( + cos_sim, ops::CosSimKernel); +REGISTER_OP_CUDA_KERNEL( + cos_sim_grad, + ops::CosSimGradKernel); diff --git a/paddle/operators/cos_sim_op.h b/paddle/operators/cos_sim_op.h index 62a4e484eceeabc4cc26e68ac54a50be1ac95df7..eadcca55f9bfc3e59f329df8ff419ad4c5a29007 100644 --- a/paddle/operators/cos_sim_op.h +++ b/paddle/operators/cos_sim_op.h @@ -1,33 +1,29 @@ /* 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 +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 + 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. */ +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/framework/eigen.h" #include "paddle/framework/op_registry.h" +#include "paddle/operators/math/cos_sim_functor.h" +#include "paddle/operators/math/math_function.h" +#include "paddle/platform/for_range.h" namespace paddle { namespace operators { using Tensor = framework::Tensor; -template -using EigenMatrix = framework::EigenMatrix; -template -using EigenVector = framework::EigenVector; -template +template class CosSimKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { @@ -41,32 +37,30 @@ class CosSimKernel : public framework::OpKernel { out_x_norm->mutable_data(context.GetPlace()); out_y_norm->mutable_data(context.GetPlace()); - // convert Tensor to Eigen Tensor int rows_x = in_x->dims()[0]; int rows_y = in_y->dims()[0]; - auto x = EigenMatrix::Reshape(*in_x, 1); - auto y = EigenMatrix::Reshape(*in_y, 1); - auto z = EigenVector::Flatten(*out_z); - auto x_norm = EigenVector::Flatten(*out_x_norm); - auto y_norm = EigenVector::Flatten(*out_y_norm); - // compute - auto place = context.GetEigenDevice(); - auto row_along = Eigen::array({{1}}); - x_norm.device(place) = x.square().sum(row_along).sqrt(); - y_norm.device(place) = y.square().sum(row_along).sqrt(); + int cols = framework::product(in_x->dims()) / rows_x; + if (rows_x == rows_y) { - auto xy = (x * y).sum(Eigen::array({{1}})); - z.device(place) = xy / x_norm / y_norm; + math::CosSimFunctor functor( + in_x->data(), in_y->data(), out_x_norm->data(), + out_y_norm->data(), out_z->data(), cols); + platform::ForRange for_range( + static_cast(context.device_context()), rows_x); + for_range(functor); } else { - Eigen::DSizes bcast(rows_x, 1); - auto xy = (x * y.broadcast(bcast)).sum(row_along); - z.device(place) = xy / x_norm / y_norm.broadcast(bcast); + math::CosSimFunctor functor( + in_x->data(), in_y->data(), out_x_norm->data(), + out_y_norm->data(), out_z->data(), cols); + platform::ForRange for_range( + static_cast(context.device_context()), rows_x); + for_range(functor); } } }; -template +template class CosSimGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { @@ -80,61 +74,54 @@ class CosSimGradKernel : public framework::OpKernel { auto* out_grad_y = context.Output(framework::GradVarName("Y")); auto* in_grad_z = context.Input(framework::GradVarName("Out")); - // convert Tensor to Eigen Tensor - auto x = EigenMatrix::Reshape(*in_x, 1); - auto y = EigenMatrix::Reshape(*in_y, 1); - auto z = EigenMatrix::Reshape(*in_z, 1); - auto x_norm = EigenMatrix::Reshape(*in_x_norm, 1); - auto y_norm = EigenMatrix::Reshape(*in_y_norm, 1); - auto dz = EigenMatrix::Reshape(*in_grad_z, 1); - // compute gradident int rows_x = in_x->dims()[0]; int rows_y = in_y->dims()[0]; int cols = framework::product(in_x->dims()) / rows_x; - Eigen::DSizes bcast_cols(1, cols); - auto z_bcast = z.broadcast(bcast_cols); - auto dz_bcast = dz.broadcast(bcast_cols); - auto x_snorm_bcast = x_norm.square().eval().broadcast(bcast_cols); - auto place = context.GetEigenDevice(); + if (rows_x == rows_y) { - auto y_snorm_bcast = y_norm.square().eval().broadcast(bcast_cols); - auto norm_prod_bcast = (x_norm * y_norm).eval().broadcast(bcast_cols); - // compute dx if (out_grad_x) { - out_grad_x->mutable_data(context.GetPlace()); - auto dx = EigenMatrix::Reshape(*out_grad_x, 1); - auto grad = y / norm_prod_bcast - z_bcast * x / x_snorm_bcast; - dx.device(place) = dz_bcast * grad; + math::CosSimGradFunctor functor( + in_x_norm->data(), in_y_norm->data(), in_x->data(), + in_y->data(), in_z->data(), in_grad_z->data(), + out_grad_x->mutable_data(context.GetPlace()), cols); + platform::ForRange for_range( + static_cast(context.device_context()), + rows_x); + for_range(functor); } - // compute dy if (out_grad_y) { - out_grad_y->mutable_data(context.GetPlace()); - auto dy = EigenMatrix::Reshape(*out_grad_y, 1); - auto grad = x / norm_prod_bcast - z_bcast * y / y_snorm_bcast; - dy.device(place) = dz_bcast * grad; + math::CosSimGradFunctor functor( + in_y_norm->data(), in_x_norm->data(), in_y->data(), + in_x->data(), in_z->data(), in_grad_z->data(), + out_grad_y->mutable_data(context.GetPlace()), cols); + platform::ForRange for_range( + static_cast(context.device_context()), + rows_x); + for_range(functor); } } else { - Eigen::DSizes bcast_rows(rows_x, 1); - Eigen::DSizes bcast_rows_cols(rows_x, cols); - auto y_bcast = y.broadcast(bcast_rows); - auto y_snorm_bcast = y_norm.square().eval().broadcast(bcast_rows_cols); - auto norm_prod_bcast = (x_norm * y_norm.eval().broadcast(bcast_rows)) - .eval() - .broadcast(bcast_cols); - // compute dx if (out_grad_x) { - out_grad_x->mutable_data(context.GetPlace()); - auto dx = EigenMatrix::Reshape(*out_grad_x, 1); - auto grad = y_bcast / norm_prod_bcast - z_bcast * x / x_snorm_bcast; - dx.device(place) = dz_bcast * grad; + math::CosSimDxFunctor functor( + in_x_norm->data(), in_y_norm->data(), in_x->data(), + in_y->data(), in_z->data(), in_grad_z->data(), + out_grad_x->mutable_data(context.GetPlace()), cols); + platform::ForRange for_range( + static_cast(context.device_context()), + rows_x); + for_range(functor); } - // compute dy if (out_grad_y) { out_grad_y->mutable_data(context.GetPlace()); - auto dy = EigenVector::Flatten(*out_grad_y); - auto grad = x / norm_prod_bcast - z_bcast * y_bcast / y_snorm_bcast; - dy.device(place) = (dz_bcast * grad).sum(Eigen::array({{0}})); + math::SetConstant set_zero; + auto& dev_ctx = context.template device_context(); + set_zero(dev_ctx, out_grad_y, static_cast(0)); + + math::CosSimDyFunctor functor; + functor(dev_ctx, in_x_norm->data(), in_y_norm->data(), + in_x->data(), in_y->data(), in_z->data(), + in_grad_z->data(), static_cast(rows_x), + static_cast(cols), out_grad_y->data()); } } } diff --git a/paddle/operators/crf_decoding_op.cc b/paddle/operators/crf_decoding_op.cc index f418f489c0ff471464a23380598e9f4c8da16ca9..024e1d061a5b2eabc27110d5379cb8226a104079 100644 --- a/paddle/operators/crf_decoding_op.cc +++ b/paddle/operators/crf_decoding_op.cc @@ -18,8 +18,7 @@ namespace paddle { namespace operators { class CRFDecodingOpMaker : public framework::OpProtoAndCheckerMaker { public: - CRFDecodingOpMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker) + CRFDecodingOpMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("Emission", "(LoDTensor, default: LoDTensor). A LoDTensor with shape " @@ -36,17 +35,18 @@ class CRFDecodingOpMaker : public framework::OpProtoAndCheckerMaker { "w. See more details in comments of the linear_chain_crf operator."); AddInput( "Label", - "(LoDTensor, LoDTensor). The ground truth with shape " + "(LoDTensor, LoDTensor). The ground truth with shape " "[N x 1]. This input is optional. See more details in the operator's " "comments.") .AsDispensable(); - AddOutput("ViterbiPath", - "(LoDTensor, LoDTensor). The decoding results. What to " - "return changes depending on whether the Input(Label) (the groud " - "truth) is given. See more details in the operator's comment."); + AddOutput( + "ViterbiPath", + "(LoDTensor, LoDTensor). The decoding results. What to " + "return changes depending on whether the Input(Label) (the ground " + "truth) is given. See more details in the operator's comment."); AddComment(R"DOC( The crf_decoding operator reads the emission feature weights and the transition -freature weights learned by the linear_chain_crf operator. It implements the +feature weights learned by the linear_chain_crf operator. It implements the Viterbi algorithm which is a dynamic programming algorithm for finding the most likely sequence of hidden states, called the Viterbi path, that results in a sequence of observed tags. @@ -60,14 +60,14 @@ operator. When Input(Label) is given, the crf_decoding operator returns a row vector with shape [N x 1] whose values are fixed to be 0, indicating an incorrect -prediction, or 1 indicating a tag is correctly predicted. Such an ouput is the +prediction, or 1 indicating a tag is correctly predicted. Such an output is the input to chunk_eval operator. 2. Input(Label) is not given: This is the standard decoding process. -The crf_decoding operator returns a row vecotr with shape [N x 1] whose values +The crf_decoding operator returns a row vector with shape [N x 1] whose values range from 0 to maximum tag number - 1. Each element indicates an index of a predicted tag. )DOC"); @@ -120,12 +120,18 @@ class CRFDecodingOp : public framework::OperatorWithKernel { } protected: - framework::OpKernelType GetKernelType( + framework::OpKernelType GetActualKernelType( const framework::ExecutionContext& ctx) const override { return framework::OpKernelType( framework::ToDataType(ctx.Input("Emission")->type()), ctx.device_context()); } + + framework::OpKernelType GetExpectedKernelType( + const framework::OpKernelType& actual_kernel_type) const override { + return framework::OpKernelType(actual_kernel_type.data_type_, + platform::CPUPlace()); + } }; } // namespace operators } // namespace paddle @@ -134,5 +140,6 @@ namespace ops = paddle::operators; REGISTER_OP_WITHOUT_GRADIENT(crf_decoding, ops::CRFDecodingOp, ops::CRFDecodingOpMaker); REGISTER_OP_CPU_KERNEL( - crf_decoding, ops::CRFDecodingOpKernel, - ops::CRFDecodingOpKernel); + crf_decoding, + ops::CRFDecodingOpKernel, + ops::CRFDecodingOpKernel); diff --git a/paddle/operators/crf_decoding_op.h b/paddle/operators/crf_decoding_op.h index 526e0c5dcb2649b35ee28f5153c8472ca7a0af7b..f6827b7b1128251b2bb7e0a6a032389e5adc1371 100644 --- a/paddle/operators/crf_decoding_op.h +++ b/paddle/operators/crf_decoding_op.h @@ -24,7 +24,7 @@ using framework::LoDTensor; using framework::LoD; using framework::Tensor; -template +template class CRFDecodingOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { @@ -43,9 +43,9 @@ class CRFDecodingOpKernel : public framework::OpKernel { const size_t level = 0; const size_t seq_num = lod[level].size() - 1; - int* path = decoded_path->mutable_data(platform::CPUPlace()); - math::SetConstant()(ctx.device_context(), - decoded_path, 0); + int64_t* path = decoded_path->mutable_data(platform::CPUPlace()); + math::SetConstant()( + ctx.template device_context(), decoded_path, 0); for (size_t i = 0; i < seq_num; ++i) { int start_pos = static_cast(lod[level][i]); int end_pos = static_cast(lod[level][i + 1]); @@ -57,7 +57,7 @@ class CRFDecodingOpKernel : public framework::OpKernel { if (label) { PADDLE_ENFORCE_EQ(label->NumLevels(), 1UL, "The Input(Label) should be a sequence."); - const int* label_value = label->data(); + const int64_t* label_value = label->data(); size_t batch_size = emission_weights->dims()[0]; for (size_t i = 0; i < batch_size; ++i) { path[i] = label_value[i] == path[i] ? 1 : 0; @@ -76,7 +76,7 @@ class CRFDecodingOpKernel : public framework::OpKernel { const T* x = emission_weights.data(); const T* w = transition_weights.data(); - int* path = decoded_path->data(); + int64_t* path = decoded_path->data(); // alpha is a memo table. An element alpha(k, v) records the score of the // best sequence of tags from position 1 to position k with v being the end diff --git a/paddle/operators/crop_op.cc b/paddle/operators/crop_op.cc index 6752eb8c1c72150b0b1cf5595211ca1d01ef2bf4..310e351443112c340054cf092cd2443b309ec49c 100644 --- a/paddle/operators/crop_op.cc +++ b/paddle/operators/crop_op.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/operators/crop_op.h" #include @@ -52,7 +52,7 @@ class CropOp : public framework::OperatorWithKernel { class CropOpMaker : public framework::OpProtoAndCheckerMaker { public: - CropOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker) + CropOpMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "The input of pad op. " @@ -88,7 +88,8 @@ There are two ways to set shape: The input should be a k-D tensor(k > 0 and k < 7). As an example: -Given: +Case 1: +Given X = [[0, 1, 2, 0, 0] [0, 3, 4, 0, 0] @@ -107,6 +108,27 @@ we get: Out = [[1, 2], [3, 4]]. + +Case 2: +Given + + X = [[0, 1, 2, 5, 0] + [0, 3, 4, 6, 0] + [0, 0, 0, 0, 0]], + +and + + offsets = [0, 1], + +and + + Y = [[0, 0, 0] + [0, 0, 0]], + +we get: + + Out = [[1, 2, 5], + [3, 4, 6]]. )DOC"); } }; @@ -133,5 +155,5 @@ class CropOpGrad : public framework::OperatorWithKernel { namespace ops = paddle::operators; REGISTER_OP(crop, ops::CropOp, ops::CropOpMaker, crop_grad, ops::CropOpGrad); REGISTER_OP_CPU_KERNEL(crop, ops::CropKernel); -REGISTER_OP_CPU_KERNEL(crop_grad, - ops::CropGradKernel); +REGISTER_OP_CPU_KERNEL( + crop_grad, ops::CropGradKernel); diff --git a/paddle/operators/crop_op.cu b/paddle/operators/crop_op.cu index f8ee18a1d6e894cbb2d71dd4b6b459abeb076817..bba5db4c6ce682cb00482e35fa1e340aba83e37f 100644 --- a/paddle/operators/crop_op.cu +++ b/paddle/operators/crop_op.cu @@ -1,21 +1,21 @@ /* 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 +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 + 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. */ +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. */ #define EIGEN_USE_GPU #include "paddle/operators/crop_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(crop, ops::CropKernel); -REGISTER_OP_GPU_KERNEL(crop_grad, - ops::CropGradKernel); +REGISTER_OP_CUDA_KERNEL(crop, ops::CropKernel); +REGISTER_OP_CUDA_KERNEL( + crop_grad, ops::CropGradKernel); diff --git a/paddle/operators/crop_op.h b/paddle/operators/crop_op.h index 2e72583d68d0acf0e2f5044637dba55de3b57209..69d1a92977250b4e8a64b47ac66444724fbc53f6 100644 --- a/paddle/operators/crop_op.h +++ b/paddle/operators/crop_op.h @@ -1,16 +1,16 @@ -/* Copyright (c) 2016 CropdleCropdle Authors. All Rights Reserve. +/* 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 +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 + 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. */ +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 @@ -49,7 +49,7 @@ class CropKernel : public framework::OpKernel { } }; -template +template void CropGradFunction(const framework::ExecutionContext& context) { auto* d_x = context.Output(framework::GradVarName("X")); if (d_x != nullptr) { @@ -63,12 +63,13 @@ void CropGradFunction(const framework::ExecutionContext& context) { } auto d_x_tensor = EigenTensor::From(*d_x); auto d_out_tensor = EigenTensor::From(*d_out); - d_x_tensor.device(context.GetEigenDevice()) = + d_x_tensor.device( + *context.template device_context().eigen_device()) = d_out_tensor.pad(paddings, 0); } } -template +template class CropGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { @@ -76,22 +77,22 @@ class CropGradKernel : public framework::OpKernel { context.Input(framework::GradVarName("Out"))->dims().size(); switch (rank) { case 1: - CropGradFunction(context); + CropGradFunction(context); break; case 2: - CropGradFunction(context); + CropGradFunction(context); break; case 3: - CropGradFunction(context); + CropGradFunction(context); break; case 4: - CropGradFunction(context); + CropGradFunction(context); break; case 5: - CropGradFunction(context); + CropGradFunction(context); break; case 6: - CropGradFunction(context); + CropGradFunction(context); break; default: PADDLE_THROW( diff --git a/paddle/operators/cross_entropy_op.cc b/paddle/operators/cross_entropy_op.cc index 1e82742eaf86711fe4f9d02d517ad1853131cf67..fe39cb481aa63fa401603d97778b337282511ab5 100644 --- a/paddle/operators/cross_entropy_op.cc +++ b/paddle/operators/cross_entropy_op.cc @@ -51,7 +51,7 @@ class CrossEntropyOp : public framework::OperatorWithKernel { protected: // Explicitly set that the data type of computation kernel of cross_entropy // is determined by its input "X". - framework::OpKernelType GetKernelType( + framework::OpKernelType GetActualKernelType( const framework::ExecutionContext& ctx) const override { return framework::OpKernelType( framework::ToDataType(ctx.Input("X")->type()), @@ -95,12 +95,13 @@ class CrossEntropyGradientOp : public framework::OperatorWithKernel { "Input(Label) should be 1."); } ctx->SetOutputDim(framework::GradVarName("X"), x_dims); + ctx->ShareLoD("X", framework::GradVarName("X")); } protected: // Explicitly set that the data type of computation kernel of cross_entropy // is determined by its input "X". - framework::OpKernelType GetKernelType( + framework::OpKernelType GetActualKernelType( const framework::ExecutionContext& ctx) const override { return framework::OpKernelType( framework::ToDataType(ctx.Input("X")->type()), @@ -110,19 +111,18 @@ class CrossEntropyGradientOp : public framework::OperatorWithKernel { class CrossEntropyOpMaker : public framework::OpProtoAndCheckerMaker { public: - CrossEntropyOpMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker) + CrossEntropyOpMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { 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. " + "(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. " "This input is a probability computed by the previous operator, " "which is almost always the result of a softmax operator."); AddInput("Label", "(Tensor), the ground truth which is a 2-D tensor. When " "soft_label is set to false, Label is a Tensor with shape " "[N x 1]. When soft_label is set to true, Label is a " - "Tensor with shape [N x K]."); + "Tensor with shape [N x D]."); AddOutput("Y", "(Tensor, default Tensor), a 2-D tensor with shape " "[N x 1]. The cross entropy loss."); diff --git a/paddle/operators/cross_entropy_op.cu b/paddle/operators/cross_entropy_op.cu index 6212e39dfde33c5943958adbd1a0a052262e119e..3b04894e6ccb08c13e2d24bb38196fdc7935bf9e 100644 --- a/paddle/operators/cross_entropy_op.cu +++ b/paddle/operators/cross_entropy_op.cu @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/operators/cross_entropy_op.h" @@ -53,8 +53,9 @@ class CrossEntropyOpCUDAKernel : public framework::OpKernel { Tensor* y = ctx.Output("Y"); y->mutable_data(ctx.GetPlace()); - math::CrossEntropyFunctor()( - ctx.device_context(), y, x, label, ctx.Attr("soft_label")); + math::CrossEntropyFunctor()( + ctx.template device_context(), y, x, label, + ctx.Attr("soft_label")); } }; @@ -80,15 +81,17 @@ class CrossEntropyGradientOpCUDAKernel : public framework::OpKernel { int block = 512; int grid = (batch_size * class_num + block - 1) / block; - auto stream = ctx.cuda_device_context().stream(); + + 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(ctx.device_context(), dx, 0); + math::SetConstant functor; + functor(dev_ctx, dx, 0); auto* label_data = label->data(); grid = (batch_size + block - 1) / block; CrossEntropyGradientKernel<<>>( @@ -101,8 +104,8 @@ class CrossEntropyGradientOpCUDAKernel : public framework::OpKernel { } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(cross_entropy, ops::CrossEntropyOpCUDAKernel, - ops::CrossEntropyOpCUDAKernel); -REGISTER_OP_GPU_KERNEL(cross_entropy_grad, - ops::CrossEntropyGradientOpCUDAKernel, - ops::CrossEntropyGradientOpCUDAKernel); +REGISTER_OP_CUDA_KERNEL(cross_entropy, ops::CrossEntropyOpCUDAKernel, + ops::CrossEntropyOpCUDAKernel); +REGISTER_OP_CUDA_KERNEL(cross_entropy_grad, + ops::CrossEntropyGradientOpCUDAKernel, + ops::CrossEntropyGradientOpCUDAKernel); diff --git a/paddle/operators/cross_entropy_op.h b/paddle/operators/cross_entropy_op.h index 37db0a930a6aea0ba333395ca9c5b9d231c07b32..5623d2ded16daaf51dd26c9d9a8c04a0ae5be5ec 100644 --- a/paddle/operators/cross_entropy_op.h +++ b/paddle/operators/cross_entropy_op.h @@ -37,8 +37,9 @@ class CrossEntropyOpKernel : public framework::OpKernel { Tensor* y = ctx.Output("Y"); y->mutable_data(ctx.GetPlace()); - math::CrossEntropyFunctor()( - ctx.device_context(), y, x, labels, ctx.Attr("soft_label")); + math::CrossEntropyFunctor()( + ctx.template device_context(), y, x, labels, + ctx.Attr("soft_label")); } }; @@ -61,7 +62,8 @@ class CrossEntropyGradientOpKernel : public framework::OpKernel { auto lbl_mat = EigenMatrix::From(*label); auto dx_mat = EigenMatrix::From(*dx); - dx_mat.device(ctx.GetEigenDevice()) = + dx_mat.device(*ctx.template device_context() + .eigen_device()) = -(lbl_mat * dy_mat.broadcast(Eigen::DSizes(1, class_num)) / x_mat); } else { @@ -70,8 +72,8 @@ class CrossEntropyGradientOpKernel : public framework::OpKernel { const T* x_data = x->data(); const int64_t* label_data = label->data(); - math::SetConstant functor; - functor(ctx.device_context(), dx, 0); + 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); diff --git a/paddle/operators/decayed_adagrad_op.cc b/paddle/operators/decayed_adagrad_op.cc index 640b4e77448d1b64bcf7375f26c07ff1d2bdeaa3..739a8d881c35817756421a3299901c9e5e7d96ba 100644 --- a/paddle/operators/decayed_adagrad_op.cc +++ b/paddle/operators/decayed_adagrad_op.cc @@ -55,8 +55,7 @@ class DecayedAdagradOp : public framework::OperatorWithKernel { class DecayedAdagradOpMaker : public framework::OpProtoAndCheckerMaker { public: - DecayedAdagradOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + DecayedAdagradOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("Param", "(Tensor) Input parameter"); AddInput("Grad", "(Tensor) Input gradient"); @@ -99,4 +98,4 @@ REGISTER_OP_WITHOUT_GRADIENT(decayed_adagrad, ops::DecayedAdagradOp, ops::DecayedAdagradOpMaker); REGISTER_OP_CPU_KERNEL( decayed_adagrad, - ops::DecayedAdagradOpKernel); + ops::DecayedAdagradOpKernel); diff --git a/paddle/operators/decayed_adagrad_op.cu b/paddle/operators/decayed_adagrad_op.cu index 6fce77fe4ec6b76cb7b0259aab6a3d55d2edb36c..7bc8161f2339572c2a9284f865846b9b7e594354 100644 --- a/paddle/operators/decayed_adagrad_op.cu +++ b/paddle/operators/decayed_adagrad_op.cu @@ -1,21 +1,21 @@ /* 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 +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 + 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. */ +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. */ #define EIGEN_USE_GPU #include "paddle/operators/decayed_adagrad_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL( +REGISTER_OP_CUDA_KERNEL( decayed_adagrad, - ops::DecayedAdagradOpKernel); + ops::DecayedAdagradOpKernel); diff --git a/paddle/operators/decayed_adagrad_op.h b/paddle/operators/decayed_adagrad_op.h index 0fe0fc5acd66c9824a864618b69097c5c063ea3f..fec9705cfc1e14e5423e23d6afb218c6c051f5a1 100644 --- a/paddle/operators/decayed_adagrad_op.h +++ b/paddle/operators/decayed_adagrad_op.h @@ -19,7 +19,7 @@ limitations under the License. */ namespace paddle { namespace operators { -template +template class DecayedAdagradOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { @@ -43,7 +43,7 @@ class DecayedAdagradOpKernel : public framework::OpKernel { auto param_out = framework::EigenVector::Flatten(*param_out_tensor); auto moment_out = framework::EigenVector::Flatten(*moment_out_tensor); - auto place = ctx.GetEigenDevice(); + auto& place = *ctx.template device_context().eigen_device(); moment_out.device(place) = decay * moment + (1 - decay) * grad * grad; Eigen::DSizes m_dsize(moment_out_tensor->numel()); diff --git a/paddle/operators/detail/CMakeLists.txt b/paddle/operators/detail/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..f6bdc63cc2cfae526fe911ee4d989675452d5c5d --- /dev/null +++ b/paddle/operators/detail/CMakeLists.txt @@ -0,0 +1 @@ +grpc_library(sendrecvop_grpc SRCS recv_impl.cc send_impl.cc PROTO send_recv.proto DEPS lod_tensor selected_rows) diff --git a/paddle/operators/detail/recv_impl.cc b/paddle/operators/detail/recv_impl.cc new file mode 100644 index 0000000000000000000000000000000000000000..b746f9df4640bac7b5c4e83091f2c2ab67e778be --- /dev/null +++ b/paddle/operators/detail/recv_impl.cc @@ -0,0 +1,76 @@ +/* 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. */ + +#include "send_recv_impl.h" + +namespace paddle { +namespace operators { +namespace detail { + +Status SendRecvServerImpl::SendVariable(ServerContext *context, + const VariableMessage *in_var, + VoidMessage *out_var) { + // TODO(typhoonzero): support different variable types. + std::istringstream iss(in_var->serialized()); + framework::LoDTensor t; + framework::DeserializeFromStream(iss, &t); + TensorWithName tensor_with_name = + std::make_pair(in_var->varname(), std::move(t)); + + var_recv_queue_.Push(std::move(tensor_with_name)); + return Status::OK; +} + +Status SendRecvServerImpl::GetVariable(ServerContext *context, + const VariableMessage *in_var, + VariableMessage *out_var) { + std::string get_var_name = in_var->varname(); + auto *var = scope_->FindVar(get_var_name); + auto tensor = var->Get(); + std::ostringstream oss; + framework::SerializeToStream(oss, tensor, platform::CPUDeviceContext()); + + std::string *varname = out_var->mutable_varname(); + *varname = get_var_name; + std::string *serialized = out_var->mutable_serialized(); + *serialized = oss.str(); + return Status::OK; +} + +Status SendRecvServerImpl::Wait(ServerContext *context, + const VoidMessage *in_var, + VoidMessage *out_var) { + { + std::unique_lock lock(this->mutex_); + condition_.wait(lock, [=] { return this->done_ == true; }); + } + return Status::OK; +} + +void SendRecvServerImpl::Reset() { + std::lock_guard lock(this->mutex_); + done_ = false; +} + +void SendRecvServerImpl::Done() { + { + std::lock_guard lock(this->mutex_); + done_ = true; + } + condition_.notify_all(); +} + +} // namespace detail +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/detail/safe_ref.h b/paddle/operators/detail/safe_ref.h index b71af17309f9f46b5c87f0f479d4e03443fa7f93..ff2a156f3d0e3c5c55354375ca81adf433a49686 100644 --- a/paddle/operators/detail/safe_ref.h +++ b/paddle/operators/detail/safe_ref.h @@ -1,16 +1,16 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +/* 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 +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 + 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. */ +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 diff --git a/paddle/operators/detail/send_impl.cc b/paddle/operators/detail/send_impl.cc new file mode 100644 index 0000000000000000000000000000000000000000..a812fcf39bc19a6e06dc594c13076696e7949372 --- /dev/null +++ b/paddle/operators/detail/send_impl.cc @@ -0,0 +1,77 @@ +/* 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. */ + +#include "send_recv_impl.h" + +namespace paddle { +namespace operators { +namespace detail { + +bool RPCClient::SendVariable(const framework::Scope& scope, + const std::string& inname) { + ClientContext context; + VariableMessage msg; + VoidMessage out_msg; + // FIXME(typhoonzero): pass device context to here. + auto ctx = platform::CPUDeviceContext(); + auto* var = scope.FindVar(inname); + PADDLE_ENFORCE(var); + // TODO(typhoonzero): support SelectedRows + PADDLE_ENFORCE(var->IsType(), + "Only support LoDTensor, %s has wrong type", inname); + const framework::LoDTensor& tensor = var->Get(); + std::ostringstream oss; + framework::SerializeToStream(oss, tensor, ctx); + msg.set_varname(inname); + msg.set_serialized(oss.str()); + Status status = stub_->SendVariable(&context, msg, &out_msg); + if (!status.ok()) { + LOG(ERROR) << "gRPC error: " << status.error_message(); + return false; + } + return true; +} + +bool RPCClient::GetVariable(const framework::Scope& scope, + const std::string& outname) { + ClientContext context; + VariableMessage call_msg, ret_msg; + call_msg.set_varname(outname); + auto ctx = platform::CPUDeviceContext(); + Status status = stub_->GetVariable(&context, call_msg, &ret_msg); + if (!status.ok()) { + LOG(ERROR) << "gRPC error: " << status.error_message(); + return false; + } + + std::istringstream iss(ret_msg.serialized()); + + framework::LoDTensor ret_tensor; + framework::DeserializeFromStream(iss, &ret_tensor); + auto* outvar = scope.FindVar(outname); + framework::LoDTensor* out_tensor = outvar->GetMutable(); + // FIXME(typhoonzero): do not copy. + framework::CopyFrom(ret_tensor, ctx.GetPlace(), ctx, out_tensor); + return true; +} + +void RPCClient::Wait() { + ClientContext context; + VoidMessage call_msg, ret_msg; + stub_->Wait(&context, call_msg, &ret_msg); +} + +} // namespace detail +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/detail/send_recv.proto b/paddle/operators/detail/send_recv.proto new file mode 100644 index 0000000000000000000000000000000000000000..95c8e708986eac2a27ece70864efd7eac5f84ce8 --- /dev/null +++ b/paddle/operators/detail/send_recv.proto @@ -0,0 +1,40 @@ +/* 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. */ + +syntax = "proto3"; + +package sendrecv; + +service SendRecvService { + // For parameter server round-robin like hashing, do not split tensors. + // Send and recv only one tensor + // TODO(typhoonzero): add streaming API + rpc SendVariable(VariableMessage) returns (VoidMessage) {} + // Argument VariableMessage for GetVariable should only contain varname. + rpc GetVariable(VariableMessage) returns (VariableMessage) {} + // wait for one execution of the program + rpc Wait(VoidMessage) returns (VoidMessage) {} +} + +// VariableMessage is serialized paddle variable message. +// It can be: +// Tensor +// LoDTensor +// SelectedRows +message VariableMessage { + string varname = 1; + bytes serialized = 2; +} + +message VoidMessage {} diff --git a/paddle/operators/detail/send_recv_impl.h b/paddle/operators/detail/send_recv_impl.h new file mode 100644 index 0000000000000000000000000000000000000000..47f730f7ae897096fbdd23a55252448cf6655fb4 --- /dev/null +++ b/paddle/operators/detail/send_recv_impl.h @@ -0,0 +1,94 @@ +/* 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 + +#include "paddle/framework/data_type.h" +#include "paddle/framework/lod_tensor.h" +#include "paddle/framework/scope.h" +#include "paddle/framework/selected_rows.h" +#include "paddle/operators/detail/simple_block_queue.h" + +#include "paddle/operators/detail/send_recv.grpc.pb.h" +#include "paddle/operators/detail/send_recv.pb.h" + +#include + +using grpc::Channel; +using grpc::Server; +using grpc::ServerContext; +using grpc::ServerReader; +using grpc::ServerBuilder; + +using grpc::ClientContext; +using grpc::ClientReader; +using grpc::ClientReaderWriter; +using grpc::ClientWriter; +using grpc::Status; +using sendrecv::SendRecvService; +using sendrecv::VariableMessage; +using sendrecv::VoidMessage; + +namespace paddle { +namespace operators { +namespace detail { + +typedef std::pair TensorWithName; + +class SendRecvServerImpl final : public SendRecvService::Service { + public: + explicit SendRecvServerImpl() {} + + Status SendVariable(ServerContext *context, const VariableMessage *in_var, + VoidMessage *out_var) override; + Status GetVariable(ServerContext *context, const VariableMessage *in_var, + VariableMessage *out_var) override; + Status Wait(ServerContext *context, const VoidMessage *in_var, + VoidMessage *out_var) override; + void Reset(); + void Done(); + void SetScope(framework::Scope *scope) { scope_ = scope; }; + + const TensorWithName Get() { return this->var_recv_queue_.Pop(); } + + void Push(const TensorWithName &msg) { this->var_recv_queue_.Push(msg); } + + private: + // received variable from RPC, operators fetch variable from this queue. + SimpleBlockQueue var_recv_queue_; + framework::Scope *scope_; + // condition of the sub program + std::mutex mutex_; + bool done_; + std::condition_variable condition_; +}; + +// RPCClient is a class to send tensors to pserver sub-network +// using different hashing methods. +class RPCClient { + public: + RPCClient(std::shared_ptr channel) + : stub_(SendRecvService::NewStub(channel)) {} + + bool SendVariable(const framework::Scope &scope, const std::string &inname); + bool GetVariable(const framework::Scope &scope, const std::string &outname); + void Wait(); + + private: + std::unique_ptr stub_; +}; + +} // namespace detail +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/detail/simple_block_queue.h b/paddle/operators/detail/simple_block_queue.h new file mode 100644 index 0000000000000000000000000000000000000000..c7f5ff4b5f494ce80b4ee792afb44841e9c8a2f2 --- /dev/null +++ b/paddle/operators/detail/simple_block_queue.h @@ -0,0 +1,52 @@ +/* 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 + +#include +#include +#include + +namespace paddle { +namespace operators { +namespace detail { + +template +class SimpleBlockQueue { + private: + std::mutex mutex_; + std::condition_variable condition_; + std::deque queue_; + + public: + void Push(T const& value) { + { + std::unique_lock lock(this->mutex_); + queue_.push_front(value); + } + this->condition_.notify_one(); + } + + T Pop() { + std::unique_lock lock(this->mutex_); + this->condition_.wait(lock, [=] { return !this->queue_.empty(); }); + T rc(std::move(this->queue_.back())); + this->queue_.pop_back(); + return rc; + } +}; + +} // namespace detail +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/detail/strided_memcpy.h b/paddle/operators/detail/strided_memcpy.h index 068c82f399316a1587d7322d8dab75823656800e..9ed524d4dcf7f8bd4607281ade34e9d56f409085 100644 --- a/paddle/operators/detail/strided_memcpy.h +++ b/paddle/operators/detail/strided_memcpy.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/framework/ddim.h" @@ -35,7 +35,7 @@ struct StridedMemcpyFunctor { memory::Copy(cpu_place, dst, cpu_place, src, sizeof(T) * dst_dim.head); } else { #ifdef PADDLE_WITH_CUDA - auto& gpu_place = boost::get(place); + auto& gpu_place = boost::get(place); auto& cuda_ctx = reinterpret_cast(dev_ctx); memory::Copy(gpu_place, dst, gpu_place, src, sizeof(T) * dst_dim.head, diff --git a/paddle/operators/detection_output_op.cc b/paddle/operators/detection_output_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..ea44cd32678d7e8a5836c1886cf9c1b4961970aa --- /dev/null +++ b/paddle/operators/detection_output_op.cc @@ -0,0 +1,89 @@ +/* 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. +Indicesou 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/operators/detection_output_op.h" +namespace paddle { +namespace operators { + +class DetectionOutputOpMaker : public framework::OpProtoAndCheckerMaker { + public: + DetectionOutputOpMaker(OpProto* proto, OpAttrChecker* op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("Loc", + "(Tensor) The input tensor of detection_output operator." + "The input predict locations" + "The format of input tensor is kNCHW. Where K is priorbox point " + "numbers," + "N is How many boxes are there on each point, " + "C is 4, H and W both are 1."); + AddInput("Conf", + "(Tensor) The input tensor of detection_output operator." + "The input priorbox confidence." + "The format of input tensor is kNCHW. Where K is priorbox point " + "numbers," + "N is How many boxes are there on each point, " + "C is the number of classes, H and W both are 1."); + AddInput("PriorBox", + "(Tensor) The input tensor of detection_output operator." + "The format of input tensor is the position and variance " + "of the boxes"); + AddOutput("Out", + "(Tensor) The output tensor of detection_output operator."); + AddAttr("background_label_id", "(int), The background class index."); + AddAttr("num_classes", "(int), The number of the classification."); + AddAttr("nms_threshold", + "(float), The Non-maximum suppression threshold."); + AddAttr("confidence_threshold", + "(float), The classification confidence threshold."); + AddAttr("top_k", "(int), The bbox number kept of the layer’s output."); + AddAttr("nms_top_k", + "(int), The bbox number kept of the NMS’s output."); + AddComment(R"DOC( + detection output for SSD(single shot multibox detector) + Apply the NMS to the output of network and compute the predict + bounding box location. The output’s shape of this layer could + be zero if there is no valid bounding box. + )DOC"); + } +}; + +class DetectionOutputOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Loc"), + "Input(X) of DetectionOutputOp" + "should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Conf"), + "Input(X) of DetectionOutputOp" + "should not be null."); + PADDLE_ENFORCE(ctx->HasInput("PriorBox"), + "Input(X) of DetectionOutputOp" + "should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of DetectionOutputOp should not be null."); + std::vector output_shape({1, 7}); + ctx->SetOutputDim("Out", framework::make_ddim(output_shape)); + } +}; +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP_WITHOUT_GRADIENT(detection_output, ops::DetectionOutputOp, + ops::DetectionOutputOpMaker); +REGISTER_OP_CPU_KERNEL( + detection_output, + ops::DetectionOutputKernel, + ops::DetectionOutputKernel); diff --git a/paddle/operators/detection_output_op.cu.cc b/paddle/operators/detection_output_op.cu.cc new file mode 100644 index 0000000000000000000000000000000000000000..4a6560e0492c559afd06e5152c34fab545b7ce61 --- /dev/null +++ b/paddle/operators/detection_output_op.cu.cc @@ -0,0 +1,21 @@ +/* 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. +Indicesou 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/operators/detection_output_op.h" + +namespace ops = paddle::operators; +REGISTER_OP_CUDA_KERNEL( + detection_output, + ops::DetectionOutputKernel, + ops::DetectionOutputKernel); diff --git a/paddle/operators/detection_output_op.h b/paddle/operators/detection_output_op.h new file mode 100644 index 0000000000000000000000000000000000000000..f8abd5b6406f05747b87fcfd464baeb705a7f7f2 --- /dev/null +++ b/paddle/operators/detection_output_op.h @@ -0,0 +1,167 @@ +/* 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. +Indicesou 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/framework/op_registry.h" +#include "paddle/framework/tensor.h" +#include "paddle/operators/math/detection_util.h" +#include "paddle/operators/math/math_function.h" +#include "paddle/operators/math/softmax.h" +#include "paddle/operators/strided_memcpy.h" +namespace paddle { +namespace operators { +template +inline void transpose_fun(const framework::ExecutionContext& context, + const framework::Tensor& src, + framework::Tensor* dst) { + int input_nums = src.dims()[0]; + int offset = 0; + for (int j = 0; j < input_nums; ++j) { + framework::Tensor in_p_tensor = src.Slice(j, j + 1); + std::vector shape_vec( + {in_p_tensor.dims()[0], in_p_tensor.dims()[1], in_p_tensor.dims()[3], + in_p_tensor.dims()[4], in_p_tensor.dims()[2]}); + framework::DDim shape(framework::make_ddim(shape_vec)); + framework::Tensor in_p_tensor_transpose; + in_p_tensor_transpose.mutable_data(shape, context.GetPlace()); + std::vector shape_axis({0, 1, 3, 4, 2}); + math::Transpose trans5; + trans5(context.template device_context(), in_p_tensor, + &in_p_tensor_transpose, shape_axis); + auto dst_stride = framework::stride(dst->dims()); + auto src_stride = framework::stride(in_p_tensor_transpose.dims()); + StridedMemcpy(context.device_context(), in_p_tensor_transpose.data(), + src_stride, in_p_tensor_transpose.dims(), dst_stride, + dst->data() + offset); + offset += in_p_tensor_transpose.dims()[4] * src_stride[4]; + } +} +template +class DetectionOutputKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + const framework::Tensor* in_loc = context.Input("Loc"); + const framework::Tensor* in_conf = context.Input("Conf"); + const framework::Tensor* in_priorbox = + context.Input("PriorBox"); + auto* out = context.Output("Out"); + int num_classes = context.template Attr("num_classes"); + int top_k = context.template Attr("top_k"); + int nms_top_k = context.template Attr("nms_top_k"); + int background_label_id = context.template Attr("background_label_id"); + float nms_threshold = context.template Attr("nms_threshold"); + float confidence_threshold = + context.template Attr("confidence_threshold"); + size_t batch_size = in_conf->dims()[1]; + int conf_sum_size = in_conf->numel(); + // for softmax + std::vector conf_shape_softmax_vec( + {conf_sum_size / num_classes, num_classes}); + framework::DDim conf_shape_softmax( + framework::make_ddim(conf_shape_softmax_vec)); + // for knchw => nhwc + std::vector loc_shape_vec({1, in_loc->dims()[1], in_loc->dims()[3], + in_loc->dims()[4], + in_loc->dims()[2] * in_loc->dims()[0]}); + std::vector conf_shape_vec( + {1, in_conf->dims()[1], in_conf->dims()[3], in_conf->dims()[4], + in_conf->dims()[2] * in_conf->dims()[0]}); + framework::DDim loc_shape(framework::make_ddim(loc_shape_vec)); + framework::DDim conf_shape(framework::make_ddim(conf_shape_vec)); + framework::Tensor loc_tensor; + framework::Tensor conf_tensor; + loc_tensor.mutable_data(loc_shape, context.GetPlace()); + conf_tensor.mutable_data(conf_shape, context.GetPlace()); + // for cpu + framework::Tensor loc_cpu; + framework::Tensor conf_cpu; + framework::Tensor priorbox_cpu; + const T* priorbox_data = in_priorbox->data(); + transpose_fun(context, *in_loc, &loc_tensor); + transpose_fun(context, *in_conf, &conf_tensor); + conf_tensor.Resize(conf_shape_softmax); + math::SoftmaxFunctor()( + context.template device_context(), &conf_tensor, + &conf_tensor); + T* loc_data = loc_tensor.data(); + T* conf_data = conf_tensor.data(); + if (platform::is_gpu_place(context.GetPlace())) { + loc_cpu.mutable_data(loc_tensor.dims(), platform::CPUPlace()); + framework::CopyFrom(loc_tensor, platform::CPUPlace(), + context.device_context(), &loc_cpu); + loc_data = loc_cpu.data(); + conf_cpu.mutable_data(conf_tensor.dims(), platform::CPUPlace()); + framework::CopyFrom(conf_tensor, platform::CPUPlace(), + context.device_context(), &conf_cpu); + conf_data = conf_cpu.data(); + priorbox_cpu.mutable_data(in_priorbox->dims(), platform::CPUPlace()); + framework::CopyFrom(*in_priorbox, platform::CPUPlace(), + context.device_context(), &priorbox_cpu); + priorbox_data = priorbox_cpu.data(); + } + // get decode bboxes + size_t num_priors = in_priorbox->numel() / 8; + std::vector>> all_decoded_bboxes; + for (size_t n = 0; n < batch_size; ++n) { + std::vector> decoded_bboxes; + for (size_t i = 0; i < num_priors; ++i) { + size_t prior_offset = i * 8; + size_t loc_pred_offset = n * num_priors * 4 + i * 4; + std::vector> prior_bbox_vec; + math::GetBBoxFromPriorData(priorbox_data + prior_offset, 1, + prior_bbox_vec); + std::vector> prior_bbox_var; + math::GetBBoxVarFromPriorData(priorbox_data + prior_offset, 1, + prior_bbox_var); + std::vector loc_pred_data; + for (size_t j = 0; j < 4; ++j) + loc_pred_data.push_back(*(loc_data + loc_pred_offset + j)); + math::BBox bbox = math::DecodeBBoxWithVar( + prior_bbox_vec[0], prior_bbox_var[0], loc_pred_data); + decoded_bboxes.push_back(bbox); + } + all_decoded_bboxes.push_back(decoded_bboxes); + } + std::vector>> all_indices; + int num_kept = math::GetDetectionIndices( + conf_data, num_priors, num_classes, background_label_id, batch_size, + confidence_threshold, nms_top_k, nms_threshold, top_k, + all_decoded_bboxes, &all_indices); + + if (num_kept <= 0) { + std::vector out_shape_vec({0, 0}); + framework::DDim out_shape(framework::make_ddim(out_shape_vec)); + out->Resize(out_shape); + return; + } + std::vector out_shape_vec({num_kept, 7}); + framework::DDim out_shape(framework::make_ddim(out_shape_vec)); + out->mutable_data(out_shape, context.GetPlace()); + framework::Tensor out_cpu; + T* out_data = out->data(); + if (platform::is_gpu_place(context.GetPlace())) { + out_cpu.mutable_data(out->dims(), platform::CPUPlace()); + out_data = out_cpu.data(); + } + math::GetDetectionOutput(conf_data, num_kept, num_priors, num_classes, + batch_size, all_indices, all_decoded_bboxes, + out_data); + if (platform::is_gpu_place(context.GetPlace())) { + framework::CopyFrom(out_cpu, platform::CUDAPlace(), + context.device_context(), out); + } + } +}; +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/dropout_op.cc b/paddle/operators/dropout_op.cc index 818146aca766cb13b93fd024c11c1209655d9e11..35cb18797ff66cb87a6658e73ce02b0bfae29baa 100644 --- a/paddle/operators/dropout_op.cc +++ b/paddle/operators/dropout_op.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/operators/dropout_op.h" @@ -25,12 +25,10 @@ class DropoutOp : public framework::OperatorWithKernel { void InferShape(framework::InferShapeContext* ctx) const override { PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) must not be null."); - PADDLE_ENFORCE_GE(ctx->Attrs().Get("dropout_prob"), 0); - PADDLE_ENFORCE_LE(ctx->Attrs().Get("dropout_prob"), 1); auto x_dims = ctx->GetInputDim("X"); ctx->SetOutputDim("Out", x_dims); - if (ctx->Attrs().Get("is_training") == true) { + if (ctx->Attrs().Get("is_test") == false) { ctx->SetOutputDim("Mask", x_dims); } ctx->ShareLoD("X", /*->*/ "Out"); @@ -40,16 +38,19 @@ class DropoutOp : public framework::OperatorWithKernel { template class DropoutOpMaker : public framework::OpProtoAndCheckerMaker { public: - DropoutOpMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker) + DropoutOpMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "The input of dropout op."); AddOutput("Out", "The output of dropout op."); AddOutput("Mask", "The random sampled dropout mask.").AsIntermediate(); AddAttr("dropout_prob", "Probability of setting units to zero.") - .SetDefault(.5f); - AddAttr("is_training", "True if in training phase.").SetDefault(true); + .SetDefault(.5f) + .AddCustomChecker([](const float& drop_p) { + PADDLE_ENFORCE(drop_p >= 0.0f && drop_p <= 1.0f, + "'dropout_prob' must be between 0.0 and 1.0."); + }); + AddAttr("is_test", "True if in test phase.").SetDefault(false); AddAttr("seed", "Dropout random seed.").SetDefault(0); AddComment(R"DOC( @@ -71,16 +72,14 @@ class DropoutOpGrad : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; void InferShape(framework::InferShapeContext* ctx) const override { - PADDLE_ENFORCE_EQ(ctx->Attrs().Get("is_training"), true, - "GradOp is only callable when is_training is true"); + PADDLE_ENFORCE_EQ(ctx->Attrs().Get("is_test"), false, + "GradOp is only callable when is_test is false"); PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) must not be null."); PADDLE_ENFORCE(ctx->HasInput("Mask"), "Mask must not be null."); PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), "Input(Out@GRAD) must not be null."); - PADDLE_ENFORCE_GE(ctx->Attrs().Get("dropout_prob"), 0); - PADDLE_ENFORCE_LE(ctx->Attrs().Get("dropout_prob"), 1); auto x_dims = ctx->GetInputDim("X"); auto out_dims = ctx->GetInputDim(framework::GradVarName("Out")); PADDLE_ENFORCE_EQ(x_dims, out_dims, @@ -100,6 +99,8 @@ namespace ops = paddle::operators; REGISTER_OP(dropout, ops::DropoutOp, ops::DropoutOpMaker, dropout_grad, ops::DropoutOpGrad); REGISTER_OP_CPU_KERNEL( - dropout, ops::CPUDropoutKernel); + dropout, + ops::CPUDropoutKernel); REGISTER_OP_CPU_KERNEL( - dropout_grad, ops::DropoutGradKernel); + dropout_grad, + ops::DropoutGradKernel); diff --git a/paddle/operators/dropout_op.cu b/paddle/operators/dropout_op.cu index 30c769000f2b98c69eaa78a4c139630dd0956386..c56930336e865079f1b96df0f35b0a051fe63a27 100644 --- a/paddle/operators/dropout_op.cu +++ b/paddle/operators/dropout_op.cu @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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. */ #define EIGEN_USE_GPU #include @@ -30,16 +30,15 @@ struct MaskGenerator { __host__ __device__ MaskGenerator(AttrType dropout_prob, int seed) : dropout_prob(dropout_prob), seed(seed) {} - __host__ __device__ T operator()(const unsigned int n) const { + inline __host__ __device__ T operator()(const unsigned int n) const { thrust::minstd_rand rng; rng.seed(seed); thrust::uniform_real_distribution dist(0, 1); rng.discard(n); if (dist(rng) < dropout_prob) { return static_cast(0); - } else { - return static_cast(1); } + return static_cast(1); } }; @@ -58,8 +57,8 @@ class GPUDropoutKernel : public framework::OpKernel { auto X = EigenMatrix::Reshape(*x, 1); auto Y = EigenMatrix::Reshape(*y, 1); - auto place = context.GetEigenDevice(); - if (context.Attr("is_training")) { + auto& place = *context.template device_context().eigen_device(); + if (!context.Attr("is_test")) { auto* mask = context.Output("Mask"); auto* mask_data = mask->mutable_data(context.GetPlace()); int size = framework::product(mask->dims()); @@ -71,7 +70,7 @@ class GPUDropoutKernel : public framework::OpKernel { auto M = EigenMatrix::Reshape(*mask, 1); Y.device(place) = X * M; } else { - Y.device(place) = X * dropout_prob; + Y.device(place) = X * (1.0f - dropout_prob); } } }; @@ -80,7 +79,9 @@ class GPUDropoutKernel : public framework::OpKernel { } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL( - dropout, ops::GPUDropoutKernel); -REGISTER_OP_GPU_KERNEL( - dropout_grad, ops::DropoutGradKernel); +REGISTER_OP_CUDA_KERNEL( + dropout, + ops::GPUDropoutKernel); +REGISTER_OP_CUDA_KERNEL( + dropout_grad, + ops::DropoutGradKernel); diff --git a/paddle/operators/dropout_op.h b/paddle/operators/dropout_op.h index 6000b75fecdff74844605215e9364ac8f8a1525a..c90b8d277eb78048c001d36a367287146b51c636 100644 --- a/paddle/operators/dropout_op.h +++ b/paddle/operators/dropout_op.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 @@ -25,7 +25,7 @@ template using EigenMatrix = framework::EigenMatrix; -template +template class CPUDropoutKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { @@ -35,7 +35,7 @@ class CPUDropoutKernel : public framework::OpKernel { auto* y_data = y->mutable_data(context.GetPlace()); float dropout_prob = context.Attr("dropout_prob"); - if (context.Attr("is_training")) { + if (!context.Attr("is_test")) { auto* mask = context.Output("Mask"); auto* mask_data = mask->mutable_data(context.GetPlace()); int seed = context.Attr("seed"); @@ -55,18 +55,19 @@ class CPUDropoutKernel : public framework::OpKernel { } else { auto X = EigenMatrix::Reshape(*x, 1); auto Y = EigenMatrix::Reshape(*y, 1); - auto place = context.GetEigenDevice(); - Y.device(place) = X * dropout_prob; + auto& place = + *context.template device_context().eigen_device(); + Y.device(place) = X * (1.0f - dropout_prob); } } }; -template +template class DropoutGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { - PADDLE_ENFORCE(context.Attr("is_training"), - "GradOp is only callable when is_training is true"); + PADDLE_ENFORCE(!context.Attr("is_test"), + "GradOp is only callable when is_test is false"); auto* grad_x = context.Output(framework::GradVarName("X")); auto* grad_y = context.Input(framework::GradVarName("Out")); @@ -77,7 +78,8 @@ class DropoutGradKernel : public framework::OpKernel { auto dX = EigenMatrix::Reshape(*grad_x, 1); auto dY = EigenMatrix::Reshape(*grad_y, 1); - auto place = context.GetEigenDevice(); + auto& place = + *context.template device_context().eigen_device(); dX.device(place) = dY * M; } }; diff --git a/paddle/operators/dynamic_recurrent_op.cc b/paddle/operators/dynamic_recurrent_op.cc deleted file mode 100644 index d48cc4e8df587708ab93e7d788145adc01c1d3e5..0000000000000000000000000000000000000000 --- a/paddle/operators/dynamic_recurrent_op.cc +++ /dev/null @@ -1,418 +0,0 @@ -/* 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. */ - -#include "paddle/operators/dynamic_recurrent_op.h" - -#include "paddle/framework/op_registry.h" - -namespace paddle { -namespace operators { - -using framework::Scope; -using framework::TensorArray; -using framework::LoDTensor; -using framework::Variable; -using framework::OperatorBase; -using framework::DySeqMetaBatch; - -namespace detail { - -inline void CreateVariables(Scope& scope, - const std::vector& var_names) { - for (const auto& name : var_names) { - scope.Var(name); - } -} - -/* - * The inputs with sequence should be reordered when they are split, so the - * boot_states should be reordered in the same order. - * - * NOTE This may require that the `pre_state` of the first time step should just - * copy the `boot_state` rather than reference it, for that the content should - * be reordered, but the RNN op should not change the `boot_state` as an input - * variable's content. - */ -inline void ReorderInitialState(const DySeqMetaBatch& metas, - const LoDTensor& boot_state, LoDTensor* tensor, - const platform::Place& dst_place) { - for (size_t seq_id = 0; seq_id < metas.size(); seq_id++) { - auto slice = tensor->Slice(seq_id, seq_id + 1); - auto boot_slice = - boot_state.Slice(metas[seq_id].ori_idx, metas[seq_id].ori_idx + 1); - // TODO(superjom) pass in device context as an argument - slice.CopyFrom(boot_slice, dst_place, platform::CPUDeviceContext()); - } -} - -inline void RestoreInitialState(const DySeqMetaBatch& metas, - const LoDTensor& tensor, LoDTensor* boot_state, - const platform::Place& dst_place) { - for (size_t seq_id = 0; seq_id < metas.size(); seq_id++) { - auto slice = tensor.Slice(seq_id, seq_id + 1); - auto boot_slice = - boot_state->Slice(metas[seq_id].ori_idx, metas[seq_id].ori_idx + 1); - boot_slice.CopyFrom(slice, dst_place, platform::CPUDeviceContext()); - } -} - -} // namespace detail - -// Implementation for forward propagation. -template <> -void RNNAlgorithm::Run( - const framework::Scope& scope, const framework::OperatorBase& op, - const platform::DeviceContext& dev_ctx) { - SetComputeMode(ComputeMode::kForward); - cache_.Init(kArgNames[mode_], op, scope, &dev_ctx, &arg_); - SplitInputs(); - CreateScopes(); - WriteStepInputs(); - InitStates(); - WriteStepOutputs(); - RunSteps(); - ConcatOutputs(); -} - -// Implementation for backward propagation. -template <> -void RNNAlgorithm::Run( - const framework::Scope& scope, const framework::OperatorBase& op, - const platform::DeviceContext& dev_ctx) { - SetComputeMode(ComputeMode::kBackward); - cache_.Init(kArgNames[mode_], op, scope, &dev_ctx, &arg_); - SplitInputs(); - WriteStepInputs(); - InitStates(); - WriteStepOutputs(); - RunSteps(); - // copy boot-states' gradients back. - for (const auto& state : arg_.states) { - ExportInitialStateGradient(state); - } - - ConcatOutputs(); -} - -void RNNAlgorithm::SplitInputs() { - // TODO(superjom) make level a config - // TODO(superjom) check all the inputs has the same LoD - int level = 0; - for (const auto& item : cache_.inputs) { - const auto& var = item.second; - const auto& tensor = var->Get(); - TensorArray& ta = step_inputs_[item.first]; - - dy_seq_metas_[item.first] = - ta.Unpack(tensor, level, true /*length_descend*/); - - if (cache_.num_steps) { - PADDLE_ENFORCE_EQ(ta.size(), cache_.num_steps, - "inputs should have the same steps"); - } else { - cache_.num_steps = ta.size(); - } - } -} - -void RNNAlgorithm::WriteStepInputs() { - for (const auto& item : cache_.inputs) { - auto ta_it = step_inputs_.find(item.first); - PADDLE_ENFORCE(ta_it != step_inputs_.end(), - "step_inputs_ not compatible with memory set"); - TensorArray& ta = ta_it->second; - for (size_t step = 0; step < ta.size(); step++) { - auto tensor = ta.Read(step); - auto& step_scope = cache_.GetScope(step); - Variable* var = step_scope.FindVar(item.first); - if (var == nullptr) { - var = step_scope.Var(item.first); - } - var->GetMutable()->ShareDataWith(tensor); - } - } -} - -void RNNAlgorithm::WriteStepOutputs() { - // initialize step outputs - for (const auto& item : cache_.outputs) { - step_outputs_.emplace(item.first, TensorArray()); - } - PADDLE_ENFORCE_GT(step_outputs_.size(), 0UL); -} - -void RNNAlgorithm::CreateScopes() { - PADDLE_ENFORCE_GT(cache_.num_steps, 0); - // resize scopes - size_t num_scopes_need_create = cache_.num_steps - cache_.scopes->size(); - for (size_t i = 0; i < num_scopes_need_create; i++) { - cache_.scopes->emplace_back(&cache_.scope->NewScope()); - } - - // init temporary inputs - PADDLE_ENFORCE_NOT_NULL(step_unit_, "stepnet should be set first"); - std::vector states; - std::vector ex_states; - std::vector step_unit_outputs; - std::transform(arg_.states.begin(), arg_.states.end(), - std::back_inserter(states), - [](const rnn::StateAttr& m) { return m.var; }); - std::transform(arg_.states.begin(), arg_.states.end(), - std::back_inserter(ex_states), - [](const rnn::StateAttr& m) { return m.pre_var; }); - for (const auto& item : step_unit_->Outputs()) { - for (const auto& var : item.second) { - step_unit_outputs.push_back(var); - } - } - - for (size_t step = 0; step < cache_.num_steps; step++) { - auto& scope = cache_.GetScope(step); - detail::CreateVariables(scope, arg_.inlinks); - detail::CreateVariables(scope, arg_.outlinks); - detail::CreateVariables(scope, states); - detail::CreateVariables(scope, ex_states); - detail::CreateVariables(scope, step_unit_outputs); - } -} - -void RNNAlgorithm::ConcatOutputs() { - // TODO(superjom) transform this to a config - int level = 0; - for (size_t step = 0; step < cache_.num_steps; step++) { - auto& scope = cache_.GetScope(step); - for (auto& item : step_outputs_) { - auto* var = scope.FindVar(item.first); - PADDLE_ENFORCE_NOT_NULL(var); - auto* tensor = var->GetMutable(); - tensor->mutable_data(platform::CPUPlace()); - item.second.WriteShared(step, *tensor); - } - } - // the inputs' lods should be the same, so randomly get one lod. - const auto& some_lod = - cache_.scope->FindVar(arg_.inlinks.front())->Get().lod(); - const auto& some_meta = dy_seq_metas_[arg_.inlinks.front()]; - for (auto& item : step_outputs_) { - auto tensor = item.second.Pack(level, some_meta, some_lod); - auto* output = cache_.outputs[item.first]->GetMutable(); - const_cast(output)->ShareDataWith(tensor); - } -} - -void RNNAlgorithm::RunSteps() { - if (IsBackward()) { - // call stepnet in all the time steps reversely - for (int step = cache_.num_steps - 1; step >= 0; step--) { - auto& step_scope = cache_.GetScope(step); - step_unit_->Run(step_scope, *cache_.dev_ctx); - } - } else { - for (size_t step = 0; step < cache_.num_steps; step++) { - auto& step_scope = cache_.GetScope(step); - step_unit_->Run(step_scope, *cache_.dev_ctx); - } - } -} - -void RNNAlgorithm::InitStates() { - for (size_t step = 0; step < cache_.num_steps; step++) { - for (const auto& state : arg_.states) { - CreateState(state, step); - LinkState(state, step); - } - } -} - -void RNNAlgorithm::CreateState(const rnn::StateAttr& state_attr, size_t step) { - auto& scope = cache_.GetScope(step); - auto& state = *cache_.GetTensor(scope, state_attr.var); - auto& boot_state = *cache_.GetTensor(*cache_.scope, state_attr.boot_var); - - size_t num_instances = - step_inputs_[arg_.inlinks.front()].Read(step).dims()[0]; - auto dims = boot_state.dims(); - dims[0] = num_instances; - - state.Resize(dims); - state.mutable_data(platform::CPUPlace()); - states_[state_attr.var].WriteShared(step, state); -} - -void RNNAlgorithm::LinkState(const rnn::StateAttr& state, size_t step) { - auto& scope = cache_.GetScope(step); - auto& state_pre = *cache_.GetTensor(scope, state.pre_var); - - // process the first state's boot-state(the 0-step in forward mode or the - // last step in backward mode) - // Only forward mode need to link the boot-state to the `pre-state` in first - // time step. In backward mode, need to copy the gradient of `pre-state` in - // first time step to the gradient of `boot-state`. - if (step == 0 && IsForward()) { - LinkInitialState(state); - } else { - size_t num_instances = - step_inputs_[arg_.inlinks.front()].Read(step).dims()[0]; - auto* pre_state = cache_.GetTensor(cache_.GetScope(step - 1), state.var); - // shink and share from previous state - auto shrinked_pre_state = pre_state->Slice(0, num_instances); - state_pre.ShareDataWith(shrinked_pre_state); - } -} - -void RNNAlgorithm::LinkInitialState(const rnn::StateAttr& state) { - // all the step_inputs' metas should be the same, just randomly select one - // and get the dyseq meta. - const auto& some_meta = dy_seq_metas_[arg_.inlinks.front()]; - auto& scope = cache_.GetScope(0); - auto& state_pre = *cache_.GetTensor(scope, state.pre_var); - auto* pre_state = cache_.GetTensor(*cache_.scope, state.boot_var); - pre_state->mutable_data(platform::CPUPlace()); - // allocate state - state_pre.Resize(pre_state->dims()); - state_pre.mutable_data(platform::CPUPlace()); - detail::ReorderInitialState(some_meta, *pre_state, &state_pre, - pre_state->place()); -} - -void RNNAlgorithm::ExportInitialStateGradient(const rnn::StateAttr& state) { - // all the step_inputs' metas should be the same, just randomly select one - // and get the dyseq meta. - const auto& some_meta = dy_seq_metas_[arg_.inlinks.front()]; - auto& scope = cache_.GetScope(0); - - auto& state_pre = *cache_.GetTensor(scope, state.pre_var); - auto& pre_state = *cache_.GetTensor(*cache_.scope, state.boot_var); - pre_state.Resize(state_pre.dims()); - detail::RestoreInitialState(some_meta, state_pre, &pre_state, - pre_state.place()); -} - -void RNNAlgorithm::ArgCache::Init(const rnn::ArgumentName& name, - const paddle::framework::OperatorBase& op, - const paddle::framework::Scope& scope, - platform::DeviceContext const* dev_ctx, - rnn::Argument* arg) { - this->scope = &scope; - InitArgument(name, op, arg); - CacheScopes(scope, *arg); - CacheInlinks(scope, arg->inlinks); - CacheOutlinks(scope, arg->outlinks); - this->dev_ctx = dev_ctx; -} - -void RNNAlgorithm::ArgCache::InitArgument(const rnn::ArgumentName& name, - const OperatorBase& op, - rnn::Argument* arg) { - rnn::InitArgument(name, arg, op, false /*is_grad*/); -} - -void RNNAlgorithm::ArgCache::CacheScopes(const Scope& scope, - const rnn::Argument& arg) { - auto scopes_var = scope.FindVar(arg.step_scopes); - PADDLE_ENFORCE(scopes_var != nullptr, - "the step_scopes output argument [%s] should be created first " - "by framework.", - arg.step_scopes); - this->scopes = scopes_var->GetMutable>(); -} - -void RNNAlgorithm::ArgCache::CacheInlinks( - const Scope& scope, const std::vector& names) { - for (auto name : names) { - auto* var = GetVariable(scope, name); - inputs[name] = var; - } -} - -void RNNAlgorithm::ArgCache::CacheOutlinks( - const Scope& scope, const std::vector& names) { - for (auto name : names) { - auto* var = GetVariable(scope, name); - outputs[name] = var; - } -} - -Variable* RNNAlgorithm::ArgCache::GetVariable(const Scope& scope, - const std::string& name) { - auto* var = scope.FindVar(name); - PADDLE_ENFORCE_NOT_NULL(var, "variable [%s] not exist in scope", name); - return var; -} - -LoDTensor* RNNAlgorithm::ArgCache::GetTensor(const framework::Scope& scope, - const std::string& name) { - auto* var = GetVariable(scope, name); - return var->GetMutable(); -} - -const std::array RNNAlgorithm::kArgNames{ - {rnn::ArgumentName{"step_unit", "step_scopes", "inputs", "outputs", - "states", "ex_states", "initial_states"}, - rnn::ArgumentName{"step_unit", "step_scopes@GRAD", "outputs@GRAD", - "inputs@GRAD", "states", "ex_states", - "initial_states@GRAD"}}}; - -void DynamicRecurrentOp::Run(const framework::Scope& scope, - const platform::DeviceContext& dev_ctx) const { - rnn.Run( - scope, *dynamic_cast(this), dev_ctx); -} - -void DynamicRecurrentGradientOp::Run( - const Scope& scope, const platform::DeviceContext& dev_ctx) const { - rnn.Run( - scope, *dynamic_cast(this), dev_ctx); -} - -class DynamicRecurrentOpProtoAndCheckerMaker - : public framework::OpProtoAndCheckerMaker { - public: - DynamicRecurrentOpProtoAndCheckerMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { - const auto& name = - RNNAlgorithm::kArgNames[RNNAlgorithm::ComputeMode::kForward]; - // inputs and outputs stored in proto - AddInput(name.inlinks, - "The inputs that need to be segmented for each step.") - .AsDuplicable(); - AddInput(name.initial_states, "Variables to initialize the states.") - .AsDuplicable(); - - AddOutput(name.outlinks, - "The outputs that need to be concatenated for all steps.") - .AsDuplicable(); - AddOutput(name.step_scopes, "step scopes"); - - // Attributes stored in AttributeMap - AddAttr>(name.ex_states, "names of ex_states"); - AddAttr>(name.states, "names of states"); - - AddComment(R"DOC( -Dynamic Recurrent Operator. - -This is a RNN operator for varience-length sequences. - -)DOC"); - } -}; - -} // namespace operators -} // namespace paddle - -REGISTER_OP(dynamic_recurrent, paddle::operators::DynamicRecurrentOp, - paddle::operators::DynamicRecurrentOpProtoAndCheckerMaker, - dynamic_recurrent_grad, - paddle::operators::DynamicRecurrentGradientOp); diff --git a/paddle/operators/dynamic_recurrent_op.h b/paddle/operators/dynamic_recurrent_op.h deleted file mode 100644 index 5b0548c3a44c9f58838ecc567ee41a587883c26a..0000000000000000000000000000000000000000 --- a/paddle/operators/dynamic_recurrent_op.h +++ /dev/null @@ -1,233 +0,0 @@ -/* 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 - -#ifdef PADDLE_WITH_TESTING -#include "gtest/gtest.h" -#endif - -#include "paddle/framework/lod_tensor.h" -#include "paddle/framework/operator.h" -#include "paddle/framework/tensor_array.h" -#include "paddle/framework/variable.h" -#include "paddle/operators/rnn/recurrent_op_utils.h" - -namespace paddle { -namespace operators { - -class RNNAlgorithm { - public: - enum ComputeMode { kForward = 0, kBackward = 1 }; - static const std::array kArgNames; - using value_type = float; - - /* - * Different `Run` method for forward and backward, `_` is just for template - * specifialization. - */ - template - void Run(const framework::Scope& scope, const framework::OperatorBase& op, - const platform::DeviceContext& dev_ctx); - /* - * Split the inputs(LoDTensors) to segments for each time step. - */ - void SplitInputs(); - - /* - * Create step-scopes to store temporary outputs in each time steps. - */ - void CreateScopes(); - - /* - * Link TensorArray steps to the corresponding variables located in - * step-scopes. - */ - void WriteStepInputs(); - - /* - * Write output of each step to the corresponding TensorArray. - */ - void WriteStepOutputs(); - - /* - * Initialize the states, each state will have a corresponding pre-state, - * which share the memory with the state in the previous time state. The - * pre-state in the first time step will be initialized with an zero tensor or - * a tensor in parent scope if is provided. - */ - void InitStates(); - - /* - * Create state variables for each time step. - */ - void CreateState(const rnn::StateAttr& state, size_t step); - - /* - * Link pre-state variable in current scope to the state variable in the - * previous time step (scope) by reference. - */ - void LinkState(const rnn::StateAttr& state, size_t step); - - /* - * Link the pre-state of the first time step to the `boot-state` in parent's - * scope. - */ - void LinkInitialState(const rnn::StateAttr& state); - - /* - * Copy the gradient from `pre-state` in the first step-scope to the - * `boot-state` in parent's scope. - */ - void ExportInitialStateGradient(const rnn::StateAttr& state); - - /* - * Calculate time steps. - */ - void RunSteps(); - - /* - * Concatenate outputs in each time step and generate a LoDTensor. - */ - void ConcatOutputs(); - - void SetComputeMode(ComputeMode mode) { mode_ = mode; } - bool IsForward() const { return mode_ == ComputeMode::kForward; } - bool IsBackward() const { return mode_ == ComputeMode::kBackward; } - - /* - * set a step unit that is created according to a RecurrentOp's step unit. - */ - void SetStepUnit(std::unique_ptr step_unit) { - PADDLE_ENFORCE_NOT_NULL(step_unit); - step_unit_ = std::move(step_unit); - } - const framework::OperatorBase& GetStepUnit() const { return *step_unit_; } - - const framework::TensorArray& state(const std::string& name) const { - auto it = states_.find(name); - PADDLE_ENFORCE(it != states_.end()); - return it->second; - } - const framework::TensorArray& step_input(const std::string& name) const { - auto it = step_inputs_.find(name); - PADDLE_ENFORCE(it != step_inputs_.end()); - return it->second; - } - const framework::TensorArray& step_output(const std::string& name) const { - auto it = step_outputs_.find(name); - PADDLE_ENFORCE(it != step_outputs_.end()); - return it->second; - } - - protected: - struct ArgCache { - framework::Scope const* scope; - std::vector* scopes; - std::map inputs; - std::map outputs; - platform::DeviceContext const* dev_ctx; - - size_t num_steps{0}; - - void Init(const rnn::ArgumentName& name, const framework::OperatorBase& op, - const framework::Scope& scope, - platform::DeviceContext const* dev_ctx, rnn::Argument* arg); - - framework::Scope& GetScope(size_t index) { - PADDLE_ENFORCE_LT(index, num_steps); - return *scopes->at(index); - } - - framework::LoDTensor* GetTensor(const framework::Scope& scope, - const std::string& name); - - private: - void InitArgument(const rnn::ArgumentName& name, - const framework::OperatorBase& op, rnn::Argument* arg); - void CacheScopes(const framework::Scope& scope, const rnn::Argument& arg); - void CacheInlinks(const framework::Scope& scope, - const std::vector& names); - void CacheOutlinks(const framework::Scope& scope, - const std::vector& names); - framework::Variable* GetVariable(const framework::Scope& scope, - const std::string& name); - }; - - private: - std::unique_ptr step_unit_; - std::map states_; - std::map step_inputs_; - std::map step_outputs_; - std::map> dy_seq_metas_; - rnn::Argument arg_; - ArgCache cache_; - ComputeMode mode_{ComputeMode::kForward}; - -#ifdef PADDLE_WITH_TESTING - // test forward - friend class RNNAlgorithmTestHelper; - FRIEND_TEST(RNNAlgorithmTestHelper, SplitInputs); - FRIEND_TEST(RNNAlgorithmTestHelper, CreateCache); - FRIEND_TEST(RNNAlgorithmTestHelper, CreateScopes); - FRIEND_TEST(RNNAlgorithmTestHelper, WriteStepInputs); - FRIEND_TEST(RNNAlgorithmTestHelper, WriteStepOutputs); - FRIEND_TEST(RNNAlgorithmTestHelper, InitStates); - FRIEND_TEST(RNNAlgorithmTestHelper, ConcatOutputs); -// TODO(superjom) test backward -#endif -}; - -class DynamicRecurrentOp : public framework::OperatorBase { - public: - DynamicRecurrentOp(const std::string& type, - const framework::VariableNameMap& inputs, - const framework::VariableNameMap& outputs, - const framework::AttributeMap& attrs) - : OperatorBase(type, inputs, outputs, attrs) {} - - DynamicRecurrentOp(const DynamicRecurrentOp& o) - : framework::OperatorBase( - static_cast(o)) { - PADDLE_THROW("Not implemented"); - } - - void Run(const framework::Scope& scope, - const platform::DeviceContext& dev_ctx) const override; - - mutable RNNAlgorithm rnn; -}; - -class DynamicRecurrentGradientOp : public framework::OperatorBase { - public: - DynamicRecurrentGradientOp(const std::string& type, - const framework::VariableNameMap& inputs, - const framework::VariableNameMap& outputs, - const framework::AttributeMap& attrs) - : OperatorBase(type, inputs, outputs, attrs) {} - - DynamicRecurrentGradientOp(const DynamicRecurrentGradientOp& o) - : framework::OperatorBase( - static_cast(o)) { - PADDLE_THROW("Not implemented"); - } - - void Run(const framework::Scope& scope, - const platform::DeviceContext& dev_ctx) const override; - - mutable RNNAlgorithm rnn; -}; - -} // namespace operators -} // namespace paddle diff --git a/paddle/operators/dynamic_recurrent_op_test.cc b/paddle/operators/dynamic_recurrent_op_test.cc deleted file mode 100644 index 8d840e259b190ead86a66df8ab31c5170db4d824..0000000000000000000000000000000000000000 --- a/paddle/operators/dynamic_recurrent_op_test.cc +++ /dev/null @@ -1,217 +0,0 @@ -#include "paddle/operators/dynamic_recurrent_op.h" - -#include - -#include "paddle/framework/ddim.h" -#include "paddle/framework/lod_tensor.h" -#include "paddle/framework/op_desc.h" -#include "paddle/framework/op_registry.h" -#include "paddle/operators/net_op.h" - -namespace paddle { -namespace operators { - -using framework::Scope; -using framework::TensorArray; -using framework::LoDTensor; -using framework::Variable; - -class TestOp : public framework::OperatorBase { - public: - using framework::OperatorBase::OperatorBase; - DEFINE_OP_CLONE_METHOD(TestOp); - void Run(const Scope& scope, - const platform::DeviceContext& dev_ctx) const override {} -}; - -void OpDescNewVar(const std::string& param_name, - std::initializer_list arguments, - paddle::framework::OpDesc::Var* var) { - var->set_parameter(param_name); - for (auto& arg_name : arguments) { - var->add_arguments(arg_name); - } -} - -// create a LoD tensor in scope with specific dims -LoDTensor* CreateVar(Scope& scope, std::string name, framework::DDim dims, - const platform::Place& place) { - auto* var = scope.Var(name); - auto* tensor = var->GetMutable(); - tensor->Resize(dims); - tensor->mutable_data(place); - return tensor; -} - -class RNNAlgorithmTestHelper : public ::testing::Test { - protected: - const rnn::ArgumentName argname = RNNAlgorithm::kArgNames[0]; - - virtual void SetUp() override { - CreateGlobalVariables(); - - auto op_desc = CreateOpDesc(); - op = paddle::framework::OpRegistry::CreateOp(op_desc); - dop = &(dynamic_cast(op.get())->rnn); - InitCacheManually(); - InitStepNet(); - } - - framework::OpDesc CreateOpDesc() { - // create op - paddle::framework::OpDesc op_desc; - op_desc.set_type("dynamic_recurrent"); - - OpDescNewVar(argname.inlinks, {"in0"}, op_desc.add_inputs()); - OpDescNewVar(argname.initial_states, {"boot_mem"}, op_desc.add_inputs()); - OpDescNewVar(argname.step_scopes, {"step_scopes"}, op_desc.add_outputs()); - OpDescNewVar(argname.outlinks, {"out0"}, op_desc.add_outputs()); - - // set pre-states - auto pre_memories = op_desc.mutable_attrs()->Add(); - pre_memories->set_name(argname.ex_states); - pre_memories->set_type(paddle::framework::AttrType::STRINGS); - auto pre_memories_item = pre_memories->add_strings(); - *pre_memories_item = "mem@pre"; - - // set states - auto memories = op_desc.mutable_attrs()->Add(); - memories->set_name(argname.states); - memories->set_type(paddle::framework::AttrType::STRINGS); - auto memories_item = memories->add_strings(); - *memories_item = "mem"; - return op_desc; - } - - void CreateGlobalVariables() { - platform::CPUPlace place; - scope.Var("step_scopes"); - CreateVar(scope, "boot_mem", framework::make_ddim({10, 20}), place); - CreateVar(scope, "out0", framework::make_ddim({10, 20}), place); - auto* in0 = CreateVar(scope, "in0", framework::make_ddim({10, 8}), place); - // 10 instanes with 4 sentences, length is 4, 3, 2, 1 respectively. - framework::LoD in0_lod(1); - for (int x : std::vector{0, 4, 7, 9, 10}) { - in0_lod[0].push_back(x); - } - in0->set_lod(in0_lod); - in0->Resize(framework::make_ddim({10, 8})); - // set the content, each sentence content is seqid.batchid - // the seqid starts from 0 - int start = 0; - for (size_t seqid = 0; seqid < in0_lod.size() - 1; seqid++) { - for (size_t batchid = 0; - batchid < in0_lod[0][seqid + 1] - in0_lod[0][seqid]; batchid++) { - float v = seqid + batchid * 0.1; - - for (size_t dim = 0; dim < 8; dim++) { - in0->data()[start * 8 + dim] = v; - } - start++; - } - } - } - - void InitCacheManually() { - dop->cache_.Init(RNNAlgorithm::kArgNames[0], *op, scope, &device_context, - &dop->arg_); - } - - void InitStepNet() { - std::unique_ptr stepnet{new NetOp}; - dynamic_cast(stepnet.get()) - ->AppendOp(std::unique_ptr(new TestOp( - "test", {{"inputs", {"in0"}}, {"initial_states", {"boot_mem"}}}, - {{"outputs", {"out0"}}, {"step_scopes", {"step_scopes"}}}, {}))); - dop->SetStepUnit(std::move(stepnet)); - } - - protected: - RNNAlgorithm* dop; - std::unique_ptr op; - paddle::platform::CPUDeviceContext device_context; - paddle::framework::Scope scope; -}; - -TEST_F(RNNAlgorithmTestHelper, CreateCache) { - const rnn::Argument& arg = dop->arg_; - ASSERT_EQ(arg.inlinks.size(), 1UL); - ASSERT_EQ(arg.outlinks.size(), 1UL); -} - -TEST_F(RNNAlgorithmTestHelper, SplitInputs) { - dop->SplitInputs(); - auto& in0_ta = dop->step_inputs_["in0"]; - ASSERT_EQ(in0_ta.size(), 4UL); - - const auto& batch0 = in0_ta.Read(0); - const auto& batch1 = in0_ta.Read(1); - const auto& batch2 = in0_ta.Read(2); - const auto& batch3 = in0_ta.Read(3); - EXPECT_EQ(batch0.dims()[0], 4); - EXPECT_EQ(batch1.dims()[0], 3); - EXPECT_EQ(batch2.dims()[0], 2); - EXPECT_EQ(batch3.dims()[0], 1); -} - -TEST_F(RNNAlgorithmTestHelper, CreateScopes) { - dop->SplitInputs(); - dop->CreateScopes(); - ASSERT_EQ(dop->cache_.num_steps, 4UL); - ASSERT_EQ(dop->cache_.scopes->size(), 4UL); -} - -TEST_F(RNNAlgorithmTestHelper, WriteStepInputs) { - dop->SplitInputs(); - dop->CreateScopes(); - dop->WriteStepInputs(); - - for (size_t step = 0; step < dop->cache_.num_steps; step++) { - auto& scope = dop->cache_.GetScope(step); - for (auto name : std::vector({"in0"})) { - ASSERT_TRUE(scope.FindVar(name) != nullptr); - } - } -} - -TEST_F(RNNAlgorithmTestHelper, WriteStepOutputs) { - dop->SplitInputs(); - dop->CreateScopes(); - dop->WriteStepInputs(); - dop->WriteStepOutputs(); - - for (size_t step = 0; step < dop->cache_.num_steps; step++) { - auto& scope = dop->cache_.GetScope(step); - for (auto name : std::vector({"out0"})) { - ASSERT_TRUE(scope.FindVar(name)); - } - } -} - -TEST_F(RNNAlgorithmTestHelper, ConcatOutputs) { - // Let's leave this test to python unittest. -} - -TEST_F(RNNAlgorithmTestHelper, InitStates) { - dop->SetComputeMode(RNNAlgorithm::ComputeMode::kForward); - dop->SplitInputs(); - dop->CreateScopes(); - dop->WriteStepInputs(); - dop->WriteStepOutputs(); - dop->InitStates(); - - for (size_t step = 0; step < dop->cache_.num_steps; step++) { - auto& scope = dop->cache_.GetScope(step); - auto state = scope.FindVar("mem"); - ASSERT_TRUE(state != nullptr); - - auto* pre_state = scope.FindVar("mem@pre"); - ASSERT_TRUE(pre_state != nullptr); - - auto* boot_state = scope.FindVar("boot_mem"); - ASSERT_TRUE(boot_state != nullptr); - } -} - -} // operators -} // namespace paddle diff --git a/paddle/operators/elementwise_add_op.cc b/paddle/operators/elementwise_add_op.cc index 432b9ba6f72f8dd11c666d5473c570bde60de995..70b7c9f2ec11bf8ad56a24324a53792955edc77d 100644 --- a/paddle/operators/elementwise_add_op.cc +++ b/paddle/operators/elementwise_add_op.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/operators/elementwise_add_op.h" #include "paddle/operators/elementwise_op.h" @@ -19,8 +19,7 @@ namespace paddle { namespace operators { class ElementwiseAddOpMaker : public ElementwiseOpMaker { public: - ElementwiseAddOpMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker) + ElementwiseAddOpMaker(OpProto* proto, OpAttrChecker* op_checker) : ElementwiseOpMaker(proto, op_checker) { SetComment("Add", "$Out = X + Y$"); AddComment(comment_); @@ -34,13 +33,13 @@ REGISTER_OP(elementwise_add, ops::ElementwiseOp, ops::ElementwiseAddOpMaker, elementwise_add_grad, ops::ElementwiseOpGrad); REGISTER_OP_CPU_KERNEL( elementwise_add, - ops::ElementwiseAddKernel, - ops::ElementwiseAddKernel, - ops::ElementwiseAddKernel, - ops::ElementwiseAddKernel); + ops::ElementwiseAddKernel, + ops::ElementwiseAddKernel, + ops::ElementwiseAddKernel, + ops::ElementwiseAddKernel); REGISTER_OP_CPU_KERNEL( elementwise_add_grad, - ops::ElementwiseAddGradKernel, - ops::ElementwiseAddGradKernel, - ops::ElementwiseAddGradKernel, - ops::ElementwiseAddGradKernel); + ops::ElementwiseAddGradKernel, + ops::ElementwiseAddGradKernel, + ops::ElementwiseAddGradKernel, + ops::ElementwiseAddGradKernel); diff --git a/paddle/operators/elementwise_add_op.cu b/paddle/operators/elementwise_add_op.cu index 7591428ac7c2f74f25f0f7d818eafcf59c8e4a4f..641cea323acee549898cb6f0245ccac4c069ce32 100644 --- a/paddle/operators/elementwise_add_op.cu +++ b/paddle/operators/elementwise_add_op.cu @@ -1,31 +1,32 @@ /* 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 +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 + 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. */ +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. */ #define EIGEN_USE_GPU #include "paddle/operators/elementwise_add_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL( +REGISTER_OP_CUDA_KERNEL( elementwise_add, - ops::ElementwiseAddKernel, - ops::ElementwiseAddKernel, - ops::ElementwiseAddKernel, - ops::ElementwiseAddKernel); -REGISTER_OP_GPU_KERNEL( + ops::ElementwiseAddKernel, + ops::ElementwiseAddKernel, + ops::ElementwiseAddKernel, + ops::ElementwiseAddKernel); +REGISTER_OP_CUDA_KERNEL( elementwise_add_grad, - ops::ElementwiseAddGradKernel, - ops::ElementwiseAddGradKernel, - ops::ElementwiseAddGradKernel, - ops::ElementwiseAddGradKernel); + ops::ElementwiseAddGradKernel, + ops::ElementwiseAddGradKernel, + ops::ElementwiseAddGradKernel, + ops::ElementwiseAddGradKernel); diff --git a/paddle/operators/elementwise_add_op.h b/paddle/operators/elementwise_add_op.h index f04fe3ec6069ab1bf227be6a3a5c10ee908e4824..59abbb57d1dcfbef6ead70e4afa9a3816d60d9b5 100644 --- a/paddle/operators/elementwise_add_op.h +++ b/paddle/operators/elementwise_add_op.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 @@ -19,11 +19,48 @@ namespace paddle { namespace operators { -template +template +struct AddFunctor { + inline HOSTDEVICE T operator()(T a, T b) const { return a + b; } +}; + +template class ElementwiseAddKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { - ElementwiseCompute(ctx); + using Tensor = framework::Tensor; + + auto* x = ctx.Input("X"); + auto* y = ctx.Input("Y"); + auto* z = ctx.Output("Out"); + z->mutable_data(ctx.GetPlace()); + TransformFunctor, T, DeviceContext> functor( + x, y, z, ctx.template device_context(), AddFunctor()); + + auto x_dims = x->dims(); + auto y_dims = y->dims(); + PADDLE_ENFORCE_GE(x_dims.size(), y_dims.size(), + "Rank of first input must >= rank of second input."); + + if (x_dims == y_dims) { + functor.Run(); + return; + } + + int axis = ctx.Attr("axis"); + axis = (axis == -1 ? x_dims.size() - y_dims.size() : axis); + PADDLE_ENFORCE(axis >= 0 && axis < x_dims.size(), + "Axis should be in range [0, x_dims)"); + + int pre, n, post; + get_mid_dims(x_dims, y_dims, axis, pre, n, post); + if (post == 1) { + functor.RunRowWise(n, pre); + return; + } else { + functor.RunMidWise(n, pre, post); + return; + } } }; @@ -100,11 +137,11 @@ struct ElementwiseAddBroadCast2GradFunctor { } }; -template +template class ElementwiseAddGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { - ElementwiseGradCompute, + ElementwiseGradCompute, ElementwiseAddOneGradFunctor, ElementwiseAddBroadCastGradFunctor, ElementwiseAddBroadCast2GradFunctor>(ctx); diff --git a/paddle/operators/elementwise_div_op.cc b/paddle/operators/elementwise_div_op.cc index 7a325199bd07e44042a4e8b3aae0ab93fae1c351..1fa960866fa2066a351ef2e65a3c77cf8b6595f7 100644 --- a/paddle/operators/elementwise_div_op.cc +++ b/paddle/operators/elementwise_div_op.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/operators/elementwise_div_op.h" #include "paddle/operators/elementwise_op.h" @@ -19,8 +19,7 @@ namespace paddle { namespace operators { class ElementwiseDivOpMaker : public ElementwiseOpMaker { public: - ElementwiseDivOpMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker) + ElementwiseDivOpMaker(OpProto* proto, OpAttrChecker* op_checker) : ElementwiseOpMaker(proto, op_checker) { SetComment("Div", "$Out = X / Y$"); AddComment(comment_); @@ -35,13 +34,13 @@ REGISTER_OP(elementwise_div, ops::ElementwiseOp, ops::ElementwiseDivOpMaker, elementwise_div_grad, ops::ElementwiseOpGrad); REGISTER_OP_CPU_KERNEL( elementwise_div, - ops::ElementwiseDivKernel, - ops::ElementwiseDivKernel, - ops::ElementwiseDivKernel, - ops::ElementwiseDivKernel); + ops::ElementwiseDivKernel, + ops::ElementwiseDivKernel, + ops::ElementwiseDivKernel, + ops::ElementwiseDivKernel); REGISTER_OP_CPU_KERNEL( elementwise_div_grad, - ops::ElementwiseDivGradKernel, - ops::ElementwiseDivGradKernel, - ops::ElementwiseDivGradKernel, - ops::ElementwiseDivGradKernel); + ops::ElementwiseDivGradKernel, + ops::ElementwiseDivGradKernel, + ops::ElementwiseDivGradKernel, + ops::ElementwiseDivGradKernel); diff --git a/paddle/operators/elementwise_div_op.cu b/paddle/operators/elementwise_div_op.cu index de4d0c33442a1fcfe0dd4c16df7ceeec737fbc6d..a0372123d6ffe7e3a90727ddd37a787c1030e0bd 100644 --- a/paddle/operators/elementwise_div_op.cu +++ b/paddle/operators/elementwise_div_op.cu @@ -1,31 +1,32 @@ /* 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 +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 + 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. */ +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. */ #define EIGEN_USE_GPU #include "paddle/operators/elementwise_div_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL( +REGISTER_OP_CUDA_KERNEL( elementwise_div, - ops::ElementwiseDivKernel, - ops::ElementwiseDivKernel, - ops::ElementwiseDivKernel, - ops::ElementwiseDivKernel); -REGISTER_OP_GPU_KERNEL( + ops::ElementwiseDivKernel, + ops::ElementwiseDivKernel, + ops::ElementwiseDivKernel, + ops::ElementwiseDivKernel); +REGISTER_OP_CUDA_KERNEL( elementwise_div_grad, - ops::ElementwiseDivGradKernel, - ops::ElementwiseDivGradKernel, - ops::ElementwiseDivGradKernel, - ops::ElementwiseDivGradKernel); + ops::ElementwiseDivGradKernel, + ops::ElementwiseDivGradKernel, + ops::ElementwiseDivGradKernel, + ops::ElementwiseDivGradKernel); diff --git a/paddle/operators/elementwise_div_op.h b/paddle/operators/elementwise_div_op.h index 8946ff3d25c2aff3dc3aa69368f0083371cd2fef..875abd313ffc8fdf910d461922ff41f65ef276e7 100644 --- a/paddle/operators/elementwise_div_op.h +++ b/paddle/operators/elementwise_div_op.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 @@ -19,11 +19,11 @@ namespace paddle { namespace operators { -template +template class ElementwiseDivKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { - ElementwiseCompute(ctx); + ElementwiseCompute(ctx); } }; @@ -102,11 +102,11 @@ struct ElementwiseDivBroadCast2GradFunctor { } }; -template +template class ElementwiseDivGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { - ElementwiseGradCompute, + ElementwiseGradCompute, ElementwiseDivGradFunctor, ElementwiseDivBroadCastGradFunctor, ElementwiseDivBroadCast2GradFunctor>(ctx); diff --git a/paddle/operators/elementwise_mul_op.cc b/paddle/operators/elementwise_mul_op.cc index 8851267a524f51773a9f86ff83943cea4cb042aa..a6d11736194cb79bdc247c721acf8bda9c81dbe5 100644 --- a/paddle/operators/elementwise_mul_op.cc +++ b/paddle/operators/elementwise_mul_op.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/operators/elementwise_mul_op.h" #include "paddle/operators/elementwise_op.h" @@ -20,8 +20,7 @@ namespace operators { class ElementwiseMulOpMaker : public ElementwiseOpMaker { public: - ElementwiseMulOpMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker) + ElementwiseMulOpMaker(OpProto* proto, OpAttrChecker* op_checker) : ElementwiseOpMaker(proto, op_checker) { SetComment("Mul", "$Out = X \\odot\\ Y$"); AddComment(comment_); @@ -36,13 +35,13 @@ REGISTER_OP(elementwise_mul, ops::ElementwiseOp, ops::ElementwiseMulOpMaker, elementwise_mul_grad, ops::ElementwiseOpGrad); REGISTER_OP_CPU_KERNEL( elementwise_mul, - ops::ElementwiseMulKernel, - ops::ElementwiseMulKernel, - ops::ElementwiseMulKernel, - ops::ElementwiseMulKernel); + ops::ElementwiseMulKernel, + ops::ElementwiseMulKernel, + ops::ElementwiseMulKernel, + ops::ElementwiseMulKernel); REGISTER_OP_CPU_KERNEL( elementwise_mul_grad, - ops::ElementwiseMulGradKernel, - ops::ElementwiseMulGradKernel, - ops::ElementwiseMulGradKernel, - ops::ElementwiseMulGradKernel); + ops::ElementwiseMulGradKernel, + ops::ElementwiseMulGradKernel, + ops::ElementwiseMulGradKernel, + ops::ElementwiseMulGradKernel); diff --git a/paddle/operators/elementwise_mul_op.cu b/paddle/operators/elementwise_mul_op.cu index b0dfdee1ccef56c6cda06ae6759017294fa5115c..f73e8afda960a89aff8568eab66b0f120db2e342 100644 --- a/paddle/operators/elementwise_mul_op.cu +++ b/paddle/operators/elementwise_mul_op.cu @@ -1,31 +1,32 @@ /* 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 +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 + 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. */ +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. */ #define EIGEN_USE_GPU #include "paddle/operators/elementwise_mul_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL( +REGISTER_OP_CUDA_KERNEL( elementwise_mul, - ops::ElementwiseMulKernel, - ops::ElementwiseMulKernel, - ops::ElementwiseMulKernel, - ops::ElementwiseMulKernel); -REGISTER_OP_GPU_KERNEL( + ops::ElementwiseMulKernel, + ops::ElementwiseMulKernel, + ops::ElementwiseMulKernel, + ops::ElementwiseMulKernel); +REGISTER_OP_CUDA_KERNEL( elementwise_mul_grad, - ops::ElementwiseMulGradKernel, - ops::ElementwiseMulGradKernel, - ops::ElementwiseMulGradKernel, - ops::ElementwiseMulGradKernel); + ops::ElementwiseMulGradKernel, + ops::ElementwiseMulGradKernel, + ops::ElementwiseMulGradKernel, + ops::ElementwiseMulGradKernel); diff --git a/paddle/operators/elementwise_mul_op.h b/paddle/operators/elementwise_mul_op.h index 4469b07eaa08a3b011a88e58f1d645dd30b10ced..3ee50207c07fa2b7ccf2c002903a4f055dbfb352 100644 --- a/paddle/operators/elementwise_mul_op.h +++ b/paddle/operators/elementwise_mul_op.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/operators/elementwise_op_function.h" @@ -18,11 +18,11 @@ namespace paddle { namespace operators { -template +template class ElementwiseMulKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { - ElementwiseCompute(ctx); + ElementwiseCompute(ctx); } }; @@ -101,11 +101,11 @@ struct ElementwiseMulBroadCast2GradFunctor { } }; -template +template class ElementwiseMulGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { - ElementwiseGradCompute, + ElementwiseGradCompute, ElementwiseMulGradFunctor, ElementwiseMulBroadCastGradFunctor, ElementwiseMulBroadCast2GradFunctor>(ctx); diff --git a/paddle/operators/elementwise_op.h b/paddle/operators/elementwise_op.h index 56e5eb69bc382a2c15d88b759fa6987f02c6cabb..f308ee05e11210540e41cda4b9a896f9f96c4730 100644 --- a/paddle/operators/elementwise_op.h +++ b/paddle/operators/elementwise_op.h @@ -35,7 +35,7 @@ class ElementwiseOp : public framework::OperatorWithKernel { auto x_dim = ctx->GetInputDim("X"); auto y_dim = ctx->GetInputDim("Y"); PADDLE_ENFORCE_GE(x_dim.size(), y_dim.size(), - "Rank of first input must >= rank of second input.") + "Rank of first input must >= rank of second input."); ctx->SetOutputDim("Out", x_dim); ctx->ShareLoD("X", /*->*/ "Out"); } @@ -43,8 +43,7 @@ class ElementwiseOp : public framework::OperatorWithKernel { class ElementwiseOpMaker : public framework::OpProtoAndCheckerMaker { public: - ElementwiseOpMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker) + ElementwiseOpMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "(Tensor) The first input tensor of elementwise op"); AddInput("Y", "(Tensor) The second input tensor of elementwise op"); @@ -120,7 +119,7 @@ class ElementwiseOpGrad : public framework::OperatorWithKernel { auto out_dims = ctx->GetInputDim(framework::GradVarName("Out")); PADDLE_ENFORCE_GE(x_dims.size(), y_dims.size(), - "Rank of first input must >= rank of second input.") + "Rank of first input must >= rank of second input."); auto x_grad_name = framework::GradVarName("X"); auto y_grad_name = framework::GradVarName("Y"); diff --git a/paddle/operators/elementwise_op_function.h b/paddle/operators/elementwise_op_function.h index 488a35aafc8600bb8bb252fc3a5161c72a2f6df1..560247cb108dce5432bfe66556b9e675a3accc27 100644 --- a/paddle/operators/elementwise_op_function.h +++ b/paddle/operators/elementwise_op_function.h @@ -1,21 +1,26 @@ /* 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 +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 + 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. */ +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/framework/eigen.h" #include "paddle/framework/op_registry.h" #include "paddle/framework/operator.h" +#include "paddle/platform/transform.h" + +#ifdef __NVCC__ +#include +#endif #include "paddle/operators/math/math_function.h" @@ -54,18 +59,175 @@ inline void get_mid_dims(const framework::DDim& x_dims, } } +template +class RowwiseTransformIterator; +template +class MidWiseTransformIterator; + +template +class RowwiseTransformIterator { + public: + RowwiseTransformIterator(const T* ptr, int n) : ptr_(ptr), i_(0), n_(n) {} + + RowwiseTransformIterator& operator++() { + ++i_; + if (UNLIKELY(i_ == n_)) { + i_ = 0; + } + return *this; + } + + bool operator==(const RowwiseTransformIterator& + rhs) const { + return (ptr_ + i_) == &(*rhs); + } + + bool operator!=(const RowwiseTransformIterator& + rhs) const { + return (ptr_ + i_) != &(*rhs); + } + + const T& operator*() { return ptr_[i_]; } + + private: + const T* ptr_; + int i_; + int64_t n_; +}; + +template +class MidWiseTransformIterator { + public: + MidWiseTransformIterator(const T* ptr, int n, int post) + : ptr_(ptr), i_(0), j_(0), n_(n), post_(post) {} + + MidWiseTransformIterator& operator++() { + ++j_; + if (UNLIKELY(j_ == post_)) { + ++i_; + j_ = 0; + if (UNLIKELY(i_ == n_)) { + i_ = 0; + } + } + return *this; + } + + bool operator==(const MidWiseTransformIterator& + rhs) const { + return (ptr_ + i_) == &(*rhs); + } + + bool operator!=(const MidWiseTransformIterator& + rhs) const { + return (ptr_ + i_) != &(*rhs); + } + + const T& operator*() { return ptr_[i_]; } + + private: + const T* ptr_; + int64_t i_; + int64_t j_; + int64_t n_; + int64_t post_; +}; + +#ifdef __NVCC__ +template +class RowwiseTransformIterator + : public thrust::iterator_adaptor< + RowwiseTransformIterator, const T*> { + public: + typedef thrust::iterator_adaptor< + RowwiseTransformIterator, const T*> + super_t; + HOSTDEVICE RowwiseTransformIterator(const T* x, int n) + : super_t(x), begin_(x), n_(n){}; + friend class thrust::iterator_core_access; + + private: + unsigned int n_; + const T* begin_; + HOSTDEVICE typename super_t::reference dereference() const { + return *(begin_ + (this->base() - begin_) % n_); + } +}; + +template +class MidWiseTransformIterator + : public thrust::iterator_adaptor< + MidWiseTransformIterator, const T*> { + public: + typedef thrust::iterator_adaptor< + MidWiseTransformIterator, const T*> + super_t; + HOSTDEVICE MidWiseTransformIterator(const T* x, int n, int post) + : super_t(x), begin_(x), n_(n), post_(post){}; + friend class thrust::iterator_core_access; + + private: + unsigned int post_; + unsigned int n_; + const T* begin_; + HOSTDEVICE typename super_t::reference dereference() const { + return *(begin_ + (((this->base() - begin_) / post_) % n_)); + } +}; +#endif + +template +class TransformFunctor { + public: + TransformFunctor(const framework::Tensor* x, const framework::Tensor* y, + framework::Tensor* z, const DeviceContext& ctx, Functor func) + : x_(x->data()), + y_(y->data()), + z_(z->mutable_data(ctx.GetPlace())), + nx_(x->numel()), + ctx_(ctx), + func_(func) {} + + inline void Run() const { + platform::Transform trans; + trans(ctx_, x_, x_ + nx_, y_, z_, func_); + } + + inline void RunRowWise(int n, int pre) const { + platform::Transform trans; + trans(ctx_, x_, x_ + nx_, RowwiseTransformIterator(y_, n), + z_, func_); + } + + inline void RunMidWise(int n, int pre, int post) const { + platform::Transform trans; + trans(ctx_, x_, x_ + nx_, + MidWiseTransformIterator(y_, n, post), z_, func_); + } + + private: + const T* x_; + const T* y_; + T* z_; + int64_t nx_; + const DeviceContext& ctx_; + Functor func_; +}; + #define EIGEN_FUNCTOR(name, eigen_op) \ struct Eigen##name##Functor { \ - template \ + template \ inline void Run(const framework::Tensor* x, const framework::Tensor* y, \ framework::Tensor* z, \ const framework::ExecutionContext& ctx) { \ auto x_e = framework::EigenVector::Flatten(*x); \ auto y_e = framework::EigenVector::Flatten(*y); \ auto z_e = framework::EigenVector::Flatten(*z); \ - z_e.device(ctx.GetEigenDevice()) = eigen_op(x_e, y_e); \ + z_e.device( \ + *ctx.template device_context().eigen_device()) = \ + eigen_op(x_e, y_e); \ } \ - template \ + template \ inline void RunBroadCast(const framework::Tensor* x, \ const framework::Tensor* y, framework::Tensor* z, \ const framework::ExecutionContext& ctx, int pre, \ @@ -76,9 +238,11 @@ inline void get_mid_dims(const framework::DDim& x_dims, auto y_bcast = y_e.reshape(Eigen::DSizes(1, n)) \ .broadcast(Eigen::DSizes(pre, 1)) \ .reshape(Eigen::DSizes(x_e.size())); \ - z_e.device(ctx.GetEigenDevice()) = eigen_op(x_e, y_bcast); \ + z_e.device( \ + *ctx.template device_context().eigen_device()) = \ + eigen_op(x_e, y_bcast); \ } \ - template \ + template \ inline void RunBroadCast2(const framework::Tensor* x, \ const framework::Tensor* y, \ framework::Tensor* z, \ @@ -90,11 +254,13 @@ inline void get_mid_dims(const framework::DDim& x_dims, auto y_bcast = y_e.reshape(Eigen::DSizes(1, n, 1)) \ .broadcast(Eigen::DSizes(pre, 1, post)) \ .reshape(Eigen::DSizes(x_e.size())); \ - z_e.device(ctx.GetEigenDevice()) = eigen_op(x_e, y_bcast); \ + z_e.device( \ + *ctx.template device_context().eigen_device()) = \ + eigen_op(x_e, y_bcast); \ } \ } -template +template void ElementwiseCompute(const framework::ExecutionContext& ctx) { using Tensor = framework::Tensor; @@ -106,11 +272,11 @@ void ElementwiseCompute(const framework::ExecutionContext& ctx) { auto x_dims = x->dims(); auto y_dims = y->dims(); PADDLE_ENFORCE_GE(x_dims.size(), y_dims.size(), - "Rank of first input must >= rank of second input.") + "Rank of first input must >= rank of second input."); if (x_dims == y_dims) { functor f; - f.template Run(x, y, z, ctx); + f.template Run(x, y, z, ctx); return; } @@ -123,11 +289,11 @@ void ElementwiseCompute(const framework::ExecutionContext& ctx) { get_mid_dims(x_dims, y_dims, axis, pre, n, post); if (post == 1) { functor f; - f.template RunBroadCast(x, y, z, ctx, pre, n); + f.template RunBroadCast(x, y, z, ctx, pre, n); return; } else { functor f; - f.template RunBroadCast2(x, y, z, ctx, pre, n, post); + f.template RunBroadCast2(x, y, z, ctx, pre, n, post); return; } } @@ -144,8 +310,9 @@ EIGEN_FUNCTOR(Mul, EIGEN_MUL); #define EIGEN_DIV(x, y) ((x) / (y)) EIGEN_FUNCTOR(Div, EIGEN_DIV); -template +template void ElementwiseGradCompute(const framework::ExecutionContext& ctx) { using Tensor = framework::Tensor; @@ -154,7 +321,7 @@ void ElementwiseGradCompute(const framework::ExecutionContext& ctx) { auto* out = ctx.Input("Out"); auto* dout = ctx.Input(framework::GradVarName("Out")); - auto place = ctx.GetEigenDevice(); + auto& place = *ctx.template device_context().eigen_device(); auto x_dims = x->dims(); auto y_dims = y->dims(); diff --git a/paddle/operators/elementwise_sub_op.cc b/paddle/operators/elementwise_sub_op.cc index 95d7979e39bfe7b484acb7771d1bd078014293a2..2a8d0845b1800277a7d3cd6ff6c5c984e92197ee 100644 --- a/paddle/operators/elementwise_sub_op.cc +++ b/paddle/operators/elementwise_sub_op.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/operators/elementwise_sub_op.h" #include "paddle/operators/elementwise_op.h" @@ -19,8 +19,7 @@ namespace paddle { namespace operators { class ElementwiseSubOpMaker : public ElementwiseOpMaker { public: - ElementwiseSubOpMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker) + ElementwiseSubOpMaker(OpProto* proto, OpAttrChecker* op_checker) : ElementwiseOpMaker(proto, op_checker) { SetComment("Sub", "$Out = X - Y$"); AddComment(comment_); @@ -34,13 +33,13 @@ REGISTER_OP(elementwise_sub, ops::ElementwiseOp, ops::ElementwiseSubOpMaker, elementwise_sub_grad, ops::ElementwiseOpGrad); REGISTER_OP_CPU_KERNEL( elementwise_sub, - ops::ElementwiseSubKernel, - ops::ElementwiseSubKernel, - ops::ElementwiseSubKernel, - ops::ElementwiseSubKernel); + ops::ElementwiseSubKernel, + ops::ElementwiseSubKernel, + ops::ElementwiseSubKernel, + ops::ElementwiseSubKernel); REGISTER_OP_CPU_KERNEL( elementwise_sub_grad, - ops::ElementwiseSubGradKernel, - ops::ElementwiseSubGradKernel, - ops::ElementwiseSubGradKernel, - ops::ElementwiseSubGradKernel); + ops::ElementwiseSubGradKernel, + ops::ElementwiseSubGradKernel, + ops::ElementwiseSubGradKernel, + ops::ElementwiseSubGradKernel); diff --git a/paddle/operators/elementwise_sub_op.cu b/paddle/operators/elementwise_sub_op.cu index ec23bec35feae26f5463c575b1ab6f58d417e100..7a2516ef6a6d5739e9f122455d289cbfeaaf2549 100644 --- a/paddle/operators/elementwise_sub_op.cu +++ b/paddle/operators/elementwise_sub_op.cu @@ -1,31 +1,32 @@ /* 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 +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 + 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. */ +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. */ #define EIGEN_USE_GPU #include "paddle/operators/elementwise_sub_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL( +REGISTER_OP_CUDA_KERNEL( elementwise_sub, - ops::ElementwiseSubKernel, - ops::ElementwiseSubKernel, - ops::ElementwiseSubKernel, - ops::ElementwiseSubKernel); -REGISTER_OP_GPU_KERNEL( + ops::ElementwiseSubKernel, + ops::ElementwiseSubKernel, + ops::ElementwiseSubKernel, + ops::ElementwiseSubKernel); +REGISTER_OP_CUDA_KERNEL( elementwise_sub_grad, - ops::ElementwiseSubGradKernel, - ops::ElementwiseSubGradKernel, - ops::ElementwiseSubGradKernel, - ops::ElementwiseSubGradKernel); + ops::ElementwiseSubGradKernel, + ops::ElementwiseSubGradKernel, + ops::ElementwiseSubGradKernel, + ops::ElementwiseSubGradKernel); diff --git a/paddle/operators/elementwise_sub_op.h b/paddle/operators/elementwise_sub_op.h index 3f40c1c5bcea5e8473765b039de4ee2a16054f0c..66edf8672d13086f883f0a2ad7ef5802317cc79a 100644 --- a/paddle/operators/elementwise_sub_op.h +++ b/paddle/operators/elementwise_sub_op.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/operators/elementwise_op_function.h" @@ -18,11 +18,11 @@ namespace paddle { namespace operators { -template +template class ElementwiseSubKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { - ElementwiseCompute(ctx); + ElementwiseCompute(ctx); } }; @@ -101,11 +101,11 @@ struct ElementwiseSubBroadCast2GradFunctor { } }; -template +template class ElementwiseSubGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { - ElementwiseGradCompute, + ElementwiseGradCompute, ElementwiseSubOneGradFunctor, ElementwiseSubBroadCastGradFunctor, ElementwiseSubBroadCast2GradFunctor>(ctx); diff --git a/paddle/operators/expand_op.cc b/paddle/operators/expand_op.cc index 282775fcda45fe3bbd72bf04a7ae828f2c840ab7..08fa91ed72aa41ed2f513c090b9085410bb5cc47 100644 --- a/paddle/operators/expand_op.cc +++ b/paddle/operators/expand_op.cc @@ -55,7 +55,7 @@ class ExpandOp : public framework::OperatorWithKernel { class ExpandOpMaker : public framework::OpProtoAndCheckerMaker { public: - ExpandOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker) + ExpandOpMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "(Tensor, default Tensor) A tensor with rank in [1, 6]." @@ -130,7 +130,8 @@ class ExpandGradOp : public framework::OperatorWithKernel { namespace ops = paddle::operators; REGISTER_OP(expand, ops::ExpandOp, ops::ExpandOpMaker, expand_grad, ops::ExpandGradOp); -REGISTER_OP_CPU_KERNEL(expand, - ops::ExpandKernel); REGISTER_OP_CPU_KERNEL( - expand_grad, ops::ExpandGradKernel); + expand, ops::ExpandKernel); +REGISTER_OP_CPU_KERNEL( + expand_grad, + ops::ExpandGradKernel); diff --git a/paddle/operators/expand_op.cu b/paddle/operators/expand_op.cu index 6744562b6c21dd8bfeb7e4cb6b809dc7913aa3a5..84e8fa567b80599d9687fed516eac6fbb308b24a 100644 --- a/paddle/operators/expand_op.cu +++ b/paddle/operators/expand_op.cu @@ -1,23 +1,24 @@ /* 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 +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 + 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. */ +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. */ #define EIGEN_USE_GPU #include "paddle/operators/expand_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(expand, - ops::ExpandKernel); -REGISTER_OP_GPU_KERNEL( - expand_grad, ops::ExpandGradKernel); +REGISTER_OP_CUDA_KERNEL( + expand, ops::ExpandKernel); +REGISTER_OP_CUDA_KERNEL( + expand_grad, + ops::ExpandGradKernel); diff --git a/paddle/operators/expand_op.h b/paddle/operators/expand_op.h index 8ae2c11a5d31dafc1b90d129054ebfabfb761bfe..1d9012cd4a4c6ad596e7d434b5c4ecea1ddcde87 100644 --- a/paddle/operators/expand_op.h +++ b/paddle/operators/expand_op.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 @@ -56,7 +56,7 @@ template using EigenTensor = framework::EigenTensor; -template +template class ExpandKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { @@ -83,12 +83,13 @@ class ExpandKernel : public framework::OpKernel { auto x = EigenTensor::From(*in0); out0->mutable_data(context.GetPlace()); auto y = EigenTensor::From(*out0); - auto place = context.GetEigenDevice(); + auto& place = + *context.template device_context().eigen_device(); y.device(place) = x.broadcast(bcast_dims); } }; -template +template class ExpandGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { @@ -125,7 +126,8 @@ class ExpandGradKernel : public framework::OpKernel { auto* in0 = context.Input(framework::GradVarName("Out")); auto* out0 = context.Output(framework::GradVarName("X")); out0->mutable_data(context.GetPlace()); - out0->CopyFrom(*in0, context.GetPlace(), context.device_context()); + framework::CopyFrom(*in0, context.GetPlace(), context.device_context(), + out0); } else { switch (dims) { REP_EXPAND_GRAD_TEMPLATE(72) @@ -163,7 +165,8 @@ class ExpandGradKernel : public framework::OpKernel { reduce_dims[i] = reduce_dims_vec[i]; } auto out_grad = EigenVector::Flatten(*in0); - x_grad.device(context.GetEigenDevice()) = + x_grad.device( + *context.template device_context().eigen_device()) = out_grad.reshape(reshape_dims).sum(reduce_dims).reshape(x.dimensions()); } }; diff --git a/paddle/operators/feed_op.cc b/paddle/operators/feed_op.cc index 0dd84cbeaafbafd45132b0a0b744554ce7475411..48da52c3b68879a1da8550a5448090f9f1e715d3 100644 --- a/paddle/operators/feed_op.cc +++ b/paddle/operators/feed_op.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/framework/feed_fetch_type.h" #include "paddle/framework/op_registry.h" @@ -25,7 +25,7 @@ class FeedOp : public framework::OperatorBase { const framework::AttributeMap &attrs) : OperatorBase(type, inputs, outputs, attrs) {} void Run(const framework::Scope &scope, - const platform::DeviceContext &dev_ctx) const override { + const platform::Place &place) const override { auto feed_var_name = Input("X"); auto *feed_var = scope.FindVar(feed_var_name); @@ -47,15 +47,19 @@ class FeedOp : public framework::OperatorBase { auto &feed_list = feed_var->Get(); auto &feed_item = feed_list.at(static_cast(col)); auto *out_item = out_var->GetMutable(); - out_item->CopyFrom(feed_item, dev_ctx.GetPlace(), dev_ctx); + + // get device context from pool + platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance(); + auto &dev_ctx = *pool.Get(place); + + framework::CopyFrom(feed_item, place, dev_ctx, out_item); out_item->set_lod(feed_item.lod()); } }; class FeedOpInfoMaker : public framework::OpProtoAndCheckerMaker { public: - FeedOpInfoMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + FeedOpInfoMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "The input of feed op"); AddOutput("Out", "The output of feed op"); diff --git a/paddle/operators/fetch_op.cc b/paddle/operators/fetch_op.cc index 8108ae69dec4bafd1c04d5ab05eef6f467d4c6e8..387d1e0a747f71d85826b52d140c2838112227f6 100644 --- a/paddle/operators/fetch_op.cc +++ b/paddle/operators/fetch_op.cc @@ -1,19 +1,20 @@ /* 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 +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 + 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. */ +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/framework/feed_fetch_type.h" #include "paddle/framework/op_registry.h" +#include "paddle/platform/device_context.h" namespace paddle { namespace operators { @@ -26,7 +27,7 @@ class FetchOp : public framework::OperatorBase { : OperatorBase(type, inputs, outputs, attrs) {} void Run(const framework::Scope &scope, - const platform::DeviceContext &dev_ctx) const override { + const platform::Place &place) const override { auto fetch_var_name = Input("X"); auto *fetch_var = scope.FindVar(fetch_var_name); PADDLE_ENFORCE(fetch_var != nullptr, @@ -51,7 +52,10 @@ class FetchOp : public framework::OperatorBase { // FIXME(yuyang18): Should we assume the fetch operator always generate // CPU outputs? - dst_item.CopyFrom(src_item, platform::CPUPlace(), dev_ctx); + platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance(); + auto &dev_ctx = *pool.Get(place); + + CopyFrom(src_item, platform::CPUPlace(), dev_ctx, &dst_item); dev_ctx.Wait(); dst_item.set_lod(src_item.lod()); @@ -61,8 +65,7 @@ class FetchOp : public framework::OperatorBase { class FetchOpInfoMaker : public framework::OpProtoAndCheckerMaker { public: - FetchOpInfoMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + FetchOpInfoMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "The input of fetch op"); AddOutput("Out", "The output of fetch op"); diff --git a/paddle/operators/fill_constant_batch_size_like_op.cc b/paddle/operators/fill_constant_batch_size_like_op.cc index 985b5d1e865e513d833bff72dcd20a8f20851d8c..852ecdfe45e7f4737a505c1f722d25457ad6ad32 100644 --- a/paddle/operators/fill_constant_batch_size_like_op.cc +++ b/paddle/operators/fill_constant_batch_size_like_op.cc @@ -49,10 +49,10 @@ class FillConstantBatchSizeLikeOp : public framework::OperatorWithKernel { } protected: - framework::OpKernelType GetKernelType( + framework::OpKernelType GetActualKernelType( const framework::ExecutionContext &ctx) const override { return framework::OpKernelType( - static_cast(ctx.Attr("data_type")), + static_cast(ctx.Attr("dtype")), ctx.device_context()); } }; @@ -60,13 +60,12 @@ class FillConstantBatchSizeLikeOp : public framework::OperatorWithKernel { class FillConstantBatchSizeLikeOpMaker : public framework::OpProtoAndCheckerMaker { public: - FillConstantBatchSizeLikeOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + FillConstantBatchSizeLikeOpMaker(OpProto *proto, OpAttrChecker *op_checker) : framework::OpProtoAndCheckerMaker(proto, op_checker) { - AddAttr("data_type", + AddAttr("dtype", "(int, default 5 (FP32)) " "Output data type") - .SetDefault(framework::DataType::FP32); + .SetDefault(framework::proto::DataType::FP32); AddInput("Input", "(Tensor) Tensor " "whose dim_idx th dimension is used to specify the batch_size"); @@ -100,8 +99,11 @@ REGISTER_OPERATOR(fill_constant_batch_size_like, ops::FillConstantBatchSizeLikeOpMaker); REGISTER_OP_CPU_KERNEL( fill_constant_batch_size_like, - ops::FillConstantBatchSizeLikeOpKernel, - ops::FillConstantBatchSizeLikeOpKernel, - ops::FillConstantBatchSizeLikeOpKernel, - ops::FillConstantBatchSizeLikeOpKernel, + ops::FillConstantBatchSizeLikeOpKernel, + ops::FillConstantBatchSizeLikeOpKernel, + ops::FillConstantBatchSizeLikeOpKernel); diff --git a/paddle/operators/fill_constant_batch_size_like_op.cu.cc b/paddle/operators/fill_constant_batch_size_like_op.cu.cc index 9e7a1eeab863c962ca72908e561e12a04d5021c5..608f4b91623e4ddf0240c37be7a8e56117dd40f2 100644 --- a/paddle/operators/fill_constant_batch_size_like_op.cu.cc +++ b/paddle/operators/fill_constant_batch_size_like_op.cu.cc @@ -1,25 +1,28 @@ /* 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 +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 + 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. */ +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/operators/fill_constant_batch_size_like_op.h" #include "paddle/framework/op_registry.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL( +REGISTER_OP_CUDA_KERNEL( fill_constant_batch_size_like, - ops::FillConstantBatchSizeLikeOpKernel, - ops::FillConstantBatchSizeLikeOpKernel, - ops::FillConstantBatchSizeLikeOpKernel, - ops::FillConstantBatchSizeLikeOpKernel, + ops::FillConstantBatchSizeLikeOpKernel, + ops::FillConstantBatchSizeLikeOpKernel, + ops::FillConstantBatchSizeLikeOpKernel); diff --git a/paddle/operators/fill_constant_batch_size_like_op.h b/paddle/operators/fill_constant_batch_size_like_op.h index 339d97a30a5819ab488e83990651ba99212239ec..66da9d0307e36db3726f30518c8c57a923e54388 100644 --- a/paddle/operators/fill_constant_batch_size_like_op.h +++ b/paddle/operators/fill_constant_batch_size_like_op.h @@ -19,7 +19,7 @@ limitations under the License. */ namespace paddle { namespace operators { -template +template class FillConstantBatchSizeLikeOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { @@ -27,8 +27,9 @@ class FillConstantBatchSizeLikeOpKernel : public framework::OpKernel { out->mutable_data(ctx.GetPlace()); auto value = ctx.Attr("value"); - math::SetConstant setter; - setter(ctx.device_context(), out, static_cast(value)); + math::SetConstant setter; + setter(ctx.template device_context(), out, + static_cast(value)); } }; diff --git a/paddle/operators/fill_constant_op.cc b/paddle/operators/fill_constant_op.cc index 818f113b90a4c239a857791fb9957e51d3287b97..dcd43a30c86b62d79f52ac640f14b295a062146c 100644 --- a/paddle/operators/fill_constant_op.cc +++ b/paddle/operators/fill_constant_op.cc @@ -15,6 +15,7 @@ limitations under the License. */ #include "paddle/framework/data_type.h" #include "paddle/framework/op_registry.h" #include "paddle/operators/math/math_function.h" +#include "paddle/platform/device_context.h" namespace paddle { namespace operators { @@ -33,8 +34,9 @@ class FillConstantOp : public framework::OperatorBase { public: using framework::OperatorBase::OperatorBase; void Run(const framework::Scope &scope, - const platform::DeviceContext &dev_ctx) const override { - auto data_type = static_cast(Attr("data_type")); + const platform::Place &dev_place) const override { + auto data_type = + static_cast(Attr("dtype")); auto value = Attr("value"); auto force_cpu = Attr("force_cpu"); auto &out = @@ -44,21 +46,23 @@ class FillConstantOp : public framework::OperatorBase { auto cpu = platform::CPUPlace(); out.mutable_data(cpu, framework::ToTypeIndex(data_type)); } else { - out.mutable_data(dev_ctx.GetPlace(), framework::ToTypeIndex(data_type)); + out.mutable_data(dev_place, framework::ToTypeIndex(data_type)); } + + platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance(); + auto &dev_ctx = *pool.Get(dev_place); math::set_constant(dev_ctx, &out, value); } }; class FillConstantOpMaker : public framework::OpProtoAndCheckerMaker { public: - FillConstantOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + FillConstantOpMaker(OpProto *proto, OpAttrChecker *op_checker) : framework::OpProtoAndCheckerMaker(proto, op_checker) { - AddAttr("data_type", + AddAttr("dtype", "(int, default 5 (FP32)) " "Output data type") - .SetDefault(framework::DataType::FP32); + .SetDefault(framework::proto::DataType::FP32); AddAttr>("shape", "(vector) The shape of the output"); AddAttr("value", "(float, default 0) The value to be filled") .SetDefault(0.0f); diff --git a/paddle/operators/fill_op.cc b/paddle/operators/fill_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..084ba1db62de0a6bf6829f8e9f4c274fb777e879 --- /dev/null +++ b/paddle/operators/fill_op.cc @@ -0,0 +1,114 @@ +/* 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. */ + +#include "paddle/framework/data_type.h" +#include "paddle/framework/op_registry.h" +#include "paddle/operators/detail/safe_ref.h" +#include "paddle/platform/device_context.h" + +namespace paddle { +namespace operators { + +struct FillOpVisitor { + FillOpVisitor(framework::LoDTensor *tensor, const std::vector &value) + : tensor_(tensor), value_(value) {} + + template + void operator()() const { + platform::CPUPlace cpu; + auto *data = tensor_->mutable_data(cpu); + std::transform(value_.data(), value_.data() + tensor_->numel(), data, + [](float dat) { return static_cast(dat); }); + } + + framework::LoDTensor *tensor_; + const std::vector &value_; +}; + +class FillOp : public framework::OperatorBase { + public: + FillOp(const std::string &type, const framework::VariableNameMap &inputs, + const framework::VariableNameMap &outputs, + const framework::AttributeMap &attrs) + : OperatorBase(type, inputs, outputs, attrs) {} + void Run(const framework::Scope &scope, + const platform::Place &place) const override { + auto &out = + detail::Ref(detail::Ref(scope.FindVar(Output("Out")), + "Cannot find variable %s", Output("Out")) + .GetMutable()); + out.Resize(framework::make_ddim(Attr>("shape"))); + auto dtype = static_cast(Attr("dtype")); + platform::CPUPlace cpu; + auto force_cpu = Attr("force_cpu"); + out.mutable_data(force_cpu ? cpu : place, framework::ToTypeIndex(dtype)); + + framework::LoDTensor tensor; + + if (force_cpu || platform::is_cpu_place(place)) { + tensor.ShareDataWith(out); + } else { + // Always make tensor in CPU memory. + tensor.Resize(out.dims()); + tensor.mutable_data(cpu, framework::ToTypeIndex(dtype)); + } + + framework::VisitDataType( + dtype, FillOpVisitor(&tensor, Attr>("value"))); + + if (!force_cpu && platform::is_gpu_place(place)) { + // Copy tensor to out + platform::DeviceContextPool &pool = + platform::DeviceContextPool::Instance(); + auto &dev_ctx = *pool.Get(place); + framework::CopyFrom(tensor, place, dev_ctx, &out); + } + } +}; + +class FillOpMaker : public framework::OpProtoAndCheckerMaker { + public: + FillOpMaker(OpProto *proto, OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddComment(R"DOC(Fill operator + +Fill an tensor with `value` and `shape`. The type of the tensor is specify by +`dtype`. +)DOC"); + AddOutput("Out", "(LoDTensor) The output tensor."); + AddAttr>( + "value", "The float values of tensor, which are flatten in row major"); + AddAttr>("shape", "The shape of output tensor"); + AddAttr("dtype", "The data type of output tensor, Default is float") + .SetDefault(framework::proto::DataType::FP32); + AddAttr("force_cpu", + "Whether the output tensor must be at CPU memory or not. " + "Default is false.") + .SetDefault(false); + } +}; + +class FillOpInferShape : public framework::InferShapeBase { + public: + void operator()(framework::InferShapeContext *context) const override { + context->SetOutputDim( + "Out", + framework::make_ddim(context->Attrs().Get>("shape"))); + } +}; + +} // namespace operators +} // namespace paddle +namespace ops = paddle::operators; +REGISTER_OPERATOR(fill, ops::FillOp, ops::FillOpInferShape, ops::FillOpMaker); diff --git a/paddle/operators/fill_zeros_like_op.cc b/paddle/operators/fill_zeros_like_op.cc index 95fb5932b8b555e1357adc9fdfb7b6e6db7da71d..b4ae1de876010effff6bf577a4e33043f6760a4f 100644 --- a/paddle/operators/fill_zeros_like_op.cc +++ b/paddle/operators/fill_zeros_like_op.cc @@ -24,20 +24,19 @@ class FillZerosLikeOp : public framework::OperatorWithKernel { void InferShape(framework::InferShapeContext *ctx) const override { PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) of FillZerosLikeOp should not be null."); - PADDLE_ENFORCE(ctx->HasOutput("Y"), - "Output(Y) of FillZerosLikeOp should not be null."); - ctx->SetOutputDim("Y", ctx->GetInputDim("X")); - ctx->ShareLoD("X", /*->*/ "Y"); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of FillZerosLikeOp should not be null."); + ctx->SetOutputDim("Out", ctx->GetInputDim("X")); + ctx->ShareLoD("X", /*->*/ "Out"); } }; class FillZerosLikeOpMaker : public framework::OpProtoAndCheckerMaker { public: - FillZerosLikeOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + FillZerosLikeOpMaker(OpProto *proto, OpAttrChecker *op_checker) : framework::OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "The input of fill-zeros-like op."); - AddOutput("Y", "The variable will be filled up with zeros."); + AddOutput("Out", "The variable will be filled up with zeros."); AddComment(R"DOC( FillZerosLike Operator. @@ -54,8 +53,9 @@ namespace ops = paddle::operators; REGISTER_OP_WITHOUT_GRADIENT(fill_zeros_like, ops::FillZerosLikeOp, ops::FillZerosLikeOpMaker); REGISTER_OP_CPU_KERNEL( - fill_zeros_like, ops::FillZerosLikeKernel, - ops::FillZerosLikeKernel, - ops::FillZerosLikeKernel, - ops::FillZerosLikeKernel, - ops::FillZerosLikeKernel); + fill_zeros_like, + ops::FillZerosLikeKernel, + ops::FillZerosLikeKernel, + ops::FillZerosLikeKernel, + ops::FillZerosLikeKernel, + ops::FillZerosLikeKernel); diff --git a/paddle/operators/fill_zeros_like_op.cu.cc b/paddle/operators/fill_zeros_like_op.cu.cc index 1501a17441072223ba0e8cf5b6c8cdd5e903a467..b7048e8f5857e646e16d5017593f5d3c6e79ea7e 100644 --- a/paddle/operators/fill_zeros_like_op.cu.cc +++ b/paddle/operators/fill_zeros_like_op.cu.cc @@ -1,24 +1,25 @@ /* 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 +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 + 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. */ +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/operators/fill_zeros_like_op.h" #include "paddle/framework/op_registry.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL( - fill_zeros_like, ops::FillZerosLikeKernel, - ops::FillZerosLikeKernel, - ops::FillZerosLikeKernel, - ops::FillZerosLikeKernel, - ops::FillZerosLikeKernel); +REGISTER_OP_CUDA_KERNEL( + fill_zeros_like, + ops::FillZerosLikeKernel, + ops::FillZerosLikeKernel, + ops::FillZerosLikeKernel, + ops::FillZerosLikeKernel, + ops::FillZerosLikeKernel); diff --git a/paddle/operators/fill_zeros_like_op.h b/paddle/operators/fill_zeros_like_op.h index 7e7d78eea2bce427d6ad4dfb77bcb4ace35cd287..351ecf8b2f1d945fabdd1d6c5ed56f76f3caae61 100644 --- a/paddle/operators/fill_zeros_like_op.h +++ b/paddle/operators/fill_zeros_like_op.h @@ -19,15 +19,16 @@ limitations under the License. */ namespace paddle { namespace operators { -template +template class FillZerosLikeKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { - auto* out = context.Output("Y"); + auto* out = context.Output("Out"); out->mutable_data(context.GetPlace()); - math::SetConstant setter; - setter(context.device_context(), out, static_cast(0)); + math::SetConstant setter; + setter(context.template device_context(), out, + static_cast(0)); } }; diff --git a/paddle/operators/ftrl_op.cc b/paddle/operators/ftrl_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..d00700823d48eb2ea4fc64d1fa2989f18c7c5f18 --- /dev/null +++ b/paddle/operators/ftrl_op.cc @@ -0,0 +1,139 @@ +/* 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. */ + +#include "paddle/operators/ftrl_op.h" + +namespace paddle { +namespace operators { + +class FTRLOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Param"), + "Input(Param) of FTRL should not be null."); + PADDLE_ENFORCE(ctx->HasInput("SquaredAccumulator"), + "Input(SquaredAccumulator) of FTRL should not be null."); + PADDLE_ENFORCE(ctx->HasInput("LinearAccumulator"), + "Input(LinearAccumulator) of FTRL should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Grad"), + "Input(Grad) of FTRL should not be null."); + PADDLE_ENFORCE(ctx->HasInput("LearningRate"), + "Input(LearningRate) of FTRL should not be null."); + + PADDLE_ENFORCE(ctx->HasOutput("ParamOut"), + "Output(ParamOut) of FTRL should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("SquaredAccumOut"), + "Output(SquaredAccumOut) of FTRL should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("LinearAccumOut"), + "Output(LinearAccumOut) of FTRL should not be null."); + + auto param_dim = ctx->GetInputDim("Param"); + PADDLE_ENFORCE_EQ(param_dim, ctx->GetInputDim("Grad"), + "Two input of FTRL Op's dimension must be same."); + + auto lr_dim = ctx->GetInputDim("LearningRate"); + PADDLE_ENFORCE_EQ(framework::product(lr_dim), 1, + "Learning Rate should be a scalar."); + + ctx->SetOutputDim("ParamOut", param_dim); + ctx->SetOutputDim("SquaredAccumOut", param_dim); + ctx->SetOutputDim("LinearAccumOut", param_dim); + } +}; + +class FTRLOpMaker : public framework::OpProtoAndCheckerMaker { + public: + FTRLOpMaker(OpProto *proto, OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("Param", + "(Tensor, default Tensor) " + "Input parameter value that has to be updated."); + AddInput("SquaredAccumulator", + "(Tensor, default Tensor) " + "Accumulator that accumulates squared gradients."); + AddInput("LinearAccumulator", + "(Tensor, default Tensor) " + "Accumulator that accumulates linear gradients."); + AddInput("Grad", + "(Tensor, default Tensor) " + "Input gradient of the parameter."); + AddInput("LearningRate", + "(Tensor, default Tensor) " + "The learning rate should be a tensor of size 1."); + + AddOutput("ParamOut", "(Tensor) Output updated parameter value."); + AddOutput("SquaredAccumOut", + "(Tensor) Output accumulated squared" + " gradients."); + AddOutput("LinearAccumOut", + "(Tensor) Output accumulated linear" + " gradients."); + + AddAttr("l1", + "(float, default 0.0) " + "L1 regularization strength.") + .SetDefault(0.0f); + AddAttr("l2", + "(float, default 0.0) " + "L2 regularization strength.") + .SetDefault(0.0f); + AddAttr("lr_power", + "(float, default -0.5f) " + "Learning Rate Power.") + .SetDefault(-0.5f); + AddComment(R"DOC( +FTRL (Follow The Regularized Leader) Operator. + +Optimizer that implements the FTRL algorithm: + +$$ +new\_accum = squared\_accum + grad^2 \\ +if (lr\_power == -0.5) { + linear\_accum += grad - (\surd(new\_accum) - \surd(squared\_accum)) / + (learning\_rate * param) \\ +} else { + linear\_accum += grad - + (new\_accum^{-lr\_power} - accum^{-lr\_power}) / + (learning\_rate * param) \\ +} + +x = (l1 * sign(linear\_accum) - linear\_accum) +if (lr\_power == -0.5) { + y = \frac{\surd(new\_accum)}{learning\_rate} + (2 * l2) \\ + pre\_shrink = \frac{x}{y} \\ + param = (abs(linear\_accum) > l1).select(pre\_shrink, 0.0) \\ +} else { + y = \frac{new\_accum^{-lr\_power}}{learning\_rate} + (2 * l2) \\ + pre\_shrink = \frac{x}{y} \\ + param = (abs(linear\_accum) > l1).select(pre\_shrink, 0.0) \\ +} +squared\_accum += grad^2; +$$ + +The paper that proposed Follow The Regularized Leader (FTRL): +(https://www.eecs.tufts.edu/~dsculley/papers/ad-click-prediction.pdf) + +)DOC"); + } +}; +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP_WITHOUT_GRADIENT(ftrl, ops::FTRLOp, ops::FTRLOpMaker); +REGISTER_OP_CPU_KERNEL( + ftrl, ops::FTRLOpKernel); diff --git a/paddle/operators/ftrl_op.cu b/paddle/operators/ftrl_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..abbbe7adbe6bd14f55f7f941c5e6740fada24910 --- /dev/null +++ b/paddle/operators/ftrl_op.cu @@ -0,0 +1,19 @@ +/* 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. */ + +#define EIGEN_USE_GPU +#include "paddle/operators/ftrl_op.h" + +namespace ops = paddle::operators; +REGISTER_OP_CUDA_KERNEL( + ftrl, ops::FTRLOpKernel); diff --git a/paddle/operators/ftrl_op.h b/paddle/operators/ftrl_op.h new file mode 100644 index 0000000000000000000000000000000000000000..4eea04cd8d61bb34fc612e0ca1765a664e329ca9 --- /dev/null +++ b/paddle/operators/ftrl_op.h @@ -0,0 +1,96 @@ +/* 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 +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; +template +using EigenVector = framework::EigenVector; + +template +class FTRLOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* param_out = ctx.Output("ParamOut"); + auto* sq_accum_out = ctx.Output("SquaredAccumOut"); + auto* lin_accum_out = ctx.Output("LinearAccumOut"); + + param_out->mutable_data(ctx.GetPlace()); + sq_accum_out->mutable_data(ctx.GetPlace()); + lin_accum_out->mutable_data(ctx.GetPlace()); + + auto grad = ctx.Input("Grad"); + + auto l1 = static_cast(ctx.Attr("l1")); + auto l2 = static_cast(ctx.Attr("l2")); + auto lr_power = static_cast(ctx.Attr("lr_power")); + + auto p = EigenVector::Flatten(*ctx.Input("Param")); + auto sq_accum = + EigenVector::Flatten(*ctx.Input("SquaredAccumulator")); + auto lin_accum = + EigenVector::Flatten(*ctx.Input("LinearAccumulator")); + auto g = EigenVector::Flatten(*grad); + auto lr = EigenVector::Flatten(*ctx.Input("LearningRate")); + + auto p_out = EigenVector::Flatten(*param_out); + auto s_acc_out = EigenVector::Flatten(*sq_accum_out); + auto l_acc_out = EigenVector::Flatten(*lin_accum_out); + auto& place = *ctx.template device_context().eigen_device(); + + Eigen::DSizes grad_dsize(grad->numel()); + + auto new_accum = sq_accum + g * g; + // Special case for lr_power = -0.5 + if (lr_power == static_cast(-0.5)) { + l_acc_out.device(place) = + lin_accum + g - + ((new_accum.sqrt() - sq_accum.sqrt()) / lr.broadcast(grad_dsize)) * p; + } else { + l_acc_out.device(place) = + lin_accum + g - + ((new_accum.pow(-lr_power) - sq_accum.pow(-lr_power)) / + lr.broadcast(grad_dsize)) * + p; + } + + auto x = (l_acc_out.constant(l1) * l_acc_out.sign() - l_acc_out); + if (lr_power == static_cast(-0.5)) { + auto y = (new_accum.sqrt() / lr.broadcast(grad_dsize)) + + l_acc_out.constant(static_cast(2) * l2); + auto pre_shrink = x / y; + p_out.device(place) = + (l_acc_out.abs() > l_acc_out.constant(l1)) + .select(pre_shrink, p.constant(static_cast(0))); + } else { + auto y = (new_accum.pow(-lr_power) / lr.broadcast(grad_dsize)) + + l_acc_out.constant(static_cast(2) * l2); + auto pre_shrink = x / y; + p_out.device(place) = + (l_acc_out.abs() > l_acc_out.constant(l1)) + .select(pre_shrink, p.constant(static_cast(0))); + } + + s_acc_out.device(place) = sq_accum + g * g; + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/gather.cu.h b/paddle/operators/gather.cu.h index 8d04ecd284226c7b4c6cdd5531915fee2d94ce61..9840c066f053e5e1cd1c756d4dd938eace1a5eb4 100644 --- a/paddle/operators/gather.cu.h +++ b/paddle/operators/gather.cu.h @@ -1,16 +1,16 @@ -/* Copyright (c) 2016 PaddlePaddle Authors All Rights Reserve. +/* 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 +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 + 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. */ +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/framework/tensor.h" @@ -20,7 +20,7 @@ namespace paddle { namespace operators { using framework::Tensor; -using platform::Place; +using platform::DeviceContext; #define CUDA_1D_KERNEL_LOOP(i, n) \ for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < (n); \ diff --git a/paddle/operators/gather_op.cc b/paddle/operators/gather_op.cc index 8f80fb162519f60fcce897b3c31a3507bbf6ba6d..45e9d8df702403e66f9100e4edaf5c17470eb20d 100644 --- a/paddle/operators/gather_op.cc +++ b/paddle/operators/gather_op.cc @@ -40,7 +40,7 @@ class GatherOp : public framework::OperatorWithKernel { } protected: - framework::OpKernelType GetKernelType( + framework::OpKernelType GetActualKernelType( const framework::ExecutionContext& ctx) const override { return framework::OpKernelType( framework::ToDataType(ctx.Input("X")->type()), @@ -57,7 +57,7 @@ class GatherGradOp : public framework::OperatorWithKernel { } protected: - framework::OpKernelType GetKernelType( + framework::OpKernelType GetActualKernelType( const framework::ExecutionContext& ctx) const override { return framework::OpKernelType( framework::ToDataType(ctx.Input("X")->type()), @@ -67,7 +67,7 @@ class GatherGradOp : public framework::OperatorWithKernel { class GatherOpMaker : public framework::OpProtoAndCheckerMaker { public: - GatherOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker) + GatherOpMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "The source input of gather op"); AddInput("Index", "The index input of gather op"); diff --git a/paddle/operators/gather_op.cu b/paddle/operators/gather_op.cu index 92219d6a433e6db0bb9886ed8670cbafaa843ff8..eec2415e1de2434de0a920567863d421d2d3032d 100644 --- a/paddle/operators/gather_op.cu +++ b/paddle/operators/gather_op.cu @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 "gather.cu.h" #include "paddle/framework/eigen.h" @@ -49,7 +49,8 @@ class GatherGradOpCUDAKernel : public framework::OpKernel { dX->mutable_data(ctx.GetPlace()); auto dxt = framework::EigenVector::Flatten(*dX); - auto place = ctx.GetEigenDevice(); + auto &place = *ctx.template device_context() + .eigen_device(); dxt.device(place) = dxt.constant(static_cast(0)); GPUScatterAssign(ctx.device_context(), *dO, *Index, dX); @@ -60,5 +61,5 @@ class GatherGradOpCUDAKernel : public framework::OpKernel { } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(gather, ops::GatherOpCUDAKernel); -REGISTER_OP_GPU_KERNEL(gather_grad, ops::GatherGradOpCUDAKernel); +REGISTER_OP_CUDA_KERNEL(gather, ops::GatherOpCUDAKernel); +REGISTER_OP_CUDA_KERNEL(gather_grad, ops::GatherGradOpCUDAKernel); diff --git a/paddle/operators/gather_op.h b/paddle/operators/gather_op.h index 8276ed0d3d8b676aafab45fae70942e78b72b8e6..1a1ba0c41aef95d3dc8cc929db72770a7bd08b18 100644 --- a/paddle/operators/gather_op.h +++ b/paddle/operators/gather_op.h @@ -53,7 +53,8 @@ class GatherGradientOpKernel : public framework::OpKernel { dX->mutable_data(ctx.GetPlace()); auto dxt = framework::EigenVector::Flatten(*dX); - auto place = ctx.GetEigenDevice(); + auto &place = *ctx.template device_context() + .eigen_device(); dxt.device(place) = dxt.constant(static_cast(0)); ScatterAssign(ctx.device_context(), *dO, *Index, dX); diff --git a/paddle/operators/gaussian_random_op.cc b/paddle/operators/gaussian_random_op.cc index 53ad86c6c48d1868f4495af51661d91b39a84f0b..9ed493a7d027e1bd3e8c5fca376678fd5fcf14f1 100644 --- a/paddle/operators/gaussian_random_op.cc +++ b/paddle/operators/gaussian_random_op.cc @@ -1,13 +1,16 @@ /* 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. */ + +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/framework/op_registry.h" @@ -57,18 +60,17 @@ class GaussianRandomOp : public framework::OperatorWithKernel { } protected: - framework::OpKernelType GetKernelType( + framework::OpKernelType GetActualKernelType( const framework::ExecutionContext& ctx) const override { return framework::OpKernelType( - static_cast(ctx.Attr("data_type")), + static_cast(ctx.Attr("dtype")), ctx.device_context()); } }; class GaussianRandomOpMaker : public framework::OpProtoAndCheckerMaker { public: - GaussianRandomOpMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker) + GaussianRandomOpMaker(OpProto* proto, OpAttrChecker* op_checker) : framework::OpProtoAndCheckerMaker(proto, op_checker) { AddOutput("Out", "Output matrix of gaussian random op"); @@ -88,10 +90,10 @@ class GaussianRandomOpMaker : public framework::OpProtoAndCheckerMaker { "Random seed of generator." "0 means use system wide seed.") .SetDefault(0); - AddAttr("data_type", + AddAttr("dtype", "(int, default 5(FP32)) " "Output data type.") - .SetDefault(framework::DataType::FP32); + .SetDefault(framework::proto::DataType::FP32); AddComment(R"DOC( GaussianRandom Operator. diff --git a/paddle/operators/gaussian_random_op.cu b/paddle/operators/gaussian_random_op.cu index 315560bf1ba8a66b9a3b7d79510d202885e845d6..8a70db17e17ebf7d5bad1e1ee6a2acdff1b85a09 100644 --- a/paddle/operators/gaussian_random_op.cu +++ b/paddle/operators/gaussian_random_op.cu @@ -1,16 +1,16 @@ /* 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. */ -#include -#include +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 "paddle/framework/op_registry.h" @@ -60,5 +60,5 @@ class GPUGaussianRandomKernel : public framework::OpKernel { } // namespace operators } // namespace paddle -REGISTER_OP_GPU_KERNEL(gaussian_random, - paddle::operators::GPUGaussianRandomKernel); +REGISTER_OP_CUDA_KERNEL(gaussian_random, + paddle::operators::GPUGaussianRandomKernel); diff --git a/paddle/operators/gru_op.cc b/paddle/operators/gru_op.cc index 5aa03f8916a67222fb0ca5781533766063e52683..76f2adefede3b4bc4035f86f8f8663eed29343ae 100644 --- a/paddle/operators/gru_op.cc +++ b/paddle/operators/gru_op.cc @@ -1,13 +1,16 @@ /* 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. */ + +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/operators/gru_op.h" @@ -67,7 +70,7 @@ class GRUOp : public framework::OperatorWithKernel { class GRUOpMaker : public framework::OpProtoAndCheckerMaker { public: - GRUOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker) + GRUOpMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("Input", "(LoDTensor) The first input is a LodTensor, which supports " @@ -213,8 +216,9 @@ class GRUGradOp : public framework::OperatorWithKernel { namespace ops = paddle::operators; REGISTER_OP(gru, ops::GRUOp, ops::GRUOpMaker, gru_grad, ops::GRUGradOp); -REGISTER_OP_CPU_KERNEL(gru, ops::GRUKernel, - ops::GRUKernel); -REGISTER_OP_CPU_KERNEL(gru_grad, - ops::GRUGradKernel, - ops::GRUGradKernel); +REGISTER_OP_CPU_KERNEL( + gru, ops::GRUKernel, + ops::GRUKernel); +REGISTER_OP_CPU_KERNEL( + gru_grad, ops::GRUGradKernel, + ops::GRUGradKernel); diff --git a/paddle/operators/gru_op.cu.cc b/paddle/operators/gru_op.cu.cc index 0ceff94ec3ddaadbd5f0ca4f5a4eebe6cb8ee3a9..9cb0cc42d5589792aae6d99cec807aac6e4991b6 100644 --- a/paddle/operators/gru_op.cu.cc +++ b/paddle/operators/gru_op.cu.cc @@ -1,22 +1,23 @@ /* 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 +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 + 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. */ +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/operators/gru_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(gru, ops::GRUKernel, - ops::GRUKernel); -REGISTER_OP_GPU_KERNEL(gru_grad, - ops::GRUGradKernel, - ops::GRUGradKernel); +REGISTER_OP_CUDA_KERNEL( + gru, ops::GRUKernel, + ops::GRUKernel); +REGISTER_OP_CUDA_KERNEL( + gru_grad, ops::GRUGradKernel, + ops::GRUGradKernel); diff --git a/paddle/operators/gru_op.h b/paddle/operators/gru_op.h index 1b18368e0e16365682520b62a7f6adab0cbb527f..b1957fb9ce6add8628cb206abf2c569d3f615c85 100644 --- a/paddle/operators/gru_op.h +++ b/paddle/operators/gru_op.h @@ -1,19 +1,20 @@ /* 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 +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 + 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. */ +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/operators/math/detail/activation_functions.h" #include "paddle/operators/math/gru_compute.h" #include "paddle/operators/math/math_function.h" #include "paddle/operators/math/sequence2batch.h" @@ -27,16 +28,16 @@ namespace operators { using LoDTensor = framework::LoDTensor; using Tensor = framework::Tensor; -template -inline void ReorderInitState(const platform::DeviceContext& ctx, +template +inline void ReorderInitState(const DeviceContext& ctx, const framework::Tensor& src, const size_t* index, framework::Tensor* dst, bool indexed_src) { - math::CopyMatrixRowsFunctor row_shuffle; + math::CopyMatrixRowsFunctor row_shuffle; dst->mutable_data(src.dims(), ctx.GetPlace()); row_shuffle(ctx, src, index, *dst, indexed_src); } -template +template class GRUKernel : public framework::OpKernel { public: void BatchCompute(const framework::ExecutionContext& context) const { @@ -60,19 +61,19 @@ class GRUKernel : public framework::OpKernel { auto hidden_dims = hidden->dims(); bool is_reverse = context.Attr("is_reverse"); - math::LoDTensor2BatchFunctor to_batch; - auto& dev_ctx = context.device_context(); + math::LoDTensor2BatchFunctor to_batch; + auto& dev_ctx = context.template device_context(); to_batch(dev_ctx, *input, *batch_gate, true, is_reverse); if (bias) { - math::RowwiseAdd add_bias; + math::RowwiseAdd add_bias; add_bias(dev_ctx, *batch_gate, *bias, batch_gate); } int frame_size = hidden_dims[1]; - math::hl_gru_value gru_value; - gru_value.gateWeight = const_cast(weight_data); - gru_value.stateWeight = + math::GRUMetaValue gru_value; + gru_value.gate_weight = const_cast(weight_data); + gru_value.state_weight = const_cast(weight_data + 2 * frame_size * frame_size); Tensor ordered_h0; const size_t* order = batch_gate->lod()[2].data(); @@ -80,14 +81,19 @@ class GRUKernel : public framework::OpKernel { // Since the batch computing for GRU reorders the input sequences // according to their length. The initialized cell state also needs // to reorder. - ReorderInitState(context.device_context(), *h0, order, - &ordered_h0, true); - gru_value.prevOutValue = ordered_h0.data(); + ReorderInitState( + context.template device_context(), *h0, order, + &ordered_h0, true); + gru_value.prev_out_value = ordered_h0.data(); } else { - gru_value.prevOutValue = nullptr; + gru_value.prev_out_value = nullptr; } auto batch_starts = batch_gate->lod()[0]; size_t num_batch = batch_starts.size() - 1; + auto active_node = math::detail::GetActivationType( + context.Attr("activation")); + auto active_gate = math::detail::GetActivationType( + context.Attr("gate_activation")); for (size_t n = 0; n < num_batch; n++) { int bstart = static_cast(batch_starts[n]); int bend = static_cast(batch_starts[n + 1]); @@ -96,17 +102,16 @@ class GRUKernel : public framework::OpKernel { Tensor gate_t = batch_gate->Slice(bstart, bend); Tensor reset_hidden_prev_t = batch_reset_hidden_prev->Slice(bstart, bend); Tensor hidden_t = batch_hidden->Slice(bstart, bend); - gru_value.outputValue = hidden_t.data(); - gru_value.gateValue = gate_t.data(); - gru_value.resetOutputValue = reset_hidden_prev_t.data(); - math::GRUUnitFunctor::compute( - dev_ctx, gru_value, frame_size, cur_batch_size, - math::ActiveType(context.Attr("activation")), - math::ActiveType(context.Attr("gate_activation"))); - gru_value.prevOutValue = gru_value.outputValue; + gru_value.output_value = hidden_t.data(); + gru_value.gate_value = gate_t.data(); + gru_value.reset_output_value = reset_hidden_prev_t.data(); + math::GRUUnitFunctor::compute( + dev_ctx, gru_value, frame_size, cur_batch_size, active_node, + active_gate); + gru_value.prev_out_value = gru_value.output_value; } - math::Batch2LoDTensorFunctor to_seq; + math::Batch2LoDTensorFunctor to_seq; batch_hidden->set_lod(batch_gate->lod()); to_seq(dev_ctx, *batch_hidden, *hidden); } @@ -116,7 +121,7 @@ class GRUKernel : public framework::OpKernel { } }; -template +template class GRUGradKernel : public framework::OpKernel { public: void BatchCompute(const framework::ExecutionContext& context) const { @@ -141,14 +146,14 @@ class GRUGradKernel : public framework::OpKernel { auto hidden_dims = hidden->dims(); int frame_size = hidden_dims[1]; - math::LoDTensor2BatchFunctor to_batch; + math::LoDTensor2BatchFunctor to_batch; LoDTensor batch_hidden_grad, batch_gate_grad, batch_reset_hidden_prev_grad; batch_hidden_grad.mutable_data(hidden_dims, context.GetPlace()); batch_gate_grad.mutable_data(gate_dims, context.GetPlace()); batch_reset_hidden_prev_grad.mutable_data(hidden_dims, context.GetPlace()); - math::SetConstant zero; - auto& dev_ctx = context.device_context(); + math::SetConstant zero; + auto& dev_ctx = context.template device_context(); zero(dev_ctx, &batch_hidden_grad, static_cast(0.0)); zero(dev_ctx, &batch_gate_grad, static_cast(0.0)); zero(dev_ctx, &batch_reset_hidden_prev_grad, static_cast(0.0)); @@ -156,85 +161,89 @@ class GRUGradKernel : public framework::OpKernel { Tensor ordered_h0, ordered_h0_grad; const size_t* order = batch_gate->lod()[2].data(); if (h0) { - ReorderInitState(context.device_context(), *h0, order, - &ordered_h0, true); + ReorderInitState(dev_ctx, *h0, order, &ordered_h0, + true); } if (h0_grad) { ordered_h0_grad.mutable_data(h0_grad->dims(), context.GetPlace()); - zero(context.device_context(), &ordered_h0_grad, static_cast(0.0)); + zero(context.template device_context(), &ordered_h0_grad, + static_cast(0.0)); } 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); - math::hl_gru_value gru_value; - gru_value.gateWeight = const_cast(weight_data); - gru_value.stateWeight = + math::GRUMetaValue gru_value; + gru_value.gate_weight = const_cast(weight_data); + gru_value.state_weight = const_cast(weight_data + 2 * frame_size * frame_size); - math::hl_gru_grad gru_grad; + math::GRUMetaGrad gru_grad; if (weight_grad) { - gru_grad.gateWeightGrad = + gru_grad.gate_weight_grad = weight_grad->mutable_data(context.GetPlace()); zero(dev_ctx, weight_grad, static_cast(0.0)); - gru_grad.stateWeightGrad = + gru_grad.state_weight_grad = weight_grad->data() + 2 * frame_size * frame_size; } else { - gru_grad.gateWeightGrad = nullptr; - gru_grad.stateWeightGrad = nullptr; + gru_grad.gate_weight_grad = nullptr; + gru_grad.state_weight_grad = nullptr; } auto batch_starts = batch_hidden_grad.lod()[0]; size_t num_batch = batch_starts.size() - 1; + auto active_node = math::detail::GetActivationType( + context.Attr("activation")); + auto active_gate = math::detail::GetActivationType( + context.Attr("gate_activation")); 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]); int cur_batch_size = bend - bstart; Tensor gate_t = batch_gate->Slice(bstart, bend); - gru_value.gateValue = gate_t.data(); + gru_value.gate_value = gate_t.data(); Tensor reset_hidden_prev_t = batch_reset_hidden_prev->Slice(bstart, bend); - gru_value.resetOutputValue = reset_hidden_prev_t.data(); + gru_value.reset_output_value = reset_hidden_prev_t.data(); Tensor hidden_grad_t = batch_hidden_grad.Slice(bstart, bend); - gru_grad.outputGrad = hidden_grad_t.data(); + gru_grad.output_grad = hidden_grad_t.data(); Tensor gate_grad_t = batch_gate_grad.Slice(bstart, bend); - gru_grad.gateGrad = gate_grad_t.data(); + gru_grad.gate_grad = gate_grad_t.data(); Tensor reset_hidden_prev_grad_t = batch_reset_hidden_prev_grad.Slice(bstart, bend); - gru_grad.resetOutputGrad = reset_hidden_prev_grad_t.data(); + gru_grad.reset_output_grad = reset_hidden_prev_grad_t.data(); if (n == 0) { - gru_value.prevOutValue = h0 ? ordered_h0.data() : nullptr; - gru_grad.prevOutGrad = + gru_value.prev_out_value = h0 ? ordered_h0.data() : nullptr; + gru_grad.prev_out_grad = h0 && h0_grad ? ordered_h0_grad.data() : nullptr; } else { int bstart_pre = static_cast(batch_starts[n - 1]); Tensor hidden_prev_t = batch_hidden->Slice(bstart_pre, bstart); - gru_value.prevOutValue = hidden_prev_t.data(); + gru_value.prev_out_value = hidden_prev_t.data(); Tensor hidden_prev_grad_t = batch_hidden_grad.Slice(bstart_pre, bstart); - gru_grad.prevOutGrad = hidden_prev_grad_t.data(); + gru_grad.prev_out_grad = hidden_prev_grad_t.data(); } - math::GRUUnitGradFunctor::compute( - dev_ctx, gru_value, gru_grad, frame_size, cur_batch_size, - math::ActiveType(context.Attr("activation")), - math::ActiveType(context.Attr("gate_activation"))); + math::GRUUnitGradFunctor::compute( + dev_ctx, gru_value, gru_grad, frame_size, cur_batch_size, active_node, + active_gate); } if (input_grad) { input_grad->mutable_data(context.GetPlace()); - math::Batch2LoDTensorFunctor to_seq; + math::Batch2LoDTensorFunctor to_seq; batch_gate_grad.set_lod(batch_gate->lod()); to_seq(dev_ctx, batch_gate_grad, *input_grad); } if (bias_grad) { bias_grad->mutable_data(context.GetPlace()); - math::ColwiseSum col_sum; + math::ColwiseSum col_sum; col_sum(dev_ctx, batch_gate_grad, bias_grad); } if (h0 && h0_grad) { - ReorderInitState(context.device_context(), ordered_h0_grad, - order, h0_grad, false); + ReorderInitState(dev_ctx, ordered_h0_grad, order, + h0_grad, false); } } diff --git a/paddle/operators/gru_unit_op.cc b/paddle/operators/gru_unit_op.cc index 89c027ff1eea93012dc5ab22b081786efc328e96..c354293be7720abd7d96b1c4311b32049a16730c 100644 --- a/paddle/operators/gru_unit_op.cc +++ b/paddle/operators/gru_unit_op.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/operators/gru_unit_op.h" @@ -71,8 +71,7 @@ class GRUUnitOp : public framework::OperatorWithKernel { class GRUUnitOpMaker : public framework::OpProtoAndCheckerMaker { public: - GRUUnitOpMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker) + GRUUnitOpMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("Input", "(Tensor) Matrix with shape [batch_size, frame_size * 3] for the " @@ -114,18 +113,19 @@ class GRUUnitOpMaker : public framework::OpProtoAndCheckerMaker { .SetDefault(sigmoid) .InEnum({identity, sigmoid, tanh, relu}); AddComment(R"DOC( -GRUUnit Operator. - -This operator implements partial calculations of the GRU unit as follows: +GRUUnit Operator implements partial calculations of the GRU unit as following: $$ -update \ gate: u_t = actGate(xu_t + W_u * hidden_{prev} + bias_u) \\ -reset \ gate: r_t = actGate(xr_t + W_r * hidden_{prev} + bias_r) \\ -output \ candidate: {h}_t = actNode({xc}_t + W_c * dot(r_t, hidden_{prev}) + bias_c) \\ -output: h_t = dot((1-u_t), {h}_t) + dot(u_t, hidden_{prev}) +update \ gate: u_t = actGate(xu_t + W_u * h_{t-1} + b_u) \\ +reset \ gate: r_t = actGate(xr_t + W_r * h_{t-1} + b_r) \\ +output \ candidate: {h}_t = actNode(xc_t + W_c * dot(r_t, h_{t-1}) + b_c) \\ +output: h_t = dot((1 - u_t), h_{t-1}) + dot(u_t, {h}_t) $$ -The rest of GRU unit can be completed by using FCOp's output as the input of GRUUnitOp. +which is same as one time step of GRU Operator. + +@note To implement the complete GRU unit, fully-connected operator must be +used before to feed xu, xr and xc as the Input of GRUUnit operator. )DOC"); } @@ -150,12 +150,6 @@ class GRUUnitGradOp : public framework::OperatorWithKernel { "ResetHiddenPrev"); PADDLE_ENFORCE(ctx->HasInput("Hidden"), "Input(%s) of GRUUnitGradOp should not be null.", "Hidden"); - PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Gate")), - "Input(%s@GRAD) of GRUUnitGradOp should not be null.", - "Gate"); - PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("ResetHiddenPrev")), - "Input(%s@GRAD) of GRUUnitGradOp should not be null.", - "ResetHiddenPrev"); PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Hidden")), "Input(%s@GRAD) of GRUUnitGradOp should not be null.", "Hidden"); @@ -206,9 +200,10 @@ class GRUUnitGradOp : public framework::OperatorWithKernel { namespace ops = paddle::operators; REGISTER_OP(gru_unit, ops::GRUUnitOp, ops::GRUUnitOpMaker, gru_unit_grad, ops::GRUUnitGradOp); -REGISTER_OP_CPU_KERNEL(gru_unit, - ops::GRUUnitKernel, - ops::GRUUnitKernel); REGISTER_OP_CPU_KERNEL( - gru_unit_grad, ops::GRUUnitGradKernel, - ops::GRUUnitGradKernel); + gru_unit, ops::GRUUnitKernel, + ops::GRUUnitKernel); +REGISTER_OP_CPU_KERNEL( + gru_unit_grad, + ops::GRUUnitGradKernel, + ops::GRUUnitGradKernel); diff --git a/paddle/operators/gru_unit_op.cu b/paddle/operators/gru_unit_op.cu index 821c8c6421771bd99474b0b2f8aa2acf04697779..95c8c23dadadf0e053012c86d44346ee31565cfc 100644 --- a/paddle/operators/gru_unit_op.cu +++ b/paddle/operators/gru_unit_op.cu @@ -1,24 +1,25 @@ /* 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 +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 + 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. */ +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. */ #define EIGEN_USE_GPU #include "paddle/operators/gru_unit_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(gru_unit, - ops::GRUUnitKernel, - ops::GRUUnitKernel); -REGISTER_OP_GPU_KERNEL( - gru_unit_grad, ops::GRUUnitGradKernel, - ops::GRUUnitGradKernel); +REGISTER_OP_CUDA_KERNEL( + gru_unit, ops::GRUUnitKernel, + ops::GRUUnitKernel); +REGISTER_OP_CUDA_KERNEL( + gru_unit_grad, + ops::GRUUnitGradKernel, + ops::GRUUnitGradKernel); diff --git a/paddle/operators/gru_unit_op.h b/paddle/operators/gru_unit_op.h index c53e7d9827e0395e6ce613302e732b2797f83cdd..a77be46718b766d9a0a8b8fb4cf2316b44687db8 100644 --- a/paddle/operators/gru_unit_op.h +++ b/paddle/operators/gru_unit_op.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 @@ -28,9 +28,13 @@ template using EigenMatrix = framework::EigenMatrix; +template +using EigenVector = framework::EigenVector; + enum GRUActivationType { identity = 0, sigmoid = 1, tanh = 2, relu = 3 }; -template +template class GRUUnitKernel : public framework::OpKernel { public: template @@ -67,7 +71,8 @@ class GRUUnitKernel : public framework::OpKernel { auto g = EigenMatrix::From(*gate); auto r_h_p = EigenMatrix::From(*reset_hidden_prev); auto h = EigenMatrix::From(*hidden); - auto place = context.GetEigenDevice(); + auto& place = + *context.template device_context().eigen_device(); // calculate unactivated gate outputs if (bias) { @@ -82,10 +87,10 @@ 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.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); + 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); // calculate activited gate Eigen::array extents({{batch_size, frame_size}}); @@ -98,11 +103,11 @@ 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.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); + 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); Eigen::array c_offsets({{0, frame_size * 2}}); ActCompute(context.Attr("activation"), place, @@ -110,11 +115,11 @@ class GRUUnitKernel : public framework::OpKernel { auto c = g.slice(c_offsets, extents); // output candidate // calculate final output - h.device(place) = u * (h_p - c) + c; + h.device(place) = u * (c - h_p) + h_p; } }; -template +template class GRUUnitGradKernel : public framework::OpKernel { public: template @@ -146,34 +151,27 @@ class GRUUnitGradKernel : public framework::OpKernel { auto* weight_grad = context.Output(framework::GradVarName("Weight")); auto* bias_grad = context.Output(framework::GradVarName("Bias")); - input_grad->mutable_data(context.GetPlace()); - hidden_prev_grad->mutable_data(context.GetPlace()); - weight_grad->mutable_data(context.GetPlace()); Tensor gate_grad; - gate_grad.mutable_data(input->dims(), context.GetPlace()); Tensor reset_hidden_prev_grad; - reset_hidden_prev_grad.mutable_data(reset_hidden_prev->dims(), - context.GetPlace()); - - int batch_size = input->dims()[0]; - int frame_size = hidden_prev->dims()[1]; const T* hidden_prev_data = hidden_prev->data(); - T* hidden_prev_grad_data = hidden_prev_grad->data(); const T* weight_data = weight->data(); - T* weight_grad_data = weight_grad->data(); - T* gate_grad_data = gate_grad.data(); + T* gate_grad_data = + gate_grad.mutable_data(input->dims(), context.GetPlace()); const T* reset_hidden_prev_data = reset_hidden_prev->data(); - T* reset_hidden_prev_grad_data = reset_hidden_prev_grad.data(); + T* reset_hidden_prev_grad_data = reset_hidden_prev_grad.mutable_data( + reset_hidden_prev->dims(), context.GetPlace()); auto h_p = EigenMatrix::From(*hidden_prev); auto g = EigenMatrix::From(*gate); auto d_h = EigenMatrix::From(*hidden_grad); - auto d_x = EigenMatrix::From(*input_grad); - auto d_h_p = EigenMatrix::From(*hidden_prev_grad); auto d_g = EigenMatrix::From(gate_grad); auto d_r_h_p = EigenMatrix::From(reset_hidden_prev_grad); - auto place = context.GetEigenDevice(); + auto& place = + *context.template device_context().eigen_device(); + + int batch_size = input->dims()[0]; + int frame_size = hidden_prev->dims()[1]; Eigen::array extents({{batch_size, frame_size}}); Eigen::array u_offsets({{0, 0}}); @@ -185,42 +183,58 @@ class GRUUnitGradKernel : public framework::OpKernel { // backward for unactivated update gate ActGradCompute(context.Attr("gate_activation"), place, u, u, - d_g.slice(u_offsets, extents), d_h * (h_p - c)); + d_g.slice(u_offsets, extents), d_h * (c - h_p)); // backward for unactivated output candidate ActGradCompute(context.Attr("activation"), place, c, c, - d_g.slice(c_offsets, extents), d_h * (u.constant(T(1)) - u)); + d_g.slice(c_offsets, extents), d_h * u); // backward for reset_hidden_prev - math::gemm(context.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); - // backward for state_weight - math::gemm( - context.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); + 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); // 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); - // backward for update_gate_weight and reset_gate_weight - math::gemm(context.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); + // backward for weight + 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); + + // 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); + } // backward for hidden_prev - d_h_p.device(place) = d_r_h_p * r + d_h * u; - math::gemm(context.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); + if (hidden_prev_grad) { + T* hidden_prev_grad_data = + 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); + } // backward for input - d_x.device(place) = d_g; + if (input_grad) { + input_grad->mutable_data(context.GetPlace()); + auto d_x = EigenMatrix::From(*input_grad); + d_x.device(place) = d_g; + } // backward for bias if (bias_grad) { bias_grad->mutable_data(context.GetPlace()); - auto d_b = EigenMatrix::From(*bias_grad); + auto d_b = EigenVector::Flatten(*bias_grad); d_b.device(place) = d_g.sum(Eigen::array({{0}})); } } diff --git a/paddle/operators/hinge_loss_op.cc b/paddle/operators/hinge_loss_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..19d2e9dc56fe11f9dfb13e8cb271a23e128bf91b --- /dev/null +++ b/paddle/operators/hinge_loss_op.cc @@ -0,0 +1,113 @@ +/* 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. */ + +#include "paddle/operators/hinge_loss_op.h" + +namespace paddle { +namespace operators { + +class HingeLossOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Logits"), + "Input(Logits) must be initialized."); + PADDLE_ENFORCE(ctx->HasInput("Labels"), + "Input(Labels) must be initialized."); + + auto pred_dims = ctx->GetInputDim("Logits"); + auto label_dims = ctx->GetInputDim("Labels"); + + PADDLE_ENFORCE_EQ(pred_dims, label_dims); + PADDLE_ENFORCE_EQ(pred_dims.size(), 2, + "The rank of Input(Logits) must be 2 and the shape is " + "[batch_size, 1]."); + PADDLE_ENFORCE_EQ(pred_dims[1], 1, + "Each row of Input(Logits) contains a real value, " + "so the 2nd dimension of Input(Logits) must be 1."); + + ctx->SetOutputDim("Loss", {pred_dims[0], 1}); + ctx->ShareLoD("Logits", "Loss"); + } +}; + +template +class HingeLossOpMaker : public framework::OpProtoAndCheckerMaker { + public: + HingeLossOpMaker(OpProto* proto, OpAttrChecker* op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("Logits", + "The input value (Logits) of Hinge loss op." + "Logits is a 2-D tensor with shape [batch_size, 1]."); + AddInput("Labels", + "The target value (Labels) of Hinge loss op." + "Labels is a 2-D tensor with shape [batch_size, 1]."); + AddOutput("Loss", + "The output tensor with shape [batch_size, 1] " + "which represents the hinge loss."); + AddComment(R"DOC( +HingeLoss Operator. + +Let x be a logit (prediction) and y be the actual label. The logit can +take any values from (-inf, inf), but the labels should be either -1 or 1. +Then, the hinge loss is computed as follows: + +$$ +L_(x, y) = max(1 - y.x, 0) +$$ + +Note that the labels passed as input will have values as either 0 or 1. + +)DOC"); + } +}; + +class HingeLossGradOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Logits"), + "Input(Logits) should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Labels"), + "Input(Labels) should not be null."); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Loss")), + "Input(Loss@GRAD) should not be null."); + PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("Logits")), + "Input(Logits@GRAD) should not be null."); + + auto pred_dims = ctx->GetInputDim("Logits"); + auto lab_dims = ctx->GetInputDim("Labels"); + auto loss_grad_dims = ctx->GetInputDim(framework::GradVarName("Loss")); + + PADDLE_ENFORCE_EQ(loss_grad_dims, pred_dims); + + auto pred_grad_name = framework::GradVarName("Logits"); + ctx->SetOutputDim(pred_grad_name, pred_dims); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP(hinge_loss, ops::HingeLossOp, ops::HingeLossOpMaker, + hinge_loss_grad, ops::HingeLossGradOp); +REGISTER_OP_CPU_KERNEL( + hinge_loss, + ops::HingeLossKernel); +REGISTER_OP_CPU_KERNEL( + hinge_loss_grad, + ops::HingeLossGradKernel); diff --git a/paddle/operators/hinge_loss_op.cu b/paddle/operators/hinge_loss_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..b9cfbc50c49c6cc902cb3667200c12c74fb5d13d --- /dev/null +++ b/paddle/operators/hinge_loss_op.cu @@ -0,0 +1,24 @@ +/* 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. */ + +#define EIGEN_USE_GPU +#include "paddle/operators/hinge_loss_op.h" + +namespace ops = paddle::operators; +REGISTER_OP_CUDA_KERNEL( + hinge_loss, + ops::HingeLossKernel); +REGISTER_OP_CUDA_KERNEL( + hinge_loss_grad, + ops::HingeLossGradKernel); diff --git a/paddle/operators/hinge_loss_op.h b/paddle/operators/hinge_loss_op.h new file mode 100644 index 0000000000000000000000000000000000000000..91369cfb8a5d4f40be9e6249b50079ba2b550003 --- /dev/null +++ b/paddle/operators/hinge_loss_op.h @@ -0,0 +1,71 @@ +/* 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 +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" + +namespace paddle { +namespace operators { + +template +class HingeLossKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto* pred = context.Input("Logits"); + auto* label = context.Input("Labels"); + auto* loss = context.Output("Loss"); + auto& place = + *context.template device_context().eigen_device(); + + auto x = framework::EigenVector::Flatten(*pred); + auto y = framework::EigenVector::Flatten(*label); + loss->mutable_data(context.GetPlace()); + auto l = framework::EigenVector::Flatten(*loss); + l.device(place) = + (static_cast(1) - x * (static_cast(2) * y - static_cast(1))) + .cwiseMax(static_cast(0)); + } +}; + +template +class HingeLossGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto* pred = context.Input("Logits"); + auto* label = context.Input("Labels"); + auto* dloss = + context.Input(framework::GradVarName("Loss")); + auto* dpred = + context.Output(framework::GradVarName("Logits")); + auto& place = + *context.template device_context().eigen_device(); + + auto x = framework::EigenVector::Flatten(*pred); + auto y = framework::EigenVector::Flatten(*label); + auto dl = framework::EigenVector::Flatten(*dloss); + + if (dpred) { + dpred->mutable_data(context.GetPlace()); + auto dx = framework::EigenVector::Flatten(*dpred); + auto alt_labels = static_cast(2) * y - static_cast(1); + dx.device(place) = + dl * ((x * alt_labels) < static_cast(1)).template cast() * + (-alt_labels); + } + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/huber_loss_op.cc b/paddle/operators/huber_loss_op.cc index 3435e74b0afb470fcbd1c0f4e06ad363352cac00..5c92f2c7b2d2f701bcc487716db41a0cce91002f 100644 --- a/paddle/operators/huber_loss_op.cc +++ b/paddle/operators/huber_loss_op.cc @@ -45,8 +45,7 @@ class HuberLossOp : public framework::OperatorWithKernel { template class HuberLossOpMaker : public framework::OpProtoAndCheckerMaker { public: - HuberLossOpMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker) + HuberLossOpMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "The input value of huber loss op." @@ -70,11 +69,18 @@ input value and Y as the target value. Huber loss can evaluate the fitness of X to Y. Different from MSE loss, Huber loss is more robust for outliers. The shape of X and Y are [batch_size, 1]. The equation is: -L_{\delta}(y, f(x)) = +$$ +Out_{\delta}(X, Y)_i = \begin{cases} -0.5 * (y - f(x))^2, \quad |y - f(x)| \leq \delta \\ -\delta * (|y - f(x)| - 0.5 * \delta), \quad otherwise +0.5 * (Y_i - X_i)^2, +\quad |Y_i - X_i| \leq \delta \\ +\delta * (|Y_i - X_i| - 0.5 * \delta), +\quad otherwise \end{cases} +$$ + +In the above equation, $Out_\delta(X, Y)_i$, $X_i$ and $Y_i$ represent the ith +element of Out, X and Y. )DOC"); } @@ -117,8 +123,9 @@ class HuberLossGradOp : public framework::OperatorWithKernel { namespace ops = paddle::operators; REGISTER_OP(huber_loss, ops::HuberLossOp, ops::HuberLossOpMaker, huber_loss_grad, ops::HuberLossGradOp); -REGISTER_OP_CPU_KERNEL(huber_loss, - ops::HuberLossKernel); +REGISTER_OP_CPU_KERNEL( + huber_loss, + ops::HuberLossKernel); REGISTER_OP_CPU_KERNEL( huber_loss_grad, - ops::HuberLossGradKernel); + ops::HuberLossGradKernel); diff --git a/paddle/operators/huber_loss_op.cu b/paddle/operators/huber_loss_op.cu index 317321dc6c495f6e9a8808d841c71bfa26b754d0..ccc83a16ba271f0aa879c5c075dce1932dd40494 100644 --- a/paddle/operators/huber_loss_op.cu +++ b/paddle/operators/huber_loss_op.cu @@ -1,23 +1,24 @@ /* 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 +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 + 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. */ +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. */ #define EIGEN_USE_GPU #include "paddle/operators/huber_loss_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(huber_loss, - ops::HuberLossKernel); -REGISTER_OP_GPU_KERNEL( +REGISTER_OP_CUDA_KERNEL( + huber_loss, + ops::HuberLossKernel); +REGISTER_OP_CUDA_KERNEL( huber_loss_grad, - ops::HuberLossGradKernel); + ops::HuberLossGradKernel); diff --git a/paddle/operators/huber_loss_op.h b/paddle/operators/huber_loss_op.h index 4e7bc5543226e19fe0d6190171cdd9c2b3d2d985..4dd20e8b080ab8bd2e61830241d64ee8546a80ec 100644 --- a/paddle/operators/huber_loss_op.h +++ b/paddle/operators/huber_loss_op.h @@ -41,7 +41,7 @@ struct HuberLossForward { T delta; }; -template +template class HuberLossKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { @@ -50,7 +50,8 @@ class HuberLossKernel : public framework::OpKernel { auto* out0 = context.Output("Residual"); auto* out1 = context.Output("Out"); auto delta = static_cast(context.Attr("delta")); - auto place = context.GetEigenDevice(); + auto& place = + *context.template device_context().eigen_device(); auto x = EigenVector::Flatten(*in0); auto y = EigenVector::Flatten(*in1); @@ -85,7 +86,7 @@ struct HuberLossBackward { T delta; }; -template +template class HuberLossGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { @@ -94,7 +95,8 @@ class HuberLossGradKernel : public framework::OpKernel { auto* out0 = context.Output(framework::GradVarName("X")); auto* out1 = context.Output(framework::GradVarName("Y")); auto delta = static_cast(context.op().Attr("delta")); - auto place = context.GetEigenDevice(); + auto& place = + *context.template device_context().eigen_device(); auto residual = EigenVector::Flatten(*in0); auto out_grad = EigenVector::Flatten(*in1); diff --git a/paddle/operators/increment_op.cc b/paddle/operators/increment_op.cc index 35efb12932f1d61fdb511b4ee2cdab3891507c61..e0b80cc4e74429dee1b9a25e41b116970ad4de2a 100644 --- a/paddle/operators/increment_op.cc +++ b/paddle/operators/increment_op.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/framework/op_registry.h" @@ -52,7 +52,7 @@ class IncrementOp : public framework::OperatorBase { : OperatorBase(type, inputs, outputs, attrs) {} void Run(const framework::Scope &scope, - const platform::DeviceContext &dev_ctx) const override { + const platform::Place &place) const override { auto &x = scope.FindVar(Input("X"))->Get(); auto &out = *scope.FindVar(Output("Out"))->GetMutable(); @@ -61,6 +61,8 @@ class IncrementOp : public framework::OperatorBase { out.Resize(x.dims()); out.mutable_data(x.place(), x.type()); float value = Attr("step"); + VLOG(10) << Output("Out") << " increase " << Input("X") << " with " + << value; framework::VisitDataType(framework::ToDataType(out.type()), IncrementFunctor(x, &out, value)); } @@ -68,8 +70,7 @@ class IncrementOp : public framework::OperatorBase { class IncrementOpMaker : public framework::OpProtoAndCheckerMaker { public: - IncrementOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + IncrementOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "(Tensor) The input tensor of increment operator"); AddOutput("Out", "(Tensor) The output tensor of increment operator."); @@ -92,13 +93,13 @@ class IncrementGradOpMaker : public framework::SingleGradOpDescMaker { public: using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; - std::unique_ptr Apply() const override { - auto *grad_op = new framework::OpDescBind(); + std::unique_ptr Apply() const override { + auto *grad_op = new framework::OpDesc(); grad_op->SetType("increment"); grad_op->SetInput("X", Output("Out")); grad_op->SetOutput("Out", Input("X")); grad_op->SetAttr("step", -boost::get(GetAttr("step"))); - return std::unique_ptr(grad_op); + return std::unique_ptr(grad_op); } }; diff --git a/paddle/operators/is_empty_op.cc b/paddle/operators/is_empty_op.cc index 54fecf44e881b5c283c81580fd161da9808d253e..492ae48845aa5aa123989e62d07f5ae899af6193 100644 --- a/paddle/operators/is_empty_op.cc +++ b/paddle/operators/is_empty_op.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/framework/op_registry.h" #include "paddle/framework/operator.h" @@ -29,7 +29,7 @@ class IsEmptyOp : public framework::OperatorBase { : OperatorBase(type, inputs, outputs, attrs) {} void Run(const framework::Scope &scope, - const platform::DeviceContext &dev_ctx) const override { + const platform::Place &place) const override { // get input auto *var = scope.FindVar(Input(kInput)); PADDLE_ENFORCE_NOT_NULL(var); @@ -47,8 +47,7 @@ class IsEmptyOp : public framework::OperatorBase { class IsEmptyOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - IsEmptyOpProtoMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + 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."); diff --git a/paddle/operators/l1_norm_op.cc b/paddle/operators/l1_norm_op.cc index 02ebf022968e95d0b20598d3c935fb51177c8841..1a5d6e19263325821dd220d8a31c0e34600b8220 100644 --- a/paddle/operators/l1_norm_op.cc +++ b/paddle/operators/l1_norm_op.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/operators/l1_norm_op.h" @@ -48,7 +48,7 @@ class L1NormGradOp : public framework::OperatorWithKernel { class L1NormOpMaker : public framework::OpProtoAndCheckerMaker { public: - L1NormOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker) + L1NormOpMaker(OpProto* proto, OpAttrChecker* op_checker) : framework::OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "(Tensor) The input of l1_norm op."); AddOutput("Out", "(Scalar) The output of l1_norm op."); @@ -69,7 +69,8 @@ $$Out = \sum{|X|}$$ namespace ops = paddle::operators; REGISTER_OP(l1_norm, ops::L1NormOp, ops::L1NormOpMaker, l1_norm_grad, ops::L1NormGradOp); -REGISTER_OP_CPU_KERNEL(l1_norm, - ops::L1NormKernel); REGISTER_OP_CPU_KERNEL( - l1_norm_grad, ops::L1NormGradKernel); + l1_norm, ops::L1NormKernel); +REGISTER_OP_CPU_KERNEL( + l1_norm_grad, + ops::L1NormGradKernel); diff --git a/paddle/operators/l1_norm_op.cu b/paddle/operators/l1_norm_op.cu index 1c206e04ccbb5f4c2cb9d45aef7bac17c62d55c5..7ecc774670a8480da46ac688f3635e04b1ab7c1f 100644 --- a/paddle/operators/l1_norm_op.cu +++ b/paddle/operators/l1_norm_op.cu @@ -1,22 +1,23 @@ /* 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 +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 + 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. */ +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. */ #define EIGEN_USE_GPU #include "paddle/operators/l1_norm_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(l1_norm, - ops::L1NormKernel); -REGISTER_OP_GPU_KERNEL( - l1_norm_grad, ops::L1NormGradKernel); +REGISTER_OP_CUDA_KERNEL( + l1_norm, ops::L1NormKernel); +REGISTER_OP_CUDA_KERNEL( + l1_norm_grad, + ops::L1NormGradKernel); diff --git a/paddle/operators/l1_norm_op.h b/paddle/operators/l1_norm_op.h index 3c60dc3dc7415f34ed9d238e6f41b197ec404883..086d42705dceaf0cbd35ce8e5115156a76a0b6e8 100644 --- a/paddle/operators/l1_norm_op.h +++ b/paddle/operators/l1_norm_op.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/framework/eigen.h" @@ -20,7 +20,7 @@ namespace paddle { namespace operators { // Out = sum(abs(X)) -template +template class L1NormKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext &context) const override { @@ -30,14 +30,15 @@ class L1NormKernel : public framework::OpKernel { auto x = framework::EigenVector::Flatten(*X); auto out = framework::EigenScalar::From(*Out); - auto place = context.GetEigenDevice(); + auto &place = + *context.template device_context().eigen_device(); out.device(place) = x.abs().sum(); } }; // dX = dout * sign(X) -template +template class L1NormGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext &context) const override { @@ -52,7 +53,8 @@ class L1NormGradKernel : public framework::OpKernel { auto x_eigen = framework::EigenVector::Flatten(*x); auto d_out_eigen = framework::EigenVector::Flatten(*d_out); auto dx_eigen = framework::EigenVector::Flatten(*dx); - auto place = context.GetEigenDevice(); + auto &place = + *context.template device_context().eigen_device(); Eigen::DSizes x_dsize(x->numel()); dx_eigen.device(place) = d_out_eigen.broadcast(x_dsize) * x_eigen.sign(); diff --git a/paddle/operators/linear_chain_crf_op.cc b/paddle/operators/linear_chain_crf_op.cc index 066bdf67aa037e9c25cfdfaff7ec8771eb59cde8..666207ea07628ca5f2a8313fa3f5febda140a294 100644 --- a/paddle/operators/linear_chain_crf_op.cc +++ b/paddle/operators/linear_chain_crf_op.cc @@ -19,8 +19,7 @@ namespace operators { class LinearChainCRFOpMaker : public framework::OpProtoAndCheckerMaker { public: - LinearChainCRFOpMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker) + LinearChainCRFOpMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("Emission", "(LoDTensor, default LoDTensor) " @@ -32,19 +31,19 @@ class LinearChainCRFOpMaker : public framework::OpProtoAndCheckerMaker { "[(D + 2) x D]. The learnable parameter for the linear_chain_crf " "operator. See more details in the operator's comments."); AddInput("Label", - "(LoDTensor, default LoDTensor) A LoDTensor with shape " + "(LoDTensor, default LoDTensor) A LoDTensor with shape " "[N x 1], where N is the total element number in a mini-batch. " "The ground truth."); AddOutput( "Alpha", "(Tensor, default Tensor) A 2-D Tensor with shape [N x D]. " - "The forward vectors for the entire batch. Denote it as \f$\alpha\f$. " - "\f$\alpha$\f is a memo table used to calculate the normalization " - "factor in CRF. \f$\alpha[k, v]$\f stores the unnormalized " + "The forward vectors for the entire batch. Denote it as $\alpha$. " + "$\alpha$ is a memo table used to calculate the normalization " + "factor in CRF. $\alpha[k, v]$ stores the unnormalized " "probabilites of all possible unfinished sequences of tags that end at " - "position \f$k$\f with tag \f$v$\f. For each \f$k$\f, " - "\f$\alpha[k, v]$\f is a vector of length \f$D$\f with a component for " - "each tag value \f$v$\f. This vector is called a forward vecotr and " + "position $k$ with tag $v$. For each $k$, " + "$\alpha[k, v]$ is a vector of length $D$ with a component for " + "each tag value $v$. This vector is called a forward vecotr and " "will also be used in backward computations.") .AsIntermediate(); AddOutput( @@ -73,9 +72,9 @@ LinearChainCRF Operator. Conditional Random Field defines an undirected probabilistic graph with nodes denoting random variables and edges denoting dependencies between these -variables. CRF learns the conditional probability \f$P(Y|X)\f$, where -\f$X = (x_1, x_2, ... , x_n)\f$ are structured inputs and -\f$Y = (y_1, y_2, ... , y_n)\f$ are labels for the inputs. +variables. CRF learns the conditional probability $P(Y|X)$, where +$X = (x_1, x_2, ... , x_n)$ are structured inputs and +$Y = (y_1, y_2, ... , y_n)$ are labels for the inputs. Linear chain CRF is a special case of CRF that is useful for sequence labeling task. Sequence labeling tasks do not assume a lot of conditional @@ -88,21 +87,22 @@ CRF. Please refer to http://www.cs.columbia.edu/~mcollins/fb.pdf and http://cseweb.ucsd.edu/~elkan/250Bwinter2012/loglinearCRFs.pdf for details. Equation: -1. Denote Input(Emission) to this operator as \f$x\f$ here. +1. Denote Input(Emission) to this operator as $x$ here. 2. The first D values of Input(Transition) to this operator are for starting -weights, denoted as \f$a\f$ here. +weights, denoted as $a$ here. 3. The next D values of Input(Transition) of this operator are for ending -weights, denoted as \f$b\f$ here. +weights, denoted as $b$ here. 4. The remaning values of Input(Transition) are for transition weights, -denoted as \f$w\f$ here. -5. Denote Input(Label) as \f$s\f$ here. - -The probability of a sequence \f$s\f$ of length \f$L\f$ is defined as: -\f$P(s) = (1/Z) \exp(a_{s_1} + b_{s_L} - + \sum_{l=1}^L x_{s_l} - + \sum_{l=2}^L w_{s_{l-1},s_l})\f$ -where \f$Z\f$ is a normalization value so that the sum of \f$P(s)\f$ over -all possible sequences is \f$1\f$, and \f$x\f$ is the emission feature weight +denoted as $w$ here. +5. Denote Input(Label) as $s$ here. + +The probability of a sequence $s$ of length $L$ is defined as: +$$P(s) = (1/Z) \exp(a_{s_1} + b_{s_L} + + \sum_{l=1}^L x_{s_l} + + \sum_{l=2}^L w_{s_{l-1},s_l})$$ + +where $Z$ is a normalization value so that the sum of $P(s)$ over +all possible sequences is 1, and $x$ is the emission feature weight to the linear chain CRF. Finally, the linear chain CRF operator outputs the logarithm of the conditional @@ -183,7 +183,7 @@ class LinearChainCRFOp : public framework::OperatorWithKernel { protected: // Explicitly set that the data type of computation kernel of linear_chain_crf // is determined by its input "Emission". - framework::OpKernelType GetKernelType( + framework::OpKernelType GetActualKernelType( const framework::ExecutionContext& ctx) const override { return framework::OpKernelType( framework::ToDataType(ctx.Input("Emission")->type()), @@ -242,7 +242,7 @@ class LinearChainCRFGradOp : public framework::OperatorWithKernel { protected: // Explicitly set that the data type of output of the linear_chain_crf_grad // operator is determined by its input: gradients of LogLikelihood. - framework::OpKernelType GetKernelType( + framework::OpKernelType GetActualKernelType( const framework::ExecutionContext& ctx) const override { return framework::OpKernelType( framework::ToDataType( @@ -260,9 +260,10 @@ REGISTER_OP(linear_chain_crf, ops::LinearChainCRFOp, ops::LinearChainCRFOpMaker, linear_chain_crf_grad, ops::LinearChainCRFGradOp); REGISTER_OP_CPU_KERNEL( linear_chain_crf, - ops::LinearChainCRFOpKernel, - ops::LinearChainCRFOpKernel); + ops::LinearChainCRFOpKernel, + ops::LinearChainCRFOpKernel); REGISTER_OP_CPU_KERNEL( linear_chain_crf_grad, - ops::LinearChainCRFGradOpKernel, - ops::LinearChainCRFGradOpKernel); + ops::LinearChainCRFGradOpKernel, + ops::LinearChainCRFGradOpKernel); diff --git a/paddle/operators/linear_chain_crf_op.cu b/paddle/operators/linear_chain_crf_op.cu index 6fc8995f4c2ce05f89ffb58129695113f89159fa..da612510b4d45d8eefabe7de303e9fd0132c5f77 100644 --- a/paddle/operators/linear_chain_crf_op.cu +++ b/paddle/operators/linear_chain_crf_op.cu @@ -1,10 +1,10 @@ /* 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 +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 + 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, @@ -16,11 +16,12 @@ limitations under the License. */ namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL( +REGISTER_OP_CUDA_KERNEL( linear_chain_crf, - ops::LinearChainCRFOpKernel, - ops::LinearChainCRFOpKernel); -REGISTER_OP_GPU_KERNEL( + ops::LinearChainCRFOpKernel, + ops::LinearChainCRFOpKernel); +REGISTER_OP_CUDA_KERNEL( linear_chain_crf_grad, - ops::LinearChainCRFGradOpKernel, - ops::LinearChainCRFGradOpKernel); + ops::LinearChainCRFGradOpKernel, + ops::LinearChainCRFGradOpKernel); diff --git a/paddle/operators/linear_chain_crf_op.h b/paddle/operators/linear_chain_crf_op.h index ddf73981751798c72cef08f2dd5c87580b45aec3..19c6715ec877dea6dcf0babc7373333a4d9eed0f 100644 --- a/paddle/operators/linear_chain_crf_op.h +++ b/paddle/operators/linear_chain_crf_op.h @@ -50,7 +50,7 @@ template using EigenMatrix = framework::EigenMatrix; -template +template class LinearChainCRFOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { @@ -137,7 +137,8 @@ class LinearChainCRFOpKernel : public framework::OpKernel { framework::make_ddim({static_cast(batch_size), 1}), platform::CPUPlace()); - auto place = ctx.GetEigenDevice(); + auto& place = *ctx.template device_context() + .eigen_device(); auto x = EigenMatrix::From(*emission_weights); auto x_row_max = EigenMatrix::From(emission_row_max); x_row_max.device(place) = @@ -195,7 +196,7 @@ class LinearChainCRFOpKernel : public framework::OpKernel { auto copyLoDTensor = [](const platform::DeviceContext& ctx, const LoDTensor& src, LoDTensor* dst) { dst->mutable_data(src.dims(), platform::CPUPlace()); - dst->CopyFrom(src, platform::CPUPlace(), ctx); + framework::CopyFrom(src, platform::CPUPlace(), ctx, dst); }; copyLoDTensor(ctx, emission_weights_src, emission_weights_dst); @@ -203,8 +204,8 @@ class LinearChainCRFOpKernel : public framework::OpKernel { transition_weights_dst->mutable_data(transition_weights_src.dims(), platform::CPUPlace()); - transition_weights_dst->CopyFrom(transition_weights_src, - platform::CPUPlace(), ctx); + framework::CopyFrom(transition_weights_src, platform::CPUPlace(), ctx, + transition_weights_dst); } void CopyOutputsToGpuMemory(const platform::DeviceContext& ctx, @@ -218,8 +219,8 @@ class LinearChainCRFOpKernel : public framework::OpKernel { // operators runs on GPU device. auto copyTensor = [](const platform::DeviceContext& ctx, const Tensor& src, Tensor* dst) { - dst->mutable_data(platform::GPUPlace()); - dst->CopyFrom(src, platform::GPUPlace(), ctx); + dst->mutable_data(platform::CUDAPlace()); + framework::CopyFrom(src, platform::CUDAPlace(), ctx, dst); }; copyTensor(ctx, emission_exps_src, emission_exps_dst); copyTensor(ctx, transition_exps_src, transition_exps_dst); @@ -271,7 +272,7 @@ class LinearChainCRFOpKernel : public framework::OpKernel { ll -= std::log(sum); // Now ll is equal to -log(Z). - const int* lbl = label.data(); + const int64_t* lbl = label.data(); PADDLE_ENFORCE_LT( static_cast(*std::max_element(lbl, lbl + seq_length)), tag_num, "An invalid tag label that execesses the largest tag number."); @@ -287,7 +288,7 @@ class LinearChainCRFOpKernel : public framework::OpKernel { } }; -template +template class LinearChainCRFGradOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { @@ -359,8 +360,7 @@ class LinearChainCRFGradOpKernel : public framework::OpKernel { emission_grad->mutable_data(platform::CPUPlace()); if (transition_grad) { transition_grad->mutable_data(platform::CPUPlace()); - math::SetConstant()(ctx.device_context(), - transition_grad, 0.); + math::set_constant(ctx.device_context(), transition_grad, 0.); } // Now, all the inputs and outputs should be on the CPU memory. @@ -384,10 +384,10 @@ class LinearChainCRFGradOpKernel : public framework::OpKernel { Tensor one_seq_beta = beta.Slice(start_pos, end_pos); Tensor one_seq_emission_grad = emission_grad->Slice(start_pos, end_pos); - BackwardOneSequence(ctx.device_context(), ll_grad[i], - one_seq_emission_exps, *transition_exps, - one_seq_alpha, one_seq_label, &one_seq_beta, - transition_grad, &one_seq_emission_grad); + BackwardOneSequence( + ctx.template device_context(), ll_grad[i], + one_seq_emission_exps, *transition_exps, one_seq_alpha, one_seq_label, + &one_seq_beta, transition_grad, &one_seq_emission_grad); } if (platform::is_gpu_place(ctx.GetPlace())) { @@ -410,12 +410,12 @@ class LinearChainCRFGradOpKernel : public framework::OpKernel { // Copy the inputs from GPU memory to CPU memory when this operators runs on // GPU device. label_dst->mutable_data(label_src.dims(), platform::CPUPlace()); - label_dst->CopyFrom(label_src, platform::CPUPlace(), ctx); + framework::CopyFrom(label_src, platform::CPUPlace(), ctx, label_dst); auto copyTensor = [](const platform::DeviceContext& ctx, const Tensor& src, Tensor* dst) { dst->mutable_data(src.dims(), platform::CPUPlace()); - dst->CopyFrom(src, platform::CPUPlace(), ctx); + framework::CopyFrom(src, platform::CPUPlace(), ctx, dst); }; copyTensor(ctx, emission_exps_src, emission_exps_dst); copyTensor(ctx, transition_exps_src, transition_exps_dst); @@ -433,23 +433,23 @@ class LinearChainCRFGradOpKernel : public framework::OpKernel { auto copyTensor = [](const platform::DeviceContext& ctx, const Tensor* src, Tensor* dst) { if (src && dst) { - dst->mutable_data(platform::GPUPlace()); - dst->CopyFrom(*src, platform::GPUPlace(), ctx); + dst->mutable_data(platform::CUDAPlace()); + framework::CopyFrom(*src, platform::CUDAPlace(), ctx, dst); } }; copyTensor(ctx, emission_grad_src, emission_grad_dst); copyTensor(ctx, transition_grad_src, transition_grad_dst); } - void BackwardOneSequence(const platform::DeviceContext& ctx, const T ll_grad, - const Tensor& emission_exps, + void BackwardOneSequence(const platform::CPUDeviceContext& ctx, + const T ll_grad, const Tensor& emission_exps, const Tensor& transition_exps, const Tensor& alpha, const Tensor& label, Tensor* beta, Tensor* transition_grad, Tensor* emission_grad) const { const T* w_exps = transition_exps.data(); const T* x_exps = emission_exps.data(); - const int* label_value = label.data(); + const int64_t* label_value = label.data(); T* beta_value = beta->data(); auto x_dims = emission_exps.dims(); @@ -481,7 +481,7 @@ class LinearChainCRFGradOpKernel : public framework::OpKernel { auto alpha_mat = EigenMatrix::From(alpha); auto beta_mat = EigenMatrix::From(*beta); - auto* place = ctx.GetEigenDevice(); + auto* place = ctx.eigen_device(); auto prob = alpha_mat * beta_mat; auto row_sum = prob.sum(Eigen::DSizes(1)) .reshape(Eigen::DSizes(seq_length, 1)) diff --git a/paddle/operators/load_op.cc b/paddle/operators/load_op.cc index b71a33a6b1ce80b545e6d7a4020dafc941dc55d2..08b972a233aab8596a5ce7f74ea903df3b8ef0f2 100644 --- a/paddle/operators/load_op.cc +++ b/paddle/operators/load_op.cc @@ -1,20 +1,20 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +/* 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 +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 + 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. */ +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/framework/op_registry.h" - -#include +#include "paddle/platform/device_context.h" namespace paddle { namespace operators { @@ -26,7 +26,7 @@ class LoadOp : public framework::OperatorBase { const framework::AttributeMap &attrs) : OperatorBase(type, inputs, outputs, attrs) {} void Run(const framework::Scope &scope, - const platform::DeviceContext &dev_ctx) const override { + const platform::Place &place) const override { auto filename = Attr("file_path"); std::ifstream fin(filename); PADDLE_ENFORCE(static_cast(fin), "Cannot open file %s for load op", @@ -38,63 +38,11 @@ class LoadOp : public framework::OperatorBase { out_var_name); auto *tensor = out_var->GetMutable(); + DeserializeFromStream(fin, tensor); - uint32_t version; - fin.read(reinterpret_cast(&version), sizeof(version)); - PADDLE_ENFORCE_EQ(version, 0U, "Only version 0 is supported"); - framework::TensorDesc desc; - { // int32_t size - // proto buffer - int32_t size; - fin.read(reinterpret_cast(&size), sizeof(size)); - std::unique_ptr buf(new char[size]); - fin.read(reinterpret_cast(buf.get()), size); - PADDLE_ENFORCE(desc.ParseFromArray(buf.get(), size), - "Cannot parse tensor desc"); - } - { // read tensor - std::vector dims; - dims.reserve(static_cast(desc.dims().size())); - std::copy(desc.dims().begin(), desc.dims().end(), - std::back_inserter(dims)); - tensor->Resize(framework::make_ddim(dims)); - - void *buf; - platform::Place cpu = platform::CPUPlace(); - switch (desc.data_type()) { - case framework::FP32: - buf = tensor->mutable_data(cpu); - break; - case framework::FP64: - buf = tensor->mutable_data(cpu); - break; - case framework::INT32: - buf = tensor->mutable_data(cpu); - break; - case framework::INT64: - buf = tensor->mutable_data(cpu); - break; - default: - PADDLE_THROW("DataType %d not supported", desc.data_type()); - } - fin.read(static_cast(buf), tensor->memory_size()); - } - { // read lod - uint64_t lod_level; - fin.read(reinterpret_cast(&lod_level), sizeof(lod_level)); - auto &lod = *tensor->mutable_lod(); - lod.resize(lod_level); - for (uint64_t i = 0; i < lod_level; ++i) { - uint64_t size; - fin.read(reinterpret_cast(&size), sizeof(size)); - std::vector tmp(size / sizeof(size_t)); - fin.read(reinterpret_cast(tmp.data()), - static_cast(size)); - lod[i] = tmp; - } - } + platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance(); + auto &dev_ctx = *pool.Get(place); - auto place = dev_ctx.GetPlace(); if (platform::is_gpu_place(place)) { // copy CPU to GPU framework::LoDTensor cpu_tensor; @@ -105,15 +53,14 @@ class LoadOp : public framework::OperatorBase { out_var->Clear(); tensor = out_var->GetMutable(); tensor->set_lod(cpu_tensor.lod()); - tensor->CopyFrom(cpu_tensor, place, dev_ctx); + CopyFrom(cpu_tensor, place, dev_ctx, tensor); } } }; class LoadOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - LoadOpProtoMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + LoadOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddOutput("Out", "(Tensor) The tensor need to be loaded"); AddAttr("file_path", diff --git a/paddle/operators/lod_array_length_op.cc b/paddle/operators/lod_array_length_op.cc index 80445eb575703be3354595672a4c064b30e0f18c..d2c52745cfdf8d0fdb168ef2d90e75a515c31015 100644 --- a/paddle/operators/lod_array_length_op.cc +++ b/paddle/operators/lod_array_length_op.cc @@ -1,16 +1,16 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +/* 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 +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 + 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. */ +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/framework/lod_tensor_array.h" #include "paddle/framework/op_registry.h" @@ -26,7 +26,7 @@ class LoDArrayLengthOp : public framework::OperatorBase { const framework::AttributeMap &attrs) : OperatorBase(type, inputs, outputs, attrs) {} void Run(const framework::Scope &scope, - const platform::DeviceContext &dev_ctx) const override { + const platform::Place &place) const override { auto &x = scope.FindVar(Input("X"))->Get(); auto &out = *scope.FindVar(Output("Out"))->GetMutable(); @@ -38,17 +38,20 @@ class LoDArrayLengthOp : public framework::OperatorBase { class LoDArrayLengthProtoMaker : public framework::OpProtoAndCheckerMaker { public: - LoDArrayLengthProtoMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + LoDArrayLengthProtoMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "(LoDTensorArray) The input tensor array."); AddOutput("Out", "(Tensor) 1x1 CPU Tensor of length, int64_t"); - AddComment(R"DOC(Get the length of lod tensor array + AddComment(R"DOC( +LoDArrayLength Operator. -Out = len(X) +This operator obtains the length of lod tensor array: + +$$Out = len(X)$$ NOTE: The output is a CPU Tensor since the control variable should be only in CPU and the length of LoDTensorArray should be used as control variables. + )DOC"); } }; diff --git a/paddle/operators/lod_rank_table_op.cc b/paddle/operators/lod_rank_table_op.cc index f7d4db1947b83fecf57575e17fafe26795c92bdd..8711dd62c886fdada083d316d6aabc93a050ff82 100644 --- a/paddle/operators/lod_rank_table_op.cc +++ b/paddle/operators/lod_rank_table_op.cc @@ -1,16 +1,16 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +/* 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 +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 + 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. */ +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/framework/lod_rank_table.h" #include "paddle/framework/op_registry.h" namespace paddle { @@ -24,19 +24,19 @@ class LoDRankTableOp : public framework::OperatorBase { const framework::AttributeMap &attrs) : OperatorBase(type, inputs, outputs, attrs) {} void Run(const framework::Scope &scope, - const platform::DeviceContext &dev_ctx) const override { + const platform::Place &dev_place) const override { auto x = scope.FindVar(Input("X"))->Get(); auto *out = scope.FindVar(Output("Out"))->GetMutable(); VLOG(10) << "Level = " << static_cast(Attr("level")); out->Reset(x.lod(), static_cast(Attr("level"))); + VLOG(10) << Input("X") << "'s lod information is " << *out; } }; class LoDRankTableOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - LoDRankTableOpProtoMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + LoDRankTableOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "(LoDTensor) input lod tensor, must contain lod information."); @@ -63,11 +63,11 @@ class LoDRankTableInferShape : public framework::InferShapeBase { class LoDRankTableInferVarType : public framework::VarTypeInference { public: - void operator()(const framework::OpDescBind &op_desc, - framework::BlockDescBind *block) const override { + void operator()(const framework::OpDesc &op_desc, + framework::BlockDesc *block) const override { for (auto &o : op_desc.Output("Out")) { block->FindRecursiveOrCreateVar(o)->SetType( - framework::VarDesc::LOD_RANK_TABLE); + framework::proto::VarDesc::LOD_RANK_TABLE); } } }; diff --git a/paddle/operators/lod_reset_op.cc b/paddle/operators/lod_reset_op.cc index 32831cb1e2cf188a507773ef1e00b22de98d82ab..f3c0badf2a74431b980abd532e51ba3d251524a1 100644 --- a/paddle/operators/lod_reset_op.cc +++ b/paddle/operators/lod_reset_op.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/operators/lod_reset_op.h" @@ -38,7 +38,7 @@ class LoDResetOp : public framework::OperatorWithKernel { } protected: - framework::OpKernelType GetKernelType( + framework::OpKernelType GetActualKernelType( const framework::ExecutionContext &ctx) const override { return framework::OpKernelType( framework::ToDataType(ctx.Input("X")->type()), @@ -48,8 +48,7 @@ class LoDResetOp : public framework::OperatorWithKernel { class LoDResetOpMaker : public framework::OpProtoAndCheckerMaker { public: - LoDResetOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + LoDResetOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "(LoDTensor) The input tensor of lod_reset operator."); AddInput("TargetLoD", @@ -98,7 +97,7 @@ class LoDResetGradOp : public framework::OperatorWithKernel { } protected: - framework::OpKernelType GetKernelType( + framework::OpKernelType GetActualKernelType( const framework::ExecutionContext &ctx) const override { return framework::OpKernelType( framework::ToDataType(ctx.Input("X")->type()), diff --git a/paddle/operators/lod_reset_op.cu b/paddle/operators/lod_reset_op.cu index 5244a17c3aad01909e3b8cf5f4d5abf8a44edc7f..910866ea6330059f8e0b04e036e3b124e920b5c4 100644 --- a/paddle/operators/lod_reset_op.cu +++ b/paddle/operators/lod_reset_op.cu @@ -1,24 +1,25 @@ /* 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 +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 + 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. */ +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/operators/lod_reset_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(lod_reset, - ops::LoDResetKernel, - ops::LoDResetKernel); -REGISTER_OP_GPU_KERNEL( - lod_reset_grad, ops::LoDResetGradKernel, - ops::LoDResetGradKernel); +REGISTER_OP_CUDA_KERNEL( + lod_reset, ops::LoDResetKernel, + ops::LoDResetKernel); +REGISTER_OP_CUDA_KERNEL( + lod_reset_grad, + ops::LoDResetGradKernel, + ops::LoDResetGradKernel); diff --git a/paddle/operators/lod_reset_op.h b/paddle/operators/lod_reset_op.h index 2bb916ccee80c83a02ea429fe95f5fafc86ccfa6..306373fb1fb6f16a0db7f0e836e38fd8c49f7e86 100644 --- a/paddle/operators/lod_reset_op.h +++ b/paddle/operators/lod_reset_op.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 @@ -20,7 +20,7 @@ namespace paddle { namespace operators { -template +template class LoDResetKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const { @@ -33,7 +33,8 @@ class LoDResetKernel : public framework::OpKernel { auto* lod = lod_t->data(); if (platform::is_gpu_place(ctx.GetPlace())) { framework::Tensor lod_cpu; - lod_cpu.CopyFrom(*lod_t, platform::CPUPlace(), ctx.device_context()); + framework::CopyFrom(*lod_t, platform::CPUPlace(), ctx.device_context(), + &lod_cpu); lod = lod_cpu.data(); } level0 = std::vector(lod, lod + lod_t->numel()); @@ -64,7 +65,7 @@ class LoDResetKernel : public framework::OpKernel { } }; -template +template class LoDResetGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const { diff --git a/paddle/operators/lod_tensor_to_array_op.cc b/paddle/operators/lod_tensor_to_array_op.cc index 58af35564d83b9699af4f7783fb6367ff9590682..8d164b4abc54722a95a176dfe8ed341f8c5125d1 100644 --- a/paddle/operators/lod_tensor_to_array_op.cc +++ b/paddle/operators/lod_tensor_to_array_op.cc @@ -1,19 +1,21 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +/* 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 +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 + 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. */ +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/framework/lod_rank_table.h" #include "paddle/framework/lod_tensor_array.h" #include "paddle/framework/op_registry.h" +#include "paddle/operators/detail/safe_ref.h" +#include "paddle/platform/device_context.h" namespace paddle { namespace operators { @@ -31,16 +33,21 @@ class LoDTensorToArrayOp : public framework::OperatorBase { const framework::AttributeMap &attrs) : OperatorBase(type, inputs, outputs, attrs) {} void Run(const framework::Scope &scope, - const platform::DeviceContext &dev_ctx) const override { - auto &x = scope.FindVar(Input("X"))->Get(); - auto &rank_table = - scope.FindVar(Input("RankTable"))->Get(); - auto &out = - *scope.FindVar(Output("Out"))->GetMutable(); - + const platform::Place &place) const override { + auto &x = detail::Ref(scope.FindVar(Input("X")), "Cannot find input %s", + Input("X")) + .Get(); + auto &rank_table = detail::Ref(scope.FindVar(Input("RankTable"))) + .Get(); + auto &out = *detail::Ref(scope.FindVar(Output("Out"))) + .GetMutable(); auto &items = rank_table.items(); auto max_seq_len = items[0].length; auto rank_level = rank_table.level(); + + PADDLE_ENFORCE_LT(rank_level, x.lod().size(), + "Input should be a LOD tensor, and size is at least %d", + rank_level + 1); out.resize(max_seq_len); std::vector> copy_ranges(max_seq_len); @@ -55,16 +62,13 @@ class LoDTensorToArrayOp : public framework::OperatorBase { size_t start_idx = x.lod()[rank_level][item.index] + t; auto lod_and_offset = framework::GetSubLoDAndAbsoluteOffset( x.lod(), start_idx, start_idx + 1, rank_level + 1); - auto &lod_length = lod_and_offset.first; framework::AppendLoD(&lod, lod_length); - size_t start_offset = lod_and_offset.second.first; size_t end_offset = lod_and_offset.second.second; copy_ranges[t].emplace_back(CopyRange{start_offset, end_offset}); } } - for (size_t i = 0; i < max_seq_len; ++i) { auto &ranges = copy_ranges[i]; size_t height = std::accumulate( @@ -81,11 +85,16 @@ class LoDTensorToArrayOp : public framework::OperatorBase { continue; } // out[i][offset: offset+len] = x[each_range.begin: each_range.end] - out[i] - .Slice(static_cast(offset), static_cast(offset + len)) - .CopyFrom(x.Slice(static_cast(each_range.begin), - static_cast(each_range.end)), - x.place(), dev_ctx); + auto slice = out[i].Slice(static_cast(offset), + static_cast(offset + len)); + + platform::DeviceContextPool &pool = + platform::DeviceContextPool::Instance(); + auto &dev_ctx = *pool.Get(place); + + framework::CopyFrom(x.Slice(static_cast(each_range.begin), + static_cast(each_range.end)), + x.place(), dev_ctx, &slice); offset += len; } } @@ -94,8 +103,7 @@ class LoDTensorToArrayOp : public framework::OperatorBase { class LoDTensorToArrayOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - LoDTensorToArrayOpProtoMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + LoDTensorToArrayOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", ""); AddInput("RankTable", ""); @@ -125,10 +133,10 @@ class LoDTensorToArrayInferShape : public framework::InferShapeBase { class LoDTensorToArrayInferVarType : public framework::VarTypeInference { public: - void operator()(const framework::OpDescBind &op_desc, - framework::BlockDescBind *block) const override { + void operator()(const framework::OpDesc &op_desc, + framework::BlockDesc *block) const override { for (auto &out_var : op_desc.Output("Out")) { - block->Var(out_var)->SetType(framework::VarDesc::LOD_TENSOR_ARRAY); + block->Var(out_var)->SetType(framework::proto::VarDesc::LOD_TENSOR_ARRAY); } } }; @@ -138,14 +146,14 @@ class LoDTensorToArrayGradMaker : public framework::SingleGradOpDescMaker { using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; protected: - std::unique_ptr Apply() const override { - auto *grad_op = new framework::OpDescBind(); + std::unique_ptr Apply() const override { + auto *grad_op = new framework::OpDesc(); grad_op->SetType("array_to_lod_tensor"); grad_op->SetInput("X", OutputGrad("Out")); grad_op->SetInput("RankTable", Input("RankTable")); grad_op->SetOutput("Out", InputGrad("X")); grad_op->SetAttrMap(Attrs()); - return std::unique_ptr(grad_op); + return std::unique_ptr(grad_op); } }; diff --git a/paddle/operators/log_loss_op.cc b/paddle/operators/log_loss_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..f714945354c5668f58e273dc8d6c7c16d51ac17d --- /dev/null +++ b/paddle/operators/log_loss_op.cc @@ -0,0 +1,115 @@ +/* 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. */ + +#include "paddle/operators/log_loss_op.h" + +namespace paddle { +namespace operators { + +class LogLossOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Predicted"), + "Input(Predicted) must be initialized."); + PADDLE_ENFORCE(ctx->HasInput("Labels"), + "Input(Labels) must be initialized."); + + auto pred_dims = ctx->GetInputDim("Predicted"); + auto label_dims = ctx->GetInputDim("Labels"); + + PADDLE_ENFORCE_EQ(pred_dims, label_dims); + PADDLE_ENFORCE_EQ(pred_dims.size(), 2, + "The rank of Input(Predicted) must be 2 and the shape is " + "[batch_size, 1]."); + PADDLE_ENFORCE_EQ(pred_dims[1], 1, + "Each row of Input(Predicted) contains a real value, " + "so the 2nd dimension of Input(X) must be 1."); + + ctx->SetOutputDim("Loss", {pred_dims[0], 1}); + ctx->ShareLoD("Predicted", "Loss"); + } +}; + +template +class LogLossOpMaker : public framework::OpProtoAndCheckerMaker { + public: + LogLossOpMaker(OpProto* proto, OpAttrChecker* op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("Predicted", + "The input value (Predicted) of Log loss op." + "Predicted is a 2-D tensor with shape [batch_size, 1]."); + AddInput("Labels", + "The target value (Labels) of Log loss op." + "Labels is a 2-D tensor with shape [batch_size, 1]."); + AddOutput("Loss", + "The output tensor with shape [batch_size, 1] " + "which represents the log loss."); + AddAttr("epsilon", "Epsilon in log loss."); + AddComment(R"DOC( +LogLoss Operator. + +Log loss is a loss function used for binary classification. Log Loss quantifies +the accuracy of a classifier by penalising false classifications. Minimising the +Log Loss is equivalent to maximising the accuracy of the classifier. We define +Predicted as the values predicted by our model and Labels as the target ground +truth value. Log loss can evaluate how close the predicted values are to the +target. The shapes of Predicted and Labels are both [batch_size, 1]. +The equation is: + +$$ +Loss = - Labels * log(Predicted + \epsilon) - + (1 - Labels) * log(1 - Predicted + \epsilon) +$$ + +)DOC"); + } +}; + +class LogLossGradOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Predicted"), + "Input(Predicted) should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Labels"), + "Input(Labels) should not be null."); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Loss")), + "Input(Loss@GRAD) should not be null."); + PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("Predicted")), + "Output(Predicted@GRAD) should not be null."); + + auto pred_dims = ctx->GetInputDim("Predicted"); + auto label_dims = ctx->GetInputDim("Labels"); + auto loss_grad_dims = ctx->GetInputDim(framework::GradVarName("Loss")); + PADDLE_ENFORCE_EQ(loss_grad_dims, pred_dims); + + auto pred_grad_name = framework::GradVarName("Predicted"); + ctx->SetOutputDim(pred_grad_name, pred_dims); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP(log_loss, ops::LogLossOp, ops::LogLossOpMaker, log_loss_grad, + ops::LogLossGradOp); +REGISTER_OP_CPU_KERNEL( + log_loss, ops::LogLossKernel); +REGISTER_OP_CPU_KERNEL( + log_loss_grad, + ops::LogLossGradKernel); diff --git a/paddle/operators/log_loss_op.cu b/paddle/operators/log_loss_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..be283e470052cc3a569be564ab4baa6bc5b75808 --- /dev/null +++ b/paddle/operators/log_loss_op.cu @@ -0,0 +1,23 @@ +/* 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. */ + +#define EIGEN_USE_GPU +#include "paddle/operators/log_loss_op.h" + +namespace ops = paddle::operators; +REGISTER_OP_CUDA_KERNEL( + log_loss, ops::LogLossKernel); +REGISTER_OP_CUDA_KERNEL( + log_loss_grad, + ops::LogLossGradKernel); diff --git a/paddle/operators/log_loss_op.h b/paddle/operators/log_loss_op.h new file mode 100644 index 0000000000000000000000000000000000000000..743eddb74004b5e87ed9b8a6ccb1b8496b8548dc --- /dev/null +++ b/paddle/operators/log_loss_op.h @@ -0,0 +1,75 @@ +/* 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 +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; +template +using EigenVector = framework::EigenVector; + +template +class LogLossKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* loss_out = ctx.Output("Loss"); + + loss_out->mutable_data(ctx.GetPlace()); + + auto epsilon = static_cast(ctx.Attr("epsilon")); + + auto prediction = EigenVector::Flatten(*ctx.Input("Predicted")); + auto label = EigenVector::Flatten(*ctx.Input("Labels")); + + auto loss = EigenVector::Flatten(*loss_out); + auto& place = *ctx.template device_context().eigen_device(); + + loss.device(place) = (-(label * (prediction + epsilon).log()) - + ((static_cast(1) - label) * + (static_cast(1) - prediction + epsilon).log())); + } +}; + +template +class LogLossGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto epsilon = static_cast(ctx.Attr("epsilon")); + + auto prediction = EigenVector::Flatten(*ctx.Input("Predicted")); + auto label = EigenVector::Flatten(*ctx.Input("Labels")); + + auto* dloss = ctx.Input(framework::GradVarName("Loss")); + auto* dpred = ctx.Output(framework::GradVarName("Predicted")); + + auto dl = EigenVector::Flatten(*dloss); + auto& place = *ctx.template device_context().eigen_device(); + + if (dpred) { + dpred->mutable_data(ctx.GetPlace()); + auto dx = framework::EigenVector::Flatten(*dpred); + dx.device(place) = dl * (-(label / (prediction + epsilon)) + + ((static_cast(1) - label) / + (static_cast(1) - prediction + epsilon))); + } + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/logical_op.cc b/paddle/operators/logical_op.cc index a37582c1d840ac11f847d8743c824ef1aef0fd66..7417192479a13ca9537e2d40f9779a3bf5f1eb61 100644 --- a/paddle/operators/logical_op.cc +++ b/paddle/operators/logical_op.cc @@ -1,16 +1,16 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +/* 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 +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 + 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. */ +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/operators/logical_op.h" #include "paddle/framework/op_registry.h" @@ -20,8 +20,7 @@ namespace operators { template class BinaryLogicalOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - BinaryLogicalOpProtoMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + BinaryLogicalOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { OpComment comment; AddInput("X", @@ -45,8 +44,7 @@ Each element of Out is calculated by %s template class UnaryLogicalOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - UnaryLogicalOpProtoMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + UnaryLogicalOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { OpComment comment; AddInput("X", string::Sprintf("(LoDTensor) Operand of %s operator", @@ -101,9 +99,9 @@ class LogicalOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - framework::OpKernelType GetKernelType( + framework::OpKernelType GetActualKernelType( const framework::ExecutionContext &ctx) const override { - framework::OpKernelType kt = OperatorWithKernel::GetKernelType(ctx); + framework::OpKernelType kt = OperatorWithKernel::GetActualKernelType(ctx); // LogicalOp kernel's device type is decided by input tensor place kt.place_ = ctx.Input("X")->place(); return kt; @@ -139,15 +137,16 @@ class LogicalOp : public framework::OperatorWithKernel { ::paddle::operators::UnaryLogicalOpInferShape<_##op_type##Comment>, \ ::paddle::framework::EmptyGradOpMaker); -REGISTER_BINARY_LOGICAL_OP(logical_and, "Out = X && Y"); +REGISTER_BINARY_LOGICAL_OP(logical_and, "$$Out = X \\&\\& Y$$"); REGISTER_BINARY_LOGICAL_KERNEL(logical_and, CPU, paddle::operators::LogicalAndFunctor); -REGISTER_BINARY_LOGICAL_OP(logical_or, "Out = X && Y"); +REGISTER_BINARY_LOGICAL_OP(logical_or, "$$Out = X || Y$$"); REGISTER_BINARY_LOGICAL_KERNEL(logical_or, CPU, paddle::operators::LogicalOrFunctor); -REGISTER_UNARY_LOGICAL_OP(logical_not, "Out = !X"); +REGISTER_UNARY_LOGICAL_OP(logical_not, "$$Out = !X$$"); REGISTER_UNARY_LOGICAL_KERNEL(logical_not, CPU, paddle::operators::LogicalNotFunctor); -REGISTER_BINARY_LOGICAL_OP(logical_xor, "Out = (X || Y) && !(X && Y)"); +REGISTER_BINARY_LOGICAL_OP(logical_xor, + "$$Out = (X || Y) \\, \\&\\& \\, !(X \\&\\& Y)$$"); REGISTER_BINARY_LOGICAL_KERNEL(logical_xor, CPU, paddle::operators::LogicalXorFunctor); diff --git a/paddle/operators/logical_op.cu b/paddle/operators/logical_op.cu index d41239b2ca43e7145ea56afcb0af69948838cc48..87f2287b8f11aabe8afe87776eff49295c1ea2ac 100644 --- a/paddle/operators/logical_op.cu +++ b/paddle/operators/logical_op.cu @@ -1,24 +1,24 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +/* 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 +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 + 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. */ +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/operators/logical_op.h" -REGISTER_BINARY_LOGICAL_KERNEL(logical_and, GPU, +REGISTER_BINARY_LOGICAL_KERNEL(logical_and, CUDA, paddle::operators::LogicalAndFunctor); -REGISTER_BINARY_LOGICAL_KERNEL(logical_or, GPU, +REGISTER_BINARY_LOGICAL_KERNEL(logical_or, CUDA, paddle::operators::LogicalOrFunctor); -REGISTER_UNARY_LOGICAL_KERNEL(logical_not, GPU, +REGISTER_UNARY_LOGICAL_KERNEL(logical_not, CUDA, paddle::operators::LogicalNotFunctor); -REGISTER_BINARY_LOGICAL_KERNEL(logical_xor, GPU, +REGISTER_BINARY_LOGICAL_KERNEL(logical_xor, CUDA, paddle::operators::LogicalXorFunctor); diff --git a/paddle/operators/logical_op.h b/paddle/operators/logical_op.h index 6e78a7d6ed87ba950886e6bc667f82118ff78904..413857685603c7b84e885135d9aadf7cc71a4f72 100644 --- a/paddle/operators/logical_op.h +++ b/paddle/operators/logical_op.h @@ -1,16 +1,16 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +/* 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 +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 + 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. */ +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 @@ -47,7 +47,7 @@ struct LogicalXorFunctor { } }; -template +template class BinaryLogicalOpKernel : public framework::OpKernel { public: @@ -57,14 +57,14 @@ class BinaryLogicalOpKernel auto* y = context.Input("Y"); auto* out = context.Output("Out"); Functor binary_func; - platform::Transform trans; - trans(context.device_context(), x->data(), x->data() + x->numel(), - y->data(), out->mutable_data(context.GetPlace()), - binary_func); + platform::Transform trans; + trans(context.template device_context(), x->data(), + x->data() + x->numel(), y->data(), + out->mutable_data(context.GetPlace()), binary_func); } }; -template +template class UnaryLogicalOpKernel : public framework::OpKernel { public: @@ -73,8 +73,9 @@ class UnaryLogicalOpKernel auto* x = context.Input("X"); auto* out = context.Output("Out"); Functor unary_func; - platform::Transform trans; - trans(context.device_context(), x->data(), x->data() + x->numel(), + platform::Transform trans; + trans(context.template device_context(), x->data(), + x->data() + x->numel(), out->mutable_data(context.GetPlace()), unary_func); } }; @@ -85,9 +86,9 @@ class UnaryLogicalOpKernel #define REGISTER_BINARY_LOGICAL_KERNEL(op_type, dev, functor) \ REGISTER_OP_##dev##_KERNEL( \ op_type, ::paddle::operators::BinaryLogicalOpKernel< \ - ::paddle::platform::dev##Place, functor>); + ::paddle::platform::dev##DeviceContext, functor>); #define REGISTER_UNARY_LOGICAL_KERNEL(op_type, dev, functor) \ REGISTER_OP_##dev##_KERNEL( \ op_type, ::paddle::operators::UnaryLogicalOpKernel< \ - ::paddle::platform::dev##Place, functor>); + ::paddle::platform::dev##DeviceContext, functor>); diff --git a/paddle/operators/lookup_table_op.cc b/paddle/operators/lookup_table_op.cc index 93e812ac5be5aea6bf3ab353d31480322c51ccbc..6e5cbd6f8cefc965d6c8d24b16eb3bafde55cc49 100644 --- a/paddle/operators/lookup_table_op.cc +++ b/paddle/operators/lookup_table_op.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/operators/lookup_table_op.h" #include "paddle/framework/var_type_inference.h" @@ -41,7 +41,7 @@ class LookupTableOp : public framework::OperatorWithKernel { } protected: - framework::OpKernelType GetKernelType( + framework::OpKernelType GetActualKernelType( const framework::ExecutionContext& ctx) const override { return framework::OpKernelType( framework::ToDataType(ctx.Input("W")->type()), @@ -51,8 +51,7 @@ class LookupTableOp : public framework::OperatorWithKernel { class LookupTableOpMaker : public framework::OpProtoAndCheckerMaker { public: - LookupTableOpMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker) + LookupTableOpMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("W", "An input represents embedding tensors, " @@ -99,7 +98,7 @@ class LookupTableOpGrad : public framework::OperatorWithKernel { } protected: - framework::OpKernelType GetKernelType( + framework::OpKernelType GetActualKernelType( const framework::ExecutionContext& ctx) const override { return framework::OpKernelType( framework::ToDataType(ctx.Input("W")->type()), @@ -109,19 +108,20 @@ class LookupTableOpGrad : public framework::OperatorWithKernel { class LookupTableOpGradVarTypeInference : public framework::VarTypeInference { public: - void operator()(const framework::OpDescBind& op_desc, - framework::BlockDescBind* block) const override { + void operator()(const framework::OpDesc& op_desc, + framework::BlockDesc* block) const override { auto out_var_name = op_desc.Output(framework::GradVarName("W")).front(); auto attr = op_desc.GetAttr("is_sparse"); bool is_sparse = boost::get(attr); if (is_sparse) { VLOG(3) << "lookup_table_grad op " << framework::GradVarName("W") << " is set to SelectedRows"; - block->Var(out_var_name)->SetType(framework::VarDesc::SELECTED_ROWS); + block->Var(out_var_name) + ->SetType(framework::proto::VarDesc::SELECTED_ROWS); } else { VLOG(3) << "lookup_table_grad op " << framework::GradVarName("W") << " is set to LoDTensor"; - block->Var(out_var_name)->SetType(framework::VarDesc::LOD_TENSOR); + block->Var(out_var_name)->SetType(framework::proto::VarDesc::LOD_TENSOR); } } }; diff --git a/paddle/operators/lookup_table_op.cu b/paddle/operators/lookup_table_op.cu index 84b044184a36a0d3a72a4105d6baf401b4774cf7..261a28da694bf551d8d9e630139680aebc4be51a 100644 --- a/paddle/operators/lookup_table_op.cu +++ b/paddle/operators/lookup_table_op.cu @@ -1,13 +1,16 @@ /* 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. */ + +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/framework/eigen.h" #include "paddle/framework/op_registry.h" @@ -85,6 +88,8 @@ template class LookupTableGradCUDAKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { + auto& dev_ctx = + context.template device_context(); bool is_sparse = context.Attr("is_sparse"); if (is_sparse) { auto* ids = context.Input("Ids"); @@ -95,11 +100,11 @@ class LookupTableGradCUDAKernel : public framework::OpKernel { auto* ids_data = ids->data(); auto ids_dim = ids->dims(); - auto stream = context.cuda_device_context().stream(); + auto stream = dev_ctx.stream(); // copy GPU memory to CPU pinned memory framework::Vector new_rows; new_rows.resize(ids_dim[0]); - auto gpu_place = boost::get(context.GetPlace()); + auto gpu_place = boost::get(context.GetPlace()); memory::Copy(platform::CPUPlace(), new_rows.data(), gpu_place, ids_data, ids_dim[0] * sizeof(int64_t), stream); @@ -129,14 +134,11 @@ class LookupTableGradCUDAKernel : public framework::OpKernel { T* d_table = d_table_t->mutable_data(context.GetPlace()); auto t = framework::EigenVector::Flatten(*d_table_t); - t.device(context.GetEigenDevice()) = - t.constant(static_cast(0)); + t.device(*dev_ctx.eigen_device()) = t.constant(static_cast(0)); dim3 threads(128, 8); dim3 grids(8, 1); - LookupTableGrad< - T, 128, 8, - 8><<>>( + LookupTableGrad<<>>( d_table, d_output, ids, N, K, D); } } @@ -146,7 +148,8 @@ class LookupTableGradCUDAKernel : public framework::OpKernel { } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(lookup_table, ops::LookupTableCUDAKernel, - ops::LookupTableCUDAKernel); -REGISTER_OP_GPU_KERNEL(lookup_table_grad, ops::LookupTableGradCUDAKernel, - ops::LookupTableGradCUDAKernel); +REGISTER_OP_CUDA_KERNEL(lookup_table, ops::LookupTableCUDAKernel, + ops::LookupTableCUDAKernel); +REGISTER_OP_CUDA_KERNEL(lookup_table_grad, + ops::LookupTableGradCUDAKernel, + ops::LookupTableGradCUDAKernel); diff --git a/paddle/operators/lookup_table_op.h b/paddle/operators/lookup_table_op.h index 99b912163b71594340d8917645dff107fd208aea..2fd3335868406455ec01f9ded6bacc7bda5e2a67 100644 --- a/paddle/operators/lookup_table_op.h +++ b/paddle/operators/lookup_table_op.h @@ -1,13 +1,16 @@ /* 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. */ + +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 diff --git a/paddle/operators/lrn_op.cc b/paddle/operators/lrn_op.cc index 00392b7967d020a7951a16a7850a2f08735baeb8..95673ba19e776b3c52eb492d0b14d761b584f807 100644 --- a/paddle/operators/lrn_op.cc +++ b/paddle/operators/lrn_op.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/operators/lrn_op.h" @@ -19,6 +19,103 @@ namespace operators { using framework::Tensor; +template +struct LRNFunctor { + void operator()(const framework::ExecutionContext& ctx, + const framework::Tensor& input, framework::Tensor* out, + framework::Tensor* mid, int N, int C, int H, int W, int n, + T k, T alpha, T beta) { + auto x_v = framework::EigenVector::Flatten(input); + + const int start = -(n - 1) / 2; + const int end = start + n; + + auto e_mid = framework::EigenTensor::From(*mid); + e_mid = e_mid.constant(k); + + auto e_x = framework::EigenTensor::From(input); + for (int m = 0; m < N; m++) { + for (int i = 0; i < C; i++) { + for (int c = start; c <= end; c++) { + int ch = i + c; + if (ch >= 0 && ch < C) { + auto s = e_mid.slice(Eigen::array({{m, i, 0, 0}}), + Eigen::array({{1, 1, H, W}})); + + auto r = e_x.slice(Eigen::array({{m, ch, 0, 0}}), + Eigen::array({{1, 1, H, W}})); + + s += alpha * r.square(); + } + } + } + } + + auto out_e = framework::EigenVector::Flatten(*out); + out_e = x_v * e_mid.reshape(Eigen::DSizes(e_mid.size())).pow(-beta); + } +}; +template struct LRNFunctor; +template struct LRNFunctor; + +template +struct LRNGradFunctor { + void operator()(const framework::ExecutionContext& ctx, + const framework::Tensor& x, const framework::Tensor& out, + const framework::Tensor& mid, framework::Tensor* x_g, + const framework::Tensor& out_g, int N, int C, int H, int W, + int n, T alpha, T beta) { + T ratio = -2 * alpha * beta; + auto x_g_e = framework::EigenVector::Flatten(*x_g); + x_g_e = x_g_e.constant(0.0); + + auto e_x = framework::EigenTensor::From(x); + auto e_x_g = framework::EigenTensor::From(*x_g); + auto e_out = framework::EigenTensor::From(out); + auto e_out_g = framework::EigenTensor::From(out_g); + auto e_mid = framework::EigenTensor::From(mid); + + const int start = -(n - 1) / 2; + const int end = start + n; + for (int m = 0; m < N; m++) { + for (int i = 0; i < C; i++) { + auto i_x = e_x.slice(Eigen::array({{m, i, 0, 0}}), + Eigen::array({{1, 1, H, W}})); + + auto i_x_g = e_x_g.slice(Eigen::array({{m, i, 0, 0}}), + Eigen::array({{1, 1, H, W}})); + + auto i_out_g = e_out_g.slice(Eigen::array({{m, i, 0, 0}}), + Eigen::array({{1, 1, H, W}})); + + auto i_mid = e_mid.slice(Eigen::array({{m, i, 0, 0}}), + Eigen::array({{1, 1, H, W}})); + + i_x_g = i_mid.pow(-beta) * i_out_g; + for (int c = start; c <= end; c++) { + int ch = i + c; + if (ch < 0 || ch >= C) { + continue; + } + + auto c_out = e_out.slice(Eigen::array({{m, ch, 0, 0}}), + Eigen::array({{1, 1, H, W}})); + + auto c_mid = e_mid.slice(Eigen::array({{m, ch, 0, 0}}), + Eigen::array({{1, 1, H, W}})); + + auto c_out_g = e_out_g.slice(Eigen::array({{m, ch, 0, 0}}), + Eigen::array({{1, 1, H, W}})); + + i_x_g += ratio * c_out_g * c_out * i_x / c_mid; + } + } + } + } +}; +template struct LRNGradFunctor; +template struct LRNGradFunctor; + class LRNOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; @@ -43,7 +140,7 @@ class LRNOp : public framework::OperatorWithKernel { template class LRNOpMaker : public framework::OpProtoAndCheckerMaker { public: - LRNOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker) + LRNOpMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "(Tensor) The input of LRN operator. " @@ -83,8 +180,8 @@ class LRNOpMaker : public framework::OpProtoAndCheckerMaker { AddComment(R"DOC( Local Response Normalization Operator. -This operator comes from the paper -"ImageNet Classification with Deep Convolutional Neural Networks". +This operator comes from the paper: +<>. The original formula is: @@ -107,7 +204,7 @@ Input(i, x, y), Output(i, x, y) represents an element in an image. C is the number of feature maps of one image. n is a hyper-parameter configured when operator is initialized. The sum in the denominator is the sum of the same positions in the neighboring maps. - + )DOC"); } }; @@ -119,8 +216,7 @@ class LRNOpGrad : public framework::OperatorWithKernel { protected: void InferShape(framework::InferShapeContext* ctx) const override { PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null"); - PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("MidOut")), - "Input(MidOut@GRAD) should not be null"); + PADDLE_ENFORCE(ctx->HasInput("MidOut"), "Input(MidOut) should not be null"); PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), "Input(Out@GRAD) should not be null"); @@ -134,6 +230,7 @@ class LRNOpGrad : public framework::OperatorWithKernel { namespace ops = paddle::operators; REGISTER_OP(lrn, ops::LRNOp, ops::LRNOpMaker, lrn_grad, ops::LRNOpGrad); -REGISTER_OP_CPU_KERNEL(lrn, ops::LRNKernel); -REGISTER_OP_CPU_KERNEL(lrn_grad, - ops::LRNGradKernel); +REGISTER_OP_CPU_KERNEL( + lrn, ops::LRNKernel); +REGISTER_OP_CPU_KERNEL( + lrn_grad, ops::LRNGradKernel); diff --git a/paddle/operators/lrn_op.cu b/paddle/operators/lrn_op.cu index 607dc6d86a72b0a0c953f52782955dc530b7478c..eb9d66a73dfe3e22f1151d73ce5e34f2eda0835e 100644 --- a/paddle/operators/lrn_op.cu +++ b/paddle/operators/lrn_op.cu @@ -1,22 +1,178 @@ /* 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 +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 + 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. */ +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. */ -#define EIGEN_USE_GPU #include "paddle/operators/lrn_op.h" -namespace ops = paddle::operators; +namespace paddle { +namespace operators { + +template +__global__ void KeCMRNormFillScale(int img_size, const T* in, T* mid, int C, + int H, int W, int size, T k, T alpha) { + const int idx = threadIdx.x + blockIdx.x * blockDim.x; + if (idx < img_size) { + const int w = idx % W; + const int h = (idx / W) % H; + const int n = idx / W / H; + const int offset = (n * C * H + h) * W + w; + + in += offset; + mid += offset; + const int step = H * W; + const int pre_pad = (size - 1) / 2; + const int post_pad = size - pre_pad - 1; + + T accum = 0; + int index = 0; + while (index < C + post_pad) { + if (index < C) { + T val = in[index * step]; + accum += val * val; + } + if (index >= size) { + T val = in[(index - size) * step]; + accum -= val * val; + } + if (index >= post_pad) { + mid[(index - post_pad) * step] = k + accum * alpha; + } + ++index; + } + } +} + +template +__global__ void KeCMRNormOutput(int input_size, const T* in, const T* mid, + T negative_beta, T* out) { + const int index = threadIdx.x + blockIdx.x * blockDim.x; + if (index < input_size) { + out[index] = in[index] * pow(mid[index], negative_beta); + } +} + +template +void CrossMapNormal(const framework::ExecutionContext& ctx, const T* inputs, + T* outputs, T* mid, int N, int C, int H, int W, int n, T k, + T alpha, T beta) { + int img_size = N * H * W; + const int block_size = 1024; + int grid_size = (img_size + block_size - 1) / block_size; + + auto& dev_ctx = ctx.template device_context(); + KeCMRNormFillScale<<>>( + img_size, inputs, mid, C, H, W, n, k, alpha); + + int input_size = N * H * W * C; + grid_size = (input_size + block_size - 1) / block_size; + KeCMRNormOutput<<>>( + input_size, inputs, mid, -beta, outputs); +} + +template +struct LRNFunctor { + void operator()(const framework::ExecutionContext& ctx, + const framework::Tensor& input, framework::Tensor* out, + framework::Tensor* mid, int N, int C, int H, int W, int n, + T k, T alpha, T beta) { + CrossMapNormal( + ctx, input.data(), out->mutable_data(ctx.GetPlace()), + mid->mutable_data(ctx.GetPlace()), N, C, H, W, n, k, alpha, beta); + } +}; + +template struct LRNFunctor; +template struct LRNFunctor; -REGISTER_OP_GPU_KERNEL(lrn, ops::LRNKernel); -REGISTER_OP_GPU_KERNEL(lrn_grad, - ops::LRNGradKernel); +template +__global__ void KeCMRNormDiff(int img_size, const T* x, const T* out, + const T* mid, T* x_g, const T* out_g, int C, + int H, int W, int size, T negative_beta, + T ratio) { + const int idx = threadIdx.x + blockIdx.x * blockDim.x; + if (idx < img_size) { + const int w = idx % W; + const int h = (idx / W) % H; + const int n = idx / W / H; + const int offset = (n * C * H + h) * W + w; + x += offset; + out += offset; + mid += offset; + out_g += offset; + x_g += offset; + + const int step = H * W; + const int pre_pad = size - (size + 1) / 2; + const int post_pad = size - pre_pad - 1; + + int index = 0; + T accum = 0; + // TODO(gongwb): optimize this with thread shared array. + while (index < C + post_pad) { + if (index < C) { + x_g[index * step] = 0.0; + accum += out_g[index * step] * out[index * step] / mid[index * step]; + } + if (index >= size) { + accum -= out_g[(index - size) * step] * out[(index - size) * step] / + mid[(index - size) * step]; + } + if (index >= post_pad) { + x_g[(index - post_pad) * step] += + out_g[(index - post_pad) * step] * + pow(mid[(index - post_pad) * step], negative_beta) - + ratio * x[(index - post_pad) * step] * accum; + } + ++index; + } + } +} + +template +void CrossMapNormalGrad(const framework::ExecutionContext& ctx, const T* x, + const T* out, const T* mid, T* x_g, const T* out_g, + int N, int C, int H, int W, int n, T alpha, T beta) { + int img_size = N * H * W; + + const int block_size = 1024; + int grid_size = (img_size + block_size - 1) / block_size; + + auto& dev_ctx = ctx.template device_context(); + KeCMRNormDiff<<>>( + img_size, x, out, mid, x_g, out_g, C, H, W, n, -beta, + 2.0f * alpha * beta); +} + +template +struct LRNGradFunctor { + void operator()(const framework::ExecutionContext& ctx, + const framework::Tensor& x, const framework::Tensor& out, + const framework::Tensor& mid, framework::Tensor* x_g, + const framework::Tensor& out_g, int N, int C, int H, int W, + int n, T alpha, T beta) { + CrossMapNormalGrad(ctx, x.data(), out.data(), mid.data(), + x_g->mutable_data(ctx.GetPlace()), out_g.data(), + N, C, H, W, n, alpha, beta); + } +}; + +template struct LRNGradFunctor; +template struct LRNGradFunctor; +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP_CUDA_KERNEL( + lrn, ops::LRNKernel); +REGISTER_OP_CUDA_KERNEL( + lrn_grad, ops::LRNGradKernel); diff --git a/paddle/operators/lrn_op.h b/paddle/operators/lrn_op.h index 606c65744303b53846c9077dfa832bdbeedb410e..ef3a2883a88ff321fb9a87ddaf31123a3b9ee90a 100644 --- a/paddle/operators/lrn_op.h +++ b/paddle/operators/lrn_op.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 @@ -21,7 +21,15 @@ namespace paddle { namespace operators { -template +template +struct LRNFunctor { + void operator()(const framework::ExecutionContext& ctx, + const framework::Tensor& input, framework::Tensor* out, + framework::Tensor* mid, int N, int C, int H, int W, int n, + T k, T alpha, T beta); +}; + +template class LRNKernel : public framework::OpKernel { public: using Tensor = framework::Tensor; @@ -31,8 +39,8 @@ class LRNKernel : public framework::OpKernel { // f(x) represents outputs void Compute(const framework::ExecutionContext& ctx) const override { // input - const Tensor* x = ctx.Input("X"); - auto x_dims = x->dims(); + const Tensor& x = *ctx.Input("X"); + auto x_dims = x.dims(); // NCHW int N = x_dims[0]; @@ -57,38 +65,20 @@ class LRNKernel : public framework::OpKernel { PADDLE_ENFORCE(beta >= 0.0, "beta should >= 0.0"); PADDLE_ENFORCE(k >= 0.0, "k should >= 0.0"); - auto x_v = framework::EigenVector::Flatten(*x); - - const int start = -(n - 1) / 2; - const int end = start + n; - - auto e_mid = framework::EigenTensor::From(*mid); - e_mid.device(ctx.GetEigenDevice()) = e_mid.constant(k); - - auto e_x = framework::EigenTensor::From(*x); - for (int m = 0; m < N; m++) { - for (int i = 0; i < C; i++) { - for (int c = start; c <= end; c++) { - int ch = i + c; - if (ch >= 0 && ch < C) { - auto s = e_mid.slice(Eigen::array({{m, i, 0, 0}}), - Eigen::array({{1, 1, H, W}})); - - auto r = e_x.slice(Eigen::array({{m, ch, 0, 0}}), - Eigen::array({{1, 1, H, W}})); - - s.device(ctx.GetEigenDevice()) += alpha * r.square(); - } - } - } - } - - auto out_e = framework::EigenVector::Flatten(*out); - out_e.device(ctx.GetEigenDevice()) = - x_v * e_mid.reshape(Eigen::DSizes(e_mid.size())).pow(-beta); + LRNFunctor f; + f(ctx, x, out, mid, N, C, H, W, n, k, alpha, beta); } }; +template +struct LRNGradFunctor { + void operator()(const framework::ExecutionContext& ctx, + const framework::Tensor& x, const framework::Tensor& out, + const framework::Tensor& mid, framework::Tensor* x_g, + const framework::Tensor& out_g, int N, int C, int H, int W, + int n, T alpha, T beta); +}; + /** * \brief Backward calculation for normalization with across maps. * @@ -97,7 +87,7 @@ class LRNKernel : public framework::OpKernel { * The implementation of this Function is derived from the * CrossMapNormalFunc implementation. * - * InputGrad = OutputGrad * denoms ^ (-beta) + * InputGrad = OutputGrad * MidOut ^ (-beta) * -- upper * + > (OutputGrad * OutputValue * (-2 * alpha * beta) / MidOut) * InputValue * -- lower @@ -108,23 +98,20 @@ class LRNKernel : public framework::OpKernel { * The upper and lower is the same as forward. The logic of the sum * is also the same as forward. */ -template +template class LRNGradKernel : public framework::OpKernel { public: using Tensor = framework::Tensor; void Compute(const framework::ExecutionContext& ctx) const override { - const Tensor* x = ctx.Input("X"); - const Tensor* out = ctx.Input("Out"); - const Tensor* out_g = ctx.Input(framework::GradVarName("Out")); - const Tensor* mid = ctx.Input("MidOut"); + const Tensor& x = *ctx.Input("X"); + const Tensor& out = *ctx.Input("Out"); + const Tensor& out_g = *ctx.Input(framework::GradVarName("Out")); + const Tensor& mid = *ctx.Input("MidOut"); auto x_g = ctx.Output(framework::GradVarName("X")); x_g->mutable_data(ctx.GetPlace()); - auto x_g_e = framework::EigenVector::Flatten(*x_g); - x_g_e.device(ctx.GetEigenDevice()) = x_g_e.constant(0.0); - - auto x_dims = x->dims(); + auto x_dims = x.dims(); int N = x_dims[0]; int C = x_dims[1]; int H = x_dims[2]; @@ -133,51 +120,9 @@ class LRNGradKernel : public framework::OpKernel { int n = ctx.Attr("n"); T alpha = ctx.Attr("alpha"); T beta = ctx.Attr("beta"); - T ratio = -2 * alpha * beta; - - auto e_x = framework::EigenTensor::From(*x); - auto e_x_g = framework::EigenTensor::From(*x_g); - auto e_out = framework::EigenTensor::From(*out); - auto e_out_g = framework::EigenTensor::From(*out_g); - auto e_mid = framework::EigenTensor::From(*mid); - - const int start = -(n - 1) / 2; - const int end = start + n; - for (int m = 0; m < N; m++) { - for (int i = 0; i < C; i++) { - auto i_x = e_x.slice(Eigen::array({{m, i, 0, 0}}), - Eigen::array({{1, 1, H, W}})); - - auto i_x_g = e_x_g.slice(Eigen::array({{m, i, 0, 0}}), - Eigen::array({{1, 1, H, W}})); - - auto i_out_g = e_out_g.slice(Eigen::array({{m, i, 0, 0}}), - Eigen::array({{1, 1, H, W}})); - - auto i_mid = e_mid.slice(Eigen::array({{m, i, 0, 0}}), - Eigen::array({{1, 1, H, W}})); - - i_x_g.device(ctx.GetEigenDevice()) = i_mid.pow(-beta) * i_out_g; - for (int c = start; c <= end; c++) { - int ch = i + c; - if (ch < 0 || ch >= C) { - continue; - } - - auto c_out = e_out.slice(Eigen::array({{m, ch, 0, 0}}), - Eigen::array({{1, 1, H, W}})); - - auto c_mid = e_mid.slice(Eigen::array({{m, ch, 0, 0}}), - Eigen::array({{1, 1, H, W}})); - - auto c_out_g = e_out_g.slice(Eigen::array({{m, ch, 0, 0}}), - Eigen::array({{1, 1, H, W}})); - - i_x_g.device(ctx.GetEigenDevice()) += - ratio * c_out_g * c_out * i_x / c_mid; - } - } - } + + LRNGradFunctor f; + f(ctx, x, out, mid, x_g, out_g, N, C, H, W, n, alpha, beta); } }; diff --git a/paddle/operators/lstm_op.cc b/paddle/operators/lstm_op.cc index 4cbb60f3fdab968e8c36d4fbad55fd3efc7b1d0d..b8fcec0f29b46e838f91ad1ee0fded8e42f27bd5 100644 --- a/paddle/operators/lstm_op.cc +++ b/paddle/operators/lstm_op.cc @@ -92,7 +92,7 @@ class LSTMOp : public framework::OperatorWithKernel { } protected: - framework::OpKernelType GetKernelType( + framework::OpKernelType GetActualKernelType( const framework::ExecutionContext& ctx) const override { return framework::OpKernelType( framework::ToDataType(ctx.Input("Input")->type()), @@ -102,7 +102,7 @@ class LSTMOp : public framework::OperatorWithKernel { class LSTMOpMaker : public framework::OpProtoAndCheckerMaker { public: - LSTMOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker) + LSTMOpMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("Input", "(LoDTensor) the first input is a LodTensor, which support " @@ -181,7 +181,7 @@ class LSTMOpMaker : public framework::OpProtoAndCheckerMaker { AddComment(R"DOC( Long-Short Term Memory (LSTM) Operator. -The defalut implementation is diagonal/peephole connection +The defalut implementation is diagonal/peephole connection (https://arxiv.org/pdf/1402.1128.pdf), the formula is as follows: $$ @@ -198,27 +198,27 @@ c_t = f_t \odot c_{t-1} + i_t \odot \tilde{c_t} \\ h_t = o_t \odot act_h(c_t) $$ -where the W terms denote weight matrices (e.g. \f$W_{xi}\f$ is the matrix -of weights from the input gate to the input), \f$W_{ic}, W_{fc}, W_{oc}\f$ +where the W terms denote weight matrices (e.g. $W_{xi}$ is the matrix +of weights from the input gate to the input), $W_{ic}, W_{fc}, W_{oc}$ are diagonal weight matrices for peephole connections. In our implementation, we use vectors to reprenset these diagonal weight matrices. The b terms -denote bias vectors (\f$b_i\f$ is the input gate bias vector), \f$\sigma\f$ +denote bias vectors ($b_i$ is the input gate bias vector), $\sigma$ is the non-line activations, such as logistic sigmoid function, and -\f$i, f, o\f$ and \f$c\f$ are the input gate, forget gate, output gate, +$i, f, o$ and $c$ are the input gate, forget gate, output gate, and cell activation vectors, respectively, all of which have the same size as -the cell output activation vector \f$h\f$. +the cell output activation vector $h$. -The \f$\odot\f$ is the element-wise product of the vectors. \f$act_g\f$ and \f$act_h\f$ +The $\odot$ is the element-wise product of the vectors. $act_g$ and $act_h$ are the cell input and cell output activation functions and `tanh` is usually -used for them. \f$\tilde{c_t}\f$ is also called candidate hidden state, +used for them. $\tilde{c_t}$ is also called candidate hidden state, which is computed based on the current input and the previous hidden state. -Set `use_peepholes` False to disable peephole connection -(http://www.bioinf.jku.at/publications/older/2604.pdf). The formula -is omitted here. +Set `use_peepholes` False to disable peephole connection. The formula +is omitted here, please refer to the paper +http://www.bioinf.jku.at/publications/older/2604.pdf for details. -Note that these \f$W_{xi}x_{t}, W_{xf}x_{t}, W_{xc}x_{t}, W_{xo}x_{t}\f$ -operations on the input \f$x_{t}\f$ are NOT included in this operator. +Note that these $W_{xi}x_{t}, W_{xf}x_{t}, W_{xc}x_{t}, W_{xo}x_{t}$ +operations on the input $x_{t}$ are NOT included in this operator. Users can choose to use fully-connect operator before LSTM operator. )DOC"); @@ -260,7 +260,7 @@ class LSTMGradOp : public framework::OperatorWithKernel { } protected: - framework::OpKernelType GetKernelType( + framework::OpKernelType GetActualKernelType( const framework::ExecutionContext& ctx) const override { return framework::OpKernelType( framework::ToDataType(ctx.Input("Input")->type()), @@ -273,8 +273,9 @@ class LSTMGradOp : public framework::OperatorWithKernel { namespace ops = paddle::operators; REGISTER_OP(lstm, ops::LSTMOp, ops::LSTMOpMaker, lstm_grad, ops::LSTMGradOp); -REGISTER_OP_CPU_KERNEL(lstm, ops::LSTMKernel, - ops::LSTMKernel); -REGISTER_OP_CPU_KERNEL(lstm_grad, - ops::LSTMGradKernel, - ops::LSTMGradKernel); +REGISTER_OP_CPU_KERNEL( + lstm, ops::LSTMKernel, + ops::LSTMKernel); +REGISTER_OP_CPU_KERNEL( + lstm_grad, ops::LSTMGradKernel, + ops::LSTMGradKernel); diff --git a/paddle/operators/lstm_op.cu.cc b/paddle/operators/lstm_op.cu.cc index 610cbb03e890203407b1489800bc17f1a196d12c..cfcc1fc92a074c9bfe83e6c32560177edef12ae9 100644 --- a/paddle/operators/lstm_op.cu.cc +++ b/paddle/operators/lstm_op.cu.cc @@ -1,22 +1,23 @@ /* 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 +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 + 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. */ +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/operators/lstm_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(lstm, ops::LSTMKernel, - ops::LSTMKernel); -REGISTER_OP_GPU_KERNEL(lstm_grad, - ops::LSTMGradKernel, - ops::LSTMGradKernel); +REGISTER_OP_CUDA_KERNEL( + lstm, ops::LSTMKernel, + ops::LSTMKernel); +REGISTER_OP_CUDA_KERNEL( + lstm_grad, ops::LSTMGradKernel, + ops::LSTMGradKernel); diff --git a/paddle/operators/lstm_op.h b/paddle/operators/lstm_op.h index 721aa42c92f2926aabbc13d0a9027b2b4e573225..c57ee414dc5b3417549c8ac3a7fd57a9c8f452df 100644 --- a/paddle/operators/lstm_op.h +++ b/paddle/operators/lstm_op.h @@ -14,6 +14,7 @@ limitations under the License. */ #pragma once #include "paddle/framework/op_registry.h" +#include "paddle/operators/math/detail/activation_functions.h" #include "paddle/operators/math/lstm_compute.h" #include "paddle/operators/math/math_function.h" #include "paddle/operators/math/sequence2batch.h" @@ -24,16 +25,16 @@ namespace operators { using LoDTensor = framework::LoDTensor; using Tensor = framework::Tensor; -template -inline void ReorderInitState(const platform::DeviceContext& ctx, +template +inline void ReorderInitState(const DeviceContext& ctx, const framework::Tensor& src, const size_t* index, framework::Tensor* dst, bool indexed_src) { - math::CopyMatrixRowsFunctor row_shuffle; + math::CopyMatrixRowsFunctor row_shuffle; dst->mutable_data(src.dims(), ctx.GetPlace()); row_shuffle(ctx, src, index, *dst, indexed_src); } -template +template class LSTMKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { @@ -52,8 +53,8 @@ class LSTMKernel : public framework::OpKernel { cell_out->mutable_data(ctx.GetPlace()); bool is_reverse = ctx.Attr("is_reverse"); - math::LoDTensor2BatchFunctor to_batch; - auto& device_ctx = ctx.device_context(); + math::LoDTensor2BatchFunctor to_batch; + auto& device_ctx = ctx.template device_context(); to_batch(device_ctx, *input, *batch_gate, true, is_reverse); auto in_dims = input->dims(); @@ -64,7 +65,7 @@ class LSTMKernel : public framework::OpKernel { Tensor b = *bias; b.Resize({bias->numel(), 1}); Tensor gate_bias = b.Slice(0, 4 * frame_size); - math::RowwiseAdd add_bias; + math::RowwiseAdd add_bias; add_bias(device_ctx, *batch_gate, gate_bias, batch_gate); } @@ -73,24 +74,24 @@ class LSTMKernel : public framework::OpKernel { T* bias_data = const_cast(bias->data()); // the code style in LstmMetaValue will be updated later. - lstm_value.checkIg = bias_data + 4 * frame_size; - lstm_value.checkFg = lstm_value.checkIg + frame_size; - lstm_value.checkOg = lstm_value.checkFg + frame_size; + lstm_value.check_ig = bias_data + 4 * frame_size; + lstm_value.check_fg = lstm_value.check_ig + frame_size; + lstm_value.check_og = lstm_value.check_fg + frame_size; } else { - lstm_value.checkIg = nullptr; - lstm_value.checkFg = nullptr; - lstm_value.checkOg = nullptr; + lstm_value.check_ig = nullptr; + lstm_value.check_fg = nullptr; + lstm_value.check_og = nullptr; } - lstm_value.prevStateValue = nullptr; + lstm_value.prev_state_value = nullptr; Tensor ordered_c0; const size_t* order = batch_gate->lod()[2].data(); if (cell_t0) { // Since the batch computing for LSTM reorders the input sequence // according to their length. The initialized cell state also needs // to reorder. - ReorderInitState(device_ctx, *cell_t0, order, &ordered_c0, - true); - lstm_value.prevStateValue = ordered_c0.data(); + ReorderInitState(device_ctx, *cell_t0, order, + &ordered_c0, true); + lstm_value.prev_state_value = ordered_c0.data(); } // Use the local variable as here. @@ -102,9 +103,12 @@ class LSTMKernel : public framework::OpKernel { auto batch_starts = batch_gate->lod()[0]; size_t num_batch = batch_starts.size() - 1; - auto gate_act = ctx.Attr("gate_activation"); - auto cell_act = ctx.Attr("cell_activation"); - auto cand_act = ctx.Attr("candidate_activation"); + auto gate_act = math::detail::GetActivationType( + ctx.Attr("gate_activation")); + auto cell_act = math::detail::GetActivationType( + ctx.Attr("cell_activation")); + auto cand_act = math::detail::GetActivationType( + ctx.Attr("candidate_activation")); for (size_t n = 0; n < num_batch; n++) { int bstart = static_cast(batch_starts[n]); @@ -121,9 +125,9 @@ 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)); + math::matmul(device_ctx, 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. @@ -133,24 +137,24 @@ class LSTMKernel : public framework::OpKernel { // according to their length. The initialized hidden state also needs // to reorder. 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)); + 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)); } - lstm_value.gateValue = gate_t.data(); - lstm_value.outputValue = out_t.data(); - lstm_value.stateValue = cell_t.data(); - lstm_value.stateActiveValue = cell_pre_act_t.data(); - math::LstmUnitFunctor::compute(device_ctx, lstm_value, - frame_size, cur_batch_size, - gate_act, cell_act, cand_act); - lstm_value.prevStateValue = lstm_value.stateValue; + lstm_value.gate_value = gate_t.data(); + lstm_value.output_value = out_t.data(); + lstm_value.state_value = cell_t.data(); + lstm_value.state_active_value = cell_pre_act_t.data(); + math::LstmUnitFunctor::compute( + device_ctx, lstm_value, frame_size, cur_batch_size, gate_act, + cell_act, cand_act); + lstm_value.prev_state_value = lstm_value.state_value; } - math::Batch2LoDTensorFunctor to_seq; + 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); @@ -161,7 +165,7 @@ class LSTMKernel : public framework::OpKernel { } }; -template +template class LSTMGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { @@ -187,8 +191,8 @@ class LSTMGradKernel : public framework::OpKernel { auto* h0_g = ctx.Output(framework::GradVarName("H0")); auto* c0_g = ctx.Output(framework::GradVarName("C0")); - auto& device_ctx = ctx.device_context(); - math::SetConstant zero; + auto& device_ctx = ctx.template device_context(); + math::SetConstant zero; if (weight_g) { weight_g->mutable_data(ctx.GetPlace()); zero(device_ctx, weight_g, static_cast(0.0)); @@ -200,7 +204,8 @@ class LSTMGradKernel : public framework::OpKernel { Tensor ordered_h0, ordered_c0, ordered_h0_g, ordered_c0_g; const size_t* order = batch_gate->lod()[2].data(); if (c0) { - ReorderInitState(device_ctx, *c0, order, &ordered_c0, true); + ReorderInitState(device_ctx, *c0, order, &ordered_c0, + true); } if (c0 && c0_g) { ordered_c0_g.mutable_data(c0_g->dims(), ctx.GetPlace()); @@ -214,13 +219,13 @@ class LSTMGradKernel : public framework::OpKernel { math::LstmMetaValue lstm_value; if (bias && ctx.Attr("use_peepholes")) { T* bias_data = const_cast(bias->data()); - lstm_value.checkIg = bias_data + 4 * frame_size; - lstm_value.checkFg = lstm_value.checkIg + frame_size; - lstm_value.checkOg = lstm_value.checkFg + frame_size; + lstm_value.check_ig = bias_data + 4 * frame_size; + lstm_value.check_fg = lstm_value.check_ig + frame_size; + lstm_value.check_og = lstm_value.check_fg + frame_size; } else { - lstm_value.checkIg = nullptr; - lstm_value.checkFg = nullptr; - lstm_value.checkOg = nullptr; + lstm_value.check_ig = nullptr; + lstm_value.check_fg = nullptr; + lstm_value.check_og = nullptr; } math::LstmMetaGrad lstm_grad; @@ -231,19 +236,19 @@ class LSTMGradKernel : public framework::OpKernel { } if (bias && bias_g && ctx.Attr("use_peepholes")) { T* bias_g_data = bias_g->data(); - lstm_grad.checkIgGrad = bias_g_data + 4 * frame_size; - lstm_grad.checkFgGrad = lstm_grad.checkIgGrad + frame_size; - lstm_grad.checkOgGrad = lstm_grad.checkFgGrad + frame_size; + lstm_grad.check_ig_grad = bias_g_data + 4 * frame_size; + lstm_grad.check_fg_grad = lstm_grad.check_ig_grad + frame_size; + lstm_grad.check_og_grad = lstm_grad.check_fg_grad + frame_size; } else { - lstm_grad.checkIgGrad = nullptr; - lstm_grad.checkFgGrad = nullptr; - lstm_grad.checkOgGrad = nullptr; + lstm_grad.check_ig_grad = nullptr; + lstm_grad.check_fg_grad = nullptr; + lstm_grad.check_og_grad = nullptr; } - math::LoDTensor2BatchFunctor to_batch; + math::LoDTensor2BatchFunctor to_batch; auto ToBatch = [&batch_gate, &to_batch]( - const platform::DeviceContext& ctx, const framework::LoDTensor& src, + const DeviceContext& ctx, const framework::LoDTensor& src, const framework::DDim& dims, framework::LoDTensor& dst) { dst.mutable_data(dims, ctx.GetPlace()); dst.set_lod(batch_gate->lod()); @@ -263,9 +268,12 @@ class LSTMGradKernel : public framework::OpKernel { batch_gate_g.mutable_data(batch_gate->dims(), ctx.GetPlace()); batch_gate_g.set_lod(batch_gate->lod()); - auto gate_act = ctx.Attr("gate_activation"); - auto cell_act = ctx.Attr("cell_activation"); - auto cand_act = ctx.Attr("candidate_activation"); + auto gate_act = math::detail::GetActivationType( + ctx.Attr("gate_activation")); + auto cell_act = math::detail::GetActivationType( + ctx.Attr("cell_activation")); + auto cand_act = math::detail::GetActivationType( + ctx.Attr("candidate_activation")); auto batch_starts = batch_gate->lod()[0]; size_t num_batch = batch_starts.size() - 1; @@ -276,30 +284,30 @@ class LSTMGradKernel : public framework::OpKernel { Tensor gate = batch_gate->Slice(bstart, bend); Tensor cell = batch_cell.Slice(bstart, bend); Tensor cell_pre_act = batch_cell_pre_act->Slice(bstart, bend); - lstm_value.gateValue = gate.data(); - lstm_value.stateValue = cell.data(); - lstm_value.stateActiveValue = cell_pre_act.data(); + lstm_value.gate_value = gate.data(); + lstm_value.state_value = cell.data(); + lstm_value.state_active_value = cell_pre_act.data(); Tensor out_g = batch_hidden_g.Slice(bstart, bend); Tensor gate_g = batch_gate_g.Slice(bstart, bend); Tensor cell_g = batch_cell_g.Slice(bstart, bend); - lstm_grad.stateGrad = cell_g.data(); - lstm_grad.gateGrad = gate_g.data(); - lstm_grad.outputGrad = out_g.data(); + lstm_grad.state_grad = cell_g.data(); + lstm_grad.gate_grad = gate_g.data(); + lstm_grad.output_grad = out_g.data(); if (n > 0) { int bstart_pre = static_cast(batch_starts[n - 1]); Tensor cell_pre = batch_cell.Slice(bstart_pre, bstart); Tensor cell_pre_g = batch_cell_g.Slice(bstart_pre, bstart); - lstm_value.prevStateValue = cell_pre.data(); - lstm_grad.prevStateGrad = cell_pre_g.data(); + lstm_value.prev_state_value = cell_pre.data(); + lstm_grad.prev_state_grad = cell_pre_g.data(); } else { - lstm_value.prevStateValue = c0 ? ordered_c0.data() : nullptr; - lstm_grad.prevStateGrad = c0_g ? ordered_c0_g.data() : nullptr; + lstm_value.prev_state_value = c0 ? ordered_c0.data() : nullptr; + lstm_grad.prev_state_grad = c0_g ? ordered_c0_g.data() : nullptr; } int cur_batch_size = bend - bstart; - math::LstmUnitGradFunctor::compute( + math::LstmUnitGradFunctor::compute( device_ctx, lstm_value, lstm_grad, frame_size, cur_batch_size, gate_act, cell_act, cand_act); @@ -307,33 +315,34 @@ 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)); + math::matmul(device_ctx, 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)); + math::matmul(device_ctx, 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)); + 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)); } 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)); + math::matmul(device_ctx, gate_g, false, *weight, + true, static_cast(1.0), + &ordered_h0_g, static_cast(0.0)); } } } - math::Batch2LoDTensorFunctor to_seq; + math::Batch2LoDTensorFunctor to_seq; if (in_g) { /* backward data */ in_g->mutable_data(ctx.GetPlace()); @@ -344,15 +353,17 @@ class LSTMGradKernel : public framework::OpKernel { Tensor b_g = *bias_g; b_g.Resize({bias_g->numel(), 1}); Tensor gate_bias_g = b_g.Slice(0, 4 * frame_size); - math::ColwiseSum col_sum; + math::ColwiseSum col_sum; col_sum(device_ctx, batch_gate_g, &gate_bias_g); } if (h0 && h0_g) { - ReorderInitState(device_ctx, ordered_h0_g, order, h0_g, false); + ReorderInitState(device_ctx, ordered_h0_g, order, h0_g, + false); } if (c0 && c0_g) { - ReorderInitState(device_ctx, ordered_c0_g, order, c0_g, false); + ReorderInitState(device_ctx, ordered_c0_g, order, c0_g, + false); } } }; diff --git a/paddle/operators/lstm_unit_op.cc b/paddle/operators/lstm_unit_op.cc index 18b9cdf2a39e8226c634194ff2cc56d169979774..c2d2c43982580c9724849d68576d42ffa44fc6b4 100644 --- a/paddle/operators/lstm_unit_op.cc +++ b/paddle/operators/lstm_unit_op.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/operators/lstm_unit_op.h" @@ -48,10 +48,12 @@ class LstmUnitOp : public framework::OperatorWithKernel { class LstmUnitOpMaker : public framework::OpProtoAndCheckerMaker { public: - LstmUnitOpMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker) + LstmUnitOpMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("X", "FC input before the non-linear activation."); + AddInput("X", + "Lstm unit only applies non-linear activations, please make sure" + "that linear tranformation has already been applied to `X`. " + "Linear tranformation can be applied by adding a `fc` layer"); AddInput( "C_prev", "The cell state tensor of last time-step in the Lstm Unit operator."); diff --git a/paddle/operators/lstm_unit_op.cu b/paddle/operators/lstm_unit_op.cu index e192283aa0afac49e8e467506f3703d1ce60d2a6..5ee5ddd280f7720c4583053e4e48a5043ab423f4 100644 --- a/paddle/operators/lstm_unit_op.cu +++ b/paddle/operators/lstm_unit_op.cu @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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. */ /* Acknowledgement: the following code is strongly inspired by https://github.com/caffe2/caffe2/blob/master/caffe2/operators/lstm_unit_op_gpu.cu @@ -98,7 +98,7 @@ class LstmUnitOpCUDAKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), - "It must use GPUPlace."); + "It must use CUDAPlace."); auto* x_tensor = ctx.Input("X"); auto* c_prev_tensor = ctx.Input("C_prev"); @@ -129,7 +129,7 @@ class LstmUnitGradOpCUDAKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), - "It must use GPUPlace."); + "It must use CUDAPlace."); auto x_tensor = ctx.Input("X"); auto c_prev_tensor = ctx.Input("C_prev"); @@ -173,7 +173,7 @@ class LstmUnitGradOpCUDAKernel : public framework::OpKernel { } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(lstm_unit, ops::LstmUnitOpCUDAKernel, - ops::LstmUnitOpCUDAKernel); -REGISTER_OP_GPU_KERNEL(lstm_unit_grad, ops::LstmUnitGradOpCUDAKernel, - ops::LstmUnitGradOpCUDAKernel); +REGISTER_OP_CUDA_KERNEL(lstm_unit, ops::LstmUnitOpCUDAKernel, + ops::LstmUnitOpCUDAKernel); +REGISTER_OP_CUDA_KERNEL(lstm_unit_grad, ops::LstmUnitGradOpCUDAKernel, + ops::LstmUnitGradOpCUDAKernel); diff --git a/paddle/operators/lstm_unit_op.h b/paddle/operators/lstm_unit_op.h index 38cb298f92a21bb5c7508761fec701d28279a85f..fa8d141bcb6ee4bfc9a29e337b7adbc5ecd3ad23 100644 --- a/paddle/operators/lstm_unit_op.h +++ b/paddle/operators/lstm_unit_op.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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. */ /* Acknowledgement: the following code is strongly inspired by https://github.com/caffe2/caffe2/blob/master/caffe2/operators/lstm_unit_op.h @@ -35,7 +35,7 @@ inline T tanh(T x) { return 2. * sigmoid(2. * x) - 1.; } -template +template class LstmUnitKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { @@ -78,7 +78,7 @@ class LstmUnitKernel : public framework::OpKernel { } }; -template +template class LstmUnitGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { diff --git a/paddle/operators/margin_rank_loss_op.cc b/paddle/operators/margin_rank_loss_op.cc index d7e8a0ea7632650203106b01531d724cf0b8e085..e0df3077742bc330ce8510bf06b0411148f669d8 100644 --- a/paddle/operators/margin_rank_loss_op.cc +++ b/paddle/operators/margin_rank_loss_op.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/operators/margin_rank_loss_op.h" @@ -42,8 +42,7 @@ class MarginRankLossOp : public framework::OperatorWithKernel { template class MarginRankLossOpMaker : public framework::OpProtoAndCheckerMaker { public: - MarginRankLossOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + MarginRankLossOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X1", "(2-D tensor with shape [batch_size x 1]) The score for " @@ -117,7 +116,7 @@ REGISTER_OP(margin_rank_loss, ops::MarginRankLossOp, ops::MarginRankLossGradOp); REGISTER_OP_CPU_KERNEL( margin_rank_loss, - ops::MarginRankLossKernel); + ops::MarginRankLossKernel); REGISTER_OP_CPU_KERNEL( margin_rank_loss_grad, - ops::MarginRankLossGradKernel); + ops::MarginRankLossGradKernel); diff --git a/paddle/operators/margin_rank_loss_op.cu b/paddle/operators/margin_rank_loss_op.cu index 3a639f25d478a712c1030d57c57d7e55de1488b5..798c3ed182b08b07a779da88924bfc05743c680e 100644 --- a/paddle/operators/margin_rank_loss_op.cu +++ b/paddle/operators/margin_rank_loss_op.cu @@ -1,24 +1,24 @@ /* 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 +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 + 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. */ +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/operators/margin_rank_loss_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL( +REGISTER_OP_CUDA_KERNEL( margin_rank_loss, - ops::MarginRankLossKernel); -REGISTER_OP_GPU_KERNEL( + ops::MarginRankLossKernel); +REGISTER_OP_CUDA_KERNEL( margin_rank_loss_grad, - ops::MarginRankLossGradKernel); + ops::MarginRankLossGradKernel); diff --git a/paddle/operators/margin_rank_loss_op.h b/paddle/operators/margin_rank_loss_op.h index 8d0830147ecc465909e8988e90125929829f6f34..7438e881e1c69c9ef1f84b21e6ee0ba093f3378a 100644 --- a/paddle/operators/margin_rank_loss_op.h +++ b/paddle/operators/margin_rank_loss_op.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 @@ -34,7 +34,7 @@ struct Heaviside { } }; -template +template class MarginRankLossKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const { @@ -56,13 +56,13 @@ class MarginRankLossKernel : public framework::OpKernel { auto x1 = framework::EigenVector::Flatten(*x1_t); auto x2 = framework::EigenVector::Flatten(*x2_t); - auto& dev = ctx.GetEigenDevice(); + auto& dev = *ctx.template device_context().eigen_device(); out.device(dev) = (-label * (x1 - x2) + margin).unaryExpr(ReLU()); act.device(dev) = out.unaryExpr(Heaviside()); } }; -template +template class MarginRankLossGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const { @@ -78,7 +78,7 @@ class MarginRankLossGradKernel : public framework::OpKernel { auto d_out = framework::EigenVector::Flatten(*d_out_t); auto act = framework::EigenVector::Flatten(*act_t); auto label = framework::EigenVector::Flatten(*label_t); - auto& dev = ctx.GetEigenDevice(); + auto& dev = *ctx.template device_context().eigen_device(); // compute d_x1 if (d_x1_t) { diff --git a/paddle/operators/math/CMakeLists.txt b/paddle/operators/math/CMakeLists.txt index 002b68fecf4f1e294387357f0346d9926a2b2b5a..7ebcfb9ab9f30e3b0f13d3646a59d008335b232d 100644 --- a/paddle/operators/math/CMakeLists.txt +++ b/paddle/operators/math/CMakeLists.txt @@ -9,11 +9,14 @@ if(WITH_GPU) nv_library(cross_entropy SRCS cross_entropy.cc cross_entropy.cu DEPS device_context) nv_library(pooling SRCS pooling.cc pooling.cu DEPS device_context) nv_library(sequence_pooling SRCS sequence_pooling.cc sequence_pooling.cu DEPS device_context math_function) - nv_library(vol2col SRCS vol2col.cc vol2col.cu DEPS device_context) + nv_library(vol2col SRCS vol2col.cc vol2col.cu DEPS device_context tensor) nv_library(context_project SRCS context_project.cc context_project.cu DEPS device_context math_function) - nv_library(sequence2batch SRCS sequence2batch.cc sequence2batch.cu DEPS device_context) + nv_library(sequence2batch SRCS sequence2batch.cc sequence2batch.cu DEPS device_context tensor) nv_library(lstm_compute SRCS lstm_compute.cc lstm_compute.cu DEPS device_context activation_functions) + nv_library(maxouting SRCS maxouting.cc maxouting.cu DEPS device_context) + nv_library(unpooling SRCS unpooling.cc unpooling.cu DEPS device_context) nv_library(gru_compute SRCS gru_compute.cc gru_compute.cu DEPS device_context activation_functions math_function) + nv_library(cos_sim_functor SRCS cos_sim_functor.cc cos_sim_functor.cu DEPS device_context) else() cc_library(math_function SRCS math_function.cc im2col.cc DEPS cblas device_context framework_proto) cc_library(selected_rows_functor SRCS selected_rows_functor.cc DEPS selected_rows math_function) @@ -21,11 +24,14 @@ else() cc_library(cross_entropy SRCS cross_entropy.cc DEPS device_context) cc_library(pooling SRCS pooling.cc DEPS device_context) cc_library(sequence_pooling SRCS sequence_pooling.cc DEPS device_context math_function) - cc_library(vol2col SRCS vol2col.cc DEPS device_context) + cc_library(vol2col SRCS vol2col.cc DEPS device_context tensor) cc_library(context_project SRCS context_project.cc DEPS device_context math_function) - cc_library(sequence2batch SRCS sequence2batch.cc DEPS device_context) + cc_library(sequence2batch SRCS sequence2batch.cc DEPS device_context tensor) cc_library(lstm_compute SRCS lstm_compute.cc DEPS device_context activation_functions) + cc_library(maxouting SRCS maxouting.cc DEPS device_context) + cc_library(unpooling SRCS unpooling.cc DEPS device_context) cc_library(gru_compute SRCS gru_compute.cc DEPS device_context activation_functions math_function) + cc_library(cos_sim_functor SRCS cos_sim_functor.cc DEPS device_context) endif() cc_test(math_function_test SRCS math_function_test.cc DEPS math_function tensor) diff --git a/paddle/operators/math/context_project.cc b/paddle/operators/math/context_project.cc index f82ea5d7bee81fd1578c46f79477bb23939e627a..980dd90df8710cdbcb760e1ca84f1492a76fdb70 100644 --- a/paddle/operators/math/context_project.cc +++ b/paddle/operators/math/context_project.cc @@ -18,8 +18,8 @@ namespace paddle { namespace operators { namespace math { -template class ContextProjectFunctor; -template class ContextProjectFunctor; +template class ContextProjectFunctor; +template class ContextProjectFunctor; } // namespace math } // namespace operators diff --git a/paddle/operators/math/context_project.cu b/paddle/operators/math/context_project.cu index 04eeed543cb165fe449d3578a951cf74b0422252..934e3df645916013b4d1fe5eb4a19be924c914d5 100644 --- a/paddle/operators/math/context_project.cu +++ b/paddle/operators/math/context_project.cu @@ -20,8 +20,8 @@ namespace paddle { namespace operators { namespace math { -template class ContextProjectFunctor; -template class ContextProjectFunctor; +template class ContextProjectFunctor; +template class ContextProjectFunctor; } // namespace math } // namespace operators diff --git a/paddle/operators/math/context_project.h b/paddle/operators/math/context_project.h index 72f4202bace4461d2597204feaa2a21e355bd1ac..4036614086e1eb724a4a647db6ef13b6fe7aaaa0 100644 --- a/paddle/operators/math/context_project.h +++ b/paddle/operators/math/context_project.h @@ -81,17 +81,17 @@ using LoDTensor = framework::LoDTensor; * */ -template +template class ContextProjectFunctor { public: - void operator()(const platform::DeviceContext& context, const LoDTensor& in, + void operator()(const DeviceContext& context, const LoDTensor& in, const Tensor& padding_data, bool padding_trainable, const int context_start, const int context_length, const int context_stride, const int up_pad, const int down_pad, Tensor* col) { auto lod_level_0 = in.lod()[0]; - math::Im2ColFunctor im2col_ocf; + math::Im2ColFunctor im2col_ocf; std::vector dilation({1, 1}); std::vector padding({up_pad, 0, down_pad, 0}); @@ -149,7 +149,7 @@ class ContextProjectFunctor { 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); - out_t_sub.CopyFrom(w_sub, context.GetPlace(), context); + framework::CopyFrom(w_sub, context.GetPlace(), context, &out_t_sub); } } if (down_pad > 0) { // add down pad @@ -179,7 +179,7 @@ class ContextProjectFunctor { (down_pad_begin_row + t) * context_length); Tensor w_sub = padding_data.Slice( up_pad + padding_idx, up_pad + padding_idx + padding_size); - out_t_sub.CopyFrom(w_sub, context.GetPlace(), context); + framework::CopyFrom(w_sub, context.GetPlace(), context, &out_t_sub); } } out_t.Resize({sequence_height, context_length * sequence_width}); @@ -188,17 +188,17 @@ class ContextProjectFunctor { } }; -template +template class ContextProjectGradFunctor { public: - void operator()(const platform::DeviceContext& context, const LoDTensor& in, + void operator()(const DeviceContext& context, const LoDTensor& in, bool padding_trainable, const int context_start, const int context_length, const int context_stride, const int up_pad, const int down_pad, bool pad_grad, bool input_grad, Tensor* padding_data, Tensor* col) { auto lod_level_0 = in.lod()[0]; - math::Col2ImFunctor col2im_ocf; + math::Col2ImFunctor col2im_ocf; std::vector dilation({1, 1}); std::vector padding({up_pad, 0, down_pad, 0}); @@ -258,8 +258,8 @@ 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()); + axpy(context, w_sub.numel(), static_cast(1), + out_t_sub.data(), w_sub.data()); } } if (down_pad > 0) { @@ -290,8 +290,8 @@ 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()); + axpy(context, 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/operators/math/cos_sim_functor.cc b/paddle/operators/math/cos_sim_functor.cc new file mode 100644 index 0000000000000000000000000000000000000000..6af9f0fcd967b4e8e9e338c155d5a8ee2866fbfa --- /dev/null +++ b/paddle/operators/math/cos_sim_functor.cc @@ -0,0 +1,48 @@ +/* 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. */ + +#include "paddle/operators/math/cos_sim_functor.h" + +namespace paddle { +namespace operators { +namespace math { + +template +struct CosSimDyFunctor { + void operator()(const platform::CPUDeviceContext& ctx, const T* x_norm, + const T* y_norm, const T* x, const T* y, const T* z, + const T* dz, const size_t rows, const size_t cols, + T* dy) const { + for (size_t row_id = 0; row_id < rows; ++row_id) { + auto xy_norm_prod = x_norm[row_id] * y_norm[0]; + auto dz_data = dz[row_id]; + auto z_data = z[row_id]; + auto* x_data = x + cols * row_id; + auto reciprocal_xy_norm_prod = 1 / xy_norm_prod; + + auto y_norm_square = y_norm[0] * y_norm[0]; + auto reciprocal_y_norm_square = 1 / y_norm_square; + for (size_t i = 0; i < cols; ++i) { + dy[i] += dz_data * (x_data[i] * reciprocal_xy_norm_prod - + z_data * y[i] * reciprocal_y_norm_square); + } + } + } +}; + +template struct CosSimDyFunctor; +template struct CosSimDyFunctor; +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/math/cos_sim_functor.cu b/paddle/operators/math/cos_sim_functor.cu new file mode 100644 index 0000000000000000000000000000000000000000..6eb0a4ea4c5b86f84c93b97615255adf55e9e042 --- /dev/null +++ b/paddle/operators/math/cos_sim_functor.cu @@ -0,0 +1,64 @@ +/* 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. */ + +#include "paddle/operators/math/cos_sim_functor.h" +#include "paddle/platform/cuda_helper.h" + +namespace paddle { +namespace operators { +namespace math { + +template +__global__ void CosSimDyKernel(const T* x_norm, const T* y_norm, const T* x, + const T* y, const T* z, const T* dz, + const size_t rows, const size_t cols, T* dy) { + int grid_size = blockDim.x * gridDim.x; + T y_norm_data = y_norm[0]; + for (int row_id = blockIdx.x * blockDim.x + threadIdx.x; row_id < rows; + row_id += grid_size) { + T xy_norm_prod = x_norm[row_id] * y_norm_data; + T dz_data = dz[row_id]; + T z_data = z[row_id]; + const T* x_data = x + cols * row_id; + T reciprocal_xy_norm_prod = 1 / xy_norm_prod; + + T y_norm_square = y_norm_data * y_norm_data; + T reciprocal_y_norm_square = 1 / y_norm_square; + for (size_t i = 0; i < cols; ++i) { + T dy_data = dz_data * (x_data[i] * reciprocal_xy_norm_prod - + z_data * y[i] * reciprocal_y_norm_square); + platform::CudaAtomicAdd(dy + i, dy_data); + } + } +} + +template +struct CosSimDyFunctor { + void operator()(const platform::CUDADeviceContext& ctx, const T* x_norm, + const T* y_norm, const T* x, const T* y, const T* z, + const T* dz, const size_t rows, const size_t cols, + T* dy) const { + const int block_size = 512; + dim3 threads(block_size, 1); + dim3 grid(1, (rows + block_size - 1) / block_size); + CosSimDyKernel<<>>( + x_norm, y_norm, x, y, z, dz, rows, cols, dy); + } +}; + +template struct CosSimDyFunctor; +template struct CosSimDyFunctor; +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/math/cos_sim_functor.h b/paddle/operators/math/cos_sim_functor.h new file mode 100644 index 0000000000000000000000000000000000000000..aae8ab5b7a937c016e8a45e34b22aba7a1df3066 --- /dev/null +++ b/paddle/operators/math/cos_sim_functor.h @@ -0,0 +1,166 @@ +/* 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 +#include +#include +#include "paddle/platform/device_context.h" +#include "paddle/platform/hostdevice.h" + +namespace paddle { +namespace operators { +namespace math { + +template +struct CosSimFunctor { + CosSimFunctor(const T* x, const T* y, T* x_norm, T* y_norm, T* z, int cols) + : x_norm_(x_norm), + y_norm_(y_norm), + x_(x), + y_(y), + z_(z), + cols_(static_cast(cols)) {} + + inline HOSTDEVICE void operator()(size_t row_id) const { + auto* x = x_ + cols_ * row_id; + T xx = 0, xy = 0, yy = 0; + if (same_row) { + auto* y = y_ + cols_ * row_id; + T tep_x, tep_y; + for (size_t i = 0; i < cols_; ++i) { + tep_x = x[i]; + tep_y = y[i]; + xx += tep_x * tep_x; + yy += tep_y * tep_y; + xy += tep_x * tep_y; + } + xx = sqrt(xx); + yy = sqrt(yy); + y_norm_[row_id] = yy; + x_norm_[row_id] = xx; + z_[row_id] = xy / (xx * yy); + } else { // This can be wrote in a better way. + T tep_x, tep_y; + for (size_t i = 0; i < cols_; ++i) { + tep_x = x[i]; + tep_y = y_[i]; + xx += tep_x * tep_x; + yy += tep_y * tep_y; + xy += tep_x * tep_y; + } + xx = sqrt(xx); + yy = sqrt(yy); + if (row_id == 0) y_norm_[0] = yy; + x_norm_[row_id] = xx; + z_[row_id] = xy / (xx * yy); + } + } + + T* x_norm_; + T* y_norm_; + const T* x_; + const T* y_; + T* z_; + const size_t cols_; +}; + +template +struct CosSimGradFunctor { + CosSimGradFunctor(const T* x_norm, const T* y_norm, const T* x, const T* y, + const T* z, const T* dz, T* dx, int cols) + : x_norm_(x_norm), + y_norm_(y_norm), + x_(x), + y_(y), + z_(z), + dz_(dz), + dx_(dx), + cols_(static_cast(cols)) {} + + inline HOSTDEVICE void operator()(size_t row_id) const { + auto x_norm_square = x_norm_[row_id] * x_norm_[row_id]; + auto xy_norm_prod = x_norm_[row_id] * y_norm_[row_id]; + auto dz = dz_[row_id]; + auto z = z_[row_id]; + + auto* dx = dx_ + cols_ * row_id; + auto* x = x_ + cols_ * row_id; + auto* y = y_ + cols_ * row_id; + + auto reciprocal_xy_norm_prod = 1 / xy_norm_prod; + auto reciprocal_x_norm_square = 1 / x_norm_square; + for (size_t i = 0; i < cols_; ++i) { + dx[i] = dz * (y[i] * reciprocal_xy_norm_prod - + z * x[i] * reciprocal_x_norm_square); + } + } + + const T* x_norm_; + const T* y_norm_; + const T* x_; + const T* y_; + const T* z_; + const T* dz_; + T* dx_; + const size_t cols_; +}; + +template +struct CosSimDxFunctor { + CosSimDxFunctor(const T* x_norm, const T* y_norm, const T* x, const T* y, + const T* z, const T* dz, T* dx, int cols) + : x_norm_(x_norm), + y_norm_(y_norm), + x_(x), + y_(y), + z_(z), + dz_(dz), + dx_(dx), + cols_(static_cast(cols)) {} + + inline HOSTDEVICE void operator()(size_t row_id) const { + auto xy_norm_prod = x_norm_[row_id] * y_norm_[0]; + auto dz = dz_[row_id]; + auto z = z_[row_id]; + auto* x = x_ + cols_ * row_id; + auto reciprocal_xy_norm_prod = 1 / xy_norm_prod; + auto x_norm_square = x_norm_[row_id] * x_norm_[row_id]; + auto* dx = dx_ + cols_ * row_id; + auto reciprocal_x_norm_square = 1 / x_norm_square; + + for (size_t i = 0; i < cols_; ++i) { + dx[i] = dz * (y_[i] * reciprocal_xy_norm_prod - + z * x[i] * reciprocal_x_norm_square); + } + } + const T* x_norm_; + const T* y_norm_; + const T* x_; + const T* y_; + const T* z_; + const T* dz_; + T* dx_; + const size_t cols_; +}; + +template +struct CosSimDyFunctor { + void operator()(const DeviceContext& ctx, const T* x_norm, const T* y_norm, + const T* x, const T* y, const T* z, const T* dz, + const size_t rows, const size_t cols, T* dy) const; +}; + +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/math/cross_entropy.cc b/paddle/operators/math/cross_entropy.cc index cf238a58e0a0b930077b0376a71dc02c5b31efe5..d9cb016fb440b6b2fe1d222812215feb5970dc4f 100644 --- a/paddle/operators/math/cross_entropy.cc +++ b/paddle/operators/math/cross_entropy.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/operators/math/cross_entropy.h" @@ -24,9 +24,9 @@ template ; template -class CrossEntropyFunctor { +class CrossEntropyFunctor { public: - void operator()(const platform::DeviceContext& ctx, framework::Tensor* out, + void operator()(const platform::CPUDeviceContext& ctx, framework::Tensor* out, const framework::Tensor* prob, const framework::Tensor* labels, const bool softLabel) { const int batch_size = prob->dims()[0]; @@ -35,7 +35,7 @@ class CrossEntropyFunctor { auto lbl = EigenMatrix::From(*labels); auto loss = EigenMatrix::From(*out); - loss.device(*ctx.GetEigenDevice()) = + loss.device(*ctx.eigen_device()) = -((lbl * in.log().unaryExpr(math::TolerableValue())) .sum(Eigen::DSizes(1)) .reshape(Eigen::DSizes(batch_size, 1))); @@ -53,8 +53,8 @@ class CrossEntropyFunctor { } }; -template class CrossEntropyFunctor; -template class CrossEntropyFunctor; +template class CrossEntropyFunctor; +template class CrossEntropyFunctor; } // namespace math } // namespace operators } // namespace paddle diff --git a/paddle/operators/math/cross_entropy.cu b/paddle/operators/math/cross_entropy.cu index 651c08f740c2991b11c210c9bf012e505adc1835..16c9e7b28ec8d453492455c8d620ba9edf130a07 100644 --- a/paddle/operators/math/cross_entropy.cu +++ b/paddle/operators/math/cross_entropy.cu @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/operators/math/cross_entropy.h" @@ -95,10 +95,10 @@ __global__ void SoftCrossEntropyKernel(T* Y, const T* X, const T* label, using Tensor = framework::Tensor; template -class CrossEntropyFunctor { +class CrossEntropyFunctor { public: - void operator()(const platform::DeviceContext& ctx, framework::Tensor* out, - const framework::Tensor* prob, + void operator()(const platform::CUDADeviceContext& ctx, + framework::Tensor* out, const framework::Tensor* prob, const framework::Tensor* labels, bool softLabel) { const T* prob_data = prob->data(); T* loss_data = out->mutable_data(ctx.GetPlace()); @@ -118,16 +118,14 @@ class CrossEntropyFunctor { const int64_t* label_data = labels->data(); int block = 512; int grid = (batch_size + block - 1) / block; - CrossEntropyKernel<<< - grid, block, 0, - reinterpret_cast(ctx).stream()>>>( + CrossEntropyKernel<<>>( loss_data, prob_data, label_data, batch_size, class_num); } } }; -template class CrossEntropyFunctor; -template class CrossEntropyFunctor; +template class CrossEntropyFunctor; +template class CrossEntropyFunctor; } // namespace math } // namespace operators } // namespace paddle diff --git a/paddle/operators/math/cross_entropy.h b/paddle/operators/math/cross_entropy.h index 70ed9ddd551bb8cb7989727c02fea870186c9f2e..b3b6d767a8b8f59e3c75e72ac6c98653a8e1c3a4 100644 --- a/paddle/operators/math/cross_entropy.h +++ b/paddle/operators/math/cross_entropy.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/framework/eigen.h" @@ -33,11 +33,11 @@ struct TolerableValue { } }; -template +template class CrossEntropyFunctor { public: - void operator()(const platform::DeviceContext& context, - framework::Tensor* out, const framework::Tensor* prob, + void operator()(const DeviceContext& context, framework::Tensor* out, + const framework::Tensor* prob, const framework::Tensor* labels, const bool softLabel); }; } // namespace math diff --git a/paddle/operators/math/detail/activation_functions.h b/paddle/operators/math/detail/activation_functions.h index a20c35d1d9dc4a3a6fae92023fd1aae787a716ec..585a0123437a39c2b610306b18fe0a970c0ed072 100644 --- a/paddle/operators/math/detail/activation_functions.h +++ b/paddle/operators/math/detail/activation_functions.h @@ -14,6 +14,7 @@ limitations under the License. */ #pragma once #include +#include "paddle/platform/enforce.h" #include "paddle/platform/hostdevice.h" #ifdef __AVX__ @@ -29,6 +30,26 @@ namespace detail { #define SIGMOID_THRESHOLD_MAX 13.0 #define EXP_MAX_INPUT 40.0 +enum ActivationType { + kSigmoid, + kReLU, + kTanh, + kIdentity, +}; + +inline ActivationType GetActivationType(const std::string &type) { + if (type == "sigmoid") { + return ActivationType::kSigmoid; + } else if (type == "relu") { + return ActivationType::kReLU; + } else if (type == "tanh") { + return ActivationType::kTanh; + } else if (type == "identity" || type == "") { + return ActivationType::kIdentity; + } + PADDLE_THROW("Not support type %s.", type); +} + namespace forward { template diff --git a/paddle/operators/math/detail/gru_cpu_kernel.h b/paddle/operators/math/detail/gru_cpu_kernel.h index 51af140cf4d5e6581765bea00033fa53d383230d..a61b232f4275d93cae1d9a71d49a779216c3555b 100644 --- a/paddle/operators/math/detail/gru_cpu_kernel.h +++ b/paddle/operators/math/detail/gru_cpu_kernel.h @@ -25,393 +25,395 @@ namespace detail { #ifndef __NVCC__ template -void hl_naive_gru_forward_reset_output(OpResetOutput opResetOutput, - T *gateValue, T *resetOutputValue, - T *prevOutputValue, int frameSize, - activation_mode_t active_gate) { - T rValueUpdateGate; - T rValueResetGate; - T rValueResetOutput; - T rPrevOut = 0; - T *updateGate = gateValue; - T *resetGate = gateValue + frameSize; - - for (int i = 0; i < frameSize; i++) { - rValueUpdateGate = updateGate[i]; - rValueResetGate = resetGate[i]; - if (prevOutputValue) { - rPrevOut = prevOutputValue[i]; +void hl_naive_gru_forward_reset_output(OpResetOutput op_reset_output, + T *gate_value, T *reset_output_value, + T *prev_output_value, int frame_size, + ActivationType active_gate) { + T r_value_update_gate; + T r_value_reset_gate; + T r_value_reset_output; + T r_prev_out = 0; + T *update_gate = gate_value; + T *reset_gate = gate_value + frame_size; + + for (int i = 0; i < frame_size; i++) { + r_value_update_gate = update_gate[i]; + r_value_reset_gate = reset_gate[i]; + if (prev_output_value) { + r_prev_out = prev_output_value[i]; } - opResetOutput(rValueUpdateGate, rValueResetGate, rPrevOut, - rValueResetOutput, active_gate); + op_reset_output(r_value_update_gate, r_value_reset_gate, r_prev_out, + r_value_reset_output, active_gate); - updateGate[i] = rValueUpdateGate; - resetGate[i] = rValueResetGate; - resetOutputValue[i] = rValueResetOutput; + update_gate[i] = r_value_update_gate; + reset_gate[i] = r_value_reset_gate; + reset_output_value[i] = r_value_reset_output; } } template -void hl_naive_gru_forward_final_output(OpFinalOutput opFinalOutput, - T *gateValue, T *prevOutputValue, - T *outputValue, int frameSize, - activation_mode_t active_node) { - T rValueUpdateGate; - T rValueFrameState; - T rPrevOut = 0; - T rOutput; - T *updateGate = gateValue; - T *frameState = gateValue + frameSize * 2; - - for (int i = 0; i < frameSize; i++) { - rValueUpdateGate = updateGate[i]; - rValueFrameState = frameState[i]; - if (prevOutputValue) { - rPrevOut = prevOutputValue[i]; +void hl_naive_gru_forward_final_output(OpFinalOutput op_final_output, + T *gate_value, T *prev_output_value, + T *output_value, int frame_size, + ActivationType active_node) { + T r_value_update_gate; + T r_value_frame_state; + T r_prev_out = 0; + T r_output; + T *update_gate = gate_value; + T *frame_state = gate_value + frame_size * 2; + + for (int i = 0; i < frame_size; i++) { + r_value_update_gate = update_gate[i]; + r_value_frame_state = frame_state[i]; + if (prev_output_value) { + r_prev_out = prev_output_value[i]; } - opFinalOutput(rValueUpdateGate, rValueFrameState, rPrevOut, rOutput, - active_node); + op_final_output(r_value_update_gate, r_value_frame_state, r_prev_out, + r_output, active_node); - frameState[i] = rValueFrameState; - outputValue[i] = rOutput; + frame_state[i] = r_value_frame_state; + output_value[i] = r_output; } } template -void hl_avx_gru_forward_reset_output(OpResetOutput opResetOutput, T *gateValue, - T *resetOutputValue, T *prevOutputValue, - int frameSize, - activation_mode_t active_gate) { +void hl_avx_gru_forward_reset_output(OpResetOutput op_reset_output, + T *gate_value, T *reset_output_value, + T *prev_output_value, int frame_size, + ActivationType active_gate) { #ifdef __AVX__ - __m256 rValueUpdateGate; - __m256 rValueResetGate; - __m256 rValueResetOutput; - __m256 rPrevOut = _mm256_set1_ps(0.0f); - __m256 *updateGate = (__m256 *)gateValue; - __m256 *resetGate = (__m256 *)(gateValue + frameSize); - - for (int i = 0; i < frameSize / 8; i++) { - rValueUpdateGate = updateGate[i]; - rValueResetGate = resetGate[i]; - if (prevOutputValue) { - rPrevOut = ((__m256 *)prevOutputValue)[i]; + __m256 r_value_update_gate; + __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); + + 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]; } - opResetOutput(rValueUpdateGate, rValueResetGate, rPrevOut, - rValueResetOutput, active_gate); + op_reset_output(r_value_update_gate, r_value_reset_gate, r_prev_out, + r_value_reset_output, active_gate); - updateGate[i] = rValueUpdateGate; - resetGate[i] = rValueResetGate; - ((__m256 *)resetOutputValue)[i] = rValueResetOutput; + update_gate[i] = r_value_update_gate; + reset_gate[i] = r_value_reset_gate; + ((__m256 *)reset_output_value)[i] = r_value_reset_output; } #endif } template -void hl_avx_gru_forward_final_output(OpFinalOutput opFinalOutput, T *gateValue, - T *prevOutputValue, T *outputValue, - int frameSize, - activation_mode_t active_node) { +void hl_avx_gru_forward_final_output(OpFinalOutput op_final_output, + T *gate_value, T *prev_output_value, + T *output_value, int frame_size, + ActivationType active_node) { #ifdef __AVX__ - __m256 rValueUpdateGate; - __m256 rValueFrameState; - __m256 rPrevOut = _mm256_set1_ps(0.0f); - __m256 rOutput; - __m256 *updateGate = (__m256 *)gateValue; - __m256 *frameState = (__m256 *)(gateValue + frameSize * 2); - - for (int i = 0; i < frameSize / 8; i++) { - rValueUpdateGate = updateGate[i]; - rValueFrameState = frameState[i]; - if (prevOutputValue) { - rPrevOut = ((__m256 *)prevOutputValue)[i]; + __m256 r_value_update_gate; + __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); + + 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]; } - opFinalOutput(rValueUpdateGate, rValueFrameState, rPrevOut, rOutput, - active_node); + op_final_output(r_value_update_gate, r_value_frame_state, r_prev_out, + r_output, active_node); - frameState[i] = rValueFrameState; - ((__m256 *)outputValue)[i] = rOutput; + frame_state[i] = r_value_frame_state; + ((__m256 *)output_value)[i] = r_output; } #endif } template -inline void forward_reset_output(OpResetOutput opResetOutput, - hl_gru_value value, int frameSize, - int batchSize, activation_mode_t active_gate) { - for (int b = 0; b < batchSize; b++) { - if (OpResetOutput::avx && !(frameSize & (8 - 1)) && (sizeof(T) == 4)) { +inline void forward_reset_output(OpResetOutput op_reset_output, + GRUMetaValue value, int frame_size, + int batch_size, ActivationType active_gate) { + for (int b = 0; b < batch_size; b++) { + if (OpResetOutput::avx && !(frame_size & (8 - 1)) && (sizeof(T) == 4)) { hl_avx_gru_forward_reset_output( - opResetOutput, value.gateValue, value.resetOutputValue, - value.prevOutValue, frameSize, active_gate); + op_reset_output, value.gate_value, value.reset_output_value, + value.prev_out_value, frame_size, active_gate); } else { hl_naive_gru_forward_reset_output( - opResetOutput, value.gateValue, value.resetOutputValue, - value.prevOutValue, frameSize, active_gate); + op_reset_output, value.gate_value, value.reset_output_value, + value.prev_out_value, frame_size, active_gate); } - value.gateValue += frameSize * 3; - value.resetOutputValue += frameSize; - if (value.prevOutValue) { - value.prevOutValue += frameSize; + value.gate_value += frame_size * 3; + value.reset_output_value += frame_size; + if (value.prev_out_value) { + value.prev_out_value += frame_size; } } } template -inline void forward_final_output(OpFinalOutput opFinalOutput, - hl_gru_value value, int frameSize, - int batchSize, activation_mode_t active_node) { - for (int b = 0; b < batchSize; b++) { - if (OpFinalOutput::avx && !(frameSize & (8 - 1)) && (sizeof(T) == 4)) { - hl_avx_gru_forward_final_output(opFinalOutput, value.gateValue, - value.prevOutValue, value.outputValue, - frameSize, active_node); +inline void forward_final_output(OpFinalOutput op_final_output, + GRUMetaValue value, int frame_size, + int batch_size, ActivationType active_node) { + for (int b = 0; b < batch_size; b++) { + if (OpFinalOutput::avx && !(frame_size & (8 - 1)) && (sizeof(T) == 4)) { + hl_avx_gru_forward_final_output(op_final_output, value.gate_value, + value.prev_out_value, value.output_value, + frame_size, active_node); } else { - hl_naive_gru_forward_final_output(opFinalOutput, value.gateValue, - value.prevOutValue, value.outputValue, - frameSize, active_node); + hl_naive_gru_forward_final_output( + op_final_output, value.gate_value, value.prev_out_value, + value.output_value, frame_size, active_node); } - value.gateValue += frameSize * 3; - value.outputValue += frameSize; - if (value.prevOutValue) { - value.prevOutValue += frameSize; + value.gate_value += frame_size * 3; + value.output_value += frame_size; + if (value.prev_out_value) { + value.prev_out_value += frame_size; } } } template -void hl_naive_gru_backward_state_grad(OpStateGrad opStateGrad, T *gateValue, - T *gateGrad, T *prevOutValue, - T *prevOutGrad, T *outputGrad, - int frameSize, - activation_mode_t active_node) { - T rUpdateGateValue; - T rUpdateGateGrad; - T rFrameStateValue; - T rFrameStateGrad; - T rOutGrad; - T rPrevOutValue = 0; - T rPrevOutGrad = 0; - T *updateGateValue = gateValue; - T *updateGateGrad = gateGrad; - T *frameStateValue = gateValue + frameSize * 2; - T *frameStateGrad = gateGrad + frameSize * 2; - - for (int i = 0; i < frameSize; i++) { - rUpdateGateValue = updateGateValue[i]; - rFrameStateValue = frameStateValue[i]; - rOutGrad = outputGrad[i]; - if (prevOutValue) { - rPrevOutValue = prevOutValue[i]; +void hl_naive_gru_backward_state_grad(OpStateGrad op_state_grad, T *gate_value, + T *gate_grad, T *prev_out_value, + T *prev_out_grad, T *output_grad, + int frame_size, + ActivationType active_node) { + T r_update_gate_value; + T r_update_gate_grad; + T r_frame_state_value; + T r_frame_state_grad; + T r_out_grad; + T r_prev_out_value = 0; + T r_prev_out_grad = 0; + T *update_gate_value = gate_value; + T *update_gate_grad = gate_grad; + T *frame_state_value = gate_value + frame_size * 2; + T *frame_state_grad = gate_grad + frame_size * 2; + + for (int i = 0; i < frame_size; i++) { + r_update_gate_value = update_gate_value[i]; + r_frame_state_value = frame_state_value[i]; + r_out_grad = output_grad[i]; + if (prev_out_value) { + r_prev_out_value = prev_out_value[i]; } - if (prevOutGrad) { - rPrevOutGrad = prevOutGrad[i]; + if (prev_out_grad) { + r_prev_out_grad = prev_out_grad[i]; } - opStateGrad(rUpdateGateValue, rUpdateGateGrad, rFrameStateValue, - rFrameStateGrad, rPrevOutValue, rPrevOutGrad, rOutGrad, - 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); - updateGateGrad[i] = rUpdateGateGrad; - frameStateGrad[i] = rFrameStateGrad; - if (prevOutGrad) { - prevOutGrad[i] = rPrevOutGrad; + update_gate_grad[i] = r_update_gate_grad; + frame_state_grad[i] = r_frame_state_grad; + if (prev_out_grad) { + prev_out_grad[i] = r_prev_out_grad; } } } template -void hl_naive_gru_backward_reset_grad(OpResetGrad opResetGrad, T *gateValue, - T *gateGrad, T *prevOutValue, - T *prevOutGrad, T *resetOutputGrad, - int frameSize, - activation_mode_t active_gate) { - T rUpdateGateValue; - T rUpdateGateGrad; - T rResetGateValue; - T rResetGateGrad; - T rResetOutputGrad = 0; - T rPrevOutValue = 0; - T rPrevOutGrad = 0; - T *updateGateValue = gateValue; - T *updateGateGrad = gateGrad; - T *resetGateValue = gateValue + frameSize; - T *resetGateGrad = gateGrad + frameSize; - - for (int i = 0; i < frameSize; i++) { - rUpdateGateValue = updateGateValue[i]; - rUpdateGateGrad = updateGateGrad[i]; - rResetGateValue = resetGateValue[i]; - - if (prevOutValue && prevOutGrad) { - rResetOutputGrad = resetOutputGrad[i]; +void hl_naive_gru_backward_reset_grad(OpResetGrad op_reset_grad, T *gate_value, + T *gate_grad, T *prev_out_value, + T *prev_out_grad, T *reset_output_grad, + int frame_size, + ActivationType active_gate) { + T r_update_gate_value; + T r_update_gate_grad; + T r_reset_gate_value; + T r_reset_gate_grad; + T r_reset_output_grad = 0; + T r_prev_out_value = 0; + T r_prev_out_grad = 0; + T *update_gate_value = gate_value; + T *update_gate_grad = gate_grad; + T *reset_gate_value = gate_value + frame_size; + T *reset_gate_grad = gate_grad + frame_size; + + for (int i = 0; i < frame_size; i++) { + r_update_gate_value = update_gate_value[i]; + r_update_gate_grad = update_gate_grad[i]; + r_reset_gate_value = reset_gate_value[i]; + + if (prev_out_value && prev_out_grad) { + r_reset_output_grad = reset_output_grad[i]; } - if (prevOutValue) { - rPrevOutValue = prevOutValue[i]; + if (prev_out_value) { + r_prev_out_value = prev_out_value[i]; } - if (prevOutGrad) { - rPrevOutGrad = prevOutGrad[i]; + if (prev_out_grad) { + r_prev_out_grad = prev_out_grad[i]; } - opResetGrad(rUpdateGateValue, rUpdateGateGrad, rResetGateValue, - rResetGateGrad, rPrevOutValue, rPrevOutGrad, rResetOutputGrad, - 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); - updateGateGrad[i] = rUpdateGateGrad; - resetGateGrad[i] = rResetGateGrad; - if (prevOutGrad) { - prevOutGrad[i] = rPrevOutGrad; + update_gate_grad[i] = r_update_gate_grad; + reset_gate_grad[i] = r_reset_gate_grad; + if (prev_out_grad) { + prev_out_grad[i] = r_prev_out_grad; } } } template -void hl_avx_gru_backward_state_grad(OpStateGrad opStateGrad, T *gateValue, - T *gateGrad, T *prevOutValue, - T *prevOutGrad, T *outputGrad, - int frameSize, - activation_mode_t active_node) { +void hl_avx_gru_backward_state_grad(OpStateGrad op_state_grad, T *gate_value, + T *gate_grad, T *prev_out_value, + T *prev_out_grad, T *output_grad, + int frame_size, + ActivationType active_node) { #ifdef __AVX__ - __m256 rUpdateGateValue; - __m256 rUpdateGateGrad; - __m256 rFrameStateValue; - __m256 rFrameStateGrad; - __m256 rOutGrad; - __m256 rPrevOutValue = _mm256_set1_ps(0.0f); - __m256 rPrevOutGrad = _mm256_set1_ps(0.0f); - __m256 *updateGateValue = (__m256 *)gateValue; - __m256 *updateGateGrad = (__m256 *)gateGrad; - __m256 *frameStateValue = (__m256 *)(gateValue + frameSize * 2); - __m256 *frameStateGrad = (__m256 *)(gateGrad + frameSize * 2); - - for (int i = 0; i < frameSize / 8; i++) { - rUpdateGateValue = updateGateValue[i]; - rFrameStateValue = frameStateValue[i]; - rOutGrad = ((__m256 *)outputGrad)[i]; - if (prevOutValue) { - rPrevOutValue = ((__m256 *)prevOutValue)[i]; + __m256 r_update_gate_value; + __m256 r_update_gate_grad; + __m256 r_frame_state_value; + __m256 r_frame_state_grad; + __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); + + 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]; + if (prev_out_value) { + r_prev_out_value = ((__m256 *)prev_out_value)[i]; } - if (prevOutGrad) { - rPrevOutGrad = ((__m256 *)prevOutGrad)[i]; + if (prev_out_grad) { + r_prev_out_grad = ((__m256 *)prev_out_grad)[i]; } - opStateGrad(rUpdateGateValue, rUpdateGateGrad, rFrameStateValue, - rFrameStateGrad, rPrevOutValue, rPrevOutGrad, rOutGrad, - 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); - updateGateGrad[i] = rUpdateGateGrad; - frameStateGrad[i] = rFrameStateGrad; - if (prevOutGrad) { - ((__m256 *)prevOutGrad)[i] = rPrevOutGrad; + 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; } } #endif } template -void hl_avx_gru_backward_reset_grad(OpResetGrad opResetGrad, T *gateValue, - T *gateGrad, T *prevOutValue, - T *prevOutGrad, T *resetOutputGrad, - int frameSize, - activation_mode_t active_gate) { +void hl_avx_gru_backward_reset_grad(OpResetGrad op_reset_grad, T *gate_value, + T *gate_grad, T *prev_out_value, + T *prev_out_grad, T *reset_output_grad, + int frame_size, + ActivationType active_gate) { #ifdef __AVX__ - __m256 rUpdateGateValue; - __m256 rUpdateGateGrad; - __m256 rResetGateValue; - __m256 rResetGateGrad; - __m256 rResetOutputGrad = _mm256_set1_ps(0.0f); - __m256 rPrevOutValue = _mm256_set1_ps(0.0f); - __m256 rPrevOutGrad = _mm256_set1_ps(0.0f); - __m256 *updateGateValue = (__m256 *)gateValue; - __m256 *updateGateGrad = (__m256 *)gateGrad; - __m256 *resetGateValue = (__m256 *)(gateValue + frameSize); - __m256 *resetGateGrad = (__m256 *)(gateGrad + frameSize); - - for (int i = 0; i < frameSize / 8; i++) { - rUpdateGateValue = updateGateValue[i]; - rUpdateGateGrad = updateGateGrad[i]; - rResetGateValue = resetGateValue[i]; - - if (prevOutValue && prevOutGrad) { - rResetOutputGrad = ((__m256 *)resetOutputGrad)[i]; + __m256 r_update_gate_value; + __m256 r_update_gate_grad; + __m256 r_reset_gate_value; + __m256 r_reset_gate_grad; + __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); + + for (int i = 0; i < frame_size / 8; i++) { + r_update_gate_value = update_gate_value[i]; + r_update_gate_grad = update_gate_grad[i]; + r_reset_gate_value = reset_gate_value[i]; + + if (prev_out_value && prev_out_grad) { + r_reset_output_grad = ((__m256 *)reset_output_grad)[i]; } - if (prevOutValue) { - rPrevOutValue = ((__m256 *)prevOutValue)[i]; + if (prev_out_value) { + r_prev_out_value = ((__m256 *)prev_out_value)[i]; } - if (prevOutGrad) { - rPrevOutGrad = ((__m256 *)prevOutGrad)[i]; + if (prev_out_grad) { + r_prev_out_grad = ((__m256 *)prev_out_grad)[i]; } - opResetGrad(rUpdateGateValue, rUpdateGateGrad, rResetGateValue, - rResetGateGrad, rPrevOutValue, rPrevOutGrad, rResetOutputGrad, - 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); - updateGateGrad[i] = rUpdateGateGrad; - resetGateGrad[i] = rResetGateGrad; - if (prevOutGrad) { - ((__m256 *)prevOutGrad)[i] = rPrevOutGrad; + 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; } } #endif } template -inline void backward_state_grad(OpStateGrad opStateGrad, hl_gru_value value, - hl_gru_grad grad, int frameSize, - int batchSize, activation_mode_t active_node) { - for (int b = 0; b < batchSize; b++) { - if (OpStateGrad::avx && !(frameSize & (8 - 1)) && (sizeof(T) == 4)) { +inline void backward_state_grad(OpStateGrad op_state_grad, + GRUMetaValue value, GRUMetaGrad grad, + int frame_size, int batch_size, + ActivationType active_node) { + for (int b = 0; b < batch_size; b++) { + if (OpStateGrad::avx && !(frame_size & (8 - 1)) && (sizeof(T) == 4)) { hl_avx_gru_backward_state_grad( - opStateGrad, value.gateValue, grad.gateGrad, value.prevOutValue, - grad.prevOutGrad, grad.outputGrad, frameSize, active_node); + op_state_grad, value.gate_value, grad.gate_grad, value.prev_out_value, + grad.prev_out_grad, grad.output_grad, frame_size, active_node); } else { hl_naive_gru_backward_state_grad( - opStateGrad, value.gateValue, grad.gateGrad, value.prevOutValue, - grad.prevOutGrad, grad.outputGrad, frameSize, active_node); + op_state_grad, value.gate_value, grad.gate_grad, value.prev_out_value, + grad.prev_out_grad, grad.output_grad, frame_size, active_node); } - value.gateValue += frameSize * 3; - if (value.prevOutValue) { - value.prevOutValue += frameSize; + value.gate_value += frame_size * 3; + if (value.prev_out_value) { + value.prev_out_value += frame_size; } - grad.gateGrad += frameSize * 3; - grad.outputGrad += frameSize; - if (grad.prevOutGrad) { - grad.prevOutGrad += frameSize; + grad.gate_grad += frame_size * 3; + grad.output_grad += frame_size; + if (grad.prev_out_grad) { + grad.prev_out_grad += frame_size; } } } template -inline void backward_reset_grad(OpResetGrad opResetGrad, hl_gru_value value, - hl_gru_grad grad, int frameSize, - int batchSize, activation_mode_t active_gate) { - for (int b = 0; b < batchSize; b++) { - if (OpResetGrad::avx && !(frameSize & (8 - 1)) && (sizeof(T) == 4)) { +inline void backward_reset_grad(OpResetGrad op_reset_grad, + GRUMetaValue value, GRUMetaGrad grad, + int frame_size, int batch_size, + ActivationType active_gate) { + for (int b = 0; b < batch_size; b++) { + if (OpResetGrad::avx && !(frame_size & (8 - 1)) && (sizeof(T) == 4)) { hl_avx_gru_backward_reset_grad( - opResetGrad, value.gateValue, grad.gateGrad, value.prevOutValue, - grad.prevOutGrad, grad.resetOutputGrad, frameSize, active_gate); + op_reset_grad, value.gate_value, grad.gate_grad, value.prev_out_value, + grad.prev_out_grad, grad.reset_output_grad, frame_size, active_gate); } else { hl_naive_gru_backward_reset_grad( - opResetGrad, value.gateValue, grad.gateGrad, value.prevOutValue, - grad.prevOutGrad, grad.resetOutputGrad, frameSize, active_gate); + op_reset_grad, value.gate_value, grad.gate_grad, value.prev_out_value, + grad.prev_out_grad, grad.reset_output_grad, frame_size, active_gate); } - value.gateValue += frameSize * 3; - if (value.prevOutValue) { - value.prevOutValue += frameSize; + value.gate_value += frame_size * 3; + if (value.prev_out_value) { + value.prev_out_value += frame_size; } - grad.gateGrad += frameSize * 3; - grad.resetOutputGrad += frameSize; - if (grad.prevOutGrad) { - grad.prevOutGrad += frameSize; + grad.gate_grad += frame_size * 3; + grad.reset_output_grad += frame_size; + if (grad.prev_out_grad) { + grad.prev_out_grad += frame_size; } } } diff --git a/paddle/operators/math/detail/gru_gpu_kernel.h b/paddle/operators/math/detail/gru_gpu_kernel.h index 6441c648b048422c110872a85aa8cb719f11a8d7..1783d46096858c874b27ce75760342082835b180 100644 --- a/paddle/operators/math/detail/gru_gpu_kernel.h +++ b/paddle/operators/math/detail/gru_gpu_kernel.h @@ -19,182 +19,180 @@ limitations under the License. */ #include "paddle/platform/cuda_helper.h" #include "paddle/platform/device_context.h" -#include - namespace paddle { namespace operators { namespace math { namespace detail { /* - * threads(framePerBlock, batchPerBlock) - * grid(frameBlocks, batchBlocks) + * threads(frame_per_block, batch_per_block) + * grid(frame_blocks, batch_blocks) */ -template -__global__ void KeGruForwardResetOutput(OpResetOutput opResetOutput, - T *gateValue, T *resetOutputValue, - T *prevOutputValue, int frameSize, - int batchSize, - activation_mode_t active_gate) { - const int frameIdx = blockIdx.x * blockDim.x + threadIdx.x; - if (frameIdx >= frameSize) return; - - int batchIdx = 0; - if (isBatch) { - batchIdx = blockIdx.y * blockDim.y + threadIdx.y; - if (batchIdx >= batchSize) return; - gateValue += batchIdx * 3 * frameSize; - resetOutputValue += batchIdx * frameSize; +template +__global__ void KeGruForwardResetOutput(OpResetOutput op_reset_output, + T *gate_value, T *reset_output_value, + T *prev_output_value, int frame_size, + int batch_size, + ActivationType active_gate) { + const int frame_idx = blockIdx.x * blockDim.x + threadIdx.x; + if (frame_idx >= frame_size) return; + + int batch_idx = 0; + if (is_batch) { + batch_idx = blockIdx.y * blockDim.y + threadIdx.y; + if (batch_idx >= batch_size) return; + gate_value += batch_idx * 3 * frame_size; + reset_output_value += batch_idx * frame_size; } - T rPrevOut = 0; - T rValueResetOutput; - T rValueUpdateGate = gateValue[frameIdx + frameSize * 0]; - T rValueResetGate = gateValue[frameIdx + frameSize * 1]; + T r_prev_out = 0; + T r_value_reset_output; + T r_value_update_gate = gate_value[frame_idx + frame_size * 0]; + T r_value_reset_gate = gate_value[frame_idx + frame_size * 1]; - if (prevOutputValue) { - if (isBatch) prevOutputValue += batchIdx * frameSize; - rPrevOut = prevOutputValue[frameIdx]; + if (prev_output_value) { + if (is_batch) prev_output_value += batch_idx * frame_size; + r_prev_out = prev_output_value[frame_idx]; } - opResetOutput(rValueUpdateGate, rValueResetGate, rPrevOut, rValueResetOutput, - active_gate); + op_reset_output(r_value_update_gate, r_value_reset_gate, r_prev_out, + r_value_reset_output, active_gate); - gateValue[frameIdx + frameSize * 0] = rValueUpdateGate; - gateValue[frameIdx + frameSize * 1] = rValueResetGate; - resetOutputValue[frameIdx] = rValueResetOutput; + gate_value[frame_idx + frame_size * 0] = r_value_update_gate; + gate_value[frame_idx + frame_size * 1] = r_value_reset_gate; + reset_output_value[frame_idx] = r_value_reset_output; } /* - * threads(framePerBlock, batchPerBlock) - * grid(frameBlocks, batchBlocks) + * threads(frame_per_block, batch_per_block) + * grid(frame_blocks, batch_blocks) */ -template -__global__ void KeGruForwardFinalOutput(OpFinalOutput opFinalOutput, - T *gateValue, T *prevOutputValue, - T *outputValue, int frameSize, - int batchSize, - activation_mode_t active_node) { - const int frameIdx = blockIdx.x * blockDim.x + threadIdx.x; - if (frameIdx >= frameSize) return; - int batchIdx = 0; - if (isBatch) { - batchIdx = blockIdx.y * blockDim.y + threadIdx.y; - if (batchIdx >= batchSize) return; - gateValue += batchIdx * 3 * frameSize; - outputValue += batchIdx * frameSize; +template +__global__ void KeGruForwardFinalOutput(OpFinalOutput op_final_output, + T *gate_value, T *prev_output_value, + T *output_value, int frame_size, + int batch_size, + ActivationType active_node) { + const int frame_idx = blockIdx.x * blockDim.x + threadIdx.x; + if (frame_idx >= frame_size) return; + int batch_idx = 0; + if (is_batch) { + batch_idx = blockIdx.y * blockDim.y + threadIdx.y; + if (batch_idx >= batch_size) return; + gate_value += batch_idx * 3 * frame_size; + output_value += batch_idx * frame_size; } - T rOutput; - T rPrevOut = 0; - T rValueUpdateGate = gateValue[frameIdx + frameSize * 0]; - T rValueFrameState = gateValue[frameIdx + frameSize * 2]; + T r_output; + T r_prev_out = 0; + T r_value_update_gate = gate_value[frame_idx + frame_size * 0]; + T r_value_frame_state = gate_value[frame_idx + frame_size * 2]; - if (prevOutputValue) { - if (isBatch) prevOutputValue += batchIdx * frameSize; - rPrevOut = prevOutputValue[frameIdx]; + if (prev_output_value) { + if (is_batch) prev_output_value += batch_idx * frame_size; + r_prev_out = prev_output_value[frame_idx]; } - opFinalOutput(rValueUpdateGate, rValueFrameState, rPrevOut, rOutput, - active_node); + op_final_output(r_value_update_gate, r_value_frame_state, r_prev_out, + r_output, active_node); - gateValue[frameIdx + frameSize * 2] = rValueFrameState; - outputValue[frameIdx] = rOutput; + gate_value[frame_idx + frame_size * 2] = r_value_frame_state; + output_value[frame_idx] = r_output; } /* - * threads(framePerBlock, batchPerBlock) - * grid(frameBlocks, batchBlocks) + * threads(frame_per_block, batch_per_block) + * grid(frame_blocks, batch_blocks) */ -template -__global__ void KeGruBackwardStateGrad(OpStateGrad opStateGrad, T *gateValue, - T *gateGrad, T *prevOutValue, - T *prevOutGrad, T *outputGrad, - int frameSize, int batchSize, - activation_mode_t active_node) { - const int frameIdx = blockIdx.x * blockDim.x + threadIdx.x; - if (frameIdx >= frameSize) return; - int batchIdx = 0; - if (isBatch) { - batchIdx = blockIdx.y * blockDim.y + threadIdx.y; - if (batchIdx >= batchSize) return; - gateValue += batchIdx * 3 * frameSize; - gateGrad += batchIdx * 3 * frameSize; - outputGrad += batchIdx * frameSize; +template +__global__ void KeGruBackwardStateGrad(OpStateGrad op_state_grad, T *gate_value, + T *gate_grad, T *prev_out_value, + T *prev_out_grad, T *output_grad, + int frame_size, int batch_size, + ActivationType active_node) { + const int frame_idx = blockIdx.x * blockDim.x + threadIdx.x; + if (frame_idx >= frame_size) return; + int batch_idx = 0; + if (is_batch) { + batch_idx = blockIdx.y * blockDim.y + threadIdx.y; + if (batch_idx >= batch_size) return; + gate_value += batch_idx * 3 * frame_size; + gate_grad += batch_idx * 3 * frame_size; + output_grad += batch_idx * frame_size; } - T rUpdateGateGrad; - T rFrameStateGrad; - T rPrevOutValue = 0; - T rPrevOutGrad = 0; - T rUpdateGateValue = gateValue[frameIdx + frameSize * 0]; - T rFrameStateValue = gateValue[frameIdx + frameSize * 2]; - T rOutGrad = outputGrad[frameIdx]; + T r_update_gate_grad; + T r_frame_state_grad; + T r_prev_out_value = 0; + T r_prev_out_grad = 0; + T r_update_gate_value = gate_value[frame_idx + frame_size * 0]; + T r_frame_state_value = gate_value[frame_idx + frame_size * 2]; + T r_out_grad = output_grad[frame_idx]; - if (prevOutValue && prevOutGrad) { - if (isBatch) prevOutValue += batchIdx * frameSize; - rPrevOutValue = prevOutValue[frameIdx]; + if (prev_out_value && prev_out_grad) { + if (is_batch) prev_out_value += batch_idx * frame_size; + r_prev_out_value = prev_out_value[frame_idx]; - if (isBatch) prevOutGrad += batchIdx * frameSize; - rPrevOutGrad = prevOutGrad[frameIdx]; + if (is_batch) prev_out_grad += batch_idx * frame_size; + r_prev_out_grad = prev_out_grad[frame_idx]; } - opStateGrad(rUpdateGateValue, rUpdateGateGrad, rFrameStateValue, - rFrameStateGrad, rPrevOutValue, rPrevOutGrad, rOutGrad, - 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); - gateGrad[frameIdx + frameSize * 0] = rUpdateGateGrad; - gateGrad[frameIdx + frameSize * 2] = rFrameStateGrad; - if (prevOutGrad) { - prevOutGrad[frameIdx] = rPrevOutGrad; + gate_grad[frame_idx + frame_size * 0] = r_update_gate_grad; + gate_grad[frame_idx + frame_size * 2] = r_frame_state_grad; + if (prev_out_grad) { + prev_out_grad[frame_idx] = r_prev_out_grad; } } /* - * threads(framePerBlock, batchPerBlock) - * grid(frameBlocks, batchBlocks) + * threads(frame_per_block, batch_per_block) + * grid(frame_blocks, batch_blocks) */ -template -__global__ void KeGruBackwardResetGrad(OpResetGrad opResetGrad, T *gateValue, - T *gateGrad, T *prevOutValue, - T *prevOutGrad, T *resetOutputGrad, - int frameSize, int batchSize, - activation_mode_t active_gate) { - const int frameIdx = blockIdx.x * blockDim.x + threadIdx.x; - if (frameIdx >= frameSize) return; - int batchIdx = 0; - if (isBatch) { - batchIdx = blockIdx.y * blockDim.y + threadIdx.y; - if (batchIdx >= batchSize) return; - gateValue += batchIdx * 3 * frameSize; - gateGrad += batchIdx * 3 * frameSize; - resetOutputGrad += batchIdx * frameSize; +template +__global__ void KeGruBackwardResetGrad(OpResetGrad op_reset_grad, T *gate_value, + T *gate_grad, T *prev_out_value, + T *prev_out_grad, T *reset_output_grad, + int frame_size, int batch_size, + ActivationType active_gate) { + const int frame_idx = blockIdx.x * blockDim.x + threadIdx.x; + if (frame_idx >= frame_size) return; + int batch_idx = 0; + if (is_batch) { + batch_idx = blockIdx.y * blockDim.y + threadIdx.y; + if (batch_idx >= batch_size) return; + gate_value += batch_idx * 3 * frame_size; + gate_grad += batch_idx * 3 * frame_size; + reset_output_grad += batch_idx * frame_size; } - T rResetGateGrad; - T rPrevOutValue = 0; - T rPrevOutGrad = 0; - T rResetOutputGrad = 0; - T rUpdateGateValue = gateValue[frameIdx + frameSize * 0]; - T rUpdateGateGrad = gateGrad[frameIdx + frameSize * 0]; - T rResetGateValue = gateValue[frameIdx + frameSize * 1]; - - if (prevOutValue && prevOutGrad) { - if (isBatch) prevOutValue += batchIdx * frameSize; - if (isBatch) prevOutGrad += batchIdx * frameSize; - rPrevOutValue = prevOutValue[frameIdx]; - rPrevOutGrad = prevOutGrad[frameIdx]; - rResetOutputGrad = resetOutputGrad[frameIdx]; + T r_reset_gate_grad; + T r_prev_out_value = 0; + T r_prev_out_grad = 0; + T r_reset_output_grad = 0; + T r_update_gate_value = gate_value[frame_idx + frame_size * 0]; + T r_update_gate_grad = gate_grad[frame_idx + frame_size * 0]; + T r_reset_gate_value = gate_value[frame_idx + frame_size * 1]; + + if (prev_out_value && prev_out_grad) { + if (is_batch) prev_out_value += batch_idx * frame_size; + if (is_batch) prev_out_grad += batch_idx * frame_size; + r_prev_out_value = prev_out_value[frame_idx]; + r_prev_out_grad = prev_out_grad[frame_idx]; + r_reset_output_grad = reset_output_grad[frame_idx]; } - opResetGrad(rUpdateGateValue, rUpdateGateGrad, rResetGateValue, - rResetGateGrad, rPrevOutValue, rPrevOutGrad, rResetOutputGrad, - 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); - gateGrad[frameIdx + frameSize * 0] = rUpdateGateGrad; - gateGrad[frameIdx + frameSize * 1] = rResetGateGrad; - if (prevOutGrad) { - prevOutGrad[frameIdx] = rPrevOutGrad; + gate_grad[frame_idx + frame_size * 0] = r_update_gate_grad; + gate_grad[frame_idx + frame_size * 1] = r_reset_gate_grad; + if (prev_out_grad) { + prev_out_grad[frame_idx] = r_prev_out_grad; } } } // namespace detail diff --git a/paddle/operators/math/detail/gru_kernel.h b/paddle/operators/math/detail/gru_kernel.h index 8a681d8d8bced72e1296f863489f6ccbc7913167..4d8245cb5d03b33edbda5d8350be02b4fa87ab95 100644 --- a/paddle/operators/math/detail/gru_kernel.h +++ b/paddle/operators/math/detail/gru_kernel.h @@ -28,23 +28,25 @@ namespace forward { template class gru_resetOutput { public: - HOSTDEVICE void operator()(T &valueUpdateGate, T &valueResetGate, T &prevOut, - T &valueResetOutput, activation_mode_t actGate) { - valueUpdateGate = activation(valueUpdateGate, actGate); - valueResetGate = activation(valueResetGate, actGate); - valueResetOutput = prevOut * valueResetGate; + 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; } #ifndef __NVCC__ #ifndef __AVX__ static const bool avx = false; #else static const bool avx = true; - HOSTDEVICE void operator()(__m256 &valueUpdateGate, __m256 &valueResetGate, - __m256 &prevOut, __m256 &valueResetOutput, - activation_mode_t actGate) { - valueUpdateGate = activation(valueUpdateGate, actGate); - valueResetGate = activation(valueResetGate, actGate); - valueResetOutput = _mm256_mul_ps(prevOut, valueResetGate); + 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); } #endif #endif @@ -53,24 +55,25 @@ class gru_resetOutput { template class gru_finalOutput { public: - HOSTDEVICE void operator()(T &valueUpdateGate, T &valueFrameState, T &prevOut, - T &valueOutput, activation_mode_t actInput) { - valueFrameState = activation(valueFrameState, actInput); - valueOutput = prevOut - (valueUpdateGate * prevOut) + - (valueUpdateGate * valueFrameState); + 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); } #ifndef __NVCC__ #ifndef __AVX__ static const bool avx = false; #else static const bool avx = true; - HOSTDEVICE void operator()(__m256 &valueUpdateGate, __m256 &valueFrameState, - __m256 &prevOut, __m256 &valueOutput, - activation_mode_t actInput) { - valueFrameState = activation(valueFrameState, actInput); - valueOutput = _mm256_add_ps( - _mm256_sub_ps(prevOut, _mm256_mul_ps(valueUpdateGate, prevOut)), - _mm256_mul_ps(valueUpdateGate, valueFrameState)); + 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 @@ -82,34 +85,37 @@ namespace backward { template class gru_stateGrad { public: - HOSTDEVICE void operator()(T &valueUpdateGate, T &gradUpdateGate, - T &valueFrameState, T &gradFrameState, - T &valuePrevOut, T &gradPrevOut, T &gradOutput, - activation_mode_t actInput) { - gradUpdateGate = (gradOutput * valueFrameState); - gradUpdateGate -= (gradOutput * valuePrevOut); - gradPrevOut -= (gradOutput * valueUpdateGate); - gradPrevOut += gradOutput; - gradFrameState = - activation(gradOutput * valueUpdateGate, valueFrameState, actInput); + 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 &valueUpdateGate, __m256 &gradUpdateGate, - __m256 &valueFrameState, __m256 &gradFrameState, - __m256 &valuePrevOut, __m256 &gradPrevOut, - __m256 &gradOutput, activation_mode_t actInput) { - gradUpdateGate = _mm256_mul_ps(gradOutput, valueFrameState); - gradUpdateGate = - _mm256_sub_ps(gradUpdateGate, _mm256_mul_ps(gradOutput, valuePrevOut)); - gradPrevOut = _mm256_add_ps( - _mm256_sub_ps(gradPrevOut, _mm256_mul_ps(gradOutput, valueUpdateGate)), - gradOutput); - gradFrameState = activation(_mm256_mul_ps(gradOutput, valueUpdateGate), - valueFrameState, actInput); + 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); } #endif #endif @@ -118,30 +124,32 @@ class gru_stateGrad { template class gru_resetGrad { public: - HOSTDEVICE void operator()(T &valueUpdateGate, T &gradUpdateGate, - T &valueResetGate, T &gradResetGate, - T &valuePrevOut, T &gradPrevOut, - T &gradResetOutput, activation_mode_t actGate) { - gradResetGate = (gradResetOutput * valuePrevOut); - gradPrevOut += (gradResetOutput * valueResetGate); - gradUpdateGate = activation(gradUpdateGate, valueUpdateGate, actGate); - gradResetGate = activation(gradResetGate, valueResetGate, actGate); + 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 &valueUpdateGate, __m256 &gradUpdateGate, - __m256 &valueResetGate, __m256 &gradResetGate, - __m256 &valuePrevOut, __m256 &gradPrevOut, - __m256 &gradResetOutput, - activation_mode_t actGate) { - gradResetGate = _mm256_mul_ps(gradResetOutput, valuePrevOut); - gradPrevOut = _mm256_add_ps(gradPrevOut, - _mm256_mul_ps(gradResetOutput, valueResetGate)); - gradUpdateGate = activation(gradUpdateGate, valueUpdateGate, actGate); - gradResetGate = activation(gradResetGate, valueResetGate, actGate); + 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); } #endif #endif diff --git a/paddle/operators/math/detail/lstm_cpu_kernel.h b/paddle/operators/math/detail/lstm_cpu_kernel.h index fc3ad0ce58aa1552ef7e717fb529c2d454b4895a..42888fcdb0a464892e3007ee73c195fcd2a431bb 100644 --- a/paddle/operators/math/detail/lstm_cpu_kernel.h +++ b/paddle/operators/math/detail/lstm_cpu_kernel.h @@ -26,278 +26,281 @@ namespace detail { template void naive_lstm_forward_one_sequence(Op op, LstmMetaValue value, - int frameSize, - activation_mode_t active_node, - activation_mode_t active_gate, - activation_mode_t active_state) { - T rValueIn; - T rValueIg; - T rValueFg; - T rValueOg; - T rCheckI; - T rCheckF; - T rCheckO; - T rState; - T rPrevState = 0; - T rStateAtv; - T rOut; - - T *valueIn = value.gateValue; - T *valueIg = value.gateValue + frameSize; - T *valueFg = value.gateValue + frameSize * 2; - T *valueOg = value.gateValue + frameSize * 3; - - for (int i = 0; i < frameSize; i++) { - rValueIn = valueIn[i]; - rValueIg = valueIg[i]; - rValueFg = valueFg[i]; - rValueOg = valueOg[i]; - rCheckI = value.checkIg ? value.checkIg[i] : 0; - rCheckF = value.checkFg ? value.checkFg[i] : 0; - rCheckO = value.checkOg ? value.checkOg[i] : 0; - - if (value.prevStateValue) { - rPrevState = value.prevStateValue[i]; + int frame_size, ActivationType active_node, + ActivationType active_gate, + ActivationType active_state) { + T r_value_in; + T r_value_ig; + T r_value_fg; + T r_value_og; + T r_checkI; + T r_checkF; + T r_checkO; + T r_state; + T r_prev_state = 0; + T r_state_atv; + T r_out; + + T *value_in = value.gate_value; + T *value_ig = value.gate_value + frame_size; + T *value_fg = value.gate_value + frame_size * 2; + T *value_og = value.gate_value + frame_size * 3; + + for (int i = 0; i < frame_size; i++) { + r_value_in = value_in[i]; + r_value_ig = value_ig[i]; + r_value_fg = value_fg[i]; + r_value_og = value_og[i]; + r_checkI = value.check_ig ? value.check_ig[i] : 0; + r_checkF = value.check_fg ? value.check_fg[i] : 0; + r_checkO = value.check_og ? value.check_og[i] : 0; + + if (value.prev_state_value) { + r_prev_state = value.prev_state_value[i]; } - op(rValueIn, rValueIg, rValueFg, rValueOg, rPrevState, rState, rStateAtv, - rOut, rCheckI, rCheckF, rCheckO, active_node, active_gate, active_state); - - valueIn[i] = rValueIn; - valueIg[i] = rValueIg; - valueFg[i] = rValueFg; - valueOg[i] = rValueOg; - value.stateValue[i] = rState; - value.stateActiveValue[i] = rStateAtv; - value.outputValue[i] = rOut; + 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; + value.state_value[i] = r_state; + value.state_active_value[i] = r_state_atv; + value.output_value[i] = r_out; } } template void naive_lstm_backward_one_sequence(Op op, LstmMetaValue value, - LstmMetaGrad grad, int frameSize, - activation_mode_t active_node, - activation_mode_t active_gate, - activation_mode_t active_state) { - T rValueIn; - T rValueIg; - T rValueFg; - T rValueOg; - T rGradIn; - T rGradIg; - T rGradFg; - T rGradOg; - T rPrevState = 0; - T rPrevStateGrad; - T rState; - T rStateGrad; - T rStateAtv; - T rOutputGrad; - T rCheckI; - T rCheckF; - T rCheckO; - T rCheckIGrad; - T rCheckFGrad; - T rCheckOGrad; - - T *valueIn = value.gateValue; - T *valueIg = value.gateValue + frameSize; - T *valueFg = value.gateValue + frameSize * 2; - T *valueOg = value.gateValue + frameSize * 3; - T *gradIn = grad.gateGrad; - T *gradIg = grad.gateGrad + frameSize; - T *gradFg = grad.gateGrad + frameSize * 2; - T *gradOg = grad.gateGrad + frameSize * 3; - - for (int i = 0; i < frameSize; i++) { - rValueIn = valueIn[i]; - rValueIg = valueIg[i]; - rValueFg = valueFg[i]; - rValueOg = valueOg[i]; - rCheckI = value.checkIg ? value.checkIg[i] : 0; - rCheckF = value.checkFg ? value.checkFg[i] : 0; - rCheckO = value.checkOg ? value.checkOg[i] : 0; - rState = value.stateValue[i]; - rStateAtv = value.stateActiveValue[i]; - rOutputGrad = grad.outputGrad[i]; - rStateGrad = grad.stateGrad[i]; - if (value.prevStateValue) { - rPrevState = value.prevStateValue[i]; + LstmMetaGrad grad, int frame_size, + ActivationType active_node, + ActivationType active_gate, + ActivationType active_state) { + T r_value_in; + T r_value_ig; + T r_value_fg; + T r_value_og; + T r_grad_in; + T r_grad_ig; + T r_grad_fg; + T r_grad_og; + T r_prev_state = 0; + T r_prev_state_grad; + T r_state; + T r_state_grad; + T r_state_atv; + T r_output_grad; + T r_checkI; + T r_checkF; + T r_checkO; + T r_checkIGrad; + T r_checkFGrad; + T r_checkOGrad; + + T *value_in = value.gate_value; + T *value_ig = value.gate_value + frame_size; + T *value_fg = value.gate_value + frame_size * 2; + T *value_og = value.gate_value + frame_size * 3; + T *grad_in = grad.gate_grad; + T *grad_ig = grad.gate_grad + frame_size; + T *grad_fg = grad.gate_grad + frame_size * 2; + T *grad_og = grad.gate_grad + frame_size * 3; + + for (int i = 0; i < frame_size; i++) { + r_value_in = value_in[i]; + r_value_ig = value_ig[i]; + r_value_fg = value_fg[i]; + r_value_og = value_og[i]; + r_checkI = value.check_ig ? value.check_ig[i] : 0; + r_checkF = value.check_fg ? value.check_fg[i] : 0; + r_checkO = value.check_og ? value.check_og[i] : 0; + r_state = value.state_value[i]; + r_state_atv = value.state_active_value[i]; + r_output_grad = grad.output_grad[i]; + r_state_grad = grad.state_grad[i]; + if (value.prev_state_value) { + r_prev_state = value.prev_state_value[i]; } - op(rValueIn, rValueIg, rValueFg, rValueOg, rGradIn, rGradIg, rGradFg, - rGradOg, rPrevState, rPrevStateGrad, rState, rStateGrad, rStateAtv, - rOutputGrad, rCheckI, rCheckF, rCheckO, rCheckIGrad, rCheckFGrad, - rCheckOGrad, active_node, active_gate, active_state); - - gradIn[i] = rGradIn; - gradIg[i] = rGradIg; - gradFg[i] = rGradFg; - gradOg[i] = rGradOg; - grad.stateGrad[i] = rStateGrad; - - if (grad.prevStateGrad) grad.prevStateGrad[i] = rPrevStateGrad; - if (value.prevStateValue) { - if (grad.checkIgGrad) grad.checkIgGrad[i] += rCheckIGrad; - if (grad.checkFgGrad) grad.checkFgGrad[i] += rCheckFGrad; + 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; + grad.state_grad[i] = r_state_grad; + + if (grad.prev_state_grad) grad.prev_state_grad[i] = r_prev_state_grad; + if (value.prev_state_value) { + if (grad.check_ig_grad) grad.check_ig_grad[i] += r_checkIGrad; + if (grad.check_fg_grad) grad.check_fg_grad[i] += r_checkFGrad; } - if (grad.checkOgGrad) grad.checkOgGrad[i] += rCheckOGrad; + if (grad.check_og_grad) grad.check_og_grad[i] += r_checkOGrad; } } template -void avx_lstm_forward_one_sequence(Op op, LstmMetaValue value, int frameSize, - activation_mode_t active_node, - activation_mode_t active_gate, - activation_mode_t active_state) { +void avx_lstm_forward_one_sequence(Op op, LstmMetaValue value, + int frame_size, ActivationType active_node, + ActivationType active_gate, + ActivationType active_state) { #ifdef __AVX__ - __m256 rValueIn; - __m256 rValueIg; - __m256 rValueFg; - __m256 rValueOg; - __m256 rCheckI = _mm256_set1_ps(0.0f); - __m256 rCheckF = _mm256_set1_ps(0.0f); - __m256 rCheckO = _mm256_set1_ps(0.0f); - __m256 rState; - __m256 rPrevState = _mm256_set1_ps(0.0f); - __m256 rStateAtv; - __m256 rOut; - - __m256 *valueIn = (__m256 *)value.gateValue; - __m256 *valueIg = (__m256 *)(value.gateValue + frameSize); - __m256 *valueFg = (__m256 *)(value.gateValue + frameSize * 2); - __m256 *valueOg = (__m256 *)(value.gateValue + frameSize * 3); - - for (int i = 0; i < frameSize / 8; i++) { - rValueIn = valueIn[i]; - rValueIg = valueIg[i]; - rValueFg = valueFg[i]; - rValueOg = valueOg[i]; - if (value.checkIg) { - rCheckI = ((__m256 *)value.checkIg)[i]; - rCheckF = ((__m256 *)value.checkFg)[i]; - rCheckO = ((__m256 *)value.checkOg)[i]; + __m256 r_value_in; + __m256 r_value_ig; + __m256 r_value_fg; + __m256 r_value_og; + __m256 r_checkI = _mm256_set1_ps(0.0f); + __m256 r_checkF = _mm256_set1_ps(0.0f); + __m256 r_checkO = _mm256_set1_ps(0.0f); + __m256 r_state; + __m256 r_prev_state = _mm256_set1_ps(0.0f); + __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); + + for (int i = 0; i < frame_size / 8; i++) { + r_value_in = value_in[i]; + r_value_ig = value_ig[i]; + 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]; } - if (value.prevStateValue) { - rPrevState = ((__m256 *)value.prevStateValue)[i]; + if (value.prev_state_value) { + r_prev_state = ((__m256 *)value.prev_state_value)[i]; } - op(rValueIn, rValueIg, rValueFg, rValueOg, rPrevState, rState, rStateAtv, - rOut, rCheckI, rCheckF, rCheckO, active_node, active_gate, active_state); - - valueIn[i] = rValueIn; - valueIg[i] = rValueIg; - valueFg[i] = rValueFg; - valueOg[i] = rValueOg; - ((__m256 *)value.stateValue)[i] = rState; - ((__m256 *)value.stateActiveValue)[i] = rStateAtv; - ((__m256 *)value.outputValue)[i] = rOut; + 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; } #endif } template void avx_lstm_backward_one_sequence(Op op, LstmMetaValue value, - LstmMetaGrad grad, int frameSize, - activation_mode_t active_node, - activation_mode_t active_gate, - activation_mode_t active_state) { + LstmMetaGrad grad, int frame_size, + ActivationType active_node, + ActivationType active_gate, + ActivationType active_state) { #ifdef __AVX__ - __m256 rValueIn; - __m256 rValueIg; - __m256 rValueFg; - __m256 rValueOg; - __m256 rGradIn; - __m256 rGradIg; - __m256 rGradFg; - __m256 rGradOg; - __m256 rPrevState = _mm256_set1_ps(0.0f); - __m256 rPrevStateGrad; - __m256 rStateGrad; - __m256 rState; - __m256 rStateAtv; - __m256 rOutputGrad; - __m256 rCheckI = _mm256_set1_ps(0.0f); - __m256 rCheckF = _mm256_set1_ps(0.0f); - __m256 rCheckO = _mm256_set1_ps(0.0f); - __m256 rCheckIGrad; - __m256 rCheckFGrad; - __m256 rCheckOGrad; - - __m256 *valueIn = (__m256 *)value.gateValue; - __m256 *valueIg = (__m256 *)(value.gateValue + frameSize); - __m256 *valueFg = (__m256 *)(value.gateValue + frameSize * 2); - __m256 *valueOg = (__m256 *)(value.gateValue + frameSize * 3); - __m256 *gradIn = (__m256 *)grad.gateGrad; - __m256 *gradIg = (__m256 *)(grad.gateGrad + frameSize); - __m256 *gradFg = (__m256 *)(grad.gateGrad + frameSize * 2); - __m256 *gradOg = (__m256 *)(grad.gateGrad + frameSize * 3); - - for (int i = 0; i < frameSize / 8; i++) { - rValueIn = valueIn[i]; - rValueIg = valueIg[i]; - rValueFg = valueFg[i]; - rValueOg = valueOg[i]; - if (value.checkIg) { - rCheckI = ((__m256 *)value.checkIg)[i]; - rCheckF = ((__m256 *)value.checkFg)[i]; - rCheckO = ((__m256 *)value.checkOg)[i]; + __m256 r_value_in; + __m256 r_value_ig; + __m256 r_value_fg; + __m256 r_value_og; + __m256 r_grad_in; + __m256 r_grad_ig; + __m256 r_grad_fg; + __m256 r_grad_og; + __m256 r_prev_state = _mm256_set1_ps(0.0f); + __m256 r_prev_state_grad; + __m256 r_state_grad; + __m256 r_state; + __m256 r_state_atv; + __m256 r_output_grad; + __m256 r_checkI = _mm256_set1_ps(0.0f); + __m256 r_checkF = _mm256_set1_ps(0.0f); + __m256 r_checkO = _mm256_set1_ps(0.0f); + __m256 r_checkIGrad; + __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); + + for (int i = 0; i < frame_size / 8; i++) { + r_value_in = value_in[i]; + r_value_ig = value_ig[i]; + 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]; } - rState = ((__m256 *)value.stateValue)[i]; - rStateAtv = ((__m256 *)value.stateActiveValue)[i]; - rOutputGrad = ((__m256 *)grad.outputGrad)[i]; - rStateGrad = ((__m256 *)grad.stateGrad)[i]; - if (value.prevStateValue) { - rPrevState = ((__m256 *)value.prevStateValue)[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]; + if (value.prev_state_value) { + r_prev_state = ((__m256 *)value.prev_state_value)[i]; } - op(rValueIn, rValueIg, rValueFg, rValueOg, rGradIn, rGradIg, rGradFg, - rGradOg, rPrevState, rPrevStateGrad, rState, rStateGrad, rStateAtv, - rOutputGrad, rCheckI, rCheckF, rCheckO, rCheckIGrad, rCheckFGrad, - rCheckOGrad, active_node, active_gate, active_state); - - gradIn[i] = rGradIn; - gradIg[i] = rGradIg; - gradFg[i] = rGradFg; - gradOg[i] = rGradOg; - ((__m256 *)grad.stateGrad)[i] = rStateGrad; - - if (grad.prevStateGrad) ((__m256 *)grad.prevStateGrad)[i] = rPrevStateGrad; - if (value.prevStateValue) { - if (grad.checkIgGrad) ((__m256 *)grad.checkIgGrad)[i] += rCheckIGrad; - if (grad.checkFgGrad) ((__m256 *)grad.checkFgGrad)[i] += rCheckFGrad; + 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; + + if (grad.prev_state_grad) + ((__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.checkOgGrad) ((__m256 *)grad.checkOgGrad)[i] += rCheckOGrad; + if (grad.check_og_grad) ((__m256 *)grad.check_og_grad)[i] += r_checkOGrad; } #endif } template -void cpu_lstm_forward(Op op, LstmMetaValue value, int frameSize, - activation_mode_t active_node, - activation_mode_t active_gate, - activation_mode_t active_state) { - if (Op::avx && !(frameSize & (8 - 1)) && (std::is_same::value)) { - avx_lstm_forward_one_sequence(op, value, frameSize, active_node, +void cpu_lstm_forward(Op op, LstmMetaValue value, int frame_size, + ActivationType active_node, ActivationType active_gate, + ActivationType active_state) { + if (Op::avx && !(frame_size & (8 - 1)) && (std::is_same::value)) { + avx_lstm_forward_one_sequence(op, value, frame_size, active_node, active_gate, active_state); } else { - naive_lstm_forward_one_sequence(op, value, frameSize, active_node, + naive_lstm_forward_one_sequence(op, value, frame_size, active_node, active_gate, active_state); } } template void cpu_lstm_backward(Op op, LstmMetaValue value, LstmMetaGrad grad, - int frameSize, activation_mode_t active_node, - activation_mode_t active_gate, - activation_mode_t active_state) { - if (Op::avx && !(frameSize & (8 - 1)) && (std::is_same::value)) { - avx_lstm_backward_one_sequence(op, value, grad, frameSize, active_node, + int frame_size, ActivationType active_node, + ActivationType active_gate, + ActivationType active_state) { + if (Op::avx && !(frame_size & (8 - 1)) && (std::is_same::value)) { + avx_lstm_backward_one_sequence(op, value, grad, frame_size, active_node, active_gate, active_state); } else { - naive_lstm_backward_one_sequence(op, value, grad, frameSize, active_node, - active_gate, active_state); + naive_lstm_backward_one_sequence(op, value, grad, frame_size, + active_node, active_gate, active_state); } } diff --git a/paddle/operators/math/detail/lstm_gpu_kernel.h b/paddle/operators/math/detail/lstm_gpu_kernel.h index d138bbe411f69929a14ad19af3e84824ac7a5d58..e31e657e8b6964c2b99f6e456545c83d8da8e7f9 100644 --- a/paddle/operators/math/detail/lstm_gpu_kernel.h +++ b/paddle/operators/math/detail/lstm_gpu_kernel.h @@ -26,189 +26,191 @@ namespace math { namespace detail { /* - * threads(framePerBlock, batchPerBlock) - * grid(frameBlocks, batchBlocks) + * threads(frame_per_block, batch_per_block) + * grid(frame_blocks, batch_blocks) */ -template -__global__ void KeLstmForward(Op op, LstmMetaValue value, int frameSize, - int batchSize, activation_mode_t active_node, - activation_mode_t active_gate, - activation_mode_t active_state) { - const int frameIdx = blockIdx.x * blockDim.x + threadIdx.x; - if (frameIdx >= frameSize) return; - - int batchIdx = 0; - if (isBatch) { - batchIdx = blockIdx.y * blockDim.y + threadIdx.y; - if (batchIdx >= batchSize) return; - value.gateValue += batchIdx * frameSize * 4; - value.outputValue += batchIdx * frameSize; - value.stateValue += batchIdx * frameSize; - value.stateActiveValue += batchIdx * frameSize; +template +__global__ void KeLstmForward(Op op, LstmMetaValue value, int frame_size, + int batch_size, ActivationType active_node, + ActivationType active_gate, + ActivationType active_state) { + const int frame_idx = blockIdx.x * blockDim.x + threadIdx.x; + if (frame_idx >= frame_size) return; + + int batch_idx = 0; + if (is_batch) { + batch_idx = blockIdx.y * blockDim.y + threadIdx.y; + if (batch_idx >= batch_size) return; + value.gate_value += batch_idx * frame_size * 4; + value.output_value += batch_idx * frame_size; + value.state_value += batch_idx * frame_size; + value.state_active_value += batch_idx * frame_size; } - T rState; - T rPrevState = 0; - T rStateAtv; - T rOut; - T rValueIn; - T rValueIg; - T rValueFg; - T rValueOg; - - T rCheckI = value.checkIg ? value.checkIg[frameIdx] : 0; - T rCheckF = value.checkFg ? value.checkFg[frameIdx] : 0; - T rCheckO = value.checkOg ? value.checkOg[frameIdx] : 0; - - rValueIn = value.gateValue[frameIdx]; - rValueIg = value.gateValue[frameIdx + frameSize]; - rValueFg = value.gateValue[frameIdx + frameSize * 2]; - rValueOg = value.gateValue[frameIdx + frameSize * 3]; - - if (value.prevStateValue) { - if (isBatch) value.prevStateValue += batchIdx * frameSize; - rPrevState = value.prevStateValue[frameIdx]; + T r_state; + T r_prev_state = 0; + T r_state_atv; + T r_out; + T r_value_in; + T r_value_ig; + T r_value_fg; + T r_value_og; + + T r_checkI = value.check_ig ? value.check_ig[frame_idx] : 0; + T r_checkF = value.check_fg ? value.check_fg[frame_idx] : 0; + T r_checkO = value.check_og ? value.check_og[frame_idx] : 0; + + r_value_in = value.gate_value[frame_idx]; + r_value_ig = value.gate_value[frame_idx + frame_size]; + r_value_fg = value.gate_value[frame_idx + frame_size * 2]; + r_value_og = value.gate_value[frame_idx + frame_size * 3]; + + if (value.prev_state_value) { + if (is_batch) value.prev_state_value += batch_idx * frame_size; + r_prev_state = value.prev_state_value[frame_idx]; } - op(rValueIn, rValueIg, rValueFg, rValueOg, rPrevState, rState, rStateAtv, - rOut, rCheckI, rCheckF, rCheckO, 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.gateValue[frameIdx] = rValueIn; - value.gateValue[frameIdx + frameSize] = rValueIg; - value.gateValue[frameIdx + frameSize * 2] = rValueFg; - value.gateValue[frameIdx + frameSize * 3] = rValueOg; + value.gate_value[frame_idx] = r_value_in; + value.gate_value[frame_idx + frame_size] = r_value_ig; + value.gate_value[frame_idx + frame_size * 2] = r_value_fg; + value.gate_value[frame_idx + frame_size * 3] = r_value_og; - value.stateValue[frameIdx] = rState; - value.stateActiveValue[frameIdx] = rStateAtv; - value.outputValue[frameIdx] = rOut; + value.state_value[frame_idx] = r_state; + value.state_active_value[frame_idx] = r_state_atv; + value.output_value[frame_idx] = r_out; } /* - * threads(framePerBlock, batchPerBlock) - * grid(frameBlocks, batchBlocks) + * threads(frame_per_block, batch_per_block) + * grid(frame_blocks, batch_blocks) */ -template +template __global__ void KeLstmBackward(Op op, LstmMetaValue value, - LstmMetaGrad grad, int frameSize, - int batchSize, activation_mode_t active_node, - activation_mode_t active_gate, - activation_mode_t active_state) { - const int frameIdx = blockIdx.x * blockDim.x + threadIdx.x; - if (frameIdx >= frameSize) return; - - int batchIdx = 0; - if (isBatch) { - batchIdx = blockIdx.y * blockDim.y + threadIdx.y; - if (batchIdx >= batchSize) return; - value.gateValue += batchIdx * frameSize * 4; - value.stateValue += batchIdx * frameSize; - value.stateActiveValue += batchIdx * frameSize; - grad.gateGrad += batchIdx * frameSize * 4; - grad.stateGrad += batchIdx * frameSize; - grad.outputGrad += batchIdx * frameSize; + LstmMetaGrad grad, int frame_size, + int batch_size, ActivationType active_node, + ActivationType active_gate, + ActivationType active_state) { + const int frame_idx = blockIdx.x * blockDim.x + threadIdx.x; + if (frame_idx >= frame_size) return; + + int batch_idx = 0; + if (is_batch) { + batch_idx = blockIdx.y * blockDim.y + threadIdx.y; + if (batch_idx >= batch_size) return; + value.gate_value += batch_idx * frame_size * 4; + value.state_value += batch_idx * frame_size; + value.state_active_value += batch_idx * frame_size; + grad.gate_grad += batch_idx * frame_size * 4; + grad.state_grad += batch_idx * frame_size; + grad.output_grad += batch_idx * frame_size; } - T rValueIn; - T rValueIg; - T rValueFg; - T rValueOg; - T rGradIn; - T rGradIg; - T rGradFg; - T rGradOg; - T rPrevState = 0; - T rPrevStateGrad; - T rState; - T rStateGrad; - T rStateAtv; - T rOutputGrad; - T rCheckI = value.checkIg ? value.checkIg[frameIdx] : 0; - T rCheckF = value.checkFg ? value.checkFg[frameIdx] : 0; - T rCheckO = value.checkOg ? value.checkOg[frameIdx] : 0; - - T rCheckIGrad; - T rCheckFGrad; - T rCheckOGrad; - - rValueIn = value.gateValue[frameIdx]; - rValueIg = value.gateValue[frameIdx + frameSize]; - rValueFg = value.gateValue[frameIdx + frameSize * 2]; - rValueOg = value.gateValue[frameIdx + frameSize * 3]; - rState = value.stateValue[frameIdx]; - rStateAtv = value.stateActiveValue[frameIdx]; - rOutputGrad = grad.outputGrad[frameIdx]; - rStateGrad = grad.stateGrad[frameIdx]; - - if (value.prevStateValue) { - if (isBatch) value.prevStateValue += batchIdx * frameSize; - rPrevState = value.prevStateValue[frameIdx]; + T r_value_in; + T r_value_ig; + T r_value_fg; + T r_value_og; + T r_grad_in; + T r_grad_ig; + T r_grad_fg; + T r_grad_og; + T r_prev_state = 0; + T r_prev_state_grad; + T r_state; + T r_state_grad; + T r_state_atv; + T r_output_grad; + T r_checkI = value.check_ig ? value.check_ig[frame_idx] : 0; + T r_checkF = value.check_fg ? value.check_fg[frame_idx] : 0; + T r_checkO = value.check_og ? value.check_og[frame_idx] : 0; + + T r_checkIGrad; + T r_checkFGrad; + T r_checkOGrad; + + r_value_in = value.gate_value[frame_idx]; + r_value_ig = value.gate_value[frame_idx + frame_size]; + r_value_fg = value.gate_value[frame_idx + frame_size * 2]; + r_value_og = value.gate_value[frame_idx + frame_size * 3]; + r_state = value.state_value[frame_idx]; + r_state_atv = value.state_active_value[frame_idx]; + r_output_grad = grad.output_grad[frame_idx]; + r_state_grad = grad.state_grad[frame_idx]; + + if (value.prev_state_value) { + if (is_batch) value.prev_state_value += batch_idx * frame_size; + r_prev_state = value.prev_state_value[frame_idx]; } - op(rValueIn, rValueIg, rValueFg, rValueOg, rGradIn, rGradIg, rGradFg, rGradOg, - rPrevState, rPrevStateGrad, rState, rStateGrad, rStateAtv, rOutputGrad, - rCheckI, rCheckF, rCheckO, rCheckIGrad, rCheckFGrad, rCheckOGrad, - active_node, active_gate, active_state); - - grad.gateGrad[frameIdx] = rGradIn; - grad.gateGrad[frameIdx + frameSize] = rGradIg; - grad.gateGrad[frameIdx + frameSize * 2] = rGradFg; - grad.gateGrad[frameIdx + frameSize * 3] = rGradOg; - grad.stateGrad[frameIdx] = rStateGrad; - if (grad.prevStateGrad) { - if (isBatch) grad.prevStateGrad += batchIdx * frameSize; - grad.prevStateGrad[frameIdx] = rPrevStateGrad; + 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; + grad.gate_grad[frame_idx + frame_size * 2] = r_grad_fg; + grad.gate_grad[frame_idx + frame_size * 3] = r_grad_og; + grad.state_grad[frame_idx] = r_state_grad; + if (grad.prev_state_grad) { + if (is_batch) grad.prev_state_grad += batch_idx * frame_size; + grad.prev_state_grad[frame_idx] = r_prev_state_grad; } - if (isBatch) { - if (value.prevStateValue) { - if (grad.checkIgGrad) - paddle::platform::CudaAtomicAdd(grad.checkIgGrad + frameIdx, - rCheckIGrad); - if (grad.checkFgGrad) - paddle::platform::CudaAtomicAdd(grad.checkFgGrad + frameIdx, - rCheckFGrad); + if (is_batch) { + if (value.prev_state_value) { + if (grad.check_ig_grad) + paddle::platform::CudaAtomicAdd(grad.check_ig_grad + frame_idx, + r_checkIGrad); + if (grad.check_fg_grad) + paddle::platform::CudaAtomicAdd(grad.check_fg_grad + frame_idx, + r_checkFGrad); } - if (grad.checkOgGrad) - paddle::platform::CudaAtomicAdd(grad.checkOgGrad + frameIdx, rCheckOGrad); + if (grad.check_og_grad) + paddle::platform::CudaAtomicAdd(grad.check_og_grad + frame_idx, + r_checkOGrad); } else { - if (value.prevStateValue) { - if (grad.checkIgGrad) grad.checkIgGrad[frameIdx] += rCheckIGrad; - if (grad.checkFgGrad) grad.checkFgGrad[frameIdx] += rCheckFGrad; + if (value.prev_state_value) { + if (grad.check_ig_grad) grad.check_ig_grad[frame_idx] += r_checkIGrad; + if (grad.check_fg_grad) grad.check_fg_grad[frame_idx] += r_checkFGrad; } - if (grad.checkOgGrad) grad.checkOgGrad[frameIdx] += rCheckOGrad; + if (grad.check_og_grad) grad.check_og_grad[frame_idx] += r_checkOGrad; } } template void gpu_lstm_forward(const platform::DeviceContext& context, Op op, - LstmMetaValue value, int frameSize, int batchSize, - activation_mode_t active_node, - activation_mode_t active_gate, - activation_mode_t active_state) { + LstmMetaValue value, int frame_size, int batch_size, + ActivationType active_node, ActivationType active_gate, + ActivationType active_state) { dim3 threads; dim3 grid; - if (batchSize == 1) { - int framePerBlock = frameSize <= 1024 ? frameSize : 1024; - int frameBlocks = (frameSize + 1024 - 1) / 1024; - threads = dim3(framePerBlock, 1); - grid = dim3(frameBlocks, 1); + if (batch_size == 1) { + int frame_per_block = frame_size <= 1024 ? frame_size : 1024; + int frame_blocks = (frame_size + 1024 - 1) / 1024; + threads = dim3(frame_per_block, 1); + grid = dim3(frame_blocks, 1); } else { - /* framePerBlock = 32 batchPerBlock = 32 */ + /* frame_per_block = 32 batch_per_block = 32 */ threads = dim3(32, 32); - grid = dim3((frameSize + 32 - 1) / 32, (batchSize + 32 - 1) / 32); + grid = dim3((frame_size + 32 - 1) / 32, (batch_size + 32 - 1) / 32); } auto stream = reinterpret_cast(context).stream(); - if (batchSize == 1) { + if (batch_size == 1) { KeLstmForward<<>>( - op, value, frameSize, batchSize, active_node, active_gate, + /* is_batch= */ false><<>>( + op, value, frame_size, batch_size, active_node, active_gate, active_state); } else { KeLstmForward<<>>( - op, value, frameSize, batchSize, active_node, active_gate, + /* is_batch= */ true><<>>( + op, value, frame_size, batch_size, active_node, active_gate, active_state); } } @@ -216,34 +218,33 @@ void gpu_lstm_forward(const platform::DeviceContext& context, Op op, template void gpu_lstm_backward(const platform::DeviceContext& context, Op op, LstmMetaValue value, LstmMetaGrad grad, - int frameSize, int batchSize, - activation_mode_t active_node, - activation_mode_t active_gate, - activation_mode_t active_state) { + int frame_size, int batch_size, + ActivationType active_node, ActivationType active_gate, + ActivationType active_state) { dim3 threads; dim3 grid; - if (batchSize == 1) { - int framePerBlock = frameSize <= 1024 ? frameSize : 1024; - int frameBlocks = (frameSize + 1024 - 1) / 1024; - threads = dim3(framePerBlock, 1); - grid = dim3(frameBlocks, 1); + if (batch_size == 1) { + int frame_per_block = frame_size <= 1024 ? frame_size : 1024; + int frame_blocks = (frame_size + 1024 - 1) / 1024; + threads = dim3(frame_per_block, 1); + grid = dim3(frame_blocks, 1); } else { - /* framePerBlock = 32 batchPerBlock = 16 */ + /* frame_per_block = 32 batch_per_block = 16 */ threads = dim3(32, 16); - grid = dim3((frameSize + 32 - 1) / 32, (batchSize + 16 - 1) / 16); + grid = dim3((frame_size + 32 - 1) / 32, (batch_size + 16 - 1) / 16); } auto stream = reinterpret_cast(context).stream(); - if (batchSize == 1) { + if (batch_size == 1) { KeLstmBackward<<>>( - op, value, grad, frameSize, batchSize, active_node, active_gate, + /* is_batch= */ false><<>>( + op, value, grad, frame_size, batch_size, active_node, active_gate, active_state); } else { KeLstmBackward<<>>( - op, value, grad, frameSize, batchSize, active_node, active_gate, + /* is_batch= */ true><<>>( + op, value, grad, frame_size, batch_size, active_node, active_gate, active_state); } } diff --git a/paddle/operators/math/detail/lstm_kernel.h b/paddle/operators/math/detail/lstm_kernel.h index 9daaf91981a8e0252374f528f0e063111bd32675..fed8f9c4ca48905ad4c524ba400e8c7bb2f7fbd1 100644 --- a/paddle/operators/math/detail/lstm_kernel.h +++ b/paddle/operators/math/detail/lstm_kernel.h @@ -27,19 +27,19 @@ namespace forward { template class lstm { public: - HOSTDEVICE void operator()(T &valueIn, T &valueIg, T &valueFg, T &valueOg, - T &prevState, T &state, T &stateAtv, T &output, + 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, - activation_mode_t active_node, - activation_mode_t active_gate, - activation_mode_t active_state) { - valueIn = activation(valueIn, active_node); - valueIg = activation(valueIg + prevState * checkI, active_gate); - valueFg = activation(valueFg + prevState * checkF, active_gate); - state = valueIn * valueIg + prevState * valueFg; - valueOg = activation(valueOg + state * checkO, active_gate); - stateAtv = activation(state, active_state); - output = valueOg * stateAtv; + 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; } #ifndef __NVCC__ #ifndef __AVX__ // If not compiled with AVX instructs. Disable AVX by default @@ -48,24 +48,27 @@ class lstm { // Only float support AVX optimization static const bool avx = std::is_same::value; - HOSTDEVICE void operator()(__m256 &valueIn, __m256 &valueIg, __m256 &valueFg, - __m256 &valueOg, __m256 &prevState, __m256 &state, - __m256 &stateAtv, __m256 &output, __m256 &checkI, + 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, - activation_mode_t active_node, - activation_mode_t active_gate, - activation_mode_t active_state) { - valueIn = activation(valueIn, active_node); - valueIg = activation( - _mm256_add_ps(valueIg, _mm256_mul_ps(prevState, checkI)), active_gate); - valueFg = activation( - _mm256_add_ps(valueFg, _mm256_mul_ps(prevState, checkF)), active_gate); - state = _mm256_add_ps(_mm256_mul_ps(valueIn, valueIg), - _mm256_mul_ps(prevState, valueFg)); - valueOg = activation(_mm256_add_ps(valueOg, _mm256_mul_ps(state, checkO)), - active_gate); - stateAtv = activation(state, active_state); - output = _mm256_mul_ps(valueOg, stateAtv); + 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); } #endif #endif @@ -78,25 +81,26 @@ namespace backward { template class lstm { public: - HOSTDEVICE void operator()(T &valueIn, T &valueIg, T &valueFg, T &valueOg, - T &gradIn, T &gradIg, T &gradFg, T &gradOg, - T &prevState, T &prevStateGrad, T &state, - T &stateGrad, T &stateAtv, T &outputGrad, + 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, - activation_mode_t active_node, - activation_mode_t active_gate, - activation_mode_t active_state) { - gradOg = activation(outputGrad * stateAtv, valueOg, active_gate); - stateGrad += activation(outputGrad * valueOg, stateAtv, active_state) + - gradOg * checkO; - gradIn = activation(stateGrad * valueIg, valueIn, active_node); - gradIg = activation(stateGrad * valueIn, valueIg, active_gate); - gradFg = activation(stateGrad * prevState, valueFg, active_gate); - prevStateGrad = gradIg * checkI + gradFg * checkF + stateGrad * valueFg; - checkIGrad = gradIg * prevState; - checkFGrad = gradFg * prevState; - checkOGrad = gradOg * state; + 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; } #ifndef __NVCC__ #ifndef __AVX__ // If not compiled with AVX instructs. Disable AVX by default @@ -105,32 +109,32 @@ class lstm { // Only float support AVX optimization static const bool avx = std::is_same::value; HOSTDEVICE void operator()( - __m256 &valueIn, __m256 &valueIg, __m256 &valueFg, __m256 &valueOg, - __m256 &gradIn, __m256 &gradIg, __m256 &gradFg, __m256 &gradOg, - __m256 &prevState, __m256 &prevStateGrad, __m256 &state, - __m256 &stateGrad, __m256 &stateAtv, __m256 &outputGrad, __m256 &checkI, - __m256 &checkF, __m256 &checkO, __m256 &checkIGrad, __m256 &checkFGrad, - __m256 &checkOGrad, activation_mode_t active_node, - activation_mode_t active_gate, activation_mode_t active_state) { - gradOg = - activation(_mm256_mul_ps(outputGrad, stateAtv), valueOg, active_gate); - stateGrad = _mm256_add_ps( - activation(_mm256_mul_ps(outputGrad, valueOg), stateAtv, active_state), - stateGrad); - stateGrad = _mm256_add_ps(_mm256_mul_ps(gradOg, checkO), stateGrad); - gradIn = - activation(_mm256_mul_ps(stateGrad, valueIg), valueIn, active_node); - gradIg = - activation(_mm256_mul_ps(stateGrad, valueIn), valueIg, active_gate); - gradFg = - activation(_mm256_mul_ps(stateGrad, prevState), valueFg, active_gate); - prevStateGrad = _mm256_add_ps(_mm256_mul_ps(gradIg, checkI), - _mm256_mul_ps(gradFg, checkF)); - prevStateGrad = - _mm256_add_ps(_mm256_mul_ps(stateGrad, valueFg), prevStateGrad); - checkIGrad = _mm256_mul_ps(gradIg, prevState); - checkFGrad = _mm256_mul_ps(gradFg, prevState); - checkOGrad = _mm256_mul_ps(gradOg, state); + __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); } #endif #endif diff --git a/paddle/operators/math/detection_util.h b/paddle/operators/math/detection_util.h new file mode 100644 index 0000000000000000000000000000000000000000..e3a3ef2badc37924d866ded8ee7a7338fbc4b2d2 --- /dev/null +++ b/paddle/operators/math/detection_util.h @@ -0,0 +1,300 @@ +/* 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 +#include +#include "paddle/framework/selected_rows.h" +#include "paddle/platform/device_context.h" + +namespace paddle { +namespace operators { +namespace math { +template +struct BBox { + BBox(T x_min, T y_min, T x_max, T y_max) + : x_min(x_min), + y_min(y_min), + x_max(x_max), + y_max(y_max), + is_difficult(false) {} + + BBox() {} + + T get_width() const { return x_max - x_min; } + + T get_height() const { return y_max - y_min; } + + T get_center_x() const { return (x_min + x_max) / 2; } + + T get_center_y() const { return (y_min + y_max) / 2; } + + T get_area() const { return get_width() * get_height(); } + + // coordinate of bounding box + T x_min; + T y_min; + T x_max; + T y_max; + // whether difficult object (e.g. object with heavy occlusion is difficult) + bool is_difficult; +}; +// KNCHW ==> NHWC +// template +template +void GetBBoxFromPriorData(const T* prior_data, const size_t num_bboxes, + std::vector>& bbox_vec); +template +void GetBBoxVarFromPriorData(const T* prior_data, const size_t num, + std::vector>& var_vec); +template +BBox DecodeBBoxWithVar(BBox& prior_bbox, + const std::vector& prior_bbox_var, + const std::vector& loc_pred_data); +template +bool SortScorePairDescend(const std::pair& pair1, + const std::pair& pair2); +template +bool SortScorePairDescend(const std::pair>& pair1, + const std::pair>& pair2); +template +T jaccard_overlap(const BBox& bbox1, const BBox& bbox2); + +template +void ApplyNmsFast(const std::vector>& bboxes, const T* conf_score_data, + size_t class_idx, size_t top_k, T conf_threshold, + T nms_threshold, size_t num_priors, size_t num_classes, + std::vector* indices); +template +int GetDetectionIndices( + const T* conf_data, const size_t num_priors, const size_t num_classes, + const size_t background_label_id, const size_t batch_size, + const T conf_threshold, const size_t nms_top_k, const T nms_threshold, + const size_t top_k, + const std::vector>>& all_decoded_bboxes, + std::vector>>* all_detection_indices); +template +BBox ClipBBox(const BBox& bbox); +template +void GetDetectionOutput( + const T* conf_data, const size_t num_kept, const size_t num_priors, + const size_t num_classes, const size_t batch_size, + const std::vector>>& all_indices, + const std::vector>>& all_decoded_bboxes, T* out_data); +template +void GetBBoxFromPriorData(const T* prior_data, const size_t num_bboxes, + std::vector>& bbox_vec) { + size_t out_offset = bbox_vec.size(); + bbox_vec.resize(bbox_vec.size() + num_bboxes); + for (size_t i = 0; i < num_bboxes; ++i) { + BBox bbox; + bbox.x_min = *(prior_data + i * 8); + bbox.y_min = *(prior_data + i * 8 + 1); + bbox.x_max = *(prior_data + i * 8 + 2); + bbox.y_max = *(prior_data + i * 8 + 3); + bbox_vec[out_offset + i] = bbox; + } +} +template +void GetBBoxVarFromPriorData(const T* prior_data, const size_t num, + std::vector>& var_vec) { + size_t out_offset = var_vec.size(); + var_vec.resize(var_vec.size() + num); + for (size_t i = 0; i < num; ++i) { + std::vector var; + var.push_back(*(prior_data + i * 8 + 4)); + var.push_back(*(prior_data + i * 8 + 5)); + var.push_back(*(prior_data + i * 8 + 6)); + var.push_back(*(prior_data + i * 8 + 7)); + var_vec[out_offset + i] = var; + } +} +template +BBox DecodeBBoxWithVar(BBox& prior_bbox, + const std::vector& prior_bbox_var, + const std::vector& loc_pred_data) { + T prior_bbox_width = prior_bbox.get_width(); + T prior_bbox_height = prior_bbox.get_height(); + T prior_bbox_center_x = prior_bbox.get_center_x(); + T prior_bbox_center_y = prior_bbox.get_center_y(); + + T decoded_bbox_center_x = + prior_bbox_var[0] * loc_pred_data[0] * prior_bbox_width + + prior_bbox_center_x; + T decoded_bbox_center_y = + prior_bbox_var[1] * loc_pred_data[1] * prior_bbox_height + + prior_bbox_center_y; + T decoded_bbox_width = + std::exp(prior_bbox_var[2] * loc_pred_data[2]) * prior_bbox_width; + T decoded_bbox_height = + std::exp(prior_bbox_var[3] * loc_pred_data[3]) * prior_bbox_height; + + BBox decoded_bbox; + decoded_bbox.x_min = decoded_bbox_center_x - decoded_bbox_width / 2; + decoded_bbox.y_min = decoded_bbox_center_y - decoded_bbox_height / 2; + decoded_bbox.x_max = decoded_bbox_center_x + decoded_bbox_width / 2; + decoded_bbox.y_max = decoded_bbox_center_y + decoded_bbox_height / 2; + + return decoded_bbox; +} +template +bool SortScorePairDescend(const std::pair& pair1, + const std::pair& pair2) { + return pair1.first > pair2.first; +} +template +T jaccard_overlap(const BBox& bbox1, const BBox& bbox2) { + if (bbox2.x_min > bbox1.x_max || bbox2.x_max < bbox1.x_min || + bbox2.y_min > bbox1.y_max || bbox2.y_max < bbox1.y_min) { + return 0.0; + } else { + T inter_x_min = std::max(bbox1.x_min, bbox2.x_min); + T inter_y_min = std::max(bbox1.y_min, bbox2.y_min); + T interX_max = std::min(bbox1.x_max, bbox2.x_max); + T interY_max = std::min(bbox1.y_max, bbox2.y_max); + + T inter_width = interX_max - inter_x_min; + T inter_height = interY_max - inter_y_min; + T inter_area = inter_width * inter_height; + + T bbox_area1 = bbox1.get_area(); + T bbox_area2 = bbox2.get_area(); + + return inter_area / (bbox_area1 + bbox_area2 - inter_area); + } +} + +template +void ApplyNmsFast(const std::vector>& bboxes, const T* conf_score_data, + size_t class_idx, size_t top_k, T conf_threshold, + T nms_threshold, size_t num_priors, size_t num_classes, + std::vector* indices) { + std::vector> scores; + for (size_t i = 0; i < num_priors; ++i) { + size_t conf_offset = i * num_classes + class_idx; + if (conf_score_data[conf_offset] > conf_threshold) + scores.push_back(std::make_pair(conf_score_data[conf_offset], i)); + } + std::stable_sort(scores.begin(), scores.end(), + SortScorePairDescend); + if (top_k > 0 && top_k < scores.size()) scores.resize(top_k); + while (scores.size() > 0) { + const size_t idx = scores.front().second; + bool keep = true; + for (size_t i = 0; i < indices->size(); ++i) { + if (keep) { + const size_t saved_idx = (*indices)[i]; + T overlap = jaccard_overlap(bboxes[idx], bboxes[saved_idx]); + keep = overlap <= nms_threshold; + } else { + break; + } + } + if (keep) indices->push_back(idx); + scores.erase(scores.begin()); + } +} +template +int GetDetectionIndices( + const T* conf_data, const size_t num_priors, const size_t num_classes, + const size_t background_label_id, const size_t batch_size, + const T conf_threshold, const size_t nms_top_k, const T nms_threshold, + const size_t top_k, + const std::vector>>& all_decoded_bboxes, + std::vector>>* all_detection_indices) { + int total_keep_num = 0; + for (size_t n = 0; n < batch_size; ++n) { + const std::vector>& decoded_bboxes = all_decoded_bboxes[n]; + size_t num_detected = 0; + std::map> indices; + size_t conf_offset = n * num_priors * num_classes; + for (size_t c = 0; c < num_classes; ++c) { + if (c == background_label_id) continue; + ApplyNmsFast(decoded_bboxes, conf_data + conf_offset, c, nms_top_k, + conf_threshold, nms_threshold, num_priors, num_classes, + &(indices[c])); + num_detected += indices[c].size(); + } + if (top_k > 0 && num_detected > top_k) { + // std::vector> score_index_pairs; + std::vector>> score_index_pairs; + for (size_t c = 0; c < num_classes; ++c) { + const std::vector& label_indices = indices[c]; + for (size_t i = 0; i < label_indices.size(); ++i) { + size_t idx = label_indices[i]; + score_index_pairs.push_back( + std::make_pair((conf_data + conf_offset)[idx * num_classes + c], + std::make_pair(c, idx))); + } + } + std::sort(score_index_pairs.begin(), score_index_pairs.end(), + SortScorePairDescend>); + score_index_pairs.resize(top_k); + std::map> new_indices; + for (size_t i = 0; i < score_index_pairs.size(); ++i) { + size_t label = score_index_pairs[i].second.first; + size_t idx = score_index_pairs[i].second.second; + new_indices[label].push_back(idx); + } + all_detection_indices->push_back(new_indices); + total_keep_num += top_k; + } else { + all_detection_indices->push_back(indices); + total_keep_num += num_detected; + } + } + return total_keep_num; +} +template +BBox ClipBBox(const BBox& bbox) { + T one = static_cast(1.0); + T zero = static_cast(0.0); + BBox clipped_bbox; + clipped_bbox.x_min = std::max(std::min(bbox.x_min, one), zero); + clipped_bbox.y_min = std::max(std::min(bbox.y_min, one), zero); + clipped_bbox.x_max = std::max(std::min(bbox.x_max, one), zero); + clipped_bbox.y_max = std::max(std::min(bbox.y_max, one), zero); + return clipped_bbox; +} +template +void GetDetectionOutput( + const T* conf_data, const size_t num_kept, const size_t num_priors, + const size_t num_classes, const size_t batch_size, + const std::vector>>& all_indices, + const std::vector>>& all_decoded_bboxes, T* out_data) { + size_t count = 0; + for (size_t n = 0; n < batch_size; ++n) { + for (std::map>::const_iterator it = + all_indices[n].begin(); + it != all_indices[n].end(); ++it) { + size_t label = it->first; + const std::vector& indices = it->second; + const std::vector>& decoded_bboxes = all_decoded_bboxes[n]; + for (size_t i = 0; i < indices.size(); ++i) { + size_t idx = indices[i]; + size_t conf_offset = n * num_priors * num_classes + idx * num_classes; + out_data[count * 7] = n; + out_data[count * 7 + 1] = label; + out_data[count * 7 + 2] = (conf_data + conf_offset)[label]; + BBox clipped_bbox = ClipBBox(decoded_bboxes[idx]); + out_data[count * 7 + 3] = clipped_bbox.x_min; + out_data[count * 7 + 4] = clipped_bbox.y_min; + out_data[count * 7 + 5] = clipped_bbox.x_max; + out_data[count * 7 + 6] = clipped_bbox.y_max; + ++count; + } + } + } +} +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/math/gru_compute.cc b/paddle/operators/math/gru_compute.cc index 125af449d3f700e24be5e4b7615c3b0e03fd4e5b..101ab859624869bf34d171cd42d46d0c5bdac29c 100644 --- a/paddle/operators/math/gru_compute.cc +++ b/paddle/operators/math/gru_compute.cc @@ -19,83 +19,85 @@ namespace operators { namespace math { template -struct GRUUnitFunctor { - static void compute(const platform::DeviceContext &context, - hl_gru_value value, int frameSize, int batchSize, - activation_mode_t active_node, - activation_mode_t active_gate) { +struct GRUUnitFunctor { + static void compute(const platform::CPUDeviceContext &context, + GRUMetaValue value, int frame_size, int batch_size, + const detail::ActivationType active_node, + const detail::ActivationType active_gate) { #ifndef __NVCC__ - if (value.prevOutValue) { - math::gemm( - context, false, false, batchSize, frameSize * 2, frameSize, 1, - value.prevOutValue, frameSize, value.gateWeight, frameSize * 2, 1, - value.gateValue, frameSize * 3); + 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); } detail::forward_reset_output(detail::forward::gru_resetOutput(), value, - frameSize, batchSize, active_gate); + frame_size, batch_size, active_gate); - if (value.prevOutValue) { - math::gemm( - context, false, false, batchSize, frameSize, frameSize, 1, - value.resetOutputValue, frameSize, value.stateWeight, frameSize, 1, - value.gateValue + frameSize * 2, frameSize * 3); + 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); } detail::forward_final_output(detail::forward::gru_finalOutput(), value, - frameSize, batchSize, active_node); + frame_size, batch_size, active_node); #endif } }; template -struct GRUUnitGradFunctor { - static void compute(const platform::DeviceContext &context, - hl_gru_value value, hl_gru_grad grad, int frameSize, - int batchSize, activation_mode_t active_node, - activation_mode_t active_gate) { +struct GRUUnitGradFunctor { + static void compute(const platform::CPUDeviceContext &context, + GRUMetaValue value, GRUMetaGrad grad, + int frame_size, int batch_size, + const detail::ActivationType active_node, + const detail::ActivationType active_gate) { #ifndef __NVCC__ detail::backward_state_grad(detail::backward::gru_stateGrad(), value, - grad, frameSize, batchSize, active_node); + grad, frame_size, batch_size, active_node); - if (value.prevOutValue && grad.prevOutGrad) { - math::gemm( - context, false, true, batchSize, frameSize, frameSize, 1, - grad.gateGrad + frameSize * 2, frameSize * 3, value.stateWeight, - frameSize, 0, grad.resetOutputGrad, frameSize); + 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); - if (grad.stateWeightGrad) { - math::gemm( - context, true, false, frameSize, frameSize, batchSize, 1, - value.resetOutputValue, frameSize, grad.gateGrad + frameSize * 2, - frameSize * 3, 1, grad.stateWeightGrad, frameSize); + 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); } } detail::backward_reset_grad(detail::backward::gru_resetGrad(), value, - grad, frameSize, batchSize, active_gate); + grad, frame_size, batch_size, active_gate); - if (grad.prevOutGrad && value.prevOutValue) { - math::gemm( - context, false, true, batchSize, frameSize, frameSize * 2, 1, - grad.gateGrad, frameSize * 3, value.gateWeight, frameSize * 2, 1, - grad.prevOutGrad, frameSize); + 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); - if (grad.gateWeightGrad) { - math::gemm( - context, true, false, frameSize, frameSize * 2, batchSize, 1, - value.prevOutValue, frameSize, grad.gateGrad, frameSize * 3, 1, - grad.gateWeightGrad, frameSize * 2); + 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); } } #endif } }; -template struct GRUUnitFunctor; -template struct GRUUnitFunctor; -template struct GRUUnitGradFunctor; -template struct GRUUnitGradFunctor; +template struct GRUUnitFunctor; +template struct GRUUnitFunctor; +template struct GRUUnitGradFunctor; +template struct GRUUnitGradFunctor; } // namespace math } // namespace operators diff --git a/paddle/operators/math/gru_compute.cu b/paddle/operators/math/gru_compute.cu index 7b9e54ac029f6aa00553338435684097d6d02b25..d5a0e630ea0eadea990988c3170395c842a91900 100644 --- a/paddle/operators/math/gru_compute.cu +++ b/paddle/operators/math/gru_compute.cu @@ -19,159 +19,159 @@ namespace operators { namespace math { template -struct GRUUnitFunctor { - static void compute(const platform::DeviceContext &context, - hl_gru_value value, int frameSize, int batchSize, - activation_mode_t active_node, - activation_mode_t active_gate) { - auto stream = - reinterpret_cast(context).stream(); +struct GRUUnitFunctor { + static void compute(const platform::CUDADeviceContext &context, + GRUMetaValue value, int frame_size, int batch_size, + const detail::ActivationType active_node, + const detail::ActivationType active_gate) { + auto stream = context.stream(); dim3 threads; dim3 grid; - if (batchSize == 1) { - int framePerBlock = frameSize <= 1024 ? frameSize : 1024; - int frameBlocks = (frameSize + 1024 - 1) / 1024; - threads = dim3(framePerBlock, 1); - grid = dim3(frameBlocks, 1); + if (batch_size == 1) { + int frame_per_block = frame_size <= 1024 ? frame_size : 1024; + int frame_blocks = (frame_size + 1024 - 1) / 1024; + threads = dim3(frame_per_block, 1); + grid = dim3(frame_blocks, 1); } else { threads = dim3(32, 32); - grid = dim3((frameSize + 32 - 1) / 32, (batchSize + 32 - 1) / 32); + grid = dim3((frame_size + 32 - 1) / 32, (batch_size + 32 - 1) / 32); } - if (value.prevOutValue) { - math::gemm( - context, false, false, batchSize, frameSize * 2, frameSize, 1, - value.prevOutValue, frameSize, value.gateWeight, frameSize * 2, 1, - value.gateValue, frameSize * 3); + 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); } - if (batchSize == 1) { + if (batch_size == 1) { detail::KeGruForwardResetOutput, - /* isBatch= */ false, + /* is_batch= */ false, T><<>>( - detail::forward::gru_resetOutput(), value.gateValue, - value.resetOutputValue, value.prevOutValue, frameSize, batchSize, - active_gate); + detail::forward::gru_resetOutput(), value.gate_value, + value.reset_output_value, value.prev_out_value, frame_size, + batch_size, active_gate); } else { detail::KeGruForwardResetOutput, - /* isBatch= */ true, + /* is_batch= */ true, T><<>>( - detail::forward::gru_resetOutput(), value.gateValue, - value.resetOutputValue, value.prevOutValue, frameSize, batchSize, - active_gate); + detail::forward::gru_resetOutput(), value.gate_value, + value.reset_output_value, value.prev_out_value, frame_size, + batch_size, active_gate); } - if (value.prevOutValue) { - math::gemm( - context, false, false, batchSize, frameSize, frameSize, 1, - value.resetOutputValue, frameSize, value.stateWeight, frameSize, 1, - value.gateValue + frameSize * 2, frameSize * 3); + 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); } - if (batchSize == 1) { + if (batch_size == 1) { detail::KeGruForwardFinalOutput, - /* isBatch= */ false, + /* is_batch= */ false, T><<>>( - detail::forward::gru_finalOutput(), value.gateValue, - value.prevOutValue, value.outputValue, frameSize, batchSize, + detail::forward::gru_finalOutput(), value.gate_value, + value.prev_out_value, value.output_value, frame_size, batch_size, active_node); } else { detail::KeGruForwardFinalOutput, - /* isBatch= */ true, + /* is_batch= */ true, T><<>>( - detail::forward::gru_finalOutput(), value.gateValue, - value.prevOutValue, value.outputValue, frameSize, batchSize, + detail::forward::gru_finalOutput(), value.gate_value, + value.prev_out_value, value.output_value, frame_size, batch_size, active_node); } } }; template -struct GRUUnitGradFunctor { - static void compute(const platform::DeviceContext &context, - hl_gru_value value, hl_gru_grad grad, int frameSize, - int batchSize, activation_mode_t active_node, - activation_mode_t active_gate) { - auto stream = - reinterpret_cast(context).stream(); +struct GRUUnitGradFunctor { + static void compute(const platform::CUDADeviceContext &context, + GRUMetaValue value, GRUMetaGrad grad, + int frame_size, int batch_size, + const detail::ActivationType active_node, + const detail::ActivationType active_gate) { + auto stream = context.stream(); dim3 threads; dim3 grid; - if (batchSize == 1) { - int framePerBlock = frameSize <= 1024 ? frameSize : 1024; - int frameBlocks = (frameSize + 1024 - 1) / 1024; - threads = dim3(framePerBlock, 1); - grid = dim3(frameBlocks, 1); + if (batch_size == 1) { + int frame_per_block = frame_size <= 1024 ? frame_size : 1024; + int frame_blocks = (frame_size + 1024 - 1) / 1024; + threads = dim3(frame_per_block, 1); + grid = dim3(frame_blocks, 1); } else { threads = dim3(32, 32); - grid = dim3((frameSize + 32 - 1) / 32, (batchSize + 32 - 1) / 32); + grid = dim3((frame_size + 32 - 1) / 32, (batch_size + 32 - 1) / 32); } - if (batchSize == 1) { + if (batch_size == 1) { detail::KeGruBackwardStateGrad< detail::backward::gru_stateGrad, - /* isBatch= */ false><<>>( - detail::backward::gru_stateGrad(), value.gateValue, grad.gateGrad, - value.prevOutValue, grad.prevOutGrad, grad.outputGrad, frameSize, - batchSize, active_node); + /* is_batch= */ false><<>>( + detail::backward::gru_stateGrad(), value.gate_value, + grad.gate_grad, value.prev_out_value, grad.prev_out_grad, + grad.output_grad, frame_size, batch_size, active_node); } else { detail::KeGruBackwardStateGrad< detail::backward::gru_stateGrad, - /* isBatch= */ true><<>>( - detail::backward::gru_stateGrad(), value.gateValue, grad.gateGrad, - value.prevOutValue, grad.prevOutGrad, grad.outputGrad, frameSize, - batchSize, active_node); + /* is_batch= */ true><<>>( + detail::backward::gru_stateGrad(), value.gate_value, + grad.gate_grad, value.prev_out_value, grad.prev_out_grad, + grad.output_grad, frame_size, batch_size, active_node); } - if (value.prevOutValue && grad.prevOutGrad) { - math::gemm( - context, false, true, batchSize, frameSize, frameSize, 1, - grad.gateGrad + frameSize * 2, frameSize * 3, value.stateWeight, - frameSize, 0, grad.resetOutputGrad, frameSize); - - if (grad.stateWeightGrad) { - math::gemm( - context, true, false, frameSize, frameSize, batchSize, 1, - value.resetOutputValue, frameSize, grad.gateGrad + frameSize * 2, - frameSize * 3, 1, grad.stateWeightGrad, frameSize); + 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); + + 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); } } - if (batchSize == 1) { + if (batch_size == 1) { detail::KeGruBackwardResetGrad< detail::backward::gru_resetGrad, - /* isBatch= */ false><<>>( - detail::backward::gru_resetGrad(), value.gateValue, grad.gateGrad, - value.prevOutValue, grad.prevOutGrad, grad.resetOutputGrad, frameSize, - batchSize, active_gate); + /* is_batch= */ false><<>>( + detail::backward::gru_resetGrad(), value.gate_value, + grad.gate_grad, value.prev_out_value, grad.prev_out_grad, + grad.reset_output_grad, frame_size, batch_size, active_gate); } else { detail::KeGruBackwardResetGrad< detail::backward::gru_resetGrad, - /* isBatch= */ true><<>>( - detail::backward::gru_resetGrad(), value.gateValue, grad.gateGrad, - value.prevOutValue, grad.prevOutGrad, grad.resetOutputGrad, frameSize, - batchSize, active_gate); + /* is_batch= */ true><<>>( + detail::backward::gru_resetGrad(), value.gate_value, + grad.gate_grad, value.prev_out_value, grad.prev_out_grad, + grad.reset_output_grad, frame_size, batch_size, active_gate); } - if (grad.prevOutGrad && value.prevOutValue) { - math::gemm( - context, false, true, batchSize, frameSize, frameSize * 2, 1, - grad.gateGrad, frameSize * 3, value.gateWeight, frameSize * 2, 1, - grad.prevOutGrad, frameSize); - - if (grad.gateWeightGrad) { - math::gemm( - context, true, false, frameSize, frameSize * 2, batchSize, 1, - value.prevOutValue, frameSize, grad.gateGrad, frameSize * 3, 1, - grad.gateWeightGrad, frameSize * 2); + 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); + + 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); } } } }; -template struct GRUUnitFunctor; -template struct GRUUnitFunctor; -template struct GRUUnitGradFunctor; -template struct GRUUnitGradFunctor; +template struct GRUUnitFunctor; +template struct GRUUnitFunctor; +template struct GRUUnitGradFunctor; +template struct GRUUnitGradFunctor; } // namespace math } // namespace operators diff --git a/paddle/operators/math/gru_compute.h b/paddle/operators/math/gru_compute.h index 1475fb38104f353857dfd968e46af98a6d52c52a..bf69147b506661692a6d71823043cd3506ea8b5d 100644 --- a/paddle/operators/math/gru_compute.h +++ b/paddle/operators/math/gru_compute.h @@ -11,7 +11,7 @@ limitations under the License. */ #pragma once -#include "paddle/operators/math/lstm_compute.h" +#include "paddle/operators/math/detail/activation_functions.h" #include "paddle/platform/device_context.h" #include "paddle/platform/enforce.h" @@ -19,41 +19,40 @@ namespace paddle { namespace operators { namespace math { -// TODO(guosheng): refine code style in gru_compute template -struct hl_gru_value { - T *gateWeight; - T *stateWeight; - T *gateValue; - T *resetOutputValue; - T *outputValue; - T *prevOutValue; +struct GRUMetaValue { + T *gate_weight; + T *state_weight; + T *gate_value; + T *reset_output_value; + T *output_value; + T *prev_out_value; }; template -struct hl_gru_grad { - T *gateWeightGrad; - T *stateWeightGrad; - T *gateGrad; - T *resetOutputGrad; - T *outputGrad; - T *prevOutGrad; +struct GRUMetaGrad { + T *gate_weight_grad; + T *state_weight_grad; + T *gate_grad; + T *reset_output_grad; + T *output_grad; + T *prev_out_grad; }; -template +template struct GRUUnitFunctor { - static void compute(const platform::DeviceContext &context, - hl_gru_value value, int frameSize, int batchSize, - activation_mode_t active_node, - activation_mode_t active_gate); + static void compute(const DeviceContext &context, GRUMetaValue value, + int frame_size, int batch_size, + const detail::ActivationType active_node, + const detail::ActivationType active_gate); }; -template +template struct GRUUnitGradFunctor { - static void compute(const platform::DeviceContext &context, - hl_gru_value value, hl_gru_grad grad, int frameSize, - int batchSize, activation_mode_t active_node, - activation_mode_t active_gate); + static void compute(const DeviceContext &context, GRUMetaValue value, + GRUMetaGrad grad, int frame_size, int batch_size, + const detail::ActivationType active_node, + const detail::ActivationType active_gate); }; } // namespace math diff --git a/paddle/operators/math/im2col.cc b/paddle/operators/math/im2col.cc index c10c44c52076c8ee56eee3a0d82c31df70a1c9c7..c2633b2e16434558d16f699a701e7b8cf1de8342 100644 --- a/paddle/operators/math/im2col.cc +++ b/paddle/operators/math/im2col.cc @@ -25,9 +25,9 @@ namespace math { */ template class Im2ColFunctor { + platform::CPUDeviceContext, T> { public: - void operator()(const platform::DeviceContext& context, + void operator()(const platform::CPUDeviceContext& context, const framework::Tensor& im, const std::vector& dilation, const std::vector& stride, const std::vector& padding, framework::Tensor* col) { @@ -61,14 +61,13 @@ class Im2ColFunctor(); T* col_data = col->data(); - for (int c = 0; c < channels_col; ++c) { int w_offset = c % filter_width; int h_offset = (c / filter_width) % filter_height; - int c_im = c / filter_width / filter_height; + int c_im = c / (filter_width * filter_height); for (int h = 0; h < col_height; ++h) { + int im_row_idx = h * stride[0] - padding[0] + h_offset * dilation[0]; for (int w = 0; w < col_width; ++w) { - int im_row_idx = h * stride[0] - padding[0] + h_offset * dilation[0]; int im_col_idx = w * stride[1] - padding[1] + w_offset * dilation[1]; int col_idx = (c * col_height + h) * col_width + w; int im_idx = (im_row_idx + c_im * im_height) * im_width + im_col_idx; @@ -90,9 +89,9 @@ class Im2ColFunctor class Col2ImFunctor { + platform::CPUDeviceContext, T> { public: - void operator()(const platform::DeviceContext& context, + void operator()(const platform::CPUDeviceContext& context, const framework::Tensor& col, const std::vector& dilation, const std::vector& stride, @@ -130,16 +129,14 @@ class Col2ImFunctor= 0 && (im_row_idx) < im_height && (im_col_idx) >= 0 && (im_col_idx) < im_width) { - im_row_idx += c_im * im_height; - im_data[im_row_idx * im_width + im_col_idx] += + im_data[(im_row_idx + c_im * im_height) * im_width + im_col_idx] += col_data[(c * col_height + h) * col_width + w]; } } @@ -149,13 +146,13 @@ class Col2ImFunctor; + platform::CPUDeviceContext, float>; template class Im2ColFunctor; + platform::CPUDeviceContext, double>; template class Col2ImFunctor; + platform::CPUDeviceContext, float>; template class Col2ImFunctor; + platform::CPUDeviceContext, double>; /* * im = [input_channels, input_height, input_width] @@ -164,9 +161,9 @@ template class Col2ImFunctor class Im2ColFunctor { + platform::CPUDeviceContext, T> { public: - void operator()(const platform::DeviceContext& context, + void operator()(const platform::CPUDeviceContext& context, const framework::Tensor& im, const std::vector& dilation, const std::vector& stride, const std::vector& padding, framework::Tensor* col) { @@ -199,12 +196,13 @@ class Im2ColFunctor class Col2ImFunctor { + platform::CPUDeviceContext, T> { public: - void operator()(const platform::DeviceContext& context, + void operator()(const platform::CPUDeviceContext& context, const framework::Tensor& col, const std::vector& dilation, const std::vector& stride, @@ -271,12 +269,13 @@ class Col2ImFunctor= 0 && im_row_offset < im_height && im_col_offset >= 0 && im_col_offset < im_width) { int im_offset = @@ -300,13 +300,13 @@ class Col2ImFunctor; + platform::CPUDeviceContext, float>; template class Im2ColFunctor; + platform::CPUDeviceContext, double>; template class Col2ImFunctor; + platform::CPUDeviceContext, float>; template class Col2ImFunctor; + platform::CPUDeviceContext, double>; } // namespace math } // namespace operators diff --git a/paddle/operators/math/im2col.cu b/paddle/operators/math/im2col.cu index bf7894243919571c2ab15d53690b1ef05bfcc6ee..a88e837b030f286cce272f99ad7991c70336e4a9 100644 --- a/paddle/operators/math/im2col.cu +++ b/paddle/operators/math/im2col.cu @@ -58,9 +58,9 @@ __global__ void im2col(const T* data_im, int num_outs, int im_height, */ template class Im2ColFunctor { + platform::CUDADeviceContext, T> { public: - void operator()(const platform::DeviceContext& context, + void operator()(const platform::CUDADeviceContext& context, const framework::Tensor& im, const std::vector& dilation, const std::vector& stride, const std::vector& padding, framework::Tensor* col) { @@ -96,9 +96,7 @@ class Im2ColFunctor<<(context) - .stream()>>>( + im2col<<>>( im.data(), num_outputs, im_height, im_width, dilation[0], dilation[1], filter_height, filter_width, stride[0], stride[1], padding[0], padding[1], col_height, col_width, col->data()); @@ -160,9 +158,9 @@ __global__ void col2im(int n, const T* data_col, int im_height, int im_width, */ template class Col2ImFunctor { + platform::CUDADeviceContext, T> { public: - void operator()(const platform::DeviceContext& context, + void operator()(const platform::CUDADeviceContext& context, const framework::Tensor& col, const std::vector& dilation, const std::vector& stride, @@ -203,9 +201,7 @@ class Col2ImFunctor<<(context) - .stream()>>>( + col2im<<>>( num_kernels, col.data(), im_height, im_width, dilation[0], dilation[1], filter_height, filter_width, stride[0], stride[1], padding[0], padding[2], col_height, col_width, im->data()); @@ -213,13 +209,13 @@ class Col2ImFunctor; + platform::CUDADeviceContext, float>; template class Im2ColFunctor; + platform::CUDADeviceContext, double>; template class Col2ImFunctor; + platform::CUDADeviceContext, float>; template class Col2ImFunctor; + platform::CUDADeviceContext, double>; template __global__ void im2colOCF(const T* im_data, int im_channels, int im_height, @@ -260,9 +256,9 @@ __global__ void im2colOCF(const T* im_data, int im_channels, int im_height, */ template class Im2ColFunctor { + platform::CUDADeviceContext, T> { public: - void operator()(const platform::DeviceContext& context, + void operator()(const platform::CUDADeviceContext& context, const framework::Tensor& im, const std::vector& dilation, const std::vector& stride, const std::vector& padding, framework::Tensor* col) { @@ -310,9 +306,7 @@ class Im2ColFunctor<<(context) - .stream()>>>( + im2colOCF<<>>( im.data(), im_channels, im_height, im_width, filter_height, filter_width, stride[0], stride[1], padding[0], padding[1], col_height, col_width, col->data()); @@ -358,9 +352,9 @@ __global__ void col2imOCF(const T* col_data, int im_channels, int im_height, */ template class Col2ImFunctor { + platform::CUDADeviceContext, T> { public: - void operator()(const platform::DeviceContext& context, + void operator()(const platform::CUDADeviceContext& context, const framework::Tensor& col, const std::vector& dilation, const std::vector& stride, @@ -409,9 +403,7 @@ class Col2ImFunctor<<(context) - .stream()>>>( + col2imOCF<<>>( col.data(), im_channels, im_height, im_width, filter_height, filter_width, stride[0], stride[1], padding[0], padding[1], col_height, col_width, im->data()); @@ -419,13 +411,13 @@ class Col2ImFunctor; + platform::CUDADeviceContext, float>; template class Im2ColFunctor; + platform::CUDADeviceContext, double>; template class Col2ImFunctor; + platform::CUDADeviceContext, float>; template class Col2ImFunctor; + platform::CUDADeviceContext, double>; } // namespace math } // namespace operators diff --git a/paddle/operators/math/im2col.h b/paddle/operators/math/im2col.h index deb60051beef56437cf75f0fa2cef90bbc0a209a..38f2c9fe0adf80a2a4355a45bebb9ba0f341d1ab 100644 --- a/paddle/operators/math/im2col.h +++ b/paddle/operators/math/im2col.h @@ -15,6 +15,7 @@ limitations under the License. */ #pragma once #include "paddle/framework/tensor.h" +#include "paddle/framework/tensor_util.h" #include "paddle/platform/device_context.h" namespace paddle { @@ -78,20 +79,19 @@ enum class ColFormat { kCFO = 0, kOCF = 1 }; * \note The caller needs to ensure that imShape.inputChannels is equal to * colShape.inputChannels. */ -template +template class Im2ColFunctor { public: - void operator()(const platform::DeviceContext& context, - const framework::Tensor& im, const std::vector& dilation, + void operator()(const DeviceContext& context, const framework::Tensor& im, + const std::vector& dilation, const std::vector& stride, const std::vector& padding, framework::Tensor* col); }; -template +template class Col2ImFunctor { public: - void operator()(const platform::DeviceContext& context, - const framework::Tensor& col, + void operator()(const DeviceContext& context, const framework::Tensor& col, const std::vector& dilation, const std::vector& stride, const std::vector& padding, framework::Tensor* im); diff --git a/paddle/operators/math/im2col_test.cc b/paddle/operators/math/im2col_test.cc index 10c28da72ba9d3b94bb59c5cf00e7f5a2f28fd06..26c038e435827b401d723ee6eef2255a89670f46 100644 --- a/paddle/operators/math/im2col_test.cc +++ b/paddle/operators/math/im2col_test.cc @@ -16,7 +16,7 @@ limitations under the License. */ #include #include -template +template void testIm2col() { paddle::framework::Tensor input_tmp; paddle::framework::Tensor input; @@ -59,22 +59,11 @@ void testIm2col() { memcpy(input_ptr, arr, 6 * sizeof(float)); auto* place = new Place(); - paddle::platform::DeviceContext* context; - if (paddle::platform::is_cpu_place(*place)) { - context = - new paddle::platform::CPUDeviceContext(paddle::platform::CPUPlace()); - } else { -#ifdef PADDLE_WITH_CUDA - context = - new paddle::platform::CUDADeviceContext(paddle::platform::GPUPlace()); -#else - PADDLE_THROW("no GPU support"); -#endif // PADDLE_WITH_CUDA - } + DeviceContext* context = new DeviceContext(*place); if (paddle::platform::is_cpu_place(*place)) { input = input_tmp; } else { - input.CopyFrom(input_tmp, *place, *context); + CopyFrom(input_tmp, *place, *context, &input); } output_cfo.mutable_data( {1, filter_size, filter_size, output_height, output_width}, *place); @@ -83,10 +72,10 @@ void testIm2col() { // Im2Col paddle::operators::math::Im2ColFunctor< - paddle::operators::math::ColFormat::kCFO, Place, float> + paddle::operators::math::ColFormat::kCFO, DeviceContext, float> im2col; paddle::operators::math::Im2ColFunctor< - paddle::operators::math::ColFormat::kOCF, Place, float> + paddle::operators::math::ColFormat::kOCF, DeviceContext, float> im2col_ocf; im2col(*context, input, dilation, stride, padding, &output_cfo); @@ -99,7 +88,7 @@ void testIm2col() { if (paddle::platform::is_cpu_place(*place)) { out_cfo_ptr = output_cfo.data(); } else { - output_tmp.CopyFrom(output_cfo, paddle::platform::CPUPlace(), *context); + CopyFrom(output_cfo, paddle::platform::CPUPlace(), *context, &output_tmp); out_cfo_ptr = output_tmp.data(); } for (int i = 0; i < 6; ++i) { @@ -110,7 +99,7 @@ void testIm2col() { if (paddle::platform::is_cpu_place(*place)) { out_ocf_ptr = output_ocf.data(); } else { - output_tmp.CopyFrom(output_ocf, paddle::platform::CPUPlace(), *context); + CopyFrom(output_ocf, paddle::platform::CPUPlace(), *context, &output_tmp); out_ocf_ptr = output_tmp.data(); } for (int i = 0; i < 6; ++i) { @@ -119,10 +108,10 @@ void testIm2col() { // Col2Im: kCFO paddle::operators::math::Col2ImFunctor< - paddle::operators::math::ColFormat::kCFO, Place, float> + paddle::operators::math::ColFormat::kCFO, DeviceContext, float> col2im; paddle::operators::math::Col2ImFunctor< - paddle::operators::math::ColFormat::kOCF, Place, float> + paddle::operators::math::ColFormat::kOCF, DeviceContext, float> col2im_ocf; float col2im_data[] = {0, 2, 2, 3, 8, 5}; @@ -130,7 +119,7 @@ void testIm2col() { if (paddle::platform::is_cpu_place(*place)) { input = input_tmp; } else { - input.CopyFrom(input_tmp, *place, *context); + CopyFrom(input_tmp, *place, *context, &input); } col2im(*context, output_cfo, dilation, stride, padding, &input); @@ -139,7 +128,7 @@ void testIm2col() { if (paddle::platform::is_cpu_place(*place)) { in_ptr = input.data(); } else { - input_tmp.CopyFrom(input, paddle::platform::CPUPlace(), *context); + CopyFrom(input, paddle::platform::CPUPlace(), *context, &input_tmp); in_ptr = input_tmp.data(); } for (int i = 0; i < 6; ++i) { @@ -151,7 +140,7 @@ void testIm2col() { if (paddle::platform::is_cpu_place(*place)) { input = input_tmp; } else { - input.CopyFrom(input_tmp, *place, *context); + CopyFrom(input_tmp, *place, *context, &input); } col2im_ocf(*context, output_ocf, dilation, stride, padding, &input); @@ -159,7 +148,7 @@ void testIm2col() { if (paddle::platform::is_cpu_place(*place)) { in_ptr = input.data(); } else { - input_tmp.CopyFrom(input, paddle::platform::CPUPlace(), *context); + CopyFrom(input, paddle::platform::CPUPlace(), *context, &input_tmp); in_ptr = input_tmp.data(); } for (int i = 0; i < 6; ++i) { @@ -168,8 +157,9 @@ void testIm2col() { } TEST(math, im2col) { - testIm2col(); + testIm2col(); #ifdef PADDLE_WITH_CUDA - testIm2col(); + testIm2col(); #endif } diff --git a/paddle/operators/math/lstm_compute.cc b/paddle/operators/math/lstm_compute.cc index 0febf8e3b70111d12f858cf6259a2801a42d9a90..d453102ecefc9d79e1f4474ba94be0eb69a87c85 100644 --- a/paddle/operators/math/lstm_compute.cc +++ b/paddle/operators/math/lstm_compute.cc @@ -21,61 +21,61 @@ namespace operators { namespace math { template -struct LstmUnitFunctor { - static void compute(const platform::DeviceContext& context, +struct LstmUnitFunctor { + static void compute(const platform::CPUDeviceContext& context, LstmMetaValue value, int frame_size, int batch_size, - const std::string& gate_act, const std::string& cell_act, - const std::string& cand_act) { + const detail::ActivationType& gate_act, + const detail::ActivationType& cell_act, + const detail::ActivationType& cand_act) { for (int b = 0; b < batch_size; b++) { detail::cpu_lstm_forward(detail::forward::lstm(), value, frame_size, - ActiveType(cand_act), ActiveType(gate_act), - ActiveType(cell_act)); - value.gateValue += frame_size * 4; - value.stateValue += frame_size; - value.stateActiveValue += frame_size; - value.outputValue += frame_size; - if (value.prevStateValue) { - value.prevStateValue += frame_size; + cand_act, gate_act, cell_act); + value.gate_value += frame_size * 4; + value.state_value += frame_size; + value.state_active_value += frame_size; + value.output_value += frame_size; + if (value.prev_state_value) { + value.prev_state_value += frame_size; } } } }; template -struct LstmUnitGradFunctor { - static void compute(const platform::DeviceContext& context, +struct LstmUnitGradFunctor { + static void compute(const platform::CPUDeviceContext& context, LstmMetaValue value, LstmMetaGrad grad, int frame_size, int batch_size, - const std::string& gate_act, const std::string& cell_act, - const std::string& cand_act) { + const detail::ActivationType& gate_act, + const detail::ActivationType& cell_act, + const detail::ActivationType& cand_act) { for (int b = 0; b < batch_size; b++) { detail::cpu_lstm_backward(detail::backward::lstm(), value, grad, - frame_size, ActiveType(cand_act), - ActiveType(gate_act), ActiveType(cell_act)); + frame_size, cand_act, gate_act, cell_act); - value.gateValue += frame_size * 4; - value.stateValue += frame_size; - value.stateActiveValue += frame_size; - value.outputValue += frame_size; - if (value.prevStateValue) { - value.prevStateValue += frame_size; + value.gate_value += frame_size * 4; + value.state_value += frame_size; + value.state_active_value += frame_size; + value.output_value += frame_size; + if (value.prev_state_value) { + value.prev_state_value += frame_size; } - grad.gateGrad += frame_size * 4; - grad.stateGrad += frame_size; - grad.stateActiveGrad += frame_size; - grad.outputGrad += frame_size; - if (grad.prevStateGrad) { - grad.prevStateGrad += frame_size; + grad.gate_grad += frame_size * 4; + grad.state_grad += frame_size; + grad.state_active_grad += frame_size; + grad.output_grad += frame_size; + if (grad.prev_state_grad) { + grad.prev_state_grad += frame_size; } } } }; -template class LstmUnitFunctor; -template class LstmUnitFunctor; -template class LstmUnitGradFunctor; -template class LstmUnitGradFunctor; +template class LstmUnitFunctor; +template class LstmUnitFunctor; +template class LstmUnitGradFunctor; +template class LstmUnitGradFunctor; } // namespace math } // namespace operators diff --git a/paddle/operators/math/lstm_compute.cu b/paddle/operators/math/lstm_compute.cu index b2122f2a5c08a6d9d53293833177f0ba2c3ab860..82065d699f760db6cc86bf3d6c56e51c583c6ace 100644 --- a/paddle/operators/math/lstm_compute.cu +++ b/paddle/operators/math/lstm_compute.cu @@ -21,34 +21,36 @@ namespace operators { namespace math { template -struct LstmUnitFunctor { - static void compute(const platform::DeviceContext& context, +struct LstmUnitFunctor { + static void compute(const platform::CUDADeviceContext& context, LstmMetaValue value, int frame_size, int batch_size, - const std::string& gate_act, const std::string& cell_act, - const std::string& cand_act) { + const detail::ActivationType& gate_act, + const detail::ActivationType& cell_act, + const detail::ActivationType& cand_act) { detail::gpu_lstm_forward(context, detail::forward::lstm(), value, - frame_size, batch_size, ActiveType(cand_act), - ActiveType(gate_act), ActiveType(cell_act)); + frame_size, batch_size, cand_act, gate_act, + cell_act); } }; template -struct LstmUnitGradFunctor { - static void compute(const platform::DeviceContext& context, +struct LstmUnitGradFunctor { + static void compute(const platform::CUDADeviceContext& context, LstmMetaValue value, LstmMetaGrad grad, int frame_size, int batch_size, - const std::string& gate_act, const std::string& cell_act, - const std::string& cand_act) { + const detail::ActivationType& gate_act, + const detail::ActivationType& cell_act, + const detail::ActivationType& cand_act) { detail::gpu_lstm_backward(context, detail::backward::lstm(), value, grad, - frame_size, batch_size, ActiveType(cand_act), - ActiveType(gate_act), ActiveType(cell_act)); + frame_size, batch_size, cand_act, gate_act, + cell_act); } }; -template class LstmUnitFunctor; -template class LstmUnitFunctor; -template class LstmUnitGradFunctor; -template class LstmUnitGradFunctor; +template class LstmUnitFunctor; +template class LstmUnitFunctor; +template class LstmUnitGradFunctor; +template class LstmUnitGradFunctor; } // namespace math } // namespace operators diff --git a/paddle/operators/math/lstm_compute.h b/paddle/operators/math/lstm_compute.h index 28d2c6fd3b0d8143da90c37f241072e37397f98b..e1ad6b64d201ef99d83eaa2a821356008dcc9c8e 100644 --- a/paddle/operators/math/lstm_compute.h +++ b/paddle/operators/math/lstm_compute.h @@ -14,6 +14,7 @@ limitations under the License. */ #pragma once +#include "paddle/operators/math/detail/activation_functions.h" #include "paddle/platform/device_context.h" #include "paddle/platform/enforce.h" @@ -21,69 +22,48 @@ namespace paddle { namespace operators { namespace math { -typedef enum { - HL_ACTIVATION_SIGMOID = 0, - HL_ACTIVATION_RELU = 1, - HL_ACTIVATION_TANH = 2, - HL_ACTIVATION_LINEAR = 3, - HL_ACTIVATION_END -} activation_mode_t; - template struct LstmMetaValue { - T *gateValue; - T *prevStateValue; - T *stateValue; - T *stateActiveValue; - T *outputValue; - T *checkIg; - T *checkFg; - T *checkOg; + T *gate_value; + T *prev_state_value; + T *state_value; + T *state_active_value; + T *output_value; + T *check_ig; + T *check_fg; + T *check_og; }; template struct LstmMetaGrad { - T *gateGrad; - T *prevStateGrad; - T *stateGrad; - T *stateActiveGrad; - T *outputGrad; - T *checkIgGrad; - T *checkFgGrad; - T *checkOgGrad; + T *gate_grad; + T *prev_state_grad; + T *state_grad; + T *state_active_grad; + T *output_grad; + T *check_ig_grad; + T *check_fg_grad; + T *check_og_grad; }; -inline activation_mode_t ActiveType(const std::string &type) { - if (type == "sigmoid") { - return HL_ACTIVATION_SIGMOID; - } else if (type == "relu") { - return HL_ACTIVATION_RELU; - } else if (type == "tanh") { - return HL_ACTIVATION_TANH; - } else if (type == "linear" || type == "identity" || type == "") { - return HL_ACTIVATION_LINEAR; - } else { - PADDLE_THROW("Do not support activation type."); - } -} - -template +template class LstmUnitFunctor { public: - static void compute(const platform::DeviceContext &context, - LstmMetaValue value, int frame_size, int batch_size, - const std::string &gate_act, const std::string &cell_act, - const std::string &cand_act); + static void compute(const DeviceContext &context, LstmMetaValue value, + int frame_size, int batch_size, + const detail::ActivationType &gate_act, + const detail::ActivationType &cell_act, + const detail::ActivationType &cand_act); }; -template +template class LstmUnitGradFunctor { public: - static void compute(const platform::DeviceContext &context, - LstmMetaValue value, LstmMetaGrad grad, - int frame_size, int batch_size, - const std::string &gate_act, const std::string &cell_act, - const std::string &cand_act); + static void compute(const DeviceContext &context, LstmMetaValue value, + LstmMetaGrad grad, int frame_size, int batch_size, + const detail::ActivationType &gate_act, + const detail::ActivationType &cell_act, + const detail::ActivationType &cand_act); }; } // namespace math diff --git a/paddle/operators/math/math_function.cc b/paddle/operators/math/math_function.cc index 2e333a8cde721f8e65dbf2cf5e3aac6272172cc0..dcf4b85e1aadf88e4b1ca70ac7e8b5416fc58cd8 100644 --- a/paddle/operators/math/math_function.cc +++ b/paddle/operators/math/math_function.cc @@ -21,13 +21,11 @@ namespace operators { namespace math { template <> -void gemm(const platform::DeviceContext& 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) { +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; @@ -36,13 +34,11 @@ void gemm(const platform::DeviceContext& context, } template <> -void gemm(const platform::DeviceContext& 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) { +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; @@ -51,35 +47,32 @@ void gemm(const platform::DeviceContext& context, } template <> -void gemm(const platform::DeviceContext& 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) { +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::DeviceContext& 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) { +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::DeviceContext& context, const framework::Tensor& matrix_a, - bool trans_a, const framework::Tensor& matrix_b, bool trans_b, float alpha, +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(); @@ -99,15 +92,16 @@ void matmul( CBLAS_TRANSPOSE transA = (trans_a == false) ? CblasNoTrans : CblasTrans; CBLAS_TRANSPOSE transB = (trans_b == false) ? CblasNoTrans : CblasTrans; - gemm( + gemm( context, transA, transB, M, N, K, alpha, matrix_a.data(), matrix_b.data(), beta, matrix_out->data()); } template <> -void matmul( - const platform::DeviceContext& context, const framework::Tensor& matrix_a, - bool trans_a, const framework::Tensor& matrix_b, bool trans_b, double alpha, +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(); @@ -127,16 +121,16 @@ void matmul( CBLAS_TRANSPOSE transA = (trans_a == false) ? CblasNoTrans : CblasTrans; CBLAS_TRANSPOSE transB = (trans_b == false) ? CblasNoTrans : CblasTrans; - gemm( + gemm( context, transA, transB, M, N, K, alpha, matrix_a.data(), matrix_b.data(), beta, matrix_out->data()); } -#ifdef PADDLE_USE_MKLML +#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::DeviceContext& context, const CBLAS_TRANSPOSE transA, +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 int strideA, const int strideB) { @@ -157,8 +151,8 @@ void batched_gemm( } template <> -void batched_gemm( - const platform::DeviceContext& context, const CBLAS_TRANSPOSE transA, +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 int strideA, const int strideB) { @@ -183,8 +177,8 @@ void batched_gemm( // functions of Intel MKL are not available. In the future, this computation // should be parallelized. template <> -void batched_gemm( - const platform::DeviceContext& context, const CBLAS_TRANSPOSE transA, +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 int strideA, const int strideB) { @@ -192,14 +186,14 @@ void batched_gemm( 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); + gemm(context, transA, transB, M, N, K, + alpha, Ak, Bk, beta, Ck); } } template <> -void batched_gemm( - const platform::DeviceContext& context, const CBLAS_TRANSPOSE transA, +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 int strideA, const int strideB) { @@ -207,55 +201,56 @@ void batched_gemm( 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); + gemm(context, transA, transB, M, N, K, + alpha, Ak, Bk, beta, Ck); } } #endif template <> -void gemv(const platform::DeviceContext& 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) { +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::DeviceContext& 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) { +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::DeviceContext& context, - const int n, const float alpha, - const float* x, float* y) { +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::DeviceContext& context, - const int n, const double alpha, - const double* x, double* y) { +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; -template struct SetConstant; -template struct SetConstant; +template struct SetConstant; +template struct SetConstant; +template struct SetConstant; +template struct SetConstant; +template struct SetConstant; -#define DEFINE_CPU_TRANS(RANK) \ - template struct Transpose; \ - template struct Transpose; +#define DEFINE_CPU_TRANS(RANK) \ + template struct Transpose; \ + template struct Transpose; \ + template struct Transpose; \ + template struct Transpose; \ + template struct Transpose; DEFINE_CPU_TRANS(1); DEFINE_CPU_TRANS(2); @@ -310,10 +305,31 @@ void set_constant(const platform::DeviceContext& context, #endif } -template struct RowwiseAdd; -template struct RowwiseAdd; -template struct ColwiseSum; -template struct ColwiseSum; +template +struct RowwiseAdd { + void operator()(const platform::CPUDeviceContext& context, + const framework::Tensor& input, + const framework::Tensor& vector, framework::Tensor* output) { + auto in_dims = input.dims(); + auto size = input.numel() / in_dims[0]; + PADDLE_ENFORCE_EQ(vector.numel(), size); + PADDLE_ENFORCE_EQ(output->dims(), in_dims); + + auto in = framework::EigenMatrix::From(input); + auto vec = framework::EigenVector::Flatten(vector); + auto out = framework::EigenMatrix::From(*output); + + for (int64_t i = 0; i < in_dims[0]; ++i) { + out.chip(i, 0) = in.chip(i, 0) + vec; + } + } +}; + +template struct RowwiseAdd; +template struct RowwiseAdd; + +template struct ColwiseSum; +template struct ColwiseSum; } // namespace math } // namespace operators diff --git a/paddle/operators/math/math_function.cu b/paddle/operators/math/math_function.cu index 58356a4b7783241ca0292829bf05dc1a8ed80c6c..d47a7f818ded61baf31e46ea3b8ae3101324111f 100644 --- a/paddle/operators/math/math_function.cu +++ b/paddle/operators/math/math_function.cu @@ -22,13 +22,11 @@ namespace operators { namespace math { template <> -void gemm(const platform::DeviceContext& 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) { +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; @@ -39,19 +37,16 @@ void gemm(const platform::DeviceContext& context, (transB == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T; PADDLE_ENFORCE(platform::dynload::cublasSgemm( - reinterpret_cast(context) - .cublas_handle(), - cuTransB, cuTransA, N, M, K, &alpha, B, ldb, A, lda, &beta, C, N)); + context.cublas_handle(), cuTransB, cuTransA, N, M, K, &alpha, B, ldb, A, + lda, &beta, C, N)); } template <> -void gemm(const platform::DeviceContext& 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) { +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; @@ -61,51 +56,45 @@ void gemm(const platform::DeviceContext& context, cublasOperation_t cuTransB = (transB == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T; PADDLE_ENFORCE(platform::dynload::cublasDgemm( - reinterpret_cast(context) - .cublas_handle(), - cuTransB, cuTransA, N, M, K, &alpha, B, ldb, A, lda, &beta, C, N)); + context.cublas_handle(), cuTransB, cuTransA, N, M, K, &alpha, B, ldb, A, + lda, &beta, C, N)); } template <> -void gemm(const platform::DeviceContext& 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) { +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( - reinterpret_cast(context) - .cublas_handle(), - cuTransB, cuTransA, N, M, K, &alpha, B, ldb, A, lda, &beta, C, ldc)); + context.cublas_handle(), cuTransB, cuTransA, N, M, K, &alpha, B, ldb, A, + lda, &beta, C, ldc)); } template <> -void gemm(const platform::DeviceContext& 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) { +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( - reinterpret_cast(context) - .cublas_handle(), - cuTransB, cuTransA, N, M, K, &alpha, B, ldb, A, lda, &beta, C, ldc)); + context.cublas_handle(), cuTransB, cuTransA, N, M, K, &alpha, B, ldb, A, + lda, &beta, C, ldc)); } template <> -void matmul( - const platform::DeviceContext& context, const framework::Tensor& matrix_a, - bool trans_a, const framework::Tensor& matrix_b, bool trans_b, float alpha, +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(); @@ -116,7 +105,7 @@ void matmul( 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 GPUPlace"); + "Matrix must all be in CUDAPlace"); int M = dim_out[0]; int N = dim_out[1]; @@ -125,15 +114,16 @@ void matmul( CBLAS_TRANSPOSE transA = (trans_a == false) ? CblasNoTrans : CblasTrans; CBLAS_TRANSPOSE transB = (trans_b == false) ? CblasNoTrans : CblasTrans; - gemm( + gemm( context, transA, transB, M, N, K, alpha, matrix_a.data(), matrix_b.data(), beta, matrix_out->data()); } template <> -void matmul( - const platform::DeviceContext& context, const framework::Tensor& matrix_a, - bool trans_a, const framework::Tensor& matrix_b, bool trans_b, double alpha, +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(); @@ -144,7 +134,7 @@ void matmul( 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 GPUPlace"); + "Matrix must all be in CUDAPlace"); int M = dim_out[0]; int N = dim_out[1]; @@ -153,14 +143,14 @@ void matmul( CBLAS_TRANSPOSE transA = (trans_a == false) ? CblasNoTrans : CblasTrans; CBLAS_TRANSPOSE transB = (trans_b == false) ? CblasNoTrans : CblasTrans; - gemm( + gemm( context, transA, transB, M, N, K, alpha, matrix_a.data(), matrix_b.data(), beta, matrix_out->data()); } template <> -void batched_gemm( - const platform::DeviceContext& context, const CBLAS_TRANSPOSE transA, +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 int strideA, const int strideB) { @@ -176,15 +166,13 @@ void batched_gemm( const int strideC = M * N; PADDLE_ENFORCE(platform::dynload::cublasSgemmStridedBatched( - reinterpret_cast(context) - .cublas_handle(), - cuTransB, cuTransA, N, M, K, &alpha, B, ldb, strideB, A, lda, strideA, - &beta, C, ldc, strideC, batchCount)); + context.cublas_handle(), cuTransB, cuTransA, N, M, K, &alpha, B, ldb, + strideB, A, lda, strideA, &beta, C, ldc, strideC, batchCount)); } template <> -void batched_gemm( - const platform::DeviceContext& context, const CBLAS_TRANSPOSE transA, +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 int strideA, const int strideB) { @@ -200,68 +188,58 @@ void batched_gemm( const int strideC = M * N; PADDLE_ENFORCE(platform::dynload::cublasDgemmStridedBatched( - reinterpret_cast(context) - .cublas_handle(), - cuTransB, cuTransA, N, M, K, &alpha, B, ldb, strideB, A, lda, strideA, - &beta, C, ldc, strideC, batchCount)); + context.cublas_handle(), cuTransB, cuTransA, N, M, K, &alpha, B, ldb, + strideB, A, lda, strideA, &beta, C, ldc, strideC, batchCount)); } template <> -void gemv(const platform::DeviceContext& 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) { +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( - reinterpret_cast(context) - .cublas_handle(), - cuTransA, N, M, &alpha, A, N, B, 1, &beta, C, 1)); + PADDLE_ENFORCE(platform::dynload::cublasSgemv(context.cublas_handle(), + cuTransA, N, M, &alpha, A, N, B, + 1, &beta, C, 1)); } template <> -void gemv(const platform::DeviceContext& 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) { +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( - reinterpret_cast(context) - .cublas_handle(), - cuTransA, N, M, &alpha, A, N, B, 1, &beta, C, 1)); + PADDLE_ENFORCE(platform::dynload::cublasDgemv(context.cublas_handle(), + cuTransA, N, M, &alpha, A, N, B, + 1, &beta, C, 1)); } template <> -void axpy(const platform::DeviceContext& context, - const int n, const float alpha, - const float* x, float* y) { - PADDLE_ENFORCE(platform::dynload::cublasSaxpy( - reinterpret_cast(context) - .cublas_handle(), - n, &alpha, x, 1, y, 1)); +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::DeviceContext& context, - const int n, const double alpha, - const double* x, double* y) { - PADDLE_ENFORCE(platform::dynload::cublasDaxpy( - reinterpret_cast(context) - .cublas_handle(), - n, &alpha, x, 1, y, 1)); +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; -template struct SetConstant; -template struct SetConstant; +template struct SetConstant; +template struct SetConstant; +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; DEFINE_GPU_TRANS(1); DEFINE_GPU_TRANS(2); @@ -277,8 +255,9 @@ struct TensorSetConstantGPU { template void operator()() const { - SetConstant functor; - functor(context_, tensor_, static_cast(value_)); + SetConstant functor; + functor(reinterpret_cast(context_), + tensor_, static_cast(value_)); } const platform::DeviceContext& context_; @@ -287,17 +266,64 @@ struct TensorSetConstantGPU { }; template <> -void set_constant_with_place( +void set_constant_with_place( const platform::DeviceContext& context, framework::Tensor* tensor, float value) { framework::VisitDataType(framework::ToDataType(tensor->type()), TensorSetConstantGPU(context, tensor, value)); } -template struct RowwiseAdd; -template struct RowwiseAdd; -template struct ColwiseSum; -template struct ColwiseSum; +template +__global__ void RowwiseAddKernel(const T* a, const T* b, T* c, int width, + int num) { + T tmp = 1.0 / width; + for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < num; + i += blockDim.x * gridDim.x) { + int h = i * tmp; + int w = i - h * width; + c[i] = a[i] + b[w]; + } +} + +template +struct RowwiseAdd { + void operator()(const platform::CUDADeviceContext& context, + const framework::Tensor& input, + const framework::Tensor& vector, framework::Tensor* output) { + auto in_dims = input.dims(); + auto size = input.numel() / in_dims[0]; + PADDLE_ENFORCE_EQ(vector.numel(), size); + PADDLE_ENFORCE_EQ(output->dims(), in_dims); + int blocks = 512; + int grids = (input.numel() + blocks - 1) / blocks; + RowwiseAddKernel<<>>( + input.data(), vector.data(), output->data(), + static_cast(in_dims[1]), static_cast(input.numel())); + } +}; + +template struct RowwiseAdd; +template struct RowwiseAdd; +template struct ColwiseSum; +// template struct ColwiseSum; +// The ColwiseSum failed in debug mode, +// and only failed for this case. So reimplemented it. +template <> +void ColwiseSum::operator()( + const platform::CUDADeviceContext& context, const framework::Tensor& input, + framework::Tensor* vector) { + auto in_dims = input.dims(); + auto size = input.numel() / in_dims[0]; + PADDLE_ENFORCE_EQ(vector->numel(), size); + framework::Tensor one; + 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()); +} } // namespace math } // namespace operators diff --git a/paddle/operators/math/math_function.h b/paddle/operators/math/math_function.h index ffb99f53808c4316ede96b04e57aec4dae4134de..8cc03c2ba0facae691a0d2b8a4f2ea768cfa5491 100644 --- a/paddle/operators/math/math_function.h +++ b/paddle/operators/math/math_function.h @@ -13,7 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once -#ifdef PADDLE_USE_MKLML +#ifdef PADDLE_WITH_MKLML #include #include #include @@ -49,6 +49,7 @@ int LAPACKE_dgetri(int matrix_layout, int n, double* a, int lda, #include "paddle/framework/eigen.h" #include "paddle/framework/tensor.h" +#include "paddle/framework/tensor_util.h" #include "paddle/platform/device_context.h" #include "paddle/platform/enforce.h" @@ -61,53 +62,51 @@ namespace math { // 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 platform::DeviceContext& context, const CBLAS_TRANSPOSE transA, +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 platform::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); +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 platform::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); +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 platform::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 int strideA, const int strideB); - -template -void gemv(const platform::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 platform::DeviceContext& context, const int n, const T alpha, - const T* x, T* y); - -template +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 int strideA, + const int 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 platform::DeviceContext& context, - const framework::Tensor& in, framework::Tensor* out, - const std::vector& axis); + void operator()(const DeviceContext& context, const framework::Tensor& in, + framework::Tensor* out, const std::vector& axis); }; -template +template struct SetConstant { - void operator()(const platform::DeviceContext& context, - framework::Tensor* tensor, T num); + void operator()(const DeviceContext& context, framework::Tensor* tensor, + T num); }; template @@ -117,17 +116,16 @@ void set_constant_with_place(const platform::DeviceContext& context, void set_constant(const platform::DeviceContext& context, framework::Tensor* tensor, float value); -template +template struct RowwiseAdd { - void operator()(const platform::DeviceContext& context, - const framework::Tensor& input, const framework::Tensor& vec, - framework::Tensor* output); + void operator()(const DeviceContext& context, const framework::Tensor& input, + const framework::Tensor& vec, framework::Tensor* output); }; -template +template struct ColwiseSum { - void operator()(const platform::DeviceContext& context, - const framework::Tensor& input, framework::Tensor* vec); + void operator()(const DeviceContext& context, const framework::Tensor& input, + framework::Tensor* vec); }; } // namespace math diff --git a/paddle/operators/math/math_function_impl.h b/paddle/operators/math/math_function_impl.h index 4dc17a4e525c52b8f696277274a7ad00a6b00a08..de591626df28e2bc3391b609f909612411398247 100644 --- a/paddle/operators/math/math_function_impl.h +++ b/paddle/operators/math/math_function_impl.h @@ -20,16 +20,17 @@ namespace paddle { namespace operators { namespace math { -template -void SetConstant::operator()(const platform::DeviceContext& context, - framework::Tensor* tensor, T num) { +template +void SetConstant::operator()(const DeviceContext& context, + framework::Tensor* tensor, + T num) { auto t = framework::EigenVector::Flatten(*tensor); - t.device(*context.GetEigenDevice()) = t.constant(static_cast(num)); + t.device(*context.eigen_device()) = t.constant(static_cast(num)); } -template -void Transpose::operator()( - const platform::DeviceContext& context, const framework::Tensor& in, +template +void Transpose::operator()( + const DeviceContext& context, const framework::Tensor& in, framework::Tensor* out, const std::vector& axis) { Eigen::array permute; for (int i = 0; i < Rank; i++) { @@ -40,43 +41,51 @@ void Transpose::operator()( auto eigen_in = framework::EigenTensor::From(in); auto eigen_out = framework::EigenTensor::From(*out); - auto* dev = context.GetEigenDevice(); + auto* dev = context.eigen_device(); eigen_out.device(*dev) = eigen_in.shuffle(permute); } -template -void RowwiseAdd::operator()(const platform::DeviceContext& context, - const framework::Tensor& input, - const framework::Tensor& vector, - framework::Tensor* output) { +template +void ColwiseSum::operator()(const DeviceContext& context, + const framework::Tensor& input, + framework::Tensor* out) { auto in_dims = input.dims(); auto size = input.numel() / in_dims[0]; - PADDLE_ENFORCE_EQ(vector.numel(), size); - PADDLE_ENFORCE_EQ(output->dims(), in_dims); + PADDLE_ENFORCE_EQ(out->numel(), size); auto in = framework::EigenMatrix::From(input); - auto vec = framework::EigenMatrix::From(vector); - auto out = framework::EigenMatrix::From(*output); - Eigen::array shape({{1, static_cast(size)}}); - Eigen::array bcast({{static_cast(in_dims[0]), 1}}); - out.device(*context.GetEigenDevice()) = - in + vec.reshape(shape).broadcast(bcast); + auto vec = framework::EigenVector::Flatten(*out); + + vec.device(*context.eigen_device()) = in.sum(Eigen::array({{0}})); } -template -void ColwiseSum::operator()(const platform::DeviceContext& context, - const framework::Tensor& input, - framework::Tensor* vector) { - auto in_dims = input.dims(); - auto size = input.numel() / in_dims[0]; - PADDLE_ENFORCE_EQ(vector->numel(), size); +// Specialize for CPU, since Eigen implement a general reduce. However, +// colwise-sum can be easily implemented. General reduce has a huge overhead in +// CPU +template +class ColwiseSum { + public: + void operator()(const platform::CPUDeviceContext& context, + const framework::Tensor& input, framework::Tensor* out) { + auto& in_dims = input.dims(); + auto height = in_dims[0]; + auto size = in_dims[1]; + PADDLE_ENFORCE_EQ(out->numel(), size); - auto vec = framework::EigenMatrix::From(*vector); - auto in = framework::EigenMatrix::From(input); - Eigen::array shape({{1, static_cast(size)}}); - vec.reshape(shape).device(*context.GetEigenDevice()) = - in.sum(Eigen::array({{0}})).reshape(shape); -} + T* out_buf = out->mutable_data(out->place()); + const T* in_buf = input.data(); + + for (size_t i = 0; i < static_cast(height); ++i) { + for (size_t j = 0; j < static_cast(size); ++j) { + if (i == 0) { + out_buf[j] = in_buf[i * size + j]; + } else { + out_buf[j] += in_buf[i * size + j]; + } + } + } + } +}; } // namespace math } // namespace operators diff --git a/paddle/operators/math/math_function_test.cc b/paddle/operators/math/math_function_test.cc index 983c9fdcffb0a67da1bc0b5b4af9420a68bd2ac1..7c6f098ca9065ded1644420a3ab47911bf7bc3b3 100644 --- a/paddle/operators/math/math_function_test.cc +++ b/paddle/operators/math/math_function_test.cc @@ -21,7 +21,7 @@ TEST(math_function, gemm_notrans_cblas) { memcpy(input3_ptr, arr3, 8 * sizeof(float)); paddle::platform::CPUDeviceContext context(*cpu_place); - paddle::operators::math::gemm( + paddle::operators::math::gemm( context, false, false, m, n, k, 1, input1_ptr, 3, input2_ptr + 1, 4, 1, input3_ptr + 1, 4); @@ -55,7 +55,7 @@ TEST(math_function, gemm_trans_clbas) { memcpy(input3_ptr, arr3, 8 * sizeof(float)); paddle::platform::CPUDeviceContext context(*cpu_place); - paddle::operators::math::gemm( + paddle::operators::math::gemm( context, false, true, m, n, k, 1, input1_ptr, 3, input2_ptr + 3, 3, 1, input3_ptr + 1, 4); @@ -74,7 +74,8 @@ TEST(math_function, zero) { auto* cpu_place = new paddle::platform::CPUPlace(); float* t = tensor.mutable_data({2, 2}, *cpu_place); paddle::platform::CPUDeviceContext context(*cpu_place); - paddle::operators::math::SetConstant + paddle::operators::math::SetConstant functor; functor(context, &tensor, 0); EXPECT_EQ(t[0], 0); @@ -110,7 +111,7 @@ void GemvTest(int m, int n, bool trans) { } paddle::platform::CPUDeviceContext context(*cpu_place); - paddle::operators::math::gemv( + paddle::operators::math::gemv( context, trans, static_cast(m), static_cast(n), 1., data_a, data_b, 0., data_c); diff --git a/paddle/operators/math/math_function_test.cu b/paddle/operators/math/math_function_test.cu index 780d17ffc6539c5f4d67ebab5476d6f646840b41..4325a79664f15cfaea48870cd503ce70cc31044f 100644 --- a/paddle/operators/math/math_function_test.cu +++ b/paddle/operators/math/math_function_test.cu @@ -13,18 +13,18 @@ TEST(math_function, notrans_mul_trans) { float arr[6] = {0, 1, 2, 3, 4, 5}; memcpy(input1_ptr, arr, 6 * sizeof(float)); - auto* gpu_place = new paddle::platform::GPUPlace(0); + auto* gpu_place = new paddle::platform::CUDAPlace(0); paddle::platform::CUDADeviceContext context(*gpu_place); - input1_gpu.CopyFrom(input1, *gpu_place, context); - input2_gpu.CopyFrom(input1, *gpu_place, context); + paddle::framework::CopyFrom(input1, *gpu_place, context, &input1_gpu); + paddle::framework::CopyFrom(input1, *gpu_place, context, &input2_gpu); out_gpu.mutable_data({2, 2}, *gpu_place); - paddle::operators::math::matmul( + paddle::operators::math::matmul( context, input1_gpu, false, input2_gpu, true, 1, &out_gpu, 0); - out.CopyFrom(out_gpu, *cpu_place, context); + paddle::framework::CopyFrom(out_gpu, *cpu_place, context, &out); float* out_ptr = out.data(); context.Wait(); @@ -47,18 +47,18 @@ TEST(math_function, trans_mul_notrans) { float arr[6] = {0, 1, 2, 3, 4, 5}; memcpy(input1_ptr, arr, 6 * sizeof(float)); - auto* gpu_place = new paddle::platform::GPUPlace(0); + auto* gpu_place = new paddle::platform::CUDAPlace(0); paddle::platform::CUDADeviceContext context(*gpu_place); - input1_gpu.CopyFrom(input1, *gpu_place, context); - input2_gpu.CopyFrom(input1, *gpu_place, context); + paddle::framework::CopyFrom(input1, *gpu_place, context, &input1_gpu); + paddle::framework::CopyFrom(input1, *gpu_place, context, &input2_gpu); out_gpu.mutable_data({3, 3}, *gpu_place); - paddle::operators::math::matmul( + paddle::operators::math::matmul( context, input1_gpu, true, input2_gpu, false, 1, &out_gpu, 0); - out.CopyFrom(out_gpu, *cpu_place, context); + paddle::framework::CopyFrom(out_gpu, *cpu_place, context, &out); float* out_ptr = out.data(); context.Wait(); @@ -96,20 +96,20 @@ TEST(math_function, gemm_notrans_cublas) { float arr3[8] = {0, 1, 2, 3, 4, 5, 6, 7}; memcpy(input3_ptr, arr3, 8 * sizeof(float)); - auto* gpu_place = new paddle::platform::GPUPlace(0); + auto* gpu_place = new paddle::platform::CUDAPlace(0); paddle::platform::CUDADeviceContext context(*gpu_place); - input1_gpu.CopyFrom(input1, *gpu_place, context); - input2_gpu.CopyFrom(input2, *gpu_place, context); - input3_gpu.CopyFrom(input3, *gpu_place, context); + paddle::framework::CopyFrom(input1, *gpu_place, context, &input1_gpu); + paddle::framework::CopyFrom(input2, *gpu_place, context, &input2_gpu); + paddle::framework::CopyFrom(input3, *gpu_place, context, &input3_gpu); float* a = input1_gpu.data(); float* b = input2_gpu.data(); float* c = input3_gpu.mutable_data(*gpu_place); - paddle::operators::math::gemm( + paddle::operators::math::gemm( context, false, false, m, n, k, 1, a, 3, b + 1, 4, 1, c + 1, 4); - input3.CopyFrom(input3_gpu, *cpu_place, context); + paddle::framework::CopyFrom(input3_gpu, *cpu_place, context, &input3); // numpy code: // a = np.arange(6).reshape(2, 3) @@ -151,20 +151,20 @@ TEST(math_function, gemm_trans_cublas) { float arr3[8] = {0, 1, 2, 3, 4, 5, 6, 7}; memcpy(input3_ptr, arr3, 8 * sizeof(float)); - auto* gpu_place = new paddle::platform::GPUPlace(0); + auto* gpu_place = new paddle::platform::CUDAPlace(0); paddle::platform::CUDADeviceContext context(*gpu_place); - input1_gpu.CopyFrom(input1, *gpu_place, context); - input2_gpu.CopyFrom(input2, *gpu_place, context); - input3_gpu.CopyFrom(input3, *gpu_place, context); + paddle::framework::CopyFrom(input1, *gpu_place, context, &input1_gpu); + paddle::framework::CopyFrom(input2, *gpu_place, context, &input2_gpu); + paddle::framework::CopyFrom(input3, *gpu_place, context, &input3_gpu); float* a = input1_gpu.data(); float* b = input2_gpu.data(); float* c = input3_gpu.mutable_data(*gpu_place); - paddle::operators::math::gemm( + paddle::operators::math::gemm( context, false, true, m, n, k, 1, a, 3, b + 3, 3, 1, c + 1, 4); - input3.CopyFrom(input3_gpu, *cpu_place, context); + paddle::framework::CopyFrom(input3_gpu, *cpu_place, context, &input3); context.Wait(); EXPECT_EQ(input3_ptr[0], 0); @@ -189,7 +189,7 @@ void GemvTest(int m, int n, bool trans) { T* data_b = vec_b.mutable_data({trans ? m : n}, *cpu_place); T* data_c = vec_c.mutable_data({trans ? n : m}, *cpu_place); - auto* gpu_place = new paddle::platform::GPUPlace(0); + auto* gpu_place = new paddle::platform::CUDAPlace(0); paddle::framework::Tensor g_mat_a; paddle::framework::Tensor g_vec_b; paddle::framework::Tensor g_vec_c; @@ -205,14 +205,15 @@ void GemvTest(int m, int n, bool trans) { } paddle::platform::CUDADeviceContext context(*gpu_place); - g_mat_a.CopyFrom(mat_a, *gpu_place, context); - g_vec_b.CopyFrom(vec_b, *gpu_place, context); + paddle::framework::CopyFrom(mat_a, *gpu_place, context, &g_mat_a); + paddle::framework::CopyFrom(vec_b, *gpu_place, context, &g_vec_b); - paddle::operators::math::gemv( + paddle::operators::math::gemv( context, trans, static_cast(m), static_cast(n), 1., g_data_a, g_data_b, 0., g_data_c); - vec_c.CopyFrom(g_vec_c, paddle::platform::CPUPlace(), context); + paddle::framework::CopyFrom(g_vec_c, paddle::platform::CPUPlace(), context, + &vec_c); if (!trans) { for (int i = 0; i < m; ++i) { diff --git a/paddle/operators/math/matmul.h b/paddle/operators/math/matmul.h index 6ba9a0ba9a70bd938f9362179990ab68fa3186ba..7048e11e6f27a075892c28681a3c4913a27b3f3e 100644 --- a/paddle/operators/math/matmul.h +++ b/paddle/operators/math/matmul.h @@ -26,13 +26,12 @@ namespace math { // // Both a & b can be 1- to 3-dimensional. Higher rank tensors are not supported // yet. -template +template class MatMulFunctor { public: - void operator()(const platform::DeviceContext& context, - const framework::Tensor& a, bool trans_a, - const framework::Tensor& b, bool trans_b, T alpha, - framework::Tensor* out, T beta) { + 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(); @@ -108,13 +107,13 @@ class MatMulFunctor { if (!batchCount) { // regular matrix multiplication - gemm(context, transA, transB, M, N, kA, alpha, a.data(), - b.data(), beta, out->data()); + 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); + batched_gemm( + context, transA, transB, M, N, kA, alpha, a.data(), b.data(), + beta, out->data(), batchCount, strideA, strideB); } } }; diff --git a/paddle/operators/math/maxouting.cc b/paddle/operators/math/maxouting.cc new file mode 100644 index 0000000000000000000000000000000000000000..fea86675f75dad99a336d795d4561ae32d58c30a --- /dev/null +++ b/paddle/operators/math/maxouting.cc @@ -0,0 +1,101 @@ +/* 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. */ + +#include "paddle/operators/math/maxouting.h" + +namespace paddle { +namespace operators { +namespace math { + +// All tensors are in NCHW format, and the groups must be greater than 1 +template +class MaxOutFunctor { + public: + void operator()(const platform::CPUDeviceContext& context, + const framework::Tensor& input, framework::Tensor* output, + int groups) { + const int batch_size = input.dims()[0]; + const int input_height = input.dims()[2]; + const int input_width = input.dims()[3]; + const int output_channels = output->dims()[1]; + int fea_size = input_height * input_width; + // c_size means the output size of each sample + int c_size = fea_size * output_channels; + const T* input_data = input.data(); + T* output_data = output->mutable_data(context.GetPlace()); + + for (int i = 0; i < batch_size; ++i) { + int new_bindex = c_size * i; + for (int c = 0; c < output_channels; ++c) { + int new_cindex = fea_size * c; + for (int f = 0; f < fea_size; ++f) { + T ele = static_cast(-FLT_MAX); + for (int ph = 0; ph < groups; ++ph) { + T x = input_data[(new_bindex + new_cindex) * groups + + ph * fea_size + f]; + ele = ele > x ? ele : x; + } + output_data[(new_bindex + new_cindex + f)] = ele; + } + } + } + } +}; + +template +class MaxOutGradFunctor { + public: + void operator()(const platform::CPUDeviceContext& context, + const framework::Tensor& input, framework::Tensor* input_grad, + const framework::Tensor& output, + const framework::Tensor& output_grad, int groups) { + const int batch_size = input.dims()[0]; + const int input_height = input.dims()[2]; + const int input_width = input.dims()[3]; + const int output_channels = output.dims()[1]; + int fea_size = input_height * input_width; + const T* input_data = input.data(); + const T* output_data = output.data(); + const T* output_grad_data = output_grad.data(); + T* input_grad_data = input_grad->mutable_data(context.GetPlace()); + + for (int i = 0; i < batch_size; ++i) { + int blen = fea_size * output_channels * i; + for (int c = 0; c < output_channels; ++c) { + int clen = fea_size * c; + for (int f = 0; f < fea_size; ++f) { + int input_idx0 = (blen + clen) * groups + f; + bool continue_match = true; + int output_idx = blen + clen + f; + for (int g = 0; g < groups && continue_match; ++g) { + int input_idx = input_idx0 + fea_size * g; + if (input_data[input_idx] == output_data[output_idx]) { + input_grad_data[input_idx] += output_grad_data[output_idx]; + continue_match = false; + } + } + } + } + } + } +}; + +template class MaxOutGradFunctor; +template class MaxOutGradFunctor; +template class MaxOutFunctor; +template class MaxOutFunctor; + +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/math/maxouting.cu b/paddle/operators/math/maxouting.cu new file mode 100644 index 0000000000000000000000000000000000000000..6056ad251c12976fe9032f03aaaeb52727da1f42 --- /dev/null +++ b/paddle/operators/math/maxouting.cu @@ -0,0 +1,147 @@ +/* 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. */ + +#include "paddle/operators/math/maxouting.h" +#include "paddle/platform/cuda_helper.h" + +namespace paddle { +namespace operators { +namespace math { + +template +__global__ void KernelMaxOut(const int nthreads, const T* input_data, + const int channels, const int input_height, + const int input_width, int groups, + T* output_data) { + const int size = input_height * input_width * channels / groups; + const int feat_len = input_height * input_width; + int index = blockIdx.x * blockDim.x + threadIdx.x; + int offset = blockDim.x * gridDim.x; + for (int i = index; i < nthreads; i += offset) { + int batch_idx = i / size; + int batch_offset = i % size; + int channel_idx = batch_offset / feat_len; + int feat_idx = batch_offset % feat_len; + int data_idx = + (batch_idx * size + channel_idx * feat_len) * groups + feat_idx; + T ele = static_cast(-FLT_MAX); + for (int g = 0; g < groups; ++g) { + T x = input_data[data_idx + g * feat_len]; + ele = ele > x ? ele : x; + } + output_data[i] = ele; + } +} +template +__global__ void KernelMaxoutGrad(const int nthreads, const T* input_data, + const T* output_data, const T* output_grad, + T* input_grad, const int channels, + const int input_height, const int input_width, + int groups) { + const int size = input_height * input_width * channels / groups; + const int feat_len = input_height * input_width; + int index = blockIdx.x * blockDim.x + threadIdx.x; + int offset = blockDim.x * gridDim.x; + for (int i = index; i < nthreads; i += offset) { + int batch_idx = i / size; + int batch_offset = i % size; + int channel_idx = batch_offset / feat_len; + int feat_idx = batch_offset % feat_len; + int data_idx = + (batch_idx * size + channel_idx * feat_len) * groups + feat_idx; + int max_index = -1; + bool continue_match = true; + for (int g = 0; g < groups && continue_match; ++g) { + if (input_data[data_idx + g * feat_len] == output_data[i]) { + max_index = data_idx + g * feat_len; + continue_match = false; + break; + } + } + if (max_index != -1) { + input_grad[max_index] += output_grad[index]; + } + } +} +/* + * All tensors are in NCHW format. + */ +template +class MaxOutFunctor { + public: + void operator()(const platform::CUDADeviceContext& context, + const framework::Tensor& input, framework::Tensor* output, + int groups) { + const int batch_size = input.dims()[0]; + const int input_channels = input.dims()[1]; + const int input_height = input.dims()[2]; + const int input_width = input.dims()[3]; + const int output_channels = output->dims()[1]; + const int output_height = output->dims()[2]; + const int output_width = output->dims()[3]; + + const T* input_data = input.data(); + T* output_data = output->mutable_data(context.GetPlace()); + int nthreads = output->numel(); + int blocks = (nthreads + 1024 - 1) / 1024; + dim3 threads(1024, 1); + dim3 grid(blocks, 1); + + KernelMaxOut<<>>( + nthreads, input_data, input_channels, input_height, input_width, groups, + output_data); + } +}; +/* + * All tensors are in NCHW format. + */ +template +class MaxOutGradFunctor { + public: + void operator()(const platform::CUDADeviceContext& context, + const framework::Tensor& input, framework::Tensor* input_grad, + const framework::Tensor& output, + const framework::Tensor& output_grad, int groups) { + const int batch_size = input.dims()[0]; + const int input_channels = input.dims()[1]; + const int input_height = input.dims()[2]; + const int input_width = input.dims()[3]; + const int output_channels = output.dims()[1]; + const int output_height = output.dims()[2]; + const int output_width = output.dims()[3]; + + const T* input_data = input.data(); + const T* output_data = output.data(); + const T* output_grad_data = output_grad.data(); + T* input_grad_data = input_grad->mutable_data(context.GetPlace()); + int nthreads = output.numel(); + int blocks = (nthreads + 1024 - 1) / 1024; + dim3 threads(1024, 1); + dim3 grid(blocks, 1); + + KernelMaxoutGrad<<>>( + nthreads, input_data, output_data, output_grad_data, input_grad_data, + input_channels, input_height, input_width, groups); + } +}; + +template class MaxOutGradFunctor; +template class MaxOutGradFunctor; + +template class MaxOutFunctor; +template class MaxOutFunctor; + +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/math/maxouting.h b/paddle/operators/math/maxouting.h new file mode 100644 index 0000000000000000000000000000000000000000..68f4743db07af0f369eb18f1a7cb6e326d469e85 --- /dev/null +++ b/paddle/operators/math/maxouting.h @@ -0,0 +1,43 @@ +/* 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 +#include "paddle/framework/tensor.h" +#include "paddle/platform/device_context.h" +#include "paddle/platform/hostdevice.h" + +namespace paddle { +namespace operators { +namespace math { + +#define FLT_MAX __FLT_MAX__ + +template +class MaxOutFunctor { + public: + void operator()(const DeviceContext& context, const framework::Tensor& input, + framework::Tensor* output, int groups); +}; + +template +class MaxOutGradFunctor { + public: + void operator()(const DeviceContext& context, const framework::Tensor& input, + framework::Tensor* input_grad, + const framework::Tensor& output, + const framework::Tensor& output_grad, int groups); +}; +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/math/pooling.cc b/paddle/operators/math/pooling.cc index ead89e146f32ef005b06f4f6f04224d691805d74..150de6fd59ef3ac0c4cb9160bf5afb1ce1064577 100644 --- a/paddle/operators/math/pooling.cc +++ b/paddle/operators/math/pooling.cc @@ -24,9 +24,9 @@ namespace math { * height and width, respectively. */ template -class Pool2dFunctor { +class Pool2dFunctor { public: - void operator()(const platform::DeviceContext& context, + 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) { @@ -84,9 +84,9 @@ class Pool2dFunctor { * and width, respectively. */ template -class Pool2dGradFunctor { +class Pool2dGradFunctor { public: - void operator()(const platform::DeviceContext& context, + void operator()(const platform::CPUDeviceContext& context, const framework::Tensor& input, const framework::Tensor& output, const framework::Tensor& output_grad, std::vector& ksize, @@ -152,9 +152,9 @@ class Pool2dGradFunctor { * height and width, respectively. */ template -class MaxPool2dGradFunctor { +class MaxPool2dGradFunctor { public: - void operator()(const platform::DeviceContext& context, + void operator()(const platform::CPUDeviceContext& context, const framework::Tensor& input, const framework::Tensor& output, const framework::Tensor& output_grad, std::vector& ksize, @@ -213,25 +213,29 @@ class MaxPool2dGradFunctor { } }; -template class MaxPool2dGradFunctor; -template class MaxPool2dGradFunctor; +template class MaxPool2dGradFunctor; +template class MaxPool2dGradFunctor; -template class Pool2dFunctor, float>; -template class Pool2dFunctor, float>; -template class Pool2dGradFunctor< - platform::CPUPlace, paddle::operators::math::MaxPoolGrad, float>; -template class Pool2dGradFunctor< - platform::CPUPlace, paddle::operators::math::AvgPoolGrad, float>; -template class Pool2dFunctor, + float>; +template class Pool2dGradFunctor, + float>; +template class Pool2dFunctor, double>; -template class Pool2dFunctor, double>; -template class Pool2dGradFunctor< - platform::CPUPlace, paddle::operators::math::MaxPoolGrad, double>; -template class Pool2dGradFunctor< - platform::CPUPlace, paddle::operators::math::AvgPoolGrad, double>; +template class Pool2dGradFunctor, + double>; +template class Pool2dGradFunctor, + double>; /* * All tensors are in NCDHW format. @@ -239,9 +243,9 @@ template class Pool2dGradFunctor< * depth, height and width, respectively. */ template -class Pool3dFunctor { +class Pool3dFunctor { public: - void operator()(const platform::DeviceContext& context, + 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) { @@ -314,9 +318,9 @@ class Pool3dFunctor { * depth, height and width, respectively. */ template -class Pool3dGradFunctor { +class Pool3dGradFunctor { public: - void operator()(const platform::DeviceContext& context, + void operator()(const platform::CPUDeviceContext& context, const framework::Tensor& input, const framework::Tensor& output, const framework::Tensor& output_grad, std::vector& ksize, @@ -398,9 +402,9 @@ class Pool3dGradFunctor { * depth, height and width, respectively. */ template -class MaxPool3dGradFunctor { +class MaxPool3dGradFunctor { public: - void operator()(const platform::DeviceContext& context, + void operator()(const platform::CPUDeviceContext& context, const framework::Tensor& input, const framework::Tensor& output, const framework::Tensor& output_grad, std::vector& ksize, @@ -473,35 +477,39 @@ class MaxPool3dGradFunctor { } }; -template class MaxPool3dGradFunctor; -template class MaxPool3dGradFunctor; +template class MaxPool3dGradFunctor; +template class MaxPool3dGradFunctor; -template class Pool3dFunctor, float>; -template class Pool3dFunctor, float>; -template class Pool3dGradFunctor< - platform::CPUPlace, paddle::operators::math::MaxPoolGrad, float>; -template class Pool3dGradFunctor< - platform::CPUPlace, paddle::operators::math::AvgPoolGrad, float>; -template class Pool3dFunctor, + float>; +template class Pool3dGradFunctor, + float>; +template class Pool3dFunctor, double>; -template class Pool3dFunctor, double>; -template class Pool3dGradFunctor< - platform::CPUPlace, paddle::operators::math::MaxPoolGrad, double>; -template class Pool3dGradFunctor< - platform::CPUPlace, paddle::operators::math::AvgPoolGrad, double>; +template class Pool3dGradFunctor, + double>; +template class Pool3dGradFunctor, + double>; /* * All tensors are in NCHW format. * Ksize, strides, paddings are two elements. These two elements represent * height and width, respectively. */ -template -class MaxPool2dWithIndexFunctor { +template +class MaxPool2dWithIndexFunctor { public: - void operator()(const platform::DeviceContext& context, + 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) { @@ -520,9 +528,9 @@ class MaxPool2dWithIndexFunctor { const int input_stride = input_height * input_width; const int output_stride = output_height * output_width; - const T* input_data = input.data(); - T* output_data = output->mutable_data(context.GetPlace()); - T* mask_data = mask->mutable_data(context.GetPlace()); + const T1* input_data = input.data(); + T1* output_data = output->mutable_data(context.GetPlace()); + T2* mask_data = mask->mutable_data(context.GetPlace()); for (int i = 0; i < batch_size; i++) { for (int c = 0; c < output_channels; ++c) { @@ -535,7 +543,7 @@ class MaxPool2dWithIndexFunctor { int wend = std::min(wstart + ksize_width, input_width); wstart = std::max(wstart, 0); - T ele = static_cast(-FLT_MAX); + T1 ele = static_cast(-FLT_MAX); int index = -1; for (int h = hstart; h < hend; ++h) { for (int w = wstart; w < wend; ++w) { @@ -563,10 +571,10 @@ class MaxPool2dWithIndexFunctor { * Ksize, strides, paddings are two elements. These two elements represent * height and width, respectively. */ -template -class MaxPool2dWithIndexGradFunctor { +template +class MaxPool2dWithIndexGradFunctor { public: - void operator()(const platform::DeviceContext& context, + void operator()(const platform::CPUDeviceContext& context, const framework::Tensor& output_grad, const framework::Tensor& mask, std::vector& ksize, std::vector& strides, std::vector& paddings, @@ -580,9 +588,9 @@ class MaxPool2dWithIndexGradFunctor { const int input_stride = input_height * input_width; const int output_stride = output_height * output_width; - const T* mask_data = mask.data(); - const T* output_grad_data = output_grad.data(); - T* input_grad_data = input_grad->mutable_data(context.GetPlace()); + const T2* mask_data = mask.data(); + const T1* output_grad_data = output_grad.data(); + T1* input_grad_data = input_grad->mutable_data(context.GetPlace()); for (int n = 0; n < batch_size; ++n) { for (int c = 0; c < output_channels; ++c) { @@ -602,20 +610,24 @@ class MaxPool2dWithIndexGradFunctor { } }; -template class MaxPool2dWithIndexFunctor; -template class MaxPool2dWithIndexGradFunctor; -template class MaxPool2dWithIndexFunctor; -template class MaxPool2dWithIndexGradFunctor; +template class MaxPool2dWithIndexFunctor; +template class MaxPool2dWithIndexGradFunctor; +template class MaxPool2dWithIndexFunctor; +template class MaxPool2dWithIndexGradFunctor; /* * All tensors are in NCDHW format. * Ksize, strides, paddings are three elements. These three elements represent * depth, height and width, respectively. */ -template -class MaxPool3dWithIndexFunctor { +template +class MaxPool3dWithIndexFunctor { public: - void operator()(const platform::DeviceContext& context, + 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) { @@ -639,9 +651,9 @@ class MaxPool3dWithIndexFunctor { const int input_stride = input_depth * input_height * input_width; const int output_stride = output_depth * output_height * output_width; - const T* input_data = input.data(); - T* output_data = output->mutable_data(context.GetPlace()); - T* mask_data = mask->mutable_data(context.GetPlace()); + const T1* input_data = input.data(); + T1* output_data = output->mutable_data(context.GetPlace()); + T2* mask_data = mask->mutable_data(context.GetPlace()); for (int i = 0; i < batch_size; i++) { for (int c = 0; c < output_channels; ++c) { @@ -659,7 +671,7 @@ class MaxPool3dWithIndexFunctor { wstart = std::max(wstart, 0); int output_idx = (pd * output_height + ph) * output_width + pw; - T ele = static_cast(-FLT_MAX); + T1 ele = static_cast(-FLT_MAX); int index = -1; for (int d = dstart; d < dend; ++d) { for (int h = hstart; h < hend; ++h) { @@ -691,10 +703,10 @@ class MaxPool3dWithIndexFunctor { * Ksize, strides, paddings are three elements. These three elements represent * depth, height and width, respectively. */ -template -class MaxPool3dWithIndexGradFunctor { +template +class MaxPool3dWithIndexGradFunctor { public: - void operator()(const platform::DeviceContext& context, + void operator()(const platform::CPUDeviceContext& context, const framework::Tensor& output_grad, const framework::Tensor& mask, std::vector& ksize, std::vector& strides, std::vector& paddings, @@ -710,9 +722,9 @@ class MaxPool3dWithIndexGradFunctor { const int input_stride = input_depth * input_height * input_width; const int output_stride = output_depth * output_height * output_width; - const T* mask_data = mask.data(); - const T* output_grad_data = output_grad.data(); - T* input_grad_data = input_grad->mutable_data(context.GetPlace()); + const T2* mask_data = mask.data(); + const T1* output_grad_data = output_grad.data(); + T1* input_grad_data = input_grad->mutable_data(context.GetPlace()); for (int n = 0; n < batch_size; ++n) { for (int c = 0; c < output_channels; ++c) { @@ -735,10 +747,14 @@ class MaxPool3dWithIndexGradFunctor { } }; -template class MaxPool3dWithIndexFunctor; -template class MaxPool3dWithIndexGradFunctor; -template class MaxPool3dWithIndexFunctor; -template class MaxPool3dWithIndexGradFunctor; +template class MaxPool3dWithIndexFunctor; +template class MaxPool3dWithIndexGradFunctor; +template class MaxPool3dWithIndexFunctor; +template class MaxPool3dWithIndexGradFunctor; } // namespace math } // namespace operators } // namespace paddle diff --git a/paddle/operators/math/pooling.cu b/paddle/operators/math/pooling.cu index 6d1138ad50cb095e85b4ceb44fa81731316f10dd..0243cf8316a2a83bfc4c091f64419574c1be2f5c 100644 --- a/paddle/operators/math/pooling.cu +++ b/paddle/operators/math/pooling.cu @@ -155,9 +155,9 @@ __global__ void KernelMaxPool2DGrad( * height and width, respectively. */ template -class Pool2dFunctor { +class Pool2dFunctor { public: - void operator()(const platform::DeviceContext& context, + 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) { @@ -183,11 +183,7 @@ class Pool2dFunctor { dim3 threads(1024, 1); dim3 grid(blocks, 1); - KernelPool2D< - PoolProcess, - T><<(context) - .stream()>>>( + KernelPool2D<<>>( nthreads, input_data, input_channels, input_height, input_width, output_height, output_width, ksize_height, ksize_width, stride_height, stride_width, padding_height, padding_width, pool_process, output_data); @@ -200,9 +196,9 @@ class Pool2dFunctor { * height and width, respectively. */ template -class Pool2dGradFunctor { +class Pool2dGradFunctor { public: - void operator()(const platform::DeviceContext& context, + void operator()(const platform::CUDADeviceContext& context, const framework::Tensor& input, const framework::Tensor& output, const framework::Tensor& output_grad, std::vector& ksize, @@ -231,11 +227,7 @@ class Pool2dGradFunctor { dim3 threads(1024, 1); dim3 grid(blocks, 1); - KernelPool2DGrad< - PoolProcess, - T><<(context) - .stream()>>>( + KernelPool2DGrad<<>>( nthreads, input_data, output_data, output_grad_data, input_channels, input_height, input_width, output_height, output_width, ksize_height, ksize_width, stride_height, stride_width, padding_height, padding_width, @@ -249,9 +241,9 @@ class Pool2dGradFunctor { * height and width, respectively. */ template -class MaxPool2dGradFunctor { +class MaxPool2dGradFunctor { public: - void operator()(const platform::DeviceContext& context, + void operator()(const platform::CUDADeviceContext& context, const framework::Tensor& input, const framework::Tensor& output, const framework::Tensor& output_grad, std::vector& ksize, @@ -281,10 +273,7 @@ class MaxPool2dGradFunctor { dim3 threads(1024, 1); dim3 grid(blocks, 1); - KernelMaxPool2DGrad< - T><<(context) - .stream()>>>( + KernelMaxPool2DGrad<<>>( nthreads, input_data, output_data, output_grad_data, input_channels, input_height, input_width, output_height, output_width, ksize_height, ksize_width, stride_height, stride_width, padding_height, padding_width, @@ -292,25 +281,29 @@ class MaxPool2dGradFunctor { } }; -template class MaxPool2dGradFunctor; -template class MaxPool2dGradFunctor; +template class MaxPool2dGradFunctor; +template class MaxPool2dGradFunctor; -template class Pool2dFunctor, float>; -template class Pool2dFunctor, float>; -template class Pool2dGradFunctor< - platform::GPUPlace, paddle::operators::math::MaxPoolGrad, float>; -template class Pool2dGradFunctor< - platform::GPUPlace, paddle::operators::math::AvgPoolGrad, float>; -template class Pool2dFunctor, + float>; +template class Pool2dGradFunctor, + float>; +template class Pool2dFunctor, double>; -template class Pool2dFunctor, double>; -template class Pool2dGradFunctor< - platform::GPUPlace, paddle::operators::math::MaxPoolGrad, double>; -template class Pool2dGradFunctor< - platform::GPUPlace, paddle::operators::math::AvgPoolGrad, double>; +template class Pool2dGradFunctor, + double>; +template class Pool2dGradFunctor, + double>; template __global__ void KernelPool3D(const int nthreads, const T* input_data, @@ -478,9 +471,9 @@ __global__ void KernelMaxPool3DGrad( * depth, height and width, respectively. */ template -class Pool3dFunctor { +class Pool3dFunctor { public: - void operator()(const platform::DeviceContext& context, + 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) { @@ -512,11 +505,7 @@ class Pool3dFunctor { dim3 threads(1024, 1); dim3 grid(blocks, 1); - KernelPool3D< - PoolProcess, - T><<(context) - .stream()>>>( + KernelPool3D<<>>( nthreads, input_data, input_channels, input_depth, input_height, input_width, output_depth, output_height, output_width, ksize_depth, ksize_height, ksize_width, stride_depth, stride_height, stride_width, @@ -531,9 +520,9 @@ class Pool3dFunctor { * depth, height and width, respectively. */ template -class Pool3dGradFunctor { +class Pool3dGradFunctor { public: - void operator()(const platform::DeviceContext& context, + void operator()(const platform::CUDADeviceContext& context, const framework::Tensor& input, const framework::Tensor& output, const framework::Tensor& output_grad, std::vector& ksize, @@ -569,11 +558,7 @@ class Pool3dGradFunctor { dim3 threads(1024, 1); dim3 grid(blocks, 1); - KernelPool3DGrad< - PoolProcess, - T><<(context) - .stream()>>>( + KernelPool3DGrad<<>>( nthreads, input_data, output_data, output_grad_data, input_channels, input_depth, input_height, input_width, output_depth, output_height, output_width, ksize_depth, ksize_height, ksize_width, stride_depth, @@ -588,9 +573,9 @@ class Pool3dGradFunctor { * depth, height and width, respectively. */ template -class MaxPool3dGradFunctor { +class MaxPool3dGradFunctor { public: - void operator()(const platform::DeviceContext& context, + void operator()(const platform::CUDADeviceContext& context, const framework::Tensor& input, const framework::Tensor& output, const framework::Tensor& output_grad, std::vector& ksize, @@ -626,10 +611,7 @@ class MaxPool3dGradFunctor { dim3 threads(1024, 1); dim3 grid(blocks, 1); - KernelMaxPool3DGrad< - T><<(context) - .stream()>>>( + KernelMaxPool3DGrad<<>>( nthreads, input_data, output_data, output_grad_data, input_channels, input_depth, input_height, input_width, output_depth, output_height, output_width, ksize_depth, ksize_height, ksize_width, stride_depth, @@ -638,33 +620,37 @@ class MaxPool3dGradFunctor { } }; -template class MaxPool3dGradFunctor; -template class MaxPool3dGradFunctor; +template class MaxPool3dGradFunctor; +template class MaxPool3dGradFunctor; -template class Pool3dFunctor, float>; -template class Pool3dFunctor, float>; -template class Pool3dGradFunctor< - platform::GPUPlace, paddle::operators::math::MaxPoolGrad, float>; -template class Pool3dGradFunctor< - platform::GPUPlace, paddle::operators::math::AvgPoolGrad, float>; -template class Pool3dFunctor, + float>; +template class Pool3dGradFunctor, + float>; +template class Pool3dFunctor, double>; -template class Pool3dFunctor, double>; -template class Pool3dGradFunctor< - platform::GPUPlace, paddle::operators::math::MaxPoolGrad, double>; -template class Pool3dGradFunctor< - platform::GPUPlace, paddle::operators::math::AvgPoolGrad, double>; - -template +template class Pool3dGradFunctor, + double>; +template class Pool3dGradFunctor, + double>; + +template __global__ void KernelMaxPool2dWithIdx( - const int nthreads, const T* input_data, const int channels, + const int nthreads, const T1* input_data, const int channels, const int input_height, const int input_width, const int output_height, const int output_width, const int ksize_height, const int ksize_width, const int stride_height, const int stride_width, const int padding_height, - const int padding_width, T* output_data, T* mask_data) { + const int padding_width, T1* output_data, T2* mask_data) { for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < nthreads; index += blockDim.x * gridDim.x) { int pw = index % output_width; @@ -681,7 +667,7 @@ __global__ void KernelMaxPool2dWithIdx( wstart = max(wstart, 0); input_data += (batch_idx * channels + c) * input_height * input_width; - T ele = -FLT_MAX; + T1 ele = -FLT_MAX; int max_index = -1; for (int h = hstart; h < hend; ++h) { for (int w = wstart; w < wend; ++w) { @@ -697,13 +683,13 @@ __global__ void KernelMaxPool2dWithIdx( } } -template +template __global__ void KernelMaxPool2DWithIdxGrad( - const int nthreads, const T* output_grad, const T* mask_data, + const int nthreads, const T1* output_grad, const T2* mask_data, const int channels, const int input_height, const int input_width, const int output_height, const int output_width, const int ksize_height, const int ksize_width, const int stride_height, const int stride_width, - const int padding_height, const int padding_width, T* input_grad) { + const int padding_height, const int padding_width, T1* input_grad) { for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < nthreads; index += blockDim.x * gridDim.x) { int w_offset = index % input_width; @@ -724,7 +710,7 @@ __global__ void KernelMaxPool2DWithIdxGrad( int pw_end = min((w_offset + padding_width) / stride_width + 1, output_width); - T gradient = 0; + T1 gradient = 0; int input_current_featuremap_idx = h_offset * input_width + w_offset; int output_idx = (batch_idx * channels + c_offset) * output_height * output_width; @@ -746,10 +732,10 @@ __global__ void KernelMaxPool2DWithIdxGrad( * Ksize, strides, paddings are two elements. These two elements represent * height and width, respectively. */ -template -class MaxPool2dWithIndexFunctor { +template +class MaxPool2dWithIndexFunctor { public: - void operator()(const platform::DeviceContext& context, + 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) { @@ -767,19 +753,16 @@ class MaxPool2dWithIndexFunctor { const int padding_height = paddings[0]; const int padding_width = paddings[1]; - const T* input_data = input.data(); - T* output_data = output->mutable_data(context.GetPlace()); - T* mask_data = mask->mutable_data(context.GetPlace()); + const T1* input_data = input.data(); + T1* output_data = output->mutable_data(context.GetPlace()); + T2* mask_data = mask->mutable_data(context.GetPlace()); int nthreads = batch_size * output_channels * output_height * output_width; int blocks = (nthreads + 1024 - 1) / 1024; dim3 threads(1024, 1); dim3 grid(blocks, 1); - KernelMaxPool2dWithIdx< - T><<(context) - .stream()>>>( + KernelMaxPool2dWithIdx<<>>( nthreads, input_data, input_channels, input_height, input_width, output_height, output_width, ksize_height, ksize_width, stride_height, stride_width, padding_height, padding_width, output_data, mask_data); @@ -791,10 +774,10 @@ class MaxPool2dWithIndexFunctor { * Ksize, strides, paddings are two elements. These two elements represent * height and width, respectively. */ -template -class MaxPool2dWithIndexGradFunctor { +template +class MaxPool2dWithIndexGradFunctor { public: - void operator()(const platform::DeviceContext& context, + void operator()(const platform::CUDADeviceContext& context, const framework::Tensor& output_grad, const framework::Tensor& mask, std::vector& ksize, std::vector& strides, std::vector& paddings, @@ -812,40 +795,41 @@ class MaxPool2dWithIndexGradFunctor { const int padding_height = paddings[0]; const int padding_width = paddings[1]; - const T* mask_data = mask.data(); - const T* output_grad_data = output_grad.data(); - T* input_grad_data = input_grad->mutable_data(context.GetPlace()); + const T2* mask_data = mask.data(); + const T1* output_grad_data = output_grad.data(); + T1* input_grad_data = input_grad->mutable_data(context.GetPlace()); int nthreads = batch_size * input_channels * input_height * input_width; int blocks = (nthreads + 1024 - 1) / 1024; dim3 threads(1024, 1); dim3 grid(blocks, 1); - KernelMaxPool2DWithIdxGrad< - T><<(context) - .stream()>>>(nthreads, output_grad_data, mask_data, - input_channels, input_height, input_width, - output_height, output_width, ksize_height, - ksize_width, stride_height, stride_width, - padding_height, padding_width, input_grad_data); + KernelMaxPool2DWithIdxGrad<<>>( + nthreads, output_grad_data, mask_data, input_channels, input_height, + input_width, output_height, output_width, ksize_height, ksize_width, + stride_height, stride_width, padding_height, padding_width, + input_grad_data); } }; -template class MaxPool2dWithIndexFunctor; -template class MaxPool2dWithIndexGradFunctor; -template class MaxPool2dWithIndexFunctor; -template class MaxPool2dWithIndexGradFunctor; +template class MaxPool2dWithIndexFunctor; +template class MaxPool2dWithIndexGradFunctor; +template class MaxPool2dWithIndexFunctor; +template class MaxPool2dWithIndexGradFunctor; -template +template __global__ void KernelMaxPool3DWithIdx( - const int nthreads, const T* input_data, const int channels, + const int nthreads, const T1* input_data, const int channels, const int input_depth, const int input_height, const int input_width, const int output_depth, const int output_height, const int output_width, const int ksize_depth, const int ksize_height, const int ksize_width, const int stride_depth, const int stride_height, const int stride_width, const int padding_depth, const int padding_height, const int padding_width, - T* output_data, T* mask_data) { + T1* output_data, T2* mask_data) { for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < nthreads; index += blockDim.x * gridDim.x) { int pw = index % output_width; @@ -865,7 +849,7 @@ __global__ void KernelMaxPool3DWithIdx( hstart = max(hstart, 0); wstart = max(wstart, 0); - T ele = -FLT_MAX; + T1 ele = -FLT_MAX; int max_index = -1; input_data += (batch_idx * channels + c) * input_depth * input_height * input_width; @@ -885,15 +869,15 @@ __global__ void KernelMaxPool3DWithIdx( } } -template +template __global__ void KernelMaxPool3DWithIdxGrad( - const int nthreads, const T* output_grad, const T* mask, const int channels, - const int input_depth, const int input_height, const int input_width, - const int output_depth, const int output_height, const int output_width, - const int ksize_depth, const int ksize_height, const int ksize_width, - const int stride_depth, const int stride_height, const int stride_width, - const int padding_depth, const int padding_height, const int padding_width, - T* input_grad) { + const int nthreads, const T1* output_grad, const T2* mask, + const int channels, const int input_depth, const int input_height, + const int input_width, const int output_depth, const int output_height, + const int output_width, const int ksize_depth, const int ksize_height, + const int ksize_width, const int stride_depth, const int stride_height, + const int stride_width, const int padding_depth, const int padding_height, + const int padding_width, T1* input_grad) { for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < nthreads; index += blockDim.x * gridDim.x) { int w_offset = index % input_width; @@ -922,7 +906,7 @@ __global__ void KernelMaxPool3DWithIdxGrad( int pw_end = min((w_offset + padding_width) / stride_width + 1, output_width); - T gradient = 0; + T1 gradient = 0; int input_current_feature_map_idx = (d_offset * input_height + h_offset) * input_width + w_offset; int output_idx = (batch_idx * channels + c_offset) * output_depth * @@ -949,10 +933,10 @@ __global__ void KernelMaxPool3DWithIdxGrad( * Ksize, strides, paddings are three elements. These three elements represent * depth, height and width, respectively. */ -template -class MaxPool3dWithIndexFunctor { +template +class MaxPool3dWithIndexFunctor { public: - void operator()(const platform::DeviceContext& context, + 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) { @@ -975,9 +959,9 @@ class MaxPool3dWithIndexFunctor { const int padding_height = paddings[1]; const int padding_width = paddings[2]; - const T* input_data = input.data(); - T* output_data = output->mutable_data(context.GetPlace()); - T* mask_data = mask->mutable_data(context.GetPlace()); + const T1* input_data = input.data(); + T1* output_data = output->mutable_data(context.GetPlace()); + T2* mask_data = mask->mutable_data(context.GetPlace()); int nthreads = batch_size * output_channels * output_depth * output_height * output_width; @@ -985,10 +969,7 @@ class MaxPool3dWithIndexFunctor { dim3 threads(1024, 1); dim3 grid(blocks, 1); - KernelMaxPool3DWithIdx< - T><<(context) - .stream()>>>( + KernelMaxPool3DWithIdx<<>>( nthreads, input_data, input_channels, input_depth, input_height, input_width, output_depth, output_height, output_width, ksize_depth, ksize_height, ksize_width, stride_depth, stride_height, stride_width, @@ -1001,10 +982,10 @@ class MaxPool3dWithIndexFunctor { * Ksize, strides, paddings are three elements. These three elements represent * depth, height and width, respectively. */ -template -class MaxPool3dWithIndexGradFunctor { +template +class MaxPool3dWithIndexGradFunctor { public: - void operator()(const platform::DeviceContext& context, + void operator()(const platform::CUDADeviceContext& context, const framework::Tensor& output_grad, const framework::Tensor& mask, std::vector& ksize, std::vector& strides, std::vector& paddings, @@ -1027,9 +1008,9 @@ class MaxPool3dWithIndexGradFunctor { const int padding_height = paddings[1]; const int padding_width = paddings[2]; - const T* output_grad_data = output_grad.data(); - const T* mask_data = mask.data(); - T* input_grad_data = input_grad->mutable_data(context.GetPlace()); + const T1* output_grad_data = output_grad.data(); + const T2* mask_data = mask.data(); + T1* input_grad_data = input_grad->mutable_data(context.GetPlace()); int nthreads = batch_size * input_channels * input_depth * input_height * input_width; @@ -1037,10 +1018,7 @@ class MaxPool3dWithIndexGradFunctor { dim3 threads(1024, 1); dim3 grid(blocks, 1); - KernelMaxPool3DWithIdxGrad< - T><<(context) - .stream()>>>( + KernelMaxPool3DWithIdxGrad<<>>( nthreads, output_grad_data, mask_data, input_channels, input_depth, input_height, input_width, output_depth, output_height, output_width, ksize_depth, ksize_height, ksize_width, stride_depth, stride_height, @@ -1049,10 +1027,14 @@ class MaxPool3dWithIndexGradFunctor { } }; -template class MaxPool3dWithIndexFunctor; -template class MaxPool3dWithIndexGradFunctor; -template class MaxPool3dWithIndexFunctor; -template class MaxPool3dWithIndexGradFunctor; +template class MaxPool3dWithIndexFunctor; +template class MaxPool3dWithIndexGradFunctor; +template class MaxPool3dWithIndexFunctor; +template class MaxPool3dWithIndexGradFunctor; } // namespace math } // namespace operators diff --git a/paddle/operators/math/pooling.h b/paddle/operators/math/pooling.h index f6719e1e628cdd2cf7445ec9cd05713bc4f14c84..2759f06cb6a51f7ceb6b8010d792030eb6ad5d3e 100644 --- a/paddle/operators/math/pooling.h +++ b/paddle/operators/math/pooling.h @@ -84,62 +84,58 @@ class AvgPoolGrad { * This is different from average pooling. So we rewrite the max_pool_grad: * MaxPool2dGradFunctor, MaxPool3dGradFunctor. */ -template +template class Pool2dFunctor { public: - void operator()(const platform::DeviceContext& context, - const framework::Tensor& input, std::vector& ksize, - std::vector& strides, std::vector& paddings, - PoolProcess pool_compute, framework::Tensor* output); + void operator()(const DeviceContext& context, const framework::Tensor& input, + std::vector& ksize, std::vector& strides, + std::vector& paddings, PoolProcess pool_compute, + framework::Tensor* output); }; -template +template class Pool2dGradFunctor { public: - void operator()(const platform::DeviceContext& context, - const framework::Tensor& input, + 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); }; -template +template class MaxPool2dGradFunctor { public: - void operator()(const platform::DeviceContext& context, - const framework::Tensor& input, + 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, framework::Tensor* input_grad); }; -template +template class Pool3dFunctor { public: - void operator()(const platform::DeviceContext& context, - const framework::Tensor& input, std::vector& ksize, - std::vector& strides, std::vector& paddings, - PoolProcess pool_compute, framework::Tensor* output); + void operator()(const DeviceContext& context, const framework::Tensor& input, + std::vector& ksize, std::vector& strides, + std::vector& paddings, PoolProcess pool_compute, + framework::Tensor* output); }; -template +template class Pool3dGradFunctor { public: - void operator()(const platform::DeviceContext& context, - const framework::Tensor& input, + 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); }; -template +template class MaxPool3dGradFunctor { public: - void operator()(const platform::DeviceContext& context, - const framework::Tensor& input, + 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, @@ -153,38 +149,38 @@ class MaxPool3dGradFunctor { * In pool2d, all tensors are in NCHW format. In pool3d, all tensors are in * NCDHW format. */ -template +template class MaxPool2dWithIndexFunctor { public: - void operator()(const platform::DeviceContext& context, - const framework::Tensor& input, std::vector& ksize, - std::vector& strides, std::vector& paddings, - framework::Tensor* output, framework::Tensor* mask); + void operator()(const DeviceContext& context, const framework::Tensor& input, + std::vector& ksize, std::vector& strides, + std::vector& paddings, framework::Tensor* output, + framework::Tensor* mask); }; -template +template class MaxPool2dWithIndexGradFunctor { public: - void operator()(const platform::DeviceContext& context, + void operator()(const DeviceContext& context, const framework::Tensor& output_grad, const framework::Tensor& mask, std::vector& ksize, std::vector& strides, std::vector& paddings, framework::Tensor* input_grad); }; -template +template class MaxPool3dWithIndexFunctor { public: - void operator()(const platform::DeviceContext& context, - const framework::Tensor& input, std::vector& ksize, - std::vector& strides, std::vector& paddings, - framework::Tensor* output, framework::Tensor* mask); + void operator()(const DeviceContext& context, const framework::Tensor& input, + std::vector& ksize, std::vector& strides, + std::vector& paddings, framework::Tensor* output, + framework::Tensor* mask); }; -template +template class MaxPool3dWithIndexGradFunctor { public: - void operator()(const platform::DeviceContext& context, + void operator()(const DeviceContext& context, const framework::Tensor& output_grad, const framework::Tensor& mask, std::vector& ksize, std::vector& strides, std::vector& paddings, diff --git a/paddle/operators/math/selected_rows_functor.cc b/paddle/operators/math/selected_rows_functor.cc index 075196b47eeaf118a588b96532d87a05e4e600c6..8a1ebb58c26578f076bf243adfbd51d10c682b99 100644 --- a/paddle/operators/math/selected_rows_functor.cc +++ b/paddle/operators/math/selected_rows_functor.cc @@ -12,15 +12,17 @@ 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/operators/math/selected_rows_functor.h" +#include + #include "paddle/operators/math/math_function.h" +#include "paddle/operators/math/selected_rows_functor.h" namespace paddle { namespace operators { namespace math { template -struct SelectedRowsAdd { - void operator()(const platform::DeviceContext& context, +struct SelectedRowsAdd { + void operator()(const platform::CPUDeviceContext& context, const framework::SelectedRows& input1, const framework::SelectedRows& input2, framework::SelectedRows* output) { @@ -67,12 +69,12 @@ struct SelectedRowsAdd { } }; -template struct SelectedRowsAdd; -template struct SelectedRowsAdd; +template struct SelectedRowsAdd; +template struct SelectedRowsAdd; template -struct SelectedRowsAddTensor { - void operator()(const platform::DeviceContext& context, +struct SelectedRowsAddTensor { + void operator()(const platform::CPUDeviceContext& context, const framework::SelectedRows& input1, const framework::Tensor& input2, framework::Tensor* output) { auto in1_height = input1.height(); @@ -88,7 +90,7 @@ struct SelectedRowsAddTensor { PADDLE_ENFORCE_EQ(in1_row_numel, input2.numel() / in1_height); PADDLE_ENFORCE_EQ(in1_row_numel, output->numel() / in1_height); - SetConstant functor; + SetConstant functor; functor(context, output, 0.0); auto* in1_data = in1_value.data(); @@ -103,17 +105,16 @@ struct SelectedRowsAddTensor { auto out_eigen = framework::EigenVector::Flatten(*output); auto in2_eigen = framework::EigenVector::Flatten(input2); - out_eigen.device(*context.GetEigenDevice()) = - out_eigen + in2_eigen; + out_eigen.device(*context.eigen_device()) = out_eigen + in2_eigen; } }; -template struct SelectedRowsAddTensor; -template struct SelectedRowsAddTensor; +template struct SelectedRowsAddTensor; +template struct SelectedRowsAddTensor; template -struct SelectedRowsAddTo { - void operator()(const platform::DeviceContext& context, +struct SelectedRowsAddTo { + void operator()(const platform::CPUDeviceContext& context, const framework::SelectedRows& input1, const int64_t input2_offset, framework::SelectedRows* input2) { @@ -143,12 +144,14 @@ struct SelectedRowsAddTo { } }; -template struct SelectedRowsAddTo; -template struct SelectedRowsAddTo; +template struct SelectedRowsAddTo; +template struct SelectedRowsAddTo; +template struct SelectedRowsAddTo; +template struct SelectedRowsAddTo; template -struct SelectedRowsAddToTensor { - void operator()(const platform::DeviceContext& context, +struct SelectedRowsAddToTensor { + void operator()(const platform::CPUDeviceContext& context, const framework::SelectedRows& input1, framework::Tensor* input2) { auto in1_height = input1.height(); @@ -173,9 +176,123 @@ struct SelectedRowsAddToTensor { } }; -template struct SelectedRowsAddToTensor; -template struct SelectedRowsAddToTensor; +template struct SelectedRowsAddToTensor; +template struct SelectedRowsAddToTensor; +template struct SelectedRowsAddToTensor; +template struct SelectedRowsAddToTensor; + +// This is a separated namespace for manipulate SelectedRows typed +// data. Like merge duplicated rows, adding two SelectedRows etc. +// +// Another group of functors is called "scatter updates", which means +// use SelectedRows to update a dense tensor with different Ops, like +// add or mul. +namespace scatter { + +size_t FindPos(const std::vector& rows, int64_t value) { + return std::find(rows.begin(), rows.end(), value) - rows.begin(); +} + +template +struct MergeAdd { + framework::SelectedRows operator()(const platform::CPUDeviceContext& context, + const framework::SelectedRows& input) { + framework::SelectedRows out; + auto input_rows = input.rows(); + std::set row_set(input_rows.begin(), input_rows.end()); + std::vector merge_rows(row_set.begin(), row_set.end()); + + auto input_width = input.value().dims()[1]; + out.set_rows(merge_rows); + out.set_height(input.height()); + out.mutable_value()->mutable_data( + framework::make_ddim( + {static_cast(merge_rows.size()), input_width}), + context.GetPlace()); + + math::SetConstant constant_functor; + constant_functor(context, out.mutable_value(), 0.0); + + auto* out_data = out.mutable_value()->data(); + auto* input_data = input.value().data(); + + for (size_t i = 0; i < input_rows.size(); i++) { + size_t out_i = FindPos(merge_rows, input_rows[i]); + for (int64_t j = 0; j < input_width; j++) { + out_data[out_i * input_width + j] += input_data[i * input_width + j]; + } + } + return out; + } +}; + +template struct MergeAdd; +template struct MergeAdd; +template struct MergeAdd; +template struct MergeAdd; + +template +struct UpdateToTensor { + void operator()(const platform::CPUDeviceContext& context, + const ScatterOps& op, const framework::SelectedRows& input1, + framework::Tensor* input2) { + auto in1_height = input1.height(); + auto in2_dims = input2->dims(); + PADDLE_ENFORCE_EQ(in1_height, in2_dims[0]); + + auto& in1_value = input1.value(); + auto& in1_rows = input1.rows(); + + int64_t in1_row_numel = in1_value.numel() / in1_rows.size(); + PADDLE_ENFORCE_EQ(in1_row_numel, input2->numel() / in1_height); + + auto* in1_data = in1_value.data(); + auto* input2_data = input2->data(); + + // FIXME(typhoonzero): use macro fix the below messy code. + switch (op) { + case ScatterOps::ASSIGN: + INLINE_FOR2(in1_rows.size(), in1_row_numel) + input2_data[in1_rows[i] * in1_row_numel + j] = + in1_data[i * in1_row_numel + j]; + break; + case ScatterOps::ADD: + INLINE_FOR2(in1_rows.size(), in1_row_numel) + input2_data[in1_rows[i] * in1_row_numel + j] += + in1_data[i * in1_row_numel + j]; + break; + case ScatterOps::SUB: + INLINE_FOR2(in1_rows.size(), in1_row_numel) + input2_data[in1_rows[i] * in1_row_numel + j] -= + in1_data[i * in1_row_numel + j]; + break; + case ScatterOps::SUBBY: + INLINE_FOR2(in1_rows.size(), in1_row_numel) + input2_data[in1_rows[i] * in1_row_numel + j] = + in1_data[i * in1_row_numel + j] - + input2_data[in1_rows[i] * in1_row_numel + j]; + break; + case ScatterOps::MUL: + INLINE_FOR2(in1_rows.size(), in1_row_numel) + input2_data[in1_rows[i] * in1_row_numel + j] *= + in1_data[i * in1_row_numel + j]; + break; + case ScatterOps::DIV: + INLINE_FOR2(in1_rows.size(), in1_row_numel) + input2_data[in1_rows[i] * in1_row_numel + j] /= + in1_data[i * in1_row_numel + j]; + break; + case ScatterOps::DIVBY: + INLINE_FOR2(in1_rows.size(), in1_row_numel) + input2_data[in1_rows[i] * in1_row_numel + j] = + in1_data[i * in1_row_numel + j] / + input2_data[in1_rows[i] * in1_row_numel + j]; + break; + } + } +}; +} // namespace scatter } // namespace math } // namespace operators } // namespace paddle diff --git a/paddle/operators/math/selected_rows_functor.cu b/paddle/operators/math/selected_rows_functor.cu index 47fe3b44a50fee9f41ae807793187258159b9f29..0ee456f9bc61436bd0f2f8ef20dd1654e7e56d56 100644 --- a/paddle/operators/math/selected_rows_functor.cu +++ b/paddle/operators/math/selected_rows_functor.cu @@ -12,6 +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 "paddle/operators/math/math_function.h" #include "paddle/operators/math/selected_rows_functor.h" #include "paddle/platform/cuda_helper.h" @@ -20,8 +22,8 @@ namespace paddle { namespace operators { namespace math { template -struct SelectedRowsAdd { - void operator()(const platform::DeviceContext& context, +struct SelectedRowsAdd { + void operator()(const platform::CUDADeviceContext& context, const framework::SelectedRows& input1, const framework::SelectedRows& input2, framework::SelectedRows* output) { @@ -58,22 +60,21 @@ struct SelectedRowsAdd { PADDLE_ENFORCE(platform::is_gpu_place(out_place)); memory::Copy( - boost::get(out_place), out_data, - boost::get(in1_place), in1_data, + boost::get(out_place), out_data, + boost::get(in1_place), in1_data, in1_value.numel() * sizeof(T), reinterpret_cast(context).stream()); auto* in2_data = in2_value.data(); - memory::Copy( - boost::get(out_place), out_data + in1_value.numel(), - boost::get(in2_place), in2_data, - in2_value.numel() * sizeof(T), - reinterpret_cast(context).stream()); + memory::Copy(boost::get(out_place), + out_data + in1_value.numel(), + boost::get(in2_place), in2_data, + in2_value.numel() * sizeof(T), context.stream()); } }; -template struct SelectedRowsAdd; -template struct SelectedRowsAdd; +template struct SelectedRowsAdd; +template struct SelectedRowsAdd; namespace { template @@ -96,8 +97,8 @@ __global__ void SelectedRowsAddTensorKernel(const T* selected_rows, } // namespace template -struct SelectedRowsAddTensor { - void operator()(const platform::DeviceContext& context, +struct SelectedRowsAddTensor { + void operator()(const platform::CUDADeviceContext& context, const framework::SelectedRows& input1, const framework::Tensor& input2, framework::Tensor* output) { auto in1_height = input1.height(); @@ -117,30 +118,28 @@ struct SelectedRowsAddTensor { auto* in2_data = input2.data(); auto* out_data = output->data(); - SetConstant functor; + SetConstant functor; functor(context, output, 0.0); const int block_size = 256; dim3 threads(block_size, 1); dim3 grid(1, in1_rows.size()); - SelectedRowsAddTensorKernel<<< - grid, threads, 0, - reinterpret_cast(context) - .stream()>>>(in1_data, in1_rows.data(), out_data, in1_row_numel); + SelectedRowsAddTensorKernel< + T, block_size><<>>( + in1_data, in1_rows.data(), out_data, in1_row_numel); auto out_eigen = framework::EigenVector::Flatten(*output); auto in2_eigen = framework::EigenVector::Flatten(input2); - out_eigen.device(*context.GetEigenDevice()) = - out_eigen + in2_eigen; + out_eigen.device(*context.eigen_device()) = out_eigen + in2_eigen; } }; -template struct SelectedRowsAddTensor; -template struct SelectedRowsAddTensor; +template struct SelectedRowsAddTensor; +template struct SelectedRowsAddTensor; template -struct SelectedRowsAddTo { - void operator()(const platform::DeviceContext& context, +struct SelectedRowsAddTo { + void operator()(const platform::CUDADeviceContext& context, const framework::SelectedRows& input1, const int64_t input2_offset, framework::SelectedRows* input2) { @@ -163,16 +162,17 @@ struct SelectedRowsAddTo { auto* in1_data = in1_value.data(); auto* in2_data = in2_value->data(); - memory::Copy( - boost::get(in2_place), in2_data + input2_offset, - boost::get(in1_place), in1_data, - in1_value.numel() * sizeof(T), - reinterpret_cast(context).stream()); + memory::Copy(boost::get(in2_place), + in2_data + input2_offset, + boost::get(in1_place), in1_data, + in1_value.numel() * sizeof(T), context.stream()); } }; -template struct SelectedRowsAddTo; -template struct SelectedRowsAddTo; +template struct SelectedRowsAddTo; +template struct SelectedRowsAddTo; +template struct SelectedRowsAddTo; +template struct SelectedRowsAddTo; namespace { template @@ -195,8 +195,8 @@ __global__ void SelectedRowsAddToTensorKernel(const T* selected_rows, } // namespace template -struct SelectedRowsAddToTensor { - void operator()(const platform::DeviceContext& context, +struct SelectedRowsAddToTensor { + void operator()(const platform::CUDADeviceContext& context, const framework::SelectedRows& input1, framework::Tensor* input2) { auto in1_height = input1.height(); @@ -214,16 +214,167 @@ struct SelectedRowsAddToTensor { const int block_size = 256; dim3 threads(block_size, 1); dim3 grid(1, in1_rows.size()); - SelectedRowsAddToTensorKernel<<< - grid, threads, 0, - reinterpret_cast(context) - .stream()>>>(in1_data, in1_rows.data(), in2_data, in1_row_numel); + SelectedRowsAddToTensorKernel< + T, block_size><<>>( + in1_data, in1_rows.data(), in2_data, in1_row_numel); } }; -template struct SelectedRowsAddToTensor; -template struct SelectedRowsAddToTensor; +template struct SelectedRowsAddToTensor; +template struct SelectedRowsAddToTensor; +template struct SelectedRowsAddToTensor; +template struct SelectedRowsAddToTensor; + +namespace scatter { +template +__global__ void MergeAddKernel(const T* input, const int64_t* input_rows, + T* out, const int64_t* out_rows, + size_t out_rows_size, int64_t row_numel) { + const int ty = blockIdx.y; + int tid = threadIdx.x; + __shared__ size_t out_idx; + + if (tid == 0) { + for (size_t i = 0; i < out_rows_size; i++) { + if (input_rows[ty] == out_rows[i]) { + out_idx = i; + } + } + } + + __syncthreads(); + + input += ty * row_numel; + out += out_idx * row_numel; + for (int index = tid; index < row_numel; index += block_size) { + paddle::platform::CudaAtomicAdd(out + index, input[index]); + } +} + +template +struct MergeAdd { + framework::SelectedRows operator()(const platform::CUDADeviceContext& context, + const framework::SelectedRows& input) { + framework::SelectedRows out; + auto input_rows = input.rows(); + std::set row_set(input_rows.begin(), input_rows.end()); + std::vector merge_rows(row_set.begin(), row_set.end()); + + auto input_width = input.value().dims()[1]; + + out.set_rows(merge_rows); + out.set_height(input.height()); + out.mutable_value()->mutable_data( + framework::make_ddim( + {static_cast(merge_rows.size()), input_width}), + context.GetPlace()); + + math::SetConstant constant_functor; + constant_functor(context, out.mutable_value(), 0.0); + + auto* out_data = out.mutable_value()->data(); + auto* input_data = input.value().data(); + + const int block_size = 256; + dim3 threads(block_size, 1); + dim3 grid1(1, input_rows.size()); + + MergeAddKernel< + T, 256><<(context) + .stream()>>>(input_data, input.rows().data(), out_data, + out.rows().data(), out.rows().size(), + input_width); + return out; + } +}; + +template struct MergeAdd; +template struct MergeAdd; +template struct MergeAdd; +template struct MergeAdd; + +template +__global__ void UpdateToTensorKernel(const T* selected_rows, + const int64_t* rows, const ScatterOps& op, + T* tensor_out, int64_t row_numel) { + const int ty = blockIdx.y; + int tid = threadIdx.x; + + selected_rows += ty * row_numel; + tensor_out += rows[ty] * row_numel; + // FIXME(typhoonzero): use macro fix the below messy code. + switch (op) { + case ScatterOps::ASSIGN: + for (int index = tid; index < row_numel; index += block_size) { + tensor_out[index] = selected_rows[index]; + } + break; + case ScatterOps::ADD: + for (int index = tid; index < row_numel; index += block_size) { + tensor_out[index] += selected_rows[index]; + } + break; + case ScatterOps::SUB: + for (int index = tid; index < row_numel; index += block_size) { + tensor_out[index] -= selected_rows[index]; + } + break; + case ScatterOps::SUBBY: + for (int index = tid; index < row_numel; index += block_size) { + tensor_out[index] = selected_rows[index] - tensor_out[index]; + } + break; + case ScatterOps::MUL: + for (int index = tid; index < row_numel; index += block_size) { + tensor_out[index] *= selected_rows[index]; + } + break; + case ScatterOps::DIV: + for (int index = tid; index < row_numel; index += block_size) { + tensor_out[index] /= selected_rows[index]; + } + break; + case ScatterOps::DIVBY: + for (int index = tid; index < row_numel; index += block_size) { + tensor_out[index] = selected_rows[index] / tensor_out[index]; + } + break; + } +} + +template +struct UpdateToTensor { + void operator()(const platform::CUDADeviceContext& context, + const ScatterOps& op, const framework::SelectedRows& input1, + framework::Tensor* input2) { + // NOTE: Use SelectedRowsAddToTensor for better performance + // no additional MergeAdd called. + MergeAdd merge_func; + auto merged_in1 = merge_func(context, input1); + + auto in1_height = merged_in1.height(); + auto in2_dims = input2->dims(); + PADDLE_ENFORCE_EQ(in1_height, in2_dims[0]); + + auto& in1_value = merged_in1.value(); + auto& in1_rows = merged_in1.rows(); + + int64_t in1_row_numel = in1_value.numel() / in1_rows.size(); + PADDLE_ENFORCE_EQ(in1_row_numel, input2->numel() / in1_height); + + auto* in1_data = in1_value.template data(); + auto* in2_data = input2->data(); + + dim3 threads(platform::PADDLE_CUDA_NUM_THREADS, 1); + dim3 grid(1, in1_rows.size()); + UpdateToTensorKernel<<< + grid, threads, 0, context.stream()>>>(in1_data, in1_rows.data(), op, + in2_data, in1_row_numel); + } +}; +} // namespace scatter } // namespace math } // namespace operators } // namespace paddle diff --git a/paddle/operators/math/selected_rows_functor.h b/paddle/operators/math/selected_rows_functor.h index d6dc6c03c941f965394d952574d309c51eb82a62..09d4631905f90f78772368ad71b11826877bdc34 100644 --- a/paddle/operators/math/selected_rows_functor.h +++ b/paddle/operators/math/selected_rows_functor.h @@ -12,46 +12,123 @@ 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/framework/eigen.h" #include "paddle/framework/selected_rows.h" #include "paddle/platform/device_context.h" +#define INLINE_FOR2(sizei, sizej) \ + for (int64_t i = 0; i < sizei; i++) \ + for (int64_t j = 0; j < sizej; j++) + namespace paddle { namespace operators { namespace math { // SelectedRows + SelectedRows will simplely concat value and rows. // The real computation happens in dealing with LoDTensor. -template +template struct SelectedRowsAdd { - void operator()(const platform::DeviceContext& context, + void operator()(const DeviceContext& context, const framework::SelectedRows& input1, const framework::SelectedRows& input2, framework::SelectedRows* output); }; -template +template struct SelectedRowsAddTensor { - void operator()(const platform::DeviceContext& context, + void operator()(const DeviceContext& context, const framework::SelectedRows& input1, const framework::Tensor& input2, framework::Tensor* output); }; // input2 = input1 + input2 -template +template struct SelectedRowsAddTo { - void operator()(const platform::DeviceContext& context, + void operator()(const DeviceContext& context, const framework::SelectedRows& input1, const int64_t input2_offset, framework::SelectedRows* input2); }; // input2 = input1 + input2 -template +template struct SelectedRowsAddToTensor { - void operator()(const platform::DeviceContext& context, + void operator()(const DeviceContext& context, + const framework::SelectedRows& input1, + framework::Tensor* input2); +}; + +namespace scatter { +// functors for manuplating SelectedRows data +template +struct MergeAdd { + // unary functor, merge by adding duplicated rows in + // the input SelectedRows object. + framework::SelectedRows operator()(const DeviceContext& context, + const framework::SelectedRows& input); +}; + +template +struct Add { + framework::SelectedRows operator()(const DeviceContext& context, + const framework::SelectedRows& input1, + const framework::SelectedRows& input2) { + framework::SelectedRows out; + out.set_rows(input1.rows()); + out.set_height(input1.height()); + out.mutable_value()->mutable_data(input1.value().dims(), + context.GetPlace()); + auto e_out = framework::EigenVector::Flatten(*(out.mutable_value())); + auto e_in1 = framework::EigenVector::Flatten(input1.value()); + auto e_in2 = framework::EigenVector::Flatten(input2.value()); + e_out.device(*context.eigen_device()) = e_in1 + e_in2; + return out; + } +}; + +template +struct Mul { + // multiply two SelectedRows + framework::SelectedRows operator()(const DeviceContext& context, + const framework::SelectedRows& input1, + const framework::SelectedRows& input2) { + framework::SelectedRows out; + out.set_rows(input1.rows()); + out.set_height(input1.height()); + out.mutable_value()->mutable_data(input1.value().dims(), + context.GetPlace()); + auto e_out = framework::EigenVector::Flatten(*(out.mutable_value())); + auto e_in1 = framework::EigenVector::Flatten(input1.value()); + auto e_in2 = framework::EigenVector::Flatten(input2.value()); + e_out.device(*context.eigen_device()) = e_in1 * e_in2; + return out; + } + // multiply scalar to SelectedRows + framework::SelectedRows operator()(const DeviceContext& context, + const framework::SelectedRows& input1, + const T input2) { + framework::SelectedRows out; + out.set_rows(input1.rows()); + out.set_height(input1.height()); + out.mutable_value()->mutable_data(input1.value().dims(), + context.GetPlace()); + auto e_out = framework::EigenVector::Flatten(*(out.mutable_value())); + auto e_in1 = framework::EigenVector::Flatten(input1.value()); + e_out.device(*context.eigen_device()) = input2 * e_in1; + return out; + } +}; + +enum class ScatterOps { ASSIGN, ADD, SUB, SUBBY, MUL, DIV, DIVBY }; + +// out = seleted_rows_in / tensor +template +struct UpdateToTensor { + void operator()(const DeviceContext& context, const ScatterOps& op, const framework::SelectedRows& input1, framework::Tensor* input2); }; +} // namespace scatter } // namespace math } // namespace operators } // namespace paddle diff --git a/paddle/operators/math/selected_rows_functor_test.cc b/paddle/operators/math/selected_rows_functor_test.cc index a3649b6875aca61ee3ceb1ca83c7f9b38dc06c42..8c74cab0a1e817f9e98fa682fe4122db7837aec9 100644 --- a/paddle/operators/math/selected_rows_functor_test.cc +++ b/paddle/operators/math/selected_rows_functor_test.cc @@ -23,7 +23,7 @@ TEST(selected_rows_functor, cpu_add) { CPUPlace cpu_place; CPUDeviceContext ctx(cpu_place); - SetConstant functor; + SetConstant functor; int64_t height = 10; int64_t row_numel = 10; @@ -47,7 +47,7 @@ TEST(selected_rows_functor, cpu_add) { // simplely concat two SelectedRows out_value->mutable_data(make_ddim({7, 10}), cpu_place); - SelectedRowsAdd add_functor; + SelectedRowsAdd add_functor; add_functor(ctx, *selected_rows1, *selected_rows2, output.get()); auto out_height = output->height(); @@ -85,7 +85,7 @@ TEST(selected_rows_functor, cpu_add) { std::unique_ptr tensor2{new Tensor()}; tensor2->mutable_data(make_ddim({height, row_numel}), cpu_place); - SelectedRowsAddTensor add_tensor_functor; + SelectedRowsAddTensor add_tensor_functor; add_tensor_functor(ctx, *output, *tensor1, tensor2.get()); auto* tensor2_data = tensor2->data(); @@ -112,7 +112,7 @@ TEST(selected_rows_functor, cpu_add_to) { CPUPlace cpu_place; CPUDeviceContext ctx(cpu_place); - SetConstant functor; + SetConstant functor; int64_t height = 10; int64_t row_numel = 10; @@ -137,7 +137,7 @@ TEST(selected_rows_functor, cpu_add_to) { // simplely concat two SelectedRows out_value->mutable_data(make_ddim({7, 10}), cpu_place); - SelectedRowsAddTo add_to_functor; + SelectedRowsAddTo add_to_functor; add_to_functor(ctx, *selected_rows1, 0, output.get()); add_to_functor(ctx, *selected_rows2, in1_value->numel(), output.get()); @@ -173,7 +173,7 @@ TEST(selected_rows_functor, cpu_add_to) { tensor1->mutable_data(make_ddim({height, row_numel}), cpu_place); functor(ctx, tensor1.get(), 3.0); - SelectedRowsAddToTensor add_to_tensor_functor; + SelectedRowsAddToTensor add_to_tensor_functor; add_to_tensor_functor(ctx, *output, tensor1.get()); auto* tensor1_data = tensor1->data(); diff --git a/paddle/operators/math/selected_rows_functor_test.cu b/paddle/operators/math/selected_rows_functor_test.cu index 09de9dc53a1de9537b5109b3cc7cf9744f9c7908..0a2e36f68acee04bd6b272d37679c18231cb8760 100644 --- a/paddle/operators/math/selected_rows_functor_test.cu +++ b/paddle/operators/math/selected_rows_functor_test.cu @@ -21,10 +21,10 @@ TEST(selected_rows_functor, gpu_add) { using namespace paddle::platform; using namespace paddle::operators::math; - GPUPlace gpu_place(0); + CUDAPlace gpu_place(0); CPUPlace cpu_place; CUDADeviceContext ctx(gpu_place); - SetConstant functor; + SetConstant functor; int64_t height = 10; int64_t row_numel = 10; @@ -48,7 +48,7 @@ TEST(selected_rows_functor, gpu_add) { // simplely concat two SelectedRows out_value->mutable_data(make_ddim({7, 10}), gpu_place); - SelectedRowsAdd add_functor; + SelectedRowsAdd add_functor; add_functor(ctx, *selected_rows1, *selected_rows2, output.get()); auto out_height = output->height(); @@ -67,7 +67,7 @@ TEST(selected_rows_functor, gpu_add) { EXPECT_EQ(out_rows[6], 9); Tensor out_cpu; - out_cpu.CopyFrom(*out_value, cpu_place, ctx); + CopyFrom(*out_value, cpu_place, ctx, &out_cpu); ctx.Wait(); auto* out_cpu_data = out_cpu.data(); @@ -90,11 +90,11 @@ TEST(selected_rows_functor, gpu_add) { std::unique_ptr tensor2{new Tensor()}; tensor2->mutable_data(make_ddim({height, row_numel}), gpu_place); - SelectedRowsAddTensor add_tensor_functor; + SelectedRowsAddTensor add_tensor_functor; add_tensor_functor(ctx, *output, *tensor1, tensor2.get()); Tensor tensor2_cpu; - tensor2_cpu.CopyFrom(*tensor2, cpu_place, ctx); + CopyFrom(*tensor2, cpu_place, ctx, &tensor2_cpu); ctx.Wait(); auto* tensor2_cpu_data = tensor2_cpu.data(); @@ -119,10 +119,10 @@ TEST(selected_rows_functor, gpu_add_to) { using namespace paddle::platform; using namespace paddle::operators::math; - GPUPlace gpu_place(0); + CUDAPlace gpu_place(0); CPUPlace cpu_place; CUDADeviceContext ctx(gpu_place); - SetConstant functor; + SetConstant functor; int64_t height = 10; int64_t row_numel = 10; @@ -147,7 +147,7 @@ TEST(selected_rows_functor, gpu_add_to) { // simplely concat two SelectedRows out_value->mutable_data(make_ddim({7, 10}), gpu_place); - SelectedRowsAddTo add_to_functor; + SelectedRowsAddTo add_to_functor; add_to_functor(ctx, *selected_rows1, 0, output.get()); add_to_functor(ctx, *selected_rows2, in1_value->numel(), output.get()); @@ -167,7 +167,7 @@ TEST(selected_rows_functor, gpu_add_to) { EXPECT_EQ(out_rows[6], 9); Tensor out_cpu; - out_cpu.CopyFrom(*out_value, cpu_place, ctx); + CopyFrom(*out_value, cpu_place, ctx, &out_cpu); ctx.Wait(); auto* out_cpu_data = out_cpu.data(); @@ -187,11 +187,11 @@ TEST(selected_rows_functor, gpu_add_to) { tensor1->mutable_data(make_ddim({height, row_numel}), gpu_place); functor(ctx, tensor1.get(), 3.0); - SelectedRowsAddToTensor add_to_tensor_functor; + SelectedRowsAddToTensor add_to_tensor_functor; add_to_tensor_functor(ctx, *output, tensor1.get()); Tensor tensor1_cpu; - tensor1_cpu.CopyFrom(*tensor1, cpu_place, ctx); + CopyFrom(*tensor1, cpu_place, ctx, &tensor1_cpu); ctx.Wait(); auto* tensor1_cpu_data = tensor1_cpu.data(); diff --git a/paddle/operators/math/sequence2batch.cc b/paddle/operators/math/sequence2batch.cc index 5b3bde02fbf981772759caa3d0054fac4a8520f9..88977be1f8c030741c3a3a8f07a4feeb1d8bb4d9 100644 --- a/paddle/operators/math/sequence2batch.cc +++ b/paddle/operators/math/sequence2batch.cc @@ -19,9 +19,9 @@ namespace operators { namespace math { template -class CopyMatrixRowsFunctor { +class CopyMatrixRowsFunctor { public: - void operator()(const platform::DeviceContext& context, + void operator()(const platform::CPUDeviceContext& context, const framework::Tensor& src, const size_t* index, framework::Tensor& dst, bool is_src_index) { auto src_dims = src.dims(); @@ -48,13 +48,13 @@ class CopyMatrixRowsFunctor { } }; -template class CopyMatrixRowsFunctor; -template class CopyMatrixRowsFunctor; +template class CopyMatrixRowsFunctor; +template class CopyMatrixRowsFunctor; -template class LoDTensor2BatchFunctor; -template class LoDTensor2BatchFunctor; -template class Batch2LoDTensorFunctor; -template class Batch2LoDTensorFunctor; +template class LoDTensor2BatchFunctor; +template class LoDTensor2BatchFunctor; +template class Batch2LoDTensorFunctor; +template class Batch2LoDTensorFunctor; } // namespace math } // namespace operators diff --git a/paddle/operators/math/sequence2batch.cu b/paddle/operators/math/sequence2batch.cu index c5d968aeb216bbb3e0e17f138b9e891494d99f75..452ae8951000872b706f7e4227a62dbf98109e7e 100644 --- a/paddle/operators/math/sequence2batch.cu +++ b/paddle/operators/math/sequence2batch.cu @@ -39,9 +39,9 @@ __global__ void CopyMatrixRowsKernel(const T* src, T* dst, const size_t* index, } template -class CopyMatrixRowsFunctor { +class CopyMatrixRowsFunctor { public: - void operator()(const platform::DeviceContext& context, + void operator()(const platform::CUDADeviceContext& context, const framework::Tensor& src, const size_t* index, framework::Tensor& dst, bool is_src_index) { auto src_dims = src.dims(); @@ -59,20 +59,19 @@ class CopyMatrixRowsFunctor { dim3 threads(128, 8); dim3 grid(8, 1); - auto stream = - reinterpret_cast(context).stream(); + auto stream = context.stream(); CopyMatrixRowsKernel<<>>( src_data, dst_data, index, height, width, is_src_index); } }; -template class CopyMatrixRowsFunctor; -template class CopyMatrixRowsFunctor; +template class CopyMatrixRowsFunctor; +template class CopyMatrixRowsFunctor; -template class LoDTensor2BatchFunctor; -template class LoDTensor2BatchFunctor; -template class Batch2LoDTensorFunctor; -template class Batch2LoDTensorFunctor; +template class LoDTensor2BatchFunctor; +template class LoDTensor2BatchFunctor; +template class Batch2LoDTensorFunctor; +template class Batch2LoDTensorFunctor; } // namespace math } // namespace operators diff --git a/paddle/operators/math/sequence2batch.h b/paddle/operators/math/sequence2batch.h index 73295ddbcb73fe80be08e732790f0ec75e94b415..a5c43a2c7d4d729c35a20a27de2a23141e6019bc 100644 --- a/paddle/operators/math/sequence2batch.h +++ b/paddle/operators/math/sequence2batch.h @@ -26,7 +26,7 @@ template using EigenMatrix = framework::EigenMatrix; -template +template class CopyMatrixRowsFunctor { public: // If is_src_index is true, @@ -34,12 +34,12 @@ class CopyMatrixRowsFunctor { // If is_src_index is false, // copy the input src to the indexed rows of output dst. // The indexed rows are based on the input index. - void operator()(const platform::DeviceContext& context, - const framework::Tensor& src, const size_t* index, - framework::Tensor& dst, bool is_src_index); + void operator()(const DeviceContext& context, const framework::Tensor& src, + const size_t* index, framework::Tensor& dst, + bool is_src_index); }; -template +template class LoDTensor2BatchFunctor { // Calculate the length of each sequence and // sort sequence index by the length. @@ -56,7 +56,7 @@ class LoDTensor2BatchFunctor { }; public: - void operator()(const platform::DeviceContext& context, + void operator()(const DeviceContext& context, const framework::LoDTensor& lod_tensor, framework::LoDTensor& batch, bool is_cal_batch_lod, bool is_reverse = false) const { @@ -65,7 +65,7 @@ class LoDTensor2BatchFunctor { PADDLE_ENFORCE_GT(lods.size(), 2UL); PADDLE_ENFORCE_EQ(lods[1].size(), static_cast(lod_tensor.dims()[0])); - CopyMatrixRowsFunctor to_batch; + CopyMatrixRowsFunctor to_batch; to_batch(context, lod_tensor, lods[1].data(), batch, true); return; } @@ -143,22 +143,22 @@ class LoDTensor2BatchFunctor { } batch.set_lod(batch_lods); - CopyMatrixRowsFunctor to_batch; + CopyMatrixRowsFunctor to_batch; to_batch(context, lod_tensor, seq2batch_idx, batch, true); } }; -template +template class Batch2LoDTensorFunctor { public: - void operator()(const platform::DeviceContext& context, + void operator()(const DeviceContext& context, const framework::LoDTensor& batch, 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])); - CopyMatrixRowsFunctor to_seq; + CopyMatrixRowsFunctor to_seq; size_t* index = in_lod[1].data(); to_seq(context, batch, index, lod_tensor, false); } diff --git a/paddle/operators/math/sequence_pooling.cc b/paddle/operators/math/sequence_pooling.cc index 5913c99fdb01100d0de44ab317124550fa626528..8fb92b1a130b8f25163d856f3f596136072180cf 100644 --- a/paddle/operators/math/sequence_pooling.cc +++ b/paddle/operators/math/sequence_pooling.cc @@ -20,9 +20,9 @@ namespace operators { namespace math { template -class MaxSeqPoolFunctor { +class MaxSeqPoolFunctor { public: - void operator()(const platform::DeviceContext& context, + void operator()(const platform::CPUDeviceContext& context, const framework::LoDTensor& input, framework::Tensor* output, framework::Tensor* index) { auto in_dims = input.dims(); @@ -60,9 +60,9 @@ class MaxSeqPoolFunctor { }; template -class MaxSeqPoolGradFunctor { +class MaxSeqPoolGradFunctor { public: - void operator()(const platform::DeviceContext& context, + void operator()(const platform::CPUDeviceContext& context, const framework::Tensor& out_grad, const framework::Tensor& index, framework::LoDTensor* in_grad) { @@ -80,7 +80,7 @@ class MaxSeqPoolGradFunctor { const int* max_index = index.data(); T* ig_data = in_grad->data(); - SetConstant set_zero; + SetConstant set_zero; set_zero(context, in_grad, static_cast(0.0)); int64_t num_seq = og_dims[0]; int64_t dim = out_grad.numel() / num_seq; @@ -93,10 +93,10 @@ class MaxSeqPoolGradFunctor { } }; -template class MaxSeqPoolFunctor; -template class MaxSeqPoolFunctor; -template class MaxSeqPoolGradFunctor; -template class MaxSeqPoolGradFunctor; +template class MaxSeqPoolFunctor; +template class MaxSeqPoolFunctor; +template class MaxSeqPoolGradFunctor; +template class MaxSeqPoolGradFunctor; } // namespace math } // namespace operators diff --git a/paddle/operators/math/sequence_pooling.cu b/paddle/operators/math/sequence_pooling.cu index 5ed951402fecba66a8960f4d024bf3785dac51c7..4c9e6b375ce7251747b9cd443d86cca0858c84ef 100644 --- a/paddle/operators/math/sequence_pooling.cu +++ b/paddle/operators/math/sequence_pooling.cu @@ -46,9 +46,9 @@ __global__ void KeMaxSequencePool(const T* input, const size_t* starts, } template -class MaxSeqPoolFunctor { +class MaxSeqPoolFunctor { public: - void operator()(const platform::DeviceContext& context, + void operator()(const platform::CUDADeviceContext& context, const framework::LoDTensor& input, framework::Tensor* output, framework::Tensor* index) { auto in_dims = input.dims(); @@ -71,8 +71,7 @@ class MaxSeqPoolFunctor { dim3 threads(256, 1); dim3 grid(num_seq, 1); - auto stream = - reinterpret_cast(context).stream(); + auto stream = context.stream(); KeMaxSequencePool<<>>( in_data, starts.data(), out_data, max_index, num_seq, dim); } @@ -91,9 +90,9 @@ __global__ void KeMaxSequencePoolGrad(const T* out_grad, const int* max_index, } template -class MaxSeqPoolGradFunctor { +class MaxSeqPoolGradFunctor { public: - void operator()(const platform::DeviceContext& context, + void operator()(const platform::CUDADeviceContext& context, const framework::Tensor& out_grad, const framework::Tensor& index, framework::LoDTensor* in_grad) { @@ -111,7 +110,7 @@ class MaxSeqPoolGradFunctor { const int* max_index = index.data(); T* ig_data = in_grad->data(); - SetConstant set_zero; + SetConstant set_zero; set_zero(context, in_grad, static_cast(0.0)); int64_t num_seq = og_dims[0]; int64_t dim = out_grad.numel() / num_seq; @@ -119,17 +118,16 @@ class MaxSeqPoolGradFunctor { unsigned int blocks = (num_seq * dim + 128 - 1) / 128; dim3 threads(128, 1); dim3 grid(blocks, 1); - auto stream = - reinterpret_cast(context).stream(); + auto stream = context.stream(); KeMaxSequencePoolGrad<<>>( og_data, max_index, ig_data, num_seq, dim); } }; -template class MaxSeqPoolFunctor; -template class MaxSeqPoolFunctor; -template class MaxSeqPoolGradFunctor; -template class MaxSeqPoolGradFunctor; +template class MaxSeqPoolFunctor; +template class MaxSeqPoolFunctor; +template class MaxSeqPoolGradFunctor; +template class MaxSeqPoolGradFunctor; } // namespace math } // namespace operators diff --git a/paddle/operators/math/sequence_pooling.h b/paddle/operators/math/sequence_pooling.h index 35dfe26de1a87a064410401244914d4e2a94176e..13ffb2ebef3a683b5e5fe64433a90237b944002e 100644 --- a/paddle/operators/math/sequence_pooling.h +++ b/paddle/operators/math/sequence_pooling.h @@ -23,18 +23,18 @@ namespace math { #define FLT_MAX __FLT_MAX__ -template +template class MaxSeqPoolFunctor { public: - void operator()(const platform::DeviceContext& context, + void operator()(const DeviceContext& context, const framework::LoDTensor& input, framework::Tensor* output, framework::Tensor* index); }; -template +template class MaxSeqPoolGradFunctor { public: - void operator()(const platform::DeviceContext& context, + void operator()(const DeviceContext& context, const framework::Tensor& out_grad, const framework::Tensor& index, framework::LoDTensor* in_grad); diff --git a/paddle/operators/math/softmax.cc b/paddle/operators/math/softmax.cc index 3e2f15d6c27f58818128f32fab0bd4c5f36b0050..72f10f35f4ef39b41fbc5e900313eafd7ba669e9 100644 --- a/paddle/operators/math/softmax.cc +++ b/paddle/operators/math/softmax.cc @@ -19,10 +19,10 @@ namespace paddle { namespace operators { namespace math { -template class SoftmaxFunctor; -template class SoftmaxFunctor; -template class SoftmaxGradFunctor; -template class SoftmaxGradFunctor; +template class SoftmaxFunctor; +template class SoftmaxFunctor; +template class SoftmaxGradFunctor; +template class SoftmaxGradFunctor; } // namespace math } // namespace operators diff --git a/paddle/operators/math/softmax.cu b/paddle/operators/math/softmax.cu index 4dbab51d46bdaaa506a6c242d0958c73687f4eb9..9e73f6a371c950ed6f81ee90216f7fd3899f73ce 100644 --- a/paddle/operators/math/softmax.cu +++ b/paddle/operators/math/softmax.cu @@ -21,10 +21,10 @@ namespace paddle { namespace operators { namespace math { -template class SoftmaxFunctor; -template class SoftmaxFunctor; -template class SoftmaxGradFunctor; -template class SoftmaxGradFunctor; +template class SoftmaxFunctor; +template class SoftmaxFunctor; +template class SoftmaxGradFunctor; +template class SoftmaxGradFunctor; } // namespace math } // namespace operators diff --git a/paddle/operators/math/softmax.h b/paddle/operators/math/softmax.h index fe1074650234c5beb5889e7efd713164769ad740..471f44d340cfd0d6305a9127c34289ef1663accb 100644 --- a/paddle/operators/math/softmax.h +++ b/paddle/operators/math/softmax.h @@ -19,19 +19,18 @@ namespace paddle { namespace operators { namespace math { -template +template class SoftmaxFunctor { public: - void operator()(const platform::DeviceContext& context, - const framework::Tensor* X, framework::Tensor* Y); + void operator()(const DeviceContext& context, const framework::Tensor* X, + framework::Tensor* Y); }; -template +template class SoftmaxGradFunctor { public: - void operator()(const platform::DeviceContext& context, - const framework::Tensor* y, const framework::Tensor* y_grad, - framework::Tensor* x_grad); + void operator()(const DeviceContext& context, const framework::Tensor* y, + const framework::Tensor* y_grad, framework::Tensor* x_grad); }; } // namespace math diff --git a/paddle/operators/math/softmax_impl.h b/paddle/operators/math/softmax_impl.h index 05793eeb3eeafaf36c301236197555b7b35e5803..82f597ff792decb1760f59e693026cd453432d05 100644 --- a/paddle/operators/math/softmax_impl.h +++ b/paddle/operators/math/softmax_impl.h @@ -32,10 +32,10 @@ struct ValueClip { } }; -template -void SoftmaxFunctor::operator()( - const platform::DeviceContext& context, const framework::Tensor* X, - framework::Tensor* Y) { +template +void SoftmaxFunctor::operator()(const DeviceContext& context, + const framework::Tensor* X, + framework::Tensor* Y) { auto logits = EigenMatrix::From(*X); auto softmax = EigenMatrix::From(*Y); @@ -56,19 +56,18 @@ void SoftmaxFunctor::operator()( .broadcast(one_by_class)) .unaryExpr(ValueClip()); - softmax.device(*context.GetEigenDevice()) = shifted_logits.exp(); - softmax.device(*context.GetEigenDevice()) = - (softmax * - softmax.sum(along_class) - .inverse() - .eval() - .reshape(batch_by_one) - .broadcast(one_by_class)); + softmax.device(*context.eigen_device()) = shifted_logits.exp(); + softmax.device(*context.eigen_device()) = (softmax * + softmax.sum(along_class) + .inverse() + .eval() + .reshape(batch_by_one) + .broadcast(one_by_class)); } -template -void SoftmaxGradFunctor::operator()( - const platform::DeviceContext& context, const framework::Tensor* y, +template +void SoftmaxGradFunctor::operator()( + const DeviceContext& context, const framework::Tensor* y, const framework::Tensor* y_grad, framework::Tensor* x_grad) { auto softmax = EigenMatrix::From(*y); auto softmax_grad = EigenMatrix::From(*y_grad); @@ -89,8 +88,7 @@ void SoftmaxGradFunctor::operator()( .eval() .reshape(batch_by_one) .broadcast(one_by_class); - logits_grad.device(*context.GetEigenDevice()) = - (softmax_grad - dot) * softmax; + logits_grad.device(*context.eigen_device()) = (softmax_grad - dot) * softmax; } } // namespace math diff --git a/paddle/operators/math/unpooling.cc b/paddle/operators/math/unpooling.cc new file mode 100644 index 0000000000000000000000000000000000000000..ecd3a647e00655a57d11c2f082bd1f81822cf92b --- /dev/null +++ b/paddle/operators/math/unpooling.cc @@ -0,0 +1,91 @@ +/* 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. */ + +#include "paddle/operators/math/unpooling.h" +namespace paddle { +namespace operators { +namespace math { +template +class Unpool2dMaxFunctor { + public: + void operator()(const platform::CPUDeviceContext& context, + const framework::Tensor& input, + const framework::Tensor& indices, framework::Tensor* output) { + const int batch_size = input.dims()[0]; + const int input_height = input.dims()[2]; + const int input_width = input.dims()[3]; + const int output_channels = output->dims()[1]; + const int output_height = output->dims()[2]; + const int output_width = output->dims()[3]; + int input_feasize = input_height * input_width; + int output_feasize = output_height * output_width; + const T* input_data = input.data(); + const int* indices_data = indices.data(); + T* output_data = output->mutable_data(context.GetPlace()); + for (int b = 0; b < batch_size; ++b) { + for (int c = 0; c < output_channels; ++c) { + for (int i = 0; i < input_feasize; ++i) { + int index = indices_data[i]; + PADDLE_ENFORCE(index < output_feasize, "err index in unpooling!"); + output_data[index] = input_data[i]; + } + input_data += input_feasize; + indices_data += input_feasize; + output_data += output_feasize; + } + } + } +}; +template +class Unpool2dMaxGradFunctor { + public: + void operator()(const platform::CPUDeviceContext& context, + const framework::Tensor& input, + const framework::Tensor& indices, + const framework::Tensor& output, + const framework::Tensor& output_grad, + 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]; + const int output_channels = output.dims()[1]; + const int output_height = output.dims()[2]; + const int output_width = output.dims()[3]; + int input_feasize = input_height * input_width; + int output_feasize = output_height * output_width; + const int* indices_data = indices.data(); + const T* output_grad_data = output_grad.data(); + T* input_grad_data = input_grad->mutable_data(context.GetPlace()); + + for (int b = 0; b < batch_size; ++b) { + for (int c = 0; c < output_channels; ++c) { + for (int i = 0; i < input_feasize; ++i) { + int index = indices_data[i]; + PADDLE_ENFORCE(index < output_feasize, "err index in unpooling!"); + input_grad_data[i] = output_grad_data[index]; + } + input_grad_data += input_feasize; + indices_data += input_feasize; + output_grad_data += output_feasize; + } + } + } +}; +template class Unpool2dMaxGradFunctor; +template class Unpool2dMaxGradFunctor; +template class Unpool2dMaxFunctor; +template class Unpool2dMaxFunctor; +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/math/unpooling.cu b/paddle/operators/math/unpooling.cu new file mode 100644 index 0000000000000000000000000000000000000000..ecbde0f6a798ba817c28714b37af8187d2e9555e --- /dev/null +++ b/paddle/operators/math/unpooling.cu @@ -0,0 +1,128 @@ +/* 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. */ + +#include "paddle/operators/math/unpooling.h" +#include "paddle/platform/cuda_helper.h" + +namespace paddle { +namespace operators { +namespace math { +template +__global__ void KernelUnpool2dMax(const int nthreads, const T* input_data, + const int* indices_data, + const int input_height, const int input_width, + const int channels, T* output_data, + const int output_height, + const int output_width) { + int in_n_stride = input_height * input_width * channels; + int in_c_stride = input_height * input_width; + int out_n_stride = output_height * output_width * channels; + int out_c_stride = output_height * output_width; + int index = blockIdx.x * blockDim.x + threadIdx.x; + int offset = blockDim.x * gridDim.x; + for (int i = index; i < nthreads; i += offset) { + int bidx = i / in_n_stride; + int boffset = i % in_n_stride; + int cidx = boffset / in_c_stride; + int out_offset = bidx * out_n_stride + cidx * out_c_stride; + int out_index = indices_data[i]; + PADDLE_ASSERT(out_index < out_c_stride); + output_data[out_offset + out_index] = input_data[i]; + } +} +template +__global__ void KernelUnpool2dMaxGrad( + const int nthreads, const T* input_data, const int* indices_data, + const int input_height, const int input_width, const int channels, + const T* output_data, const T* output_grad, const int output_height, + const int output_width, T* input_grad) { + int in_n_stride = input_height * input_width * channels; + int in_c_stride = input_height * input_width; + int out_n_stride = output_height * output_width * channels; + int out_c_stride = output_height * output_width; + int index = blockIdx.x * blockDim.x + threadIdx.x; + int offset = blockDim.x * gridDim.x; + for (int i = index; i < nthreads; i += offset) { + int bidx = i / in_n_stride; + int boffset = i % in_n_stride; + int cidx = boffset / in_c_stride; + int out_offset = bidx * out_n_stride + cidx * out_c_stride; + int out_index = indices_data[i]; + PADDLE_ASSERT(out_index < out_c_stride); + input_grad[i] = output_grad[out_offset + out_index]; + } +} +/* + * All tensors are in NCHW format. + */ +template +class Unpool2dMaxFunctor { + public: + void operator()(const platform::CUDADeviceContext& context, + const framework::Tensor& input, + const framework::Tensor& indices, framework::Tensor* output) { + const int batch_size = input.dims()[0]; + const int input_height = input.dims()[2]; + const int input_width = input.dims()[3]; + const int output_channels = output->dims()[1]; + const int output_height = output->dims()[2]; + const int output_width = output->dims()[3]; + const T* input_data = input.data(); + const int* indices_data = indices.data(); + T* output_data = output->mutable_data(context.GetPlace()); + int threads = 1024; + int grid = (input.numel() + threads - 1) / threads; + KernelUnpool2dMax<<>>( + input.numel(), input_data, indices_data, input_height, input_width, + output_channels, output_data, output_height, output_width); + } +}; +/* + * All tensors are in NCHW format. + */ +template +class Unpool2dMaxGradFunctor { + public: + void operator()(const platform::CUDADeviceContext& context, + const framework::Tensor& input, + const framework::Tensor& indices, + const framework::Tensor& output, + const framework::Tensor& output_grad, + 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]; + const int output_channels = output.dims()[1]; + const int output_height = output.dims()[2]; + const int output_width = output.dims()[3]; + const T* input_data = input.data(); + const int* indices_data = indices.data(); + const T* output_data = output.data(); + const T* output_grad_data = output_grad.data(); + T* input_grad_data = input_grad->mutable_data(context.GetPlace()); + int threads = 1024; + int grid = (input.numel() + threads - 1) / threads; + KernelUnpool2dMaxGrad<<>>( + input.numel(), input_data, indices_data, input_height, input_width, + output_channels, output_data, output_grad_data, output_height, + output_width, input_grad_data); + } +}; +template class Unpool2dMaxGradFunctor; +template class Unpool2dMaxGradFunctor; +template class Unpool2dMaxFunctor; +template class Unpool2dMaxFunctor; +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/math/unpooling.h b/paddle/operators/math/unpooling.h new file mode 100644 index 0000000000000000000000000000000000000000..0f0ff1371ebea8c7501aee1c7c45bc6a79de397e --- /dev/null +++ b/paddle/operators/math/unpooling.h @@ -0,0 +1,38 @@ +/* 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 +#include "paddle/framework/tensor.h" + +namespace paddle { +namespace operators { +namespace math { +template +class Unpool2dMaxFunctor { + public: + void operator()(const DeviceContext& context, const framework::Tensor& input, + const framework::Tensor& indices, framework::Tensor* output); +}; +template +class Unpool2dMaxGradFunctor { + public: + void operator()(const DeviceContext& context, const framework::Tensor& input, + const framework::Tensor& indices, + const framework::Tensor& output, + const framework::Tensor& output_grad, + framework::Tensor* input_grad); +}; +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/math/vol2col.cc b/paddle/operators/math/vol2col.cc index 99eb7fd46de42400a915d86706580d15b08a74a2..d574ed9234304d992a6e4a10fce0816aee7fa40a 100644 --- a/paddle/operators/math/vol2col.cc +++ b/paddle/operators/math/vol2col.cc @@ -25,9 +25,9 @@ namespace math { * output_depth, output_height, output_width] */ template -class Vol2ColFunctor { +class Vol2ColFunctor { public: - void operator()(const platform::DeviceContext& context, + void operator()(const platform::CPUDeviceContext& context, const framework::Tensor& vol, const std::vector& dilations, const std::vector& strides, @@ -111,9 +111,9 @@ class Vol2ColFunctor { * output_depth, output_height, output_width] */ template -class Col2VolFunctor { +class Col2VolFunctor { public: - void operator()(const platform::DeviceContext& context, + void operator()(const platform::CPUDeviceContext& context, const framework::Tensor& col, const std::vector& dilations, const std::vector& strides, @@ -190,10 +190,10 @@ class Col2VolFunctor { } }; -template class Vol2ColFunctor; -template class Vol2ColFunctor; -template class Col2VolFunctor; -template class Col2VolFunctor; +template class Vol2ColFunctor; +template class Vol2ColFunctor; +template class Col2VolFunctor; +template class Col2VolFunctor; } // namespace math } // namespace operators diff --git a/paddle/operators/math/vol2col.cu b/paddle/operators/math/vol2col.cu index dae3be858e9f47d0133aa37e8a5f90a0addf1dfd..b029442fe48dd27232d322aadec5864760e1b9ff 100644 --- a/paddle/operators/math/vol2col.cu +++ b/paddle/operators/math/vol2col.cu @@ -68,9 +68,9 @@ __global__ void vol2col(int num_kernels, const T* data_vol, int depth, * output_depth, output_height, output_width] */ template -class Vol2ColFunctor { +class Vol2ColFunctor { public: - void operator()(const platform::DeviceContext& context, + void operator()(const platform::CUDADeviceContext& context, const framework::Tensor& vol, const std::vector& dilations, const std::vector& strides, @@ -117,9 +117,7 @@ class Vol2ColFunctor { const int threads = 1024; const int blocks = (num_outputs + 1024 - 1) / 1024; - vol2col<<(context) - .stream()>>>( + vol2col<<>>( num_outputs, vol.data(), input_depth, input_height, input_width, dilations[0], dilations[1], dilations[2], filter_depth, filter_height, filter_width, strides[0], strides[1], strides[2], paddings[0], @@ -196,9 +194,9 @@ __global__ void col2vol(int num_kernels, const T* data_col, int depth, * output_depth, output_height, output_width] */ template -class Col2VolFunctor { +class Col2VolFunctor { public: - void operator()(const platform::DeviceContext& context, + void operator()(const platform::CUDADeviceContext& context, const framework::Tensor& col, const std::vector& dilations, const std::vector& strides, @@ -245,9 +243,7 @@ class Col2VolFunctor { const int threads = 1024; const int blocks = (num_kernels + 1024 - 1) / 1024; - col2vol<<(context) - .stream()>>>( + col2vol<<>>( num_kernels, col.data(), input_depth, input_height, input_width, dilations[0], dilations[1], dilations[2], filter_depth, filter_height, filter_width, strides[0], strides[1], strides[2], paddings[0], @@ -256,10 +252,10 @@ class Col2VolFunctor { } }; -template class Vol2ColFunctor; -template class Vol2ColFunctor; -template class Col2VolFunctor; -template class Col2VolFunctor; +template class Vol2ColFunctor; +template class Vol2ColFunctor; +template class Col2VolFunctor; +template class Col2VolFunctor; } // namespace math } // namespace operators diff --git a/paddle/operators/math/vol2col.h b/paddle/operators/math/vol2col.h index cbc30bd754608dd6e6def1a4097d69bdf0c942c3..dcd80370e8516d34b764b1ab3b0b98516e738bf6 100644 --- a/paddle/operators/math/vol2col.h +++ b/paddle/operators/math/vol2col.h @@ -15,6 +15,7 @@ limitations under the License. */ #pragma once #include "paddle/framework/tensor.h" +#include "paddle/framework/tensor_util.h" #include "paddle/platform/device_context.h" namespace paddle { @@ -62,22 +63,20 @@ namespace math { * \note The caller needs to ensure that volShape.inputChannels is equal to * colShape.inputChannels. */ -template +template class Vol2ColFunctor { public: - void operator()(const platform::DeviceContext& context, - const framework::Tensor& vol, + void operator()(const DeviceContext& context, const framework::Tensor& vol, const std::vector& dilations, const std::vector& strides, const std::vector& paddings, framework::Tensor* col) const; }; -template +template class Col2VolFunctor { public: - void operator()(const platform::DeviceContext& context, - const framework::Tensor& col, + void operator()(const DeviceContext& context, const framework::Tensor& col, const std::vector& dilations, const std::vector& strides, const std::vector& paddings, diff --git a/paddle/operators/math/vol2col_test.cc b/paddle/operators/math/vol2col_test.cc index c31c716842f30de67c29b803866b8c82ddcf4a41..3794f0e52d200a08253a979991da04ec564cae47 100644 --- a/paddle/operators/math/vol2col_test.cc +++ b/paddle/operators/math/vol2col_test.cc @@ -16,7 +16,7 @@ limitations under the License. */ #include #include -template +template void testVol2col() { paddle::framework::Tensor input; paddle::framework::Tensor input_tmp; @@ -24,18 +24,7 @@ void testVol2col() { paddle::framework::Tensor output_tmp; auto* place = new Place(); - paddle::platform::DeviceContext* context; - if (paddle::platform::is_cpu_place(*place)) { - context = - new paddle::platform::CPUDeviceContext(paddle::platform::CPUPlace()); - } else { -#ifdef PADDLE_WITH_CUDA - context = - new paddle::platform::CUDADeviceContext(paddle::platform::GPUPlace()); -#else - PADDLE_THROW("no GPU support"); -#endif // PADDLE_WITH_CUDA - } + DeviceContext* context = new DeviceContext(*place); /** * input = [[0, 1, 2, @@ -82,13 +71,13 @@ void testVol2col() { if (paddle::platform::is_cpu_place(*place)) { input = input_tmp; } else { - input.CopyFrom(input_tmp, *place, *context); + CopyFrom(input_tmp, *place, *context, &input); } output.mutable_data({1, filter_size, filter_size, filter_size, output_depth, output_height, output_width}, *place); - paddle::operators::math::Vol2ColFunctor vol2col; + paddle::operators::math::Vol2ColFunctor vol2col; vol2col(*context, input, dilations, strides, paddings, &output); float vol_2_col[] = {0, 1, 1, 2, 3, 4, 4, 5, 6, 7, 7, 8, 9, 10, 10, 11}; @@ -96,7 +85,7 @@ void testVol2col() { if (paddle::platform::is_cpu_place(*place)) { out_cfo_ptr = output.data(); } else { - output_tmp.CopyFrom(output, paddle::platform::CPUPlace(), *context); + CopyFrom(output, paddle::platform::CPUPlace(), *context, &output_tmp); out_cfo_ptr = output_tmp.data(); } @@ -110,17 +99,17 @@ void testVol2col() { if (paddle::platform::is_cpu_place(*place)) { input = input_tmp; } else { - input.CopyFrom(input_tmp, *place, *context); + CopyFrom(input_tmp, *place, *context, &input); } - paddle::operators::math::Col2VolFunctor col2vol; + paddle::operators::math::Col2VolFunctor col2vol; col2vol(*context, output, dilations, strides, paddings, &input); float* in_ptr; if (paddle::platform::is_cpu_place(*place)) { in_ptr = input.data(); } else { - input_tmp.CopyFrom(input, paddle::platform::CPUPlace(), *context); + CopyFrom(input, paddle::platform::CPUPlace(), *context, &input_tmp); in_ptr = input_tmp.data(); } @@ -130,8 +119,9 @@ void testVol2col() { } TEST(math, vol2col) { - testVol2col(); + testVol2col(); #ifdef PADDLE_WITH_CUDA - testVol2col(); + testVol2col(); #endif // PADDLE_WITH_CUDA } diff --git a/paddle/operators/matmul_op.cc b/paddle/operators/matmul_op.cc index 5a1a6154203d40186f1e41491194b19612931b1f..fd65d894d5749c97f860d614de354e89f6d9441d 100644 --- a/paddle/operators/matmul_op.cc +++ b/paddle/operators/matmul_op.cc @@ -130,7 +130,7 @@ class MatMulOp : public framework::OperatorWithKernel { class MatMulOpMaker : public framework::OpProtoAndCheckerMaker { public: - MatMulOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker) + MatMulOpMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "The first input of MatMul op"); AddInput("Y", "The second input of MatMul op"); @@ -206,7 +206,8 @@ class MatMulOpGrad : public framework::OperatorWithKernel { namespace ops = paddle::operators; REGISTER_OP(matmul, ops::MatMulOp, ops::MatMulOpMaker, matmul_grad, ops::MatMulOpGrad); -REGISTER_OP_CPU_KERNEL(matmul, - ops::MatMulKernel); REGISTER_OP_CPU_KERNEL( - matmul_grad, ops::MatMulGradKernel); + matmul, ops::MatMulKernel); +REGISTER_OP_CPU_KERNEL( + matmul_grad, + ops::MatMulGradKernel); diff --git a/paddle/operators/matmul_op.cu.cc b/paddle/operators/matmul_op.cu.cc index b7e66382f00445b087e14103e7a148d450b37405..d28d12164e493786c5bdafb157795d797ee87b91 100644 --- a/paddle/operators/matmul_op.cu.cc +++ b/paddle/operators/matmul_op.cu.cc @@ -1,21 +1,22 @@ /* 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 +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 + 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. */ +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/operators/matmul_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(matmul, - ops::MatMulKernel); -REGISTER_OP_GPU_KERNEL( - matmul_grad, ops::MatMulGradKernel); +REGISTER_OP_CUDA_KERNEL( + matmul, ops::MatMulKernel); +REGISTER_OP_CUDA_KERNEL( + matmul_grad, + ops::MatMulGradKernel); diff --git a/paddle/operators/matmul_op.h b/paddle/operators/matmul_op.h index 1e4aa48b7018d8e3d6f02591fbca2877ddbd3c5d..78adc64f76f45afce64c49bcf734647e0db2d6b3 100644 --- a/paddle/operators/matmul_op.h +++ b/paddle/operators/matmul_op.h @@ -1,16 +1,16 @@ -/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserve. +/* 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 +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 + 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. */ +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 @@ -27,7 +27,7 @@ using DDim = framework::DDim; using framework::make_ddim; using framework::vectorize; -template +template class MatMulKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { @@ -38,8 +38,9 @@ class MatMulKernel : public framework::OpKernel { bool transpose_x = context.Attr("transpose_X"); bool transpose_y = context.Attr("transpose_Y"); - math::MatMulFunctor()(context.device_context(), x, transpose_x, y, - transpose_y, T(1), out, T(0)); + math::MatMulFunctor()( + context.template device_context(), x, transpose_x, y, + transpose_y, T(1), out, T(0)); } }; @@ -68,17 +69,16 @@ Tensor CombineBatchAndM(const Tensor& input) { // 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 framework::ExecutionContext& context, - const Tensor& input) { +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.device_context(), input, &output, axis); + 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 { @@ -112,7 +112,7 @@ Tensor CombineBatchAndN(const framework::ExecutionContext& context, // // 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 +template class MatMulGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { @@ -178,24 +178,23 @@ class MatMulGradKernel : public framework::OpKernel { 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(context, dOut) + ? CombineBatchAndN(dev_ctx, dOut) : dOut; if (x_dims.size() == 2 && y_dims.size() == 3) { Y = transpose_y ? CombineBatchAndM(Y) - : CombineBatchAndN(context, Y); + : CombineBatchAndN(dev_ctx, Y); } if (transpose_x) { - math::MatMulFunctor()(context.device_context(), Y, - transpose_y, dOut_for_dX, transpose_x, - T(1), dx, T(0)); + math::MatMulFunctor()( + dev_ctx, Y, transpose_y, dOut_for_dX, transpose_x, T(1), dx, T(0)); } else { - math::MatMulFunctor()(context.device_context(), dOut_for_dX, - transpose_x, Y, !transpose_y, T(1), dx, - T(0)); + math::MatMulFunctor()( + dev_ctx, dOut_for_dX, transpose_x, Y, !transpose_y, T(1), dx, T(0)); } } @@ -205,18 +204,16 @@ class MatMulGradKernel : public framework::OpKernel { ? CombineBatchAndM(dOut) : dOut; if (y_dims.size() == 2 && x_dims.size() == 3) { - X = transpose_x ? CombineBatchAndN(context, X) + X = transpose_x ? CombineBatchAndN(dev_ctx, X) : CombineBatchAndM(X); dOut = CombineBatchAndM(dOut); } if (transpose_y) { - math::MatMulFunctor()(context.device_context(), dOut_for_dY, - transpose_y, X, transpose_x, T(1), dy, - T(0)); + math::MatMulFunctor()( + dev_ctx, dOut_for_dY, transpose_y, X, transpose_x, T(1), dy, T(0)); } else { - math::MatMulFunctor()(context.device_context(), X, - !transpose_x, dOut_for_dY, transpose_y, - T(1), dy, T(0)); + math::MatMulFunctor()( + dev_ctx, X, !transpose_x, dOut_for_dY, transpose_y, T(1), dy, T(0)); } } } diff --git a/paddle/operators/max_sequence_len_op.cc b/paddle/operators/max_sequence_len_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..019150e4914e8bd34a5e8b7d37318aee43942fcc --- /dev/null +++ b/paddle/operators/max_sequence_len_op.cc @@ -0,0 +1,65 @@ +/* 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. */ + +#include "paddle/framework/lod_rank_table.h" +#include "paddle/framework/op_registry.h" +#include "paddle/framework/operator.h" + +namespace paddle { +namespace operators { + +class MaxSeqenceLenOp : public framework::OperatorBase { + public: + MaxSeqenceLenOp(const std::string &type, + const framework::VariableNameMap &inputs, + const framework::VariableNameMap &outputs, + const framework::AttributeMap &attrs) + : OperatorBase(type, inputs, outputs, attrs) {} + + void Run(const framework::Scope &scope, + const platform::Place &dev_place) const override { + auto &rank_table = + scope.FindVar(Input("RankTable"))->Get(); + auto *out = + scope.FindVar(Output("Out"))->GetMutable(); + int64_t *out_ptr = out->mutable_data({1}, platform::CPUPlace()); + *out_ptr = rank_table.items()[0].length; + } +}; + +class MaxSeqenceLenOpProtoMaker : public framework::OpProtoAndCheckerMaker { + public: + MaxSeqenceLenOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("RankTable", "The lod_rank_table."); + AddOutput("Out", "The max sequence length."); + AddComment( + R"DOC(Calculate the max sequence length through lod_rank_table.)DOC"); + } +}; + +class MaxSeqenceLenInferShape : public framework::InferShapeBase { + public: + void operator()(framework::InferShapeContext *context) const override { + PADDLE_ENFORCE(context->HasInput("RankTable")); + context->SetOutputDim("Out", {1}); + } +}; +} // namespace operators +} // namespace paddle + +REGISTER_OPERATOR(max_sequence_len, paddle::operators::MaxSeqenceLenOp, + paddle::operators::MaxSeqenceLenOpProtoMaker, + paddle::operators::MaxSeqenceLenInferShape, + paddle::framework::EmptyGradOpMaker); diff --git a/paddle/operators/maxout_op.cc b/paddle/operators/maxout_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..3ee32269417e80cd14a6ff0f8e52c0b2dec4b8be --- /dev/null +++ b/paddle/operators/maxout_op.cc @@ -0,0 +1,108 @@ +/* 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. */ + +#include "paddle/operators/maxout_op.h" +namespace paddle { +namespace operators { + +using framework::Tensor; + +class MaxOutOpMaker : public framework::OpProtoAndCheckerMaker { + public: + MaxOutOpMaker(OpProto* proto, OpAttrChecker* op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput( + "X", + "(Tensor) The input tensor of maxout operator. " + "The format of input tensor is NCHW. Where N is batch size, C is the " + "number of channels, H and W is the height and width of feature."); + AddOutput("Out", + "(Tensor) The output tensor of maxout operator." + "The format of output tensor is also NCHW." + "Where N is batch size, C is " + "the number of channels, H and W is the height and " + "width of feature."); + AddAttr( + "groups", + R"DOC("Specifies how many groups the input tensor will be split" + "in the channel dimension. And the number of output channel is " + "the number of channels divided by groups.." + )DOC"); + AddComment(R"DOC( +MaxOut Operator. + +Assumed the input shape is (N, Ci, H, W). +The output shape is (N, Co, H, W). +Then $Co = Ci / groups$ and the operator formula is as follows: + +$$ +y_{si+j} = \max_k x_{gsi + sk + j} \\ +g = groups \\ +s = \frac{input.size}{num\_channels} \\ +0 \le i < \frac{num\_channels}{groups} \\ +0 \le j < s \\ +0 \le k < groups +$$ + +Please refer to Paper: + - Maxout Networks: http://www.jmlr.org/proceedings/papers/v28/goodfellow13.pdf + - Multi-digit Number Recognition from Street View \ + Imagery using Deep Convolutional Neural Networks: \ + https://arxiv.org/pdf/1312.6082v4.pdf + +)DOC"); + } +}; + +class MaxOutOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of MaxoutOp" + "should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of MaxoutOp should not be null."); + auto in_x_dims = ctx->GetInputDim("X"); + int groups = ctx->Attrs().Get("groups"); + // check groups > 1 + PADDLE_ENFORCE_GT(groups, 1, "groups should be larger than 1 in maxoutop"); + std::vector output_shape({in_x_dims[0], in_x_dims[1] / groups}); + output_shape.push_back(in_x_dims[2]); + output_shape.push_back(in_x_dims[3]); + ctx->SetOutputDim("Out", framework::make_ddim(output_shape)); + } +}; + +class MaxOutOpGrad : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) must not be null."); + PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("X")), + "Input(X@GRAD) should not be null."); + ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X")); + } +}; +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP(maxout, ops::MaxOutOp, ops::MaxOutOpMaker, maxout_grad, + ops::MaxOutOpGrad); +REGISTER_OP_CPU_KERNEL( + maxout, ops::MaxOutKernel); +REGISTER_OP_CPU_KERNEL( + maxout_grad, + ops::MaxOutGradKernel); diff --git a/paddle/operators/maxout_op.cu.cc b/paddle/operators/maxout_op.cu.cc new file mode 100644 index 0000000000000000000000000000000000000000..c4a2d676d3aca4d59d0bfa8c75aa0c249e202ab5 --- /dev/null +++ b/paddle/operators/maxout_op.cu.cc @@ -0,0 +1,24 @@ +/* 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. */ + +#include "paddle/operators/maxout_op.h" + +namespace ops = paddle::operators; +REGISTER_OP_CUDA_KERNEL( + maxout, ops::MaxOutKernel, + ops::MaxOutKernel); +REGISTER_OP_CUDA_KERNEL( + maxout_grad, + ops::MaxOutGradKernel, + ops::MaxOutGradKernel); diff --git a/paddle/operators/maxout_op.h b/paddle/operators/maxout_op.h new file mode 100644 index 0000000000000000000000000000000000000000..e8b12552b9ff39e23702de17abc9825a527f02aa --- /dev/null +++ b/paddle/operators/maxout_op.h @@ -0,0 +1,62 @@ +/* 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 + +#include "paddle/framework/op_registry.h" +#include "paddle/operators/math/math_function.h" +#include "paddle/operators/math/maxouting.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; + +template +class MaxOutKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + const Tensor* in_x = context.Input("X"); + Tensor* out = context.Output("Out"); + int groups = context.template Attr("groups"); + + math::MaxOutFunctor maxout_forward; + maxout_forward(context.template device_context(), *in_x, out, + groups); + } +}; + +template +class MaxOutGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + const Tensor* in_x = context.Input("X"); + const Tensor* out = context.Input("Out"); + const Tensor* out_grad = + context.Input(framework::GradVarName("Out")); + Tensor* in_x_grad = context.Output(framework::GradVarName("X")); + int groups = context.template Attr("groups"); + auto& device_ctx = context.template device_context(); + math::SetConstant zero; + if (in_x_grad) { + in_x_grad->mutable_data(context.GetPlace()); + zero(device_ctx, in_x_grad, static_cast(0.0)); + math::MaxOutGradFunctor maxout_backward; + maxout_backward(device_ctx, *in_x, in_x_grad, *out, *out_grad, groups); + } + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/mean_op.cc b/paddle/operators/mean_op.cc index dcc5b4286f4ac833268a779a9a7edd2ed119ffff..411f4d14efbfa5a8ee6dd7da645a044b191bf006 100644 --- a/paddle/operators/mean_op.cc +++ b/paddle/operators/mean_op.cc @@ -32,7 +32,7 @@ class MeanOp : public framework::OperatorWithKernel { class MeanOpMaker : public framework::OpProtoAndCheckerMaker { public: - MeanOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker) + MeanOpMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "The input of mean op"); AddOutput("Out", "The output of mean op"); @@ -60,13 +60,13 @@ class MeanGradMaker : public framework::SingleGradOpDescMaker { using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; protected: - std::unique_ptr Apply() const override { - auto* grad_op = new framework::OpDescBind(); + std::unique_ptr Apply() const override { + auto* grad_op = new framework::OpDesc(); grad_op->SetType("mean_grad"); grad_op->SetInput("X", Input("X")); grad_op->SetInput(framework::GradVarName("Out"), OutputGrad("Out")); grad_op->SetOutput(framework::GradVarName("X"), InputGrad("X")); - return std::unique_ptr(grad_op); + return std::unique_ptr(grad_op); } }; @@ -76,8 +76,9 @@ class MeanGradMaker : public framework::SingleGradOpDescMaker { namespace ops = paddle::operators; REGISTER_OPERATOR(mean, ops::MeanOp, ops::MeanOpMaker, ops::MeanGradMaker); REGISTER_OPERATOR(mean_grad, ops::MeanGradOp); -REGISTER_OP_CPU_KERNEL(mean, ops::MeanKernel, - ops::MeanKernel); -REGISTER_OP_CPU_KERNEL(mean_grad, - ops::MeanGradKernel, - ops::MeanGradKernel); +REGISTER_OP_CPU_KERNEL( + mean, ops::MeanKernel, + ops::MeanKernel); +REGISTER_OP_CPU_KERNEL( + mean_grad, ops::MeanGradKernel, + ops::MeanGradKernel); diff --git a/paddle/operators/mean_op.cu b/paddle/operators/mean_op.cu index ca089938c048f7aa5bd561f57c093aa74cce4e11..212d4481138c1478f6e3aa684008f9e42c5a3870 100644 --- a/paddle/operators/mean_op.cu +++ b/paddle/operators/mean_op.cu @@ -1,24 +1,25 @@ /* 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 +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 + 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. */ +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. */ #define EIGEN_USE_GPU #include "paddle/operators/mean_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(mean, ops::MeanKernel, - ops::MeanKernel); -REGISTER_OP_GPU_KERNEL(mean_grad, - ops::MeanGradKernel, - ops::MeanGradKernel); +REGISTER_OP_CUDA_KERNEL( + mean, ops::MeanKernel, + ops::MeanKernel); +REGISTER_OP_CUDA_KERNEL( + mean_grad, ops::MeanGradKernel, + ops::MeanGradKernel); diff --git a/paddle/operators/mean_op.h b/paddle/operators/mean_op.h index c99286a5b928f1edcd845b01b21b95654c25db07..351b34595974b1771d9f4ae5232e0b3a33491104 100644 --- a/paddle/operators/mean_op.h +++ b/paddle/operators/mean_op.h @@ -27,7 +27,7 @@ template using EigenVector = framework::EigenVector; -template +template class MeanKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { @@ -38,13 +38,14 @@ class MeanKernel : public framework::OpKernel { auto X = EigenVector::Flatten(*input); auto y = EigenScalar::From(*output); - auto& place = context.GetEigenDevice(); + auto& place = + *context.template device_context().eigen_device(); y.device(place) = X.mean(); } }; -template +template class MeanGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { @@ -56,7 +57,8 @@ class MeanGradKernel : public framework::OpKernel { T ig_size = static_cast(IG->numel()); Eigen::DSizes bcast(ig_size); - EigenVector::Flatten(*IG).device(context.GetEigenDevice()) = + EigenVector::Flatten(*IG).device( + *context.template device_context().eigen_device()) = (EigenVector::From(*OG) / ig_size).broadcast(bcast); } }; diff --git a/paddle/operators/merge_lod_tensor_op.cc b/paddle/operators/merge_lod_tensor_op.cc index 80460c476921b63ec5228a9780880c7db3c85217..3f999e404f8afe6bded09c820509fa0f36d30bf6 100644 --- a/paddle/operators/merge_lod_tensor_op.cc +++ b/paddle/operators/merge_lod_tensor_op.cc @@ -28,7 +28,11 @@ class MergeLoDTensorOp : public framework::OperatorBase { const framework::AttributeMap &attrs) : OperatorBase(type, inputs, outputs, attrs) {} void Run(const framework::Scope &scope, - const platform::DeviceContext &dev_ctx) const override { + const platform::Place &dev_place) const override { + // get device context from pool + platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance(); + auto &dev_ctx = *pool.Get(dev_place); + auto &x = scope.FindVar(Input("X"))->Get(); auto &mask = scope.FindVar(Input("Mask"))->Get(); auto &in_true = scope.FindVar(Input("InTrue"))->Get(); @@ -45,7 +49,7 @@ class MergeLoDTensorOp : public framework::OperatorBase { cpu_mask->ShareDataWith(mask); } else if (platform::is_gpu_place(mask.place())) { #ifdef PADDLE_WITH_CUDA - cpu_mask->CopyFrom(mask, platform::CPUPlace(), dev_ctx); + framework::CopyFrom(mask, platform::CPUPlace(), dev_ctx, cpu_mask.get()); #else PADDLE_THROW("Not supported GPU, Please compile WITH_GPU option"); #endif @@ -99,8 +103,9 @@ class MergeLoDTensorOp : public framework::OperatorBase { if (len == 0) { continue; } - out->Slice(out_offset, out_offset + len) - .CopyFrom(input->Slice(start_offset, end_offset), place, dev_ctx); + auto slice = out->Slice(out_offset, out_offset + len); + framework::CopyFrom(input->Slice(start_offset, end_offset), place, + dev_ctx, &slice); out_offset += len; (*in_idx) += 1; } @@ -113,8 +118,7 @@ class MergeLoDTensorOp : public framework::OperatorBase { class MergeLoDTensorOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - MergeLoDTensorOpProtoMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + MergeLoDTensorOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "The input LoDTensor, contains complete lod information to " @@ -161,15 +165,15 @@ class MergeLoDTensorGradMaker : public framework::SingleGradOpDescMaker { using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; protected: - std::unique_ptr Apply() const override { - auto *grad_op = new framework::OpDescBind(); + std::unique_ptr Apply() const override { + auto *grad_op = new framework::OpDesc(); grad_op->SetType("split_lod_tensor"); grad_op->SetInput("X", OutputGrad("Out")); grad_op->SetInput("Mask", Input("Mask")); grad_op->SetOutput("OutTrue", InputGrad("InTrue")); grad_op->SetOutput("OutFalse", InputGrad("InFalse")); grad_op->SetAttrMap(Attrs()); - return std::unique_ptr(grad_op); + return std::unique_ptr(grad_op); } }; diff --git a/paddle/operators/minus_op.cc b/paddle/operators/minus_op.cc index 4684c20208501a3239fd57b35428946bb52af4a0..3d7742dd4bc2a3c727279bc1e6c7dd47b96eefa3 100644 --- a/paddle/operators/minus_op.cc +++ b/paddle/operators/minus_op.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/operators/minus_op.h" #include "paddle/operators/net_op.h" @@ -46,7 +46,7 @@ class MinusOp : public framework::OperatorWithKernel { class MinusOpMaker : public framework::OpProtoAndCheckerMaker { public: - MinusOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + MinusOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "The left tensor of minus operator."); AddInput("Y", "The right tensor of minus operator."); @@ -70,12 +70,11 @@ class MinusGradMaker : public framework::GradOpDescMakerBase { public: using framework::GradOpDescMakerBase::GradOpDescMakerBase; - std::vector> operator()() - const override { - std::vector> ops; + std::vector> operator()() const override { + std::vector> ops; auto x_g = InputGrad("X"); if (!x_g.empty()) { - auto *x_g_op = new framework::OpDescBind(); + auto *x_g_op = new framework::OpDesc(); x_g_op->SetType("scale"); x_g_op->SetInput("X", OutputGrad("Out")); x_g_op->SetOutput("Out", x_g); @@ -85,7 +84,7 @@ class MinusGradMaker : public framework::GradOpDescMakerBase { auto y_g = InputGrad("Y"); if (!y_g.empty()) { - auto *y_g_op = new framework::OpDescBind(); + auto *y_g_op = new framework::OpDesc(); y_g_op->SetType("scale"); y_g_op->SetInput("X", OutputGrad("Out")); y_g_op->SetOutput("Out", y_g); @@ -102,5 +101,5 @@ class MinusGradMaker : public framework::GradOpDescMakerBase { namespace ops = paddle::operators; REGISTER_OPERATOR(minus, ops::MinusOp, ops::MinusOpMaker, ops::MinusGradMaker); -REGISTER_OP_CPU_KERNEL(minus, - ops::MinusKernel); +REGISTER_OP_CPU_KERNEL( + minus, ops::MinusKernel); diff --git a/paddle/operators/minus_op.cu b/paddle/operators/minus_op.cu index a8375cc6301b2c1a917299c3933b03226bb72907..80cd9f7c16845904b7b46ae1597ce9558c32f46a 100644 --- a/paddle/operators/minus_op.cu +++ b/paddle/operators/minus_op.cu @@ -1,18 +1,19 @@ /* 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 +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 + 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. */ +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/operators/minus_op.h" -REGISTER_OP_GPU_KERNEL( - minus, paddle::operators::MinusKernel); +REGISTER_OP_CUDA_KERNEL( + minus, + paddle::operators::MinusKernel); diff --git a/paddle/operators/minus_op.h b/paddle/operators/minus_op.h index bd9a2790aa2b208c2d3dfc792031283eb6c42397..20760b8cd5bd2f74ed8469addda8f67f11f4545c 100644 --- a/paddle/operators/minus_op.h +++ b/paddle/operators/minus_op.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/framework/eigen.h" @@ -19,7 +19,7 @@ namespace paddle { namespace operators { -template +template class MinusKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { @@ -28,7 +28,8 @@ class MinusKernel : public framework::OpKernel { auto* out_tensor = context.Output("Out"); out_tensor->mutable_data(context.GetPlace()); - auto& dev = context.GetEigenDevice(); + auto& dev = + *context.template device_context().eigen_device(); framework::EigenVector::Flatten(*out_tensor).device(dev) = framework::EigenVector::Flatten(*left_tensor) - framework::EigenVector::Flatten(*right_tensor); diff --git a/paddle/operators/modified_huber_loss_op.cc b/paddle/operators/modified_huber_loss_op.cc index 28528848af1f467bf38be53f9d05fee6ca3f93cc..f5d69071a86e3f8037840c091cf5b7683e4eeb96 100644 --- a/paddle/operators/modified_huber_loss_op.cc +++ b/paddle/operators/modified_huber_loss_op.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/operators/modified_huber_loss_op.h" @@ -39,8 +39,7 @@ class ModifiedHuberLossOp : public framework::OperatorWithKernel { class ModifiedHuberLossOpMaker : public framework::OpProtoAndCheckerMaker { public: - ModifiedHuberLossOpMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker) + ModifiedHuberLossOpMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "The input tensor of modified huber loss op. " @@ -115,6 +114,6 @@ REGISTER_OP(modified_huber_loss, ops::ModifiedHuberLossOp, REGISTER_OP_CPU_KERNEL( modified_huber_loss, - ops::ModifiedHuberLossKernel); + ops::ModifiedHuberLossKernel); REGISTER_OP_CPU_KERNEL(modified_huber_loss_grad, ops::ModifiedHuberLossGradCPUKernel); diff --git a/paddle/operators/modified_huber_loss_op.cu b/paddle/operators/modified_huber_loss_op.cu index 8854e166cd99ce914d7f9f9bcead3234b0649506..3d2a5562e8cc2117b0b460496d9ba8e96823fbfb 100644 --- a/paddle/operators/modified_huber_loss_op.cu +++ b/paddle/operators/modified_huber_loss_op.cu @@ -1,16 +1,16 @@ /* 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. */ -#include -#include +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 "paddle/framework/op_registry.h" @@ -71,8 +71,8 @@ class ModifiedHuberLossGradGPUKernel : public framework::OpKernel { } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL( +REGISTER_OP_CUDA_KERNEL( modified_huber_loss, - ops::ModifiedHuberLossKernel); -REGISTER_OP_GPU_KERNEL(modified_huber_loss_grad, - ops::ModifiedHuberLossGradGPUKernel); + ops::ModifiedHuberLossKernel); +REGISTER_OP_CUDA_KERNEL(modified_huber_loss_grad, + ops::ModifiedHuberLossGradGPUKernel); diff --git a/paddle/operators/modified_huber_loss_op.h b/paddle/operators/modified_huber_loss_op.h index aba75efad9c19e3e113b4f09bc1fbd4732f4e187..6ce86feee574efca8811f316f47f1c3fbbdd0bf9 100644 --- a/paddle/operators/modified_huber_loss_op.h +++ b/paddle/operators/modified_huber_loss_op.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 @@ -46,7 +46,7 @@ struct ModifiedHuberLossForward { } }; -template +template class ModifiedHuberLossKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { @@ -57,7 +57,8 @@ class ModifiedHuberLossKernel : public framework::OpKernel { out0->mutable_data(context.GetPlace()); out1->mutable_data(context.GetPlace()); - auto place = context.GetEigenDevice(); + auto& place = + *context.template device_context().eigen_device(); auto x = EigenVector::Flatten(*in0); auto y = EigenVector::Flatten(*in1); diff --git a/paddle/operators/momentum_op.cc b/paddle/operators/momentum_op.cc index 19954006195c1e9fd34328b52ed2a9eade526235..15b8b80776732f43c3ef4f8b80cffedf5c2a76fd 100644 --- a/paddle/operators/momentum_op.cc +++ b/paddle/operators/momentum_op.cc @@ -54,8 +54,7 @@ class MomentumOp : public framework::OperatorWithKernel { class MomentumOpMaker : public framework::OpProtoAndCheckerMaker { public: - MomentumOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + MomentumOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("Param", "(Tensor, default Tensor) " @@ -71,8 +70,12 @@ class MomentumOpMaker : public framework::OpProtoAndCheckerMaker { "(Tensor, default Tensor) " "Input learning rate"); - AddOutput("ParamOut", "(Tensor) Output updated parameter"); - AddOutput("VelocityOut", "(Tensor) Output updated velocity"); + AddOutput("ParamOut", + "(Tensor) This output is updated parameter. " + "It shared memory with Input(Param)."); + AddOutput("VelocityOut", + "(Tensor) This output is updated velocity. " + "It shared memory with Input(Velocity)."); AddAttr("mu", "(float) Momentum coefficient"); AddAttr("use_nesterov", @@ -101,5 +104,5 @@ $$ namespace ops = paddle::operators; REGISTER_OP_WITHOUT_GRADIENT(momentum, ops::MomentumOp, ops::MomentumOpMaker); -REGISTER_OP_CPU_KERNEL( - momentum, ops::MomentumOpKernel); +REGISTER_OP_CPU_KERNEL(momentum, ops::MomentumOpKernel, + ops::MomentumOpKernel); diff --git a/paddle/operators/momentum_op.cu b/paddle/operators/momentum_op.cu index efc24e795e05951024009f0b3258769c352df344..2b9314162e6f10b3791c913203c732d2822861ab 100644 --- a/paddle/operators/momentum_op.cu +++ b/paddle/operators/momentum_op.cu @@ -1,20 +1,78 @@ /* 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 +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 + 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. */ +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. */ -#define EIGEN_USE_GPU -#include "paddle/operators/momentum_op.h" +#include "paddle/framework/op_registry.h" + +namespace paddle { +namespace operators { + +template +__global__ void MomentumKernel(const T* p, const T* g, const T* v, + const T* learning_rate, const T mu, + const int64_t num, bool use_nesterov, T* p_out, + T* v_out) { + T lr = learning_rate[0]; + if (use_nesterov) { + for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < num; + i += blockDim.x * gridDim.x) { + T g_val = g[i]; + T v_new = v[i] * mu + g_val; + v_out[i] = v_new; + p_out[i] = p[i] - (g_val - v_new * mu) * lr; + } + } else { + for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < num; + i += blockDim.x * gridDim.x) { + T v_new = v[i] * mu + g[i]; + v_out[i] = v_new; + p_out[i] = p[i] - lr * v_new; + } + } +} + +template +class MomentumOpCUDAKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto param_out = ctx.Output("ParamOut"); + auto velocity_out = ctx.Output("VelocityOut"); + auto param = ctx.Input("Param"); + auto velocity = ctx.Input("Velocity"); + auto grad = ctx.Input("Grad"); + auto learning_rate = ctx.Input("LearningRate"); + + T* p_out = param_out->mutable_data(ctx.GetPlace()); + T* v_out = velocity_out->mutable_data(ctx.GetPlace()); + + T mu = static_cast(ctx.Attr("mu")); + bool use_nesterov = ctx.Attr("use_nesterov"); + + auto* p = param->data(); + auto* v = velocity->data(); + auto* g = grad->data(); + auto* lr = learning_rate->data(); + + int block = 512; + int grid = (param->numel() + block - 1) / block; + MomentumKernel<<>>( + p, g, v, lr, mu, param->numel(), use_nesterov, p_out, v_out); + } +}; + +} // namespace operators +} // namespace paddle namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL( - momentum, ops::MomentumOpKernel); +REGISTER_OP_CUDA_KERNEL(momentum, ops::MomentumOpCUDAKernel, + ops::MomentumOpCUDAKernel); diff --git a/paddle/operators/momentum_op.h b/paddle/operators/momentum_op.h index 8f7f5eb5c21c0342f57a47b85d28f4454f4566c2..da69532ea58bad8d3908770d82dbcc6e6b108fce 100644 --- a/paddle/operators/momentum_op.h +++ b/paddle/operators/momentum_op.h @@ -19,7 +19,7 @@ limitations under the License. */ namespace paddle { namespace operators { -template +template class MomentumOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { @@ -33,7 +33,7 @@ class MomentumOpKernel : public framework::OpKernel { param_out->mutable_data(ctx.GetPlace()); velocity_out->mutable_data(ctx.GetPlace()); - float mu = ctx.Attr("mu"); + T mu = static_cast(ctx.Attr("mu")); bool use_nesterov = ctx.Attr("use_nesterov"); auto p_out = framework::EigenVector::Flatten(*param_out); @@ -42,18 +42,13 @@ class MomentumOpKernel : public framework::OpKernel { auto p = framework::EigenVector::Flatten(*param); auto v = framework::EigenVector::Flatten(*velocity); auto g = framework::EigenVector::Flatten(*grad); - auto lr = framework::EigenVector::Flatten(*learning_rate); + auto* lr = learning_rate->data(); - auto place = ctx.GetEigenDevice(); - - Eigen::DSizes grad_dsize(grad->numel()); - - v_out.device(place) = v * mu + g; + v_out = v * mu + g; if (use_nesterov) { - p_out.device(place) = p - g * lr.broadcast(grad_dsize) + - v_out * mu * lr.broadcast(grad_dsize); + p_out = p - (g - v_out * mu) * lr[0]; } else { - p_out.device(place) = p - lr.broadcast(grad_dsize) * v_out; + p_out = p - lr[0] * v_out; } } }; diff --git a/paddle/operators/mul_op.cc b/paddle/operators/mul_op.cc index 3c39ae10dc50084cff284c307167c33c9208a3ce..c923e988a55b43ebb7ba6256e7b72a85c124f360 100644 --- a/paddle/operators/mul_op.cc +++ b/paddle/operators/mul_op.cc @@ -71,41 +71,52 @@ class MulOpShapeInference : public framework::InferShapeBase { class MulOpMaker : public framework::OpProtoAndCheckerMaker { public: - MulOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker) + MulOpMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("X", "The first input of mul op"); - AddInput("Y", "The second input of mul op"); - AddOutput("Out", "The output of mul op"); + 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."); AddAttr( "x_num_col_dims", - "(int, default 1) " - R"DOC(mul_op can take tensors with more than two dimensions as input `X`, - in that case, tensors will be reshaped to a matrix. The matrix's first - dimension(column length) will be the product of tensor's last - `num_col_dims` dimensions, and the matrix's second dimension(row length) - will be the product of tensor's first `rank - num_col_dims` dimensions. + R"DOC((int, default 1), The mul_op can take tensors with more than two + dimensions as its inputs. If the input $X$ is a tensor with more + than two dimensions, $X$ will be flattened into a two-dimensional + matrix first. The flattening rule is: the first `num_col_dims` + will be flattened to form the first dimension of the final matrix + (the height of the matrix), and the rest `rank(X) - num_col_dims` + dimensions are flattened to form the second dimension of the final + matrix (the width of the matrix). As a result, height of the + flattened matrix is equal to the product of $X$'s first + `x_num_col_dims` dimensions' sizes, and width of the flattened + matrix is equal to the product of $X$'s last `rank(x) - num_col_dims` + dimensions' size. For example, suppose $X$ is a 6-dimensional + tensor with the shape [2, 3, 4, 5, 6], and `x_num_col_dims` = 3. + Thus, the flattened matrix will have a shape [2 x 3 x 4, 5 x 6] = + [24, 30]. )DOC") .SetDefault(1) .EqualGreaterThan(1); AddAttr( "y_num_col_dims", - "(int, default 1) " - R"DOC(mul_op can take tensors with more than two dimensions as input `Y`, - in that case, tensors will be reshaped to a matrix. Just like input `X`. + R"DOC((int, default 1), The mul_op can take tensors with more than two, + dimensions as its inputs. If the input $Y$ is a tensor with more + than two dimensions, $Y$ will be flattened into a two-dimensional + matrix first. The attribute `y_num_col_dims` determines how $Y$ is + flattened. See comments of `x_num_col_dims` for more details. )DOC") .SetDefault(1) .EqualGreaterThan(1); AddComment(R"DOC( -Mul Operator. +Mul Operator. -This operator is used to perform matrix multiplication for input X and Y. +This operator is used to perform matrix multiplication for input $X$ and $Y$. The equation is: - $$Out = X * Y$$ +$$Out = X * Y$$ -Both the input `X` and `Y` can carry the LoD (Level of Details) information, -or not. But the output only shares the LoD information with input `X`. +Both the input $X$ and $Y$ can carry the LoD (Level of Details) information, +or not. But the output only shares the LoD information with input $X$. )DOC"); } @@ -149,6 +160,7 @@ REGISTER_OPERATOR(mul, paddle::framework::OperatorWithKernel, ops::MulOpMaker, ops::MulOpShapeInference, paddle::framework::DefaultGradOpDescMaker); REGISTER_OPERATOR(mul_grad, ops::MulOpGrad); -REGISTER_OP_CPU_KERNEL(mul, ops::MulKernel); -REGISTER_OP_CPU_KERNEL(mul_grad, - ops::MulGradKernel); +REGISTER_OP_CPU_KERNEL( + mul, ops::MulKernel); +REGISTER_OP_CPU_KERNEL( + mul_grad, ops::MulGradKernel); diff --git a/paddle/operators/mul_op.cu.cc b/paddle/operators/mul_op.cu.cc index 66dc3d6d106a18640adad413d4e967fa101abcfc..43de9a719499e4e0e8fd2e5fcc6771d717ce6522 100644 --- a/paddle/operators/mul_op.cu.cc +++ b/paddle/operators/mul_op.cu.cc @@ -1,20 +1,21 @@ /* 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 +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 + 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. */ +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/operators/mul_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(mul, ops::MulKernel); -REGISTER_OP_GPU_KERNEL(mul_grad, - ops::MulGradKernel); +REGISTER_OP_CUDA_KERNEL( + mul, ops::MulKernel); +REGISTER_OP_CUDA_KERNEL( + mul_grad, ops::MulGradKernel); diff --git a/paddle/operators/mul_op.h b/paddle/operators/mul_op.h index 0eb9df41e9415845f88af283de63856158b447f9..1fb0569b49cce80c3f1e408fb57b5f5cf7033a27 100644 --- a/paddle/operators/mul_op.h +++ b/paddle/operators/mul_op.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 @@ -23,7 +23,7 @@ namespace operators { using Tensor = framework::Tensor; -template +template class MulKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { @@ -46,15 +46,16 @@ class MulKernel : public framework::OpKernel { if (z_dim.size() != 2) { z->Resize({x_matrix.dims()[0], y_matrix.dims()[1]}); } - math::matmul(context.device_context(), x_matrix, false, y_matrix, - false, 1, z, 0); + math::matmul( + context.template device_context(), x_matrix, false, + y_matrix, false, 1, z, 0); if (z_dim.size() != 2) { z->Resize(z_dim); } } }; -template +template class MulGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { @@ -77,6 +78,7 @@ 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(); if (dx) { dx->mutable_data(ctx.GetPlace()); Tensor dx_matrix = dx->dims().size() > 2 @@ -84,8 +86,8 @@ class MulGradKernel : public framework::OpKernel { : *dx; // dx = dout * y'. dx: M x K, dout : M x N, y : K x N - math::matmul(ctx.device_context(), dout_mat, false, y_matrix, - true, 1, &dx_matrix, 0); + math::matmul(dev_ctx, dout_mat, false, y_matrix, true, + 1, &dx_matrix, 0); } if (dy) { dy->mutable_data(ctx.GetPlace()); @@ -93,8 +95,8 @@ 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(ctx.device_context(), x_matrix, true, dout_mat, - false, 1, &dy_matrix, 0); + math::matmul(dev_ctx, x_matrix, true, dout_mat, false, + 1, &dy_matrix, 0); } } }; diff --git a/paddle/operators/multiplex_op.cc b/paddle/operators/multiplex_op.cc index f8527dfab3f3c42f430c433a11351f12b8dfae8b..11e047b5d57b6bc18e6d6f4a1d122e18dfc6e357 100644 --- a/paddle/operators/multiplex_op.cc +++ b/paddle/operators/multiplex_op.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/operators/multiplex_op.h" @@ -51,7 +51,7 @@ class MultiplexOp : public framework::OperatorWithKernel { } protected: - framework::OpKernelType GetKernelType( + framework::OpKernelType GetActualKernelType( const framework::ExecutionContext& ctx) const override { return framework::OpKernelType( framework::ToDataType(ctx.MultiInput("X")[0]->type()), @@ -61,8 +61,7 @@ class MultiplexOp : public framework::OperatorWithKernel { class MultiplexOpMaker : public framework::OpProtoAndCheckerMaker { public: - MultiplexOpMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker) + MultiplexOpMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("Ids", "The index tensor of multiplex operator."); AddInput("X", "The candidate tensors of multiplex operator.") @@ -99,17 +98,11 @@ class MultiplexGradOp : public framework::OperatorWithKernel { "Output(X@Grad) should not be null."); PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), "Input(Out@GRAD) should not be null."); - std::vector d_ins; - auto ins = ctx->GetInputsDim("X"); - // No need to compute gradient for Input(Ids) - for (size_t i = 0; i < ins.size(); i++) { - d_ins.push_back(ins[i]); - } - ctx->SetOutputsDim(framework::GradVarName("X"), d_ins); + ctx->SetOutputsDim(framework::GradVarName("X"), ctx->GetInputsDim("X")); } protected: - framework::OpKernelType GetKernelType( + framework::OpKernelType GetActualKernelType( const framework::ExecutionContext& ctx) const override { return framework::OpKernelType( framework::ToDataType(ctx.MultiInput("X")[0]->type()), @@ -125,7 +118,8 @@ REGISTER_OPERATOR(multiplex, ops::MultiplexOp, ops::MultiplexOpMaker, paddle::framework::DefaultGradOpDescMaker); REGISTER_OPERATOR(multiplex_grad, ops::MultiplexGradOp); REGISTER_OP_CPU_KERNEL( - multiplex, ops::MultiplexCPUKernel); + multiplex, + ops::MultiplexCPUKernel); REGISTER_OP_CPU_KERNEL( multiplex_grad, - ops::MultiplexGradCPUKernel); + ops::MultiplexGradCPUKernel); diff --git a/paddle/operators/multiplex_op.cu b/paddle/operators/multiplex_op.cu index 49ed8a8879527fd32dd8b001ea256e46a0353487..f49ee71f104b72f5c8ea5fb1d49999528c21832e 100644 --- a/paddle/operators/multiplex_op.cu +++ b/paddle/operators/multiplex_op.cu @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/framework/op_registry.h" #include "paddle/operators/multiplex_op.h" @@ -33,10 +33,10 @@ class MultiplexGPUKernel : public framework::OpKernel { auto cols = ins[0]->numel() / rows; // copy index to cpu Tensor index_t_cpu; - index_t_cpu.CopyFrom(*ids, platform::CPUPlace(), ctx.device_context()); + CopyFrom(*ids, platform::CPUPlace(), ctx.device_context(), &index_t_cpu); auto* index = index_t_cpu.data(); auto stream = ctx.cuda_device_context().stream(); - Place place = boost::get(ctx.GetPlace()); + platform::CUDAPlace place = boost::get(ctx.GetPlace()); for (auto i = 0; i < rows; i++) { int32_t k = index[i]; PADDLE_ENFORCE_GE(k, 0, "index must be nonnegative."); @@ -60,7 +60,8 @@ class MultiplexGradGPUKernel : public framework::OpKernel { if (d_ins[i]) { d_ins[i]->mutable_data(ctx.GetPlace()); auto t = framework::EigenVector::Flatten(*d_ins[i]); - t.device(ctx.GetEigenDevice()) = t.constant(static_cast(0)); + t.device(*ctx.template device_context().eigen_device()) = + t.constant(static_cast(0)); } } @@ -68,11 +69,11 @@ class MultiplexGradGPUKernel : public framework::OpKernel { auto cols = ins[0]->numel() / rows; // copy index to cpu Tensor index_t_cpu; - index_t_cpu.CopyFrom(*ids, platform::CPUPlace(), ctx.device_context()); + CopyFrom(*ids, platform::CPUPlace(), ctx.device_context(), &index_t_cpu); auto* index = index_t_cpu.data(); auto stream = ctx.cuda_device_context().stream(); - Place place = boost::get(ctx.GetPlace()); + platform::CUDAPlace place = boost::get(ctx.GetPlace()); for (auto i = 0; i < rows; i++) { size_t k = static_cast(index[i]); if (d_ins[k]) { @@ -87,8 +88,9 @@ class MultiplexGradGPUKernel : public framework::OpKernel { namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL( - multiplex, ops::MultiplexGPUKernel); -REGISTER_OP_GPU_KERNEL( +REGISTER_OP_CUDA_KERNEL( + multiplex, + ops::MultiplexGPUKernel); +REGISTER_OP_CUDA_KERNEL( multiplex_grad, - ops::MultiplexGradGPUKernel); + ops::MultiplexGradGPUKernel); diff --git a/paddle/operators/multiplex_op.h b/paddle/operators/multiplex_op.h index ab3cafaa324a29d6f249cf1f73db92e1364eebc8..ef66be5556ee613a037de13286ecc66b53885c1f 100644 --- a/paddle/operators/multiplex_op.h +++ b/paddle/operators/multiplex_op.h @@ -1,17 +1,16 @@ - /* 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 +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 + 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. */ +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 @@ -22,7 +21,7 @@ namespace paddle { namespace operators { -template +template class MultiplexCPUKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const { @@ -35,7 +34,7 @@ class MultiplexCPUKernel : public framework::OpKernel { auto rows = ins[0]->dims()[0]; auto cols = ins[0]->numel() / rows; auto index = ids->data(); - Place place = boost::get(ctx.GetPlace()); + platform::CPUPlace place = boost::get(ctx.GetPlace()); for (auto i = 0; i < rows; i++) { int32_t k = index[i]; PADDLE_ENFORCE_GE(k, 0, "index must be nonnegative."); @@ -47,7 +46,7 @@ class MultiplexCPUKernel : public framework::OpKernel { } }; -template +template class MultiplexGradCPUKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const { @@ -60,14 +59,15 @@ class MultiplexGradCPUKernel : public framework::OpKernel { if (d_ins[i]) { d_ins[i]->mutable_data(ctx.GetPlace()); auto t = framework::EigenVector::Flatten(*d_ins[i]); - t.device(ctx.GetEigenDevice()) = t.constant(static_cast(0)); + t.device(*ctx.template device_context().eigen_device()) = + t.constant(static_cast(0)); } } auto rows = ins[0]->dims()[0]; auto cols = ins[0]->numel() / rows; auto* index = ids->data(); - Place place = boost::get(ctx.GetPlace()); + platform::CPUPlace place = boost::get(ctx.GetPlace()); for (auto i = 0; i < rows; i++) { size_t k = static_cast(index[i]); if (d_ins[k]) { diff --git a/paddle/operators/nccl/nccl_gpu_common.cc b/paddle/operators/nccl/nccl_gpu_common.cc index 6be735e4c731f79684e0bdac3d69a30b328fed84..1602a3d9b54dd64813770a7162f8d4f3dd0e791a 100644 --- a/paddle/operators/nccl/nccl_gpu_common.cc +++ b/paddle/operators/nccl/nccl_gpu_common.cc @@ -1,13 +1,16 @@ /* 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. */ + +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/operators/nccl/nccl_gpu_common.h" #include "paddle/platform/gpu_info.h" diff --git a/paddle/operators/nccl/nccl_gpu_common.h b/paddle/operators/nccl/nccl_gpu_common.h index 48e322f99398a7f1d6af9cab653d0cc92d981fe0..5173996f2020ec7a94643277e8c7a532d41d9045 100644 --- a/paddle/operators/nccl/nccl_gpu_common.h +++ b/paddle/operators/nccl/nccl_gpu_common.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 diff --git a/paddle/operators/nccl_op.cc b/paddle/operators/nccl_op.cc index 66fcc09bc877867e66a37adc73230d8dabf4cbed..9d51153b0631b988c9297f395672be67e18ee3f9 100644 --- a/paddle/operators/nccl_op.cc +++ b/paddle/operators/nccl_op.cc @@ -1,13 +1,16 @@ /* 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. */ + +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/framework/op_registry.h" #include "paddle/operators/nccl/nccl_gpu_common.h" @@ -24,7 +27,7 @@ class NCCLInitOp : public framework::OperatorBase { : OperatorBase(type, inputs, outputs, attrs) {} void Run(const framework::Scope &scope, - const platform::DeviceContext &dev_ctx) const override { + const platform::Place &place) const override { const auto &name = Output("Communicator"); PADDLE_ENFORCE_NOT_NULL(scope.FindVar(name), "Can not find variable '%s' in the scope.", name); @@ -43,16 +46,15 @@ class NCCLInitOp : public framework::OperatorBase { class NCCLInitOpMaker : public framework::OpProtoAndCheckerMaker { public: - NCCLInitOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + NCCLInitOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddOutput("Communicator", "Create Communicator for communicating between gpus"); AddAttr>("gpus", "(vector) GPU id lists"); - AddAttr("data_type", + AddAttr("dtype", "(int, default 5 (FP32)) " "Output data type") - .SetDefault(framework::DataType::FP32); + .SetDefault(framework::proto::DataType::FP32); AddComment(R"DOC( NCCLInit Operator. @@ -141,8 +143,7 @@ class NCCLBcastOp : public framework::OperatorWithKernel { // AllreduceOp class NCCLAllReduceOpMaker : public framework::OpProtoAndCheckerMaker { public: - NCCLAllReduceOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + NCCLAllReduceOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "The input of AllReduce op"); AddInput("Communicator", "Communicator for communicating between gpus"); @@ -163,8 +164,7 @@ AllReduce the input tensors. // ReduceOp class NCCLReduceOpMaker : public framework::OpProtoAndCheckerMaker { public: - NCCLReduceOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + NCCLReduceOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "The input of Reduce op"); AddInput("Communicator", "Communicator for communicating between gpus"); @@ -190,8 +190,7 @@ Reduce the tensors. // BcastOp class NCCLBcastOpMaker : public framework::OpProtoAndCheckerMaker { public: - NCCLBcastOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + NCCLBcastOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "The input of BcastSend op"); AddInput("Communicator", "Communicator for communicating between gpus"); diff --git a/paddle/operators/nccl_op.cu.cc b/paddle/operators/nccl_op.cu.cc index 4f0a2a79edb9f24c7758fc91483d374425b36853..1b986a13650de7d77f4828d71798ee00d61c1284 100644 --- a/paddle/operators/nccl_op.cu.cc +++ b/paddle/operators/nccl_op.cu.cc @@ -67,7 +67,7 @@ class NCCLAllReduceKernel : public framework::OpKernel { auto stream = ctx.cuda_device_context().stream(); // device id - int gpu_id = boost::get(ctx.GetPlace()).GetDeviceId(); + int gpu_id = boost::get(ctx.GetPlace()).GetDeviceId(); int idx = comm->GetCommId(gpu_id); for (size_t i = 0; i < ins.size(); ++i) { @@ -120,7 +120,7 @@ class NCCLReduceKernel : public framework::OpKernel { ctx.device_context()) .stream(); // device id - int gpu_id = boost::get(ctx.GetPlace()).GetDeviceId(); + int gpu_id = boost::get(ctx.GetPlace()).GetDeviceId(); int idx = comm->GetCommId(gpu_id); auto ins_names = ctx.Inputs("X"); @@ -164,7 +164,7 @@ class NCCLBcastKernel : public framework::OpKernel { ctx.device_context()) .stream(); // device id - int gpu_id = boost::get(ctx.GetPlace()).GetDeviceId(); + int gpu_id = boost::get(ctx.GetPlace()).GetDeviceId(); int idx = comm->GetCommId(gpu_id); if (idx == root) { @@ -204,6 +204,6 @@ class NCCLBcastKernel : public framework::OpKernel { } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(ncclAllReduce, ops::NCCLAllReduceKernel); -REGISTER_OP_GPU_KERNEL(ncclBcast, ops::NCCLBcastKernel); -REGISTER_OP_GPU_KERNEL(ncclReduce, ops::NCCLReduceKernel); +REGISTER_OP_CUDA_KERNEL(ncclAllReduce, ops::NCCLAllReduceKernel); +REGISTER_OP_CUDA_KERNEL(ncclBcast, ops::NCCLBcastKernel); +REGISTER_OP_CUDA_KERNEL(ncclReduce, ops::NCCLReduceKernel); diff --git a/paddle/operators/nccl_op_test.cu.cc b/paddle/operators/nccl_op_test.cu.cc index 56ba57854955c08031214d1f751c17fbb8bb882c..6546096069d4c3fbc4908a16c2dba2ac6d7e6421 100644 --- a/paddle/operators/nccl_op_test.cu.cc +++ b/paddle/operators/nccl_op_test.cu.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 @@ -22,6 +22,7 @@ #include #include "paddle/framework/block_desc.h" +#include "paddle/framework/init.h" #include "paddle/framework/op_desc.h" #include "paddle/framework/op_registry.h" #include "paddle/framework/program_desc.h" @@ -33,9 +34,9 @@ #include "paddle/platform/place.h" USE_NO_KERNEL_OP(ncclInit); -USE_GPU_ONLY_OP(ncclAllReduce); -USE_GPU_ONLY_OP(ncclReduce); -USE_GPU_ONLY_OP(ncclBcast); +USE_CUDA_ONLY_OP(ncclAllReduce); +USE_CUDA_ONLY_OP(ncclReduce); +USE_CUDA_ONLY_OP(ncclBcast); namespace f = paddle::framework; namespace p = paddle::platform; @@ -49,9 +50,9 @@ const f::DDim kDims = {100, 100}; class NCCLTester : public ::testing::Test { public: virtual void SetUp() override { - cpu_ctx = new p::CPUDeviceContext(p::CPUPlace()); + paddle::platform::CPUPlace cpu_place; for (size_t i = 0; i < gpu_list.size(); ++i) { - p::GPUPlace place(i); + p::CUDAPlace place(i); dev_ctxs.emplace_back(new p::CUDADeviceContext(place)); } @@ -65,7 +66,8 @@ class NCCLTester : public ::testing::Test { } void NCCLInitOp() { - std::unique_ptr op1(new f::OpDescBind); + paddle::platform::CPUPlace cpu_place; + std::unique_ptr op1(new f::OpDesc); op1->SetType("ncclInit"); op1->SetOutput("Communicator", {"comm"}); @@ -76,17 +78,16 @@ class NCCLTester : public ::testing::Test { auto op = f::OpRegistry::CreateOp(*op1); VLOG(1) << "invoke NCCLInitOp."; - op->Run(g_scope, *cpu_ctx); + op->Run(g_scope, cpu_place); VLOG(1) << "NCCLInitOp finished."; } template - void PerThreadProgram(int gpu_id, const f::OpDescBind &op_desc, - f::Scope *scope) { + void PerThreadProgram(int gpu_id, const f::OpDesc &op_desc, f::Scope *scope) { std::unique_lock lk(mu); - const f::OpDescBind *op1 = &op_desc; + const f::OpDesc *op1 = &op_desc; - p::GPUPlace place(gpu_id); + p::CUDAPlace place(gpu_id); auto &ctx = dev_ctxs.at(gpu_id); auto *send_tensor = scope->Var("st")->GetMutable(); @@ -97,7 +98,7 @@ class NCCLTester : public ::testing::Test { send_tensor->mutable_data(kDims, place); std::vector send_vector(f::product(kDims), gpu_id); - send_tensor->CopyFromVector(send_vector, *ctx); + paddle::framework::CopyFromVector(send_vector, *ctx, send_tensor); ctx->Wait(); VLOG(1) << "Send Tensor filled with elements " << send_tensor->numel(); } @@ -112,40 +113,39 @@ class NCCLTester : public ::testing::Test { VLOG(1) << "Device : " << gpu_id << " invoke " << op_desc.Type(); VLOG(1) << " send_tensor : " << send_tensor->numel() << " recv_tensor : " << recv_tensor->numel(); - op->Run(*scope, *ctx); + op->Run(*scope, place); VLOG(1) << "Device : " << gpu_id << " finished " << op_desc.Type(); } public: std::vector dev_ctxs; - p::DeviceContext *cpu_ctx; f::Scope g_scope; std::mutex mu; }; // ncclInitOp with desc TEST(NCCL, ncclInitOp) { - std::unique_ptr op_desc(new f::OpDescBind); + std::unique_ptr op_desc(new f::OpDesc); op_desc->SetType("ncclInit"); op_desc->SetOutput("Communicator", {"x1"}); op_desc->SetAttr("gpus", {gpu_list}); f::Scope g_scope; - std::unique_ptr ctx(new p::CPUDeviceContext(p::CPUPlace())); + paddle::platform::CPUPlace cpu_place; auto *var = g_scope.Var("x1"); var->GetMutable(); auto op = f::OpRegistry::CreateOp(*op_desc); VLOG(1) << "invoke NCCLInitOp."; - op->Run(g_scope, *ctx.get()); + op->Run(g_scope, cpu_place); VLOG(1) << "NCCLInitOp finished."; } // ncclAllReduceOp with desc TEST_F(NCCLTester, ncclAllReduceOp) { - std::unique_ptr op2(new f::OpDescBind); + std::unique_ptr op2(new f::OpDesc); op2->SetType("ncclAllReduce"); op2->SetInput("X", {"st"}); op2->SetInput("Communicator", {"comm"}); @@ -171,7 +171,7 @@ TEST_F(NCCLTester, ncclAllReduceOp) { for (size_t i = 0; i < dev_scopes.size(); ++i) { p::CPUPlace cpu_place; - p::GPUPlace gpu_place(gpu_list[i]); + p::CUDAPlace gpu_place(gpu_list[i]); auto &recv_tensor = dev_scopes[i]->FindVar("rt")->Get(); auto *rt = recv_tensor.data(); @@ -180,7 +180,7 @@ TEST_F(NCCLTester, ncclAllReduceOp) { auto *ct = result_tensor->mutable_data(cpu_place); paddle::memory::Copy( - cpu_place, ct, p::GPUPlace(gpu_list[i]), rt, + cpu_place, ct, p::CUDAPlace(gpu_list[i]), rt, recv_tensor.numel() * sizeof(float), static_cast(dev_ctxs[i])->stream()); @@ -192,7 +192,7 @@ TEST_F(NCCLTester, ncclAllReduceOp) { // ncclReduceOp with desc TEST_F(NCCLTester, ncclReduceOp) { - std::unique_ptr op2(new f::OpDescBind); + std::unique_ptr op2(new f::OpDesc); const int kRoot = 0; op2->SetType("ncclReduce"); op2->SetInput("X", {"st"}); @@ -219,7 +219,7 @@ TEST_F(NCCLTester, ncclReduceOp) { float result = std::accumulate(gpu_list.begin(), gpu_list.end(), 0); p::CPUPlace cpu_place; - p::GPUPlace gpu_place(gpu_list[kRoot]); + p::CUDAPlace gpu_place(gpu_list[kRoot]); auto &recv_tensor = dev_scopes[kRoot]->FindVar("rt")->Get(); auto *rt = recv_tensor.data(); @@ -229,7 +229,7 @@ TEST_F(NCCLTester, ncclReduceOp) { auto *ct = result_tensor->mutable_data(cpu_place); paddle::memory::Copy( - cpu_place, ct, p::GPUPlace(gpu_list[kRoot]), rt, + cpu_place, ct, p::CUDAPlace(gpu_list[kRoot]), rt, recv_tensor.numel() * sizeof(float), static_cast(dev_ctxs[kRoot])->stream()); @@ -240,7 +240,7 @@ TEST_F(NCCLTester, ncclReduceOp) { // ncclBcastOp with desc TEST_F(NCCLTester, ncclBcastOp) { - std::unique_ptr op2(new f::OpDescBind); + std::unique_ptr op2(new f::OpDesc); const int kRoot = 5; op2->SetType("ncclBcast"); op2->SetInput("X", {"st"}); @@ -268,7 +268,7 @@ TEST_F(NCCLTester, ncclBcastOp) { float result = kRoot; p::CPUPlace cpu_place; - p::GPUPlace gpu_place(gpu_list[idx]); + p::CUDAPlace gpu_place(gpu_list[idx]); auto &recv_tensor = dev_scopes[idx]->FindVar("rt")->Get(); auto *rt = recv_tensor.data(); @@ -277,7 +277,7 @@ TEST_F(NCCLTester, ncclBcastOp) { auto *ct = result_tensor->mutable_data(cpu_place); paddle::memory::Copy( - cpu_place, ct, p::GPUPlace(gpu_list[idx]), rt, + cpu_place, ct, p::CUDAPlace(gpu_list[idx]), rt, recv_tensor.numel() * sizeof(float), static_cast(dev_ctxs[idx])->stream()); @@ -295,9 +295,18 @@ int main(int argc, char **argv) { return 0; } - for (int i = 0; i < dev_count; ++i) { + std::vector places; + + places.emplace_back(paddle::platform::CPUPlace()); + int count = paddle::platform::GetCUDADeviceCount(); + for (int i = 0; i < count; ++i) { + places.emplace_back(paddle::platform::CUDAPlace(i)); gpu_list.emplace_back(i); } + + VLOG(0) << " DeviceCount " << count; + paddle::platform::DeviceContextPool::Init(places); + testing::InitGoogleTest(&argc, argv); // device context should be release before scope. diff --git a/paddle/operators/nce_op.cc b/paddle/operators/nce_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..d39ca87d53518963f652f7b8c8cb289a6fef70fd --- /dev/null +++ b/paddle/operators/nce_op.cc @@ -0,0 +1,186 @@ +/* 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. */ + +#include "paddle/operators/nce_op.h" + +namespace paddle { +namespace operators { + +using framework::Tensor; + +class NCEOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Input")); + PADDLE_ENFORCE(ctx->HasInput("Label")); + PADDLE_ENFORCE(ctx->HasInput("Weight")); + PADDLE_ENFORCE(ctx->HasOutput("Cost")); + PADDLE_ENFORCE(ctx->HasOutput("SampleLogits")); + PADDLE_ENFORCE(ctx->HasOutput("SampleLabels")); + + auto x_dims = ctx->GetInputDim("Input"); + auto label_dims = ctx->GetInputDim("Label"); + PADDLE_ENFORCE_EQ(x_dims[0], label_dims[0]); + int num_true_classes = label_dims.size() == 2 ? label_dims[1] : 1; + if (ctx->HasInput("Bias")) { + PADDLE_ENFORCE_EQ(ctx->GetInputDim("Weight")[0], + ctx->GetInputDim("Bias")[0]); + } + auto num_neg_samples = ctx->Attrs().Get("num_neg_samples"); + auto num_total_classes = ctx->Attrs().Get("num_total_classes"); + std::vector custom_neg_classes = + ctx->Attrs().Get>("custom_neg_classes"); + PADDLE_ENFORCE_EQ(num_total_classes, ctx->GetInputDim("Weight")[0]); + if (custom_neg_classes.size() > 0) { + PADDLE_ENFORCE_EQ(custom_neg_classes.size(), + static_cast(num_neg_samples)); + } + // set dims of output(Out) + std::vector out_dims; + out_dims.push_back(x_dims[0]); + out_dims.push_back(1); + ctx->SetOutputDim("Cost", framework::make_ddim(out_dims)); + + // set dims of output(SampleOut) + std::vector sample_out_dims; + sample_out_dims.push_back(x_dims[0]); + sample_out_dims.push_back(num_neg_samples + num_true_classes); + ctx->SetOutputDim("SampleLogits", framework::make_ddim(sample_out_dims)); + ctx->SetOutputDim("SampleLabels", framework::make_ddim(sample_out_dims)); + } + + protected: + framework::OpKernelType GetActualKernelType( + const framework::ExecutionContext& ctx) const override { + return framework::OpKernelType( + framework::ToDataType(ctx.Input("Input")->type()), + ctx.GetPlace()); + } +}; + +class NCEOpMaker : public framework::OpProtoAndCheckerMaker { + public: + NCEOpMaker(OpProto* proto, OpAttrChecker* op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("Input", "(Tensor) A tensor of shape [batch_size, dim]."); + AddInput( + "Label", + "(Tensor) A tensor of shape [batch_size, num_true_class]. " + "'num_true_class' is the number of target classes in each sample." + "The number of target classes per sample should be same. " + "If you have a variable number of target classes, " + "you can pad them out to a constant number by either repeating them" + " or by padding with an otherwise unused class.)"); + AddInput("Weight", + "(Tensor) A tensor of shape [num_class, dim]. 'num_class' is the " + "total number of class."); + AddInput( + "Bias", + "(Tensor) A tensor of shape [num_class, 1]. 'num_class' is the total " + "number of class. It is a dispensable input.") + .AsDispensable(); + AddInput("SampleWeight", + "(Tensor) A tensor of shape [batch_size, 1] storing a weight for " + "each sample. And it is a dispensable input. The default value of " + "sample is 1.") + .AsDispensable(); + AddOutput("Cost", + "(Tensor) A tensor of shape [batch_size, 1]. Cost of samples."); + AddOutput("SampleLogits", + "An intermediate tensor of shape[batch_size, num_neg_samples + " + "num_pos_samples]." + "This tensor is output of forward kernel and used in backward " + "kernel to compute grads." + "Given X is the dot product of input tensor and sampled labels' " + "weights." + "Then 'SampleLogits' is sigmoid(X).") + .AsIntermediate(); + AddOutput("SampleLabels", + "An intermediate tensor of shape[batch_size, num_neg_samples + " + "num_pos_samples]." + "This tensor is output of forward kernel and used in backward " + "kernel to compute grads." + "") + .AsIntermediate(); + AddAttr("num_total_classes", + "Total number of classes in all samples."); + AddAttr("num_neg_samples", + "The number of negative classes. The default value is 10.") + .SetDefault(10); + AddAttr>("custom_neg_classes", + "This attribute only be used in unitest. Classes " + "in this list wiil be used as negative classes " + "for every samples. Under normal conditions, " + "user should avoid setting this attribute."); + AddComment(R"DOC( +Compute and return the noise-contrastive estimation training loss. +See [Noise-contrastive estimation: A new estimation principle for unnormalized statistical models](http://www.jmlr.org/proceedings/papers/v9/gutmann10a/gutmann10a.pdf). +By default this operator uses a uniform distribution for sampling. +)DOC"); + } +}; + +class NCEOpGrad : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Input")); + PADDLE_ENFORCE(ctx->HasInput("Weight")); + PADDLE_ENFORCE(ctx->HasInput("Cost")); + PADDLE_ENFORCE(ctx->HasInput("SampleLogits")); + PADDLE_ENFORCE(ctx->HasInput("SampleLabels")); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Cost")), + "The input(Out@GRAD) should not be null."); + + auto x_dims = ctx->GetInputDim("Input"); + auto x_grad_name = framework::GradVarName("Input"); + if (ctx->HasOutput(x_grad_name)) { + ctx->SetOutputDim(x_grad_name, x_dims); + } + + auto w_dims = ctx->GetInputDim("Weight"); + auto w_grad_name = framework::GradVarName("Weight"); + if (ctx->HasOutput(w_grad_name)) { + ctx->SetOutputDim(w_grad_name, w_dims); + } + + auto bias_grad_name = framework::GradVarName("Bias"); + if (ctx->HasOutput(bias_grad_name)) { + auto bias_dims = ctx->GetInputDim("Bias"); + ctx->SetOutputDim(bias_grad_name, bias_dims); + } + } + + protected: + framework::OpKernelType GetActualKernelType( + const framework::ExecutionContext& ctx) const override { + return framework::OpKernelType( + framework::ToDataType(ctx.Input("Input")->type()), + ctx.GetPlace()); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP(nce, ops::NCEOp, ops::NCEOpMaker, nce_grad, ops::NCEOpGrad); +REGISTER_OP_CPU_KERNEL(nce, ops::NCEKernel, + ops::NCEKernel); +REGISTER_OP_CPU_KERNEL(nce_grad, + ops::NCEGradKernel, + ops::NCEGradKernel); diff --git a/paddle/operators/nce_op.h b/paddle/operators/nce_op.h new file mode 100644 index 0000000000000000000000000000000000000000..e6b496f7896dcb412be8ff096fdccb2f0b682369 --- /dev/null +++ b/paddle/operators/nce_op.h @@ -0,0 +1,211 @@ +/* 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 + +#include +#include +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" +#include "unsupported/Eigen/CXX11/Tensor" +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; + +template +using EigenMatrix = framework::EigenMatrix; + +template +void PrepareSamples(const framework::ExecutionContext& context) { + auto label = context.Input("Label"); + const int64_t* label_data = label->data(); + auto label_dims = label->dims(); + int num_total_classes = context.Attr("num_total_classes"); + // for unitest + std::vector custom_neg_classes = + context.Attr>("custom_neg_classes"); + // random machine + std::random_device rd; + std::mt19937 rng(rd()); + std::uniform_int_distribution rand(0, num_total_classes - 1); + + auto sample_labels = context.Output("SampleLabels"); + auto sample_labels_dims = sample_labels->dims(); + int64_t* sample_labels_data = + sample_labels->mutable_data(context.GetPlace()); + + int num_label = label_dims.size() == 2 ? label_dims[1] : 1; + int index = 0; + for (int64_t i = 0; i < label_dims[0]; ++i) { + int j = 0; + for (; j < num_label; ++j) { + sample_labels_data[index++] = label_data[i * num_label + j]; + } + if (custom_neg_classes.size() > 0) { + for (auto label : custom_neg_classes) { + sample_labels_data[index++] = label; + } + } else { + for (; j < sample_labels_dims[1]; ++j) { + // TODO(wanghaoshuang): support more distribution sampling + sample_labels_data[index++] = rand(rng); + } + } + } +} + +template +class NCEKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + PrepareSamples(context); + auto sample_labels = context.Output("SampleLabels"); + const int64_t* sample_labels_data = sample_labels->data(); + auto sample_out = context.Output("SampleLogits"); + T* sample_out_data = sample_out->mutable_data(context.GetPlace()); + auto label = context.Input("Label"); + auto sample_weight = context.Input("SampleWeight"); + const T* sample_weight_data = nullptr; + if (sample_weight != nullptr) { + sample_weight_data = sample_weight->data(); + } + auto out = context.Output("Cost"); + T* out_data = out->mutable_data(context.GetPlace()); + int num_neg_samples = context.Attr("num_neg_samples"); + int num_total_classes = context.Attr("num_total_classes"); + int64_t num_true_class = 1; + if (label != nullptr) { + num_true_class = label->dims()[1]; + } + T b = 1. / num_total_classes * num_neg_samples; + // forward bias + auto bias = context.Input("Bias"); + if (bias != nullptr) { + const T* bias_data = bias->data(); + for (int64_t i = 0; i < sample_labels->numel(); ++i) { + sample_out_data[i] = bias_data[sample_labels_data[i]]; + } + } else { + for (int64_t i = 0; i < sample_labels->numel(); ++i) { + sample_out_data[i] = 0; + } + } + // forward mul + auto input_mat = EigenMatrix::From(*(context.Input("Input"))); + auto weight_mat = EigenMatrix::From(*(context.Input("Weight"))); + for (int64_t i = 0; i < sample_labels->numel(); ++i) { + Eigen::Tensor result = + (input_mat.chip((int)(i / sample_labels->dims()[1]), 0) * + weight_mat.chip(sample_labels_data[i], 0)) + .sum(); + sample_out_data[i] += result(0); + sample_out_data[i] = (1. / (1. + exp(-sample_out_data[i]))); + } + // forward cost + for (int64_t i = 0; i < sample_labels->dims()[0]; ++i) { + int64_t j = 0; + out_data[i] = 0; + T w = sample_weight == nullptr ? 1. : sample_weight_data[i]; + // for true classes + for (; j < num_true_class; ++j) { + T o = sample_out_data[i * sample_out->dims()[1] + j]; + T cost = -log(o / (o + b)); + out_data[i] += w * cost; + } + // for sampled neg classes + for (; j < sample_labels->dims()[1]; ++j) { + T o = sample_out_data[i * sample_out->dims()[1] + j]; + T cost = -log(b / (o + b)); + out_data[i] += w * cost; + } + } + } +}; + +template +class NCEGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto d_out = context.Input(framework::GradVarName("Cost")); + const T* d_out_data = d_out->data(); + auto label = context.Input("Label"); + auto sample_out = context.Input("SampleLogits"); + const T* sample_out_data = sample_out->data(); + auto sample_labels = context.Input("SampleLabels"); + const int64_t* sample_labels_data = sample_labels->data(); + auto sample_weight = context.Input("SampleWeight"); + const T* sample_weight_data = nullptr; + if (sample_weight != nullptr) { + sample_weight_data = sample_weight->data(); + } + int num_neg_samples = context.Attr("num_neg_samples"); + int num_total_classes = context.Attr("num_total_classes"); + int num_true_class = 1; + if (label != nullptr) { + num_true_class = label->dims()[1]; + } + T b = 1. / num_total_classes * num_neg_samples; + Tensor sample_grad; // tmp tensor + T* sample_grad_data = + sample_grad.mutable_data(sample_labels->dims(), context.GetPlace()); + // backward cost + for (int64_t i = 0; i < sample_labels->numel(); ++i) { + T o = sample_out_data[i]; + T w = sample_weight == nullptr + ? 1 + : sample_weight_data[i / sample_labels->dims()[1]]; + sample_grad_data[i] = (i % sample_labels->dims()[1]) < num_true_class + ? w * (b / (o + b)) * (o - 1) + : w * (o * (1 - o) / (o + b)); + sample_grad_data[i] *= d_out_data[i / sample_labels->dims()[1]]; + } + // get d_bias + auto d_bias = context.Output(framework::GradVarName("Bias")); + if (d_bias != nullptr) { + T* d_bias_data = d_bias->mutable_data(context.GetPlace()); + std::fill(d_bias_data, d_bias_data + d_bias->numel(), 0.0); + for (int64_t i = 0; i < sample_labels->numel(); ++i) { + d_bias_data[sample_labels_data[i]] += sample_grad_data[i]; + } + } + // get d_w + auto d_w = context.Output(framework::GradVarName("Weight")); + if (d_w != nullptr) { + auto d_w_data = d_w->mutable_data(context.GetPlace()); + std::fill(d_w_data, d_w_data + d_w->numel(), 0.0); + auto d_w_matrix = EigenMatrix::From(*d_w); + auto x_matrix = EigenMatrix::From(*(context.Input("Input"))); + for (int64_t i = 0; i < sample_labels->numel(); ++i) { + d_w_matrix.chip(sample_labels_data[i], 0) += + x_matrix.chip((int)(i / sample_labels->dims()[1]), 0) * + sample_grad_data[i]; + } + } + // get d_x + auto d_x = context.Output(framework::GradVarName("Input")); + if (d_x != nullptr) { + d_x->mutable_data(context.GetPlace()); + auto d_x_matrix = EigenMatrix::From(*d_x); + auto w_matrix = EigenMatrix::From(*(context.Input("Weight"))); + for (int64_t i = 0; i < sample_labels->numel(); ++i) { + d_x_matrix.chip((int)(i / sample_labels->dims()[1]), 0) += + w_matrix.chip(sample_labels_data[i], 0) * sample_grad_data[i]; + } + } + } +}; +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/net_op.h b/paddle/operators/net_op.h index ebeb262d9621fa35c870b6407992f6b6d2bf7c70..85d0153b32c0ba53bfe0912fc2682c8b635ba172 100644 --- a/paddle/operators/net_op.h +++ b/paddle/operators/net_op.h @@ -38,7 +38,10 @@ namespace operators { class NetOp : public framework::OperatorBase { public: static const char kAll[]; - NetOp() : framework::OperatorBase("plain_net", {}, {}, {}) {} + NetOp() + : framework::OperatorBase("plain_net", framework::VariableNameMap{}, + framework::VariableNameMap{}, + framework::AttributeMap{}) {} NetOp(const std::string& type, const framework::VariableNameMap& inputs, const framework::VariableNameMap& outputs, @@ -62,9 +65,9 @@ class NetOp : public framework::OperatorBase { * will be used. */ void Run(const framework::Scope& scope, - const platform::DeviceContext& dev_ctx) const override { + const platform::Place& place) const override { for (auto& op : ops_) { - op->Run(scope, dev_ctx); + op->Run(scope, place); } } diff --git a/paddle/operators/net_op_test.cc b/paddle/operators/net_op_test.cc index 63bebd5b44719868a38ddf2b023955d1ab05245c..dfd86546e83a6276aedd198eaeb6fad2c50944df 100644 --- a/paddle/operators/net_op_test.cc +++ b/paddle/operators/net_op_test.cc @@ -13,8 +13,7 @@ class TestOp : public framework::OperatorBase { public: using framework::OperatorBase::OperatorBase; DEFINE_OP_CLONE_METHOD(TestOp); - void Run(const Scope& scope, - const platform::DeviceContext& dev_ctx) const override { + void Run(const Scope& scope, const platform::Place& place) const override { ++run_cnt; } }; @@ -38,10 +37,10 @@ TEST(OpKernel, all) { net->AppendOp(std::unique_ptr( new TestOp("test", {{"X", {"x"}}, {"W", {"w1"}}, {"b", {"b1"}}}, - {{"Out", {"y"}}}, {}))); + {{"Out", {"y"}}}, framework::AttributeMap{}))); net->AppendOp(std::unique_ptr( new TestOp("test", {{"X", {"y"}}, {"W", {"w2"}}, {"b", {"b2"}}}, - {{"Out", {"z"}}}, {}))); + {{"Out", {"z"}}}, framework::AttributeMap{}))); net->CompleteAddOp(); AssertSameVectorWithoutOrder({"x", "w1", "b1", "w2", "b2"}, @@ -58,7 +57,7 @@ TEST(NetOp, insert_op) { NetOp net; auto op1 = std::unique_ptr( new framework::NOP("empty", {{"X", {"x"}}, {"W", {"w1"}}, {"b", {"b1"}}}, - {{"Out", {"y"}}}, {})); + {{"Out", {"y"}}}, framework::AttributeMap{})); net.AppendOp(*op1); net.InsertOp(0, *op1); ASSERT_EQ(2UL, net.ops_.size()); @@ -68,10 +67,12 @@ TEST(NetOp, insert_op) { TEST(NetOp, Clone) { NetOp net; - net.AppendOp( - std::unique_ptr(new framework::NOP{"empty", {}, {}, {}})); - net.AppendOp(std::unique_ptr( - new framework::NOP{"empty2", {}, {}, {}})); + net.AppendOp(std::unique_ptr(new framework::NOP{ + "empty", framework::VariableNameMap{}, framework::VariableNameMap{}, + framework::AttributeMap{}})); + net.AppendOp(std::unique_ptr(new framework::NOP{ + "empty2", framework::VariableNameMap{}, framework::VariableNameMap{}, + framework::AttributeMap{}})); net.CompleteAddOp(true); auto new_net_op = net.Clone(); ASSERT_NE(new_net_op, nullptr); diff --git a/paddle/operators/batch_norm_op.md b/paddle/operators/op_documentation/batch_norm_op.md similarity index 100% rename from paddle/operators/batch_norm_op.md rename to paddle/operators/op_documentation/batch_norm_op.md diff --git a/paddle/operators/name_convention.md b/paddle/operators/op_documentation/name_convention.md similarity index 96% rename from paddle/operators/name_convention.md rename to paddle/operators/op_documentation/name_convention.md index b5cb176e003b4584321142ac9f1c3380b7010936..a02b356f058da68442516c2705d0bac140f8ef18 100644 --- a/paddle/operators/name_convention.md +++ b/paddle/operators/op_documentation/name_convention.md @@ -35,8 +35,8 @@ Here we give some examples to show how these rules will be used. ```c++ class AccumulateOpMaker : public framework::OpProtoAndCheckerMaker { public: - AccumulateOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + AccumulateOpMaker(OpProto *proto, + OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "(Tensor) The input tensor that has to be accumulated to the output tensor. If the output size is not the same as input size, diff --git a/paddle/operators/net_op_design.md b/paddle/operators/op_documentation/net_op_design.md similarity index 100% rename from paddle/operators/net_op_design.md rename to paddle/operators/op_documentation/net_op_design.md diff --git a/paddle/operators/op_documentation/op_markdown_format.md b/paddle/operators/op_documentation/op_markdown_format.md new file mode 100644 index 0000000000000000000000000000000000000000..0ee804d592252c727622cbe59b0644813db3c4fd --- /dev/null +++ b/paddle/operators/op_documentation/op_markdown_format.md @@ -0,0 +1,64 @@ +# Standard Markdown Format for Operators +The following should be the standard format for documentation for all the operators that will get rendered in the `html`: + +``` +Operator Name (In PaddlePaddle) + +Operator Name (Standard) + +Operator description. + +LaTeX equation of how the operator performs an update. + +The signature of the operator. +``` + +Each section mentioned above has been covered in further detail in the rest of the document. + +# PaddlePaddle Operator Name +This should be in all small letters, in case of multiple words, we separate them with an underscore. For example: +`array to lod tensor` should be written as `array_to_lod_tensor`. + +This naming convention should be standard across all PaddlePaddle operators. + +# Standard Operator Name +This is the standard name of the operator as used in the community. The general standard is usually: +- Standard abbreviations like `SGD` are written in all capital letters. +- Operator names that have multiple words inside a single word use `camelCase` (capitalize word boundaries inside of a word). +- Keep numbers inside a word as is, with no boundary delimiters. +- Follow the name of the operator with the keyword: `Activation Operator.` + +# Operator description +This section should contain the description of what the operator does, including the operation performed, the literature from where it comes and was introduced first, and other important details. The relevant paper/article including the hyperlink should be cited in this section. + +# LaTeX equation +This section should contain an overall equation of the update or operation that the operator performs. The variables used in the equation should follow the naming convention of operators as described [here](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/name_convention.md). Two words in the same word should be separated by an underscore (`_`). + +# The signature +This section describes the signature of the operator. A list of Inputs and Outputs, each of which have a small description of what the variable represents and the type of variable. The variable names follow the `CamelCase` naming convention. The proposed format for this is: +`Section : +VariableName : (VariableType) VariableDescription +... +... +` + + +The following example for an `sgd` operator covers the above mentioned sections as they would ideally look like in the `html`: + +``` +sgd + +SGD operator + +This operator implements one step of the stochastic gradient descent algorithm. + +param_out = param_learning_rate * grad + +Inputs: +Param : (Tensor) Input parameter +LearningRate : (Tensor) Learning rate of SGD +Grad : (Tensor) Input gradient + +Outputs: +ParamOut : (Tensor) Output parameter +``` diff --git a/paddle/operators/rnn_design.md b/paddle/operators/op_documentation/rnn_design.md similarity index 100% rename from paddle/operators/rnn_design.md rename to paddle/operators/op_documentation/rnn_design.md diff --git a/paddle/operators/pad_op.cc b/paddle/operators/pad_op.cc index adb75df6ef10c59fc6f3db4d36e1ffb1ae0b4b1e..90c53bd17732aa046beceb3ac0a3b8c0d69994f3 100644 --- a/paddle/operators/pad_op.cc +++ b/paddle/operators/pad_op.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/operators/pad_op.h" @@ -48,7 +48,7 @@ class PadOp : public framework::OperatorWithKernel { class PadOpMaker : public framework::OpProtoAndCheckerMaker { public: - PadOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker) + PadOpMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "The input of pad op. " @@ -116,14 +116,14 @@ class PadOpGradMaker : public framework::SingleGradOpDescMaker { using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; protected: - std::unique_ptr Apply() const override { - auto* bind = new framework::OpDescBind(); + std::unique_ptr Apply() const override { + auto* bind = new framework::OpDesc(); bind->SetInput("X", Input("X")); bind->SetInput(framework::GradVarName("Out"), OutputGrad("Out")); bind->SetOutput(framework::GradVarName("X"), InputGrad("X")); bind->SetAttrMap(Attrs()); bind->SetType("pad_grad"); - return std::unique_ptr(bind); + return std::unique_ptr(bind); } }; @@ -134,6 +134,7 @@ namespace ops = paddle::operators; REGISTER_OPERATOR(pad, ops::PadOp, ops::PadOpMaker, ops::PadOpGradMaker); REGISTER_OPERATOR(pad_grad, ops::PadOpGrad); -REGISTER_OP_CPU_KERNEL(pad, ops::PadKernel); -REGISTER_OP_CPU_KERNEL(pad_grad, - ops::PadGradKernel); +REGISTER_OP_CPU_KERNEL( + pad, ops::PadKernel); +REGISTER_OP_CPU_KERNEL( + pad_grad, ops::PadGradKernel); diff --git a/paddle/operators/pad_op.cu b/paddle/operators/pad_op.cu index 555a7dba23c6fa2659cabf4858b42ff70d74bf18..433b5f1112a27b36edbe6d99fcdd4fc8395bc2e8 100644 --- a/paddle/operators/pad_op.cu +++ b/paddle/operators/pad_op.cu @@ -1,21 +1,22 @@ /* 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 +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 + 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. */ +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. */ #define EIGEN_USE_GPU #include "paddle/operators/pad_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(pad, ops::PadKernel); -REGISTER_OP_GPU_KERNEL(pad_grad, - ops::PadGradKernel); +REGISTER_OP_CUDA_KERNEL( + pad, ops::PadKernel); +REGISTER_OP_CUDA_KERNEL( + pad_grad, ops::PadGradKernel); diff --git a/paddle/operators/pad_op.h b/paddle/operators/pad_op.h index 9534dbf54529e3b9ae2b6640d51fe291e9521927..fdf91a5776620485c38a8b2c5f8b26039e438d0c 100644 --- a/paddle/operators/pad_op.h +++ b/paddle/operators/pad_op.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 @@ -26,7 +26,7 @@ template using EigenTensor = framework::EigenTensor; -template +template void PadFunction(const framework::ExecutionContext& context) { auto pads = context.Attr>("paddings"); Eigen::array, D> paddings; @@ -42,33 +42,34 @@ void PadFunction(const framework::ExecutionContext& context) { auto x_tensor = EigenTensor::From(*x); auto out_tensor = EigenTensor::From(*out); - auto place = context.GetEigenDevice(); + auto& place = + *context.template device_context().eigen_device(); out_tensor.device(place) = x_tensor.pad(paddings, pad_value); } -template +template class PadKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { int rank = context.Input("X")->dims().size(); switch (rank) { case 1: - PadFunction(context); + PadFunction(context); break; case 2: - PadFunction(context); + PadFunction(context); break; case 3: - PadFunction(context); + PadFunction(context); break; case 4: - PadFunction(context); + PadFunction(context); break; case 5: - PadFunction(context); + PadFunction(context); break; case 6: - PadFunction(context); + PadFunction(context); break; default: PADDLE_THROW( @@ -77,7 +78,7 @@ class PadKernel : public framework::OpKernel { } }; -template +template void PadGradFunction(const framework::ExecutionContext& context) { auto pads = context.Attr>("paddings"); Eigen::array, D> paddings; @@ -91,12 +92,13 @@ void PadGradFunction(const framework::ExecutionContext& context) { d_x->mutable_data(context.GetPlace()); auto d_x_tensor = EigenTensor::From(*d_x); auto d_out_tensor = EigenTensor::From(*d_out); - auto place = context.GetEigenDevice(); + auto& place = + *context.template device_context().eigen_device(); d_x_tensor.device(place) = d_out_tensor.pad(paddings, 0); } } -template +template class PadGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { @@ -104,22 +106,22 @@ class PadGradKernel : public framework::OpKernel { context.Input(framework::GradVarName("Out"))->dims().size(); switch (rank) { case 1: - PadGradFunction(context); + PadGradFunction(context); break; case 2: - PadGradFunction(context); + PadGradFunction(context); break; case 3: - PadGradFunction(context); + PadGradFunction(context); break; case 4: - PadGradFunction(context); + PadGradFunction(context); break; case 5: - PadGradFunction(context); + PadGradFunction(context); break; case 6: - PadGradFunction(context); + PadGradFunction(context); break; default: PADDLE_THROW( diff --git a/paddle/operators/pool_cudnn_op.cc b/paddle/operators/pool_cudnn_op.cc index f962d9e3e6abde14ce21eb0102f10d139fdb160e..77407f5cdf7e4ef7b76c38ef8992517b4bd1c5fe 100644 --- a/paddle/operators/pool_cudnn_op.cc +++ b/paddle/operators/pool_cudnn_op.cc @@ -19,7 +19,21 @@ namespace ops = paddle::operators; REGISTER_OP(pool2d_cudnn, ops::PoolOp, ops::Pool2dOpMaker, pool2d_cudnn_grad, ops::PoolOpGrad); -REGISTER_OP_CPU_KERNEL(pool2d_cudnn, - ops::PoolKernel); -REGISTER_OP_CPU_KERNEL(pool2d_cudnn_grad, - ops::PoolGradKernel) +REGISTER_OP_CPU_KERNEL( + pool2d_cudnn, ops::PoolKernel, + ops::PoolKernel); +REGISTER_OP_CPU_KERNEL( + pool2d_cudnn_grad, + ops::PoolGradKernel, + ops::PoolGradKernel) + +REGISTER_OP(pool3d_cudnn, ops::PoolOp, ops::Pool3dOpMaker, pool3d_cudnn_grad, + ops::PoolOpGrad); + +REGISTER_OP_CPU_KERNEL( + pool3d_cudnn, ops::PoolKernel, + ops::PoolKernel); +REGISTER_OP_CPU_KERNEL( + pool3d_cudnn_grad, + ops::PoolGradKernel, + ops::PoolGradKernel) diff --git a/paddle/operators/pool_cudnn_op.cu.cc b/paddle/operators/pool_cudnn_op.cu.cc index f9d8af3e1c5db49873979fdfeb17a32d16341a1a..2d0001ba1184c99d9fc642f60c97ba89cec97ccd 100644 --- a/paddle/operators/pool_cudnn_op.cu.cc +++ b/paddle/operators/pool_cudnn_op.cu.cc @@ -29,7 +29,7 @@ class PoolCudnnOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext &ctx) const override { PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), - "It must use GPUPlace."); + "It must use CUDAPlace."); const Tensor *input = ctx.Input("X"); Tensor *output = ctx.Output("Out"); @@ -52,7 +52,13 @@ class PoolCudnnOpKernel : public framework::OpKernel { ScopedTensorDescriptor input_desc; ScopedTensorDescriptor output_desc; ScopedPoolingDescriptor pool_desc; - DataLayout layout = DataLayout::kNCHW; + DataLayout layout; + + if (strides.size() == 2U) { + layout = DataLayout::kNCHW; + } else { + layout = DataLayout::kNCDHW; + } cudnnTensorDescriptor_t cudnn_input_desc = input_desc.descriptor( layout, framework::vectorize2int(input->dims())); @@ -84,7 +90,7 @@ class PoolCudnnGradOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext &ctx) const override { PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), - "It must use GPUPlace."); + "It must use CUDAPlace."); const Tensor *input = ctx.Input("X"); const Tensor *output = ctx.Input("Out"); @@ -112,7 +118,13 @@ class PoolCudnnGradOpKernel : public framework::OpKernel { ScopedTensorDescriptor input_desc; ScopedTensorDescriptor output_desc; ScopedPoolingDescriptor pool_desc; - DataLayout layout = DataLayout::kNCHW; + DataLayout layout; + + if (strides.size() == 2U) { + layout = DataLayout::kNCHW; + } else { + layout = DataLayout::kNCDHW; + } cudnnTensorDescriptor_t cudnn_input_desc = input_desc.descriptor( layout, framework::vectorize2int(input->dims())); @@ -150,5 +162,12 @@ class PoolCudnnGradOpKernel : public framework::OpKernel { namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(pool2d_cudnn, ops::PoolCudnnOpKernel); -REGISTER_OP_GPU_KERNEL(pool2d_cudnn_grad, ops::PoolCudnnGradOpKernel); +REGISTER_OP_CUDA_KERNEL(pool2d_cudnn, ops::PoolCudnnOpKernel, + ops::PoolCudnnOpKernel); +REGISTER_OP_CUDA_KERNEL(pool2d_cudnn_grad, ops::PoolCudnnGradOpKernel, + ops::PoolCudnnGradOpKernel); + +REGISTER_OP_CUDA_KERNEL(pool3d_cudnn, ops::PoolCudnnOpKernel, + ops::PoolCudnnOpKernel); +REGISTER_OP_CUDA_KERNEL(pool3d_cudnn_grad, ops::PoolCudnnGradOpKernel, + ops::PoolCudnnGradOpKernel); diff --git a/paddle/operators/pool_op.cc b/paddle/operators/pool_op.cc index f3963b1995ef8767786f0bf230b134afc69aa99d..50057eb6483e9c9e745bc07dee26a0bbbbb5a48c 100644 --- a/paddle/operators/pool_op.cc +++ b/paddle/operators/pool_op.cc @@ -67,8 +67,7 @@ void PoolOpGrad::InferShape(framework::InferShapeContext *ctx) const { ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X")); } -Pool2dOpMaker::Pool2dOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) +Pool2dOpMaker::Pool2dOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput( "X", @@ -105,7 +104,7 @@ Pool2dOpMaker::Pool2dOpMaker(framework::OpProto *proto, // TypedAttrChecker don't support vector type.) AddAttr>( "paddings", - "(vector, defalut {0,0}), paddings(height, width) of pooling " + "(vector, default {0,0}), paddings(height, width) of pooling " "operator." "If global_pooling = true, paddings and ksize will be ignored.") .SetDefault({0, 0}); // TODO(Chengduo): Add checker. (Currently, @@ -122,22 +121,21 @@ Parameters(ksize, strides, paddings) are two elements. These two elements represent height and width, respectively. The input(X) size and output(Out) size may be different. -Example: +Example: Input: X shape: $(N, C, H_{in}, W_{in})$ Output: Out shape: $(N, C, H_{out}, W_{out})$ - where + Where $$ - H_{out} = (H_{in} - ksize[0] + 2 * paddings[0]) / strides[0] + 1 \\ - W_{out} = (W_{in} - ksize[1] + 2 * paddings[1]) / strides[1] + 1 + H_{out} = \frac{(H_{in} - ksize[0] + 2 * paddings[0])}{strides[0]} + 1 \\ + W_{out} = \frac{(W_{in} - ksize[1] + 2 * paddings[1])}{strides[1]} + 1 $$ )DOC"); } -Pool3dOpMaker::Pool3dOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) +Pool3dOpMaker::Pool3dOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "(Tensor) The input tensor of pooling operator. " @@ -177,7 +175,7 @@ Pool3dOpMaker::Pool3dOpMaker(framework::OpProto *proto, // TypedAttrChecker don't support vector type.) AddAttr>( "paddings", - "(vector, defalut {0,0,0}), paddings(depth, height, " + "(vector, default {0,0,0}), paddings(depth, height, " "width) of pooling operator. " "If global_pooling = true, ksize and paddings will be ignored.") .SetDefault({0, 0, 0}); // TODO(Chengduo): Add checker. (Currently, @@ -199,12 +197,12 @@ Example: X shape: $(N, C, D_{in}, H_{in}, W_{in})$ Output: Out shape: $(N, C, D_{out}, H_{out}, W_{out})$ - where - $$ - D_{out} = (D_{in} - ksize[0] + 2 * paddings[0]) / strides[0] + 1 \\ - H_{out} = (H_{in} - ksize[1] + 2 * paddings[1]) / strides[1] + 1 \\ - W_{out} = (W_{in} - ksize[2] + 2 * paddings[2]) / strides[2] + 1 - $$ + Where + $$ + D_{out} = \frac{(D_{in} - ksize[0] + 2 * paddings[0])}{strides[0]} + 1 \\ + H_{out} = \frac{(H_{in} - ksize[1] + 2 * paddings[1])}{strides[1]} + 1 \\ + W_{out} = \frac{(W_{in} - ksize[2] + 2 * paddings[2])}{strides[2]} + 1 + $$ )DOC"); } @@ -216,15 +214,19 @@ namespace ops = paddle::operators; REGISTER_OP(pool2d, ops::PoolOp, ops::Pool2dOpMaker, pool2d_grad, ops::PoolOpGrad); -REGISTER_OP_CPU_KERNEL(pool2d, - ops::PoolKernel); -REGISTER_OP_CPU_KERNEL(pool2d_grad, - ops::PoolGradKernel) +REGISTER_OP_CPU_KERNEL( + pool2d, ops::PoolKernel, + ops::PoolKernel); +REGISTER_OP_CPU_KERNEL( + pool2d_grad, ops::PoolGradKernel, + ops::PoolGradKernel) REGISTER_OP(pool3d, ops::PoolOp, ops::Pool3dOpMaker, pool3d_grad, ops::PoolOpGrad); -REGISTER_OP_CPU_KERNEL(pool3d, - ops::PoolKernel); -REGISTER_OP_CPU_KERNEL(pool3d_grad, - ops::PoolGradKernel); +REGISTER_OP_CPU_KERNEL( + pool3d, ops::PoolKernel, + ops::PoolKernel); +REGISTER_OP_CPU_KERNEL( + pool3d_grad, ops::PoolGradKernel, + ops::PoolGradKernel); diff --git a/paddle/operators/pool_op.cu.cc b/paddle/operators/pool_op.cu.cc index 0e3b80868f7b9d1697d619889160856d65ad59a3..39a9dfbf794b3dbaf81e2435f8609014dc27f3af 100644 --- a/paddle/operators/pool_op.cu.cc +++ b/paddle/operators/pool_op.cu.cc @@ -16,12 +16,18 @@ limitations under the License. */ namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(pool2d, - ops::PoolKernel); -REGISTER_OP_GPU_KERNEL(pool2d_grad, - ops::PoolGradKernel); +REGISTER_OP_CUDA_KERNEL( + pool2d, ops::PoolKernel, + ops::PoolKernel); +REGISTER_OP_CUDA_KERNEL( + pool2d_grad, + ops::PoolGradKernel, + ops::PoolGradKernel); -REGISTER_OP_GPU_KERNEL(pool3d, - ops::PoolKernel); -REGISTER_OP_GPU_KERNEL(pool3d_grad, - ops::PoolGradKernel); +REGISTER_OP_CUDA_KERNEL( + pool3d, ops::PoolKernel, + ops::PoolKernel); +REGISTER_OP_CUDA_KERNEL( + pool3d_grad, + ops::PoolGradKernel, + ops::PoolGradKernel); diff --git a/paddle/operators/pool_op.h b/paddle/operators/pool_op.h index 63492a89e8d4e44a036bc3c2b16cc54c7e77b534..3860e295f4b4dbeb2d60cfb304847de39083f1e1 100644 --- a/paddle/operators/pool_op.h +++ b/paddle/operators/pool_op.h @@ -40,17 +40,15 @@ class PoolOpGrad : public framework::OperatorWithKernel { class Pool2dOpMaker : public framework::OpProtoAndCheckerMaker { public: - Pool2dOpMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker); + Pool2dOpMaker(OpProto* proto, OpAttrChecker* op_checker); }; class Pool3dOpMaker : public framework::OpProtoAndCheckerMaker { public: - Pool3dOpMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker); + Pool3dOpMaker(OpProto* proto, OpAttrChecker* op_checker); }; -template +template class PoolKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { @@ -67,41 +65,41 @@ class PoolKernel : public framework::OpKernel { ksize[i] = static_cast(in_x->dims()[i + 2]); } } - + auto& dev_ctx = context.template device_context(); switch (ksize.size()) { case 2: { if (pooling_type == "max") { paddle::operators::math::Pool2dFunctor< - Place, paddle::operators::math::MaxPool, T> + DeviceContext, paddle::operators::math::MaxPool, T> pool2d_forward; paddle::operators::math::MaxPool pool_process; - pool2d_forward(context.device_context(), *in_x, ksize, strides, - paddings, pool_process, out); + pool2d_forward(dev_ctx, *in_x, ksize, strides, paddings, pool_process, + out); } else if (pooling_type == "avg") { paddle::operators::math::Pool2dFunctor< - Place, paddle::operators::math::AvgPool, T> + DeviceContext, paddle::operators::math::AvgPool, T> pool2d_forward; paddle::operators::math::AvgPool pool_process; - pool2d_forward(context.device_context(), *in_x, ksize, strides, - paddings, pool_process, out); + pool2d_forward(dev_ctx, *in_x, ksize, strides, paddings, pool_process, + out); } } break; case 3: { if (pooling_type == "max") { paddle::operators::math::Pool3dFunctor< - Place, paddle::operators::math::MaxPool, T> + DeviceContext, paddle::operators::math::MaxPool, T> pool3d_forward; paddle::operators::math::MaxPool pool_process; - pool3d_forward(context.device_context(), *in_x, ksize, strides, - paddings, pool_process, out); + pool3d_forward(dev_ctx, *in_x, ksize, strides, paddings, pool_process, + out); } else if (pooling_type == "avg") { paddle::operators::math::Pool3dFunctor< - Place, paddle::operators::math::AvgPool, T> + DeviceContext, paddle::operators::math::AvgPool, T> pool3d_forward; paddle::operators::math::AvgPool pool_process; - pool3d_forward(context.device_context(), *in_x, ksize, strides, - paddings, pool_process, out); + pool3d_forward(dev_ctx, *in_x, ksize, strides, paddings, pool_process, + out); } } break; default: { PADDLE_THROW("Pool op only supports 2D and 3D input."); } @@ -109,7 +107,7 @@ class PoolKernel : public framework::OpKernel { } }; -template +template class PoolGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { @@ -130,42 +128,43 @@ class PoolGradKernel : public framework::OpKernel { ksize[i] = static_cast(in_x->dims()[i + 2]); } } - + auto& dev_ctx = context.template device_context(); if (in_x_grad) { in_x_grad->mutable_data(context.GetPlace()); auto temp = framework::EigenVector::Flatten(*in_x_grad); - temp.device(context.GetEigenDevice()) = + temp.device( + *context.template device_context().eigen_device()) = temp.constant(static_cast(0)); switch (ksize.size()) { case 2: { if (pooling_type == "max") { - paddle::operators::math::MaxPool2dGradFunctor + paddle::operators::math::MaxPool2dGradFunctor pool2d_backward; - pool2d_backward(context.device_context(), *in_x, *out, *out_grad, - ksize, strides, paddings, in_x_grad); + pool2d_backward(dev_ctx, *in_x, *out, *out_grad, ksize, strides, + paddings, in_x_grad); } else if (pooling_type == "avg") { paddle::operators::math::Pool2dGradFunctor< - Place, paddle::operators::math::AvgPoolGrad, T> + DeviceContext, paddle::operators::math::AvgPoolGrad, T> pool2d_backward; paddle::operators::math::AvgPoolGrad pool_process; - pool2d_backward(context.device_context(), *in_x, *out, *out_grad, - ksize, strides, paddings, pool_process, in_x_grad); + pool2d_backward(dev_ctx, *in_x, *out, *out_grad, ksize, strides, + paddings, pool_process, in_x_grad); } } break; case 3: { if (pooling_type == "max") { - paddle::operators::math::MaxPool3dGradFunctor + paddle::operators::math::MaxPool3dGradFunctor pool3d_backward; - pool3d_backward(context.device_context(), *in_x, *out, *out_grad, - ksize, strides, paddings, in_x_grad); + pool3d_backward(dev_ctx, *in_x, *out, *out_grad, ksize, strides, + paddings, in_x_grad); } else if (pooling_type == "avg") { paddle::operators::math::Pool3dGradFunctor< - Place, paddle::operators::math::AvgPoolGrad, T> + DeviceContext, paddle::operators::math::AvgPoolGrad, T> pool3d_backward; paddle::operators::math::AvgPoolGrad pool_process; - pool3d_backward(context.device_context(), *in_x, *out, *out_grad, - ksize, strides, paddings, pool_process, in_x_grad); + pool3d_backward(dev_ctx, *in_x, *out, *out_grad, ksize, strides, + paddings, pool_process, in_x_grad); } } break; default: { PADDLE_THROW("Pool op only supports 2D and 3D input."); } diff --git a/paddle/operators/pool_with_index_op.cc b/paddle/operators/pool_with_index_op.cc index 1df36e965abab3549aeb88bf682b712033c4d79c..76c5123527c3ff5b7e6c7eec39f4eb1d612759d3 100644 --- a/paddle/operators/pool_with_index_op.cc +++ b/paddle/operators/pool_with_index_op.cc @@ -29,11 +29,11 @@ class MaxPoolWithIndexOp : public framework::OperatorWithKernel { void InferShape(framework::InferShapeContext *ctx) const override { PADDLE_ENFORCE(ctx->HasInput("X"), - "X(Input) of Pooling should not be null."); + "Input(X) of Pooling should not be null."); PADDLE_ENFORCE(ctx->HasOutput("Out"), - "Out(Output) of Pooling should not be null."); + "Output(Out) of Pooling should not be null."); PADDLE_ENFORCE(ctx->HasOutput("Mask"), - "Mask(Output) of Pooling should not be null."); + "Output(Mask) of Pooling should not be null."); auto in_x_dims = ctx->GetInputDim("X"); @@ -67,6 +67,14 @@ class MaxPoolWithIndexOp : public framework::OperatorWithKernel { ctx->SetOutputDim("Out", framework::make_ddim(output_shape)); ctx->SetOutputDim("Mask", framework::make_ddim(output_shape)); } + + protected: + framework::OpKernelType GetActualKernelType( + const framework::ExecutionContext &ctx) const override { + return framework::OpKernelType( + framework::ToDataType(ctx.Input("X")->type()), + ctx.device_context()); + } }; class MaxPoolWithIndexOpGrad : public framework::OperatorWithKernel { @@ -80,12 +88,19 @@ class MaxPoolWithIndexOpGrad : public framework::OperatorWithKernel { "Input(X@GRAD) should not be null."); ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X")); } + + protected: + framework::OpKernelType GetActualKernelType( + const framework::ExecutionContext &ctx) const override { + return framework::OpKernelType( + framework::ToDataType(ctx.Input("X")->type()), + ctx.device_context()); + } }; class MaxPool2dWithIndexOpMaker : public framework::OpProtoAndCheckerMaker { public: - MaxPool2dWithIndexOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + MaxPool2dWithIndexOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput( "X", @@ -116,7 +131,7 @@ class MaxPool2dWithIndexOpMaker : public framework::OpProtoAndCheckerMaker { // TypedAttrChecker don't support vector type.) AddAttr( "global_pooling", - "(bool, default false) Whether to use the global pooling. " + "(bool, default:false) Whether to use the global pooling. " "If global_pooling = true, ksize and paddings will be ignored.") .SetDefault(false); AddAttr>("strides", @@ -126,7 +141,7 @@ class MaxPool2dWithIndexOpMaker : public framework::OpProtoAndCheckerMaker { // TypedAttrChecker don't support vector type.) AddAttr>( "paddings", - "(vector, defalut {0, 0}), paddings(height, width) of pooling " + "(vector, default:{0, 0}), paddings(height, width) of pooling " "operator. " "If global_pooling = true, paddings and will be ignored.") .SetDefault({0, 0}); // TODO(Chengduo): Add checker. (Currently, @@ -150,10 +165,10 @@ Example: Output: Out shape: $(N, C, H_{out}, W_{out})$ Mask shape: $(N, C, H_{out}, W_{out})$ - where + Where $$ - H_{out} = (H_{in} - ksize[0] + 2 * paddings[0]) / strides[0] + 1 \\ - W_{out} = (W_{in} - ksize[1] + 2 * paddings[1]) / strides[1] + 1 + H_{out} = \frac{(H_{in} - ksize[0] + 2 * paddings[0])}{strides[0]} + 1 \\ + W_{out} = \frac{(W_{in} - ksize[1] + 2 * paddings[1])}{strides[1]} + 1 $$ )DOC"); @@ -162,8 +177,7 @@ Example: class MaxPool3dWithIndexOpMaker : public framework::OpProtoAndCheckerMaker { public: - MaxPool3dWithIndexOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + MaxPool3dWithIndexOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "(Tensor) The input tensor of pooling operator. " @@ -204,7 +218,7 @@ class MaxPool3dWithIndexOpMaker : public framework::OpProtoAndCheckerMaker { // TypedAttrChecker don't support vector type.) AddAttr>( "paddings", - "(vector, defalut {0,0,0}), paddings(depth, " + "(vector, default {0,0,0}), paddings(depth, " "height, width) of pooling operator. " "If global_pooling = true, paddings and ksize will be ignored.") .SetDefault({0, 0, 0}); // TODO(Chengduo): Add checker. (Currently, @@ -228,11 +242,11 @@ Example: Output: Out shape: $(N, C, D_{out}, H_{out}, W_{out})$ Mask shape: $(N, C, D_{out}, H_{out}, W_{out})$ - where + Where $$ - D_{out} = (D_{in} - ksize[0] + 2 * paddings[0]) / strides[0] + 1 \\ - H_{out} = (H_{in} - ksize[1] + 2 * paddings[1]) / strides[1] + 1 \\ - W_{out} = (W_{in} - ksize[2] + 2 * paddings[2]) / strides[2] + 1 + D_{out} = \frac{(D_{in} - ksize[0] + 2 * paddings[0])}{strides[0]} + 1 \\ + H_{out} = \frac{(H_{in} - ksize[1] + 2 * paddings[1])}{strides[1]} + 1 \\ + W_{out} = \frac{(W_{in} - ksize[2] + 2 * paddings[2])}{strides[2]} + 1 $$ )DOC"); @@ -250,10 +264,15 @@ REGISTER_OP(max_pool2d_with_index, ops::MaxPoolWithIndexOp, REGISTER_OP_CPU_KERNEL( max_pool2d_with_index, - ops::MaxPoolWithIndexKernel); + ops::MaxPoolWithIndexKernel, + ops::MaxPoolWithIndexKernel); REGISTER_OP_CPU_KERNEL( max_pool2d_with_index_grad, - ops::MaxPoolWithIndexGradKernel) + ops::MaxPoolWithIndexGradKernel, + ops::MaxPoolWithIndexGradKernel) REGISTER_OP(max_pool3d_with_index, ops::MaxPoolWithIndexOp, ops::MaxPool3dWithIndexOpMaker, max_pool3d_with_index_grad, @@ -261,7 +280,12 @@ REGISTER_OP(max_pool3d_with_index, ops::MaxPoolWithIndexOp, REGISTER_OP_CPU_KERNEL( max_pool3d_with_index, - ops::MaxPoolWithIndexKernel); + ops::MaxPoolWithIndexKernel, + ops::MaxPoolWithIndexKernel); REGISTER_OP_CPU_KERNEL( max_pool3d_with_index_grad, - ops::MaxPoolWithIndexGradKernel) + ops::MaxPoolWithIndexGradKernel, + ops::MaxPoolWithIndexGradKernel) diff --git a/paddle/operators/pool_with_index_op.cu.cc b/paddle/operators/pool_with_index_op.cu.cc index 287657d4b1c57f354ef050885f71261092bdc062..4c9804da639e3ad44f90963b53948cd8b755a6ac 100644 --- a/paddle/operators/pool_with_index_op.cu.cc +++ b/paddle/operators/pool_with_index_op.cu.cc @@ -16,16 +16,28 @@ limitations under the License. */ namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL( +REGISTER_OP_CUDA_KERNEL( max_pool2d_with_index, - ops::MaxPoolWithIndexKernel); -REGISTER_OP_GPU_KERNEL( + ops::MaxPoolWithIndexKernel, + ops::MaxPoolWithIndexKernel); +REGISTER_OP_CUDA_KERNEL( max_pool2d_with_index_grad, - ops::MaxPoolWithIndexGradKernel) + ops::MaxPoolWithIndexGradKernel, + ops::MaxPoolWithIndexGradKernel) -REGISTER_OP_GPU_KERNEL( +REGISTER_OP_CUDA_KERNEL( max_pool3d_with_index, - ops::MaxPoolWithIndexKernel); -REGISTER_OP_GPU_KERNEL( + ops::MaxPoolWithIndexKernel, + ops::MaxPoolWithIndexKernel); +REGISTER_OP_CUDA_KERNEL( max_pool3d_with_index_grad, - ops::MaxPoolWithIndexGradKernel) + ops::MaxPoolWithIndexGradKernel, + ops::MaxPoolWithIndexGradKernel) diff --git a/paddle/operators/pool_with_index_op.h b/paddle/operators/pool_with_index_op.h index a081607edce335f0265388ab01238d584bcf3ead..4f4087d1dd36d6e91cdd9a9253dd72a71735e136 100644 --- a/paddle/operators/pool_with_index_op.h +++ b/paddle/operators/pool_with_index_op.h @@ -24,8 +24,8 @@ namespace operators { using Tensor = framework::Tensor; -template -class MaxPoolWithIndexKernel : public framework::OpKernel { +template +class MaxPoolWithIndexKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { const Tensor* in_x = context.Input("X"); @@ -35,6 +35,8 @@ class MaxPoolWithIndexKernel : public framework::OpKernel { std::vector ksize = context.Attr>("ksize"); std::vector strides = context.Attr>("strides"); std::vector paddings = context.Attr>("paddings"); + + auto& dev_ctx = context.template device_context(); if (context.Attr("global_pooling")) { for (size_t i = 0; i < ksize.size(); ++i) { paddings[i] = 0; @@ -44,24 +46,24 @@ class MaxPoolWithIndexKernel : public framework::OpKernel { switch (ksize.size()) { case 2: { - paddle::operators::math::MaxPool2dWithIndexFunctor + paddle::operators::math::MaxPool2dWithIndexFunctor pool2d_forward; - pool2d_forward(context.device_context(), *in_x, ksize, strides, - paddings, out, mask); + pool2d_forward(dev_ctx, *in_x, ksize, strides, paddings, out, mask); } break; case 3: { - paddle::operators::math::MaxPool3dWithIndexFunctor + paddle::operators::math::MaxPool3dWithIndexFunctor pool3d_forward; - pool3d_forward(context.device_context(), *in_x, ksize, strides, - paddings, out, mask); + pool3d_forward(dev_ctx, *in_x, ksize, strides, paddings, out, mask); } break; default: { PADDLE_THROW("Pool op only supports 2D and 3D input."); } } } }; -template -class MaxPoolWithIndexGradKernel : public framework::OpKernel { +template +class MaxPoolWithIndexGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { const Tensor* mask = context.Input("Mask"); @@ -80,19 +82,21 @@ class MaxPoolWithIndexGradKernel : public framework::OpKernel { } if (in_x_grad) { - in_x_grad->mutable_data(context.GetPlace()); - auto& device_ctx = context.device_context(); + in_x_grad->mutable_data(context.GetPlace()); + auto& device_ctx = context.template device_context(); math::set_constant(device_ctx, in_x_grad, 0); switch (ksize.size()) { case 2: { - paddle::operators::math::MaxPool2dWithIndexGradFunctor + paddle::operators::math::MaxPool2dWithIndexGradFunctor pool2d_backward; pool2d_backward(device_ctx, *out_grad, *mask, ksize, strides, paddings, in_x_grad); } break; case 3: { - paddle::operators::math::MaxPool3dWithIndexGradFunctor + paddle::operators::math::MaxPool3dWithIndexGradFunctor pool3d_backward; pool3d_backward(device_ctx, *out_grad, *mask, ksize, strides, paddings, in_x_grad); diff --git a/paddle/operators/positive_negative_pair_op.cc b/paddle/operators/positive_negative_pair_op.cc index 4ba40a62ec5f696ad980c2913f7e162879a557e2..a6b23c995b8b9104f2da9d3d29ceb3eb88e7da63 100644 --- a/paddle/operators/positive_negative_pair_op.cc +++ b/paddle/operators/positive_negative_pair_op.cc @@ -85,7 +85,7 @@ class PositiveNegativePairOp : public framework::OperatorWithKernel { } protected: - framework::OpKernelType GetKernelType( + framework::OpKernelType GetActualKernelType( const framework::ExecutionContext &ctx) const override { return framework::OpKernelType( framework::ToDataType(ctx.Input("Score")->type()), @@ -95,8 +95,7 @@ class PositiveNegativePairOp : public framework::OperatorWithKernel { class PositiveNegativePairOpMaker : public framework::OpProtoAndCheckerMaker { public: - PositiveNegativePairOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + PositiveNegativePairOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("Score", "(Tensor, float) Model Score on an item (with " @@ -155,13 +154,14 @@ class PositiveNegativePairOpMaker : public framework::OpProtoAndCheckerMaker { "Noting that reducing on the first dim will make the LoD info lost.") .SetDefault(0); AddComment(R"DOC( - PositiveNegativePairOp can be used to evaluate Learning To Rank(LTR) - model performance. - Within some context, e.g. the "query", a LTR model generates scores - for a list of items, which gives a partial order of the items. - PositiveNegativePairOp takes a list of reference rank order - (Input("Label")) and the model generated scores (Input(Score)) as - inputs and counts the pairs that ranked correctly and incorrectly. +PositiveNegativePairOp can be used to evaluate Learning To Rank(LTR) model's +performance. + +Within some context, e.g. the "query", a LTR model generates scores for a list +of items, which gives a partial order of the items. PositiveNegativePairOp +takes a list of reference rank order (Input("Label")) and the model generated +scores (Input(Score)) as inputs and counts the pairs that ranked correctly +and incorrectly. )DOC"); } }; diff --git a/paddle/operators/positive_negative_pair_op.h b/paddle/operators/positive_negative_pair_op.h index 2efd3777e04c17b27c07bccde524de5785af35fe..977e59b7d2f771fc4c3412f0092f1eba92ef22da 100644 --- a/paddle/operators/positive_negative_pair_op.h +++ b/paddle/operators/positive_negative_pair_op.h @@ -22,7 +22,7 @@ namespace operators { using Tensor = framework::Tensor; using LoDTensor = framework::LoDTensor; -template +template class PositiveNegativePairKernel : public framework::OpKernel { public: struct PredictionResult { diff --git a/paddle/operators/precision_recall_op.cc b/paddle/operators/precision_recall_op.cc index 1ace4f2a5935dcb4239526c42599a42d288ff552..c5753147effd17c012683e1058e34af46288f366 100644 --- a/paddle/operators/precision_recall_op.cc +++ b/paddle/operators/precision_recall_op.cc @@ -80,7 +80,7 @@ class PrecisionRecallOp : public framework::OperatorWithKernel { } protected: - framework::OpKernelType GetKernelType( + framework::OpKernelType GetActualKernelType( const framework::ExecutionContext &ctx) const override { return framework::OpKernelType( framework::ToDataType(ctx.Input("MaxProbs")->type()), @@ -90,8 +90,7 @@ class PrecisionRecallOp : public framework::OperatorWithKernel { class PrecisionRecallOpMaker : public framework::OpProtoAndCheckerMaker { public: - PrecisionRecallOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + PrecisionRecallOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("MaxProbs", "(Tensor, default Tensor) A 2-D tensor with shape N x 1, " diff --git a/paddle/operators/precision_recall_op.h b/paddle/operators/precision_recall_op.h index 4a871ce6741469cf9af409ec90215f721d52f36c..c0d55405a362809f414b8dc3b12ed692f96c24e9 100644 --- a/paddle/operators/precision_recall_op.h +++ b/paddle/operators/precision_recall_op.h @@ -26,7 +26,7 @@ using EigenMatrix = framework::EigenMatrix; enum StateVariable { TP = 0, FP, TN, FN }; -template +template class PrecisionRecallKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { diff --git a/paddle/operators/prelu_op.cc b/paddle/operators/prelu_op.cc index 055c471b4561e5fd3c7a65c6f81d66cdce1a5578..ddc21a657024dcc800726475fa6242f8e6576ad1 100644 --- a/paddle/operators/prelu_op.cc +++ b/paddle/operators/prelu_op.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/operators/prelu_op.h" #include "paddle/operators/net_op.h" @@ -38,7 +38,7 @@ class PReluOp : public framework::OperatorWithKernel { class PReluOpMaker : public framework::OpProtoAndCheckerMaker { public: - PReluOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + PReluOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "The input tensor of prelu operator."); AddInput("Alpha", "The alpha weight of prelu operator."); @@ -85,7 +85,8 @@ namespace ops = paddle::operators; REGISTER_OP(prelu, ops::PReluOp, ops::PReluOpMaker, prelu_grad, ops::PReluGradOp); -REGISTER_OP_CPU_KERNEL(prelu, - ops::PReluKernel); -REGISTER_OP_CPU_KERNEL(prelu_grad, - ops::PReluGradKernel); +REGISTER_OP_CPU_KERNEL( + prelu, ops::PReluKernel); +REGISTER_OP_CPU_KERNEL( + prelu_grad, + ops::PReluGradKernel); diff --git a/paddle/operators/prelu_op.cu b/paddle/operators/prelu_op.cu index 9e391dabae735cc8a740b46b50d31d271f99b65d..1718bb5cd65f48eba391023e4374a30e405a164d 100644 --- a/paddle/operators/prelu_op.cu +++ b/paddle/operators/prelu_op.cu @@ -1,21 +1,22 @@ /* 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 +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 + 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. */ +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/operators/prelu_op.h" -REGISTER_OP_GPU_KERNEL( - prelu, paddle::operators::PReluKernel); -REGISTER_OP_GPU_KERNEL( - prelu_grad, - paddle::operators::PReluGradKernel); +REGISTER_OP_CUDA_KERNEL( + prelu, + paddle::operators::PReluKernel); +REGISTER_OP_CUDA_KERNEL(prelu_grad, + paddle::operators::PReluGradKernel< + paddle::platform::CUDADeviceContext, float>); diff --git a/paddle/operators/prelu_op.h b/paddle/operators/prelu_op.h index 5ad31c2203ae6c9bf6f48bb9ecf9a714597e7da8..56f9a553ec12d5bfa745af63ec0570ad30910628 100644 --- a/paddle/operators/prelu_op.h +++ b/paddle/operators/prelu_op.h @@ -39,7 +39,7 @@ class PReluFunctor { const T* alpha_; }; -template +template class PReluKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { @@ -54,9 +54,9 @@ class PReluKernel : public framework::OpKernel { int numel = x->numel(); - Transform trans; - trans(context.device_context(), x_ptr, x_ptr + numel, o_ptr, - PReluFunctor(alpha_ptr)); + Transform trans; + trans(context.template device_context(), x_ptr, + x_ptr + numel, o_ptr, PReluFunctor(alpha_ptr)); } }; @@ -76,7 +76,7 @@ class PReluGradFunctor { const T* alpha_; }; -template +template class PReluGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { @@ -92,9 +92,9 @@ class PReluGradKernel : public framework::OpKernel { const T* out_ptr = out->data(); int numel = dx->numel(); - Transform trans; - trans(context.device_context(), out_ptr, out_ptr + numel, dout_ptr, dx_ptr, - PReluGradFunctor(alpha_ptr)); + Transform trans; + trans(context.template device_context(), out_ptr, + out_ptr + numel, dout_ptr, dx_ptr, PReluGradFunctor(alpha_ptr)); // TODO(Zhuoyuan): add dalpha upgrade when GPU kernels ready } diff --git a/paddle/operators/proximal_adagrad_op.cc b/paddle/operators/proximal_adagrad_op.cc index 36e460103ab46bf6f1408840a0699793e2be134d..b92f46b5bd4e48a25f8c87873c5df53f1753b71b 100644 --- a/paddle/operators/proximal_adagrad_op.cc +++ b/paddle/operators/proximal_adagrad_op.cc @@ -59,8 +59,7 @@ class ProximalAdagradOp : public framework::OperatorWithKernel { class ProximalAdagradOpMaker : public framework::OpProtoAndCheckerMaker { public: - ProximalAdagradOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + ProximalAdagradOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("Param", "(Tensor, default Tensor) " @@ -114,4 +113,4 @@ REGISTER_OP_WITHOUT_GRADIENT(proximal_adagrad, ops::ProximalAdagradOp, ops::ProximalAdagradOpMaker); REGISTER_OP_CPU_KERNEL( proximal_adagrad, - ops::ProximalAdagradOpKernel); + ops::ProximalAdagradOpKernel); diff --git a/paddle/operators/proximal_adagrad_op.cu b/paddle/operators/proximal_adagrad_op.cu index d0ae0395184ae4f794565f2e28c57f960f0ccbeb..42a178f94b94c8e80ec8f9b5e6471b75878b65d1 100644 --- a/paddle/operators/proximal_adagrad_op.cu +++ b/paddle/operators/proximal_adagrad_op.cu @@ -15,6 +15,6 @@ specific language governing permissions and limitations under the License. */ #include "paddle/operators/proximal_adagrad_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL( +REGISTER_OP_CUDA_KERNEL( proximal_adagrad, - ops::ProximalAdagradOpKernel); + ops::ProximalAdagradOpKernel); diff --git a/paddle/operators/proximal_adagrad_op.h b/paddle/operators/proximal_adagrad_op.h index 7a1560e8cb339a306ab19513808aab165f82cc8a..523924d80e127d9ad2483e6b239fb948aa72200c 100644 --- a/paddle/operators/proximal_adagrad_op.h +++ b/paddle/operators/proximal_adagrad_op.h @@ -24,7 +24,7 @@ template using EigenVector = framework::EigenVector; -template +template class ProximalAdagradOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { @@ -45,20 +45,20 @@ class ProximalAdagradOpKernel : public framework::OpKernel { auto p_out = EigenVector::Flatten(*param_out); auto m_out = EigenVector::Flatten(*moment_out); - auto place = ctx.GetEigenDevice(); + auto* place = ctx.template device_context().eigen_device(); Eigen::DSizes grad_dsize(grad->numel()); - m_out.device(place) = m + g * g; + m_out.device(*place) = m + g * g; auto prox_param = p - lr.broadcast(grad_dsize) * g / m_out.sqrt(); if (l1 > static_cast(0)) { - p_out.device(place) = + p_out.device(*place) = prox_param.sign() * (((prox_param.abs() - (lr * l1).broadcast(grad_dsize)) .cwiseMax(static_cast(0.0))) / (static_cast(1.0) + (lr * l2).broadcast(grad_dsize))); } else { - p_out.device(place) = + p_out.device(*place) = prox_param / (static_cast(1.0) + (lr * l2).broadcast(grad_dsize)); } } diff --git a/paddle/operators/proximal_gd_op.cc b/paddle/operators/proximal_gd_op.cc index 5693d0ec9ebf4c470dfa5141b6eeee431f24f2ea..2d3bbdaf320a4d6bdf18ec92230a81ad98371498 100644 --- a/paddle/operators/proximal_gd_op.cc +++ b/paddle/operators/proximal_gd_op.cc @@ -47,8 +47,7 @@ class ProximalGDOp : public framework::OperatorWithKernel { class ProximalGDOpMaker : public framework::OpProtoAndCheckerMaker { public: - ProximalGDOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + ProximalGDOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("Param", "(Tensor, default Tensor) " @@ -94,4 +93,5 @@ namespace ops = paddle::operators; REGISTER_OP_WITHOUT_GRADIENT(proximal_gd, ops::ProximalGDOp, ops::ProximalGDOpMaker); REGISTER_OP_CPU_KERNEL( - proximal_gd, ops::ProximalGDOpKernel); + proximal_gd, + ops::ProximalGDOpKernel); diff --git a/paddle/operators/proximal_gd_op.cu b/paddle/operators/proximal_gd_op.cu index 26f4ebaa0f43620fee7ece2d71755be94a0e01a5..b7dd840d19a13cd3329fb68563693a80d22291ca 100644 --- a/paddle/operators/proximal_gd_op.cu +++ b/paddle/operators/proximal_gd_op.cu @@ -15,5 +15,6 @@ specific language governing permissions and limitations under the License. */ #include "paddle/operators/proximal_gd_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL( - proximal_gd, ops::ProximalGDOpKernel); +REGISTER_OP_CUDA_KERNEL( + proximal_gd, + ops::ProximalGDOpKernel); diff --git a/paddle/operators/proximal_gd_op.h b/paddle/operators/proximal_gd_op.h index bebda0204173ec5c3ec9a7a9da6fb623171f4cea..64648b3ccaf9615c995d65464607105d87c04198 100644 --- a/paddle/operators/proximal_gd_op.h +++ b/paddle/operators/proximal_gd_op.h @@ -24,7 +24,7 @@ template using EigenVector = framework::EigenVector; -template +template class ProximalGDOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { @@ -42,7 +42,7 @@ class ProximalGDOpKernel : public framework::OpKernel { auto lr = EigenVector::Flatten(*ctx.Input("LearningRate")); auto p_out = EigenVector::Flatten(*param_out); - auto place = ctx.GetEigenDevice(); + auto& place = *ctx.template device_context().eigen_device(); Eigen::DSizes grad_dsize(grad->numel()); diff --git a/paddle/operators/rank_loss_op.cc b/paddle/operators/rank_loss_op.cc index 061e82412ea5f4f17fd26a7094e68b97138cc09c..f2164a0f80519ed9c2490ab3aa6809dc84c6070d 100644 --- a/paddle/operators/rank_loss_op.cc +++ b/paddle/operators/rank_loss_op.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/operators/rank_loss_op.h" @@ -35,23 +35,30 @@ class RankLossOp : public framework::OperatorWithKernel { auto right_dims = ctx->GetInputDim("Right"); PADDLE_ENFORCE((label_dims == left_dims) && (left_dims == right_dims), - "All inputs must have the same size"); - PADDLE_ENFORCE((label_dims.size() == 2) && (label_dims[1] == 1), - "All inputs must be row vector with size batch_size x 1."); + "All inputs must have the same size."); + PADDLE_ENFORCE( + (label_dims.size() == 2) && (label_dims[1] == 1), + "All inputs must be 2-D tensors with shape [batch_size x 1]."); ctx->SetOutputDim("Out", label_dims); } }; class RankLossOpMaker : public framework::OpProtoAndCheckerMaker { public: - RankLossOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + RankLossOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("Label", - "The label indicating A ranked higher than B or not, row vector."); - AddInput("Left", "The output of RankNet for doc A, vector."); - AddInput("Right", "The output of RankNet for doc B, vetor."); - AddOutput("Out", "The output loss of RankLoss operator, vector."); + "(2-D Tensor with shape [batch_size x 1]) " + "The label indicating A ranked higher than B or not."); + AddInput("Left", + "(2-D Tensor with shape [batch_size x 1]) " + "The output of RankNet for doc A."); + AddInput("Right", + "(2-D Tensor with shape [batch_size x 1]) " + "The output of RankNet for doc B."); + AddOutput("Out", + "(2-D Tensor with shape [batch_size x 1]) " + "The output loss of RankLoss operator."); AddComment(R"DOC( RankLoss Operator. @@ -65,16 +72,17 @@ P = {0, 1} or {0, 0.5, 1}, where 0.5 means no information about the rank of the input pair. The RankLoss operator takes three inputs: Left (o_i), Right (o_j) and Label -(P_{i,j}), which represent the output of RankNet for the two docs and the label, -respectively, and yields the rank loss C_{i,j} using the following equation: +(P_{i,j}), which represent the output score of RankNet for the two docs and +the label respectively, and yields the rank loss C_{i,j} using the following +equation: -\f$$ - C_{i,j} = -\tilde{P_{ij}} * o_{i,j} + log(1 + e^{o_{i,j}}) \\ +$$ + C_{i,j} = -\tilde{P_{ij}} * o_{i,j} + \log(1 + e^{o_{i,j}}) \\ o_{i,j} = o_i - o_j \\ \tilde{P_{i,j}} = \left \{0, 0.5, 1 \right \} \ or \ \left \{0, 1 \right \} -\f$$ +$$ -The operator can take inputs of one sample or in batch. +The operator can take batch inputs with size batch_size (batch_size >= 1). )DOC"); } @@ -114,7 +122,8 @@ namespace ops = paddle::operators; REGISTER_OP(rank_loss, ops::RankLossOp, ops::RankLossOpMaker, rank_loss_grad, ops::RankLossGradOp); -REGISTER_OP_CPU_KERNEL(rank_loss, - ops::RankLossKernel); REGISTER_OP_CPU_KERNEL( - rank_loss_grad, ops::RankLossGradKernel); + rank_loss, ops::RankLossKernel); +REGISTER_OP_CPU_KERNEL( + rank_loss_grad, + ops::RankLossGradKernel); diff --git a/paddle/operators/rank_loss_op.cu b/paddle/operators/rank_loss_op.cu index 779588ff36c792b8925a535d60f1cfbbe3c66d86..294b22738347b17dd67df05291ac496bfb608323 100644 --- a/paddle/operators/rank_loss_op.cu +++ b/paddle/operators/rank_loss_op.cu @@ -1,22 +1,22 @@ /* 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 +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 + 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. */ +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/operators/rank_loss_op.h" -REGISTER_OP_GPU_KERNEL( - rank_loss, - paddle::operators::RankLossKernel); -REGISTER_OP_GPU_KERNEL( - rank_loss_grad, - paddle::operators::RankLossGradKernel); +REGISTER_OP_CUDA_KERNEL(rank_loss, + paddle::operators::RankLossKernel< + paddle::platform::CUDADeviceContext, float>); +REGISTER_OP_CUDA_KERNEL(rank_loss_grad, + paddle::operators::RankLossGradKernel< + paddle::platform::CUDADeviceContext, float>); diff --git a/paddle/operators/rank_loss_op.h b/paddle/operators/rank_loss_op.h index f184d6efcb496a1d7f38540712b6c431f816482e..bd0c49ca6e42bcb1a25c53421c0e672cecbb3a15 100644 --- a/paddle/operators/rank_loss_op.h +++ b/paddle/operators/rank_loss_op.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 @@ -20,7 +20,7 @@ namespace paddle { namespace operators { -template +template class RankLossKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const { @@ -35,13 +35,13 @@ class RankLossKernel : public framework::OpKernel { auto left = framework::EigenVector::Flatten(*left_t); auto right = framework::EigenVector::Flatten(*right_t); - auto& dev = ctx.GetEigenDevice(); + auto& dev = *ctx.template device_context().eigen_device(); out.device(dev) = (1. + (left - right).exp()).log() - label * (left - right); } }; -template +template class RankLossGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const { @@ -55,7 +55,7 @@ class RankLossGradKernel : public framework::OpKernel { auto* left_t = ctx.Input("Left"); auto* right_t = ctx.Input("Right"); - auto& dev = ctx.GetEigenDevice(); + auto& dev = *ctx.template device_context().eigen_device(); auto d_out = framework::EigenVector::Flatten(*d_out_t); auto label = framework::EigenVector::Flatten(*label_t); auto left = framework::EigenVector::Flatten(*left_t); diff --git a/paddle/operators/recurrent_op.cc b/paddle/operators/recurrent_op.cc index 0075ccd24271bf83f139e121efad00c2316cc11b..056fa46949cd623845956521b068109085a8795e 100644 --- a/paddle/operators/recurrent_op.cc +++ b/paddle/operators/recurrent_op.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/framework/executor.h" @@ -25,7 +25,7 @@ constexpr char kOutputs[] = "outputs"; constexpr char kStepScopes[] = "step_scopes"; constexpr char kExStates[] = "ex_states"; constexpr char kStates[] = "states"; -constexpr char kStepBlock[] = "step_block"; +constexpr char kStepBlock[] = "sub_block"; constexpr char kReverse[] = "reverse"; constexpr char kIsTrain[] = "is_train"; #define GRAD_SUFFIX "@GRAD" @@ -227,14 +227,15 @@ class RecurrentOp : public RecurrentBase { : RecurrentBase(type, inputs, outputs, attrs) {} void Run(const framework::Scope &scope, - const platform::DeviceContext &dev_ctx) const override { + const platform::Place &place) const override { auto seq_len = static_cast(this->GetSequenceLength(scope)); VLOG(3) << "Static RNN input sequence length = " << seq_len; StepScopes scopes = CreateStepScopes(scope, seq_len); auto reverse = Attr(kReverse); - framework::Executor executor(dev_ctx); - auto *block = Attr(kStepBlock); + framework::Executor executor(place); + auto *block = Attr(kStepBlock); + auto *program = block->Program(); for (size_t i = 0; i < seq_len; ++i) { @@ -270,6 +271,11 @@ class RecurrentOp : public RecurrentBase { executor.Run(*program, &cur_scope, block->ID(), false /*create_local_scope*/); + // get device context from pool + platform::DeviceContextPool &pool = + platform::DeviceContextPool::Instance(); + auto &dev_ctx = *pool.Get(place); + // Copy inside::output -> outside::output // outside::output[seq_offset: seq_offset + 1] = inside::output this->LinkTensorWithCallback( @@ -278,13 +284,13 @@ class RecurrentOp : public RecurrentBase { framework::LoDTensor *dst_tensor) { if (i == 0) { // create output tensor at begin dst_tensor->Resize(PrependDims(seq_len, src_tensor.dims())); - dst_tensor->mutable_data(dev_ctx.GetPlace(), src_tensor.type()); + dst_tensor->mutable_data(place, src_tensor.type()); } auto dst_out = dst_tensor->Slice(seq_offset, seq_offset + 1); // Explicit copy output since the local RNN scope can be destroyed // early. - dst_out.CopyFrom(src_tensor, dev_ctx.GetPlace(), dev_ctx); + framework::CopyFrom(src_tensor, place, dev_ctx, &dst_out); }); scopes.Next(); @@ -310,15 +316,20 @@ class RecurrentGradOp : public RecurrentBase { : RecurrentBase(type, inputs, outputs, attrs) {} void Run(const framework::Scope &scope, - const platform::DeviceContext &dev_ctx) const override { + const platform::Place &place) const override { auto seq_len = static_cast(GetSequenceLength(scope)); StepScopes scopes = CreateStepScopes(scope, seq_len); auto reverse = Attr(kReverse); - framework::Executor executor(dev_ctx); - auto *block = Attr(kStepBlock); + framework::Executor executor(place); + auto *block = Attr(kStepBlock); + auto *program = block->Program(); + // get device context from pool + platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance(); + auto &dev_ctx = *pool.Get(place); + for (size_t step_id = 0; step_id < seq_len; ++step_id) { size_t seq_offset = reverse ? step_id : seq_len - step_id - 1; VLOG(3) << "Recurrent backward operate at the time step " << seq_offset; @@ -365,7 +376,7 @@ class RecurrentGradOp : public RecurrentBase { auto *cur_grad_var = cur_scope.Var(cur_grad); auto cur_grad_tensor = cur_grad_var->GetMutable(); - cur_grad_tensor->CopyFrom(ex_tensor, dev_ctx.GetPlace(), dev_ctx); + framework::CopyFrom(ex_tensor, place, dev_ctx, cur_grad_tensor); } } @@ -401,13 +412,14 @@ class RecurrentGradOp : public RecurrentBase { auto &inside_tensor = cur_scope.FindVar(inside_grad_name) ->Get(); framework::AttributeMap attrs; - attrs["data_type"] = framework::ToDataType(inside_tensor.type()); + attrs["dtype"] = framework::ToDataType(inside_tensor.type()); attrs["shape"] = framework::vectorize2int(inside_tensor.dims()); attrs["value"] = 0.0f; auto zero_op = framework::OpRegistry::CreateOp( - "fill_constant", {}, {{"Out", {pg_names[param_id]}}}, attrs); - zero_op->Run(scope, dev_ctx); + "fill_constant", framework::VariableNameMap{}, + {{"Out", {pg_names[param_id]}}}, attrs); + zero_op->Run(scope, place); } auto new_inside_name = cur_scope.Rename(inside_grad_name); @@ -415,8 +427,8 @@ class RecurrentGradOp : public RecurrentBase { auto sum_op = framework::OpRegistry::CreateOp( "sum", {{"X", {pg_names[param_id], new_inside_name}}}, - {{"Out", {pg_names[param_id]}}}, {}); - sum_op->Run(cur_scope, dev_ctx); + {{"Out", {pg_names[param_id]}}}, framework::AttributeMap{}); + sum_op->Run(cur_scope, place); cur_scope.Rename(new_inside_name, inside_grad_name); } @@ -434,11 +446,11 @@ class RecurrentGradOp : public RecurrentBase { } if (step_id == 0) { // alloc memory outside->Resize(PrependDims(seq_len, inside.dims())); - outside->mutable_data(dev_ctx.GetPlace(), inside.type()); + outside->mutable_data(place, inside.type()); } auto dst = outside->Slice(seq_offset, seq_offset + 1); - dst.CopyFrom(inside, dev_ctx.GetPlace(), dev_ctx); + framework::CopyFrom(inside, place, dev_ctx, &dst); }); VLOG(5) << "Link outside gradient finished "; @@ -450,8 +462,8 @@ class RecurrentGradOp : public RecurrentBase { [&](const framework::LoDTensor &inside, framework::LoDTensor *outside) { outside->Resize(inside.dims()); - outside->mutable_data(dev_ctx.GetPlace(), inside.type()); - outside->CopyFrom(inside, dev_ctx.GetPlace(), dev_ctx); + outside->mutable_data(place, inside.type()); + framework::CopyFrom(inside, place, dev_ctx, outside); }); VLOG(5) << "Link initialize state gradient finished "; } @@ -480,7 +492,7 @@ class RecurrentGradOp : public RecurrentBase { std::unordered_set LocalVarNames( const framework::Scope &scope) const { - return this->List2Set(scope.GetAllNames(false)); + return this->List2Set(scope.LocalVarNames()); } static std::vector GradVarLists( const std::vector &var_names) { @@ -494,8 +506,7 @@ class RecurrentGradOp : public RecurrentBase { class RecurrentOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - RecurrentOpProtoMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + RecurrentOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput(kInputs, "rnn inputs").AsDuplicable(); AddInput(kInitialStates, "rnn initial states").AsDuplicable(); @@ -520,8 +531,7 @@ The ex-state means the state value in the ex-timestep or the previous time step string::Sprintf( "The state variable names. [%s, %s, %s] must be the same order", kExStates, kStates, kInitStateGrads)); - AddAttr(kStepBlock, - "The step block inside RNN"); + AddAttr(kStepBlock, "The step block inside RNN"); AddAttr(kReverse, R"DOC(Calculate RNN reversely or not. By default reverse=False @@ -563,13 +573,13 @@ class RecurrentGradOpDescMaker : public framework::SingleGradOpDescMaker { using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; protected: - virtual std::unique_ptr Apply() const { - auto *grad = new framework::OpDescBind(); + virtual std::unique_ptr Apply() const { + auto *grad = new framework::OpDesc(); grad->SetType("recurrent_grad"); for (auto &input_param : this->InputNames()) { grad->SetInput(input_param, this->Input(input_param)); grad->SetOutput(framework::GradVarName(input_param), - this->InputGrad(input_param)); + this->InputGrad(input_param, false)); } for (auto &output_param : this->OutputNames()) { @@ -586,7 +596,7 @@ class RecurrentGradOpDescMaker : public framework::SingleGradOpDescMaker { grad->SetAttrMap(this->Attrs()); grad->SetBlockAttr(kStepBlock, *grad_block_[0]); - return std::unique_ptr(grad); + return std::unique_ptr(grad); } }; @@ -597,7 +607,9 @@ class RecurrentGradOpShapeInference : public framework::InferShapeBase { std::vector output{kOutputs}; for (auto &s : input) { PADDLE_ENFORCE(ctx->HasInputs(s)); - PADDLE_ENFORCE(ctx->HasOutputs(framework::GradVarName(s))); + PADDLE_ENFORCE(ctx->HasOutputs(framework::GradVarName(s)), + "Cannot find the gradient variable %s", + framework::GradVarName(s)); } for (auto &s : output) { PADDLE_ENFORCE(ctx->HasInputs(s)); diff --git a/paddle/operators/recv_op.cc b/paddle/operators/recv_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..322f8571cfd4341f064e8f9df512a8d74b91ed9d --- /dev/null +++ b/paddle/operators/recv_op.cc @@ -0,0 +1,198 @@ +/* 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. */ + +#include +#include +#include +#include + +#include + +#include "paddle/framework/data_type.h" +#include "paddle/framework/executor.h" +#include "paddle/framework/framework.pb.h" +#include "paddle/framework/lod_tensor.h" +#include "paddle/framework/op_registry.h" +#include "paddle/framework/proto_desc.h" +#include "paddle/operators/detail/send_recv_impl.h" +#include "paddle/operators/detail/simple_block_queue.h" + +#define LISTEN_TERMINATE_MESSAGE "TERMINATE@RECV" + +namespace paddle { +namespace operators { + +void RunServer(Server **rpc_server, + std::shared_ptr service, + const std::string &server_address) { + ServerBuilder builder; + builder.AddListeningPort(server_address, grpc::InsecureServerCredentials()); + builder.RegisterService(service.get()); + std::unique_ptr server(builder.BuildAndStart()); + *rpc_server = server.get(); + LOG(INFO) << "Server listening on " << server_address; + server->Wait(); +} + +class RecvOp : public framework::OperatorBase { + public: + RecvOp(const std::string &type, const framework::VariableNameMap &inputs, + const framework::VariableNameMap &outputs, + const framework::AttributeMap &attrs) + : OperatorBase(type, inputs, outputs, attrs) { + if (!rpc_service_) { + rpc_service_.reset(new detail::SendRecvServerImpl()); + std::string endpoint = Attr("endpoint"); + server_thread_.reset( + new std::thread(RunServer, &rpc_server_, rpc_service_, endpoint)); + } + } + + void Stop() override { + detail::TensorWithName term_msg; + term_msg.first = LISTEN_TERMINATE_MESSAGE; + rpc_service_->Push(term_msg); + rpc_server_->Shutdown(); + server_thread_->join(); + } + + std::string GetGradVarNameForTrainer(const std::string &varname) const { + if (grads_counter_.find(varname) == grads_counter_.end()) { + grads_counter_[varname] = 0; + } + char ret[256]; + snprintf(ret, sizeof(ret), "%s.trainer_%d", varname.c_str(), + grads_counter_[varname]++); + return std::string(ret); + } + + void Run(const framework::Scope &scope, + const platform::Place &dev_place) const override { + // FIXME(typhoonzero): no new scopes for every run. + framework::Scope &recv_scope = scope.NewScope(); + rpc_service_->SetScope(&recv_scope); + auto param_list = Attr>("ParamList"); + auto grad_list = Attr>("GradList"); + auto trainer_count = Attr("Trainers"); + size_t param_count = param_list.size(); + rpc_service_->Reset(); + // TODO(typhoonzero): change this to a while_op for every cluster-batch. + bool exit_flag = false; + while (!exit_flag) { + // Get from multiple trainers, we don't care about order in which + // the gradient arrives, just add suffix 0~n then average the gradient. + for (size_t i = 0; i < param_count * trainer_count; ++i) { + // blocking get one var from client. + const detail::TensorWithName &v = rpc_service_->Get(); + auto grad_var_name = v.first; + if (grad_var_name == LISTEN_TERMINATE_MESSAGE) { + exit_flag = true; + break; + } + auto it = std::find(grad_list.begin(), grad_list.end(), grad_var_name); + std::string param_var_name; + if (it != grad_list.end()) { + param_var_name = param_list[it - grad_list.begin()]; + } else { + LOG(ERROR) << "grad have no paired param found!"; + } + VLOG(3) << "recved grad: " << grad_var_name + << " updating param: " << param_var_name; + auto *merged_grad = recv_scope.FindVar(grad_var_name); + if (merged_grad == nullptr) { + // create output of merged var. + auto merged_var = recv_scope.Var(grad_var_name); + merged_var->GetMutable(); + } + + if (trainer_count > 1) { + grad_var_name = this->GetGradVarNameForTrainer(grad_var_name); + } + + auto *var = recv_scope.Var(grad_var_name); + auto *tensor = var->GetMutable(); + // FIXME(typhoonzero): do not copy + platform::DeviceContextPool &pool = platform::DeviceContextPool::Get(); + auto &dev_ctx = *pool.Borrow(dev_place); + framework::CopyFrom(v.second, dev_place, dev_ctx, tensor); + } + if (exit_flag) { + break; + } + rpc_service_->Reset(); + + std::string program_str = Attr("OptimizeProgram"); + framework::proto::ProgramDesc program_desc; + program_desc.ParseFromString(program_str); + framework::ProgramDesc program(program_desc); + framework::Executor executor(dev_place); + // Run sub graph to get optimized tensor + try { + executor.Run(program, &recv_scope, 0, /*global_block*/ + false /*create_local_scope*/, false /*create_vars*/); + } catch (std::exception &e) { + LOG(ERROR) << "run sub program error " << e.what(); + } + rpc_service_->Done(); + grads_counter_.clear(); + } // while(true) + } + + protected: + // grpc server instance to track status and gracefully shutdown. + // borrow an pointer from server thread. + Server *rpc_server_{nullptr}; + // grpc send/recv service implement to register. + std::shared_ptr rpc_service_; + std::shared_ptr server_thread_; + mutable std::unordered_map grads_counter_; +}; + +class RecvOpMaker : public framework::OpProtoAndCheckerMaker { + public: + RecvOpMaker(OpProto *proto, OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("RX", "(Tensor) Input tensor to be optimized").AsDuplicable(); + AddComment(R"DOC( +Recv operator + +This operator will recv tensor from send_op +)DOC"); + AddAttr("endpoint", + "(string, default 127.0.0.1:6164)" + "IP address to listen on.") + .SetDefault("127.0.0.1:6164") + .AddCustomChecker([](const std::string &ip) { return !ip.empty(); }); + AddAttr("OptimizeProgram", "type string", + "Serialized ProgramDesc string for recv to run."); + AddAttr>( + "ParamList", "type list of string", + "grad->param name mapping to find which param to optimize.") + .SetDefault({}); + AddAttr>( + "GradList", "type list of string", + "grad->param name mapping to find which param to optimize.") + .SetDefault({}); + AddAttr("Trainers", "type int", + "Number of trainers in the current cluster job") + .SetDefault(1); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; + +REGISTER_OPERATOR(recv, ops::RecvOp, ops::RecvOpMaker); diff --git a/paddle/operators/reduce_op.cc b/paddle/operators/reduce_op.cc index 2589a54cfc7fc5bc11ae983797d480a134e0eb25..a3ff4a6ca0ef30be42e7801386a3561930638a8a 100644 --- a/paddle/operators/reduce_op.cc +++ b/paddle/operators/reduce_op.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/operators/reduce_op.h" #include "paddle/operators/net_op.h" @@ -37,18 +37,23 @@ class ReduceOp : public framework::OperatorWithKernel { PADDLE_ENFORCE_LT( dim, x_rank, "The dim should be in the range [-rank(input), rank(input))."); - bool keep_dim = ctx->Attrs().Get("keep_dim"); - auto dims_vector = vectorize(x_dims); - if (keep_dim || x_rank == 1) { - dims_vector[dim] = 1; + bool reduce_all = ctx->Attrs().Get("reduce_all"); + if (reduce_all) { + ctx->SetOutputDim("Out", {1}); } else { - dims_vector.erase(dims_vector.begin() + dim); - } - auto out_dims = framework::make_ddim(dims_vector); - ctx->SetOutputDim("Out", out_dims); - if (dim != 0) { - // Only pass LoD when not reducing on the first dim. - ctx->ShareLoD("X", /*->*/ "Out"); + bool keep_dim = ctx->Attrs().Get("keep_dim"); + auto dims_vector = vectorize(x_dims); + if (keep_dim || x_rank == 1) { + dims_vector[dim] = 1; + } else { + dims_vector.erase(dims_vector.begin() + dim); + } + auto out_dims = framework::make_ddim(dims_vector); + ctx->SetOutputDim("Out", out_dims); + if (dim != 0) { + // Only pass LoD when not reducing on the first dim. + ctx->ShareLoD("X", /*->*/ "Out"); + } } } }; @@ -78,7 +83,7 @@ class ReduceGradOp : public framework::OperatorWithKernel { class ReduceOpMaker : public framework::OpProtoAndCheckerMaker { public: - ReduceOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + ReduceOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "(Tensor) The input tensor. Tensors with rank at most 6 are " @@ -95,11 +100,16 @@ class ReduceOpMaker : public framework::OpProtoAndCheckerMaker { "(bool, default false) " "If true, retain the reduced dimension with length 1.") .SetDefault(false); + AddAttr("reduce_all", + "(bool, default false) " + "If true, output a scalar reduced along all dimensions.") + .SetDefault(false); comment_ = R"DOC( {ReduceOp} Operator. This operator computes the {reduce} 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_); @@ -125,8 +135,7 @@ The result tensor has 1 fewer dimension than the input unless keep_dim is true. class ReduceSumOpMaker : public ReduceOpMaker { public: - ReduceSumOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + ReduceSumOpMaker(OpProto *proto, OpAttrChecker *op_checker) : ReduceOpMaker(proto, op_checker) { SetComment("ReduceSum", "sum"); AddComment(comment_); @@ -135,8 +144,7 @@ class ReduceSumOpMaker : public ReduceOpMaker { class ReduceMeanOpMaker : public ReduceOpMaker { public: - ReduceMeanOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + ReduceMeanOpMaker(OpProto *proto, OpAttrChecker *op_checker) : ReduceOpMaker(proto, op_checker) { SetComment("ReduceMean", "mean"); AddComment(comment_); @@ -145,8 +153,7 @@ class ReduceMeanOpMaker : public ReduceOpMaker { class ReduceMaxOpMaker : public ReduceOpMaker { public: - ReduceMaxOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + ReduceMaxOpMaker(OpProto *proto, OpAttrChecker *op_checker) : ReduceOpMaker(proto, op_checker) { SetComment("ReduceMax", "max"); AddComment(comment_); @@ -155,8 +162,7 @@ class ReduceMaxOpMaker : public ReduceOpMaker { class ReduceMinOpMaker : public ReduceOpMaker { public: - ReduceMinOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + ReduceMinOpMaker(OpProto *proto, OpAttrChecker *op_checker) : ReduceOpMaker(proto, op_checker) { SetComment("ReduceMin", "min"); AddComment(comment_); @@ -180,12 +186,13 @@ REGISTER_OP(reduce_max, ops::ReduceOp, ops::ReduceMaxOpMaker, reduce_max_grad, REGISTER_OP(reduce_min, ops::ReduceOp, ops::ReduceMinOpMaker, reduce_min_grad, ops::ReduceGradOp); -#define REGISTER_REDUCE_CPU_KERNEL(reduce_type, functor, grad_functor) \ - REGISTER_OP_CPU_KERNEL( \ - reduce_type, \ - ops::ReduceKernel); \ - REGISTER_OP_CPU_KERNEL(reduce_type##_grad, \ - ops::ReduceGradKernel); +#define REGISTER_REDUCE_CPU_KERNEL(reduce_type, functor, grad_functor) \ + REGISTER_OP_CPU_KERNEL(reduce_type, \ + ops::ReduceKernel); \ + REGISTER_OP_CPU_KERNEL( \ + reduce_type##_grad, \ + ops::ReduceGradKernel); FOR_EACH_KERNEL_FUNCTOR(REGISTER_REDUCE_CPU_KERNEL); diff --git a/paddle/operators/reduce_op.cu b/paddle/operators/reduce_op.cu index d306e1a24096d737438d71d4d4abc35328d160cb..1dd948ed8a79cce8468f2fe210b5636e7dd1f99e 100644 --- a/paddle/operators/reduce_op.cu +++ b/paddle/operators/reduce_op.cu @@ -1,28 +1,29 @@ /* 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 +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 + 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. */ +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. */ #define EIGEN_USE_GPU #include "paddle/operators/reduce_op.h" namespace ops = paddle::operators; -#define REGISTER_REDUCE_GPU_KERNEL(reduce_type, functor, grad_functor) \ - REGISTER_OP_GPU_KERNEL( \ - reduce_type, \ - ops::ReduceKernel); \ - REGISTER_OP_GPU_KERNEL(reduce_type##_grad, \ - ops::ReduceGradKernel); +#define REGISTER_REDUCE_GPU_KERNEL(reduce_type, functor, grad_functor) \ + REGISTER_OP_CUDA_KERNEL( \ + reduce_type, ops::ReduceKernel); \ + REGISTER_OP_CUDA_KERNEL( \ + reduce_type##_grad, \ + ops::ReduceGradKernel); FOR_EACH_KERNEL_FUNCTOR(REGISTER_REDUCE_GPU_KERNEL); diff --git a/paddle/operators/reduce_op.h b/paddle/operators/reduce_op.h index dd6547542d16b0fe336184a0c09a8498027db6ea..da5f3977769990a45c94db21f5dbd01ac70ac06e 100644 --- a/paddle/operators/reduce_op.h +++ b/paddle/operators/reduce_op.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 @@ -26,61 +26,63 @@ using DDim = framework::DDim; template using EigenTensor = framework::EigenTensor; - template using EigenScalar = framework::EigenScalar; +template +using EigenVector = framework::EigenVector; struct SumFunctor { - template - void operator()(const Place& place, X& x, Y& y, const Dim& dim) { + template + void operator()(const DeviceContext& place, X& x, Y& y, const Dim& dim) { y.device(place) = x.sum(dim); } }; struct SumGradFunctor { - template - void operator()(const Place& place, X& x, Y& y, DX& dx, DY& dy, + template + void operator()(const DeviceContext& place, X& x, Y& y, DX& dx, DY& dy, const Dim& dim, int size) { dx.device(place) = dy.broadcast(dim); } }; struct MeanFunctor { - template - void operator()(const Place& place, X& x, Y& y, const Dim& dim) { + template + void operator()(const DeviceContext& place, X& x, Y& y, const Dim& dim) { y.device(place) = x.mean(dim); } }; struct MeanGradFunctor { - template - void operator()(const Place& place, X& x, Y& y, DX& dx, DY& dy, + template + void operator()(const DeviceContext& place, X& x, Y& y, DX& dx, DY& dy, const Dim& dim, int size) { dx.device(place) = dy.broadcast(dim) / dx.constant(size); } }; struct MaxFunctor { - template - void operator()(const Place& place, X& x, Y& y, const Dim& dim) { + template + void operator()(const DeviceContext& place, X& x, Y& y, const Dim& dim) { y.device(place) = x.maximum(dim); } }; struct MinFunctor { - template - void operator()(const Place& place, X& x, Y& y, const Dim& dim) { + template + void operator()(const DeviceContext& place, X& x, Y& y, const Dim& dim) { y.device(place) = x.minimum(dim); } }; struct MaxOrMinGradFunctor { - template - void operator()(const Place& place, X& x, Y& y, DX& dx, DY& dy, + template + void operator()(const DeviceContext& place, X& x, Y& y, DX& dx, DY& dy, const Dim& dim, int size) { auto equals = x == y.broadcast(dim); auto ones = dx.constant(1); @@ -91,30 +93,45 @@ struct MaxOrMinGradFunctor { } }; -template +template class ReduceKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { - 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; + bool reduce_all = context.Attr("reduce_all"); + if (reduce_all) { + // Flatten and reduce 1-D tensor + auto* input = context.Input("X"); + auto* output = context.Output("Out"); + output->mutable_data(context.GetPlace()); + auto x = EigenVector::Flatten(*input); + auto out = EigenScalar::From(*output); + auto& place = + *context.template device_context().eigen_device(); + auto reduce_dim = Eigen::array({{0}}); + 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; + } } } @@ -139,7 +156,8 @@ class ReduceKernel : public framework::OpKernel { dims = framework::make_ddim(dims_vector); } - auto& place = context.GetEigenDevice(); + auto& place = + *context.template device_context().eigen_device(); Functor functor; if (D == 1) { @@ -152,30 +170,50 @@ class ReduceKernel : public framework::OpKernel { } }; -template +template class ReduceGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { - int rank = context.Input("X")->dims().size(); - switch (rank) { - case 1: - ReduceGradCompute<1>(context); - break; - case 2: - ReduceGradCompute<2>(context); - break; - case 3: - ReduceGradCompute<3>(context); - break; - case 4: - ReduceGradCompute<4>(context); - break; - case 5: - ReduceGradCompute<5>(context); - break; - case 6: - ReduceGradCompute<6>(context); - break; + bool reduce_all = context.Attr("reduce_all"); + if (reduce_all) { + auto* input0 = context.Input("X"); + auto* input1 = context.Input("Out"); + auto* input2 = context.Input(framework::GradVarName("Out")); + auto* output = context.Output(framework::GradVarName("X")); + output->mutable_data(context.GetPlace()); + auto x = EigenVector::Flatten(*input0); + auto x_reduce = EigenVector::From(*input1); + auto x_reduce_grad = EigenVector::From(*input2); + auto x_grad = EigenVector::Flatten(*output); + auto& place = + *context.template device_context().eigen_device(); + auto broadcast_dim = + Eigen::array({{static_cast(input0->numel())}}); + Functor functor; + functor(place, x, x_reduce, x_grad, x_reduce_grad, broadcast_dim, + broadcast_dim[0]); + } else { + int rank = context.Input("X")->dims().size(); + switch (rank) { + case 1: + ReduceGradCompute<1>(context); + break; + case 2: + ReduceGradCompute<2>(context); + break; + case 3: + ReduceGradCompute<3>(context); + break; + case 4: + ReduceGradCompute<4>(context); + break; + case 5: + ReduceGradCompute<5>(context); + break; + case 6: + ReduceGradCompute<6>(context); + break; + } } } @@ -201,7 +239,8 @@ class ReduceGradKernel : public framework::OpKernel { Eigen::array broadcast_dim; for (size_t i = 0; i < D; ++i) broadcast_dim[i] = 1; broadcast_dim[dim] = input0->dims()[dim]; - auto& place = context.GetEigenDevice(); + 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]); diff --git a/paddle/operators/reorder_lod_tensor_by_rank_op.cc b/paddle/operators/reorder_lod_tensor_by_rank_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..8d652ff806461cea3d0e8d3bd70704b4b6bc2173 --- /dev/null +++ b/paddle/operators/reorder_lod_tensor_by_rank_op.cc @@ -0,0 +1,235 @@ +/* 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. */ + +#include "paddle/framework/lod_rank_table.h" +#include "paddle/framework/op_registry.h" +#include "paddle/operators/detail/safe_ref.h" +#include "paddle/platform/device_context.h" + +namespace paddle { +namespace operators { + +class ReorderLoDTensorByRankTableOpProtoMaker + : public framework::OpProtoAndCheckerMaker { + public: + ReorderLoDTensorByRankTableOpProtoMaker(OpProto *proto, + OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "(LoDTensor) the input lod tensor need to be reordered."); + AddInput("RankTable", + "(LoDRankTable) the rank table that input need follow"); + AddOutput("Out", "(LoDTensor) reordered lod tensor"); + AddComment(R"DOC(ReorderLoDTensorByRankTable + +Reorder the input X by the rank of `RankTable`. If `RankTable` is ordered by +index [3, 0, 2, 1]. Input X will reorder its sequence, the third sequence of +X will be the first sequence of Output. + +NOTE: The RankTable does not need to be calculated by X. + +For example: +The X = [Seq0, Seq1, Seq2, Seq3]. The indices of RankTable are [3, 0, 2, 1]. + +The Out = [Seq3, Seq0, Seq2, Seq1] with correct LoD information. +)DOC"); + } +}; + +class ReorderLoDTensorByRankTableBase : public framework::OperatorBase { + public: + ReorderLoDTensorByRankTableBase(const std::string &type, + const framework::VariableNameMap &inputs, + const framework::VariableNameMap &outputs, + const framework::AttributeMap &attrs) + : OperatorBase(type, inputs, outputs, attrs) {} + void Run(const framework::Scope &scope, + const platform::Place &place) const override { + auto &x = + detail::Ref(scope.FindVar(Input("X")), + "Cannot find input lod tensor variable %s", Input("X")) + .Get(); + auto &rank_table = detail::Ref(scope.FindVar(Input("RankTable")), + "Cannot find input rank table variable %s", + Input("RankTable")) + .Get(); + auto &out = + *detail::Ref(scope.FindVar(Output("Out")), + "Cannot find output lod tensor variable %s", Output("Out")) + .GetMutable(); + + out.Resize(x.dims()); + out.mutable_data(x.place(), x.type()); + this->process(place, x, rank_table, &out); + } + + protected: + virtual void process(const platform::Place &place, + const framework::LoDTensor &x, + const framework::LoDRankTable &rank_table, + framework::LoDTensor *out) const = 0; + + struct AbsoluteRankTableItem { + size_t offset; // the absolute/accumulated offset. + size_t length; // the length + framework::LoD lod; + }; + + std::vector GetAbsoluteOffsetAndLengthByLoDRankTable( + const framework::LoDTensor &x) const { + std::vector absolute_table; + size_t level = 0; + size_t size = x.lod()[level].size(); + + for (size_t i = 0; i < size - 1; ++i) { + auto lod_offset = + framework::GetSubLoDAndAbsoluteOffset(x.lod(), i, i + 1, level); + + auto &offset = lod_offset.second; + + absolute_table.emplace_back(); + absolute_table.back().length = offset.second - offset.first; + absolute_table.back().offset = offset.first; + absolute_table.back().lod = lod_offset.first; + } + return absolute_table; + } + + size_t CopyTensorAndLod(const platform::Place &place, + const AbsoluteRankTableItem &item, + const framework::LoDTensor &x, + framework::LoDTensor *out, size_t out_offset) const { + auto &out_lod = *out->mutable_lod(); + auto len = item.length; + auto x_offset = item.offset; + + if (out_lod.empty()) { + for (size_t i = 0; i < item.lod.size(); ++i) { + out_lod.push_back(std::vector({0})); + } + } + + for (size_t i = 0; i < out_lod.size(); ++i) { + auto &out_v = out_lod[i]; + auto &new_lod_v = item.lod[i]; + + for (auto &detail : new_lod_v) { + out_v.push_back(out_v.back() + detail); + } + } + + auto x_sliced = x.Slice(x_offset, x_offset + len); + auto out_sliced = out->Slice(out_offset, out_offset + len); + + platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance(); + auto &dev_ctx = *pool.Get(place); + framework::CopyFrom(x_sliced, out_sliced.place(), dev_ctx, &out_sliced); + out_offset += len; + return out_offset; + } +}; + +class ReorderLoDTensorByRankTableOp : public ReorderLoDTensorByRankTableBase { + public: + ReorderLoDTensorByRankTableOp(const std::string &type, + const framework::VariableNameMap &inputs, + const framework::VariableNameMap &outputs, + const framework::AttributeMap &attrs) + : ReorderLoDTensorByRankTableBase(type, inputs, outputs, attrs) {} + + protected: + void process(const platform::Place &place, const framework::LoDTensor &x, + const framework::LoDRankTable &rank_table, + framework::LoDTensor *out) const override { + auto absolute_table = GetAbsoluteOffsetAndLengthByLoDRankTable(x); + size_t out_offset = 0; + out->mutable_lod()->clear(); + for (auto &item : rank_table.items()) { + PADDLE_ENFORCE_LT(item.index, absolute_table.size()); + out_offset = CopyTensorAndLod(place, absolute_table[item.index], x, out, + out_offset); + } + } +}; + +class IdentityInferShape : public framework::InferShapeBase { + public: + void operator()(framework::InferShapeContext *context) const override { + context->SetOutputDim("Out", context->GetInputDim("X")); + } +}; + +class ReorderLodTensorByRankGradOpMaker + : public framework::SingleGradOpDescMaker { + public: + using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; + + protected: + std::unique_ptr Apply() const override { + auto *grad_op = new framework::OpDesc(); + grad_op->SetType("reorder_lod_tensor_by_rank_grad"); + grad_op->SetInput("X", OutputGrad("Out")); + grad_op->SetOutput("Out", InputGrad("X")); + grad_op->SetInput("RankTable", Input("RankTable")); + return std::unique_ptr(grad_op); + } +}; + +class ReorderLoDTensorByRankGradOp : public ReorderLoDTensorByRankTableBase { + public: + ReorderLoDTensorByRankGradOp(const std::string &type, + const framework::VariableNameMap &inputs, + const framework::VariableNameMap &outputs, + const framework::AttributeMap &attrs) + : ReorderLoDTensorByRankTableBase(type, inputs, outputs, attrs) {} + + protected: + void process(const platform::Place &place, const framework::LoDTensor &x, + const framework::LoDRankTable &rank_table, + framework::LoDTensor *out) const override { + auto absolute_table = GetAbsoluteOffsetAndLengthByLoDRankTable(x); + + // offsets = enumerate([item.index for item in rank_table.items()]) + std::vector> offsets; + offsets.reserve(rank_table.items().size()); + for (size_t i = 0; i < rank_table.items().size(); ++i) { + offsets.push_back({i, rank_table.items()[i].index}); + } + + // offsets.sort(key=lambda x: x[1]) + std::sort( + offsets.begin(), offsets.end(), + [](const std::pair &a, + const std::pair &b) { return a.second < b.second; }); + + // Copy TensorAndLod + size_t out_offset = 0; + for (auto &offset : offsets) { + out_offset = this->CopyTensorAndLod(place, absolute_table[offset.first], + x, out, out_offset); + } + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; + +REGISTER_OPERATOR(reorder_lod_tensor_by_rank, + ops::ReorderLoDTensorByRankTableOp, + ops::ReorderLodTensorByRankGradOpMaker, + ops::ReorderLoDTensorByRankTableOpProtoMaker, + ops::IdentityInferShape); +REGISTER_OPERATOR(reorder_lod_tensor_by_rank_grad, + ops::ReorderLoDTensorByRankGradOp, ops::IdentityInferShape); diff --git a/paddle/operators/reshape_op.cc b/paddle/operators/reshape_op.cc index ba774ec2160c0460867de42f7ad9d5cd65ad8d6a..58e8fd6124d8c076337ae9bb2f5103e7a3cb7ff0 100644 --- a/paddle/operators/reshape_op.cc +++ b/paddle/operators/reshape_op.cc @@ -1,17 +1,16 @@ - /* 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 +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 + 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. */ +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/operators/reshape_op.h" @@ -35,21 +34,33 @@ class ReshapeOp : public framework::OperatorWithKernel { auto shape = ctx->Attrs().Get>("shape"); PADDLE_ENFORCE(shape.size() > 0, "Attr(shape) shouldn't be empty."); auto x_dims = ctx->GetInputDim("X"); - // TODO(qiao) change batch_size - for (size_t i = 1; i < shape.size(); ++i) { - PADDLE_ENFORCE(shape[i] > 0, - "Each dimension of shape " - "must be positiv except the first."); - } - if (shape[0] < 0) { - shape[0] = x_dims[0]; + + std::vector neg_dims_idx; + // set some dimension to -1 if it is unknown + const int unknown_size = -1; + for (size_t i = 0; i < shape.size(); ++i) { + PADDLE_ENFORCE(shape[i] > 0 || shape[i] == unknown_size, + "Each dimension of Attr(shape) must be positive or %d.", + unknown_size); + if (shape[i] == unknown_size) { + neg_dims_idx.push_back(i); + PADDLE_ENFORCE(neg_dims_idx.size() <= 1, + "Only one dimension of Attr(shape) can be unknown."); + } } - // capacity check + int64_t capacity = std::accumulate(shape.begin(), shape.end(), 1, std::multiplies()); int64_t in_size = framework::product(x_dims); - PADDLE_ENFORCE_EQ(capacity, in_size, - "The size of Input(X) mismatches with Attr(shape)."); + if (neg_dims_idx.size() == 1) { + // dim infer + shape[neg_dims_idx[0]] = in_size / (-capacity); + // recalculate capacity + capacity = shape[neg_dims_idx[0]] * (-capacity); + } + // capacity check + PADDLE_ENFORCE(capacity == in_size, + "The size of Input(X) mismatches with Attr(shape)."); // resize output std::vector shape_int64(shape.size(), 0); std::transform(shape.begin(), shape.end(), shape_int64.begin(), @@ -66,8 +77,7 @@ class ReshapeOp : public framework::OperatorWithKernel { class ReshapeOpMaker : public framework::OpProtoAndCheckerMaker { public: - ReshapeOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + ReshapeOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "The input tensor of reshape operator."); AddOutput("Out", "The output tensor of reshape operator."); @@ -85,10 +95,13 @@ Given a 2-D tensor X with 2 rows and 2 columns [[1, 2], [3, 4]] and target shape = [1, 4], the reshape operator will transform -the tensor X into a 1-D tensor: +the tensor X into a 2-D tensor: - [1, 2, 3, 4] + [[1, 2, 3, 4]] +One dimension in the target shape can be set -1, representing that its +size is unknown. In this case, the real dimension will be infered from +the original shape of Input(X) and other dimensions in the target shape. )DOC"); } }; diff --git a/paddle/operators/reshape_op.cu b/paddle/operators/reshape_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..f487e43b99d5be2af299a9edd91dcda0c4eb7b99 --- /dev/null +++ b/paddle/operators/reshape_op.cu @@ -0,0 +1,22 @@ +/* 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. */ + +#include "paddle/operators/reshape_op.h" + +REGISTER_OP_CUDA_KERNEL( + reshape, + paddle::operators::ReshapeKernel); +REGISTER_OP_CUDA_KERNEL( + reshape_grad, + paddle::operators::ReshapeGradKernel); diff --git a/paddle/operators/reshape_op.cu.cc b/paddle/operators/reshape_op.cu.cc deleted file mode 100644 index 23dbe089d3b37aabedf9ef166f7bbfbf67da7e0a..0000000000000000000000000000000000000000 --- a/paddle/operators/reshape_op.cu.cc +++ /dev/null @@ -1,22 +0,0 @@ -/* 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. */ - -#include "paddle/operators/reshape_op.h" - -REGISTER_OP_GPU_KERNEL( - reshape, - paddle::operators::ReshapeKernel); -REGISTER_OP_GPU_KERNEL( - reshape_grad, - paddle::operators::ReshapeGradKernel); diff --git a/paddle/operators/reshape_op.h b/paddle/operators/reshape_op.h index beb951713ae2a9fd83fe7c1a5e97ee8c642158a8..a4eb34a0ad1230b6257cd299c8ed563acb054367 100644 --- a/paddle/operators/reshape_op.h +++ b/paddle/operators/reshape_op.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 @@ -20,7 +20,7 @@ namespace paddle { namespace operators { -template +template class ReshapeKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const { @@ -28,12 +28,12 @@ class ReshapeKernel : public framework::OpKernel { auto* in = ctx.Input("X"); auto out_dims = out->dims(); out->mutable_data(ctx.GetPlace()); - out->CopyFrom(*in, ctx.GetPlace(), ctx.device_context()); + framework::CopyFrom(*in, ctx.GetPlace(), ctx.device_context(), out); out->Resize(out_dims); } }; -template +template class ReshapeGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const { @@ -42,7 +42,7 @@ class ReshapeGradKernel : public framework::OpKernel { d_x->mutable_data(ctx.GetPlace()); auto in_dims = d_x->dims(); - d_x->CopyFrom(*d_out, ctx.GetPlace(), ctx.device_context()); + framework::CopyFrom(*d_out, ctx.GetPlace(), ctx.device_context(), d_x); d_x->Resize(in_dims); } }; diff --git a/paddle/operators/rmsprop_op.cc b/paddle/operators/rmsprop_op.cc index a9c45f639c6728ff2fd6de6fcdadfe5032a705d7..f7c250bf913b9213e7d7e2cca9ecadf74cac91a1 100644 --- a/paddle/operators/rmsprop_op.cc +++ b/paddle/operators/rmsprop_op.cc @@ -63,8 +63,7 @@ class RmspropOp : public framework::OperatorWithKernel { class RmspropOpMaker : public framework::OpProtoAndCheckerMaker { public: - RmspropOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + RmspropOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("Param", "(Tensor, default Tensor) " @@ -116,5 +115,5 @@ http://www.cs.toronto.edu/~tijmen/csc321/slides/lecture_slides_lec6.pdf) namespace ops = paddle::operators; REGISTER_OP_WITHOUT_GRADIENT(rmsprop, ops::RmspropOp, ops::RmspropOpMaker); -REGISTER_OP_CPU_KERNEL(rmsprop, - ops::RmspropOpKernel); +REGISTER_OP_CPU_KERNEL( + rmsprop, ops::RmspropOpKernel); diff --git a/paddle/operators/rmsprop_op.cu b/paddle/operators/rmsprop_op.cu index 52634a54816bcd5ad0ba82a56f1df95110112265..0295dc262f095a2b58ab34e3c2ec9f5440e4bfca 100644 --- a/paddle/operators/rmsprop_op.cu +++ b/paddle/operators/rmsprop_op.cu @@ -1,20 +1,20 @@ /* 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 +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 + 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. */ +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. */ #define EIGEN_USE_GPU #include "paddle/operators/rmsprop_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(rmsprop, - ops::RmspropOpKernel); +REGISTER_OP_CUDA_KERNEL( + rmsprop, ops::RmspropOpKernel); diff --git a/paddle/operators/rmsprop_op.h b/paddle/operators/rmsprop_op.h index 7bf2129010f994966d79ef11d5cec30159b47068..16a561835d02457cf2268f713289001773e63d6c 100644 --- a/paddle/operators/rmsprop_op.h +++ b/paddle/operators/rmsprop_op.h @@ -24,7 +24,7 @@ template using EigenVector = framework::EigenVector; -template +template class RmspropOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { @@ -51,7 +51,7 @@ class RmspropOpKernel : public framework::OpKernel { auto p_out = EigenVector::Flatten(*param_out); auto mom_out = EigenVector::Flatten(*moment_out); auto ms_out = EigenVector::Flatten(*mean_square_out); - auto place = ctx.GetEigenDevice(); + auto& place = *ctx.template device_context().eigen_device(); Eigen::DSizes grad_dsize(grad->numel()); diff --git a/paddle/operators/rnn/recurrent_op_utils.cc b/paddle/operators/rnn/recurrent_op_utils.cc deleted file mode 100644 index ee61ea300c33722471189d06eb09f67a083d2a4d..0000000000000000000000000000000000000000 --- a/paddle/operators/rnn/recurrent_op_utils.cc +++ /dev/null @@ -1,134 +0,0 @@ -/* 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. */ - -#include "paddle/operators/rnn/recurrent_op_utils.h" - -namespace paddle { -namespace operators { -namespace rnn { - -namespace f = paddle::framework; - -using Tensor = framework::Tensor; -using LoDTensor = framework::LoDTensor; - -void SegmentInputs(const std::vector& step_scopes, - const std::vector& inlinks, - const size_t seq_len) { - PADDLE_ENFORCE(!inlinks.empty(), "no in links are provided."); - for (size_t i = 0; i < inlinks.size(); ++i) { - // global inputs - auto input_var = step_scopes[0]->parent().FindVar(inlinks[i]); - PADDLE_ENFORCE_NOT_NULL(input_var, "input link [%s] is not in scope.", - inlinks[i]); - - LoDTensor* input = input_var->GetMutable(); - f::DDim dims = input->dims(); - PADDLE_ENFORCE_EQ(static_cast(dims[0]), seq_len, - "all the inputs be the same length"); - f::DDim step_dims = slice_ddim(dims, 1, dims.size()); - for (size_t j = 0; j < seq_len; j++) { - Tensor* step_input = - step_scopes[j]->Var(inlinks[i])->GetMutable(); - // The input of operators of each step is Tensor here. - // Maybe need to modify Slice function. - *step_input = input->Slice(j, j + 1); - step_input->Resize(step_dims); - } - } -} - -void ConcatOutputs(const std::vector& step_scopes, - const std::vector& outlinks, - const size_t seq_len, const platform::DeviceContext& ctx) { - for (size_t i = 0; i < outlinks.size(); i++) { - auto* output_var = step_scopes[0]->parent().FindVar(outlinks[i]); - PADDLE_ENFORCE_NOT_NULL(output_var, "output link [%s] is not in scope.", - outlinks[i]); - LoDTensor* output = output_var->GetMutable(); - - auto* step_scope_var = step_scopes[0]->FindVar(outlinks[i]); - PADDLE_ENFORCE_NOT_NULL(step_scope_var, "%s not in scope", outlinks[i]); - f::DDim step_dims = - step_scope_var->template GetMutable()->dims(); - std::vector dims_vec = vectorize(step_dims); - dims_vec.insert(dims_vec.begin(), seq_len); - output->Resize(f::make_ddim(dims_vec)); - output->mutable_data(platform::CPUPlace()); - for (size_t j = 0; j < seq_len; j++) { - LoDTensor* step_output = - step_scopes[j]->FindVar(outlinks[i])->GetMutable(); - // TODO(luotao02) data type and platform::DeviceContext() should set - // correctly - (output->Slice(j, j + 1)) - .CopyFrom(*step_output, platform::CPUPlace(), ctx); - } - } -} - -void LinkMemories(const std::vector& scopes, - const std::vector& memories, - const size_t step_id, const int offset) { - PADDLE_ENFORCE_LT(step_id, scopes.size(), - "step [%d] is out of range of step scopes' size [%d]", - step_id, scopes.size()); - PADDLE_ENFORCE_GE(static_cast(step_id) + offset, 0, - "offset [%d] must be large than -[%d]", offset, step_id); - PADDLE_ENFORCE_LT( - step_id + offset, scopes.size(), - "offset [%d] is out of range, it must be less than (%d - %d)", offset, - scopes.size(), step_id); - auto* scope = scopes[step_id]; - auto* linked_scope = scopes[step_id + offset]; - for (auto& attr : memories) { - auto* mem = scope->FindVar(attr.pre_var)->GetMutable(); - auto* linked_mem = linked_scope->FindVar(attr.var)->GetMutable(); - mem->Resize(linked_mem->dims()); - mem->ShareDataWith(*linked_mem); - } -} - -void InitArgument(const ArgumentName& name, Argument* arg, - const framework::OperatorBase& op, bool is_grad) { - arg->step_scopes = - is_grad ? op.Input(name.step_scopes) : op.Output(name.step_scopes); - arg->inlinks = op.Inputs(name.inlinks); - arg->outlinks = op.Outputs(name.outlinks); - - auto& boot_memories = is_grad ? op.Outputs(name.initial_states) - : op.Inputs(name.initial_states); - // attributes - auto& memories = op.Attr>(name.states); - auto& pre_memories = op.Attr>(name.ex_states); - - PADDLE_ENFORCE(memories.size() == boot_memories.size(), - "the size of states, initial_states don't match:%d,%d", - memories.size(), boot_memories.size()); - PADDLE_ENFORCE(pre_memories.size() == boot_memories.size(), - "the size of ex_states, initial_states don't match:%d,%d", - pre_memories.size(), boot_memories.size()); - PADDLE_ENFORCE(memories.size() > 0, "more than 1 states should be set"); - - for (size_t i = 0; i < memories.size(); ++i) { - rnn::StateAttr mem_attr; - mem_attr.var = memories[i]; - mem_attr.pre_var = pre_memories[i]; - mem_attr.boot_var = boot_memories[i]; - (arg->states).push_back(mem_attr); - } -} - -} // namespace rnn -} // namespace operators -} // namespace paddle diff --git a/paddle/operators/rnn/recurrent_op_utils.h b/paddle/operators/rnn/recurrent_op_utils.h deleted file mode 100644 index fb0e158e07745d58c6211d33e385b324e492b95e..0000000000000000000000000000000000000000 --- a/paddle/operators/rnn/recurrent_op_utils.h +++ /dev/null @@ -1,85 +0,0 @@ -/* 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 - -#include - -#include "paddle/framework/operator.h" - -namespace paddle { -namespace operators { -namespace rnn { - -using Scope = framework::Scope; - -/** - * Memory of a RNN (same as the role of `Momory` in PaddlePaddle). - * - * Memory attributes cached by this op, dims will be infered from - * boot memories in father scope. Other attributes are copied from Op's proto - * attributes. - */ -struct StateAttr { - // name of current state variable - std::string var; - // name of previous step's state variable - std::string pre_var; - // name of the variables to init this memory (same role of `boot_layer` in - // PaddlePaddle), which is store in father's scope. - std::string boot_var; -}; - -struct Argument { - std::string step_net; - std::string step_scopes; - std::vector inlinks; - std::vector outlinks; - std::vector states; -}; - -struct ArgumentName { - std::string step_net; - std::string step_scopes; - std::string inlinks; - std::string outlinks; - std::string states; // the memory name - std::string ex_states; // the previous memory name - std::string initial_states; // the boot memory name -}; - -/** - * Prepare inputs for each step net. - */ -void SegmentInputs(const std::vector& step_scopes, - const std::vector& inlinks, - const size_t seq_len); - -/** - * Process outputs of step nets and merge to variables. - */ -void ConcatOutputs(const std::vector& step_scopes, - const std::vector& outlinks, - const size_t seq_len, const platform::DeviceContext& ctx); - -void LinkMemories(const std::vector& step_scopes, - const std::vector& memories, const size_t step_id, - const int offset); - -void InitArgument(const ArgumentName& name, Argument* arg, - const framework::OperatorBase& op, bool is_grad = false); - -} // namespace rnn -} // namespace operators -} // namespace paddle diff --git a/paddle/operators/rnn_memory_helper_op.cc b/paddle/operators/rnn_memory_helper_op.cc index b621c7f1ba3f9e9613dea5bc98ef74c7c6dae9a0..eb55ed6a05b51d7a6c63d16fcf5aff73f6744903 100644 --- a/paddle/operators/rnn_memory_helper_op.cc +++ b/paddle/operators/rnn_memory_helper_op.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/framework/op_registry.h" #include "paddle/framework/operator.h" @@ -25,7 +25,7 @@ class RNNMemoryHelperOp : public framework::OperatorBase { const framework::AttributeMap &attrs) : OperatorBase(type, inputs, outputs, attrs) {} void Run(const framework::Scope &scope, - const platform::DeviceContext &dev_ctx) const override { + const platform::Place &dev_place) const override { auto mem_var_name = Input("X"); auto *mem_var = scope.FindVar(mem_var_name); PADDLE_ENFORCE(mem_var != nullptr, @@ -57,15 +57,14 @@ class RNNMemoryHelperOpShapeInference : public framework::InferShapeBase { class RNNMemoryHelperOpInfoMaker : public framework::OpProtoAndCheckerMaker { public: - RNNMemoryHelperOpInfoMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + RNNMemoryHelperOpInfoMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", ""); AddOutput("Out", ""); - AddAttr("data_type", + AddAttr("dtype", "(int, default 5 (FP32)) " "Output data type") - .SetDefault(framework::DataType::FP32); + .SetDefault(framework::proto::DataType::FP32); AddComment(""); } }; @@ -78,7 +77,7 @@ class RNNMemoryHelperGradOp : public framework::OperatorBase { const framework::AttributeMap &attrs) : OperatorBase(type, inputs, outputs, attrs) {} void Run(const framework::Scope &scope, - const platform::DeviceContext &dev_ctx) const override { + const platform::Place &dev_place) const override { auto out_grad_var_name = Input(framework::GradVarName("Out")); auto *out_grad_var = scope.FindVar(out_grad_var_name); @@ -95,13 +94,13 @@ class RNNMemoryHelperGradOp : public framework::OperatorBase { auto &in_var_tensor = in_var->Get(); framework::AttributeMap attrs; - attrs["data_type"] = framework::ToDataType(in_var_tensor.type()); + attrs["dtype"] = framework::ToDataType(in_var_tensor.type()); attrs["shape"] = framework::vectorize2int(in_var_tensor.dims()); attrs["value"] = 0.0f; auto zero_op = framework::OpRegistry::CreateOp( "fill_constant", {}, {{"Out", {in_grad_var_name}}}, attrs); - zero_op->Run(scope, dev_ctx); + zero_op->Run(scope, dev_place); } else { auto &out_grad_tensor = out_grad_var->Get(); auto *in_grad_tensor = in_grad_var->GetMutable(); @@ -114,17 +113,16 @@ class RNNMemoryHelperGradOp : public framework::OperatorBase { class RNNMemoryHelperGradOpInfoMaker : public framework::OpProtoAndCheckerMaker { public: - RNNMemoryHelperGradOpInfoMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + RNNMemoryHelperGradOpInfoMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput(framework::GradVarName("Out"), ""); AddInput("X", ""); AddInput("Out", ""); AddOutput(framework::GradVarName("X"), ""); - AddAttr("data_type", + AddAttr("dtype", "(int, default 5 (FP32)) " "Output data type") - .SetDefault(framework::DataType::FP32); + .SetDefault(framework::proto::DataType::FP32); AddComment(""); } }; diff --git a/paddle/operators/roi_pool_op.cc b/paddle/operators/roi_pool_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..ef1804d9762200686ac8537140af046c21443779 --- /dev/null +++ b/paddle/operators/roi_pool_op.cc @@ -0,0 +1,165 @@ +/* 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. */ + +#include "paddle/operators/roi_pool_op.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; + +static constexpr int kROISize = 5; + +class ROIPoolOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of ROIPoolOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("ROIs"), + "Input(ROIs) of ROIPoolOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of ROIPoolOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Argmax"), + "Output(Argmax) of ROIPoolOp should not be null."); + auto input_dims = ctx->GetInputDim("X"); + auto rois_dims = ctx->GetInputDim("ROIs"); + + 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], …]."); + 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], …]."); + + int pooled_height = ctx->Attrs().Get("pooled_height"); + int pooled_width = ctx->Attrs().Get("pooled_width"); + float spatial_scale = ctx->Attrs().Get("spatial_scale"); + + PADDLE_ENFORCE_GT(pooled_height, 0, + "The pooled output height must greater than 0"); + PADDLE_ENFORCE_GT(pooled_width, 0, + "The pooled output width must greater than 0"); + PADDLE_ENFORCE_GT(spatial_scale, 0.0f, + "The spatial scale must greater than 0"); + + auto out_dims = input_dims; + out_dims[0] = rois_dims[0]; + out_dims[1] = input_dims[1]; + out_dims[2] = pooled_height; + out_dims[3] = pooled_width; + + ctx->SetOutputDim("Out", out_dims); + ctx->SetOutputDim("Argmax", out_dims); + } + + protected: + framework::OpKernelType GetActualKernelType( + const framework::ExecutionContext& ctx) const override { + return framework::OpKernelType( + framework::ToDataType(ctx.Input("X")->type()), + ctx.device_context()); + } +}; + +class ROIPoolGradOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "The gradient of Out should not be null."); + PADDLE_ENFORCE(ctx->HasOutputs(framework::GradVarName("X")), + "The gradient of X should not be null."); + ctx->SetOutputsDim(framework::GradVarName("X"), ctx->GetInputsDim("X")); + } + + protected: + framework::OpKernelType GetActualKernelType( + const framework::ExecutionContext& ctx) const override { + return framework::OpKernelType( + framework::ToDataType(ctx.Input("X")->type()), + ctx.device_context()); + } +}; + +class ROIPoolOpMaker : public framework::OpProtoAndCheckerMaker { + public: + ROIPoolOpMaker(OpProto* proto, OpAttrChecker* op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", + "(Tensor), " + "the input of ROIPoolOp. " + "The format of input tensor is NCHW. Where N is batch size, " + "C is the number of input channels, " + "H is the height of the feature, and " + "W is the width of the feature."); + AddInput("ROIs", + "(Tensor), " + "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], …]. " + "Where batch_id is the id of the data, " + "(x1, y1) is the top left coordinates, and " + "(x2, y2) is the bottom right coordinates."); + AddOutput("Out", + "(Tensor), " + "The output of ROIPoolOp is a 4-D tensor with shape " + "(num_rois, channels, pooled_h, pooled_w)."); + AddOutput("Argmax", + "(Tensor), " + "Argmaxes corresponding to indices in X used " + "for gradient computation. Only output " + "if arg “is_test” is false.") + .AsIntermediate(); + AddAttr("spatial_scale", + "(float, default 1.0), " + "Multiplicative spatial scale factor " + "to translate ROI coords from their input scale " + "to the scale used when pooling.") + .SetDefault(1.0); + AddAttr("pooled_height", + "(int, default 1), " + "The pooled output height.") + .SetDefault(1); + AddAttr("pooled_width", + "(int, default 1), " + "The pooled output width.") + .SetDefault(1); + AddComment(R"DOC( +ROIPool operator + +ROI Pooling for Faster-RCNN. The link below is a further introduction: +https://stackoverflow.com/questions/43430056/what-is-roi-layer-in-fast-rcnn + )DOC"); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP(roi_pool, ops::ROIPoolOp, ops::ROIPoolOpMaker, roi_pool_grad, + ops::ROIPoolGradOp); +REGISTER_OP_CPU_KERNEL( + roi_pool, + ops::CPUROIPoolOpKernel, + ops::CPUROIPoolOpKernel); +REGISTER_OP_CPU_KERNEL( + roi_pool_grad, + ops::CPUROIPoolGradOpKernel, + ops::CPUROIPoolOpKernel); diff --git a/paddle/operators/roi_pool_op.cu b/paddle/operators/roi_pool_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..a874befe4d12029afa9ce55230da22cb048000aa --- /dev/null +++ b/paddle/operators/roi_pool_op.cu @@ -0,0 +1,209 @@ +/* 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. */ + +#include "paddle/operators/roi_pool_op.h" +#include "paddle/platform/cuda_helper.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; + +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, + kNumMaxinumNumBlocks); +} + +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) { + 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; + + 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_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)); + + 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); + wend = min(max(wend + roi_start_w, 0), width); + bool is_empty = (hend <= hstart) || (wend <= wstart); + + T maxval = is_empty ? 0 : -std::numeric_limits::max(); + int maxidx = -1; + const T* offset_input_data = + input_data + (roi_batch_ind * channels + c) * height * width; + for (int h = hstart; h < hend; ++h) { + for (int w = wstart; w < wend; ++w) { + int input_data_index = h * width + w; + if (offset_input_data[input_data_index] > maxval) { + maxval = offset_input_data[input_data_index]; + maxidx = input_data_index; + } + } + } + output_data[index] = maxval; + if (argmax_data) { + argmax_data[index] = maxidx; + } + } +} + +template +__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) { + 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; + + const int64_t* offset_input_rois = input_rois + n * kROISize; + int roi_batch_ind = offset_input_rois[0]; + 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; + T* offset_input_grad = input_grad + input_offset; + const int64_t* offset_argmax_data = argmax_data + output_offset; + + int argmax = offset_argmax_data[ph * pooled_width + pw]; + if (argmax != -1) { + platform::CudaAtomicAdd( + offset_input_grad + argmax, + static_cast(offset_output_grad[ph * pooled_width + pw])); + } + } +} + +template +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* out = ctx.Output("Out"); + auto* argmax = ctx.Output("Argmax"); + + auto pooled_height = ctx.Attr("pooled_height"); + auto pooled_width = ctx.Attr("pooled_width"); + auto spatial_scale = ctx.Attr("spatial_scale"); + + auto in_dims = in->dims(); + 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]; + if (rois_num == 0) return; + + int output_size = out->numel(); + int blocks = NumBlocks(output_size); + int threads = kNumCUDAThreads; + + GPUROIPoolForward< + T><<>>( + output_size, in->data(), rois->data(), spatial_scale, + channels, height, width, pooled_height, pooled_width, + out->mutable_data(ctx.GetPlace()), + argmax->mutable_data(ctx.GetPlace())); + } +}; + +template +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* argmax = ctx.Input("Argmax"); + + auto* out_grad = ctx.Input(framework::GradVarName("Out")); + auto* x_grad = ctx.Output(framework::GradVarName("X")); + + auto pooled_height = ctx.Attr("pooled_height"); + auto pooled_width = ctx.Attr("pooled_width"); + auto spatial_scale = ctx.Attr("spatial_scale"); + + size_t rois_num = rois->dims()[0]; + int channels = in->dims()[1]; + int height = in->dims()[2]; + int width = in->dims()[3]; + + if (x_grad) { + x_grad->mutable_data(ctx.GetPlace()); + math::SetConstant set_zero; + set_zero(ctx.cuda_device_context(), x_grad, static_cast(0)); + + int output_grad_size = out_grad->numel(); + int blocks = NumBlocks(output_grad_size); + int threads = kNumCUDAThreads; + + if (output_grad_size > 0) { + GPUROIPoolBackward< + T><<>>( + output_grad_size, rois->data(), out_grad->data(), + argmax->data(), rois_num, spatial_scale, channels, height, + width, pooled_height, pooled_width, + x_grad->mutable_data(ctx.GetPlace())); + } + } + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP_CUDA_KERNEL( + roi_pool, + ops::GPUROIPoolOpKernel, + ops::GPUROIPoolOpKernel); +REGISTER_OP_CUDA_KERNEL( + roi_pool_grad, + ops::GPUROIPoolGradOpKernel, + ops::GPUROIPoolOpKernel); diff --git a/paddle/operators/roi_pool_op.h b/paddle/operators/roi_pool_op.h new file mode 100644 index 0000000000000000000000000000000000000000..09a9d3d870c1066f1c6f780c4b3203679e9e7505 --- /dev/null +++ b/paddle/operators/roi_pool_op.h @@ -0,0 +1,184 @@ +/* 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 +#include "paddle/framework/op_registry.h" +#include "paddle/operators/math/math_function.h" + +namespace paddle { +namespace operators { + +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* out = ctx.Output("Out"); + auto* argmax = ctx.Output("Argmax"); + + auto pooled_height = ctx.Attr("pooled_height"); + auto pooled_width = ctx.Attr("pooled_width"); + auto spatial_scale = ctx.Attr("spatial_scale"); + + auto in_dims = in->dims(); + int batch_size = in_dims[0]; + int channels = in_dims[1]; + int height = in_dims[2]; + int width = in_dims[3]; + int rois_num = rois->dims()[0]; + + auto in_stride = framework::stride(in_dims); + auto argmax_stride = framework::stride(argmax->dims()); + auto roi_stride = framework::stride(rois->dims()); + 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]; + } + + 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); + + // Force malformed ROIs to be 1x1 + int roi_height = std::max(roi_end_h - roi_start_h + 1, 1); + int roi_width = std::max(roi_end_w - roi_start_w + 1, 1); + + const float bin_size_h = + static_cast(roi_height) / static_cast(pooled_height); + const float bin_size_w = + static_cast(roi_width) / static_cast(pooled_width); + + const T* batch_data = input_data + roi_batch_id * in_stride[0]; + + for (int c = 0; c < channels; ++c) { + for (int ph = 0; ph < pooled_height; ++ph) { + for (int pw = 0; pw < pooled_width; ++pw) { + // Compute pooling region for this output unit: + // start (included) = floor(ph * roi_height / pooled_height_) + // end (excluded) = ceil((ph + 1) * roi_height / pooled_height_) + 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)); + + hstart = std::min(std::max(hstart + roi_start_h, 0), height); + hend = std::min(std::max(hend + roi_start_h, 0), height); + wstart = std::min(std::max(wstart + roi_start_w, 0), width); + wend = std::min(std::max(wend + roi_start_w, 0), width); + + const int pool_index = ph * pooled_width + pw; + + // Define an empty pooling region to be zero + bool is_empty = (hend <= hstart) || (wend <= wstart); + output_data[pool_index] = + is_empty ? 0 : -std::numeric_limits::max(); + argmax_data[pool_index] = -1; + + for (int h = hstart; h < hend; ++h) { + for (int w = wstart; w < wend; ++w) { + const int index = h * width + w; + if (batch_data[index] > output_data[pool_index]) { + output_data[pool_index] = batch_data[index]; + argmax_data[pool_index] = index; + } + } + } + } + } + + batch_data += in_stride[1]; + output_data += out_stride[1]; + argmax_data += argmax_stride[1]; + } + // Increment ROI data pointer + rois_data += roi_stride[0]; + } + return; + } +}; + +template +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* argmax = ctx.Input("Argmax"); + auto* out_grad = + ctx.Input(framework::GradVarName("Out")); + auto* in_grad = ctx.Output(framework::GradVarName("X")); + + auto pooled_height = ctx.Attr("pooled_height"); + auto pooled_width = ctx.Attr("pooled_width"); + + if (in_grad) { + const int64_t* rois_data = rois->data(); + const T* out_grad_data = out_grad->data(); + const int64_t* argmax_data = argmax->data(); + T* in_grad_data = in_grad->mutable_data(ctx.GetPlace()); + math::SetConstant set_zero; + set_zero(ctx.template device_context(), in_grad, + static_cast(0)); + + auto in_stride = framework::stride(in->dims()); + auto argmax_stride = framework::stride(argmax->dims()); + 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]; + 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) { + for (int pw = 0; pw < pooled_width; ++pw) { + int pool_index = ph * pooled_width + pw; + if (argmax_data[pool_index] >= 0) { + auto index = argmax_data[pool_index]; + batch_grad_data[index] += out_grad_data[pool_index]; + } + } + } + batch_grad_data += in_stride[1]; + out_grad_data += out_stride[1]; + argmax_data += argmax_stride[1]; + } + rois_data += roi_stride[0]; + } + } + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/row_conv_op.cc b/paddle/operators/row_conv_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..68f4e3531566fd346055404d45651c2b53ebe31b --- /dev/null +++ b/paddle/operators/row_conv_op.cc @@ -0,0 +1,259 @@ +/* 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. */ + +#include "paddle/operators/row_conv_op.h" +#include "paddle/framework/eigen.h" + +namespace paddle { +namespace operators { + +using LoDTensor = framework::LoDTensor; +using framework::Tensor; + +template +using EigenMatrix = framework::EigenMatrix; + +class RowConvOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of RowConvOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Filter"), + "Input(Filter) of RowConvOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of RowConvOp should not be null."); + + auto x_dims = ctx->GetInputDim("X"); + auto filter_dims = ctx->GetInputDim("Filter"); + PADDLE_ENFORCE_EQ(x_dims.size(), 2, "Input(X)'s rank should be 2."); + PADDLE_ENFORCE_EQ(filter_dims.size(), 2, "Input(Y)'s rank should be 2."); + PADDLE_ENFORCE_EQ( + x_dims[1], filter_dims[1], + "The 2nd dimension of Input(X) and Input(Filter) should be same."); + ctx->SetOutputDim("Out", x_dims); + ctx->ShareLoD("X", "Out"); + } +}; + +class RowConvGradOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Filter"), + "Input(Filter) should not be null."); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Gradient of output(Out) should not be null."); + + auto x_grad_name = framework::GradVarName("X"); + if (ctx->HasOutput(x_grad_name)) { + auto x_dims = ctx->GetInputDim("X"); + ctx->SetOutputDim(x_grad_name, x_dims); + } + + auto filter_grad_name = framework::GradVarName("Filter"); + if (ctx->HasOutput(filter_grad_name)) { + auto filter_dims = ctx->GetInputDim("Filter"); + ctx->SetOutputDim(filter_grad_name, filter_dims); + } + } +}; + +class RowConvOpMaker : public framework::OpProtoAndCheckerMaker { + public: + RowConvOpMaker(OpProto *proto, OpAttrChecker *op_checker) + : framework::OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", + "(LoDTensor), the input(X) is a LodTensor, which supports " + "variable time-length input sequences. The underlying tensor " + "in this LoDTensor is a matrix with shape (T x N), where T " + "is the total time steps in this mini-batch and N is the input " + "data dimension."); + AddInput("Filter", + "(Tensor), the input(Filter) is a learnable parameter. It " + "is a 2-D tensor with shape (future_context x N), where, " + "future_context is the future context length and N is the data " + "dimension."); + AddOutput("Out", + "(LoDTensor), the output(Out) is a LodTensor, which supports " + "variable time-length input sequences. The underlying tensor " + "in this LodTensor is a matrix with shape T x N, i.e., the " + "same shape as X."); + AddComment(R"DOC( +Row-convolution Operator. + +The row convolution is called lookahead convolution. This operator was +introduced in the following paper for DeepSpeech2: +http://www.cs.cmu.edu/~dyogatam/papers/wang+etal.iclrworkshop2016.pdf + +The main motivation is that a bidirectional RNN, useful in DeepSpeech +like speech models, learns representation for a sequence by performing a +forward and a backward pass through the entire sequence. However, unlike +unidirectional RNNs, bidirectional RNNs are challenging to deploy in an online +and low-latency setting. The lookahead convolution incorporates information +from future subsequences in a computationally efficient manner to improve +unidirectional recurrent neural networks. The row convolution operator is +different from the 1D sequence convolution, and is computed as follows: + +Given an input sequence $in$ of length $t$ and input dimension $d$, +and a filter ($W$) of size $context \times d$, +the output sequence is convolved as: + +$$ +out_{i, :} = \sum_{j=i}^{i + context} in_{j,:} \dot W_{i-j, :} +$$ + +)DOC"); + } +}; + +template +class RowConvKernel + : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &context) const override { + auto *x = context.Input("X"); + auto *filter = context.Input("Filter"); + auto *out = context.Output("Out"); + + out->mutable_data(context.GetPlace()); + + auto batch_indices = x->lod()[0]; + auto input_dim = x->dims()[1]; // 'in' is of size T x N + size_t num_sequence = batch_indices.size() - 1; + + auto future_context = filter->dims()[0]; + auto weights = EigenMatrix::From(*filter); + + for (size_t i = 0; i < num_sequence; i++) { + int start = static_cast(batch_indices[i]); + int end = static_cast(batch_indices[i + 1]); + int current_timesteps = end - start; + Tensor cur_input_sequence = + x->Slice(start, end); // Current input sequence + Tensor cur_output_sequence = + out->Slice(start, end); // Current output sequence + auto cip_seq = EigenMatrix::From(cur_input_sequence); + auto cot_seq = EigenMatrix::From(cur_output_sequence); + + for (int k = 0; k < current_timesteps; + k++) { // For different time steps in the same sequence + for (int w = 0; (w < future_context) && ((k + w) < current_timesteps); + w++) { + for (int d = 0; d < input_dim; d++) { + if (w == 0) { + cot_seq(k, d) = weights(w, d) * cip_seq(k + w, d); + } else { + cot_seq(k, d) += weights(w, d) * cip_seq(k + w, d); + } + } + } + } + } + } +}; + +template +class RowConvGradKernel + : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &context) const override { + auto *x = context.Input("X"); + auto *filter = context.Input("Filter"); + auto *d_out = context.Input(framework::GradVarName("Out")); + auto *dx = context.Output(framework::GradVarName("X")); + auto *d_filter = context.Output(framework::GradVarName("Filter")); + + auto input_dim = x->dims()[1]; // 'x' is of size T x N + auto batch_indices = x->lod()[0]; + size_t num_sequence = batch_indices.size() - 1; + auto future_context = filter->dims()[0]; + + if (d_filter) { + d_filter->mutable_data(context.GetPlace()); + auto dweights = + EigenMatrix::From(*d_filter); // Gradient of weight matrix + dweights.setZero(); + + for (size_t i = 0; i < num_sequence; i++) { // For different sequences + int start = static_cast(batch_indices[i]); + int end = static_cast(batch_indices[i + 1]); + + Tensor cur_input = x->Slice(start, end); // Current input sequence + Tensor cur_doutput = + d_out->Slice(start, end); // Current output grad sequence + + auto cur_ip = EigenMatrix::From(cur_input); + auto cur_dout = EigenMatrix::From(cur_doutput); + int current_timesteps = end - start; + + for (int k = 0; k < current_timesteps; + k++) { // For different time steps in the same sequence + for (int w = 0; (w < future_context) && ((k + w) < current_timesteps); + w++) { + // For dweights (Updating the gradient of weight matrix) + for (int d = 0; d < input_dim; d++) { + dweights(w, d) += cur_ip(k + w, d) * cur_dout(k, d); + } + } + } + } + } + + if (dx) { + dx->mutable_data(context.GetPlace()); + auto weights = EigenMatrix::From(*filter); + for (size_t i = 0; i < num_sequence; i++) { // For different sequences + int start = static_cast(batch_indices[i]); + int end = static_cast(batch_indices[i + 1]); + + Tensor cur_doutput = + d_out->Slice(start, end); // Current output grad sequence + Tensor cur_dinput = + dx->Slice(start, end); // Current input grad sequence + + auto cur_dout = EigenMatrix::From(cur_doutput); + auto cur_dip = EigenMatrix::From(cur_dinput); + cur_dip.setZero(); + int current_timesteps = end - start; + + for (int k = 0; k < current_timesteps; + k++) { // For different time steps in the same sequence + for (int w = 0; (w < future_context) && ((k + w) < current_timesteps); + w++) { + // For dinput (Updating the gradient wrt input) + for (int d = 0; d < input_dim; d++) { + cur_dip(k + w, d) += weights(w, d) * cur_dout(k, d); + } + } + } + } + } + } +}; +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP(row_conv, ops::RowConvOp, ops::RowConvOpMaker, row_conv_grad, + ops::RowConvGradOp); +REGISTER_OP_CPU_KERNEL( + row_conv, ops::RowConvKernel); +REGISTER_OP_CPU_KERNEL( + row_conv_grad, + ops::RowConvGradKernel); diff --git a/paddle/operators/row_conv_op.cu b/paddle/operators/row_conv_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..41f2c5b9de91ade15b4010f56377675cfd1b611c --- /dev/null +++ b/paddle/operators/row_conv_op.cu @@ -0,0 +1,410 @@ +/* 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. */ + +#include "paddle/operators/math/math_function.h" +#include "paddle/operators/row_conv_op.h" +#include "paddle/platform/cuda_helper.h" + +namespace paddle { +namespace operators { + +using LoDTensor = framework::LoDTensor; +using framework::Tensor; + +namespace { + +inline int DivUp(int x, int y) { return (x + y - 1) / y; } + +// Forward prop (shared memory version, for small future_context) +template +__global__ void RowConvForwardSharedMemory(const T *in, const T *wt, + int num_sequence, int input_dim, + int future_context, + const size_t *batch_indices, + T *out) { + int blx = blockDim.x; + int bly = blockDim.y; + int thx = threadIdx.x; + int thy = threadIdx.y; + int d = blockIdx.x * blx + thx; // index along input dim + + extern __shared__ T mem[]; + T *sw = mem; + + if (thy < future_context) { + sw[thy * blx + thx] = + (d < input_dim) ? wt[thy * input_dim + d] : static_cast(0); + } + __syncthreads(); + + for (size_t i = 0; i < num_sequence; i++) { + int start = static_cast(batch_indices[i]); + int end = static_cast(batch_indices[i + 1]); + int current_timesteps = end - start; + for (int k = thy; k < current_timesteps; k += bly) { + T sum = 0; + for (int w = 0; (w < future_context) && ((k + w) < current_timesteps); + w++) { + sum += (d < input_dim) + ? sw[w * blx + thx] * in[(start + k + w) * input_dim + d] + : static_cast(0); + } + if (d < input_dim) { + out[(start + k) * input_dim + d] = sum; + } + } + } +} + +// Forward prop (naive version) +template +__global__ void RowConvForward(const T *in, const T *wt, int num_sequence, + int input_dim, int future_context, + const size_t *batch_indices, T *out) { + int d = blockIdx.x * blockDim.x + threadIdx.x; // index along input_dim + int bly = blockDim.y; + int thy = threadIdx.y; + + if (d >= input_dim) return; + + for (size_t i = 0; i < num_sequence; i++) { + int start = static_cast(batch_indices[i]); + int end = static_cast(batch_indices[i + 1]); + int current_timesteps = end - start; + for (int k = thy; k < current_timesteps; k += bly) { + T sum = 0; + for (int w = 0; (w < future_context) && ((k + w) < current_timesteps); + w++) { + sum += (wt[w * input_dim + d] * in[(start + k + w) * input_dim + d]); + } + out[(start + k) * input_dim + d] = sum; + } + } +} + +// Compute input gradient (shared memory version, for small future_context) +template +__global__ void RowConvGradInputSharedMemory(const T *dout, const T *wt, + int num_sequence, int input_dim, + int future_context, + const size_t *batch_indices, + T *din) { + int blx = blockDim.x; + int bly = blockDim.y; + int thx = threadIdx.x; + int thy = threadIdx.y; + int d = blockIdx.x * blx + thx; // index along input dim + + extern __shared__ T mem[]; + T *sw = mem; + if (thy < future_context) { + sw[thy * blx + thx] = + (d < input_dim) ? wt[thy * input_dim + d] : static_cast(0); + } + __syncthreads(); + + for (int i = 0; i < num_sequence; i++) { + int start = static_cast(batch_indices[i]); + int end = static_cast(batch_indices[i + 1]); + int current_timesteps = end - start; + for (int k = thy; k < current_timesteps; k += bly) { + T sum = 0; + for (int w = 0; (w < future_context) && ((k - w) >= 0); w++) { + sum += (d < input_dim) + ? (sw[w * blx + thx] * dout[(k + start - w) * input_dim + d]) + : static_cast(0); + } + if (d < input_dim) { + din[(k + start) * input_dim + d] = sum; + } + } + } +} + +// Compute input gradient (Naive version) +template +__global__ void RowConvGradInput(const T *dout, const T *wt, int num_sequence, + int input_dim, int future_context, + const size_t *batch_indices, T *din) { + int d = blockIdx.x * blockDim.x + threadIdx.x; // index along input_dim + int bly = blockDim.y; + int thy = threadIdx.y; + + if (d >= input_dim) return; + for (int i = 0; i < num_sequence; i++) { + int start = static_cast(batch_indices[i]); + int end = static_cast(batch_indices[i + 1]); + int current_timesteps = end - start; + for (int k = thy; k < current_timesteps; k += bly) { + T sum = 0; + for (int w = 0; (w < future_context) && ((k - w) >= 0); w++) { + sum += (wt[w * input_dim + d] * dout[(k + start - w) * input_dim + d]); + } + din[(k + start) * input_dim + d] = sum; + } + } +} + +// Compute W gradient (small future_context version) +template +__global__ void RowConvGradFilterImproved(const T *in, const T *dout, + int num_sequence, int input_dim, + int future_context, int block_x, + int block_y, + const size_t *batch_indices, + T *dfilter) { + int blx = blockDim.x; + int bly = blockDim.y; + int thx = threadIdx.x; + int thy = threadIdx.y; + int gx = blockIdx.x * blx; + int d = gx + thx; // index along input dim + + extern __shared__ T mem[]; + + int xdim_sh_in = block_y; + int xdim_sh_dout = block_y; + // int xdim_sh_dfilter = future_context; + int ydim_sh_in = block_x; + int ydim_sh_dout = block_x + future_context - 1; + int ydim_sh_dfilter = block_y; + + T *sh_in = mem; + T *sh_dout = &mem[xdim_sh_in * ydim_sh_in]; + T *sh_dfilter = &mem[xdim_sh_in * ydim_sh_in + xdim_sh_dout * ydim_sh_dout]; + + if (thy < future_context) { + sh_dfilter[thy * ydim_sh_dfilter + thx] = static_cast(0); + } + __syncthreads(); + + for (int i = 0; i < num_sequence; i++) { + int start = static_cast(batch_indices[i]); + int end = static_cast(batch_indices[i + 1]); + int current_timesteps = end - start; + int scaled_cur_steps = + ((current_timesteps + block_x - 1) / block_x) * block_x; + + for (int k = thy; k < scaled_cur_steps; k += block_x) { + int pos = start + k; + sh_in[thx * ydim_sh_in + thy] = + (d < input_dim && pos < end) ? in[pos * input_dim + d] : T(0); + sh_dout[thx * ydim_sh_dout + thy + future_context - 1] = + (d < input_dim && pos < end) ? dout[pos * input_dim + d] : T(0); + __syncthreads(); + + if (thy < future_context - 1) { + int pos_offset = pos - future_context + 1; + sh_dout[thx * ydim_sh_dout + thy] = + (d < input_dim && pos_offset >= start) + ? dout[pos_offset * input_dim + d] + : T(0); + } + __syncthreads(); + + for (int w = 0; w < future_context; w++) { + T val = sh_in[thy * ydim_sh_in + thx] * + sh_dout[thy * ydim_sh_dout + thx + future_context - 1 - w]; + __syncthreads(); + + for (int offset = 16; offset > 0; + offset = offset / 2) { // blockDim.x is 32. + val += __shfl_down(val, offset); + } + __syncthreads(); + + if (thx == 0) { + sh_dfilter[w * ydim_sh_dfilter + thy] += val; + } + __syncthreads(); + } + } + } + for (int w = thy; (w < future_context) && (d < input_dim); w += bly) { + dfilter[w * input_dim + d] += sh_dfilter[w * ydim_sh_dfilter + thx]; + } +} + +// Compute weight(filter) gradient +template +__global__ void RowConvGradFilter(const T *in, const T *dout, int num_sequence, + int input_dim, int future_context, + int block_x, int block_y, + const size_t *batch_indices, T *dfilter) { + int blx = blockDim.x; + int thx = threadIdx.x; + int thy = threadIdx.y; + int gx = blockIdx.x * blx; + int d = gx + thx; // index along input dim + extern __shared__ T mem[]; + T *sh_in = mem; + T *sh_dout = &mem[block_x * block_y]; + + for (int i = 0; i < num_sequence; i++) { + int start = static_cast(batch_indices[i]); + int end = static_cast(batch_indices[i + 1]); + int current_timesteps = end - start; + int scaled_cur_steps = + ((current_timesteps + block_x - 1) / block_x) * block_x; + + for (int k = thy; k < scaled_cur_steps; k += block_x) { + int pos = start + k; + sh_in[thx * block_y + thy] = + (d < input_dim && pos < end) ? in[pos * input_dim + d] : 0.0; + __syncthreads(); + + for (int w = 0; w < future_context; w++) { + sh_dout[thx * block_y + thy] = + (d < input_dim && (k - w) >= 0 && (k - w) < current_timesteps) + ? dout[(pos - w) * input_dim + d] + : 0.0; + __syncthreads(); + + T val = sh_in[thy * block_y + thx] * sh_dout[thy * block_y + thx]; + __syncthreads(); + + for (int offset = 16; offset > 0; + offset = offset / 2) { // blockDim.x is 32. + val += __shfl_down(val, offset); + } + __syncthreads(); + + if (thx == 0 && (gx + thy) < input_dim) { + dfilter[w * input_dim + gx + thy] += val; + } + } + } + } +} + +} // namespace + +template +class RowConvKernel + : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &context) const override { + auto *X = context.Input("X"); + auto *Filter = context.Input("Filter"); + auto *Out = context.Output("Out"); + + const T *in = X->data(); + const T *weight = Filter->data(); + T *out = Out->mutable_data(context.GetPlace()); + + auto batch_indices = X->lod()[0]; + int input_dim = X->dims()[1]; + int num_sequence = batch_indices.size() - 1; + int future_context = Filter->dims()[0]; + size_t *idx = batch_indices.data(); + auto stream = context.cuda_device_context().stream(); + + if (future_context <= 32) { + dim3 block_dim = dim3(32, 32); + dim3 grid_dim = dim3(DivUp(input_dim, block_dim.x), 1); + int mem_per_block = (future_context * block_dim.x) * sizeof(T); + RowConvForwardSharedMemory< + T><<>>( + in, weight, num_sequence, input_dim, future_context, idx, out); + } else { + dim3 block_dim = dim3(32, 32); + dim3 grid_dim = dim3(DivUp(input_dim, block_dim.x), 1); + RowConvForward<<>>( + in, weight, num_sequence, input_dim, future_context, idx, out); + } + } +}; + +template +class RowConvGradKernel + : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &context) const override { + auto *X = context.Input("X"); + auto *Filter = context.Input("Filter"); + auto *dOut = context.Input(framework::GradVarName("Out")); + const T *in = X->data(); + const T *weights = Filter->data(); + const T *dout = dOut->data(); + + Tensor *dX = context.Output(framework::GradVarName("X")); + Tensor *dFilter = context.Output(framework::GradVarName("Filter")); + + auto batch_indices = X->lod()[0]; + int input_dim = X->dims()[1]; + int num_sequence = batch_indices.size() - 1; + int future_context = Filter->dims()[0]; + size_t *idx = batch_indices.data(); + + auto &device_ctx = context.cuda_device_context(); + math::SetConstant zero; + + if (dFilter) { + T *dfilter = dFilter->mutable_data(context.GetPlace()); + zero(device_ctx, dFilter, static_cast(0.0)); + + if (future_context <= 32) { + dim3 block_dim = dim3(32, 32); + dim3 grid_dim = dim3(DivUp(input_dim, block_dim.x), 1); + int block_x = block_dim.x; + int block_y = block_dim.y; + int mem_per_block = + (block_y * block_x + block_y * (block_x + future_context - 1) + + future_context * block_y) * + sizeof(T); + RowConvGradFilterImproved< + T><<>>( + in, dout, num_sequence, input_dim, future_context, block_x, block_y, + idx, dfilter); + } else { + dim3 block_dim = dim3(32, 32); + dim3 grid_dim = dim3(DivUp(input_dim, block_dim.x), 1); + int block_x = block_dim.x; + int block_y = block_dim.y; + int mem_per_block = + (block_x * block_y * 2) * sizeof(T); // For 2 arrays of size 32x32 + RowConvGradFilter< + T><<>>( + in, dout, num_sequence, input_dim, future_context, block_x, block_y, + idx, dfilter); + } + } + + if (dX) { + T *din = dX->mutable_data(context.GetPlace()); + if (future_context <= 32) { + dim3 block_dim = dim3(32, 32); + dim3 grid_dim = dim3(DivUp(input_dim, block_dim.x), 1); + int mem_per_block = (future_context * block_dim.x) * sizeof(T); + RowConvGradInputSharedMemory< + T><<>>( + dout, weights, num_sequence, input_dim, future_context, idx, din); + } else { + dim3 block_dim = dim3(32, 32); + dim3 grid_dim = dim3(DivUp(input_dim, block_dim.x), 1); + RowConvGradInput<<>>( + dout, weights, num_sequence, input_dim, future_context, idx, din); + } + } + } +}; +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP_CUDA_KERNEL( + row_conv, ops::RowConvKernel); +REGISTER_OP_CUDA_KERNEL( + row_conv_grad, + ops::RowConvGradKernel); diff --git a/paddle/operators/row_conv_op.h b/paddle/operators/row_conv_op.h new file mode 100644 index 0000000000000000000000000000000000000000..10d435ab080851713ee08a491c43aad1549f6fbb --- /dev/null +++ b/paddle/operators/row_conv_op.h @@ -0,0 +1,33 @@ +/* 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 +#include "paddle/framework/op_registry.h" + +namespace paddle { +namespace operators { + +template +class RowConvKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &context) const override; +}; + +template +class RowConvGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &context) const override; +}; +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/save_load_op_test.cc b/paddle/operators/save_load_op_test.cc index a57466a48d4d6016fe2618d19fdca4c4f667124a..40103d864fb58804b39ca5f3c63e802a430ce886 100644 --- a/paddle/operators/save_load_op_test.cc +++ b/paddle/operators/save_load_op_test.cc @@ -1,16 +1,16 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +/* 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 +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 + 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. */ +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 "gtest/gtest.h" #include "paddle/framework/op_registry.h" @@ -21,7 +21,7 @@ USE_NO_KERNEL_OP(load); TEST(SaveLoadOp, CPU) { paddle::framework::Scope scope; paddle::platform::CPUPlace place; - paddle::platform::CPUDeviceContext ctx(place); + auto var = scope.Var("test_var"); auto tensor = var->GetMutable(); tensor->Resize({10, 10}); @@ -42,13 +42,13 @@ TEST(SaveLoadOp, CPU) { auto save_op = paddle::framework::OpRegistry::CreateOp( "save", {{"X", {"test_var"}}}, {}, attrs); - save_op->Run(scope, ctx); + 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, ctx); + load_op->Run(scope, place); int* actual = target->data(); for (int64_t i = 0; i < tensor->numel(); ++i) { EXPECT_EQ(expect[i], actual[i]); diff --git a/paddle/operators/save_op.cc b/paddle/operators/save_op.cc index 56909fb65f44ad00314103e21bee9535fbd59317..4b1cbe88836e340c94f797806243a6768410ed3d 100644 --- a/paddle/operators/save_op.cc +++ b/paddle/operators/save_op.cc @@ -1,16 +1,16 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +/* 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 +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 + 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. */ +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 @@ -21,6 +21,7 @@ #include "paddle/framework/framework.pb.h" #include "paddle/framework/lod_tensor.h" #include "paddle/framework/op_registry.h" +#include "paddle/platform/device_context.h" namespace paddle { namespace operators { @@ -62,7 +63,7 @@ class SaveOp : public framework::OperatorBase { const framework::AttributeMap &attrs) : OperatorBase(type, inputs, outputs, attrs) {} void Run(const framework::Scope &scope, - const platform::DeviceContext &dev_ctx) const override { + const platform::Place &place) const override { auto filename = Attr("file_path"); auto overwrite = Attr("overwrite"); @@ -89,79 +90,17 @@ class SaveOp : public framework::OperatorBase { auto &tensor = var->Get(); - { // the 1st field, uint32_t version - constexpr uint32_t version = 0; - fout.write(reinterpret_cast(&version), sizeof(version)); - } - { // the 2nd field, tensor description - // int32_t size - // void* protobuf message - framework::TensorDesc desc; - desc.set_data_type(framework::ToDataType(tensor.type())); - auto dims = framework::vectorize(tensor.dims()); - auto *pb_dims = desc.mutable_dims(); - pb_dims->Resize(static_cast(dims.size()), 0); - std::copy(dims.begin(), dims.end(), pb_dims->begin()); - int32_t size = desc.ByteSize(); - fout.write(reinterpret_cast(&size), sizeof(size)); - auto out = desc.SerializeAsString(); - fout.write(out.data(), size); - } - { // the 3rd field, tensor data - uint64_t size = tensor.memory_size(); - auto *data_ptr = tensor.data(); - PADDLE_ENFORCE(size < std::numeric_limits::max(), - "Index overflow when writing tensor"); - if (platform::is_gpu_place(tensor.place())) { -#ifdef PADDLE_WITH_CUDA - constexpr size_t kBufSize = 1024 * 1024 * 64; // 64MB - std::unique_ptr buf(new char[kBufSize]); - auto &gpu_dev_ctx = - static_cast(dev_ctx); - platform::CPUPlace cpu; - uintptr_t data = reinterpret_cast(data_ptr); - while (size != 0) { - size_t size_to_write = std::min(kBufSize, static_cast(size)); - memory::Copy(cpu, buf.get(), - boost::get(tensor.place()), - reinterpret_cast(data), size_to_write, - gpu_dev_ctx.stream()); - gpu_dev_ctx.Wait(); - fout.write(buf.get(), size_to_write); - data += size_to_write; - size -= size_to_write; - } -#else - PADDLE_THROW("Unexpected branch"); -#endif - } else { - fout.write(static_cast(data_ptr), - static_cast(size)); - } - } - { // the 4th field, lod information - // uint64_t lod_level - // uint64_t lod_level_1 size in byte. - // int* lod_level_1 data - // ... - auto lod = tensor.lod(); - uint64_t size = lod.size(); - fout.write(reinterpret_cast(&size), sizeof(size)); - - for (auto &each : lod) { - size = each.size() * sizeof(framework::LoD::value_type::value_type); - fout.write(reinterpret_cast(&size), sizeof(size)); - fout.write(reinterpret_cast(each.data()), - static_cast(size)); - } - } + // get device context from pool + platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance(); + auto &dev_ctx = *pool.Get(place); + + framework::SerializeToStream(fout, tensor, dev_ctx); } }; class SaveOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - SaveOpProtoMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + SaveOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "(Tensor ) Input tensor to be saved"); AddComment(R"DOC( diff --git a/paddle/operators/scale_op.cc b/paddle/operators/scale_op.cc index 5745580504fb9bda551f21665bff5c65ae82aeb9..f634ebe9a2a4648bd08f00af635ef22e8d86a8de 100644 --- a/paddle/operators/scale_op.cc +++ b/paddle/operators/scale_op.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/operators/scale_op.h" #include "paddle/operators/net_op.h" @@ -38,7 +38,7 @@ class ScaleOp : public framework::OperatorWithKernel { template class ScaleOpMaker : public framework::OpProtoAndCheckerMaker { public: - ScaleOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + ScaleOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "(Tensor) Input tensor of scale operator."); AddOutput("Out", "(Tensor) Output tensor of scale operator."); @@ -58,13 +58,13 @@ class ScaleGradMaker : public framework::SingleGradOpDescMaker { public: using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; - std::unique_ptr Apply() const override { - auto *grad_op = new framework::OpDescBind(); + std::unique_ptr Apply() const override { + auto *grad_op = new framework::OpDesc(); grad_op->SetType("scale"); grad_op->SetInput("X", OutputGrad("Out")); grad_op->SetOutput("Out", InputGrad("X")); grad_op->SetAttr("scale", GetAttr("scale")); - return std::unique_ptr(grad_op); + return std::unique_ptr(grad_op); } }; @@ -75,6 +75,8 @@ namespace ops = paddle::operators; REGISTER_OPERATOR(scale, ops::ScaleOp, ops::ScaleOpMaker, ops::ScaleGradMaker); -REGISTER_OP_CPU_KERNEL(scale, - ops::ScaleKernel, - ops::ScaleKernel); +REGISTER_OP_CPU_KERNEL( + scale, ops::ScaleKernel, + ops::ScaleKernel, + ops::ScaleKernel, + ops::ScaleKernel); diff --git a/paddle/operators/scale_op.cu b/paddle/operators/scale_op.cu index 820fd4e6855bb192ec3292ea6983d5ecae73b6e6..7202c0de707ff0b0b3ad966d9d1d3a7c0a89e880 100644 --- a/paddle/operators/scale_op.cu +++ b/paddle/operators/scale_op.cu @@ -1,19 +1,23 @@ /* 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 +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 + 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. */ +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/operators/scale_op.h" -REGISTER_OP_GPU_KERNEL( - scale, paddle::operators::ScaleKernel, - paddle::operators::ScaleKernel); +REGISTER_OP_CUDA_KERNEL( + scale, + paddle::operators::ScaleKernel, + paddle::operators::ScaleKernel, + paddle::operators::ScaleKernel, + paddle::operators::ScaleKernel); diff --git a/paddle/operators/scale_op.h b/paddle/operators/scale_op.h index 4931294c9d3661f4c53798bd0895a5cd38ae4501..395268c2eee40c187f5d211317ca8b28d35a71e0 100644 --- a/paddle/operators/scale_op.h +++ b/paddle/operators/scale_op.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 @@ -19,7 +19,7 @@ namespace paddle { namespace operators { -template +template class ScaleKernel : public framework::OpKernel { public: virtual void Compute(const framework::ExecutionContext& context) const { @@ -31,7 +31,8 @@ class ScaleKernel : public framework::OpKernel { auto eigen_out = framework::EigenVector::Flatten(*tensor); auto eigen_in = framework::EigenVector::Flatten(*in); - auto& dev = context.GetEigenDevice(); + auto& dev = + *context.template device_context().eigen_device(); eigen_out.device(dev) = scale * eigen_in; } }; diff --git a/paddle/operators/scatter.cu.h b/paddle/operators/scatter.cu.h index d95436be4f25b9df4aaef57ddb249ecf944f0666..55555300fc3219c0651583d8540b47189c8d3f13 100644 --- a/paddle/operators/scatter.cu.h +++ b/paddle/operators/scatter.cu.h @@ -1,16 +1,16 @@ -/* Copyright (c) 2016 PaddlePaddle Authors All Rights Reserve. +/* 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 +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 + 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. */ +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/framework/tensor.h" diff --git a/paddle/operators/scatter_op.cc b/paddle/operators/scatter_op.cc index ce4b794bc35aca0912d89a4ae81a9aa0c73a2104..806dccc6ca78bf64da828fe13e08e043097bd939 100644 --- a/paddle/operators/scatter_op.cc +++ b/paddle/operators/scatter_op.cc @@ -49,7 +49,7 @@ class ScatterOp : public framework::OperatorWithKernel { } protected: - framework::OpKernelType GetKernelType( + framework::OpKernelType GetActualKernelType( const framework::ExecutionContext& ctx) const override { return framework::OpKernelType( framework::ToDataType(ctx.Input("Ref")->type()), @@ -68,7 +68,7 @@ class ScatterGradOp : public framework::OperatorWithKernel { } protected: - framework::OpKernelType GetKernelType( + framework::OpKernelType GetActualKernelType( const framework::ExecutionContext& ctx) const override { return framework::OpKernelType( framework::ToDataType(ctx.Input("Ref")->type()), @@ -78,8 +78,7 @@ class ScatterGradOp : public framework::OperatorWithKernel { class ScatterOpMaker : public framework::OpProtoAndCheckerMaker { public: - ScatterOpMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker) + ScatterOpMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("Ref", "The source input of scatter op"); AddInput("Index", @@ -87,10 +86,15 @@ class ScatterOpMaker : public framework::OpProtoAndCheckerMaker { AddInput("Updates", "The updated value of updates op"); AddOutput("Out", "The output of add op"); AddComment(R"DOC( -Scatter Operator by selecting from the first axis, +Scatter Operator. -Out = Ref +This operator obtains output by updating the input on selected indices on the first axis: + +$$ +Out = Ref \\ Out[Index] = Ref[Index] + Updates +$$ + )DOC"); } }; diff --git a/paddle/operators/scatter_op.cu b/paddle/operators/scatter_op.cu index 3b32ae2fb77a5d3d4c558742ec469c74d15eee07..0c198d225890882ab6697d3a8b3d17e034c06cc4 100644 --- a/paddle/operators/scatter_op.cu +++ b/paddle/operators/scatter_op.cu @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 "gather.cu.h" #include "paddle/operators/gather_op.h" @@ -59,5 +59,5 @@ class ScatterGradOpCUDAKernel : public framework::OpKernel { } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(scatter, ops::ScatterOpCUDAKernel); -REGISTER_OP_GPU_KERNEL(scatter_grad, ops::ScatterGradOpCUDAKernel); +REGISTER_OP_CUDA_KERNEL(scatter, ops::ScatterOpCUDAKernel); +REGISTER_OP_CUDA_KERNEL(scatter_grad, ops::ScatterGradOpCUDAKernel); diff --git a/paddle/operators/send_op.cc b/paddle/operators/send_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..95c207221a7b34732eca4cfd07fed0a8f1671981 --- /dev/null +++ b/paddle/operators/send_op.cc @@ -0,0 +1,101 @@ +/* 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. */ + +#include + +#include "paddle/framework/data_type.h" +#include "paddle/framework/framework.pb.h" +#include "paddle/framework/lod_tensor.h" +#include "paddle/framework/op_registry.h" + +#include "paddle/operators/detail/send_recv_impl.h" +#include "paddle/operators/detail/simple_block_queue.h" + +namespace paddle { +namespace operators { + +// TODO(typhoonzero): this is a simple implementation which only send +// one tensor +class SendOp : public framework::OperatorBase { + public: + SendOp(const std::string &type, const framework::VariableNameMap &inputs, + const framework::VariableNameMap &outputs, + const framework::AttributeMap &attrs) + : OperatorBase(type, inputs, outputs, attrs) { + // init client when the operator is created at runtime. + std::vector endpoints = + Attr>("endpoints"); + for (auto ep : endpoints) { + client_map_[ep].reset(new detail::RPCClient( + grpc::CreateChannel(ep, grpc::InsecureChannelCredentials()))); + } + } + + void Run(const framework::Scope &scope, + const platform::Place &dev_place) const override { + auto ins = Inputs("X"); + auto outs = Outputs("Out"); + std::vector epmap = Attr>("epmap"); + // TODO(typhoonzero): use async calls to send multiple variable asyncly. + for (size_t i = 0; i < ins.size(); ++i) { + bool ret = client_map_[epmap[i]]->SendVariable(scope, ins[i]); + if (!ret) { + LOG(ERROR) << "send variable error: " << ins[i]; + } + } + // TODO(typhoonzero): support async optimization + client_map_[epmap[0]]->Wait(); + for (size_t i = 0; i < outs.size(); ++i) { + bool ret = client_map_[epmap[i]]->GetVariable(scope, outs[i]); + if (!ret) { + LOG(ERROR) << "GetVariable error: " << outs[i]; + } + } + } + + protected: + mutable std::unordered_map> + client_map_; +}; + +class SendOpMaker : public framework::OpProtoAndCheckerMaker { + public: + SendOpMaker(OpProto *proto, OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "(Tensor) Input tensor to be send").AsDuplicable(); + AddOutput("Out", "(Tensor) Output tensor to get from server") + .AsDuplicable(); + AddComment(R"DOC( +Recv operator + +This operator will send tensor to recv_op. +)DOC"); + AddAttr>("endpoints", + "(string vector, default 127.0.0.1:6164)" + "Server endpoints to send variables to.") + .SetDefault({}); + AddAttr>("epmap", + "(string vector, default 127.0.0.1:6164)" + "Server endpoints in the order of input " + "variables for mapping") + .SetDefault({}); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; + +REGISTER_OPERATOR(send, ops::SendOp, ops::SendOpMaker); diff --git a/paddle/operators/send_recv_op_test.cc b/paddle/operators/send_recv_op_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..108e2dec6b3eecadd431fd25f9a31ec17a874b6b --- /dev/null +++ b/paddle/operators/send_recv_op_test.cc @@ -0,0 +1,131 @@ +/* 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. */ + +#include +#include +#include + +#include "gtest/gtest.h" +#include "paddle/framework/op_registry.h" +#include "paddle/framework/operator.h" +#include "paddle/framework/program_desc.h" +#include "paddle/string/printf.h" + +USE_NO_KERNEL_OP(send); +USE_NO_KERNEL_OP(recv); +USE_OP(sum); + +// global for simplicity. +std::unique_ptr recv_op; + +void InitTensorsInScope(paddle::framework::Scope &scope, + paddle::platform::CPUPlace &place) { + paddle::platform::CPUDeviceContext ctx(place); + for (int i = 0; i < 2; ++i) { + auto var_name = paddle::string::Sprintf("x%d", i); + auto var = scope.Var(var_name); + auto tensor = var->GetMutable(); + tensor->Resize({10, 10}); + float *expect = tensor->mutable_data(place); + for (int64_t i = 0; i < tensor->numel(); ++i) { + expect[i] = static_cast(i); + } + } + + auto out_var = scope.Var("Out"); + auto out_tensor = out_var->GetMutable(); + out_tensor->Resize({10, 10}); + out_tensor->mutable_data(place); // allocate +} + +void AddOp(const std::string &type, + const paddle::framework::VariableNameMap &inputs, + const paddle::framework::VariableNameMap &outputs, + paddle::framework::AttributeMap attrs, + paddle::framework::BlockDesc *block) { + // insert output + for (auto kv : outputs) { + for (auto v : kv.second) { + auto var = block->Var(v); + var->SetDataType(paddle::framework::proto::DataType::FP32); + } + } + + // insert op + auto op = block->AppendOp(); + op->SetType(type); + for (auto &kv : inputs) { + op->SetInput(kv.first, kv.second); + } + for (auto &kv : outputs) { + op->SetOutput(kv.first, kv.second); + } + op->SetAttrMap(attrs); +} + +void StartServerNet() { + paddle::framework::Scope scope; + paddle::platform::CPUPlace place; + InitTensorsInScope(scope, place); + + // sub program run in recv_op, for simple test we use sum + paddle::framework::ProgramDesc program; + paddle::framework::BlockDesc *block = program.MutableBlock(0); + // X for server side tensors, RX for received tensers, must be of same shape. + AddOp("sum", {{"X", {"x0", "x1"}}}, {{"Out", {"x0"}}}, {}, block); + + paddle::framework::AttributeMap attrs; + attrs.insert({"endpoint", std::string("127.0.0.1:6174")}); + attrs.insert({"ParamList", std::vector({"x0"})}); + attrs.insert({"GradList", std::vector({"x1"})}); + std::string program_proto; + PADDLE_ENFORCE(program.Proto()->SerializeToString(&program_proto)); + + attrs.insert({"OptimizeProgram", program_proto}); + recv_op = paddle::framework::OpRegistry::CreateOp("recv", {{"RX", {"x1"}}}, + {}, attrs); + recv_op->Run(scope, place); +} + +TEST(SendRecvOp, CPU) { + std::thread server_thread(StartServerNet); + sleep(5); // wait server to start + // local net + paddle::framework::Scope scope; + paddle::platform::CPUPlace place; + InitTensorsInScope(scope, place); + + paddle::framework::AttributeMap attrs; + attrs.insert({"endpoints", std::vector({"127.0.0.1:6174"})}); + attrs.insert({"epmap", std::vector({"127.0.0.1:6174"})}); + auto send_op = paddle::framework::OpRegistry::CreateOp( + "send", {{"X", {"x1"}}}, {{"Out", {"x0"}}}, attrs); + send_op->Run(scope, place); + + auto in_var = scope.Var("x1"); + auto tensor = in_var->GetMutable(); + float *expected = tensor->data(); + auto out_var = scope.Var("x0"); + auto target = out_var->GetMutable(); + // x1 * 2 == x0 + EXPECT_NE(target->memory_size(), size_t(0)); + float *actual = target->data(); + for (int64_t i = 0; i < target->numel(); ++i) { + EXPECT_EQ(expected[i] * 2, actual[i]); + } + + recv_op->Stop(); + server_thread.join(); + // recv_op.reset(); +} diff --git a/paddle/operators/seq_expand_op.cu b/paddle/operators/seq_expand_op.cu deleted file mode 100644 index f1e4b82a76e628c4d9fb83bc93f3dcfd2f98ea5b..0000000000000000000000000000000000000000 --- a/paddle/operators/seq_expand_op.cu +++ /dev/null @@ -1,23 +0,0 @@ -/* 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. */ - -#define EIGEN_USE_GPU -#include "paddle/operators/seq_expand_op.h" - -namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(seq_expand, - ops::SeqExpandKernel); -REGISTER_OP_GPU_KERNEL( - seq_expand_grad, - ops::SeqExpandGradKernel); diff --git a/paddle/operators/sequence_concat_op.cc b/paddle/operators/sequence_concat_op.cc index d1de0b444712a8c304c33bd194e306dfe3c41f02..2f0aad2003e48952ca26ca27573bc45386a4e585 100644 --- a/paddle/operators/sequence_concat_op.cc +++ b/paddle/operators/sequence_concat_op.cc @@ -43,8 +43,7 @@ class SequenceConcatOp : public framework::OperatorWithKernel { class SequenceConcatOpMaker : public framework::OpProtoAndCheckerMaker { public: - SequenceConcatOpMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker) + SequenceConcatOpMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "(LodTensorArray) Input is a vector of LoDTensor, " @@ -68,12 +67,12 @@ class SequenceConcatOpMaker : public framework::OpProtoAndCheckerMaker { "The level should be less than the level number of inputs.") .SetDefault(0); AddComment(R"DOC( -The sequence_concat operator concatenates multiple LoDTensors. -It only supports sequence (LoD Tensor with level number is 1) +The sequence_concat operator concatenates multiple LoDTensors. +It only supports sequence (LoD Tensor with level number is 1) or a nested sequence (LoD tensor with level number is 2) as its input. - Case1: If the axis is other than 0(here, axis is 1 and level is 1), - each input should have the same LoD information and the LoD + each input should have the same LoD information and the LoD information of the output keeps the same as the input. LoD(x0) = {{0,2,4}, {0,1,2,3,4}}; Dims(x0) = (4,3,4) @@ -81,7 +80,7 @@ or a nested sequence (LoD tensor with level number is 2) as its input. LoD(Out) = {{0,2,4}, {0,1,2,3,4}}; Dims(Out) = (4,7,4) - Case2: - If the axis is 0(here, leve is 0), the inputs are concatenated along + If the axis is 0(here, leve is 0), the inputs are concatenated along time steps, the LoD information of the output need to re-compute. The LoD information of level-1 should be same. @@ -125,11 +124,12 @@ class SequenceConcatGradOp : public framework::OperatorWithKernel { } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP(sequence_concat, ops::SequenceConcatOp, ops::SequenceConcatOpMaker, - sequence_concat_grad, ops::SequenceConcatGradOp); +REGISTER_OP_EX(sequence_concat, ops::SequenceConcatOp, + ops::SequenceConcatOpMaker, sequence_concat_grad, + ops::SequenceConcatGradOp, false); REGISTER_OP_CPU_KERNEL( sequence_concat, - ops::SequenceConcatOpKernel); + ops::SequenceConcatOpKernel); REGISTER_OP_CPU_KERNEL( sequence_concat_grad, - ops::SequenceConcatGradOpKernel); + ops::SequenceConcatGradOpKernel); diff --git a/paddle/operators/sequence_concat_op.cu.cc b/paddle/operators/sequence_concat_op.cu.cc index 9ca99c2258f547e6f9c23be0d394bc3ea2bb6678..144bdb5af635b0cb75bcd1f654700041186dae46 100644 --- a/paddle/operators/sequence_concat_op.cu.cc +++ b/paddle/operators/sequence_concat_op.cu.cc @@ -15,9 +15,9 @@ limitations under the License. */ #include "paddle/operators/sequence_concat_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL( +REGISTER_OP_CUDA_KERNEL( sequence_concat, - ops::SequenceConcatOpKernel); -REGISTER_OP_GPU_KERNEL( - sequence_concat_grad, - ops::SequenceConcatGradOpKernel); + ops::SequenceConcatOpKernel); +REGISTER_OP_CUDA_KERNEL(sequence_concat_grad, + ops::SequenceConcatGradOpKernel< + paddle::platform::CUDADeviceContext, float>); diff --git a/paddle/operators/sequence_concat_op.h b/paddle/operators/sequence_concat_op.h index 09212070aa90b0f080f6140a312924229162aaec..8445224f46aba6110280783c9080ed4691266b8b 100644 --- a/paddle/operators/sequence_concat_op.h +++ b/paddle/operators/sequence_concat_op.h @@ -59,7 +59,7 @@ LoD ConcatLoD(const std::vector ins, const size_t level) { return out_lod; } -template +template class SequenceConcatOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { @@ -119,7 +119,7 @@ class SequenceConcatOpKernel : public framework::OpKernel { } }; -template +template class SequenceConcatGradOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { diff --git a/paddle/operators/sequence_conv_op.cc b/paddle/operators/sequence_conv_op.cc index 41cadce4c603a9c14db79e2f6b30f8664cf72a38..c5b7c81bd7c6e1110aa9e2ced629bea5d88832d1 100644 --- a/paddle/operators/sequence_conv_op.cc +++ b/paddle/operators/sequence_conv_op.cc @@ -100,8 +100,7 @@ class SequenceConvGradOp : public framework::OperatorWithKernel { class SequenceConvOpMaker : public framework::OpProtoAndCheckerMaker { public: - SequenceConvOpMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker) + SequenceConvOpMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput( "X", @@ -179,7 +178,10 @@ REGISTER_OP(sequence_conv, ops::SequenceConvOp, ops::SequenceConvOpMaker, sequence_conv_grad, ops::SequenceConvGradOp); REGISTER_OP_CPU_KERNEL( - sequence_conv, ops::SequenceConvKernel); + sequence_conv, + ops::SequenceConvKernel, + ops::SequenceConvKernel); REGISTER_OP_CPU_KERNEL( sequence_conv_grad, - ops::SequenceConvGradKernel); + ops::SequenceConvGradKernel, + ops::SequenceConvGradKernel); diff --git a/paddle/operators/sequence_conv_op.cu.cc b/paddle/operators/sequence_conv_op.cu.cc index 6106b0e46c0ab96e01dfc344055f23dbf4a1a2c3..0b8f2c695564f19cf71ecc56a60e707c3703af36 100644 --- a/paddle/operators/sequence_conv_op.cu.cc +++ b/paddle/operators/sequence_conv_op.cu.cc @@ -1,22 +1,25 @@ /* 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 +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 + 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. */ +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/operators/sequence_conv_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL( - sequence_conv, ops::SequenceConvKernel); -REGISTER_OP_GPU_KERNEL( +REGISTER_OP_CUDA_KERNEL( + sequence_conv, + ops::SequenceConvKernel, + ops::SequenceConvKernel); +REGISTER_OP_CUDA_KERNEL( sequence_conv_grad, - ops::SequenceConvGradKernel); + ops::SequenceConvGradKernel, + ops::SequenceConvGradKernel); diff --git a/paddle/operators/sequence_conv_op.h b/paddle/operators/sequence_conv_op.h index b8fbe2647c4338a2fa16aa655ebab64dd8d5417d..bb584b7bfa5fb8f6eb0a452468d24ca034be6f1b 100644 --- a/paddle/operators/sequence_conv_op.h +++ b/paddle/operators/sequence_conv_op.h @@ -23,7 +23,7 @@ namespace operators { using Tensor = framework::Tensor; using LoDTensor = framework::LoDTensor; -template +template class SequenceConvKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { @@ -56,21 +56,23 @@ class SequenceConvKernel : public framework::OpKernel { Tensor col; col.mutable_data(col_shape, context.GetPlace()); // Because if padding_trainable is false, padding data should be zeros. - math::SetConstant set_zero; - set_zero(context.device_context(), &col, static_cast(0)); + math::SetConstant set_zero; + auto& dev_ctx = context.template device_context(); + set_zero(dev_ctx, &col, static_cast(0)); - math::ContextProjectFunctor seq_project_functor; + math::ContextProjectFunctor seq_project_functor; - seq_project_functor(context.device_context(), *in, *padding_data, - padding_trainable, context_start, context_length, - context_stride, up_pad, down_pad, &col); + seq_project_functor(dev_ctx, *in, *padding_data, padding_trainable, + context_start, context_length, context_stride, up_pad, + down_pad, &col); - math::matmul(context.device_context(), col, false, filter, false, - static_cast(1.0), out, static_cast(0.0)); + math::matmul(dev_ctx, col, false, filter, false, + static_cast(1.0), out, + static_cast(0.0)); } }; -template +template class SequenceConvGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { @@ -95,7 +97,8 @@ class SequenceConvGradKernel : public framework::OpKernel { int down_pad = std::max(0, context_start + context_length - 1); int sequence_width = static_cast(in->dims()[1]); - math::SetConstant set_zero; + math::SetConstant set_zero; + auto& dev_ctx = context.template device_context(); // use col_shape in the im2col calculation framework::DDim col_shape = {in->dims()[0], sequence_width * context_length}; @@ -104,38 +107,36 @@ class SequenceConvGradKernel : public framework::OpKernel { if (in_g || filter_g || (padding_trainable && padding_data_g)) { col.mutable_data(col_shape, context.GetPlace()); // Because if padding_trainable is false, padding data should be zeros. - set_zero(context.device_context(), &col, static_cast(0)); - math::matmul(context.device_context(), *out_g, false, *filter, - true, T(1.0), &col, T(1.0)); + set_zero(dev_ctx, &col, static_cast(0)); + math::matmul(dev_ctx, *out_g, false, *filter, true, + T(1.0), &col, T(1.0)); } - math::ContextProjectFunctor seq_project_functor; - math::ContextProjectGradFunctor seq_project_grad_functor; + math::ContextProjectFunctor seq_project_functor; + math::ContextProjectGradFunctor seq_project_grad_functor; if (in_g) { in_g->mutable_data(context.GetPlace()); in_g->set_lod(in->lod()); - set_zero(context.device_context(), in_g, static_cast(0)); + set_zero(dev_ctx, in_g, static_cast(0)); - seq_project_grad_functor(context.device_context(), *in_g, - padding_trainable, context_start, context_length, - context_stride, up_pad, down_pad, false, true, - padding_data_g, &col); + seq_project_grad_functor(dev_ctx, *in_g, padding_trainable, context_start, + context_length, context_stride, up_pad, down_pad, + false, true, padding_data_g, &col); } if (padding_trainable && padding_data_g) { padding_data_g->mutable_data(context.GetPlace()); - set_zero(context.device_context(), padding_data_g, static_cast(0)); + set_zero(dev_ctx, padding_data_g, static_cast(0)); LoDTensor* input = const_cast(in); - seq_project_grad_functor(context.device_context(), *input, - padding_trainable, context_start, context_length, - context_stride, up_pad, down_pad, true, false, - padding_data_g, &col); + seq_project_grad_functor( + dev_ctx, *input, padding_trainable, context_start, context_length, + context_stride, up_pad, down_pad, true, false, padding_data_g, &col); } if (filter_g) { filter_g->mutable_data(context.GetPlace()); - set_zero(context.device_context(), filter_g, static_cast(0)); + set_zero(dev_ctx, filter_g, static_cast(0)); Tensor filter_grad = *filter_g; LoDTensor out_grad = *out_g; @@ -145,12 +146,12 @@ class SequenceConvGradKernel : public framework::OpKernel { padding_data = context.Input("PaddingData"); } - seq_project_functor(context.device_context(), *in, *padding_data, - padding_trainable, context_start, context_length, - context_stride, up_pad, down_pad, &col); + seq_project_functor(dev_ctx, *in, *padding_data, padding_trainable, + context_start, context_length, context_stride, up_pad, + down_pad, &col); - math::matmul(context.device_context(), col, true, out_grad, - false, T(1.0), &filter_grad, T(1.0)); + math::matmul(dev_ctx, col, true, out_grad, false, + T(1.0), &filter_grad, T(1.0)); } } }; diff --git a/paddle/operators/seq_expand_op.cc b/paddle/operators/sequence_expand_op.cc similarity index 72% rename from paddle/operators/seq_expand_op.cc rename to paddle/operators/sequence_expand_op.cc index b862056ad400290a60e8a75a23dceeb1d4422ea4..b40ec617e42110e0ab5168a8ac675adaf760fb3c 100644 --- a/paddle/operators/seq_expand_op.cc +++ b/paddle/operators/sequence_expand_op.cc @@ -1,25 +1,25 @@ /* 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 +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 + 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. */ +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/operators/seq_expand_op.h" +#include "paddle/operators/sequence_expand_op.h" namespace paddle { namespace operators { using framework::Tensor; -class SeqExpandOp : public framework::OperatorWithKernel { +class SequenceExpandOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; @@ -35,25 +35,24 @@ class SeqExpandOp : public framework::OperatorWithKernel { } }; -class SeqExpandOpMaker : public framework::OpProtoAndCheckerMaker { +class SequenceExpandOpMaker : public framework::OpProtoAndCheckerMaker { public: - SeqExpandOpMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker) + SequenceExpandOpMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "(Tensor or LoDTensor) The input(X) of this operator can be a " "LoDTensor or a base Tensor."); AddInput("Y", - "(LoDTensor)The reference input(Y) of seq_expand op." + "(LoDTensor)The reference input(Y) of sequence_expand op." "It must be a LoDTensor with k-level(k>0)." "The input(X) will be expanded according to LOD of input(Y)." "The element numbers of last level in input(Y) " "must be equal to dims[0] of input(X)."); AddOutput("Out", - "(LodTensor)The output of seq_expand op." + "(LodTensor)The output of sequence_expand op." "The lod of output will be as same as input(Y)'s lod."); AddComment(R"DOC( -Seq Expand Operator. +Sequence Expand Operator. This operator expands input(X) according to LOD of input(Y). Following are cases to better explain how this works: @@ -124,7 +123,7 @@ then we get 2-level LoDTensor } }; -class SeqExpandOpGrad : public framework::OperatorWithKernel { +class SequenceExpandOpGrad : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; @@ -146,10 +145,11 @@ class SeqExpandOpGrad : public framework::OperatorWithKernel { } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP(seq_expand, ops::SeqExpandOp, ops::SeqExpandOpMaker, - seq_expand_grad, ops::SeqExpandOpGrad); -REGISTER_OP_CPU_KERNEL(seq_expand, - ops::SeqExpandKernel); +REGISTER_OP(sequence_expand, ops::SequenceExpandOp, ops::SequenceExpandOpMaker, + sequence_expand_grad, ops::SequenceExpandOpGrad); REGISTER_OP_CPU_KERNEL( - seq_expand_grad, - ops::SeqExpandGradKernel); + sequence_expand, + ops::SequenceExpandKernel); +REGISTER_OP_CPU_KERNEL( + sequence_expand_grad, + ops::SequenceExpandGradKernel); diff --git a/paddle/operators/sequence_expand_op.cu b/paddle/operators/sequence_expand_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..0b9638b2ce60f73c95e1ccbcfb16cef7b5351073 --- /dev/null +++ b/paddle/operators/sequence_expand_op.cu @@ -0,0 +1,24 @@ +/* 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. */ + +#define EIGEN_USE_GPU +#include "paddle/operators/sequence_expand_op.h" + +namespace ops = paddle::operators; +REGISTER_OP_CUDA_KERNEL( + sequence_expand, + ops::SequenceExpandKernel); +REGISTER_OP_CUDA_KERNEL( + sequence_expand_grad, + ops::SequenceExpandGradKernel); diff --git a/paddle/operators/seq_expand_op.h b/paddle/operators/sequence_expand_op.h similarity index 73% rename from paddle/operators/seq_expand_op.h rename to paddle/operators/sequence_expand_op.h index 4ef0d02cf85c43e95335660be65a67df66b4f55c..2ba628e9c37278025e31779ab0468db46f2ff40a 100644 --- a/paddle/operators/seq_expand_op.h +++ b/paddle/operators/sequence_expand_op.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 @@ -23,8 +23,8 @@ namespace operators { using LoDTensor = framework::LoDTensor; -template -class SeqExpandKernel : public framework::OpKernel { +template +class SequenceExpandKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto* x = context.Input("X"); @@ -37,7 +37,8 @@ class SeqExpandKernel : public framework::OpKernel { "The size of last lod level in Input(Y)" "must be equal to dims[0] of Input(X)."); out->set_lod(y->lod()); - auto place = context.GetEigenDevice(); + auto* place = + context.template device_context().eigen_device(); size_t element_len = framework::product(x_dims) / x_dims[0]; T* out_data = out->mutable_data(context.GetPlace()); auto out_starts = out->lod().back(); @@ -50,7 +51,7 @@ class SeqExpandKernel : public framework::OpKernel { Eigen::TensorMap> out_t(out_data, scale, element_len); Eigen::array cast({{scale, 1}}); - out_t.device(place) = x_t.broadcast(cast); + out_t.device(*place) = x_t.broadcast(cast); x_data += element_len; out_data += element_len * scale; } @@ -69,8 +70,8 @@ class SeqExpandKernel : public framework::OpKernel { * Grad(X).lod = Input(X).lod * * */ -template -class SeqExpandGradKernel : public framework::OpKernel { +template +class SequenceExpandGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto* d_out = context.Input(framework::GradVarName("Out")); @@ -89,8 +90,9 @@ class SeqExpandGradKernel : public framework::OpKernel { d_out_t(d_out_data, static_cast(repeat), element_len); Eigen::TensorMap> d_x_t(d_x_data, static_cast(element_len)); - auto place = context.GetEigenDevice(); - d_x_t.device(place) = d_out_t.sum(Eigen::array({{0}})); + auto place = + context.template device_context().eigen_device(); + d_x_t.device(*place) = d_out_t.sum(Eigen::array({{0}})); d_out_data += (repeat * element_len); d_x_data += element_len; } diff --git a/paddle/operators/sequence_pool_op.cc b/paddle/operators/sequence_pool_op.cc index 2a000ac60b176737277605c3ac812ea65a0e03fc..aea98744d8fc1fc59a07250d57f76f26fb9f3634 100644 --- a/paddle/operators/sequence_pool_op.cc +++ b/paddle/operators/sequence_pool_op.cc @@ -37,8 +37,7 @@ class SequencePoolOp : public framework::OperatorWithKernel { class SequencePoolOpMaker : public framework::OpProtoAndCheckerMaker { public: - SequencePoolOpMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker) + SequencePoolOpMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "(LoDTensor) The variable-length input of SequencePoolOp"); AddOutput("Out", @@ -50,7 +49,7 @@ class SequencePoolOpMaker : public framework::OpProtoAndCheckerMaker { .AsIntermediate(); AddAttr( "pooltype", - "(int, default AVERAGE) the pooling pooltype of SequencePoolOp.") + "(string, default 'AVERAGE') the pooling pooltype of SequencePoolOp.") .SetDefault("AVERAGE") .InEnum({"AVERAGE", "SUM", "SQRT", "LAST", "FIRST", "MAX"}); AddComment(R"DOC( @@ -58,12 +57,12 @@ Sequence Pool Operator. The SequencePoolOp pools features of all time-steps of each instance. It supports six pooling types: -1. AVERAGE: Out[i] = $$avg(X_i)$$ -2. SUM: Out[i] = $$\sum_jX_{ij}$$ -3. SQRT: Out[i] = $$\frac{\sum_jX_{ij}}{\sqrt{len(X_i)}}$$ +1. AVERAGE: $$Out[i] = \frac{\sum_i X_i}{N}$$ +2. SUM: $$Out[i] = \sum_jX_{ij}$$ +3. SQRT: $$Out[i] = \frac{\sum_jX_{ij}}{\sqrt{len(X_i)}}$$ 4. LAST: Out[i] = last instance in i-th sequence X[i] 5. FIRST: Out[i] = first instance in i-th sequence X[i] -6. MAX: Out[i] = $$max(X_i)$$ +6. MAX: $$Out[i] = max(X_i)$$ The following example explains how this works: For a mini-batch of 3 variable-length sentences, @@ -104,10 +103,11 @@ class SequencePoolGradOp : public framework::OperatorWithKernel { PADDLE_ENFORCE_EQ(og_dims[i], x_dims[i], "The dimension mismatch."); } ctx->SetOutputDim(framework::GradVarName("X"), x_dims); + ctx->ShareLoD("X", framework::GradVarName("X")); } protected: - framework::OpKernelType GetKernelType( + framework::OpKernelType GetActualKernelType( const framework::ExecutionContext& ctx) const override { return framework::OpKernelType( framework::ToDataType(ctx.Input("X")->type()), @@ -122,7 +122,8 @@ namespace ops = paddle::operators; REGISTER_OP(sequence_pool, ops::SequencePoolOp, ops::SequencePoolOpMaker, sequence_pool_grad, ops::SequencePoolGradOp); REGISTER_OP_CPU_KERNEL( - sequence_pool, ops::SequencePoolKernel); + sequence_pool, + ops::SequencePoolKernel); REGISTER_OP_CPU_KERNEL( sequence_pool_grad, - ops::SequencePoolGradKernel); + ops::SequencePoolGradKernel); diff --git a/paddle/operators/sequence_pool_op.cu b/paddle/operators/sequence_pool_op.cu index 66850772d501f873cf754205c19e9d0c0090370a..265f695935236236f98c2dd2062072756e9c8b14 100644 --- a/paddle/operators/sequence_pool_op.cu +++ b/paddle/operators/sequence_pool_op.cu @@ -1,24 +1,25 @@ /* 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 +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 + 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. */ +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. */ #define EIGEN_USE_GPU #include "paddle/operators/sequence_pool_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL( - sequence_pool, ops::SequencePoolKernel); -REGISTER_OP_GPU_KERNEL( +REGISTER_OP_CUDA_KERNEL( + sequence_pool, + ops::SequencePoolKernel); +REGISTER_OP_CUDA_KERNEL( sequence_pool_grad, - ops::SequencePoolGradKernel); + ops::SequencePoolGradKernel); diff --git a/paddle/operators/sequence_pool_op.h b/paddle/operators/sequence_pool_op.h index 7f136d8cf0e1eaae7b4de32988b60ae8a5034cc6..7519aa1d7208b9832f7a3d3afbc59a2eb4e8e13a 100644 --- a/paddle/operators/sequence_pool_op.h +++ b/paddle/operators/sequence_pool_op.h @@ -30,7 +30,7 @@ template using EigenMatrix = framework::EigenMatrix; -template +template class SequencePoolKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { @@ -54,17 +54,18 @@ class SequencePoolKernel : public framework::OpKernel { auto lod_level_0 = lod[0]; out->mutable_data(context.GetPlace()); - + auto& dev_ctx = context.template device_context(); if (pooltype == "MAX") { - math::MaxSeqPoolFunctor max_pool; + math::MaxSeqPoolFunctor max_pool; auto* index = context.Output("MaxIndex"); index->Resize({dims}); index->mutable_data(context.GetPlace()); - max_pool(context.device_context(), *in, out, index); + max_pool(dev_ctx, *in, out, index); return; } - auto place = context.GetEigenDevice(); + auto& place = + *context.template device_context().eigen_device(); for (int i = 0; i < static_cast(lod_level_0.size()) - 1; ++i) { Tensor in_t = in->Slice(static_cast(lod_level_0[i]), static_cast(lod_level_0[i + 1])); @@ -91,7 +92,7 @@ class SequencePoolKernel : public framework::OpKernel { } }; -template +template class SequencePoolGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { @@ -105,20 +106,23 @@ class SequencePoolGradKernel : public framework::OpKernel { int64_t w = in->numel() / dims[0]; in_g->mutable_data(context.GetPlace()); + auto& dev_ctx = context.template device_context(); if (pooltype == "MAX") { - math::MaxSeqPoolGradFunctor max_pool_grad; + math::MaxSeqPoolGradFunctor max_pool_grad; auto* index = context.Input("MaxIndex"); - max_pool_grad(context.device_context(), *out_g, *index, in_g); + max_pool_grad(dev_ctx, *out_g, *index, in_g); return; } if (pooltype == "LAST" || pooltype == "FIRST") { // set X@Grad be zero at first when pooltype is LAST/FIRST - math::SetConstant functor; - functor(context.device_context(), in_g, 0); + math::SetConstant functor; + functor(dev_ctx, in_g, 0); } - auto place = context.GetEigenDevice(); + auto& place = + *context.template device_context().eigen_device(); + for (int i = 0; i < static_cast(lod.size()) - 1; ++i) { auto in_g_t = in_g->Slice(static_cast(lod[i]), static_cast(lod[i + 1])); diff --git a/paddle/operators/sequence_slice_op.cc b/paddle/operators/sequence_slice_op.cc old mode 100755 new mode 100644 index cbe0b4233160dd1f3ebdf6db8b5f6df392efdfe7..98bd8854903e5abf6d27432a2af0aaae980c0b1d --- a/paddle/operators/sequence_slice_op.cc +++ b/paddle/operators/sequence_slice_op.cc @@ -45,10 +45,10 @@ class SequenceSliceOp : public framework::OperatorWithKernel { // Initialize the output's dims to maximum, // and re-set to real dims by the value of Offset and Length at kernel ctx->SetOutputDim("Out", input_dims); - } + } protected: - framework::OpKernelType GetKernelType( + framework::OpKernelType GetActualKernelType( const framework::ExecutionContext& ctx) const override { return framework::OpKernelType( framework::ToDataType(ctx.Input("X")->type()), @@ -69,7 +69,7 @@ class SequenceSliceGradOp : public framework::OperatorWithKernel { } protected: - framework::OpKernelType GetKernelType( + framework::OpKernelType GetActualKernelType( const framework::ExecutionContext& ctx) const override { return framework::OpKernelType( framework::ToDataType(ctx.Input("X")->type()), @@ -79,8 +79,7 @@ class SequenceSliceGradOp : public framework::OperatorWithKernel { class SequenceSliceOpMaker : public framework::OpProtoAndCheckerMaker { public: - SequenceSliceOpMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker) + SequenceSliceOpMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "(LoDTensor), " @@ -93,8 +92,7 @@ class SequenceSliceOpMaker : public framework::OpProtoAndCheckerMaker { "(Tensor), " "a vector to describe the length of every input sequence for " "sub sequence item."); - AddOutput("Out", - "(LoDTensor), the output of SequenceSliceOp."); + AddOutput("Out", "(LoDTensor), the output of SequenceSliceOp."); AddComment(R"DOC( Sequence slice operator @@ -126,7 +124,7 @@ REGISTER_OP(sequence_slice, ops::SequenceSliceOp, ops::SequenceSliceOpMaker, sequence_slice_grad, ops::SequenceSliceGradOp); REGISTER_OP_CPU_KERNEL( sequence_slice, - ops::SequenceSliceOpKernel); + ops::SequenceSliceOpKernel); REGISTER_OP_CPU_KERNEL( sequence_slice_grad, - ops::SequenceSliceGradOpKernel); + ops::SequenceSliceGradOpKernel); diff --git a/paddle/operators/sequence_slice_op.cu b/paddle/operators/sequence_slice_op.cu index a9f59dadba74d900fa5cc0601fb5b264ea19e34d..43a21d619f4116874c329eb968f09dc230975c05 100755 --- a/paddle/operators/sequence_slice_op.cu +++ b/paddle/operators/sequence_slice_op.cu @@ -15,9 +15,9 @@ limitations under the License. */ #include "paddle/operators/sequence_slice_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL( +REGISTER_OP_CUDA_KERNEL( sequence_slice, - ops::SequenceSliceOpKernel); -REGISTER_OP_GPU_KERNEL( + ops::SequenceSliceOpKernel); +REGISTER_OP_CUDA_KERNEL( sequence_slice_grad, - ops::SequenceSliceGradOpKernel); + ops::SequenceSliceGradOpKernel); diff --git a/paddle/operators/sequence_slice_op.h b/paddle/operators/sequence_slice_op.h old mode 100755 new mode 100644 index 2c9b8464a1236a054cf1a38b9dc1d73588f8dd38..14bcaebbb402cb47507f1bf60035bc2d37f9baf7 --- a/paddle/operators/sequence_slice_op.h +++ b/paddle/operators/sequence_slice_op.h @@ -26,7 +26,7 @@ using LoD = framework::LoD; template inline LoD SequenceSliceLoD(const T& in, const int64_t* offset_data, - const int64_t* length_data) { + const int64_t* length_data) { auto out_lod = in.lod(); size_t lod_offset = 0; @@ -34,12 +34,12 @@ inline LoD SequenceSliceLoD(const T& in, const int64_t* offset_data, out_lod[0][0] = 0; for (size_t i = 0; i < n; ++i) { lod_offset += length_data[i]; - out_lod[0][i+1] = lod_offset; + out_lod[0][i + 1] = lod_offset; } return out_lod; } -template +template class SequenceSliceOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { @@ -51,14 +51,13 @@ class SequenceSliceOpKernel : public framework::OpKernel { auto lod = in->lod(); auto n = lod[0].size() - 1; - PADDLE_ENFORCE_EQ(lod.size(), 1UL, - "Only support one level sequence now."); + PADDLE_ENFORCE_EQ(lod.size(), 1UL, "Only support one level sequence now."); PADDLE_ENFORCE_EQ( n, static_cast(length->dims()[0]), - "The size of input-sequence and length-array should be the same") + "The size of input-sequence and length-array should be the same"); PADDLE_ENFORCE_EQ( n, static_cast(offset->dims()[0]), - "The size of input-sequence and offset-array should be the same") + "The size of input-sequence and offset-array should be the same"); const int64_t* offset_data = offset->data(); const int64_t* length_data = length->data(); @@ -67,23 +66,23 @@ class SequenceSliceOpKernel : public framework::OpKernel { if (platform::is_gpu_place(ctx.GetPlace())) { offset_cpu.mutable_data(offset->dims(), platform::CPUPlace()); - offset_cpu.CopyFrom(*offset, platform::CPUPlace(), ctx.device_context()); + framework::CopyFrom(*offset, platform::CPUPlace(), ctx.device_context(), + &offset_cpu); offset_data = offset_cpu.data(); length_cpu.mutable_data(length->dims(), platform::CPUPlace()); - length_cpu.CopyFrom(*length, platform::CPUPlace(), ctx.device_context()); + framework::CopyFrom(*length, platform::CPUPlace(), ctx.device_context(), + &length_cpu); length_data = length_cpu.data(); } for (size_t i = 0; i < n; ++i) { PADDLE_ENFORCE_LT(0, offset_data[i], - "The offset[%d] must greater than zero.", i) + "The offset[%d] must greater than zero.", i); PADDLE_ENFORCE_LT(0, length_data[i], - "The length[%d] must greater than zero.", i) - PADDLE_ENFORCE_LT( - lod[0][i] + offset_data[i] + length_data[i], - lod[0][i + 1], - "The target tensor's length overflow.") + "The length[%d] must greater than zero.", i); + PADDLE_ENFORCE_LT(lod[0][i] + offset_data[i] + length_data[i], + lod[0][i + 1], "The target tensor's length overflow."); } out->mutable_data(ctx.GetPlace()); @@ -98,20 +97,18 @@ class SequenceSliceOpKernel : public framework::OpKernel { size_t out_offset = 0; for (size_t i = 0; i < n; ++i) { - Tensor in_t = - in->Slice(static_cast(lod[0][i] + offset_data[i]), - static_cast(lod[0][i] + offset_data[i] + - length_data[i])); - - StridedMemcpy(ctx.device_context(), in_t.data(), - in_stride, in_t.dims(), out_stride, - out->data() + out_offset); + Tensor in_t = in->Slice( + static_cast(lod[0][i] + offset_data[i]), + static_cast(lod[0][i] + offset_data[i] + length_data[i])); + + StridedMemcpy(ctx.device_context(), in_t.data(), in_stride, + in_t.dims(), out_stride, out->data() + out_offset); out_offset += length_data[i] * in_stride[0]; } } }; -template +template class SequenceSliceGradOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { @@ -130,11 +127,13 @@ class SequenceSliceGradOpKernel : public framework::OpKernel { if (platform::is_gpu_place(ctx.GetPlace())) { offset_cpu.mutable_data(offset->dims(), platform::CPUPlace()); - offset_cpu.CopyFrom(*offset, platform::CPUPlace(), ctx.device_context()); + framework::CopyFrom(*offset, platform::CPUPlace(), ctx.device_context(), + &offset_cpu); offset_data = offset_cpu.data(); length_cpu.mutable_data(length->dims(), platform::CPUPlace()); - length_cpu.CopyFrom(*length, platform::CPUPlace(), ctx.device_context()); + framework::CopyFrom(*length, platform::CPUPlace(), ctx.device_context(), + &length_cpu); length_data = length_cpu.data(); } @@ -144,8 +143,9 @@ class SequenceSliceGradOpKernel : public framework::OpKernel { if (x_grad) { x_grad->mutable_data(ctx.GetPlace()); x_grad->set_lod(in->lod()); - math::SetConstant set_zero; - set_zero(ctx.device_context(), x_grad, static_cast(0)); + math::SetConstant set_zero; + set_zero(ctx.template device_context(), x_grad, + static_cast(0)); auto out_grad_stride = framework::stride(out_grad->dims()); @@ -162,8 +162,8 @@ class SequenceSliceGradOpKernel : public framework::OpKernel { static_cast(lod[0][i] + offset_data[i] + length_data[i])); StridedMemcpy(ctx.device_context(), out_grad_t.data(), - out_grad_stride, out_grad_t.dims(), x_grad_stride, - x_grad_t.data()); + out_grad_stride, out_grad_t.dims(), x_grad_stride, + x_grad_t.data()); } } } diff --git a/paddle/operators/sequence_softmax_op.cc b/paddle/operators/sequence_softmax_op.cc index 32c15025660ebf0baf317e269a33c047e6844219..b74766f012e333cc2a317e6efe17c5b60238924a 100644 --- a/paddle/operators/sequence_softmax_op.cc +++ b/paddle/operators/sequence_softmax_op.cc @@ -33,8 +33,7 @@ class SequenceSoftmaxOp : public framework::OperatorWithKernel { class SequenceSoftmaxOpMaker : public framework::OpProtoAndCheckerMaker { public: - SequenceSoftmaxOpMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker) + SequenceSoftmaxOpMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "(LoDTensor) 1-D or 2-D input LoDTensor with the 2-nd dimension " @@ -51,10 +50,14 @@ input Tensor can be either [N, 1] or [N], where N is the sum of the length of all sequences. The algorithm works as follows: + for i-th sequence in a mini-batch: - $$Out(X[lod[i]:lod[i+1]], :) = - \frac{\exp(X[lod[i]:lod[i+1], :])} - {\sum(\exp(X[lod[i]:lod[i+1], :]))}$$ + +$$ +Out(X[lod[i]:lod[i+1]], :) = \ +\frac{\exp(X[lod[i]:lod[i+1], :])} \ +{\sum(\exp(X[lod[i]:lod[i+1], :]))} +$$ For example, for a mini-batch of 3 sequences with variable-length, each containing 2, 3, 2 time-steps, the lod of which is [0, 2, 5, 7], @@ -99,7 +102,7 @@ REGISTER_OP(sequence_softmax, ops::SequenceSoftmaxOp, ops::SequenceSoftmaxGradOp); REGISTER_OP_CPU_KERNEL( sequence_softmax, - ops::SequenceSoftmaxKernel); + ops::SequenceSoftmaxKernel); REGISTER_OP_CPU_KERNEL( sequence_softmax_grad, - ops::SequenceSoftmaxGradKernel); + ops::SequenceSoftmaxGradKernel); diff --git a/paddle/operators/sequence_softmax_op.cu.cc b/paddle/operators/sequence_softmax_op.cu.cc index 7023795a3b5777c250a9323a304a54849d763e9e..5f65b4daf97cf025b975d2d95212375b5fca01f8 100644 --- a/paddle/operators/sequence_softmax_op.cu.cc +++ b/paddle/operators/sequence_softmax_op.cu.cc @@ -15,9 +15,9 @@ limitations under the License. */ #include "paddle/operators/sequence_softmax_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL( +REGISTER_OP_CUDA_KERNEL( sequence_softmax, - ops::SequenceSoftmaxKernel) -REGISTER_OP_GPU_KERNEL( + ops::SequenceSoftmaxKernel) +REGISTER_OP_CUDA_KERNEL( sequence_softmax_grad, - ops::SequenceSoftmaxGradKernel); + ops::SequenceSoftmaxGradKernel); diff --git a/paddle/operators/sequence_softmax_op.h b/paddle/operators/sequence_softmax_op.h index 1b68dd0662ddfffc57b187945fe131e202c55174..e889e88cb34719b6648e3032754645fbb2807741 100644 --- a/paddle/operators/sequence_softmax_op.h +++ b/paddle/operators/sequence_softmax_op.h @@ -23,7 +23,7 @@ namespace operators { using Tensor = framework::Tensor; using LoDTensor = framework::LoDTensor; -template +template class SequenceSoftmaxKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { @@ -52,12 +52,13 @@ class SequenceSoftmaxKernel : public framework::OpKernel { framework::DDim dims_i = framework::make_ddim({1UL, end_pos - start_pos}); x_i.Resize(dims_i); out_i.Resize(dims_i); - math::SoftmaxFunctor()(ctx.device_context(), &x_i, &out_i); + math::SoftmaxFunctor()( + ctx.template device_context(), &x_i, &out_i); } } }; -template +template class SequenceSoftmaxGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { @@ -83,8 +84,9 @@ class SequenceSoftmaxGradKernel : public framework::OpKernel { out_i.Resize(dims_i); out_grad_i.Resize(dims_i); x_grad_i.Resize(dims_i); - math::SoftmaxGradFunctor()(ctx.device_context(), &out_i, - &out_grad_i, &x_grad_i); + math::SoftmaxGradFunctor()( + ctx.template device_context(), &out_i, &out_grad_i, + &x_grad_i); } } }; diff --git a/paddle/operators/sgd_op.cc b/paddle/operators/sgd_op.cc index 72f4e4d5cbcd692423fa2a3e9ec8e7033b552c3c..a11c9624ce5e8485449dd6b420ad1f23ff3550c7 100644 --- a/paddle/operators/sgd_op.cc +++ b/paddle/operators/sgd_op.cc @@ -43,7 +43,7 @@ class SGDOp : public framework::OperatorWithKernel { class SGDOpMaker : public framework::OpProtoAndCheckerMaker { public: - SGDOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker) + SGDOpMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("Param", "(Tensor) Input parameter"); AddInput("LearningRate", "(Tensor) Learning rate of SGD"); @@ -55,48 +55,15 @@ SGD operator This operator implements one step of the stochastic gradient descent algorithm. -$$param_out = param - learning_rate * grad$$ +$$param\_out = param - learning\_rate * grad$$ )DOC"); } }; -template -struct SparseSGDFunctor { - void operator()(const platform::DeviceContext& context, - const framework::SelectedRows& input, - const framework::Tensor& learning_rate, - framework::Tensor* output) { - auto in_height = input.height(); - auto out_dims = output->dims(); - PADDLE_ENFORCE_EQ(in_height, out_dims[0]); - - auto& in_value = input.value(); - auto& in_rows = input.rows(); - - int64_t in_row_numel = in_value.numel() / in_rows.size(); - PADDLE_ENFORCE_EQ(in_row_numel, output->numel() / in_height); - - auto* in_data = in_value.data(); - auto* out_data = output->data(); - auto* lr = learning_rate.data(); - - for (size_t i = 0; i < in_rows.size(); i++) { - for (int64_t j = 0; j < in_row_numel; j++) { - out_data[in_rows[i] * in_row_numel + j] -= - lr[0] * in_data[i * in_row_numel + j]; - } - } - } -}; - -template struct SparseSGDFunctor; -template struct SparseSGDFunctor; - } // namespace operators } // namespace paddle namespace ops = paddle::operators; REGISTER_OP_WITHOUT_GRADIENT(sgd, ops::SGDOp, ops::SGDOpMaker); -REGISTER_OP_CPU_KERNEL(sgd, ops::SGDOpKernel, - ops::SGDOpKernel); +REGISTER_OP_CPU_KERNEL(sgd, ops::SGDOpKernel, ops::SGDOpKernel); diff --git a/paddle/operators/sgd_op.cu b/paddle/operators/sgd_op.cu index 7b6c5ec30628b521b594ceaa3b7f1e0e03e497e4..42f8f8b2f072f9d204dfadcd732926b5c98dc617 100644 --- a/paddle/operators/sgd_op.cu +++ b/paddle/operators/sgd_op.cu @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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. */ #define EIGEN_USE_GPU #include "paddle/operators/sgd_op.h" @@ -20,6 +20,19 @@ namespace paddle { namespace operators { namespace { + +template +__global__ void SGDKernel(const T* g, const T* p, const T* learning_rate, + const int num, T* p_out) { + T lr = learning_rate[0]; + int grid_size = blockDim.x * gridDim.x; + for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < num; i += grid_size) { + T g_data = g[i]; + T p_data = p[i]; + p_out[i] = p_data - lr * g_data; + } +} + template __global__ void SparseSGDFunctorKernel(const T* selected_rows, const int64_t* rows, @@ -41,42 +54,65 @@ __global__ void SparseSGDFunctorKernel(const T* selected_rows, } // namespace template -struct SparseSGDFunctor { - void operator()(const platform::DeviceContext& context, - const framework::SelectedRows& input, - const framework::Tensor& learning_rate, - framework::Tensor* output) { - auto in_height = input.height(); - auto out_dims = output->dims(); - PADDLE_ENFORCE_EQ(in_height, out_dims[0]); - - auto& in_value = input.value(); - auto& in_rows = input.rows(); - - int64_t in_row_numel = in_value.numel() / in_rows.size(); - PADDLE_ENFORCE_EQ(in_row_numel, output->numel() / in_height); - - auto* in_data = in_value.data(); - auto* out_data = output->data(); - - const int block_size = 256; - dim3 threads(block_size, 1); - dim3 grid(1, in_rows.size()); - SparseSGDFunctorKernel< - T, 256><<(context) - .stream()>>>(in_data, in_rows.data(), - learning_rate.data(), out_data, - in_row_numel); +class SGDOpCUDAKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* param = ctx.Input("Param"); + auto* param_out = ctx.Output("ParamOut"); + auto* learning_rate = ctx.Input("LearningRate"); + + auto* grad_var = ctx.InputVar("Grad"); + // Actually, all tensors are LoDTensor except SelectedRows. + if (grad_var->IsType()) { + param_out->mutable_data(ctx.GetPlace()); + auto* grad = ctx.Input("Grad"); + auto* grad_data = grad->data(); + auto* param_data = param->data(); + auto* param_out_data = param_out->data(); + + int block = 512; + int grid = (param->numel() + block - 1) / block; + + SGDKernel<<>>( + grad_data, param_data, learning_rate->data(), param->numel(), + param_out_data); + + } else if (grad_var->IsType()) { + // TODO(qijun): In Sparse SGD operator, in-place update is enforced. + // This manual optimization brings difficulty to track data dependency. + // It's better to find a more elegant solution. + PADDLE_ENFORCE_EQ(param, param_out); + auto* grad = ctx.Input("Grad"); + + auto in_height = grad->height(); + auto out_dims = param_out->dims(); + PADDLE_ENFORCE_EQ(in_height, out_dims[0]); + + auto& in_value = grad->value(); + auto& in_rows = grad->rows(); + + int64_t in_row_numel = in_value.numel() / in_rows.size(); + PADDLE_ENFORCE_EQ(in_row_numel, param_out->numel() / in_height); + + auto* in_data = in_value.data(); + auto* out_data = param_out->data(); + + const int block_size = 256; + dim3 threads(block_size, 1); + dim3 grid(1, in_rows.size()); + SparseSGDFunctorKernel< + T, 256><<>>( + in_data, in_rows.data(), learning_rate->data(), out_data, + in_row_numel); + + } else { + PADDLE_THROW("Unsupported Variable Type of Grad"); + } } }; - -template struct SparseSGDFunctor; -template struct SparseSGDFunctor; - } // namespace operators } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(sgd, ops::SGDOpKernel, - ops::SGDOpKernel); +REGISTER_OP_CUDA_KERNEL(sgd, ops::SGDOpCUDAKernel, + ops::SGDOpCUDAKernel); diff --git a/paddle/operators/sgd_op.h b/paddle/operators/sgd_op.h index 78b595fc6c63d775b627f23cafa9458f1dadd4e5..a6c544591e1172320f6cf7192bf640ff25225b99 100644 --- a/paddle/operators/sgd_op.h +++ b/paddle/operators/sgd_op.h @@ -20,15 +20,7 @@ limitations under the License. */ namespace paddle { namespace operators { -template -struct SparseSGDFunctor { - void operator()(const platform::DeviceContext& context, - const framework::SelectedRows& input, - const framework::Tensor& learning_rate, - framework::Tensor* output); -}; - -template +template class SGDOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { @@ -45,19 +37,36 @@ class SGDOpKernel : public framework::OpKernel { auto p = framework::EigenVector::Flatten(*param); auto g = framework::EigenVector::Flatten(*grad); auto o = framework::EigenVector::Flatten(*param_out); - auto lr = framework::EigenVector::Flatten(*learning_rate); - auto place = ctx.GetEigenDevice(); + auto* lr = learning_rate->data(); - Eigen::DSizes grad_dsize(grad->numel()); - o.device(place) = p - lr.broadcast(grad_dsize) * g; + o = p - lr[0] * g; } else if (grad_var->IsType()) { // TODO(qijun): In Sparse SGD operator, in-place update is enforced. // This manual optimization brings difficulty to track data dependency. // It's better to find a more elegant solution. PADDLE_ENFORCE_EQ(param, param_out); auto* grad = ctx.Input("Grad"); - SparseSGDFunctor functor; - functor(ctx.device_context(), *grad, *learning_rate, param_out); + + auto in_height = grad->height(); + auto out_dims = param_out->dims(); + PADDLE_ENFORCE_EQ(in_height, out_dims[0]); + + auto& in_value = grad->value(); + auto& in_rows = grad->rows(); + + int64_t in_row_numel = in_value.numel() / in_rows.size(); + PADDLE_ENFORCE_EQ(in_row_numel, param_out->numel() / in_height); + + auto* in_data = in_value.data(); + auto* out_data = param_out->data(); + auto* lr = learning_rate->data(); + + for (size_t i = 0; i < in_rows.size(); i++) { + for (int64_t j = 0; j < in_row_numel; j++) { + out_data[in_rows[i] * in_row_numel + j] -= + lr[0] * in_data[i * in_row_numel + j]; + } + } } else { PADDLE_THROW("Unsupported Variable Type of Grad"); } diff --git a/paddle/operators/shrink_rnn_memory_op.cc b/paddle/operators/shrink_rnn_memory_op.cc index 65bccc0c81d0ad9674649933a20ec7b09fec5b37..b37269b471b4d71b42c41641fd14c7a64d2719d6 100644 --- a/paddle/operators/shrink_rnn_memory_op.cc +++ b/paddle/operators/shrink_rnn_memory_op.cc @@ -1,16 +1,16 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +/* 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 +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 + 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. */ +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/framework/lod_rank_table.h" #include "paddle/operators/array_operator.h" #include "paddle/operators/math/math_function.h" @@ -27,11 +27,11 @@ class ShrinkRNNMemoryOp : public ArrayOp { : ArrayOp(type, inputs, outputs, attrs) {} void Run(const framework::Scope &scope, - const platform::DeviceContext &dev_ctx) const override { + const platform::Place &place) const override { auto *x_var = scope.FindVar(Input("X")); PADDLE_ENFORCE(x_var != nullptr, "Input X must be set"); auto &x_tensor = x_var->Get(); - size_t offset = this->GetOffset(scope, dev_ctx); + size_t offset = this->GetOffset(scope, place); auto *rank_table_var = scope.FindVar(Input("RankTable")); PADDLE_ENFORCE(rank_table_var != nullptr, "RankTable must be set"); auto &rank_table = rank_table_var->Get(); @@ -54,14 +54,23 @@ class ShrinkRNNMemoryOp : public ArrayOp { class ShrinkRNNMemoryOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - ShrinkRNNMemoryOpProtoMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + ShrinkRNNMemoryOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("X", ""); - AddInput("RankTable", ""); - AddInput("I", ""); - AddOutput("Out", ""); - AddComment(""); + AddInput("X", "(LoDTensor) The RNN step memory to be shrinked."); + AddInput("RankTable", "(LoDRankTable) The lod_rank_table of dynamic RNN."); + AddInput("I", + "(LoDTensor) The step index. The RNN step memory 'X' will be " + "shrinked to match the size of the input of the index'th step."); + AddOutput("Out", "(LoDTensor) The shrinked RNN step memory."); + AddComment( + R"DOC( + In dynamic RNN, we are able to handle sequences of different lengths. + Because of the multiple lengths, the size of each step input can be + different, which may lead to a mismatching between the input of + the current step and the memory generated by the previous one. This + operator shrinks memory according to the size of the next step input, + to make sure that they can match each other. + )DOC"); } }; @@ -84,7 +93,7 @@ class ShrinkRNNMemoryGradOp : public ArrayOp { : ArrayOp(type, inputs, outputs, attrs) {} void Run(const framework::Scope &scope, - const platform::DeviceContext &dev_ctx) const override { + const platform::Place &place) const override { auto *dout_var = scope.FindVar(Input(framework::GradVarName("Out"))); auto *dx_var = scope.FindVar(Output(framework::GradVarName("X"))); PADDLE_ENFORCE(dx_var != nullptr, "Input Gradient should not be nullptr"); @@ -96,16 +105,20 @@ class ShrinkRNNMemoryGradOp : public ArrayOp { dx_tensor.Resize(x_tensor.dims()); dx_tensor.mutable_data(x_tensor.place(), x_tensor.type()); + // get device context from pool + platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance(); + auto &dev_ctx = *pool.Get(place); + if (dout_var == nullptr) { // dx_tensor fill zero math::set_constant(dev_ctx, &dx_tensor, 0.0f); } else { auto &dout_tensor = dout_var->Get(); auto height = dout_tensor.dims()[0]; - dx_tensor.Slice(0, static_cast(height)) - .CopyFrom(dout_tensor, dout_tensor.place(), dev_ctx); - if (dx_tensor.dims()[0] < height) { + auto slice = dx_tensor.Slice(0, static_cast(height)); + framework::CopyFrom(dout_tensor, dout_tensor.place(), dev_ctx, &slice); + if (dx_tensor.dims()[0] > height) { auto rest_tensor = dx_tensor.Slice( - static_cast(height), static_cast(dout_tensor.dims()[0])); + static_cast(height), static_cast(dx_tensor.dims()[0])); math::set_constant(dev_ctx, &rest_tensor, 0.0f); } } @@ -127,14 +140,14 @@ class ShrinkRNNGradOpMaker : public framework::SingleGradOpDescMaker { using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; protected: - std::unique_ptr Apply() const override { - auto *op = new framework::OpDescBind(); + std::unique_ptr Apply() const override { + auto *op = new framework::OpDesc(); op->SetType("shrink_rnn_memory_grad"); op->SetInput("X", Input("X")); op->SetInput(framework::GradVarName("Out"), OutputGrad("Out")); op->SetOutput(framework::GradVarName("X"), InputGrad("X")); op->SetAttrMap(Attrs()); - return std::unique_ptr(op); + return std::unique_ptr(op); } }; diff --git a/paddle/operators/sigmoid_cross_entropy_with_logits_op.cc b/paddle/operators/sigmoid_cross_entropy_with_logits_op.cc index d9e40546523c60b0a7eec2e0593446258996ba58..c526a88a127da12a6384777bca31b60873844d94 100644 --- a/paddle/operators/sigmoid_cross_entropy_with_logits_op.cc +++ b/paddle/operators/sigmoid_cross_entropy_with_logits_op.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/operators/sigmoid_cross_entropy_with_logits_op.h" @@ -25,20 +25,19 @@ class SigmoidCrossEntropyWithLogitsOp : public framework::OperatorWithKernel { void InferShape(framework::InferShapeContext* ctx) const override { PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should be not null."); - PADDLE_ENFORCE(ctx->HasInput("Labels"), - "Input(Labels) should be not null."); + PADDLE_ENFORCE(ctx->HasInput("Label"), "Input(Label) should be not null."); PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) should be not null."); auto x_dims = ctx->GetInputDim("X"); - auto labels_dims = ctx->GetInputDim("Labels"); + auto labels_dims = ctx->GetInputDim("Label"); PADDLE_ENFORCE_EQ(x_dims.size(), 2, "Input(X)'s rank should be 2."); PADDLE_ENFORCE_EQ(labels_dims.size(), 2, - "Input(Labels)'s rank should be 2."); + "Input(Label)'s rank should be 2."); PADDLE_ENFORCE_EQ(x_dims[0], labels_dims[0], - "The 1st dimension of Input(X) and Input(Labels) should " + "The 1st dimension of Input(X) and Input(Label) should " "be equal."); PADDLE_ENFORCE_EQ(x_dims[1], labels_dims[1], - "The 2nd dimension of Input(X) and Input(Labels) should " + "The 2nd dimension of Input(X) and Input(Label) should " "be equal."); ctx->SetOutputDim("Out", x_dims); @@ -53,26 +52,25 @@ class SigmoidCrossEntropyWithLogitsGradOp void InferShape(framework::InferShapeContext* ctx) const override { PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should be not null."); - PADDLE_ENFORCE(ctx->HasInput("Labels"), - "Input(Labels) should be not null."); + PADDLE_ENFORCE(ctx->HasInput("Label"), "Input(Label) should be not null."); PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), "Input(Out@GRAD) shoudl be not null."); PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("X")), "Output(X@GRAD) should be not null."); auto x_dims = ctx->GetInputDim("X"); - auto labels_dims = ctx->GetInputDim("Labels"); + auto labels_dims = ctx->GetInputDim("Label"); auto dout_dims = ctx->GetInputDim(framework::GradVarName("Out")); PADDLE_ENFORCE_EQ(x_dims.size(), 2, "Input(X)'s rank should be 2."); PADDLE_ENFORCE_EQ(labels_dims.size(), 2, - "Input(Labels)'s rank should be 2."); + "Input(Label)'s rank should be 2."); PADDLE_ENFORCE_EQ(dout_dims.size(), 2, "Input(Out@Grad)'s rank should be 2."); PADDLE_ENFORCE_EQ(x_dims[0], labels_dims[0], - "The 1st dimension of Input(X) and Input(Labels) should " + "The 1st dimension of Input(X) and Input(Label) should " "be equal."); PADDLE_ENFORCE_EQ(x_dims[1], labels_dims[1], - "The 2nd dimension of Input(X) and Input(Labels) should " + "The 2nd dimension of Input(X) and Input(Label) should " "be equal."); PADDLE_ENFORCE_EQ(x_dims[0], dout_dims[0], "The 1st dimension of Input(X) and Input(Out@Grad) " @@ -88,8 +86,8 @@ class SigmoidCrossEntropyWithLogitsGradOp class SigmoidCrossEntropyWithLogitsOpMaker : public framework::OpProtoAndCheckerMaker { public: - SigmoidCrossEntropyWithLogitsOpMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker) + SigmoidCrossEntropyWithLogitsOpMaker(OpProto* proto, + OpAttrChecker* op_checker) : framework::OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "(Tensor, default Tensor), a 2-D tensor with shape N x D, " @@ -97,7 +95,7 @@ class SigmoidCrossEntropyWithLogitsOpMaker "This input is a tensor of logits computed by the previous " " operator. Logits are unscaled log probabilities given as " "log(p/(1-p))."); - AddInput("Labels", + AddInput("Label", "(Tensor, default Tensor), a 2-D tensor of the same type " "and shape as X. This input is a tensor of probabalistic labels " "for each logit"); @@ -144,7 +142,7 @@ REGISTER_OP(sigmoid_cross_entropy_with_logits, ops::SigmoidCrossEntropyWithLogitsGradOp); REGISTER_OP_CPU_KERNEL(sigmoid_cross_entropy_with_logits, ops::SigmoidCrossEntropyWithLogitsKernel< - paddle::platform::CPUPlace, float>); + paddle::platform::CPUDeviceContext, float>); REGISTER_OP_CPU_KERNEL(sigmoid_cross_entropy_with_logits_grad, ops::SigmoidCrossEntropyWithLogitsGradKernel< - paddle::platform::CPUPlace, float>); + paddle::platform::CPUDeviceContext, float>); diff --git a/paddle/operators/sigmoid_cross_entropy_with_logits_op.cu b/paddle/operators/sigmoid_cross_entropy_with_logits_op.cu index 32a39956a14a206373b7b4c141dad19577d171f0..3f393265f48b428dca8703ff77688de979fb63df 100644 --- a/paddle/operators/sigmoid_cross_entropy_with_logits_op.cu +++ b/paddle/operators/sigmoid_cross_entropy_with_logits_op.cu @@ -1,24 +1,24 @@ /* 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 +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 + 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. */ +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. */ #define EIGEN_USE_GPU #include "paddle/operators/sigmoid_cross_entropy_with_logits_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(sigmoid_cross_entropy_with_logits, - ops::SigmoidCrossEntropyWithLogitsKernel< - paddle::platform::GPUPlace, float>); -REGISTER_OP_GPU_KERNEL(sigmoid_cross_entropy_with_logits_grad, - ops::SigmoidCrossEntropyWithLogitsGradKernel< - paddle::platform::GPUPlace, float>); +REGISTER_OP_CUDA_KERNEL(sigmoid_cross_entropy_with_logits, + ops::SigmoidCrossEntropyWithLogitsKernel< + paddle::platform::CUDADeviceContext, float>); +REGISTER_OP_CUDA_KERNEL(sigmoid_cross_entropy_with_logits_grad, + ops::SigmoidCrossEntropyWithLogitsGradKernel< + paddle::platform::CUDADeviceContext, float>); diff --git a/paddle/operators/sigmoid_cross_entropy_with_logits_op.h b/paddle/operators/sigmoid_cross_entropy_with_logits_op.h index 41c619f181c878f08959a8ca461c60af5ffdff2a..b78bcc436e9fa5c5d4db3fbb22224e328c3bc3c2 100644 --- a/paddle/operators/sigmoid_cross_entropy_with_logits_op.h +++ b/paddle/operators/sigmoid_cross_entropy_with_logits_op.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/framework/eigen.h" @@ -20,20 +20,19 @@ namespace paddle { namespace operators { // Out = max(X, 0) - X * Labels + log(1 + exp(-abs(X))) -template +template class SigmoidCrossEntropyWithLogitsKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext &context) const override { const framework::Tensor *X = context.Input("X"); - const framework::Tensor *Labels = - context.Input("Labels"); + const framework::Tensor *Labels = context.Input("Label"); framework::Tensor *Out = context.Output("Out"); Out->mutable_data(context.GetPlace()); auto x = framework::EigenVector::Flatten(*X); auto labels = framework::EigenVector::Flatten(*Labels); auto out = framework::EigenVector::Flatten(*Out); - auto place = context.GetEigenDevice(); + auto &place = *context.device_context().eigen_device(); // term1 = max(x, 0) auto term1 = x.cwiseMax(static_cast(0)); @@ -47,13 +46,12 @@ class SigmoidCrossEntropyWithLogitsKernel : public framework::OpKernel { }; // dX = sigmoid(X) - labels -template +template class SigmoidCrossEntropyWithLogitsGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext &context) const override { const framework::Tensor *X = context.Input("X"); - const framework::Tensor *Labels = - context.Input("Labels"); + const framework::Tensor *Labels = context.Input("Label"); const framework::Tensor *dOut = context.Input(framework::GradVarName("Out")); framework::Tensor *dX = @@ -64,7 +62,8 @@ class SigmoidCrossEntropyWithLogitsGradKernel : public framework::OpKernel { auto labels = framework::EigenVector::Flatten(*Labels); auto dout = framework::EigenVector::Flatten(*dOut); auto dx = framework::EigenVector::Flatten(*dX); - auto place = context.GetEigenDevice(); + auto &place = + *context.template device_context().eigen_device(); auto sigmoid_x = static_cast(1) / (static_cast(1) + (-x).exp()); dx.device(place) = dout * (sigmoid_x - labels); diff --git a/paddle/operators/sign_op.cc b/paddle/operators/sign_op.cc index 08bf2e4e7cc101a3bcc907d3b40ee82347b39f80..f63eaa4464cc668acdb8e5b8a74ad5bba936db44 100644 --- a/paddle/operators/sign_op.cc +++ b/paddle/operators/sign_op.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/operators/sign_op.h" @@ -34,7 +34,7 @@ class SignOp : public framework::OperatorWithKernel { template class SignOpMaker : public framework::OpProtoAndCheckerMaker { public: - SignOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + SignOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "(Tensor) Input tensor of sign operator."); AddOutput("Out", "(Tensor) Output tensor of sign operator."); @@ -50,13 +50,13 @@ class SignGradMaker : public framework::SingleGradOpDescMaker { public: using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; - std::unique_ptr Apply() const override { - auto *grad_op = new framework::OpDescBind(); + std::unique_ptr Apply() const override { + auto *grad_op = new framework::OpDesc(); grad_op->SetType("scale"); grad_op->SetInput("X", OutputGrad("Out")); grad_op->SetOutput("Out", InputGrad("X")); grad_op->SetAttr("scale", 0.0f); - return std::unique_ptr(grad_op); + return std::unique_ptr(grad_op); } }; @@ -67,5 +67,5 @@ namespace ops = paddle::operators; REGISTER_OPERATOR(sign, ops::SignOp, ops::SignOpMaker, ops::SignGradMaker); -REGISTER_OP_CPU_KERNEL(sign, - ops::SignKernel); +REGISTER_OP_CPU_KERNEL( + sign, ops::SignKernel); diff --git a/paddle/operators/sign_op.cu b/paddle/operators/sign_op.cu index 4d0638cb97d84bf650fb23e4d2a201adc51a4b68..f224880cffb2154a7c46b8a4701d7357e67bb70c 100644 --- a/paddle/operators/sign_op.cu +++ b/paddle/operators/sign_op.cu @@ -1,18 +1,19 @@ /* 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 +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 + 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. */ +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/operators/sign_op.h" -REGISTER_OP_GPU_KERNEL( - sign, paddle::operators::SignKernel); +REGISTER_OP_CUDA_KERNEL( + sign, + paddle::operators::SignKernel); diff --git a/paddle/operators/sign_op.h b/paddle/operators/sign_op.h index ab5cd4bac019d602c63ea51629fb85fa7e206841..9fe49ae1a2161d9f1472eef830c11b0f8305c568 100644 --- a/paddle/operators/sign_op.h +++ b/paddle/operators/sign_op.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 @@ -19,7 +19,7 @@ namespace paddle { namespace operators { -template +template class SignKernel : public framework::OpKernel { public: virtual void Compute(const framework::ExecutionContext& context) const { @@ -29,7 +29,8 @@ class SignKernel : public framework::OpKernel { auto eigen_out = framework::EigenVector::Flatten(*out); auto eigen_in = framework::EigenVector::Flatten(*in); - auto& place = context.GetEigenDevice(); + auto& place = + *context.template device_context().eigen_device(); eigen_out.device(place) = eigen_in.sign(); } }; diff --git a/paddle/operators/smooth_l1_loss_op.cc b/paddle/operators/smooth_l1_loss_op.cc index ebf7b43700a7498aa18b5f648b0b8c2c4e7b442b..dcb18d729da69beaa556e4b93129dafb08b72c06 100644 --- a/paddle/operators/smooth_l1_loss_op.cc +++ b/paddle/operators/smooth_l1_loss_op.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/operators/smooth_l1_loss_op.h" @@ -22,22 +22,20 @@ class SmoothL1LossOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; void InferShape(framework::InferShapeContext* ctx) const override { - PADDLE_ENFORCE(ctx->HasInput("X"), "X must be initialized."); - PADDLE_ENFORCE(ctx->HasInput("Y"), "Y must be initialized."); + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Y"), "Input(Y) should not be null."); auto x_dims = ctx->GetInputDim("X"); auto y_dims = ctx->GetInputDim("Y"); - PADDLE_ENFORCE_EQ(x_dims, y_dims, "The shape of X and Y must be the same."); + PADDLE_ENFORCE_EQ(x_dims, y_dims); PADDLE_ENFORCE_GE(x_dims.size(), 2, - "The tensor rank of X must be at least 2."); + "The tensor rank of Input(X) should not be less than 2."); if (ctx->HasInput("InsideWeight")) { PADDLE_ENFORCE(ctx->HasInput("OutsideWeight"), "If weights are provided, must specify both " "inside and outside weights."); - PADDLE_ENFORCE_EQ(ctx->GetInputDim("InsideWeight"), x_dims, - "The shape of InsideWeight must be same as X."); - PADDLE_ENFORCE_EQ(ctx->GetInputDim("OutsideWeight"), x_dims, - "The shape of OutsideWeight must be same as X."); + PADDLE_ENFORCE_EQ(ctx->GetInputDim("InsideWeight"), x_dims); + PADDLE_ENFORCE_EQ(ctx->GetInputDim("OutsideWeight"), x_dims); } ctx->SetOutputDim("Diff", x_dims); @@ -49,29 +47,32 @@ class SmoothL1LossOp : public framework::OperatorWithKernel { template class SmoothL1LossOpMaker : public framework::OpProtoAndCheckerMaker { public: - SmoothL1LossOpMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker) + SmoothL1LossOpMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", - "The input tensor of smooth l1 loss op." - "The rank should be greater or equal to 2 with shape " - "[batch_size, value_dim1, value_dim2, ..., value_dimN]"); + "(Tensor, default Tensor) A tensor with rank at least 2. " + "The input value of smooth l1 loss op with shape " + "[batch_size, dim1, ..., dimN]."); AddInput("Y", - "The target tensor of smooth l1 loss op " - "with the same shape as X."); + "(Tensor, default Tensor) A tensor with rank at least 2. " + "The target value of smooth l1 loss op with same shape as X."); AddInput("InsideWeight", - "Optional input tensor of smooth l1 loss op with the same shape " - "as X. If provided, the result of (X - Y) will be multiplied " + "(Tensor, default Tensor) 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.") .AsDispensable(); AddInput("OutsideWeight", - "Optinal input of smooth l1 loss op with the same shape as X." - "If provided, the output smooth l1 loss will be multiplied by " - "this tensor element by element.") + "(Tensor, default Tensor) 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 by element.") .AsDispensable(); - AddOutput("Diff", "Intermediate variable to cache InsideWeight*(X-Y).") + AddOutput("Diff", "Intermediate variable to cache InsideWeight * (X - Y).") .AsIntermediate(); - AddOutput("Out", "Smooth l1 loss."); + AddOutput("Out", + "(Tensor, default Tensor) A tensor with rank be 2. " + "The output smooth l1 loss with shape [batch_size, 1]."); AddAttr("sigma", "Hyper parameter of smooth l1 loss op." "A float scalar with default value 3.0.") @@ -79,15 +80,23 @@ class SmoothL1LossOpMaker : public framework::OpProtoAndCheckerMaker { AddComment(R"DOC( Smooth L1 Loss Operator. -This operator computes the smooth l1 loss for input and target. -The operator takes the first dimension of input as the batch size. +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 -and then sums all the losses. So the resulting output shape -is [batch_size, 1]. +and then sums all the losses. So the shape of Out is [batch_size, 1]. The equation is: -loss = $$0.5 * (\sigma * (x-y))^2$$ if $$|x - y| < 1 /({\sigma}^2)$$ - $$\frac{|x - y| - 0.5}{{\sigma}^2}$$ otherwise +$$ +Out_{\sigma}(X, Y)_i = \begin{cases} +0.5 * (\sigma * (X_i - Y_i)) ^ 2 +\quad |X_i - Y_i| \lt \frac{1} {{\sigma} ^ 2} \\ +\frac{|X_i - Y_i| - 0.5}{{\sigma}^2}, +\quad otherwise +\end{cases} +$$ + +In the above equation, $Out_{\sigma}(X, Y)_i$, $X_i$ and $Y_i$ represent the ith +element of Out, X and Y. )DOC"); } @@ -128,7 +137,8 @@ REGISTER_OP(smooth_l1_loss, ops::SmoothL1LossOp, ops::SmoothL1LossOpMaker, smooth_l1_loss_grad, ops::SmoothL1LossGradOp); REGISTER_OP_CPU_KERNEL( - smooth_l1_loss, ops::SmoothL1LossKernel); + smooth_l1_loss, + ops::SmoothL1LossKernel); REGISTER_OP_CPU_KERNEL( smooth_l1_loss_grad, - ops::SmoothL1LossGradKernel); + ops::SmoothL1LossGradKernel); diff --git a/paddle/operators/smooth_l1_loss_op.cu b/paddle/operators/smooth_l1_loss_op.cu index 1c3172f43867741cd1f26979a366b2425f326321..213429bc370ef0d5b493b3a448df1b3bf0e4e87c 100644 --- a/paddle/operators/smooth_l1_loss_op.cu +++ b/paddle/operators/smooth_l1_loss_op.cu @@ -1,24 +1,25 @@ /* 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 +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 + 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. */ +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. */ #define EIGEN_USE_GPU #include "paddle/operators/smooth_l1_loss_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL( - smooth_l1_loss, ops::SmoothL1LossKernel); -REGISTER_OP_GPU_KERNEL( +REGISTER_OP_CUDA_KERNEL( + smooth_l1_loss, + ops::SmoothL1LossKernel); +REGISTER_OP_CUDA_KERNEL( smooth_l1_loss_grad, - ops::SmoothL1LossGradKernel); + ops::SmoothL1LossGradKernel); diff --git a/paddle/operators/smooth_l1_loss_op.h b/paddle/operators/smooth_l1_loss_op.h index 39d0070b6c8909b8f433de48038240e851d9d6cf..3facfae116d711f86ea5c193562c20ea60a2efc9 100644 --- a/paddle/operators/smooth_l1_loss_op.h +++ b/paddle/operators/smooth_l1_loss_op.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/framework/eigen.h" @@ -44,7 +44,7 @@ struct SmoothL1LossForward { T sigma2; }; -template +template class SmoothL1LossKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { @@ -57,7 +57,8 @@ class SmoothL1LossKernel : public framework::OpKernel { out0->mutable_data(context.GetPlace()); out1->mutable_data(context.GetPlace()); - auto place = context.GetEigenDevice(); + auto* place = + context.template device_context().eigen_device(); auto sigma = static_cast(context.Attr("sigma")); T sigma2 = sigma * sigma; @@ -67,12 +68,12 @@ class SmoothL1LossKernel : public framework::OpKernel { auto y = EigenVector::Flatten(*in1); auto diff = EigenVector::Flatten(*out0); - diff.device(place) = x - y; + diff.device(*place) = x - y; // multiply inside weight if (has_weight) { auto inside_weight = EigenVector::Flatten(*in2); // cache diff, reused in bp - diff.device(place) = diff * inside_weight; + diff.device(*place) = diff * inside_weight; } auto in_counts = in0->numel(); @@ -81,12 +82,12 @@ class SmoothL1LossKernel : public framework::OpKernel { context.GetPlace()); auto errors = EigenVector::Flatten(ptensor_errors); // apply smooth l1 forward - errors.device(place) = diff.unaryExpr(SmoothL1LossForward(sigma2)); + errors.device(*place) = diff.unaryExpr(SmoothL1LossForward(sigma2)); // multiply outside weight if (has_weight) { auto outside_weight = EigenVector::Flatten(*in3); - errors.device(place) = errors * outside_weight; + errors.device(*place) = errors * outside_weight; } auto loss = EigenVector::Flatten(*out1); // first dimension of 'X' is the number of samples @@ -94,7 +95,7 @@ class SmoothL1LossKernel : public framework::OpKernel { framework::make_ddim({static_cast(in0->dims()[0]), static_cast(in_counts / in0->dims()[0])}); auto errors_mat_view = EigenMatrix::From(ptensor_errors, mat_dims); - loss.device(place) = errors_mat_view.sum(Eigen::array({{1}})); + loss.device(*place) = errors_mat_view.sum(Eigen::array({{1}})); } }; @@ -114,7 +115,7 @@ struct SmoothL1LossBackward { T sigma2; }; -template +template class SmoothL1LossGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { @@ -126,7 +127,8 @@ class SmoothL1LossGradKernel : public framework::OpKernel { T sigma2 = sigma * sigma; bool has_weight = (in0 != nullptr) && (in1 != nullptr); - auto place = context.GetEigenDevice(); + auto* place = + context.template device_context().eigen_device(); auto in_dims = in2->dims(); auto counts = in2->numel(); @@ -139,7 +141,7 @@ class SmoothL1LossGradKernel : public framework::OpKernel { context.GetPlace()); auto diff = EigenVector::Flatten(ptensor_diff); // apply smooth l1 backwoard - diff.device(place) = EigenVector::Flatten(*in2).unaryExpr( + diff.device(*place) = EigenVector::Flatten(*in2).unaryExpr( SmoothL1LossBackward(sigma2)); // compute weights @@ -147,11 +149,11 @@ class SmoothL1LossGradKernel : public framework::OpKernel { ptensor_weights.mutable_data(mat_dims, context.GetPlace()); auto weights = EigenMatrix::From(ptensor_weights); // initialize to 1.0 - weights.device(place) = weights.constant(static_cast(1.0)); + weights.device(*place) = weights.constant(static_cast(1.0)); if (has_weight) { auto inside_weight = EigenMatrix::From(*in0, mat_dims); auto outside_weight = EigenMatrix::From(*in1, mat_dims); - weights.device(place) = inside_weight * outside_weight; + weights.device(*place) = inside_weight * outside_weight; } // compute gradients @@ -167,13 +169,13 @@ class SmoothL1LossGradKernel : public framework::OpKernel { if (out0) { out0->mutable_data(context.GetPlace()); auto x_grad = EigenMatrix::From(*out0, mat_dims); - x_grad.device(place) = gradients; + x_grad.device(*place) = gradients; } if (out1) { out1->mutable_data(context.GetPlace()); auto y_grad = EigenMatrix::From(*out1, mat_dims); - y_grad.device(place) = -1 * gradients; + y_grad.device(*place) = -1 * gradients; } } }; diff --git a/paddle/operators/softmax_op.cc b/paddle/operators/softmax_op.cc index 93f89e33a73c5f4c6c0e5a8793a0abe7c692b656..e7306bc5f13377813e0bd49846bc834d501602eb 100644 --- a/paddle/operators/softmax_op.cc +++ b/paddle/operators/softmax_op.cc @@ -24,25 +24,24 @@ class SoftmaxOp : public framework::OperatorWithKernel { void InferShape(framework::InferShapeContext* ctx) const override { PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) of SoftmaxOp should not be null."); - PADDLE_ENFORCE(ctx->HasOutput("Y"), - "Output(Y) of SoftmaxOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of SoftmaxOp should not be null."); auto x_dims = ctx->GetInputDim("X"); PADDLE_ENFORCE(x_dims.size() == 2UL, "The input of softmax op must be a matrix."); - ctx->SetOutputDim("Y", x_dims); + ctx->SetOutputDim("Out", x_dims); } }; class SoftmaxOpMaker : public framework::OpProtoAndCheckerMaker { public: - SoftmaxOpMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker) + SoftmaxOpMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "The input tensor of softmax. " "2-D with shape [batch_size, input_feature_dimensions]."); - AddOutput("Y", "The normalized values with the same shape as X."); + AddOutput("Out", "The normalized values with the same shape as X."); AddComment(R"DOC( Softmax Operator. @@ -59,8 +58,8 @@ Then the ratio of the exponential of the given dimension and the sum of exponential values of all the other dimensions is the output of the softmax operator. -For each row `i` and each column `j` in input X, we have: - $$Y[i, j] = \frac{\exp(X[i, j])}{\sum_j(exp(X[i, j])}$$ +For each row $i$ and each column $j$ in Input(X), we have: + $$Out[i, j] = \frac{\exp(X[i, j])}{\sum_j(exp(X[i, j])}$$ )DOC"); } @@ -71,12 +70,12 @@ class SoftmaxOpGrad : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; void InferShape(framework::InferShapeContext* ctx) const override { - PADDLE_ENFORCE(ctx->HasInput("Y"), "Input(Y) should be not null."); - PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Y")), - "Input(Y@GRAD) should be not null."); - PADDLE_ENFORCE_EQ(ctx->GetInputDim("Y"), - ctx->GetInputDim(framework::GradVarName("Y")), - "Input(Y) and its gradients should have a same shape."); + PADDLE_ENFORCE(ctx->HasInput("Out"), "Input(Out) should be not null."); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Input(Out@GRAD) should be not null."); + PADDLE_ENFORCE_EQ(ctx->GetInputDim("Out"), + ctx->GetInputDim(framework::GradVarName("Out")), + "Input(Out) and its gradients should have a same shape."); ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X")); } @@ -89,7 +88,8 @@ namespace ops = paddle::operators; REGISTER_OP(softmax, ops::SoftmaxOp, ops::SoftmaxOpMaker, softmax_grad, ops::SoftmaxOpGrad); -REGISTER_OP_CPU_KERNEL(softmax, - ops::SoftmaxKernel); REGISTER_OP_CPU_KERNEL( - softmax_grad, ops::SoftmaxGradKernel); + softmax, ops::SoftmaxKernel); +REGISTER_OP_CPU_KERNEL( + softmax_grad, + ops::SoftmaxGradKernel); diff --git a/paddle/operators/softmax_op.cu.cc b/paddle/operators/softmax_op.cu.cc index 013ace19ae3d4a1af29b570ba33fea3e4595fe5b..e7da40f3e82d5db858a795e9634abf57b884d6a2 100644 --- a/paddle/operators/softmax_op.cu.cc +++ b/paddle/operators/softmax_op.cu.cc @@ -1,22 +1,23 @@ -/* Copyright (c) 2016 PaddlePaddle Authors All Rights Reserve. +/* 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 +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 + 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. */ +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/operators/softmax_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(softmax, - ops::SoftmaxKernel); -REGISTER_OP_GPU_KERNEL( - softmax_grad, ops::SoftmaxGradKernel); +REGISTER_OP_CUDA_KERNEL( + softmax, ops::SoftmaxKernel); +REGISTER_OP_CUDA_KERNEL( + softmax_grad, + ops::SoftmaxGradKernel); diff --git a/paddle/operators/softmax_op.h b/paddle/operators/softmax_op.h index 44d1e63f1bb4798144218cd1caf01f133825bcff..63e379a3b31a6c75aab0c56b4ce1b988fa7f0318 100644 --- a/paddle/operators/softmax_op.h +++ b/paddle/operators/softmax_op.h @@ -21,32 +21,34 @@ namespace operators { using Tensor = framework::Tensor; -template +template class SoftmaxKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto* X = context.Input("X"); - auto* Y = context.Output("Y"); + auto* Out = context.Output("Out"); // allocate memory on device. - Y->mutable_data(context.GetPlace()); + Out->mutable_data(context.GetPlace()); - math::SoftmaxFunctor()(context.device_context(), X, Y); + math::SoftmaxFunctor()( + context.template device_context(), X, Out); } }; -template +template class SoftmaxGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { - auto* Y = context.Input("Y"); - auto* dY = context.Input(framework::GradVarName("Y")); + auto* Out = context.Input("Out"); + auto* dOut = context.Input(framework::GradVarName("Out")); auto* dX = context.Output(framework::GradVarName("X")); // allocate memory on device. dX->mutable_data(context.GetPlace()); - math::SoftmaxGradFunctor()(context.device_context(), Y, dY, dX); + math::SoftmaxGradFunctor()( + context.template device_context(), Out, dOut, dX); } }; diff --git a/paddle/operators/softmax_with_cross_entropy_op.cc b/paddle/operators/softmax_with_cross_entropy_op.cc index 3dbb62d2e571eb92025c1b3fc0a6653c7cda007a..41e65b701e62bd2e671f3590869a5d7fed90701c 100644 --- a/paddle/operators/softmax_with_cross_entropy_op.cc +++ b/paddle/operators/softmax_with_cross_entropy_op.cc @@ -1,10 +1,10 @@ /* 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 +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 + 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, @@ -13,7 +13,6 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/operators/softmax_with_cross_entropy_op.h" -#include namespace paddle { namespace operators { @@ -21,8 +20,7 @@ namespace operators { class SoftmaxWithCrossEntropyOpMaker : public framework::OpProtoAndCheckerMaker { public: - SoftmaxWithCrossEntropyOpMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker) + SoftmaxWithCrossEntropyOpMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("Logits", "(Tensor, default: Tensor), The unscaled log probabilities " @@ -67,15 +65,15 @@ The equation is as follows: 1) Hard label (one-hot label, so every sample has exactly one class) -$$Loss_j = \f$ -\text{Logit}_{Label_j} + +$$Loss_j = -\text{Logit}_{Label_j} + \log\left(\sum_{i=0}^{K}\exp(\text{Logit}_i)\right), -j = 1, ..., K $\f$$ +j = 1,..., K$$ 2) Soft label (each sample can have a distribution over all classes) -$$Loss_j = \f$ -\sum_{i=0}^{K}\text{Label}_i\left(\text{Logit}_i - +$$Loss_j = -\sum_{i=0}^{K}\text{Label}_i \left(\text{Logit}_i - \log\left(\sum_{i=0}^{K}\exp(\text{Logit}_i)\right)\right), -j = 1,...,K $\f$$ +j = 1,...,K$$ )DOC"); } @@ -120,7 +118,7 @@ class SoftmaxWithCrossEntropyOp : public framework::OperatorWithKernel { } protected: - framework::OpKernelType GetKernelType( + framework::OpKernelType GetActualKernelType( const framework::ExecutionContext& ctx) const override { return framework::OpKernelType( framework::ToDataType(ctx.Input("Logits")->type()), @@ -161,7 +159,7 @@ class SoftmaxWithCrossEntropyOpGrad : public framework::OperatorWithKernel { } protected: - framework::OpKernelType GetKernelType( + framework::OpKernelType GetActualKernelType( const framework::ExecutionContext& ctx) const override { return framework::OpKernelType( framework::ToDataType( @@ -175,8 +173,8 @@ class SoftmaxGradMaker : public framework::SingleGradOpDescMaker { using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; protected: - std::unique_ptr Apply() const override { - auto* grad_op = new framework::OpDescBind(); + std::unique_ptr Apply() const override { + auto* grad_op = new framework::OpDesc(); grad_op->SetType("softmax_with_cross_entropy_grad"); grad_op->SetInput("Label", Input("Label")); grad_op->SetInput("Softmax", Output("Softmax")); @@ -185,7 +183,7 @@ class SoftmaxGradMaker : public framework::SingleGradOpDescMaker { grad_op->SetInput(framework::GradVarName("Loss"), OutputGrad("Loss")); grad_op->SetOutput(framework::GradVarName("Logits"), InputGrad("Logits")); grad_op->SetAttrMap(Attrs()); - return std::unique_ptr(grad_op); + return std::unique_ptr(grad_op); } }; diff --git a/paddle/operators/softmax_with_cross_entropy_op.cu b/paddle/operators/softmax_with_cross_entropy_op.cu index b1faddac3fd21aaf817caf9d3e57e664f4e0e2d5..61583c6161c3bbc62788dc8b6940ddcc29b2302a 100644 --- a/paddle/operators/softmax_with_cross_entropy_op.cu +++ b/paddle/operators/softmax_with_cross_entropy_op.cu @@ -1,10 +1,10 @@ /* 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 +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 + 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, @@ -69,10 +69,10 @@ class SoftmaxWithCrossEntropyCUDAKernel : public framework::OpKernel { softmax->mutable_data(context.GetPlace()); loss->mutable_data(context.GetPlace()); - math::SoftmaxFunctor()(context.device_context(), - logits, softmax); - math::CrossEntropyFunctor()( - context.device_context(), loss, softmax, labels, + math::SoftmaxFunctor()( + context.cuda_device_context(), logits, softmax); + math::CrossEntropyFunctor()( + context.cuda_device_context(), loss, softmax, labels, context.Attr("soft_label")); } }; @@ -98,18 +98,18 @@ class SoftmaxWithCrossEntropyGradCUDAKernel : public framework::OpKernel { if (context.Attr("soft_label")) { const T* label_data = labels->data(); - SoftCrossEntropyGradientKernel<<< - grid, block, 0, reinterpret_cast( - context.device_context()) - .stream()>>>(logit_grad_data, loss_grad_data, - label_data, batch_size, class_num); + SoftCrossEntropyGradientKernel< + T><<() + .stream()>>>(logit_grad_data, loss_grad_data, label_data, + batch_size, class_num); } else { const int64_t* label_data = labels->data(); - CrossEntropyGrad<<< - grid, block, 0, reinterpret_cast( - context.device_context()) - .stream()>>>(logit_grad_data, loss_grad_data, - label_data, batch_size, class_num); + CrossEntropyGrad< + T><<() + .stream()>>>(logit_grad_data, loss_grad_data, label_data, + batch_size, class_num); } } }; @@ -118,9 +118,9 @@ class SoftmaxWithCrossEntropyGradCUDAKernel : public framework::OpKernel { } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(softmax_with_cross_entropy, - ops::SoftmaxWithCrossEntropyCUDAKernel, - ops::SoftmaxWithCrossEntropyCUDAKernel); -REGISTER_OP_GPU_KERNEL(softmax_with_cross_entropy_grad, - ops::SoftmaxWithCrossEntropyGradCUDAKernel, - ops::SoftmaxWithCrossEntropyGradCUDAKernel); +REGISTER_OP_CUDA_KERNEL(softmax_with_cross_entropy, + ops::SoftmaxWithCrossEntropyCUDAKernel, + ops::SoftmaxWithCrossEntropyCUDAKernel); +REGISTER_OP_CUDA_KERNEL(softmax_with_cross_entropy_grad, + ops::SoftmaxWithCrossEntropyGradCUDAKernel, + ops::SoftmaxWithCrossEntropyGradCUDAKernel); diff --git a/paddle/operators/softmax_with_cross_entropy_op.h b/paddle/operators/softmax_with_cross_entropy_op.h index c4ab3f74b4b07d13957d99e01aa4868fac719f61..6bde0f37e06ccf7d81487e0c99227287787c5d72 100644 --- a/paddle/operators/softmax_with_cross_entropy_op.h +++ b/paddle/operators/softmax_with_cross_entropy_op.h @@ -1,10 +1,10 @@ /* 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 +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 + 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, @@ -40,11 +40,12 @@ class SoftmaxWithCrossEntropyKernel : public framework::OpKernel { softmax->mutable_data(context.GetPlace()); loss->mutable_data(context.GetPlace()); - math::SoftmaxFunctor()(context.device_context(), - logits, softmax); - math::CrossEntropyFunctor()( - context.device_context(), loss, softmax, labels, - context.Attr("soft_label")); + auto& dev_ctx = + context.template device_context(); + math::SoftmaxFunctor()(dev_ctx, logits, + softmax); + math::CrossEntropyFunctor()( + dev_ctx, loss, softmax, labels, context.Attr("soft_label")); } }; @@ -62,14 +63,15 @@ class SoftmaxWithCrossEntropyGradKernel : public framework::OpKernel { const int class_num = logit_grad->dims()[1]; auto out_grad_mat = EigenMatrix::From(*out_grad); auto logit_grad_mat = EigenMatrix::From(*logit_grad); - + auto& place = *context.template device_context() + .eigen_device(); if (context.Attr("soft_label")) { auto lbl_mat = EigenMatrix::From(*labels); - logit_grad_mat.device(context.GetEigenDevice()) = + logit_grad_mat.device(place) = out_grad_mat.broadcast(Eigen::DSizes(1, class_num)) * (logit_grad_mat - lbl_mat); } else { - logit_grad_mat.device(context.GetEigenDevice()) = + logit_grad_mat.device(place) = logit_grad_mat * out_grad_mat.broadcast(Eigen::DSizes(1, class_num)); diff --git a/paddle/operators/split_lod_tensor_op.cc b/paddle/operators/split_lod_tensor_op.cc index db635f2ba0804143c9a2e04ff006dfbc8744f3fc..2d8787d740c70f1d4696fdec381b572ecf031f57 100644 --- a/paddle/operators/split_lod_tensor_op.cc +++ b/paddle/operators/split_lod_tensor_op.cc @@ -14,6 +14,7 @@ limitations under the License. */ #include "paddle/framework/op_registry.h" #include "paddle/memory/memcpy.h" +#include "paddle/platform/device_context.h" namespace paddle { namespace operators { @@ -33,7 +34,7 @@ class SplitLoDTensorOp : public framework::OperatorBase { const framework::AttributeMap &attrs) : OperatorBase(type, inputs, outputs, attrs) {} void Run(const framework::Scope &scope, - const platform::DeviceContext &dev_ctx) const override { + const platform::Place &dev_place) const override { auto &x = scope.FindVar(Input("X"))->Get(); auto &mask = scope.FindVar(Input("Mask"))->Get(); auto *out_true = @@ -44,12 +45,15 @@ class SplitLoDTensorOp : public framework::OperatorBase { auto &x_lod = x.lod(); auto &mask_dim = mask.dims(); + platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance(); + auto &dev_ctx = *pool.Get(dev_place); + std::unique_ptr cpu_mask{new framework::LoDTensor()}; if (platform::is_cpu_place(mask.place())) { cpu_mask->ShareDataWith(mask); } else if (platform::is_gpu_place(mask.place())) { #ifdef PADDLE_WITH_CUDA - cpu_mask->CopyFrom(mask, platform::CPUPlace(), dev_ctx); + framework::CopyFrom(mask, platform::CPUPlace(), dev_ctx, cpu_mask.get()); #else PADDLE_THROW("Not supported GPU, Please compile WITH_GPU option"); #endif @@ -105,10 +109,11 @@ class SplitLoDTensorOp : public framework::OperatorBase { continue; } // out[offset: offset+len] = x[each_range.begin: each_range.end] - out->Slice(static_cast(offset), static_cast(offset + len)) - .CopyFrom(x.Slice(static_cast(each_range.begin), - static_cast(each_range.end)), - x.place(), dev_ctx); + auto slice = out->Slice(static_cast(offset), + static_cast(offset + len)); + framework::CopyFrom(x.Slice(static_cast(each_range.begin), + static_cast(each_range.end)), + x.place(), dev_ctx, &slice); offset += len; } } @@ -117,8 +122,7 @@ class SplitLoDTensorOp : public framework::OperatorBase { class SplitLoDTensorOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - SplitLoDTensorOpProtoMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + SplitLoDTensorOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "The input LoDTensor"); AddInput("Mask", "A bool column vector which mask the input"); @@ -163,8 +167,8 @@ class SplitLoDTensorArrayGradMaker : public framework::SingleGradOpDescMaker { using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; protected: - std::unique_ptr Apply() const override { - auto *grad_op = new framework::OpDescBind(); + std::unique_ptr Apply() const override { + auto *grad_op = new framework::OpDesc(); grad_op->SetType("merge_lod_tensor"); grad_op->SetInput("InTrue", OutputGrad("OutTrue")); grad_op->SetInput("InFalse", OutputGrad("OutFalse")); @@ -172,7 +176,7 @@ class SplitLoDTensorArrayGradMaker : public framework::SingleGradOpDescMaker { grad_op->SetInput("X", Input("X")); grad_op->SetOutput("Out", InputGrad("X")); grad_op->SetAttrMap(Attrs()); - return std::unique_ptr(grad_op); + return std::unique_ptr(grad_op); } }; diff --git a/paddle/operators/split_op.cc b/paddle/operators/split_op.cc index 275b25e96aa75fdbcb7275e272c49ea8d278d2c8..4dfae043cb1091c9491d89aec4d1415d4741e013 100644 --- a/paddle/operators/split_op.cc +++ b/paddle/operators/split_op.cc @@ -65,7 +65,7 @@ class SplitOp : public framework::OperatorWithKernel { class SplitOpMaker : public framework::OpProtoAndCheckerMaker { public: - SplitOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + SplitOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "(Tensor) Input tensor of the split operator."); AddOutput("Out", "(Tensor) Output tensors of the split operator.") @@ -108,13 +108,13 @@ class SplitGradMaker : public framework::SingleGradOpDescMaker { using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; protected: - std::unique_ptr Apply() const override { - auto op = new framework::OpDescBind(); + std::unique_ptr Apply() const override { + auto op = new framework::OpDesc(); op->SetType("concat"); op->SetInput("X", OutputGrad("Out")); op->SetOutput("Out", InputGrad("X")); op->SetAttrMap(Attrs()); - return std::unique_ptr(op); + return std::unique_ptr(op); } }; diff --git a/paddle/operators/split_op.cu.cc b/paddle/operators/split_op.cu.cc index 93d1fc3c44cbc146c945c51af1abe6494572d1ae..dbad0bbf68d7924cfba80721bb3294b7e0cfac00 100644 --- a/paddle/operators/split_op.cu.cc +++ b/paddle/operators/split_op.cu.cc @@ -14,5 +14,5 @@ limitations under the License. */ #include "paddle/operators/split_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(split, - ops::SplitOpKernel); +REGISTER_OP_CUDA_KERNEL( + split, ops::SplitOpKernel); diff --git a/paddle/operators/split_op.h b/paddle/operators/split_op.h index fa26e5f677b18c84b45dd583004d02cab4c1d375..a38c435d531d7da2a1a60eb2c455bc1782c1cd4c 100644 --- a/paddle/operators/split_op.h +++ b/paddle/operators/split_op.h @@ -21,7 +21,7 @@ limitations under the License. */ namespace paddle { namespace operators { -template +template class SplitOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { diff --git a/paddle/operators/spp_op.cc b/paddle/operators/spp_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..c0aa87b0f06ca9c7d156dfdf8df188da68ac1450 --- /dev/null +++ b/paddle/operators/spp_op.cc @@ -0,0 +1,99 @@ +/* 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. +Indicesou 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/operators/spp_op.h" +namespace paddle { +namespace operators { + +class SppOpMaker : public framework::OpProtoAndCheckerMaker { + public: + SppOpMaker(OpProto* proto, OpAttrChecker* op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput( + "X", + "(Tensor) The input tensor of spp operator. " + "The format of input tensor is NCHW. Where N is batch size, C is the " + "number of channels, H and W is the height and width of feature."); + AddOutput("Out", + "(Tensor) The output tensor of spp operator." + "N * M." + "M = C * H * W"); + AddAttr("pyramid_height", "(int), multi level pooling"); + AddAttr( + "pooling_type", + "(string), pooling type, can be \"max\" for max-pooling " + "and \"avg\" for average-pooling.") + .InEnum({"max", "avg"}); + AddComment(R"DOC( + "With spatial pyramid pooling, the input image can + be of any sizes. This not only allows arbitrary aspect + ratios, but also allows arbitrary scales. We can resize + the input image to any scale (e.g., min(w, h)=180, 224, + ...) and apply the same deep network. When the + input image is at different scales, the network (with + the same filter sizes) will extract features at different + scales. The scales play important roles in traditional + methods. + Input shape: $(N, C_{in}, H_{in}, W_{in})$ + Output shape: $(H_{out}, W_{out})$ + Where + $$ + H_{out} = N \\ + W_{out} = (((4^pyramid_height) - 1) / (4 - 1))$ * C_{in} + $$ + paper https://arxiv.org/pdf/1406.4729v4.pdf + )DOC"); + } +}; + +class SppOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of SppOp" + "should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of SppOp should not be null."); + auto in_x_dims = ctx->GetInputDim("X"); + int pyramid_height = ctx->Attrs().Get("pyramid_height"); + PADDLE_ENFORCE(in_x_dims.size() == 4, + "Spping intput must be of 4-dimensional."); + int outlen = ((std::pow(4, pyramid_height) - 1) / (4 - 1)) * in_x_dims[1]; + std::vector output_shape({in_x_dims[0], outlen}); + ctx->SetOutputDim("Out", framework::make_ddim(output_shape)); + } +}; + +class SppOpGrad : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) must not be null."); + PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("X")), + "Input(X@GRAD) should not be null."); + ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X")); + } +}; +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP(spp, ops::SppOp, ops::SppOpMaker, spp_grad, ops::SppOpGrad); +REGISTER_OP_CPU_KERNEL( + spp, ops::SppKernel, + ops::SppKernel); +REGISTER_OP_CPU_KERNEL( + spp_grad, ops::SppGradKernel, + ops::SppGradKernel); diff --git a/paddle/operators/spp_op.cu.cc b/paddle/operators/spp_op.cu.cc new file mode 100644 index 0000000000000000000000000000000000000000..761e4d6c4a9639898ba548d56bed3c8817720c1b --- /dev/null +++ b/paddle/operators/spp_op.cu.cc @@ -0,0 +1,23 @@ +/* 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. +Indicesou 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/operators/spp_op.h" + +namespace ops = paddle::operators; +REGISTER_OP_CUDA_KERNEL( + spp, ops::SppKernel, + ops::SppKernel); +REGISTER_OP_CUDA_KERNEL( + spp_grad, ops::SppGradKernel, + ops::SppGradKernel); diff --git a/paddle/operators/spp_op.h b/paddle/operators/spp_op.h new file mode 100644 index 0000000000000000000000000000000000000000..f35b305d02c73bcae6e72b8afa5ce55148ea98b8 --- /dev/null +++ b/paddle/operators/spp_op.h @@ -0,0 +1,161 @@ +/* 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. +Indicesou 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/framework/op_registry.h" +#include "paddle/operators/math/math_function.h" +#include "paddle/operators/math/pooling.h" +#include "paddle/operators/strided_memcpy.h" + +namespace paddle { +namespace operators { +template +class SppKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + const framework::Tensor* in_x = context.Input("X"); + auto* out = context.Output("Out"); + int pyramid_height = context.template Attr("pyramid_height"); + std::string pooling_type = + context.template Attr("pooling_type"); + out->mutable_data(context.GetPlace()); + auto out_stride = framework::stride(out->dims()); + int input_h = in_x->dims()[2]; + int input_w = in_x->dims()[3]; + size_t output_offset = 0; + for (int p = 0; p < pyramid_height; ++p) { + int bins = std::pow(2, p); + int kernel_size_h = std::ceil(input_h / static_cast(bins)); + int kernel_size_w = std::ceil(input_w / static_cast(bins)); + int padding_h = (kernel_size_h * bins - input_h + 1) / 2; + int padding_w = (kernel_size_w * bins - input_w + 1) / 2; + std::vector kernel_size({kernel_size_h, kernel_size_w}); + std::vector strides({kernel_size_h, kernel_size_w}); + std::vector paddings({padding_h, padding_w}); + // pooling output shape + framework::Tensor out_level; + std::vector output_shape_vec( + {in_x->dims()[0], in_x->dims()[1], bins, bins}); + framework::DDim output_shape(framework::make_ddim(output_shape_vec)); + out_level.mutable_data(output_shape, context.GetPlace()); + // pooling + if (pooling_type == "max") { + math::Pool2dFunctor, T> pool_forward; + math::MaxPool max_process; + pool_forward(context.template device_context(), *in_x, + kernel_size, strides, paddings, max_process, &out_level); + } else if (pooling_type == "avg") { + math::Pool2dFunctor, T> pool_forward; + math::AvgPool avg_process; + pool_forward(context.template device_context(), *in_x, + kernel_size, strides, paddings, avg_process, &out_level); + } + // flatten pooling output shape + int output_flatten_w = in_x->dims()[1] * bins * bins; + std::vector output_flatten_shape_vec( + {in_x->dims()[0], output_flatten_w}); + framework::DDim output_flatten_shape( + framework::make_ddim(output_flatten_shape_vec)); + out_level.Resize(output_flatten_shape); + // concat + auto out_level_stride = framework::stride(out_level.dims()); + StridedMemcpy(context.template device_context(), + out_level.data(), out_level_stride, out_level.dims(), + out_stride, out->data() + output_offset); + output_offset += out_level.dims()[1] * out_level_stride[1]; + } + } +}; +template +class SppGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + const framework::Tensor* in_x = context.Input("X"); + const framework::Tensor* out = context.Input("Out"); + const framework::Tensor* out_grad = + context.Input(framework::GradVarName("Out")); + framework::Tensor* in_x_grad = + context.Output(framework::GradVarName("X")); + int pyramid_height = context.template Attr("pyramid_height"); + std::string pooling_type = + context.template Attr("pooling_type"); + auto& device_ctx = context.template device_context(); + math::SetConstant zero; + in_x_grad->mutable_data(context.GetPlace()); + zero(device_ctx, in_x_grad, static_cast(0)); + auto out_stride = framework::stride(out->dims()); + int input_h = in_x->dims()[2]; + int input_w = in_x->dims()[3]; + size_t out_offset = 0; + for (int p = 0; p < pyramid_height; ++p) { + int bins = std::pow(2, p); + int kernel_size_h = std::ceil(input_h / static_cast(bins)); + int kernel_size_w = std::ceil(input_w / static_cast(bins)); + int padding_h = (kernel_size_h * bins - input_h + 1) / 2; + int padding_w = (kernel_size_w * bins - input_w + 1) / 2; + std::vector kernel_size({kernel_size_h, kernel_size_w}); + std::vector strides({kernel_size_h, kernel_size_w}); + std::vector paddings({padding_h, padding_w}); + // split out and outgrad ... to flatten + framework::Tensor out_level; + framework::Tensor outgrad_level; + int out_flatten_w = in_x->dims()[1] * bins * bins; + std::vector out_flatten_shape_vec( + {in_x->dims()[0], out_flatten_w}); + framework::DDim out_flatten_shape( + framework::make_ddim(out_flatten_shape_vec)); + out_level.mutable_data(out_flatten_shape, context.GetPlace()); + outgrad_level.mutable_data(out_flatten_shape, context.GetPlace()); + auto flatten_stride = framework::stride(out_level.dims()); + // memcpy + StridedMemcpy(context.template device_context(), + out->data() + out_offset, out_stride, + out_level.dims(), flatten_stride, out_level.data()); + + StridedMemcpy(context.template device_context(), + out_grad->data() + out_offset, out_stride, + outgrad_level.dims(), flatten_stride, + outgrad_level.data()); + out_offset += out_level.dims()[1] * out_stride[1]; + // flatten backward to nchw + + std::vector out_shape_vec({in_x->dims()[0], in_x->dims()[1]}); + out_shape_vec.push_back( + (input_h - kernel_size_h + 2 * padding_h) / kernel_size_h + 1); + out_shape_vec.push_back( + (input_w - kernel_size_w + 2 * padding_w) / kernel_size_w + 1); + framework::DDim out_shape(framework::make_ddim(out_shape_vec)); + out_level.ShareDataWith(out_level); + out_level.Resize(out_shape); + outgrad_level.ShareDataWith(outgrad_level); + outgrad_level.Resize(out_shape); + // pooling backward + if (pooling_type == "max") { + math::MaxPool2dGradFunctor pool2d_backward; + pool2d_backward(context.template device_context(), *in_x, + *&out_level, *&outgrad_level, kernel_size, strides, + paddings, in_x_grad); + } else if (pooling_type == "avg") { + math::Pool2dGradFunctor, T> + pool_backward; + math::AvgPoolGrad avg_process; + pool_backward(context.template device_context(), *in_x, + *&out_level, *&outgrad_level, kernel_size, strides, + paddings, avg_process, in_x_grad); + } + } + } +}; +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/squared_l2_distance_op.cc b/paddle/operators/squared_l2_distance_op.cc index bec2a2c18ae8da892ee7d71f45afe53c887c0f57..9e097176f3434e81e31f2ecf4093af47b654e816 100644 --- a/paddle/operators/squared_l2_distance_op.cc +++ b/paddle/operators/squared_l2_distance_op.cc @@ -56,8 +56,7 @@ class SquaredL2DistanceOp : public framework::OperatorWithKernel { class SquaredL2DistanceOpMaker : public framework::OpProtoAndCheckerMaker { public: - SquaredL2DistanceOpMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker) + SquaredL2DistanceOpMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "(Tensor) Input of SquaredL2DistanceOp."); AddInput("Y", "(Tensor) Target of SquaredL2DistanceOp."); @@ -115,7 +114,7 @@ REGISTER_OP(squared_l2_distance, ops::SquaredL2DistanceOp, ops::SquaredL2DistanceGradOp); REGISTER_OP_CPU_KERNEL( squared_l2_distance, - ops::SquaredL2DistanceKernel); -REGISTER_OP_CPU_KERNEL( - squared_l2_distance_grad, - ops::SquaredL2DistanceGradKernel); + ops::SquaredL2DistanceKernel); +REGISTER_OP_CPU_KERNEL(squared_l2_distance_grad, + ops::SquaredL2DistanceGradKernel< + paddle::platform::CPUDeviceContext, float>); diff --git a/paddle/operators/squared_l2_distance_op.cu b/paddle/operators/squared_l2_distance_op.cu index 3fe62f1a9cb56722ea544b0fed052ac384e799aa..f2648dde5eb9c56aed3fad81521e6207dc2d973e 100644 --- a/paddle/operators/squared_l2_distance_op.cu +++ b/paddle/operators/squared_l2_distance_op.cu @@ -1,25 +1,25 @@ /* 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 +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 + 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. */ +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. */ #define EIGEN_USE_GPU #include "paddle/operators/squared_l2_distance_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL( +REGISTER_OP_CUDA_KERNEL( squared_l2_distance, - ops::SquaredL2DistanceKernel); -REGISTER_OP_GPU_KERNEL( - squared_l2_distance_grad, - ops::SquaredL2DistanceGradKernel); + ops::SquaredL2DistanceKernel); +REGISTER_OP_CUDA_KERNEL(squared_l2_distance_grad, + ops::SquaredL2DistanceGradKernel< + paddle::platform::CUDADeviceContext, float>); diff --git a/paddle/operators/squared_l2_distance_op.h b/paddle/operators/squared_l2_distance_op.h index 259ef4029646914f83a112b9c6d7fdf8401483f6..5bd5f4819a35966b73038f433d38c06031e18715 100644 --- a/paddle/operators/squared_l2_distance_op.h +++ b/paddle/operators/squared_l2_distance_op.h @@ -27,7 +27,7 @@ template using EigenMatrix = framework::EigenMatrix; -template +template class SquaredL2DistanceKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { @@ -51,7 +51,8 @@ class SquaredL2DistanceKernel : public framework::OpKernel { auto sub_result = EigenMatrix::From(*out0); auto z = EigenVector::Flatten(*out1); - auto place = context.GetEigenDevice(); + auto& place = + *context.template device_context().eigen_device(); auto x_dims = x.dimensions(); auto y_dims = y.dimensions(); // buffer the substraction result @@ -67,7 +68,7 @@ class SquaredL2DistanceKernel : public framework::OpKernel { } }; -template +template class SquaredL2DistanceGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { @@ -89,7 +90,8 @@ class SquaredL2DistanceGradKernel : public framework::OpKernel { sub_result; // propagate back to input - auto eigen_place = context.GetEigenDevice(); + auto& eigen_place = + *context.template device_context().eigen_device(); if (x_g) { x_g->mutable_data(context.GetPlace()); // eigen matrix diff --git a/paddle/operators/squared_l2_norm_op.cc b/paddle/operators/squared_l2_norm_op.cc index 3c10e6159f44bc8c21b1e79aefaa962c7a2b64ed..6626bf0375548eac457f960105ce33e63e1a3706 100644 --- a/paddle/operators/squared_l2_norm_op.cc +++ b/paddle/operators/squared_l2_norm_op.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/operators/squared_l2_norm_op.h" @@ -48,8 +48,7 @@ class SquaredL2NormGradOp : public framework::OperatorWithKernel { class SquaredL2NormOpMaker : public framework::OpProtoAndCheckerMaker { public: - SquaredL2NormOpMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker) + SquaredL2NormOpMaker(OpProto* proto, OpAttrChecker* op_checker) : framework::OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "(Tensor) The input of squared_l2_norm op."); AddOutput("Out", "(Scalar) The output of squared_l2_norm op."); @@ -72,7 +71,7 @@ REGISTER_OP(squared_l2_norm, ops::SquaredL2NormOp, ops::SquaredL2NormOpMaker, squared_l2_norm_grad, ops::SquaredL2NormGradOp); REGISTER_OP_CPU_KERNEL( squared_l2_norm, - ops::SquaredL2NormKernel); + ops::SquaredL2NormKernel); REGISTER_OP_CPU_KERNEL( squared_l2_norm_grad, - ops::SquaredL2NormGradKernel); + ops::SquaredL2NormGradKernel); diff --git a/paddle/operators/squared_l2_norm_op.cu b/paddle/operators/squared_l2_norm_op.cu index d384e9c28c9150fa901404478739ff809f29126f..b222113a8c82061bd841da440a714b66f6c1fb9c 100644 --- a/paddle/operators/squared_l2_norm_op.cu +++ b/paddle/operators/squared_l2_norm_op.cu @@ -1,24 +1,24 @@ /* 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 +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 + 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. */ +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. */ #define EIGEN_USE_GPU #include "paddle/operators/squared_l2_norm_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL( +REGISTER_OP_CUDA_KERNEL( squared_l2_norm, - ops::SquaredL2NormKernel); -REGISTER_OP_GPU_KERNEL( + ops::SquaredL2NormKernel); +REGISTER_OP_CUDA_KERNEL( squared_l2_norm_grad, - ops::SquaredL2NormGradKernel); + ops::SquaredL2NormGradKernel); diff --git a/paddle/operators/squared_l2_norm_op.h b/paddle/operators/squared_l2_norm_op.h index 48d7b1c2d56882f04330dbf27b0a92e37cb8874c..1ce26c775ed5700cee73d00f4c51d58a692a1152 100644 --- a/paddle/operators/squared_l2_norm_op.h +++ b/paddle/operators/squared_l2_norm_op.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/framework/eigen.h" @@ -20,7 +20,7 @@ namespace paddle { namespace operators { // Out = sum(square(X)) -template +template class SquaredL2NormKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext &context) const override { @@ -30,14 +30,15 @@ class SquaredL2NormKernel : public framework::OpKernel { auto x = framework::EigenVector::Flatten(*X); auto out = framework::EigenScalar::From(*Out); - auto place = context.GetEigenDevice(); + auto *place = + context.template device_context().eigen_device(); - out.device(place) = x.square().sum(); + out.device(*place) = x.square().sum(); } }; // dX = X -template +template class SquaredL2NormGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext &context) const override { @@ -53,10 +54,11 @@ class SquaredL2NormGradKernel : public framework::OpKernel { auto x = framework::EigenVector::Flatten(*X); auto dout = framework::EigenVector::Flatten(*dOut); auto dx = framework::EigenVector::Flatten(*dX); - auto place = context.GetEigenDevice(); + auto *place = + context.template device_context().eigen_device(); Eigen::DSizes x_dsize(X->numel()); - dx.device(place) = (dout.broadcast(x_dsize) * x) * static_cast(2.0); + dx.device(*place) = (dout.broadcast(x_dsize) * x) * static_cast(2.0); } }; diff --git a/paddle/operators/strided_memcpy.h b/paddle/operators/strided_memcpy.h index c9dd80518424017d9834a2bf7aee14caa56c9d79..735cabcd973a3b5ea1ab8ab57091eae14e23b89b 100644 --- a/paddle/operators/strided_memcpy.h +++ b/paddle/operators/strided_memcpy.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/operators/detail/strided_memcpy.h" diff --git a/paddle/operators/strided_memcpy_test.cc b/paddle/operators/strided_memcpy_test.cc index 68f064eaee5851333ddf9767b7138da83a28503d..06d81188558aad85c41c56ddefad3617d48da74c 100644 --- a/paddle/operators/strided_memcpy_test.cc +++ b/paddle/operators/strided_memcpy_test.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/operators/strided_memcpy.h" #include "gtest/gtest.h" @@ -82,11 +82,13 @@ TEST(StridedMemcpy, GPUCrop) { }; // clang-format on - platform::GPUPlace gpu0(0); + platform::CUDAPlace gpu0(0); platform::CPUPlace cpu; + platform::CUDADeviceContext ctx(gpu0); + int* gpu_src = reinterpret_cast(memory::Alloc(gpu0, sizeof(src))); - memory::Copy(gpu0, gpu_src, cpu, src, sizeof(src)); + memory::Copy(gpu0, gpu_src, cpu, src, sizeof(src), ctx.stream()); framework::DDim src_stride({5, 1}); @@ -96,7 +98,6 @@ TEST(StridedMemcpy, GPUCrop) { framework::DDim dst_dim({2, 2}); framework::DDim dst_stride({2, 1}); - platform::CUDADeviceContext ctx(gpu0); StridedMemcpy(ctx, gpu_src + 1, src_stride, dst_dim, dst_stride, gpu_dst); @@ -120,11 +121,12 @@ TEST(StridedMemcpy, GPUConcat) { }; // clang-format on - platform::GPUPlace gpu0(0); + platform::CUDAPlace gpu0(0); platform::CPUPlace cpu; + platform::CUDADeviceContext ctx(gpu0); int* gpu_src = reinterpret_cast(memory::Alloc(gpu0, sizeof(src))); - memory::Copy(gpu0, gpu_src, cpu, src, sizeof(src)); + memory::Copy(gpu0, gpu_src, cpu, src, sizeof(src), ctx.stream()); int dst[8]; int* gpu_dst = reinterpret_cast(memory::Alloc(gpu0, sizeof(dst))); @@ -132,7 +134,6 @@ TEST(StridedMemcpy, GPUConcat) { framework::DDim src_stride({2, 1}); framework::DDim dst_dim({2, 2}); framework::DDim dst_stride({4, 1}); - platform::CUDADeviceContext ctx(gpu0); StridedMemcpy(ctx, gpu_src, src_stride, dst_dim, dst_stride, gpu_dst); StridedMemcpy(ctx, gpu_src, src_stride, dst_dim, dst_stride, diff --git a/paddle/operators/sum_op.cc b/paddle/operators/sum_op.cc index c2b7632b2865a3ef66051d815d7722a08c6a8cbd..b86e8266425ca094a51d224fd39ce33700057f13 100644 --- a/paddle/operators/sum_op.cc +++ b/paddle/operators/sum_op.cc @@ -29,7 +29,7 @@ class SumOp : public framework::OperatorWithKernel { "Output(Out) of SumOp should not be null."); if (ctx->IsRuntime() && ctx->GetOutputsVarType("Out")[0] == - framework::VarDesc::LOD_TENSOR_ARRAY) { + framework::proto::VarDesc::LOD_TENSOR_ARRAY) { return; // skip runtime infershape when is tensor array; } @@ -37,23 +37,43 @@ class SumOp : public framework::OperatorWithKernel { size_t N = x_dims.size(); PADDLE_ENFORCE_GT(N, 1, "Input tensors count should > 1."); - auto in_dim = x_dims[0]; - for (size_t i = 1; i < N; i++) { - auto dim = x_dims[i]; - PADDLE_ENFORCE_EQ(in_dim, dim, "Input tensors must have same shape"); + framework::DDim in_dim({0}); + for (auto& x_dim : x_dims) { + if (framework::product(x_dim) == 0) { + continue; + } + if (framework::product(in_dim) == 0) { + in_dim = x_dim; + } else { + PADDLE_ENFORCE_EQ(in_dim, x_dim, "Input tensors must have same shape"); + } } ctx->SetOutputDim("Out", in_dim); ctx->ShareLoD("X", /*->*/ "Out"); } protected: - framework::OpKernelType GetKernelType( + framework::OpKernelType GetActualKernelType( const framework::ExecutionContext& ctx) const override { auto x_vars = ctx.MultiInputVar("X"); if (x_vars[0]->IsType()) { + int dtype = -1; + for (auto& x_var : x_vars) { + auto& lod_tensor = x_var->Get(); + if (lod_tensor.numel() == 0) { + continue; + } + if (dtype == -1) { + dtype = framework::ToDataType(lod_tensor.type()); + } else { + PADDLE_ENFORCE_EQ(dtype, framework::ToDataType(lod_tensor.type())); + } + } + PADDLE_ENFORCE_NE(dtype, -1, + "Sum operator should have at least one tensor"); + return framework::OpKernelType( - framework::ToDataType(x_vars[0]->Get().type()), - ctx.device_context()); + static_cast(dtype), ctx.device_context()); } else if (x_vars[0]->IsType()) { return framework::OpKernelType( framework::ToDataType( @@ -78,7 +98,7 @@ class SumOp : public framework::OperatorWithKernel { class SumOpMaker : public framework::OpProtoAndCheckerMaker { public: - SumOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker) + SumOpMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "(vector) The input tensors of sum operator.") .AsDuplicable(); @@ -86,8 +106,8 @@ class SumOpMaker : public framework::OpProtoAndCheckerMaker { AddComment(R"DOC( Sum operator. -This operators sums the input tensors. All the inputs can carry the -LoD (Level of Details) information. However, the output only shares +This operators sums the input tensors. All the inputs can carry the +LoD (Level of Details) information. However, the output only shares the LoD information with the first input. )DOC"); } @@ -95,10 +115,10 @@ the LoD information with the first input. class SumOpVarTypeInference : public framework::VarTypeInference { public: - void operator()(const framework::OpDescBind& op_desc, - framework::BlockDescBind* block) const override { + void operator()(const framework::OpDesc& op_desc, + framework::BlockDesc* block) const override { auto& inputs = op_desc.Input("X"); - auto var_type = framework::VarDesc::SELECTED_ROWS; + auto var_type = framework::proto::VarDesc::SELECTED_ROWS; for (auto& name : op_desc.Input("X")) { VLOG(10) << name << " " @@ -108,12 +128,12 @@ class SumOpVarTypeInference : public framework::VarTypeInference { bool any_input_is_lod_tensor = std::any_of( inputs.begin(), inputs.end(), [block](const std::string& name) { return block->FindRecursiveOrCreateVar(name)->GetType() == - framework::VarDesc::LOD_TENSOR; + framework::proto::VarDesc::LOD_TENSOR; }); auto is_tensor_array = [block](const std::string& name) { return detail::Ref(block->FindRecursiveOrCreateVar(name)).GetType() == - framework::VarDesc::LOD_TENSOR_ARRAY; + framework::proto::VarDesc::LOD_TENSOR_ARRAY; }; bool any_input_is_tensor_array = @@ -132,9 +152,9 @@ class SumOpVarTypeInference : public framework::VarTypeInference { PADDLE_ENFORCE(all_inputs_are_tensor_array, "Not all inputs are tensor array:\n%s", os.str()); } - var_type = framework::VarDesc::LOD_TENSOR_ARRAY; + var_type = framework::proto::VarDesc::LOD_TENSOR_ARRAY; } else if (any_input_is_lod_tensor) { - var_type = framework::VarDesc::LOD_TENSOR; + var_type = framework::proto::VarDesc::LOD_TENSOR; } auto out_var_name = op_desc.Output("Out").front(); @@ -149,20 +169,19 @@ class SumGradMaker : public framework::GradOpDescMakerBase { public: using framework::GradOpDescMakerBase::GradOpDescMakerBase; - std::vector> operator()() - const override { - auto x_grads = InputGrad("X"); - std::vector> grad_ops; + std::vector> operator()() const override { + auto x_grads = InputGrad("X", false); + std::vector> grad_ops; grad_ops.reserve(x_grads.size()); auto og = OutputGrad("Out"); std::transform(x_grads.begin(), x_grads.end(), std::back_inserter(grad_ops), [&og](const std::string& x_grad) { - auto* grad_op = new framework::OpDescBind(); + auto* grad_op = new framework::OpDesc(); grad_op->SetType("scale"); grad_op->SetInput("X", og); grad_op->SetOutput("Out", {x_grad}); grad_op->SetAttr("scale", 1.0f); - return std::unique_ptr(grad_op); + return std::unique_ptr(grad_op); }); return grad_ops; } @@ -175,5 +194,8 @@ namespace ops = paddle::operators; REGISTER_OPERATOR(sum, ops::SumOp, ops::SumOpMaker, ops::SumGradMaker, ops::SumOpVarTypeInference); -REGISTER_OP_CPU_KERNEL(sum, ops::SumKernel, - ops::SumKernel); +REGISTER_OP_CPU_KERNEL( + sum, ops::SumKernel, + ops::SumKernel, + ops::SumKernel, + ops::SumKernel); diff --git a/paddle/operators/sum_op.cu b/paddle/operators/sum_op.cu index 5cf05b876b6d6a2ce61d9e10b7ec52ed3cef57d7..873155076c179d5280a418e25fd39fdaf4b0a2b2 100644 --- a/paddle/operators/sum_op.cu +++ b/paddle/operators/sum_op.cu @@ -13,5 +13,8 @@ limitations under the License. */ #include "paddle/operators/sum_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(sum, ops::SumKernel, - ops::SumKernel); +REGISTER_OP_CUDA_KERNEL( + sum, ops::SumKernel, + ops::SumKernel, + ops::SumKernel, + ops::SumKernel); diff --git a/paddle/operators/sum_op.h b/paddle/operators/sum_op.h index 4ca15611392b3117aa6c92cba95911eb8bebeb15..552b48f608b7e0248f03dbea940a83f112a67712 100644 --- a/paddle/operators/sum_op.h +++ b/paddle/operators/sum_op.h @@ -26,7 +26,7 @@ template using EigenVector = framework::EigenVector; -template +template class SumKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext &context) const override { @@ -37,27 +37,32 @@ class SumKernel : public framework::OpKernel { bool in_place = out_var == in_vars[0]; if (out_var->IsType()) { - auto *out = context.Output("Out"); - out->mutable_data(context.GetPlace()); - + auto *out = context.Output("Out"); + if (!in_place) { + out->mutable_data(context.GetPlace()); + } auto result = EigenVector::Flatten(*out); - if (!in_place) { - math::SetConstant constant_functor; - constant_functor(context.device_context(), out, 0.0); + math::SetConstant constant_functor; + constant_functor(context.template device_context(), out, + 0.0); } - math::SelectedRowsAddToTensor functor; - auto place = context.GetEigenDevice(); + math::SelectedRowsAddToTensor functor; + auto &place = + *context.template device_context().eigen_device(); // If in_place, just skip the first tensor for (int i = in_place ? 1 : 0; i < N; i++) { if (in_vars[i]->IsType()) { auto &in_t = in_vars[i]->Get(); + if (in_t.numel() == 0) { + continue; + } auto in = EigenVector::Flatten(in_t); result.device(place) = result + in; } else if (in_vars[i]->IsType()) { auto &in_t = in_vars[i]->Get(); - functor(context.device_context(), in_t, out); + functor(context.template device_context(), in_t, out); } else { PADDLE_THROW("Variable type must be LoDTensor/SelectedRows."); } @@ -79,14 +84,14 @@ class SumKernel : public framework::OpKernel { out_value->Resize(framework::make_ddim(in_dim_vec)); out_value->mutable_data(context.GetPlace()); - math::SelectedRowsAddTo functor; + math::SelectedRowsAddTo functor; int64_t offset = 0; for (int i = 0; i < N; i++) { PADDLE_ENFORCE_EQ(out->height(), - in_vars[i]->Get().height()) - functor(context.device_context(), in_vars[i]->Get(), - offset, out); + in_vars[i]->Get().height()); + functor(context.template device_context(), + in_vars[i]->Get(), offset, out); offset += in_vars[i]->Get().value().numel(); } } else if (out_var->IsType()) { @@ -102,14 +107,15 @@ class SumKernel : public framework::OpKernel { out_array.resize(i + 1); } if (out_array[i].numel() == 0) { - out_array[i].CopyFrom(in_array[i], in_array[i].place(), - context.device_context()); + framework::CopyFrom(in_array[i], in_array[i].place(), + context.device_context(), &out_array[i]); out_array[i].set_lod(in_array[i].lod()); } else { PADDLE_ENFORCE(out_array[i].lod() == in_array[i].lod()); auto in = EigenVector::Flatten(in_array[i]); auto result = EigenVector::Flatten(out_array[i]); - result.device(context.GetEigenDevice()) = result + in; + result.device(*context.template device_context() + .eigen_device()) = result + in; } } } diff --git a/paddle/operators/tensor.save b/paddle/operators/tensor.save new file mode 100644 index 0000000000000000000000000000000000000000..c24308a7d0131b84c28c0a9857cce4949afb2091 Binary files /dev/null and b/paddle/operators/tensor.save differ diff --git a/paddle/operators/tensor_array_read_write_op.cc b/paddle/operators/tensor_array_read_write_op.cc index ae1b48d7a8e3d573a5134a822a2ed5ef70511077..d5ff3e3fce29b1a888b2cd4d307c2655669e3e4c 100644 --- a/paddle/operators/tensor_array_read_write_op.cc +++ b/paddle/operators/tensor_array_read_write_op.cc @@ -1,16 +1,16 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +/* 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 +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 + 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. */ +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/operators/array_operator.h" #include "paddle/operators/detail/safe_ref.h" namespace paddle { @@ -25,11 +25,11 @@ class WriteToArrayOp : public ArrayOp { : ArrayOp(type, inputs, outputs, attrs) {} void Run(const framework::Scope &scope, - const platform::DeviceContext &dev_ctx) const override { + const platform::Place &place) const override { auto *x = scope.FindVar(Input("X")); - PADDLE_ENFORCE(x != nullptr, "X must be set"); + if (x == nullptr) return; auto &x_tensor = x->Get(); - size_t offset = GetOffset(scope, dev_ctx); + size_t offset = GetOffset(scope, place); auto *out = scope.FindVar(Output("Out"))->GetMutable(); if (offset >= out->size()) { @@ -37,16 +37,26 @@ class WriteToArrayOp : public ArrayOp { << " to " << offset + 1; out->resize(offset + 1); } - auto *out_tensor = &out->at(offset); - out_tensor->CopyFrom(x_tensor, dev_ctx.GetPlace(), dev_ctx); - out_tensor->set_lod(x_tensor.lod()); + if (x_tensor.memory_size() > 0) { + auto *out_tensor = &out->at(offset); + + platform::DeviceContextPool &pool = + platform::DeviceContextPool::Instance(); + auto &dev_ctx = *pool.Get(place); + + CopyFrom(x_tensor, place, dev_ctx, out_tensor); + out_tensor->set_lod(x_tensor.lod()); + } else { + VLOG(10) << "WARNING: The input tensor 'x_tensor' holds no memory, so " + "nothing has been written to output array[" + << offset << "]."; + } } }; class WriteToArrayOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - WriteToArrayOpProtoMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + WriteToArrayOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "(LoDTensor) the tensor will be written to tensor array"); AddInput( @@ -54,12 +64,16 @@ class WriteToArrayOpProtoMaker : public framework::OpProtoAndCheckerMaker { "(Tensor) the subscript index in tensor array. The number of element " "should be 1"); AddOutput("Out", "(TensorArray) the tensor array will be written"); - AddComment(R"DOC(Write a LoDTensor to a LoDTensor array. + AddComment(R"DOC( +WriteToArray Operator. + +This operator writes a LoDTensor to a LoDTensor array. -Assume T is LoDTensor, i is the subscript of the array, and A is the array. The +Assume $T$ is LoDTensor, $i$ is the subscript of the array, and $A$ is the array. The equation is -A[i] = T +$$A[i] = T$$ + )DOC"); } }; @@ -70,7 +84,9 @@ class WriteToArrayInferShape : public framework::InferShapeBase { PADDLE_ENFORCE(context->HasInput("I"), "Must set the subscript index"); PADDLE_ENFORCE_EQ(framework::product(context->GetInputDim("I")), 1, "The number of element of subscript index must be 1"); - PADDLE_ENFORCE(context->HasInput("X"), NotHasXError()); + if (!context->HasInput("X")) { + return; + } PADDLE_ENFORCE(context->HasOutput("Out"), NotHasOutError()); context->SetOutputDim("Out", context->GetInputDim("X")); } @@ -85,17 +101,18 @@ class WriteToArrayInferShape : public framework::InferShapeBase { class WriteToArrayInferVarType : public framework::VarTypeInference { public: - void operator()(const framework::OpDescBind &op_desc, - framework::BlockDescBind *block) const override { + void operator()(const framework::OpDesc &op_desc, + framework::BlockDesc *block) const override { auto x_name = op_desc.Input("X")[0]; auto out_name = op_desc.Output("Out")[0]; VLOG(10) << "Set Variable " << out_name << " as LOD_TENSOR_ARRAY"; auto &out = detail::Ref(block->FindRecursiveOrCreateVar(out_name), "Cannot found %s", out_name); - out.SetType(framework::VarDesc::LOD_TENSOR_ARRAY); - auto &x = - detail::Ref(block->FindVarRecursive(x_name), "Cannot found %s", x_name); - out.SetDataType(x.GetDataType()); + out.SetType(framework::proto::VarDesc::LOD_TENSOR_ARRAY); + auto *x = block->FindVarRecursive(x_name); + if (x != nullptr) { + out.SetDataType(x->GetDataType()); + } } }; @@ -107,36 +124,45 @@ class ReadFromArrayOp : public ArrayOp { const framework::AttributeMap &attrs) : ArrayOp(type, inputs, outputs, attrs) {} void Run(const framework::Scope &scope, - const platform::DeviceContext &dev_ctx) const override { + const platform::Place &place) const override { auto *x = scope.FindVar(Input("X")); PADDLE_ENFORCE(x != nullptr, "X must be set"); auto &x_array = x->Get(); auto *out = scope.FindVar(Output("Out")); PADDLE_ENFORCE(out != nullptr, "Out must be set"); - auto *out_tensor = out->GetMutable(); - size_t offset = GetOffset(scope, dev_ctx); - PADDLE_ENFORCE_LT(offset, x_array.size()); - out_tensor->CopyFrom(x_array[offset], dev_ctx.GetPlace(), dev_ctx); - out_tensor->set_lod(x_array[offset].lod()); + size_t offset = GetOffset(scope, place); + if (offset < x_array.size()) { + auto *out_tensor = out->GetMutable(); + platform::DeviceContextPool &pool = + platform::DeviceContextPool::Instance(); + auto &dev_ctx = *pool.Get(place); + framework::CopyFrom(x_array[offset], place, dev_ctx, out_tensor); + out_tensor->set_lod(x_array[offset].lod()); + } else { + VLOG(10) << "offset " << offset << " >= " << x_array.size(); + } } }; class ReadFromArrayProtoMaker : public framework::OpProtoAndCheckerMaker { public: - ReadFromArrayProtoMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + ReadFromArrayProtoMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "(TensorArray) the array will be read from."); AddInput("I", "(Tensor) the subscript index in tensor array. The number of " "element should be 1"); AddOutput("Out", "(LoDTensor) the tensor will be read from."); - AddComment(R"DOC(Read a LoDTensor from a LoDTensor Array + AddComment(R"DOC( +ReadFromArray Operator. -Assume T is LoDTensor, i is th e subscript of the array, and A is the array. The +Read a LoDTensor from a LoDTensor Array. + +Assume $T$ is LoDTensor, $i$ is the subscript of the array, and $A$ is the array. The equation is -T = A[i] +$$T = A[i]$$ + )DOC"); } }; @@ -156,14 +182,14 @@ class WriteToArrayGradMaker : public framework::SingleGradOpDescMaker { using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; protected: - std::unique_ptr Apply() const override { - auto *grad_op = new framework::OpDescBind(); + std::unique_ptr Apply() const override { + auto *grad_op = new framework::OpDesc(); grad_op->SetType("read_from_array"); grad_op->SetInput("I", Input("I")); grad_op->SetInput("X", OutputGrad("Out")); grad_op->SetOutput("Out", InputGrad("X")); grad_op->SetAttrMap(Attrs()); - return std::unique_ptr(grad_op); + return std::unique_ptr(grad_op); } }; @@ -172,14 +198,14 @@ class ReadFromArrayGradMaker : public framework::SingleGradOpDescMaker { using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; protected: - std::unique_ptr Apply() const override { - auto *grad_op = new framework::OpDescBind(); + std::unique_ptr Apply() const override { + auto *grad_op = new framework::OpDesc(); grad_op->SetType("write_to_array"); grad_op->SetInput("I", Input("I")); grad_op->SetInput("X", OutputGrad("Out")); grad_op->SetOutput("Out", InputGrad("X")); grad_op->SetAttrMap(Attrs()); - return std::unique_ptr(grad_op); + return std::unique_ptr(grad_op); } }; diff --git a/paddle/operators/top_k_op.cc b/paddle/operators/top_k_op.cc index 16ae925eb5cab1c05f3bc376972cabadc4367d20..bb72210bb67f925af3e450961069f0737dbde35e 100644 --- a/paddle/operators/top_k_op.cc +++ b/paddle/operators/top_k_op.cc @@ -46,7 +46,7 @@ class TopkOp : public framework::OperatorWithKernel { class TopkOpMaker : public framework::OpProtoAndCheckerMaker { public: - TopkOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + TopkOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "(Tensor) The input of Topk op"); AddOutput("Out", "(Tensor) The output tensor of Topk op"); diff --git a/paddle/operators/top_k_op.cu b/paddle/operators/top_k_op.cu index 7851c71bbe9fe73402968ce14f6db0df523cd6d3..f7bf58e7218cb9b94526cb9346bc5f9aa971038a 100644 --- a/paddle/operators/top_k_op.cu +++ b/paddle/operators/top_k_op.cu @@ -1,16 +1,16 @@ -/* Copyright (c) 2016 PaddlePaddle Authors All Rights Reserve. +/* 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 +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 + 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. */ +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/framework/op_registry.h" #include "paddle/platform/assert.h" @@ -283,7 +283,7 @@ class TopkOpCUDAKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), - "It must use GPUPlace."); + "It must use CUDAPlace."); auto* input = ctx.Input("X"); auto* output = ctx.Output("Out"); auto* indices = ctx.Output("Indices"); @@ -317,4 +317,4 @@ class TopkOpCUDAKernel : public framework::OpKernel { } // namespace operators } // namespace paddle -REGISTER_OP_GPU_KERNEL(top_k, paddle::operators::TopkOpCUDAKernel); +REGISTER_OP_CUDA_KERNEL(top_k, paddle::operators::TopkOpCUDAKernel); diff --git a/paddle/operators/top_k_op.h b/paddle/operators/top_k_op.h index bc8563717a21bd5b3d8fc87f689657990066957b..e9cd9bbd4d964c28f305fb4ab4c4733ed27ebfff 100644 --- a/paddle/operators/top_k_op.h +++ b/paddle/operators/top_k_op.h @@ -27,7 +27,7 @@ template using EigenMatrix = framework::EigenMatrix; -template +template class TopkKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { diff --git a/paddle/operators/transpose_op.cc b/paddle/operators/transpose_op.cc index 94de3d5069017a7ca818e246ad574c4db92d8006..11615d806a61b3525d2ed50f5ea5940e8d61c8f8 100644 --- a/paddle/operators/transpose_op.cc +++ b/paddle/operators/transpose_op.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/operators/transpose_op.h" @@ -55,8 +55,7 @@ class TransposeOp : public framework::OperatorWithKernel { class TransposeOpMaker : public framework::OpProtoAndCheckerMaker { public: - TransposeOpMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker) + TransposeOpMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput( "X", @@ -71,18 +70,31 @@ class TransposeOpMaker : public framework::OpProtoAndCheckerMaker { Transpose Operator. The input tensor will be permuted according to the axis values given. -The op functions similar to how numpy.transpose works in python. +The op functions is similar to how numpy.transpose works in python. + For example: - >> input = numpy.arange(6).reshape((2,3)) - >> input - array([[0, 1, 2], - [3, 4, 5]]) - >> axis = [1, 0] - >> output = input.transpose(axis) - >> output - array([[0, 3], - [1, 4], - [2, 5]]) + + .. code-block:: text + + input = numpy.arange(6).reshape((2,3)) + + the input is: + + array([[0, 1, 2], + [3, 4, 5]]) + + given axis is: + + [1, 0] + + output = input.transpose(axis) + + then the output is: + + array([[0, 3], + [1, 4], + [2, 5]]) + So, given a input tensor of shape(N, C, H, W) and the axis is {0, 2, 3, 1}, the output tensor shape will be (N, H, W, C) @@ -112,8 +124,8 @@ class TransposeOpGrad : public framework::OperatorWithKernel { namespace ops = paddle::operators; REGISTER_OP(transpose, ops::TransposeOp, ops::TransposeOpMaker, transpose_grad, ops::TransposeOpGrad); -REGISTER_OP_CPU_KERNEL(transpose, - ops::TransposeKernel); +REGISTER_OP_CPU_KERNEL( + transpose, ops::TransposeKernel); REGISTER_OP_CPU_KERNEL( transpose_grad, - ops::TransposeGradKernel); + ops::TransposeGradKernel); diff --git a/paddle/operators/transpose_op.cu.cc b/paddle/operators/transpose_op.cu.cc index af3f581462c919bbd2dd1067e536cc638f9c267d..281c4468cc267c659befd238c9a286dd23eaf16d 100644 --- a/paddle/operators/transpose_op.cu.cc +++ b/paddle/operators/transpose_op.cu.cc @@ -1,22 +1,23 @@ /* 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 +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 + 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. */ +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/operators/transpose_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(transpose, - ops::TransposeKernel); -REGISTER_OP_GPU_KERNEL( +REGISTER_OP_CUDA_KERNEL( + transpose, + ops::TransposeKernel); +REGISTER_OP_CUDA_KERNEL( transpose_grad, - ops::TransposeGradKernel); + ops::TransposeGradKernel); diff --git a/paddle/operators/transpose_op.h b/paddle/operators/transpose_op.h index e296032f4147f9f8338148f9e4fef100c7cf816f..b9686a2db3f76dbb9b8ebdba5e243f5e5a3c571a 100644 --- a/paddle/operators/transpose_op.h +++ b/paddle/operators/transpose_op.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 @@ -20,33 +20,33 @@ namespace paddle { namespace operators { -template -inline void TransCompute(const int dim, const platform::DeviceContext& dev_ctx, +template +inline void TransCompute(const int dim, const DeviceContext& dev_ctx, const framework::Tensor& in, framework::Tensor* out, const std::vector& axis) { switch (dim) { case 1: - math::Transpose trans1; + math::Transpose trans1; trans1(dev_ctx, in, out, axis); break; case 2: - math::Transpose trans2; + math::Transpose trans2; trans2(dev_ctx, in, out, axis); break; case 3: - math::Transpose trans3; + math::Transpose trans3; trans3(dev_ctx, in, out, axis); break; case 4: - math::Transpose trans4; + math::Transpose trans4; trans4(dev_ctx, in, out, axis); break; case 5: - math::Transpose trans5; + math::Transpose trans5; trans5(dev_ctx, in, out, axis); break; case 6: - math::Transpose trans6; + math::Transpose trans6; trans6(dev_ctx, in, out, axis); break; default: @@ -54,7 +54,7 @@ inline void TransCompute(const int dim, const platform::DeviceContext& dev_ctx, } } -template +template class TransposeKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { @@ -64,12 +64,12 @@ class TransposeKernel : public framework::OpKernel { std::vector axis = context.Attr>("axis"); int ndims = axis.size(); - auto& dev_ctx = context.device_context(); - TransCompute(ndims, dev_ctx, *x, out, axis); + auto& dev_ctx = context.template device_context(); + TransCompute(ndims, dev_ctx, *x, out, axis); } }; -template +template class TransposeGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { @@ -88,8 +88,9 @@ class TransposeGradKernel : public framework::OpKernel { } int ndims = axis.size(); - auto& dev_ctx = context.device_context(); - TransCompute(ndims, dev_ctx, *out_grad, x_grad, reversed_axis); + auto& dev_ctx = context.template device_context(); + TransCompute(ndims, dev_ctx, *out_grad, x_grad, + reversed_axis); } }; diff --git a/paddle/operators/uniform_random_op.cc b/paddle/operators/uniform_random_op.cc index 7975efc7cf134aaf591385a6866254a9c5f2a0bb..4d5dd86cb8103a76247913cc088db4cff6b6ff43 100644 --- a/paddle/operators/uniform_random_op.cc +++ b/paddle/operators/uniform_random_op.cc @@ -1,16 +1,16 @@ /* 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. */ -#include -#include +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/framework/op_registry.h" #include "paddle/framework/operator.h" @@ -63,18 +63,17 @@ class UniformRandomOp : public framework::OperatorWithKernel { } protected: - framework::OpKernelType GetKernelType( + framework::OpKernelType GetActualKernelType( const framework::ExecutionContext& ctx) const override { return framework::OpKernelType( - static_cast(ctx.Attr("data_type")), - ctx.device_context()); + static_cast(ctx.Attr("dtype")), + ctx.GetPlace()); } }; class UniformRandomOpMaker : public framework::OpProtoAndCheckerMaker { public: - UniformRandomOpMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker) + UniformRandomOpMaker(OpProto* proto, OpAttrChecker* op_checker) : framework::OpProtoAndCheckerMaker(proto, op_checker) { AddOutput("Out", "(Tensor) The output tensor of uniform random op"); AddComment(R"DOC( @@ -99,8 +98,8 @@ uniform distribution. "Random seed used for generating samples. " "0 means use a seed generated by the system.") .SetDefault(0); - AddAttr("data_type", "(int, default 5(FP32)) Output tensor data type") - .SetDefault(framework::DataType::FP32); + AddAttr("dtype", "(int, default 5(FP32)) Output tensor data type") + .SetDefault(framework::proto::DataType::FP32); } }; } // namespace operators diff --git a/paddle/operators/uniform_random_op.cu b/paddle/operators/uniform_random_op.cu index 8b20bb8287807aca673817c503fee6db04b55753..719d0872a7cba55dc97c95a73a7f86614ab3f4b7 100644 --- a/paddle/operators/uniform_random_op.cu +++ b/paddle/operators/uniform_random_op.cu @@ -1,16 +1,16 @@ /* 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. */ -#include -#include +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 "paddle/framework/op_registry.h" @@ -63,6 +63,6 @@ class GPUUniformRandomKernel : public framework::OpKernel { } // namespace operators } // namespace paddle -REGISTER_OP_GPU_KERNEL(uniform_random, - paddle::operators::GPUUniformRandomKernel, - paddle::operators::GPUUniformRandomKernel); +REGISTER_OP_CUDA_KERNEL(uniform_random, + paddle::operators::GPUUniformRandomKernel, + paddle::operators::GPUUniformRandomKernel); diff --git a/paddle/operators/unpool_op.cc b/paddle/operators/unpool_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..aeed9679b2a3cce744189809c223a5b0d581ccdd --- /dev/null +++ b/paddle/operators/unpool_op.cc @@ -0,0 +1,141 @@ +/* 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. +Indicesou 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/operators/unpool_op.h" +namespace paddle { +namespace operators { + +class Unpool2dOpMaker : public framework::OpProtoAndCheckerMaker { + public: + Unpool2dOpMaker(OpProto* proto, OpAttrChecker* op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput( + "X", + "(Tensor) The input tensor of unpool operator. " + "The format of input tensor is NCHW. Where N is batch size, C is the " + "number of channels, H and W is the height and width of feature."); + AddInput( + "Indices", + "(Tensor) The input tensor of the indices given out by MaxPool2d. " + "The format of input tensor is NCHW. Where N is batch size, C is the " + "number of channels, H and W is the height and width of feature."); + AddOutput("Out", + "(Tensor) The output tensor of unpool operator." + "The format of output tensor is also NCHW." + "Where N is batch size, C is " + "the number of channels, H and W is the height and " + "width of feature."); + AddAttr>( + "ksize", + "(vector), the unpooling window size(height, width) " + "of unpooling operator."); + AddAttr>("strides", + "(vector, default:{1, 1}), " + "strides (height, width) of unpooling operator.") + .SetDefault({1, 1}); + AddAttr>("paddings", + "(vector defalut:{0,0}), " + "paddings (height, width) of unpooling operator.") + .SetDefault({0, 0}); + AddAttr( + "unpooling_type", + "(string), unpooling type, can be \"max\" for max-unpooling ") + .InEnum({"max"}); + AddComment(R"DOC( +Input shape is: $(N, C_{in}, H_{in}, W_{in})$, Output shape is: +$(N, C_{out}, H_{out}, W_{out})$, where +$$ +H_{out} = (H_{in}−1) * strides[0] − 2 * paddings[0] + ksize[0] \\ +W_{out} = (W_{in}−1) * strides[1] − 2 * paddings[1] + ksize[1] +$$ +Paper: http://www.matthewzeiler.com/wp-content/uploads/2017/07/iccv2011.pdf +)DOC"); + } +}; + +int OutputSize(int input_size, int ksize, int padding, int stride) { + int output_size = (input_size - 1) * stride - 2 * padding + ksize; + return output_size; +} + +class UnpoolOp : public framework::OperatorWithKernel { + protected: + framework::OpKernelType GetActualKernelType( + const framework::ExecutionContext& ctx) const override { + return framework::OpKernelType( + framework::ToDataType(ctx.Input("X")->type()), + ctx.device_context()); + } + + public: + using framework::OperatorWithKernel::OperatorWithKernel; + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of UnpoolOp" + "should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Indices"), + "Input(Indices) of UnpoolOp" + "should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of UnpoolOp should not be null."); + auto in_x_dims = ctx->GetInputDim("X"); + auto in_y_dims = ctx->GetInputDim("Indices"); + std::string unpooling_type = + ctx->Attrs().Get("unpooling_type"); + std::vector ksize = ctx->Attrs().Get>("ksize"); + std::vector strides = ctx->Attrs().Get>("strides"); + std::vector paddings = ctx->Attrs().Get>("paddings"); + PADDLE_ENFORCE(in_x_dims.size() == 4, + "Unpooling intput must be of 4-dimensional."); + PADDLE_ENFORCE_EQ(in_x_dims, in_y_dims); + std::vector output_shape({in_x_dims[0], in_x_dims[1]}); + for (size_t i = 0; i < ksize.size(); ++i) { + output_shape.push_back( + OutputSize(in_x_dims[i + 2], ksize[i], paddings[i], strides[i])); + } + ctx->SetOutputDim("Out", framework::make_ddim(output_shape)); + } +}; + +class UnpoolOpGrad : public framework::OperatorWithKernel { + protected: + framework::OpKernelType GetActualKernelType( + const framework::ExecutionContext& ctx) const override { + return framework::OpKernelType( + framework::ToDataType(ctx.Input("X")->type()), + ctx.device_context()); + } + + public: + using framework::OperatorWithKernel::OperatorWithKernel; + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) must not be null."); + PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("X")), + "Input(X@GRAD) should not be null."); + ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X")); + } +}; +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP(unpool, ops::UnpoolOp, ops::Unpool2dOpMaker, unpool_grad, + ops::UnpoolOpGrad); +REGISTER_OP_CPU_KERNEL( + unpool, ops::UnpoolKernel, + ops::UnpoolKernel); +REGISTER_OP_CPU_KERNEL( + unpool_grad, + ops::UnpoolGradKernel, + ops::UnpoolGradKernel); diff --git a/paddle/operators/unpool_op.cu.cc b/paddle/operators/unpool_op.cu.cc new file mode 100644 index 0000000000000000000000000000000000000000..9b002e35c434db561114dbbafce2f3f934daaf6a --- /dev/null +++ b/paddle/operators/unpool_op.cu.cc @@ -0,0 +1,24 @@ +/* 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. +Indicesou 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/operators/unpool_op.h" + +namespace ops = paddle::operators; +REGISTER_OP_CUDA_KERNEL( + unpool, ops::UnpoolKernel, + ops::UnpoolKernel); +REGISTER_OP_CUDA_KERNEL( + unpool_grad, + ops::UnpoolGradKernel, + ops::UnpoolGradKernel); diff --git a/paddle/operators/unpool_op.h b/paddle/operators/unpool_op.h new file mode 100644 index 0000000000000000000000000000000000000000..ee18b118c957c7933890000bbe934e6ffdc9e56f --- /dev/null +++ b/paddle/operators/unpool_op.h @@ -0,0 +1,71 @@ +/* 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. +Indicesou 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/framework/op_registry.h" +#include "paddle/operators/math/math_function.h" +#include "paddle/operators/math/unpooling.h" + +namespace paddle { +namespace operators { +template +class UnpoolKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + const framework::Tensor* in_x = context.Input("X"); + const framework::Tensor* in_y = context.Input("Indices"); + auto* out = context.Output("Out"); + std::string unpooling_type = context.Attr("unpooling_type"); + std::vector ksize = context.Attr>("ksize"); + std::vector strides = context.Attr>("strides"); + std::vector paddings = context.Attr>("paddings"); + T* output_data = out->mutable_data(context.GetPlace()); + auto& dev_ctx = context.template device_context(); + if (output_data) { + math::SetConstant set_zero; + set_zero(dev_ctx, out, static_cast(0)); + } + math::Unpool2dMaxFunctor unpool2d_max_forward; + unpool2d_max_forward(dev_ctx, *in_x, *in_y, out); + } +}; +template +class UnpoolGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + const framework::Tensor* in_x = context.Input("X"); + const framework::Tensor* in_y = context.Input("Indices"); + const framework::Tensor* out = context.Input("Out"); + const framework::Tensor* out_grad = + context.Input(framework::GradVarName("Out")); + framework::Tensor* in_x_grad = + context.Output(framework::GradVarName("X")); + std::string unpooling_type = context.Attr("unpooling_type"); + std::vector ksize = context.Attr>("ksize"); + std::vector strides = context.Attr>("strides"); + std::vector paddings = context.Attr>("paddings"); + + auto& device_ctx = context.template device_context(); + math::SetConstant zero; + if (in_x_grad) { + in_x_grad->mutable_data(context.GetPlace()); + zero(device_ctx, in_x_grad, static_cast(0)); + } + math::Unpool2dMaxGradFunctor unpool2d_max_backward; + unpool2d_max_backward(device_ctx, *in_x, *in_y, *out, *out_grad, in_x_grad); + } +}; +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/while_op.cc b/paddle/operators/while_op.cc index dcc59f5ff2ae3a8ca999d72a20cfd5c759987d89..728ef6079465d57f54dab383aac5e2bb750fe113 100644 --- a/paddle/operators/while_op.cc +++ b/paddle/operators/while_op.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/framework/executor.h" @@ -25,7 +25,7 @@ namespace operators { using StepScopeVar = std::vector; using LoDTensor = framework::LoDTensor; -constexpr char kStepBlock[] = "step_block"; +constexpr char kStepBlock[] = "sub_block"; constexpr char kCondition[] = "Condition"; constexpr char kStepScopes[] = "StepScopes"; constexpr char kParameters[] = "X"; @@ -40,13 +40,14 @@ class WhileOp : public framework::OperatorBase { : framework::OperatorBase(type, inputs, outputs, attrs) {} void Run(const framework::Scope &scope, - const platform::DeviceContext &dev_ctx) const override { + 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_ctx); - auto *block = Attr(kStepBlock); + framework::Executor executor(dev_place); + auto *block = Attr(kStepBlock); + auto *program = block->Program(); auto step_scopes = @@ -64,7 +65,7 @@ class WhileOp : public framework::OperatorBase { class WhileOpMaker : public framework::OpProtoAndCheckerMaker { public: - WhileOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + WhileOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput(kParameters, "A set of variables, which are required by operators inside the " @@ -82,8 +83,8 @@ class WhileOpMaker : public framework::OpProtoAndCheckerMaker { "(StepScopeVar) A vector of local scope, which size equals the " "step number of While Op. The i'th scope storages temporary " "variables generated in the i'th step."); - AddAttr(kStepBlock, - "The step block inside WhileOp"); + AddAttr(kStepBlock, + "The step block inside WhileOp"); AddComment(R"DOC( )DOC"); } @@ -97,11 +98,9 @@ class WhileGradOp : public framework::OperatorBase { : framework::OperatorBase(type, inputs, outputs, attrs) {} void Run(const framework::Scope &scope, - const platform::DeviceContext &dev_ctx) const override { - // PADDLE_ENFORCE(...) - - framework::Executor executor(dev_ctx); - auto *block = Attr(kStepBlock); + const platform::Place &dev_place) const override { + framework::Executor executor(dev_place); + auto *block = Attr(kStepBlock); auto *program = block->Program(); auto *step_scopes = @@ -124,8 +123,12 @@ class WhileGradOp : public framework::OperatorBase { auto inside_og_name = inside_og_names[i]; VLOG(10) << "Linking outside " << outside_og_name << " --> inside " << inside_og_name; - auto &og_outside = detail::Ref(scope.FindVar(outside_og_name)); - auto &og_inside = detail::Ref(cur_scope.Var(inside_og_name)); + auto &og_outside = + detail::Ref(scope.FindVar(outside_og_name), + "Cannot find Outside Gradient %s", outside_og_name); + auto &og_inside = + detail::Ref(cur_scope.Var(inside_og_name), + "Cannot find inside gradient %s", inside_og_name); if (og_outside.Type().hash_code() == typeid(framework::LoDTensor).hash_code()) { auto &outside_tensor = og_outside.Get(); @@ -160,7 +163,7 @@ class WhileGradOp : public framework::OperatorBase { PADDLE_ENFORCE_EQ(pg_names.size(), p_names.size()); for (size_t param_id = 0; param_id < pg_names.size(); ++param_id) { if (pg_names[param_id] == framework::kEmptyVarName) { - continue; // iterator doesn't have gradient + continue; // parameter doesn't have gradient } auto inside_grad_name = framework::GradVarName(p_names[param_id]); @@ -180,22 +183,22 @@ class WhileGradOp : public framework::OperatorBase { if (var->IsType()) { auto &inside_tensor = var->Get(); framework::AttributeMap attrs; - attrs["data_type"] = framework::ToDataType(inside_tensor.type()); + attrs["dtype"] = framework::ToDataType(inside_tensor.type()); attrs["shape"] = framework::vectorize2int(inside_tensor.dims()); attrs["value"] = 0.0f; auto zero_op = framework::OpRegistry::CreateOp( - "fill_constant", {}, {{"Out", {pg_names[param_id]}}}, attrs); - zero_op->Run(scope, dev_ctx); + "fill_constant", framework::VariableNameMap{}, + {{"Out", {pg_names[param_id]}}}, attrs); + zero_op->Run(scope, dev_place); } } - // sum gradient auto new_inside_name = cur_scope.Rename(inside_grad_name); auto sum_op = framework::OpRegistry::CreateOp( "sum", {{"X", {pg_names[param_id], new_inside_name}}}, - {{"Out", {pg_names[param_id]}}}, {}); - sum_op->Run(cur_scope, dev_ctx); + {{"Out", {pg_names[param_id]}}}, framework::AttributeMap{}); + sum_op->Run(cur_scope, dev_place); cur_scope.Rename(new_inside_name, inside_grad_name); } } @@ -207,18 +210,35 @@ class WhileGradOpDescMaker : public framework::SingleGradOpDescMaker { using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; protected: - virtual std::unique_ptr Apply() const { - auto *grad = new framework::OpDescBind(); + std::unique_ptr Apply() const override { + auto *grad = new framework::OpDesc(); grad->SetType("while_grad"); grad->SetInput(kParameters, Input(kParameters)); - grad->SetOutput( - framework::GradVarName(kParameters), - InputGrad(kParameters, /*do not drop empty gradient*/ false)); + + // Not all of IGs will be generated by inner gradient operators of while op. + // Ignore IGs that is not generated by the inside block. + auto igs = InputGrad(kParameters, /*do not drop empty gradient*/ false); + std::unordered_set all_outs; + for (size_t i = 0; i < grad_block_[0]->OpSize(); ++i) { + for (auto &oname : grad_block_[0]->Op(i)->OutputArgumentNames()) { + all_outs.insert(oname); + } + } + for (auto &each_ig : igs) { + if (all_outs.find(each_ig) == all_outs.end()) { + VLOG(10) << "Ignore " << each_ig; + each_ig = framework::kEmptyVarName; + } + } + + grad->SetOutput(framework::GradVarName(kParameters), igs); + grad->SetInput(kOutputs, Output(kOutputs)); // OG should be re-calculated by step blocks, since many outputs of while op // do not need to calculate gradients. std::unordered_set block_ins; + auto *fwd_block = this->grad_block_[0]->ParentBlock(); { for (auto &p : Input(kParameters)) { block_ins.insert(p); @@ -233,6 +253,13 @@ class WhileGradOpDescMaker : public framework::SingleGradOpDescMaker { if (block_ins.find(input_name) != block_ins.end()) { continue; } + + // If the input of Op is generated by the forward block, do not make it + // as input again. + if (fwd_block->FindVar(input_name) != nullptr) { + continue; + } + extra_inputs.insert(input_name); } @@ -253,14 +280,14 @@ class WhileGradOpDescMaker : public framework::SingleGradOpDescMaker { // while operator could be renamed. grad->SetAttr("original_output_grad", extra_inputs_list); - return std::unique_ptr(grad); + return std::unique_ptr(grad); } }; class WhileGradOpVarTypeInference : public framework::VarTypeInference { public: - void operator()(const framework::OpDescBind &op_desc, - framework::BlockDescBind *block) const override { + void operator()(const framework::OpDesc &op_desc, + framework::BlockDesc *block) const override { auto p_names = op_desc.Input(kParameters); auto pg_names = op_desc.Output(framework::GradVarName(kParameters)); @@ -287,7 +314,6 @@ class WhileGradOpShapeInference : public framework::InferShapeBase { auto p_names = ctx->Inputs(kParameters); auto pg_names = ctx->Outputs(kParamGrads); - auto dims = ctx->GetInputsDim(kParameters); auto var_types = ctx->GetInputsVarType(kParameters); std::vector names_to_set; std::vector dims_to_set; @@ -295,13 +321,14 @@ class WhileGradOpShapeInference : public framework::InferShapeBase { if (pg_names[i] == framework::kEmptyVarName) { continue; } - if (var_types[i] == framework::VarDesc::LOD_TENSOR) { + auto dims = ctx->GetInputsElementDim(kParameters, i); + if (var_types[i] == framework::proto::VarDesc::LOD_TENSOR) { names_to_set.push_back(pg_names[i]); - dims_to_set.push_back(dims[i]); - } else if (var_types[i] == framework::VarDesc::LOD_TENSOR_ARRAY) { + dims_to_set.push_back(dims); + } else if (var_types[i] == framework::proto::VarDesc::LOD_TENSOR_ARRAY) { // not sure how to set the dim of LOD_TENSOR_ARRAY names_to_set.push_back(pg_names[i]); - dims_to_set.push_back(dims[i]); + dims_to_set.push_back(dims); } } ctx->SetDims(names_to_set, dims_to_set); diff --git a/paddle/optimizer/adadelta_optimizer.cc b/paddle/optimizer/adadelta_optimizer.cc index 5cc7c47d4486c3d149c37fd6e312780f3d44eda8..8ca048257e53af21f327ee1b8f24b9e2e5d54060 100644 --- a/paddle/optimizer/adadelta_optimizer.cc +++ b/paddle/optimizer/adadelta_optimizer.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 "adadelta_optimizer.h" #include diff --git a/paddle/optimizer/adadelta_optimizer.h b/paddle/optimizer/adadelta_optimizer.h index 6aab1ad553b15ebbd2d04c9323c5e56e1b8f60f5..48f1ae175041cc1cfa11787c44317dd173430928 100644 --- a/paddle/optimizer/adadelta_optimizer.h +++ b/paddle/optimizer/adadelta_optimizer.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 diff --git a/paddle/optimizer/adagrad_optimizer.cc b/paddle/optimizer/adagrad_optimizer.cc index c981996bab1b2e7ae5d6e2d858a73efde12e32f3..c6d39a366adcd40e7298e1f1bed48c15f517a217 100644 --- a/paddle/optimizer/adagrad_optimizer.cc +++ b/paddle/optimizer/adagrad_optimizer.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 diff --git a/paddle/optimizer/adagrad_optimizer.h b/paddle/optimizer/adagrad_optimizer.h index 447b7c7547d5bad7436df6f3b3582b4a219f08c8..b0cff061f5c311d2356bab4965ec65c5eca3dbb8 100644 --- a/paddle/optimizer/adagrad_optimizer.h +++ b/paddle/optimizer/adagrad_optimizer.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 diff --git a/paddle/optimizer/adam_optimizer.cc b/paddle/optimizer/adam_optimizer.cc index 6dc2d749708d0e2a7f36734d89eec30d4576842e..8a384b59c47df321438e3a3bc245a2398a7ef6d1 100644 --- a/paddle/optimizer/adam_optimizer.cc +++ b/paddle/optimizer/adam_optimizer.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 "adam_optimizer.h" #include diff --git a/paddle/optimizer/adam_optimizer.h b/paddle/optimizer/adam_optimizer.h index 37ab53afc37a5f749a2909de12c7871ed926583f..7df40064df31536cead9eba4ce827452cf0dce38 100644 --- a/paddle/optimizer/adam_optimizer.h +++ b/paddle/optimizer/adam_optimizer.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 diff --git a/paddle/optimizer/optimizer.cc b/paddle/optimizer/optimizer.cc index faa23764522cef03bae1359adbf58d10ee7809ac..3af44484363654befe85c8511d7c0791d5a33ab8 100644 --- a/paddle/optimizer/optimizer.cc +++ b/paddle/optimizer/optimizer.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 "optimizer.h" #include diff --git a/paddle/optimizer/optimizer.h b/paddle/optimizer/optimizer.h index e6fa12a4d250ccb078358704b0131942ea6ab039..516e612167f557f043d3fe29789bec4b73405755 100644 --- a/paddle/optimizer/optimizer.h +++ b/paddle/optimizer/optimizer.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 diff --git a/paddle/optimizer/parameter_optimizer.cc b/paddle/optimizer/parameter_optimizer.cc index da92c2d01cc2a27d1fadd51a338d23b01e0cb0bc..1603e5fdc8a69d8b117f02b9c950a5374625c632 100644 --- a/paddle/optimizer/parameter_optimizer.cc +++ b/paddle/optimizer/parameter_optimizer.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 "adadelta_optimizer.h" diff --git a/paddle/optimizer/parameter_optimizer.h b/paddle/optimizer/parameter_optimizer.h index 99d0416e751c4ca6695d6ed77396e18d48fc86b8..1f501c49e1d7a8018f3274ae1128ca230b9298bb 100644 --- a/paddle/optimizer/parameter_optimizer.h +++ b/paddle/optimizer/parameter_optimizer.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 diff --git a/paddle/optimizer/parameter_optimizer_test.cc b/paddle/optimizer/parameter_optimizer_test.cc index f29e5317120642e3790a6f6c1976bdda67093a0c..83757a391784453341f22eca73bc73c14ce4174f 100644 --- a/paddle/optimizer/parameter_optimizer_test.cc +++ b/paddle/optimizer/parameter_optimizer_test.cc @@ -127,8 +127,3 @@ TEST_F(OptimizerTest, TestGetWeight) { TestGetWeight(); } TEST_F(OptimizerTest, TestUpdate) { TestUpdate(); } TEST_F(OptimizerTest, TestCheckPoint) { TestCheckPoint(); } - -int main(int argc, char** argv) { - testing::InitGoogleTest(&argc, argv); - return RUN_ALL_TESTS(); -} diff --git a/paddle/optimizer/serialization_test.cc b/paddle/optimizer/serialization_test.cc index 4c416f55ee0bd70f9ec6e288b08a5399d8b2bf39..940e941e9042d8a37363311867df5bb477b3dac0 100644 --- a/paddle/optimizer/serialization_test.cc +++ b/paddle/optimizer/serialization_test.cc @@ -46,8 +46,3 @@ TEST(TensorToProto, Case2) { EXPECT_EQ(t1[i], t[i]); } } - -int main(int argc, char** argv) { - testing::InitGoogleTest(&argc, argv); - return RUN_ALL_TESTS(); -} diff --git a/paddle/optimizer/sgd_optimizer.cc b/paddle/optimizer/sgd_optimizer.cc index c150144ac24b8375d08691a98be680b6bf5d1e7f..ee80f543fc4c1d87ba56540fc581816df8a39793 100644 --- a/paddle/optimizer/sgd_optimizer.cc +++ b/paddle/optimizer/sgd_optimizer.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 "sgd_optimizer.h" #include "serialization.h" diff --git a/paddle/optimizer/sgd_optimizer.h b/paddle/optimizer/sgd_optimizer.h index 0b1da0aa27d98e8d6a8d9fd7a1ebe355acb2a1f4..16a4df9973ef6a599a8ecee6be5903f13799771c 100644 --- a/paddle/optimizer/sgd_optimizer.h +++ b/paddle/optimizer/sgd_optimizer.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 diff --git a/paddle/parameter/FirstOrderOptimizer.h b/paddle/parameter/FirstOrderOptimizer.h index f157188a4f736319ea187052b90a17f8be9e9edb..5b0c52a30dfbc34f0db4a03366e31da54b86c9fc 100644 --- a/paddle/parameter/FirstOrderOptimizer.h +++ b/paddle/parameter/FirstOrderOptimizer.h @@ -38,7 +38,7 @@ public: real torch_learningRate = optConfig_.learning_method() == "torch_momentum" ? 1.0 - paraConfig.momentum() : 1.0; -#ifdef PADDLE_USE_MKLDNN +#ifdef PADDLE_WITH_MKLDNN sgdUpdate(learningRate_ * paraConfig.learning_rate() * (firstTime_ ? 1.0 : torch_learningRate), paraConfig.momentum(), diff --git a/paddle/parameter/ParameterUpdateFunctions.cpp b/paddle/parameter/ParameterUpdateFunctions.cpp index 1898598e49652a2829e57329bab6017304cec662..d60cb363830ff26a1f5054fb4cebf37afdfd1d03 100644 --- a/paddle/parameter/ParameterUpdateFunctions.cpp +++ b/paddle/parameter/ParameterUpdateFunctions.cpp @@ -30,7 +30,7 @@ void sgdUpdateCpu(real learningRate, const real* grad, real* momentumVec) { decayRate *= learningRate; -#ifdef PADDLE_USE_MKLML +#ifdef PADDLE_WITH_MKLML #pragma omp parallel for #endif for (size_t i = 0; i < size; ++i) { diff --git a/paddle/platform/CMakeLists.txt b/paddle/platform/CMakeLists.txt index bd86a9fe268c277065cd450f91b544def6c4d32f..f0a0ea70a0aa14e1db959e4e6ace2a44363d0c35 100644 --- a/paddle/platform/CMakeLists.txt +++ b/paddle/platform/CMakeLists.txt @@ -1,15 +1,20 @@ -cc_library(cpu_info SRCS cpu_info.cc DEPS gflags glog) +if(WITH_GPU) + cc_library(enforce SRCS enforce.cc DEPS nccl) +else() + cc_library(enforce SRCS enforce.cc) +endif() +cc_test(enforce_test SRCS enforce_test.cc DEPS stringpiece enforce) + +cc_library(cpu_info SRCS cpu_info.cc DEPS gflags glog enforce) cc_test(cpu_info_test SRCS cpu_info_test.cc DEPS cpu_info) -nv_library(gpu_info SRCS gpu_info.cc DEPS gflags glog) +nv_library(gpu_info SRCS gpu_info.cc DEPS gflags glog enforce) -cc_library(place SRCS place.cc) +cc_library(place SRCS place.cc DEPS enforce) cc_test(place_test SRCS place_test.cc DEPS place glog gflags) add_subdirectory(dynload) -cc_test(enforce_test SRCS enforce_test.cc DEPS stringpiece) - IF(WITH_GPU) set(GPU_CTX_DEPS dynload_cuda dynamic_loader) ELSE() @@ -20,7 +25,7 @@ ENDIF() # avoiding cycle dependencies cc_library(device_context SRCS device_context.cc DEPS memory buddy_allocator system_allocator memory_block meta_data meta_cache place eigen3 ${GPU_CTX_DEPS}) -nv_test(device_context_test SRCS device_context_test.cc DEPS device_context gpu_info) +nv_test(device_context_test SRCS device_context_test.cu DEPS device_context gpu_info) nv_test(cudnn_helper_test SRCS cudnn_helper_test.cc DEPS dynload_cuda) nv_test(transform_test SRCS transform_test.cu DEPS paddle_memory place device_context) diff --git a/paddle/platform/call_once.h b/paddle/platform/call_once.h index d9f49527dcf150fcb35d3af512088f75dec0b5c6..00337a7f051758559a0f8012d8c78dbe8e3457a6 100644 --- a/paddle/platform/call_once.h +++ b/paddle/platform/call_once.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 diff --git a/paddle/platform/cuda_helper.h b/paddle/platform/cuda_helper.h index a7d99cde106a0a66f122a8c43f49717c03e60dec..376bb0e6887c797c3c1019e92f738a62d01a9c51 100644 --- a/paddle/platform/cuda_helper.h +++ b/paddle/platform/cuda_helper.h @@ -31,6 +31,16 @@ constexpr int PADDLE_CUDA_NUM_THREADS = 512; // For atomicAdd. USE_CUDA_ATOMIC(Add, float); +USE_CUDA_ATOMIC(Add, int); +USE_CUDA_ATOMIC(Add, unsigned int); +USE_CUDA_ATOMIC(Add, unsigned long long int); + +CUDA_ATOMIC_WRAPPER(Add, int64_t) { + static_assert(sizeof(int64_t) == sizeof(long long int), + "long long should be int64"); + return CudaAtomicAdd(reinterpret_cast(address), + static_cast(val)); +} #if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 600 USE_CUDA_ATOMIC(Add, double); diff --git a/paddle/platform/cuda_profiler.h b/paddle/platform/cuda_profiler.h new file mode 100644 index 0000000000000000000000000000000000000000..67d5f626d41c0fae280062533d6c1d1074341164 --- /dev/null +++ b/paddle/platform/cuda_profiler.h @@ -0,0 +1,37 @@ +/* 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 +#include +#include +#include +#include + +namespace paddle { +namespace platform { + +void CudaProfilerInit(std::string output_file, std::string output_mode, + std::string config_file) { + PADDLE_ENFORCE(output_mode == "kvp" || output_mode == "csv"); + cudaOutputMode_t mode = output_mode == "csv" ? cudaCSV : cudaKeyValuePair; + PADDLE_ENFORCE( + cudaProfilerInitialize(config_file.c_str(), output_file.c_str(), mode)); +} + +void CudaProfilerStart() { PADDLE_ENFORCE(cudaProfilerStart()); } + +void CudaProfilerStop() { PADDLE_ENFORCE(cudaProfilerStop()); } + +} // namespace platform +} // namespace paddle diff --git a/paddle/platform/cudnn_helper.h b/paddle/platform/cudnn_helper.h index dd48605b9ed688e4656d4cd1ddf1f298d0a50a9e..80a4c9bb4bbcd03cf849d86118db4e502382f031 100644 --- a/paddle/platform/cudnn_helper.h +++ b/paddle/platform/cudnn_helper.h @@ -116,7 +116,7 @@ inline cudnnTensorFormat_t GetCudnnTensorFormat( case DataLayout::kNCHW: return CUDNN_TENSOR_NCHW; case DataLayout::kNCDHW: - return CUDNN_TENSOR_NCHW; // TODO(chengduoZH) : add CUDNN_TENSOR_NCDHW + return CUDNN_TENSOR_NCHW; // NOTE: cudnn treat NdTensor as the same default: PADDLE_THROW("Unknown cudnn equivalent for order"); } @@ -143,7 +143,7 @@ class ScopedTensorDescriptor { strides[i] = dims[i + 1] * strides[i + 1]; } // Update tensor descriptor dims setting if groups > 1 - // FIXME(typhoonzero): Assume using NCHW or NCDHW order + // NOTE: Assume using NCHW or NCDHW order std::vector dims_with_group(dims.begin(), dims.end()); // copy if (groups > 1) { dims_with_group[1] = dims_with_group[1] / groups; @@ -186,7 +186,6 @@ class ScopedFilterDescriptor { // width of the filter. std::vector kernel_with_group(kernel.begin(), kernel.end()); if (groups > 1) { - // M /= groups kernel_with_group[0] /= groups; // NOTE: input filter(C) of the filter is already asserted to be C/groups. } @@ -224,13 +223,15 @@ class ScopedConvolutionDescriptor { PADDLE_ENFORCE_EQ(pads.size(), strides.size()); PADDLE_ENFORCE_EQ(pads.size(), dilations.size()); -#if CUDNN_VERSION < 6000 +#if !CUDNN_VERSION_MIN(6, 0, 0) // cudnn v5 does not support dilation conv, the argument is called upscale // instead of dilations and it is must be one. for (size_t i = 0; i < dilations.size(); ++i) { PADDLE_ENFORCE_EQ( dilations[i], 1, - "Dilations conv is not supported in this cuDNN version"); + "Dilations conv is not supported in this cuDNN version(%d.%d.%d).", + CUDNN_VERSION / 1000, CUDNN_VERSION % 1000 / 100, + CUDNN_VERSION % 100); } #endif diff --git a/paddle/platform/cudnn_helper_test.cc b/paddle/platform/cudnn_helper_test.cc index 6bd85ae1ca8b47b203e0321e9d9224d5cfd3a586..427359f69713b961c4730b697d3ccde5f7085838 100644 --- a/paddle/platform/cudnn_helper_test.cc +++ b/paddle/platform/cudnn_helper_test.cc @@ -38,6 +38,26 @@ TEST(CudnnHelper, ScopedTensorDescriptor) { EXPECT_EQ(strides[2], 6); EXPECT_EQ(strides[1], 36); EXPECT_EQ(strides[0], 144); + + // test tensor5d: ScopedTensorDescriptor + ScopedTensorDescriptor tensor5d_desc; + std::vector shape_5d = {2, 4, 6, 6, 6}; + auto desc_5d = tensor5d_desc.descriptor(DataLayout::kNCDHW, shape_5d); + + std::vector dims_5d(5); + std::vector strides_5d(5); + paddle::platform::dynload::cudnnGetTensorNdDescriptor( + desc_5d, 5, &type, &nd, dims_5d.data(), strides_5d.data()); + + EXPECT_EQ(nd, 5); + for (size_t i = 0; i < dims_5d.size(); ++i) { + EXPECT_EQ(dims_5d[i], shape_5d[i]); + } + EXPECT_EQ(strides_5d[4], 1); + EXPECT_EQ(strides_5d[3], 6); + EXPECT_EQ(strides_5d[2], 36); + EXPECT_EQ(strides_5d[1], 216); + EXPECT_EQ(strides_5d[0], 864); } TEST(CudnnHelper, ScopedFilterDescriptor) { @@ -60,6 +80,20 @@ TEST(CudnnHelper, ScopedFilterDescriptor) { for (size_t i = 0; i < shape.size(); ++i) { EXPECT_EQ(kernel[i], shape[i]); } + + ScopedFilterDescriptor filter_desc_4d; + std::vector shape_4d = {2, 3, 3, 3}; + auto desc_4d = filter_desc.descriptor(DataLayout::kNCDHW, shape_4d); + + std::vector kernel_4d(4); + paddle::platform::dynload::cudnnGetFilterNdDescriptor( + desc_4d, 4, &type, &format, &nd, kernel_4d.data()); + + EXPECT_EQ(GetCudnnTensorFormat(DataLayout::kNCHW), format); + EXPECT_EQ(nd, 4); + for (size_t i = 0; i < shape_4d.size(); ++i) { + EXPECT_EQ(kernel_4d[i], shape_4d[i]); + } } TEST(CudnnHelper, ScopedConvolutionDescriptor) { diff --git a/paddle/platform/device_context.cc b/paddle/platform/device_context.cc index 7afcdfce9371e29aad968a1729931173fb2309b5..ea07f2e002cb76d09a11f7a5305c2d45b780e7bd 100644 --- a/paddle/platform/device_context.cc +++ b/paddle/platform/device_context.cc @@ -15,17 +15,46 @@ limitations under the License. */ namespace paddle { namespace platform { -template <> -Eigen::DefaultDevice* DeviceContext::GetEigenDevice< - platform::CPUPlace, Eigen::DefaultDevice>() const { - return reinterpret_cast(this)->eigen_device(); +DeviceContextPool* DeviceContextPool::pool = nullptr; + +const platform::DeviceContext* DeviceContextPool::Get( + const platform::Place& place) { + auto it = device_contexts_.find(place); + if (it == device_contexts_.end()) { + PADDLE_THROW( + "'Place' is not supported, Please re-compile with WITH_GPU " + "option"); + } + return it->second; +} + +DeviceContextPool::DeviceContextPool( + const std::vector& places) { + PADDLE_ENFORCE_GT(places.size(), 0); + for (size_t i = 0; i < places.size(); i++) { + if (platform::is_cpu_place(places[i])) { + device_contexts_.emplace(places[i], + new platform::CPUDeviceContext( + boost::get(places[i]))); + } else if (platform::is_gpu_place(places[i])) { +#ifdef PADDLE_WITH_CUDA + device_contexts_.emplace(places[i], + new platform::CUDADeviceContext( + boost::get(places[i]))); +#else + PADDLE_THROW( + "'CUDAPlace' is not supported, Please re-compile with WITH_GPU " + "option"); +#endif + } + } } CPUDeviceContext::CPUDeviceContext() { eigen_device_.reset(new Eigen::DefaultDevice()); } -CPUDeviceContext::CPUDeviceContext(CPUPlace place) { +CPUDeviceContext::CPUDeviceContext(CPUPlace place) : place_(place) { eigen_device_.reset(new Eigen::DefaultDevice()); } @@ -33,16 +62,10 @@ Eigen::DefaultDevice* CPUDeviceContext::eigen_device() const { return eigen_device_.get(); } -Place CPUDeviceContext::GetPlace() const { return CPUPlace(); } +Place CPUDeviceContext::GetPlace() const { return place_; } #ifdef PADDLE_WITH_CUDA -template <> -Eigen::GpuDevice* -DeviceContext::GetEigenDevice() const { - return reinterpret_cast(this)->eigen_device(); -} - class EigenCudaStreamDevice : public Eigen::StreamInterface { public: EigenCudaStreamDevice() : scratch_(nullptr), semaphore_(nullptr) { @@ -50,7 +73,7 @@ class EigenCudaStreamDevice : public Eigen::StreamInterface { } ~EigenCudaStreamDevice() override {} - void Reinitialize(const cudaStream_t* cuda_stream, GPUPlace place) { + void Reinitialize(const cudaStream_t* cuda_stream, CUDAPlace place) { stream_ = cuda_stream; place_ = place; device_prop_ = &Eigen::m_deviceProperties[place.device]; @@ -89,14 +112,14 @@ class EigenCudaStreamDevice : public Eigen::StreamInterface { } private: - GPUPlace place_; + CUDAPlace place_; const cudaStream_t* stream_; // not owned; const cudaDeviceProp* device_prop_; // not owned; mutable void* scratch_; mutable unsigned int* semaphore_; }; -CUDADeviceContext::CUDADeviceContext(GPUPlace place) : place_(place) { +CUDADeviceContext::CUDADeviceContext(CUDAPlace place) : place_(place) { SetDeviceId(place_.device); PADDLE_ENFORCE(cudaStreamCreate(&stream_)); eigen_stream_.reset(new EigenCudaStreamDevice()); @@ -122,10 +145,6 @@ Place CUDADeviceContext::GetPlace() const { return place_; } void CUDADeviceContext::Wait() const { PADDLE_ENFORCE(cudaStreamSynchronize(stream_)); -} - -void CUDADeviceContext::Finish() const { - Wait(); PADDLE_ENFORCE(cudaGetLastError()); } @@ -141,6 +160,20 @@ cudnnHandle_t CUDADeviceContext::cudnn_handle() const { return cudnn_handle_; } cudaStream_t CUDADeviceContext::stream() const { return stream_; } +CUDNNDeviceContext::CUDNNDeviceContext(CUDAPlace place) + : CUDADeviceContext(place) { + PADDLE_ENFORCE(dynload::cudnnCreate(&cudnn_handle_)); + PADDLE_ENFORCE(dynload::cudnnSetStream(cudnn_handle_, stream())); +} + +CUDNNDeviceContext::~CUDNNDeviceContext() { + SetDeviceId(boost::get(GetPlace()).device); + Wait(); + PADDLE_ENFORCE(dynload::cudnnDestroy(cudnn_handle_)); +} + +cudnnHandle_t CUDNNDeviceContext::cudnn_handle() const { return cudnn_handle_; } + #endif } // namespace platform diff --git a/paddle/platform/device_context.h b/paddle/platform/device_context.h index 526d089e35da9c9f89a3852095ad3a4c82d4d85d..2b366e6383d23e2d31a194edd04412892a8311eb 100644 --- a/paddle/platform/device_context.h +++ b/paddle/platform/device_context.h @@ -11,8 +11,8 @@ limitations under the License. */ #pragma once -#include "paddle/platform/enforce.h" -#include "paddle/platform/place.h" +#include +#include #ifdef PADDLE_WITH_CUDA #include "paddle/platform/dynload/cublas.h" @@ -20,34 +20,22 @@ limitations under the License. */ #include "paddle/platform/gpu_info.h" #define EIGEN_USE_GPU #endif -#include + +#include "paddle/platform/enforce.h" #include "paddle/platform/place.h" #include "unsupported/Eigen/CXX11/Tensor" +#include "glog/logging.h" + namespace paddle { namespace platform { -template -struct EigenDeviceConverter; - -template <> -struct EigenDeviceConverter { - using EigenDeviceType = Eigen::DefaultDevice; -}; - class DeviceContext { public: virtual ~DeviceContext() {} virtual Place GetPlace() const = 0; - template ::EigenDeviceType> - DeviceType* GetEigenDevice() const; - virtual void Wait() const {} - - virtual void Finish() const {} }; class CPUDeviceContext : public DeviceContext { @@ -60,28 +48,30 @@ class CPUDeviceContext : public DeviceContext { Place GetPlace() const override; private: + CPUPlace place_; std::unique_ptr eigen_device_; }; -#ifdef PADDLE_WITH_CUDA +template +struct DefaultDeviceContextType; + template <> -struct EigenDeviceConverter { - using EigenDeviceType = Eigen::GpuDevice; +struct DefaultDeviceContextType { + using TYPE = CPUDeviceContext; }; +#ifdef PADDLE_WITH_CUDA + class EigenCudaStreamDevice; class CUDADeviceContext : public DeviceContext { public: - explicit CUDADeviceContext(GPUPlace place); + explicit CUDADeviceContext(CUDAPlace place); virtual ~CUDADeviceContext(); /*! \brief Wait for all operations completion in the stream. */ void Wait() const override; - /*! \brief Check potential errors for the cuda kernel calls. */ - void Finish() const override; - /*! \brief Return place in the device context. */ Place GetPlace() const override; @@ -98,7 +88,7 @@ class CUDADeviceContext : public DeviceContext { cudaStream_t stream() const; private: - GPUPlace place_; + CUDAPlace place_; std::unique_ptr eigen_device_; std::unique_ptr eigen_stream_; @@ -108,7 +98,71 @@ class CUDADeviceContext : public DeviceContext { cublasHandle_t cublas_handle_; }; +template <> +struct DefaultDeviceContextType { + using TYPE = CUDADeviceContext; +}; + +class CUDNNDeviceContext : public CUDADeviceContext { + public: + explicit CUDNNDeviceContext(CUDAPlace place); + virtual ~CUDNNDeviceContext(); + + /*! \brief Return cudnn handle in the device context. */ + cudnnHandle_t cudnn_handle() const; + + private: + cudnnHandle_t cudnn_handle_; +}; + #endif +/*! \brief device context pool singleton */ +class DeviceContextPool { + public: + explicit DeviceContextPool(const std::vector& places); + + static DeviceContextPool& Instance() { + PADDLE_ENFORCE_NOT_NULL(pool, "Need to Create DeviceContextPool first!"); + return *pool; + } + + /*! \brief Create should only called by Init function */ + static DeviceContextPool& Init(const std::vector& places) { + if (pool == nullptr) { + pool = new DeviceContextPool(places); + } + return *pool; + } + + /*! \brief Return handle of single device context. */ + const platform::DeviceContext* Get(const platform::Place& place); + + template + const typename DefaultDeviceContextType::TYPE* GetByPlace( + const Place& place) { + return reinterpret_cast< + const typename DefaultDeviceContextType::TYPE*>(Get(place)); + } + + private: + static DeviceContextPool* pool; + constexpr static int LEFT_SHIFT = 8; + struct Hash { + std::hash hash_; + size_t operator()(const platform::Place& place) const { + int pre_hash = place.which() + (1 << LEFT_SHIFT); + if (platform::is_gpu_place(place)) { + pre_hash += boost::get(place).GetDeviceId(); + } + return hash_(pre_hash); + } + }; + std::unordered_map + device_contexts_; + DISABLE_COPY_AND_ASSIGN(DeviceContextPool); +}; + } // namespace platform } // namespace paddle diff --git a/paddle/platform/device_context_test.cc b/paddle/platform/device_context_test.cc deleted file mode 100644 index 8bf5174c4a5579f6f5602dd38e5a87ed3ef444a7..0000000000000000000000000000000000000000 --- a/paddle/platform/device_context_test.cc +++ /dev/null @@ -1,49 +0,0 @@ -/* 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. */ - -#include "paddle/platform/device_context.h" -#include "gtest/gtest.h" - -TEST(Device, Init) { - using paddle::platform::DeviceContext; - using paddle::platform::CUDADeviceContext; - using paddle::platform::GPUPlace; - - int count = paddle::platform::GetCUDADeviceCount(); - for (int i = 0; i < count; i++) { - DeviceContext* device_context = new CUDADeviceContext(GPUPlace(i)); - Eigen::GpuDevice* gpu_device = - device_context->template GetEigenDevice(); - ASSERT_NE(nullptr, gpu_device); - delete device_context; - } -} - -TEST(Device, CUDADeviceContext) { - using paddle::platform::CUDADeviceContext; - using paddle::platform::GPUPlace; - - int count = paddle::platform::GetCUDADeviceCount(); - for (int i = 0; i < count; i++) { - CUDADeviceContext* device_context = new CUDADeviceContext(GPUPlace(i)); - Eigen::GpuDevice* gpu_device = device_context->eigen_device(); - ASSERT_NE(nullptr, gpu_device); - cudnnHandle_t cudnn_handle = device_context->cudnn_handle(); - ASSERT_NE(nullptr, cudnn_handle); - cublasHandle_t cublas_handle = device_context->cublas_handle(); - ASSERT_NE(nullptr, cublas_handle); - ASSERT_NE(nullptr, device_context->stream()); - delete device_context; - } -} diff --git a/paddle/platform/device_context_test.cu b/paddle/platform/device_context_test.cu new file mode 100644 index 0000000000000000000000000000000000000000..ca10cf34639376798bf5ba05970c9c734e5a1ef8 --- /dev/null +++ b/paddle/platform/device_context_test.cu @@ -0,0 +1,101 @@ +/* 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. */ + +#include "gtest/gtest.h" +#include "paddle/platform/device_context.h" + +#include "glog/logging.h" + +TEST(Device, Init) { + using paddle::platform::DeviceContext; + using paddle::platform::CUDADeviceContext; + using paddle::platform::CUDAPlace; + + int count = paddle::platform::GetCUDADeviceCount(); + for (int i = 0; i < count; i++) { + CUDADeviceContext* device_context = new CUDADeviceContext(CUDAPlace(i)); + Eigen::GpuDevice* gpu_device = device_context->eigen_device(); + ASSERT_NE(nullptr, gpu_device); + delete device_context; + } +} + +TEST(Device, CUDADeviceContext) { + using paddle::platform::CUDADeviceContext; + using paddle::platform::CUDAPlace; + + int count = paddle::platform::GetCUDADeviceCount(); + for (int i = 0; i < count; i++) { + CUDADeviceContext* device_context = new CUDADeviceContext(CUDAPlace(i)); + Eigen::GpuDevice* gpu_device = device_context->eigen_device(); + ASSERT_NE(nullptr, gpu_device); + cudnnHandle_t cudnn_handle = device_context->cudnn_handle(); + ASSERT_NE(nullptr, cudnn_handle); + cublasHandle_t cublas_handle = device_context->cublas_handle(); + ASSERT_NE(nullptr, cublas_handle); + ASSERT_NE(nullptr, device_context->stream()); + delete device_context; + } +} + +TEST(Device, CUDNNDeviceContext) { + using paddle::platform::CUDNNDeviceContext; + using paddle::platform::CUDAPlace; + if (paddle::platform::dynload::HasCUDNN()) { + int count = paddle::platform::GetCUDADeviceCount(); + for (int i = 0; i < count; ++i) { + CUDNNDeviceContext* device_context = new CUDNNDeviceContext(CUDAPlace(i)); + cudnnHandle_t cudnn_handle = device_context->cudnn_handle(); + ASSERT_NE(nullptr, cudnn_handle); + ASSERT_NE(nullptr, device_context->stream()); + delete device_context; + } + } +} + +TEST(Device, DeviceContextPool) { + using paddle::platform::DeviceContextPool; + using paddle::platform::CUDADeviceContext; + using paddle::platform::Place; + using paddle::platform::CPUPlace; + using paddle::platform::CUDAPlace; + + DeviceContextPool& pool = DeviceContextPool::Instance(); + auto cpu_dev_ctx1 = pool.Get(CPUPlace()); + auto cpu_dev_ctx2 = pool.Get(CPUPlace()); + ASSERT_EQ(cpu_dev_ctx2, cpu_dev_ctx1); + + std::vector gpu_places; + int count = paddle::platform::GetCUDADeviceCount(); + for (int i = 0; i < count; ++i) { + auto dev_ctx = pool.Get(CUDAPlace(i)); + ASSERT_NE(dev_ctx, nullptr); + } +} + +int main(int argc, char** argv) { + std::vector places; + + places.emplace_back(paddle::platform::CPUPlace()); + int count = paddle::platform::GetCUDADeviceCount(); + for (int i = 0; i < count; ++i) { + places.emplace_back(paddle::platform::CUDAPlace(i)); + } + + VLOG(0) << " DeviceCount " << count; + paddle::platform::DeviceContextPool::Init(places); + + testing::InitGoogleTest(&argc, argv); + return RUN_ALL_TESTS(); +} diff --git a/paddle/platform/dynload/CMakeLists.txt b/paddle/platform/dynload/CMakeLists.txt index bb3fec1be9e811c26cc6851314e960e96fc366b3..f4fda65907dc26e9edb91ee46f3b8bd2de7b3f3a 100644 --- a/paddle/platform/dynload/CMakeLists.txt +++ b/paddle/platform/dynload/CMakeLists.txt @@ -1,3 +1,3 @@ -cc_library(dynamic_loader SRCS dynamic_loader.cc DEPS glog gflags) +cc_library(dynamic_loader SRCS dynamic_loader.cc DEPS glog gflags enforce) nv_library(dynload_cuda SRCS cublas.cc cudnn.cc curand.cc nccl.cc DEPS dynamic_loader nccl) diff --git a/paddle/platform/dynload/cudnn.cc b/paddle/platform/dynload/cudnn.cc index d3e4cb567d71b987724366b6a0896f5df0eb6055..76ec82e10840751a654c7d7f57da8d5570d2a9ce 100644 --- a/paddle/platform/dynload/cudnn.cc +++ b/paddle/platform/dynload/cudnn.cc @@ -12,7 +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 "paddle/platform/dynload/cudnn.h" +#include "paddle/platform/enforce.h" namespace paddle { namespace platform { @@ -37,6 +38,25 @@ CUDNN_DNN_ROUTINE_EACH_AFTER_R4(DEFINE_WRAP); CUDNN_DNN_ROUTINE_EACH_R5(DEFINE_WRAP); #endif +#ifdef CUDNN_DNN_ROUTINE_EACH_R7 +CUDNN_DNN_ROUTINE_EACH_R7(DEFINE_WRAP); +#endif + +#ifdef PADDLE_USE_DSO +bool HasCUDNN() { + std::call_once(cudnn_dso_flag, GetCudnnDsoHandle, &cudnn_dso_handle); + return cudnn_dso_handle != nullptr; +} + +void EnforceCUDNNLoaded(const char* fn_name) { + PADDLE_ENFORCE(cudnn_dso_handle != nullptr, + "Cannot load cudnn shared library. Cannot invoke method %s", + fn_name); +} +#else +bool HasCUDNN() { return true; } +#endif + } // namespace dynload } // namespace platform } // namespace paddle diff --git a/paddle/platform/dynload/cudnn.h b/paddle/platform/dynload/cudnn.h index b2d69da93bcd4a5c8e694a18ca648ddc4bd947af..8c937b37d714a06c623f4e204bd572fdd200ea5d 100644 --- a/paddle/platform/dynload/cudnn.h +++ b/paddle/platform/dynload/cudnn.h @@ -25,9 +25,11 @@ namespace dynload { extern std::once_flag cudnn_dso_flag; extern void* cudnn_dso_handle; +extern bool HasCUDNN(); #ifdef PADDLE_USE_DSO +extern void EnforceCUDNNLoaded(const char* fn_name); #define DECLARE_DYNAMIC_LOAD_CUDNN_WRAP(__name) \ struct DynLoad__##__name { \ template \ @@ -36,6 +38,7 @@ extern void* cudnn_dso_handle; std::call_once(cudnn_dso_flag, \ paddle::platform::dynload::GetCudnnDsoHandle, \ &cudnn_dso_handle); \ + EnforceCUDNNLoaded(#__name); \ void* p_##__name = dlsym(cudnn_dso_handle, #__name); \ return reinterpret_cast(p_##__name)(args...); \ } \ @@ -135,6 +138,12 @@ CUDNN_DNN_ROUTINE_EACH_AFTER_R4(DECLARE_DYNAMIC_LOAD_CUDNN_WRAP) CUDNN_DNN_ROUTINE_EACH_R5(DECLARE_DYNAMIC_LOAD_CUDNN_WRAP) #endif +#if CUDNN_VERSION >= 7001 +#define CUDNN_DNN_ROUTINE_EACH_R7(__macro) \ + __macro(cudnnSetConvolutionGroupCount); +CUDNN_DNN_ROUTINE_EACH_R7(DECLARE_DYNAMIC_LOAD_CUDNN_WRAP) +#endif + } // namespace dynload } // namespace platform } // namespace paddle diff --git a/paddle/platform/dynload/dynamic_loader.cc b/paddle/platform/dynload/dynamic_loader.cc index 6feba42c0d9d618d27da12e6a6752058b296995e..7a82d06a0acbfa44386d40df97f6b0e43ed46577 100644 --- a/paddle/platform/dynload/dynamic_loader.cc +++ b/paddle/platform/dynload/dynamic_loader.cc @@ -78,12 +78,11 @@ static inline void GetDsoHandleFromDefaultPath(std::string& dso_path, *dso_handle = dlopen(dso_path.c_str(), dynload_flags); if (nullptr == *dso_handle) { if (dso_path == "libcudnn.dylib") { - PADDLE_ENFORCE(true, - "Note: [Recommend] copy cudnn into /usr/local/cuda/ \n " - "For instance, sudo tar -xzf " - "cudnn-7.5-osx-x64-v5.0-ga.tgz -C /usr/local \n sudo " - "chmod a+r /usr/local/cuda/include/cudnn.h " - "/usr/local/cuda/lib/libcudnn*"); + LOG(WARNING) << "Note: [Recommend] copy cudnn into /usr/local/cuda/ \n " + "For instance, sudo tar -xzf " + "cudnn-7.5-osx-x64-v5.0-ga.tgz -C /usr/local \n sudo " + "chmod a+r /usr/local/cuda/include/cudnn.h " + "/usr/local/cuda/lib/libcudnn*"; } } } @@ -92,7 +91,8 @@ static inline void GetDsoHandleFromDefaultPath(std::string& dso_path, static inline void GetDsoHandleFromSearchPath(const std::string& search_root, const std::string& dso_name, - void** dso_handle) { + void** dso_handle, + bool throw_on_error = true) { int dynload_flags = RTLD_LAZY | RTLD_LOCAL; *dso_handle = nullptr; @@ -111,15 +111,19 @@ static inline void GetDsoHandleFromSearchPath(const std::string& search_root, GetDsoHandleFromDefaultPath(dlPath, dso_handle, dynload_flags); } } - PADDLE_ENFORCE(nullptr != *dso_handle, - "Failed to find dynamic library: %s ( %s ) \n Please specify " - "its path correctly using following ways: \n Method. set " - "environment variable LD_LIBRARY_PATH on Linux or " - "DYLD_LIBRARY_PATH on Mac OS. \n For instance, issue command: " - "export LD_LIBRARY_PATH=... \n Note: After Mac OS 10.11, " - "using the DYLD_LIBRARY_PATH is impossible unless System " - "Integrity Protection (SIP) is disabled.", - dlPath, dlerror()); + auto error_msg = + "Failed to find dynamic library: %s ( %s ) \n Please specify " + "its path correctly using following ways: \n Method. set " + "environment variable LD_LIBRARY_PATH on Linux or " + "DYLD_LIBRARY_PATH on Mac OS. \n For instance, issue command: " + "export LD_LIBRARY_PATH=... \n Note: After Mac OS 10.11, " + "using the DYLD_LIBRARY_PATH is impossible unless System " + "Integrity Protection (SIP) is disabled."; + if (throw_on_error) { + PADDLE_ENFORCE(nullptr != *dso_handle, error_msg, dlPath, dlerror()); + } else if (nullptr == *dso_handle) { + LOG(WARNING) << string::Sprintf(error_msg, dlPath, dlerror()); + } } void GetCublasDsoHandle(void** dso_handle) { @@ -132,9 +136,10 @@ void GetCublasDsoHandle(void** dso_handle) { void GetCudnnDsoHandle(void** dso_handle) { #if defined(__APPLE__) || defined(__OSX__) - GetDsoHandleFromSearchPath(FLAGS_cudnn_dir, "libcudnn.dylib", dso_handle); + GetDsoHandleFromSearchPath(FLAGS_cudnn_dir, "libcudnn.dylib", dso_handle, + false); #else - GetDsoHandleFromSearchPath(FLAGS_cudnn_dir, "libcudnn.so", dso_handle); + GetDsoHandleFromSearchPath(FLAGS_cudnn_dir, "libcudnn.so", dso_handle, false); #endif } diff --git a/paddle/platform/dynload/nccl.cc b/paddle/platform/dynload/nccl.cc index 8f92b8d94d56047b7d3fb43b15e3c06575c8d57b..4cec829a8ad8994d4a7643613331881e3a397b9a 100644 --- a/paddle/platform/dynload/nccl.cc +++ b/paddle/platform/dynload/nccl.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/platform/dynload/nccl.h" @@ -25,6 +25,11 @@ void *nccl_dso_handle; NCCL_RAND_ROUTINE_EACH(DEFINE_WRAP); +void LoadNCCLDSO() { + platform::call_once(nccl_dso_flag, + [] { GetNCCLDsoHandle(&nccl_dso_handle); }); +} + } // namespace dynload } // namespace platform } // namespace paddle diff --git a/paddle/platform/dynload/nccl.h b/paddle/platform/dynload/nccl.h index 981b2ab258a34ce92f02ee12b5957f88ba61d1c0..6c776afc97a53c964f0bc2b2a8abf2c29f474d3f 100644 --- a/paddle/platform/dynload/nccl.h +++ b/paddle/platform/dynload/nccl.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 @@ -28,18 +28,18 @@ extern std::once_flag nccl_dso_flag; extern void* nccl_dso_handle; #ifdef PADDLE_USE_DSO -#define DECLARE_DYNAMIC_LOAD_NCCL_WRAP(__name) \ - struct DynLoad__##__name { \ - template \ - auto operator()(Args... args) -> decltype(__name(args...)) { \ - using nccl_func = decltype(__name(args...)) (*)(Args...); \ - platform::call_once(nccl_dso_flag, \ - paddle::platform::dynload::GetNCCLDsoHandle, \ - &nccl_dso_handle); \ - void* p_##__name = dlsym(nccl_dso_handle, #__name); \ - return reinterpret_cast(p_##__name)(args...); \ - } \ - }; \ +extern void LoadNCCLDSO(); + +#define DECLARE_DYNAMIC_LOAD_NCCL_WRAP(__name) \ + struct DynLoad__##__name { \ + template \ + auto operator()(Args... args) -> decltype(__name(args...)) { \ + using nccl_func = decltype(__name(args...)) (*)(Args...); \ + paddle::platform::dynload::LoadNCCLDSO(); \ + void* p_##__name = dlsym(nccl_dso_handle, #__name); \ + return reinterpret_cast(p_##__name)(args...); \ + } \ + }; \ extern DynLoad__##__name __name #else #define DECLARE_DYNAMIC_LOAD_NCCL_WRAP(__name) \ @@ -63,6 +63,8 @@ extern void* nccl_dso_handle; __macro(ncclAllReduce); \ __macro(ncclBcast); \ __macro(ncclAllGather); \ + __macro(ncclGroupStart); \ + __macro(ncclGroupEnd); \ __macro(ncclReduce); \ __macro(ncclGetErrorString); diff --git a/paddle/platform/enforce.cc b/paddle/platform/enforce.cc new file mode 100644 index 0000000000000000000000000000000000000000..e8d31bc782ec3cddd18ceaedf88fe5e7b4aed2cc --- /dev/null +++ b/paddle/platform/enforce.cc @@ -0,0 +1,19 @@ +/* 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. */ + +#include "paddle/platform/enforce.h" + +namespace paddle { +namespace platform {} // namespace platform +} // namespace paddle diff --git a/paddle/platform/enforce.h b/paddle/platform/enforce.h index bfe708748a62ff9ac5d151bc652142e1f4925c83..d1c7be0790b5e11d6273efe6c08cdb7bf22425c6 100644 --- a/paddle/platform/enforce.h +++ b/paddle/platform/enforce.h @@ -22,6 +22,7 @@ limitations under the License. */ #include #include +#include "paddle/platform/macros.h" #include "paddle/string/printf.h" #include "paddle/string/to_string.h" @@ -49,7 +50,6 @@ limitations under the License. */ namespace paddle { namespace platform { -namespace { #ifdef __GNUC__ inline std::string demangle(std::string name) { int status = -4; // some arbitrary value to eliminate the compiler warning @@ -60,7 +60,6 @@ inline std::string demangle(std::string name) { #else inline std::string demangle(std::string name) { return name; } #endif -} struct EnforceNotMet : public std::exception { std::exception_ptr exp_; @@ -236,16 +235,24 @@ inline void throw_on_error(T e) { __PADDLE_BINARY_COMPARE(__VAL0, __VAL1, <, >=, __VA_ARGS__) #define PADDLE_ENFORCE_LE(__VAL0, __VAL1, ...) \ __PADDLE_BINARY_COMPARE(__VAL0, __VAL1, <=, >, __VA_ARGS__) -#define PADDLE_ENFORCE_NOT_NULL(__VAL, ...) \ - PADDLE_ENFORCE(nullptr != (__VAL), #__VAL " should not be null\n%s", \ - paddle::string::Sprintf("" __VA_ARGS__)); - -#define __PADDLE_BINARY_COMPARE(__VAL0, __VAL1, __CMP, __INV_CMP, ...) \ - PADDLE_ENFORCE(__VAL0 __CMP __VAL1, \ - "enforce %s " #__CMP " %s failed, %s " #__INV_CMP " %s\n%s", \ - #__VAL0, #__VAL1, paddle::string::to_string(__VAL0), \ - paddle::string::to_string(__VAL1), \ - paddle::string::Sprintf("" __VA_ARGS__)); +#define PADDLE_ENFORCE_NOT_NULL(__VAL, ...) \ + do { \ + if (UNLIKELY(nullptr == (__VAL))) { \ + PADDLE_THROW(#__VAL " should not be null\n%s", \ + paddle::string::Sprintf("" __VA_ARGS__)); \ + } \ + } while (0) + +#define __PADDLE_BINARY_COMPARE(__VAL0, __VAL1, __CMP, __INV_CMP, ...) \ + do { \ + if (UNLIKELY(!((__VAL0)__CMP(__VAL1)))) { \ + PADDLE_THROW("enforce %s " #__CMP " %s failed, %s " #__INV_CMP \ + " %s\n%s", \ + #__VAL0, #__VAL1, paddle::string::to_string(__VAL0), \ + paddle::string::to_string(__VAL1), \ + paddle::string::Sprintf("" __VA_ARGS__)); \ + } \ + } while (0) } // namespace platform } // namespace paddle diff --git a/paddle/platform/for_range.h b/paddle/platform/for_range.h new file mode 100644 index 0000000000000000000000000000000000000000..694a66d9ac4eb6ad02daf1931806fa1287de7cab --- /dev/null +++ b/paddle/platform/for_range.h @@ -0,0 +1,85 @@ +/* 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 +#include "paddle/platform/device_context.h" + +namespace paddle { +namespace platform { + +template +struct ForRange { + ForRange(const DeviceContext& dev_ctx, size_t limit); + + template + void operator()(Function func) const; +}; + +template <> +struct ForRange { + ForRange(const CPUDeviceContext& dev_ctx, size_t limit) : limit_(limit) {} + + template + void operator()(Function func) const { + for (size_t i = 0; i < limit_; ++i) { + func(i); + } + } + + size_t limit_; +}; + +#ifdef __NVCC__ +template +__global__ static void ForRangeElemwiseOpGridIsOne(Function func) { + size_t idx = static_cast(threadIdx.x); + func(idx); +} + +template +__global__ static void ForRangeElemwiseOp(Function func, int limit) { + size_t idx = static_cast(blockIdx.x * blockDim.x + threadIdx.x); + if (idx < limit) { + func(idx); + } +} + +template <> +struct ForRange { + ForRange(const CUDADeviceContext& dev_ctx, size_t limit) + : dev_ctx_(dev_ctx), limit_(static_cast(limit)) {} + + template + inline void operator()(Function func) const { + constexpr int num_threads = 1024; + int block_size = limit_ <= num_threads ? limit_ : num_threads; + int grid_size = (limit_ + num_threads - 1) / num_threads; + + if (grid_size == 1) { + ForRangeElemwiseOpGridIsOne<<<1, block_size, 0, dev_ctx_.stream()>>>( + func); + } else { + ForRangeElemwiseOp<<>>( + func, limit_); + } + } + + const CUDADeviceContext& dev_ctx_; + int limit_; +}; + +#endif + +} // namespace platform +} // namespace paddle diff --git a/paddle/platform/gpu_info.cc b/paddle/platform/gpu_info.cc index 36b216d872138d49bfd5ab6e3499d15d49ebd0ca..7037551d7544d6fea54e2f4bf887309b7dc5a52e 100644 --- a/paddle/platform/gpu_info.cc +++ b/paddle/platform/gpu_info.cc @@ -18,8 +18,8 @@ limitations under the License. */ #include "paddle/platform/enforce.h" -DEFINE_double(fraction_of_gpu_memory_to_use, 0.95, - "Default use 95% of GPU memory for PaddlePaddle," +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"); namespace paddle { @@ -73,17 +73,22 @@ size_t GpuMaxChunkSize() { size_t available = 0; GpuMemoryUsage(available, total); + VLOG(10) << "GPU Usage " << available / 1024 / 1024 << "M/" + << total / 1024 / 1024 << "M"; + size_t reserving = static_cast(0.05 * total); + // If available less than minimum chunk size, no usable memory exists. + available = + std::min(std::max(available, GpuMinChunkSize()) - GpuMinChunkSize(), + total - reserving); // Reserving the rest memory for page tables, etc. - size_t reserving = (1 - FLAGS_fraction_of_gpu_memory_to_use) * total; - // If available less than minimum chunk size, no usable memory exists. - available = std::max(available, GpuMinChunkSize()) - GpuMinChunkSize(); + size_t allocating = static_cast(FLAGS_fraction_of_gpu_memory_to_use * + (total - reserving)); - // If available less than reserving, no usable memory exists. - size_t usable = std::max(available, reserving) - reserving; + PADDLE_ENFORCE_LE(allocating, available); - return usable; + return allocating; } void GpuMemcpyAsync(void *dst, const void *src, size_t count, @@ -92,17 +97,6 @@ void GpuMemcpyAsync(void *dst, const void *src, size_t count, "cudaMemcpyAsync failed in paddle::platform::GpuMemcpyAsync"); } -void GpuMemcpySync(void *dst, const void *src, size_t count, - enum cudaMemcpyKind kind) { - PADDLE_ENFORCE(cudaMemcpy(dst, src, count, kind), - "cudaMemcpy failed in paddle::platform::GpuMemcpySync"); - // note: cudaMemcpy may actually be asynchronous with respect to the caller, - // block on stream 0 to make sure the copy has completed - PADDLE_ENFORCE( - cudaStreamSynchronize(0), - "cudaStreamSynchronize failed in paddle::platform::GpuMemcpySync"); -} - void GpuMemcpyPeer(void *dst, int dst_device, const void *src, int src_device, size_t count, cudaStream_t stream) { PADDLE_ENFORCE( diff --git a/paddle/platform/gpu_info.h b/paddle/platform/gpu_info.h index db961f3838af73855312d4cf6a80e2355306e08f..d05131fa4196057d19a8ae57bf4574c666e409cf 100644 --- a/paddle/platform/gpu_info.h +++ b/paddle/platform/gpu_info.h @@ -52,10 +52,6 @@ size_t GpuMaxChunkSize(); void GpuMemcpyAsync(void *dst, const void *src, size_t count, enum cudaMemcpyKind kind, cudaStream_t stream); -//! Copy memory from address src to dst synchronously. -void GpuMemcpySync(void *dst, const void *src, size_t count, - enum cudaMemcpyKind kind); - //! Copy memory from one device to another device. void GpuMemcpyPeer(void *dst, int dst_device, const void *src, int src_device, size_t count, cudaStream_t stream); diff --git a/paddle/platform/nccl_test.cu b/paddle/platform/nccl_test.cu index c99dae68bef67c58d3efea42fef45e84bb3d9255..ef6d845874745af1150e4425f8d6be416cc44ece 100644 --- a/paddle/platform/nccl_test.cu +++ b/paddle/platform/nccl_test.cu @@ -1,28 +1,30 @@ /* 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 +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 + 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. */ +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 "glog/logging.h" #include "gtest/gtest.h" + +#include "paddle/framework/init.h" #include "paddle/platform/device_context.h" #include "paddle/platform/dynload/nccl.h" #include "paddle/platform/enforce.h" #include "paddle/platform/gpu_info.h" -#include -#include -#include - static int dev_count = 0; namespace paddle { @@ -32,6 +34,7 @@ TEST(NCCL, init) { std::vector comms; comms.resize(dev_count); PADDLE_ENFORCE(dynload::ncclCommInitAll(comms.data(), dev_count, nullptr)); + for (int i = 0; i < dev_count; ++i) { dynload::ncclCommDestroy(comms[i]); } @@ -47,7 +50,7 @@ struct PerThreadData { T* RecvBuff() { return thrust::raw_pointer_cast(recv_buff.data()); } - PerThreadData(int gpu_id, size_t size) : dev_ctx(GPUPlace(gpu_id)) { + PerThreadData(int gpu_id, size_t size) : dev_ctx(CUDAPlace(gpu_id)) { send_buff.resize(size); for (size_t i = 0; i < size; ++i) { send_buff[i] = static_cast(i); @@ -62,7 +65,7 @@ TEST(NCCL, all_reduce) { std::vector comms; comms.resize(dev_count); VLOG(1) << "Initializing ncclComm"; - PADDLE_ENFORCE(dynload::ncclCommInitAll(comms.data(), dev_count, nullptr)); + dynload::ncclCommInitAll(comms.data(), dev_count, nullptr); VLOG(1) << "ncclComm initialized"; VLOG(1) << "Creating thread data"; std::vector>> data; @@ -131,6 +134,18 @@ int main(int argc, char** argv) { << dev_count; return 0; } + + std::vector places; + + places.emplace_back(paddle::platform::CPUPlace()); + int count = paddle::platform::GetCUDADeviceCount(); + for (int i = 0; i < count; ++i) { + places.emplace_back(paddle::platform::CUDAPlace(i)); + } + + VLOG(0) << " DeviceCount " << count; + paddle::platform::DeviceContextPool::Init(places); + testing::InitGoogleTest(&argc, argv); return RUN_ALL_TESTS(); } diff --git a/paddle/platform/place.cc b/paddle/platform/place.cc index 856e54df89c1c18ade040957188a2fbda0901473..249527e3e136992970033c44ad490a1744bfed35 100644 --- a/paddle/platform/place.cc +++ b/paddle/platform/place.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/platform/place.h" @@ -23,7 +23,9 @@ class PlacePrinter : public boost::static_visitor<> { public: explicit PlacePrinter(std::ostream &os) : os_(os) {} void operator()(const CPUPlace &) { os_ << "CPUPlace"; } - void operator()(const GPUPlace &p) { os_ << "GPUPlace(" << p.device << ")"; } + void operator()(const CUDAPlace &p) { + os_ << "CUDAPlace(" << p.device << ")"; + } private: std::ostream &os_; @@ -36,16 +38,15 @@ static Place the_default_place; void set_place(const Place &place) { the_default_place = place; } const Place &get_place() { return the_default_place; } -const GPUPlace default_gpu() { return GPUPlace(0); } +const CUDAPlace default_gpu() { return CUDAPlace(0); } const CPUPlace default_cpu() { return CPUPlace(); } bool is_gpu_place(const Place &p) { - return boost::apply_visitor(IsGPUPlace(), p); -} -bool is_cpu_place(const Place &p) { - return !boost::apply_visitor(IsGPUPlace(), p); + return boost::apply_visitor(IsCUDAPlace(), p); } +bool is_cpu_place(const Place &p) { return !is_gpu_place(p); } + bool places_are_same_class(const Place &p1, const Place &p2) { return p1.which() == p2.which(); } diff --git a/paddle/platform/place.h b/paddle/platform/place.h index 5370360a7de26e409a1545182a12d3df1f37658b..76b5c502cc48431a4e9b13b07505978884576e1d 100644 --- a/paddle/platform/place.h +++ b/paddle/platform/place.h @@ -15,7 +15,7 @@ limitations under the License. */ #pragma once #include - +#include "paddle/platform/enforce.h" #include "paddle/platform/variant.h" namespace paddle { @@ -31,39 +31,31 @@ struct CPUPlace { inline bool operator!=(const CPUPlace &) const { return false; } }; -struct GPUPlace { - GPUPlace() : GPUPlace(0) {} - explicit GPUPlace(int d) : device(d) {} +struct CUDAPlace { + CUDAPlace() : CUDAPlace(0) {} + explicit CUDAPlace(int d) : device(d) {} inline int GetDeviceId() const { return device; } // needed for variant equality comparison - inline bool operator==(const GPUPlace &o) const { return device == o.device; } - inline bool operator!=(const GPUPlace &o) const { return !(*this == o); } + inline bool operator==(const CUDAPlace &o) const { + return device == o.device; + } + inline bool operator!=(const CUDAPlace &o) const { return !(*this == o); } int device; }; -struct IsGPUPlace : public boost::static_visitor { +struct IsCUDAPlace : public boost::static_visitor { bool operator()(const CPUPlace &) const { return false; } - bool operator()(const GPUPlace &gpu) const { return true; } + bool operator()(const CUDAPlace &gpu) const { return true; } }; -// Define the max number of Place in bit length. i.e., the max number of places -// should be less equal than 2^(NUM_PLACE_TYPE_LIMIT_IN_BIT) -#define NUM_PLACE_TYPE_LIMIT_IN_BIT 4 - -typedef boost::variant Place; - -// static check number of place types is less equal than -// 2^(NUM_PLACE_TYPE_LIMIT_IN_BIT) -BOOST_MPL_ASSERT((boost::mpl::less_equal< - Place::types::size, - boost::mpl::long_<1 << NUM_PLACE_TYPE_LIMIT_IN_BIT>>)); +typedef boost::variant Place; void set_place(const Place &); const Place &get_place(); -const GPUPlace default_gpu(); +const CUDAPlace default_gpu(); const CPUPlace default_cpu(); bool is_gpu_place(const Place &); @@ -72,5 +64,31 @@ bool places_are_same_class(const Place &, const Place &); std::ostream &operator<<(std::ostream &, const Place &); +template +struct PlaceVisitorWrapper + : public boost::static_visitor { + const Visitor &visitor_; + explicit PlaceVisitorWrapper(const Visitor &visitor) : visitor_(visitor) {} + + typename Visitor::result_type operator()(const CPUPlace &cpu) const { + return visitor_(cpu); + } + + typename Visitor::result_type operator()(const CUDAPlace &cuda) const { +#ifdef PADDLE_WITH_CUDA + return visitor_(cuda); +#else + PADDLE_THROW("Paddle is not compiled with CUDA. Cannot visit cuda device"); + return typename Visitor::result_type(); +#endif + } +}; + +template +typename Visitor::result_type VisitPlace(const Place &place, + const Visitor &visitor) { + return boost::apply_visitor(PlaceVisitorWrapper(visitor), place); +} + } // namespace platform } // namespace paddle diff --git a/paddle/platform/place_test.cc b/paddle/platform/place_test.cc index 33e2e5a439ce6801c02daba4bcbd462a74d7a614..4f1eba01df5531529ad3c79648b5e7f8651df619 100644 --- a/paddle/platform/place_test.cc +++ b/paddle/platform/place_test.cc @@ -4,7 +4,7 @@ TEST(Place, Equality) { paddle::platform::CPUPlace cpu; - paddle::platform::GPUPlace g0(0), g1(1), gg0(0); + paddle::platform::CUDAPlace g0(0), g1(1), gg0(0); EXPECT_EQ(cpu, cpu); EXPECT_EQ(g0, g0); @@ -22,6 +22,7 @@ TEST(Place, Default) { EXPECT_TRUE(paddle::platform::is_gpu_place(paddle::platform::default_gpu())); EXPECT_TRUE(paddle::platform::is_cpu_place(paddle::platform::default_cpu())); + EXPECT_FALSE(paddle::platform::is_cpu_place(paddle::platform::get_place())); paddle::platform::set_place(paddle::platform::CPUPlace()); EXPECT_TRUE(paddle::platform::is_cpu_place(paddle::platform::get_place())); } @@ -29,8 +30,8 @@ TEST(Place, Default) { TEST(Place, Print) { { std::stringstream ss; - ss << paddle::platform::GPUPlace(1); - EXPECT_EQ("GPUPlace(1)", ss.str()); + ss << paddle::platform::CUDAPlace(1); + EXPECT_EQ("CUDAPlace(1)", ss.str()); } { std::stringstream ss; diff --git a/paddle/platform/transform.h b/paddle/platform/transform.h index bb9d59ec0a18ce013632f128c9b5d230255f1ac4..a88902b164c7a705cae09662e724b35cd8c8b4bf 100644 --- a/paddle/platform/transform.h +++ b/paddle/platform/transform.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 @@ -31,7 +31,7 @@ namespace paddle { namespace platform { // Transform on host or device. It provides the same API in std library. -template +template struct Transform { template void operator()(const DeviceContext& context, InputIter first, InputIter last, @@ -45,16 +45,16 @@ struct Transform { }; template <> -struct Transform { +struct Transform { template - void operator()(const DeviceContext& context, InputIter first, InputIter last, - OutputIter result, UnaryOperation op) { + void operator()(const platform::CPUDeviceContext& context, InputIter first, + InputIter last, OutputIter result, UnaryOperation op) { std::transform(first, last, result, op); } template - void operator()(const DeviceContext& context, InputIter1 first1, + void operator()(const platform::CPUDeviceContext& context, InputIter1 first1, InputIter1 last1, InputIter2 first2, OutputIter result, BinaryOperation op) { std::transform(first1, last1, first2, result, op); @@ -63,27 +63,25 @@ struct Transform { #ifdef __NVCC__ template <> -struct Transform { +struct Transform { template - void operator()(const DeviceContext& context, InputIter first, InputIter last, - OutputIter result, UnaryOperation op) { + void operator()(const platform::CUDADeviceContext& context, InputIter first, + InputIter last, OutputIter result, UnaryOperation op) { auto place = context.GetPlace(); PADDLE_ENFORCE(is_gpu_place(place), "It must use GPU place."); - auto& ctx = reinterpret_cast(context); - thrust::transform(thrust::cuda::par.on(ctx.stream()), + thrust::transform(thrust::cuda::par.on(context.stream()), details::DevPtrCast(first), details::DevPtrCast(last), details::DevPtrCast(result), op); } template - void operator()(const DeviceContext& context, InputIter1 first1, + void operator()(const platform::CUDADeviceContext& context, InputIter1 first1, InputIter1 last1, InputIter2 first2, OutputIter result, BinaryOperation op) { auto place = context.GetPlace(); PADDLE_ENFORCE(is_gpu_place(place), "It must use GPU place."); - auto& ctx = reinterpret_cast(context); - thrust::transform(thrust::cuda::par.on(ctx.stream()), + thrust::transform(thrust::cuda::par.on(context.stream()), details::DevPtrCast(first1), details::DevPtrCast(last1), details::DevPtrCast(first2), details::DevPtrCast(result), op); diff --git a/paddle/platform/transform_test.cu b/paddle/platform/transform_test.cu index c76cab80e4b0e8df98a7be15f86699cfb6f93af2..af9204a0a7b6e09fcfdacb8ba985e269665b4034 100644 --- a/paddle/platform/transform_test.cu +++ b/paddle/platform/transform_test.cu @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/memory/memcpy.h" @@ -39,7 +39,7 @@ TEST(Transform, CPUUnary) { using namespace paddle::platform; CPUDeviceContext ctx; float buf[4] = {0.1, 0.2, 0.3, 0.4}; - Transform trans; + Transform trans; trans(ctx, buf, buf + 4, buf, Scale(10)); for (int i = 0; i < 4; ++i) { ASSERT_NEAR(buf[i], static_cast(i + 1), 1e-5); @@ -49,15 +49,15 @@ TEST(Transform, CPUUnary) { TEST(Transform, GPUUnary) { using namespace paddle::platform; using namespace paddle::memory; - GPUPlace gpu0(0); + CUDAPlace gpu0(0); CUDADeviceContext ctx(gpu0); float cpu_buf[4] = {0.1, 0.2, 0.3, 0.4}; float* gpu_buf = static_cast(Alloc(gpu0, sizeof(float) * 4)); - Copy(gpu0, gpu_buf, CPUPlace(), cpu_buf, sizeof(cpu_buf)); - Transform trans; + Copy(gpu0, gpu_buf, CPUPlace(), cpu_buf, sizeof(cpu_buf), ctx.stream()); + Transform trans; trans(ctx, gpu_buf, gpu_buf + 4, gpu_buf, Scale(10)); ctx.Wait(); - Copy(CPUPlace(), cpu_buf, gpu0, gpu_buf, sizeof(cpu_buf)); + Copy(CPUPlace(), cpu_buf, gpu0, gpu_buf, sizeof(cpu_buf), ctx.stream()); Free(gpu0, gpu_buf); for (int i = 0; i < 4; ++i) { ASSERT_NEAR(cpu_buf[i], static_cast(i + 1), 1e-5); @@ -68,7 +68,7 @@ TEST(Transform, CPUBinary) { using namespace paddle::platform; using namespace paddle::memory; int buf[4] = {1, 2, 3, 4}; - Transform trans; + Transform trans; CPUDeviceContext ctx; trans(ctx, buf, buf + 4, buf, buf, Multiply()); for (int i = 0; i < 4; ++i) { @@ -80,14 +80,14 @@ TEST(Transform, GPUBinary) { using namespace paddle::platform; using namespace paddle::memory; int buf[4] = {1, 2, 3, 4}; - GPUPlace gpu0(0); + CUDAPlace gpu0(0); CUDADeviceContext ctx(gpu0); int* gpu_buf = static_cast(Alloc(gpu0, sizeof(buf))); - Copy(gpu0, gpu_buf, CPUPlace(), buf, sizeof(buf)); - Transform trans; + Copy(gpu0, gpu_buf, CPUPlace(), buf, sizeof(buf), ctx.stream()); + Transform trans; trans(ctx, gpu_buf, gpu_buf + 4, gpu_buf, gpu_buf, Multiply()); ctx.Wait(); - Copy(CPUPlace(), buf, gpu0, gpu_buf, sizeof(buf)); + Copy(CPUPlace(), buf, gpu0, gpu_buf, sizeof(buf), ctx.stream()); Free(gpu0, gpu_buf); for (int i = 0; i < 4; ++i) { ASSERT_EQ((i + 1) * (i + 1), buf[i]); diff --git a/paddle/platform/variant.h b/paddle/platform/variant.h index 619897ca19eb2e6f4dbfd9160edf8c4bc58c89a9..ea6ef8fddf5171731fa1fd323983bf7d90ecdd9a 100644 --- a/paddle/platform/variant.h +++ b/paddle/platform/variant.h @@ -1,19 +1,32 @@ /* 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 +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 + 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. */ +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 __CUDACC__ +#ifdef __CUDACC_VER_MAJOR__ +// CUDA 9 define `__CUDACC_VER__` as a warning message, manually define +// __CUDACC_VER__ instead. +#undef __CUDACC_VER__ + +#define __CUDACC_VER__ \ + (__CUDACC_VER_MAJOR__ * 10000 + __CUDACC_VER_MINOR__ * 100 + \ + __CUDACC_VER_BUILD__) +#endif + +#endif + #include #ifdef PADDLE_WITH_CUDA diff --git a/paddle/pserver/CMakeLists.txt b/paddle/pserver/CMakeLists.txt index ccfc0e76020c7b4f54a493cc4048e7571379ec1a..f75475a88f7224ee3889827795088c8aa920b63b 100644 --- a/paddle/pserver/CMakeLists.txt +++ b/paddle/pserver/CMakeLists.txt @@ -49,7 +49,7 @@ if(WITH_TESTING) add_subdirectory(test) endif() -if(NOT WITH_C_API) +if(NOT MOBILE_INFERENCE) add_executable(paddle_pserver_main ${PSERVER_MAIN_SOURCES}) link_paddle_exe(paddle_pserver_main) diff --git a/paddle/pybind/CMakeLists.txt b/paddle/pybind/CMakeLists.txt index a9bcc474387513a8ca019bc9382b88c93e08ff8d..7b374307071d2da91a677361b404448f1a3816b0 100644 --- a/paddle/pybind/CMakeLists.txt +++ b/paddle/pybind/CMakeLists.txt @@ -1,8 +1,13 @@ if(WITH_PYTHON) cc_library(paddle_pybind SHARED - SRCS pybind.cc exception.cc protobuf.cc - DEPS pybind python backward proto_desc tensor_array paddle_memory executor prune + SRCS pybind.cc exception.cc protobuf.cc const_value.cc + DEPS pybind python backward proto_desc paddle_memory executor prune init ${GLOB_OP_LIB}) + if(NOT APPLE AND NOT ANDROID) + target_link_libraries(paddle_pybind rt) + endif(NOT APPLE AND NOT ANDROID) endif(WITH_PYTHON) -cc_binary(print_operators_doc SRCS print_operators_doc.cc DEPS ${GLOB_OP_LIB} tensor_array) +if(WITH_DOC) + cc_binary(print_operators_doc SRCS print_operators_doc.cc DEPS ${GLOB_OP_LIB}) +endif(WITH_DOC) diff --git a/paddle/pybind/const_value.cc b/paddle/pybind/const_value.cc new file mode 100644 index 0000000000000000000000000000000000000000..761635aa5e5eac445c2ec8331b0dc37ffd11248c --- /dev/null +++ b/paddle/pybind/const_value.cc @@ -0,0 +1,34 @@ +/* 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. */ + +#include "const_value.h" +#include "paddle/framework/operator.h" + +namespace paddle { +namespace pybind { + +void BindConstValue(pybind11::module& m) { + m.def("kEmptyVarName", [] { return framework::kEmptyVarName; }); + m.def("kTempVarName", [] { return framework::kTempVarName; }); + m.def("kGradVarSuffix", [] { return framework::kGradVarSuffix; }); + m.def("kZeroVarSuffix", [] { return framework::kZeroVarSuffix; }); + + // for kernel_hint key + m.def("kUseCPU", [] { return framework::kUseCPU; }); + m.def("kUseCUDNN", [] { return framework::kUseCUDNN; }); + m.def("kUseMKLDNN", [] { return framework::kUseMKLDNN; }); +} + +} // namespace pybind +} // namespace paddle diff --git a/paddle/pybind/const_value.h b/paddle/pybind/const_value.h new file mode 100644 index 0000000000000000000000000000000000000000..3d57c972a9d5339c0e155fa1d6395af9face8744 --- /dev/null +++ b/paddle/pybind/const_value.h @@ -0,0 +1,26 @@ +/* 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 +#include +#include "paddle/platform/enforce.h" +#include "pybind11/pybind11.h" + +namespace py = pybind11; + +namespace paddle { +namespace pybind { +extern void BindConstValue(pybind11::module& m); +} // namespace pybind +} // namespace paddle diff --git a/paddle/pybind/exception.cc b/paddle/pybind/exception.cc index ff79b12ee4b28c53ee04f4c170b5bca9ca28d14a..e29ac3ebab760a011a3798f8e4be46270d6b80cc 100644 --- a/paddle/pybind/exception.cc +++ b/paddle/pybind/exception.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/pybind/exception.h" diff --git a/paddle/pybind/exception.h b/paddle/pybind/exception.h index 70beac146046f74e23f747bab130483901a7d443..436ddd5707ace37a5668c8d4401c1bdcf2dadfe3 100644 --- a/paddle/pybind/exception.h +++ b/paddle/pybind/exception.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 diff --git a/paddle/pybind/print_operators_doc.cc b/paddle/pybind/print_operators_doc.cc index 24f2a9383f7a069f1a8c7ed2bf3da46720470efa..f4f281229e611a6c9c8e9ecd54e0097ab683bbf3 100644 --- a/paddle/pybind/print_operators_doc.cc +++ b/paddle/pybind/print_operators_doc.cc @@ -31,31 +31,32 @@ std::string Escape(const std::string& s) { return r; } -std::string AttrType(paddle::framework::AttrType at) { +std::string AttrType(paddle::framework::proto::AttrType at) { switch (at) { - case paddle::framework::INT: + case paddle::framework::proto::INT: return "int"; - case paddle::framework::FLOAT: + case paddle::framework::proto::FLOAT: return "float"; - case paddle::framework::STRING: + case paddle::framework::proto::STRING: return "string"; - case paddle::framework::BOOLEAN: + case paddle::framework::proto::BOOLEAN: return "bool"; - case paddle::framework::INTS: + case paddle::framework::proto::INTS: return "int array"; - case paddle::framework::FLOATS: + case paddle::framework::proto::FLOATS: return "float array"; - case paddle::framework::STRINGS: + case paddle::framework::proto::STRINGS: return "string array"; - case paddle::framework::BOOLEANS: + case paddle::framework::proto::BOOLEANS: return "bool array"; - case paddle::framework::BLOCK: + case paddle::framework::proto::BLOCK: return "block id"; } return "UNKNOWN"; // not possible } -void PrintVar(const paddle::framework::OpProto::Var& v, std::stringstream& ss) { +void PrintVar(const paddle::framework::proto::OpProto::Var& v, + std::stringstream& ss) { ss << " { " << "\n" << " \"name\" : \"" << Escape(v.name()) << "\",\n" @@ -65,7 +66,7 @@ void PrintVar(const paddle::framework::OpProto::Var& v, std::stringstream& ss) { << " },"; } -void PrintAttr(const paddle::framework::OpProto::Attr& a, +void PrintAttr(const paddle::framework::proto::OpProto::Attr& a, std::stringstream& ss) { ss << " { " << "\n" @@ -81,7 +82,7 @@ void PrintOpProto(const std::string& type, std::stringstream& ss) { std::cerr << "Processing " << type << "\n"; - const paddle::framework::OpProto* p = opinfo.proto_; + const paddle::framework::proto::OpProto* p = opinfo.proto_; if (p == nullptr) { return; // It is possible that an operator doesn't have OpProto. } diff --git a/paddle/pybind/protobuf.cc b/paddle/pybind/protobuf.cc index 5a1ff9b7976abbe4a37f8366181d9d1ae78ea4a0..07292d47e9c165c67fe4a30ee7d851c350beb2e0 100644 --- a/paddle/pybind/protobuf.cc +++ b/paddle/pybind/protobuf.cc @@ -108,21 +108,21 @@ static py::bytes SerializeMessage(T &self) { // Bind Methods void BindProgramDesc(py::module &m) { - py::class_(m, "ProgramDesc", "") + py::class_(m, "ProgramDesc", "") .def(py::init<>()) .def("__init__", - [](ProgramDescBind &self, const ProgramDescBind &other) { - new (&self) ProgramDescBind(other); + [](ProgramDesc &self, const ProgramDesc &other) { + new (&self) ProgramDesc(other); }) .def("__init__", - [](ProgramDescBind &self, const py::bytes &binary_str) { + [](ProgramDesc &self, const py::bytes &binary_str) { std::string str(binary_str); - new (&self) ProgramDescBind(str); + new (&self) ProgramDesc(str); }) - .def("append_block", &ProgramDescBind::AppendBlock, + .def("append_block", &ProgramDesc::AppendBlock, py::return_value_policy::reference) .def("append_backward", - [](ProgramDescBind &program_desc, const VarDescBind &target, + [](ProgramDesc &program_desc, const VarDesc &target, const std::unordered_set &no_grad_vars) { ParamGradInfoMap param_grad_map = AppendBackward(program_desc, target, no_grad_vars); @@ -138,13 +138,13 @@ void BindProgramDesc(py::module &m) { } return retv; }) - .def("block", &ProgramDescBind::MutableBlock, + .def("block", &ProgramDesc::MutableBlock, py::return_value_policy::reference) - .def("num_blocks", &ProgramDescBind::Size) - .def("serialize_to_string", SerializeMessage) + .def("num_blocks", &ProgramDesc::Size) + .def("serialize_to_string", SerializeMessage) .def("parse_from_string", - [](ProgramDescBind &program_desc, const std::string &data) { - ProgramDesc *desc = program_desc.Proto(); + [](ProgramDesc &program_desc, const std::string &data) { + proto::ProgramDesc *desc = program_desc.Proto(); PADDLE_ENFORCE(desc->ParseFromString(data), "Fail to parse ProgramDesc from string. This could " "be a bug of Paddle."); @@ -152,109 +152,134 @@ void BindProgramDesc(py::module &m) { } void BindBlockDesc(py::module &m) { - py::class_(m, "BlockDesc", "") - .def_property_readonly("id", &BlockDescBind::ID) - .def_property_readonly("parent", &BlockDescBind::Parent) - .def("append_op", &BlockDescBind::AppendOp, + py::class_(m, "BlockDesc", "") + .def_property_readonly("id", &BlockDesc::ID) + .def_property_readonly("parent", &BlockDesc::Parent) + .def("append_op", &BlockDesc::AppendOp, py::return_value_policy::reference) - .def("prepend_op", &BlockDescBind::PrependOp, + .def("prepend_op", &BlockDesc::PrependOp, py::return_value_policy::reference) + .def("remove_op", &BlockDesc::RemoveOp) .def("var", - [](BlockDescBind &self, py::bytes byte_name) { + [](BlockDesc &self, py::bytes byte_name) { std::string name = byte_name; return self.Var(name); }, py::return_value_policy::reference) .def("has_var", - [](BlockDescBind &self, py::bytes byte_name) { + [](BlockDesc &self, py::bytes byte_name) { std::string name = byte_name; return self.HasVar(name); }) + .def("has_var_recursive", + [](BlockDesc &self, py::bytes byte_name) { + std::string name = byte_name; + return self.HasVarRecursive(name); + }) .def("find_var", - [](BlockDescBind &self, py::bytes byte_name) { + [](BlockDesc &self, py::bytes byte_name) { std::string name = byte_name; return self.FindVar(name); }, py::return_value_policy::reference) - .def("all_vars", &BlockDescBind::AllVars, + .def("find_var_recursive", + [](BlockDesc &self, py::bytes byte_name) { + std::string name = byte_name; + return self.FindVarRecursive(name); + }, py::return_value_policy::reference) - .def("op_size", &BlockDescBind::OpSize) - .def("op", &BlockDescBind::Op, py::return_value_policy::reference) - .def("serialize_to_string", SerializeMessage); + .def("all_vars", &BlockDesc::AllVars, py::return_value_policy::reference) + .def("op_size", &BlockDesc::OpSize) + .def("op", &BlockDesc::Op, py::return_value_policy::reference) + .def("serialize_to_string", SerializeMessage); } void BindVarDsec(py::module &m) { - py::enum_(m, "DataType", "") - .value("BOOL", DataType::BOOL) - .value("INT16", DataType::INT16) - .value("INT32", DataType::INT32) - .value("INT64", DataType::INT64) - .value("FP16", DataType::FP16) - .value("FP32", DataType::FP32) - .value("FP64", DataType::FP64); + py::enum_(m, "DataType", "") + .value("BOOL", proto::DataType::BOOL) + .value("INT16", proto::DataType::INT16) + .value("INT32", proto::DataType::INT32) + .value("INT64", proto::DataType::INT64) + .value("FP16", proto::DataType::FP16) + .value("FP32", proto::DataType::FP32) + .value("FP64", proto::DataType::FP64); - py::class_ var_desc(m, "VarDesc", ""); + py::class_ var_desc(m, "VarDesc", ""); var_desc .def("name", - [](const VarDescBind &self) { + [](const VarDesc &self) { py::bytes name = self.Name(); return name; }, py::return_value_policy::reference) - .def("set_shape", &VarDescBind::SetShape) - .def("set_data_type", &VarDescBind::SetDataType) - .def("shape", &VarDescBind::Shape, py::return_value_policy::reference) - .def("data_type", &VarDescBind::GetDataType) - .def("lod_level", &VarDescBind::GetLodLevel) - .def("set_lod_level", &VarDescBind::SetLoDLevel) - .def("type", &VarDescBind::GetType) - .def("set_type", &VarDescBind::SetType) - .def("serialize_to_string", SerializeMessage) - .def("persistable", &VarDescBind::Persistable) - .def("set_persistable", &VarDescBind::SetPersistable); + .def("set_shape", &VarDesc::SetShape) + .def("set_dtype", &VarDesc::SetDataType) + .def("shape", &VarDesc::Shape, py::return_value_policy::reference) + .def("dtype", &VarDesc::GetDataType, py::return_value_policy::reference) + .def("lod_level", &VarDesc::GetLodLevel) + .def("set_lod_level", &VarDesc::SetLoDLevel) + .def("type", &VarDesc::GetType) + .def("set_type", &VarDesc::SetType) + .def("serialize_to_string", SerializeMessage) + .def("persistable", &VarDesc::Persistable) + .def("set_persistable", &VarDesc::SetPersistable); - py::enum_(var_desc, "VarType", "") - .value("LOD_TENSOR", VarDesc::LOD_TENSOR) - .value("SELECTED_ROWS", VarDesc::SELECTED_ROWS) - .value("FEED_MINIBATCH", VarDesc::FEED_MINIBATCH) - .value("FETCH_LIST", VarDesc::FETCH_LIST) - .value("STEP_SCOPES", VarDesc::STEP_SCOPES) - .value("LOD_RANK_TABLE", VarDesc::LOD_RANK_TABLE) - .value("LOD_TENSOR_ARRAY", VarDesc::LOD_TENSOR_ARRAY); + py::enum_(var_desc, "VarType", "") + .value("LOD_TENSOR", proto::VarDesc::LOD_TENSOR) + .value("SELECTED_ROWS", proto::VarDesc::SELECTED_ROWS) + .value("FEED_MINIBATCH", proto::VarDesc::FEED_MINIBATCH) + .value("FETCH_LIST", proto::VarDesc::FETCH_LIST) + .value("STEP_SCOPES", proto::VarDesc::STEP_SCOPES) + .value("LOD_RANK_TABLE", proto::VarDesc::LOD_RANK_TABLE) + .value("LOD_TENSOR_ARRAY", proto::VarDesc::LOD_TENSOR_ARRAY); } void BindOpDesc(py::module &m) { - py::enum_(m, "AttrType", "") - .value("INT", AttrType::INT) - .value("INTS", AttrType::INTS) - .value("FLOAT", AttrType::FLOAT) - .value("FLOATS", AttrType::FLOATS) - .value("STRING", AttrType::STRING) - .value("STRINGS", AttrType::STRINGS) - .value("BOOL", AttrType::BOOLEAN) - .value("BOOLS", AttrType::BOOLEANS) - .value("BLOCK", AttrType::BLOCK); + py::enum_(m, "AttrType", "") + .value("INT", proto::AttrType::INT) + .value("INTS", proto::AttrType::INTS) + .value("FLOAT", proto::AttrType::FLOAT) + .value("FLOATS", proto::AttrType::FLOATS) + .value("STRING", proto::AttrType::STRING) + .value("STRINGS", proto::AttrType::STRINGS) + .value("BOOL", proto::AttrType::BOOLEAN) + .value("BOOLS", proto::AttrType::BOOLEANS) + .value("BLOCK", proto::AttrType::BLOCK); - py::class_ op_desc(m, "OpDesc", ""); - op_desc.def("type", &OpDescBind::Type) - .def("set_type", &OpDescBind::SetType) - .def("input", &OpDescBind::Input) - .def("input_names", &OpDescBind::InputNames) - .def("set_input", &OpDescBind::SetInput) - .def("output", &OpDescBind::Output) - .def("output_names", &OpDescBind::OutputNames) - .def("set_output", &OpDescBind::SetOutput) - .def("has_attr", &OpDescBind::HasAttr) - .def("attr_type", &OpDescBind::GetAttrType) - .def("attr_names", &OpDescBind::AttrNames) - .def("set_attr", &OpDescBind::SetAttr) - .def("attr", &OpDescBind::GetAttr) - .def("set_block_attr", &OpDescBind::SetBlockAttr) - .def("block_attr", &OpDescBind::GetBlockAttr) - .def("check_attrs", &OpDescBind::CheckAttrs) - .def("infer_shape", &OpDescBind::InferShape) - .def("infer_var_type", &OpDescBind::InferVarType) - .def("serialize_to_string", SerializeMessage); + py::class_ op_desc(m, "OpDesc", ""); + op_desc + .def("__init__", [](OpDesc &self) { new (&self) OpDesc(); }, + py::return_value_policy::reference) + .def("copy_from", &OpDesc::CopyFrom) + .def("type", &OpDesc::Type) + .def("set_type", &OpDesc::SetType) + .def("input", &OpDesc::Input) + .def("input_names", &OpDesc::InputNames) + .def("output", &OpDesc::Output) + .def("output_names", &OpDesc::OutputNames) + .def("set_input", &OpDesc::SetInput) + .def("set_output", &OpDesc::SetOutput) + .def("input_arg_names", &OpDesc::InputArgumentNames) + .def("output_arg_names", &OpDesc::OutputArgumentNames) + .def("rename_input", &OpDesc::RenameInput) + .def("rename_output", &OpDesc::RenameOutput) + .def("has_attr", &OpDesc::HasAttr) + .def("attr_type", &OpDesc::GetAttrType) + .def("attr_names", &OpDesc::AttrNames) + .def("set_attr", &OpDesc::SetAttr) + .def("attr", &OpDesc::GetAttr) + .def("set_block_attr", &OpDesc::SetBlockAttr) + .def("set_serialized_attr", + [](OpDesc &self, const std::string &name, + const py::bytes &seriralized) { + std::string ser(seriralized); + self.SetAttr(name, ser); + }) + .def("block_attr", &OpDesc::GetBlockAttr) + .def("check_attrs", &OpDesc::CheckAttrs) + .def("infer_shape", &OpDesc::InferShape) + .def("infer_var_type", &OpDesc::InferVarType) + .def("serialize_to_string", SerializeMessage); } } // namespace pybind diff --git a/paddle/pybind/pybind.cc b/paddle/pybind/pybind.cc index 3d8d3f1d2fd3977f945928c723db5fcafffeae85..04485ce7c1ab87f8655b0e6cbaecc36b3382f647 100644 --- a/paddle/pybind/pybind.cc +++ b/paddle/pybind/pybind.cc @@ -16,22 +16,21 @@ limitations under the License. */ #include // for call_once #include -#include "gflags/gflags.h" #include "paddle/framework/backward.h" #include "paddle/framework/executor.h" #include "paddle/framework/feed_fetch_method.h" #include "paddle/framework/framework.pb.h" +#include "paddle/framework/init.h" #include "paddle/framework/lod_rank_table.h" #include "paddle/framework/lod_tensor.h" #include "paddle/framework/lod_tensor_array.h" #include "paddle/framework/prune.h" #include "paddle/framework/selected_rows.h" -#include "paddle/framework/tensor_array.h" #include "paddle/operators/cond_op.h" -#include "paddle/operators/dynamic_recurrent_op.h" #include "paddle/operators/net_op.h" #include "paddle/platform/enforce.h" #include "paddle/platform/place.h" +#include "paddle/pybind/const_value.h" #include "paddle/pybind/exception.h" #include "paddle/pybind/pybind.h" #include "paddle/pybind/tensor_py.h" @@ -39,6 +38,7 @@ limitations under the License. */ #ifdef PADDLE_WITH_CUDA #include "paddle/operators/nccl/nccl_gpu_common.h" +#include "paddle/platform/cuda_profiler.h" #include "paddle/platform/gpu_info.h" #endif @@ -52,24 +52,6 @@ static size_t UniqueIntegerGenerator(const std::string &prefix) { return generators[prefix].fetch_add(1); } -std::once_flag gflags_init_flag; - -// TODO(qijun) move init gflags to init.cc -void InitGflags(std::vector &argv) { - std::call_once(gflags_init_flag, [&]() { - int argc = argv.size(); - char **arr = new char *[argv.size()]; - std::string line; - for (size_t i = 0; i < argv.size(); i++) { - arr[i] = &argv[i][0]; - line += argv[i]; - line += ' '; - } - google::ParseCommandLineFlags(&argc, &arr, true); - VLOG(1) << "Init commandline: " << line; - }); -} - bool IsCompileGPU() { #ifndef PADDLE_WITH_CUDA return false; @@ -96,8 +78,12 @@ PYBIND11_PLUGIN(core) { [](Tensor &self, const std::vector &dim) { self.Resize(make_ddim(dim)); }) + .def("set_layout", + [](Tensor &self, const std::string &layout) { + self.set_layout(StringToDataLayout(layout)); + }) .def("alloc_float", - [](Tensor &self, paddle::platform::GPUPlace &place) { + [](Tensor &self, paddle::platform::CUDAPlace &place) { self.mutable_data(place); }) .def("alloc_float", @@ -109,7 +95,7 @@ PYBIND11_PLUGIN(core) { self.mutable_data(place); }) .def("alloc_int", - [](Tensor &self, paddle::platform::GPUPlace &place) { + [](Tensor &self, paddle::platform::CUDAPlace &place) { self.mutable_data(place); }) .def("set", PyCPUTensorSetFromArray) @@ -283,16 +269,39 @@ All parameter, weight, gradient are variables in Paddle. } return ret_values; }); - m.def("prune", [](const ProgramDescBind &origin, + m.def( + "get_grad_op_desc", [](const OpDesc &op_desc, + const std::unordered_set &no_grad_set, + const std::vector &grad_sub_block) { + std::unordered_map grad_to_var; + std::vector> grad_op_descs = + framework::OpInfoMap::Instance() + .Get(op_desc.Type()) + .GradOpMaker()(op_desc, no_grad_set, &grad_to_var, + grad_sub_block); + std::vector grad_op_desc_ptrs(grad_op_descs.size()); + std::transform(grad_op_descs.begin(), grad_op_descs.end(), + grad_op_desc_ptrs.begin(), + [](std::unique_ptr &p) { return p.release(); }); + return std::make_pair(grad_op_desc_ptrs, grad_to_var); + }); + m.def("prune", [](const ProgramDesc &origin, const std::vector> &targets) { - ProgramDescBind prog_with_targets(origin); + ProgramDesc prog_with_targets(origin); for (const auto &t : targets) { prog_with_targets.MutableBlock(t[0])->Op(t[1])->MarkAsTarget(); } - ProgramDesc pruned_desc; + proto::ProgramDesc pruned_desc; Prune(*prog_with_targets.Proto(), &pruned_desc); - return new ProgramDescBind(pruned_desc); + return new ProgramDesc(pruned_desc); }); + m.def("inference_optimize", [](ProgramDesc &origin) { + proto::ProgramDesc pruned_desc; + InferenceOptimize(*(origin.Proto()), &pruned_desc); + return new ProgramDesc(pruned_desc); + }); + m.def("empty_var_name", []() { return framework::kEmptyVarName; }); + m.def("grad_var_suffix", []() { return framework::kGradVarSuffix; }); m.def_submodule( "var_names", "The module will return special predefined variable name in Paddle") @@ -306,10 +315,10 @@ All parameter, weight, gradient are variables in Paddle. return new paddle::platform::CPUDeviceContext(); }) .def_static("create", - [](paddle::platform::GPUPlace& place) + [](paddle::platform::CUDAPlace& place) -> paddle::platform::DeviceContext* { #ifndef PADDLE_WITH_CUDA - PADDLE_THROW("GPUPlace is not supported in CPU device."); + PADDLE_THROW("CUDAPlace is not supported in CPU device."); #else return new paddle::platform::CUDADeviceContext(place); #endif @@ -319,9 +328,9 @@ All parameter, weight, gradient are variables in Paddle. #ifdef PADDLE_WITH_CUDA py::class_(m, "Communicator").def(py::init<>()); #endif - py::class_(m, "GPUPlace") + py::class_(m, "CUDAPlace") .def(py::init()) - .def("__str__", string::to_string); + .def("__str__", string::to_string); py::class_(m, "CPUPlace") .def(py::init<>()) @@ -334,14 +343,14 @@ All parameter, weight, gradient are variables in Paddle. self = cpu_place; }) .def("set_place", - [](platform::Place &self, const platform::GPUPlace &gpu_place) { + [](platform::Place &self, const platform::CUDAPlace &gpu_place) { self = gpu_place; }); py::class_(m, "Operator") .def_static("create", [](py::bytes protobin) { - OpDesc desc; + proto::OpDesc desc; PADDLE_ENFORCE(desc.ParsePartialFromString(protobin), "Cannot parse user input to OpDesc"); PADDLE_ENFORCE(desc.IsInitialized(), @@ -356,10 +365,10 @@ All parameter, weight, gradient are variables in Paddle. }) .def("run", [](OperatorBase &self, const Scope &scope, - const platform::DeviceContext &dev_ctx) { - self.Run(scope, dev_ctx); - dev_ctx.Wait(); - }) + const platform::CPUPlace &place) { self.Run(scope, place); }) + .def("run", + [](OperatorBase &self, const Scope &scope, + const platform::CUDAPlace &place) { self.Run(scope, place); }) .def("type", [](const OperatorBase &op) -> std::string { return op.Type(); }) .def("outputs", @@ -390,88 +399,11 @@ All parameter, weight, gradient are variables in Paddle. self->CompleteAddOp(); }); - py::class_(m, "TensorArray") - .def("__init__", - [](TensorArray &instance) { new (&instance) TensorArray(); }) - .def("read", - [](TensorArray &self, size_t index) { return self.Read(index); }) - .def("write", [](TensorArray &self, size_t index, - LoDTensor &value) { self.Write(index, value); }) - .def("write_shared", - [](TensorArray &self, size_t index, const LoDTensor &value) { - self.WriteShared(index, value); - }) - .def("size", [](TensorArray &self) { return self.size(); }) - .def("pack", - [](TensorArray &self, size_t level, - const std::vector> &meta_info, - const std::vector> &lod) { - std::vector meta; - for (auto &info : meta_info) { - PADDLE_ENFORCE_EQ(info.size(), 3UL); - meta.emplace_back(info[0], info[1], info[2]); - } -#ifndef PADDLE_WITH_CUDA - return self.Pack(level, meta, lod); -#else - LoD new_lod; - new_lod.reserve(lod.size()); - std::copy(lod.begin(), lod.end(), std::back_inserter(new_lod)); - return self.Pack(level, meta, new_lod); -#endif - }) - .def("unpack", - [](TensorArray &self, const LoDTensor &source, int level, - bool length_descend) { - auto metas = self.Unpack(source, level, length_descend); - std::vector> meta_info; - for (auto meta : metas) { - meta_info.emplace_back( - std::vector({meta.begin, meta.end, meta.ori_idx})); - } - return meta_info; - }) - .def("stack", [](TensorArray &self) { return self.Stack(); }) - .def("unstack", - [](TensorArray &self, const LoDTensor &source) { - return self.Unstack(source); - }) - .def("unstack_shared", [](TensorArray &self, const LoDTensor &source) { - return self.UnstackShared(source); - }); - - py::class_(m, - "DynamicRecurrentOp") - .def_static("create", - [](py::bytes protobin) -> operators::DynamicRecurrentOp * { - OpDesc desc; - PADDLE_ENFORCE(desc.ParsePartialFromString(protobin), - "Cannot parse user input to OpDesc"); - PADDLE_ENFORCE(desc.IsInitialized(), - "User OpDesc is not initialized, reason %s", - desc.InitializationErrorString()); - auto rnn_op = OpRegistry::CreateOp(desc); - return static_cast( - rnn_op.release()); - }) - .def("set_step_unit", - [](operators::DynamicRecurrentOp &self, const operators::NetOp &net) - -> void { self.rnn.SetStepUnit(net.Clone()); }) - .def("get_state", - [](operators::DynamicRecurrentOp &self, const std::string &name) - -> const TensorArray & { return self.rnn.state(name); }) - .def("get_step_input", - [](operators::DynamicRecurrentOp &self, const std::string &name) - -> const TensorArray & { return self.rnn.step_input(name); }) - .def("get_step_output", - [](operators::DynamicRecurrentOp &self, const std::string &name) - -> const TensorArray & { return self.rnn.step_output(name); }); - // cond_op py::class_(m, "CondOp") .def_static("create", [](py::bytes protobin) -> operators::CondOp * { - OpDesc desc; + proto::OpDesc desc; PADDLE_ENFORCE(desc.ParsePartialFromString(protobin), "Cannot parse user input to OpDesc"); PADDLE_ENFORCE(desc.IsInitialized(), @@ -490,11 +422,12 @@ All parameter, weight, gradient are variables in Paddle. }); py::class_(m, "Executor") - .def(py::init &>()) + .def(py::init()) .def("run", &Executor::Run); m.def("unique_integer", UniqueIntegerGenerator); - m.def("init_gflags", InitGflags); + m.def("init_gflags", framework::InitGflags); + m.def("init_devices", &framework::InitDevices); m.def("is_compile_gpu", IsCompileGPU); m.def("set_feed_variable", framework::SetFeedVariable); @@ -504,6 +437,7 @@ All parameter, weight, gradient are variables in Paddle. BindBlockDesc(m); BindVarDsec(m); BindOpDesc(m); + BindConstValue(m); py::class_(m, "LodRankTable") .def("items", [](framework::LoDRankTable &table) { @@ -534,6 +468,10 @@ All parameter, weight, gradient are variables in Paddle. m.def("op_support_gpu", OpSupportGPU); #ifdef PADDLE_WITH_CUDA m.def("get_cuda_device_count", platform::GetCUDADeviceCount); + + m.def("nvprof_init", platform::CudaProfilerInit); + m.def("nvprof_start", platform::CudaProfilerStart); + m.def("nvprof_stop", platform::CudaProfilerStop); #endif return m.ptr(); diff --git a/paddle/pybind/tensor_py.h b/paddle/pybind/tensor_py.h index 41fa658502d341fe9653a3e99b58498fcaeada47..6b4290972bade585d1a0c2ae919a2e712bdf308c 100644 --- a/paddle/pybind/tensor_py.h +++ b/paddle/pybind/tensor_py.h @@ -1,21 +1,22 @@ /* 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 +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 + 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. */ +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/framework/tensor.h" #include "paddle/memory/memcpy.h" +#include "paddle/platform/device_context.h" #include "pybind11/numpy.h" #include "pybind11/pybind11.h" @@ -61,21 +62,25 @@ struct CastToPyBufferImpl { auto *src_ptr = static_cast(tensor.data()); auto *dst_ptr = static_cast(dst_tensor.mutable_data( tensor.dims(), platform::CPUPlace())); - // TODO(qijun): Here we use default CUDA stream to set GPU Tensor to - // a Python numpy array. It's better to manage CDUA stream unifiedly. - paddle::platform::GpuMemcpySync(dst_ptr, src_ptr, - sizeof(CUR_TYPE) * tensor.numel(), - cudaMemcpyDeviceToHost); + + 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()); #else - PADDLE_THROW("'GPUPlace' is not supported in CPU only device."); + PADDLE_THROW("'CUDAPlace' is not supported in CPU only device."); #endif } else if (paddle::platform::is_cpu_place(tensor.place())) { dst_tensor = tensor; } - return py::buffer_info( - dst_tensor.mutable_data(dst_tensor.place()), - sizeof(CUR_TYPE), py::format_descriptor::format(), - (size_t)framework::arity(dst_tensor.dims()), dims_outside, strides); + return py::buffer_info(dst_tensor.data(), sizeof(CUR_TYPE), + py::format_descriptor::format(), + (size_t)framework::arity(dst_tensor.dims()), + dims_outside, strides); } else { constexpr bool less = I + 1 < std::tuple_size>::value; return CastToPyBufferImpl()(tensor); @@ -123,7 +128,7 @@ template void PyCUDATensorSetFromArray( framework::Tensor &self, py::array_t array, - paddle::platform::GPUPlace &place) { + paddle::platform::CUDAPlace &place) { std::vector dims; dims.reserve(array.ndim()); for (size_t i = 0; i < array.ndim(); ++i) { @@ -132,10 +137,12 @@ void PyCUDATensorSetFromArray( self.Resize(framework::make_ddim(dims)); auto *dst = self.mutable_data(place); - // TODO(qijun): Here we use default CUDA stream to set a Python numpy - // array to a GPU Tensor. It's better to manage CDUA stream unifiedly. - paddle::platform::GpuMemcpySync(dst, array.data(), sizeof(T) * array.size(), - cudaMemcpyHostToDevice); + + 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()); } #endif diff --git a/paddle/scripts/CMakeLists.txt b/paddle/scripts/CMakeLists.txt index a52f06fe497dac467e4ef2543ebda7a423ca326d..68cb5a19f99ab5148b04d193eb2356588bdc5a59 100644 --- a/paddle/scripts/CMakeLists.txt +++ b/paddle/scripts/CMakeLists.txt @@ -5,11 +5,3 @@ configure_file(submit_local.sh.in install(FILES ${CMAKE_CURRENT_BINARY_DIR}/paddle DESTINATION bin PERMISSIONS OWNER_EXECUTE OWNER_WRITE OWNER_READ GROUP_EXECUTE GROUP_READ WORLD_EXECUTE WORLD_READ) - -configure_file(tools/usage_stat/usage.sh - paddle_usage - @ONLY) - -install(FILES ${CMAKE_CURRENT_BINARY_DIR}/paddle_usage DESTINATION opt/paddle/bin - PERMISSIONS OWNER_EXECUTE OWNER_WRITE OWNER_READ - GROUP_EXECUTE GROUP_READ WORLD_EXECUTE WORLD_READ) diff --git a/paddle/scripts/check_env.sh b/paddle/scripts/check_env.sh new file mode 100755 index 0000000000000000000000000000000000000000..af16b84ca8a18151f0fa36d39fd201d3cab21a5f --- /dev/null +++ b/paddle/scripts/check_env.sh @@ -0,0 +1,261 @@ +#!/bin/bash + +if [ "`uname -s`" != "Linux" ]; then + echo "Current scenario only support in Linux yet!" + exit 0 +fi + +echo "========================= Hardware Information =========================" +sockets=`grep 'physical id' /proc/cpuinfo | sort -u | wc -l` +cores_per_socket=`grep 'core id' /proc/cpuinfo | sort -u | wc -l` +ht=`lscpu |grep "per core" |awk -F':' '{print $2}'|xargs` +physical_cores=$((sockets * cores_per_socket)) +virtual_cores=`grep 'processor' /proc/cpuinfo | sort -u | wc -l` +numa_nodes=`lscpu |grep "NUMA node(s)"|awk -F':' '{print $2}'|xargs` +echo "CPU Name : `cat /proc/cpuinfo |grep -i "model name" |uniq |awk -F ':' '{print $2}'|xargs`" +echo "CPU Family : `lscpu |grep \"CPU family\" |awk -F':' '{print $2}'|xargs`" +echo "Socket Number : $sockets" +echo "Cores Per Socket : $cores_per_socket" +echo "Total Physical Cores : $physical_cores" +echo "Total Virtual Cores : $virtual_cores" +if [ $ht -eq 1 ]; then + echo "Hyper Threading : OFF" + if [ $physical_cores -ne $virtual_cores ]; then + echo "Error: HT logical error" + fi +else + echo "Hyper Threading : ON" + if [ $physical_cores -ge $virtual_cores ]; then + echo "Error: HT logical error" + fi +fi +echo "NUMA Nodes : $numa_nodes" +if [ $numa_nodes -lt $sockets ]; then + echo "Warning: NUMA node is not enough for the best performance,\ + at least $sockets" +fi + +echo "-------------------------- Memory Information --------------------------" +# dmidecode support start from 2.11 +dmi_ver=`dmidecode --version|awk -F '.' '{print $1}'|xargs` +if [ $dmi_ver -lt 2 ]; then + echo "Error: dmidecode unknown or version is too old" + exit 0 +fi +if [ `dmidecode | grep -ic "Permission denied"` -ne 0 ]; then + echo "Error: need root to run dmidecode" + exit 0 +fi +max_dimms=0 +num_dimms_installed=0 +for dimm_id in `dmidecode |grep Locator|sort -u | awk -F ':' '{print $2}'`; do + num_refered=`dmidecode |grep -wc "$dimm_id"` + # the actual dimm id should be refered only once + if [ $num_refered -eq 1 ]; then + num_unknown=`dmidecode | awk '/'$dimm_id'/ {s=1; f=0}; + /Unknown/ {f=1}; + /Manufacturer/ {if (s==1) {print f; exit 0;}};'` + if [ $num_unknown -eq 0 ]; then + dimms_installed="$dimms_installed \n $dimm_id" + ((num_dimms_installed++)) + else + dimms_uninstalled="$dimms_uninstalled \n $dimm_id" + fi + ((max_dimms++)) + fi +done +echo "Installed DIMM number : $num_dimms_installed" +num_dimms_mapped=`dmidecode | grep "Memory Device Mapped" | wc -l` +if [ $num_dimms_installed -ne $num_dimms_mapped ]; then + echo "Error: The installed DIMMs number does ont match the mapped memory device: $num_dimms_mapped" +fi +num_clock_configed=`dmidecode | grep -i "Configured Clock Speed" |grep -ic "Hz"` +if [ $num_dimms_installed -ne $num_clock_configed ]; then + echo "Error: The installed DIMMs number does ont match configured clocks: $num_clock_configed" +fi +echo -e "Installed DIMMs Locator: $dimms_installed" +echo -e "Not installed DIMMs : $dimms_uninstalled" +max_dimm_slots=`dmidecode | grep -c "Bank Locator"` +echo "DIMMs max slots : $max_dimm_slots" +if [ $max_dimms -ne $max_dimm_slots ]; then + echo "Error: The max dimm slots do not match the max dimms: $max_dimms" +fi +free_ver_main=`free -V|awk -F ' ' '{print $NF}'|awk -F '.' '{print $1}'` +free_ver_sub=`free -V|awk -F ' ' '{print $NF}'|awk -F '.' '{print $2}'` +if [ $free_ver_main -lt 3 ] || [ $free_ver_sub -lt 3 ]; then + mem_sz=`free |grep -i mem |awk -F' ' '{print $2}'|xargs` + swap_sz=`free |grep -i swap |awk -F' ' '{print $2}'|xargs` + total_sz=`free -t |grep -i total |tail -n 1| awk -F' ' '{print $2}'|xargs` + mem_sz="`awk 'BEGIN{printf "%.1f\n",('$mem_sz'/1024/1024)}'` GB" + swap_sz="`awk 'BEGIN{printf "%.1f\n",('$swap_sz'/1024/1024)}'` GB" + total_sz="`awk 'BEGIN{printf "%.1f\n",('$total_sz'/1024/1024)}'` GB" +else + mem_sz=`free -h |grep -i mem |awk -F' ' '{print $2}'|xargs` + swap_sz=`free -h |grep -i swap |awk -F' ' '{print $2}'|xargs` + total_sz=`free -th |grep -i total |tail -n 1| awk -F' ' '{print $2}'|xargs` +fi +echo "Memory Size : $mem_sz" +echo "Swap Memory Size : $swap_sz" +echo "Total Memory Size : $total_sz" +echo "Max Memory Capacity : `dmidecode |grep -i \"maximum capacity\"|sort -u|awk -F':' '{print $2}'|xargs`" +# DIMMs fequency +clock_speeds=`dmidecode | grep -i "Configured Clock Speed" | grep -i "Hz" |sort -u | awk -F':' '{print $2}'|xargs` +echo "Configed Clock Speed : $clock_speeds" +num_clock_type=`dmidecode | grep -i "Configured Clock Speed" | grep -i "Hz" |sort -u | wc -l` +if [ $num_clock_type -ne 1 ]; then + echo "Warning: Have more than 1 speed type, all DIMMs should have same fequency: $clock_speeds" +fi + +echo "-------------------------- Turbo Information --------------------------" +scaling_drive=`cat /sys/devices/system/cpu/cpu0/cpufreq/scaling_driver` +echo "Scaling Driver : $scaling_drive" +if [ $scaling_drive == "intel_pstate" ] && [ -e /sys/devices/system/cpu/intel_pstate/no_turbo ]; then + turbo=`cat /sys/devices/system/cpu/intel_pstate/no_turbo` + if [ $turbo -eq 1 ]; then + echo "Turbo Status : OFF" + else + echo "Turbo Status : ON" + fi +else + echo "Warning: Scaling driver is not intel_pstarte, maybe should enable it in BIOS" + echo "Turbo Status : Unknown" +fi +# cpu frequency +num_max_freq=`cat /sys/devices/system/cpu/cpu*/cpufreq/scaling_max_freq| sort -u |wc -l` +num_min_freq=`cat /sys/devices/system/cpu/cpu*/cpufreq/scaling_min_freq| sort -u |wc -l` +if [ $num_max_freq -ne 1 ]; then + echo "Error: the max_frequency of all CPU should be equal" +fi +if [ $num_min_freq -ne 1 ]; then + echo "Error: the min_frequency of all CPU should be equal" +fi +max_freq=`cat /sys/devices/system/cpu/cpu*/cpufreq/scaling_max_freq| uniq|xargs` # kHz +max_freq=`awk 'BEGIN{printf "%.2f",('$max_freq' / 1000000)}'` # GHz +min_freq=`cat /sys/devices/system/cpu/cpu*/cpufreq/scaling_min_freq| uniq|xargs` # kHz +min_freq=`awk 'BEGIN{printf "%.2f",('$min_freq' / 1000000)}'` # GHz +echo "CPU Max Frequency : $max_freq GHz" +echo "CPU Min Frequency : $min_freq GHz" +# cpu governor +num_governor=`cat /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor| sort -u |wc -l` +if [ $num_governor -ne 1 ]; then + echo "Error: the governor of all CPU should be the same" +fi +governor=`cat /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor| sort -u |uniq` +echo "CPU Freq Governor : $governor" + + +echo "========================= Software Information =========================" +echo "BIOS Release Date : `dmidecode | grep "Release Date"|awk -F ':' '{print $2}'|xargs`" +echo "OS Version : `cat /etc/redhat-release`" +echo "Kernel Release Version : `uname -r`" +echo "Kernel Patch Version : `uname -v`" +echo "GCC Version :`gcc --version | head -n 1|awk -F '\\\(GCC\\\)' '{print $2}'`" +if command -v cmake >/dev/null 2>&1; then + cmake_ver=`cmake --version | head -n 1 | awk -F 'version' '{print $2}'` +else + cmake_ver=" Not installed" +fi +echo "CMake Version :$cmake_ver" +echo "------------------ Environment Variables Information -------------------" +kmp_affinity=`env | grep KMP_AFFINITY` +omp_dynamic=`env | grep OMP_DYNAMIC` +omp_nested=`env | grep OMP_NESTED` +omp_num_threads=`env | grep OMP_NUM_THREADS` +mkl_num_threads=`env | grep MKL_NUM_THREADS` +mkl_dynamic=`env | grep MKL_DYNAMIC` +if [ ! $kmp_affinity ]; then kmp_affinity="unset"; fi +if [ ! $omp_dynamic ]; then omp_dynamic="unset"; fi +if [ ! $omp_nested ]; then omp_nested="unset"; fi +if [ ! $omp_num_threads ]; then omp_num_threads="unset"; fi +if [ ! $mkl_num_threads ]; then mkl_num_threads="unset"; fi +if [ ! $mkl_dynamic ]; then mkl_dynamic="unset"; fi +echo "KMP_AFFINITY : $kmp_affinity" +echo "OMP_DYNAMIC : $omp_dynamic" +echo "OMP_NESTED : $omp_nested" +echo "OMP_NUM_THREADS : $omp_num_threads" +echo "MKL_NUM_THREADS : $mkl_num_threads" +echo "MKL_DYNAMIC : $mkl_dynamic" +# Check if any MKL related libraries have been installed in LD_LIBRARY_PATH +for path in `echo $LD_LIBRARY_PATH | awk -F ':' '{for(i=1;i<=NF;++i)print $i}'`; do + mkldnn_found=`find $path -name "libmkldnn.so"` + if [ "$mkldnn_found" ]; then + echo "Found MKL-DNN : $mkldnn_found" + fi + mklml_found=`find $path -name "libmklml_intel.so"` + if [ "$mklml_found" ]; then + echo "Found MKLML : $mklml_found" + fi + iomp_found=`find $path -name "libiomp5.so"` + if [ "$iomp_found" ]; then + echo "Found IOMP : $iomp_found" + fi +done + +# dump all details for fully check +lscpu > lscpu.dump +dmidecode > dmidecode.dump + +# The expected result would be like: +# ========================= Hardware Information ========================= +# CPU Name : Intel(R) Xeon(R) Gold 6148M CPU @ 2.40GHz +# CPU Family : 6 +# Socket Number : 2 +# Cores Per Socket : 20 +# Total Physical Cores : 40 +# Total Virtual Cores : 40 +# Hyper Threading : OFF +# NUMA Nodes : 2 +# -------------------------- Memory Information -------------------------- +# Installed DIMM number : 12 +# Installed DIMMs Locator: +# CPU1_DIMM_A1 +# CPU1_DIMM_B1 +# CPU1_DIMM_C1 +# CPU1_DIMM_D1 +# CPU1_DIMM_E1 +# CPU1_DIMM_F1 +# CPU2_DIMM_A1 +# CPU2_DIMM_B1 +# CPU2_DIMM_C1 +# CPU2_DIMM_D1 +# CPU2_DIMM_E1 +# CPU2_DIMM_F1 +# Not installed DIMMs : +# CPU1_DIMM_A2 +# CPU1_DIMM_B2 +# CPU1_DIMM_C2 +# CPU1_DIMM_D2 +# CPU1_DIMM_E2 +# CPU1_DIMM_F2 +# CPU2_DIMM_A2 +# CPU2_DIMM_B2 +# CPU2_DIMM_C2 +# CPU2_DIMM_D2 +# CPU2_DIMM_E2 +# CPU2_DIMM_F2 +# DIMMs max slots : 24 +# Memory Size : 376G +# Swap Memory Size : 4.0G +# Total Memory Size : 380G +# Max Memory Capacity : 2304 GB +# Configed Clock Speed : 2666 MHz +# -------------------------- Turbo Information -------------------------- +# Scaling Driver : intel_pstate +# Turbo Status : ON +# CPU Max Frequency : 3.70 GHz +# CPU Min Frequency : 1.00 GHz +# CPU Freq Governor : performance +# ========================= Software Information ========================= +# BIOS Release Date : 03/10/2017 +# OS Version : CentOS Linux release 7.3.1611 (Core) +# Kernel Release Version : 3.10.0-514.el7.x86_64 +# Kernel Patch Version : #1 SMP Tue Nov 22 16:42:41 UTC 2016 +# GCC Version : 4.8.5 20150623 (Red Hat 4.8.5-11) +# CMake Version : 3.5.2 +# ------------------ Environment Variables Information ------------------- +# KMP_AFFINITY : unset +# OMP_DYNAMIC : unset +# OMP_NESTED : unset +# OMP_NUM_THREADS : unset +# MKL_NUM_THREADS : unset +# MKL_DYNAMIC : unset diff --git a/paddle/scripts/cluster_train_v2/openmpi/docker_cluster/Dockerfile b/paddle/scripts/cluster_train_v2/openmpi/docker_cluster/Dockerfile index 1a2d19e823541750830fcaa25f65b2f8e1ea2b49..c2f631bdf4ed52a5dfa3fbcf1157d0abbdeadb9b 100644 --- a/paddle/scripts/cluster_train_v2/openmpi/docker_cluster/Dockerfile +++ b/paddle/scripts/cluster_train_v2/openmpi/docker_cluster/Dockerfile @@ -1,7 +1,7 @@ # Build this image: docker build -t mpi . # -FROM paddledev/paddle:0.10.0rc3 +FROM paddlepaddle/paddle:0.10.0rc3 ENV DEBIAN_FRONTEND noninteractive diff --git a/paddle/scripts/docker/README.md b/paddle/scripts/docker/README.md index f3a6f1dba7588c6b29c1dcae26ec134c1a7f937d..f0620498cfa6775ce2949cc02fa9f6c9529dec2e 100644 --- a/paddle/scripts/docker/README.md +++ b/paddle/scripts/docker/README.md @@ -20,7 +20,7 @@ binaries. ## Run The Build -### Build Evironments +### Build Environments The pre-built build environment images are: @@ -192,7 +192,7 @@ For developers who are interested in the C++ source code, please use -e "WOBOQ=O - The following command builds PaddlePaddle, generates HTML pages from C++ source code, and writes HTML pages into `$HOME/woboq_out` on the host: ```bash -docker run -v $PWD:/paddle -v $HOME/woboq_out:/woboq_out -e "WITH_GPU=OFF" -e "WITH_AVX=ON" -e "WITH_TEST=ON" -e "WOBOQ=ON" paddlepaddle/paddle:latest-dev +docker run -v $PWD:/paddle -v $HOME/woboq_out:/woboq_out -e "WITH_GPU=OFF" -e "WITH_AVX=ON" -e "WITH_TESTING=ON" -e "WOBOQ=ON" paddlepaddle/paddle:latest-dev ``` - You can open the generated HTML files in your Web browser. Or, if you want to run a Nginx container to serve them for a wider audience, you can run: diff --git a/paddle/scripts/docker/build.sh b/paddle/scripts/docker/build.sh index 595d25fd4830b6e69b9a1080803771b0464741db..92039ec6b05d224e702f0ba5dc05c057a492287e 100644 --- a/paddle/scripts/docker/build.sh +++ b/paddle/scripts/docker/build.sh @@ -16,11 +16,13 @@ function cmake_gen() { 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" @@ -34,6 +36,7 @@ function cmake_gen() { ${PYTHON_FLAGS} -DWITH_DOC=OFF -DWITH_GPU=${WITH_GPU:-OFF} + -DWITH_DISTRIBUTE=${WITH_DISTRIBUTE:-OFF} -DWITH_MKL=${WITH_MKL:-ON} -DWITH_AVX=${WITH_AVX:-OFF} -DWITH_GOLANG=${WITH_GOLANG:-ON} @@ -55,6 +58,7 @@ EOF ${PYTHON_FLAGS} \ -DWITH_DOC=OFF \ -DWITH_GPU=${WITH_GPU:-OFF} \ + -DWITH_DISTRIBUTE=${WITH_DISTRIBUTE:-OFF} \ -DWITH_MKL=${WITH_MKL:-ON} \ -DWITH_AVX=${WITH_AVX:-OFF} \ -DWITH_GOLANG=${WITH_GOLANG:-ON} \ @@ -109,7 +113,10 @@ EOF -DWITH_SWIG_PY=ON \ -DWITH_STYLE_CHECK=OFF make -j `nproc` gen_proto_py + make -j `nproc` paddle_python make -j `nproc` paddle_docs paddle_docs_cn + make -j `nproc` print_operators_doc + paddle/pybind/print_operators_doc > doc/en/html/operators.json popd fi @@ -144,7 +151,7 @@ function gen_dockerfile() { DOCKERFILE_GPU_ENV="" DOCKERFILE_CUDNN_DSO="" if [[ ${WITH_GPU:-OFF} == 'ON' ]]; then - DOCKERFILE_GPU_ENV="ENV LD_LIBRARY_PATH /usr/lib/x86_64-linux-gnu:${LD_LIBRARY_PATH}" + DOCKERFILE_GPU_ENV="ENV LD_LIBRARY_PATH /usr/lib/x86_64-linux-gnu:\${LD_LIBRARY_PATH}" DOCKERFILE_CUDNN_DSO="RUN ln -s /usr/lib/x86_64-linux-gnu/libcudnn.so.5 /usr/lib/x86_64-linux-gnu/libcudnn.so" fi @@ -171,7 +178,7 @@ EOF # run paddle version to install python packages first RUN apt-get update &&\ ${NCCL_DEPS}\ - apt-get install -y wget python-pip && pip install -U pip && \ + apt-get install -y wget python-pip dmidecode python-tk && pip install -U pip && \ pip install /*.whl; apt-get install -f -y && \ apt-get clean -y && \ rm -f /*.whl && \ diff --git a/paddle/scripts/submit_local.sh.in b/paddle/scripts/submit_local.sh.in index d71cb84df3785008ea5793519fc26a174e1b95f7..8a352b0078d701f797f7202c85bd0e08201ac9b8 100755 --- a/paddle/scripts/submit_local.sh.in +++ b/paddle/scripts/submit_local.sh.in @@ -71,9 +71,7 @@ function threads_config() { # auto set OMP_NUM_THREADS and MKL_NUM_THREADS # according to trainer_count and total processors # only when MKL enabled - if [ "@WITH_MKL@" == "OFF" ]; then - return 0 - fi + # auto set OPENBLAS_NUM_THREADS when do not use MKL processors=`grep "processor" /proc/cpuinfo|sort -u|wc -l` trainers=`grep -Eo 'trainer_count.[0-9]+' <<< "$@" |grep -Eo '[0-9]+'|xargs` if [ -z $trainers ]; then @@ -83,12 +81,19 @@ function threads_config() { if [ $threads -eq 0 ]; then threads=1 fi - if [ -z "$OMP_NUM_THREADS" ]; then - export OMP_NUM_THREADS=$threads - fi - if [ -z "$MKL_NUM_THREADS" ]; then - export MKL_NUM_THREADS=$threads + if [ "@WITH_MKL@" == "ON" ]; then + if [ -z "$OMP_NUM_THREADS" ]; then + export OMP_NUM_THREADS=$threads + fi + if [ -z "$MKL_NUM_THREADS" ]; then + export MKL_NUM_THREADS=$threads + fi + else + if [ -z "$OPENBLAS_NUM_THREADS" ]; then + export OPENBLAS_NUM_THREADS=$threads + fi fi + } PADDLE_CONF_HOME="$HOME/.config/paddle" @@ -140,13 +145,17 @@ else: sys.exit(0) EOF -cpu_config +if [ "`uname -s`" == "Linux" ]; then + # only support on linux yet, with mac can use v2 + cpu_config +fi + # echo $KMP_AFFINITY $OMP_DYNAMIC case "$1" in "train") threads_config $@ - # echo $OMP_NUM_THREADS $MKL_NUM_THREADS + # echo $OMP_NUM_THREADS $MKL_NUM_THREADS $OPENBLAS_NUM_THREADS ${DEBUGGER} $PADDLE_BIN_PATH/paddle_trainer ${@:2} ;; "merge_model") @@ -161,9 +170,6 @@ case "$1" in "make_diagram") python -m paddle.utils.make_model_diagram ${@:2} ;; - "usage") - $PADDLE_BIN_PATH/paddle_usage ${@:2} - ;; "version") version ;; diff --git a/paddle/scripts/tools/build_docs/build_docs.sh b/paddle/scripts/tools/build_docs/build_docs.sh index c6cbbc4eef94fb2e2fc3c1ce71734fbb23fc22d7..f9bc8bf63ae9afdfca1ff660bc83e62e71f03005 100755 --- a/paddle/scripts/tools/build_docs/build_docs.sh +++ b/paddle/scripts/tools/build_docs/build_docs.sh @@ -5,4 +5,4 @@ docker run --rm \ -e "WITH_AVX=ON" \ -e "WITH_DOC=ON" \ -e "WOBOQ=ON" \ - ${1:-"paddledev/paddle:dev"} + ${1:-"paddlepaddle/paddle:latest-dev"} diff --git a/paddle/scripts/tools/usage_stat/usage.sh b/paddle/scripts/tools/usage_stat/usage.sh deleted file mode 100755 index 7dbd1f58842f50ea1df0e2476c4a493569b1dda9..0000000000000000000000000000000000000000 --- a/paddle/scripts/tools/usage_stat/usage.sh +++ /dev/null @@ -1,168 +0,0 @@ -#!/bin/bash - -ARGPARSE=`getopt -o u:vin:l:e: --long git-user:,help,dry-run,task-name:,log-file:,exit-code: -- "$@"` -KEEP_ANONYMOUS="A_USER_DOES_NOT_TELL_US" -# paddle config home dir, same as paddle -PADDLE_CONF_HOME="$HOME/.config/paddle" -# api url, mirror url(s) will be append later -PD_URLS="http://api.paddlepaddle.org/version" - -usage() -{ - echo "Usage: `basename $0` [options]" - echo "Options:" - echo " -e, --exit-code=EXIT_CODE The train/predict process's exit code" - echo " -l, --log-file=LOG_FILE_PATH Read which log file to get the duration of process" - echo " -n, --task-name=TASK_NAME The name of demo or example" - echo " -u, --git-user=GITHUB_USER provide contact info, like username or email" - echo " -v, -i Verbose output and interact with user when necessary" - echo " --help display this help message" -} - -eval set -- "${ARGPARSE}" -while true; do - case "$1" in - -l|--log-file) - log_file=$2 - shift 2 - ;; - -e|--exit-code) - exit_code=$2 - shift 2 - ;; - -u|--git-user) - github_user=$2 - shift 2 - ;; - -n|--task-name) - task=$2 - shift 2 - ;; - -v|-i) - v=1 - shift - ;; - --dry-run) - dry_run=1 - shift - ;; - --) - shift - break - ;; - --help) - usage - exit 0 - ;; - *) - echo "Invalid option $1" - usage - exit 1 - ;; - esac -done - -# parse the log_file to get the time costs -if [ -s "${log_file}" ]; then - duration=`awk 'BEGIN{day=0;last_sec=0;min_sec=0;max_sec=0;} - {if(index($2,":")==3){ - t=substr($2,1,8); - sec=day*86400+substr(t,1,2)*3600+substr(t,4,2)*60+substr(t,7,2); - if(secsec){min_sec=sec;} - if(max_sec==0 || max_sec/dev/null` - git_url=`git config --get remote.origin.url 2>/dev/null` - if [ "`echo ${git_url} | cut -b 1-19`" = "https://github.com/" ]; then - # under a git url, like https://github.com/user_xxx/proj_yyy.git - if [ "${v}" = "1" ]; then echo " from github url..."; fi - github_user=`echo ${git_url} | cut -d "/" -f 4` - if [ "${github_user}" = "PaddlePaddle" ]; then - github_user= - fi - fi - if [ -n "${git_username}" -a -z "${github_user}" ]; then - if [ "${v}" = "1" ]; then echo " from global git username..."; fi - github_user=${git_username} - fi - fi -fi -# allow user to set the user name, if it's not found -if [ -z "${github_user}" -a "${v}" = "1" ]; then - read -p "Please input your github username or email, or just return to keep this feedback anonymous:" - github_user=${REPLY} - if [ -z "${github_user}" ]; then - # empty input, consider as one anonymous user - github_user="${KEEP_ANONYMOUS}" - fi -fi -if [ -n "${github_user}" -a -z "${dry_run}" ]; then - # valid user and not in dry-run mode, then save to cache - mkdir -p ${PADDLE_CONF_HOME} - echo "${github_user}" >${PADDLE_CONF_HOME}/github_user -fi -if [ "${v}" = "1" ]; then echo "username: ${github_user}"; fi -if [ "${github_user}" = "${KEEP_ANONYMOUS}" ]; then - # anonymous user should keep the var empty. - github_user= -fi - -# read local paddle version -paddle_version=`paddle version | grep PaddlePaddle | head -n1 | cut -d " " -f 2 | cut -d "," -f 1` -if [ "${v}" = "1" ]; then echo "version:${paddle_version}"; fi - -# read local system time -system_time=`date "+%Y%m%d%H%M%S"` -if [ "${v}" = "1" ]; then echo "system time:${system_time}"; fi - -# make empty job_name as default value. -if [ -z "${task}" ]; then - task="(unknown_task)" -fi -if [ "${v}" = "1" ]; then echo "task: ${task}"; fi - -# concat the curl command -params="content={\"data_type\":\"usage\",\ -\"system_time\":${system_time},\"paddle_version\":\"${paddle_version}\",\ -\"github_user\":\"${github_user}\",\"job_name\":\"${task}\",\ -\"duration\":${duration},\"exit_code\":\"${exit_code}\"\ -}&type=1" -curl_cmd_prefix="curl -m 5 -X POST -d ${params}\ - -b ${PADDLE_CONF_HOME}/paddle.cookie -c ${PADDLE_CONF_HOME}/paddle.cookie " - -if [ "${dry_run}" = "1" ]; then - first_url=`echo ${PD_URLS} | cut -d " " -f 1` - echo "(dry-run mode)curl command: ${curl_cmd_prefix} ${first_url}" - exit 0 -else - for u in ${PD_URLS}; do - curl_cmd="${curl_cmd_prefix} ${u}" - if [ "${v}" = "1" ]; then echo "run: ${curl_cmd}"; fi - ${curl_cmd} >/dev/null 2>&1 - if [ $? -eq 0 ]; then - if [ "${v}" = "1" ]; then echo "upload OK!"; fi - exit 0 - else - if [ "${v}" = "1" ]; then echo "upload failed...try next"; fi - fi - done - if [ "${v}" = "1" ]; then echo "all urls tried but all failed...exit"; fi - exit 1 -fi diff --git a/paddle/scripts/travis/build_doc.sh b/paddle/scripts/travis/build_doc.sh index 28d82343ed32273740d0c52d0451681e43b3675e..0db8d33bbcb5278ed0dd5584b5822502b719ede9 100755 --- a/paddle/scripts/travis/build_doc.sh +++ b/paddle/scripts/travis/build_doc.sh @@ -8,7 +8,10 @@ cd $TRAVIS_BUILD_DIR/build # Compile Documentation only. cmake .. -DCMAKE_BUILD_TYPE=Debug -DWITH_GPU=OFF -DWITH_MKL=OFF -DWITH_DOC=ON make -j `nproc` gen_proto_py +make -j `nproc` paddle_python make -j `nproc` paddle_docs paddle_docs_cn +make -j `nproc` print_operators_doc +paddle/pybind/print_operators_doc > doc/en/html/operators.json # check websites for broken links linkchecker doc/en/html/index.html diff --git a/paddle/string/to_string.h b/paddle/string/to_string.h index 4f478b6a36b23bdba8ef3ddae94b3eadf18716c2..3b3bcc69a478045156225728236174fd601461dd 100644 --- a/paddle/string/to_string.h +++ b/paddle/string/to_string.h @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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 diff --git a/paddle/string/to_string_test.cc b/paddle/string/to_string_test.cc index 971484dd0c073762e99f3926576eb21b96197769..4956bd96fad5fd1decaad0a367135cb7d7ecaf6e 100644 --- a/paddle/string/to_string_test.cc +++ b/paddle/string/to_string_test.cc @@ -1,16 +1,16 @@ /* 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 +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 + 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. */ +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/string/to_string.h" #include diff --git a/paddle/testing/CMakeLists.txt b/paddle/testing/CMakeLists.txt index 4245df5ab72bf0fd67261818b307f0babdb5d685..77f84cd43bdf35ae6f54b0db2b5f720d24872878 100644 --- a/paddle/testing/CMakeLists.txt +++ b/paddle/testing/CMakeLists.txt @@ -5,4 +5,7 @@ if(WITH_TESTING) add_dependencies(paddle_test_main paddle_proto ${external_project_dependencies}) add_library(paddle_test_util STATIC TestUtil.cpp) add_dependencies(paddle_test_util paddle_proto ${external_project_dependencies}) + if(NOT MOBILE_INFERENCE) + cc_library(paddle_gtest_main SRCS paddle_gtest_main.cc DEPS init paddle_memory gtest gflags) + endif() endif() diff --git a/paddle/testing/paddle_gtest_main.cc b/paddle/testing/paddle_gtest_main.cc new file mode 100644 index 0000000000000000000000000000000000000000..108ff335bf6b920c648d4bfebbd6a40ffb6fd939 --- /dev/null +++ b/paddle/testing/paddle_gtest_main.cc @@ -0,0 +1,44 @@ +/* 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. */ + +#include + +#include "gflags/gflags.h" +#include "gtest/gtest.h" +#include "paddle/framework/init.h" +#include "paddle/memory/memory.h" + +int main(int argc, char** argv) { + std::vector new_argv; + std::string gflags_env; + new_argv.push_back(argv[0]); +#ifdef PADDLE_WITH_CUDA + new_argv.push_back( + strdup("--tryfromenv=fraction_of_gpu_memory_to_use,use_pinned_memory")); +#else + new_argv.push_back(strdup("--tryfromenv=use_pinned_memory")); +#endif + int new_argc = static_cast(new_argv.size()); + char** new_argv_address = new_argv.data(); + google::ParseCommandLineFlags(&new_argc, &new_argv_address, false); + testing::InitGoogleTest(&argc, argv); + paddle::memory::Used(paddle::platform::CPUPlace()); + std::vector devs = {"CPU"}; +#ifdef PADDLE_WITH_CUDA + paddle::memory::Used(paddle::platform::CUDAPlace(0)); + devs.push_back("GPU:0"); +#endif + paddle::framework::InitDevices(devs); + return RUN_ALL_TESTS(); +} diff --git a/paddle/trainer/CMakeLists.txt b/paddle/trainer/CMakeLists.txt index 3d471a0c01ca17cb98272159baf6d489c18824d5..72911695bd4959d73d783897b0c5e674454c30bc 100644 --- a/paddle/trainer/CMakeLists.txt +++ b/paddle/trainer/CMakeLists.txt @@ -54,7 +54,7 @@ if(WITH_TESTING) add_subdirectory(tests) endif() -if(NOT WITH_C_API) +if(NOT MOBILE_INFERENCE) add_paddle_exe(paddle_trainer TrainerMain.cpp) add_paddle_exe(paddle_merge_model MergeModel.cpp) @@ -74,7 +74,5 @@ endif() if(WITH_GOLANG) add_dependencies(paddle_trainer_lib paddle_pserver_cclient) target_link_libraries(paddle_trainer_lib paddle_pserver_cclient) - if(NOT WITH_C_API) - target_link_libraries(paddle_trainer paddle_pserver_cclient) - endif() + target_link_libraries(paddle_trainer paddle_pserver_cclient) endif(WITH_GOLANG) diff --git a/paddle/trainer/Trainer.cpp b/paddle/trainer/Trainer.cpp index 88e684849df6fbfe4042b92bdb76ef98159eecea..3e4a2b5fa8a3981f6362edc1dc61ae1616e257ef 100644 --- a/paddle/trainer/Trainer.cpp +++ b/paddle/trainer/Trainer.cpp @@ -138,7 +138,7 @@ void Trainer::init(const std::shared_ptr& config, } if (FLAGS_use_mkldnn) { - CHECK_EQ(FLAGS_trainer_count, 1UL) << "MKLDNN only need 1 trainer"; + CHECK_EQ(FLAGS_trainer_count, 1) << "MKLDNN only need 1 trainer"; } if (testing) { diff --git a/paddle/trainer/tests/CMakeLists.txt b/paddle/trainer/tests/CMakeLists.txt index 80665551ec51214d90b866f0c7b2abb2fdee5f39..bd518d8598f5aa7c32298ed2110a96a2743536b3 100644 --- a/paddle/trainer/tests/CMakeLists.txt +++ b/paddle/trainer/tests/CMakeLists.txt @@ -1,20 +1,17 @@ -################# test_Compare ############################ -add_unittest_without_exec(test_Compare - test_Compare.cpp) -add_test(NAME test_Compare - COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python - ${CMAKE_CURRENT_BINARY_DIR}/test_Compare - WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/) +set(PYTHON_PATH + ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d + ${PADDLE_SOURCE_DIR}/python/:${PADDLE_SOURCE_DIR}/paddle/trainer/tests) +function(trainer_test TARGET) + add_unittest_without_exec(${TARGET} ${TARGET}.cpp) + add_test(NAME ${TARGET} + COMMAND ${PYTHON_PATH} ${CMAKE_CURRENT_BINARY_DIR}/${TARGET} + WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/) +endfunction() -################# test_Trainer ########################### -add_unittest_without_exec(test_Trainer - test_Trainer.cpp) -add_test(NAME test_Trainer - COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python/ - ${PYTHON_EXECUTABLE} ${PADDLE_SOURCE_DIR}/paddle/trainer/tests/gen_proto_data.py && - ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python/ - ${CMAKE_CURRENT_BINARY_DIR}/test_Trainer - WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/) +trainer_test(test_Compare) +trainer_test(test_PyDataProviderWrapper) +trainer_test(test_recurrent_machine_generation) +trainer_test(test_Trainer) ############### test_TrainerOnePass ########################## if(WITH_PYTHON) @@ -23,32 +20,13 @@ if(WITH_PYTHON) add_unittest_without_exec(test_TrainerOnePass test_TrainerOnePass.cpp) add_test(NAME test_TrainerOnePass - COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d - ${PADDLE_SOURCE_DIR}/python/:${PADDLE_SOURCE_DIR}/paddle/trainer/tests - ${PADDLE_SOURCE_DIR}/paddle/.set_port.sh -p port ${CMAKE_CURRENT_BINARY_DIR}/test_TrainerOnePass + COMMAND ${PYTHON_PATH} ${PADDLE_SOURCE_DIR}/paddle/.set_port.sh -p port + ${CMAKE_CURRENT_BINARY_DIR}/test_TrainerOnePass WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/) endif() -################# test_recurrent_machine_generation ############### -add_unittest_without_exec(test_recurrent_machine_generation - test_recurrent_machine_generation.cpp) -add_test(NAME test_recurrent_machine_generation - COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python/ - ${CMAKE_CURRENT_BINARY_DIR}/test_recurrent_machine_generation - WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/) - -#################### test_PyDataProviderWrapper ######################### -add_unittest_without_exec(test_PyDataProviderWrapper - test_PyDataProviderWrapper.cpp) - -add_test(NAME test_PyDataProviderWrapper - COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d - ${PADDLE_SOURCE_DIR}/python/:${PADDLE_SOURCE_DIR}/paddle/trainer/tests - ${CMAKE_CURRENT_BINARY_DIR}/test_PyDataProviderWrapper - WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/) - #################### test_config_parser ######################### add_test(NAME test_config_parser - COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python/ - ${PYTHON_EXECUTABLE} ${PADDLE_SOURCE_DIR}/paddle/trainer/tests/config_parser_test.py + COMMAND ${PYTHON_PATH} ${PYTHON_EXECUTABLE} + ${PADDLE_SOURCE_DIR}/paddle/trainer/tests/config_parser_test.py WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/) diff --git a/paddle/trainer/tests/chunking.conf b/paddle/trainer/tests/chunking.conf deleted file mode 100644 index d88df919df8fee9209336ffa29d724dabe6af31b..0000000000000000000000000000000000000000 --- a/paddle/trainer/tests/chunking.conf +++ /dev/null @@ -1,125 +0,0 @@ -#edit-mode: -*- python -*- -# 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. - -#Todo(luotao02) This config is only used for unitest. It is out of date now, and will be updated later. - -TrainData(ProtoData( - files = 'trainer/tests/train_files.txt', - usage_ratio = 1.0, -)) - -TestData(ProtoData( - files = 'trainer/tests/test_files.txt' -)) - -default_initial_std(1) -default_decay_rate(4e-4) -default_device(0) - -Inputs("features", "word", "pos", "chunk") - -Outputs("crf") - -Layer( - name = "features", - type = "data", - size = 4339, -) - -Layer( - name = "word", - type = "data", - size = 478, -) - -Layer( - name = "pos", - type = "data", - size = 45 -) - -Layer( - name = "chunk", - type = "data", - size = 23 -) - -Layer( - name = "output", - type = "mixed", - size = 23, - bias = False, - device = -1, - inputs = [ - FullMatrixProjection("features", parameter_name="feature_weights"), - # TableProjection("word"), - # TableProjection("pos"), - ], -) - -Layer( - name = "crf", - type = "crf", - size = 23, - device = -1, - inputs = [ - Input("output", parameter_name="crfw"), - "chunk" - ] -) - -Layer( - name = "crf_decoding", - type = "crf_decoding", - size = 23, - device = -1, - inputs = [ - Input("output", parameter_name="crfw"), - "chunk" - ] -) - -Evaluator( - name = "error", - type = "sum", - inputs = "crf_decoding", -) - -''' -# chuck evaluator cannot be used for GPU training -Evaluator( - name = "chunk_f1", - type = "chunk", - inputs = ["crf_decoding", "chunk"], - chunk_scheme = "IOB", - num_chunk_types = 11, -) -''' - -Settings( - algorithm = 'sgd', - batch_size = 100, - average_window = 0.5, - max_average_window = 2500, - learning_rate = 1e-1, - learning_rate_decay_a = 5e-7, - learning_rate_decay_b = 0.75, - l1weight = 0, - l2weight = 1, - c1 = 0.0001, - backoff = 0.5, - owlqn_steps = 100, - max_backoff = 5, -) diff --git a/paddle/trainer/tests/compare_sparse_data b/paddle/trainer/tests/compare_sparse_data deleted file mode 100644 index 18fc6541383d8e8e1687b8fe1abd57aece3d4cfc..0000000000000000000000000000000000000000 Binary files a/paddle/trainer/tests/compare_sparse_data and /dev/null differ diff --git a/paddle/trainer/tests/data_bin_part b/paddle/trainer/tests/data_bin_part deleted file mode 100644 index 66ede391b0cffe6bc9611d3616b7b626864f5c3e..0000000000000000000000000000000000000000 --- a/paddle/trainer/tests/data_bin_part +++ /dev/null @@ -1,214 +0,0 @@ -F -X -X -X -X -X -X -X -X -HC=TFTIַ;H=TFTIYW.8T˔I͚48TN8TE98TW8T&6ͅTTHC=TFTIַ;><.8˔I͚48+E98W8&68H=TFTIHC=TFTIַ;H=TFTI86HC=TFTIַ;W8T;8TJJ8T&$H=TFTIW8Ю+JJ8HC=TFTIַ;H=TFTI HC=TFTIַ;@?H=TFTI@HC=TFTIַ;H=TFTI868T8T&9C6HC=TFTIַ;BT&$88&Ӗ5H=TFTIBTHC=TFTIַ;H=TFTIVTHC=TFTIַ;8T8TͅTT8T&86;8T@N8T8T;9H=TFTI8888&86;8@N88HC=TFTIַ;H=TFTIMKHC=TFTIַ;ٟ@17ȣ8Gȣ8/>7;BAUQUT0A?H=TFTIٟ@17G/>7;BAUQUT0HC=TFTIַ;H=TFTIHC=TFTIַ;H=TFTIHC=TFTIַ;H=TFTI.8T˔I͚48TN8TE98TW8T&6ͅTTHC=TFTIַ;'JA-EJ@8T-Eބ248TYW.8˔I͚48+E98W8&68H=TFTIAM18Mބ248HC=TFTIַ;H=TFTIYW.8T˔I͚48TN8TE98TW8T&6ͅTTHC=TFTIַ;><.8˔I͚48+E98W8&68H=TFTIHC=TFTIַ;H=TFTI HC=TFTIַ;@KH=TFTI@KHC=TFTIַ;H=TFTI HC=TFTIַ;@?H=TFTI@HC=TFTIַ;H=TFTI#!14UƕT6.Q8T@Ԛ<14ƕT6.Q8@Ԛ<HC=TFTIַ;H=TFTIVTHC=TFTIַ;8T8TͅTT8T&86;8T@N8T8T;9H=TFTI8888&86;8@N88HC=TFTIַ;H=TFTIHC=TFTIַ;ܥ6H=TFTIܥ6HC=TFTIַ;H=TFTIHC=TFTIַ;H=TFTIHC=TFTIַ;H=TFTI;9HC=TFTIַ;Q;B !H=TFTIQBHC=TFTIַ;H=TFTIYW.8T˔I͚48TN8TE98TW8T&6ͅTTHC=TFTIַ;><.8˔I͚48+E98W8&68H=TFTIHC=TFTIַ;H=TFTI53HW8T;8T8THC=TFTIַ;#!HW8Ю+8H=TFTIHC=TFTIַ;H=TFTI HC=TFTIַ;@?H=TFTI@HC=TFTIַ;H=TFTI&$HC=TFTIַ;VGD; H=TFTIVGD;  ̣ OG  ̣ OG&$Eʌ3OXMQ̣ Jʌ3D4T#!Eʌ3OXMQ̣ Jʌ3UT  ̣ OG  ̣ OGG͡S<%&б ̣ Fۧ11ņAǧ1ņAņA<6ҥ3߫UVKTVU6>VMUF>M5%̋'wuG͡S<% ̣ Fۧ11ņAǧ1ņAņA<6UVKTV6>VMUF>ʶM%̋'  ̣ OG  ̣ OG&$Eʌ3OXMQ̣ Jʌ3D4T#!Eʌ3OXMQ̣ Jʌ3UT  ̣ OG  ̣ OG̣ '@@@  @@  ̣ OG  ̣ OG&$Eʌ3OXMQ̣ Jʌ3D4T#!Eʌ3OXMQ̣ Jʌ3UT  ̣ OG  ̣ OG&$O4=ӪN/>K/;8,T O4=ӪN/>K;,T  ̣ OG  ̣ OG&$Eʌ3OXMQ̣ Jʌ3D4T#!Eʌ3OXMQ̣ Jʌ3UT  ̣ OG  ̣ OG><,9O8.̣ TB0O!./WDSW53,9O8.TB0O!./WDSW  ̣ OG  ̣ OG&$Eʌ3OXMQ̣ Jʌ3D4T#!Eʌ3OXMQ̣ Jʌ3UT  ̣ OG  ̣ OG:=X̣ QUTG܂=X̣ QTG  ̣ OG  ̣ OG&$Eʌ3OXMQ̣ Jʌ3D4T#!Eʌ3OXMQ̣ Jʌ3UT  ̣ OG  ̣ OG)'= 0̣ M6ͅTO,@Ԛ<#!=ؐ0̣ M6ͅTO,@Ԛ<  ̣ OG  ̣ OG&$Eʌ3OXMQ̣ Jʌ3D4T#!Eʌ3OXMQ̣ Jʌ3UT  ̣ OG  ̣ OG/-= 0̣ M6ͅTO,DSDA)'=ؐ0̣ M6ͅTO,DSDA  ̣ OG  ̣ OG&$Eʌ3OXMQ̣ Jʌ3D4T#!Eʌ3OXMQ̣ Jʌ3UT  ̣ OG  ̣ OG  ̣ Ҧ)GG4Q>.ŞGGщQ4Q>.GщQ 4Q.ŞG6P6T4Q.6P64Q>.ŞGGщQ4Q>.GщQ204AQ.ŞGщQHAVTJD8DAP&$4AQ.щQHAVTD8A4Q>.ŞGGщQ4Q>.GщQ 4Q.ŞG6P6T4Q.6P64Q>.ŞGGщQ4Q>.GщQ&$R4Q>.ŞGGщQ6?@Ԛ<#!R4Q>.GщQ6?@Ԛ<4Q>.ŞGGщQ4Q>.GщQ 4Q.ŞG6P6T4Q.6P64Q>.ŞGGщQ4Q>.GщQ&$4Q.ŞGJIGщQDSDA#!4Q.JIGщQDSDA4Q>.ŞGGщQ4Q>.GщQ 4Q.ŞG6P6T4Q.6P64Q>.ŞGGщQ4Q>.GщQ&$.ŞGٟ@6G5IGщQA7B.ٟ@6G5IGщQ+4Q>.ŞGGщQ4Q>.GщQ 4Q.ŞG6P6T4Q.6P64Q>.ŞGGщQ4Q>.GщQ534Q>.ŞGDAP;0T?6T)! 4Q>.A;T6T)4Q>.ŞGGщQ4Q>.GщQ 4Q.ŞG6P6T4Q.6P64Q>.ŞGGщQ4Q>.GщQ534Q>.ŞGDAP;0T?6T)! 4Q>.A;T6T)4Q>.ŞGGщQ4Q>.GщQ 4Q.ŞG6P6T4Q.6P64Q>.ŞGGщQ4Q>.GщQ><49KQ.ŞGRGD9HOKJA.ŞG=RJ/-4-Q.RGD9HKJA.RJ4Q>.ŞGGщQ4Q>.GщQ 4Q.ŞG6P6T4Q.6P64Q>.ŞGGщQ4Q>.GщQ534AIQ.ŞGщQHAVTJD8DAP)'4AIQ.щQHAVTD8A4Q>.ŞGGщQ4Q>.GщQ 4Q.ŞG6P6T4Q.6P64Q>.ŞGGщQ4Q>.GщQ/-4=R4Q>AE.ŞGC/W99 4R4Q>C/W9CPH5CPH5;9H91GRFP.ܤKHUA6)ʪ86H1GRFP.ܤKHUA6)ʪCPH5CPH5UPH>G@Ԛ<UPH>G@Ԛ<CPH5CPH5&$CPHA>GDSPԮK߀3#!CPHA>GDSPٮKCPH5CPH5AHACPG@Ԛ<AHACP@Ԛ<CPH5CPH5;9H91GRFP.ܤKHUA6)ʪ86H1GRFP.ܤKHUA6)ʪCPH5CPH5MKHFșK>7QKH.CQR>“JMB>WMLG,@Ԛ<MKHFșK>7QKH.CQR>“JMB>WMLG,@Ԛ<CPH5CPH5&$CPHA>GDSPԮK߀3#!CPHA>GDSPٮKCPH5CPH553AHMDP58Qٟ@H3/A@@@/-AHMDP8Qٟ@H3/A@@CPH5CPH5;9H91GRFP.ܤKHUA6)ʪ86H1GRFP.ܤKHUA6)ʪCPH5CPH5#!AHACPGDSDA AHACPDSDACPH5CPH5&$CPHA>GDSPԮK߀3#!CPHA>GDSPٮKCPH5CPH5YWI==R>H//GM>ϪJRK22U׵AHTUA6)ʪYWI==R>H//GM>ϪJRK22U׵AHTUA6)ʪCPH5CPH5;9H91GRFP.ܤKHUA6)ʪ86H1GRFP.ܤKHUA6)ʪCPH5CPH5 6PH>5HOAB 6PH>5HOABCPH5CPH5&$CPHA>GDSPԮK߀3#!CPHA>GDSPٮKCPH5CPH5HG22A@@@HG22A@@ B߹-O B߹-O߹-BTCO@L:߹-BCO@L: B߹-O B߹-O20 N߹-7BO1ַ;L߹-NA7OIַ;)' N߹-7BO1;߹-NA7I B߹-O B߹-O߹-BTCO@L:߹-BCO@L: B߹-O B߹-O,* N߹-BO߹-7O߹-ַ;OʈF<4)' N߹-BO߹-7߹-ַ;OʈF<4 B߹-O B߹-O߹-BTCO@L:߹-BCO@L: B߹-O B߹-O&$A N߹-BO>8ֽHٟ@@Ԛ<#!A N߹-BO>8ٟ@@Ԛ< B߹-O B߹-O߹-BTCO@L:߹-BCO@L: B߹-O B߹-O/- - N߹-C7FBOR1:?T)' - Nں-7BOR1:?T B߹-O B߹-O߹-BTCO@L:߹-BCO@L: B߹-O B߹-O ߹-7O߹-BT ߹-7߹-B B߹-O B߹-O߹-BTCO@L:߹-BCO@L: B߹-O B߹-O/- N߹-BO7FOO?L߹-OǧBT)' N߹-BO7OO?L߹-OT B߹-O B߹-O߹-BTCO@L:߹-BCO@L: B߹-O B߹-O><߹- NLB7FOQӮDDA40AT(",*߹- NLOQӮDDA0AT B߹-O B߹-O߹-BTCO@L:߹-BCO@L: B߹-O B߹-O,* ߹-7BOİU1>CBBUQ4,* ߹-7BOİU1>CBBUQ4 L17A¶7J/ L17NJ/GE/1RLA¶7CʡH =;>W=ѾC -:K48?:T86/1LNCʡH =.=ѾC -:48?:T L17A¶7J/ L17NJ/>< - /@ʡH9H1RLA¶7/JDO8,T#!N91LN/JD,T L17A¶7J/ L17NJ/GE/1RLA¶7CʡH =;>W=ѾC -:K48?:T86/1LNCʡH =.=ѾC -:48?:T L17A¶7J/ L17NJ/b`1RLDA¶7/ - J0EKB8//OEKю2E,/WT)ʪDB1LDN/J0KB8/OEю2E)ʪ L17A¶7J/ L17NJ/GE/1RLA¶7CʡH =;>W=ѾC -:K48?:T86/1LNCʡH =.=ѾC -:48?:T L17A¶7J/ L17NJ/20 - 1RLA¶7/J0EO@K&$1LN/J0EO@K L17A¶7J/ L17NJ/GE/1RLA¶7CʡH =;>W=ѾC -:K48?:T86/1LNCʡH =.=ѾC -:48?:T L17A¶7J/ L17NJ/>T7O=P; >7=P L17A¶7J/ L17NJ/GE/1RLA¶7CʡH =;>W=ѾC -:K48?:T86/1LNCʡH =.=ѾC -:48?:T L17A¶7J/ L17NJ//-DA¶7/1RLJʡHWWT%! DN/1LJʡHWWՄO L17A¶7J/ L17NJ/GE/1RLA¶7CʡH =;>W=ѾC -:K48?:T86/1LNCʡH =.=ѾC -:48?:T L17A¶7J/ L17NJ/>< - N1RLA¶7CH231RLA¶7//&$N1LN޻/231LN/ L17A¶7J/ L17NJ/GE/1RLA¶7CʡH =;>W=ѾC -:K48?:T86/1LNCʡH =.=ѾC -:48?:T L17A¶7J/ L17NJ//- LGR1¶7/17>>G>GW=ѾC -:K48?:T86/1LNCʡH =.=ѾC -:48?:T L17A¶7J/ L17NJ/JHA¶7/C1RLH7/N=,::84SQH9T86N/C1L+N=,ў84SQH9T L17A¶7J/ L17NJ/GE/1RLA¶7CʡH =;>W=ѾC -:K48?:T86/1LNCʡH =.=ѾC -:48?:T L17A¶7J/ L17NJ/DB - /@ʡH9H1RLA¶7/JDOEJ< NT΂:8/CT΂:KT΂:WJT΂:ì,UWJ&$ NTCT:Tژ< NT΂:8/CT΂:KT΂:WJT΂:ì,UWJ&$ NTCT:TژBDJ99щQ#!#HK9>BDJ99щQ&$#%9T@A6WDPDA #9@A6WDPDA)'#%HK9T>BDJ99щQ#!#HK9>BDJ99щQ\ZRBDJ99щQ#!#HK9>BDJ99щQ&$#%9T@A6WDPDA #9@A6WDPDA)'#%HK9T>BDJ99щQ#!#HK9>BDJ99щQ;9>R>%B>ڜ>A9TK91A#%@@@20>R>%B>ڜ>A9K91A#@@)'#%HK9T>BDJ99щQ#!#HK9>BDJ99щQ&$#%9T@A6WDPDA #9@A6WDPDA)'#%HK9T>BDJ99щQ#!#HK9>BDJ99щQ#!#%9TKڜ>BEIUT#9Kڜ>BEIU)'#%HK9T>BDJ99щQ#!#HK9>BDJ99щQ&$#%9T@A6WDPDA #9@A6WDPDA)'#%HK9T>BDJ99щQ#!#HK9>BDJ99щQ#!#%K9TD06O@Ԛ<#K9D06@Ԛ<)'#%HK9T>BDJ99щQ#!#HK9>BDJ99щQ&$#%9T@A6WDPDA #9@A6WDPDA)'#%HK9T>BDJ99щQ#!#HK9>BDJ99щQ#%9TCۚK@Ԛ<#9CۚK@Ԛ<)'#%HK9T>BDJ99щQ#!#HK9>BDJ99щQ&$#%9T@A6WDPDA #9@A6WDPDA)'#%HK9T>BDJ99щQ#!#HK9>BDJ99щQGE6W#%>9T?#%6O/OO/U!'B8>ڜ>;96W#>9?#6O/O/U!'B8>ڜ>)'#%HK9T>BDJ99щQ#!#HK9>BDJ99щQ&$#%9T@A6WDPDA #9@A6WDPDA)'#%HK9T>BDJ99щQ#!#HK9>BDJ99щQYW#%9T>K-A96TWB:OSRQ9#%ѾCHTL6LTJH#9>KA96TWB:OSRQ9#%5L6LT,*E6FA6ܤKJV8=B>S,,*E6FA6ܤKJV8=B>S, ؓ =BܤKS/C8Tœ =BܤKS8T,*E6FA6ܤKJV8=B>S,,*E6FA6ܤKJV8=B>S,DBGDG>W-3M8F=Bٟ@6S9ܤKȟN U686GDG>W38F=B5S9ܤKȟN U,*E6FA6ܤKJV8=B>S,,*E6FA6ܤKJV8=B>S,>THH8@9FFSA@Ԛ<53ER=B67>HH8@9FFA@Ԛ<,*E6FA6ܤKJV8=B>S,,*E6FA6ܤKJV8=B>S,PN84C81=BRVT6CAE/:6LUUNԛL@;6GDB8C81=BRVTCAE:6LUUNԛL@6G,*E6FA6ܤKJV8=B>S,,*E6FA6ܤKJV8=B>S,JHH=B/-8>ܤKDA9=S˱U8QTָUJ)ʪDBH=B/8>ܤKDA9=S˱U8QTU)ʪ,*E6FA6ܤKJV8=B>S,,*E6FA6ܤKJV8=B>S,GEABRBE9A6BϜ>8=B6ץRRDO6ө ۆ ;9ABRBE9A6BϜ>8=B6ץR6ө ,*E6FA6ܤKJV8=B>S,,*E6FA6ܤKJV8=B>S,;9RQSAEM8=B>ץR9)NU6!GJ53RQSAEC=B>ץR9)NU6!1,*E6FA6ܤKJV8=B>S,,*E6FA6ܤKJV8=B>S,/-VJV18=BR6?#%@@@)'VJV18=BR6?#@@,*E6FA6ܤKJV8=B>S,,*E6FA6ܤKJV8=B>S,D>EȊ56RT8JF=BKT:8J=BRFK,34DH@CӽDҾWK?>S@99ISDPDAzxD>Eˊ5RT8S=BАT:8J=BRF,34DH@CӽDҾWK?>S@9ISDPDA,*E6FA6ܤKJV8=B>S,,*E6FA6ܤKJV8=B>S,V68BXʉ5=B>ܤK%&Ξ)ʉ5VTVEXGVXGV8G&Ξ)VEBVƔ>XVU8—P=ۚKC>JU̟KO4>LV68BX=B>ܤK%&Ξ)ʉ5VTVEXVXV8G&Ξ)VEBV۔>VU8=CJ.4>HD6߻WXHD6߻WXC߻WX@Ԛ<C߻WX@Ԛ<HD6߻WXHD6߻WXDCGR@NDCG@NHD6߻WXHD6߻WXC߻WX@Ԛ<C߻WX@Ԛ<HD6߻WXHD6߻WX#!6CGDʉ5>R#!6CGDʉ5>RHD6߻WXHD6߻WXC߻WX@Ԛ<C߻WX@Ԛ<HD6߻WXHD6߻WX86GR>RP>R699VADSDA20GR>RP>R69VADSDAHD6߻WXHD6߻WXC߻WX@Ԛ<C߻WX@Ԛ<HD6߻WXHD6߻WX#!DR߻W99@@@DR߻W99@@HD6߻WXHD6߻WXC߻WX@Ԛ<C߻WX@Ԛ<HD6߻WXHD6߻WXUV1;2X4UV1;2XHD6߻WXHD6߻WXC߻WX@Ԛ<C߻WX@Ԛ<HD6߻WXHD6߻WXnl>A6߻W$6XT6/ҥ3)T:6X-6ME@EU%!)!MK>A6߻W$6‰XɺRҥ3?:6X-6E@E )PHD6߻WXHD6߻WXC߻WX@Ԛ<C߻WX@Ԛ<HD6߻WXHD6߻WXA?6=C߻WED>3K֟MȬTT(#$!,*6=C߻WED>3K֟MȬTT HD6߻WXHD6߻WXC߻WX@Ԛ<C߻WX@Ԛ<HD6߻WXHD6߻WX;966GȂ3ʉ5>R>BCT6;3D5366GȂ3ʉ5>R>BCT;3DHD6߻WXHD6߻WXC߻WX@Ԛ<C߻WX@Ԛ<HD6߻WXHD6߻WXDC߻WR1@KDCW1@K,*SPKO—P=D9RB5966#!SPO=DRB5966@@@,*SPKO—P=D9RB5966#!SPO=DRB5966GE6/KOٟ@—P=>8E9RBHAVTJD8DAP536/Oٟ@=>8ERBHAVTD8A,*SPKO—P=D9RB5966#!SPO=DRB5966@@@,*SPKO—P=D9RB5966#!SPO=DRB5966&$CKOI9RB2SCI9COIRB2SC9,*SPKO—P=D9RB5966#!SPO=DRB5966@@@,*SPKO—P=D9RB5966#!SPO=DRB5966)'LPKO9RB6P6T LPORB6P6,*SPKO—P=D9RB5966#!SPO=DRB5966@@@,*SPKO—P=D9RB5966#!SPO=DRB5966PN6KO9RBEIT6>SK?KI—P=>KI90C9T><6ORBEIT6>SK?K=>K90CT,*SPKO—P=D9RB5966#!SPO=DRB5966@@@,*SPKO—P=D9RB5966#!SPO=DRB5966,*кBPKOK=9F9RHG8T#!кBPOK=9FRHG8,*SPKO—P=D9RB5966#!SPO=DRB5966@@@,*SPKO—P=D9RB5966#!SPO=DRB5966JHHKO>6/—P=9RH>DAP;0T?6T)!/-HO>6/=RH>A;T6T),*SPKO—P=D9RB5966#!SPO=DRB5966@@@,*SPKO—P=D9RB5966#!SPO=DRB5966MKKO6/—P=KORDB6OKKO696KO6щQ@Ԛ<53O6/=ORDB6KO9O6щQ@Ԛ<,*SPKO—P=D9RB5966#!SPO=DRB5966@@@,*SPKO—P=D9RB5966#!SPO=DRB5966,*6/KO9RBDǬP/-C9AT0?9-8ٟ@6EE>PC9AT0?=C9AT0?=#!C9AT0?9-8@Ԛ<#!C9AT0?9-8@Ԛ<C9AT0?=C9AT0?=20ʻ?0?9<9=C9ATVB$/?BRÙKBTA?D>0?9<9=C9ATVB$/BEBC9AT0?=C9AT0?=#!C9AT0?9-8@Ԛ<#!C9AT0?9-8@Ԛ<C9AT0?=C9AT0?=0?6TU7ח>\ZHS=HˮD>7KOUJҲ.щQHT-:66(UʡH966SQHS=HˮD>7KOUJҲ.щQHT:6(UʡH966RTU7ח>6TU7ח>B7Uח>D6@Ԛ<B7Uח>D6@Ԛ<6RTU7ח>6TU7ח>\ZHS=HˮD>7KOUJҲ.щQHT-:66(UʡH966SQHS=HˮD>7KOUJҲ.щQHT:6(UʡH966RTU7ח>6TU7ח> Uח>@K Uח>@K6RTU7ח>6TU7ח>\ZHS=HˮD>7KOUJҲ.щQHT-:66(UʡH966SQHS=HˮD>7KOUJҲ.щQHT:6(UʡH966RTU7ח>6TU7ח>B7Uח>ͦBOERB7Uח>ͦBOER6RTU7ח>6TU7ח>\ZHS=HˮD>7KOUJҲ.щQHT-:66(UʡH966SQHS=HˮD>7KOUJҲ.щQHT:6(UʡH966RTU7ח>6TU7ח>B7Uח>8;BٖTTB7Uח>8;BT6RTU7ח>6TU7ח>\ZHS=HˮD>7KOUJҲ.щQHT-:66(UʡH966SQHS=HˮD>7KOUJҲ.щQHT:6(UʡH966RTU7ח>6TU7ח>86AHFS=@=՞RU70ח>GDSPԮK߀320AHF=@=՞RU70ח>GDSPٮK6RTU7ח>6TU7ח>\ZHS=HˮD>7KOUJҲ.щQHT-:66(UʡH966SQHS=HˮD>7KOUJҲ.щQHT:6(UʡH966RTU7ח>6TU7ח> B7Uח>DT("B7Uח>DT6RTU7ח>6TU7ח>\ZHS=HˮD>7KOUJҲ.щQHT-:66(UʡH966SQHS=HˮD>7KOUJҲ.щQHT:6(UʡH966RTU7ח>6TU7ח>6RTU7HˮDDA6TU7HˮDDA6RTU7ח>6TU7ח>\ZHS=HˮD>7KOUJҲ.щQHT-:66(UʡH966SQHS=HˮD>7KOUJҲ.щQHT:6(UʡH966RTU7ח>6TU7ח> B7Uח> B7Uח>6RTU7ח>6TU7ח>\ZHS=HˮD>7KOUJҲ.щQHT-:66(UʡH966SQHS=HˮD>7KOUJҲ.щQHT:6(UʡH966RTU7ח>6TU7ח>20AHFS=@=՞RU70ח>GPB6,*AHF=@=՞RU70ח>GPB  ;GB;9ӱQL4ߩ75Q-<>;G  48@@@<ߩ7>48@@  F9Q?WɤKIԊX>F9Q?WɤK 3ϊXQK  ԊXQK,*3ϊX17Q7G/׆N8GF̛<ԊX13G/NGF 3ϊXQK  ԊXQK&$R3ϊX46߻WLQG8@Ԛ< RԊX46߻WLQG@Ԛ< 3ϊXQK  ԊXQK,*3ϊX17Q7G/׆N8GF̛<ԊX13G/NGF 3ϊXQK  ԊXQK><3ϊXR7Q7@475@:ȥB@AT/-ԊXR3@475@:ȥBA 3ϊXQK  ԊXQK,*3ϊX17Q7G/׆N8GF̛<ԊX13G/NGF 3ϊXQK  ԊXQK"!F>"FN߀3/ڶ>F7N߀3/ڶ>F7,*DN.ی'79Ԛ<=/ڶ>J7@Ԛ<)'DN.ی'71=/ڶ>J7@Ԛ<N߀3/ڶ>F7N߀3/ڶ>F720DN/EL>7Aڶ>F7CDƹ;@Ԛ<,*DNȜML>7Aڶ>F7C4@Ԛ<N߀3/ڶ>F7N߀3/ڶ>F7,*DN.ی'79Ԛ<=/ڶ>J7@Ԛ<)'DN.ی'71=/ڶ>J7@Ԛ<N߀3/ڶ>F7N߀3/ڶ>F7)'HN/KN/ڶ>F7=A7B#!HN/KN/ڶ>F7=+N߀3/ڶ>F7N߀3/ڶ>F7,*DN.ی'79Ԛ<=/ڶ>J7@Ԛ<)'DN.ی'71=/ڶ>J7@Ԛ<N߀3/ڶ>F7N߀3/ڶ>F7 H/67—P=DG@KH/67=D@KN߀3/ڶ>F7N߀3/ڶ>F7,*DN.ی'79Ԛ<=/ڶ>J7@Ԛ<)'DN.ی'71=/ڶ>J7@Ԛ<N߀3/ڶ>F7N߀3/ڶ>F7ܤKKA7B  ܤKK+N߀3/ڶ>F7N߀3/ڶ>F7,*DN.ی'79Ԛ<=/ڶ>J7@Ԛ<)'DN.ی'71=/ڶ>J7@Ԛ<N߀3/ڶ>F7N߀3/ڶ>F77>1T֛7ٟ@9F6U>ʔ71/>ٟ@6LD7>/I/>=щQDDHIN./59Ԛ<ڶ>S-=DN@UW=-щQܭDHTDS=DSDA7>1֛7ٟ@9F6U>ʔ71/>5LD>/I/>=щQDDHIN./51S-=DN@UW=-щQܭDHTDS=DSDAN߀3/ڶ>F7N߀3/ڶ>F7,*DN.ی'79Ԛ<=/ڶ>J7@Ԛ<)'DN.ی'71=/ڶ>J7@Ԛ<N߀3/ڶ>F7N߀3/ڶ>F7,*DN=8T=4ڶ>F7S@@@)'DN=8T=4ڶ>F7S@@N߀3/ڶ>F7N߀3/ڶ>F7,*DN.ی'79Ԛ<=/ڶ>J7@Ԛ<)'DN.ی'71=/ڶ>J7@Ԛ<N߀3/ڶ>F7N߀3/ڶ>F7 H/67Dƹ;DG@KH/674D@KN߀3/ڶ>F7N߀3/ڶ>F7,*DN.ی'79Ԛ<=/ڶ>J7@Ԛ<)'DN.ی'71=/ڶ>J7@Ԛ<N߀3/ڶ>F7N߀3/ڶ>F7>S=>7ʗ74=>SB7ST86D1ƹ;T4>S=>7ʗ74>SB7STN߀3/ڶ>F7N߀3/ڶ>F7,*DN.ی'79Ԛ<=/ڶ>J7@Ԛ<)'DN.ی'71=/ڶ>J7@Ԛ<N߀3/ڶ>F7N߀3/ڶ>F7V/67=DG@KV/67=D@K$5H149A$5H149Aec$/4UR5RH$>#=1,1>Bٟ@T9ALKٟ@6J=@Ԛ<\Z$/4U5RH$>#=1,1>@T9ALKٟ@6=@Ԛ<$5H149A$5H149A>EѾCT86VOTBA?$US/6T9A6APɺDEXET8VOTB$5H149A$5H149Aec$/4UR5RH$>#=1,1>Bٟ@T9ALKٟ@6J=@Ԛ<\Z$/4U5RH$>#=1,1>@T9ALKٟ@6=@Ԛ<$5H149A$5H149A53ER91@5H1Bٟ@49AE@@@/-ER91@5H1@49AE@@$5H149A$5H149A/-$U5/8=49Aٟ@5DSDA/-$U5/8=49Aٟ@5DSDA$5H149A$5H149A,*$9656549Q5؂=@Ԛ<,*$9656549Q5؂=@Ԛ<$5H149A$5H149Aec$/4UR5RH$>#=1,1>Bٟ@T9ALKٟ@6J=@Ԛ<\Z$/4U5RH$>#=1,1>@T9ALKٟ@6=@Ԛ<$5H149A$5H149ADB"Ξ)69$R549AIٟ@TN>CJ@@Ԛ<><"69$R549A@TN>CJ@@Ԛ<$5H149A$5H149A/-$U5/8=49Aٟ@5DSDA/-$U5/8=49Aٟ@5DSDA$5H149A$5H149A;9E4WN$RB5H4LDLIĪNCS@K;9E4WN$RB5H4LDLIĪNCS@K$5H149A$5H149Aec$/4UR5RH$>#=1,1>Bٟ@T9ALKٟ@6J=@Ԛ<\Z$/4U5RH$>#=1,1>@T9ALKٟ@6=@Ԛ<$5H149A$5H149A53@;5RH$ULT9A6DPDA/-@5RHULT9A6DPDA$5H149A$5H149A/-$U5/8=49Aٟ@5DSDA/-$U5/8=49Aٟ@5DSDA$5H149A$5H149A86DP>E5H"$ĪNL=496A7B/-P>E5H"$ĪNL=496+:/SʡH99SH :S9HDBSWJ9?9?:/SʡH99:/SʡH995ܛ?M)'WJ99:S9:S95ܛ?M:/SʡH99SH :S9H/-:/SʡH999?99?D6T:S9999D6:/SʡH99SH :S9H&$SV:/SʡH99S6TV:S96:/SʡH99SH :S9H#!S:/SʡH999?Έ;F:S99Έ;F:/SʡH99SH :S9HDBSWJ9?9?:/SʡH99:/SʡH995ܛ?M)'WJ99:S9:S95ܛ?M:/SʡH99SH :S9H&$SV:/SʡH999?<>KDH><>KJHRD>HHHHHH<>KDH><>Kwu7RDH><>K,07R2 -.TʆL@ϡS4,ܢEM,.O2J6MKR2 -.TʆL@ϡS4,E,.OJ6DH><>KDH><>KJHRD>HHHHHH<>KDH><>KMKRDH><>K,0IO9491یV0—P=—PH>.E6A?RH><>K,0IO94V0=—PH>.E6DH><>KDH><>KJHRD>HHHHHH<>KDH><>K#!RD>HH<>KDH><>KJHRD>HHHHHH<>KDH><>K/-DH>K=<,D6R=4,@Ԛ<&$D54,D6R=4,@Ԛ<DH><>KDH><>KJHRD>HHHHHH<>KDH><>K7RDH><>K2>7.ʆJ6ʆG1?—P=1?I2K7>>MGMߎM6>JRʆ.J6~.ʆJ6ʆG1?=1?IK7>MGMߎM6>JRʆ.J6DH><>KDH><>KJHRD>HHHHHH<>KDH><>K20RDH><>K2>J6/;IN9,*RH><>K2>J6/;N9DH><>KDH><>KJHRD>HHHHHH<>KDH><>K_]RDH><>K2>ʆ>I2́N4TȇN4TI(—Pބ2>N4ʆN4GERH><>K2>ʆ>I(N4ȇN4I(܉2>NʆNDH><>KDH><>KJHRD>HHHHHH<>KDH><>KGERDH><>K2>J>I2ˏR3˰(IB>—P3ˏR2;9RH><>K2>JIˏR3˰(IB>3ˏR2ʰDBNMG> BMG>JHɵO9FDSC4ʰDBN5>35-=9O2:@@@53ɵO9DSC4B5>I-=O2G@@ʰDBNMG> BMG> LNLBʰDBN@@@LNLB@@ʰDBNMG> BMG>JHɵO9FDSC4ʰDBN5>35-=9O2:@@@53ɵO9DSC4B5>I-=O2G@@ʰDBNMG> BMG>)'$";0Q8ҐJ9ҽ6WH)'$";0Q8ҐJ9ҽ6WHCARVCWOAWCARVCWOA4ARQJOA4ARQJOCARVCWOAWCARVCWOA 4AR=J DG@K4AR= D@KCARVCWOAWCARVCWOA4ARQJOA4ARQJOCARVCWOAWCARVCWOA/4ARQ=JB4/4ARQ=BCARVCWOAWCARVCWOA4ARQJOA4ARQJOCARVCWOAWCARVCWOA><İFE1;TVL8ARO8L0AWН?/Н?T,*İFBTVL8ARO8LAН?-CARVCWOAWCARVCWOA4ARQJOA4ARQJOCARVCWOAWCARVCWOA20İFE1;TVL8ARO8L0AW&$İFBTVL8ARO8LACARVCWOAWCARVCWOA4ARQJOA4ARQJOCARVCWOAWCARVCWOAPNİFE1;TVL8ARO8L0AWWН?W?UUWTН?>;9İFBTVL8ARO8LAWW?UUW?CARVCWOAWCARVCWOA4ARQJOA4ARQJOCARVCWOAWCARVCWOA,*İFE1;TVL8O3߫UТ@HT&$İFBTVL8O3߫UТ@HTCARVCWOAWCARVCWOA4ARQJOA4ARQJOCARVCWOAWCARVCWOA AR4J AR4JCARVCWOAWCARVCWOA4ARQJOA4ARQJOCARVCWOAWCARVCWOAL/4ARQ>L/4ARQ>CARVCWOAWCARVCWOA4ARQJOA4ARQJOCARVCWOAWCARVCWOA&$ŷ5/BAR4JX>BHH9;>B 9XR9  9R9PNMRF=:9X94.б H>N̛<;TTН?T("'!53MRF=:994.б H>N;Tܞ? 9XR9  9R9CR9Xnj8@Ԛ<CR9nj8@Ԛ< 9XR9  9R9&$KX/9CR=U93ATX9CR=U93A 9XR9  9R99XUTI9XNS;UOIַ;URIIIKIHBOF;F;N̛<;TTН?T("'!53MRF=:994.б H>N;Tܞ? 9XR9  9R9&$9X9C5I91ӛ?69; 9F5I91ӛ?69; 9XR9  9R9&$KX/9CR=U93ATX9CR=U93A 9XR9  9R99XUC;- 9UC- 9XR9  9R9PNMRF=:9X94.б H>N̛<;TTН?T("'!53MRF=:994.б H>N;Tܞ? 9XR9  9R9\ZG9XWF5ԎB@JP11.3>72PNG9WF5BJP11.3>72 9XR9  9R9&$KX/9CR=U93ATX9CR=U93A 9XR9  9R9869XB9ԎB@@OLWFR9B9N̛<;TTН?T("'!53MRF=:994.б H>N;Tܞ? 9XR9  9R9&$9X9X59QCͦ(!995ƋQC 9XR9  9R9&$KX/9CR=U93ATX9CR=U93A 9XR9  9R9 9X@?9@ 9XR9  9R9PNMRF=:9X94.б H>N̛<;TTН?T("'!53MRF=:994.б H>N;Tܞ? 9XR9  9R9YW9XB9ԎB@>54WFR9B9 IC70FŔ6ADMIַ;70DB9B9B>54WFR9B9 IC0FŔ61I7 9XR9  9R9&$KX/9CR=U93ATX9CR=U93A 9XR9  9R9DB9XCK29R5>9XWA/1C2ODKOD539C2R5>9WA1C2ODKOD 9XR9  9R9PNMRF=:9X94.б H>N̛<;TTН?T("'!53MRF=:994.б H>N;Tܞ? 9XR9  9R9&$9X>KTCΚIRН?>AT9>KCΚIR?A 9XR9  9R9&$KX/9CR=U93ATX9CR=U93A 9XR9  9R920CCTC7VCEICַ;C;-CTCCCVĸIַ;C-C 9XR9  9R9PNMRF=:9X94.б H>N̛<;TTН?T("'!53MRF=:994.б H>N;Tܞ? 9XR9  9R9DBHW:9XB9ԎB@=ʼnEDWFR9B99XCT86HW:9B9B=ʼnEDWFR9B99C 9XR9  9R9&$KX/9CR=U93ATX9CR=U93A 9XR9  9R99X@T9XR0ܥ69@T9Rܥ6 9XR9  9R9PNMRF=:9X94.б H>N̛<;TTН?T("'!53MRF=:994.б H>N;Tܞ? 9XR9  9R9G7;CT G7;C 9XR9  9R9&$KX/9CR=U93ATX9CR=U93A 9XR9  9R9A?Hʜ2RA@RS9@>9X3>)כ$>;GB;9Hʜ2RA@RS9@>93>)כ$>;G 9XR9  9R9PNMRF=:9X94.б H>N̛<;TTН?T("'!53MRF=:994.б H>N;Tܞ? 9XR9  9R9><NR=9XC9S99׵AAKEAABC/;9NR=9C9S99׵AAKEAABC/=@KE= =@E=><@Q0H@KûAQH@KûAQ,HPHCB020@0H@ûAQH@ûAQ,HPHB0=@KE= =@E==@J@KI5@=@J@I5@=@KE= =@E=/-7ûAK3@3@K7KK3!#!7ûAK3@3@7K3=@KE= =@E=86=@KAKCK-3O?3377CT)'=@AKCK-.?.7C=@KE= =@E=/-K6S5@KE=4I,S@@@)'K6S5@E=4I,S@@=@KE= =@E= @K@?@@=@KE= =@E=)'C@ַ;C@GC@K=@AB&$C@ַ;C@GC@=@AB=@KE= =@E=DBIK@KQOַ;OE6V=ԋ J>JT7LJ653IK@QOַ;OE6V=JJ7LJ6ԃP;ܢE4JAˑ+86Q FM1UܢE4NԃP;O4HН?U,T#!Q FM1UAOH,ԃP;ܢE4JAˑ+,*ԃP;Q8ȘIK5ܢE4N>4OJAQ8K5>4OԃP;ܢE4JAˑ+ ԃP;1ܢE4NН?̛4׶K21T)'AHQ8K5C>4׶K21ԃP;ܢE4JAˑ+86ԃP;HQ8ȘIK5ܢE4NC>4б XQT)'AHQ8K5C>4б XQԃP;ܢE4JAˑ+><ԃP;HQ8ȘIK5ܢE4NC>4HН?Н?>HT,*AHQ8K5C>4H?HTԃP;ܢE4JAˑ+MKԃP;HQ8ȘIK5NC>4ԃP;Q:33ȘIJ82THA>4AQ:33ȘIJ82HԃP;ܢE4JAˑ+53ԃP;HQ8ȘIK5ܢE4NC>4TН?T&$AHQ8K5C>4Tܞ?ԃP;ܢE4JAˑ+/-ԃP;HL-TܢE4NC41TН?> AHL-TC41?ԃP;ܢE4JAˑ+Dֈ;0OFԃP;ܢE4JAˑ+,*ԃP;HQ8ȘIK5ܢE4NC>4 AHQ8K5C>4ԃP;ܢE4JAˑ+86Q FM1UܢE4NԃP;O4HН?U,T#!Q FM1UAOH,ԃP;ܢE4JAˑ+)'ԃP;E72TܢE4NŇ7̛4б 3QT)'AHQ8K5C>4б 3QԃP;ܢE4JAˑ+86ԃP;HQ8ȘIK5ܢE4NC>4׶K21T)'AHQ8K5C>4׶K21ԃP;ܢE4JAˑ+86ԃP;HQ8ȘIK5ܢE4NCT?TCܢE0&$AHQ8K5CT?T/ԃP;ܢE4JAˑ+><ԃP;HQ8ȘIK5ܢE4NC>4HН?Н?>HT,*AHQ8K5C>4H?HTԃP;ܢE4JAˑ+86ԃP;HQ8ȘIK5ܢE4NC>4XН?2J&$AHQ8K5C>4X2ԃP;ܢE4JAˑ+53ԃP;HQ8ȘIK5ܢE4NC>4TН?T&$AHQ8K5C>4Tܞ?ԃP;ܢE4JAˑ+86ԃP;HQ8ȘIK5ܢE4NC>4Н?̛4?ETԃP;ܢE4JAˑ+Dֈ;0OFԃP;ܢE4JAˑ+86ԃP;HQ8ȘIK5ܢE4NC>4Н?̛4?UԃP;ܢE4JAˑ+86Q FM1UܢE4NԃP;O4HН?U,T#!Q FM1UAOH,ԃP;ܢE4JAˑ+#!Q1NÚQ8ȘIKTԃP;4Q1N8KTA4ԃP;ܢE4JAˑ+ ԃP;1ܢE4NН?̛C1A1J>=)'-AHC;>C1Aܹ1>=ԃP;ܢE4JAˑ+86ԃP;HQ8ȘIK5ܢE4NC>4׶K21T)'AHQ8K5C>4׶K21ԃP;ܢE4JAˑ+~4NU.̤3@>ϥJ=T.-0ܢE4N5H01ԃP;R:?=N.̤3@>PTT>JFF8G3b`4NU.LϥJ=T.-05H01AR:=N.LPT>JFF8GԃP;ܢE4JAˑ+><ԃP;HQ8ȘIK5ܢE4NC>4HН?Н?>HT,*AHQ8K5C>4H?HTԃP;ܢE4JAˑ+20ԃP;߽4Q8ȘIK5ܢE4N,4U/T&$A߽4Q8K5,4U/TԃP;ܢE4JAˑ+53ԃP;HQ8ȘIK5ܢE4NC>4TН?T&$AHQ8K5C>4Tܞ?ԃP;ܢE4JAˑ+,*Q1ʡH9BXTܢE4NН?̛<7TQ19XT?7ԃP;ܢE4JAˑ+Dֈ;0OFԃP;ܢE4JAˑ+86ԃP;HQ8ȘIK5ܢE4NC>4Н?>RT&$AHQ8K5C>4?RԃP;ܢE4JAˑ+86Q FM1UܢE4NԃP;O4HН?U,T#!Q FM1UAOH,ԃP;ܢE4JAˑ+20ԃP;߽4Q8ȘIK5ܢE4N,4XQT#!A߽4Q8K5,4XQԃP;ܢE4JAˑ+ ԃP;1ܢE4NН?̛4C-HН?̛<&##!)'AHQ8K5C>4*? ԃP;ܢE4JAˑ+86ԃP;HQ8ȘIK5ܢE4NC>4׶K21T)'AHQ8K5C>4׶K21ԃP;ܢE4JAˑ+SQԃP;HQʡHɤUBUHMܢE4NCT۹/8HMT>JT8:G3>JT:GԃP;ܢE4JAˑ+><ԃP;HQ8ȘIK5ܢE4NC>4HН?Н?>HT,*AHQ8K5C>4H?HTԃP;ܢE4JAˑ+ecԃP;߽4Q8ȘIK5ܢE4N,4ԃP;ܢE4N5NģCF4QO1MJEа.TН?>;9A߽4Q8K5,4AQO-Eа.T?ԃP;ܢE4JAˑ+53ԃP;HQ8ȘIK5ܢE4NC>4TН?T&$AHQ8K5C>4Tܞ?ԃP;ܢE4JAˑ+/-ԃP;HUܢE4NCRKD?TيR̛<&$AHUCRKD?TيR̛<ԃP;ܢE4JAˑ+Dֈ;0OFԃP;ܢE4JAˑ+,*7ԃP;E72TܢE4NН?>AT7AE7T?AԃP;ܢE4JAˑ+86Q FM1UܢE4NԃP;O4HН?U,T#!Q FM1UAOH,ԃP;ܢE4JAˑ+86߹-JН?̛<ԃP;HQ8ȘIK5ܢE4NC>4&$-?AHQ8K5C>4ԃP;ܢE4JAˑ+ ԃP;1ܢE4NН?̛4б XQT)'AHQ8K5C>4б XQԃP;ܢE4JAˑ+86ԃP;HQ8ȘIK5ܢE4NC>4׶K21T)'AHQ8K5C>4׶K21ԃP;ܢE4JAˑ+#!ԃP;߽4UL6.TܢE4NA߽4UL6TԃP;ܢE4JAˑ+><ԃP;HQ8ȘIK5ܢE4NC>4HН?Н?>HT,*AHQ8K5C>4H?HTԃP;ܢE4JAˑ+20ԃP;߽4U72TܢE4NԃP;߽4TН?T A߽4U7TA߽4Tܞ?ԃP;ܢE4JAˑ+53ԃP;HQ8ȘIK5ܢE4NC>4TН?T&$AHQ8K5C>4Tܞ?ԃP;ܢE4JAˑ+;9ԃP;HQ8ȘIK5ܢE4NC>4 0̛4 0QTIOT TIOT 53RP4JTIOT> Sߢ?U>9@Ԛ<,*R4TIOT> S?>9@Ԛ<TIOT TIOT /-TIOTބ2BJ768T7P4J#!TIOTބ2BќJ6874TIOT TIOT &$TIOT0Q7J6J7&$TIOT0Q7J6J7TIOT TIOT JHRTIOT4/ >BԚԚ U@Ԛ< RTIOT> U@Ԛ<TIOT TIOT 53RP4JTIOT> Sߢ?U>9@Ԛ<,*R4TIOT> S?>9@Ԛ<TIOT TIOT 53RP4JTIOT> Sߢ?U>9@Ԛ<,*R4TIOT> S?>9@Ԛ<TIOT TIOT #!TIOTބ2B>TV>T#!TIOTބ2B>TV>TTIOT TIOT &$TIOT0Q7J6J7&$TIOT0Q7J6J7TIOT TIOT SQRP4D3TMɾSBTIOTL;U$ N,%!@Ԛ<;9R4D3TMBTIOTL;U N,@Ԛ<TIOT TIOT  RTIOT> U@Ԛ< RTIOT> U@Ԛ<TIOT TIOT 86R9TIOT> BK1١-JL;@@@/-R9TIOT> BK1١-8@@TIOT TIOT 53RP4JTIOT> Sߢ?U>9@Ԛ<,*R4TIOT> S?>9@Ԛ<TIOT TIOT trT>IOTմ2O̤@ROWBǞV<>MɾS3D UJDP>W>5ֈD,DL9ADSDAkiT>IOTմ2@ROWBȞV>M3D UJDP>W>5ֈD,DL9ADSDATIOT TIOT &$TIOT0Q7J6J7&$TIOT0Q7J6J7TIOT TIOT 53TIOT*B6J768T7P4J2)'TIOT*B6ќJ68742TIOT TIOT  RTIOT> U@Ԛ< RTIOT> U@Ԛ<TIOT TIOT ,*TIOT> ,:%!@Ԛ< TIOT> ,:@Ԛ<TIOT TIOT 53RP4JTIOT> Sߢ?U>9@Ԛ<,*R4TIOT> S?>9@Ԛ<TIOT TIOT A?T7IOT> 3D,R,SUUP4J@@@53T7IOT> 3D,R,SU4@@TIOT TIOT &$TIOT0Q7J6J7&$TIOT0Q7J6J7TIOT TIOT 53RP4JTIOT> Sߢ?U>9@Ԛ<,*R4TIOT> S?>9@Ԛ<TIOT TIOT  RTIOT> U@Ԛ< RTIOT> U@Ԛ<TIOT TIOT 86RT>IOTK>SF> P4J@@@)'RT>IOTKS> 4@@TIOT TIOT 53RP4JTIOT> Sߢ?U>9@Ԛ<,*R4TIOT> S?>9@Ԛ<TIOT TIOT /-TIOTB62LCP4J>T#!TIOTB62C4>TIOT TIOT &$TIOT0Q7J6J7&$TIOT0Q7J6J7TIOT TIOT MK9QDT7IOT>SFDU>F> ;/?BRÙKBT><9QDT7IOTSDU>F> ;/BEBTIOT TIOT  RTIOT> U@Ԛ< RTIOT> U@Ԛ<TIOT TIOT 20P4JTIOTSUXߢ?U,6XT&$4TIOTSUX?6XTTIOT TIOT 53RP4JTIOT> Sߢ?U>9@Ԛ<,*R4TIOT> S?>9@Ԛ<TIOT TIOT 20TIOT47>4 3DFDSDA,*TIOT4> 3DFDSDATIOT TIOT &$TIOT0Q7J6J7&$TIOT0Q7J6J7TIOT TIOT 20TIOTB6J768T7P4BT)'TIOTB6ќJ6874BTTIOT TIOT  RTIOT> U@Ԛ< RTIOT> U@Ԛ<TIOT TIOT hfRT>IOT> UP4>4—P=AN,:L%!**P4>٬J=$@Ԛ<SQRT>IOT> U4>4=AN,:L**4>٬J=$@Ԛ<TIOT TIOT 53RP4JTIOT> Sߢ?U>9@Ԛ<,*R4TIOT> S?>9@Ԛ<TIOT TIOT DBRP4JTIOT>MKJIOTKK DPDA>MKJIOTKK DPDAA,G߇;G߇;%>MA,G߇;G߇;%>M\ZAPIDK4,G,G,G߇;5>,VCʿ7NPI>>>V0>@Ԛ<\ZAPIDK4,G,G,G߇;5>,VCʿ7NPI>>>V0>@Ԛ<A,G߇;G߇;%>MA,G߇;G߇;%>M%A%AG  %AAA,G߇;G߇;%>MA,G߇;G߇;%>M%A%A%AAA,G߇;G߇;%>MA,G߇;G߇;%>M&'%IIA$ۏ"&'%IIAG&'%II :AGD3AT(%!AG}{&'%IIA&'%IIA&'%II :AD3ATVAA,G߇;G߇;%>MA,G߇;G߇;%>M\ZAPIDK4,G,G,G߇;5>,VCʿ7NPI>>>V0>@Ԛ<\ZAPIDK4,G,G,G߇;5>,VCʿ7NPI>>>V0>@Ԛ<A,G߇;G߇;%>MA,G߇;G߇;%>M%A%A %AA A,G߇;G߇;%>MA,G߇;G߇;%>M%A%A%AAA,G߇;G߇;%>MA,G߇;G߇;%>M20%CV2%0J%2CWFTOWW)'%CV2%0%2WFTO9A,G߇;G߇;%>MA,G߇;G߇;%>M\ZAPIDK4,G,G,G߇;5>,VCʿ7NPI>>>V0>@Ԛ<\ZAPIDK4,G,G,G߇;5>,VCʿ7NPI>>>V0>@Ԛ<A,G߇;G߇;%>MA,G߇;G߇;%>M20%BF%JW DG%AG@F:=#!%<%J D%A@:=A,G߇;G߇;%>MA,G߇;G߇;%>M%A%A%AAA,G߇;G߇;%>MA,G߇;G߇;%>MJHD9GM>AQٟ@DBU,G߇;G3MVٟ@6DPDA>AQٟ@DK,G߇;G3MV5DPDAA,G߇;G߇;%>MA,G߇;G߇;%>M\ZAPIDK4,G,G,G߇;5>,VCʿ7NPI>>>V0>@Ԛ<\ZAPIDK4,G,G,G߇;5>,VCʿ7NPI>>>V0>@Ԛ<A,G߇;G߇;%>MA,G߇;G߇;%>M/-AG%;̽>MŹ(Źʿ@@@)'AG%;>Ź(Źʿ@@A,G߇;G߇;%>MA,G߇;G߇;%>M%A%A%AAA,G߇;G߇;%>MA,G߇;G߇;%>M20%DJW.>=V%JW G%A)'%DJW.>=V%J GA 2EֈD$& 2EֈD)'$ 2̙EֈD>ܤK"6"&#!$ 2̙EֈD>ܤK"6" 2EֈD$& 2EֈD86$& C2̙EϪJֈDT9J9@AB/- C2̙EϪJֈDTJ9@AB 2EֈD$& 2EֈD)'$ 2̙EֈD>ܤK"6"&#!$ 2̙EֈD>ܤK"6" 2EֈD$& 2EֈD)' 2EC$&E̛<0>WT 2ECE0>W 2EֈD$& 2EֈD)'$ 2̙EֈD>ܤK"6"&#!$ 2̙EֈD>ܤK"6" 2EֈD$& 2EֈD)'$& C2GE9ֈD@Ԛ<#! C2GE9ֈD@Ԛ< 2EֈD$& 2EֈD)'$ 2̙EֈD>ܤK"6"&#!$ 2̙EֈD>ܤK"6" 2EֈD$& 2EֈD;9Sޡ8$&>&2̙E ֈD>ܤK$'&9Q')'S>&2̙E ֈD>ܤKƋQ' 2EֈD$& 2EֈD)'$ 2̙EֈD>ܤK"6"&#!$ 2̙EֈD>ܤK"6" 2EֈD$& 2EֈD~6AB6T 2EۈXD:ۈX>ў7&B$&,&ίB>T7>KUVJJKUQTI1R/0Qec6AB6T 2EۈXD:ۈX>ў7&B,&ίB>T7KVQI1R/Q 2EֈD$& 2EֈD)'$ 2̙EֈD>ܤK"6"&#!$ 2̙EֈD>ܤK"6" 2EֈD$& 2EֈD$&2@ 8,T2@ ,T 2EֈD$& 2EֈD)'$ 2̙EֈD>ܤK"6"&#!$ 2̙EֈD>ܤK"6" 2EֈD$& 2EֈD_]$֗>AS 19EŹ4(>&24 EB߻WֈD1H%,9: >I\Z$֗>AS 19EŹ4(>&24 EB߻WֈD1H%,: >I 2EֈD$& 2EֈD)'$ 2̙EֈD>ܤK"6"&#!$ 2̙EֈD>ܤK"6" 2EֈD$& 2EֈDA?$& ۈX2@QTWNEܾW,;PT,T86 ۈX2@QTWNEܾW,;ٱP,T 2EֈD$& 2EֈD)'$ 2̙EֈD>ܤK"6"&#!$ 2̙EֈD>ܤK"6" 2EֈD$& 2EֈD53ޥ0CE$&0> 2EֈDJ<=@,*ޥ0CE0> 2EֈDJ=@;1>DH ;1>D,*;>DH66;DH9FA@Ԛ<#!;>D6;D9FA@Ԛ<;1>DH ;1>D;DHؕ7;EE@;Dؕ7;EE@;1>DH ;1>D,*;>DH66;DH9FA@Ԛ<#!;>D6;D9FA@Ԛ<;1>DH ;1>DXŷ5D/D/ Xŷ5DD;1>DH ;1>D,*;>DH66;DH9FA@Ԛ<#!;>D6;D9FA@Ԛ<;1>DH ;1>D ;DHBU>UW6T;DΑB>U6;1>DH ;1>D,*;>DH66;DH9FA@Ԛ<#!;>D6;D9FA@Ԛ<;1>DH ;1>D;DHDHDHT;DDDT;1>DH ;1>D,*;>DH66;DH9FA@Ԛ<#!;>D6;D9FA@Ԛ<;1>DH ;1>D ;DH>  ;D>;1>DH ;1>D,*;>DH66;DH9FA@Ԛ<#!;>D6;D9FA@Ԛ<;1>DH ;1>D BD/>  BD>;1>DH ;1>D,*;>DH66;DH9FA@Ԛ<#!;>D6;D9FA@Ԛ<;1>DH ;1>D;DH=DH ;1>D,*;>DH66;DH9FA@Ԛ<#!;>D6;D9FA@Ԛ<;1>DH ;1>D;ӈ5UD>DHDH;5D>DDE1?0;E1?0;ַ;E1?,;@Ԛ<ַ;E1?,;@Ԛ<E1?0;E1?0;;9K6>HE1K/Q4DGKIAB86K6>HE1K/Q4GKIABE1?0;E1?0;GEDKOFHE1K/Q4DGKOJܤK>6DG@K20HE1K/Q4GܤK>6D@KE1?0;E1?0;#!DE1ߢ?08IDE1?1BT/>׆B/1/69IPTR;I@Ԛ<MKܤ5ַ;>E1?1BT/>׆B/1/69IPTR;I@Ԛ<E1?0;E1?0;GEDKOFHE1K/Q4DGKOJܤK>6DG@K20HE1K/Q4GܤK>6D@KE1?0;E1?0;A?A׆B?KUEI3R>7DE1?P;66@Ԛ<;9A׆B?KUEI3>7DE1?P;6@Ԛ<QE1?0;E1?0;1A?Iַ;  1AIE1?0;E1?0;ַ;E1?,;@Ԛ<ַ;E1?,;@Ԛ<E1?0;E1?0;53AUE1AIٟ@;N?985D@@@/-AUE1A@;N?985D@@G=ݰFBSF G=FF#!BN0ݰFBSF2Uа.TBNFF2*G=ݰFBSF G=FF)'AOݰFBFASF>LS2 AOFFAF>LSG=ݰFBSF G=FF#!BN0ݰFBSF2Uа.TBNFF2*G=ݰFBSF G=FF86ݰFBSFQBJ768T7QݰFBSFB&$FFQBќJ687QFFBG=ݰFBSF G=FF#!BN0ݰFBSF2Uа.TBNFF2*G=ݰFBSF G=FF AסET/ݰFBٟ@3@Ԛ<ATFٟ@3@Ԛ<G=ݰFBSF G=FF#!BN0ݰFBSF2Uа.TBNFF2*G=ݰFBSF G=FFSFUR7T FU7T11F֎TPAJ11F֎TPAJ20ڶ>S:—PG2&**11F֎T@Ԛ<,*S:I2&**11F֎T@Ԛ<11F֎TPAJ11F֎TPAJ11F֎T V>б 11F֎T Vб 11F֎TPAJ11F֎TPAJA?7D2T:֎T11F֎T=?N;TTK;2072T:֎T11F֎TTTK;11F֎TPAJ11F֎TPAJ/-SAS11F֎T=>щQCE@@@,*SAS11F֎T=>щQCE@@11F֎TPAJ11F֎TPAJ20ڶ>S:—PG2&**11F֎T@Ԛ<,*S:I2&**11F֎T@Ԛ<11F֎TPAJ11F֎TPAJ)'11F֎T=?N;78K11F֎T7K11F֎TPAJ11F֎TPAJA?7D2T:֎T11F֎T=?N;TTK;2072T:֎T11F֎TTTK;11F֎TPAJ11F֎TPAJ=?N;C;MC;M11F֎TPAJ11F֎TPAJ20ڶ>S:—PG2&**11F֎T@Ԛ<,*S:I2&**11F֎T@Ԛ<11F֎TPAJ11F֎TPAJJH$U-£-E7-Ҳ0AʡH9DS&11F֎T7J6!A?$U-£-E7-Ҳ0AʡH9DS&11F֎T7611F֎TPAJ11F֎TPAJA?7D2T:֎T11F֎T=?N;TTK;2072T:֎T11F֎TTTK;11F֎TPAJ11F֎TPAJ;911F֎TBJHį-HUHڶ>2>AR@Ԛ<;911F֎TBJHį-HUHڶ>2>AR@Ԛ<11F֎TPAJ11F֎TPAJ20ڶ>S:—PG2&**11F֎T@Ԛ<,*S:I2&**11F֎T@Ԛ<11F֎TPAJ11F֎TPAJ ӪN11F֎TE@@@ӪN11F֎TE@@11F֎TPAJ11F֎TPAJA?7D2T:֎T11F֎T=?N;TTK;2072T:֎T11F֎TTTK;11F֎TPAJ11F֎TPAJ8611F֎T03VCJ768T711F֎T2011F֎T03VCќJ68711F֎T11F֎TPAJ11F֎TPAJ20ڶ>S:—PG2&**11F֎T@Ԛ<,*S:I2&**11F֎T@Ԛ<11F֎TPAJ11F֎TPAJ11F֎T@?11F֎T@11F֎TPAJ11F֎TPAJA?7D2T:֎T11F֎T=?N;TTK;2072T:֎T11F֎TTTK;11F֎TPAJ11F֎TPAJSQDR07>I8Ҳ02AXڃN>11F֎TAKAٟ@HDPDAPNDR07>8Ҳ02AXڃN>11F֎TAKAٟ@HDPDA11F֎TPAJ11F֎TPAJ20ڶ>S:—PG2&**11F֎T@Ԛ<,*S:I2&**11F֎T@Ԛ<11F֎TPAJ11F֎TPAJ5311F֎TW")$IK46)'11F֎TW")I411F֎TPAJ11F֎TPAJA?7D2T:֎T11F֎T=?N;TTK;2072T:֎T11F֎TTTK;11F֎TPAJ11F֎TPAJJH7&:֎T11F֎TTTT=?N;T!537&:֎T11F֎TTTTTK11F֎TPAJ11F֎TPAJ20ڶ>S:—PG2&**11F֎T@Ԛ<,*S:I2&**11F֎T@Ԛ<11F֎TPAJ11F֎TPAJ&$CE>11F֎T@0=@Ԛ<&$CE>11F֎T@0=@Ԛ<11F֎TPAJ11F֎TPAJA?7D2T:֎T11F֎T=?N;TTK;2072T:֎T11F֎TTTK;11F֎TPAJ11F֎TPAJPN11F֎T=?N;7=?N;GTTT - !.,11F֎T7GTTT+11F֎TPAJ11F֎TPAJ20ڶ>S:—PG2&**11F֎T@Ԛ<,*S:I2&**11F֎T@Ԛ<11F֎TPAJ11F֎TPAJki11F֎TKSħ;S C9>>4K.TRҲ0AGB@>=?N;)ʪ\Z11F֎TKSS Cޖ>>4K.TRҲ0AGB@>)ʪ11F֎TPAJ11F֎TPAJA?7D2T:֎T11F֎T=?N;TTK;2072T:֎T11F֎TTTK;11F֎TPAJ11F֎TPAJ53&11F֎TRBOEVCE@@@,*&11F֎TRBOECE@@11F֎TPAJ11F֎TPAJ20ڶ>S:—PG2&**11F֎T@Ԛ<,*S:I2&**11F֎T@Ԛ<11F֎TPAJ11F֎TPAJ,*7DT11F֎T6U=?N;7T11F֎T611F֎TPAJ11F֎TPAJA?7D2T:֎T11F֎T=?N;TTK;2072T:֎T11F֎TTTK;11F֎TPAJ11F֎TPAJ_]$U-£-E7-Ҳ0AʡH9DS&11F֎T$U-CɤUTҲ0AB!YW$U-£-E7-Ҳ0AʡH9DS&11F֎T$U-CɤUTҲ0AB86X,19CK/ - NW=HDEģCKX1KNW/DCVTX,19CʡH97/ - NW=HDEģCKGģC:7BWT53X1ʡH97NW/DCGģC7BW86X,19CK/ - NW=HDEģCKX1KNW/DCJHX,19CʡH97/ - NW=HDE8KDG@K/-X1ʡH97΂NW/D8KD@K86X,19CK/ - NW=HDEģCKX1KNW/DCGEX,19CʡH97/ - NWCHDEģCK΂:6T)'X1ʡH97NW޻/DC΂:686X,19CK/ - NW=HDEģCKX1KNW/DC\Z-AX,19CʡH97/ - NW=HDEģCK -:K48?:T><-AX1ʡH97NW/DC -:48?:T86X,19CK/ - NW=HDEģCKX1KNW/DC;9X,19CK/ - NW=HDEGI#!X1KNW/DGI86X,19CK/ - NW=HDEģCKX1KNW/DCb`X,19CʡH97/ - NW=HDEK?IU>DE?T΂:C̛<A?X1ʡH97NW/DE?IUD?΂:C̛<86X,19CK/ - NW=HDEģCKX1KNW/DC86X,19CK/ - NW=HDEģCKX1KNW/DC86X,19CK/ - NW=HDEģCKX1KNW/DCGEX,19CʡH97/ - NW=HDE8K΂:4T/-X1ʡH97NW/D8K΂:4T86X,19CK/ - NW=HDEģCKX1KNW/DCVTX,19CʡH97/ - NW=HDEģCKGģC:7BWT53X1ʡH97NW/DCGģC7BW86X,19CK/ - NW=HDEģCKX1KNW/DC86X,19CK/ - NW=HDE>KX1KNW/D>86X,19CK/ - NW=HDEģCKX1KNW/DCGEX,19CʡH97/ - NWCHDEģCK΂:6T)'X1ʡH97NW޻/DC΂:686X,19CK/ - NW=HDEģCKX1KNW/DCSQX,19CʡH97/ - NW=HDEOKDOGDO6G20X1ʡH97NW/DODGD6G86X,19CK/ - NW=HDEģCKX1KNW/DC;9X,19CK/ - NW=HDEGI#!X1KNW/DGI86X,19CK/ - NW=HDEģCKX1KNW/DCJHX,19CʡH97/ - /@CHWDEģCKùBNL,*X1ʡH97N޻/WDCùBNL86X,19CK/ - NW=HDEģCKX1KNW/DC86X,19CK/ - NW=HDEģCKX1KNW/DC86X,19CK/ - NW=HDEģCKX1KNW/DCqoX,19CʡH97/ - NW=HDE>KL28AWT6O0U—PD7>6;PNX1ʡH97NW/D>LPAW6O0U—PD7>6;86X,19CK/ - NW=HDEģCKX1KNW/DCVTX,19CʡH97/ - NW=HDEģCKGģC:7BWT53X1ʡH97NW/DCGģC7BW86X,19CK/ - NW=HDEģCKX1KNW/DC>E6DSDA53-II6I6I66U>E6DSDANB-<66N-<66#!NB-<66ODSDAN-<66DSDANB-<66N-<66_]-I6DD9D66>=/,ֈ;N?KCL3;ނB/6/7TNؕ7؄/ESQ-I6D966>=/,ֈ;N?KCL3ނB/6/7TNڕ7ENB-<66N-<66><-I66OE60FǂSHAVTJD8DAP/--I66E6FǂSHAVTD8ANB-<66N-<6686-II6I6I66OU>E6DSDA53-II6I6I66U>E6DSDANB-<66N-<66#!NB-<6OC8A99N-<6OC8A9NB-<66N-<66_]-I6DD9D66>=/,ֈ;N?KCL3;ނB/6/7TNؕ7؄/ESQ-I6D966>=/,ֈ;N?KCL3ނB/6/7TNڕ7ENB-<66N-<66E6DSDA53-II6I6I66U>E6DSDANB-<66N-<66)'D-IHD6/E6-116)'D-IHD6/E6-116NB-<66N-<66_]-I6DD9D66>=/,ֈ;N?KCL3;ނB/6/7TNؕ7؄/ESQ-I6D966>=/,ֈ;N?KCL3ނB/6/7TNڕ7ENB-<66N-<66 кB-<ԋ/C66JƱCTкB-<ԋ/C66JϱCNB-<66N-<6686-II6I6I66OU>E6DSDA53-II6I6I66U>E6DSDANB-<66N-<66,*NB-=/,ֈ;N?KCL3;ނB/6/7TNؕ7؄/ESQ-I6D966>=/,ֈ;N?KCL3ނB/6/7TNڕ7ENB-<66N-<66865-Н?T  R>ܞ? İU7/ İU7/204UİU7/5.W@ßNWF/ÐWW/-4UİU7/5.W@ßNW/ÐWW İU7/ İU7//-UİU7/.W@ßN1T7̛<,*UİU7/.W@ßN17̛< İU7/ İU7/  -NUİU7/.@K  -NUİU7/.@K İU7/ İU7/534İU7/5:S9İU:4K"!,*4İU7/5:S9İU:4K" İU7/ İU7/86T14UİU7/5.:S9İUAWAT20T14UİU7/5.:S9İUAA İU7/ İU7/;94UİU7/5.W@ßNWF?9GHН?T204UİU7/5.W@ßNW?9G/ İU7/ İU7/204UİU7/5.W@ßNWF/ÐWW/-4UİU7/5.W@ßNW/ÐWW İU7/ İU7/GEUİU7/.W@ßNWF/ɴ9Н?Tɴ9ʡH9?/T;9UİU7/.W@ßNW/ɴ9ܞ?ɴ99/T İU7/ İU7/  -NUİU7/.@K  -NUİU7/.@K İU7/ İU7/#!4UİU7/5.W@ßN#!4UİU7/5.W@ßN İU7/ İU7/86T14UİU7/5.:S9İUAWAT20T14UİU7/5.:S9İUAA İU7/ İU7/864UİU7/5.W@ßNWF/̝5̛FˎWBDIKT)ʪ/-KFEڶ>FˎWBDIK)ʪį-KEˎWٟ@6֬4Jį-KEˎW5֬4J-K-Kį-KEˎWٟ@6֬4Jį-KEˎW5֬4J&$Sį-K>JNTCTT#!Sį-K>JϞNCTTį-KEˎWٟ@6֬4Jį-KEˎW5֬4J-K-Kį-KEˎWٟ@6֬4Jį-KEˎW5֬4J кB6Sį-KIKT:KкB6Sį-KIK:Kį-KEˎWٟ@6֬4Jį-KEˎW5֬4J-K-Kį-KEˎWٟ@6֬4Jį-KEˎW5֬4J#!;Kʗ,/Sտ7PC@;B ;Kʗ,/Sտ7PC;Bį-KEˎWٟ@6֬4Jį-KEˎW5֬4J-K-Kį-KEˎWٟ@6֬4Jį-KEˎW5֬4J203BBDK6S9A@S@060T203BBDK6S9A@S@060Tį-KEˎWٟ@6֬4Jį-KEˎW5֬4J-K-Kį-KEˎWٟ@6֬4Jį-KEˎW5֬4J/-;Kʗ,/—PL>CBFRKAKB,*;Kʗ,/—PL>CBFRKAKį-KEˎWٟ@6֬4Jį-KEˎW5֬4J-K-Kį-KEˎWٟ@6֬4Jį-KEˎW5֬4J)';Kʗ,/SKD͙7IRN͙7T&$;Kʗ,/SKDIRN͙7Tį-KEˎWٟ@6֬4Jį-KEˎW5֬4J-K-Kį-KEˎWٟ@6֬4Jį-KEˎW5֬4J#!SKб J768T7U>SKб ќJ687U>R/,B;R/,B;)'NЃB;W$,BΞ)9"@@@&$NЃB;W$,BΞ)9"@@R/,B;R/,B;,BR/>47,BR/>47R/,B;R/,B;)'NЃB;W$,BΞ)9"@@@&$NЃB;W$,BΞ)9"@@R/,B;R/,B;;9ѹ67,BƸ=DJ7.K/B9A=B@@@&$չ6,BƸ=DJ*/BA@@R/,B;R/,B;)'NЃB;W$,BΞ)9"@@@&$NЃB;W$,BΞ)9"@@R/,B;R/,B;;9R/B,B.P԰'0VAUѹ6FG,*R/B,B.P0VA"Uݹ6GR/,B;R/,B;)'NЃB;W$,BΞ)9"@@@&$NЃB;W$,BΞ)9"@@R/,B;R/,B;zxR/,BCMR/@BBR-P2KONJ768T7;2/ޟEŮß1QİL R/Ξ),BWβI3I@K/->ß1QİL R/Ξ),BWβI3I@R/,B;R/,B;)'NЃB;W$,BΞ)9"@@@&$NЃB;W$,BΞ)9"@@R/,B;R/,B;\Z(<7N6B=G;3>7K  #!<K  R/,B;R/,B;)'NЃB;W$,BΞ)9"@@@&$NЃB;W$,BΞ)9"@@R/,B;R/,B;)'N6@4,BHAR/D@Ԛ<&$N@4,BHAR/D@Ԛ<R/,B;R/,B;)'NЃB;W$,BΞ)9"@@@&$NЃB;W$,BΞ)9"@@R/,B;R/,B;/-R/,B@Hٟ@ʜ2IAN6@@@)'R/,B@Hٟ@ʜ2IAN@@R/,B;R/,B;)'NЃB;W$,BΞ)9"@@@&$NЃB;W$,BΞ)9"@@R/,B;R/,B;,*/>,BJ>,BJ>,BAB,*/>,BJ>,BJ>,BABB78;U B8;UB;U>C@KB;U>C@KB78;U B8;U,*ʡH9=7B;U>CEJCEJC7CC78N@>;GB B;>8N@>;GB78;U B8;U,*B7;>8N@Ɓ-67Ɓ-6HT#!B;>8N@ȁ-7ȁ-HTB78;U B8;U7B;U>C8,T7B;U>C,TB78;U B8;UB;ULC8,TB;ULC,TB78;U B8;UB;U>C@KB;U>C@KB78;U B8;U)'7B;U>CBU8JCBU8JC7CC7;UN8C.VI<7; B>;UN8C.I7FU/J.ʭB/ FJ.ϭBMKDVD:JTʭB/>ڶ>9ԚGJE@Ԛ<A?DVD:JTϭB>9ԚGJE@Ԛ<FU/J.ʭB/ FJ.ϭBJ.ʭB/@? J.ϭB@FU/J.ʭB/ FJ.ϭB#!J.ʭB/L FUO@KJ.ϭBL FO@KFU/J.ʭB/ FJ.ϭB)'J.ʭB/L F;F?8,T J.ϭBL F;F,TFU/J.ʭB/ FJ.ϭBMKDVD:JTʭB/>ڶ>9ԚGJE@Ԛ<A?DVD:JTϭB>9ԚGJE@Ԛ<FU/J.ʭB/ FJ.ϭB,*J.ʭB/L FUO'GNOC&$J.ϭBL FO'GNOCFU/J.ʭB/ FJ.ϭB#!J.ʭB/L FUO@KJ.ϭBL FO@KFU/J.ʭB/ FJ.ϭB,*J.ʭB/L FUOLBڶ>9ԚGJE@Ԛ<A?DVD:JTϭB>9ԚGJE@Ԛ<FU/J.ʭB/ FJ.ϭBJ.ʭB/>LJ.ϭB>LFU/J.ʭB/ FJ.ϭB#!J.ʭB/L FUO@KJ.ϭBL FO@KFU/J.ʭB/ FJ.ϭB53J.ʭB/8NJ.ʭB/G>98F>T,*J.ϭB8NJ.ϭBG>98F>FU/J.ʭB/ FJ.ϭBMKDVD:JTʭB/>ڶ>9ԚGJE@Ԛ<A?DVD:JTϭB>9ԚGJE@Ԛ<FU/J.ʭB/ FJ.ϭB/-J.ʭB/8IC¨03?;9<>TJ.ϭB8IϨ0-<>FU/J.ʭB/ FJ.ϭB#!J.ʭB/L FUO@KJ.ϭBL FO@KFU/J.ʭB/ FJ.ϭBJ.ʭB/;J6J.ϭB;J6G8ԓ4BWC=G8ԓ4BWC=SQDŽPB;8>׽RG>G8;?Sԓ459D0ԓ4B=R/AEATMKDŽPB;8>׽RG>G8;?Sԓ45D0ԓ4B=R/AEAG8ԓ4BWC=G8ԓ4BWC=/-?;8WB=&;WɾS2SCI9)'?;8WB=&;W2SC9G8ԓ4BWC=G8ԓ4BWC=SQDŽPB;8>׽RG>G8;?Sԓ459D0ԓ4B=R/AEATMKDŽPB;8>׽RG>G8;?Sԓ45D0ԓ4B=R/AEAG8ԓ4BWC=G8ԓ4BWC=/-H޽B;8AE0WB=щQUP.T,*H޽B;8AE0WB=щQUP.G8ԓ4BWC=G8ԓ4BWC=SQDŽPB;8>׽RG>G8;?Sԓ459D0ԓ4B=R/AEATMKDŽPB;8>׽RG>G8;?Sԓ45D0ԓ4B=R/AEAG8ԓ4BWC=G8ԓ4BWC=JHWBRPI9=50׽RG>G8;?Sԓ459D0ԓ4B=R/AEATMKDŽPB;8>׽RG>G8;?Sԓ45D0ԓ4B=R/AEAG8ԓ4BWC=G8ԓ4BWC=20PG,DNG806WB=C=S7,*PG,DNG85WB=CS7G8ԓ4BWC=G8ԓ4BWC=SQDŽPB;8>׽RG>G8;?Sԓ459D0ԓ4B=R/AEATMKDŽPB;8>׽RG>G8;?Sԓ45D0ԓ4B=R/AEAG8ԓ4BWC=G8ԓ4BWC=GEW=D,?R;G0G8DN@WG7ӽDIECӽDI>׽RG>G8;?Sԓ459D0ԓ4B=R/AEATMKDŽPB;8>׽RG>G8;?Sԓ45D0ԓ4B=R/AEAG8ԓ4BWC=G8ԓ4BWC=20޽BR0WB>=M>I?;8щQ@Ԛ<20޽BR0WB>=M>I?;8щQ@Ԛ<G8ԓ4BWC=G8ԓ4BWC=SQDŽPB;8>׽RG>G8;?Sԓ459D0ԓ4B=R/AEATMKDŽPB;8>׽RG>G8;?Sԓ45D0ԓ4B=R/AEAG8ԓ4BWC=G8ԓ4BWC=)';80WB=D>щQDSDA&$;80WB=ӗ>щQDSDAG8ԓ4BWC=G8ԓ4BWC=SQDŽPB;8>׽RG>G8;?Sԓ459D0ԓ4B=R/AEATMKDŽPB;8>׽RG>G8;?Sԓ45D0ԓ4B=R/AEAG8ԓ4BWC=G8ԓ4BWC=JHԓ459D0ԓ4B=SRJ>E;86ST!!";9ԓ45D0ԓ4B=SRJ>E;86STXG8ԓ4BWC=G8ԓ4BWC=SQDŽPB;8>׽RG>G8;?Sԓ459D0ԓ4B=R/AEATMKDŽPB;8>׽RG>G8;?Sԓ45D0ԓ4B=R/AEAG8ԓ4BWC=G8ԓ4BWC=#!;8>E6QWB=@N ;>E6QWB=@NG8ԓ4BWC=G8ԓ4BWC=SQDŽPB;8>׽RG>G8;?Sԓ459D0ԓ4B=R/AEATMKDŽPB;8>׽RG>G8;?Sԓ45D0ԓ4B=R/AEAG8ԓ4BWC=G8ԓ4BWC=R8G8>=>PR8G8>=>PG8ԓ4BWC=G8ԓ4BWC=SQDŽPB;8>׽RG>G8;?Sԓ459D0ԓ4B=R/AEATMKDŽPB;8>׽RG>G8;?Sԓ45D0ԓ4B=R/AEAG8ԓ4BWC=G8ԓ4BWC=VTԓ459D0ԓ4B=O׽RG6ST!!"DBԓ45D0ԓ4B=O׽RG6STXG8ԓ4BWC=G8ԓ4BWC=SQDŽPB;8>׽RG>G8;?Sԓ459D0ԓ4B=R/AEATMKDŽPB;8>׽RG>G8;?Sԓ45D0ԓ4B=R/AEAG8ԓ4BWC=G8ԓ4BWC=DBG׫;@2>H8GK0G8WB=F?HG,H,DBG׫;@2>H8GK0G8WB=F?HG,H,G8ԓ4BWC=G8ԓ4BWC=SQDŽPB;8>׽RG>G8;?Sԓ459D0ԓ4B=R/AEATMKDŽPB;8>׽RG>G8;?Sԓ45D0ԓ4B=R/AEAG8ԓ4BWC=G8ԓ4BWC=\Z7WCȻ22HG/CNK08W=ߌ,3=GGև9>TYW7WCȻ22HG/CNK08W=ߌ,3=GGև9>G8ԓ4BWC=G8ԓ4BWC=SQDŽPB;8>׽RG>G8;?Sԓ459D0ԓ4B=R/AEATMKDŽPB;8>׽RG>G8;?Sԓ45D0ԓ4B=R/AEAG8ԓ4BWC=G8ԓ4BWC=H$,GG88W-BGHHH$,GG88W-BGHHQH$,GG88W-BGHHDETLBL=,KH$,GG88W-BGHH$,GG88W-BGHQH$,GG88W-BGHDETLBL,KG8ԓ4BWC=G8ԓ4BWC=SQDŽPB;8>׽RG>G8;?Sԓ459D0ԓ4B=R/AEATMKDŽPB;8>׽RG>G8;?Sԓ45D0ԓ4B=R/AEAG8ԓ4BWC=G8ԓ4BWC=86G,DNG806WB=C=Pֈ;̛׽RG>G8;?Sԓ459D0ԓ4B=R/AEATMKDŽPB;8>׽RG>G8;?Sԓ45D0ԓ4B=R/AEAG8ԓ4BWC=G8ԓ4BWC=DBTCRJG<8QG8O60G6U<8Gڶ>S=86CJG<8QG8O60G6<8GS=G8ԓ4BWC=G8ԓ4BWC=SQDŽPB;8>׽RG>G8;?Sԓ459D0ԓ4B=R/AEATMKDŽPB;8>׽RG>G8;?Sԓ45D0ԓ4B=R/AEAG8ԓ4BWC=G8ԓ4BWC=DBS9I/CD<8JGԓ4GWB-RN= -KF7DBS9I/CD<8JGԓ4GWB-RN= -KF7 ʉ5  ʉ5ʉ5 @K ʉ5@K ʉ5  ʉ520 Ͳ4ʉ5/%DHGAAOC4ˉ5%DHAAOC ʉ5  ʉ5ʉ5 @K ʉ5@K ʉ5  ʉ5  ʉ5ޚTDG@K5D@K ʉ5  ʉ5ʉ5 @K ʉ5@K ʉ5  ʉ5&$ ۚKʉ5RG̛<"&ۚK݉5G̛<" ʉ5  ʉ5ʉ5 @K ʉ5@K ʉ5  ʉ5ʉ5 8,T ʉ5,T ʉ5  ʉ5ʉ5 @K ʉ5@K ʉ5  ʉ5207 ʉ5ޚT4L/ȈXʉ5B-AB#!H6=>ʉ5B-AB ʉ5  ʉ5ʉ5 @K ʉ5@K ʉ5  ʉ5;9 ۚK4ʉ5G8OE>έ;LSDʡH9;,*ۚK4ʉ5GOE>٭;SDʡH9; ʉ5  ʉ5ʉ5 @K ʉ5@K ʉ5  ʉ5#!@ >ʉ5DSDA@>ʉ5DSDA ʉ5  ʉ5ʉ5 @K ʉ5@K ʉ5  ʉ5&$$6 6ʉ5@Ԛ<$66ʉ5@Ԛ<,*,BA@D64AE54A6O&$,BA@64AE54A6/-D4A,HB54A6OUP.T)'D4A,HB54A6UP.,*,BA@D64AE54A6O&$,BA@64AE54A6 @K-; @K-;,*,BA@D64AE54A6O&$,BA@64AE54A6/-D4A,HB54A6OUP.T)'D4A,HB54A6UP.,*,BA@D64AE54A6O&$,BA@64AE54A653.HB@M64A6OI0щQUP.T/-.HB@M64A6I0щQUP.,*,BA@D64AE54A6O&$,BA@64AE54A6/-D4A,HB54A6OUP.T)'D4A,HB54A6UP.,*,BA@D64AE54A6O&$,BA@64AE54A6><.HB@M64A6OHAVTJD8DAP20.HB@M64A6HAVTD8A,*,BA@D64AE54A6O&$,BA@64AE54A6/-D4A,HB54A6OUP.T)'D4A,HB54A6UP.,*,BA@D64AE54A6O&$,BA@64AE54A6)'.49B3I6OFUPUT#!.49B3I6FUPU,*,BA@D64AE54A6O&$,BA@64AE54A6/-D4A,HB54A6OUP.T)'D4A,HB54A6UP.,*,BA@D64AE54A6O&$,BA@64AE54A6A?ڤ55D>.1B@D4A= @6OG;P20ܤ5D>.1B@4A= @6G;P,*,BA@D64AE54A6O&$,BA@64AE54A6/-D4A,HB54A6OUP.T)'D4A,HB54A6UP.,*,BA@D64AE54A6O&$,BA@64AE54A6GE.JS=HB@DH4ADAP;0T?6T)!,*.SHB@H4AA;T6T),*,BA@D64AE54A6O&$,BA@64AE54A6/-D4A,HB54A6OUP.T)'D4A,HB54A6UP.,*,BA@D64AE54A6O&$,BA@64AE54A620.BKM4AHAVTJD8DAP)'.BKM4AHAVTD8A,*,BA@D64AE54A6O&$,BA@64AE54A6/-D4A,HB54A6OUP.T)'D4A,HB54A6UP.,*,BA@D64AE54A6O&$,BA@64AE54A6/-.HB@D4A=6OGUP9T#!.HB@4ASGUP9 @GMT  @GM>  BIɤU1.@GMTC3G9/-VN>BIɤU1.@GMC3G9 @GMT  @GMDB@G.MTA/B@G.MTQ8ޚTNGKTOT,*@G6A/@G6Q8+KTO @GMT  @GM;9@G.MT,;MT73;E=57TIַ;)'@G6,;M7;E57TI @GMT  @GM)'@G@MT/-56P9?ַ;#!@G@M/-56P9? @GMT  @GM@G.MTG@=@GMT.@MTC3G3G9ܞNTTOC3G98Iַ;@G.MTG@=@GMT.@MTki@G6G=@GM.@MC3G3G9NTC3G9I@G6G=@GM.@M @GMT  @GM86@G.4@ϚL4MT;M4߹-WHԓ6Iַ;&$@G.@4M6߹-WHԓ6I @GMT  @GM&$@G.MT߹-5TOOIַ;@G6߹-5TOI @GMT  @GM.@MT.MT@MTܞND>.MTE=.MT=.MTIϪJ1.M@G.@MTDC3G98Iַ;\Z.@M6@MN>6E=6=6IϪJ1.M@G.@MDC3G9I @GMT  @GM&$.M@GMTJ-U@ؙDT#!.M@GMJ-U@ؙDT @GMT  @GM)''=.@GMTIB.<.M@GM6.@M@MEM=6C3G9I @GMT  @GM/-@G=@G.MT=.MTIG@ @G=@G6=6IG @GMT  @GM>  BIɤU1.@GMTC3G9/-VN>BIɤU1.@GMC3G9 @GMT  @GMDBMU@G@MT@MTMTMU,HP5ѳBʈFP?53M@G@M@MMM,HP5ѳBʈFP? @GMT  @GM;9@G.MT,;MT73;E=57TIַ;)'@G6,;M7;E57TI @GMT  @GM53@G.MTַ;@G.MTD,BPַ;Υ6&$@G6ַ;@G6D,Pַ;Υ6cI6;0ڳQ  +0ڳQ I6;ٟ@9ٟ@0A@Ԛ<+90A@Ԛ<KI6;0ڳQ  +0ڳQI6;-N  +-NI6;0ڳQ  +0ڳQ20I6;0ʭBќ:-WI6;I6>S2&$+0ʭBќ:-WI6I6>SoI6;0ڳQ  +0ڳQ&$UII6;-N1D@@@UI+-N1ځD@I6;0ڳQ  +0ڳQSQI6;096WI-:PUPޜFTI—PRMTI6ޜF6JH+096WI-:PUPޜFTIRMTI6ޜF6I6;0ڳQ  +0ڳQA?Q2?EC=E@.=9QCB9QCͦ(!)'Q2?EC=@ƋQCBƋQCiI6;0ڳQ  +0ڳQ .IWI6;8TAB.IW+8TABI6;0ڳQ  +0ڳQ86I6;6U=9=>C<ʡH6IHC<ʡH6IHTI6>6;DPDA86>I6>6;DPDA<I6>6;DPDA86>I6>6;DPDA,*ä=FBNLI6>6;DPDA86>I6>6;DPDA3PϪJBE҄JJ9R>9ֈDCSW9ٟ@192D>9ED>9@S6;,DP>=/UP.T.M@D>3PϪJBEԄJ9R>DCSW@192D>BD>9@S6;,DP>=/UP.I6>6;DPDA86>I6>6;DPDAI6>6;DPDA86>I6>6;DPDA6E>6E>I6>6;DPDA86>I6>6;DPDAI6>6;DPDA86>I6>6;DPDASE>C=,B/7Ȼ;T=.LGENA=C,B/7Ȼ;T=LȥW> 1ڶ>SGȥW> 1SG,*A> Q5=Qڶ>SȥW@@@&$A> Q5=QSȥW@@ȥW> 1ڶ>SGȥW> 1SG53>W5CȥWG8E<=?N;†M8T)'>W5CȥWG8E<†M8ȥW> 1ڶ>SGȥW> 1SGDB> @GWC;9Q66BW4 ȥW@@@><> @GWC;9Q6BW4 ȥW@@ȥW> 1ڶ>SGȥW> 1SGhf;>>WȥW,:K>;=?N;7=?N;GTTT - !FD;>>WȥW,:K>;7GTTT+ȥW> 1ڶ>SGȥW> 1SG,*A> Q5=Qڶ>SȥW@@@&$A> Q5=QSȥW@@ȥW> 1ڶ>SGȥW> 1SGDB>W5CWȥWG8E<=?N;TTTG8̛<86>W5CWȥWG8E<TTTG8ȥW> 1ڶ>SGȥW> 1SGDB> @GWC;9Q66BW4 ȥW@@@><> @GWC;9Q6BW4 ȥW@@ȥW> 1ڶ>SGȥW> 1SG20ȥW>W2G/I֣.ŞG9/;7;20ȥW>W2G/I֣.ŞG9/;7;ȥW> 1ڶ>SGȥW> 1SG,*A> Q5=Qڶ>SȥW@@@&$A> Q5=QSȥW@@ȥW> 1ڶ>SGȥW> 1SG20> >QR@8S֗T7ȥW@@@/-> >QR@8S֗T7ȥW@@ȥW> 1ڶ>SGȥW> 1SGDB> @GWC;9Q66BW4 ȥW@@@><> @GWC;9Q6BW4 ȥW@@ȥW> 1ڶ>SGȥW> 1SG20> ȥWS8D0;T=?N;)'> ȥWS8D0;TȥW> 1ڶ>SGȥW> 1SG,*A> Q5=Qڶ>SȥW@@@&$A> Q5=QSȥW@@ȥW> 1ڶ>SGȥW> 1SG#!ȥWȥWKȥW,:ĝ ȥWȥWKȥW,:؝ȥW> 1ڶ>SGȥW> 1SGDB> @GWC;9Q66BW4 ȥW@@@><> @GWC;9Q6BW4 ȥW@@ȥW> 1ڶ>SGȥW> 1SG_]N9UL=>˾3ȥW> G/NIǡ6TTT=?N;T!DBNU=>˾3ȥW> G/NIǡ6TTTTK  ?J=  ?J=)'VHDJ>4=5D3Ȼ;>T VD>4=5D3Ȼ;>  ?J=  ?J=DJ>?=DJ>?=}  ?J=  ?J=?J=Uа.T ?J=*  ?J=  ?J= J?,= J?,=  ?J=  ?J=;9?EJ=׍Q7E70 NʡH -H064T53?EJ=׍Q,0 NʡH -H064T  ?J=  ?J=DJ>?=GĊA>TDJ>?=GĊA>  ?J=  ?J=/-D9DDG?>J>,NDSDA#!9G?>J>=DSDA  ?J=  ?J=?EJ׍QDG@K?EJ׍QD@K  ?J=  ?J= D/F;  DF;  ?J=  ?J=,*DJ>?=E?NKLF9@K)'DJ>?=E?NKLF9@  ?J=  ?J=?=EJ=׍QPB6?=EJ=׍QPB  ?J=  ?J=;?1KEJ>=׍QCPDCK9K>ٟ@9@9W>4R/ҾWB1.O>NB9KJK>N9͝,ڪ3.WȻBDEA¶7ģC:Q;?1KEJ>=׍QCPDC9>ٟ@9@9W>4R/ҾWB1.O>NB8J>N9Ν,.WȻBDENģC:Q  ?J=  ?J=20?>?J>,N166==@Ԛ<)'?>?J>=16=@Ԛ<  ?J=  ?J=DJ>?=@KDJ>?=@K  ?J=  ?J=>?=4FSCܞN/OJ-0E/-DJ>?=4FSNOJ7E  ?J=  ?J=?J=4Н?A3AT?J=4AA  ?J=  ?J=)'VHDJ>4=5D3Ȼ;>T VD>4=5D3Ȼ;>  ?J=  ?J= ?EJ=׍QFK AB ?EJ=׍QFK AB  ?J=  ?J=?J=Uа.T ?J=*  ?J=  ?J=)'J>?=ʡH۩RV-T.6.T&$J>?=ʡH۩RV-T.6.  ?J=  ?J=;9?EJ=׍Q7E70 NʡH -H064T53?EJ=׍Q,0 NʡH -H064T  ?J=  ?J=20UWX=6?KJJ=3WН?>AT,*UWX=6?KJJ=3W?A  ?J=  ?J=/-D9DDG?>J>,NDSDA#!9G?>J>=DSDA  ?J=  ?J=864?߸3ѝ6B5-0IJ?߸3==I̛=F>>@>T#!DJ7>=F>>@>IFET> IFT>_]IFE71UC56K7WE>VWA75SJS24.@7Uև9>TVTIF71UC6K7WE>VWA75SJS24.@7Uև9>IFET> IFT>&$1FEWK.WKC:ET1FWKWKC:EIFET> IFT>,*IFED6AS1F՟?>>@Ԛ<#!IFD6Aū1?>>@Ԛ<IFET> IFT>;9IFEAW̋?6FF1UK>626::@20IFA̋?6.1UK>626::@IFET> IFT>_]IFE71UC56K7WE>VWA75SJS24.@7Uև9>TVTIF71UC6K7WE>VWA75SJS24.@7Uև9>IFET> IFT>&$IKMFE->CϨHQRTIKMF-CΨQRTIFET> IFT>,*IFED6AS1F՟?>>@Ԛ<#!IFD6Aū1?>>@Ԛ<IFET> IFT>20IFED6AS1F՟?>>DSDA)'IFD6Aū1?>>DSDAIFET> IFT>_]IFE71UC56K7WE>VWA75SJS24.@7Uև9>TVTIF71UC6K7WE>VWA75SJS24.@7Uև9>IFET> IFT>53FE>>M*ɬI*I*55TH>M*ɬI*I*5THTIFET> IFT>,*IFED6AS1F՟?>>@Ԛ<#!IFD6Aū1?>>@Ԛ<IFET> IFT>53HFE>>@IU>J-F>TLP20HF>>@IU>J-F>TLPIFET> IFT>_]IFE71UC56K7WE>VWA75SJS24.@7Uև9>TVTIF71UC6K7WE>VWA75SJS24.@7Uև9>IFET> IFT>20I—P=E>>FEDH>QIB,ܔN)'I=E>>FDH>QIBG DEO; DEO;:O;J@Ԛ<:;J@Ԛ< DEO; DEO;:O;4P@Ԛ<:;4P@Ԛ< DEO; DEO;  N:O;BF8@K N:;BF8@K DEO; DEO; -:O;WL/?T -:;W. DEO; DEO;:O;J@Ԛ<:;J@Ԛ< DEO; DEO;O:4;DG@KO:;D@K DEO; DEO;  N:O;BF8@K N:;BF8@K DEO; DEO;:OD>;@K:D>;@K DEO; DEO;:O;J@Ԛ<:;J@Ԛ< DEO; DEO; DO;2  D;2 DEO; DEO;  N:O;BF8@K N:;BF8@K DEO; DEO; :O;2,LDG@K:;2,D@K DEO; DEO;:O;J@Ԛ<:;J@Ԛ< DEO; DEO; :O;2  :;2 DEO; DEO;  N:O;BF8@K N:;BF8@K DEO; DEO;:O;28,T:;2,T DEO; DEO;:O;J@Ԛ<:;J@Ԛ< DEO; DEO;&$ :OƔ>;21ET!! :Ɣ>;21ET DEO; DEO;  N:O;BF8@K N:;BF8@K DEO; DEO;:O;28,T:;2,T DEO; DEO;:O;J@Ԛ<:;J@Ԛ< DEO; DEO; P:O8;:I̺@:TP:8;:@ DEO; DEO;  N:O;BF8@K N:;BF8@K DEO; DEO;#!:O;J:O4974T:;J:474T DEO; DEO;:O;J@Ԛ<:;J@Ԛ< DEO; DEO;DO;2:TD;2:T DEO; DEO;  N:O;BF8@K N:;BF8@K DEO; DEO;:O;2DG@K:;2D@K DEO; DEO;:O;J@Ԛ<:;J@Ԛ< DEO; DEO;:O;27Cͺ?@Ԛ<(TR;>ͺ?@Ԛ< (T;ͺ? (T;ͺ?&$/IMTS;ͺ?ٟ@6A7BITS;ͺ?5+ (T;ͺ? (T;ͺ?(TR;>ͺ?@Ԛ<(TR;>ͺ?@Ԛ< (T;ͺ? (T;ͺ?&$(TS;>6/IM@@@(TS;>6I@@ (T;ͺ? (T;ͺ?(TR;>ͺ?@Ԛ<(TR;>ͺ?@Ԛ< (T;ͺ? (T;ͺ?86/IMTR;>>VBͺ?C7=V-AB)'ITR;>>Bͺ?C7VAB (T;ͺ? (T;ͺ?(TR;>ͺ?@Ԛ<(TR;>ͺ?@Ԛ< (T;ͺ? (T;ͺ?(TS64ͺ?(TS64ͺ? (T;ͺ? (T;ͺ?(TR;>ͺ?@Ԛ<(TR;>ͺ?@Ԛ< (T;ͺ? (T;ͺ? /IMT;ͺ?DSDAIT;ͺ?DSDA (T;ͺ? (T;ͺ?(TR;>ͺ?@Ԛ<(TR;>ͺ?@Ԛ< (T;ͺ? (T;ͺ?A?/IMPD;Fͺ?M7K/1I-I-@Ԛ<53IPD;Fͺ?MK/I-I-@Ԛ< (T;ͺ? (T;ͺ?(TR;>ͺ?@Ԛ<(TR;>ͺ?@Ԛ< (T;ͺ? (T;ͺ? /IMF̽>S6>NBIF̽>S6>NB (T;ͺ? (T;ͺ?(TR;>ͺ?@Ԛ<(TR;>ͺ?@Ԛ< (T;ͺ? (T;ͺ?53;ͺ?9T./I/J@/TA/IMT,*;ͺ?9T.I/J@/TAIT (T;ͺ? (T;ͺ?(TR;>ͺ?@Ԛ<(TR;>ͺ?@Ԛ< (T;ͺ? (T;ͺ?86/IM̺ٟ@6ʔ7;Vͺ?2(/IMI@)'I̺5ʔ7;Vͺ?2(II@G>SEU G>S8/-L44ȣ8G>EUSIBEU̍ L4ȣ8G>8SIB8G>SEU G>S8&$UJG>SIBEU3H8UG>SIB8H8G>SEU G>S8/-L44ȣ8G>EUSIBEU̍ L4ȣ8G>8SIB8G>SEU G>S8zx7HܞNDG>SEU7HܞNDG>SEUQ7HܞNDG>SEUDET߹-8Lԓ6Iַ;C=.b`7HNG>S87HNG>S8Q7HNG>S8DET߹-8Lԓ6IC=G>SEU G>S8/-L44ȣ8G>EUSIBEU̍ L4ȣ8G>8SIB8G>SEU G>S8ki/K@G>SEUSTSUQ=WBSEUSIBEU߹-=EMSIַ;BU1TPN/KG>S8SŘSEU G>S8/-L44ȣ8G>EUSIBEU̍ L4ȣ8G>8SIB8G>SEU G>S8DB-ܞNDG>! )SEUQ-Q;ۓRTCG0/--NG>S8Q-Q;ۓRCG0G>SEU G>S8/-L44ȣ8G>EUSIBEU̍ L4ȣ8G>8SIB8G>SEU G>S886RNUG>SEUIBSEU) :/B#!NG>S8IBS8:/G>SEU G>S8/-L44ȣ8G>EUSIBEU̍ L4ȣ8G>8SIB8G>SEU G>S8/-ܞNDG>SIBEU;SIBEU&$NG>SIB8;SIB8G>SEU G>S8/-L44ȣ8G>EUSIBEU̍ L4ȣ8G>8SIB8G>SEU G>S8SQ-ܞNDG>SEUQD2VFȣ84XIUҔB<֗TI7Iַ;ŒATJH-NG>S8QD2VFȣ84XIUҔB<֗TI7IŒATG>SEU G>S8/-L44ȣ8G>EUSIBEU̍ L4ȣ8G>8SIB8G>SEU G>S8 P PG>SEU G>S8/-L44ȣ8G>EUSIBEU̍ L4ȣ8G>8SIB8G>SEU G>S820DGIBEUSEUV;EUBEU#!DGIB8S8V8B8 DBCDIٟ@964>DBR@54>,*BDCD94>,6ODPDA&$BDR94>,6DPDA DBCDIٟ@964>DBR@54>#!BCDO94>6O@Ԛ<BRO94>6@Ԛ< DBCDIٟ@964>DBR@54>DBDBD>CD.NA>%>R6Iٟ@97DSDA53B>R.NA>%>R6@4DSDA DBCDIٟ@964>DBR@54>20DCD97UDE4Oٟ@6AA7B DR5UDMOٟ@6+ DBCDIٟ@964>DBR@54>,*BDCD94>,6ODPDA&$BDR94>,6DPDA DBCDIٟ@964>DBR@54>PNDBCDIٟ@964E>йSDK9ٟ@9SM>BU-щQ@Ԛ<>޹S @9SM>BU-щQ@Ԛ< DBCDIٟ@964>DBR@54>DBDBD>CD.NA>%>R6Iٟ@97DSDA53B>R.NA>%>R6@4DSDA DBCDIٟ@964>DBR@54>PNOD6>D=7ADBDCD=9>DIٟ@OD2O@@@>D=7ABR9>D@OD2O@@ DBCDIٟ@964>DBR@54>,*BDCD94>,6ODPDA&$BDR94>,6DPDA DBCDIٟ@964>DBR@54>b`DBDCDCٟ@9ɤKE7>RɤK/ϪJ>H=Q996ɤKA>A910TDPNBR@9ɤKE7>RɤK/ϪJ>H=Q95ɤKA>A10TD DBCDIٟ@964>DBR@54>DBDBD>CD.NA>%>R6Iٟ@97DSDA53B>R.NA>%>R6@4DSDA DBCDIٟ@964>DBR@54>_]DBCDN59OH348BD4R4O@4WOŮPO4/TDOTDBDBRN5O38BD4MO@4WX޵+TOT DBCDIٟ@964>DBR@54>,*BDCD94>,6ODPDA&$BDR94>,6DPDA DBCDIٟ@964>DBR@54>#!BCD94>A6O@Ԛ<BR94>A6@Ԛ< DBCDIٟ@964>DBR@54>DBDBD>CD.NA>%>R6Iٟ@97DSDA53B>R.NA>%>R6@4DSDA DBCDIٟ@964>DBR@54> UCD94>A6?,UR94>A6?, BU06˩5FE91PBU06FE1PA?BTS6˩5؇9?˩5OMR9I1FUFFPJ86BS6؇9?˩5OMR9I1FUFPJ BU06˩5FE91PBU06FE1P/-B؇96˩5ֲR1FQ?ٟ@SPG3&$B؇96ֲR1FQ?ٟ@SG BU06˩5FE91PBU06FE1P20UC>B06˩5NR31SFщQ@Ԛ</-UC>B06NR31SFщQ@Ԛ< BU06˩5FE91PBU06FE1P#!BOFR6˩5֛7>3PJBOFR673PJ BU06˩5FE91PBU06FE1PA?BTS6˩5؇9?˩5OMR9I1FUFFPJ86BS6؇9?˩5OMR9I1FUFPJ BU06˩5FE91PBU06FE1PkiBTS6˩50QN?9H9RIJIН?TXLI/I/I/B=6I6B=-0YWBS60QN?9H9RIJIܞ?ɜXI/II/B=6I6B=0 BU06˩5FE91PBU06FE1P20UC>B06˩5NR31SFщQ@Ԛ</-UC>B06NR31SFщQ@Ԛ< BU06˩5FE91PBU06FE1PDB05OȨKFD9IVBTELȨKF9IV:TН?>/-05OӨKDIVBELӨKIV:? BU06˩5FE91PBU06FE1PA?BTS6˩5؇9?˩5OMR9I1FUFFPJ86BS6؇9?˩5OMR9I1FUFPJ BU06˩5FE91PBU06FE1P20BTS6˩50BT6˩51T7H;T#!BS60B617H;T BU06˩5FE91PBU06FE1P20UC>B06˩5NR31SFщQ@Ԛ</-UC>B06NR31SFщQ@Ԛ< BU06˩5FE91PBU06FE1P BT66˩50QGН?>B660QG? BU06˩5FE91PBU06FE1PA?BTS6˩5؇9?˩5OMR9I1FUFFPJ86BS6؇9?˩5OMR9I1FUFPJ BU06˩5FE91PBU06FE1PBR6˩51?FBTBR61?BT BU06˩5FE91PBU06FE1P20UC>B06˩5NR31SFщQ@Ԛ</-UC>B06NR31SFщQ@Ԛ< BU06˩5FE91PBU06FE1P BT؇96˩5M5RFFB؇96M5RFFCE.TGTK6TCE.TGK6T><.,9T>GTP6ڜ>KDTWAПCDSDA;9.,9T>GP6ڜ>KDTWAПCDSDACE.TGTK6TCE.TGK6T.TGT6>?>P.TG6>?>PCE.TGTK6TCE.TGK6T><.,9T>GTP6ڜ>KDTWAПCDSDA;9.,9T>GP6ڜ>KDTWAПCDSDACE.TGTK6TCE.TGK6T@N>PC @N>PCE.TGTK6TCE.TGK6T><.,9T>GTP6ڜ>KDTWAПCDSDA;9.,9T>GP6ڜ>KDTWAПCDSDACE.TGTK6TCE.TGK6T;9G߹-.TGTޚT>9BKR9KϋIL‡KAB20G߹-.TGޚT>BR9KϋIL‡KABCE.TGTK6TCE.TGK6T><.,9T>GTP6ڜ>KDTWAПCDSDA;9.,9T>GP6ڜ>KDTWAПCDSDACE.TGTK6TCE.TGK6T)'.TGT6>7KM?U>T .TG6>7KM?,CE.TGTK6TCE.TGK6T><.,9T>GTP6ڜ>KDTWAПCDSDA;9.,9T>GP6ڜ>KDTWAПCDSDACE.TGTK6TCE.TGK6T20.T9Kʉ55>A>BK=U;Н?T)'.T9Kʉ55>A>BU;ܞ?CE.TGTK6TCE.TGK6T><.,9T>GTP6ڜ>KDTWAПCDSDA;9.,9T>GP6ڜ>KDTWAПCDSDACE.TGTK6TCE.TGK6T/-.T9Kʉ55>A>BK=3RT&$.T9Kʉ55>A>B3ҔRCE.TGTK6TCE.TGK6T><.,9T>GTP6ڜ>KDTWAПCDSDA;9.,9T>GP6ڜ>KDTWAПCDSDACE.TGTK6TCE.TGK6TDBO<>TRIO.TRIOVTIOB<ȬTIQ>86O<>TRO.TROVTIOB<ЬTQCE.TGTK6TCE.TGK6T><.,9T>GTP6ڜ>KDTWAПCDSDA;9.,9T>GP6ڜ>KDTWAПCDSDACE.TGTK6TCE.TGK6TMK.TGT9.DS>9>AK@—PB@ A6O:@@@><.TG9.DS>IAK@B@ A6:@@CE.TGTK6TCE.TGK6T><.,9T>GTP6ڜ>KDTWAПCDSDA;9.,9T>GP6ڜ>KDTWAПCDSDACE.TGTK6TCE.TGK6T.TGT6>@Ԛ<.TG6>@Ԛ<̾-,AJT0ޡ8;-AJT0AJT?@KAJT?@K̾-,AJT0ޡ8;-AJT0;9̾-,4FE4AJT54T?1WQ̛<7T/-;-4E4AJT54T?WQ̛<7̾-,AJT0ޡ8;-AJT0\Z̾--,̾-,6.ΩWH8443AJT4Q۹/85?1WK40ޡ8>1MK̾--,;-6ΩWH8443AJT4Q۹/85?WåK0>1̾-,AJT0ޡ8;-AJT0PN̾--H,̾-,6.ΩW4L5/B4W5H/OTANя7>1GE̾--H,;-6ΩW4L5/B4W5H/OTAN>1̾-,AJT0ޡ8;-AJT0,*7H984AJT54>0ޡ87̾-,AJT0ޡ8;-AJT0 ̾-/XT ̾-/XT̾-,AJT0ޡ8;-AJT0  ̾-CT  ̾-CT̾-,AJT0ޡ8;-AJT0,*H84AJT540ޡ8>1@K)'H84AJT540>1@K̾-,AJT0ޡ8;-AJT0AJT?@KAJT?@K̾-,AJT0ޡ8;-AJT0\ZH,̾-,XΩW84ALT540ޡ8>11DD>7U ̾-X̾-X-TMKH,;-X84ALT540>11D>7U ̾-X̾-X-̾-,AJT0ޡ8;-AJT0\Z̾--,̾-,6.ΩWH8443AJT4Q۹/85?1WK40ޡ8>1MK̾--,;-6ΩWH8443AJT4Q۹/85?WåK0>1̾-,AJT0ޡ8;-AJT0JH8AJTOC6̾-,84L5/TS:-1QBU/;868AJTOC6;-84L5/S:-ڠ#/̾-,AJT0ޡ8;-AJT0,*7H984AJT54>0ޡ87̾-,AJT0ޡ8;-AJT0/-̾-//?84AJT5T;U/T&$̾-//84AJT5T;*̾-,AJT0ޡ8;-AJT0  ̾-CT  ̾-CT̾-,AJT0ޡ8;-AJT0;9̾-,EAJTTDɍPMA:7.U/T/-;-EAJTTDӍPA:7.*̾-,AJT0ޡ8;-AJT0AJT?@KAJT?@K̾-,AJT0ޡ8;-AJT0ILIL̾-,AJT0ޡ8;-AJT0\Z̾--,̾-,6.ΩWH8443AJT4Q۹/85?1WK40ޡ8>1MK̾--,;-6ΩWH8443AJT4Q۹/85?WåK0>1̾-,AJT0ޡ8;-AJT0&$̾-4AT95/?V/?T ̾-4AT95/@?T)'FBUQDND6S?F: FBUQND6S?,*7F:BP1ND?F:@Ԛ<#!7:BP1ND?@Ԛ<)'FBUQDND6S?F: FBUQND6S?R6!8,TR6,T)'FBUQDND6S?F: FBUQND6S?207F:BP1ND?F:6S@Ԛ<)'7:BP1ND?6S@Ԛ<)'FBUQDND6S?F: FBUQND6S?531K>QP?F:Bб 4D=3-AB,*1K>QP?Bб 4D=-AB)'FBUQDND6S?F: FBUQND6S?,*7F:BP1ND?F:@Ԛ<#!7:BP1ND?@Ԛ<)'FBUQDND6S?F: FBUQND6S?;94F:̔6BUPV715CS?F:@Ԛ<204:̔6BUPV715CS?@Ԛ<)'FBUQDND6S?F: FBUQND6S?207F:BP1ND?F:6S@Ԛ<)'7:BP1ND?6S@Ԛ<)'FBUQDND6S?F: FBUQND6S?&$̔6ַ;IBUVԋ/CS?F:1IBUVԋ/CS?)'FBUQDND6S?F: FBUQND6S?,*7F:BP1ND?F:@Ԛ<#!7:BP1ND?@Ԛ<)'FBUQDND6S?F: FBUQND6S?GEF:̔6BU>ȣ89071KK6S?F:DSDA><:̔6BU>ȣ89071KK6S?DSDA)'FBUQDND6S?F: FBUQND6S?207F:BP1ND?F:6S@Ԛ<)'7:BP1ND?6S@Ԛ<)'FBUQDND6S?F: FBUQND6S? DA7O=—PRߑ4PTDA7=Rߑ4PT)'FBUQDND6S?F: FBUQND6S?,*7F:BP1ND?F:@Ԛ<#!7:BP1ND?@Ԛ<)'FBUQDND6S?F: FBUQND6S? ?F:6S>JK2@Ԛ<JHU2QNDHF/@SKDND SC>K2@Ԛ<UNDHF/UDHF/#!0-0:Nٟ@HFVFT0-:ٟ@HFFTUNDHF/UDHF/GEM:5UND8F/?PS6 1B>UDF?PS6UNDHF/UDHF/PNU2QN5DHF/Bٟ@SKDND SC>K2@Ԛ<JHU2QNDHF/@SKDND SC>K2@Ԛ<UNDHF/UDHF//-ӟ;N@R>8FS/"ҥ3!@;6&$ӟ;NR>8FS"ҥ3!@6UNDHF/UDHF/GEM:5K2@Ԛ<JHU2QNDHF/@SKDND SC>K2@Ԛ<UNDHF/UDHF/&$ FS5/ FS/UNDHF/UDHF/GEM:5K2@Ԛ<JHU2QNDHF/@SKDND SC>K2@Ԛ<UNDHF/UDHF/>/@K8FENܜ>@K  ,ݠ.A,A_]O70CT,ݠ.7>DGܤKP04TVAV07>?Q;GEO߫B>GK04TVAV07>?Q;  ,ݠ.A,A86,ݠ.>O/19O616ABTGA7B/-,>O/19O616ABTG+  ,ݠ.A,A,ݠ.ݠ.O ,ݠ.O  ,ݠ.A,AO,ݠ.B:DG@KO,BD@K  ,ݠ.A,A_]O70CT,ݠ.7>DGܤKP04TVAV07>?Q;GEO߫B>GK04TVAV07>?Q;  ,ݠ.A,A CN,ݠ.QADPDACN,QADPDA  ,ݠ.A,A,ݠ.ݠ.O ,ݠ.O  ,ݠ.A,A)'Iַ;DN0CT,ݠ.AщQ@Ԛ<Iַ;DNAщQ@Ԛ<  ,ݠ.A,A_]O70CT,ݠ.7>DGܤKP04TVAV07>?Q;GEO߫B>GK04TVAV07>?Q;  ,ݠ.A,A,*>T,ݠ.9ABAA4˛5DA4>,9ABA˛5DAn  ,ݠ.A,A,ݠ.ݠ.O ,ݠ.O  ,ݠ.A,A NT,ݠ.Nĵ*  ,ݠ.A,A_]O70CT,ݠ.7>DGܤKP04TVAV07>?Q;GEO߫B>GK04TVAV07>?Q;  ,ݠ.A,APNX>T9;;>X>QA7AO7RN;X7:U>E8DBX>;>X>QA7AO7N;X7:U>E8  ,ݠ.A,A,ݠ.ݠ.O ,ݠ.O  ,ݠ.A,AA?O߹-5,ݠ.߹-,ݠ.:߹-HİUMANC)O8,T53O߹-5,߹-,:߹-HMANC)O,TFIֈD:0DFI:0DIֈD:0@? I:0@FIֈD:0DFI:0D IֈDN0D:DG@KIN0DD@KFIֈD:0DFI:0D20D3ԚIֈD0>D:DSDA#!3IF>I>DDSDAFIֈD:0DFI:0DIֈDGC?DIGC?DFIֈD:0DFI:0DIֈD:0@? I:0@FIֈD:0DFI:0DD:IֈD14  DI1FIֈD:0DFI:0D20D3ԚIֈD0>D:DSDA#!3IF>I>DDSDAFIֈD:0DFI:0D20D3ԚIֈD0>D:DSDA#!3IF>I>DDSDAFIֈD:0DFI:0DIֈD:0@? I:0@FIֈD:0DFI:0D><يRIֈD:0DيR4IֈD:0BIيR4TC,>)'يRI:0DRI:0BIRCFIֈD:0DFI:0D20D3ԚIֈD0>D:DSDA#!3IF>I>DDSDAFIֈD:0DFI:0D;9IֈD>0EFR4:0>ğCѭDӮD:ٟ@H@Ԛ</-I>0EFM:0>ɟCܮDٟ@H@Ԛ<FIֈD:0DFI:0DIֈD:0@? I:0@FIֈD:0DFI:0D,*IֈD:0ߢ?DT7N79UAT#!I:0ߢ?D7N79UAFIֈD:0DFI:0D20D3ԚIֈD0>D:DSDA#!3IF>I>DDSDAFIֈD:0DFI:0D><0IֈD:0D04IֈD:0BI04TC,>/-0I:0D04I:0BI04CC@N= ->>C@N= ->>DB - -HE07!7LJ677/-H0LJ677C@N= ->>C@N= ->>;9E87CC@N.H˱U=FCסE@@@2087CC@N.H˱U=FC@@C@N= ->>C@N= ->>DB - -HE07!7LJ677/-H0LJ677C@N= ->>C@N= ->>GE= ->C@N7U0>ٟ@6MVIW>EDSDA><= ->C@NU0>ٟ@6V=>EDSDAC@N= ->>C@N= ->>DB - -HE07!7LJ677/-H0LJ677C@N= ->>C@N= ->>A?C@N= -F>EMӛ?ߤ8>4FC@N@@@;9C@N= -F>EMӛ?ߤ8>4C@N@@C@N= ->>C@N= ->>DB - -HE07!7LJ677/-H0LJ677C@N= ->>C@N= ->>DB= ->CסEC@NDE0**ԑ49A*/@@@;9= ->CC@ND0**ԑ49A*/@@C@N= ->>C@N= ->>DB - -HE07!7LJ677/-H0LJ677C@N= ->>C@N= ->>b`B˩55=>O*7C@N7C@ĕ6TFR/HFH4ĕ6TPNB˩55=>O*C@ĕ6TFR/HF4ĕ6TC@N= ->>C@N= ->>DB - -HE07!7LJ677/-H0LJ677C@N= ->>C@N= ->>_]6ɵO=>C@Nð.A ->>ٟ@9ٟ@DDܢESܤKA@CסESܤKA@Ԛ<SQ6ɵO=>C@Nð.A ->>9DܢESܤKA@CSܤKA@Ԛ<C@N= ->>C@N= ->>DB - -HE07!7LJ677/-H0LJ677C@N= ->>C@N= ->>/-7C@N7 -=FSÐW7#! -=FSÐW7C@N= ->>C@N= ->>DB - -HE07!7LJ677/-H0LJ677C@N= ->>C@N= ->>><= -F>C@NPEMӛ?M>>Fٟ@6@Ԛ<;9= -F>C@NPEMӛ?M>>F5@Ԛ<C@N= ->>C@N= ->>DB - -HE07!7LJ677/-H0LJ677C@N= ->>C@N= ->>)'C@N= -F(D>M@Ԛ<)'C@N= -F(D>M@Ԛ<HS/ON4/:HSON4/VT7H>S/OB4ʡH9ɰ5ȥ7/:į?I EL/ EHS/ON4/:HSON4/JHùBLW¶7/J7H>/B/WȥOB4784/:ĹBN/J84/HS/ON4/:HSON4/,*HS/ON4/:ҁX?L-T#!HSON4/ҁX?L-HS/ON4/:HSON4/&$HW>S/94/:AƭIHW>S94/AHS/ON4/:HSON4/)'7H/B/>ON47/:/HS/ON4/:HSON4/;9H>S/>4NO/://9¶7JùBL/#!>4NO//NJĹB/HS/ON4/:HSON4/;97H>S/>OB47BR:0ABBR:0ABHS/ON4/:HSON4/20/:7H>S/OB47/://HS/ON4/:HSON4/VT7H>S/OB4ʡH9ɰ5ȥ7/:į?I EL/ EHS/ON4/:HSON4/&$H>S/OB4">OB4>HS/ON4/:HSON4/,*HS/ON4/:ҁX?L-T#!HSON4/ҁX?L-HS/ON4/:HSON4/PN7HS/47/: ȥǶ,W¶7/>;GB20/ ȥǶ,N/>;GHS/ON4/:HSON4/)'7H/B/>ON47/:/HS/ON4/:HSON4/kiùBL9¶7/J7HS/9ȥ4NO7/:9¶7//:66ȈX4&20ĹBN/J/N//66ȈX4HS/ON4/:HSON4/;97H>S/>OB47BR:0ABBR:0ABHS/ON4/:HSON4/><ʡHU٨I7HS/47/::,AF> ʡHU٨I/:,>HS/ON4/:HSON4/VT7H>S/OB4ʡH9ɰ5ȥ7/:į?I EL/ EHS/ON4/:HSON4/HN1,;TLH1,;TLX:86˩5 X:86&$X:86˩5>X:86˩57 X:86>X:867X:86˩5 X:8686X:86˩54X:864VDT(!)'X:864X:86VDTX:86˩5 X:86&$X:86˩5>X:86˩57 X:86>X:867X:86˩5 X:86)'E8:X66˩5H38@@@#!E8:X66H38@@X:86˩5 X:86&$X:86˩5>X:86˩57 X:86>X:867X:86˩5 X:8620X:86ӻBOX:86˩5Q464T#!X:80X:86Q6TX:86˩5 X:86&$X:86˩5>X:86˩57 X:86>X:867X:86˩5 X:86GEX:86ӻBOX:86ӻBOX:86H6T$!20X:80X:86OX:866TX:86˩5 X:86&$X:86˩5>X:86˩57 X:86>X:867X:86˩5 X:86#!5X:8>6˩56R@Ԛ< 5X:8>66R@Ԛ<߹-U.:D>߹-U.:D>_]AU.?:D>>Rٟ@6U>G1@:?;0@W;0:GȻ;T=.LYWAU.?:D>>Rٟ@6>G1@:?;0@W;0:GȻ;T=L߹-U.:D>߹-U.:D> ߹-U.8 ߹-U.8߹-U.:D>߹-U.:D>#!߹-U.:DM=6S@Ԛ<#!߹-U.:DM=6S@Ԛ<߹-U.:D>߹-U.:D>/-U.>DP?14:щQȻ;T=.L)'U.>D?14:щQȻ;T=L߹-U.:D>߹-U.:D>ec1O .9.RU.,D6/EDCD>ڤ55J@C;ϵ>͵ATJ.PTVT1O.9.RU.,D6/EDCD>ܤ5J@Aϵ>͵ATJ.P߹-U.:D>߹-U.:D>)'кBU.6:DP߇;Ȼ;T=.L#!кBU.6:D߇;Ȼ;T=L߹-U.:D>߹-U.:D>_]AU.?:D>>Rٟ@6U>G1@:?;0@W;0:GȻ;T=.LYWAU.?:D>>Rٟ@6>G1@:?;0@W;0:GȻ;T=L߹-U.:D>߹-U.:D> U.8߹-U..ʺBPT U.8߹-U..ʺBPT߹-U.:D>߹-U.:D>#!߹-U.:DM=6S@Ԛ<#!߹-U.:DM=6S@Ԛ<߹-U.:D>߹-U.:D> ߹-U.:/0EFT6 ߹-U.:/0EFT6߹-U.:D>߹-U.:D>ec1O .9.RU.,D6/EDCD>ڤ55J@C;ϵ>͵ATJ.PTVT1O.9.RU.,D6/EDCD>ܤ5J@Aϵ>͵ATJ.P߹-U.:D>߹-U.:D>&$U.VP1B,, 5&$U.VP1B,, 5߹-U.:D>߹-U.:D>_]AU.?:D>>Rٟ@6U>G1@:?;0@W;0:GȻ;T=.LYWAU.?:D>>Rٟ@6>G1@:?;0@W;0:GȻ;T=L߹-U.:D>߹-U.:D>߹-V;T64߹-V;T6߹-U.:D>߹-U.:D>#!߹-U.:DM=6S@Ԛ<#!߹-U.:DM=6S@Ԛ<߹-U.:D>߹-U.:D>MK߹-U.6>P5,A߹-U.˭V6,3T߹-˭V6܈IU?90GE߹-U.6>P59߹-U.˭V6,3T߹-˭V6߈I?90߹-U.:D>߹-U.:D>ec1O .9.RU.,D6/EDCD>ڤ55J@C;ϵ>͵ATJ.PTVT1O.9.RU.,D6/EDCD>ܤ5J@Aϵ>͵ATJ.P߹-U.:D>߹-U.:D>MK N =.H= F0BU.$D: N =MPMPJH N =U= F0BU.$D: N =MPMP߹-U.:D>߹-U.:D>_]AU.?:D>>Rٟ@6U>G1@:?;0@W;0:GȻ;T=.LYWAU.?:D>>Rٟ@6>G1@:?;0@W;0:GȻ;T=L߹-U.:D>߹-U.:D>/-߹-U.DI429-DIV=RJ#!߹-U.D429-+RJ߹-U.:D>߹-U.:D>#!߹-U.:DM=6S@Ԛ<#!߹-U.:DM=6S@Ԛ<߹-U.:D>߹-U.:D>53߹-U./0Bб DD7=EUT۹/UD,*߹-U./0Bб DD7,U/D߹-U.:D>߹-U.:D>ec1O .9.RU.,D6/EDCD>ڤ55J@C;ϵ>͵ATJ.PTVT1O.9.RU.,D6/EDCD>ܤ5J@Aϵ>͵ATJ.P߹-U.:D>߹-U.:D>;9߹-U.0NUOބ2E   =ĪC'AB&$߹-U.0NUOǷ. =AB߹-U.:D>߹-U.:D>_]AU.?:D>>Rٟ@6U>G1@:?;0@W;0:GȻ;T=.LYWAU.?:D>>Rٟ@6>G1@:?;0@W;0:GȻ;T=L߹-U.:D>߹-U.:D>SQ:Aб =>U.=9V>D>9ԚU.V>D>1OISÄN989FT6߹-U.:D>߹-U.:D>#!߹-U.:DM=6S@Ԛ<#!߹-U.:DM=6S@Ԛ<߹-U.:D>߹-U.:D>)'U.>DP6:,Ȼ;T=.L#!U.>D6:,Ȼ;T=L߹-U.:D>߹-U.:D>ec1O .9.RU.,D6/EDCD>ڤ55J@C;ϵ>͵ATJ.PTVT1O.9.RU.,D6/EDCD>ܤ5J@Aϵ>͵ATJ.P߹-U.:D>߹-U.:D> ߹-1U.9TDSDA ߹-1U.9TDSDA߹-U.:D>߹-U.:D>_]AU.?:D>>Rٟ@6U>G1@:?;0@W;0:GȻ;T=.LYWAU.?:D>>Rٟ@6>G1@:?;0@W;0:GȻ;T=L߹-U.:D>߹-U.:D>GE3Ԛ9VC=6RMK -PT61TP6531U.D9VC6RK -PT61TڀP߹-U.:D>߹-U.:D>#!߹-U.:DM=6S@Ԛ<#!߹-U.:DM=6S@Ԛ<߹-U.:D>߹-U.:D>;9߹-U.:SM?B;BɵOMSB#**.T53߹-U.:SM?B;BֵOSB#**.  U8SUS/-1۠N -FɹKU=S5ۓR:ϡSFAT#!ޠN -FU=S5ۓR:ݡSA  U8SUS&$FMGMM>.3ˠS87TFMGMM>.87  U8SUSA?N,ˏR0#>ˌD3U=SNۥN&7><N,ˏR0#>ьDU=SNۥN&7  U8SUS)'VX?AM—PS>SM8GJ#!VX?AM—PS>SٶM1  U8SUS/-1۠N -FɹKU=S5ۓR:ϡSFAT#!ޠN -FU=S5ۓR:ݡSA  U8SUS/-U=Sб .65J?O4ʄ/&87&$U=Sб .6JO4ʄ/&8  U8SUSA?N,ˏR0#>ˌD3U=SNۥN&7><N,ˏR0#>ьDU=SNۥN&7  U8SUSVT70:7KU6A8>C¾98T—P7—PX>¾987;>C¾98—P7—PX>¾98;  U8SUS/-1۠N -FɹKU=S5ۓR:ϡSFAT#!ޠN -FU=S5ۓR:ݡSA  U8SUSA?U=Sб 7̛<87E7C77̛<(!53U=Sб 7̛<87E7C77̛<  U8SUSA?N,ˏR0#>ˌD3U=SNۥN&7><N,ˏR0#>ьDU=SNۥN&7  U8SUS86W7IU8>SESٟ@M߫U@U'@@@20W7IU>SESٟ@M߫U@U'@@  U8SUS/-1۠N -FɹKU=S5ۓR:ϡSFAT#!ޠN -FU=S5ۓR:ݡSA  U8SUS>< -4J6 NLF;8T786 -4J6 NLF87 N/,QEO. N/+O.53 N/,QI/E4OL/.DQET)' N/Q@E4OL/.T N/,QEO. N/+O.;9/,<7F NCN:QEI/4O5.L,*/<7F NN:+@4O5L N/,QEO. N/+O.53 N/,QI/E4OL/.DQET)' N/Q@E4OL/.T N/,QEO. N/+O.b`/QET N/QEVK/QEL9O/Լ=ET/QE/4/VQE1WJH/+T N/+V/+L9O/=T/+//V71W N/,QEO. N/+O.53 N/,QI/E4OL/.DQET)' N/Q@E4OL/.T N/,QEO. N/+O.&$ NCN/QEL)5E6>  )@?)@)ٟ@6E6>)5E6>)'UC9S;ٟ@>6E6>@Ԛ<)'UC9S;ٟ@>6E6>@Ԛ<)ٟ@6E6>)5E6>  )@?)@)ٟ@6E6>)5E6>&$VX,)EBٟ@&EϜVQTVX)E@&EϜVQ)ٟ@6E6>)5E6>  )@?)@)ٟ@6E6>)5E6>DBDԚ<(!ٟ@6ES>)%"6"&)'D5ES>)%"6")ٟ@6E6>)5E6>  )@?)@)ٟ@6E6>)5E6>#!;ښL)E6??OKT;ښL)E6?AT)ٟ@6E6>)5E6>  )@?)@)ٟ@6E6>)5E6>&$8V1)ٟ@>6E6>@Ԛ<#!81)ٟ@>6E6>@Ԛ<)ٟ@6E6>)5E6>  )@?)@)ٟ@6E6>)5E6>53K:S;ٟ@Sٟ@>6E66>GA7B/-K:S;ٟ@Sٟ@>6E66>G+)ٟ@6E6>)5E6>  )@?)@)ٟ@6E6>)5E6>\Z$B)KFE6>RDI6PGH>R5K9>66;NDSPԮK߀3VT$B)KFE6>RDI6PGH>RK9>66;NDSPٮK)ٟ@6E6>)5E6>  )@?)@)ٟ@6E6>)5E6>/-R—Pٟ@)%ٟ@6E6>DPDA&$R—Pٟ@)5E6>DPDA)ٟ@6E6>)5E6>  )@?)@)ٟ@6E6>)5E6>GE!Rٟ@6E6>ٟ@щQKB)B$&9U>щQ@Ԛ<;9!R5E6>ٟ@щQKB)BU>щQ@Ԛ< ?9T$ɞ9$A?W6J/?9T$8:W6J,HPHCI9I20WF/ɞ9$8:WF,HPHI9I ?9T$ɞ9$hf$>I?9TWO$8$>I?9TWO$8Q$>I?9TWO$8,9PMK$>ɞ9WO$8$>ɞ9WO$8Q$>ɞ9WO$8,9P ?9T$ɞ9$A?W6J/?9T$8:W6J,HPHCI9I20WF/ɞ9$8:WF,HPHI9I ?9T$ɞ9$)'֥>$8?9TW8QH.T#!֥>$8ɞ9W8QH.T ?9T$ɞ9$A?W6J/?9T$8:W6J,HPHCI9I20WF/ɞ9$8:WF,HPHI9I ?9T$ɞ9$)'U"҈$4T޲F?9TU"4T޲Fɞ9 ?9T$ɞ9$A?W6J/?9T$8:W6J,HPHCI9I20WF/ɞ9$8:WF,HPHI9I ?9T$ɞ9$trL:V1T>B;W8׫B!UH?I?9T$8CWO?98W8ɳQWQBHO_]L:V1T>;W8׫B!UH?Iɞ9$8CWOǞ9W8ɳQWвQH ?9T$ɞ9$A?W6J/?9T$8:W6J,HPHCI9I20WF/ɞ9$8:WF,HPHI9I ?9T$ɞ9$;9µ$?9Tµ$?9T@M@>KT@/Bɞ9ɞ9@ܱM>K@/ ?9T$ɞ9$A?W6J/?9T$8:W6J,HPHCI9I20WF/ɞ9$8:WF,HPHI9I ?9T$ɞ9$86<4T޲Fɞ9"A/4T޲Fɞ9"Q8+KTO ?9T$ɞ9$A?W6J/?9T$8:W6J,HPHCI9I20WF/ɞ9$8:WF,HPHI9I ?9T$ɞ9$&$?9TQ0"lj:?9TɳQQɞ9Q0"lj:ɞ9ɳQQ ?9T$ɞ9$A?W6J/?9T$8:W6J,HPHCI9I20WF/ɞ9$8:WF,HPHI9I ?9T$ɞ9$ec?9T88I?9T$8WO888O?98QD2CI0C98>ŒATSQɞ988Iɞ9$8WO888O?98QD2C0C9>ŒAT ?9T$ɞ9$A?W6J/?9T$8:W6J,HPHCI9I20WF/ɞ9$8:WF,HPHI9I ?9T$ɞ9$20?9T$8ܞND֥>W8ݶ;UW89T#!ɞ9$8N֥>W΀8U89T ?9T$ɞ9$A?W6J/?9T$8:W6J,HPHCI9I20WF/ɞ9$8:WF,HPHI9I ?9T$ɞ9$,*"҈$4T޲F?9Tlj:""4T޲Fɞ9lj:" ?9T$ɞ9$A?W6J/?9T$8:W6J,HPHCI9I20WF/ɞ9$8:WF,HPHI9I ?9T$ɞ9$53"҈$4T޲F?9TQD2DT#!"4T޲Fɞ9QD2DT ?9T$ɞ9$A?W6J/?9T$8:W6J,HPHCI9I20WF/ɞ9$8:WF,HPHI9I ?9T$ɞ9$A?µ$?9TA=ULEQ?DZ.߰4>N.X8BLFJO;5ORܠ9ݩ5N.FJO:ݩ5ʅ>߰4>OXFJORܠ9љ55KUS̛SQT07>S7S˩5W˩5U˩5ORܠ9ݩ5JUS̛<N4NX8BLFաO5OMݩ5NFO:ݩ54OXFOMљ55US̛<աO3US̛SQT07>S7S˩5W˩5UOMߩ5US̛</N/40O;  N0աOMK/N/4ʅ>߰4>N.OX,FJO:9/N/4@@@,*N4NOX,FO:9N@@/N/40O;  N0աO/N/4ʅ>߰4>N.X8BLFJO;5ORܠ9ݩ5N.FJO:ݩ5ʅ>߰4>OXFJORܠ9љ55KUS̛SQT07>S7S˩5W˩5U˩5ORܠ9ݩ5JUS̛<N4NX8BLFաO5OMݩ5NFO:ݩ54OXFOMљ55US̛<աO3US̛SQT07>S7S˩5W˩5UOMߩ5US̛</N/40O;  N0աO>߰4>N.Xҥ3߫UBWOFJUQJ&$N4NXUBWOFUQJ/N/40O;  N0աO/N/4ʅ>߰4>N.X8BLFJO;5ORܠ9ݩ5N.FJO:ݩ5ʅ>߰4>OXFJORܠ9љ55KUS̛SQT07>S7S˩5W˩5U˩5ORܠ9ݩ5JUS̛<N4NX8BLFաO5OMݩ5NFO:ݩ54OXFOMљ55US̛<աO3US̛SQT07>S7S˩5W˩5UOMߩ5US̛</N/40O;  N0աO6Mӛ?6Mӛ?O;O/N/47>6744B9HS1HŞ1Kį?Dߋ5 Gބ24PK ۥNɿCR S2ބ2B@Bބ2ͩ- ۥN BDBM/N7߰4>N.X8BLFJO;5ORܠ9ݩ5N.FJO:ݩ5ʅ>߰4>OXFJORܠ9љ55KUS̛SQT07>S7S˩5W˩5U˩5ORܠ9ݩ5JUS̛<N4NX8BLFաO5OMݩ5NFO:ݩ54OXFOMљ55US̛<աO3US̛SQT07>S7S˩5W˩5UOMߩ5US̛</N/40O;  N0աO86/N/4ʅ>߰4>N.XWBOFJUQJ#!N4NXWBOFUQJ/N/40O;  N0աO/N/4ʅ>߰4>N.X8BLFJO;5ORܠ9ݩ5N.FJO:ݩ5ʅ>߰4>OXFJORܠ9љ55KUS̛SQT07>S7S˩5W˩5U˩5ORܠ9ݩ5JUS̛<N4NX8BLFաO5OMݩ5NFO:ݩ54OXFOMљ55US̛<աO3US̛SQT07>S7S˩5W˩5UOMߩ5US̛</N/40O;  N0աOSQ4/N/45F>JFJIݩ5ORܠ94/N/45F>JFJ,*4N54FIݩ5OM4N54F/N/40O;  N0աO/N/4ʅ>߰4>N.X8BLFJO;5ORܠ9ݩ5N.FJO:ݩ5ʅ>߰4>OXFJORܠ9љ55KUS̛SQT07>S7S˩5W˩5U˩5ORܠ9ݩ5JUS̛<N4NX8BLFաO5OMݩ5NFO:ݩ54OXFOMљ55US̛<աO3US̛SQT07>S7S˩5W˩5UOMߩ5US̛</N/40O;  N0աO 00,B4.Iַ;@?0,B4.I@/N/40O;  N0աO/N/4ʅ>߰4>N.X8BLFJO;5ORܠ9ݩ5N.FJO:ݩ5ʅ>߰4>OXFJORܠ9љ55KUS̛SQT07>S7S˩5W˩5U˩5ORܠ9ݩ5JUS̛<N4NX8BLFաO5OMݩ5NFO:ݩ54OXFOMљ55US̛<աO3US̛SQT07>S7S˩5W˩5UOMߩ5US̛</N/40O;  N0աO&$9:9;2—PX>9:;#!9:9;—PX>9:;,6BJ>P7BJ>P><76NJF3P;7N@N;JT;JQ;J/-7NJF3P;7NN;T;Q;,6BJ>P7BJ>P)'76BJPT;<̖@@TML&$7BJPT;<̖@@TML,6BJ>P7BJ>P/-176NJǭ;J2=>PQ@@@)'17NJǭ;J2=>PQ@@,6BJ>P7BJ>P,*CF76BJԿ7;˨OO/JIַ;&$CF7BJԿ7;˨OO/JI,6BJ>P7BJ>PUFJB76͎?/UFJB7͎?/,6BJ>P7BJ>P—PHIL2COJ—PHIL2COJسSB6BJDʿ7E>P—PHIL2COJ—PHIL2COJQ0N>>KJNBIL2COJBIL2COJ۳S6BJϿ7E>PBIL2COJBIL2COJQ0N׎>KɏJ,6BJ>P7BJ>P&$76BJ>PP/MGQT 7BJ>PPMGQT,6BJ>P7BJ>P,*JRJCJD6PV.6;JT)'JRJCJD6PV.6;T,6BJ>P7BJ>P 6BJDʿ7E>P@Ԛ<6BJϿ7E>P@Ԛ<,6BJ>P7BJ>P)'ARJBJD6PщQU;7P&$ARJBJD6PщQU;7,6BJ>P7BJ>P/-76BJF6F,QVMG.D6,*7BJF6F,QVMG.D6,6BJ>P7BJ>P53BܥNFCS7B76BR6HJ>AP/-BܥNFCS7B7B7HJ>AP,6BJ>P7BJ>P><76NJF3P;7N@N;JT;JQ;J/-7NJF3P;7NN;T;Q;,6BJ>P7BJ>PYWJǭ;N,6>PJǭ;DƂGщQJǭ;D@щQ,6>G3.ٟ@DƂGщQ@Ԛ<SQJǭ;N7>PJǭ;DƂGщQJǭ;D@щQ7>G3.ٟ@DƂGщQ@Ԛ<,6BJ>P7BJ>P/-176NJǭ;J2=>PQ@@@)'17NJǭ;J2=>PQ@@,6BJ>P7BJ>P/-FJō/NJD0PL36>;GB&$JNJD0PL36>;G,6BJ>P7BJ>PUFJB76͎?/UFJB7͎?/,6BJ>P7BJ>P#!6ǭ;>Q6NJ>P;7 6>Q6NJ>P;7,6BJ>P7BJ>P&$76BJ>PP/MGQT 7BJ>PPMGQT,6BJ>P7BJ>PCFJB6ǭ;@Ԛ<CFJB6@Ԛ<,6BJ>P7BJ>P 6BJDʿ7E>P@Ԛ<6BJϿ7E>P@Ԛ<,6BJ>P7BJ>P 6BJD6E>P@Ԛ< 6BJD6E>P@Ԛ<,6BJ>P7BJ>P/-76BJF6F,QVMG.D6,*7BJF6F,QVMG.D6,6BJ>P7BJ>P 76BJ>PHڶ>@Ԛ<7BJ>PHڶ>@Ԛ<,6BJ>P7BJ>P><76NJF3P;7N@N;JT;JQ;J/-7NJF3P;7NN;T;Q;,6BJ>P7BJ>P Lǭ;BϨHJ>PA7BLBϨHJ>P+,6BJ>P7BJ>P/-176NJǭ;J2=>PQ@@@)'17NJǭ;J2=>PQ@@,6BJ>P7BJ>P#!Lǭ;BϨHJ>PDG@KLBϨHJ>PD@K,6BJ>P7BJ>PUFJB76͎?/UFJB7͎?/,6BJ>P7BJ>P ;-M= ;-M=,6BJ>P7BJ>P&$76BJ>PP/MGQT 7BJ>PPMGQT,6BJ>P7BJ>P ;-M= ;-M=,6BJ>P7BJ>P 6BJDʿ7E>P@Ԛ<6BJϿ7E>P@Ԛ<,6BJ>P7BJ>P/-ϨHJō/BJ>PϨHJō/BJڶ>F=/-ϨHJō/BJ>PϨHJō/BJڶ>F=,6BJ>P7BJ>P/-76BJF6F,QVMG.D6,*7BJF6F,QVMG.D6,6BJ>P7BJ>P2076BCJ>P/G=Q>BD>ÐW,*7BCJ>PG=Q>BD>ÐW,6BJ>P7BJ>P><76NJF3P;7N@N;JT;JQ;J/-7NJF3P;7NN;T;Q;,6BJ>P7BJ>P Lǭ;BϨHJ>P:ÐW4LBϨHJ>P:ÐW44T7@<<>47@<>JH>CT<7@<6R>16R>7,O9ϪJ1<>@Ԛ<;9>CT<7@6>16>7,91<>@Ԛ<4T7@<<>47@<> <6>7T<@9:T<6>7T?9:4T7@<<>47@<>JH>CT<7@<6R>16R>7,O9ϪJ1<>@Ԛ<;9>CT<7@6>16>7,91<>@Ԛ<4T7@<<>47@<> M4TCT7@<@Ԛ<M4CT7@@Ԛ<4T7@<<>47@<>JH>CT<7@<6R>16R>7,O9ϪJ1<>@Ԛ<;9>CT<7@6>16>7,91<>@Ԛ<4T7@<<>47@<>539TB@>TK7<:7@<ǭ;?AB,*9TB>TK7<:7@ՄNAB4T7@<<>47@<>JH>CT<7@<6R>16R>7,O9ϪJ1<>@Ԛ<;9>CT<7@6>16>7,91<>@Ԛ<4T7@<<>47@<>A?>T<@>/26SCSET<@>-/7B6;9>T?>/26SCSET?>-/7B64T7@<<>47@<>JH>CT<7@<6R>16R>7,O9ϪJ1<>@Ԛ<;9>CT<7@6>16>7,91<>@Ԛ<4T7@<<>47@<>/-T@47@<>JH>CT<7@<6R>16R>7,O9ϪJ1<>@Ԛ<;9>CT<7@6>16>7,91<>@Ԛ<4T7@<<>47@<>,*4T<@HAVTJD8DAP4?HAVTD8A4T7@<<>47@<>JH>CT<7@<6R>16R>7,O9ϪJ1<>@Ԛ<;9>CT<7@6>16>7,91<>@Ԛ<4T7@<<>47@<>4T47@<>JH>CT<7@<6R>16R>7,O9ϪJ1<>@Ԛ<;9>CT<7@6>16>7,91<>@Ԛ<4T7@<<>47@<>;94TRF7@<5@2D0O6P6T,*4RI@5@20O6P64T7@<<>47@<>JH>CT<7@<6R>16R>7,O9ϪJ1<>@Ԛ<;9>CT<7@6>16>7,91<>@Ԛ<4T7@<<>47@<>,*C>8T<7@<1>DPDA&$C>8<7@1>DPDA6NBUC6O  @U66NBV1UC6O@V1U66NBUC6O  @U6866NBUC-9ԚETBET&$@6OCN3>ETBET,*@CӽD=HK:=-Ƈ>O=9=&$@CӽD=HK:=-Ƈ>O9 VHLć?OD6L@Ԛ< VHLć?OD6L@Ԛ<,*@CӽD=HK:=-Ƈ>O=9=&$@CӽD=HK:=-Ƈ>O9/-DHLKD‡?OAO6:,A7B)'DHLKD‡?OAO6:,+,*@CӽD=HK:=-Ƈ>O=9=&$@CӽD=HK:=-Ƈ>O9 VHLć?OD6L@Ԛ< VHLć?OD6L@Ԛ<,*@CӽD=HK:=-Ƈ>O=9=&$@CӽD=HK:=-Ƈ>O986H:!DƇ>O-8WHOWK-4=RJ53H:!DƇ>O-8WHOWK-4RJ,*@CӽD=HK:=-Ƈ>O=9=&$@CӽD=HK:=-Ƈ>O9 VHLć?OD6L@Ԛ< VHLć?OD6L@Ԛ<,*@CӽD=HK:=-Ƈ>O=9=&$@CӽD=HK:=-Ƈ>O986NA9=H5D‡?OJٟ@6:G2@@@,*N9H5D‡?OJ5:G2@@,*@CӽD=HK:=-Ƈ>O=9=&$@CӽD=HK:=-Ƈ>O9 VHLć?OD6L@Ԛ< VHLć?OD6L@Ԛ<,*@CӽD=HK:=-Ƈ>O=9=&$@CӽD=HK:=-Ƈ>O9SQDǭ;DQDUH:DO>&DB7DOOJDIPAFE>6MKDǭ;DQDUH:DO>&DB7DOOJDPAF>6,*@CӽD=HK:=-Ƈ>O=9=&$@CӽD=HK:=-Ƈ>O9 VHLć?OD6L@Ԛ< VHLć?OD6L@Ԛ<,*@CӽD=HK:=-Ƈ>O=9=&$@CӽD=HK:=-Ƈ>O9&$H=Dć?O=9=ϷAH@H=Dć?O9A@,*@CӽD=HK:=-Ƈ>O=9=&$@CӽD=HK:=-Ƈ>O9 VHLć?OD6L@Ԛ< VHLć?OD6L@Ԛ<,*@CӽD=HK:=-Ƈ>O=9=&$@CӽD=HK:=-Ƈ>O9A?,O-HDBٟ@;?=1PK@‡?O=9=@Ԛ<53,O-HD@?=1PK@‡?O9@Ԛ<,*@CӽD=HK:=-Ƈ>O=9=&$@CӽD=HK:=-Ƈ>O9 VHLć?OD6L@Ԛ< VHLć?OD6L@Ԛ<,*@CӽD=HK:=-Ƈ>O=9=&$@CӽD=HK:=-Ƈ>O9JHH=WK=:B:D‡?O:D1=@9=D9D5@Ԛ<>O=9=&$@CӽD=HK:=-Ƈ>O9 VHLć?OD6L@Ԛ< VHLć?OD6L@Ԛ<,*@CӽD=HK:=-Ƈ>O=9=&$@CӽD=HK:=-Ƈ>O9&$H=Dć?O=9=DSDA H=Dć?O9DSDA,*@CӽD=HK:=-Ƈ>O=9=&$@CӽD=HK:=-Ƈ>O9 VHLć?OD6L@Ԛ< VHLć?OD6L@Ԛ<,*@CӽD=HK:=-Ƈ>O=9=&$@CӽD=HK:=-Ƈ>O9>1#!H8V27Cî7/T>1GE2,߀3՟?4H8V2,7C7G/T>1?TJQ>;92߀3՟?4H8V27Cî7/T>1?TJQ)'H8V2,7C7G/T>1#!H8V27Cî7/T>120H,82,ֈ;04VC7G/T>1)'H,82ڈ;4VCî7/T>1)'H8V2,7C7G/T>1#!H8V27Cî7/T>1GE2,߀3՟?4H8V2,7C7G/T>1?TJQ>;92߀3՟?4H8V27Cî7/T>1?TJQ)'H8V2,7C7G/T>1#!H8V27Cî7/T>1,*V@,1V2,7C7G.V@ M,1V27Cî7.M)'H8V2,7C7G/T>1#!H8V27Cî7/T>1GE2,߀3՟?4H8V2,7C7G/T>1?TJQ>;92߀3՟?4H8V27Cî7/T>1?TJQ)'H8V2,7C7G/T>1#!H8V27Cî7/T>153H82,7C7G/T>1?TJQ>,*H827Cî7/T>1?TJQ)'H8V2,7C7G/T>1#!H8V27Cî7/T>1GE2,߀3՟?4H8V2,7C7G/T>1?TJQ>;92߀3՟?4H8V27Cî7/T>1?TJQ)'H8V2,7C7G/T>1#!H8V27Cî7/T>1/-W?A;OV2,7C7GA.T#!W?ҞMOV27Cî7A.)'H8V2,7C7G/T>1#!H8V27Cî7/T>1GE2,߀3՟?4H8V2,7C7G/T>1?TJQ>;92߀3՟?4H8V27Cî7/T>1?TJQ)'H8V2,7C7G/T>1#!H8V27Cî7/T>12,>B-45J2>B-5)'H8V2,7C7G/T>1#!H8V27Cî7/T>1GE2,߀3՟?4H8V2,7C7G/T>1?TJQ>;92߀3՟?4H8V27Cî7/T>1?TJQ)'H8V2,7C7G/T>1#!H8V27Cî7/T>1hfH8 -N2,ԓ4DC7G77BK;9/T>1KL/U5 -5>2,WFVTH8 -N24Cî77BK;9/T>1KL/U5 -5>2W)'H8V2,7C7G/T>1#!H8V27Cî7/T>1GE2,߀3՟?4H8V2,7C7G/T>1?TJQ>;92߀3՟?4H8V27Cî7/T>1?TJQ)'H8V2,7C7G/T>1#!H8V27Cî7/T>186H,7H82,RNVC7G/T7>1 H,>1)'H8V2,7C7G/T>1#!H8V27Cî7/T>1GE2,߀3՟?4H8V2,7C7G/T>1?TJQ>;92߀3՟?4H8V27Cî7/T>1?TJQ)'H8V2,7C7G/T>1#!H8V27Cî7/T>12C2C)'H8V2,7C7G/T>1#!H8V27Cî7/T>1GE2,߀3՟?4H8V2,7C7G/T>1?TJQ>;92߀3՟?4H8V27Cî7/T>1?TJQ)'H8V2,7C7G/T>1#!H8V27Cî7/T>1V2,7C7G¶;V27Cî7¶;BRADK BRADDBRADKC5—P=—Pބ2RADKC58:-9ET86RADC5=܉2RADC58:-9ETBRADK BRAD><@GD5ՂPRA5HDKϲLK2!QH9T;9@GD5ՂPRA5HDϲLK2!QH9TBRADK BRADJHH 5ՂP2CDKLARAK3DKMK5DKև9>TA?H 5ՂP2RKLARAK3DMK5DKև9>BRADK BRAD,*F7CPL߫WA=RADKS7)'F7CPL߫WA=RADS7BRADK BRADPNՂPLE;ߏGKCRADKCBAMKCK?KCCPD7LRABADK><,9;DR؇9U8ȴS>CPD7LRABADBRADK BRADDBRADKC5—P=—Pބ2RADKC58:-9ET86RADC5=܉2RADC58:-9ETBRADK BRAD\Z:DKCLCBCL5?LFL>HDKCRAK?MKߏGKCBùFPN:DCLCBC5?LL>HDCRAK?MKߏGKCBùFBRADK BRADJHH 5ՂP2CDKLARAK3DKMK5DKև9>TA?H 5ՂP2RKLARAK3DMK5DKև9>BRADK BRADqo -2CDKՂPLARAK3DKMKߏGK HӒC,NDK5=TUߋ5,,=>:J_] -2RKՂPLARAK3DMKߏGK HӒC,D5=TUߋ5,=>: 6ѤI; 6ѤI; ѤI;@?  ѤI;@ 6ѤI; 6ѤI;ѤI;;KFABѤI;;KFAB 6ѤI; 6ѤI; ѤI;@?  ѤI;@ 6ѤI; 6ѤI; ѤI;S>>ٟ@6;@Ԛ<ѤI;S>>5;@Ԛ< 6ѤI; 6ѤI; ѤI;@?  ѤI;@ 6ѤI; 6ѤI;JH3Ԛ<ѤI;>6;6SF;.TTD6;6SF;.TDTMSѤI22E7>>2OD@TDBѤI;A @69>TMSѤI22E7>>2OD@T 6ѤI; 6ѤI; ѤI;@?  ѤI;@ 6ѤI; 6ѤI;><ѤI;B2ѤI;2ѤI;0ѤI;SNѤI;NOFT><ѤI;B2ѤI;2ѤI;0ѤI;SNѤI;NOFT 6ѤI; 6ѤI; ѤI;@?  ѤI;@ 6ѤI; 6ѤI;53ѤI;>>;U0>;D6PGDSDA53ѤI;>>;U0>;D6PGDSDA 6ѤI; 6ѤI; ѤI;@?  ѤI;@ 6ѤI; 6ѤI;ѤI;M@KѤI;M@K 6ѤI; 6ѤI; ѤI;@?  ѤI;@ 6ѤI; 6ѤI; ѤI;8ٟ@>6CA7BѤI;8ٟ@>6C+&$5D>B7Dڹ32:TCG#!5D>B7Dڹ32:CG,*9EN39>ڹ3T21M1T)'9EN39>ڹ3T2M1T&$5D>B7Dڹ32:TCG#!5D>B7Dڹ32:CG,*:B7>B31CTCCԃP-C#!:B7>B31CCƠB7Dڹ32:TCG#!5D>B7Dڹ32:CG531TН?>/3>ND3>2HTC.:)'1?/3ND3>2HC.:&$5D>B7Dڹ32:TCG#!5D>B7Dڹ32:CG/-Ԋ/BNP92K1W>2Ԋ/Lؒ.=#!Ԋ/N9K1W>2Lؒ.=&$5D>B7Dڹ32:TCG#!5D>B7Dڹ32:CG,*V>NDͯ?ڹ3F1ȇN;9FGB)'V>NDͯ?ڹ3F1ׇN9FGB&$5D>B7Dڹ32:TCG#!5D>B7Dڹ32:CGDBNW>ڹ321%K9E?AFF?DJEʡH9?/86NW>ڹ32%K9E?AFFDJE9/&$5D>B7Dڹ32:TCG#!5D>B7Dڹ32:CG20;@7:TCUDTڹ3>NщQA7B#!;7:CUDڹ3>NщQ+&$5D>B7Dڹ32:TCG#!5D>B7Dڹ32:CG)'AFF?9E1ڹ321KW(#!AFF9E1ڹ32KW(&$5D>B7Dڹ32:TCG#!5D>B7Dڹ32:CG53EEO35B5Iٟ@7A:5Gς16T&$E>3B5@7A:5G+&$5D>B7Dڹ32:TCG#!5D>B7Dڹ32:CG86>>8RVGBڹ3;2F5>HK7<653>>8RVGBڹ3;2F5>H7<6&$5D>B7Dڹ32:TCG#!5D>B7Dڹ32:CGA?5BJH:ɚK73GHAVTJD8DAP865BJH:ɚK73GHAVTD8A&$5D>B7Dڹ32:TCG#!5D>B7Dڹ32:CGDBDNADV93>R9B>:D:TCS-@@@86DADV93>R9B>:D:CS@@&$5D>B7Dڹ32:TCG#!5D>B7Dڹ32:CGJB7>J3/:J7>J3/:&$5D>B7Dڹ32:TCG#!5D>B7Dڹ32:CG20AF?9C1NWڹ321K:&87)'AF9C1NWڹ32K:&8&$5D>B7Dڹ32:TCG#!5D>B7Dڹ32:CG,*9EN39>ڹ3T21M1T)'9EN39>ڹ3T2M1T&$5D>B7Dڹ32:TCG#!5D>B7Dڹ32:CGSQ9DBB3ҾW19659D:QTC2ʶU>3.ٟ@6ǽ=G@Ԛ<A?DBB3ҾW1965ՔDQC2ʶU>3ٟ@6G@Ԛ<&$5D>B7Dڹ32:TCG#!5D>B7Dڹ32:CG531TН?>/3>ND3>2HTC.:)'1?/3ND3>2HC.:&$5D>B7Dڹ32:TCG#!5D>B7Dڹ32:CGDB53>RD>B7HLTD>B7:LGDSDADB53>RD>B7HLTD>B7:LGDSDA&$5D>B7Dڹ32:TCG#!5D>B7Dڹ32:CG,*V>NDͯ?ڹ3F1ȇN;9FGB)'V>NDͯ?ڹ3F1ׇN9FGB&$5D>B7Dڹ32:TCG#!5D>B7Dڹ32:CG,*5DBڹ3G><97>?LS:)'5DBڹ3G><97>FS:&$5D>B7Dڹ32:TCG#!5D>B7Dڹ32:CG20;@7:TCUDTڹ3>NщQA7B#!;7:CUDڹ3>NщQ+&$5D>B7Dڹ32:TCG#!5D>B7Dڹ32:CGkiRV>NDڹ32į?1T9Fܫ7MN6K9D,K69.1R3RFBOBTec/>NDڹ32į?1T9Fܫ7MN6K9D,K69.1R3RFBOB&$5D>B7Dڹ32:TCG#!5D>B7Dڹ32:CG53EEO35B5Iٟ@7A:5Gς16T&$E>3B5@7A:5G+&$5D>B7Dڹ32:TCG#!5D>B7Dڹ32:CGSQ63P7S4DT9I871Dڹ32:TCRٍBKЅJCG>6DB63P7SCT871Dڹ32:CRٍBKЅJC>6&$5D>B7Dڹ32:TCG#!5D>B7Dڹ32:CGA?5BJH:ɚK73GHAVTJD8DAP865BJH:ɚK73GHAVTD8A&$5D>B7Dڹ32:TCG#!5D>B7Dڹ32:CG&$>>8RVGBڹ3;2F5&$>>8RVGBڹ3;2F5يR28يR28DBOHD-6J=FHيR28>DН?>QTيR453OD6J=FHيR28D?QيR4يR28يR2886يR2A8>يR2A8>8J-IN=JT)'يR2ŞيR2Ş8J-I=JTيR28يR28DBOHD-6J=FHيR28>DН?>QTيR453OD6J=FHيR28D?QيR4يR28يR28_]>7JЁH?ʡHWOUA7J1HN=FFHيR28>G@K\Z>7JЁH?ʡHWOUA7J1HN=FFHيR28G@KيR28يR28DBOHD-6J=FHيR28>DН?>QTيR453OD6J=FHيR28D?QيR4يR28يR28zx(" -UA7J1H -N= FFHيR28>DA7JUA7DН?>QTيR453OD6J=FHيR28D?QيR4يR28يR28><يRJTيR8T يRDН?>QTيR453OD6J=FHيR28D?QيR4يR28يR2886يR2A8>يR2A8>8J-WN8T,*يR2ŞيR2Ş8J-WN8TيR28يR28DBOHD-6J=FHيR28>DН?>QTيR453OD6J=FHيR28D?QيR4يR28يR28VT -UA7J1H= FFHيR28>107 A@H۰M3AMK -UA7J1H= FFHيR28107 @H3AيR28يR28DBOHD-6J=FHيR28>DН?>QTيR453OD6J=FHيR28D?QيR4يR28يR28;9يR28>9KA8D6P>JщQN.6@Ԛ<53يR289KA8DP>JщQN.6@Ԛ<يR28يR28DBOHD-6J=FHيR28>DН?>QTيR453OD6J=FHيR28D?QيR4يR28يR28b` UA7J1H N=б FFHيR28>DA7JUA7D UP/ŕD/PPQAP,9P/PPAP,9PUP/ڶ>D UP/ŕD/UPG,NKQM/UPG,KQMUP/ڶ>D UP/ŕD/PPQAP,9P/PPAP,9PUP/ڶ>D UP/ŕD /JPC98?UPT/JP98?UPTUP/ڶ>D UP/ŕD/PPQAP,9P/PPAP,9PUP/ڶ>D UP/ŕDA?//P//PO—P=-//PC?KP//Pĩ8>D UP/ŕD/PPQAP,9P/PPAP,9PUP/ڶ>D UP/ŕD>9S9Ԛ9S1/9RK@Ԛ<UP/ڶ>D UP/ŕD/PPQAP,9P/PPAP,9PUP/ڶ>D UP/ŕD2059P/ַ;/P/PA/P?PF7,*59P/ַ;/P/P/P?PFUP/ڶ>D UP/ŕD/PPQAP,9P/PPAP,9PUP/ڶ>D UP/ŕDPNUP?İUHP.F-S51SSAPK85G6)ʪJHUP?İUH1F-S51SSAPK8G6)ʪUP/ڶ>D UP/ŕD/PPQAP,9P/PPAP,9PUP/ڶ>D UP/ŕD UPʡH98CCH/TUP9CCH/TUP/ڶ>D UP/ŕD/PPQAP,9P/PPAP,9PUP/ڶ>D UP/ŕDPʰD/Fַ; PʰD/1 ܷT18W  U8WܷT18W@? U8W@ ܷT18W  U8WGE>ܷT18W>/26SCSEܷT18W>-/7B6A?>U8W>/26SCSEU8W>-/7B6 ܷT18W  U8WܷT18W@? U8W@ ܷT18W  U8WܷT1W>/>/CSܷT1W>/USܷT1W>/ܷT1W>/>8M6@66>ќ:0F6267(%!"~UW>/>/CSUW>/USUW>/UW>/>8M6@66>ќ:0F6267 ܷT18W  U8WܷT18W@? U8W@ ܷT18W  U8WGE>ܷT18W>/26SCSEܷT18W>-/7B6A?>U8W>/26SCSEU8W>-/7B6 ܷT18W  U8WܷT18W@? U8W@ ܷT18W  U8WA?6DQ66NیVOH2ܷT18W/Q66;6=;96Q66NیVOH2U8W/Q66;6= ܷT18W  U8WܷT18W@? U8W@ ܷT18W  U8W8ܷT1OW=;8UOW=; ܷT18W  U8WܷT18W@? U8W@ ܷT18W  U8W ܷT1WFM>Л6;@KUWF>Л6;@ ܷT18W  U8WܷT18W@? U8W@ ܷT18W  U8W20A89QEܷT1G4WE>FWAB)'A8ƋQEUG4WE>FAB:?9WΚI=X:?9WΚI=X86:?9>WΚI5DXK8WщQ# @@@/-:?̖>WΚI5DXK8WщQ#@@:?9WΚI=X:?9WΚI=XMK:?:?LIMW#DE=XWΚI5DXK8WщQ# @@@/-:?̖>WΚI5DXK8WщQ#@@:?9WΚI=X:?9WΚI=X WR:?9Iʉ5X@Ԛ<WR:?Iʉ5X@Ԛ<:?9WΚI=X:?9WΚI=X86:?9>WΚI5DXK8WщQ# @@@/-:?̖>WΚI5DXK8WщQ#@@:?9WΚI=X:?9WΚI=XVT9WI>:?щQV46V6#6#%6# 6$ 6#88GE9WI>:?щQV6V6#6#6#66#88:?9WΚI=X:?9WΚI=X86:?9>WΚI5DXK8WщQ# @@@/-:?̖>WΚI5DXK8WщQ#@@:?9WΚI=X:?9WΚI=X9Wʉ5X@N9Wʉ5X@N:?9WΚI=X:?9WΚI=X86:?9>WΚI5DXK8WщQ# @@@/-:?̖>WΚI5DXK8WщQ#@@:?9WΚI=X:?9WΚI=X>ܤKV#%ѾCHTL6LT53WR:?IG>ܤKV#%5L6LT:?9WΚI=X:?9WΚI=X86:?9>WΚI5DXK8WщQ# @@@/-:?̖>WΚI5DXK8WщQ#@@:?9WΚI=X:?9WΚI=X;9:?б 9WڶU5PRT53:?9WڶU5PRT:?9WΚI=X:?9WΚI=X86:?9>WΚI5DXK8WщQ# @@@/-:?̖>WΚI5DXK8WщQ#@@:?9WΚI=X:?9WΚI=X53EWN6=A9S:?9I>WΚI5DXK8WщQ# @@@/-:?̖>WΚI5DXK8WщQ#@@:?9WΚI=X:?9WΚI=X86:?Gʉ5BW/UX7AE3D#&&$:?GBW/UXAED,?RFD,?RF;9,?RFQUBDAP;0T?6T)!&$,?RFQUA;T6T)D,?RFD,?RF/-?,FR>,62ɀ?EBP22>)'8FR>,62ɀ?EBP22D,?RFD,?RFMKD=D3Dٟ@FR?,1@?>19Kٟ@9ٟ@-4,@Ԛ<>19K9-4,@Ԛ<D,?RFD,?RF,*RF>BϨH,@?,6DPDA&$RF>B؋8@86DPDAD,?RFD,?RF>EщQ@Ԛ<53D,.F?H.JV9S6>EщQ@Ԛ<D,?RFD,?RF,*RF,?HAVTJD8DAP R8?HAVTD8AD,?RFD,?RFGED3DRIF,2?.@PیVDHAVDAPT>T0N6Q20S9Xֈ?NXIWN,ڶ>T0N6Q9CXֈ?NXIַ;9Xֈ?NXI,*9CXֈ?NXIַ;M/TۓR7K 9Xֈ?NXIMTۓR79CXֈ?NXIַ;9Xֈ?NXI)'9CXֈ?NXIַ;B<ނBB@2A6ODPDA&$17>B@2A6DPDA 7BO֊2  7BOPNDD7OC-SO֊2>SDɵO689HAVTJD8DAPA?DD7C-SO>SDɵO689HAVTD8A 7BO֊2  7BO)'17>B@2A6ODPDA&$17>B@2A6DPDA 7BO֊2  7BOA?78BBر/D2ѺKٟ@6TCMUB@2A6ODPDA&$17>B@2A6DPDA 7BO֊2  7BO 57:CDO֊2ѺK@Ԛ<57:CDOѺK@Ԛ< 7BO֊2  7BO)'17>B@2A6ODPDA&$17>B@2A6DPDA 7BO֊2  7BODBкB9N7:CO֊2>ٟ@6߇;1G3F7;Q67;QT><кB9N7:CO>5߇;1G3F7;Q67;QT 7BO֊2  7BO)'17>B@2A6ODPDA&$17>B@2A6DPDA 7BO֊2  7BODB7OB62>R@2A57;QԚ<7;QT7N3>M><7OB62>R@2A57;QԚ<7;QT7N> 7BO֊2  7BO)'17>B@2A6ODPDA&$17>B@2A6DPDA 7BO֊2  7BO)'VDD7BC92>/67T)'VDD7BC92>/67T 7BO֊2  7BO)'17>B@2A6ODPDA&$17>B@2A6DPDA 7BO֊2  7BO&$V7JR1:2R<@@@#!V7JR1:2R<@@28>1S6Mſ2>15M;9KS6MԚ<28DAP;0T?6T)!#!K5MԚ<ſ2A;T6T)28>1S6Mſ2>15M;928=S0M28GN06W,646T20ſ2=S0Mſ2GN06W,64628>1S6Mſ2>15M><28ٟ@6Q>D1.S6M>=6P6T20ſ25Q>D1.5M>=6P628>1S6Mſ2>15M/-28>M281S6MۓR9T,K&$ƿ2>Mƿ215MۓR9T,K28>1S6Mſ2>15M><281S6M>1H3PیVDHAVDAPT20ſ215M>1H3PیVDHAVAT28>1S6Mſ2>15M53A28=χ71S6MN1SщQχ7=RJ,*Aſ2=χ715MN1SщQχ7RJ28>1S6Mſ2>15M;9KS6MԚ<28DAP;0T?6T)!#!K5MԚ<ſ2A;T6T)28>1S6Mſ2>15MS6MES28@Ԛ<5MESſ2@Ԛ<28>1S6Mſ2>15M><28ٟ@6Q>D1.S6M>=6P6T20ſ25Q>D1.5M>=6P628>1S6Mſ2>15MYWDNԚ<281S6M1H3Vٟ@281DAP;0T?6T)!86Nſ215M1H3Vٟ@ſ21A;T6T)28>1S6Mſ2>15M><281S6M>1H3PیVDHAVDAPT20ſ215M>1H3PیVDHAVAT28>1S6Mſ2>15M53MVٟ@28DAP;0T?6T)!Mٟ@ſ2A;T6T)28>1S6Mſ2>15M;9KS6MԚ<28DAP;0T?6T)!#!K5MԚ<ſ2A;T6T)28>1S6Mſ2>15M#!A281S6MDSDAAƿ215MDSDA —PJ>RJЍ—PJ>RJ,*M: D>J6߻WDSDA,*M: D>J6߻WDSDA —PJ>RJЍ—PJ>RJA?VNN,̥6:D9SJ6OQNέ;LSDʡH9;86VN,̥6:D9SJ6QN٭;SDʡH9; —PJ>RJЍ—PJ>RJ/-R߻W—PۃJ>JR6߻W,ƛK9@Ԛ</-R߻W—PۃJ>JR6߻W,ƛK9@Ԛ< —PJ>RJЍ—PJ>RJ53DȂ3@>Q—PJ—PJ>RCRA7B/-DȂ3@>Q—PJ—PJ>RCR+ —PJ>RJЍ—PJ>RJqo=>QH,<5Wį?;>—PJB—PۃJD9SIF>J6RN7>809DSPԮK߀3hf=>QH,5Wį?;>—PJB—PۃJD9SIF>J6R7>809DSPٮK —PJ>RJЍ—PJ>RJ;95$,U, 6D>:5JЂJDJA7B/-5$,,Ѝ6D>:5JЂJDJ+ —PJ>RJЍ—PJ>RJPN P;>LCD9J9KBDL=Ė16ǽ=EX>PGEЍP;>LCD9J9KBDL=Ė16E>P —PJ>RJЍ—PJ>RJG7TQ-G7TQ- —PJ>RJЍ—PJ>RJ)'Q—PJR69:ADSDA#!Q—PJR9ADSDA —PJ>RJЍ—PJ>RJDBA2ûR9?A>;BTûR9?A>5653TA2ûR9?A>;BûR9?A>56 —PJ>RJЍ—PJ>RJ,*M: D>J6߻WDSDA,*M: D>J6߻WDSDA —PJ>RJЍ—PJ>RJ><6JD9SJ6ȻW̑-9ٟ@—PJ>RJ@Ԛ<866JD9SJ6ȻW̑- @—PJ>RJ@Ԛ< —PJ>RJЍ—PJ>RJ/-R߻W—PۃJ>JR6߻W,ƛK9@Ԛ</-R߻W—PۃJ>JR6߻W,ƛK9@Ԛ< —PJ>RJЍ—PJ>RJMK66DJQ—PL>JRJJQ—PL@BJ9Uڤ55@Ԛ<GE66DJQ—PL>JRJJQ—PL@BJUܤ5@Ԛ< —PJ>RJЍ—PJ>RJqo=>QH,<5Wį?;>—PJB—PۃJD9SIF>J6RN7>809DSPԮK߀3hf=>QH,5Wį?;>—PJB—PۃJD9SIF>J6R7>809DSPٮK —PJ>RJЍ—PJ>RJ4B4յGWGXF4B4WX —PJ>RJЍ—PJ>RJPN P;>LCD9J9KBDL=Ė16ǽ=EX>PGEЍP;>LCD9J9KBDL=Ė16E>P —PJ>RJЍ—PJ>RJ,* R:D>ڝJRK2DG@K&$ЍR:D>ڝJRK2D@K —PJ>RJЍ—PJ>RJ)'Q—PJR69:ADSDA#!Q—PJR9ADSDA —PJ>RJЍ—PJ>RJ;95$,U, 6D>:5JЂJ9WWC/-5$,,Ѝ6D>:5JЂJ9āRA —PJ>RJЍ—PJ>RJ,*M: D>J6߻WDSDA,*M: D>J6߻WDSDA —PJ>RJЍ—PJ>RJ AF8,TЍAF,T —PJ>RJЍ—PJ>RJ/-R߻W—PۃJ>JR6߻W,ƛK9@Ԛ</-R߻W—PۃJ>JR6߻W,ƛK9@Ԛ< —PJ>RJЍ—PJ>RJ  4B  4B —PJ>RJЍ—PJ>RJqo=>QH,<5Wį?;>—PJB—PۃJD9SIF>J6RN7>809DSPԮK߀3hf=>QH,5Wį?;>—PJB—PۃJD9SIF>J6R7>809DSPٮK —PJ>RJЍ—PJ>RJ/-UCUTʡH>/X>>A2  UUʡH>/X>2Ѝ —PJ>RJЍ—PJ>RJPN P;>LCD9J9KBDL=Ė16ǽ=EX>PGEЍP;>LCD9J9KBDL=Ė16E>P —PJ>RJЍ—PJ>RJ AFPCRJЍ—PJ>RJ)'Q—PJR69:ADSDA#!Q—PJR9ADSDA —PJ>RJЍ—PJ>RJ&$UR:D>JB/T&$UR:D>JB/TPC11ȯBPC1ȯB&$&DC1ȯBI91PI@Ԛ<#!&DC1ȯBI1PI@Ԛ<PC11ȯBPC1ȯB 2<;> 2<;>PC11ȯBPC1ȯB&$&DC1ȯBI91PI@Ԛ<#!&DC1ȯBI1PI@Ԛ<PC11ȯBPC1ȯB)'V2PKC4EȯB-;J6&$V2PKC4EȯB-;ϜJPC11ȯBPC1ȯB&$&DC1ȯBI91PI@Ԛ<#!&DC1ȯBI1PI@Ԛ<PC11ȯBPC1ȯB20T3=C;D9>:CO-֛7:CO-2P:ȯBK6NKDSDA,*>2P:ȯBK6NKDSDAPC11ȯBPC1ȯB&$&DC1ȯBI91PI@Ԛ<#!&DC1ȯBI1PI@Ԛ<PC11ȯBPC1ȯB#!PNȯB>9H-BV6#!PNȯB>9H-BV6PC11ȯBPC1ȯB&$&DC1ȯBI91PI@Ԛ<#!&DC1ȯBI1PI@Ԛ<PC11ȯBPC1ȯBJHDPRȯBIH,56:LIB,I;9V;KXܤK$GEDPRȯBIH,56:LIB,I;V;KXܤK$ X˩5R9: X˩5ֲ9/-' ڲ߹-:X>˩56I:,@Ԛ<,* ڲ߹-:X>˩56I:,@Ԛ< X˩5R9: X˩5ֲ9,*ȏBҲU>Rɸ˩5G@Ԛ<DBݩ5C B6:X/ݩ5ٟ@5UI: .X>˩5G@Ԛ< X˩5R9: X˩5ֲ9GE B߹-;:XܷT6˩5J˩54 B߹-;:XܷT6˩5/7;9 B-:XܷT6J˩54 B-:XܷT6/7 X˩5R9: X˩5ֲ9DB: >˩5AKB: >˩5AK X˩5R9: X˩5ֲ96T' ߹-X6˩56T ߹-X6 X˩5R9: X˩5ֲ9/-' ڲ߹-:X>˩56I:,@Ԛ<,* ڲ߹-:X>˩56I:,@Ԛ< X˩5R9: X˩5ֲ9A?Uٟ@5߹-:XD˩5I:XB9D˩5ƛK6@@@;9Uٟ@5߹-:XD˩5I:XBD˩5ƛK6@@ X˩5R9: X˩5ֲ9GEݩ5TC B6:X/ݩ5ٟ@5UI: .X>˩5G@Ԛ<DBݩ5C B6:X/ݩ5ٟ@5UI: .X>˩5G@Ԛ< X˩5R9: X˩5ֲ9GE BI;:XܷT6˩5J˩54 BI;:XܷT6˩5/7A? BI;:XܷT6J˩54 BI;:XܷT6/7 X˩5R9: X˩5ֲ9DB: >˩5AKB: >˩5AK X˩5R9: X˩5ֲ9;95˱U̾-C3CIQ:> :X>6˩5,;865˱U̾-C3CIQ:> :X>6,; X˩5R9: X˩5ֲ9/-' ڲ߹-:X>˩56I:,@Ԛ<,* ڲ߹-:X>˩56I:,@Ԛ< X˩5R9: X˩5ֲ9\Z BH 6ӻBO ߹-:XܷTBH ߹-XܷT;W; N= FJ˩54YW BH 6O ߹-:XܷTBH ߹-XܷT;W; N= FJ˩54 X˩5R9: X˩5ֲ9GEݩ5TC B6:X/ݩ5ٟ@5UI: .X>˩5G@Ԛ<DBݩ5C B6:X/ݩ5ٟ@5UI: .X>˩5G@Ԛ< X˩5R9: X˩5ֲ9)' >6˩55WR9:DSDA  >65Wֲ9DSDA X˩5R9: X˩5ֲ9DB: >˩5AKB: >˩5AK X˩5R9: X˩5ֲ9A?б = F߹-=X B˩5HFST:TʡH?CگD/86б = F߹-=X BHFSTT9CگD/ X˩5R9: X˩5ֲ9/-' ڲ߹-:X>˩56I:,@Ԛ<,* ڲ߹-:X>˩56I:,@Ԛ< X˩5R9: X˩5ֲ9nlRA߹-:X> B6˩51D0;Hٟ@R9:KBB>5IBEKRFTD>6@Ԛ<_]RA߹-:X> B610Hٟ@ֲ9KBB>5IBEKRFTD>6@Ԛ< X˩5R9: X˩5ֲ9GEݩ5TC B6:X/ݩ5ٟ@5UI: .X>˩5G@Ԛ<DBݩ5C B6:X/ݩ5ٟ@5UI: .X>˩5G@Ԛ< X˩5R9: X˩5ֲ9GE KB6NEI:X5R9:˩5UIR>:DSDAA? KB6NEI:X5ֲ9˩5UIR>:DSDA X˩5R9: X˩5ֲ9DB: >˩5AKB: >˩5AK X˩5R9: X˩5ֲ9DB߹-:XܷT6H߹-:XܷT6˩5Q' ѲB6ӻBO453߹-:XܷT6߹-:XܷT6Q ѲB04 X˩5R9: X˩5ֲ9/-' ڲ߹-:X>˩56I:,@Ԛ<,* ڲ߹-:X>˩56I:,@Ԛ< X˩5R9: X˩5ֲ9/-0:X6˩50:X6˩5>464T&$0:X60:X6>46T X˩5R9: X˩5ֲ9GEݩ5TC B6:X/ݩ5ٟ@5UI: .X>˩5G@Ԛ<DBݩ5C B6:X/ݩ5ٟ@5UI: .X>˩5G@Ԛ< X˩5R9: X˩5ֲ9A?б = F߹-=X B˩5HFST:TʡH?CگD/86б = F߹-=X BHFSTT9CگD/ X˩5R9: X˩5ֲ9DB: >˩5AKB: >˩5AK X˩5R9: X˩5ֲ9\ZDRA9į?߹-=X>6˩5H0-DE06EщQI.6щQ22DSDAYWDRA9į?߹-=X>6H0-DE06EщQI.6щQ22DSDA NF;WH4EK NF;WH4EK&$ NF;WHEK9ݠ.ET#! NF;WHEK9ET NF;WH4EK NF;WH4EK#! N;WHEK9ݠ.ET N;WHE9ET NF;WH4EK NF;WH4EK&$ NF;WHEK9ݠ.ET#! NF;WHEK9ET NF;WH4EK NF;WH4EK&$ N;WHEKK9ݠ.ET N;WHE9ET NF;WH4EK NF;WH4EK&$ NF;WHEK9ݠ.ET#! NF;WHEK9ET NF;WH4EK NF;WH4EK)' N9ݠ.;WKE 9ݠ.ET  N9;WK 9ET NF;WH4EK NF;WH4EK&$ NF;WHEK9ݠ.ET#! NF;WHEK9ET NF;WH4EK NF;WH4EK#! N;WKEK9ݠ.ET N;WKK9ET NF;WH4EK NF;WH4EK&$ NF;WHEK9ݠ.ET#! NF;WHEK9ET NF;WH4EK NF;WH4EK&$ NF;WHEK9ݠ.ET#! NF;WHEK9ET NF;WH4EK NF;WH4EK&$ NF;WHEK9ݠ.ET#! NF;WHEK9ET NF;WH4EK NF;WH4EK#! NF;WEK9ݠ.ET  NF;WEK9ET NF;WH4EK NF;WH4EK&$ NF;WHEK9ݠ.ET#! NF;WHEK9ET NF;WH4EK NF;WH4EK)' NF;W2T9K9ݠ.ET&$ NF;W2T9K9ET NF;WH4EK NF;WH4EK&$ NF;WHEK9ݠ.ET#! NF;WHEK9ET NF;WH4EK NF;WH4EK#! NF;WHK9ݠ.ET  NF;WHK9ET NF;WH4EK NF;WH4EK&$ NF;WHEK9ݠ.ET#! NF;WHEK9ET NF;WH4EK NF;WH4EK)' N9ݠ.;WHE 9ݠ.ET#! N9;WHE 9ET:CT67Iַ;:T67I 7EU:CT67Iַ;78:T67I:CT67Iַ;:T67I/-:T678:T67Iַ;—P=8-)':T678:T67I=8-:CT67Iַ;:T67I 7EU:CT67Iַ;78:T67I:CT67Iַ;:T67I T18:CT67Iַ;T18:T67I:CT67Iַ;:T67I 7EU:CT67Iַ;78:T67I:CT67Iַ;:T67Iܥ60T67ȣ8Iַ;ܥ60T67I:CT67Iַ;:T67I 7EU:CT67Iַ;78:T67I:CT67Iַ;:T67I,*:0EUPU,I:T67Iַ;#!:08P,I:T67I:CT67Iַ;:T67I 7EU:CT67Iַ;78:T67I:CT67Iַ;:T67I0T67ȣ8Iַ;0T67I:CT67Iַ;:T67I 7EU:CT67Iַ;78:T67I:CT67Iַ;:T67I:CT67Iַ;@?:T67I@:CT67Iַ;:T67I 7EU:CT67Iַ;78:T67I:CT67Iַ;:T67I/-:CT67Iַ;:CT67Iַ;;#!:T67I:T67I;:CT67Iַ;:T67I 7EU:CT67Iַ;78:T67I:CT67Iַ;:T67I,*0EU4J8:CT67Iַ;ܥ6 084J:T67Iܥ6:CT67Iַ;:T67I 7EU:CT67Iַ;78:T67I:CT67Iַ;:T67I)':CT67Iַ;M/TۓR7K:T67IMTۓR7:CT67Iַ;:T67I 7EU:CT67Iַ;78:T67I:CT67Iַ;:T67I)':CT67Iַ;M/TۓR7K:T67IMTۓR7:CT67Iַ;:T67I 7EU:CT67Iַ;78:T67I:CT67Iַ;:T67I P P4X>E1; 4X>B E1;>XHMʭBWTB>XHMʭBW4X>E1; 4X>B:D>75.T:D>75.T4X>E1; 4X>B/--XE1;7߹-WD7ՕNծH-XBՕNծH4X>E1; 4X>BMK%X6Xޡ8XSX8XNX.XCXCXFX2X4XCA?%X6Xޡ8XX8XNXXXXFX2X4XC4X>E1; 4X>B E1;/64DG@KB/6D@K4X>E1; 4X>BSQİFE1;/64X۹/>OX۹/>TʭBS>OʭBS>TU>6K53İFB/6X>OX>B>OB>U>64X>E1; 4X>BkiE1;MIB>ю2/4AT23WS;XIю2Xю2>ю2UATXIX)PNBMI>ю2/4F23WS;X22ю2UFXI)4X>E1; 4X>B/4?BOBT/4?BOB4X>E1; 4X>B E1;>XHMʭBWTB>XHMʭBW4X>E1; 4X>B)'/43>L??HF? FT#!/43>L?HF FT4X>E1; 4X>B/--XE1;7߹-WD7ՕNծH-XBՕNծH4X>E1; 4X>BGE˛59/=T4>X?ޡ8RV4>E1;6T44K2,*-4>/4>B6T5K24X>E1; 4X>B E1;/64DG@KB/6D@K4X>E1; 4X>BVTDG:/4X>3?X?FBTF?ޡ8H?.:FʭB.4?F6>3?XFBTFޡ8H.FF64X>E1; 4X>BkiE1;MIB>ю2/4AT23WS;XIю2Xю2>ю2UATXIX)PNBMI>ю2/4F23WS;X22ю2UFXI)4X>E1; 4X>BVTE1;>C6PKH,-X?71E70NʡHH064TDBB>C6PKH,-X0NʡHH064T4X>E1; 4X>B E1;>XHMʭBWTB>XHMʭBW4X>E1; 4X>B#!U/4X>3B?8,T U/4X>3B?,T4X>E1; 4X>B/--XE1;7߹-WD7ՕNծH-XBՕNծH4X>E1; 4X>B/--446M; ->>@W>W>)'-56M; ->>@>W>4X>E1; 4X>B E1;/64DG@KB/6D@K4X>E1; 4X>B53?41K>F7>>D<(6հL3T53?41K>F7>>D<(6հL3T4X>E1; 4X>BkiE1;MIB>ю2/4AT23WS;XIю2Xю2>ю2UATXIX)PNBMI>ю2/4F23WS;X22ю2UFXI)4X>E1; 4X>B_]E1;1-X?P@4BS?H-M>ԁ:FT,;J8L0(!MKB1-X?P@4BS?H-M>ԁ:FT,;J8L0X4X>E1; 4X>B E1;>XHMʭBWTB>XHMʭBW4X>E1; 4X>B)'/43?>LHJX/ENB#!/43?>LHJXEN4X>E1; 4X>B/--XE1;7߹-WD7ՕNծH-XBՕNծH4X>E1; 4X>B3H24CM΄/ǟ9=Tޡ8?$ڻ($!3H24CMτ/-8$ڻ($!4X>E1; 4X>B E1;/64DG@KB/6D@K4X>E1; 4X>B,*/43?>LHJX/E1ʞ:-)'/43?>LHJXE1ʞ:-4X>E1; 4X>BkiE1;MIB>ю2/4AT23WS;XIю2Xю2>ю2UATXIX)PNBMI>ю2/4F23WS;X22ю2UFXI)4X>E1; 4X>B,*HE1;>/4H?LB/4HLBBQT2>FWA>BL>F)'4>BWAQT2>JF@Ԛ< 4>BWAL>S@Ԛ<WA>BQT2>FWA>BL>F869GOBQT2>PVP.5AJ>P)'9GOBL>PP.J>PWA>BQT2>FWA>BL>F20R0W6>BйSQT2>FDSDA,*R0W6>BйSL>FDSDAWA>BQT2>FWA>BL>FA?COW>M>BWAQT296O8GDSDA;9COW>M>BWAL96O8GDSDAWA>BQT2>FWA>BL>FMKR9GMWWAI>BN==9=A>MNS9=A7B86RךGWWAI>BN==>MNS9=+WA>BQT2>FWA>BL>F209G>BQT2>VJ768T7=&$9G>BL>VќJ687=WA>BQT2>FWA>BL>F;9D9DI>BWRQT2>SQU>V@Ԛ<&$9I>BWRLS>V@Ԛ<WA>BQT2>FWA>BL>F/- -FWLSJ$2AB/- -FWLSJ$2ABWA>BQT2>FWA>BL>F86DS8G>BWAQT2>M@?@@@/-DS8G>BWAL>M@?@@WA>BQT2>FWA>BL>F)'NFHFOFOVVA4@K&$NFHFOFOVVA4@WA>BQT2>FWA>BL>F&$W>V>BWAQT2@Ԛ< W>V>BWAL@Ԛ<WA>BQT2>FWA>BL>F86>BQT2>΂PF;/U N5LUٶ,*>BL>΂PF;/U N5LUWA>BQT2>FWA>BL>F)'4>BWAQT2>JF@Ԛ< 4>BWAL>S@Ԛ<WA>BQT2>FWA>BL>F#!>BWBQT2>F@Ԛ<>BWBL>F@Ԛ<WA>BQT2>FWA>BL>F20R0W6>BйSQT2>FDSDA,*R0W6>BйSL>FDSDAWA>BQT2>FWA>BL>F#!W2E>DQT2ϩNFBWE>DLϩNFWA>BQT2>FWA>BL>FMKR9GMWWAI>BN==9=A>MNS9=A7B86RךGWWAI>BN==>MNS9=+WA>BQT2>FWA>BL>F8F5R.UES28F=.UESWA>BQT2>FWA>BL>F;9D9DI>BWRQT2>SQU>V@Ԛ<&$9I>BWRLS>V@Ԛ<WA>BQT2>FWA>BL>F)'J>R8"FK%FJ>RF%FWA>BQT2>FWA>BL>F86DS8G>BWAQT2>M@?@@@/-DS8G>BWAL>M@?@@WA>BQT2>FWA>BL>F53D96MEK>BQT2>VD@@@&$D96EK>BL>V@@WA>BQT2>FWA>BL>F&$W>V>BWAQT2@Ԛ< W>V>BWAL@Ԛ<WA>BQT2>FWA>BL>F ֖F>PMމ6J6J7+Mމ6J6J7WA>BQT2>FWA>BL>F)'4>BWAQT2>JF@Ԛ< 4>BWAL>S@Ԛ<WA>BQT2>FWA>BL>F/-UMӛ?1?7F,7MRQ#!UMӛ?1?MRQWA>BQT2>FWA>BL>F20R0W6>BйSQT2>FDSDA,*R0W6>BйSL>FDSDAWA>BQT2>FWA>BL>F/-D96M>BWAQT2DSDA&$D96>BWALDSDAIDT0I ID0IDBDT30IDT30I4DT30IDZ.>4I?86D30ID30I4D30IDZ.>4I2IDT0I ID0I,*IDT01,ICDT0IDT#!IDT01ID0IDTIDT0I ID0I&$RIDTN0I0I4@Ԛ<RID00I4@Ԛ<IDT0I ID0I/-0I9Q6S=KI8KI:@Ԛ</-0I9Q6S=KI8KI:@Ԛ<IDT0I ID0I0I5I?0I5I2IDT0I ID0IA?DT3I58DT38I5DT3X58I?53D3I58D38I5D3X58I2IDT0I ID0I0IIַ;4DG@K0II4D@KIDT0I ID0I&$IDTPDN0I0I@Ԛ<IDPD00I@Ԛ<IDT0I ID0I#!I0IػKI0I4I0I#!I0IػKI0I4I0IIDT0I ID0I/-0IDTFַ;8- 8T !0D18- 8IDT0I ID0I0I4@Ԛ<0I4@Ԛ<IDT0I ID0I20D0ID0I4D0IDZ.>4I?/-D0ID0I4D0IDZ.>4I2IDT0I ID0I0IDZ.>4@Ԛ<0IDZ.>4@Ԛ<IDT0I ID0I20IDT,;01,ICDT0IDT)'IDT,;01ID0IDTIDT0I ID0IDBDT30IDT30I4DT30IDZ.>4I?86D30ID30I4D30IDZ.>4I2IDT0I ID0IDB,TܷT0I,TܷT0I4,TܷT0IDZ.>4I?86,ܷT0I,ܷT0I4,ܷT0IDZ.>4I2IDT0I ID0I&$RIDTN0I0I4@Ԛ<RID00I4@Ԛ<IDT0I ID0IMKDT3N0IDT3N0I4DT3N0IDZ.>4I?/-D30D304D30DZ.>4I2IDT0I ID0I0I5I?0I5I2IDT0I ID0I0IػK4@K0IػK4@KIDT0I ID0I0IIַ;4DG@K0II4D@KIDT0I ID0I0IDG@K0ID@KIDT0I ID0I#!I0IػKI0I4I0I#!I0IػKI0I4I0IIDT0I ID0I53DT3IDT3Iַ;DT3OII?#!D3ID3ID3OI2L7ٟ@8 Lٟ@8 L7@?L@L7ٟ@8 Lٟ@8 L7B6  LB6L7ٟ@8 Lٟ@8,*6—P,L7ٟ@8H7@K7@Ԛ<#!6ٟ@8H7@K7@Ԛ<L7ٟ@8 Lٟ@8 L7@K  L@KL7ٟ@8 Lٟ@8 L7@?L@L7ٟ@8 Lٟ@8 L76?  L6?L7ٟ@8 Lٟ@8,*6—P,L7ٟ@8H7@K7@Ԛ<#!6ٟ@8H7@K7@Ԛ<L7ٟ@8 Lٟ@8,*6L78>ٟ@HF@F76>P)'6L8>ٟ@HF@F76>PL7ٟ@8 Lٟ@8 L7@?L@L7ٟ@8 Lٟ@820A7L7Hٟ@8EP;:PO@@@,*A7LHٟ@8EP;:PO@@L7ٟ@8 Lٟ@8,*6—P,L7ٟ@8H7@K7@Ԛ<#!6ٟ@8H7@K7@Ԛ<L7ٟ@8 Lٟ@8—P,L7?60 ?60L7ٟ@8 Lٟ@8 L7@?L@L7ٟ@8 Lٟ@8/-L7ٟ@8AR>:6>NDSDA,*Lٟ@8AR>:6>NDSDAL7ٟ@8 Lٟ@8,*6—P,L7ٟ@8H7@K7@Ԛ<#!6ٟ@8H7@K7@Ԛ<L7ٟ@8 Lٟ@8wuL7DF6L7B7L76<6—P,L7 -Gٟ@867@75L78>ٟ@;FJ>N1S_]LDF6LB7L6Ǥ< -Gٟ@867@75L8>ٟ@;FJ>N1S¨0A=Tɾ=S0=Tɾ=S20¨0A=Tɾ=SN.W0AT("0=T̗<.0AT¨0A=Tɾ=S0=Tɾ=S,*¨0A=Tɾ=CPI/C/9?T#!0=PI/C/9?T¨0A=Tɾ=S0=Tɾ=S¨0ʽ=>=Tɾ=R@Ԛ<0>=Tɾ=R@Ԛ<¨0A=Tɾ=S0=Tɾ=S86¨0ʽ==Tɾ=C6=Tɾ=C6AANTAT#!0=T6=T6AATA¨0A=Tɾ=S0=Tɾ=S20¨0ʽ=Dٟ@ޢ7C7C=Tɾ=CѲ/DT("0D>=Ѳ/DT¨0A=Tɾ=S0=Tɾ=S/-=Tɾ=>¨0ʽ=ʇXQޢ0ʇXQޢΉX˛5¨0A/TD¨0A/A4J53-ʇXDQ=>ΉX˛50/TD0/AJ¨0A=Tɾ=S0=Tɾ=S¨0ʽ=>=Tɾ=R@Ԛ<0>=Tɾ=R@Ԛ<¨0A=Tɾ=S0=Tɾ=S)'¨0A=Tɾ=W9L/͒A4T0=Tɾ=W/͒A4¨0A=Tɾ=S0=Tɾ=S20¨0ʽ=Dٟ@ޢ=Tɾ=CѲ/DT("0D>=Ѳ/DT¨0A=Tɾ=S0=Tɾ=S,*¨0A=Tɾ=6=T3OTDA4&$0=Tɾ=6=T3OTDA¨0A=Tɾ=S0=Tɾ=S53¨0A=Tɾ=CDA4AATUʡH9A/,*0=T˾=DAAATUʡH9A/¨0A=Tɾ=S0=Tɾ=S ¨0A=Tɾ=68,T0=Tɾ=6,T¨0A=Tɾ=S0=Tɾ=S20¨0A=Tɾ=SN.W0AT("0=T̗<.0AT¨0A=Tɾ=S0=Tɾ=SDBS48¨0ʽ=P=Tɾ=Cϛ)ϛ)))QTɾ=C98KT/-ФO8-=ϛ)ϛ)))Q98KT¨0A=Tɾ=S0=Tɾ=S¨0ʽ=>=Tɾ=R@Ԛ<0>=Tɾ=R@Ԛ<¨0A=Tɾ=S0=Tɾ=S#!E=¨0ʽ==Tɾ=.8?̛<=0=Tɾ=.?¨0A=Tɾ=S0=Tɾ=S20¨0ʽ=Dٟ@ޢ=Tɾ=CѲ/DT("0D>=Ѳ/DT¨0A=Tɾ=S0=Tɾ=S ¨0AD>=Tɾ=C@K0D>=@K¨0A=Tɾ=S0=Tɾ=S53¨0A=Tɾ=CDA4AATUʡH9A/,*0=T˾=DAAATUʡH9A/¨0A=Tɾ=S0=Tɾ=S=Tɾ=C6¨0AT=T60T  S1/W/߹-CʡH97Qן9ں-ʡH97Qן9  S1/W/)'//Æ.J:NLJS1/GB //Æ.J:NLW/G  S1/W/S1/B;AATW/BAAT  S1/W/ S1D?  WD?  S1/W/߹-CʡH97Qן9ں-ʡH97Qן9  S1/W/53S1/E70C/77S1/AB#!W/E7W/ABw  S1/W/S1/B;AATW/BAAT  S1/W/ABAB  S1/W/߹-CʡH97Qן9ں-ʡH97Qן9  S1/W/209J/?ſQ5ߕJCMCRURН?QT)'9J/?ſQ5ߕJCMCRQ  S1/W/S1/B;AATW/BAAT  S1/W/>9@VWF?Wַ;;E-S1Н?>AT20B>9@VWF?Wַ;;E-W?A  S1/W/߹-CʡH97Qן9ں-ʡH97Qן9  S1/W/S1/IA6W/IA6EG?>-EG?>-/-G?R142TN5=7@P:J#!G?142T5=@PJEG?>-EG?>-)'G?>-PL΅/Bڶ>SJ@Ԛ< G?>-΅/BSJ@Ԛ<EG?>-EG?>- G?T4 G?T4EG?>-EG?>- GW-TG*EG?>-EG?>-E,G?/-"D:EG?/-"DEG?>-EG?>-20G?>-G64?9ʉ5;˫N¶;PNT,*G?>-G4?9ʉ5;ΫNPNTEG?>-EG?>-86G?>-22΅/8B?¶7ģCCщQDPDA,*G?>-΅/8BNCщQDPDAEG?>-EG?>-20NE,G?>-?¶7ʡHWB:ģCO#!NEG?>-NW:CEG?>-EG?>-&$G?>-8G?>-4-2#!G?>-G?>-4-2EG?>-EG?>- G?>-/.BʭBѡ8¶;G?>-/BʭBѡ8¶;>ׄ9?ϪJJ1>>ׄ9?ϪJJ1>;9>؞Cׄ9?B:9ڶ>ST=O>I,TJ@Ԛ<53ρ>ׄ9?B:9ST=O>I,TJ@Ԛ<>ׄ9?ϪJJ1>>ׄ9?ϪJJ1>,*K=9:ׄ9?DϪJP>؞C@@@ =:ׄ9?DϪJPρ>@@>ׄ9?ϪJJ1>>ׄ9?ϪJJ1>;9>؞Cׄ9?B:9ڶ>ST=O>I,TJ@Ԛ<53ρ>ׄ9?B:9ST=O>I,TJ@Ԛ<>ׄ9?ϪJJ1>>ׄ9?ϪJJ1>>؞C19Tׄ9?@Ԛ<ρ>19Tׄ9?@Ԛ<>ׄ9?ϪJJ1>>ׄ9?ϪJJ1>;9>؞Cׄ9?B:9ڶ>ST=O>I,TJ@Ԛ<53ρ>ׄ9?B:9ST=O>I,TJ@Ԛ<>ׄ9?ϪJJ1>>ׄ9?ϪJJ1>Ư8Hׄ9?>؞C@@@Ư8Hׄ9?ρ>@@>ׄ9?ϪJJ1>>ׄ9?ϪJJ1>;9>؞Cׄ9?B:9ڶ>ST=O>I,TJ@Ԛ<53ρ>ׄ9?B:9ST=O>I,TJ@Ԛ<>ׄ9?ϪJJ1>>ׄ9?ϪJJ1> ׄ9?=7 ׄ9?=7>ׄ9?ϪJJ1>>ׄ9?ϪJJ1>;9>؞Cׄ9?B:9ڶ>ST=O>I,TJ@Ԛ<53ρ>ׄ9?B:9ST=O>I,TJ@Ԛ<>ׄ9?ϪJJ1>>ׄ9?ϪJJ1>&$>؞Cׄ9?6R1TDPDA ρ>ׄ9?61TDPDA>ׄ9?ϪJJ1>>ׄ9?ϪJJ1>;9>؞Cׄ9?B:9ڶ>ST=O>I,TJ@Ԛ<53ρ>ׄ9?B:9ST=O>I,TJ@Ԛ<>ׄ9?ϪJJ1>>ׄ9?ϪJJ1>DB=>19Tׄ9?ׄ9B9>>Ư8I>؞Cб †M86><=>19Tׄ9?ׄ9B9>>Ư8Iρ>б †M8>ׄ9?ϪJJ1>>ׄ9?ϪJJ1>;9>؞Cׄ9?B:9ڶ>ST=O>I,TJ@Ԛ<53ρ>ׄ9?B:9ST=O>I,TJ@Ԛ<>ׄ9?ϪJJ1>>ׄ9?ϪJJ1>#!U—P۴2>MN,BMСGTUP>MNBMСGT>ׄ9?ϪJJ1>>ׄ9?ϪJJ1>;9>؞Cׄ9?B:9ڶ>ST=O>I,TJ@Ԛ<53ρ>ׄ9?B:9ST=O>I,TJ@Ԛ<>ׄ9?ϪJJ1>>ׄ9?ϪJJ1>hf1 TSׄ9?AJ9JOT,Q SF>T9P,1R>؞Cб :6)ʪ_]1 TSׄ9?A˱9OT,Q SF>T9P,1Rρ>б :6)ʪ>ׄ9?ϪJJ1>>ׄ9?ϪJJ1>;9>؞Cׄ9?B:9ڶ>ST=O>I,TJ@Ԛ<53ρ>ׄ9?B:9ST=O>I,TJ@Ԛ<>ׄ9?ϪJJ1>>ׄ9?ϪJJ1>A?:91STׄ9?9M,.T>BϪJ9>؞C@@@;9:91STׄ9?9M,.T>BϪJ9ρ>@@#!&6D>49@P>2#!&6D>49@P>26942A7B694+#!&6D>49@P>2#!&6D>49@P>2#!C14>@D2>@Ԛ<#!C14>@D2>@Ԛ<#!&6D>49@P>2#!&6D>49@P>2;9&FD6D249@D2>1XJVV53&FD6D249@D2>1JV#!&6D>49@P>2#!&6D>49@P>2SQ&L492IщQ—P=&1X4BD71XG:&T6GEީ L492IщQ=&14BD71G:&T6#!&6D>49@P>2#!&6D>49@P>26942A7B694+#!&6D>49@P>2#!&6D>49@P>2 Cڜ>42K.B@KCڜ>4K.@K#!&6D>49@P>2#!&6D>49@P>2;9&FD6D249@D2>1XJVV53&FD6D249@D2>1JV#!&6D>49@P>2#!&6D>49@P>2;961&6P>429Q1@&@@@2061&6P>4ƋQ1@&@@#!&6D>49@P>2#!&6D>49@P>26942A7B694+#!&6D>49@P>2#!&6D>49@P>2#! 6E424ڜ>2AЍ6E44ڜ>2A#!&6D>49@P>2#!&6D>49@P>2;9&FD6D249@D2>1XJVV53&FD6D249@D2>1JV#!&6D>49@P>2#!&6D>49@P>26>42EX@N6>4E@N#!&6D>49@P>2#!&6D>49@P>26942A7B694+#!&6D>49@P>2#!&6D>49@P>2>4ڜ>F5@Ԛ<>4ڜ>F5@Ԛ<#!&6D>49@P>2#!&6D>49@P>2;9&FD6D249@D2>1XJVV53&FD6D249@D2>1JV#!&6D>49@P>2#!&6D>49@P>2;9CRW6?۱URT:R&6D>62486CRW6?۱URT:R&6D>D4UEϨHWV@8Ǡ2ϨHWV@8>V;>Wٟ@2>6@2>6>؞C@@@53Ǡ2>V;>Wٟ@2>6@2>6ρ>@@UEϨHWV@8Ǡ2ϨHWV@8MKUE;6֊2>W6,ϨH@FL6,B,TE;>A7BDBǠ2;6֊2>W6,ϨH@FL6,B,TE;>+UEϨHWV@8Ǡ2ϨHWV@8>V;>Wٟ@2>6@2>6>؞C@@@53Ǡ2>V;>Wٟ@2>6@2>6ρ>@@UEϨHWV@8Ǡ2ϨHWV@853DUE>W@P21HSV9;W@Ԛ<,*DǠ2>W@71HSV9=@Ԛ<UEϨHWV@8Ǡ2ϨHWV@8>V;>Wٟ@2>6@2>6>؞C@@@53Ǡ2>V;>Wٟ@2>6@2>6ρ>@@UEϨHWV@8Ǡ2ϨHWV@886UE2V=L296T=ȟN2DS>؞CԚ<20Ǡ22V=L296T=ȟN2DSρ>Ԛ<UEϨHWV@8Ǡ2ϨHWV@8>V;>Wٟ@2>6@2>6>؞C@@@53Ǡ2>V;>Wٟ@2>6@2>6ρ>@@UEϨHWV@8Ǡ2ϨHWV@8JH>؞CXAN;WSV626DUE=WL6,6@Ԛ<A?ρ>XAN=SV626DǠ2=WL6,6@Ԛ<UEϨHWV@8Ǡ2ϨHWV@8>V;>Wٟ@2>6@2>6>؞C@@@53Ǡ2>V;>Wٟ@2>6@2>6ρ>@@UEϨHWV@8Ǡ2ϨHWV@8PN9;2UEDSV16=GB<6>؞Cб :6)ʪDB9;2Ǡ2DSV16=G<ρ>б :6)ʪUEϨHWV@8Ǡ2ϨHWV@8>V;>Wٟ@2>6@2>6>؞C@@@53Ǡ2>V;>Wٟ@2>6@2>6ρ>@@UEϨHWV@8Ǡ2ϨHWV@8DBD7>UE;ASVϨH,ϨHW;62>T6@Ԛ<>Ǡ2;ASV؋8ϨHW;62>T6@Ԛ<UEϨHWV@8Ǡ2ϨHWV@8>V;>Wٟ@2>6@2>6>؞C@@@53Ǡ2>V;>Wٟ@2>6@2>6ρ>@@UEϨHWV@8Ǡ2ϨHWV@8;9>UE;FWOT7,>A8SVDPDA20>Ǡ2;FW37,>ASVDPDAUEϨHWV@8Ǡ2ϨHWV@8>V;>Wٟ@2>6@2>6>؞C@@@53Ǡ2>V;>Wٟ@2>6@2>6ρ>@@UEϨHWV@8Ǡ2ϨHWV@886UE616=V6>6L=>؞C@@@/-Ǡ2616=V6>6L=ρ>@@UEϨHWV@8Ǡ2ϨHWV@8>V;>Wٟ@2>6@2>6>؞C@@@53Ǡ2>V;>Wٟ@2>6@2>6ρ>@@UEϨHWV@8Ǡ2ϨHWV@8)'>؞CUE;V626DSDA#!ρ>Ǡ2;V626DSDA# UDT #UDT@@@# UDT #UDTJHR6># HL6M9ٟ@UVUӁGDܤK8<# @@@;9R6>#HL6M@UVUӁGA8<#@@# UDT #UDT@@@# UDT #UDT UUD,A#%@@@UUD,A#@@# UDT #UDT@@@# UDT #UDT20UN.T5ƛK,6I16#%@@@,*UN.T5ƛK,6I16#@@# UDT #UDT@@@# UDT #UDTDB9Ԛ<6ϪJ># >Q@D9DFҾWSܤK# @@@,*16>#>Q@9FҾWS#@@# UDT #UDT@@@# UDT #UDT20#%>UӁGD9D.7>#%@@@#!#>UӁG9.7>#@@# UDT #UDT@@@# UDT #UDTMKR, 9S=ɵOʡH9B>UUD=UL9TM# @@@;9R, SɵO9>UUD=UL9TM#@@# UDT #UDT@@@# UDT #UDT&$DПC,UӁGDܤK# @@@DПC,UӁGA#@@# UDT #UDT@@@# UDT #UDT&$E>F# UDK0@@@ E>F#UDK0@@# UDT #UDT@@@# UDT #UDT,*# UDK-щQRQ# @@@#!#UDK-щQRQ#@@T;JC;XH-T;C;XH-DB;JIٟ@FXH-EDܤKV3ET)ʪ86;@FحXH-EDV3ET)ʪT;JC;XH-T;C;XH-YW9T:B7ٟ@)X-;J%)ѾCTO7%T87FD0A?9T:B7ٟ@X-;)ѾCTO7%T8F0T;JC;XH-T;C;XH-86;J85SXH-8E6O@Ԛ<&$;85SXH-8E6@Ԛ<T;JC;XH-T;C;XH-&$5M;J.B7H1R@Ԛ<#!5M;.B7H1R@Ԛ<T;JC;XH-T;C;XH-><;J85SXH-8E6ODSDA,*;85SXH-8E6DSDAT;JC;XH-T;C;XH-PN;JA5DN8R8EBS;76XH-NFK,DPDAA?;A5DNRNBS;5XH-NF,DPDAT;JC;XH-T;C;XH-DB;JIٟ@FXH-EDܤKV3ET)ʪ86;@FحXH-EDV3ET)ʪT;JC;XH-T;C;XH-_]5M;JDCٟ@F26K:X-RB9S8@D69>ҾWD,DPDAPN5M;D@F26K:X-RB9S8@6ߖ>D,DPDAT;JC;XH-T;C;XH-86;J85SXH-8E6O@Ԛ<&$;85SXH-8E6@Ԛ<T;JC;XH-T;C;XH-;J١-ܤKS/@N;١-ܤKS@NT;JC;XH-T;C;XH-><;J85SXH-8E6ODSDA,*;85SXH-8E6DSDAT;JC;XH-T;C;XH-866C; X-NWHT;J)ʪ/-6C;X-NWH;)ʪ/-5DR9D93A8RR7.,*5DR9D93A8RR720R9D93AR7.6ǽ=DPDA,*R9D93AR76DPDA/-5DR9D93A8RR7.,*5DR9D93A8RR7GE5DR93A7.8RAƛK2TH?T!HA†M86A?5DR93A78RAƛK2TH?T!HA†M8/-5DR9D93A8RR7.,*5DR9D93A8RR720R9D93AR7.6ǽ=DPDA,*R9D93AR76DPDA/-5DR9D93A8RR7.,*5DR9D93A8RR7865RAб D93A7.8RA!@@@205RAб D93A78RA!@@/-5DR9D93A8RR7.,*5DR9D93A8RR720R9D93AR7.6ǽ=DPDA,*R9D93AR76DPDA/-5DR9D93A8RR7.,*5DR9D93A8RR7><ʡH9BR93AV7.RAϪJHA@@@209R93AV7RAϪJHA@@/-5DR9D93A8RR7.,*5DR9D93A8RR720R9D93AR7.6ǽ=DPDA,*R9D93AR76DPDA/-5DR9D93A8RR7.,*5DR9D93A8RR7;9!HA5DR9L9BR7.RϪJ,@Ԛ<86!HA5DR9L9BR7RϪJ,@Ԛ</-5DR9D93A8RR7.,*5DR9D93A8RR720R9D93AR7.6ǽ=DPDA,*R9D93AR76DPDA/-5DR9D93A8RR7.,*5DR9D93A8RR7ki5D93AJR7.BRFD3Bٟ@75Dٟ@7>HAKADP!HA†M86ec5D93AJR7BRFD3Bٟ@75Dٟ@7>HAKADP!HA†M8/-5DR9D93A8RR7.,*5DR9D93A8RR720R9D93AR7.6ǽ=DPDA,*R9D93AR76DPDA/-5DR9D93A8RR7.,*5DR9D93A8RR7b`5DR9L9DR7.3>3RQKUDA-D3D!HA†M86\Z5DR9L9DR73>3RQKUDA-D3D!HA†M8/-5DR9D93A8RR7.,*5DR9D93A8RR720R9D93AR7.6ǽ=DPDA,*R9D93AR76DPDA/-5DR9D93A8RR7.,*5DR9D93A8RR7/-5D9L9D7.RƭI!@@@)'5D9L9D7RƭI!@@/-5DR9D93A8RR7.,*5DR9D93A8RR720R9D93AR7.6ǽ=DPDA,*R9D93AR76DPDA/-5DR9D93A8RR7.,*5DR9D93A8RR7hf95L9DRG7.3AWDEWKѾCHT7HA7:6)ʪVT95L9DRG73AWDEWK5:6)ʪ/-5DR9D93A8RR7.,*5DR9D93A8RR720R9D93AR7.6ǽ=DPDA,*R9D93AR76DPDA/-5DR9D93A8RR7.,*5DR9D93A8RR7><ϪJAHARA9D93ARADϪJ7.K5;9ϪJAHARA9D93ARADϪJ7K5.6O<-  .O</-<-N<-%%O܊7<0>T&$<-N<%%O܊70>.6O<-  .O<DBW<-7R:.6O/1EPٟ@9ٟ@MBʔ77>P/-W<7R:1EP9MBݔ7>P.6O<-  .O</-<-N<-%%O܊7<0>T&$<-N<%%O܊70>.6O<-  .O<R-R-.6O<-  .O</-<-N<-%%O܊7<0>T&$<-N<%%O܊70>.6O<-  .O<539.T&$<-N<%%O܊70>.6O<-  .O<20.6O37;0G .6.6T&$<-N<%%O܊70>.6O<-  .O<.6O8I6T.O8I6T.6O<-  .O</-<-N<-%%O܊7<0>T&$<-N<%%O܊70>.6O<-  .O<.6OTK6.OTK6.6O<-  .O</-<-N<-%%O܊7<0>T&$<-N<%%O܊70>.6O<-  .O<20 .6O/EED.6O/EE"W#! ED.6O/E"W.6O<-  .O</-<-N<-%%O܊7<0>T&$<-N<%%O܊70>.6O<-  .O<DBW<-7R:.6O/1EPٟ@9ٟ@MBDSDA20W<7R:1EP9MBDSDA.6O<-  .O</-<-N<-%%O܊7<0>T&$<-N<%%O܊70>.6O<-  .O<>4M54Н?A3AT UD1ձM4M54AA  UD1  UD153UD1ۓRD;16ǁRK3K"'!)'UD1ۓRD;16ǁRK3K  UD1  UD1,*UD1CT%8>9S1ME;)'UD1CT%8>9S1M;  UD1  UD153UD1ۓRD;16ǁRK3K"'!)'UD1ۓRD;16ǁRK3K  UD1  UD1 UD1U3ʡHWRDU UD1U3ʡHWRDU  UD1  UD153UD1ۓRD;16ǁRK3K"'!)'UD1ۓRD;16ǁRK3K  UD1  UD1 UD1M>4M5G3UD1ձM4M5G  UD1  UD153UD1ۓRD;16ǁRK3K"'!)'UD1ۓRD;16ǁRK3K  UD1  UD1/-U8JD1UHAʡH RGM=T,*U8JD1UHAʡH RGM=  UD1  UD153UD1ۓRD;16ǁRK3K"'!)'UD1ۓRD;16ǁRK3K  UD1  UD120UD1FBLL¶7JѾC4W,M4;#!UD1BNJCW14;  UD1  UD153UD1ۓRD;16ǁRK3K"'!)'UD1ۓRD;16ǁRK3K  UD1  UD1&$UD1ʡHR:DGAʈO>6#!UD1ʡHR:DGA>6  UD1  UD153UD1ۓRD;16ǁRK3K"'!)'UD1ۓRD;16ǁRK3K  UD1  UD1)'UȂ3.1PD>J١-- AB&$UȂ3.1P>J١-- AB  UD1  UD153UD1ۓRD;16ǁRK3K"'!)'UD1ۓRD;16ǁRK3K  UD1  UD1SQ>KU9D1M.OGUʡH9>9U199>U69IT@Ԛ<DB>KU9D1M.ǼOU>9U19>U6IT@Ԛ<  UD1  UD153UD1ۓRD;16ǁRK3K"'!)'UD1ۓRD;16ǁRK3K  UD1  UD1#!UD19ҧK1B—PϪJ>D UD19ҧK1BJ>D  UD1  UD153UD1ۓRD;16ǁRK3K"'!)'UD1ۓRD;16ǁRK3K  UD1  UD1#!UD19ҧK1B—PϪJ>D UD19ҧK1BJ>D  UD1  UD153UD1ۓRD;16ǁRK3K"'!)'UD1ۓRD;16ǁRK3K  UD1  UD1)'UȂ31MC—PQ>DԃPEAB U͂3M—PQ>DUAB  UD1  UD153UD1ۓRD;16ǁRK3K"'!)'UD1ۓRD;16ǁRK3K  UD1  UD1)'UD1—PRޚ6HU49QÐWB&$UD1Rޚ6HU49QÐWB  UD1  UD153UD1ۓRD;16ǁRK3K"'!)'UD1ۓRD;16ǁRK3K  UD1  UD1&$UD1W>β7UщQDG@K UD1W>ƴ7щQD@K  UD1  UD153UD1ۓRD;16ǁRK3K"'!)'UD1ۓRD;16ǁRK3K  UD1  UD1UD1MʡHRHUUD1MʡHRHU  UD1  UD153UD1ۓRD;16ǁRK3K"'!)'UD1ۓRD;16ǁRK3K  UD1  UD1/-UD1@1GM3̛<:9T!#!UD1@1GM3:TW  UD1  UD153UD1ۓRD;16ǁRK3K"'!)'UD1ۓRD;16ǁRK3K  UD1  UD1UD1>NVNFUD1>NVNF  UD1  UD153UD1ۓRD;16ǁRK3K"'!)'UD1ۓRD;16ǁRK3K  UD1  UD1,*UD1MʡHWRHUJ6J7,*UD1MʡHWRHUJ6J7  UD1  UD153UD1ۓRD;16ǁRK3K"'!)'UD1ۓRD;16ǁRK3K  UD1  UD1JHUBMBUD19ҧK1B—PϪJ>DS1UBDBN@Ԛ<>Dū1UDBN@Ԛ<  UD1  UD153UD1ۓRD;16ǁRK3K"'!)'UD1ۓRD;16ǁRK3K  UD1  UD120AʋMQU,D1U>4,3T5=T&$AQU,D1U>435=  UD1  UD153UD1ۓRD;16ǁRK3K"'!)'UD1ۓRD;16ǁRK3K  UD1  UD1534UD1M/5S7H47 N H)'4UD1M5S N H  UD1  UD153UD1ۓRD;16ǁRK3K"'!)'UD1ۓRD;16ǁRK3K  UD1  UD1)'UD1U/VӲU>/=WQT UD1*ӲU>/=WQ  UD1  UD153UD1ۓRD;16ǁRK3K"'!)'UD1ۓRD;16ǁRK3K  UD1  UD1DBUD1UʡHWR6U>G=SU/T()!/-UD1UʡHWR6U>GSUT  UD1  UD153UD1ۓRD;16ǁRK3K"'!)'UD1ۓRD;16ǁRK3K  UD1  UD1#!UD1M>U.61D3̛<2/ќ6HQT&$UD1KU>1D3/HQ  UD1  UD153UD1ۓRD;16ǁRK3K"'!)'UD1ۓRD;16ǁRK3K  UD1  UD1DBUD1ۓR4H5B—PϪJ>D3KT((!)'UD1ۓR4H5BJ>D3KEԼOR@C/8EC/8#!HԼOR@C/8>ٟ@@Ԛ<HC/8>ٟ@@Ԛ<EԼOR@C/8EC/8qoHԼOR@C/D8>ٟ@8 P@NLΊ;J@>@BΊ;RP@NLΊ;J@>DOָ:?ThfHC/D8>ٟ@8 P@NLΊ;J@>@BΊ;RP@NLΊ;J@>Dָ:?TEԼOR@C/8EC/8ԼO@K@K OK@KEԼOR@C/8EC/8nlHԼOR@C/8>ٟ@8PHۇLBDCɕH5ǟVGRPHGLBD=ږH5DOָ:?TecHC/8>ٟ@8PHۇLBDCɕH5ǟVGRPHGLBD=ږH5Dָ:?TEԼOR@C/8EC/8><ԼOR@E>1>THIԓ4C/8>ٟ@A7B,*E>1>TC/8>ٟ@+EԼOR@C/8EC/8}{M:İU;ԼOR@?R5BD5@E7K՞RWKD5C/8>ٟ@>:T(!K;86_]:?R5BD5@EG՞RʼGD5C/8>ٟ@>:TK;86EԼOR@C/8EC/8#!HԼOR@C/8>ٟ@@Ԛ<HC/8>ٟ@@Ԛ<EԼOR@C/8EC/8#!ß<:Dć?ԼO@C/8>ٟ@ ß<:Dć?OC/8>ٟ@EԼOR@C/8EC/8ԼO@K@K OK@KEԼOR@C/8EC/8,*HԼOR@NܒM̺2C/D8>ٟ@#!HNMC/D8>ٟ@EԼOR@C/8EC/8><ԼOR@E>1>THIԓ4C/8>ٟ@A7B,*E>1>TC/8>ٟ@+EԼOR@C/8EC/853HԼOR@C/D8>ٟ@RE@>DW/-HC/D8>ٟ@RE@>DWEԼOR@C/8EC/8#!HԼOR@C/8>ٟ@@Ԛ<HC/8>ٟ@@Ԛ<EԼOR@C/8EC/8&$EԼO@C/8>ٟ@DPDA#!EOC/8>ٟ@DPDAEԼOR@C/8EC/8ԼO@K@K OK@KEԼOR@C/8EC/8HԼOR@C/8>ٟ@D>AIH!D>HIH:@>Hٟ@/ў7:@՞R.ٟ@/ў79OEEXqoHC/8>ٟ@D>AIH!D>HIH:>Hٟ@/ў7:>ٟ@/ў79EEXEԼOR@C/8EC/8><ԼOR@E>1>THIԓ4C/8>ٟ@A7B,*E>1>TC/8>ٟ@+EԼOR@C/8EC/8;9HԼO@C/8>ٟ@H2992653HOC/8>ٟ@H2926EԼOR@C/8EC/8#!HԼOR@C/8>ٟ@@Ԛ<HC/8>ٟ@@Ԛ<EԼOR@C/8EC/820H?RSHIԓ4>ԼO@‹7C/8>ٟ@)'H?RS>O‹7C/8>ٟ@ \ No newline at end of file diff --git a/paddle/trainer/tests/gen_proto_data.py b/paddle/trainer/tests/gen_proto_data.py deleted file mode 100644 index 8cc6d44673b9f992c28ae95cc06db5ea5aca0642..0000000000000000000000000000000000000000 --- a/paddle/trainer/tests/gen_proto_data.py +++ /dev/null @@ -1,279 +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. - -from cStringIO import StringIO - -import paddle.proto.DataFormat_pb2 as DataFormat -from google.protobuf.internal.encoder import _EncodeVarint - -import logging -import pprint - -logging.basicConfig( - format='[%(levelname)s %(asctime)s %(filename)s:%(lineno)s] %(message)s', ) -logger = logging.getLogger('paddle') -logger.setLevel(logging.INFO) - -OOV_POLICY_IGNORE = 0 -OOV_POLICY_USE = 1 -OOV_POLICY_ERROR = 2 - -num_original_columns = 3 - -# Feature combination patterns. -# [[-1,0], [0,0]] means previous token at column 0 and current token at -# column 0 are combined as one feature. -patterns = [ - [[-2, 0]], - [[-1, 0]], - [[0, 0]], - [[1, 0]], - [[2, 0]], - [[-1, 0], [0, 0]], - [[0, 0], [1, 0]], - [[-2, 1]], - [[-1, 1]], - [[0, 1]], - [[1, 1]], - [[2, 1]], - [[-2, 1], [-1, 1]], - [[-1, 1], [0, 1]], - [[0, 1], [1, 1]], - [[1, 1], [2, 1]], - [[-2, 1], [-1, 1], [0, 1]], - [[-1, 1], [0, 1], [1, 1]], - [[0, 1], [1, 1], [2, 1]], -] - - -def make_features(sequence): - length = len(sequence) - num_features = len(sequence[0]) - - def get_features(pos): - if pos < 0: - return ['#B%s' % -pos] * num_features - if pos >= length: - return ['#E%s' % (pos - length + 1)] * num_features - return sequence[pos] - - for i in xrange(length): - for pattern in patterns: - fname = '/'.join([get_features(i + pos)[f] for pos, f in pattern]) - sequence[i].append(fname) - - -''' -Source file format: -Each line is for one timestep. The features are separated by space. -An empty line indicates end of a sequence. - -cutoff: a list of numbers. If count of a feature is smaller than this, - it will be ignored. -if oov_policy[i] is OOV_POLICY_USE, id 0 is reserved for OOV features of -i-th column. - -return a list of dict for each column -''' - - -def create_dictionaries(filename, cutoff, oov_policy): - def add_to_dict(sequence, dicts): - num_features = len(dicts) - for features in sequence: - l = len(features) - assert l == num_features, "Wrong number of features " + line - for i in xrange(l): - if features[i] in dicts[i]: - dicts[i][features[i]] += 1 - else: - dicts[i][features[i]] = 1 - - num_features = len(cutoff) - dicts = [] - for i in xrange(num_features): - dicts.append(dict()) - - f = open(filename, 'rb') - - sequence = [] - - for line in f: - line = line.strip() - if not line: - make_features(sequence) - add_to_dict(sequence, dicts) - sequence = [] - continue - features = line.split(' ') - sequence.append(features) - - for i in xrange(num_features): - dct = dicts[i] - n = 1 if oov_policy[i] == OOV_POLICY_USE else 0 - todo = [] - for k, v in dct.iteritems(): - if v < cutoff[i]: - todo.append(k) - else: - dct[k] = n - n += 1 - - if oov_policy[i] == OOV_POLICY_USE: - # placeholder so that len(dct) will be the number of features - # including OOV - dct['#OOV#'] = 0 - - logger.info('column %d dict size=%d, ignored %d' % (i, n, len(todo))) - for k in todo: - del dct[k] - - f.close() - return dicts - - -def encode_varint(v): - out = StringIO() - _EncodeVarint(out.write, v) - return out.getvalue() - - -def write_proto(file, message): - s = message.SerializeToString() - packed_len = encode_varint(len(s)) - file.write(packed_len + s) - - -''' -if oov_policy[i] == OOV_POLICY_USE, features in i-th column which are not -existed in dicts[i] will be assigned to id 0. -if oov_policy[i] == OOV_POLICY_ERROR, all features in i-th column MUST exist -in dicts[i]. -''' - - -def gen_proto_file(input_file, dicts, oov_policy, output_file): - def write_sequence(out, sequence): - num_features = len(dicts) - is_beginning = True - for features in sequence: - assert len(features) == num_features, \ - "Wrong number of features: " + line - sample = DataFormat.DataSample() - for i in xrange(num_original_columns): - id = dicts[i].get(features[i], -1) - if id != -1: - sample.id_slots.append(id) - elif oov_policy[i] == OOV_POLICY_IGNORE: - sample.id_slots.append(0xffffffff) - elif oov_policy[i] == OOV_POLICY_ERROR: - logger.fatal("Unknown token: %s" % features[i]) - else: - sample.id_slots.append(0) - - if patterns: - dim = 0 - vec = sample.vector_slots.add() - for i in xrange(num_original_columns, num_features): - id = dicts[i].get(features[i], -1) - if id != -1: - vec.ids.append(dim + id) - elif oov_policy[i] == OOV_POLICY_IGNORE: - pass - elif oov_policy[i] == OOV_POLICY_ERROR: - logger.fatal("Unknown token: %s" % features[i]) - else: - vec.ids.append(dim + 0) - - dim += len(dicts[i]) - - sample.is_beginning = is_beginning - is_beginning = False - write_proto(out, sample) - - num_features = len(dicts) - f = open(input_file, 'rb') - out = open(output_file, 'wb') - - header = DataFormat.DataHeader() - if patterns: - slot_def = header.slot_defs.add() - slot_def.type = DataFormat.SlotDef.VECTOR_SPARSE_NON_VALUE - slot_def.dim = sum( - [len(dicts[i]) for i in xrange(num_original_columns, len(dicts))]) - logger.info("feature_dim=%s" % slot_def.dim) - - for i in xrange(num_original_columns): - slot_def = header.slot_defs.add() - slot_def.type = DataFormat.SlotDef.INDEX - slot_def.dim = len(dicts[i]) - - write_proto(out, header) - - num_sequences = 0 - sequence = [] - for line in f: - line = line.strip() - if not line: - make_features(sequence) - write_sequence(out, sequence) - sequence = [] - num_sequences += 1 - continue - features = line.split(' ') - sequence.append(features) - - f.close() - out.close() - - logger.info("num_sequences=%s" % num_sequences) - - -dict2 = { - 'B-ADJP': 0, - 'I-ADJP': 1, - 'B-ADVP': 2, - 'I-ADVP': 3, - 'B-CONJP': 4, - 'I-CONJP': 5, - 'B-INTJ': 6, - 'I-INTJ': 7, - 'B-LST': 8, - 'I-LST': 9, - 'B-NP': 10, - 'I-NP': 11, - 'B-PP': 12, - 'I-PP': 13, - 'B-PRT': 14, - 'I-PRT': 15, - 'B-SBAR': 16, - 'I-SBAR': 17, - 'B-UCP': 18, - 'I-UCP': 19, - 'B-VP': 20, - 'I-VP': 21, - 'O': 22 -} - -if __name__ == '__main__': - cutoff = [3, 1, 0] - cutoff += [3] * len(patterns) - oov_policy = [OOV_POLICY_IGNORE, OOV_POLICY_ERROR, OOV_POLICY_ERROR] - oov_policy += [OOV_POLICY_IGNORE] * len(patterns) - dicts = create_dictionaries('trainer/tests/train.txt', cutoff, oov_policy) - dicts[2] = dict2 - gen_proto_file('trainer/tests/train.txt', dicts, oov_policy, - 'trainer/tests/train_proto.bin') - gen_proto_file('trainer/tests/test.txt', dicts, oov_policy, - 'trainer/tests/test_proto.bin') diff --git a/paddle/trainer/tests/test.txt b/paddle/trainer/tests/test.txt deleted file mode 100644 index 3ad503b34f2e1a84c632d0894f180b5cf9ac550a..0000000000000000000000000000000000000000 --- a/paddle/trainer/tests/test.txt +++ /dev/null @@ -1,1000 +0,0 @@ -Confidence NN B-NP -in IN B-PP -the DT B-NP -pound NN I-NP -is VBZ B-VP -widely RB I-VP -expected VBN I-VP -to TO I-VP -take VB I-VP -another DT B-NP -sharp JJ I-NP -dive NN I-NP -if IN B-SBAR -trade NN B-NP -figures NNS I-NP -for IN B-PP -September NNP B-NP -, , O -due JJ B-ADJP -for IN B-PP -release NN B-NP -tomorrow NN B-NP -, , O -fail VB B-VP -to TO I-VP -show VB I-VP -a DT B-NP -substantial JJ I-NP -improvement NN I-NP -from IN B-PP -July NNP B-NP -and CC I-NP -August NNP I-NP -'s POS B-NP -near-record JJ I-NP -deficits NNS I-NP -. . O - -Chancellor NNP O -of IN B-PP -the DT B-NP -Exchequer NNP I-NP -Nigel NNP B-NP -Lawson NNP I-NP -'s POS B-NP -restated VBN I-NP -commitment NN I-NP -to TO B-PP -a DT B-NP -firm NN I-NP -monetary JJ I-NP -policy NN I-NP -has VBZ B-VP -helped VBN I-VP -to TO I-VP -prevent VB I-VP -a DT B-NP -freefall NN I-NP -in IN B-PP -sterling NN B-NP -over IN B-PP -the DT B-NP -past JJ I-NP -week NN I-NP -. . O - -But CC O -analysts NNS B-NP -reckon VBP B-VP -underlying VBG B-NP -support NN I-NP -for IN B-PP -sterling NN B-NP -has VBZ B-VP -been VBN I-VP -eroded VBN I-VP -by IN B-PP -the DT B-NP -chancellor NN I-NP -'s POS B-NP -failure NN I-NP -to TO B-VP -announce VB I-VP -any DT B-NP -new JJ I-NP -policy NN I-NP -measures NNS I-NP -in IN B-PP -his PRP$ B-NP -Mansion NNP I-NP -House NNP I-NP -speech NN I-NP -last JJ B-NP -Thursday NNP I-NP -. . O - -This DT B-NP -has VBZ B-VP -increased VBN I-VP -the DT B-NP -risk NN I-NP -of IN B-PP -the DT B-NP -government NN I-NP -being VBG B-VP -forced VBN I-VP -to TO I-VP -increase VB I-VP -base NN B-NP -rates NNS I-NP -to TO B-PP -16 CD B-NP -% NN I-NP -from IN B-PP -their PRP$ B-NP -current JJ I-NP -15 CD I-NP -% NN I-NP -level NN I-NP -to TO B-VP -defend VB I-VP -the DT B-NP -pound NN I-NP -, , O -economists NNS B-NP -and CC O -foreign JJ B-NP -exchange NN I-NP -market NN I-NP -analysts NNS I-NP -say VBP B-VP -. . O - -`` `` O -The DT B-NP -risks NNS I-NP -for IN B-PP -sterling NN B-NP -of IN B-PP -a DT B-NP -bad JJ I-NP -trade NN I-NP -figure NN I-NP -are VBP B-VP -very RB B-ADVP -heavily RB I-ADVP -on IN B-PP -the DT B-NP -down JJ I-NP -side NN I-NP -, , O -'' '' O -said VBD B-VP -Chris NNP B-NP -Dillow NNP I-NP -, , O -senior JJ B-NP -U.K. NNP I-NP -economist NN I-NP -at IN B-PP -Nomura NNP B-NP -Research NNP I-NP -Institute NNP I-NP -. . O - -`` `` O -If IN B-SBAR -there EX B-NP -is VBZ B-VP -another DT B-NP -bad JJ I-NP -trade NN I-NP -number NN I-NP -, , O -there EX B-NP -could MD B-VP -be VB I-VP -an DT B-NP -awful JJ I-NP -lot NN I-NP -of IN B-PP -pressure NN B-NP -, , O -'' '' O -noted VBD B-VP -Simon NNP B-NP -Briscoe NNP I-NP -, , O -U.K. NNP B-NP -economist NN I-NP -for IN B-PP -Midland NNP B-NP -Montagu NNP I-NP -, , O -a DT B-NP -unit NN I-NP -of IN B-PP -Midland NNP B-NP -Bank NNP I-NP -PLC NNP I-NP -. . O - -Forecasts NNS B-NP -for IN B-PP -the DT B-NP -trade NN I-NP -figures NNS I-NP -range VBP B-VP -widely RB B-ADVP -, , O -but CC O -few JJ B-NP -economists NNS I-NP -expect VBP B-VP -the DT B-NP -data NNS I-NP -to TO B-VP -show VB I-VP -a DT B-NP -very RB I-NP -marked VBN I-NP -improvement NN I-NP -from IN B-PP -the DT O -# # O -2 CD O -billion CD O --LRB- ( O -$ $ B-ADJP -3.2 CD O -billion CD O --RRB- ) O -deficit NN B-NP -in IN B-PP -the DT B-NP -current JJ I-NP -account NN I-NP -reported VBD B-VP -for IN B-PP -August NNP B-NP -. . O - -The DT B-NP -August NNP I-NP -deficit NN I-NP -and CC O -the DT B-NP -# # I-NP -2.2 CD I-NP -billion CD I-NP -gap NN I-NP -registered VBN B-VP -in IN B-PP -July NNP B-NP -are VBP B-VP -topped VBN I-VP -only RB B-ADVP -by IN B-PP -the DT B-NP -# # I-NP -2.3 CD I-NP -billion CD I-NP -deficit NN I-NP -of IN B-PP -October NNP B-NP -1988 CD I-NP -. . O - -Sanjay NNP B-NP -Joshi NNP I-NP -, , O -European JJ B-NP -economist NN I-NP -at IN B-PP -Baring NNP B-NP -Brothers NNPS I-NP -& CC I-NP -Co. NNP I-NP -, , O -said VBD B-VP -there EX B-NP -is VBZ B-VP -no DT B-NP -sign NN I-NP -that IN B-SBAR -Britain NNP B-NP -'s POS B-NP -manufacturing NN I-NP -industry NN I-NP -is VBZ B-VP -transforming VBG I-VP -itself PRP B-NP -to TO B-VP -boost VB I-VP -exports NNS B-NP -. . O - -At IN B-PP -the DT B-NP -same JJ I-NP -time NN I-NP -, , O -he PRP B-NP -remains VBZ B-VP -fairly RB B-ADJP -pessimistic JJ I-ADJP -about IN B-PP -the DT B-NP -outlook NN I-NP -for IN B-PP -imports NNS B-NP -, , O -given VBN B-PP -continued VBD B-NP -high JJ I-NP -consumer NN I-NP -and CC I-NP -capital NN I-NP -goods NNS I-NP -inflows NNS I-NP -. . O - -He PRP B-NP -reckons VBZ B-VP -the DT B-NP -current JJ I-NP -account NN I-NP -deficit NN I-NP -will MD B-VP -narrow VB I-VP -to TO B-PP -only RB B-NP -# # I-NP -1.8 CD I-NP -billion CD I-NP -in IN B-PP -September NNP B-NP -. . O - -However RB B-ADVP -, , O -Mr. NNP B-NP -Dillow NNP I-NP -said VBD B-VP -he PRP B-NP -believes VBZ B-VP -that IN B-SBAR -a DT B-NP -reduction NN I-NP -in IN B-PP -raw JJ B-NP -material NN I-NP -stockbuilding VBG I-NP -by IN B-PP -industry NN B-NP -could MD B-VP -lead VB I-VP -to TO B-PP -a DT B-NP -sharp JJ I-NP -drop NN I-NP -in IN B-PP -imports NNS B-NP -. . O - -Combined VBN B-PP -with IN B-PP -at IN B-ADVP -least JJS I-ADVP -some DT B-NP -rebound NN I-NP -in IN B-PP -exports NNS B-NP -after IN B-PP -August NNP B-NP -'s POS B-NP -unexpected JJ I-NP -decline NN I-NP -, , O -the DT B-NP -deficit NN I-NP -could MD B-VP -narrow VB I-VP -to TO B-PP -as RB B-NP -little JJ I-NP -as IN I-NP -# # I-NP -1.3 CD I-NP -billion CD I-NP -. . O - -Mr. NNP B-NP -Briscoe NNP I-NP -, , O -who WP B-NP -also RB B-ADVP -forecasts VBZ B-VP -a DT B-NP -# # I-NP -1.3 CD I-NP -billion CD I-NP -current JJ I-NP -account NN I-NP -gap NN I-NP -, , O -warns VBZ B-VP -that IN B-SBAR -even RB B-SBAR -if IN I-SBAR -the DT B-NP -trade NN I-NP -figures NNS I-NP -are VBP B-VP -bullish JJ B-ADJP -for IN B-PP -sterling NN B-NP -, , O -the DT B-NP -currency NN I-NP -wo MD B-VP -n't RB I-VP -advance VB I-VP -much JJ B-NP -because IN B-SBAR -investors NNS B-NP -will MD B-VP -want VB I-VP -to TO I-VP -see VB I-VP -further JJ B-NP -evidence NN I-NP -of IN B-PP -the DT B-NP -turnaround NN I-NP -before IN B-PP -adjusting VBG B-VP -positions NNS B-NP -. . O - -Nevertheless RB B-ADVP -, , O -he PRP B-NP -noted VBD B-VP -, , O -`` `` O -No DT B-NP -one PRP I-NP -will MD B-VP -want VB I-VP -to TO I-VP -go VB I-VP -into IN B-PP -the DT B-NP -trade NN I-NP -figures NNS I-NP -without IN B-PP -a DT B-NP -flat JJ I-NP -position NN I-NP -'' '' O -in IN B-PP -the DT B-NP -pound NN I-NP -. . O - -Meanwhile RB B-ADVP -, , O -overall JJ B-NP -evidence NN I-NP -on IN B-PP -the DT B-NP -economy NN I-NP -remains VBZ B-VP -fairly RB B-ADJP -clouded VBN I-ADJP -. . O - -In IN B-PP -his PRP$ B-NP -Mansion NNP I-NP -House NNP I-NP -speech NN I-NP -, , O -Mr. NNP B-NP -Lawson NNP I-NP -warned VBD B-VP -that IN B-SBAR -a DT B-NP -further JJ I-NP -slowdown NN I-NP -can MD B-VP -be VB I-VP -expected VBN I-VP -as IN B-SBAR -the DT B-NP -impact NN I-NP -of IN B-PP -the DT B-NP -last JJ I-NP -rise NN I-NP -in IN B-PP -interest NN B-NP -rates NNS I-NP -earlier RBR B-NP -this DT I-NP -month NN I-NP -takes VBZ B-VP -effect NN B-NP -. . O - -U.K. JJ B-NP -base NN I-NP -rates NNS I-NP -are VBP B-VP -at IN B-PP -their PRP$ B-NP -highest JJS I-NP -level NN I-NP -in IN B-PP -eight CD B-NP -years NNS I-NP -. . O - -But CC O -consumer NN B-NP -expenditure NN I-NP -data NNS I-NP -released VBD B-VP -Friday NNP B-NP -do VBP B-VP -n't RB I-VP -suggest VB I-VP -that IN B-SBAR -the DT B-NP -U.K. NNP I-NP -economy NN I-NP -is VBZ B-VP -slowing VBG I-VP -that DT B-ADVP -quickly RB I-ADVP -. . O - -The DT B-NP -figures NNS I-NP -show VBP B-VP -that DT O -spending NN B-NP -rose VBD B-VP -0.1 CD B-NP -% NN I-NP -in IN B-PP -the DT B-NP -third JJ I-NP -quarter NN I-NP -from IN B-PP -the DT B-NP -second JJ I-NP -quarter NN I-NP -and CC O -was VBD B-VP -up IN B-ADVP -3.8 CD B-NP -% NN I-NP -from IN B-PP -a DT B-NP -year NN I-NP -ago RB B-ADVP -. . O - -This DT B-NP -compares VBZ B-VP -with IN B-PP -a DT B-NP -1.6 CD I-NP -% NN I-NP -rise NN I-NP -in IN B-PP -the DT B-NP -second NN I-NP -from IN B-PP -the DT B-NP -first JJ I-NP -quarter NN I-NP -and CC O -a DT B-NP -5.4 CD I-NP -% NN I-NP -increase NN I-NP -from IN B-PP -the DT B-NP -second JJ I-NP -quarter NN I-NP -of IN B-PP -1988 CD B-NP -. . O - -Mr. NNP B-NP -Dillow NNP I-NP -said VBD B-VP -the DT B-NP -data NNS I-NP -show VBP B-VP -the DT B-NP -economy NN I-NP -`` `` O -is VBZ B-VP -still RB B-ADVP -quite RB B-ADJP -strong JJ I-ADJP -, , O -'' '' O -but CC O -suggestions NNS B-NP -that IN B-SBAR -much NN B-NP -of IN B-PP -the DT B-NP -spending NN I-NP -went VBD B-VP -on IN B-PP -services NNS B-NP -rather RB B-PP -than IN I-PP -consumer NN B-NP -goods NNS I-NP -should MD B-VP -reduce VB I-VP -fears NNS B-NP -of IN B-PP -more JJR B-NP -import NN I-NP -rises NNS I-NP -. . O - -Certainly RB B-ADVP -, , O -the DT B-NP -chancellor NN I-NP -has VBZ B-VP -made VBN I-VP -it PRP B-NP -clear JJ B-ADJP -that IN B-SBAR -he PRP B-NP -is VBZ B-VP -prepared VBN I-VP -to TO I-VP -increase VB I-VP -interest NN B-NP -rates NNS I-NP -again RB B-ADVP -if IN B-SBAR -necessary JJ B-ADJP -to TO B-VP -both DT I-VP -ensure VB I-VP -that IN B-SBAR -a DT B-NP -substantial JJ I-NP -slowdown NN I-NP -does VBZ B-VP -take VB I-VP -place NN B-NP -and CC O -that DT O -sterling NN B-NP -does VBZ B-VP -n't RB I-VP -decline VB I-VP -further JJ B-ADVP -. . O - -Thursday NNP B-NP -, , O -he PRP B-NP -reminded VBD B-VP -his PRP$ B-NP -audience NN I-NP -that IN B-SBAR -the DT B-NP -government NN I-NP -`` `` O -can MD B-VP -not RB I-VP -allow VB I-VP -the DT B-NP -necessary JJ I-NP -rigor NN I-NP -of IN B-PP -monetary JJ B-NP -policy NN I-NP -to TO B-VP -be VB I-VP -undermined VBN I-VP -by IN B-PP -exchange NN B-NP -rate NN I-NP -weakness NN I-NP -. . O -'' '' O - -Analysts NNS B-NP -agree VBP B-VP -there EX B-NP -is VBZ B-VP -little JJ B-NP -holding NN B-VP -sterling NN B-NP -firm NN B-ADJP -at IN B-PP -the DT B-NP -moment NN I-NP -other JJ B-ADJP -than IN B-PP -Mr. NNP B-NP -Lawson NNP I-NP -'s POS B-NP -promise NN I-NP -that IN B-SBAR -rates NNS B-NP -will MD B-VP -be VB I-VP -pushed VBN I-VP -higher JJR B-ADJP -if IN B-SBAR -necessary JJ B-ADJP -. . O - -And CC O -, , O -they PRP B-NP -warn VBP B-VP -, , O -any DT B-NP -further JJ I-NP -drop NN I-NP -in IN B-PP -the DT B-NP -government NN I-NP -'s POS B-NP -popularity NN I-NP -could MD B-VP -swiftly RB I-VP -make VB I-VP -this DT B-NP -promise NN I-NP -sound NN B-VP -hollow JJ B-ADJP -. . O - -Sterling NNP B-NP -was VBD B-VP -already RB I-VP -showing VBG I-VP -some DT B-NP -signs NNS I-NP -of IN B-PP -a DT B-NP -lack NN I-NP -of IN B-PP -confidence NN B-NP -in IN B-PP -Mr. NNP B-NP -Lawson NNP I-NP -'s POS B-NP -promise NN I-NP -Friday NNP B-NP -. . O - -In IN B-PP -European JJ B-NP -trading NN I-NP -it PRP B-NP -declined VBD B-VP -to TO B-PP -$ $ B-NP -1.5890 CD I-NP -and CC O -2.9495 CD B-NP -marks NNS I-NP -from IN B-PP -$ $ B-NP -1.5940 CD I-NP -and CC O -2.9429 CD B-NP -marks NNS I-NP -late JJ B-NP -Thursday NNP I-NP -. . O - -Economists NNS B-NP -suggested VBD B-VP -that IN B-SBAR -if IN B-SBAR -the DT B-NP -pound NN I-NP -falls VBZ B-VP -much JJ B-NP -below IN B-PP -2.90 CD B-NP -marks NNS I-NP -, , O -the DT B-NP -government NN I-NP -will MD B-VP -be VB I-VP -forced VBN I-VP -to TO I-VP -increase VB I-VP -rates NNS B-NP -to TO B-PP -16 CD B-NP -% NN I-NP -, , O -both DT B-VP -to TO I-VP -halt VB B-VP -any DT B-NP -further JJ I-NP -decline NN I-NP -and CC O -ensure VB B-VP -that IN B-SBAR -the DT B-NP -balance NN I-NP -of IN B-PP -monetary JJ B-NP -policy NN I-NP -remains VBZ B-VP -unchanged JJ B-ADJP -. . O - -Friday NNP B-NP -'s POS B-NP -Market NNP I-NP -Activity NN I-NP - -The DT B-NP -dollar NN I-NP -posted VBD B-VP -gains NNS B-NP -in IN B-PP -quiet JJ B-NP -trading NN I-NP -as IN B-SBAR -concerns NNS B-NP -about IN B-PP -equities NNS B-NP -abated VBN B-VP -. . O - -Foreign JJ B-NP -exchange NN I-NP -dealers NNS I-NP -said VBD B-VP -that IN B-SBAR -the DT B-NP -currency NN I-NP -market NN I-NP -has VBZ B-VP -begun VBN I-VP -to TO I-VP -distance VB I-VP -itself PRP B-NP -from IN B-PP -the DT B-NP -volatile JJ I-NP -stock NN I-NP -exchange NN I-NP -, , O -which WDT B-NP -has VBZ B-VP -preoccupied VBN I-VP -the DT B-NP -market NN I-NP -since IN B-PP -Oct. NNP B-NP -13 CD I-NP -, , O -when WRB B-ADVP -the DT B-NP -Dow NNP I-NP -Jones NNP I-NP -Industrial NNP I-NP -Average NNP I-NP -plunged VBD B-VP -more JJR B-NP -than IN I-NP -190 CD I-NP -points NNS I-NP -. . O - -Currency NN B-NP -analysts NNS I-NP -predict VBP B-VP -that IN B-SBAR -in IN B-PP -the DT B-NP -coming VBG I-NP -week NN I-NP -the DT B-NP -foreign JJ I-NP -exchange NN I-NP -market NN I-NP -will MD B-VP -shift VB I-VP -its PRP$ B-NP -focus NN I-NP -back RB B-ADVP -to TO B-PP -economic JJ B-NP -fundamentals NNS I-NP -, , O -keeping VBG B-VP -a DT B-NP -close NN I-NP -eye NN I-NP -out IN B-ADVP -for IN B-PP -any DT B-NP -signs NNS I-NP -of IN B-PP -monetary JJ B-NP -easing NN I-NP -by IN B-PP -U.S. NNP B-NP -Federal NNP I-NP -Reserve NNP I-NP -. . O - -Late RB B-ADVP -in IN B-PP -the DT B-NP -New NNP I-NP -York NNP I-NP -trading NN I-NP -day NN I-NP -, , O -the DT B-NP -dollar NN I-NP -was VBD B-VP -quoted VBN I-VP -at IN B-PP -1.8578 CD B-NP -marks NNS I-NP -, , O -up IN B-ADVP -from IN B-PP -1.8470 CD B-NP -marks NNS I-NP -late JJ B-NP -Thursday NNP I-NP -in IN B-PP -New NNP B-NP -York NNP I-NP -. . O - -The DT B-NP -U.S. NNP I-NP -currency NN I-NP -was VBD B-VP -also RB I-VP -changing VBG I-VP -hands NNS B-NP -at IN B-PP -142.43 CD B-NP -yen NN I-NP -, , O -up IN B-ADVP -from IN B-PP -141.70 CD B-NP -yen NN I-NP -in IN B-PP -New NNP B-NP -York NNP I-NP -late JJ B-NP -Thursday NNP I-NP -. . O - -In IN B-PP -Tokyo NNP B-NP -on IN B-PP -Monday NNP B-NP -, , O -the DT B-NP -U.S. NNP I-NP -currency NN I-NP -opened VBD B-VP -for IN B-PP -trading NN B-NP -at IN B-PP -141.95 CD B-NP -yen NN I-NP -, , O -up IN B-ADVP -from IN B-PP -Friday NNP B-NP -'s POS B-NP -Tokyo NNP I-NP diff --git a/paddle/trainer/tests/test_Trainer.cpp b/paddle/trainer/tests/test_Trainer.cpp index 425b3d10a38086463784ba2a18db1293efe96e92..394038cf730f13cb957fbbc5ae0e5719b8fe9db6 100644 --- a/paddle/trainer/tests/test_Trainer.cpp +++ b/paddle/trainer/tests/test_Trainer.cpp @@ -24,7 +24,6 @@ using namespace std; // NOLINT static const string& configFile1 = "trainer/tests/sample_trainer_config.conf"; static const string& configFile2 = "trainer/tests/sample_trainer_config_hsigmoid.conf"; -static const string& configFile3 = "trainer/tests/chunking.conf"; static const string& configFile4 = "trainer/tests/sample_trainer_config_parallel.conf"; @@ -95,13 +94,6 @@ TEST(checkGradient, multi) { TEST(checkGradient, hsigmoid) { checkGradientTest(configFile2, false, false); } -TEST(checkGradient, chunk) { - checkGradientTest(configFile3, false, false); -#ifdef PADDLE_WITH_CUDA - checkGradientTest(configFile3, true, true); -#endif -} - TEST(checkGradient, non_parallel) { checkGradientTest(configFile4, false, false); } diff --git a/paddle/trainer/tests/test_config.conf b/paddle/trainer/tests/test_config.conf index d1bb9b877fe26702948586dbe90b9ff0ee27c1d6..2f86aaa75316fa2a5a28edfef31c01e15a44b3d0 100644 --- a/paddle/trainer/tests/test_config.conf +++ b/paddle/trainer/tests/test_config.conf @@ -15,12 +15,7 @@ from paddle.trainer_config_helpers import * -TrainData(ProtoData( - files = "dummy_list", - constant_slots = [1.0], - async_load_data = True)) - -TestData(SimpleData( +TrainData(SimpleData( files = "trainer/tests/sample_filelist.txt", feat_dim = 3, context_len = 0, diff --git a/paddle/trainer/tests/test_files.txt b/paddle/trainer/tests/test_files.txt deleted file mode 100644 index 49002677a848c499610d5e869ce61efb2105e3c8..0000000000000000000000000000000000000000 --- a/paddle/trainer/tests/test_files.txt +++ /dev/null @@ -1 +0,0 @@ -trainer/tests/test_proto.bin diff --git a/paddle/trainer/tests/train.list b/paddle/trainer/tests/train.list deleted file mode 100644 index f41e8e8893de6068deb43b08ec6a3bcdd4039326..0000000000000000000000000000000000000000 --- a/paddle/trainer/tests/train.list +++ /dev/null @@ -1 +0,0 @@ -trainer/tests/data_bin_part diff --git a/paddle/trainer/tests/train.txt b/paddle/trainer/tests/train.txt deleted file mode 100644 index 2313aee987ba71ba7ea779d3cf7705478e7fbde2..0000000000000000000000000000000000000000 --- a/paddle/trainer/tests/train.txt +++ /dev/null @@ -1,5000 +0,0 @@ -Confidence NN B-NP -in IN B-PP -the DT B-NP -pound NN I-NP -is VBZ B-VP -widely RB I-VP -expected VBN I-VP -to TO I-VP -take VB I-VP -another DT B-NP -sharp JJ I-NP -dive NN I-NP -if IN B-SBAR -trade NN B-NP -figures NNS I-NP -for IN B-PP -September NNP B-NP -, , O -due JJ B-ADJP -for IN B-PP -release NN B-NP -tomorrow NN B-NP -, , O -fail VB B-VP -to TO I-VP -show VB I-VP -a DT B-NP -substantial JJ I-NP -improvement NN I-NP -from IN B-PP -July NNP B-NP -and CC I-NP -August NNP I-NP -'s POS B-NP -near-record JJ I-NP -deficits NNS I-NP -. . O - -Chancellor NNP O -of IN B-PP -the DT B-NP -Exchequer NNP I-NP -Nigel NNP B-NP -Lawson NNP I-NP -'s POS B-NP -restated VBN I-NP -commitment NN I-NP -to TO B-PP -a DT B-NP -firm NN I-NP -monetary JJ I-NP -policy NN I-NP -has VBZ B-VP -helped VBN I-VP -to TO I-VP -prevent VB I-VP -a DT B-NP -freefall NN I-NP -in IN B-PP -sterling NN B-NP -over IN B-PP -the DT B-NP -past JJ I-NP -week NN I-NP -. . O - -But CC O -analysts NNS B-NP -reckon VBP B-VP -underlying VBG B-NP -support NN I-NP -for IN B-PP -sterling NN B-NP -has VBZ B-VP -been VBN I-VP -eroded VBN I-VP -by IN B-PP -the DT B-NP -chancellor NN I-NP -'s POS B-NP -failure NN I-NP -to TO B-VP -announce VB I-VP -any DT B-NP -new JJ I-NP -policy NN I-NP -measures NNS I-NP -in IN B-PP -his PRP$ B-NP -Mansion NNP I-NP -House NNP I-NP -speech NN I-NP -last JJ B-NP -Thursday NNP I-NP -. . O - -This DT B-NP -has VBZ B-VP -increased VBN I-VP -the DT B-NP -risk NN I-NP -of IN B-PP -the DT B-NP -government NN I-NP -being VBG B-VP -forced VBN I-VP -to TO I-VP -increase VB I-VP -base NN B-NP -rates NNS I-NP -to TO B-PP -16 CD B-NP -% NN I-NP -from IN B-PP -their PRP$ B-NP -current JJ I-NP -15 CD I-NP -% NN I-NP -level NN I-NP -to TO B-VP -defend VB I-VP -the DT B-NP -pound NN I-NP -, , O -economists NNS B-NP -and CC O -foreign JJ B-NP -exchange NN I-NP -market NN I-NP -analysts NNS I-NP -say VBP B-VP -. . O - -`` `` O -The DT B-NP -risks NNS I-NP -for IN B-PP -sterling NN B-NP -of IN B-PP -a DT B-NP -bad JJ I-NP -trade NN I-NP -figure NN I-NP -are VBP B-VP -very RB B-ADVP -heavily RB I-ADVP -on IN B-PP -the DT B-NP -down JJ I-NP -side NN I-NP -, , O -'' '' O -said VBD B-VP -Chris NNP B-NP -Dillow NNP I-NP -, , O -senior JJ B-NP -U.K. NNP I-NP -economist NN I-NP -at IN B-PP -Nomura NNP B-NP -Research NNP I-NP -Institute NNP I-NP -. . O - -`` `` O -If IN B-SBAR -there EX B-NP -is VBZ B-VP -another DT B-NP -bad JJ I-NP -trade NN I-NP -number NN I-NP -, , O -there EX B-NP -could MD B-VP -be VB I-VP -an DT B-NP -awful JJ I-NP -lot NN I-NP -of IN B-PP -pressure NN B-NP -, , O -'' '' O -noted VBD B-VP -Simon NNP B-NP -Briscoe NNP I-NP -, , O -U.K. NNP B-NP -economist NN I-NP -for IN B-PP -Midland NNP B-NP -Montagu NNP I-NP -, , O -a DT B-NP -unit NN I-NP -of IN B-PP -Midland NNP B-NP -Bank NNP I-NP -PLC NNP I-NP -. . O - -Forecasts NNS B-NP -for IN B-PP -the DT B-NP -trade NN I-NP -figures NNS I-NP -range VBP B-VP -widely RB B-ADVP -, , O -but CC O -few JJ B-NP -economists NNS I-NP -expect VBP B-VP -the DT B-NP -data NNS I-NP -to TO B-VP -show VB I-VP -a DT B-NP -very RB I-NP -marked VBN I-NP -improvement NN I-NP -from IN B-PP -the DT O -# # O -2 CD O -billion CD O --LRB- ( O -$ $ B-ADJP -3.2 CD O -billion CD O --RRB- ) O -deficit NN B-NP -in IN B-PP -the DT B-NP -current JJ I-NP -account NN I-NP -reported VBD B-VP -for IN B-PP -August NNP B-NP -. . O - -The DT B-NP -August NNP I-NP -deficit NN I-NP -and CC O -the DT B-NP -# # I-NP -2.2 CD I-NP -billion CD I-NP -gap NN I-NP -registered VBN B-VP -in IN B-PP -July NNP B-NP -are VBP B-VP -topped VBN I-VP -only RB B-ADVP -by IN B-PP -the DT B-NP -# # I-NP -2.3 CD I-NP -billion CD I-NP -deficit NN I-NP -of IN B-PP -October NNP B-NP -1988 CD I-NP -. . O - -Sanjay NNP B-NP -Joshi NNP I-NP -, , O -European JJ B-NP -economist NN I-NP -at IN B-PP -Baring NNP B-NP -Brothers NNPS I-NP -& CC I-NP -Co. NNP I-NP -, , O -said VBD B-VP -there EX B-NP -is VBZ B-VP -no DT B-NP -sign NN I-NP -that IN B-SBAR -Britain NNP B-NP -'s POS B-NP -manufacturing NN I-NP -industry NN I-NP -is VBZ B-VP -transforming VBG I-VP -itself PRP B-NP -to TO B-VP -boost VB I-VP -exports NNS B-NP -. . O - -At IN B-PP -the DT B-NP -same JJ I-NP -time NN I-NP -, , O -he PRP B-NP -remains VBZ B-VP -fairly RB B-ADJP -pessimistic JJ I-ADJP -about IN B-PP -the DT B-NP -outlook NN I-NP -for IN B-PP -imports NNS B-NP -, , O -given VBN B-PP -continued VBD B-NP -high JJ I-NP -consumer NN I-NP -and CC I-NP -capital NN I-NP -goods NNS I-NP -inflows NNS I-NP -. . O - -He PRP B-NP -reckons VBZ B-VP -the DT B-NP -current JJ I-NP -account NN I-NP -deficit NN I-NP -will MD B-VP -narrow VB I-VP -to TO B-PP -only RB B-NP -# # I-NP -1.8 CD I-NP -billion CD I-NP -in IN B-PP -September NNP B-NP -. . O - -However RB B-ADVP -, , O -Mr. NNP B-NP -Dillow NNP I-NP -said VBD B-VP -he PRP B-NP -believes VBZ B-VP -that IN B-SBAR -a DT B-NP -reduction NN I-NP -in IN B-PP -raw JJ B-NP -material NN I-NP -stockbuilding VBG I-NP -by IN B-PP -industry NN B-NP -could MD B-VP -lead VB I-VP -to TO B-PP -a DT B-NP -sharp JJ I-NP -drop NN I-NP -in IN B-PP -imports NNS B-NP -. . O - -Combined VBN B-PP -with IN B-PP -at IN B-ADVP -least JJS I-ADVP -some DT B-NP -rebound NN I-NP -in IN B-PP -exports NNS B-NP -after IN B-PP -August NNP B-NP -'s POS B-NP -unexpected JJ I-NP -decline NN I-NP -, , O -the DT B-NP -deficit NN I-NP -could MD B-VP -narrow VB I-VP -to TO B-PP -as RB B-NP -little JJ I-NP -as IN I-NP -# # I-NP -1.3 CD I-NP -billion CD I-NP -. . O - -Mr. NNP B-NP -Briscoe NNP I-NP -, , O -who WP B-NP -also RB B-ADVP -forecasts VBZ B-VP -a DT B-NP -# # I-NP -1.3 CD I-NP -billion CD I-NP -current JJ I-NP -account NN I-NP -gap NN I-NP -, , O -warns VBZ B-VP -that IN B-SBAR -even RB B-SBAR -if IN I-SBAR -the DT B-NP -trade NN I-NP -figures NNS I-NP -are VBP B-VP -bullish JJ B-ADJP -for IN B-PP -sterling NN B-NP -, , O -the DT B-NP -currency NN I-NP -wo MD B-VP -n't RB I-VP -advance VB I-VP -much JJ B-NP -because IN B-SBAR -investors NNS B-NP -will MD B-VP -want VB I-VP -to TO I-VP -see VB I-VP -further JJ B-NP -evidence NN I-NP -of IN B-PP -the DT B-NP -turnaround NN I-NP -before IN B-PP -adjusting VBG B-VP -positions NNS B-NP -. . O - -Nevertheless RB B-ADVP -, , O -he PRP B-NP -noted VBD B-VP -, , O -`` `` O -No DT B-NP -one PRP I-NP -will MD B-VP -want VB I-VP -to TO I-VP -go VB I-VP -into IN B-PP -the DT B-NP -trade NN I-NP -figures NNS I-NP -without IN B-PP -a DT B-NP -flat JJ I-NP -position NN I-NP -'' '' O -in IN B-PP -the DT B-NP -pound NN I-NP -. . O - -Meanwhile RB B-ADVP -, , O -overall JJ B-NP -evidence NN I-NP -on IN B-PP -the DT B-NP -economy NN I-NP -remains VBZ B-VP -fairly RB B-ADJP -clouded VBN I-ADJP -. . O - -In IN B-PP -his PRP$ B-NP -Mansion NNP I-NP -House NNP I-NP -speech NN I-NP -, , O -Mr. NNP B-NP -Lawson NNP I-NP -warned VBD B-VP -that IN B-SBAR -a DT B-NP -further JJ I-NP -slowdown NN I-NP -can MD B-VP -be VB I-VP -expected VBN I-VP -as IN B-SBAR -the DT B-NP -impact NN I-NP -of IN B-PP -the DT B-NP -last JJ I-NP -rise NN I-NP -in IN B-PP -interest NN B-NP -rates NNS I-NP -earlier RBR B-NP -this DT I-NP -month NN I-NP -takes VBZ B-VP -effect NN B-NP -. . O - -U.K. JJ B-NP -base NN I-NP -rates NNS I-NP -are VBP B-VP -at IN B-PP -their PRP$ B-NP -highest JJS I-NP -level NN I-NP -in IN B-PP -eight CD B-NP -years NNS I-NP -. . O - -But CC O -consumer NN B-NP -expenditure NN I-NP -data NNS I-NP -released VBD B-VP -Friday NNP B-NP -do VBP B-VP -n't RB I-VP -suggest VB I-VP -that IN B-SBAR -the DT B-NP -U.K. NNP I-NP -economy NN I-NP -is VBZ B-VP -slowing VBG I-VP -that DT B-ADVP -quickly RB I-ADVP -. . O - -The DT B-NP -figures NNS I-NP -show VBP B-VP -that DT O -spending NN B-NP -rose VBD B-VP -0.1 CD B-NP -% NN I-NP -in IN B-PP -the DT B-NP -third JJ I-NP -quarter NN I-NP -from IN B-PP -the DT B-NP -second JJ I-NP -quarter NN I-NP -and CC O -was VBD B-VP -up IN B-ADVP -3.8 CD B-NP -% NN I-NP -from IN B-PP -a DT B-NP -year NN I-NP -ago RB B-ADVP -. . O - -This DT B-NP -compares VBZ B-VP -with IN B-PP -a DT B-NP -1.6 CD I-NP -% NN I-NP -rise NN I-NP -in IN B-PP -the DT B-NP -second NN I-NP -from IN B-PP -the DT B-NP -first JJ I-NP -quarter NN I-NP -and CC O -a DT B-NP -5.4 CD I-NP -% NN I-NP -increase NN I-NP -from IN B-PP -the DT B-NP -second JJ I-NP -quarter NN I-NP -of IN B-PP -1988 CD B-NP -. . O - -Mr. NNP B-NP -Dillow NNP I-NP -said VBD B-VP -the DT B-NP -data NNS I-NP -show VBP B-VP -the DT B-NP -economy NN I-NP -`` `` O -is VBZ B-VP -still RB B-ADVP -quite RB B-ADJP -strong JJ I-ADJP -, , O -'' '' O -but CC O -suggestions NNS B-NP -that IN B-SBAR -much NN B-NP -of IN B-PP -the DT B-NP -spending NN I-NP -went VBD B-VP -on IN B-PP -services NNS B-NP -rather RB B-PP -than IN I-PP -consumer NN B-NP -goods NNS I-NP -should MD B-VP -reduce VB I-VP -fears NNS B-NP -of IN B-PP -more JJR B-NP -import NN I-NP -rises NNS I-NP -. . O - -Certainly RB B-ADVP -, , O -the DT B-NP -chancellor NN I-NP -has VBZ B-VP -made VBN I-VP -it PRP B-NP -clear JJ B-ADJP -that IN B-SBAR -he PRP B-NP -is VBZ B-VP -prepared VBN I-VP -to TO I-VP -increase VB I-VP -interest NN B-NP -rates NNS I-NP -again RB B-ADVP -if IN B-SBAR -necessary JJ B-ADJP -to TO B-VP -both DT I-VP -ensure VB I-VP -that IN B-SBAR -a DT B-NP -substantial JJ I-NP -slowdown NN I-NP -does VBZ B-VP -take VB I-VP -place NN B-NP -and CC O -that DT O -sterling NN B-NP -does VBZ B-VP -n't RB I-VP -decline VB I-VP -further JJ B-ADVP -. . O - -Thursday NNP B-NP -, , O -he PRP B-NP -reminded VBD B-VP -his PRP$ B-NP -audience NN I-NP -that IN B-SBAR -the DT B-NP -government NN I-NP -`` `` O -can MD B-VP -not RB I-VP -allow VB I-VP -the DT B-NP -necessary JJ I-NP -rigor NN I-NP -of IN B-PP -monetary JJ B-NP -policy NN I-NP -to TO B-VP -be VB I-VP -undermined VBN I-VP -by IN B-PP -exchange NN B-NP -rate NN I-NP -weakness NN I-NP -. . O -'' '' O - -Analysts NNS B-NP -agree VBP B-VP -there EX B-NP -is VBZ B-VP -little JJ B-NP -holding NN B-VP -sterling NN B-NP -firm NN B-ADJP -at IN B-PP -the DT B-NP -moment NN I-NP -other JJ B-ADJP -than IN B-PP -Mr. NNP B-NP -Lawson NNP I-NP -'s POS B-NP -promise NN I-NP -that IN B-SBAR -rates NNS B-NP -will MD B-VP -be VB I-VP -pushed VBN I-VP -higher JJR B-ADJP -if IN B-SBAR -necessary JJ B-ADJP -. . O - -And CC O -, , O -they PRP B-NP -warn VBP B-VP -, , O -any DT B-NP -further JJ I-NP -drop NN I-NP -in IN B-PP -the DT B-NP -government NN I-NP -'s POS B-NP -popularity NN I-NP -could MD B-VP -swiftly RB I-VP -make VB I-VP -this DT B-NP -promise NN I-NP -sound NN B-VP -hollow JJ B-ADJP -. . O - -Sterling NNP B-NP -was VBD B-VP -already RB I-VP -showing VBG I-VP -some DT B-NP -signs NNS I-NP -of IN B-PP -a DT B-NP -lack NN I-NP -of IN B-PP -confidence NN B-NP -in IN B-PP -Mr. NNP B-NP -Lawson NNP I-NP -'s POS B-NP -promise NN I-NP -Friday NNP B-NP -. . O - -In IN B-PP -European JJ B-NP -trading NN I-NP -it PRP B-NP -declined VBD B-VP -to TO B-PP -$ $ B-NP -1.5890 CD I-NP -and CC O -2.9495 CD B-NP -marks NNS I-NP -from IN B-PP -$ $ B-NP -1.5940 CD I-NP -and CC O -2.9429 CD B-NP -marks NNS I-NP -late JJ B-NP -Thursday NNP I-NP -. . O - -Economists NNS B-NP -suggested VBD B-VP -that IN B-SBAR -if IN B-SBAR -the DT B-NP -pound NN I-NP -falls VBZ B-VP -much JJ B-NP -below IN B-PP -2.90 CD B-NP -marks NNS I-NP -, , O -the DT B-NP -government NN I-NP -will MD B-VP -be VB I-VP -forced VBN I-VP -to TO I-VP -increase VB I-VP -rates NNS B-NP -to TO B-PP -16 CD B-NP -% NN I-NP -, , O -both DT B-VP -to TO I-VP -halt VB B-VP -any DT B-NP -further JJ I-NP -decline NN I-NP -and CC O -ensure VB B-VP -that IN B-SBAR -the DT B-NP -balance NN I-NP -of IN B-PP -monetary JJ B-NP -policy NN I-NP -remains VBZ B-VP -unchanged JJ B-ADJP -. . O - -Friday NNP B-NP -'s POS B-NP -Market NNP I-NP -Activity NN I-NP - -The DT B-NP -dollar NN I-NP -posted VBD B-VP -gains NNS B-NP -in IN B-PP -quiet JJ B-NP -trading NN I-NP -as IN B-SBAR -concerns NNS B-NP -about IN B-PP -equities NNS B-NP -abated VBN B-VP -. . O - -Foreign JJ B-NP -exchange NN I-NP -dealers NNS I-NP -said VBD B-VP -that IN B-SBAR -the DT B-NP -currency NN I-NP -market NN I-NP -has VBZ B-VP -begun VBN I-VP -to TO I-VP -distance VB I-VP -itself PRP B-NP -from IN B-PP -the DT B-NP -volatile JJ I-NP -stock NN I-NP -exchange NN I-NP -, , O -which WDT B-NP -has VBZ B-VP -preoccupied VBN I-VP -the DT B-NP -market NN I-NP -since IN B-PP -Oct. NNP B-NP -13 CD I-NP -, , O -when WRB B-ADVP -the DT B-NP -Dow NNP I-NP -Jones NNP I-NP -Industrial NNP I-NP -Average NNP I-NP -plunged VBD B-VP -more JJR B-NP -than IN I-NP -190 CD I-NP -points NNS I-NP -. . O - -Currency NN B-NP -analysts NNS I-NP -predict VBP B-VP -that IN B-SBAR -in IN B-PP -the DT B-NP -coming VBG I-NP -week NN I-NP -the DT B-NP -foreign JJ I-NP -exchange NN I-NP -market NN I-NP -will MD B-VP -shift VB I-VP -its PRP$ B-NP -focus NN I-NP -back RB B-ADVP -to TO B-PP -economic JJ B-NP -fundamentals NNS I-NP -, , O -keeping VBG B-VP -a DT B-NP -close NN I-NP -eye NN I-NP -out IN B-ADVP -for IN B-PP -any DT B-NP -signs NNS I-NP -of IN B-PP -monetary JJ B-NP -easing NN I-NP -by IN B-PP -U.S. NNP B-NP -Federal NNP I-NP -Reserve NNP I-NP -. . O - -Late RB B-ADVP -in IN B-PP -the DT B-NP -New NNP I-NP -York NNP I-NP -trading NN I-NP -day NN I-NP -, , O -the DT B-NP -dollar NN I-NP -was VBD B-VP -quoted VBN I-VP -at IN B-PP -1.8578 CD B-NP -marks NNS I-NP -, , O -up IN B-ADVP -from IN B-PP -1.8470 CD B-NP -marks NNS I-NP -late JJ B-NP -Thursday NNP I-NP -in IN B-PP -New NNP B-NP -York NNP I-NP -. . O - -The DT B-NP -U.S. NNP I-NP -currency NN I-NP -was VBD B-VP -also RB I-VP -changing VBG I-VP -hands NNS B-NP -at IN B-PP -142.43 CD B-NP -yen NN I-NP -, , O -up IN B-ADVP -from IN B-PP -141.70 CD B-NP -yen NN I-NP -in IN B-PP -New NNP B-NP -York NNP I-NP -late JJ B-NP -Thursday NNP I-NP -. . O - -In IN B-PP -Tokyo NNP B-NP -on IN B-PP -Monday NNP B-NP -, , O -the DT B-NP -U.S. NNP I-NP -currency NN I-NP -opened VBD B-VP -for IN B-PP -trading NN B-NP -at IN B-PP -141.95 CD B-NP -yen NN I-NP -, , O -up IN B-ADVP -from IN B-PP -Friday NNP B-NP -'s POS B-NP -Tokyo NNP I-NP -close NN I-NP -of IN B-PP -141.35 CD B-NP -yen NN I-NP -. . O - -On IN B-PP -the DT B-NP -Commodity NNP I-NP -Exchange NNP I-NP -in IN B-PP -New NNP B-NP -York NNP I-NP -, , O -gold NN B-NP -for IN B-PP -current JJ B-NP -delivery NN I-NP -settled VBD B-VP -at IN B-PP -$ $ B-NP -367.30 CD I-NP -an DT B-NP -ounce NN I-NP -, , O -up IN B-ADVP -20 CD B-NP -cents NNS I-NP -. . O - -Estimated VBN B-NP -volume NN I-NP -was VBD B-VP -a DT B-NP -light NN I-NP -2.4 CD I-NP -million CD I-NP -ounces NNS I-NP -. . O - -In IN B-PP -early JJ B-NP -trading NN I-NP -in IN B-PP -Hong NNP B-NP -Kong NNP I-NP -Monday NNP B-NP -, , O -gold NN B-NP -was VBD B-VP -quoted VBN I-VP -at IN B-PP -$ $ B-NP -366.50 CD I-NP -an DT B-NP -ounce NN I-NP -. . O - -East NNP B-NP -Rock NNP I-NP -Partners NNP I-NP -Limited NNP I-NP -Partnership NNP I-NP -said VBD B-VP -it PRP B-NP -proposed VBD B-VP -to TO I-VP -acquire VB I-VP -A.P. NNP B-NP -Green NNP I-NP -Industries NNP I-NP -Inc. NNP I-NP -for IN B-PP -$ $ B-NP -40 CD I-NP -a DT B-NP -share NN I-NP -. . O - -In IN B-PP -an DT B-NP -Oct. NNP I-NP -19 CD I-NP -letter NN I-NP -to TO B-PP -A.P. NNP B-NP -Green NNP I-NP -'s POS B-NP -board NN I-NP -, , O -East NNP B-NP -Rock NNP I-NP -said VBD B-VP -the DT B-NP -offer NN I-NP -is VBZ B-VP -subject NN B-ADJP -to TO B-PP -the DT B-NP -signing NN I-NP -of IN B-PP -a DT B-NP -merger NN I-NP -agreement NN I-NP -by IN B-PP -no DT B-ADVP -later RB I-ADVP -than IN B-PP -Oct. NNP B-NP -31 CD I-NP -. . O - -The DT B-NP -letter NN I-NP -, , O -attached VBN B-VP -to TO B-PP -a DT B-NP -filing NN I-NP -with IN B-PP -the DT B-NP -Securities NNP I-NP -and CC I-NP -Exchange NNP I-NP -Commission NNP I-NP -, , O -said VBD B-VP -the DT B-NP -approval NN I-NP -is VBZ B-VP -also RB B-ADVP -contingent JJ B-ADJP -upon IN B-PP -obtaining VBG B-VP -satisfactory JJ B-NP -financing NN I-NP -. . O - -An DT B-NP -A.P. NNP I-NP -Green NNP I-NP -official NN I-NP -declined VBD B-VP -to TO I-VP -comment VB I-VP -on IN B-PP -the DT B-NP -filing NN I-NP -. . O - -The DT B-NP -$ $ I-NP -40-a-share JJ I-NP -proposal NN I-NP -values VBZ B-VP -the DT B-NP -company NN I-NP -at IN B-PP -about RB B-NP -$ $ I-NP -106.6 CD I-NP -million CD I-NP -. . O - -A.P. NNP B-NP -Green NNP I-NP -currently RB B-ADVP -has VBZ B-VP -2,664,098 CD B-NP -shares NNS I-NP -outstanding JJ B-ADJP -. . O - -Its PRP$ B-NP -stock NN I-NP -closed VBD B-VP -at IN B-PP -$ $ B-NP -38 CD I-NP -, , O -up IN B-ADVP -$ $ B-NP -1.875 CD I-NP -, , O -in IN B-PP -national JJ B-NP -over-the-counter JJ I-NP -trading NN I-NP -. . O - -The DT B-NP -company NN I-NP -is VBZ B-VP -a DT B-NP -Mexico NNP I-NP -, , I-NP -Mo. NNP I-NP -, , I-NP -maker NN I-NP -of IN B-PP -refractory JJ B-NP -products NNS I-NP -. . O - -East NNP B-NP -Rock NNP I-NP -also RB B-ADVP -said VBD B-VP -in IN B-PP -the DT B-NP -filing NN I-NP -that IN B-SBAR -it PRP B-NP -boosted VBD B-VP -its PRP$ B-NP -stake NN I-NP -in IN B-PP -A.P. NNP B-NP -Green NNP I-NP -to TO B-PP -8.7 CD B-NP -% NN I-NP -. . O - -It PRP B-NP -now RB B-ADVP -holds VBZ B-VP -233,000 CD B-NP -A.P. NNP I-NP -Green NNP I-NP -common JJ I-NP -shares NNS I-NP -, , O -including VBG B-PP -30,000 CD B-NP -shares NNS I-NP -bought VBD B-VP -last JJ B-NP -Thursday NNP I-NP -for IN B-PP -$ $ B-NP -35.50 CD I-NP -to TO I-NP -$ $ I-NP -36.50 CD I-NP -a DT B-NP -share NN I-NP -. . O - -New NNP B-NP -York-based JJ I-NP -John NNP I-NP -Kuhns NNP I-NP -and CC I-NP -Robert NNP I-NP -MacDonald NNP I-NP -control NN B-VP -East NNP B-NP -Rock NNP I-NP -Partners NNP I-NP -Inc. NNP I-NP -, , O -the DT B-NP -sole JJ I-NP -general JJ I-NP -partner NN I-NP -of IN B-PP -East NNP B-NP -Rock NNP I-NP -Partners NNP I-NP -L.P NNP I-NP -. . O - -The DT B-NP -sole JJ I-NP -limited JJ I-NP -partner NN I-NP -of IN B-PP -the DT B-NP -partnership NN I-NP -is VBZ B-VP -Westwood NNP B-NP -Brick NNP I-NP -Lime NNP I-NP -Inc. NNP I-NP -, , O -an DT B-NP -indirect JJ I-NP -subsidiary NN I-NP -of IN B-PP -Westwood NNP B-NP -Group NNP I-NP -Inc NNP I-NP -. . O - -Both DT B-NP -Westwood NNP B-NP -Brick NNP I-NP -and CC O -Westwood NNP B-NP -Group NNP I-NP -are VBP B-VP -based VBN I-VP -in IN B-PP -Boston NNP B-NP -. . O - -Freight NN B-NP -rates NNS I-NP -, , O -declining VBG B-VP -for IN B-PP -most RBS B-NP -of IN B-PP -the DT B-NP -decade NN I-NP -because IN B-PP -of IN I-PP -competition NN B-NP -spurred VBN B-VP -by IN B-PP -deregulation NN B-NP -, , O -are VBP B-VP -bottoming VBG I-VP -out IN B-PRT -, , O -turning VBG B-VP -upward RB B-ADVP -and CC O -threatening VBG B-VP -to TO I-VP -fuel VB I-VP -inflation NN B-NP -. . O - -Trucking NNP B-NP -, , I-NP -shipping VBG I-NP -and CC I-NP -air-freight NN I-NP -companies NNS I-NP -have VBP B-VP -announced VBN I-VP -rate NN B-NP -increases NNS I-NP -, , O -scheduled VBN B-VP -for IN B-PP -this DT B-NP -fall NN I-NP -or CC O -early JJ B-NP -next JJ I-NP -year NN I-NP -, , O -reflecting VBG B-VP -higher JJR B-NP -costs NNS I-NP -and CC O -tightened VBD B-NP -demand NN I-NP -for IN B-PP -freight NN B-NP -transport NN I-NP -. . O - -Major JJ B-NP -shippers NNS I-NP -say VBP B-VP -they PRP B-NP -expect VBP B-VP -freight NN B-NP -rates NNS I-NP -to TO B-VP -rise VB I-VP -at IN B-ADVP -least JJS I-ADVP -as RB B-ADVP -fast RB I-ADVP -as IN B-PP -inflation NN B-NP -and CC B-ADVP -maybe RB I-ADVP -faster RBR B-ADVP -in IN B-PP -the DT B-NP -next JJ I-NP -few JJ I-NP -years NNS I-NP -. . O - -That DT B-NP -'s VBZ B-VP -a DT B-NP -big JJ I-NP -change NN I-NP -from IN B-PP -recent JJ B-NP -years NNS I-NP -when WRB B-ADVP -freight NN B-NP -haulage NN I-NP -was VBD B-VP -a DT B-NP -bright JJ I-NP -spot NN I-NP -for IN B-PP -U.S. NNP B-NP -productivity NN I-NP -, , O -helping VBG B-VP -to TO I-VP -restrain VB I-VP -inflation NN B-NP -and CC O -make VB B-VP -U.S. NNP B-NP -industry NN I-NP -more RBR B-ADJP -competitive JJ I-ADJP -abroad RB B-ADVP -. . O - -`` `` O -Demand NN B-NP -has VBZ B-VP -caught VBN I-VP -up IN B-PRT -with IN B-PP -the DT B-NP -supply NN I-NP -of IN B-PP -certain JJ B-NP -types NNS I-NP -of IN B-PP -freight NN B-NP -transportation NN I-NP -, , O -and CC O -rates NNS B-NP -are VBP B-VP -starting VBG I-VP -to TO I-VP -move VB I-VP -up IN B-ADVP -'' '' O -at IN B-PP -a DT B-NP -rate NN I-NP -`` `` O -close RB B-ADJP -to TO B-PP -or CC O -slightly RB B-ADJP -more JJR I-ADJP -than IN B-PP -the DT B-NP -inflation NN I-NP -rate NN I-NP -, , O -'' '' O -said VBD B-VP -Clifford NNP B-NP -Sayre NNP I-NP -, , O -director NN B-NP -of IN B-PP -logistics NNS B-NP -at IN B-PP -Du NNP B-NP -Pont NNP I-NP -Co NNP I-NP -. . O - -Shippers NNS B-NP -surveyed VBN B-VP -recently RB B-ADVP -by IN B-PP -Ohio NNP B-NP -State NNP I-NP -University NNP I-NP -said VBD B-VP -they PRP B-NP -expect VBP B-VP -their PRP$ B-NP -freight-transport JJ I-NP -, , I-NP -storage NN I-NP -and CC I-NP -distribution NN I-NP -costs NNS I-NP -to TO B-VP -rise VB I-VP -about IN B-NP -4 CD I-NP -% NN I-NP -this DT B-NP -year NN I-NP -. . O - -Only RB B-NP -10 CD I-NP -% NN I-NP -of IN B-PP -the DT B-NP -250 CD I-NP -shippers NNS I-NP -polled VBN B-VP -expected VBN B-VP -their PRP$ B-NP -freight-transport JJ I-NP -costs NNS I-NP -to TO B-VP -decrease VB I-VP -, , O -compared VBN B-PP -with IN B-PP -30 CD B-NP -% NN I-NP -who WP B-NP -had VBD B-VP -looked VBN I-VP -to TO B-PP -freight VB B-NP -transport NN I-NP -to TO B-VP -reduce VB I-VP -costs NNS B-NP -in IN B-PP -past JJ B-NP -years NNS I-NP -. . O - -`` `` O -This DT B-NP -is VBZ B-VP -the DT B-NP -first JJ I-NP -year NN I-NP -since IN B-PP -transportation NN B-NP -deregulation NN I-NP -in IN B-PP -1980 CD B-NP -that IN B-ADVP -we PRP B-NP -have VBP B-VP -had VBN I-VP -such JJ B-NP -a DT I-NP -dramatic JJ I-NP -and CC I-NP -broad-based JJ I-NP -upturn NN I-NP -in IN B-PP -perceived VBN B-NP -transportation NN I-NP -rates NNS I-NP -, , O -'' '' O -said VBD B-VP -Bernard NNP B-NP -LaLonde NNP I-NP -, , O -a DT B-NP -transportation NN I-NP -logistics NNS I-NP -professor NN I-NP -at IN B-PP -Ohio NNP B-NP -State NNP I-NP -in IN B-PP -Columbus NNP B-NP -. . O - -The DT B-NP -deregulation NN I-NP -of IN B-PP -railroads NNS B-NP -and CC I-NP -trucking NN I-NP -companies NNS I-NP -that WDT B-NP -began VBD B-VP -in IN B-PP -1980 CD B-NP -enabled VBD B-VP -shippers NNS B-NP -to TO B-VP -bargain VB I-VP -for IN B-PP -transportation NN B-NP -. . O - -Carriers NNP B-NP -could MD B-VP -use VB I-VP -their PRP$ B-NP -equipment NN I-NP -more RBR B-ADVP -efficiently RB I-ADVP -, , O -leading VBG B-VP -to TO B-PP -overcapacity NN B-NP -they PRP B-NP -were VBD B-VP -eager JJ B-ADJP -to TO B-VP -fill VB I-VP -. . O - -Shippers NNS B-NP -cut VBP B-VP -about RB B-NP -$ $ I-NP -35 CD I-NP -billion CD I-NP -from IN B-PP -their PRP$ B-NP -annual JJ I-NP -, , I-NP -inter-city JJ I-NP -truck NN I-NP -and CC I-NP -rail NN I-NP -costs NNS I-NP -, , O -to TO B-PP -about RB B-NP -$ $ I-NP -150 CD I-NP -billion CD I-NP -, , O -or CC O -about IN B-NP -6.4 CD I-NP -% NN I-NP -of IN B-PP -gross JJ B-NP -national JJ I-NP -product NN I-NP -, , O -down RB B-ADVP -from IN B-PP -8 CD B-NP -% NN I-NP -of IN B-PP -GNP NNP B-NP -in IN B-PP -1981 CD B-NP -. . O - -But CC O -with IN B-PP -much NN B-NP -of IN B-PP -the DT B-NP -inefficiency NN I-NP -squeezed VBN B-VP -out IN B-PP -of IN B-PP -the DT B-NP -freight-transport JJ I-NP -system NN I-NP -, , O -rising VBG B-NP -costs NNS I-NP -are VBP B-VP -likely JJ B-ADJP -to TO B-VP -be VB I-VP -reflected VBN I-VP -directly RB B-ADVP -in IN B-PP -higher JJR B-NP -freight NN I-NP -rates NNS I-NP -. . O - -`` `` O -Shippers NNS B-NP -are VBP B-VP -saying VBG I-VP -` `` O -the DT B-NP -party NN I-NP -'s POS B-VP -over IN B-ADJP -, , O -' '' O -'' '' O -said VBD B-VP -Mr. NNP B-NP -LaLonde NNP I-NP -. . O - -`` `` O -Shippers NNS B-NP -wo MD B-VP -n't RB I-VP -be VB I-VP -able JJ B-ADJP -to TO B-VP -look VB I-VP -for IN B-PP -transportation-cost JJ B-NP -savings NNS I-NP -as IN B-SBAR -they PRP B-NP -have VBP B-VP -for IN B-PP -the DT B-NP -last JJ I-NP -eight CD I-NP -or CC I-NP -nine CD I-NP -years NNS I-NP -. . O - -Transport NN B-NP -rates NNS I-NP -wo MD B-VP -n't RB I-VP -be VB I-VP -an DT B-NP -opportunity NN I-NP -for IN B-PP -offsetting VBG B-VP -cost NN B-NP -increases NNS I-NP -in IN B-PP -other JJ B-NP -segments NNS I-NP -of IN B-PP -the DT B-NP -economy NN I-NP -. . O -'' '' O - -Robert NNP B-NP -Delaney NNP I-NP -, , O -a DT B-NP -consultant NN I-NP -at IN B-PP -Arthur NNP B-NP -D. NNP I-NP -Little NNP I-NP -Inc. NNP I-NP -, , O -Cambridge NNP B-NP -, , O -Mass. NNP B-NP -, , O -said VBD B-VP -`` `` O -We PRP B-NP -'ve VBP B-VP -gotten VBN I-VP -all PDT B-NP -the DT I-NP -benefits NNS I-NP -of IN B-PP -deregulation NN B-NP -in IN B-PP -freight-cost JJ B-NP -reductions NNS I-NP -. . O - -Now RB B-ADVP -we PRP B-NP -are VBP B-VP -starting VBG I-VP -to TO I-VP -see VB I-VP -real JJ B-NP -freight-rate JJ I-NP -increases NNS I-NP -as IN B-SBAR -carriers NNS B-NP -replace VBP B-VP -equipment NN B-NP -, , O -pay VB B-VP -higher JJR B-NP -fuel NN I-NP -costs NNS I-NP -and CC O -pay VB B-VP -more JJR B-NP -for IN B-PP -labor NN B-NP -. . O - -You PRP B-NP -'ll MD B-VP -see VB I-VP -carriers NNS B-NP -try VB B-VP -to TO I-VP -recoup VB I-VP -some DT B-NP -of IN B-PP -the DT B-NP -price NN I-NP -cutting VBG I-NP -that WDT B-NP -occurred VBD B-VP -previously RB B-ADVP -. . O -'' '' O - -Not RB B-NP -everyone NN I-NP -believes VBZ B-VP -that IN B-SBAR -the DT B-NP -good JJ I-NP -times NNS I-NP -are VBP B-VP -over IN B-ADJP -for IN B-PP -shippers NNS B-NP -. . O - -`` `` O -There EX B-NP -'s VBZ B-VP -still RB B-ADVP -a DT B-NP -lot NN I-NP -of IN B-PP -pressure NN B-NP -on IN B-PP -rates NNS B-NP -in IN B-PP -both DT B-NP -rail NN I-NP -and CC I-NP -truck NN I-NP -, , O -'' '' O -said VBD B-VP -Gerard NNP B-NP -McCullough NNP I-NP -, , O -lecturer NN B-NP -in IN B-PP -transportation NN B-NP -at IN B-PP -Massachusetts NNP B-NP -Institute NNP I-NP -of IN B-PP -Technology NNP B-NP -. . O - -Less-than-truckload JJ B-NP -companies NNS I-NP -, , O -which WDT B-NP -carry VBP B-VP -the DT B-NP -freight NN I-NP -of IN B-PP -several JJ B-NP -shippers NNS I-NP -in IN B-PP -each DT B-NP -truck NN I-NP -trailer NN I-NP -, , O -discounted VBD B-VP -away RB B-ADVP -a DT B-NP -4.7 CD I-NP -% NN I-NP -rate NN I-NP -increase NN I-NP -implemented VBD B-VP -last JJ B-NP -April NNP I-NP -. . O - -The DT B-NP -carriers NNS I-NP -were VBD B-VP -competing VBG I-VP -fiercely RB B-ADVP -for IN B-PP -market NN B-NP -share NN I-NP -. . O - -Railroad-rate JJ B-NP -increases NNS I-NP -are VBP B-VP -likely JJ B-ADJP -to TO B-VP -be VB I-VP -restrained VBN I-VP -by IN B-PP -weakening VBG B-NP -rail-traffic JJ I-NP -levels NNS I-NP -and CC O -keen JJ B-NP -competition NN I-NP -for IN B-PP -freight NN B-NP -from IN B-PP -trucks NNS B-NP -. . O - -An DT B-NP -official NN I-NP -at IN B-PP -Consolidated NNP B-NP -Freightways NNP I-NP -Inc. NNP I-NP -, , O -a DT B-NP -Menlo NNP I-NP -Park NNP I-NP -, , I-NP -Calif. NNP I-NP -, , I-NP -less-than-truckload JJ I-NP -carrier NN I-NP -, , O -said VBD B-VP -rate NN B-NP -discounting NN I-NP -in IN B-PP -that DT B-NP -industry NN I-NP -has VBZ B-VP -begun VBN I-VP -to TO I-VP -`` `` O -stabilize VB B-VP -. . O -'' '' O - -Consolidated NNP B-NP -Freightways NNP I-NP -plans VBZ B-VP -to TO I-VP -raise VB I-VP -its PRP$ B-NP -rates NNS I-NP -5.3 CD B-NP -% NN I-NP -late JJ B-NP -this DT I-NP -year NN I-NP -or CC O -early JJ B-NP -next JJ I-NP -year NN I-NP -, , O -and CC O -at IN B-NP -least JJS I-NP -two CD I-NP -competitors NNS I-NP -have VBP B-VP -announced VBN I-VP -similar JJ B-NP -increases NNS I-NP -. . O - -Truckers NNS B-NP -are VBP B-VP -`` `` O -trying VBG B-VP -to TO I-VP -send VB I-VP -signals NNS B-NP -that IN B-SBAR -they PRP B-NP -need VBP B-VP -to TO I-VP -stop VB I-VP -the DT B-NP -bloodletting NN I-NP -, , O -forget VB B-VP -about IN B-PP -market NN B-NP -share NN I-NP -and CC O -go VB B-VP -for IN B-PP -higher JJR B-NP -rates NNS I-NP -, , O -'' '' O -said VBD B-VP -Michael NNP B-NP -Lloyd NNP I-NP -, , O -an DT B-NP -analyst NN I-NP -at IN B-PP -Salomon NNP B-NP -Bros NNP I-NP -. . O - -And CC O -`` `` O -shippers NNS B-NP -are VBP B-VP -getting VBG I-VP -the DT B-NP -feeling NN I-NP -that IN B-SBAR -they PRP B-NP -have VBP B-VP -played VBN I-VP -one CD B-NP -trucker NN I-NP -off IN B-ADVP -against IN B-PP -another DT B-NP -as RB B-NP -much JJ I-NP -as IN B-SBAR -they PRP B-NP -can MD B-VP -, , O -'' '' O -he PRP B-NP -said VBD B-VP -. . O - -Air-freight NN B-NP -carriers NNS I-NP -raised VBD B-VP -their PRP$ B-NP -rates NNS I-NP -for IN B-PP -U.S. NNP B-NP -products NNS I-NP -going VBG B-VP -across IN B-PP -the DT B-NP -Pacific NNP I-NP -to TO B-PP -Asia NNP B-NP -by IN B-PP -about IN B-NP -20 CD I-NP -% NN I-NP -earlier RBR B-NP -this DT I-NP -month NN I-NP -. . O - -And CC O -Japan NNP B-NP -Air NNP I-NP -Lines NNPS I-NP -said VBD B-VP -it PRP B-NP -plans VBZ B-VP -to TO I-VP -boost VB I-VP -its PRP$ B-NP -rates NNS I-NP -a DT B-NP -further JJ I-NP -25 CD I-NP -% NN I-NP -over IN B-PP -the DT B-NP -next JJ I-NP -two CD I-NP -years NNS I-NP -. . O - -Such JJ B-NP -rate NN I-NP -increases NNS I-NP -`` `` O -will MD B-VP -increase VB I-VP -the DT B-NP -total JJ I-NP -cost NN I-NP -of IN B-PP -U.S. NNP B-NP -products NNS I-NP -and CC O -slow JJ B-VP -down RP B-PRT -the DT B-NP -rate NN I-NP -of IN B-PP -increase NN B-NP -of IN B-PP -U.S. NNP B-NP -exports NNS I-NP -, , O -'' '' O -said VBD B-VP -Richard NNP B-NP -Connors NNP I-NP -, , O -a DT B-NP -senior JJ I-NP -vice NN I-NP -president NN I-NP -of IN B-PP -Yusen NNP B-NP -Air NNP I-NP -& CC I-NP -Sea NNP I-NP -Service NNP I-NP -U.S.A. NNP I-NP -Inc. NNP I-NP -, , O -the DT B-NP -U.S. NNP I-NP -air-freight-forwarding JJ I-NP -subsidiary NN I-NP -of IN B-PP -Nippon NNP B-NP -Yusen NNP I-NP -Kaisha NNP I-NP -of IN B-PP -Japan NNP B-NP -. . O - -Ship NN B-NP -companies NNS I-NP -carrying VBG B-VP -bulk NN B-NP -commodities NNS I-NP -, , O -such JJ B-PP -as IN I-PP -oil NN B-NP -, , O -grain NN B-NP -, , O -coal NN B-NP -and CC O -iron NN B-NP -ore NN I-NP -, , O -have VBP B-VP -been VBN I-VP -able JJ B-ADJP -to TO B-VP -increase VB I-VP -their PRP$ B-NP -rates NNS I-NP -in IN B-PP -the DT B-NP -last JJ I-NP -couple NN I-NP -of IN B-PP -years NNS B-NP -. . O - -Some DT B-NP -bulk NN I-NP -shipping VBG I-NP -rates NNS I-NP -have VBP B-VP -increased VBN I-VP -`` `` O -3 CD B-NP -% NN I-NP -to TO I-NP -4 CD I-NP -% NN I-NP -in IN B-PP -the DT B-NP -past JJ I-NP -few JJ I-NP -months NNS I-NP -, , O -'' '' O -said VBD B-VP -Salomon NNP B-NP -'s POS B-NP -Mr. NNP I-NP -Lloyd NNP I-NP -. . O - -And CC O -ship NN B-NP -lines NNS I-NP -carrying VBG B-VP -containers NNS B-NP -are VBP B-VP -also RB I-VP -trying VBG I-VP -to TO I-VP -raise VB I-VP -their PRP$ B-NP -rates NNS I-NP -. . O - -Carriers NNP B-NP -boosted VBD B-VP -rates NNS B-NP -more JJR B-NP -than IN I-NP -10 CD I-NP -% NN I-NP -in IN B-PP -the DT B-NP -North NNP I-NP -Atlantic NNP I-NP -between IN B-PP -the DT B-NP -U.S. NNP I-NP -and CC O -Europe NNP B-NP -last JJ B-NP -September NNP I-NP -, , O -hoping VBG B-VP -to TO I-VP -partly RB I-VP -restore VB I-VP -rates NNS B-NP -to TO B-PP -earlier JJR B-NP -levels NNS I-NP -. . O - -Ship NN B-NP -lines NNS I-NP -operating VBG B-VP -in IN B-PP -the DT B-NP -Pacific NNP I-NP -plan NN B-VP -to TO I-VP -raise VB I-VP -rates NNS B-NP -on IN B-PP -containers NNS B-NP -carrying VBG B-VP -U.S. NNP B-NP -exports NNS I-NP -to TO B-PP -Asia NNP B-NP -about IN B-NP -10 CD I-NP -% NN I-NP -, , O -effective JJ B-ADJP -next JJ B-NP -April NNP I-NP -. . O - -MGM NNP B-NP -Grand NNP I-NP -Inc. NNP I-NP -said VBD B-VP -it PRP B-NP -filed VBD B-VP -a DT B-NP -registration NN I-NP -statement NN I-NP -with IN B-PP -the DT B-NP -Securities NNP I-NP -and CC I-NP -Exchange NNP I-NP -Commission NNP I-NP -for IN B-PP -a DT B-NP -public JJ I-NP -offering NN I-NP -of IN B-PP -six CD B-NP -million CD I-NP -common JJ I-NP -shares NNS I-NP -. . O - -The DT B-NP -Beverly NNP I-NP -Hills NNP I-NP -, , I-NP -Calif.-based JJ I-NP -company NN I-NP -said VBD B-VP -it PRP B-NP -would MD B-VP -have VB I-VP -26.9 CD B-NP -million CD I-NP -common JJ I-NP -shares NNS I-NP -outstanding JJ B-ADJP -after IN B-PP -the DT B-NP -offering NN I-NP -. . O - -The DT B-NP -hotel NN I-NP -and CC I-NP -Gaming NNP I-NP -company NN I-NP -said VBD B-VP -Merrill NNP B-NP -Lynch NNP I-NP -Capital NNP I-NP -Markets NNPS I-NP -will MD B-VP -lead VB I-VP -the DT B-NP -underwriters NNS I-NP -. . O - -Proceeds NNS B-NP -from IN B-PP -the DT B-NP -sale NN I-NP -will MD B-VP -be VB I-VP -used VBN I-VP -for IN B-PP -remodeling VBG B-NP -and CC I-NP -refurbishing VBG I-NP -projects NNS I-NP -, , B-PP -as RB I-PP -well RB I-PP -as IN I-PP -for IN B-PP -the DT B-NP -planned VBN I-NP -MGM NNP I-NP -Grand NNP I-NP -hotel\/casino NN I-NP -and CC I-NP -theme NN I-NP -park NN I-NP -. . O - -Bob NNP B-NP -Stone NNP I-NP -stewed JJ B-VP -over IN B-PP -a DT B-NP -letter NN I-NP -from IN B-PP -his PRP$ B-NP -manager NN I-NP -putting VBG B-VP -him PRP B-NP -on IN B-PP -probation NN B-NP -for IN B-PP -insubordination NN B-NP -. . O - -Mr. NNP B-NP -Stone NNP I-NP -thought VBD B-VP -the DT B-NP -discipline NN I-NP -was VBD B-VP -unfair JJ B-ADJP -; : O -he PRP B-NP -believed VBD B-VP -that IN B-SBAR -his PRP$ B-NP -manager NN I-NP -wanted VBD B-VP -to TO I-VP -get VB I-VP -rid JJ B-ADJP -of IN B-PP -him PRP B-NP -for IN B-PP -personal JJ B-NP -reasons NNS I-NP -. . O - -Unable JJ B-ADJP -to TO B-VP -persuade VB I-VP -the DT B-NP -manager NN I-NP -to TO B-VP -change VB I-VP -his PRP$ B-NP -decision NN I-NP -, , O -he PRP B-NP -went VBD B-VP -to TO B-PP -a DT B-NP -`` `` I-NP -company NN I-NP -court NN I-NP -'' '' O -for IN B-PP -a DT B-NP -hearing NN I-NP -. . O - -At IN B-PP -the DT B-NP -scheduled VBN I-NP -time NN I-NP -, , O -Mr. NNP B-NP -Stone NNP I-NP -entered VBD B-VP -a DT B-NP -conference NN I-NP -room NN I-NP -in IN B-PP -a DT B-NP -building NN I-NP -near IN B-PP -where WRB B-ADVP -he PRP B-NP -worked VBD B-VP -. . O - -After IN B-SBAR -the DT B-NP -three CD I-NP -members NNS I-NP -of IN B-PP -the DT B-NP -court NN I-NP -introduced VBD B-VP -themselves PRP B-NP -, , O -the DT B-NP -chairman NN I-NP -of IN B-PP -the DT B-NP -panel NN I-NP -said VBD B-VP -: : O -`` `` O -Go VB B-VP -ahead RB B-ADVP -and CC O -tell VB B-VP -us PRP B-NP -what WP B-NP -happened VBD B-VP -. . O - -We PRP B-NP -may MD B-VP -ask VB I-VP -questions NNS B-NP -as IN B-SBAR -you PRP B-NP -go VBP B-VP -along IN B-PRT -, , O -or CC O -we PRP B-NP -may MD B-VP -wait VB I-VP -until IN B-PP -the DT B-NP -end NN I-NP -. . O -'' '' O - -No DT B-NP -lawyers NNS I-NP -or CC I-NP -tape NN I-NP -recorders NNS I-NP -were VBD B-VP -present JJ B-ADJP -. . O - -The DT B-NP -only RB I-NP -extra JJ I-NP -people NNS I-NP -were VBD B-VP -a DT B-NP -couple NN I-NP -of IN B-PP -personnel NNS B-NP -specialists NNS I-NP -, , O -one CD B-NP -of IN B-PP -whom WP B-NP -knew VBD B-VP -Mr. NNP B-NP -Stone NNP I-NP -'s POS B-NP -case NN I-NP -intimately RB B-ADVP -and CC O -would MD B-VP -help VB I-VP -fill VB I-VP -in IN B-PRT -any DT B-NP -facts NNS I-NP -needed VBN B-VP -to TO B-VP -give VB I-VP -the DT B-NP -court NN I-NP -the DT B-NP -full JJ I-NP -picture NN I-NP -. . O - -Over IN B-PP -a DT B-NP -cup NN I-NP -of IN B-PP -coffee NN B-NP -, , O -Mr. NNP B-NP -Stone NNP I-NP -told VBD B-VP -his PRP$ B-NP -story NN I-NP -. . O - -He PRP B-NP -talked VBD B-VP -about IN B-NP -20 CD I-NP -minutes NNS I-NP -. . O - -When WRB B-ADVP -he PRP B-NP -was VBD B-VP -through IN B-ADJP -, , O -the DT B-NP -court NN I-NP -members NNS I-NP -asked VBD B-VP -many JJ B-NP -questions NNS I-NP -, , O -then RB B-ADVP -the DT B-NP -chairman NN I-NP -said VBD B-VP -they PRP B-NP -would MD B-VP -like VB I-VP -to TO I-VP -hear VB I-VP -his PRP$ B-NP -manager NN I-NP -'s POS B-NP -side NN I-NP -and CC O -talk VB B-VP -to TO B-PP -witnesses NNS B-NP -. . O - -The DT B-NP -chairman NN I-NP -promised VBD B-VP -Mr. NNP B-NP -Stone NNP I-NP -a DT B-NP -decision NN I-NP -within IN B-PP -two CD B-NP -weeks NNS I-NP -. . O - -Bob NNP B-NP -Stone NNP I-NP -is VBZ B-VP -a DT B-NP -fictional JJ I-NP -name NN I-NP -, , O -but CC O -the DT B-NP -incident NN I-NP -described VBN B-VP -is VBZ B-VP -real JJ B-ADJP -. . O - -It PRP B-NP -happened VBD B-VP -at IN B-PP -Northrop NNP B-NP -Corp. NNP I-NP -in IN B-PP -Los NNP B-NP -Angeles NNP I-NP -. . O - -The DT B-NP -court NN I-NP -is VBZ B-VP -called VBN I-VP -the DT B-NP -Management NNP I-NP -Appeals NNP I-NP -Committee NNP I-NP -, , O -or CC O -just RB B-NP -`` `` I-NP -MAC NNP I-NP -, , O -'' '' O -and CC O -it PRP B-NP -is VBZ B-VP -likely JJ B-ADJP -to TO B-VP -hear VB I-VP -a DT B-NP -couple NN I-NP -of IN I-NP -dozen NN I-NP -cases VBZ I-NP -a DT B-NP -year NN I-NP -. . O - -Alter VB B-VP -some DT B-NP -details NNS I-NP -of IN B-PP -this DT B-NP -example NN I-NP -and CC O -it PRP B-NP -could MD B-VP -be VB I-VP -taking VBG I-VP -place NN B-NP -today NN B-ADVP -at IN B-PP -Federal NNP B-NP -Express NNP I-NP -in IN B-PP -Memphis NNP B-NP -, , O -the DT B-NP -Defense NNP I-NP -and CC I-NP -Underseas NNP I-NP -Systems NNP I-NP -divisions NNS I-NP -of IN B-PP -Honeywell NNP B-NP -in IN B-PP -Minneapolis NNP B-NP -, , O -a DT B-NP -General NNP I-NP -Electric NNP I-NP -plant NN I-NP -in IN B-PP -Columbia NNP B-NP -, , O -Md. NNP B-NP -, , O -or CC O -a DT B-NP -number NN I-NP -of IN B-PP -other JJ B-NP -companies NNS I-NP -. . O - -These DT B-NP -firms NNS I-NP -are VBP B-VP -pioneers NNS B-NP -in IN B-PP -a DT B-NP -significant JJ I-NP -new JJ I-NP -trend NN I-NP -in IN B-PP -the DT B-NP -corporate JJ I-NP -world NN I-NP -: : O -the DT B-NP -rise NN I-NP -of IN B-PP -what WP B-NP -I PRP B-NP -call VBP B-VP -corporate JJ B-NP -due JJ I-NP -process NN I-NP -. . O - -Although IN B-SBAR -corporate JJ B-NP -due JJ I-NP -process NN I-NP -is VBZ B-VP -practiced VBN I-VP -today NN B-NP -in IN B-PP -few JJ B-NP -companies NNS I-NP --- : O -perhaps RB B-ADVP -40 CD B-NP -to TO I-NP -60 CD I-NP --- : O -it PRP B-NP -is VBZ B-VP -one CD B-NP -of IN B-PP -the DT B-NP -fastest JJS I-NP -developing VBG I-NP -trends NNS I-NP -in IN B-PP -industry NN B-NP -. . O - -In IN B-PP -the DT B-NP -coming VBG I-NP -decade NN I-NP -a DT B-NP -majority NN I-NP -of IN B-PP -people-oriented JJ B-NP -companies NNS I-NP -are VBP B-VP -likely JJ B-ADJP -to TO B-VP -adopt VB I-VP -it PRP B-NP -. . O - -Corporate JJ B-NP -due JJ I-NP -process NN I-NP -appeals NNS B-VP -to TO B-PP -management NN B-NP -for IN B-PP -a DT B-NP -variety NN I-NP -of IN B-PP -reasons NNS B-NP -. . O - -It PRP B-NP -reduces VBZ B-VP -lawsuits NNS B-NP -from IN B-PP -disgruntled JJ B-NP -employees NNS I-NP -and CC I-NP -ex-employees NNS I-NP -, , O -with IN B-PP -all DT B-NP -that WDT B-NP -means VBZ B-VP -for IN B-PP -reduced VBN B-NP -legal JJ I-NP -costs NNS I-NP -and CC O -better RBR B-NP -public JJ I-NP -relations NNS I-NP -. . O - -It PRP B-NP -helps VBZ B-VP -to TO I-VP -keep VB I-VP -out IN B-PRT -unions NNS B-NP -. . O - -It PRP B-NP -increases VBZ B-VP -employee NN B-NP -commitment NN I-NP -to TO B-PP -the DT B-NP -company NN I-NP -, , O -with IN B-PP -all DT B-NP -that WDT B-NP -means VBZ B-VP -for IN B-PP -efficiency NN B-NP -and CC O -quality NN B-NP -control NN I-NP -. . O - -What WP B-NP -must MD O -your PRP$ B-NP -management NN I-NP -team NN I-NP -do VBP B-VP -to TO B-VP -establish VB I-VP -corporate JJ B-NP -due JJ I-NP -process NN I-NP -? . O - -Here RB B-ADVP -are VBP B-VP -four CD B-NP -key JJ I-NP -steps NNS I-NP -: : O - -1 CD B-LST -. . O -Make VB B-VP -sure JJ B-ADJP -you PRP B-NP -have VBP B-VP -a DT B-NP -strong JJ I-NP -personnel NNS I-NP -department NN I-NP -. . O - -It PRP B-NP -must MD B-VP -be VB I-VP -able JJ B-ADJP -to TO B-VP -handle VB I-VP -most RBS B-NP -of IN B-PP -the DT B-NP -complaints NNS I-NP -that WDT B-NP -can MD B-VP -not RB I-VP -be VB I-VP -solved VBN I-VP -in IN B-PP -the DT B-NP -trenches NNS I-NP -by IN B-PP -managers NNS B-NP -and CC O -their PRP$ B-NP -subordinates NNS I-NP -, , O -else RB B-ADVP -the DT B-NP -company NN I-NP -court NN I-NP -or CC I-NP -adjudicators NNS I-NP -will MD B-VP -be VB B-VP -inundated VBN I-VP -with IN B-PP -cases NNS B-NP -. . O - -At IN B-PP -Polaroid NNP B-NP -, , O -the DT B-NP -Personnel NNP I-NP -Policy NNP I-NP -Planning NNP I-NP -Committee NNP I-NP -may MD B-VP -hear VB I-VP -only RB B-NP -about IN I-NP -20 CD I-NP -cases VBZ I-NP -a DT B-NP -year NN I-NP -; : O -the DT B-NP -rest NN I-NP -of IN B-PP -the DT B-NP -many JJ I-NP -hundreds NNS I-NP -of IN B-PP -complaints NNS B-NP -are VBP B-VP -resolved VBN I-VP -at IN B-PP -earlier JJR B-NP -stages NNS I-NP -. . O - -At IN B-PP -TWA NNP B-NP -, , O -the DT B-NP -System NNP I-NP -Board NNP I-NP -of IN B-PP -Adjustment NNP B-NP -hears VBZ B-VP -50 CD B-NP -to TO I-NP -75 CD I-NP -cases VBZ I-NP -a DT B-NP -year NN I-NP -, , O -only RB B-NP -a DT I-NP -fraction NN I-NP -of IN B-PP -the DT B-NP -complaints NNS I-NP -brought VBN B-VP -to TO B-PP -personnel NNS B-NP -specialists NNS I-NP -. . O - -At IN B-PP -Citicorp NNP B-NP -, , O -the DT B-NP -Problem NNP I-NP -Review NNP I-NP -Board NNP I-NP -may MD B-VP -hear VB I-VP -only RB B-NP -12 CD I-NP -or CC I-NP -so RB I-NP -cases VBZ I-NP -because IN B-PP -of IN I-PP -personnel NNS B-NP -'s POS B-NP -skill NN I-NP -in IN B-PP -complaint-resolution NN B-NP -. . O - -In IN B-PP -a DT B-NP -typical JJ I-NP -year NN I-NP -, , O -up IN B-NP -to TO I-NP -20 CD I-NP -% NN I-NP -of IN B-PP -the DT B-NP -work NN I-NP -force NN I-NP -goes VBZ B-VP -to TO B-PP -personnel NNS B-NP -specialists NNS I-NP -with IN B-PP -complaints NNS B-NP -of IN B-PP -unfair JJ B-NP -treatment NN I-NP -. . O - -In IN B-PP -a DT B-NP -large JJ I-NP -company NN I-NP -that WDT B-NP -means VBZ B-VP -many JJ B-NP -hundreds NNS I-NP -of IN B-PP -complaints NNS B-NP -for IN B-PP -personnel NNS B-NP -to TO B-VP -handle VB I-VP -. . O - -2 CD B-LST -. . O -Formally RB B-ADVP -or CC I-ADVP -informally RB I-ADVP -, , O -train NN B-VP -all DT B-NP -your PRP$ I-NP -managers NNS I-NP -and CC I-NP -supervisors NNS I-NP -in IN B-PP -the DT B-NP -company NN I-NP -'s POS B-NP -due-process NN I-NP -approach NN I-NP -. . O - -See VB B-VP -that IN B-SBAR -they PRP B-NP -know VBP B-VP -company NN B-NP -personnel NNS I-NP -policy NN I-NP -backwards RB B-ADVP -and CC I-ADVP -forwards RB I-ADVP -, , O -for IN O -it PRP B-NP -is VBZ B-VP -the DT B-NP -`` `` I-NP -law NN I-NP -'' '' O -governing VBG B-VP -company NN B-NP -courts NNS I-NP -and CC I-NP -adjudicators NNS I-NP -. . O - -Coach NNP B-VP -them PRP B-NP -in IN B-PP -handling NN B-VP -complaints NNS B-NP -so RB B-SBAR -that IN I-SBAR -they PRP B-NP -can MD B-VP -resolve VB I-VP -problems NNS B-NP -immediately RB B-ADVP -. . O - -In IN B-SBAR -case NN O -managers NNS B-NP -and CC O -personnel NNS B-NP -specialists NNS I-NP -are VBP B-VP -unsuccessful JJ B-ADJP -and CC O -subordinates NNS B-NP -take VBP B-VP -their PRP$ B-NP -complaints NNS I-NP -to TO B-PP -a DT B-NP -company NN I-NP -court NN I-NP -or CC I-NP -adjudicator NN I-NP -, , O -teach VB B-VP -managers NNS B-NP -to TO B-VP -accept VB I-VP -reversals NNS B-NP -as IN B-PP -a DT B-NP -fact NN I-NP -of IN B-PP -business NN B-NP -life NN I-NP -, , O -for IN O -in IN B-PP -a DT B-NP -good JJ I-NP -due-process NN I-NP -system NN I-NP -they PRP B-NP -are VBP B-VP -bound VBN I-VP -to TO I-VP -happen VB I-VP -. . O - -In IN B-PP -the DT B-NP -15 CD I-NP -companies NNS I-NP -I PRP B-NP -studied VBD B-VP -, , O -reversal NN B-NP -rates NNS I-NP -range VBP B-VP -on IN B-PP -the DT B-NP -average NN I-NP -from IN B-PP -20 CD B-NP -% NN I-NP -to TO B-PP -40 CD B-NP -% NN I-NP -. . O - -3 CD B-LST -. . O -Decide VB B-VP -whether IN O -you PRP B-NP -want VBP B-VP -a DT B-NP -panel NN I-NP -system NN I-NP -or CC O -a DT B-NP -single JJ I-NP -adjudicator NN I-NP -. . O - -A DT B-NP -panel NN I-NP -system NN I-NP -like IN B-PP -that DT B-NP -in NN B-PP -the DT B-NP -Bob NNP I-NP -Stone NNP I-NP -example NN I-NP -enjoys VBZ B-VP -such JJ B-NP -advantages NNS I-NP -as IN B-PP -high JJ B-NP -credibility NN I-NP -and CC O -, , O -for IN B-PP -the DT B-NP -panelists NNS I-NP -, , O -mutual JJ B-NP -support NN I-NP -. . O - -An DT B-NP -adjudicator NN I-NP -system NN I-NP --- : O -that DT B-INTJ -is VBZ I-INTJ -, , O -an DT B-NP -investigator NN I-NP -who WP B-NP -acts VBZ B-VP -first JJ B-ADVP -as IN B-PP -a DT B-NP -fact-finder NN I-NP -and CC O -then RB O -switches VBZ B-VP -hats NNS B-NP -and CC O -arbitrates VBZ B-VP -the DT B-NP -facts NNS I-NP --- : O -has VBZ B-VP -such JJ B-NP -advantages NNS I-NP -as IN B-PP -speed NN B-NP -, , O -flexibility NN B-NP -and CC O -maximum JJ B-NP -privacy NN I-NP -. . O - -International NNP B-NP -Business NNP I-NP -Machines NNPS I-NP -and CC O -Bank NNP B-NP -of IN B-PP -America NNP B-NP -are VBP B-VP -among IN B-PP -the DT B-NP -companies NNS I-NP -using VBG B-VP -the DT B-NP -single-adjudicator JJ I-NP -approach NN I-NP -. . O - -4 CD B-LST -. . O -Make VB B-VP -your PRP$ B-NP -due-process NN I-NP -system NN I-NP -visible JJ B-ADJP -. . O - -It PRP B-NP -wo MD B-VP -n't RB I-VP -do VB I-VP -any DT B-NP -good NN I-NP -for IN B-PP -anybody NN B-NP -unless IN B-SBAR -employees NNS B-NP -know VBP B-VP -about IN B-PP -it PRP B-NP -. . O - -Most JJS B-NP -managements NNS I-NP -hesitate VBP B-VP -to TO I-VP -go VB I-VP -all DT B-ADVP -out NN I-ADVP -in IN B-PP -advertising VBG B-VP -their PRP$ B-NP -due-process NN I-NP -systems NNS I-NP -for IN B-PP -fear NN B-NP -of IN B-PP -encouraging VBG B-VP -cranks NNS B-NP -and CC O -chronic JJ B-NP -soreheads NNS I-NP -to TO B-VP -file VB I-VP -complaints NNS B-NP -. . O - -On IN B-PP -the DT B-NP -other JJ I-NP -hand NN I-NP -, , O -they PRP B-NP -make VBP B-VP -sure JJ B-ADJP -at IN B-PP -a DT B-NP -minimum NN I-NP -that IN B-SBAR -their PRP$ B-NP -systems NNS I-NP -are VBP B-VP -described VBN I-VP -in IN B-PP -their PRP$ B-NP -employee NN I-NP -handbooks NNS I-NP -and CC O -talked VBD B-VP -up IN B-PRT -by IN B-PP -personnel NNS B-NP -specialists NNS I-NP -. . O - -Smith-Kline NNP B-NP -Beecham NNP I-NP -goes VBZ B-VP -further JJ B-ADVP -and CC O -sometimes RB B-VP -features VBZ I-VP -its PRP$ B-NP -grievance NN I-NP -procedure NN I-NP -in IN B-PP -closed-circuit JJ B-NP -TV NN I-NP -programs NNS I-NP -. . O - -Naturally RB B-ADVP -, , O -one CD B-NP -of IN B-PP -the DT B-NP -best JJS I-NP -ways NNS I-NP -to TO B-VP -guarantee VB I-VP -visibility NN B-NP -for IN B-PP -your PRP$ B-NP -due-process NN I-NP -system NN I-NP -is VBZ B-VP -for IN B-SBAR -top JJ B-NP -management NN I-NP -to TO B-VP -support VB I-VP -it PRP B-NP -. . O - -At IN B-PP -IBM NNP B-NP -, , O -the DT B-NP -company NN I-NP -'s POS B-NP -Open NNP I-NP -Door NNP I-NP -system NN I-NP -is VBZ B-VP -sometimes RB B-ADVP -the DT B-NP -subject NN I-NP -of IN B-PP -memorandums NNS B-NP -from IN B-PP -the DT B-NP -chief JJ I-NP -executive NN I-NP -. . O - -Federal NNP B-NP -Express NNP I-NP -goes VBZ B-VP -further JJ B-ADVP -in IN B-PP -this DT B-NP -respect NN I-NP -than IN B-PP -any DT B-NP -company NN I-NP -I PRP B-NP -know VBP B-VP -of IN B-PP -with IN B-PP -both DT B-NP -Frederick NNP B-NP -Smith NNP I-NP -and CC O -James NNP B-NP -Barksdale NNP I-NP -, , O -chief JJ B-NP -executive NN I-NP -and CC O -chief JJ B-NP -operating VBG I-NP -officer NN I-NP -, , O -respectively RB B-ADVP -, , O -sitting VBG B-VP -in IN B-PRT -on IN B-PP -the DT B-NP -Appeals NNP I-NP -Board NNP I-NP -almost RB B-NP -every DT I-NP -Tuesday NNP I-NP -to TO B-VP -decide VB I-VP -cases NNS B-NP -. . O - -Mr. NNP B-NP -Ewing NNP I-NP -is VBZ B-VP -a DT B-NP -consultant NN I-NP -based VBN B-VP -in IN B-PP -Winchester NNP B-NP -, , O -Mass. NNP B-NP -, , O -and CC O -author NN B-NP -of IN B-PP -`` `` O -Justice NNP B-NP -on IN B-PP -the DT B-NP -Job NNP I-NP -: : O -Resolving NNP B-VP -Grievances NNP B-NP -in IN B-PP -the DT B-NP -Nonunion NNP I-NP -Workplace NN I-NP -'' '' O --LRB- ( O -Harvard NNP B-NP -Business NNP I-NP -School NNP I-NP -Press NNP I-NP -, , O -1989 CD B-NP --RRB- ) O -. . O - -Tokyo NNP B-NP -stocks NNS I-NP -closed VBD B-VP -higher JJR B-ADVP -in IN B-PP -active JJ B-NP -trading NN I-NP -Friday NNP B-NP -, , O -marking VBG B-VP -the DT B-NP -fourth JJ I-NP -consecutive JJ I-NP -daily JJ I-NP -gain NN I-NP -since IN B-PP -Monday NNP B-NP -'s POS B-NP -sharp JJ I-NP -fall NN I-NP -. . O - -London JJ B-NP -shares NNS I-NP -closed VBD B-VP -moderately RB B-ADVP -lower JJR I-ADVP -in IN B-PP -thin JJ B-NP -trading NN I-NP -. . O - -At IN B-PP -Tokyo NNP B-NP -, , O -the DT B-NP -Nikkei NNP I-NP -index NN I-NP -of IN B-PP -225 CD B-NP -selected VBN I-NP -issues NNS I-NP -was VBD B-VP -up IN B-ADVP -112.16 CD B-NP -points NNS I-NP -to TO B-PP -35486.38 CD B-NP -. . O - -The DT B-NP -index NN I-NP -advanced VBD B-VP -266.66 CD B-NP -points NNS I-NP -Thursday NNP B-NP -. . O - -In IN B-PP -early JJ B-NP -trading NN I-NP -in IN B-PP -Tokyo NNP B-NP -Monday NNP B-NP -, , O -the DT B-NP -Nikkei NNP I-NP -index NN I-NP -rose VBD B-VP -101.98 CD B-NP -points NNS I-NP -to TO B-PP -35588.36 CD B-NP -. . O - -Friday NNP B-NP -'s POS B-NP -volume NN I-NP -on IN B-PP -the DT B-NP -First NNP I-NP -Section NN I-NP -was VBD B-VP -estimated VBN I-VP -at IN B-PP -one CD B-NP -billion CD I-NP -shares NNS I-NP -, , O -up IN B-ADVP -from IN B-PP -862 CD B-NP -million CD I-NP -Thursday NNP B-NP -. . O - -Winners NNS B-NP -outpaced VBD B-VP -losers NNS B-NP -, , O -572 CD B-ADVP -to TO I-ADVP -368 CD I-ADVP -, , O -while IN B-SBAR -181 CD B-NP -issues NNS I-NP -remained VBD B-VP -unchanged JJ B-ADJP -. . O - -With IN B-SBAR -investors NNS B-NP -relieved VBN B-ADJP -at IN B-PP -the DT B-NP -overnight JJ I-NP -gain NN I-NP -in IN B-PP -New NNP B-NP -York NNP I-NP -stocks NNS I-NP -, , O -small-lot JJ B-NP -buying NN I-NP -orders NNS I-NP -streamed VBD B-VP -into IN B-PP -the DT B-NP -market NN I-NP -from IN B-PP -early JJ B-NP -morning NN I-NP -, , O -making VBG B-VP -traders NNS B-NP -believe VBP B-VP -the DT B-NP -market NN I-NP -was VBD B-VP -back RB B-ADVP -to TO B-PP -normal JJ B-NP -. . O - -The DT B-NP -Nikkei NNP I-NP -, , O -which WDT B-NP -reached VBD B-VP -as RB B-ADJP -high JJ I-ADJP -as IN B-PP -35611.38 CD B-NP -right NN B-ADVP -after IN B-PP -the DT B-NP -opening NN I-NP -, , O -surrendered VBD B-VP -part NN B-NP -of IN B-PP -its PRP$ B-NP -early JJ I-NP -advance NN I-NP -toward IN B-PP -the DT B-NP -end NN I-NP -of IN B-PP -the DT B-NP -day NN I-NP -because IN B-PP -of IN I-PP -profit-taking NN B-NP -. . O - -`` `` O -Investors NNS B-NP -, , B-NP -especially RB I-NP -dealers NNS B-NP -, , O -do VBP B-VP -n't RB I-VP -want VB I-VP -to TO I-VP -hold VB I-VP -a DT B-NP -position NN I-NP -over IN B-PP -the DT B-NP -weekend NN I-NP -, , O -'' '' O -a DT B-NP -trader NN I-NP -at IN B-PP -Dai-ichi NNP B-NP -Securities NNP I-NP -said VBD B-VP -, , O -adding VBG B-VP -, , O -though RB B-ADVP -, , O -that IN B-SBAR -the DT B-NP -trading NN I-NP -mood NN I-NP -remained VBD B-VP -positive JJ B-ADJP -through IN B-PP -the DT B-NP -afternoon NN I-NP -session NN I-NP -. . O - -The DT B-NP -Tokyo NNP I-NP -Stock NNP I-NP -Price NNP I-NP -Index NNP I-NP --LRB- ( O -Topix NNP B-NP --RRB- ) O -of IN B-PP -all DT B-NP -issues NNS I-NP -listed VBN B-VP -in IN B-PP -the DT B-NP -First NNP I-NP -Section NN I-NP -, , O -which WDT B-NP -gained VBD B-VP -22.78 CD B-NP -points NNS I-NP -Thursday NNP B-NP -, , O -was VBD B-VP -up IN B-ADVP -14.06 CD B-NP -points NNS I-NP -, , O -or CC O -0.53 CD B-NP -% NN I-NP -, , O -at IN B-PP -2679.72 CD B-NP -. . O - -The DT B-NP -Second JJ I-NP -Section NN I-NP -index NN I-NP -, , O -which WDT B-NP -rose VBD B-VP -15.72 CD B-NP -points NNS I-NP -Thursday NNP B-NP -, , O -was VBD B-VP -up IN B-ADVP -11.88 CD B-NP -points NNS I-NP -, , O -or CC O -0.32 CD B-NP -% NN I-NP -, , O -to TO B-VP -close VB I-VP -at IN B-PP -3717.46 CD B-NP -. . O - -Volume NN B-NP -in IN B-PP -the DT B-NP -second JJ I-NP -section NN I-NP -was VBD B-VP -estimated VBN I-VP -at IN B-PP -30 CD B-NP -million CD I-NP -shares NNS I-NP -, , O -up IN B-ADVP -from IN B-PP -28 CD B-NP -million CD I-NP -Thursday NNP B-NP -. . O - -In IN B-PP -turmoil NN B-NP -caused VBN B-VP -by IN B-PP -the DT O -previous JJ B-NP -Friday NNP I-NP -'s POS B-NP -plunge NN I-NP -in IN B-PP -New NNP B-NP -York NNP I-NP -stocks NNS I-NP -, , O -the DT B-NP -Nikkei NNP I-NP -marked VBD B-VP -a DT B-NP -sharp JJ I-NP -647.33-point JJ I-NP -fall NN I-NP -Monday NNP B-NP -. . O - -But CC O -the DT B-NP -Nikkei NNP I-NP -fell VBD B-VP -an DT B-NP -overall JJ I-NP -1.8 CD I-NP -% NN I-NP -in IN B-PP -value NN B-NP -that DT B-NP -day NN I-NP -compared VBN B-PP -with IN B-PP -Wall NNP B-NP -Street NNP I-NP -'s POS I-NP -far RB B-ADJP -sharper JJR I-ADJP -6.9 CD B-ADJP -% NN I-ADJP -drop NN B-NP -on IN B-PP -Oct. NNP B-NP -13 CD I-NP -. . O - -The DT B-NP -Tokyo NNP I-NP -market NN I-NP -'s POS B-NP -resiliency NN I-NP -helped VBD B-VP -participants NNS B-NP -to TO B-VP -regain VB I-VP -confidence NN B-NP -gradually RB B-ADVP -as IN B-SBAR -they PRP B-NP -spent VBD B-VP -more JJR B-NP -time NN I-NP -on IN B-PP -analyzing VBG B-VP -factors NNS B-NP -that WDT B-NP -caused VBD B-VP -the DT B-NP -Friday NNP I-NP -plunge NN I-NP -and CC O -realized VBD B-VP -these DT B-NP -problems NNS I-NP -were VBD B-VP -unique JJ B-ADJP -to TO B-PP -New NNP B-NP -York NNP I-NP -stocks NNS I-NP -and CC B-ADJP -not RB I-ADJP -directly RB B-ADJP -related VBN I-ADJP -to TO B-PP -Tokyo NNP B-NP -. . O - -The DT B-NP -Nikkei NNP I-NP -continued VBD B-VP -to TO I-VP -gain VB I-VP -for IN B-PP -the DT B-NP -rest NN I-NP -of IN B-PP -the DT B-NP -week NN I-NP -, , O -adding VBG B-VP -1017.69 CD B-NP -points NNS I-NP -in IN B-PP -four CD B-NP -days NNS I-NP --- : O -more JJR B-VP -than IN I-VP -erasing VBG I-VP -Monday NNP B-NP -'s POS B-NP -losses NNS I-NP -. . O - -But CC O -further JJ B-NP -major JJ I-NP -advances NNS I-NP -on IN B-PP -the DT B-NP -Nikkei NNP I-NP -are VBP B-VP -n't RB I-VP -foreseen VBN I-VP -this DT B-NP -week NN I-NP -by IN B-PP -market NN B-NP -observers NNS I-NP -. . O - -Investors NNS B-NP -are VBP B-VP -still RB I-VP -waiting VBG I-VP -to TO I-VP -see VB I-VP -how WRB B-ADVP -the DT B-NP -U.S. NNP I-NP -government NN I-NP -will MD B-VP -decide VB I-VP -on IN B-PP -interest NN B-NP -rates NNS I-NP -and CC O -how WRB B-ADVP -the DT B-NP -dollar NN I-NP -will MD B-VP -be VB I-VP -stabilized VBN I-VP -. . O - -Some DT B-NP -high-priced JJ I-NP -issues NNS I-NP -made VBD B-VP -a DT B-NP -comeback NN I-NP -Friday NNP B-NP -. . O - -Pioneer NNP B-NP -surged VBD B-VP -450 CD B-NP -yen NN I-NP --LRB- ( O -$ $ B-NP -3.16 CD I-NP --RRB- ) O -to TO B-PP -6,050 CD B-NP -yen NN I-NP --LRB- ( O -$ $ B-NP -42.60 CD I-NP --RRB- ) O -. . O - -Kyocera NNP B-NP -advanced VBD B-VP -80 CD B-NP -yen NN I-NP -to TO B-PP -5,440 CD B-NP -. . O - -Fanuc NNP B-NP -gained VBD B-VP -100 CD B-NP -to TO B-PP -7,580 CD B-NP -. . O - -Breweries NNP B-NP -attracted VBD B-VP -investors NNS B-NP -because IN B-PP -of IN I-PP -their PRP$ B-NP -land NN I-NP -property NN I-NP -holdings NNS I-NP -that WDT B-NP -could MD B-VP -figure VB I-VP -in IN B-PP -development NN B-NP -or CC O -other JJ B-NP -plans NNS I-NP -, , O -traders NNS B-NP -said VBD B-VP -. . O - -Sapporo NNP B-NP -gained VBD B-VP -80 CD B-NP -to TO B-PP -1,920 CD B-NP -and CC O -Kirin NNP B-NP -added VBD B-VP -60 CD B-NP -to TO B-PP -2,070 CD B-NP -. . O - -Housings NNS B-NP -, , I-NP -constructions NNS I-NP -and CC I-NP -pharmaceuticals NNS I-NP -continued VBD B-VP -to TO I-VP -be VB I-VP -bought VBN I-VP -following VBG B-PP -Thursday NNP B-NP -'s POS B-NP -gains NNS I-NP -because IN B-PP -of IN I-PP -strong JJ B-NP -earnings NNS I-NP -outlooks NNS I-NP -. . O - -Daiwa NNP B-NP -House NNP I-NP -gained VBD B-VP -50 CD B-NP -to TO B-PP -2,660 CD B-NP -. . O - -Misawa NNP B-NP -Homes NNP I-NP -was VBD B-VP -up IN B-ADVP -20 CD B-NP -at IN B-PP -2,960 CD B-NP -. . O - -Kajima NNP B-NP -advanced VBD B-VP -40 CD B-NP -to TO B-PP -2,120 CD B-NP -and CC O -Ohbayashi NNP B-NP -added VBD B-VP -50 CD B-NP -to TO B-PP -1,730 CD B-NP -. . O - -Fujisawa NNP B-NP -added VBD B-VP -80 CD B-NP -to TO B-PP -2,010 CD B-NP -and CC O -Mochida NNP B-NP -advanced VBD B-VP -230 CD B-NP -to TO B-PP -4,400 CD B-NP -. . O - -London JJ B-NP -share NN I-NP -prices NNS I-NP -were VBD B-VP -influenced VBN I-VP -largely RB B-ADVP -by IN B-PP -declines NNS B-NP -on IN B-PP -Wall NNP B-NP -Street NNP I-NP -and CC O -weakness NN B-NP -in IN B-PP -the DT B-NP -British JJ I-NP -pound NN I-NP -. . O - -The DT B-NP -key JJ I-NP -Financial NNP I-NP -Times-Stock NNP I-NP -Exchange NNP I-NP -100-share JJ I-NP -index NN I-NP -ended VBD B-VP -10.2 CD B-NP -points NNS I-NP -lower JJR B-ADVP -at IN B-PP -2179.1 CD B-NP -, , O -above IN B-ADVP -its PRP$ B-NP -intraday JJ I-NP -low NN I-NP -of IN B-PP -2176.9 CD B-NP -, , B-ADVP -but CC I-ADVP -off IN B-ADVP -the DT B-NP -day NN I-NP -'s POS I-NP -high NN B-NP -of IN B-PP -2189 CD B-NP -. . O - -The DT B-NP -index NN I-NP -finished VBD B-VP -2.4 CD B-NP -% NN I-NP -under IN B-PP -its PRP$ B-NP -close NN I-NP -of IN B-PP -2233.9 CD B-NP -the DT B-NP -previous JJ I-NP -Friday NNP I-NP -, , O -although IN B-SBAR -it PRP B-NP -recouped VBD B-VP -some DT B-NP -of IN B-PP -the DT B-NP -sharp JJ I-NP -losses NNS I-NP -staged VBD B-VP -early JJ B-NP -last JJ I-NP -week NN I-NP -on IN B-PP -the DT B-NP -back RB I-NP -of IN B-PP -Wall NNP B-NP -Street NNP I-NP -'s POS B-NP -fall NN I-NP -. . O - -London NNP B-NP -was VBD B-VP -weak JJ B-ADJP -throughout IN B-PP -Friday NNP B-NP -'s POS B-NP -trading NN I-NP -, , O -however RB B-ADVP -, , O -on IN B-PP -what WP B-NP -dealers NNS B-NP -attributed VBD B-VP -to TO B-PP -generally RB B-NP -thin JJ I-NP -interest NN I-NP -ahead RB B-ADVP -of IN B-PP -the DT B-NP -weekend NN I-NP -and CC O -this DT B-NP -week NN I-NP -'s POS I-NP -potentially RB B-ADJP -important JJ I-ADJP -U.K. NNP B-NP -trade NN I-NP -figures NNS I-NP -for IN B-PP -September NNP B-NP -. . O - -The DT B-NP -FT-SE NNP I-NP -100 CD I-NP -largely RB B-ADVP -remained VBD B-VP -within IN B-PP -an DT B-NP -11-point JJ I-NP -range NN I-NP -establshed VBN B-VP -within IN B-PP -the DT B-NP -first JJ I-NP -hour NN I-NP -of IN B-PP -trading NN B-NP -before IN B-PP -it PRP B-NP -eased VBD B-VP -to TO B-PP -an DT B-NP -intraday JJ I-NP -low JJ I-NP -late RB B-ADVP -in IN B-PP -the DT B-NP -session NN I-NP -when WRB B-ADVP -a DT B-NP -flurry NN I-NP -of IN B-PP -program NN B-NP -selling VBG I-NP -pushed VBN B-VP -Wall NNP B-NP -Street NNP I-NP -lower JJR B-ADVP -. . O - -The DT B-NP -FT NNP I-NP -30-share JJ I-NP -index NN I-NP -closed VBD B-VP -11.0 CD B-NP -points NNS I-NP -lower JJR B-ADVP -at IN B-PP -1761.0 CD B-NP -. . O - -Volume NN B-NP -was VBD B-VP -extremely RB B-ADJP -thin JJ I-ADJP -at IN B-PP -351.3 CD B-NP -million CD I-NP -shares NNS I-NP -, , O -the DT B-NP -lightest JJS I-NP -volume NN I-NP -of IN B-PP -the DT B-NP -week NN I-NP -and CC O -modestly RB B-ADVP -under IN B-PP -Thursday NNP B-NP -'s POS B-NP -387.4 CD I-NP -million CD I-NP -shares NNS I-NP -. . O - -Dealers NNS B-NP -said VBD B-VP -the DT B-NP -day NN I-NP -'s POS B-NP -action NN I-NP -was VBD B-VP -featureless JJ B-ADJP -outside IN B-PP -some DT B-NP -response NN I-NP -to TO B-PP -sterling NN B-NP -'s POS B-NP -early JJ I-NP -weakness NN I-NP -against IN B-PP -the DT B-NP -mark NN I-NP -, , O -and CC O -fears NNS B-NP -that IN B-SBAR -Wall NNP B-NP -Street NNP I-NP -might MD B-VP -open RB I-VP -lower JJR B-ADVP -after IN B-PP -its PRP$ B-NP -strong JJ I-NP -leap NN I-NP -forward RB B-ADVP -Thursday NNP B-NP -. . O - -They PRP B-NP -added VBD B-VP -that IN B-SBAR -market-makers NNS B-NP -were VBD B-VP -largely RB I-VP -sidelined VBN I-VP -after IN B-PP -aggressively RB B-VP -supporting VBG I-VP -the DT B-NP -market NN I-NP -Thursday NNP B-NP -in IN B-PP -their PRP$ B-NP -quest NN I-NP -to TO B-VP -cover VB I-VP -internal JJ B-NP -shortages NNS I-NP -of IN B-PP -FT-SE NNP B-NP -100 CD I-NP -shares NNS I-NP -. . O - -Interest NN B-NP -may MD B-VP -remain VB I-VP -limited JJ B-ADJP -into IN B-PP -tomorrow NN B-NP -'s POS B-NP -U.K. NNP I-NP -trade NN I-NP -figures NNS I-NP -, , O -which WDT B-NP -the DT B-NP -market NN I-NP -will MD B-VP -be VB I-VP -watching VBG I-VP -closely RB B-ADVP -to TO B-VP -see VB I-VP -if IN B-SBAR -there EX B-NP -is VBZ B-VP -any DT B-NP -improvement NN I-NP -after IN B-PP -disappointing JJ B-NP -numbers NNS I-NP -in IN B-PP -the DT B-NP -previous JJ I-NP -two CD I-NP -months NNS I-NP -. . O - -The DT B-NP -key JJ I-NP -corporate JJ I-NP -news NN I-NP -of IN B-PP -the DT B-NP -day NN I-NP -was VBD B-VP -that IN B-SBAR -British JJ B-NP -Airways NNPS I-NP -decided VBD B-VP -to TO I-VP -withdraw VB I-VP -from IN B-PP -a DT B-NP -management-led JJ I-NP -bid NN I-NP -for IN B-PP -UAL NNP B-NP -Corp. NNP I-NP -, , O -the DT B-NP -parent NN I-NP -of IN B-PP -United NNP B-NP -Airlines NNPS I-NP -. . O - -British JJ B-NP -Airways NNPS I-NP -rose VBD B-VP -initially RB B-ADVP -after IN B-PP -announcing VBG B-VP -its PRP$ B-NP -withdrawal NN I-NP -from IN B-PP -the DT B-NP -UAL NNP I-NP -deal NN I-NP -. . O - -Dealers NNS B-NP -said VBD B-VP -they PRP B-NP -viewed VBD B-VP -the DT O -initial JJ O -# # O -390-million CD O --LRB- ( O -$ $ B-ADJP -622 CD O -million CD O --RRB- ) O -outlay NN B-NP -for IN B-PP -a DT B-NP -15 CD I-NP -% NN I-NP -stake NN I-NP -in IN B-PP -the DT B-NP -airline NN I-NP -as IN B-PP -a DT B-NP -bit NN I-NP -much JJ I-NP -. . O - -Its PRP$ B-NP -shares NNS I-NP -slid VBD B-VP -in IN B-PP -late JJ B-NP -dealings NNS I-NP -to TO B-VP -close VB I-VP -a DT B-NP -penny NN I-NP -per IN B-PP -share NN B-NP -lower JJR B-ADVP -at IN B-PP -197 CD B-NP -pence NN I-NP -. . O - -The DT B-NP -airline NN I-NP -was VBD B-VP -the DT B-NP -most RBS I-NP -active JJ I-NP -FT-SE NNP I-NP -100 CD I-NP -at IN B-PP -8.2 CD B-NP -million CD I-NP -shares NNS I-NP -traded VBN B-VP -. . O - -The DT B-NP -next JJ I-NP -most RBS I-NP -active JJ I-NP -top-tier JJ I-NP -stock NN I-NP -was VBD B-VP -B.A.T NNP B-NP -Industries NNPS I-NP -, , O -the DT B-NP -target NN I-NP -of IN B-PP -Sir NNP B-NP -James NNP I-NP -Goldsmith NNP I-NP -'s POS B-NP -# # B-ADJP -13.4 CD O -billion CD O -bid NN B-NP -. . O - -The DT B-NP -company NN I-NP -gained VBD B-VP -shareholder NN B-NP -approval NN I-NP -Thursday NNP B-NP -to TO B-VP -restructure VB I-VP -in IN B-PP -a DT B-NP -bid NN I-NP -to TO B-VP -fend VB I-VP -off IN B-PRT -the DT B-NP -hostile JJ I-NP -takeover NN I-NP -. . O - -Sir NNP B-NP -James NNP I-NP -said VBD B-VP -Thursday NNP B-NP -night NN I-NP -that IN B-SBAR -his PRP$ B-NP -plans NNS I-NP -for IN B-PP -the DT B-NP -takeover NN I-NP -had VBD B-VP -n't RB I-VP -changed VBN I-VP -. . O - -B.A.T NNP B-NP -ended VBD B-VP -the DT B-NP -day NN I-NP -at IN B-PP -778 CD B-NP -, , O -down JJ B-ADVP -5 NN B-NP -, , O -on IN B-PP -turnover NN B-NP -of IN B-PP -7.5 CD B-NP -million CD I-NP -shares NNS I-NP -. . O - -Dealers NNS B-NP -said VBD B-VP -it PRP B-NP -was VBD B-VP -hit VBN I-VP -by IN B-PP -some DT B-NP -profit-taking NN I-NP -after IN B-PP -gains NNS B-NP -since IN B-PP -mid-week NN B-NP -. . O - -In IN B-PP -other JJ B-NP -active JJ I-NP -shares NNS I-NP -, , O -Trusthouse NNP B-NP -Forte NNP I-NP -shed VB B-VP -10 CD B-NP -to TO B-PP -294 CD B-NP -on IN B-PP -volume NN B-NP -of IN B-PP -6.4 CD B-NP -million CD I-NP -shares NNS I-NP -after IN B-PP -a DT B-NP -Barclays NNP I-NP -De NNP I-NP -Zoete NNP I-NP -Wedd NNP I-NP -downgrading NN I-NP -, , O -while IN B-SBAR -Hillsdown NNP B-NP -Holdings NNP I-NP -, , O -a DT B-NP -food NN I-NP -products NNS I-NP -concern VBP I-NP -, , O -was VBD B-VP -boosted VBN I-VP -2 CD B-NP -to TO B-PP -271 CD B-NP -after IN O -it PRP B-NP -disclosed VBD B-VP -it PRP B-NP -would MD B-VP -seek VB I-VP -shareholder NN B-NP -approval NN I-NP -to TO B-VP -begin VB I-VP -share NN B-NP -repurchases NNS I-NP -. . O - -Elsewhere RB B-ADVP -in IN B-PP -Europe NNP B-NP -, , O -share NN B-NP -prices NNS I-NP -closed VBD B-VP -higher JJR B-ADVP -in IN B-PP -Stockholm NNP B-NP -, , I-NP -Brussels NNP I-NP -and CC I-NP -Milan NNP I-NP -. . O - -Prices NNS B-NP -were VBD B-VP -lower JJR B-ADJP -in IN B-PP -Frankfurt NNP B-NP -, , I-NP -Zurich NNP I-NP -, , I-NP -Paris NNP I-NP -and CC I-NP -Amsterdam NNP I-NP -. . O - -South JJ B-NP -African JJ I-NP -gold NN I-NP -stocks NNS I-NP -closed VBD B-VP -moderately RB B-ADVP -lower JJR I-ADVP -. . O - -Share NN B-NP -prices NNS I-NP -closed VBD B-VP -higher JJR B-ADVP -in IN B-PP -Sydney NNP B-NP -, , O -Taipei NNP B-NP -, , O -Wellington NNP B-NP -, , O -Manila NNP B-NP -, , O -Hong NNP B-NP -Kong NNP I-NP -and CC O -Singapore NNP B-NP -and CC O -were VBD B-VP -lower JJR B-ADJP -in IN B-PP -Seoul NNP B-NP -. . O - -Here RB B-ADVP -are VBP B-VP -price NN B-NP -trends NNS I-NP -on IN B-PP -the DT B-NP -world NN I-NP -'s POS B-NP -major JJ I-NP -stock NN I-NP -markets NNS I-NP -, , O -as IN B-SBAR -calculated VBN B-VP -by IN B-PP -Morgan NNP B-NP -Stanley NNP I-NP -Capital NNP I-NP -International NNP I-NP -Perspective NNP I-NP -, , O -Geneva NNP B-NP -. . O - -To TO B-VP -make VB I-VP -them PRP B-NP -directly RB B-ADJP -comparable JJ I-ADJP -, , O -each DT B-NP -index NN I-NP -is VBZ B-VP -based VBN I-VP -on IN B-PP -the DT B-NP -close NN I-NP -of IN B-PP -1969 CD B-NP -equaling VBG B-VP -100 CD B-NP -. . O - -The DT B-NP -percentage NN I-NP -change NN I-NP -is VBZ B-VP -since IN B-PP -year-end NN B-NP -. . O - -The DT B-NP -U.S. NNP I-NP -is VBZ B-VP -required VBN I-VP -to TO I-VP -notify VB I-VP -foreign JJ B-NP -dictators NNS I-NP -if IN B-SBAR -it PRP B-NP -knows VBZ B-VP -of IN B-PP -coup NN B-NP -plans NNS I-NP -likely JJ B-ADJP -to TO B-VP -endanger VB I-VP -their PRP$ B-NP -lives NNS I-NP -, , O -government NN B-NP -officials NNS I-NP -said VBD B-VP -. . O - -The DT B-NP -notification NN I-NP -policy NN I-NP -was VBD B-VP -part NN B-NP -of IN B-PP -a DT B-NP -set NN I-NP -of IN B-PP -guidelines NNS B-NP -on IN B-PP -handling NN B-VP -coups NNS B-NP -outlined VBN B-VP -in IN B-PP -a DT B-NP -secret JJ I-NP -1988 CD I-NP -exchange NN I-NP -of IN B-PP -letters NNS B-NP -between IN B-PP -the DT B-NP -Reagan NNP I-NP -administration NN I-NP -and CC O -the DT B-NP -Senate NNP I-NP -Intelligence NNP I-NP -Committee NNP I-NP -. . O - -The DT B-NP -existence NN I-NP -of IN B-PP -the DT B-NP -guidelines NNS I-NP -has VBZ B-VP -become VBN I-VP -known VBN I-VP -since IN B-SBAR -President NNP B-NP -Bush NNP I-NP -disclosed VBD B-VP -them PRP B-NP -privately RB B-ADVP -to TO B-PP -seven CD B-NP -Republican NNP I-NP -senators NNS I-NP -at IN B-PP -a DT B-NP -White NNP I-NP -House NNP I-NP -meeting NN I-NP -last JJ B-NP -Monday NNP I-NP -. . O - -Officials NNS B-NP -familiar JJ B-ADJP -with IN B-PP -the DT B-NP -meeting NN I-NP -said VBD B-VP -Mr. NNP B-NP -Bush NNP I-NP -cited VBD B-VP -the DT B-NP -policy NN I-NP -as IN B-PP -an DT B-NP -example NN I-NP -of IN B-PP -the DT B-NP -sort NN I-NP -of IN B-PP -congressional JJ B-NP -requirements NNS I-NP -the DT B-NP -administration NN I-NP -contends VBZ B-VP -contribute VB B-VP -to TO B-PP -the DT B-NP -failure NN I-NP -of IN B-PP -such JJ B-NP -covert JJ I-NP -actions NNS I-NP -as IN B-PP -this DT B-NP -month NN I-NP -'s POS B-NP -futile JJ I-NP -effort NN I-NP -to TO B-VP -oust VB I-VP -Panamanian JJ B-NP -dictator NN I-NP -Manuel NNP I-NP -Noriega NNP I-NP -. . O - -According VBG B-PP -to TO B-PP -the DT B-NP -officials NNS I-NP -, , O -Mr. NNP B-NP -Bush NNP I-NP -even RB B-ADVP -read VB B-VP -to TO B-PP -the DT B-NP -senators NNS I-NP -selections NNS B-NP -from IN B-PP -a DT B-NP -highly RB I-NP -classified VBN I-NP -letter NN I-NP -from IN B-PP -the DT B-NP -committee NN I-NP -to TO B-PP -the DT B-NP -White NNP I-NP -House NNP I-NP -discussing VBG B-VP -the DT B-NP -guidelines NNS I-NP -. . O - -They PRP B-NP -said VBD B-VP -the DT B-NP -president NN I-NP -conceded VBD B-VP -the DT B-NP -notification NN I-NP -requirement NN I-NP -did VBD B-VP -n't RB I-VP -affect VB I-VP -his PRP$ B-NP -decision NN I-NP -to TO B-VP -lend VB I-VP -only RB B-NP -minor JJ I-NP -support NN I-NP -to TO B-PP -this DT B-NP -month NN I-NP -'s POS B-NP -Panama NNP I-NP -coup NN I-NP -effort NN I-NP -. . O - -No DT B-NP -notification NN I-NP -was VBD B-VP -ever RB I-VP -considered VBN I-VP -, , O -officials NNS B-NP -said VBD B-VP -, , O -apparently RB B-ADVP -because IN B-SBAR -the DT B-NP -U.S. NNP I-NP -did VBD B-VP -n't RB I-VP -think VB I-VP -the DT B-NP -coup NN I-NP -plotters NNS I-NP -intended VBN B-VP -to TO I-VP -kill VB I-VP -Mr. NNP B-NP -Noriega NNP I-NP -, , O -but CC O -merely RB B-VP -sought VBD I-VP -to TO I-VP -imprison VB I-VP -him PRP B-NP -. . O - -What WP B-NP -'s VBZ B-VP -more JJR B-NP -, , O -both DT B-NP -administration NN B-NP -and CC O -congressional JJ B-NP -officials NNS I-NP -hint VBP B-VP -that IN B-SBAR -the DT B-NP -notification NN I-NP -requirement NN I-NP -is VBZ B-VP -likely JJ B-ADJP -to TO B-VP -be VB I-VP -dropped VBN I-VP -from IN B-PP -the DT B-NP -guidelines NNS I-NP -on IN B-PP -coup NN B-NP -attempts NNS I-NP -that WDT B-NP -are VBP B-VP -being VBG I-VP -rewritten VBN I-VP -by IN B-PP -the DT B-NP -panel NN I-NP -and CC O -the DT B-NP -White NNP I-NP -House NNP I-NP -. . O - -The DT B-NP -rewriting VBG I-NP -was VBD B-VP -launched VBN I-VP -at IN B-PP -a DT B-NP -meeting NN I-NP -between IN B-PP -Mr. NNP B-NP -Bush NNP I-NP -and CC O -intelligence NN B-NP -committee NN I-NP -leaders NNS I-NP -Oct. NNP B-NP -12 CD I-NP -, , O -a DT B-NP -few JJ I-NP -days NNS I-NP -before IN B-PP -the DT B-NP -meeting NN I-NP -at IN B-PP -which WDT B-NP -the DT B-NP -president NN I-NP -complained VBD B-VP -about IN B-PP -the DT B-NP -rules NNS I-NP -. . O - -However RB B-ADVP -, , O -the DT B-NP -disclosure NN I-NP -of IN B-PP diff --git a/paddle/trainer/tests/train_files.txt b/paddle/trainer/tests/train_files.txt deleted file mode 100644 index 1c268914953ff090ae47c56051fcf1cad0e1707b..0000000000000000000000000000000000000000 --- a/paddle/trainer/tests/train_files.txt +++ /dev/null @@ -1 +0,0 @@ -trainer/tests/train_proto.bin diff --git a/paddle/trainer/tests/train_sparse.list b/paddle/trainer/tests/train_sparse.list deleted file mode 100644 index 6ea020e2202f8464f8a647cd96c84a9d17a03ae3..0000000000000000000000000000000000000000 --- a/paddle/trainer/tests/train_sparse.list +++ /dev/null @@ -1 +0,0 @@ -trainer/tests/compare_sparse_data diff --git a/paddle/utils/Flags.cpp b/paddle/utils/Flags.cpp index 8f100f02e90bcbc7fdcf6f053aec6f95cfb09c1a..9a7dc0e35622383a190f8b3a80736e6b42c9c959 100644 --- a/paddle/utils/Flags.cpp +++ b/paddle/utils/Flags.cpp @@ -20,7 +20,7 @@ DEFINE_bool(use_gpu, false, "Only support CPU training"); DEFINE_bool(use_gpu, true, "Whether to use GPU for training"); #endif -#ifdef PADDLE_USE_MKLDNN +#ifdef PADDLE_WITH_MKLDNN // TODO(TJ): change to true when MKLDNN layers support multi-inputs DEFINE_bool(use_mkldnn, false, "Default still keep use CPU training"); #else diff --git a/proto/ModelConfig.proto b/proto/ModelConfig.proto index 2c2cc6245932d4af56a68d6399ce31f008bf3748..1fbdd5bbd82a0ae15f620fa11a14ac3126e52838 100644 --- a/proto/ModelConfig.proto +++ b/proto/ModelConfig.proto @@ -139,6 +139,8 @@ message PoolConfig { optional uint32 output_z = 16 [ default = 1 ]; optional uint32 img_size_z = 17 [ default = 1 ]; optional uint32 padding_z = 18 [ default = 1 ]; + + optional bool exclude_mode = 19; } message SppConfig { @@ -540,6 +542,13 @@ message LayerConfig { // for switch order layer optional ReshapeConfig reshape_conf = 59; + + // for batch normalization layer + // The small constant added to the variance to improve numeric stability. + optional double epsilon = 60 [ default = 0.00001 ]; + + // for factorization machine layer + optional uint32 factor_size = 61; } message EvaluatorConfig { diff --git a/python/.gitignore b/python/.gitignore index cc7d0ece4acaba2a3fa38a89110587fe8dffb992..1ba1d4c9b0301ed920f5303089e75dd3a8e4e3fa 100644 --- a/python/.gitignore +++ b/python/.gitignore @@ -2,6 +2,7 @@ build dist paddle.egg-info +paddlepaddle_gpu.egg-info .idea paddle/proto/*.py paddle/proto/*.pyc diff --git a/python/CMakeLists.txt b/python/CMakeLists.txt index c8632295a25b160513a8e154bf1a5453c0005031..6f589e916979584897863db72924c885c258d4b2 100644 --- a/python/CMakeLists.txt +++ b/python/CMakeLists.txt @@ -33,6 +33,12 @@ if(WITH_MKLDNN) list(APPEND MKL_DEPENDS mkldnn) endif() +if(WITH_GPU) + SET(PACKAGE_NAME "paddlepaddle-gpu") +else() + SET(PACKAGE_NAME "paddlepaddle") +endif() + configure_file(${CMAKE_CURRENT_SOURCE_DIR}/setup.py.in ${CMAKE_CURRENT_BINARY_DIR}/setup.py) diff --git a/python/paddle/__init__.py b/python/paddle/__init__.py index f662d6826321eb840739382558f76327d27b5847..1030c94e16376c326cb8b32926b8c47625cd38f0 100644 --- a/python/paddle/__init__.py +++ b/python/paddle/__init__.py @@ -11,3 +11,11 @@ # 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. +try: + from version import full_version as __version__ + from version import commit as __git_commit__ +except ImportError: + import sys + sys.stderr.write('''Warning with import paddle: you should not + import paddle from the source directory; please install paddlepaddle*.whl firstly.''' + ) diff --git a/python/paddle/trainer/config_parser.py b/python/paddle/trainer/config_parser.py index 0b523ac7e0bf5231398778ea69270c883ac112d2..239fe4204b20a37a0869ba1e0e99adf4293dac7e 100644 --- a/python/paddle/trainer/config_parser.py +++ b/python/paddle/trainer/config_parser.py @@ -1116,35 +1116,6 @@ def PyData(files=None, return data_config -@config_func -def ProtoData(files=None, - type=None, - file_group_queue_capacity=None, - load_file_count=None, - constant_slots=None, - load_thread_num=None, - **xargs): - data_config = create_data_config_proto(**xargs) - if type is None: - data_config.type = 'proto' - else: - data_config.type = type - data_config.files = files - - # When type="proto_group", one data provider contains at most - # load_file_count files, and there are at most - # (queue_capacity + load_thread_num + 1) data providers in memory - if file_group_queue_capacity is not None: - data_config.file_group_conf.queue_capacity = file_group_queue_capacity - if load_file_count is not None: - data_config.file_group_conf.load_file_count = load_file_count - if load_thread_num is not None: - data_config.file_group_conf.load_thread_num = load_thread_num - if constant_slots: - data_config.constant_slots.extend(constant_slots) - return data_config - - #real data for training is actually provided by "sub_data" data providers. @config_func def MultiData(sub_data=[]): @@ -1262,7 +1233,7 @@ def parse_bilinear(bilinear, input_layer_name, bilinear_conf): bilinear_conf.out_size_y = bilinear.out_size_y -def parse_pool(pool, input_layer_name, pool_conf, ceil_mode): +def parse_pool(pool, input_layer_name, pool_conf, ceil_mode, exclude_mode): pool_conf.pool_type = pool.pool_type config_assert(pool.pool_type in [ 'max-projection', 'avg-projection', 'max-pool-with-mask', 'cudnn-max-pool', 'cudnn-avg-pool' @@ -1291,6 +1262,8 @@ def parse_pool(pool, input_layer_name, pool_conf, ceil_mode): pool_conf.output_y = cnn_output_size(pool_conf.img_size_y, pool_conf.size_y, pool_conf.padding_y, pool_conf.stride_y, not ceil_mode) + if exclude_mode != None: + pool_conf.exclude_mode = exclude_mode def parse_pool3d(pool, input_layer_name, pool_conf, ceil_mode): @@ -2066,13 +2039,20 @@ class ParameterReluLayer(LayerBase): def __init__(self, name, inputs, partial_sum=1, **args): super(ParameterReluLayer, self).__init__( name, self.layer_type, 0, inputs=inputs, **args) + input_layer = self.get_input_layer(0) config_assert(len(self.inputs) == 1, "prelu layer has only one input.") config_assert(input_layer.size % partial_sum == 0, "a wrong setting for partial_sum") + + dims = [1, input_layer.size / partial_sum] self.set_layer_size(input_layer.size) self.config.partial_sum = partial_sum - self.create_input_parameter(0, input_layer.size / partial_sum) + self.create_input_parameter(0, input_layer.size / partial_sum, dims) + + self.set_layer_height_width(self.get_input_layer(0).height, \ + self.get_input_layer(0).width) + self.set_layer_depth(self.get_input_layer(0).depth) @config_layer('conv') @@ -2309,11 +2289,17 @@ class Conv3DLayer(Conv3DLayerBase): class NormLayer(LayerBase): def __init__(self, name, inputs, **xargs): super(NormLayer, self).__init__(name, 'norm', 0, inputs=inputs, **xargs) + use_mkldnn = bool(int(g_command_config_args.get("use_mkldnn", 0))) + use_mkldnn = True if use_mkldnn and self.inputs[ + 0].norm.norm_type == 'cmrnorm-projection' else False + self.config.type = 'mkldnn_lrn' if use_mkldnn else self.config.type for input_index in xrange(len(self.inputs)): input_layer = self.get_input_layer(input_index) norm_conf = self.config.inputs[input_index].norm_conf parse_norm(self.inputs[input_index].norm, input_layer.name, norm_conf) + norm_conf.scale = self.inputs[ + input_index].norm.scale if use_mkldnn else norm_conf.scale self.set_cnn_layer(name, norm_conf.output_y, norm_conf.output_x, norm_conf.channels, False) if norm_conf.norm_type == "cross-channel-norm": @@ -2325,7 +2311,8 @@ class NormLayer(LayerBase): class PoolLayer(LayerBase): layer_type = 'pool' - def __init__(self, name, inputs, ceil_mode=True, **xargs): + def __init__(self, name, inputs, ceil_mode=True, exclude_mode=None, + **xargs): use_mkldnn = int(g_command_config_args.get("use_mkldnn", 0)) if self.layer_type == "mkldnn_pool": config_assert(use_mkldnn, "mkldnn_pool only support MKLDNN") @@ -2336,7 +2323,7 @@ class PoolLayer(LayerBase): input_layer = self.get_input_layer(input_index) pool_conf = self.config.inputs[input_index].pool_conf parse_pool(self.inputs[input_index].pool, input_layer.name, - pool_conf, ceil_mode) + pool_conf, ceil_mode, exclude_mode) self.set_cnn_layer(name, pool_conf.output_y, pool_conf.output_x, pool_conf.channels) @@ -2422,6 +2409,14 @@ class CropLayer(LayerBase): image_conf.img_size_y = input_layer.height image_conf.channels = input_layer.size / (input_layer.width * input_layer.height) + # only support for 4-dims inputs and NCHW order + if (len(self.config.inputs) == 2): + self.set_layer_height_width( + self.get_input_layer(1).height, self.get_input_layer(1).width) + self.set_layer_size(self.get_input_layer(1).size) + else: + self.set_layer_height_width(shape[-2], shape[-1]) + self.set_layer_size(reduce(lambda x, y: x * y, shape[1:])) @config_layer('batch_norm') @@ -2434,6 +2429,7 @@ class BatchNormLayer(LayerBase): bias=True, img3D=False, use_global_stats=True, + epsilon=1e-5, moving_average_fraction=0.9, batch_norm_type=None, mean_var_names=None, @@ -2482,6 +2478,9 @@ class BatchNormLayer(LayerBase): self.config.use_global_stats = use_global_stats if moving_average_fraction is not None: self.config.moving_average_fraction = moving_average_fraction + if epsilon is not None: + assert epsilon >= 1e-5, "epsilon must be no less than 1e-5." + self.config.epsilon = epsilon input_layer = self.get_input_layer(0) image_conf = self.config.inputs[0].image_conf @@ -2714,7 +2713,7 @@ Usage: max_sort_size = -1, inputs = ["output", "score"]) Input data: Samples of the same query should be loaded as a sequence, - by ProtoDataProvider or PyDataProvider etc.. User should provide + by PyDataProvider etc.. User should provide scores for each sample. The score slot should be the 2nd input of lambdaRank layer. @@ -2816,19 +2815,18 @@ class AddToLayer(LayerBase): name, self.layer_type, 0, inputs=inputs, **xargs) config_assert(len(inputs) > 0, 'inputs cannot be empty for AddToLayer') - if len(self.inputs) > 1: - for input_index in xrange(len(self.inputs)): - assert self.get_input_layer(0).height == self.get_input_layer( - input_index).height - assert self.get_input_layer(0).width == self.get_input_layer( - input_index).width - assert self.get_input_layer(0).depth == self.get_input_layer( - input_index).depth + layer_size = self.get_input_layer(0).size + # To reserve heght, width, depth. + layer_with_hwc = self.get_input_layer(0) + for input_index in xrange(len(self.inputs)): + input_layer = self.get_input_layer(input_index) + assert layer_size == input_layer.size + if input_layer.height and input_layer.height and input_layer.height: + layer_with_hwc = input_layer - self.set_layer_size(self.get_input_layer(0).size) - self.set_layer_height_width(self.get_input_layer(0).height, \ - self.get_input_layer(0).width) - self.set_layer_depth(self.get_input_layer(0).depth) + self.set_layer_size(layer_with_hwc.size) + self.set_layer_height_width(layer_with_hwc.height, layer_with_hwc.width) + self.set_layer_depth(layer_with_hwc.depth) self.create_bias_parameter(bias, self.config.size) @@ -3868,6 +3866,26 @@ class SwitchOrderLayer(LayerBase): name, 'switch_order', 0, inputs=inputs, **xargs) self.config.reshape_conf.height_axis.extend(reshape['height']) self.config.reshape_conf.width_axis.extend(reshape['width']) + input_layer = self.get_input_layer(0) + if reshape is None: + self.set_layer_size(input_layer.size) + else: + in_h = input_layer.height + in_w = input_layer.width + out_dims = None + if input_layer.has_depth(): + in_d = input_layer.depth + in_c = input_layer.size / in_h / in_w / in_d + # batch_size, depth, height, width, channel + out_dims = [0, in_d, in_h, in_w, in_c] + else: + in_c = input_layer.size / in_h / in_w + # batch_size, height, width, channel + out_dims = [0, in_h, in_w, in_c] + # Because (reshape['width'][0] > 0) always be true. + # So out_dims[0] won't be used. + size = reduce(lambda x, y: x * y, out_dims[reshape['width'][0]:]) + self.set_layer_size(size) @config_layer('scale_sub_region') @@ -3889,6 +3907,21 @@ class ScaleSubRegionLayer(LayerBase): image_conf.channels) +@config_layer('factorization_machine') +class FactorizationMachineLayer(LayerBase): + def __init__(self, name, inputs, factor_size, **xargs): + super(FactorizationMachineLayer, self).__init__( + name, 'factorization_machine', size=1, inputs=inputs, **xargs) + config_assert( + len(self.inputs) == 1, + 'factorization machine layer must have one and only one input.') + self.config.factor_size = factor_size + input_layer = self.get_input_layer(0) + psize = input_layer.size * factor_size + dims = [input_layer.size, factor_size] + self.create_input_parameter(0, psize, dims) + + # Deprecated, use a new layer specific class instead @config_func def Layer(name, type, **xargs): diff --git a/python/paddle/trainer_config_helpers/activations.py b/python/paddle/trainer_config_helpers/activations.py index c749fa827fea4a808ab715dcb3442aa24d06a4d2..00efc01c0592107314f5b23c951706d039d49a88 100644 --- a/python/paddle/trainer_config_helpers/activations.py +++ b/python/paddle/trainer_config_helpers/activations.py @@ -17,7 +17,8 @@ __all__ = [ "IdentityActivation", "LinearActivation", 'SequenceSoftmaxActivation', 'ExpActivation', "ReluActivation", "BReluActivation", "SoftReluActivation", "STanhActivation", "AbsActivation", "SquareActivation", "BaseActivation", - "LogActivation", "SqrtActivation", "ReciprocalActivation" + "LogActivation", "SqrtActivation", "ReciprocalActivation", + "SoftSignActivation" ] @@ -243,8 +244,20 @@ class ReciprocalActivation(BaseActivation): Reciprocal Activation. .. math:: - f(z) = 1/z + f(z)=\\frac{1}{z} """ def __init__(self): BaseActivation.__init__(self, 'reciprocal', False) + + +class SoftSignActivation(BaseActivation): + """ + SoftSign Activation. + + .. math:: + f(z)=\\frac{z}{1 + |z|} + """ + + def __init__(self): + BaseActivation.__init__(self, 'softsign', False) diff --git a/python/paddle/trainer_config_helpers/evaluators.py b/python/paddle/trainer_config_helpers/evaluators.py index 57979db4de08989ab583b0ab41589c09789a0921..95797fba8f67bacb421f5c2813ad6332bc53cbc9 100644 --- a/python/paddle/trainer_config_helpers/evaluators.py +++ b/python/paddle/trainer_config_helpers/evaluators.py @@ -297,7 +297,7 @@ def auc_evaluator( def pnpair_evaluator( input, label, - info, + query_id, weight=None, name=None, ): """ @@ -308,16 +308,20 @@ def pnpair_evaluator( .. code-block:: python - eval = pnpair_evaluator(input, label, info) + eval = pnpair_evaluator(input, label, query_id) :param input: Input Layer name. The output prediction of network. :type input: LayerOutput :param label: Label layer name. :type label: LayerOutput - :param info: Info layer name. (TODO, explaination) - :type info: LayerOutput + :param query_id: Query_id layer name. Query_id indicates that which query + each sample belongs to. Its shape should be + the same as output of Label layer. + :type query_id: LayerOutput :param weight: Weight Layer name. It should be a matrix with size - [sample_num, 1]. (TODO, explaination) + [sample_num, 1] which indicates the weight of each sample. + The default weight of sample is 1 if the weight layer is None. + And the pair weight is the mean of the two samples' weight. :type weight: LayerOutput :param name: Evaluator name. :type name: None|basestring @@ -326,8 +330,8 @@ def pnpair_evaluator( input = [input] if label: input.append(label) - if info: - input.append(info) + if query_id: + input.append(query_id) evaluator_base( input=input, type="pnpair", diff --git a/python/paddle/trainer_config_helpers/layers.py b/python/paddle/trainer_config_helpers/layers.py index 14cdee4c5564f7a1cc4ff7a19f4f7ac02b08f21c..19e2ab1b7da7b1ceacd6842f2d74ac551497c77b 100644 --- a/python/paddle/trainer_config_helpers/layers.py +++ b/python/paddle/trainer_config_helpers/layers.py @@ -21,7 +21,7 @@ from .activations import LinearActivation, SigmoidActivation, TanhActivation, \ ReluActivation, IdentityActivation, SoftmaxActivation, BaseActivation from .evaluators import * from .poolings import MaxPooling, AvgPooling, MaxWithMaskPooling, BasePoolingType, \ - CudnnAvgPooling, CudnnMaxPooling + CudnnAvgPooling, CudnnAvgInclPadPooling, CudnnMaxPooling from .attrs import * from .default_decorators import * @@ -148,6 +148,7 @@ __all__ = [ 'resize_layer', 'sub_seq_layer', 'scale_sub_region_layer', + 'factorization_machine', ] @@ -264,10 +265,12 @@ class LayerType(object): SCALE_SUB_REGION_LAYER = 'scale_sub_region' + FACTORIZATION_MACHINE = 'factorization_machine' + @staticmethod def is_layer_type(type_name): """ - If type_name is a layer type. + Whether type_name is a layer type. :param type_name: layer type name. Because layer type enumerations are strings. @@ -438,7 +441,7 @@ def full_matrix_projection(input, size=0, param_attr=None): with mixed_layer(size=100) as m: m += full_matrix_projection(input=layer) - 2. When used as an independant object like this, you must set the size: + 2. When used as an independent object like this, you must set the size: .. code-block:: python @@ -448,11 +451,11 @@ def full_matrix_projection(input, size=0, param_attr=None): :param input: The input of this layer. :type input: LayerOutput - :param size: The parameter size. Means the width of parameter. + :param size: The dimension of this layer. :type size: int - :param param_attr: Parameter config, None if use default. + :param param_attr: The parameter attribute. See ParameterAttribute for details. :type param_attr: ParameterAttribute - :return: A FullMatrixProjection Object. + :return: FullMatrixProjection Object. :rtype: FullMatrixProjection """ proj = FullMatrixProjection( @@ -465,12 +468,12 @@ def full_matrix_projection(input, size=0, param_attr=None): def trans_full_matrix_projection(input, size=0, param_attr=None): """ Different from full_matrix_projection, this projection performs matrix - multiplication, using transpose of weight. + multiplication, using the transpose of weight. .. math:: out.row[i] += in.row[i] * w^\mathrm{T} - :math:`w^\mathrm{T}` means transpose of weight. + :math:`w^\mathrm{T}` means the transpose of weight. The simply usage is: .. code-block:: python @@ -486,9 +489,9 @@ def trans_full_matrix_projection(input, size=0, param_attr=None): :type input: LayerOutput :param size: The parameter size. Means the width of parameter. :type size: int - :param param_attr: Parameter config, None if use default. + :param param_attr: The parameter attribute. See ParameterAttribute for details. :type param_attr: ParameterAttribute - :return: A TransposedFullMatrixProjection Object. + :return: TransposedFullMatrixProjection Object. :rtype: TransposedFullMatrixProjection """ proj = TransposedFullMatrixProjection( @@ -518,7 +521,7 @@ def table_projection(input, size=0, param_attr=None): with mixed_layer(size=100) as m: m += table_projection(input=layer) - 2. When used as an independant object like this, you must set the size: + 2. When used as an independent object like this, you must set the size: .. code-block:: python @@ -529,11 +532,11 @@ def table_projection(input, size=0, param_attr=None): :param input: The input of this layer, which must contains id fields. :type input: LayerOutput - :param size: The parameter size. Means the width of parameter. + :param size: The dimension of the output. :type size: int - :param param_attr: Parameter config, None if use default. + :param param_attr: The parameter attribute. See ParameterAttribute for details. :type param_attr: ParameterAttribute - :return: A TableProjection Object. + :return: TableProjection Object. :rtype: TableProjection """ proj = TableProjection( @@ -544,7 +547,7 @@ def table_projection(input, size=0, param_attr=None): def identity_projection(input, offset=None, size=None): """ - 1. IdentityProjection if offset=None. It performs: + 1. If offset=None, it performs IdentityProjection as follows: .. math:: out.row[i] += in.row[i] @@ -556,9 +559,8 @@ def identity_projection(input, offset=None, size=None): proj = identity_projection(input=layer) - 2. IdentityOffsetProjection if offset!=None. It likes IdentityProjection, - but layer size may be smaller than input size. - It select dimesions [offset, offset+layer_size) from input: + 2. If offset!=None, It executes IdentityOffsetProjection and takes the + elements of the input in the range [offset, offset+size) as output. .. math:: out.row[i] += in.row[i + \\textrm{offset}] @@ -570,14 +572,20 @@ def identity_projection(input, offset=None, size=None): proj = identity_projection(input=layer, offset=10) - Note that both of two projections should not have any parameter. + Note that neither of the projections have trainable parameter. :param input: The input of this layer. :type input: LayerOutput - :param offset: Offset, None if use default. + :param offset: The offset from the start of the input. The input's + elements in the range [offset, offset+size) will be + taken as output. If this parameter is not set or set + to None, the output will be the same as the input. :type offset: int - :return: A IdentityProjection or IdentityOffsetProjection object - :rtype: IdentityProjection or IdentityOffsetProjection + :param size: The dimension of this layer. It will be neglected + when offset is None or not set. + :type size: int + :return: IdentityProjection or IdentityOffsetProjection object + :rtype: IdentityProjection | IdentityOffsetProjection """ if offset is None: proj = IdentityProjection(input_layer_name=input.name) @@ -593,8 +601,8 @@ def identity_projection(input, offset=None, size=None): def slice_projection(input, slices): """ - slice_projection can slice the input value into multiple parts, - and then select some of them to merge into a new output. + slice_projection slices the input value into multiple parts, + then selects and merges some of them into a new output. .. math:: output = [input.slices()] @@ -605,15 +613,13 @@ def slice_projection(input, slices): proj = slice_projection(input=layer, slices=[(0, 10), (20, 30)]) - Note that slice_projection should not have any parameter. + Note that slice_projection has no trainable parameter. :param input: The input of this layer. :type input: LayerOutput - :param slices: An array of slice parameters. - Each slice contains the start and end offsets based - on the input. - :type slices: pair of int - :return: A SliceProjection object + :param slices: A list of start and end offsets of each slice. + :type slices: list of tuple + :return: SliceProjection object. :rtype: SliceProjection """ assert len(slices) >= 1 @@ -633,8 +639,7 @@ def slice_projection(input, slices): @wrap_param_attr_default() def scaling_projection(input, param_attr=None): """ - scaling_projection multiplies the input with a scalar parameter and add to - the output. + scaling_projection multiplies the input with a scalar parameter. .. math:: out += w * in @@ -647,9 +652,9 @@ def scaling_projection(input, param_attr=None): :param input: The input of this layer. :type input: LayerOutput - :param param_attr: Parameter config, None if use default. + :param param_attr: The parameter attribute. See ParameterAttribute for details. :type param_attr: ParameterAttribute - :return: A ScalingProjection object + :return: ScalingProjection object. :rtype: ScalingProjection """ proj = ScalingProjection(input_layer_name=input.name, **param_attr.attr) @@ -660,8 +665,8 @@ def scaling_projection(input, param_attr=None): @wrap_param_attr_default() def dotmul_projection(input, param_attr=None): """ - DotMulProjection with a layer as input. - It performs element-wise multiplication with weight. + DotMulProjection takes a layer as input and performs + element-wise multiplication with weight. .. math:: out.row[i] += in.row[i] .* weight @@ -676,9 +681,9 @@ def dotmul_projection(input, param_attr=None): :param input: The input of this layer. :type input: LayerOutput - :param param_attr: Parameter config, None if use default. + :param param_attr: The parameter attribute. See ParameterAttribute for details. :type param_attr: ParameterAttribute - :return: A DotMulProjection Object. + :return: DotMulProjection object. :rtype: DotMulProjection """ proj = DotMulProjection( @@ -695,7 +700,7 @@ def dotmul_operator(a=None, b=None, scale=1, **kwargs): out.row[i] += scale * (a.row[i] .* b.row[i]) where :math:`.*` means element-wise multiplication, and - scale is a config scalar, its default value is one. + scale is a config scalar, its default value is 1. The example usage is: @@ -703,13 +708,13 @@ def dotmul_operator(a=None, b=None, scale=1, **kwargs): op = dotmul_operator(a=layer1, b=layer2, scale=0.5) - :param a: Input layer1 + :param a: The first input of this layer. :type a: LayerOutput - :param b: Input layer2 + :param b: The second input of this layer. :type b: LayerOutput - :param scale: config scalar, default value is one. + :param scale: A scalar to scale the product. Its default value is 1. :type scale: float - :return: A DotMulOperator Object. + :return: DotMulOperator object. :rtype: DotMulOperator """ if 'x' in kwargs or 'y' in kwargs: @@ -735,28 +740,29 @@ def context_projection(input, """ Context Projection. - It just simply reorganizes input sequence, combines "context_len" sequence - to one context from context_start. "context_start" will be set to - -(context_len - 1) / 2 by default. If context position out of sequence + It just reorganizes input sequence, combines "context_len" elements of the + sequence to one context from context_start. "context_start" will be set to + -(context_len - 1) / 2 by default. When context position is out of sequence length, padding will be filled as zero if padding_attr = False, otherwise it is trainable. - For example, origin sequence is [A B C D E F G], context len is 3, then - after context projection and not set padding_attr, sequence will + For example, origin sequence is [A B C D E F G], context len is 3, padding_attr + is not set, then after context projection, sequence will be [ 0AB ABC BCD CDE DEF EFG FG0 ]. :param input: The input of this layer, which should be a sequence. :type input: LayerOutput - :param context_len: context length. + :param context_len: The length of the context. :type context_len: int - :param context_start: context start position. Default is + :param context_start: The start position of the context. The default value is -(context_len - 1)/2 :type context_start: int - :param padding_attr: Padding Parameter Attribute. If false, it means padding - always be zero. Otherwise Padding is learnable, and - parameter attribute is set by this parameter. + :param padding_attr: Parameter attribute of the padding. If the parameter is + set to False, padding will be zero. In other cases, the + padding is trainable, and its parameter attribute is set + by this parameter. :type padding_attr: bool | ParameterAttribute - :return: Projection + :return: Projection object. :rtype: Projection """ context_start = -( @@ -788,10 +794,9 @@ class MixedLayerType(LayerOutput): def __init__(self, name, size, act, bias_attr, layer_attr, parents=None): """ - Ctor. - :param name: layer name. + :param name: The name of this layer. :type name: basestring - :param size: layer size. + :param size: The dimension of this layer. :type size: int :param act: Activation type. :type act: BaseActivation @@ -799,8 +804,9 @@ class MixedLayerType(LayerOutput): whose type is not ParameterAttribute, no bias is defined. If the parameter is set to True, the bias is initialized to zero. :type bias_attr: ParameterAttribute | None | bool | Any - :param layer_attr: Extra Layer Attribute. - :type layer_attr: ExtraLayerAttribute or None + :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for + details. + :type layer_attr: ExtraLayerAttribute | None """ LayerOutput.__init__( self, @@ -865,12 +871,12 @@ def mixed_layer(size=0, bias_attr=False, layer_attr=None): """ - Mixed Layer. A mixed layer will add all inputs together, then activate. - Each inputs is a projection or operator. + Mixed Layer. A mixed layer will add all inputs together, then activate the sum. + Each input is a projection or operator. There are two styles of usages. - 1. When not set inputs parameter, use mixed_layer like this: + 1. When the parameter input is not set, use mixed_layer like this: .. code-block:: python @@ -886,21 +892,21 @@ def mixed_layer(size=0, input=[full_matrix_projection(input=layer1), full_matrix_projection(input=layer2)]) - :param name: mixed layer name. Can be referenced by other layer. + :param name: The name of this layer. It is optional. :type name: basestring - :param size: layer size. + :param size: The dimension of this layer. :type size: int - :param input: The input of this layer. It is an optional parameter. If set, - then this function will just return layer's name. + :param input: The input of this layer. It is an optional parameter. :param act: Activation Type. LinearActivation is the default activation. :type act: BaseActivation :param bias_attr: The bias attribute. If the parameter is set to False or an object whose type is not ParameterAttribute, no bias is defined. If the parameter is set to True, the bias is initialized to zero. :type bias_attr: ParameterAttribute | None | bool | Any - :param layer_attr: The extra layer config. Default is None. + :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for + details. :type layer_attr: ExtraLayerAttribute - :return: MixedLayerType object can add inputs or layer name. + :return: MixedLayerType object. :rtype: MixedLayerType """ @@ -935,14 +941,15 @@ def data_layer(name, size, depth=None, height=None, width=None, :param name: The name of this layer. :type name: basestring - :param size: Size of this data layer. + :param size: The dimension of this data layer. :type size: int - :param height: Height of this data layer, used for image + :param height: The height of the input image data. :type height: int | None - :param width: Width of this data layer, used for image + :param width: The width of the input image data. :type width: int | None - :param layer_attr: Extra Layer Attribute. - :type layer_attr: ExtraLayerAttribute. + :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for + details. + :type layer_attr: ExtraLayerAttribute :return: LayerOutput object. :rtype: LayerOutput """ @@ -975,14 +982,15 @@ def embedding_layer(input, size, name=None, param_attr=None, layer_attr=None): :param name: The name of this layer. It is optional. :type name: basestring - :param input: The input of this layer, which must be Index Data. + :param input: The input of this layer, whose type must be Index Data. :type input: LayerOutput - :param size: The embedding dimension. + :param size: The dimension of the embedding vector. :type size: int :param param_attr: The embedding parameter attribute. See ParameterAttribute for details. - :type param_attr: ParameterAttribute | None - :param layer_attr: Extra layer Config. Default is None. + :type param_attr: ParameterAttribute + :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for + details. :type layer_attr: ExtraLayerAttribute | None :return: LayerOutput object. :rtype: LayerOutput @@ -1010,7 +1018,7 @@ def fc_layer(input, bias_attr=None, layer_attr=None): """ - Helper for declare fully connected layer. + The fully connected layer. The example usage is: @@ -1032,17 +1040,18 @@ def fc_layer(input, :type name: basestring :param input: The input of this layer. :type input: LayerOutput | list | tuple - :param size: The layer dimension. + :param size: The dimension of this layer. :type size: int :param act: Activation Type. TanhActivation is the default activation. :type act: BaseActivation - :param param_attr: The Parameter Attribute|list. + :param param_attr: The parameter attribute. See ParameterAttribute for details. :type param_attr: ParameterAttribute :param bias_attr: The bias attribute. If the parameter is set to False or an object whose type is not ParameterAttribute, no bias is defined. If the parameter is set to True, the bias is initialized to zero. :type bias_attr: ParameterAttribute | None | bool | Any - :param layer_attr: Extra Layer config. + :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for + details. :type layer_attr: ExtraLayerAttribute | None :return: LayerOutput object. :rtype: LayerOutput @@ -1083,13 +1092,15 @@ def fc_layer(input, @wrap_name_default("print") def printer_layer(input, format=None, name=None): """ - Print the output value of input layers. This layer is useful for debugging. + Print the output value of the layers specified by the parameter input. + This layer is useful for debugging. :param name: The name of this layer. It is optional. :type name: basestring :param input: The input of this layer. :type input: LayerOutput | list | tuple - :return: LayerOutput + :return: LayerOutput object. + :rtype: LayerOutput """ if isinstance(input, LayerOutput): input = [input] @@ -1132,11 +1143,12 @@ def priorbox_layer(input, :param aspect_ratio: The aspect ratio. :type aspect_ratio: list :param variance: The bounding box variance. - :type min_size: The min size of the priorbox width/height. + :type min_size: The minimum size of the priorbox width/height. :param min_size: list - :type max_size: The max size of the priorbox width/height. Could be NULL. + :type max_size: The maximum size of the priorbox width/height. It could be NULL. :param max_size: list - :return: LayerOutput + :return: LayerOutput object. + :rtype: LayerOutput """ # plus one for ratio 1. num_filters = (len(aspect_ratio) * 2 + 1 + len(max_size)) * 4 @@ -1174,7 +1186,7 @@ def multibox_loss_layer(input_loc, :param name: The name of this layer. It is optional. :type name: basestring - :param input_loc: The input predict locations. + :param input_loc: The input predicted locations. :type input_loc: LayerOutput | List of LayerOutput :param input_conf: The input priorbox confidence. :type input_conf: LayerOutput | List of LayerOutput @@ -1186,13 +1198,15 @@ def multibox_loss_layer(input_loc, :type num_classes: int :param overlap_threshold: The threshold of the overlap. :type overlap_threshold: float - :param neg_pos_ratio: The ratio of the negative bbox to the positive bbox. + :param neg_pos_ratio: The ratio of the negative bounding box to + the positive bounding box. :type neg_pos_ratio: float - :param neg_overlap: The negative bbox overlap threshold. + :param neg_overlap: The negative bounding box overlap threshold. :type neg_overlap: float :param background_id: The background class index. :type background_id: int - :return: LayerOutput + :return: LayerOutput object. + :rtype: LayerOutput """ if isinstance(input_loc, LayerOutput): input_loc = [input_loc] @@ -1255,19 +1269,20 @@ def detection_output_layer(input_loc, :type input_conf: LayerOutput | List of LayerOutput. :param priorbox: The input priorbox location and the variance. :type priorbox: LayerOutput - :param num_classes: The number of the classification. + :param num_classes: The number of the classes. :type num_classes: int :param nms_threshold: The Non-maximum suppression threshold. :type nms_threshold: float - :param nms_top_k: The bbox number kept of the NMS's output + :param nms_top_k: The bounding boxes number kept of the NMS's output. :type nms_top_k: int - :param keep_top_k: The bbox number kept of the layer's output + :param keep_top_k: The bounding boxes number kept of the layer's output. :type keep_top_k: int - :param confidence_threshold: The classification confidence threshold + :param confidence_threshold: The classification confidence threshold. :type confidence_threshold: float :param background_id: The background class index. :type background_id: int - :return: LayerOutput + :return: LayerOutput object. + :rtype: LayerOutput """ if isinstance(input_loc, LayerOutput): input_loc = [input_loc] @@ -1323,7 +1338,7 @@ def roi_pool_layer(input, A layer used by Fast R-CNN to extract feature maps of ROIs from the last feature map. - :param name: The Layer Name. + :param name: The name of this layer. It is optional. :type name: basestring :param input: The input layer. :type input: LayerOutput. @@ -1335,9 +1350,10 @@ def roi_pool_layer(input, :type pooled_height: int :param spatial_scale: The spatial scale between the image and feature map. :type spatial_scale: float - :param num_channels: number of input channel. + :param num_channels: The number of the input channels. :type num_channels: int - :return: LayerOutput + :return: LayerOutput object. + :rtype: LayerOutput """ if num_channels is None: assert input.num_filters is not None @@ -1358,18 +1374,19 @@ def roi_pool_layer(input, @wrap_name_default("cross_channel_norm") def cross_channel_norm_layer(input, name=None, param_attr=None): """ - Normalize a layer's output. This layer is necessary for ssd. - This layer applys normalize across the channels of each sample to - a conv layer's output and scale the output by a group of trainable - factors which dimensions equal to the channel's number. + Normalize a layer's output. This layer is necessary for ssd. This + layer applys normalization across the channels of each sample to + a convolutional layer's output and scales the output by a group of + trainable factors whose dimensions equal to the channel's number. :param name: The name of this layer. It is optional. :type name: basestring :param input: The input of this layer. :type input: LayerOutput - :param param_attr: The Parameter Attribute|list. + :param param_attr: The parameter attribute. See ParameterAttribute for details. :type param_attr: ParameterAttribute - :return: LayerOutput + :return: LayerOutput object. + :rtype: LayerOutput """ assert input.num_filters is not None Layer( @@ -1410,12 +1427,9 @@ def pooling_layer(input, Pooling layer for sequence inputs, not used for Image. If stride > 0, this layer slides a window whose size is determined by stride, - and return the pooling value of the window as the output. Thus, a long sequence - will be shorten. - - The parameter stride specifies the intervals at which to apply the pooling - operation. Note that for sequence with sub-sequence, the default value - of stride is -1. + and returns the pooling value of the sequence in the window as the output. Thus, + a long sequence will be shortened. Note that for sequence with sub-sequence, the + default value of stride is -1. The example usage is: @@ -1432,16 +1446,16 @@ def pooling_layer(input, :type name: basestring :param input: The input of this layer. :type input: LayerOutput - :param pooling_type: Type of pooling, MaxPooling(default), AvgPooling, - SumPooling, SquareRootNPooling. + :param pooling_type: Type of pooling. MaxPooling is the default pooling. :type pooling_type: BasePoolingType | None :param stride: The step size between successive pooling regions. - :type stride: Int + :type stride: int :param bias_attr: The bias attribute. If the parameter is set to False or an object whose type is not ParameterAttribute, no bias is defined. If the parameter is set to True, the bias is initialized to zero. :type bias_attr: ParameterAttribute | None | bool | Any - :param layer_attr: The Extra Attributes for layer, such as dropout. + :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for + details. :type layer_attr: ExtraLayerAttribute | None :return: LayerOutput object. :rtype: LayerOutput @@ -1516,34 +1530,33 @@ def lstmemory(input, NOTE: This is a low level user interface. You can use network.simple_lstm to config a simple plain lstm layer. - Please refer to **Generating Sequences With Recurrent Neural Networks** for - more details about LSTM. - - Link_ goes as below. - - .. _Link: http://arxiv.org/abs/1308.0850 + Reference: + `Generating Sequences With Recurrent Neural Networks + `_ - :param name: The lstmemory layer name. + :param name: The name of this layer. It is optional. :type name: basestring - :param size: DEPRECATED. size of the lstm cell + :param size: DEPRECATED. The dimension of the lstm cell. :type size: int :param input: The input of this layer. :type input: LayerOutput - :param reverse: is sequence process reversed or not. + :param reverse: Whether the input sequence is processed in a reverse order. :type reverse: bool :param act: Activation type. TanhActivation is the default activation. :type act: BaseActivation - :param gate_act: gate activation type, SigmoidActivation by default. + :param gate_act: Activation type of this layer's gates. SigmoidActivation is the + default activation. :type gate_act: BaseActivation - :param state_act: state activation type, TanhActivation by default. + :param state_act: Activation type of the state. TanhActivation is the default activation. :type state_act: BaseActivation :param bias_attr: The bias attribute. If the parameter is set to False or an object whose type is not ParameterAttribute, no bias is defined. If the parameter is set to True, the bias is initialized to zero. :type bias_attr: ParameterAttribute | None | bool | Any - :param param_attr: Parameter Attribute. - :type param_attr: ParameterAttribute | None | False - :param layer_attr: Extra Layer attribute + :param param_attr: The parameter attribute. See ParameterAttribute for details. + :type param_attr: ParameterAttribute + :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for + details. :type layer_attr: ExtraLayerAttribute | None :return: LayerOutput object. :rtype: LayerOutput @@ -1632,14 +1645,14 @@ def grumemory(input, h_t = (1 - z_t) h_{t-1} + z_t {\\tilde{h_t}} NOTE: In PaddlePaddle's implementation, the multiplication operations - :math:`W_{r}x_{t}`, :math:`W_{z}x_{t}` and :math:`W x_t` are not computed in - gate_recurrent layer. Consequently, an additional mixed_layer with + :math:`W_{r}x_{t}`, :math:`W_{z}x_{t}` and :math:`W x_t` are not performed + in gate_recurrent layer. Consequently, an additional mixed_layer with full_matrix_projection or a fc_layer must be included before grumemory is called. - More details can be found by referring to `Empirical Evaluation of Gated - Recurrent Neural Networks on Sequence Modeling. - `_ + Reference: + `Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling + `_ The simple usage is: @@ -1647,28 +1660,29 @@ def grumemory(input, gru = grumemory(input) - :param name: The gru layer name. - :type name: None | basestring + :param name: The name of this layer. It is optional. + :type name: basestring :param input: The input of this layer. :type input: LayerOutput. - :param size: DEPRECATED. size of the gru cell + :param size: DEPRECATED. The dimension of the gru cell. :type size: int - :param reverse: Whether sequence process is reversed or not. + :param reverse: Whether the input sequence is processed in a reverse order. :type reverse: bool :param act: Activation type, TanhActivation is the default. This activation affects the :math:`{\\tilde{h_t}}`. :type act: BaseActivation - :param gate_act: gate activation type, SigmoidActivation by default. - This activation affects the :math:`z_t` and :math:`r_t`. It is the - :math:`\\sigma` in the above formula. + :param gate_act: Activation type of this layer's two gates. SigmoidActivation is + the default activation. This activation affects the :math:`z_t` + and :math:`r_t`. It is the :math:`\\sigma` in the above formula. :type gate_act: BaseActivation :param bias_attr: The bias attribute. If the parameter is set to False or an object whose type is not ParameterAttribute, no bias is defined. If the parameter is set to True, the bias is initialized to zero. :type bias_attr: ParameterAttribute | None | bool | Any - :param param_attr: Parameter Attribute. - :type param_attr: ParameterAttribute | None | False - :param layer_attr: Extra Layer attribute + :param param_attr: The parameter attribute. See ParameterAttribute for details. + :type param_attr: ParameterAttribute + :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for + details. :type layer_attr: ExtraLayerAttribute | None :return: LayerOutput object. :rtype: LayerOutput @@ -1712,10 +1726,10 @@ def last_seq(input, """ Get Last Timestamp Activation of a sequence. - If stride > 0, this layer slides a window whose size is determined by stride, - and return the last value of the window as the output. Thus, a long sequence - will be shorten. Note that for sequence with sub-sequence, the default value - of stride is -1. + If stride > 0, this layer will slide a window whose size is determined by stride, + and return the last value of the sequence in the window as the output. Thus, a + long sequence will be shortened. Note that for sequence with sub-sequence, the + default value of stride is -1. The simple usage is: @@ -1724,14 +1738,16 @@ def last_seq(input, seq = last_seq(input=layer) :param agg_level: Aggregated level + :type agg_level: AggregateLevel :param name: The name of this layer. It is optional. :type name: basestring :param input: The input of this layer. :type input: LayerOutput :param stride: The step size between successive pooling regions. - :type stride: Int - :param layer_attr: extra layer attributes. - :type layer_attr: ExtraLayerAttribute. + :type stride: int + :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for + details. + :type layer_attr: ExtraLayerAttribute :return: LayerOutput object. :rtype: LayerOutput """ @@ -1768,10 +1784,10 @@ def first_seq(input, """ Get First Timestamp Activation of a sequence. - If stride > 0, this layer slides a window whose size is determined by stride, - and return the first value of the window as the output. Thus, a long sequence - will be shorten. Note that for sequence with sub-sequence, the default value - of stride is -1. + If stride > 0, this layer will slide a window whose size is determined by stride, + and return the first value of the sequence in the window as the output. Thus, a + long sequence will be shortened. Note that for sequence with sub-sequence, the + default value of stride is -1. The simple usage is: @@ -1780,13 +1796,15 @@ def first_seq(input, seq = first_seq(input=layer) :param agg_level: aggregation level + :type agg_level: AggregateLevel :param name: The name of this layer. It is optional. :type name: basestring :param input: The input of this layer. :type input: LayerOutput :param stride: The step size between successive pooling regions. - :type stride: Int - :param layer_attr: extra layer attributes. + :type stride: int + :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for + details. :type layer_attr: ExtraLayerAttribute. :return: LayerOutput object. :rtype: LayerOutput @@ -1844,8 +1862,8 @@ def expand_layer(input, expand_level=ExpandLevel.FROM_NO_SEQUENCE, layer_attr=None): """ - A layer for "Expand Dense data or (sequence data where the length of each - sequence is one) to sequence data." + A layer for expanding dense data or (sequence data where the length of each + sequence is one) to sequence data. The example usage is: @@ -1857,7 +1875,9 @@ def expand_layer(input, :param input: The input of this layer. :type input: LayerOutput - :param expand_as: Expand as this layer's sequence info. + :param expand_as: Expand the input according to this layer's sequence infomation. And + after the operation, the input expanded will have the same number of + elememts as this layer. :type expand_as: LayerOutput :param name: The name of this layer. It is optional. :type name: basestring @@ -1865,9 +1885,10 @@ def expand_layer(input, whose type is not ParameterAttribute, no bias is defined. If the parameter is set to True, the bias is initialized to zero. :type bias_attr: ParameterAttribute | None | bool | Any - :param expand_level: whether input layer is timestep(default) or sequence. + :param expand_level: Whether the input layer is a sequence or the element of a sequence. :type expand_level: ExpandLevel - :param layer_attr: extra layer attributes. + :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for + details. :type layer_attr: ExtraLayerAttribute. :return: LayerOutput object. :rtype: LayerOutput @@ -1900,9 +1921,12 @@ def repeat_layer(input, A layer for repeating the input for num_repeats times. If as_row_vector: + .. math:: y = [x_1,\cdots, x_n, \cdots, x_1, \cdots, x_n] + If not as_row_vector: + .. math:: y = [x_1,\cdots, x_1, \cdots, x_n, \cdots, x_n] @@ -1915,19 +1939,19 @@ def repeat_layer(input, :param input: The input of this layer. :type input: LayerOutput - :param num_repeats: Repeat the input so many times + :param num_repeats: The times of repeating the input. :type num_repeats: int :param name: The name of this layer. It is optional. - :param as_row_vector: True for treating input as row vector and repeating - in the column direction. This is equivalent to apply - concat_layer() with num_repeats same input. - False for treating input as column vector and repeating - in the row direction. + :type name: basestring + :param as_row_vector: Whether to treat the input as row vectors or not. If + the parameter is set to True, the repeating operation + will be performed in the column direction. Otherwise, + it will be performed in the row direction. :type as_row_vector: bool :param act: Activation type. IdentityActivation is the default activation. :type act: BaseActivation - :type name: basestring - :param layer_attr: extra layer attributes. + :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for + details. :type layer_attr: ExtraLayerAttribute. :return: LayerOutput object. :rtype: LayerOutput @@ -1974,13 +1998,14 @@ def seq_reshape_layer(input, :param input: The input of this layer. :type input: LayerOutput - :param reshape_size: the size of reshaped sequence. + :param reshape_size: The dimension of the reshaped sequence. :type reshape_size: int :param name: The name of this layer. It is optional. :type name: basestring :param act: Activation type. IdentityActivation is the default activation. :type act: BaseActivation - :param layer_attr: extra layer attributes. + :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for + details. :type layer_attr: ExtraLayerAttribute. :param bias_attr: The bias attribute. If the parameter is set to False or an object whose type is not ParameterAttribute, no bias is defined. If the @@ -2008,7 +2033,7 @@ def seq_reshape_layer(input, @layer_support() def interpolation_layer(input, weight, name=None, layer_attr=None): """ - This layer is for linear interpolation with two inputs, + This layer performs linear interpolation on two inputs, which is used in NEURAL TURING MACHINE. .. math:: @@ -2030,7 +2055,8 @@ def interpolation_layer(input, weight, name=None, layer_attr=None): :type weight: LayerOutput :param name: The name of this layer. It is optional. :type name: basestring - :param layer_attr: extra layer attributes. + :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for + details. :type layer_attr: ExtraLayerAttribute. :return: LayerOutput object. :rtype: LayerOutput @@ -2064,7 +2090,7 @@ def bilinear_interp_layer(input, name=None, layer_attr=None): """ - This layer is to implement bilinear interpolation on conv layer output. + This layer implements bilinear interpolation on convolutional layer's output. Please refer to Wikipedia: https://en.wikipedia.org/wiki/Bilinear_interpolation @@ -2074,18 +2100,19 @@ def bilinear_interp_layer(input, bilinear = bilinear_interp_layer(input=layer1, out_size_x=64, out_size_y=64) - :param input: A input layer. - :type input: LayerOutput. - :param out_size_x: bilinear interpolation output width. - :type out_size_x: int | None - :param out_size_y: bilinear interpolation output height. - :type out_size_y: int | None - :param name: The layer's name, which cna not be specified. - :type name: None | basestring - :param layer_attr: Extra Layer attribute. - :type layer_attr: ExtraLayerAttribute + :param input: The input of this layer. + :type input: LayerOutput. + :param out_size_x: The width of the output. + :type out_size_x: int + :param out_size_y: The height of the output. + :type out_size_y: int + :param name: The name of this layer. It is optional. + :type name: basestring + :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for + details. + :type layer_attr: ExtraLayerAttribute :return: LayerOutput object. - :rtype: LayerOutput + :rtype: LayerOutput """ assert input.layer_type == LayerType.CONV_LAYER assert isinstance(input.activation, LinearActivation) @@ -2120,8 +2147,8 @@ def power_layer(input, weight, name=None, layer_attr=None): .. math:: y = x^w - where :math:`x` is a input vector, :math:`w` is scalar weight, - and :math:`y` is a output vector. + where :math:`x` is an input vector, :math:`w` is a scalar exponent, + and :math:`y` is an output vector. The example usage is: @@ -2131,11 +2158,12 @@ def power_layer(input, weight, name=None, layer_attr=None): :param input: The input of this layer. :type input: LayerOutput - :param weight: Weight layer. + :param weight: The exponent of the power. :type weight: LayerOutput :param name: The name of this layer. It is optional. :type name: basestring - :param layer_attr: extra layer attributes. + :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for + details. :type layer_attr: ExtraLayerAttribute. :return: LayerOutput object. :rtype: LayerOutput @@ -2175,11 +2203,12 @@ def scaling_layer(input, weight, name=None, layer_attr=None): :param input: The input of this layer. :type input: LayerOutput - :param weight: Weight layer. + :param weight: The weight of each sample. :type weight: LayerOutput :param name: The name of this layer. It is optional. :type name: basestring - :param layer_attr: extra layer attributes. + :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for + details. :type layer_attr: ExtraLayerAttribute. :return: LayerOutput object. :rtype: LayerOutput @@ -2217,7 +2246,8 @@ def trans_layer(input, name=None, layer_attr=None): :type input: LayerOutput :param name: The name of this layer. It is optional. :type name: basestring - :param layer_attr: extra layer attributes. + :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for + details. :type layer_attr: ExtraLayerAttribute. :return: LayerOutput object. :rtype: LayerOutput @@ -2253,11 +2283,14 @@ def rotate_layer(input, height, width, name=None, layer_attr=None): :param input: The input of this layer. :type input: LayerOutput - :param height: The height of the sample matrix + :param height: The height of the sample matrix. :type height: int + :param width: The width of the sample matrix. + :type width: int :param name: The name of this layer. It is optional. :type name: basestring - :param layer_attr: extra layer attributes. + :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for + details. :type layer_attr: ExtraLayerAttribute. :return: LayerOutput object. :rtype: LayerOutput @@ -2302,15 +2335,15 @@ def cos_sim(a, b, scale=1, size=1, name=None, layer_attr=None): :param name: The name of this layer. It is optional. :type name: basestring - :param a: input layer a + :param a: The first input of this layer. :type a: LayerOutput - :param b: input layer b + :param b: The second input of this layer. :type b: LayerOutput - :param scale: scale for cosine value. default is 5. + :param scale: The scale of the cosine similarity. 1 is the default value. :type scale: float - :param size: layer size. NOTE size_a * size should equal size_b. + :param size: The dimension of this layer. NOTE size_a * size should equal size_b. :type size: int - :param layer_attr: Extra Layer Attribute. + :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for details. :type layer_attr: ExtraLayerAttribute :return: LayerOutput object. :rtype: LayerOutput @@ -2395,8 +2428,10 @@ def hsigmoid(input, """ Organize the classes into a binary tree. At each node, a sigmoid function is used to calculate the probability of belonging to the right branch. - This idea is from "F. Morin, Y. Bengio (AISTATS 05): - Hierarchical Probabilistic Neural Network Language Model." + + Reference: + `Hierarchical Probabilistic Neural Network Language Model + `_ The example usage is: @@ -2407,19 +2442,21 @@ def hsigmoid(input, :param input: The input of this layer. :type input: LayerOutput | list | tuple - :param label: Label layer. + :param label: The input label. :type label: LayerOutput - :param num_classes: number of classes. - :type num_classes: int | None + :param num_classes: The number of classes. And it should be larger than 2. If the parameter + is not set or set to None, its actual value will be automatically set to + the number of labels. + :type num_classes: int :param name: The name of this layer. It is optional. :type name: basestring :param bias_attr: The bias attribute. If the parameter is set to False or an object whose type is not ParameterAttribute, no bias is defined. If the parameter is set to True, the bias is initialized to zero. :type bias_attr: ParameterAttribute | None | bool | Any - :param param_attr: Parameter Attribute. None means default parameter. - :type param_attr: ParameterAttribute | None - :param layer_attr: Extra Layer Attribute. + :param param_attr: The parameter attribute. See ParameterAttribute for details. + :type param_attr: ParameterAttribute + :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for details. :type layer_attr: ExtraLayerAttribute :return: LayerOutput object. :rtype: LayerOutput @@ -2507,12 +2544,12 @@ def img_conv_layer(input, input is raw pixels of image(mono or RGB), or it may be the previous layer's num_filters * num_group. - There are several group of filter in PaddlePaddle implementation. - Each group will process some channel of the inputs. For example, if an input + There are several groups of filters in PaddlePaddle implementation. + Each group will process some channels of the input. For example, if num_channel = 256, group = 4, num_filter=32, the PaddlePaddle will create - 32*4 = 128 filters to process inputs. The channels will be split into 4 - pieces. First 256/4 = 64 channels will process by first 32 filters. The - rest channels will be processed by rest group of filters. + 32*4 = 128 filters to process the input. The channels will be split into 4 + pieces. First 256/4 = 64 channels will be processed by first 32 filters. The + rest channels will be processed by the rest groups of filters. The example usage is: @@ -2528,53 +2565,68 @@ def img_conv_layer(input, :type name: basestring :param input: The input of this layer. :type input: LayerOutput - :param filter_size: The x dimension of a filter kernel. Or input a tuple for - two image dimension. + :param filter_size: The dimensions of the filter kernel. If the parameter is + set to one integer, the two dimensions on x and y axises + will be same when filter_size_y is not set. If it is set + to a list, the first element indicates the dimension on + the x axis, and the second is used to specify the dimension + on the y axis when filter_size_y is not provided. :type filter_size: int | tuple | list - :param filter_size_y: The y dimension of a filter kernel. Since PaddlePaddle - currently supports rectangular filters, the filter's - shape will be (filter_size, filter_size_y). - :type filter_size_y: int | None + :param filter_size_y: The dimension of the filter kernel on the y axis. If the parameter + is not set, it will be set automatically according to filter_size. + :type filter_size_y: int :param num_filters: Each filter group's number of filter :param act: Activation type. ReluActivation is the default activation. :type act: BaseActivation - :param groups: Group size of filters. + :param groups: The group number. 1 is the default group number. :type groups: int - :param stride: The x dimension of the stride. Or input a tuple for two image - dimension. + :param stride: The strides. If the parameter is set to one integer, the strides + on x and y axises will be same when stride_y is not set. If it is + set to a list, the first element indicates the stride on the x axis, + and the second is used to specify the stride on the y axis when + stride_y is not provided. 1 is the default value. :type stride: int | tuple | list - :param stride_y: The y dimension of the stride. + :param stride_y: The stride on the y axis. :type stride_y: int - :param padding: The x dimension of the padding. Or input a tuple for two - image dimension + :param padding: The padding sizes. If the parameter is set to one integer, the padding + sizes on x and y axises will be same when padding_y is not set. If it + is set to a list, the first element indicates the padding size on the + x axis, and the second is used to specify the padding size on the y axis + when padding_y is not provided. 0 is the default padding size. :type padding: int | tuple | list - :param padding_y: The y dimension of the padding. + :param padding_y: The padding size on the y axis. :type padding_y: int - :param dilation: The x dimension of the dilation. Or input a tuple for two - image dimension + :param dilation: The dimensions of the dilation. If the parameter is set to one integer, + the two dimensions on x and y axises will be same when dilation_y is not + set. If it is set to a list, the first element indicates the dimension + on the x axis, and the second is used to specify the dimension on the y + axis when dilation_y is not provided. 1 is the default dimension. :type dilation: int | tuple | list - :param dilation_y: The y dimension of the dilation. + :param dilation_y: The dimension of the dilation on the y axis. :type dilation_y: int :param bias_attr: The bias attribute. If the parameter is set to False or an object whose type is not ParameterAttribute, no bias is defined. If the parameter is set to True, the bias is initialized to zero. :type bias_attr: ParameterAttribute | None | bool | Any - :param num_channels: number of input channels. If None will be set - automatically from previous output. + :param num_channels: The number of input channels. If the parameter is not set or + set to None, its actual value will be automatically set to + the channel number of the input. :type num_channels: int - :param param_attr: Convolution param attribute. None means default attribute + :param param_attr: The parameter attribute. See ParameterAttribute for + details. :type param_attr: ParameterAttribute - :param shared_biases: Is biases will be shared between filters or not. + :param shared_biases: Whether biases will be shared between filters or not. :type shared_biases: bool - :param layer_attr: Layer Extra Attribute. + :param layer_attr: The extra layer attributes. See ExtraLayerAttribute for + details. :type layer_attr: ExtraLayerAttribute - :param trans: true if it is a convTransLayer, false if it is a convLayer + :param trans: True if it is a convTransLayer, False if it is a convLayer :type trans: bool - :param layer_type: specify the layer_type, default is None. If trans=True, - layer_type has to be "exconvt" or "cudnn_convt", - otherwise layer_type has to be either "exconv" or - "cudnn_conv" - :type layer_type: String + :param layer_type: Specify the layer type. If the dilation's dimension on one axis is + larger than 1, layer_type has to be "cudnn_conv" or "cudnn_convt". + If trans=True, layer_type has to be "exconvt" or "cudnn_convt", + otherwise layer_type has to be either "exconv" or "cudnn_conv". + :type layer_type: basestring :return: LayerOutput object. :rtype: LayerOutput """ @@ -2675,11 +2727,12 @@ def img_pool_layer(input, pool_size_y=None, stride_y=None, padding_y=None, - ceil_mode=True): + ceil_mode=True, + exclude_mode=None): """ Image pooling Layer. - The details of pooling layer, please refer ufldl's pooling_ . + The details of pooling layer, please refer to ufldl's pooling_ . .. _pooling: http://ufldl.stanford.edu/tutorial/supervised/Pooling/ @@ -2687,15 +2740,17 @@ def img_pool_layer(input, .. math:: - w = 1 + int(ceil(input\_width + 2 * padding - pool\_size) / float(stride)) - h = 1 + int(ceil(input\_height + 2 * padding\_y - pool\_size\_y) / float(stride\_y)) + w & = 1 + \\frac{ceil(input\_width + 2 * padding - pool\_size)}{stride} + + h & = 1 + \\frac{ceil(input\_height + 2 * padding\_y - pool\_size\_y)}{stride\_y} - ceil_mode=False: .. math:: - w = 1 + int(floor(input\_width + 2 * padding - pool\_size) / float(stride)) - h = 1 + int(floor(input\_height + 2 * padding\_y - pool\_size\_y) / float(stride\_y)) + w & = 1 + \\frac{floor(input\_width + 2 * padding - pool\_size)}{stride} + + h & = 1 + \\frac{floor(input\_height + 2 * padding\_y - pool\_size\_y)}{stride\_y} The example usage is: @@ -2711,33 +2766,43 @@ def img_pool_layer(input, padding_y=2, pool_type=MaxPooling()) - :param padding: pooling padding width. + :param padding: The padding size on the x axis. 0 is the default padding size. :type padding: int - :param padding_y: pooling padding height. It's equal to padding by default. - :type padding_y: int | None - :param name: name of pooling layer - :type name: basestring. + :param padding_y: The padding size on the y axis. If the parameter is not set + or set to None, it will be set to 'padding' automatically. + :param name: The name of this layer. It is optional. + :type name: basestring :param input: The input of this layer. :type input: LayerOutput - :param pool_size: pooling window width + :param pool_size: The pooling window length on the x axis. :type pool_size: int - :param pool_size_y: pooling window height. It's eaqual to pool_size by default. - :type pool_size_y: int | None - :param num_channels: number of input channel. + :param pool_size_y: The pooling window length on the y axis. If the parameter is + not set or set to None, its actual value will be automatically + set to pool_size. + :type pool_size_y: int + :param num_channels: The number of input channels. If the parameter is not set or + set to None, its actual value will be automatically set to + the channels number of the input. :type num_channels: int - :param pool_type: pooling type. MaxPooling or AvgPooling. Default is - MaxPooling. + :param pool_type: Pooling type. MaxPooling is the default pooling. :type pool_type: BasePoolingType - :param stride: stride width of pooling. + :param stride: The stride on the x axis. 1 is the default value. :type stride: int - :param stride_y: stride height of pooling. It is equal to stride by default. - :type stride_y: int | None - :param layer_attr: Extra Layer attribute. + :param stride_y: The stride on the y axis. If the parameter is not set or set to + None, its actual value will be automatically set to 'stride'. + :type stride_y: int + :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for + details. :type layer_attr: ExtraLayerAttribute - :param ceil_mode: Wether to use ceil mode to calculate output height and with. - Defalut is True. If set false, Otherwise use floor. - + :param ceil_mode: Whether to use the ceil function to calculate output height and width. + True is the default. If it is set to False, the floor function will + be used. :type ceil_mode: bool + :param exclude_mode: Whether to exclude the padding cells when calculating, but only + work when pool_type is AvgPooling. If None, also exclude the padding + cells. If use cudnn, use CudnnAvgPooling or CudnnAvgInclPadPooling + as pool_type to identify the mode. + :type exclude_mode: bool :return: LayerOutput object. :rtype: LayerOutput """ @@ -2751,7 +2816,7 @@ def img_pool_layer(input, pool_type.name = 'avg' assert type(pool_type) in [AvgPooling, MaxPooling, MaxWithMaskPooling, CudnnAvgPooling, - CudnnMaxPooling], \ + CudnnMaxPooling, CudnnAvgInclPadPooling], \ "only (Cudnn)AvgPooling, (Cudnn)MaxPooling, MaxWithMaskPooling are supported" type_name = pool_type.name + '-projection' \ @@ -2780,6 +2845,7 @@ def img_pool_layer(input, padding_y=padding_y)) ], ceil_mode=ceil_mode, + exclude_mode=exclude_mode, **ExtraLayerAttribute.to_kwargs(layer_attr)) return LayerOutput( name, @@ -2817,17 +2883,21 @@ def img_pool3d_layer(input, .. math:: - w = 1 + int(ceil(input\_width + 2 * padding - pool\_size) / float(stride)) - h = 1 + int(ceil(input\_height + 2 * padding\_y - pool\_size\_y) / float(stride\_y)) - d = 1 + int(ceil(input\_depth + 2 * padding\_z - pool\_size\_z) / float(stride\_z)) + w & = 1 + \\frac{ceil(input\_width + 2 * padding - pool\_size)}{stride} + + h & = 1 + \\frac{ceil(input\_height + 2 * padding\_y - pool\_size\_y)}{stride\_y} + + d & = 1 + \\frac{ceil(input\_depth + 2 * padding\_z - pool\_size\_z)}{stride\_z} - ceil_mode=False: .. math:: - w = 1 + int(floor(input\_width + 2 * padding - pool\_size) / float(stride)) - h = 1 + int(floor(input\_height + 2 * padding\_y - pool\_size\_y) / float(stride\_y)) - d = 1 + int(floor(input\_depth + 2 * padding\_z - pool\_size\_z) / float(stride\_z)) + w & = 1 + \\frac{floor(input\_width + 2 * padding - pool\_size)}{stride} + + h & = 1 + \\frac{floor(input\_height + 2 * padding\_y - pool\_size\_y)}{stride\_y} + + d & = 1 + \\frac{floor(input\_depth + 2 * padding\_z - pool\_size\_z)}{stride\_z} The example usage is: @@ -2842,24 +2912,32 @@ def img_pool3d_layer(input, :param padding: pooling padding width. :type padding: int | tuple | list - :param name: name of pooling layer + :param name: The name of this layer. It is optional. :type name: basestring. :param input: The input of this layer. :type input: LayerOutput - :param pool_size: pooling window width + :param pool_size: The pooling window lengths along three axises. If the parameter + is set to one integer, the three lengths will be same. :type pool_size: int | tuple | list - :param num_channels: number of input channel. + :param num_channels: The number of input channels. If the parameter is not set or + set to None, its actual value will be automatically set to + the channels number of the input. :type num_channels: int - :param pool_type: pooling type. MaxPooling or AvgPooling. Default is - MaxPooling. + :param pool_type: Pooling type. MaxPooling is the default pooling. :type pool_type: BasePoolingType - :param stride: stride width of pooling. + :param stride: The strides of the pooling along three axises. If the parameter + is set to one integer, the three strides will be same. 1 is the + default value. :type stride: int | tuple | list - :param layer_attr: Extra Layer attribute. + :param padding: The sizes of padding along three axises. If the parameter is set to + one integer, they will be same. 0 is the default padding size. + :type padding: int | tuple | list + :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for + details. :type layer_attr: ExtraLayerAttribute - :param ceil_mode: Wether to use ceil mode to calculate output height and with. - Defalut is True. If set false, Otherwise use floor. - + :param ceil_mode: Wether to use the ceil function to calculate output height and width. + True is the default. If it is set to False, the floor function will + be used. :type ceil_mode: bool :return: LayerOutput object. :rtype: LayerOutput @@ -2938,9 +3016,11 @@ def spp_layer(input, pyramid_height=None, layer_attr=None): """ - Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition. - The details please refer to - `Kaiming He's paper `_. + A layer performs spatial pyramid pooling. + + Reference: + `Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition + `_ The example usage is: @@ -2955,13 +3035,16 @@ def spp_layer(input, :type name: basestring :param input: The input of this layer. :type input: LayerOutput - :param num_channels: number of input channel. + :param num_channels: The number of input channels. If the parameter is not set or + set to None, its actual value will be automatically set to + the channels number of the input. :type num_channels: int - :param pool_type: Pooling type. MaxPooling or AveragePooling. Default is MaxPooling. + :param pool_type: Pooling type. MaxPooling is the default pooling. :type scale: BasePoolingType - :param pyramid_height: pyramid height. + :param pyramid_height: The pyramid height of this pooling. :type pyramid_height: int - :param layer_attr: Extra Layer Attribute. + :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for + details. :type layer_attr: ExtraLayerAttribute :return: LayerOutput object. :rtype: LayerOutput @@ -3038,8 +3121,8 @@ def img_cmrnorm_layer(input, Response normalization across feature maps. Reference: - ImageNet Classification with Deep Convolutional Neural Networks - http://www.cs.toronto.edu/~fritz/absps/imagenet.pdf + `ImageNet Classification with Deep Convolutional Neural Networks + `_ The example usage is: @@ -3085,6 +3168,7 @@ def batch_norm_layer(input, param_attr=None, layer_attr=None, batch_norm_type=None, + epsilon=1e-5, moving_average_fraction=0.9, use_global_stats=None, mean_var_names=None): @@ -3104,9 +3188,9 @@ def batch_norm_layer(input, y_i &\\gets \\gamma \\hat{x_i} + \\beta \\qquad &//\ scale\ and\ shift Reference: - Batch Normalization: Accelerating Deep Network Training by Reducing + `Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift - http://arxiv.org/abs/1502.03167 + `_ The example usage is: @@ -3155,6 +3239,8 @@ def batch_norm_layer(input, will use the mean and variance of the current batch of test data. :type use_global_stats: bool | None. + :param epsilon: The small constant added to the variance to improve numeric stability. + :type epsilon: float. :param moving_average_fraction: Factor used in the moving average computation. :math:`runningMean = newMean*(1-factor) + runningMean*factor` :type moving_average_fraction: float. @@ -3172,6 +3258,7 @@ def batch_norm_layer(input, assert (batch_norm_type is None) or (batch_norm_type == "batch_norm") or \ (batch_norm_type == "mkldnn_batch_norm") or \ (batch_norm_type == "cudnn_batch_norm") + l = Layer( name=name, img3D=img3D, @@ -3181,6 +3268,7 @@ def batch_norm_layer(input, type=LayerType.BATCH_NORM_LAYER, batch_norm_type=batch_norm_type, bias=ParamAttr.to_bias(bias_attr), + epsilon=epsilon, moving_average_fraction=moving_average_fraction, use_global_stats=use_global_stats, mean_var_names=mean_var_names, @@ -3240,7 +3328,7 @@ def row_l2_norm_layer(input, name=None, layer_attr=None): A layer for L2-normalization in each row. .. math:: - out[i] = \frac{in[i]}{\sqrt{\sum_{k=1}^N in[k]^{2}}} + out[i] = \\frac{in[i]} {\\sqrt{\\sum_{k=1}^N in[k]^{2}}} where the size of :math:`in` is (batchSize x dataDim) , and the size of :math:`out` is a (batchSize x dataDim) . @@ -4203,7 +4291,7 @@ def dot_prod_layer(input1, input2, name=None, layer_attr=None): :param name: The name of this layer. It is optional. :type name: basestring :param input1: The first input layer. - :type input: LayerOutput + :type input1: LayerOutput :param input2: The second input layer. :type input2: LayerOutput :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for @@ -4694,7 +4782,7 @@ def conv_projection(input, will be same when filter_size_y is not set. If it is set to a list, the first element indicates the dimension on the x axis, and the second is used to specify the dimension - on the y axis when filter_size is not provided. + on the y axis when filter_size_y is not provided. :type filter_size: int | tuple | list :param filter_size_y: The dimension of the filter kernel on the y axis. If the parameter is not set, it will be set automatically according to filter_size. @@ -5359,18 +5447,28 @@ def maxout_layer(input, groups, num_channels=None, name=None, layer_attr=None): to be devided by groups. Reference: - Maxout Networks - http://www.jmlr.org/proceedings/papers/v28/goodfellow13.pdf - Multi-digit Number Recognition from Street View Imagery using Deep Convolutional Neural Networks - https://arxiv.org/pdf/1312.6082v4.pdf + `Maxout Networks + `_ + `Multi-digit Number Recognition from Street View Imagery using Deep Convolutional Neural Networks + `_ + .. math:: - y_{si+j} = \max_k x_{gsi + sk + j} - g = groups - s = input.size / num_channels - 0 \le i < num_channels / groups - 0 \le j < s - 0 \le k < groups + + & out = \max_k (in[n, k, o_c , s]) + + & out_{i * s + j} = \max_k in_{ k * o_{c} * s + i * s + j} + + & s = \\frac{input.size}{ num\_channels} + + & o_{c} = \\frac{num\_channels}{groups} + + & 0 \le i < o_{c} + + & 0 \le j < s + + & 0 \le k < groups + The simple usage is: @@ -5427,9 +5525,9 @@ def ctc_layer(input, alignment between the inputs and the target labels is unknown. Reference: - Connectionist Temporal Classification: Labelling Unsegmented Sequence Data + `Connectionist Temporal Classification: Labelling Unsegmented Sequence Data with Recurrent Neural Networks - http://machinelearning.wustl.edu/mlpapers/paper_files/icml2006_GravesFGS06.pdf + `_ Note: Considering the 'blank' label needed by CTC, you need to use (num_classes + 1) @@ -5501,9 +5599,9 @@ def warp_ctc_layer(input, install it to :code:`third_party/install/warpctc` directory. Reference: - Connectionist Temporal Classification: Labelling Unsegmented Sequence Data + `Connectionist Temporal Classification: Labelling Unsegmented Sequence Data with Recurrent Neural Networks - http://machinelearning.wustl.edu/mlpapers/paper_files/icml2006_GravesFGS06.pdf + `_ Note: - Let num_classes represents the category number. Considering the 'blank' @@ -5723,8 +5821,8 @@ def nce_layer(input, Noise-contrastive estimation. Reference: - A fast and simple algorithm for training neural probabilistic language - models. https://www.cs.toronto.edu/~amnih/papers/ncelm.pdf + `A fast and simple algorithm for training neural probabilistic language + models. `_ The example usage is: @@ -5839,8 +5937,8 @@ def rank_cost(left, A cost Layer for learning to rank using gradient descent. Reference: - Learning to Rank using Gradient Descent - http://research.microsoft.com/en-us/um/people/cburges/papers/ICML_ranking.pdf + `Learning to Rank using Gradient Descent + `_ .. math:: @@ -6107,9 +6205,11 @@ def huber_regression_cost(input, Given a prediction f(x), a label y and :math:`\delta`, the loss function is defined as: - .. math: - loss = 0.5*\left ( y-f(x) \right )^2, \left | y-f(x) \right |\leq \delta - loss = \delta \left | y-f(x) \right |-0.5\delta ^2, otherwise + .. math:: + + loss = 0.5*(y-f(x))^{2}, | y-f(x) | < \delta + + loss = \delta | y-f(x) | - 0.5 \delta ^2, otherwise The example usage is: @@ -6156,12 +6256,14 @@ def huber_classification_cost(input, """ For classification purposes, a variant of the Huber loss called modified Huber is sometimes used. Given a prediction f(x) (a real-valued classifier score) and - a true binary class label :math:`y\in \left \{-1, 1 \right \}`, the modified Huber + a true binary class label :math:`y\in \{-1, 1 \}`, the modified Huber loss is defined as: .. math: - loss = \max \left ( 0, 1-yf(x) \right )^2, yf(x)\geq 1 - loss = -4yf(x), \text{otherwise} + + loss = \max ( 0, 1-yf(x) )^2, yf(x) \geq -1 + + loss = -4yf(x), otherwise The example usage is: @@ -6375,8 +6477,8 @@ def smooth_l1_cost(input, label, name=None, coeff=1.0, layer_attr=None): smooth_{L1}(x) = \\begin{cases} 0.5x^2& \\text{if} \\ |x| < 1 \\\\ |x|-0.5& \\text{otherwise} \end{cases} Reference: - Fast R-CNN - https://arxiv.org/pdf/1504.08083v2.pdf + `Fast R-CNN + `_ The example usage is: @@ -6527,7 +6629,7 @@ def row_conv_layer(input, .. math:: r_{t,r} = \sum_{j=1}^{k + 1} {w_{i,j}h_{t+j-1, i}} - \quad \text{for} \quad (1 \leq i \leq d) + \quad \\text{for} \quad (1 \leq i \leq d) Note: The `context_len` is `k + 1`. That is to say, the lookahead step @@ -6571,18 +6673,19 @@ def row_conv_layer(input, @layer_support() @wrap_name_default() -@wrap_param_attr_default() def prelu_layer(input, name=None, partial_sum=1, + channel_shared=None, + num_channels=None, param_attr=None, layer_attr=None): """ The Parametric Relu activation that actives outputs with a learnable weight. Reference: - Delving Deep into Rectifiers: Surpassing Human-Level Performance on - ImageNet Classification http://arxiv.org/pdf/1502.01852v1.pdf + `Delving Deep into Rectifiers: Surpassing Human-Level Performance on + ImageNet Classification `_ .. math:: z_i &\\quad if \\quad z_i > 0 \\\\ @@ -6605,6 +6708,14 @@ def prelu_layer(input, - partial_sum = number of outputs, indicates all elements share the same weight. :type partial_sum: int + :param channel_shared: whether or not the parameter are shared across channels. + + - channel_shared = True, we set the partial_sum to the number of outputs. + - channel_shared = False, we set the partial_sum to the number of elements in one channel. + + :type channel_shared: bool + :param num_channels: number of input channel. + :type num_channels: int :param param_attr: The parameter attribute. See ParameterAttribute for details. :type param_attr: ParameterAttribute :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for @@ -6615,7 +6726,25 @@ def prelu_layer(input, """ assert isinstance(input, LayerOutput), 'prelu_layer accepts only one input.' - assert isinstance(param_attr, ParameterAttribute) + + if not param_attr: + param_attr = ParamAttr(initial_mean=0.25, initial_std=0.0) + else: + assert isinstance(param_attr, ParameterAttribute) + + if num_channels is None: + assert input.num_filters is not None, \ + 'the input channel cannot be detected, please specify the num_channels parameter' + num_channels = input.num_filters + + if channel_shared is not None: + assert isinstance(channel_shared, bool) + assert (input.height != 0 and input.width != 0), \ + 'input height and widht must be setted' + if channel_shared: + partial_sum = input.height * input.width * num_channels + else: + partial_sum = input.height * input.width l = Layer( name=name, @@ -6627,6 +6756,7 @@ def prelu_layer(input, name=name, layer_type=LayerType.PRELU, parents=input, + num_filters=num_channels, size=l.config.size) @@ -6648,11 +6778,11 @@ def gated_unit_layer(input, The gated unit layer implements a simple gating mechanism over the input. The input :math:`X` is first projected into a new space :math:`X'`, and it is also used to produce a gate weight :math:`\sigma`. Element-wise - product between :match:`X'` and :math:`\sigma` is finally returned. + product between :math:`X'` and :math:`\sigma` is finally returned. Reference: - Language Modeling with Gated Convolutional Networks - https://arxiv.org/abs/1612.08083 + `Language Modeling with Gated Convolutional Networks + `_ .. math:: y=\\text{act}(X \cdot W + b)\otimes \sigma(X \cdot V + c) @@ -6788,6 +6918,7 @@ def crop_layer(input, offset, axis=2, shape=None, name=None, layer_attr=None): :param input: The input of this layer. If two inputs are given, the second one will be regarded as the reference. + And the input must be 4-dims and in NCHW order. :type input: LayerOutput | Sequence :param offset: The crop offset. :type offset: Sequence @@ -6877,7 +7008,7 @@ def clip_layer(input, min, max, name=None): .. math:: - out[i] = \min\left(\max\left(in[i],p_{1}\right),p_{2}\right) + out[i] = \min (\max (in[i],p_{1} ),p_{2} ) .. code-block:: python @@ -7076,7 +7207,7 @@ def img_conv3d_layer(input, :type layer_attr: ExtraLayerAttribute :param trans: True if it is a convTransLayer, False if it is a convLayer :type trans: bool - :param layer_type: Specify the layer_type. If the parameter is set, it must be "deconv3d" + :param layer_type: Specify the layer type. If the parameter is set, it must be "deconv3d" when trans=True. If not set, it will be automatically set to "deconv3d" when trans=True and "conv3d" when trans=False. :type layer_type: basestring @@ -7321,3 +7452,73 @@ def scale_sub_region_layer(input, indices, value, name=None): parents=[input, indices], num_filters=input.num_filters, size=input.size) + + +@wrap_name_default() +@wrap_act_default(act=LinearActivation()) +@wrap_param_attr_default() +@layer_support() +def factorization_machine(input, + factor_size, + act=None, + name=None, + param_attr=None, + layer_attr=None): + """ + The Factorization Machine models pairwise feature interactions as inner + product of the learned latent vectors corresponding to each input feature. + The Factorization Machine can effectively capture feature interactions + especially when the input is sparse. + + This implementation only consider the 2-order feature interactions using + Factorization Machine with the formula: + + .. math:: + y = \sum_{i=1}^{n-1}\sum_{j=i+1}^n\langle v_i, v_j \\rangle x_i x_j + + Note: + X is the input vector with size n. V is the factor matrix. Each row of V + is the latent vector corresponding to each input dimesion. The size of + each latent vector is k. + + For details of Factorization Machine, please refer to the paper: + Factorization machines. + + .. code-block:: python + first_order = paddle.layer.fc(input=input, + size=1, + act=paddle.activation.Linear()) + second_order = paddle.layer.factorization_machine(input=input, + factor_size=10) + fm = paddle.layer.addto(input=[first_order, second_order], + act=paddle.activation.Linear(), + bias_attr=False) + + :param input: The input layer. Supported input types: all input data types + on CPU, and only dense input types on GPU. + :type input: LayerOutput + :param factor_size: The hyperparameter that defines the dimensionality of + the latent vector size. + :type context_len: int + :param act: Activation Type. Default is linear activation. + :type act: BaseActivation + :param param_attr: The parameter attribute. See ParameterAttribute for + details. + :type param_attr: ParameterAttribute + :param layer_attr: Extra Layer config. + :type layer_attr: ExtraLayerAttribute|None + :return: LayerOutput object. + :rtype: LayerOutput + """ + assert isinstance(input, LayerOutput) + assert factor_size > 0, "the factor_size must be greater than 0." + + Layer( + inputs=[Input(input.name, **param_attr.attr)], + name=name, + factor_size=factor_size, + type=LayerType.FACTORIZATION_MACHINE, + active_type=act.name, + **ExtraLayerAttribute.to_kwargs(layer_attr)) + return LayerOutput( + name, LayerType.FACTORIZATION_MACHINE, input, activation=act, size=1) diff --git a/python/paddle/trainer_config_helpers/networks.py b/python/paddle/trainer_config_helpers/networks.py index 9776ae18057d57dd994fac8b62090258252922c6..b5cde7bac779ee1d54395b68941df2693e1ed0f5 100644 --- a/python/paddle/trainer_config_helpers/networks.py +++ b/python/paddle/trainer_config_helpers/networks.py @@ -25,10 +25,10 @@ from paddle.trainer.config_parser import * __all__ = [ 'sequence_conv_pool', 'simple_lstm', "simple_img_conv_pool", "img_conv_bn_pool", 'lstmemory_group', 'lstmemory_unit', 'small_vgg', - 'img_conv_group', 'vgg_16_network', 'gru_unit', 'gru_group', 'simple_gru', - 'simple_attention', 'dot_product_attention', 'multi_head_attention', - 'simple_gru2', 'bidirectional_gru', 'text_conv_pool', 'bidirectional_lstm', - 'inputs', 'outputs' + 'img_conv_group', 'img_separable_conv', 'vgg_16_network', 'gru_unit', + 'gru_group', 'simple_gru', 'simple_attention', 'dot_product_attention', + 'multi_head_attention', 'simple_gru2', 'bidirectional_gru', + 'text_conv_pool', 'bidirectional_lstm', 'inputs', 'outputs' ] ###################################################### @@ -251,13 +251,13 @@ def img_conv_bn_pool(input, pool_layer_attr=None): """ Convolution, batch normalization, pooling group. - + Img input => Conv => BN => Pooling => Output. :param name: group name. :type name: basestring :param input: input layer. - :type input: LayerOutput + :type input: LayerOutput :param filter_size: see img_conv_layer for details. :type filter_size: int :param num_filters: see img_conv_layer for details. @@ -435,6 +435,85 @@ def img_conv_group(input, input=tmp, stride=pool_stride, pool_size=pool_size, pool_type=pool_type) +@wrap_name_default("separable_conv") +def img_separable_conv(input, + num_channels, + num_out_channels, + filter_size, + stride=1, + padding=0, + depth_multiplier=1, + act=None, + bias_attr=None, + param_attr=None, + shared_bias=True, + layer_type='exconv', + name=None): + """ + Separable Convolution. + + The separable convolution module is consisted of a depthwise convolution + that acts separately on input channels, followed by a pointwise convolution + with 1*1 kernels that mixes channels. It is used for Xception: + https://arxiv.org/pdf/1610.02357.pdf + + :param input: input layer. + :type input: LayerOutput + :param num_channels: the number of input channels. + :type num_channels: int + :param num_out_channels: the number of output channels. + :type num_out_channels: int + :param filter_size: the filter size for the depthwise convolution. + :type filter_size: int|tuple + :param stride: the stride size for the depthwise convolution. + :type stride: int|tuple + :param padding: the padding size for the depthwise convolution. + :type padding: int|tuple + :param depth_multiplier: the number of filter for one channel in the + depthwize convolution. + :type depth_multiplier: int + :param act: the activation function for the output. + :type act: BaseActivation + :param bias_attr: see img_conv_layer for details. + :type bias_attr: ParameterAttribute + :param param_attr: see img_conv_layer for details. + :type param_attr: ParameterAttribute + :param shared_bias: see img_conv_layer for details. + :type shared_bias: bool + :param layer_type: see img_conv_layer for details. + :type layer_type: bool + :return: layer's output + :rtype: LayerOutput + """ + __depthwise_conv__ = img_conv_layer( + name="%s_depthwise_conv" % name, + input=input, + num_channels=num_channels, + num_filters=num_channels * depth_multiplier, + groups=num_channels, + filter_size=filter_size, + stride=stride, + padding=padding, + act=LinearActivation(), + bias_attr=bias_attr, + param_attr=param_attr, + shared_biases=shared_bias, + layer_type=layer_type) + __pointwise_conv__ = img_conv_layer( + name="%s_pointwise_conv" % name, + input=__depthwise_conv__, + num_channels=num_channels * depth_multiplier, + num_filters=num_out_channels, + filter_size=1, + stride=1, + padding=0, + act=act, + bias_attr=bias_attr, + param_attr=param_attr, + shared_biases=shared_bias) + return __pointwise_conv__ + + def small_vgg(input_image, num_channels, num_classes): def __vgg__(ipt, num_filter, times, dropouts, num_channels_=None): return img_conv_group( @@ -648,7 +727,7 @@ def lstmemory_unit(input, lstm_bias_attr=None, lstm_layer_attr=None): """ - lstmemory_unit defines the caculation process of a LSTM unit during a + lstmemory_unit defines the caculation process of a LSTM unit during a single time step. This function is not a recurrent layer, so it can not be directly used to process sequence input. This function is always used in recurrent_group (see layers.py for more details) to implement attention @@ -869,7 +948,7 @@ def gru_unit(input, gru_layer_attr=None, naive=False): """ - gru_unit defines the calculation process of a gated recurrent unit during a single + gru_unit defines the calculation process of a gated recurrent unit during a single time step. This function is not a recurrent layer, so it can not be directly used to process sequence input. This function is always used in the recurrent_group (see layers.py for more details) to implement attention @@ -1012,7 +1091,7 @@ def simple_gru(input, simple_gru in network.py. The reason why there are so many interfaces is that we have two ways to implement recurrent neural network. One way is to use one complete layer to implement rnn (including simple rnn, gru and lstm) - with multiple time steps, such as recurrent_layer, lstmemory, grumemory. But + with multiple time steps, such as recurrent_layer, lstmemory, grumemory. But the multiplication operation :math:`W x_t` is not computed in these layers. See details in their interfaces in layers.py. The other implementation is to use an recurrent group which can ensemble a @@ -1116,11 +1195,12 @@ def simple_gru2(input, :type act: BaseActivation :param gate_act: gate activiation type of gru :type gate_act: BaseActivation - :param gru_bias_attr: bias parameter attribute of gru layer, + :param gru_bias_attr: bias parameter attribute of gru layer, False means no bias, None means default bias. :type gru_bias_attr: ParameterAttribute|False|None - :param gru_layer_attr: Extra attribute of the gru layer. - :type gru_layer_attr: ExtraLayerAttribute + :param gru_param_attr: param parameter attribute of gru layer, + None means default param. + :type gru_param_attr: ParameterAttribute|None :return: the gru group. :rtype: LayerOutput """ @@ -1188,7 +1268,7 @@ def bidirectional_gru(input, :type size: int :param return_seq: If set False, the last time step of output are concatenated and returned. - If set True, the entire output sequences in forward + If set True, the entire output sequences in forward and backward directions are concatenated and returned. :type return_seq: bool :return: LayerOutput object. @@ -1277,7 +1357,7 @@ def bidirectional_lstm(input, :type size: int :param return_seq: If set False, the last time step of output are concatenated and returned. - If set True, the entire output sequences in forward + If set True, the entire output sequences in forward and backward directions are concatenated and returned. :type return_seq: bool :return: LayerOutput object. diff --git a/python/paddle/trainer_config_helpers/poolings.py b/python/paddle/trainer_config_helpers/poolings.py index f45616551bcd4822c668234c3afaf6aa35cd2953..e0aeb311b3ae842aee337dbbf869e2f947d22bd9 100644 --- a/python/paddle/trainer_config_helpers/poolings.py +++ b/python/paddle/trainer_config_helpers/poolings.py @@ -16,7 +16,8 @@ __all__ = [ "BasePoolingType", "MaxPooling", "AvgPooling", "MaxWithMaskPooling", - "CudnnMaxPooling", "CudnnAvgPooling", "SumPooling", "SquareRootNPooling" + "CudnnMaxPooling", "CudnnAvgPooling", "CudnnAvgInclPadPooling", + "SumPooling", "SquareRootNPooling" ] @@ -88,6 +89,16 @@ class CudnnAvgPooling(BasePoolingType): BasePoolingType.__init__(self, "cudnn-avg-pool") +class CudnnAvgInclPadPooling(BasePoolingType): + """ + Cudnn average pooling only support GPU. Return the average value in the + pooling window taking into account the padding cells. + """ + + def __init__(self): + BasePoolingType.__init__(self, "cudnn-avg-incl-pad-pool") + + class AvgPooling(BasePoolingType): """ Average pooling. diff --git a/python/paddle/trainer_config_helpers/tests/configs/file_list.sh b/python/paddle/trainer_config_helpers/tests/configs/file_list.sh index a21f67a2d99e7eab39708e2a571d30d7e9f20ce6..10c941f707498ec45e79bed9d3f8054eea19887d 100755 --- a/python/paddle/trainer_config_helpers/tests/configs/file_list.sh +++ b/python/paddle/trainer_config_helpers/tests/configs/file_list.sh @@ -11,6 +11,7 @@ test_recursive_topology test_gated_unit_layer test_clip_layer test_row_l2_norm_l test_kmax_seq_socre_layer test_sub_nested_seq_select_layer test_scale_shift_layer test_seq_slice_layer test_cross_entropy_over_beam test_roi_pool_layer test_pooling3D_layer test_conv3d_layer test_deconv3d_layer test_BatchNorm3D test_resize_layer -test_scale_sub_region_layer test_dot_prod_layer test_l2_distance_layer) +test_scale_sub_region_layer test_dot_prod_layer test_l2_distance_layer +test_factorization_machine) export whole_configs=(test_split_datasource) diff --git a/python/paddle/trainer_config_helpers/tests/configs/protostr/img_layers.protostr b/python/paddle/trainer_config_helpers/tests/configs/protostr/img_layers.protostr index b14121e82cb7d9516c4771fc896b9b3b9e01d1c8..3e0f957648879d4350d662b336c953273bac1378 100644 --- a/python/paddle/trainer_config_helpers/tests/configs/protostr/img_layers.protostr +++ b/python/paddle/trainer_config_helpers/tests/configs/protostr/img_layers.protostr @@ -65,6 +65,7 @@ layers { height: 227 width: 227 depth: 1 + epsilon: 1e-05 } layers { name: "__crmnorm_0__" diff --git a/python/paddle/trainer_config_helpers/tests/configs/protostr/img_trans_layers.protostr b/python/paddle/trainer_config_helpers/tests/configs/protostr/img_trans_layers.protostr index c7a487a11231cba6182b654108773037bdb0ec35..a18a4652e14c0cfc4dbca87e67d31aa663ee756b 100644 --- a/python/paddle/trainer_config_helpers/tests/configs/protostr/img_trans_layers.protostr +++ b/python/paddle/trainer_config_helpers/tests/configs/protostr/img_trans_layers.protostr @@ -65,6 +65,7 @@ layers { height: 256 width: 256 depth: 1 + epsilon: 1e-05 } layers { name: "__crmnorm_0__" diff --git a/python/paddle/trainer_config_helpers/tests/configs/protostr/test_BatchNorm3D.protostr b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_BatchNorm3D.protostr index 832ed24a31dd2bedba9a4fce77d7a088d1796fdb..9b69ae4a3b3cbcc7c0c69a2d5b3728e2f0204f33 100644 --- a/python/paddle/trainer_config_helpers/tests/configs/protostr/test_BatchNorm3D.protostr +++ b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_BatchNorm3D.protostr @@ -36,6 +36,7 @@ layers { height: 6 width: 20 depth: 3 + epsilon: 1e-05 } parameters { name: "___batch_norm_0__.w0" diff --git a/python/paddle/trainer_config_helpers/tests/configs/protostr/test_factorization_machine.protostr b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_factorization_machine.protostr new file mode 100644 index 0000000000000000000000000000000000000000..4f3002b19942ed58970bfd64e5978c1601273992 --- /dev/null +++ b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_factorization_machine.protostr @@ -0,0 +1,39 @@ +type: "nn" +layers { + name: "data" + type: "data" + size: 1024 + active_type: "" +} +layers { + name: "__factorization_machine_0__" + type: "factorization_machine" + size: 1 + active_type: "" + inputs { + input_layer_name: "data" + input_parameter_name: "___factorization_machine_0__.w0" + } + factor_size: 10 +} +parameters { + name: "___factorization_machine_0__.w0" + size: 10240 + initial_mean: 0.0 + initial_std: 0.03125 + dims: 1024 + dims: 10 + initial_strategy: 0 + initial_smart: true +} +input_layer_names: "data" +output_layer_names: "__factorization_machine_0__" +sub_models { + name: "root" + layer_names: "data" + layer_names: "__factorization_machine_0__" + input_layer_names: "data" + output_layer_names: "__factorization_machine_0__" + is_recurrent_layer_group: false +} + diff --git a/python/paddle/trainer_config_helpers/tests/configs/protostr/test_prelu_layer.protostr b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_prelu_layer.protostr index 94ad56cab063df9e6a11bb1c293727fb9dec810f..63fb38c6508675d379f577b965ea17ad4c3b4942 100644 --- a/python/paddle/trainer_config_helpers/tests/configs/protostr/test_prelu_layer.protostr +++ b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_prelu_layer.protostr @@ -4,6 +4,8 @@ layers { type: "data" size: 300 active_type: "" + height: 10 + width: 10 } layers { name: "__prelu_layer_0__" @@ -15,6 +17,9 @@ layers { input_parameter_name: "___prelu_layer_0__.w0" } partial_sum: 1 + height: 10 + width: 10 + depth: 1 } layers { name: "__prelu_layer_1__" @@ -26,6 +31,9 @@ layers { input_parameter_name: "___prelu_layer_1__.w0" } partial_sum: 1 + height: 10 + width: 10 + depth: 1 } layers { name: "__prelu_layer_2__" @@ -37,41 +45,100 @@ layers { input_parameter_name: "___prelu_layer_2__.w0" } partial_sum: 5 + height: 10 + width: 10 + depth: 1 +} +layers { + name: "__prelu_layer_3__" + type: "prelu" + size: 300 + active_type: "" + inputs { + input_layer_name: "input" + input_parameter_name: "___prelu_layer_3__.w0" + } + partial_sum: 300 + height: 10 + width: 10 + depth: 1 +} +layers { + name: "__prelu_layer_4__" + type: "prelu" + size: 300 + active_type: "" + inputs { + input_layer_name: "input" + input_parameter_name: "___prelu_layer_4__.w0" + } + partial_sum: 100 + height: 10 + width: 10 + depth: 1 } parameters { name: "___prelu_layer_0__.w0" size: 300 - initial_mean: 0.0 - initial_std: 0.057735026919 + initial_mean: 0.25 + initial_std: 0.0 + dims: 1 + dims: 300 initial_strategy: 0 - initial_smart: true + initial_smart: false } parameters { name: "___prelu_layer_1__.w0" size: 300 - initial_mean: 0.0 - initial_std: 0.057735026919 + initial_mean: 0.25 + initial_std: 0.0 + dims: 1 + dims: 300 initial_strategy: 0 - initial_smart: true + initial_smart: false } parameters { name: "___prelu_layer_2__.w0" size: 60 - initial_mean: 0.0 - initial_std: 0.129099444874 + initial_mean: 0.25 + initial_std: 0.0 + dims: 1 + dims: 60 + initial_strategy: 0 + initial_smart: false +} +parameters { + name: "___prelu_layer_3__.w0" + size: 1 + initial_mean: 0.25 + initial_std: 0.0 + dims: 1 + dims: 1 + initial_strategy: 0 + initial_smart: false +} +parameters { + name: "___prelu_layer_4__.w0" + size: 3 + initial_mean: 0.25 + initial_std: 0.0 + dims: 1 + dims: 3 initial_strategy: 0 - initial_smart: true + initial_smart: false } input_layer_names: "input" -output_layer_names: "__prelu_layer_2__" +output_layer_names: "__prelu_layer_4__" sub_models { name: "root" layer_names: "input" layer_names: "__prelu_layer_0__" layer_names: "__prelu_layer_1__" layer_names: "__prelu_layer_2__" + layer_names: "__prelu_layer_3__" + layer_names: "__prelu_layer_4__" input_layer_names: "input" - output_layer_names: "__prelu_layer_2__" + output_layer_names: "__prelu_layer_4__" is_recurrent_layer_group: false } diff --git a/python/paddle/trainer_config_helpers/tests/configs/test_factorization_machine.py b/python/paddle/trainer_config_helpers/tests/configs/test_factorization_machine.py new file mode 100644 index 0000000000000000000000000000000000000000..b249de0fee3c8ca4ad0520872fa2497c493d31b5 --- /dev/null +++ b/python/paddle/trainer_config_helpers/tests/configs/test_factorization_machine.py @@ -0,0 +1,7 @@ +from paddle.trainer_config_helpers import * + +data = data_layer(name='data', size=1024) + +fm = factorization_machine(input=data, factor_size=10) + +outputs(fm) diff --git a/python/paddle/trainer_config_helpers/tests/configs/test_prelu_layer.py b/python/paddle/trainer_config_helpers/tests/configs/test_prelu_layer.py index aae90fab32db78a70c2169ed8fafb930433f4136..45b02fbf325bb63b057bbbf64d59af8debf0bc9d 100644 --- a/python/paddle/trainer_config_helpers/tests/configs/test_prelu_layer.py +++ b/python/paddle/trainer_config_helpers/tests/configs/test_prelu_layer.py @@ -1,8 +1,10 @@ from paddle.trainer_config_helpers import * -data = data_layer(name='input', size=300) -prelu = prelu_layer(input=data) -prelu = prelu_layer(input=data, partial_sum=1) -prelu = prelu_layer(input=data, partial_sum=5) +data = data_layer(name='input', size=300, height=10, width=10) +prelu = prelu_layer(input=data, num_channels=3) +prelu = prelu_layer(input=data, partial_sum=1, num_channels=3) +prelu = prelu_layer(input=data, partial_sum=5, num_channels=3) +prelu = prelu_layer(input=data, channel_shared=True, num_channels=3) +prelu = prelu_layer(input=data, channel_shared=False, num_channels=3) outputs(prelu) diff --git a/python/paddle/v2/__init__.py b/python/paddle/v2/__init__.py index 7bbe3eaaa67a117bc53571e6571365c3a26814c1..70f61e84997efdbe3d6f268d249be8bac15b9ecd 100644 --- a/python/paddle/v2/__init__.py +++ b/python/paddle/v2/__init__.py @@ -62,21 +62,15 @@ __all__ = [ cp.begin_parse() -def init(**kwargs): - import py_paddle.swig_paddle as api - args = [] - args_dict = {} - # NOTE: append arguments if they are in ENV - for ek, ev in os.environ.iteritems(): - if ek.startswith("PADDLE_INIT_"): - args_dict[ek.replace("PADDLE_INIT_", "").lower()] = str(ev) +def set_omp_mkl_env_vars(trainer_count): + '''Auto set CPU environment if have not set before. + export KMP_AFFINITY, OMP_DYNAMIC according to the Hyper Threading status. + export OMP_NUM_THREADS, MKL_NUM_THREADS according to trainer_count. + ''' + import platform + if not platform.system() in ['Linux', 'Darwin']: + return - args_dict.update(kwargs) - # NOTE: overwrite arguments from ENV if it is in kwargs - for key in args_dict.keys(): - args.append('--%s=%s' % (key, str(args_dict[key]))) - - # auto set cpu environment def set_env(key, value): '''If the key has not been set in the environment, set it with value.''' assert isinstance(key, str) @@ -85,22 +79,58 @@ def init(**kwargs): if envset is None: os.environ[key] = value - ht = os.popen("lscpu |grep \"per core\"|awk -F':' '{print $2}'|xargs") - ht = int(ht.read()) - if ht == 1: # ht is off - set_env("OMP_DYNAMIC", "false") - set_env("KMP_AFFINITY", "granularity=fine,compact,0,0") - else: + def num_physical_cores(): + '''Get the number of physical cores''' + if platform.system() == "Linux": + num_sockets = int( + os.popen("grep 'physical id' /proc/cpuinfo | sort -u | wc -l") + .read()) + num_cores_per_socket = int( + os.popen("grep 'core id' /proc/cpuinfo | sort -u | wc -l") + .read()) + return num_sockets * num_cores_per_socket + else: + cmds = {"Darwin": "sysctl -n hw.physicalcpu"} + return int(os.popen(cmds.get(platform.system(), "expr 1")).read()) + + def num_logical_processors(): + '''Get the number of logical processors''' + cmds = { + "Linux": "grep \"processor\" /proc/cpuinfo|sort -u|wc -l", + "Darwin": "sysctl -n hw.logicalcpu" + } + return int(os.popen(cmds.get(platform.system(), "expr 1")).read()) + + num_cores = num_physical_cores() + num_processors = num_logical_processors() + if num_processors > num_cores: # Hyper Threading is enabled set_env("OMP_DYNAMIC", "true") set_env("KMP_AFFINITY", "granularity=fine,compact,1,0") - processors = os.popen("grep \"processor\" /proc/cpuinfo|sort -u|wc -l") - processors = int(processors.read()) - trainers = kwargs.get('trainer_count', 1) - threads = processors / trainers + else: + set_env("OMP_DYNAMIC", "false") + set_env("KMP_AFFINITY", "granularity=fine,compact,0,0") + threads = num_processors / trainer_count threads = '1' if threads < 1 else str(threads) set_env("OMP_NUM_THREADS", threads) set_env("MKL_NUM_THREADS", threads) + +def init(**kwargs): + import py_paddle.swig_paddle as api + args = [] + args_dict = {} + # NOTE: append arguments if they are in ENV + for ek, ev in os.environ.iteritems(): + if ek.startswith("PADDLE_INIT_"): + args_dict[ek.replace("PADDLE_INIT_", "").lower()] = str(ev) + + args_dict.update(kwargs) + # NOTE: overwrite arguments from ENV if it is in kwargs + for key in args_dict.keys(): + args.append('--%s=%s' % (key, str(args_dict[key]))) + + set_omp_mkl_env_vars(kwargs.get('trainer_count', 1)) + if 'use_gpu' in kwargs: cp.g_command_config_args['use_gpu'] = kwargs['use_gpu'] if 'use_mkldnn' in kwargs: diff --git a/python/paddle/v2/dataset/common.py b/python/paddle/v2/dataset/common.py index e31e501ce93c5dc20693a8724ee7dd864f9aef55..191d9ecfb127c1851a392bc9ec83734d630d0ac4 100644 --- a/python/paddle/v2/dataset/common.py +++ b/python/paddle/v2/dataset/common.py @@ -71,7 +71,7 @@ def download(url, module_name, md5sum): if retry < retry_limit: retry += 1 else: - raise RuntimeError("Cannot download {0} within retry limit {2}". + raise RuntimeError("Cannot download {0} within retry limit {1}". format(url, retry_limit)) print "Cache file %s not found, downloading %s" % (filename, url) r = requests.get(url, stream=True) diff --git a/python/paddle/v2/dataset/flowers.py b/python/paddle/v2/dataset/flowers.py index 634388094c804827657dc83d5c205e680625b156..7bdddeaabec733ef26b3f766c6437f5c53d65044 100644 --- a/python/paddle/v2/dataset/flowers.py +++ b/python/paddle/v2/dataset/flowers.py @@ -44,7 +44,7 @@ __all__ = ['train', 'test', 'valid'] DATA_URL = 'http://www.robots.ox.ac.uk/~vgg/data/flowers/102/102flowers.tgz' LABEL_URL = 'http://www.robots.ox.ac.uk/~vgg/data/flowers/102/imagelabels.mat' SETID_URL = 'http://www.robots.ox.ac.uk/~vgg/data/flowers/102/setid.mat' -DATA_MD5 = '52808999861908f626f3c1f4e79d11fa' +DATA_MD5 = '33bfc11892f1e405ca193ae9a9f2a118' LABEL_MD5 = 'e0620be6f572b9609742df49c70aed4d' SETID_MD5 = 'a5357ecc9cb78c4bef273ce3793fc85c' # In official 'readme', tstid is the flag of test data diff --git a/python/paddle/v2/dataset/imdb.py b/python/paddle/v2/dataset/imdb.py index cfc1c886e1389c15e3f803c341b6f62dd7b4bf41..21ed7f7a5ce279f5bc65e5b008f14a1b0ff97343 100644 --- a/python/paddle/v2/dataset/imdb.py +++ b/python/paddle/v2/dataset/imdb.py @@ -23,10 +23,9 @@ Besides, this module also provides API for building dictionary. import paddle.v2.dataset.common import collections import tarfile -import Queue import re import string -import threading +import random __all__ = ['build_dict', 'train', 'test', 'convert'] @@ -74,47 +73,21 @@ def build_dict(pattern, cutoff): return word_idx -def reader_creator(pos_pattern, neg_pattern, word_idx, buffer_size): +def reader_creator(pos_pattern, neg_pattern, word_idx): UNK = word_idx[''] + INS = [] - qs = [Queue.Queue(maxsize=buffer_size), Queue.Queue(maxsize=buffer_size)] - - def load(pattern, queue): + def load(pattern, out, label): for doc in tokenize(pattern): - queue.put(doc) - queue.put(None) + out.append(([word_idx.get(w, UNK) for w in doc], label)) + + load(pos_pattern, INS, 0) + load(neg_pattern, INS, 1) + random.shuffle(INS) def reader(): - # Creates two threads that loads positive and negative samples - # into qs. - t0 = threading.Thread( - target=load, args=( - pos_pattern, - qs[0], )) - t0.daemon = True - t0.start() - - t1 = threading.Thread( - target=load, args=( - neg_pattern, - qs[1], )) - t1.daemon = True - t1.start() - - # Read alternatively from qs[0] and qs[1]. - i = 0 - doc = qs[i].get() - while doc != None: - yield [word_idx.get(w, UNK) for w in doc], i % 2 - i += 1 - doc = qs[i % 2].get() - - # If any queue is empty, reads from the other queue. - i += 1 - doc = qs[i % 2].get() - while doc != None: - yield [word_idx.get(w, UNK) for w in doc], i % 2 - doc = qs[i % 2].get() + for doc, label in INS: + yield doc, label return reader @@ -133,7 +106,7 @@ def train(word_idx): """ return reader_creator( re.compile("aclImdb/train/pos/.*\.txt$"), - re.compile("aclImdb/train/neg/.*\.txt$"), word_idx, 1000) + re.compile("aclImdb/train/neg/.*\.txt$"), word_idx) def test(word_idx): @@ -150,7 +123,7 @@ def test(word_idx): """ return reader_creator( re.compile("aclImdb/test/pos/.*\.txt$"), - re.compile("aclImdb/test/neg/.*\.txt$"), word_idx, 1000) + re.compile("aclImdb/test/neg/.*\.txt$"), word_idx) def word_dict(): diff --git a/python/paddle/v2/dataset/uci_housing.py b/python/paddle/v2/dataset/uci_housing.py index 98b97c75ca72f11c105535e0f2a5fa0201db5d42..f10bf7e42a1ead09b3eba0d61e55701215e4360f 100644 --- a/python/paddle/v2/dataset/uci_housing.py +++ b/python/paddle/v2/dataset/uci_housing.py @@ -38,6 +38,7 @@ UCI_TEST_DATA = None URL_MODEL = 'https://github.com/PaddlePaddle/book/raw/develop/01.fit_a_line/fit_a_line.tar' MD5_MODEL = '52fc3da8ef3937822fcdd87ee05c0c9b' + def feature_range(maximums, minimums): import matplotlib matplotlib.use('Agg') @@ -114,7 +115,8 @@ def test(): def model(): - tar_file = paddle.v2.dataset.common.download(URL_MODEL, 'fit_a_line.tar', MD5_MODEL) + tar_file = paddle.v2.dataset.common.download(URL_MODEL, 'fit_a_line.tar', + MD5_MODEL) with open(tar_file, 'r') as f: parameters = Parameters.from_tar(f) return parameters diff --git a/python/paddle/v2/fluid/__init__.py b/python/paddle/v2/fluid/__init__.py index 5df612bf3530c843c16b337f2b8f83445fcf39b5..225b41c5043b5792abb90bbad53cbbfce9a3156e 100644 --- a/python/paddle/v2/fluid/__init__.py +++ b/python/paddle/v2/fluid/__init__.py @@ -1,11 +1,51 @@ -import sys -import core -__all__ = ['proto'] -argv = [] -if core.is_compile_gpu(): - argv = list(sys.argv) + [ - "--tryfromenv=fraction_of_gpu_memory_to_use,use_pinned_memory" - ] -else: - argv = list(sys.argv) + ["--tryfromenv=use_pinned_memory"] -core.init_gflags(argv) +# import all class inside framework into fluid module +import framework +from framework import * +# import all class inside executor into fluid module +import executor +from executor import * + +import io +import evaluator +import initializer +import layers +import nets +import optimizer +import backward +import regularizer +from param_attr import ParamAttr +from data_feeder import DataFeeder +from core import LoDTensor, CPUPlace, CUDAPlace +from distribute_transpiler import DistributeTranspiler +import clip + +Tensor = LoDTensor +__all__ = framework.__all__ + executor.__all__ + [ + 'io', 'initializer', 'layers', 'nets', 'optimizer', 'backward', + 'regularizer', 'LoDTensor', 'CPUPlace', 'CUDAPlace', 'Tensor', 'ParamAttr' + 'DataFeeder', 'clip', 'DistributeTranspiler' +] + + +def __read_gflags_from_env__(): + """ + Enable reading gflags from environment variables. + + Returns: + None + """ + import sys + import core + read_env_flags = ['use_pinned_memory', 'check_nan_inf'] + if core.is_compile_gpu(): + read_env_flags.append('fraction_of_gpu_memory_to_use') + core.init_gflags([sys.argv[0]] + + ["--tryfromenv=" + ",".join(read_env_flags)]) + + if core.is_compile_gpu(): + core.init_devices(["CPU", "GPU:0"]) + else: + core.init_devices(["CPU"]) + + +__read_gflags_from_env__() diff --git a/python/paddle/v2/fluid/backward.py b/python/paddle/v2/fluid/backward.py index f188582178f667125ec95cd230100fdb10ce7e88..f11c83f59c930784ca355acc3193c3c352db10e5 100644 --- a/python/paddle/v2/fluid/backward.py +++ b/python/paddle/v2/fluid/backward.py @@ -1,54 +1,323 @@ from paddle.v2.fluid import framework as framework +from . import core +import collections -__all__ = ['append_backward_ops'] +__all__ = ['append_backward'] -def append_backward_ops(loss, parameter_list=None, no_grad_set=None): +def _rename_arg_(op_descs, old_name, new_name, begin_idx=None, end_idx=None): """ - Create and add gradient Operators in BlockDesc to compute - gradients of `loss` for parameters in parameter_list + Traverse all ops in op_descs[begin_idx : end_idx], + if any op has inputs/outputs named "old_name", rename it as 'new_name' + """ + if begin_idx is None: + begin_idx = 0 + if end_idx is None: + end_idx = len(op_descs) + for i in range(begin_idx, end_idx): + op_desc = op_descs[i] + if isinstance(op_desc, tuple): + op_desc = op_desc[0] + op_desc.rename_input(old_name, new_name) + op_desc.rename_output(old_name, new_name) + + +def _create_op_desc_(op_type, inputs, outputs, attrs): + """ + Create a C++ OpDesc object with specified inputs, outputs and attributes. + """ + op_desc = core.OpDesc() + op_desc.set_type(op_type) + for para, args in inputs.iteritems(): + op_desc.set_input(para, args) + for para, args in outputs.iteritems(): + op_desc.set_output(para, args) + for name, val in attrs.iteritems(): + if isinstance(val, framework.Block): + op_desc.set_block_attr(name, val.desc) + else: + op_desc.set_attr(name, val) + return op_desc + + +def _infer_var_data_type_(grad_var_name, block): + """ + Infer the data type of given grad variable + """ + grad_var = block.desc.find_var(grad_var_name.encode("ascii")) + fwd_name = _strip_grad_suffix_(grad_var_name.encode("ascii")) + if block.desc.has_var_recursive(fwd_name): + fwd_var = block.desc.find_var_recursive(fwd_name.encode("ascii")) + grad_var.set_dtype(fwd_var.dtype()) + else: + grad_var.set_dtype(core.DataType.FP32) + + +def _all_in_set_(cands, s): + """ + Test if all elements of 'cands' are in set 's' + """ + for c in cands: + if not c in s: + return False + return True + + +def _strip_grad_suffix_(name): + """ + Strip the grad suffix from the given varibale name + e.g. x@GRAD ==> x + y@GRAD@RENAME@1 ==> y + """ + pos = name.find(core.grad_var_suffix()) + return name[:pos] if pos != -1 else name + + +def _append_grad_suffix_(name): + """ + Append grad suffix to the given variable name + e.g. x ==> x@GRAD + """ + return name + core.grad_var_suffix() + + +def _addup_repetitive_outputs_(op_descs): + """ + In backward part, an variable may be the output of more than one ops. + In this case, the variable should be the accumulation of all the outputs. + `sum_op`s are added to implement the accumulate. + """ + pending_sum_ops = [] + var_rename_count = collections.defaultdict(int) + renamed_vars = collections.defaultdict(list) + for idx, op_desc in enumerate(op_descs): + for var_name in op_desc.input_arg_names(): + if len(renamed_vars[var_name]) > 1: + pending_sum_ops.append( + (_create_op_desc_("sum", {"X": renamed_vars[var_name]}, + {"Out": [var_name]}, {}), idx)) + renamed_vars[var_name] = [var_name] + for var_name in op_desc.output_arg_names(): + if var_name == core.empty_var_name( + ) or var_name in op_desc.input_arg_names(): + # empty variable or inplace op + continue + if len(renamed_vars[var_name]) == 0: + # it's the first time we get the variable + renamed_vars[var_name] = [var_name] + else: + if len(renamed_vars[var_name]) == 1: + new_name = var_name + "@RENAME@" + \ + str(var_rename_count[var_name]) + var_rename_count[var_name] += 1 + # rename original var_name + renamed_vars[var_name][0] = new_name + _rename_arg_(op_descs, var_name, new_name, 0, idx) + _rename_arg_(pending_sum_ops, var_name, new_name) + + new_name = var_name + "@RENAME@" + \ + 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) + for var_name, inputs in renamed_vars.iteritems(): + if len(inputs) > 1: + pending_sum_ops.append((_create_op_desc_( + "sum", {"X": inputs}, {"Out": [var_name]}, {}), len(op_descs))) + # sum_op descs are sorted according to their insert position + for p in reversed(pending_sum_ops): + op_descs.insert(p[1], p[0]) + + return op_descs - :param loss: an variable generated by cost function. - :type loss: Variable - :param no_grad_set: variable that should not create gradient - :type no_grad_set: set - :param parameter_list: parameters that need to compute gradient and - update to optimize the lost. - :type: list - :return: list of (parameters, gradients) pair. - :rtype: list[Variable] + +def _remove_no_grad_branch_(op_descs, no_grad_set): + """ + Remove unnecessary grad ops + A grad op can be removed in two cases: + 1. all outputs of the grad op are in 'no_grad_set' + 2. (TODO) all grad inputs of the grad op are in 'no_grad_set' + """ + # Remove ops whose outputs are all in no_grad_dict + op_descs = filter( + lambda op_desc: not _all_in_set_(op_desc.output_arg_names(), no_grad_set), + op_descs) + # Insert fill_zeros_like_op + to_insert = [] + for idx, op_desc in enumerate(op_descs): + for arg in op_desc.input_arg_names(): + if core.grad_var_suffix() in arg and arg in no_grad_set: + to_insert.append((_create_op_desc_("fill_zeros_like", { + "X": [_strip_grad_suffix_(arg)] + }, {"Y": [arg]}, {}), idx)) + + map(lambda p: op_descs.insert(p[1], p[0]), reversed(to_insert)) + + return op_descs + + +def _append_backward_ops_(target, + block, + target_block, + no_grad_dict, + grad_to_var, + callback=None): + """ + Create all grad ops, and insert them into given block + + Args: + target(Variable): the target variable of forward pass + block(Block): the block where forward ops are + target_block(Block): the block which is going to hold new generated grad ops + no_grad_dict(dict): + key(int) block index + val(set) a set of varibale names. These varibales have no gradient + grad_to_var(dict)(output argument): + key(str): grad variable name + val(str): corresponding forward variable name + """ + # grad_op_descs holds created grad_op, and will be appended to target_block + grad_op_descs = [] + program = block.program + for op in reversed(block.ops): + grad_sub_block_list = [] + # If the op has its own sub-block, deal with the sub-block first + if op.has_attr("sub_block"): + sub_block = program.block(op.block_attr("sub_block")) + grad_sub_block = program.create_block(parent_idx=sub_block.idx) + _append_backward_ops_(target, sub_block, grad_sub_block, + no_grad_dict, grad_to_var, callback) + grad_sub_block_list.append(grad_sub_block.desc) + + # Getting op's corresponding grad_op + grad_op_desc, op_grad_to_var = core.get_grad_op_desc( + op.desc, no_grad_dict[block.idx], grad_sub_block_list) + grad_op_descs.extend(grad_op_desc) + grad_to_var.update(op_grad_to_var) + + grad_op_descs = _addup_repetitive_outputs_(grad_op_descs) + + grad_op_descs = _remove_no_grad_branch_(grad_op_descs, + no_grad_dict[block.idx]) + + if target_block.idx == 0: + grad_op_descs.insert( + 0, + _create_op_desc_("fill_constant", {}, { + "Out": [_append_grad_suffix_(target.name)] + }, {"shape": [1], + "value": 1.0, + "dtype": target.dtype})) + # append op_desc in grad_op_descs to target_block + for op_desc in grad_op_descs: + new_op_desc = target_block.desc.append_op() + new_op_desc.copy_from(op_desc) + + +def _append_backward_vars_(block, start_op_idx, grad_to_var, grad_info_map): + """ + Create new variables required by backward pass. + + Args: + block(Block): the block where new variables will be created + start_op_idx(int): Only variables required by ops in block.ops[start_op_idx : ] will be created + grad_to_var(dict): + key(str): grad variable name + val(str): corresponding forward variable name + In most cases, this dict is generated by _append_backward_ops_() + grad_info_map(dict)(output argument): + key(str): forward variable name + val(tuple): a tuple of (str, int), str is the corresponding grad name, int is the block index + """ + for op_idx in range(start_op_idx, block.desc.op_size()): + op_desc = block.desc.op(op_idx) + if op_desc.has_attr("sub_block"): + sub_block = block.program.block(op_desc.block_attr("sub_block")) + _append_backward_vars_(sub_block, 0, grad_to_var, grad_info_map) + new_vars = set() + # create new gradient variables + for grad_var_name in op_desc.output_arg_names(): + grad_var_name = grad_var_name.encode("ascii") + if block.desc.has_var_recursive( + grad_var_name) or grad_var_name == core.empty_var_name(): + continue + block.desc.var(grad_var_name) + new_vars.add(grad_var_name) + if not grad_to_var.has_key(grad_var_name): + continue + grad_info_map[grad_to_var[grad_var_name]] = (grad_var_name, block) + # infer_shape and infer_type + op_desc.infer_var_type(block.desc) + op_desc.infer_shape(block.desc) + for arg in op_desc.output_arg_names(): + if arg in new_vars: + _infer_var_data_type_(arg, block) + + +def append_backward(loss, parameter_list=None, no_grad_set=None): + """ + Append backward part to main_program + + Args: + loss(Variable): The variable generated by cost function. + parameter_list(list): Parameters that need to be updated by optimizer. + If None, it means all parameters need to be updated. + no_grad_set(set): Variables that have no gradients in Block 0. + If None, the set will be generated inside the function and + contains all variables with `step_gradient=True` from all blocks. + + Return: + (list[Variable]): list of (parameters, gradients) pair. """ assert isinstance(loss, framework.Variable) + program = loss.block.program + no_grad_dict = dict() if no_grad_set is None: - program = loss.block.program assert isinstance(program, framework.Program) - no_grad_set = list() for block in program.blocks: assert isinstance(block, framework.Block) + block_no_grad_set = set() for var in block.vars.itervalues(): assert isinstance(var, framework.Variable) if var.stop_gradient: - no_grad_set.append(var.name) - no_grad_set = set(no_grad_set) + block_no_grad_set.add(_append_grad_suffix_(var.name)) + no_grad_dict[block.idx] = block_no_grad_set + elif isinstance(no_grad_set, set): + no_grad_dict = {0: no_grad_set} + else: + raise ValueError("'no_grad_set' should be a set or None.") + + grad_info_map = dict() + root_block = program.block(0) + + fwd_op_num = root_block.desc.op_size() + current_block_idx = program.current_block_idx + grad_to_var = dict() + + _append_backward_ops_(loss, root_block, root_block, no_grad_dict, + grad_to_var) + _append_backward_vars_(root_block, fwd_op_num, grad_to_var, grad_info_map) + + program.current_block_idx = current_block_idx + program.sync_with_cpp() - param_grad_map = loss.block.program.append_backward(loss, no_grad_set) if parameter_list is not None: parameters = parameter_list else: - params = loss.block.program.global_block().all_parameters() + params = program.global_block().all_parameters() parameters = [param.name for param in params] params_and_grads = [] for param in parameters: - if param not in param_grad_map: + if param not in grad_info_map: raise ValueError("param %s is not in map" % param) - grad_info = param_grad_map[param] - grad_block = loss.block.program.block(grad_info[1]) + grad_info = grad_info_map[param] + grad_block = grad_info[1] if not grad_block.has_var(grad_info[0]): raise ValueError("grad block[{0}] did not have grad var {1}".format( grad_info[1], grad_info[0])) # Get the param var from the global block - param_var = loss.block.program.global_block().var(param) + param_var = program.global_block().var(param) grad_var = grad_block.var(grad_info[0]) if loss.block.has_var(grad_info[0]): params_and_grads.append((param_var, grad_var)) diff --git a/python/paddle/v2/fluid/clip.py b/python/paddle/v2/fluid/clip.py new file mode 100644 index 0000000000000000000000000000000000000000..d7ec2fbe13fe6d9158345099b8668afc5c7d4571 --- /dev/null +++ b/python/paddle/v2/fluid/clip.py @@ -0,0 +1,61 @@ +import functools +import layers + +__all__ = ['GradientClipByValue', 'append_gradient_clip_ops'] + + +class BaseGradientClipAttr(object): + def process_context(self, context, p_g): + raise NotImplementedError() + + def create_operators(self, param, grad): + raise NotImplementedError() + + +class NullGradientClipAttr(BaseGradientClipAttr): + def process_context(self, context, p_g): + pass + + def create_operators(self, param, grad): + return param, grad + + +class GradientClipByValue(BaseGradientClipAttr): + def __init__(self, max, min=None): + max = float(max) + if min is None: + min = -max + else: + min = float(min) + self.max = max + self.min = min + + def process_context(self, context, p_g): + pass + + def create_operators(self, param, grad): + new_grad = layers.clip(x=grad, min=self.min, max=self.max) + return param, new_grad + + +def append_gradient_clip_ops(param_grad): + context = dict() + create_op_callbacks = [] + for p, g in param_grad: + clip_attr = getattr(p, '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 BaseGradientClippingAttr" + ) + + clip_attr.process_context(context=context, p_g=param_grad) + create_op_callbacks.append( + functools.partial( + clip_attr.create_operators, param=p, grad=g)) + + return [each_callback() for each_callback in create_op_callbacks] + + +ClipByValue = GradientClipByValue diff --git a/python/paddle/v2/fluid/data_feeder.py b/python/paddle/v2/fluid/data_feeder.py new file mode 100644 index 0000000000000000000000000000000000000000..24036c3e75b9594ba58cccb02825ab8020d1e107 --- /dev/null +++ b/python/paddle/v2/fluid/data_feeder.py @@ -0,0 +1,101 @@ +from __future__ import print_function +import core +import numpy +import six.moves as six + +from framework import Variable, default_main_program + +__all__ = ['DataFeeder'] + + +class DataToLoDTensorConverter(object): + def __init__(self, place, lod_level, shape, dtype): + self.place = place + self.lod_level = lod_level + self.shape = shape + if dtype == core.DataType.FP32: + self.dtype = 'float32' + elif dtype == core.DataType.INT64: + self.dtype = 'int64' + elif dtype == core.DataType.FP64: + self.dtype = 'float64' + elif dtype == core.DataType.INT32: + self.dtype = 'int32' + else: + raise ValueError("dtype must be any of [int32, float32, int64, " + "float64]") + + self.data = [] + self.lod = [] + + for i in six.range(lod_level): + self.lod.append([0]) + + def feed(self, data): + self._feed_impl_(data, self.lod, self.lod_level) + + def _feed_impl_(self, data, lod, lod_level): + if lod_level == 0: + self.data.append(data) + else: + cur_lod_len = len(data) + lod[-1].append(lod[-1][-1] + cur_lod_len) + for each_data in data: + self._feed_impl_(each_data, lod[:-1], lod_level - 1) + + def done(self): + arr = numpy.array(self.data, dtype=self.dtype).reshape(self.shape) + t = core.LoDTensor() + t.set(arr, self.place) + if self.lod_level > 0: + t.set_lod(self.lod) + return t + + +class DataFeeder(object): + def __init__(self, feed_list, place, program=None): + self.feed_dtypes = [] + self.feed_names = [] + self.feed_shapes = [] + self.feed_lod_level = [] + if program is None: + program = default_main_program() + for each_var in feed_list: + if isinstance(each_var, basestring): + each_var = program.block(0).var(each_var) + if not isinstance(each_var, Variable): + raise TypeError("Feed list should contain a list of variable") + self.feed_dtypes.append(each_var.dtype) + self.feed_names.append(each_var.name) + shape = each_var.shape + batch_size_dim = -1 + for i, s in enumerate(shape): + if s < 0: + batch_size_dim = i + break + if batch_size_dim == -1: + raise ValueError("Variable {0} must has a batch size dimension", + each_var.name) + self.feed_lod_level.append(each_var.lod_level) + self.feed_shapes.append(shape) + + self.place = place + + def feed(self, iterable): + converter = [] + for lod_level, shape, dtype in six.zip( + self.feed_lod_level, self.feed_shapes, self.feed_dtypes): + converter.append( + DataToLoDTensorConverter( + place=self.place, + lod_level=lod_level, + shape=shape, + dtype=dtype)) + + for each_sample in iterable: + for each_converter, each_slot in six.zip(converter, each_sample): + each_converter.feed(each_slot) + ret_dict = {} + for each_name, each_converter in six.zip(self.feed_names, converter): + ret_dict[each_name] = each_converter.done() + return ret_dict diff --git a/python/paddle/v2/fluid/distribute_transpiler.py b/python/paddle/v2/fluid/distribute_transpiler.py new file mode 100644 index 0000000000000000000000000000000000000000..49ece7b725e318d7526d58fe54c97cbe20200a7d --- /dev/null +++ b/python/paddle/v2/fluid/distribute_transpiler.py @@ -0,0 +1,242 @@ +import framework +from framework import Program, default_main_program, Parameter, Variable +import optimizer +from layer_helper import LayerHelper + + +def hash_name_to_server(params_grads, pserver_endpoints): + """ + :param param_grads: + :return: a map of pserver endpoint -> + params -> [param list] + grads -> [grad list] + """ + + def _hash_param(param_name, total): + return hash(param_name) % total + + param_grad_map = dict() + for param, grad in params_grads: + if param.trainable is True and grad is not None: + server_id = _hash_param(param.name, len(pserver_endpoints)) + server_for_param = pserver_endpoints[server_id] + if not param_grad_map.has_key(server_for_param): + param_grad_map[server_for_param] = {"params": [], "grads": []} + param_grad_map[server_for_param]["params"].append(param) + param_grad_map[server_for_param]["grads"].append(grad) + + return param_grad_map + + +def round_robin(params_grads, pserver_endpoints): + assert (len(params_grads) > len(pserver_endpoints)) + + param_grad_map = dict() + pserver_idx = 0 + for param, grad in params_grads: + if param.trainable is True: + server_for_param = pserver_endpoints[pserver_idx] + if not param_grad_map.has_key(server_for_param): + param_grad_map[server_for_param] = {"params": [], "grads": []} + + param_grad_map[server_for_param]["params"].append(param) + param_grad_map[server_for_param]["grads"].append(grad) + + pserver_idx += 1 + if pserver_idx >= len(pserver_endpoints): + pserver_idx = 0 + return param_grad_map + + +class DistributeTranspiler: + def transpile(self, + optimize_ops, + params_grads, + program=None, + pservers="127.0.0.1:6174", + trainers=1, + split_method=round_robin): + """ + Transpile the program to a distributed data-parallelism programs. + + The main_program will be transform to use a remote parameter server + to do parameter optimization. And the optimization graph will be put + in to a parameter server program. + + Use different methods to split trainable varialbles to different + parameter servers. + + Example to run: + + exe = fluid.Executor(place) + t = fluid.DistributeTranspiler() + t.transpile(optimize_ops, params_grads, pservers="127.0.0.1:6174", trainers=1) + + pserver_endpoint = os.getenv("PSERVER") + if pserver_endpoint: + pserver_prog = t.get_pserver_program(pserver_endpoint, optimize_ops) + exe.run(fluid.default_startup_program()) + exe.run(pserver_prog) + else: + feeder = fluid.DataFeeder(feed_list=[images, label], place=place) + exe.run(fluid.default_startup_program()) + + for pass_id in range(PASS_NUM): + ... + + :param optimize_ops: op list of optimization, should be the + return value of Optimizer.minimize + :type optimize_ops: list + :param program: program to optimize, default default_main_program + :param pservers: parameter server endpoints like "m1:6174,m2:6174" + :type pservers: string + + :return: return a list of programs + """ + if program is None: + program = default_main_program() + self.program = program + self.trainers = trainers + self.optimize_ops = optimize_ops + self._optimize_distributed( + optimize_ops, + program, + params_grads, + pservers=pservers, + trainers=trainers, + split_method=split_method) + + def _clone_param(self, block, v): + assert isinstance(v, Parameter) + new_p = Parameter( + block=block, + shape=v.shape, + dtype=v.dtype, + type=v.type, + lod_level=v.lod_level, + stop_gradient=v.stop_gradient, + trainable=v.trainable, + optimize_attr=v.optimize_attr, + regularizer=v.regularizer, + name=v.name) + block.vars[new_p.name] = new_p + + def _clone_var(self, block, var): + assert isinstance(var, Variable) + return block.create_var( + name=var.name, + shape=var.shape, + dtype=var.dtype, + type=var.type, + lod_level=var.lod_level, + persistable=var.persistable) + + def _optimize_distributed(self, optimize_ops, program, params_and_grads, + **kwargs): + if kwargs.has_key("split_method"): + split_method = kwargs["split_method"] + else: + split_method = round_robin + + assert (callable(split_method)) + pserver_endpoints = kwargs["pservers"].split(",") + self.param_grad_map = split_method(params_and_grads, pserver_endpoints) + + send_op_ordered_inputs = [] + send_op_ordered_outputs = [] + epmap = [] + for ep, v in self.param_grad_map.iteritems(): + send_op_ordered_inputs.extend(v["grads"]) + send_op_ordered_outputs.extend(v["params"]) + for i in v["grads"]: + epmap.append(ep) + send_op = program.global_block().append_op( + type="send", + inputs={"X": send_op_ordered_inputs + }, # inputs is a list of tensors to be send + outputs={"Out": send_op_ordered_outputs}, + attrs={"endpoints": pserver_endpoints, + "epmap": epmap}) + + def get_trainer_program(self): + # remove optimize ops and add a send op to main_program + self.program.global_block().delete_ops(self.optimize_ops) + return self.program + + def _create_var_for_trainers(self, block, var, trainers): + var_list = [] + for i in xrange(trainers): + var_each = block.create_var( + name="%s.trainer_%d" % (var.name, i), + psersistable=var.persistable, + dtype=var.dtype, + shape=var.shape) + var_list.append(var_each) + return var_list + + def get_pserver_program(self, endpoint, optimize_ops): + pserver_program = Program() + for v in self.param_grad_map[endpoint]["params"]: + self._clone_param(pserver_program.global_block(), v) + + optimize_sub_program = Program() + grad_var_names = [ + var.name for var in self.param_grad_map[endpoint]["grads"] + ] + for opt_op in optimize_ops: + for _, var in opt_op.inputs.iteritems(): + # NOTE: append operators to merge gradients from multiple + # trainers. If trainers == 1, this is not needed. + if self.trainers > 1 and var.name in grad_var_names: + vars2merge = self._create_var_for_trainers( + optimize_sub_program.global_block(), var, self.trainers) + merged_var = optimize_sub_program.global_block().create_var( + name=var.name, + persistable=var.persistable, + dtype=var.dtype, + shape=var.shape) + optimize_sub_program.global_block().append_op( + type="sum", + inputs={"X": vars2merge}, + outputs={"Out": merged_var}) + optimize_sub_program.global_block().append_op( + type="scale", + inputs={"X": merged_var}, + outputs={"Out": merged_var}, + attrs={"scale": 1.0 / float(self.trainers)}) + else: + optimize_sub_program.global_block().create_var( + name=var.name, + persistable=var.persistable, + dtype=var.dtype, + shape=var.shape) + + if opt_op.inputs.has_key("Grad"): + if opt_op.inputs["Grad"].name in grad_var_names: + optimize_sub_program.global_block().append_op( + type=opt_op.type, + inputs=opt_op.inputs, + outputs=opt_op.outputs, + attrs=opt_op.attrs) + else: + optimize_sub_program.global_block().append_op( + type=opt_op.type, + inputs=opt_op.inputs, + outputs=opt_op.outputs, + attrs=opt_op.attrs) + pserver_program.global_block().append_op( + type="recv", + inputs={"RX": + self.param_grad_map[endpoint]["grads"]}, # grads to recv + outputs={}, + attrs={ + "OptimizeProgram": optimize_sub_program.desc, + "endpoint": endpoint, + "ParamList": + [p.name for p in self.param_grad_map[endpoint]["params"]], + "GradList": + [p.name for p in self.param_grad_map[endpoint]["grads"]], + "Trainers": self.trainers + }) + pserver_program.sync_with_cpp() + return pserver_program diff --git a/python/paddle/v2/fluid/evaluator.py b/python/paddle/v2/fluid/evaluator.py index 3a8f1831cf2c44c81aee62c6ee172942db188217..e186ee96c387acf24471d4e26ce020c4ecac8d19 100644 --- a/python/paddle/v2/fluid/evaluator.py +++ b/python/paddle/v2/fluid/evaluator.py @@ -1,14 +1,18 @@ import numpy as np -from paddle.v2.fluid.framework import Program, g_main_program, unique_name, Variable -import paddle.v2.fluid.core as core +import layers +from framework import Program, unique_name, Variable, program_guard +from layer_helper import LayerHelper -def _clone_var_in_block_(block, var): +__all__ = ['Accuracy', 'ChunkEvaluator'] + + +def _clone_var_(block, var): assert isinstance(var, Variable) return block.create_var( name=var.name, shape=var.shape, - dtype=var.data_type, + dtype=var.dtype, type=var.type, lod_level=var.lod_level, persistable=True) @@ -16,172 +20,171 @@ def _clone_var_in_block_(block, var): class Evaluator(object): """ - Evalutor Base class. - - create metric states - add mini-batch evaluator caculate operator - add increment operator to accumulate the metric states + Base Class for all evaluators + + Args: + name(str): The name of evaluator. such as, "accuracy". Used for generate + temporary variable name. + main_program(Program, optional): The evaluator should be added to this + main_program. Default default_main_program() + startup_program(Program, optional):The parameter should be added to this + startup_program. Default default_startup_program() + + Attributes: + states(list): The list of state variables. states will be reset to zero + when `reset` is invoked. + metrics(list): The list of metrics variables. They will be calculate + every mini-batch """ def __init__(self, name, **kwargs): - """ - init the global states - """ - self._states = {} - if kwargs.has_key("main_program"): - self._main_program = kwargs.get("main_program") - else: - self._main_program = g_main_program - - def _update_ops(self, *args, **kwargs): - """ - append update ops to the global states - """ - raise NotImplementedError() + self.states = [] + self.metrics = [] + self.helper = LayerHelper(name, **kwargs) def reset(self, executor, reset_program=None): """ - Clear metric states at the begin of each pass/user specified batch + reset metric states at the begin of each pass/user specified batch """ - if reset_program == None: + if reset_program is None: reset_program = Program() - else: - reset_program = program - block = reset_program.global_block() - for k, var in self._states.iteritems(): - g_var = _clone_var_in_block_(block, var) - zeros = block.create_var(dtype="float32", persistable=True) - block.append_op( - type="fill_constant", - outputs={"Out": [zeros]}, - attrs={ - "shape": g_var.shape, - "value": .0, - "data_type": 5, - }) - block.append_op( - type="scale", inputs={"X": zeros}, outputs={"Out": g_var}) - executor.run(reset_program, fetch_list=self._states.values()) + + with program_guard(main_program=reset_program): + for var in self.states: + assert isinstance(var, Variable) + g_var = _clone_var_(reset_program.current_block(), var) + layers.fill_constant( + shape=g_var.shape, value=0.0, dtype=g_var.dtype, out=g_var) + + executor.run(reset_program) def eval(self, executor, eval_program=None): """ - Merge the mini-batch statistics to form the evaluation result for multiple mini-batches. + Evaluate the statistics merged by multiple mini-batches. """ raise NotImplementedError() + def create_state(self, suffix, dtype, shape): + """ + Create state variable. + + NOTE: It is not a public API. + + Args: + suffix(str): the state suffix. + dtype(str|core.DataType): the state data type + shape(tuple|list): the shape of state + + Returns: State variable + + """ + state = self.helper.create_variable( + name="_".join([unique_name(self.helper.name), suffix]), + persistable=True, + dtype=dtype, + shape=shape) + self.states.append(state) + return state + class Accuracy(Evaluator): """ - Accuracy need two state variable Total, Correct + Average Accuracy for multiple mini-batches. """ - def __init__(self, *args, **kwargs): + def __init__(self, input, label, k=1, **kwargs): super(Accuracy, self).__init__("accuracy", **kwargs) - block = self._main_program.global_block() - g_total = block.create_var( - name=unique_name("Total"), - persistable=True, - dtype="int64", - shape=[1]) - g_correct = block.create_var( - name=unique_name("Correct"), - persistable=True, - dtype="int64", - shape=[1]) - self._states["Total"] = g_total - self._states["Correct"] = g_correct - - def _update_ops(self, input, label, k=1, **kwargs): - block = self._main_program.global_block() - topk_out = block.create_var(dtype=input.data_type) - topk_indices = block.create_var(dtype="int64") - block.append_op( - type="top_k", - inputs={"X": [input]}, - outputs={"Out": [topk_out], - "Indices": [topk_indices]}, - attrs={"k": k}) - acc_out = block.create_var(dtype=kwargs.get("out_dtype", "float32")) - correct = block.create_var(dtype="int64", persistable=True) - total = block.create_var(dtype="int64", persistable=True) - block.append_op( - type="accuracy", - inputs={ - "Out": [topk_out], - "Indices": [topk_indices], - "Label": [label] - }, - outputs={ - "Accuracy": [acc_out], - "Correct": [correct], - "Total": [total], - }) - - block.append_op( - type="cast", - inputs={"X": [self._states["Total"]]}, - outputs={"Out": [self._states["Total"]]}, - attrs={ - "in_data_type": 5, # float32 - "out_data_type": 2, #int32 - }) - block.append_op( - type="cast", - inputs={"X": [self._states["Correct"]]}, - outputs={"Out": [self._states["Correct"]]}, - attrs={ - "in_data_type": 5, - "out_data_type": 2, - }) - - block.append_op( - type="elementwise_add", - inputs={"X": [self._states["Total"]], - "Y": [total]}, - outputs={"Out": [self._states["Total"]]}) - block.append_op( - type="elementwise_add", - inputs={"X": [self._states["Correct"]], - "Y": [correct]}, - outputs={"Out": [self._states["Correct"]]}) - - return acc_out + main_program = self.helper.main_program + if main_program.current_block().idx != 0: + raise ValueError("You can only invoke Evaluator in root block") + + self.total = self.create_state(dtype='int64', shape=[1], suffix='total') + self.correct = self.create_state( + dtype='int64', shape=[1], suffix='correct') + total = self.helper.create_tmp_variable(dtype='int') + correct = self.helper.create_tmp_variable(dtype='int') + acc = layers.accuracy( + input=input, label=label, k=k, total=total, correct=correct) + total = layers.cast(x=total, dtype='int64') + correct = layers.cast(x=correct, dtype='int64') + layers.sums(input=[self.total, total], out=self.total) + layers.sums(input=[self.correct, correct], out=self.correct) + + self.metrics.append(acc) + + def eval(self, executor, eval_program=None): + if eval_program is None: + eval_program = Program() + block = eval_program.current_block() + with program_guard(main_program=eval_program): + total = _clone_var_(block, self.total) + correct = _clone_var_(block, self.correct) + total = layers.cast(total, dtype='float32') + correct = layers.cast(correct, dtype='float32') + out = layers.elementwise_div(x=correct, y=total) + return np.array(executor.run(eval_program, fetch_list=[out])[0]) + + +class ChunkEvaluator(Evaluator): + """ + Accumulate counter numbers output by chunk_eval from mini-batches and + compute the precision recall and F1-score using the accumulated counter + numbers. + """ + + def __init__( + self, + input, + label, + chunk_scheme, + num_chunk_types, + excluded_chunk_types=None, ): + super(ChunkEvaluator, self).__init__("chunk_eval") + main_program = self.helper.main_program + if main_program.current_block().idx != 0: + raise ValueError("You can only invoke Evaluator in root block") + + self.num_infer_chunks = self.create_state( + dtype='int64', shape=[1], suffix='num_infer_chunks') + self.num_label_chunks = self.create_state( + dtype='int64', shape=[1], suffix='num_label_chunks') + self.num_correct_chunks = self.create_state( + dtype='int64', shape=[1], suffix='num_correct_chunks') + precision, recall, f1_score, num_infer_chunks, num_label_chunks, num_correct_chunks = layers.chunk_eval( + input=input, + label=label, + chunk_scheme=chunk_scheme, + num_chunk_types=num_chunk_types, + excluded_chunk_types=excluded_chunk_types, ) + layers.sums( + input=[self.num_infer_chunks, num_infer_chunks], + out=self.num_infer_chunks) + layers.sums( + input=[self.num_label_chunks, num_label_chunks], + out=self.num_label_chunks) + layers.sums( + input=[self.num_correct_chunks, num_correct_chunks], + out=self.num_correct_chunks) + + self.metrics.extend([precision, recall, f1_score]) def eval(self, executor, eval_program=None): - if eval_program != None: - eval_program = eval_program - else: + if eval_program is None: eval_program = Program() - block = eval_program.global_block() - eval_out = block.create_var(dtype=self._states["Total"].data_type) - e_total = _clone_var_in_block_(block, self._states["Total"]) - e_correct = _clone_var_in_block_(block, self._states["Correct"]) - block.append_op( - type="cast", - inputs={"X": [e_total]}, - outputs={"Out": [e_total]}, - attrs={ - "in_data_type": 2, #int32 - "out_data_type": 5, #float32 - }) - block.append_op( - type="cast", - inputs={"X": [e_correct]}, - outputs={"Out": [e_correct]}, - attrs={ - "in_data_type": 2, - "out_data_type": 5, - }) - block.append_op( - type="elementwise_div", - inputs={"X": e_correct, - "Y": e_total}, - outputs={"Out": eval_out}) - out = executor.run(eval_program, fetch_list=[eval_out]) - return np.array(out[0]) - - -def accuracy(*args, **kwargs): - cls = Accuracy(*args, **kwargs) - out = cls._update_ops(*args, **kwargs) - return cls, out + block = eval_program.current_block() + num_infer_chunks, num_label_chunks, num_correct_chunks = executor.run( + eval_program, + fetch_list=[_clone_var_(block, state) for state in self.states]) + num_infer_chunks = num_infer_chunks[0] + num_label_chunks = num_label_chunks[0] + num_correct_chunks = num_correct_chunks[0] + precision = float( + num_correct_chunks) / num_infer_chunks if num_infer_chunks else 0 + recall = float( + num_correct_chunks) / num_label_chunks if num_label_chunks else 0 + f1_score = float(2 * precision * recall) / ( + precision + recall) if num_correct_chunks else 0 + return np.array( + [precision], dtype='float32'), np.array( + [recall], dtype='float32'), np.array( + [f1_score], dtype='float32') diff --git a/python/paddle/v2/fluid/executor.py b/python/paddle/v2/fluid/executor.py index ed1c2c06daa7ede97e138049a1f7044d071c31e8..1d6c594b41a2c295e3818fb119362d1daba1de33 100644 --- a/python/paddle/v2/fluid/executor.py +++ b/python/paddle/v2/fluid/executor.py @@ -1,9 +1,59 @@ -import paddle.v2.fluid.core as core -from paddle.v2.fluid.framework import Block, Program, g_main_program +import numpy as np +import contextlib +from framework import Program, default_main_program +from . import core + +__all__ = ['Executor', 'global_scope', 'scope_guard', 'switch_scope'] g_scope = core.Scope() +def global_scope(): + return g_scope + + +def switch_scope(scope): + global g_scope + ex = g_scope + g_scope = scope + return ex + + +@contextlib.contextmanager +def scope_guard(scope): + ex = switch_scope(scope) + yield + switch_scope(ex) + + +def as_numpy(tensor): + if isinstance(tensor, list): + return [as_numpy(t) for t in tensor] + assert isinstance(tensor, core.LoDTensor) + lod = tensor.lod() + tensor_data = np.array(tensor) + if len(lod) == 0: + ans = tensor_data + else: + raise RuntimeError("LoD Calculate lacks unit tests and buggy") + # elif len(lod) == 1: + # ans = [] + # idx = 0 + # while idx < len(lod) - 1: + # ans.append(tensor_data[lod[idx]:lod[idx + 1]]) + # idx += 1 + # else: + # for l in reversed(lod): + # ans = [] + # idx = 0 + # while idx < len(l) - 1: + # ans.append(tensor_data[l[idx]:l[idx + 1]]) + # idx += 1 + # tensor_data = ans + # ans = tensor_data + return ans + + class Executor(object): def __init__(self, places): if not isinstance(places, list) and not isinstance(places, tuple): @@ -15,7 +65,56 @@ class Executor(object): p.set_place(each) act_places.append(p) - self.executor = core.Executor(act_places) + # TODO(dzhwinter) : consider that our fluid tests all written in + # CUDAPlace(gpu_id), this will be changed in the future + if core.is_compile_gpu(): + core.init_devices(["CPU", "GPU:0"]) + else: + core.init_devices(["CPU"]) + + # TODO(dzhwinter) : only use the first place + self.executor = core.Executor(act_places[0]) + self.places = places + + def aslodtensor(self, data): + def accumulate(data): + if not isinstance(data, list): + return 1 + return sum([accumulate(sub) for sub in data]) + + def parselod(data): + seq_lens = [accumulate(seq) for seq in data] + cur_len = 0 + lod = [cur_len] + for l in seq_lens: + cur_len += l + lod.append(cur_len) + return lod + + assert len(self.places) != 0 + if not isinstance(data, list): + # pure tensor case + tensor = core.LoDTensor() + tensor.set(data, self.places[0]) + return tensor + else: + raise RuntimeError("Current implementation lacks unittests") + # lodtensor case + lod = [] + if not isinstance(data[0], list): + lod.append(parselod(data)) + flattened_data = np.concatenate(data, axis=0).astype("int64") + else: + while isinstance(data[0], list): + lod.append(parselod(seq)) + flattened_data = [item for seq in data for item in seq] + data = flattened_data + flattened_data = np.concatenate(data, axis=0).astype("int64") + flattened_data = flattened_data.reshape([len(flattened_data), 1]) + tensor = core.LoDTensor() + tensor.set(flattened_data, self.places[0]) + tensor.set_lod(lod) + return tensor def run(self, program=None, @@ -23,20 +122,21 @@ class Executor(object): fetch_list=None, feed_var_name='feed', fetch_var_name='fetch', - scope=None): + scope=None, + return_numpy=True): if feed is None: feed = {} if fetch_list is None: fetch_list = [] if program is None: - program = g_main_program + program = default_main_program() if not isinstance(program, Program): raise TypeError() if scope is None: - scope = g_scope + scope = global_scope() program = program.clone() global_block = program.global_block() @@ -52,7 +152,10 @@ class Executor(object): inputs={'X': [feed_var]}, outputs={'Out': [out]}, attrs={'col': i}) - core.set_feed_variable(scope, feed[name], feed_var.name, i) + cur_feed = feed[name] + if not isinstance(cur_feed, core.LoDTensor): + cur_feed = self.aslodtensor(cur_feed) + core.set_feed_variable(scope, cur_feed, feed_var.name, i) fetch_var = global_block.create_var( name=fetch_var_name, @@ -65,8 +168,12 @@ class Executor(object): outputs={'Out': [fetch_var]}, attrs={'col': i}) - self.executor.run(program.desc, scope, 0, True) - return [ + self.executor.run(program.desc, scope, 0, True, True) + outs = [ core.get_fetch_variable(scope, fetch_var_name, i) for i in xrange(len(fetch_list)) ] + + if return_numpy: + outs = as_numpy(outs) + return outs diff --git a/python/paddle/v2/fluid/framework.py b/python/paddle/v2/fluid/framework.py index acca6ba35ced8674d4eec7dc57e41673c90cf8f8..b66a8bce5f4f15539007876c113afd3f878b00bc 100644 --- a/python/paddle/v2/fluid/framework.py +++ b/python/paddle/v2/fluid/framework.py @@ -1,21 +1,104 @@ -import paddle.v2.fluid.core as core -import paddle.v2.fluid.proto.framework_pb2 as framework_pb2 import collections +import contextlib + import numpy as np -import copy + +import proto.framework_pb2 as framework_pb2 +from . import core __all__ = [ 'Block', 'Variable', 'Program', 'Operator', 'default_startup_program', - 'default_main_program' + 'default_main_program', 'program_guard', 'switch_startup_program', + 'switch_main_program' ] +EMPTY_VAR_NAME = core.kEmptyVarName() +TEMP_VAR_NAME = core.kTempVarName() +GRAD_VAR_SUFFIX = core.kGradVarSuffix() +ZERO_VAR_SUFFIX = core.kZeroVarSuffix() + +USE_CPU = core.kUseCPU() +USE_CUDNN = core.kUseMKLDNN() +USE_MKLDNN = core.kUseMKLDNN() + + +def grad_var_name(var_name): + """ + return gradient name for a certain var name + """ + return var_name + GRAD_VAR_SUFFIX + def unique_name(prefix): + """ + Generate unique names with prefix + + Args: + prefix(str): The prefix of return string + + Returns(str): A unique string with the prefix + + """ uid = core.unique_integer(prefix) # unique during whole process. return "_".join([prefix, str(uid)]) +def convert_np_dtype_to_dtype_(np_dtype): + """ + Convert the data type in numpy to the data type in Paddle + Args: + np_dtype(np.dtype): the data type in numpy + + Returns(core.DataType): the data type in Paddle + + """ + dtype = np.dtype(np_dtype) + if dtype == np.float32: + return core.DataType.FP32 + elif dtype == np.float64: + return core.DataType.FP64 + elif dtype == np.float16: + return core.DataType.FP16 + elif dtype == np.int32: + return core.DataType.INT32 + elif dtype == np.int16: + return core.DataType.INT16 + elif dtype == np.int64: + return core.DataType.INT64 + elif dtype == np.bool: + return core.DataType.BOOL + else: + raise ValueError("Not supported numpy dtype " + str(dtype)) + + +def dtype_is_floating(dtype): + """ + Check the data type is floating or not. + Args: + dtype(np.dtype|core.DataType): data type. + Could be numpy format or Paddle format + + Returns(bool): True if data type is a float value + + """ + if not isinstance(dtype, core.DataType): + dtype = convert_np_dtype_to_dtype_(dtype) + + return dtype in [core.DataType.FP16, core.DataType.FP32, core.DataType.FP64] + + def _debug_string_(proto, throw_on_error=True): + """ + Get the debug string of a protobuf message. The message could be not + initialized. + Args: + proto(google.protobuf.message.Message): The protobuf message + throw_on_error(bool): True if raise an error when the protobuf message + is not initialized. + + Returns(str): The debug string of the protobuf message + + """ error_fields = list() if not proto.IsInitialized(error_fields) and throw_on_error: raise ValueError("{0} are not initialized\nThe message is {1}".format( @@ -24,6 +107,38 @@ def _debug_string_(proto, throw_on_error=True): class Variable(object): + """ + Python variable. Every input and output of an operator is a variable. Every + variable belongs to a block. The variable has a name and two variables in + different blocks could have the same name. + + There are many kinds of variables. Please reference the framework.proto for + details. + + Notes: The constructor of Variable should not be invoked directly. Please + use `Block.create_var` to create a variable. + + >>> cur_program = Program() + >>> cur_block = cur_program.current_block() + >>> new_variable = cur_block.create_var( + >>> name="X", shape=[-1, 23, 48], dtype='float32') + + Args: + block(Block): The associated block. It will be passed by + `Block.create_var` automatically. + type(core.VarDesc.VarType): Variable type. Please reference the + framework.proto for details. + shape(tuple|list|None): The shape of variable. -1 means the batch size. + Some kinds of variable do not contain shape, just set it to None. + dtype(np.dtype|core.DataType|str): The data type of variable. + lod_level(int): The level of lod tensor. 0 means there is not a time + series data. + persistable(bool): True if the variable should be saved as check point. + Defaults to False. + stop_gradient(bool): True if the variable will stop to calculate + gradients when backward. Defaults to False. + """ + def __init__(self, block, type=core.VarDesc.VarType.LOD_TENSOR, @@ -66,11 +181,11 @@ class Variable(object): "matched.".format(self.name, old_shape, shape)) if dtype is not None: if not isinstance(dtype, core.DataType): - dtype = Variable._convert_np_dtype_to_dtype_(dtype) + dtype = convert_np_dtype_to_dtype_(dtype) if is_new_var: - self.desc.set_data_type(dtype) + self.desc.set_dtype(dtype) else: - old_dtype = self.data_type + old_dtype = self.dtype if dtype != old_dtype: raise ValueError("Variable {0} has been created before. " "The previous data type is {1}; the new " @@ -107,6 +222,16 @@ class Variable(object): return self.to_string(True) def to_string(self, throw_on_error): + """ + Get debug string. + + Args: + throw_on_error(bool): True if raise an exception when self is not + intialized. + + Returns(str): The debug string. + + """ protostr = self.desc.serialize_to_string() proto = framework_pb2.VarDesc.FromString(str(protostr)) return _debug_string_(proto, throw_on_error) @@ -131,8 +256,8 @@ class Variable(object): return tuple(self.desc.shape()) @property - def data_type(self): - return self.desc.data_type() + def dtype(self): + return self.desc.dtype() @property def lod_level(self): @@ -148,31 +273,13 @@ class Variable(object): uid = core.unique_integer(prefix) # unique during whole process. return "_".join([prefix, str(uid)]) - @staticmethod - def _convert_np_dtype_to_dtype_(np_dtype): - dtype = np.dtype(np_dtype) - if dtype == np.float32: - return core.DataType.FP32 - elif dtype == np.float64: - return core.DataType.FP64 - elif dtype == np.float16: - return core.DataType.FP16 - elif dtype == np.int32: - return core.DataType.INT32 - elif dtype == np.int16: - return core.DataType.INT16 - elif dtype == np.int64: - return core.DataType.INT64 - elif dtype == np.bool: - return core.DataType.BOOL - else: - raise ValueError("Not supported numpy dtype " + str(dtype)) - def get_all_op_protos(): """ Get all registered op proto from PaddlePaddle C++ end. - :return: A list of registered OpProto. + + Returns(list): list of OpProto + """ protostrs = core.get_all_op_protos() ret_values = [] @@ -183,6 +290,10 @@ def get_all_op_protos(): class OpProtoHolder(object): + """ + A global variable to hold all OpProtos from C++ as a map + """ + @classmethod def instance(cls): if not hasattr(cls, '_instance'): @@ -199,12 +310,26 @@ class OpProtoHolder(object): self.op_proto_map[proto.type] = proto def get_op_proto(self, type): + """ + Get OpProto by a type string. + Args: + type(str): The type that operator registered in C++ side. + + Returns(framework_pb2.OpProto): The OpProto + + """ if type not in self.op_proto_map: raise ValueError("Operator \"%s\" has not been registered." % type) return self.op_proto_map[type] class Operator(object): + """ + Python Operator class. The operator represents the build in instructs in a + Block. Users can use the build in instructs to describe their neural + network. + """ + def __init__(self, block, desc, @@ -212,8 +337,36 @@ class Operator(object): inputs=None, outputs=None, attrs=None): + """ + Constructor. + + Notes: The constructor of operator should not be invoked directly. Use + Block.append_op or Block.prepend_op instead. + + >>> cur_program = Program() + >>> cur_block = cur_program.current_block() + >>> # var1 += var2 + var3 + >>> cur_block.append_op(type="sum", + >>> inputs={"X": [var1, var2, var3]}, + >>> outputs={"Out": [var1]}) + + Args: + block(Block): The block has the current operator + desc(core.OpDesc): The protobuf description + type(str): The type of operator. + inputs(dict): The input dictionary. Key is the input parameter name. + Value is a list of variables. + outputs(dict): The output dictionary. Has same format with inputs + attrs(dict): The attributes dictionary. Key is attribute name. Value + is the attribute value. The attribute type should be as same as + the type registered in C++ + """ self.block = block self.desc = desc + # for clone a new operator + self.inputs = inputs + self.outputs = outputs + self.attrs = attrs if len(self.desc.type()) != 0: return if type is None: @@ -224,7 +377,7 @@ class Operator(object): def find_name(var_list, name): for var_name in var_list: - if var_name == name: + if var_list[var_name] is not None and var_name == name: return True return False @@ -244,7 +397,10 @@ class Operator(object): % (in_proto.name, len(in_args))) in_arg_names = [] for arg in in_args: - in_arg_names.append(arg.name) + if isinstance(arg, basestring): + in_arg_names.append(arg) + else: + in_arg_names.append(arg.name) self.desc.set_input(in_proto.name, in_arg_names) else: self.desc.set_input(in_proto.name, []) @@ -285,19 +441,33 @@ class Operator(object): 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): + self.desc.set_serialized_attr( + attr_name, attrs[attr_name].serialize_to_string()) else: self.desc.set_attr(attr_name, attrs[attr_name]) self.desc.check_attrs() no_kernel_op_set = { 'feed', 'fetch', 'save', 'load', 'recurrent', - 'rnn_memory_helper_grad', 'conditional_block', 'while' + 'rnn_memory_helper_grad', 'conditional_block', 'while', 'send', + 'recv' } if type not in no_kernel_op_set: self.desc.infer_var_type(self.block.desc) self.desc.infer_shape(self.block.desc) def to_string(self, throw_on_error): + """ + To debug string. + Args: + throw_on_error(bool): raise exception when self is not initialized + when throw_on_error is True + + Returns(str): The debug string. + + """ protostr = self.desc.serialize_to_string() proto = framework_pb2.OpDesc.FromString(str(protostr)) return _debug_string_(proto, throw_on_error) @@ -312,21 +482,55 @@ class Operator(object): return self.desc.type() def input(self, name): + """ + Get input arguments by the input parameter name + Args: + name(str): The input parameter name + + Returns(list): return the list of argument names associated with the + specific parameter name. + + """ return self.desc.input(name) @property def input_names(self): + """ + Get all input parameter names + Returns(list): return a list of input parameter names + + """ return self.desc.input_names() def output(self, name): + """ + Get output arguments by the output parameter name + Args: + name(str): The output parameter name + + Returns(list): return the list of argument names associated with the + specific parameter name. + + """ return self.desc.output(name) @property def output_names(self): + """ + Get all output parameter names + Returns(list): return a list of output parameter names + + """ return self.desc.output_names() @property def idx(self): + """ + Return the array index of current operator. + Returns(int): The array index in block.ops array + Raises: + ValueError: when the operator is not found. + """ for i, op in enumerate(self.block.ops): if op == self: return i @@ -334,19 +538,57 @@ class Operator(object): "Can't find op itself in it's block. It could be a bug of Paddle.") def has_attr(self, name): + """ + operator has the attribute with name or not. + Args: + name(str): the attribute name + + Returns(bool): True if has this attribute. + + """ return self.desc.has_attr(name) def attr_type(self, name): + """ + Get the type of attribute by attribute name + Args: + name(str): the attribute name + + Returns(core.AttrType): the attribute type + + """ return self.desc.attr_type(name) @property def attr_names(self): + """ + Get all attribute names + Returns(list): The list of attribute name + + """ return self.desc.attr_names() def attr(self, name): + """ + Get attribute by name + Args: + name(str): the attribute name + + Returns(bool|int|str|float|list): The attribute value. The return value + can be any valid attribute type. + + """ return self.desc.attr(name) def block_attr(self, name): + """ + Get the block attribute by name + Args: + name(str): the attribute name + + Returns(int): the block index + + """ return self.desc.block_attr(name) @@ -356,6 +598,7 @@ class Block(object): self.vars = dict() # var_name --> var self.ops = collections.deque() # operator list self.program = program + self.removed_vars = dict() def __str__(self): return self.to_string(True) @@ -384,7 +627,11 @@ class Block(object): return v def all_parameters(self): - return {v for k, v in self.vars.iteritems() if isinstance(v, Parameter)} + return list(self.iter_parameters()) + + def iter_parameters(self): + return (item[1] for item in self.vars.iteritems() + if isinstance(item[1], Parameter)) def create_var(self, *args, **kwargs): var = Variable(self, *args, **kwargs) @@ -408,6 +655,16 @@ class Block(object): self.ops.append(op) return op + def delete_ops(self, ops): + # remove from cpp + # FIXME(typhoonzero): remove only the first occuracy. + 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 prepend_op(self, *args, **kwargs): op_desc = self.desc.prepend_op() op = Operator(self, op_desc, *args, **kwargs) @@ -458,12 +715,45 @@ class Block(object): for index in range(len(self.ops)): assert self.ops[index].desc == ops_in_cpp[index] + def copy_param_info_from(self, other): + """ + Copy the information of parameters from other block + Args: + other(Block): other block + + Returns: + None + """ + if not isinstance(other, Block): + raise TypeError("copy_param_info_from should be invoked with Block") + for p in other.iter_parameters(): + assert isinstance(p, Parameter) + v = self.vars.get(p.name, None) + if v is None: + raise ValueError("copy_param_info_from should be invoked with " + "same topology") + assert isinstance(v, Variable) + new_p = Parameter( + block=self, + shape=v.shape, + dtype=v.dtype, + type=v.type, + lod_level=v.lod_level, + stop_gradient=p.stop_gradient, + trainable=p.trainable, + optimize_attr=p.optimize_attr, + regularizer=p.regularizer, + clip_attr=p.clip_attr, + name=v.name) + self.vars[new_p.name] = new_p + class Program(object): def __init__(self): self.desc = core.ProgramDesc() self.blocks = [Block(self, 0)] self.current_block_idx = 0 + self._seed = 0 def __str__(self): return self.to_string(True) @@ -478,6 +768,7 @@ class Program(object): p.desc = core.ProgramDesc(self.desc) p.blocks = [Block(p, i) for i in xrange(self.desc.num_blocks())] p.sync_with_cpp() + p.copy_param_info_from(self) return p def prune(self, targets): @@ -500,6 +791,13 @@ class Program(object): res.sync_with_cpp() return res + def inference_optimize(self): + res = Program() + res.desc = core.inference_optimize(self.desc) + res.blocks = [Block(res, i) for i in xrange(res.desc.num_blocks())] + res.sync_with_cpp() + return res + @staticmethod def parse_from_string(binary_str): p = Program() @@ -508,6 +806,16 @@ class Program(object): p.sync_with_cpp() return p + @property + def random_seed(self): + return self._seed + + @random_seed.setter + def random_seed(self, seed): + if not isinstance(seed, int): + raise ValueError("Seed must be a integer.") + self._seed = seed + def __repr__(self): return str(self) @@ -538,9 +846,11 @@ class Program(object): self.sync_with_cpp() return param_to_grad_info - def create_block(self): + def create_block(self, parent_idx=None): new_block_idx = len(self.blocks) - self.desc.append_block(self.current_block().desc) + parent = self.current_block() if parent_idx is None else self.block( + parent_idx) + self.desc.append_block(parent.desc) self.current_block_idx = new_block_idx self.blocks.append(Block(self, self.current_block_idx)) return self.current_block() @@ -554,6 +864,24 @@ class Program(object): for block in self.blocks: block.sync_with_cpp() + def copy_param_info_from(self, other): + """ + Copy the information of parameters 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") + self.global_block().copy_param_info_from(other.global_block()) + def list_vars(self): for each_block in self.blocks: for each_var in each_block.vars.itervalues(): @@ -580,15 +908,92 @@ class Parameter(Variable): self.regularizer = kwargs.get('regularizer', None) + self.clip_attr = kwargs.get('clip_attr', None) + # program is a global instance. -g_main_program = Program() -g_startup_program = Program() +_main_program_ = Program() +_startup_program_ = Program() def default_startup_program(): - return g_startup_program + """ + Get default startup program. In startup program, Paddle will initialize + parameters, initialize nccl handle, etc. + + Returns: + Program: startup program + """ + return _startup_program_ def default_main_program(): - return g_main_program + """ + Get default main program. The main program is used for training or testing. + + Returns: + Program: main program + """ + return _main_program_ + + +def switch_main_program(program): + """ + Switch the main program to a new program. + + Args: + program(Program): The new main program + + Returns: + Program: The previous main program + """ + global _main_program_ + prev_program = _main_program_ + _main_program_ = program + return prev_program + + +def switch_startup_program(program): + """ + Switch the startup program to a new program + Args: + program(Program): The new startup program + + Returns: + Program: The previous startup program + """ + global _startup_program_ + prev_program = _startup_program_ + _startup_program_ = program + return prev_program + + +@contextlib.contextmanager +def program_guard(main_program, startup_program=None): + """ + Switch program with `with` statement + + Examples: + >>> with program_guard(Program()): + >>> data = fluid.layers.data(...) + >>> hidden = fluid.layers.fc(...) + + Args: + main_program(Program): New main program inside `with` statement + startup_program(Program): New startup program inside `with` statement. + None means do not change startup program. + + Returns: + None + """ + if not isinstance(main_program, Program): + raise TypeError("main_program should be Program") + main_program = switch_main_program(main_program) + if startup_program is not None: + if not isinstance(startup_program, Program): + raise TypeError("startup_program should be Program") + startup_program = switch_startup_program(startup_program) + yield + switch_main_program(main_program) + if startup_program is not None: + switch_startup_program(startup_program) diff --git a/python/paddle/v2/fluid/initializer.py b/python/paddle/v2/fluid/initializer.py index ded144ecd5db83ce50ca0dc6243fdc52ac0b7a2f..c0839caaf2bb5bc43a76a13b5782cc519a4afe63 100644 --- a/python/paddle/v2/fluid/initializer.py +++ b/python/paddle/v2/fluid/initializer.py @@ -1,10 +1,7 @@ -import paddle.v2.fluid.framework as framework +import framework import numpy as np -__all__ = [ - 'ConstantInitializer', 'UniformInitializer', 'NormalInitializer', - 'XavierInitializer' -] +__all__ = ['Constant', 'Uniform', 'Normal', 'Xavier'] class Initializer(object): @@ -93,7 +90,7 @@ class ConstantInitializer(Initializer): outputs={"Out": var}, attrs={ "shape": var.shape, - "data_type": int(var.data_type), + "dtype": int(var.dtype), "value": self._value }) var.op = op @@ -135,12 +132,14 @@ class UniformInitializer(Initializer): assert isinstance(var, framework.Variable) assert isinstance(block, framework.Block) # Initialization Ops should be prepended and not appended + if self._seed == 0: + self._seed = block.program.random_seed op = block.prepend_op( type="uniform_random", outputs={"Out": var}, attrs={ "shape": var.shape, - "data_type": int(var.data_type), + "dtype": int(var.dtype), "min": self._low, "max": self._high, "seed": self._seed @@ -183,12 +182,14 @@ class NormalInitializer(Initializer): assert isinstance(var, framework.Variable) assert isinstance(block, framework.Block) # Initialization Ops should be prepended and not appended + if self._seed == 0: + self._seed = block.program.random_seed op = block.prepend_op( type="gaussian_random", outputs={"Out": var}, attrs={ "shape": var.shape, - "data_type": int(var.data_type), + "dtype": int(var.dtype), "mean": self._mean, "std": self._std_dev, "seed": self._seed @@ -258,6 +259,9 @@ class XavierInitializer(Initializer): fan_in = f_in if self._fan_in is None else self._fan_in fan_out = f_out if self._fan_out is None else self._fan_out + if self._seed == 0: + self._seed = block.program.random_seed + if self._uniform: limit = np.sqrt(6.0 / float(fan_in + fan_out)) op = block.prepend_op( @@ -265,7 +269,7 @@ class XavierInitializer(Initializer): outputs={"Out": var}, attrs={ "shape": var.shape, - "data_type": int(var.data_type), + "dtype": int(var.dtype), "min": -limit, "max": limit, "seed": self._seed @@ -278,10 +282,112 @@ class XavierInitializer(Initializer): outputs={"Out": var}, attrs={ "shape": var.shape, - "data_type": int(var.data_type), + "dtype": int(var.dtype), "mean": 0.0, "std": std, "seed": self._seed }) var.op = op return op + + +class MSRAInitializer(Initializer): + """Implements the MSRA initializer a.k.a. Kaiming Initializer + + This class implements the weight initialization from the paper + Delving Deep into Rectifiers: Surpassing Human-Level Performance on + ImageNet Classification[1] by Kaiming He, Xiangyu Zhang, Shaoqing Ren + and Jian Sun. This is a robust initialization method that particularly + considers the rectifier nonlinearities. In case of Uniform distribution, + the range is [-x, x], where x = sqrt(6 / fan_in). In case of Normal + distribution, the mean is 0 and the standard deviation + is sqrt(2/ fan_in). + + References: + [1] Delving Deep into Rectifiers: Surpassing Human-Level Performance + on ImageNet Classification + (https://arxiv.org/abs/1502.01852) + """ + + def __init__(self, uniform=True, fan_in=None, seed=0): + """Constructor for MSRAInitializer + + Args: + uniform: whether to use uniform or normal distribution + fan_in: fan_in for MSRAInitializer. If None, it is + inferred from the variable. + seed: random seed + + Note: It is recommended to set fan_in to None for most cases. + """ + assert uniform is not None + assert seed is not None + super(MSRAInitializer, self).__init__() + self._uniform = uniform + self._fan_in = fan_in + self._seed = seed + + def __call__(self, var, block): + """Add MSRA initialization ops for a variable + + Args: + var: Variable that needs to be initialized + block: The block in which initialization ops + should be added + + Returns: + the initialization op + """ + assert isinstance(var, framework.Variable) + assert isinstance(block, framework.Block) + f_in, f_out = self._compute_fans(var) + + # If fan_in is passed, use it + fan_in = f_in if self._fan_in is None else self._fan_in + + if self._seed == 0: + self._seed = block.program.random_seed + + if self._uniform: + limit = np.sqrt(6.0 / float(fan_in)) + op = block.prepend_op( + type="uniform_random", + outputs={"Out": var}, + attrs={ + "shape": var.shape, + "dtype": int(var.dtype), + "min": -limit, + "max": limit, + "seed": self._seed + }) + + else: + std = np.sqrt(2.0 / float(fan_in)) + op = block.prepend_op( + type="gaussian_random", + outputs={"Out": var}, + attrs={ + "shape": var.shape, + "dtype": int(var.dtype), + "mean": 0.0, + "std": std, + "seed": self._seed + }) + var.op = op + return op + + +# We short the class name, since users will use the initializer with the package +# name. The sample code: +# +# import paddle.fluid as fluid +# +# hidden = fluid.layers.fc(..., +# param_attr=ParamAttr(fluid.initializer.Xavier())) +# +# It is no need to add an `Initializer` as the class suffix +Constant = ConstantInitializer +Uniform = UniformInitializer +Normal = NormalInitializer +Xavier = XavierInitializer +MSRA = MSRAInitializer diff --git a/python/paddle/v2/fluid/io.py b/python/paddle/v2/fluid/io.py index 2d070814eef0b099ba71bef223596e30388ac48a..c47ce82aba7fa5ac42ac26cd25fa3ebc93e96cb2 100644 --- a/python/paddle/v2/fluid/io.py +++ b/python/paddle/v2/fluid/io.py @@ -1,12 +1,12 @@ import os import cPickle as pickle -from paddle.v2.fluid.framework import Program, Parameter, g_main_program, \ - Variable +from paddle.v2.fluid.framework import Program, Parameter, default_main_program, Variable __all__ = [ 'save_vars', 'save_params', 'save_persistables', 'load_vars', 'load_params', - 'load_persistables', "save_inference_model", "load_inference_model" + 'load_persistables', "save_inference_model", "load_inference_model", + "get_inference_program" ] @@ -23,7 +23,7 @@ def _clone_var_in_block_(block, var): return block.create_var( name=var.name, shape=var.shape, - dtype=var.data_type, + dtype=var.dtype, type=var.type, lod_level=var.lod_level, persistable=True) @@ -45,7 +45,7 @@ def save_vars(executor, dirname, main_program=None, vars=None, predicate=None): """ if vars is None: if main_program is None: - main_program = g_main_program + main_program = default_main_program() if not isinstance(main_program, Program): raise TypeError("program should be as Program type or None") @@ -97,7 +97,7 @@ def load_vars(executor, dirname, main_program=None, vars=None, predicate=None): :param executor: executor that save variable :param dirname: directory path :param main_program: program. If vars is None, then filter all variables in this - program which fit `predicate`. Default g_program. + program which fit `predicate`. Default default_main_program(). :param predicate: The Predicate describes a callable that returns a variable as a bool. If it returns true, the variables will be loaded. :param vars: variables need to be loaded. If specify vars, program & @@ -106,7 +106,7 @@ def load_vars(executor, dirname, main_program=None, vars=None, predicate=None): """ if vars is None: if main_program is None: - main_program = g_main_program + main_program = default_main_program() if not isinstance(main_program, Program): raise TypeError("program's type should be Program") @@ -151,6 +151,17 @@ def load_persistables(executor, dirname, main_program=None): predicate=is_persistable) +def get_inference_program(target_vars, main_program=None): + if main_program is None: + main_program = default_main_program() + if not isinstance(target_vars, list): + target_vars = [target_vars] + + pruned_program = main_program.prune(targets=target_vars) + inference_program = pruned_program.inference_optimize() + return inference_program + + def save_inference_model(dirname, feeded_var_names, target_vars, @@ -165,25 +176,38 @@ def save_inference_model(dirname, :param target_vars: Variables from which we can get inference results. :param executor: executor that save inference model :param main_program: original program, which will be pruned to build the inference model. - Default g_main_program. + Default default_main_program(). :return: None """ - if main_program is None: - main_program = g_main_program - if not isinstance(target_vars, list): + 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 isinstance(target_vars, Variable): target_vars = [target_vars] + else: + if not (bool(target_vars) and all( + isinstance(var, Variable) for var in target_vars)): + raise ValueError("'target_vars' should be a list of Variable.") + + if main_program is None: + main_program = default_main_program() if not os.path.isdir(dirname): os.makedirs(dirname) - pruned_program = main_program.prune(target_vars) + pruned_program = main_program.prune(targets=target_vars) + inference_program = pruned_program.inference_optimize() fetch_var_names = [v.name for v in target_vars] model_file_name = dirname + "/__model__" with open(model_file_name, "w") as f: pickle.dump({ - "program_desc_str": pruned_program.desc.serialize_to_string(), + "program_desc_str": inference_program.desc.serialize_to_string(), "feed_var_names": feeded_var_names, "fetch_var_names": fetch_var_names }, f, -1) @@ -259,10 +283,10 @@ def get_parameter_value_by_name(name, executor, program=None): :param executor: executor for retrieving the value :param name: the name of the parameter :param program: the program where the variable is found - Default g_main_program. + Default default_main_program(). :return: the LoDTensor for the variable """ if program is None: - program = g_main_program + program = default_main_program() var = program.global_block().var(name) return get_parameter_value(var, executor) diff --git a/python/paddle/v2/fluid/layer_helper.py b/python/paddle/v2/fluid/layer_helper.py index a97e07982bd89be72386970f28a0dd049f82372d..4469f7285efe1c31d0955c6dd4ba3ecac08070af 100644 --- a/python/paddle/v2/fluid/layer_helper.py +++ b/python/paddle/v2/fluid/layer_helper.py @@ -1,10 +1,10 @@ import copy import itertools -from paddle.v2.fluid.framework import Variable, g_main_program, \ - g_startup_program, unique_name, Program -from paddle.v2.fluid.initializer import ConstantInitializer, \ - UniformInitializer, XavierInitializer +from framework import Variable, Parameter, default_main_program, default_startup_program, \ + unique_name, dtype_is_floating +from paddle.v2.fluid.initializer import Constant, Xavier +from param_attr import ParamAttr class LayerHelper(object): @@ -21,19 +21,11 @@ class LayerHelper(object): @property def main_program(self): - prog = self.kwargs.get('main_program', None) - if prog is None: - return g_main_program - else: - return prog + return default_main_program() @property def startup_program(self): - prog = self.kwargs.get('startup_program', None) - if prog is None: - return g_startup_program - else: - return prog + return default_startup_program() def append_op(self, *args, **kwargs): return self.main_program.current_block().append_op(*args, **kwargs) @@ -61,31 +53,15 @@ class LayerHelper(object): @property def param_attr(self): - default = {'name': None, 'initializer': XavierInitializer()} - actual = self.kwargs.get('param_attr', None) - if actual is None: - actual = default - for default_field in default.keys(): - if default_field not in actual: - actual[default_field] = default[default_field] - return actual + return ParamAttr.to_attr(self.kwargs.get('param_attr', None)) @property def bias_attr(self): - default = {'name': None, 'initializer': ConstantInitializer()} - bias_attr = self.kwargs.get('bias_attr', None) - if bias_attr is None: - bias_attr = default - - if isinstance(bias_attr, dict): - for default_field in default.keys(): - if default_field not in bias_attr: - bias_attr[default_field] = default[default_field] - return bias_attr + return ParamAttr.to_attr(self.kwargs.get('bias_attr', None)) def multiple_param_attr(self, length): param_attr = self.param_attr - if isinstance(param_attr, dict): + if isinstance(param_attr, ParamAttr): param_attr = [param_attr] if len(param_attr) != 1 and len(param_attr) != length: @@ -108,23 +84,41 @@ class LayerHelper(object): dtype = None for each in inputs: if dtype is None: - dtype = each.data_type - elif dtype != each.data_type: + dtype = each.dtype + elif dtype != each.dtype: raise ValueError("Data Type mismatch") return dtype - def create_parameter(self, attr, shape, dtype, suffix='w', - initializer=None): + def create_parameter(self, + attr, + shape, + dtype, + is_bias=False, + default_initializer=None): # Deepcopy the attr so that parameters can be shared in program - attr_copy = copy.deepcopy(attr) - if initializer is not None: - attr_copy['initializer'] = initializer - if attr_copy['name'] is None: - attr_copy['name'] = unique_name(".".join([self.name, suffix])) + assert isinstance(attr, ParamAttr) + suffix = 'b' if is_bias else 'w' + + if default_initializer is None: + if is_bias: + attr.set_default_bias_initializer() + else: + attr.set_default_param_initializer() + else: + attr.set_default_initializer(default_initializer) + if attr.name is None: + attr.name = unique_name(".".join([self.name, suffix])) + self.startup_program.global_block().create_parameter( - dtype=dtype, shape=shape, **attr_copy) + dtype=dtype, shape=shape, **attr.to_kwargs(with_initializer=True)) return self.main_program.global_block().create_parameter( - name=attr_copy['name'], dtype=dtype, shape=shape) + dtype=dtype, shape=shape, **attr.to_kwargs()) + + def get_parameter(self, name): + param = self.main_program.global_block().var(name) + if not isinstance(param, Parameter): + raise ValueError("no Parameter name %s found" % name) + return param def create_tmp_variable(self, dtype): return self.main_program.current_block().create_var( @@ -144,7 +138,7 @@ class LayerHelper(object): self.startup_program.global_block().create_var( name=var.name, type=var.type, - dtype=var.data_type, + dtype=var.dtype, shape=var.shape, persistable=True, initializer=initializer) @@ -154,8 +148,10 @@ class LayerHelper(object): Append bias operator and return its output. If the user does not set bias_attr, append_bias_op will return input_var - :param input_var: the input variable. The len(input_var.shape) is larger - or equal than 2. + :param input_var: the input variable. The len(input_var.shape) is + larger or equal than 2. + :bias_initializer: an instance of a subclass of Initializer used to + initialize the bias :param dim_start: :param dim_end: the shape of the bias will be input_var.shape[dim_start:dim_end]. The bias is broadcasted to other @@ -167,8 +163,8 @@ class LayerHelper(object): return input_var b = self.create_parameter( - attr=bias_attr, shape=size, dtype=input_var.data_type, suffix='b') - tmp = self.create_tmp_variable(dtype=input_var.data_type) + attr=bias_attr, shape=size, dtype=input_var.dtype, is_bias=True) + tmp = self.create_tmp_variable(dtype=input_var.dtype) self.append_op( type='elementwise_add', inputs={'X': [input_var], @@ -183,11 +179,24 @@ class LayerHelper(object): return input_var if isinstance(act, basestring): act = {'type': act} - tmp = self.create_tmp_variable(dtype=input_var.data_type) + tmp = self.create_tmp_variable(dtype=input_var.dtype) act_type = act.pop('type') self.append_op( type=act_type, inputs={"X": [input_var]}, - outputs={"Y": [tmp]}, + outputs={"Out": [tmp]}, attrs=act) return tmp + + def _get_default_initializer(self, dtype): + if dtype is None or dtype_is_floating(dtype) is True: + return Xavier() + else: + # For integer and boolean types, initialize with all zeros + return Constant() + + def is_instance(self, param_name, cls): + param = self.kwargs.get(param_name, None) + if not isinstance(param, cls): + raise TypeError("The input {0} parameter of method {1} must be {2}", + param_name, self.layer_type, cls.__name__) diff --git a/python/paddle/v2/fluid/layers.py b/python/paddle/v2/fluid/layers.py deleted file mode 100644 index 1789d2f82a8813331b3610fc69f8447925cd7501..0000000000000000000000000000000000000000 --- a/python/paddle/v2/fluid/layers.py +++ /dev/null @@ -1,1525 +0,0 @@ -import paddle.v2.fluid.core as core -import paddle.v2.fluid.proto.framework_pb2 as framework_pb2 -from paddle.v2.fluid.framework import OpProtoHolder, Variable, Program, \ - Operator -from paddle.v2.fluid.initializer import ConstantInitializer, \ - NormalInitializer -from paddle.v2.fluid.layer_helper import LayerHelper, unique_name -import re -import cStringIO - -__all__ = [ - 'fc', 'data', 'cross_entropy', 'conv2d', 'pool2d', 'embedding', 'concat', - 'StaticRNN', 'cast', 'sequence_conv', 'sequence_pool', 'sums', 'cos_sim', - 'batch_norm', 'accuracy', 'split_lod_tensor' -] - - -def fc(input, - size, - param_attr=None, - bias_attr=None, - name=None, - act=None, - num_flatten_dims=1, - main_program=None, - startup_program=None): - """ - Fully Connected Layer. - - Args: - input: The input tensor to the function - size: The size of the layer - param_attr: The parameters/weights to the FC Layer - bias_attr: The bias parameter for the FC layer - name: Name/alias of the function - act: Activation to be applied to the output of FC layer - num_flatten_dims: Number of columns in input - main_program: Name of the main program that calls this - startup_program: Name of the startup program - - This function can take in multiple inputs and performs the Fully Connected - function (linear transformation) on top of each of them. - So for input x, the output will be : Wx + b. Where W is the parameter, - b the bias and x is the input. - - The function also applies an activation (non-linearity) on top of the - output, if activation is passed in the input. - - All the input variables of this function are passed in as local variables - to the LayerHelper constructor. - - """ - helper = LayerHelper('fc', **locals()) - - dtype = helper.input_dtype() - - mul_results = [] - for input_var, param_attr in helper.iter_inputs_and_params(): - input_shape = input_var.shape - param_shape = [ - reduce(lambda a, b: a * b, input_shape[num_flatten_dims:], 1) - ] + [size] - w = helper.create_parameter( - attr=param_attr, shape=param_shape, dtype=dtype) - tmp = helper.create_tmp_variable(dtype) - helper.append_op( - type="mul", - inputs={ - "X": input_var, - "Y": w, - }, - outputs={"Out": tmp}, - attrs={'x_num_col_dims': num_flatten_dims, - 'y_num_col_dims': 1}) - mul_results.append(tmp) - - # sum - if len(mul_results) == 1: - pre_bias = mul_results[0] - else: - pre_bias = helper.create_tmp_variable(dtype) - helper.append_op( - type="sum", inputs={"X": mul_results}, outputs={"Out": pre_bias}) - # add bias - pre_activation = helper.append_bias_op(pre_bias) - # add activation - return helper.append_activation(pre_activation) - - -def embedding(input, - size, - data_type='float32', - is_sparse=False, - param_attr=None, - main_program=None, - startup_program=None): - """ - Embedding Layer. - - Args: - input: The input to the function - size: The size of the layer - data_type: The type of data : float32, float_16, int etc - is_sparse: A flag that decleares whether the input is sparse - param_attr: Parameters for this layer - main_program: Name of the main program that calls this - startup_program: Name of the startup program - - This function can take in the input (which is a vector of IDs) and - performs a lookup in the lookup_table using these IDs, to result into - the embedding of each ID in the input. - - All the input variables of this function are passed in as local variables - to the LayerHelper constructor. - - """ - helper = LayerHelper('embedding', **locals()) - w = helper.create_parameter( - attr=helper.param_attr, shape=size, dtype=data_type) - tmp = helper.create_tmp_variable(data_type) - helper.append_op( - type='lookup_table', - inputs={'Ids': input, - 'W': w}, - outputs={'Out': tmp}, - attrs={'is_sparse': is_sparse}) - return tmp - - -# TODO(qijun): expose H0 and C0 -def dynamic_lstm(input, - size, - data_type='float32', - param_attr=None, - bias_attr=None, - use_peepholes=True, - is_reverse=False, - gate_activation='sigmoid', - cell_activation='tanh', - candidate_activation='tanh', - main_program=None, - startup_program=None): - helper = LayerHelper('lstm', **locals()) - size = size / 4 - weight = helper.create_parameter( - attr=helper.param_attr, shape=[size, 4 * size], dtype=data_type) - bias_size = [1, 7 * size] - if not use_peepholes: - bias_size[1] = 4 * size - bias = helper.create_parameter( - attr=helper.bias_attr, shape=bias_size, dtype=data_type, suffix='b') - - hidden = helper.create_tmp_variable(data_type) - cell = helper.create_tmp_variable(data_type) - batch_gate = helper.create_tmp_variable(data_type) - batch_cell_pre_act = helper.create_tmp_variable(data_type) - - helper.append_op( - type='lstm', - inputs={'Input': input, - 'Weight': weight, - 'Bias': bias}, - outputs={ - 'Hidden': hidden, - 'Cell': cell, - 'BatchGate': batch_gate, - 'BatchCellPreAct': batch_cell_pre_act - }, - attrs={ - 'use_peepholes': use_peepholes, - 'is_reverse': is_reverse, - 'gate_activation': gate_activation, - 'cell_activation': cell_activation, - 'candidate_activation': candidate_activation - }) - return hidden, cell - - -def data(name, - shape, - data_type='float32', - type=core.VarDesc.VarType.LOD_TENSOR, - append_batch_size=True, - main_program=None, - startup_program=None, - stop_gradient=True): - """ - Data Layer. - - Args: - name: The name/alias of the function - shape: Tuple declaring the shape. - data_type: The type of data : float32, float_16, int etc - type: The output type. By default it is LOD_TENSOR. - append_batch_size: Whether or not to append the data as a batch. - main_program: Name of the main program that calls this - startup_program: Name of the startup program - stop_gradient: A boolean that mentions whether gradient should flow. - - This function takes in input and based on whether data has - to be returned back as a minibatch, it creates the global variable using - the helper functions. The global variables can be accessed by all the - following operations and layers in the graph. - - All the input variables of this function are passed in as local variables - to the LayerHelper constructor. - - """ - helper = LayerHelper('data', **locals()) - shape = list(shape) - for i in xrange(len(shape)): - if shape[i] is None: - shape[i] = -1 - append_batch_size = False - elif shape[i] < 0: - append_batch_size = False - - if append_batch_size: - shape = [-1] + shape # append batch size as -1 - - return helper.create_global_variable( - name=name, - shape=shape, - dtype=data_type, - type=type, - stop_gradient=stop_gradient) - - -def create_tensor(dtype, name=None, main_program=None): - helper = LayerHelper("create_tensor", **locals()) - return helper.create_variable(name=helper.name, dtype=dtype) - - -def _convert_(name): - """ - Formatting. - - Args: - name: The name/alias - - This function takes in a name and converts it to a standard format of - group1_group2. Where as per the regular expression, group1 can have - alphabets and numbers and group2 has capital alphabets. - - """ - s1 = re.sub('(.)([A-Z][a-z]+)', r'\1_\2', name) - return re.sub('([a-z0-9])([A-Z])', r'\1_\2', s1).lower() - - -def _generate_doc_string_(op_proto): - """ - Generate docstring by OpProto - - Args: - op_proto (framework_pb2.OpProto): a protobuf message typed OpProto - - Returns: - str: the document string - """ - - def _type_to_str_(tp): - return framework_pb2.AttrType.Name(tp) - - if not isinstance(op_proto, framework_pb2.OpProto): - raise TypeError("OpProto should be `framework_pb2.OpProto`") - - buf = cStringIO.StringIO() - buf.write(op_proto.comment) - buf.write('\nArgs:\n') - for each_input in op_proto.inputs: - line_begin = ' {0}: '.format(_convert_(each_input.name)) - buf.write(line_begin) - buf.write(each_input.comment) - buf.write('\n') - buf.write(' ' * len(line_begin)) - buf.write('Duplicable: ') - buf.write(str(each_input.duplicable)) - buf.write(' Optional: ') - buf.write(str(each_input.dispensable)) - buf.write('\n') - - for each_attr in op_proto.attrs: - buf.write(' ') - buf.write(each_attr.name) - buf.write(' (') - buf.write(_type_to_str_(each_attr.type)) - buf.write('): ') - buf.write(each_attr.comment) - buf.write('\n') - - if len(op_proto.outputs) != 0: - buf.write('\nReturns:\n') - buf.write(' ') - for each_opt in op_proto.outputs: - if not each_opt.intermediate: - break - buf.write(each_opt.comment) - - return buf.getvalue() - - -def _create_op_func_(op_type): - """ - Create an Operator for a Function. - - Args: - op_type: The name of the operator to be created - - This function takes in the operator type (sigmoid, mean , average etc) and - creates the operator functionality. - - """ - op_proto = OpProtoHolder.instance().get_op_proto(op_type) - not_intermediate_outputs = \ - filter(lambda output: not output.intermediate, op_proto.outputs) - intermediate_outputs = \ - filter(lambda output: output.intermediate, op_proto.outputs) - - if len(not_intermediate_outputs) != 1: - raise ValueError("Only one non intermediate output operator can be", - "automatically generated") - - if not_intermediate_outputs[0].duplicable: - raise ValueError( - "Only non duplicable op can be automatically generated") - - for output in intermediate_outputs: - if output.duplicable: - raise ValueError("The op can be automatically generated only when ", - "all intermediate ops are not duplicable") - - o_name = not_intermediate_outputs[0].name - intermediate_output_names = [output.name for output in intermediate_outputs] - - def infer_and_check_data_type(op_proto, **kwargs): - """ - This function performs the sanity check for data_type and - instance type. - """ - dtype = None - for ipt in op_proto.inputs: - name = _convert_(ipt.name) - val = kwargs.pop(name, []) - if not isinstance(val, list) and not isinstance(val, tuple): - val = [val] - for each in val: - if not isinstance(each, Variable): - raise ValueError("input of {0} must be variable".format( - op_type)) - - if dtype is None: - dtype = each.data_type - elif dtype != each.data_type: - raise ValueError( - "operator {0} must input same dtype".format(op_type)) - - return dtype - - def func(**kwargs): - helper = LayerHelper(op_type, **kwargs) - - dtype = infer_and_check_data_type(op_proto, **kwargs) - - inputs = dict() - for ipt in op_proto.inputs: - name = _convert_(ipt.name) - val = kwargs.pop(name, []) - if not isinstance(val, list) and not isinstance(val, tuple): - val = [val] - inputs[ipt.name] = val - - outputs = dict() - out = helper.create_tmp_variable(dtype=dtype) - outputs[o_name] = [out] - for name in intermediate_output_names: - outputs[name] = [helper.create_tmp_variable(dtype=dtype)] - helper.append_op( - type=op_type, inputs=inputs, outputs=outputs, attrs=kwargs) - return helper.append_activation(out) - - func.__name__ = op_type - globals()[op_type] = func - func.__doc__ = _generate_doc_string_(op_proto) - global __all__ - __all__.append(op_type) - - -_create_op_func_('mean') -_create_op_func_('mul') -_create_op_func_('elementwise_add') -_create_op_func_('dropout') -_create_op_func_('reshape') -_create_op_func_('elementwise_add') -_create_op_func_('sigmoid') -_create_op_func_('scale') -_create_op_func_('reshape') -_create_op_func_('transpose') - - -def fill_constant(data_type, shape, value=None, program=None): - """ - This function creates a tensor , with shape as mentioned in the input and - specified data_type and fills this up with a constant value that - comes in the input. - """ - helper = LayerHelper('fill_constant', **locals()) - out = helper.create_tmp_variable(dtype=data_type) - helper.append_op( - type='fill_constant', - outputs={'Out': [out]}, - attrs={'data_type': data_type, - 'shape': shape, - 'value': value}) - return out - - -def cast(x, data_type, main_program=None): - """ - This function takes in the input with input_data_type - and casts it to the output_data_type as the output. - """ - helper = LayerHelper('cast', **locals()) - out = helper.create_tmp_variable(dtype=data_type) - helper.append_op( - type='cast', - inputs={'X': [x]}, - outputs={'Out': [out]}, - attrs={'in_data_type': x.data_type, - 'out_data_type': out.data_type}) - return out - - -def concat(input, axis, main_program=None, startup_program=None): - """ - This function concats the input along the axis mentioned - and returns that as the output. - """ - helper = LayerHelper('concat', **locals()) - out = helper.create_tmp_variable(dtype=helper.input_dtype()) - helper.append_op( - type='concat', - inputs={'X': input}, - outputs={'Out': [out]}, - attrs={'axis': axis}) - return out - - -def sums(input, main_program=None, startup_program=None): - """ - This function takes in the input and performs the sum operation on it - and returns that as the output. - """ - helper = LayerHelper('sum', **locals()) - out = helper.create_tmp_variable(dtype=helper.input_dtype()) - helper.append_op(type='sum', inputs={'X': input}, outputs={'Out': out}) - return out - - -def assign(input, output, main_program=None): - helper = LayerHelper('assign', **locals()) - helper.append_op( - type='scale', - inputs={'X': [input]}, - outputs={'Out': [output]}, - attrs={'scale': 1.0}) - return output - - -def split_lod_tensor(input, - mask, - level, - main_program=None, - startup_program=None): - helper = LayerHelper('split_lod_tensor', **locals()) - out_true = helper.create_tmp_variable(dtype=input.data_type) - out_false = helper.create_tmp_variable(dtype=input.data_type) - helper.append_op( - type='split_lod_tensor', - inputs={ - 'X': input, - 'Mask': mask, - }, - outputs={'OutTrue': out_true, - 'OutFalse': out_false}, - attrs={'level': level}) - return out_true, out_false - - -def merge_lod_tensor(in_true, - in_false, - x, - mask, - level, - main_program=None, - startup_program=None): - helper = LayerHelper('merge_lod_tensor', **locals()) - out = helper.create_tmp_variable(dtype=x.data_type) - helper.append_op( - type='merge_lod_tensor', - inputs={'X': x, - 'Mask': mask, - 'InTrue': in_true, - 'InFalse': in_false}, - outputs={'Out': out}, - attrs={'level': level}) - return out - - -def cos_sim(X, Y, **kwargs): - """ - This function performs the cosine similarity between two tensors - X and Y and returns that as the output. - """ - helper = LayerHelper('cos_sim', **kwargs) - out = helper.create_tmp_variable(dtype=X.data_type) - xnorm = helper.create_tmp_variable(dtype=X.data_type) - ynorm = helper.create_tmp_variable(dtype=X.data_type) - helper.append_op( - type='cos_sim', - inputs={'X': [X], - 'Y': [Y]}, - outputs={'Out': [out], - 'XNorm': [xnorm], - 'YNorm': [ynorm]}) - return out - - -def cross_entropy(input, label, **kwargs): - """ - This function computes cross_entropy using the input and label. - """ - helper = LayerHelper('cross_entropy', **kwargs) - out = helper.create_tmp_variable(dtype=input.data_type) - helper.append_op( - type='cross_entropy', - inputs={'X': [input], - 'Label': [label]}, - outputs={'Y': [out]}, - attrs=kwargs) - return out - - -def square_error_cost(input, label, **kwargs): - """ - This functions returns the squared error cost using the input and label. - The output is appending the op to do the above. - """ - helper = LayerHelper('square_error_cost', **kwargs) - minus_out = helper.create_tmp_variable(dtype=input.data_type) - helper.append_op( - type='elementwise_sub', - inputs={'X': [input], - 'Y': [label]}, - outputs={'Out': [minus_out]}) - - square_out = helper.create_tmp_variable(dtype=input.data_type) - helper.append_op( - type='square', inputs={'X': [minus_out]}, outputs={'Y': [square_out]}) - return square_out - - -def accuracy(input, label, k=1, **kwargs): - """ - This function computes the accuracy using the input and label. - The output is the top_k inputs and their indices. - """ - helper = LayerHelper("accuracy", **kwargs) - topk_out = helper.create_tmp_variable(dtype=input.data_type) - topk_indices = helper.create_tmp_variable(dtype="int64") - helper.append_op( - type="top_k", - inputs={"X": [input]}, - outputs={"Out": [topk_out], - "Indices": [topk_indices]}, - attrs={"k": k}) - acc_out_dtype = kwargs.get("out_dtype", "float32") - acc_out = helper.create_tmp_variable(dtype="float32") - correct = helper.create_tmp_variable(dtype="int64") - total = helper.create_tmp_variable(dtype="int64") - helper.append_op( - type="accuracy", - inputs={ - "Out": [topk_out], - "Indices": [topk_indices], - "Label": [label] - }, - outputs={ - "Accuracy": [acc_out], - "Correct": [correct], - "Total": [total], - }) - return acc_out - - -def sequence_conv(input, - num_filters, - filter_size=3, - filter_stride=1, - act=None, - padding=None, - bias_attr=None, - param_attr=None, - main_program=None, - startup_program=None): - """ - This function creates the op for sequence_conv, using the inputs and - other convolutional configurations for the filters and stride as given - in the input parameters to the function. - """ - # FIXME(dzh) : want to unify the argument of python layer - # function. So we ignore some unecessary attributes. - # such as, padding_trainable, context_start. - - helper = LayerHelper('sequence_conv', **locals()) - dtype = helper.input_dtype() - - filter_shape = [filter_size * input.shape[1], num_filters] - filter = helper.create_parameter( - attr=helper.param_attr, shape=filter_shape, dtype=dtype) - pre_bias = helper.create_tmp_variable(dtype) - - helper.append_op( - type='sequence_conv', - inputs={ - 'X': [input], - 'Filter': [filter], - }, - outputs={"Out": pre_bias}, - attrs={ - 'contextStride': filter_stride, - 'contextStart': -int(filter_size / 2), - 'contextLength': filter_size - }) - pre_act = helper.append_bias_op(pre_bias) - return helper.append_activation(pre_act) - - -def conv2d(input, - num_filters, - name=None, - filter_size=[1, 1], - act=None, - groups=None, - stride=[1, 1], - padding=None, - bias_attr=None, - param_attr=None, - main_program=None, - startup_program=None): - """ - This function creates the op for a 2-dimensional Convolution. - This is performed using the parameters of filters(size, dimensionality etc) - , stride and other configurations for a Convolution operation. - This funciton can also append an activation on top of the - conv-2d output, if mentioned in the input parameters. - """ - helper = LayerHelper('conv2d', **locals()) - dtype = helper.input_dtype() - - num_channels = input.shape[1] - if groups is None: - num_filter_channels = num_channels - else: - if num_channels % groups is not 0: - raise ValueError("num_channels must be divisible by groups.") - num_filter_channels = num_channels / groups - - if isinstance(filter_size, int): - filter_size = [filter_size, filter_size] - if isinstance(stride, int): - stride = [stride, stride] - if isinstance(padding, int): - padding = [padding, padding] - - input_shape = input.shape - filter_shape = [num_filters, num_filter_channels] + filter_size - - std = (2.0 / (filter_size[0]**2 * num_channels))**0.5 - filter = helper.create_parameter( - attr=helper.param_attr, - shape=filter_shape, - dtype=dtype, - initializer=NormalInitializer(0.0, std, 0)) - pre_bias = helper.create_tmp_variable(dtype) - - helper.append_op( - type='conv2d', - inputs={ - 'Input': input, - 'Filter': filter, - }, - outputs={"Output": pre_bias}, - attrs={'strides': stride, - 'paddings': padding, - 'groups': groups}) - - pre_act = helper.append_bias_op(pre_bias, dim_start=1, dim_end=2) - - return helper.append_activation(pre_act) - - -def sequence_pool(input, pool_type, **kwargs): - """ - This function add the operator for sequence pooling. - This is applied on top of the input using pool_type mentioned - in the parameters. - """ - helper = LayerHelper('sequence_pool', input=input, **kwargs) - dtype = helper.input_dtype() - pool_out = helper.create_tmp_variable(dtype) - max_index = helper.create_tmp_variable(dtype) - - helper.append_op( - type="sequence_pool", - inputs={"X": input}, - outputs={"Out": pool_out, - "MaxIndex": max_index}, - attrs={"pooltype": pool_type.upper()}) - - return pool_out - - -def pool2d(input, - pool_size, - pool_type, - pool_stride=[1, 1], - pool_padding=[0, 0], - global_pooling=False, - main_program=None, - startup_program=None): - """ - This function adds the operator for pooling in 2 dimensions, using the - pooling configurations mentioned in input parameters. - """ - if pool_type not in ["max", "avg"]: - raise ValueError( - "Unknown pool_type: '%s'. It can only be 'max' or 'avg'.", - str(pool_type)) - if isinstance(pool_size, int): - pool_size = [pool_size, pool_size] - if isinstance(pool_stride, int): - pool_stride = [pool_stride, pool_stride] - if isinstance(pool_padding, int): - pool_padding = [pool_padding, pool_padding] - - helper = LayerHelper('pool2d', **locals()) - dtype = helper.input_dtype() - pool_out = helper.create_tmp_variable(dtype) - - helper.append_op( - type="pool2d", - inputs={"X": input}, - outputs={"Out": pool_out}, - attrs={ - "pooling_type": pool_type, - "ksize": pool_size, - "global_pooling": global_pooling, - "strides": pool_stride, - "paddings": pool_padding - }) - - return pool_out - - -def batch_norm(input, - act=None, - is_test=False, - momentum=0.9, - epsilon=1e-05, - param_attr=None, - bias_attr=None, - data_layout='NCHW', - main_program=None, - startup_program=None): - """ - This function helps create an operator to implement - the BatchNorm layer using the configurations from the input parameters. - """ - helper = LayerHelper('batch_norm', **locals()) - dtype = helper.input_dtype() - - input_shape = input.shape - if data_layout == 'NCHW': - channel_num = input_shape[1] - else: - if data_layout == 'NHWC': - channel_num = input_shape[-1] - else: - raise ValueError("unsupported data layout:" + data_layout) - - param_shape = [channel_num] - - # create parameter - scale = helper.create_parameter( - attr=helper.param_attr, - shape=param_shape, - dtype=dtype, - initializer=ConstantInitializer(1.0)) - bias = helper.create_parameter( - attr=helper.param_attr, - shape=param_shape, - dtype=dtype, - initializer=ConstantInitializer(0.0)) - - mean = helper.create_global_variable( - dtype=input.data_type, shape=param_shape, persistable=True) - helper.set_variable_initializer( - var=mean, initializer=ConstantInitializer(0.0)) - - variance = helper.create_global_variable( - dtype=input.data_type, shape=param_shape, persistable=True) - helper.set_variable_initializer( - var=variance, initializer=ConstantInitializer(1.0)) - - # create output - # mean and mean_out share the same memory - mean_out = mean - # variance and variance out share the same memory - variance_out = variance - saved_mean = helper.create_tmp_variable(dtype) - saved_variance = helper.create_tmp_variable(dtype) - - batch_norm_out = helper.create_tmp_variable(dtype) - - helper.append_op( - type="batch_norm", - inputs={ - "X": input, - "Scale": scale, - "Bias": bias, - "Mean": mean, - "Variance": variance - }, - outputs={ - "Y": batch_norm_out, - "MeanOut": mean_out, - "VarianceOut": variance_out, - "SavedMean": saved_mean, - "SavedVariance": saved_variance - }, - attrs={"momentum": momentum, - "epsilon": epsilon, - "is_test": is_test}) - - return helper.append_activation(batch_norm_out) - - -def beam_search_decode(ids, scores, main_program=None, startup_program=None): - helper = LayerHelper('beam_search_decode', **locals()) - sentence_ids = helper.create_tmp_variable(dtype=ids.data_type) - sentence_scores = helper.create_tmp_variable(dtype=ids.data_type) - - helper.append_op( - type="beam_search_decode", - inputs={"Ids": ids, - "Scores": scores}, - outputs={ - "SentenceIds": sentence_ids, - "SentenceScores": sentence_scores - }) - - return sentence_ids, sentence_scores - - -class BlockGuard(object): - """ - BlockGuard class. - - BlockGuard class is used to create a sub-block in a program by - using the Python `with` keyword. - """ - - def __init__(self, main_program): - if not isinstance(main_program, Program): - raise TypeError("BlockGuard takes a program") - self.main_program = main_program - - def __enter__(self): - self.main_program.create_block() - - def __exit__(self, exc_type, exc_val, exc_tb): - self.main_program.rollback() - if exc_type is not None: - return False # re-raise exception - return True - - -class StaticRNNGuard(BlockGuard): - """ - StaticRNNGuard class. - - StaticRNNGuard class is used to create a StaticRNN block in a program. - """ - - def __init__(self, rnn): - if not isinstance(rnn, StaticRNN): - raise TypeError("StaticRNNGuard takes a StaticRNN") - super(StaticRNNGuard, self).__init__(rnn.helper.main_program) - self.rnn = rnn - - def __enter__(self): - self.rnn.status = StaticRNN.IN_RNN_BLOCK - return super(StaticRNNGuard, self).__enter__() - - def __exit__(self, exc_type, exc_val, exc_tb): - if exc_type is not None: - return False - self.rnn.status = StaticRNN.AFTER_RNN_BLOCK - self.rnn.complete_rnn_op() - return super(StaticRNNGuard, self).__exit__(exc_type, exc_val, exc_tb) - - -class StaticRNNMemoryLink(object): - """ - StaticRNNMemoryLink class. - - Args: - init: the initial variable for Memory - init: Variable - pre_mem: the memory variable in previous time step - pre_mem: Variable - mem: the memory variable in current time step - mem: Variable - - StaticRNNMemoryLink class is used to create a link between two - memory cells of a StaticRNN. - """ - - def __init__(self, init, pre_mem, mem=None): - self.init = init - self.pre_mem = pre_mem - self.mem = mem - - -class StaticRNN(object): - """ - StaticRNN class. - - StaticRNN class is used to create a StaticRNN. The RNN will have its - own parameters like inputs, outputs, memories, status and length. - """ - BEFORE_RNN_BLOCK = 0 - IN_RNN_BLOCK = 1 - AFTER_RNN_BLOCK = 2 - - def __init__(self, name=None, main_program=None): - self.helper = LayerHelper( - "static_rnn", name=name, main_program=main_program) - self.memories = {} # memory map, from pre_mem.name --> MemoryLink - self.inputs = [] # input variable list in current block - self.outputs = [] # output variable list in parent block - self.status = StaticRNN.BEFORE_RNN_BLOCK # status flag. - # sequence length, since it is a static RNN, sequence length are fixed. - self.seq_len = None - - def step(self): - return StaticRNNGuard(self) - - def _assert_in_rnn_block_(self, method): - if self.status != StaticRNN.IN_RNN_BLOCK: - raise ValueError("You must invoke {0} in rnn block".format(method)) - - def memory(self, - init=None, - shape=None, - batch_ref=None, - init_value=0.0, - init_batch_dim_idx=0, - ref_batch_dim_idx=1): - """ - Args: - init: boot memory, if not set, a shape, batch_ref must be provided - shape: shape of the boot memory - batch_ref: batch size reference variable - init_value: the init value of boot memory - init_batch_dim_idx: the index of batch size in init's dimension - ref_batch_dim_idx: the index of batch size in batch_ref's dimension - """ - self._assert_in_rnn_block_('memory') - if init is None: - if shape is None or batch_ref is None: - raise ValueError( - "if init is None, memory at least need shape and batch_ref") - parent_block = self.parent_block() - var_name = unique_name("@".join([self.helper.name, "memory_boot"])) - boot_var = parent_block.create_var( - name=var_name, - shape=shape, - dtype=batch_ref.data_type, - persistable=False) - - parent_block.append_op( - type="fill_constant_batch_size_like", - inputs={'Input': [batch_ref]}, - outputs={'Out': [boot_var]}, - attrs={ - 'value': init_value, - 'shape': boot_var.shape, - 'data_type': boot_var.data_type, - 'input_dim_idx': ref_batch_dim_idx, - 'output_dim_idx': init_batch_dim_idx - }) - - return self.memory(init=boot_var) - else: - pre_mem = self.helper.create_variable( - name=unique_name("@".join([self.helper.name, "mem"])), - dtype=init.data_type, - shape=init.shape) - self.memories[pre_mem.name] = StaticRNNMemoryLink( - init=init, pre_mem=pre_mem) - return pre_mem - - def step_input(self, x): - self._assert_in_rnn_block_('step_input') - if not isinstance(x, Variable): - raise TypeError("step input takes a Variable") - if self.seq_len is None: - self.seq_len = x.shape[0] - elif self.seq_len != x.shape[0]: - raise ValueError("Static RNN only take fix seq_len input") - - ipt = self.helper.create_variable( - name=x.name, - dtype=x.data_type, - shape=list(x.shape[1:]), - type=x.type) - self.inputs.append(ipt) - return ipt - - def step_output(self, o): - self._assert_in_rnn_block_('step_output') - if not isinstance(o, Variable): - raise TypeError("step output takes a Variable") - - tmp_o = self.helper.create_tmp_variable(dtype=o.data_type) - self.helper.append_op( - type='rnn_memory_helper', - inputs={'X': [o]}, - outputs={'Out': tmp_o}, - attrs={'data_type': o.data_type}) - - out_var = self.parent_block().create_var( - name=tmp_o.name, - shape=[self.seq_len] + list(tmp_o.shape), - dtype=tmp_o.data_type) - - self.outputs.append(out_var) - - def output(self, *outputs): - for each in outputs: - self.step_output(each) - - def update_memory(self, mem, var): - if not isinstance(mem, Variable) or not isinstance(var, Variable): - raise TypeError("update memory should take variables") - self.memories[mem.name].mem = var - - def parent_block(self): - prog = self.helper.main_program - parent_idx = prog.current_block().parent_idx - assert parent_idx >= 0 - parent_block = prog.block(parent_idx) - return parent_block - - def __call__(self, *args, **kwargs): - if self.status != StaticRNN.AFTER_RNN_BLOCK: - raise ValueError("RNN output can only be retrieved after rnn block") - if len(self.outputs) == 0: - raise ValueError("RNN has no output") - elif len(self.outputs) == 1: - return self.outputs[0] - else: - return self.outputs - - def complete_rnn_op(self): - main_program = self.helper.main_program - rnn_block = main_program.current_block() - parent_block = self.parent_block() - - local_inputs = set() - - for op in rnn_block.ops: - assert isinstance(op, Operator) - for oname in op.output_names: - for out_var_name in op.output(oname): - local_inputs.add(out_var_name) - - for var in self.inputs: - local_inputs.add(var.name) - for m in self.memories: - local_inputs.add(m) - - params = list() - for op in rnn_block.ops: - assert isinstance(op, Operator) - for iname in op.input_names: - for in_var_name in op.input(iname): - if in_var_name not in local_inputs: - params.append(in_var_name) - - parameters = [parent_block.var(name) for name in params] - - step_scope = parent_block.create_var( - type=core.VarDesc.VarType.STEP_SCOPES) - - inlinks = [parent_block.var(i.name) for i in self.inputs] - outlinks = self.outputs - - boot_memories = [] - pre_memories = [] - memories = [] - for _, mem in self.memories.iteritems(): - boot_memories.append(mem.init) - pre_memories.append(mem.pre_mem.name) - mem_var = rnn_block.var(mem.mem.name) - assert isinstance(mem_var, Variable) - new_mem = self.helper.create_tmp_variable(dtype=mem_var.data_type) - - rnn_block.append_op( - type='rnn_memory_helper', - inputs={'X': [mem_var]}, - outputs={'Out': [new_mem]}, - attrs={'data_type': mem_var.data_type}) - - memories.append(new_mem.name) - - parent_block.append_op( - type='recurrent', - inputs={ - 'inputs': inlinks, - 'initial_states': boot_memories, - 'parameters': parameters - }, - outputs={'outputs': outlinks, - 'step_scopes': [step_scope]}, - attrs={ - 'ex_states': pre_memories, - 'states': memories, - 'step_block': rnn_block - }) - - -class WhileGuard(BlockGuard): - def __init__(self, while_op): - if not isinstance(while_op, While): - raise TypeError("WhileGuard takes a while op") - super(WhileGuard, self).__init__(while_op.helper.main_program) - self.while_op = while_op - - def __enter__(self): - self.while_op.status = While.IN_WHILE_BLOCK - return super(WhileGuard, self).__enter__() - - def __exit__(self, exc_type, exc_val, exc_tb): - if exc_type is not None: - return False - self.while_op.status = While.AFTER_WHILE_BLOCK - self.while_op.complete() - return super(WhileGuard, self).__exit__(exc_type, exc_val, exc_tb) - - -class While(object): - BEFORE_WHILE_BLOCK = 0 - IN_WHILE_BLOCK = 1 - AFTER_WHILE_BLOCK = 2 - - def __init__(self, cond, name=None, main_program=None): - self.helper = LayerHelper("while", name=name, main_program=main_program) - self.status = While.BEFORE_WHILE_BLOCK - if not isinstance(cond, Variable): - raise TypeError("condition should be a variable") - assert isinstance(cond, Variable) - if cond.data_type != core.DataType.BOOL: - raise TypeError("condition should be a bool variable") - if reduce(lambda a, b: a * b, cond.shape, 1) != 1: - raise TypeError("condition should be a bool scalar") - self.cond_var = cond - - def block(self): - return WhileGuard(self) - - def complete(self): - main_program = self.helper.main_program - while_block = main_program.current_block() - parent_block = main_program.block(main_program.current_block() - .parent_idx) - - inner_outputs = {self.cond_var.name} - x_name_list = set() - for op in while_block.ops: - for iname in op.input_names: - for in_var_name in op.input(iname): - if in_var_name not in inner_outputs: - x_name_list.add(in_var_name) - - for oname in op.output_names: - for out_var_name in op.output(oname): - inner_outputs.add(out_var_name) - - out_vars = [] - for inner_out_name in inner_outputs: - if inner_out_name in parent_block.vars: - out_vars.append(parent_block.var(inner_out_name)) - - step_scope = parent_block.create_var( - type=core.VarDesc.VarType.STEP_SCOPES) - - parent_block.append_op( - type='while', - inputs={ - 'X': [parent_block.var(x_name) for x_name in x_name_list], - 'Condition': [self.cond_var] - }, - outputs={'Out': out_vars, - 'StepScopes': [step_scope]}, - attrs={'step_block': while_block}) - - -def lstm(x, - c_pre_init, - hidden_dim, - forget_bias=None, - main_program=None, - startup_program=None): - """ - This function helps create an operator for the LSTM (Long Short Term - Memory) cell that can be used inside an RNN. - """ - helper = LayerHelper('lstm_unit', **locals()) - rnn = StaticRNN() - with rnn.step(): - c_pre = rnn.memory(init=c_pre_init) - x_t = rnn.step_input(x) - - before_fc = concat( - input=[x_t, c_pre], - axis=1, - main_program=main_program, - startup_program=startup_program) - after_fc = fc(input=before_fc, - size=hidden_dim * 4, - main_program=main_program, - startup_program=startup_program) - - data_type = x.data_type - c = helper.create_tmp_variable(data_type) - h = helper.create_tmp_variable(data_type) - - helper.append_op( - type='lstm_unit', - inputs={"X": after_fc, - "C_prev": c_pre}, - outputs={"C": c, - "H": h}, - attrs={"forget_bias": forget_bias}) - - rnn.update_memory(c_pre, c) - rnn.output(h) - - return rnn() - - -def lod_rank_table(x, level=0, main_program=None): - """ - This function creates an operator for creating a LOD_RANK_TABLE - using the input x. - """ - helper = LayerHelper("lod_rank_table", **locals()) - table = helper.create_variable( - type=core.VarDesc.VarType.LOD_RANK_TABLE, - name=unique_name("lod_rank_table")) - helper.append_op( - type='lod_rank_table', - inputs={'X': x}, - outputs={'Out': table}, - attrs={'level': level}) - return table - - -def lod_tensor_to_array(x, table, main_program=None): - """ - This function creates an operator to convert an LOD_Tensor to - an array. - """ - helper = LayerHelper("lod_tensor_to_array", **locals()) - array = helper.create_variable( - name=unique_name("lod_tensor_to_array"), - type=core.VarDesc.VarType.LOD_TENSOR_ARRAY, - dtype=x.data_type) - helper.append_op( - type='lod_tensor_to_array', - inputs={'X': x, - 'RankTable': table}, - outputs={'Out': array}) - return array - - -def array_to_lod_tensor(x, table, main_program=None): - """ - This function creates an operator to convert an array to a - LOD_Tensor. - """ - helper = LayerHelper("array_to_lod_tensor", **locals()) - tmp = helper.create_tmp_variable(dtype=x.data_type) - helper.append_op( - type="array_to_lod_tensor", - inputs={'X': x, - 'RankTable': table}, - outputs={'Out': tmp}) - return tmp - - -def fill_constant(shape, dtype, value, main_program=None): - """ - This function creates a tensor , with shape as mentioned in the input and - specified data_type and fills this up with a constant value that - comes in the input. It also sets the stop_gradient to be True. - """ - helper = LayerHelper("fill_constant", **locals()) - out = helper.create_tmp_variable(dtype=dtype) - helper.append_op( - type='fill_constant', - inputs={}, - outputs={'Out': [out]}, - attrs={ - 'shape': shape, - 'data_type': out.data_type, - 'value': float(value) - }) - out.stop_gradient = True - return out - - -def ones(shape, dtype, main_program=None): - """ - This function performs the same function as fill_constant() declared above - with the constant value being 1.0. - """ - return fill_constant(value=1.0, **locals()) - - -def zeros(shape, dtype, main_program=None): - """ - This function performs the same function as fill_constant() declared above - with the constant value being 0.0. - """ - return fill_constant(value=0.0, **locals()) - - -def increment(x, value=1.0, in_place=True, main_program=None): - """ - This function creates an operator to increment each value in the input - `x` by an amount: `value` as mentioned in the input parameter. This - operation is performed in-place by default. - """ - helper = LayerHelper("increment", **locals()) - if not in_place: - out = helper.create_tmp_variable(dtype=x.data_type) - else: - out = x - helper.append_op( - type='increment', - inputs={'X': [x]}, - outputs={'Out': [out]}, - attrs={'step': value}) - return out - - -def array_write(x, i, array=None, main_program=None): - """ - This function creates an operator to write the data out as a - LOD_TENSOR_ARRAY. - """ - helper = LayerHelper('array_write', **locals()) - if array is None: - array = helper.create_variable( - name="{0}.out".format(helper.name), - type=core.VarDesc.VarType.LOD_TENSOR_ARRAY, - dtype=x.data_type) - helper.append_op( - type='write_to_array', - inputs={'X': [x], - 'I': [i]}, - outputs={'Out': [array]}) - return array - - -def create_array(dtype, main_program=None): - helper = LayerHelper("array", **locals()) - return helper.create_variable( - name="{0}.out".format(helper.name), - type=core.VarDesc.VarType.LOD_TENSOR_ARRAY, - dtype=dtype) - - -def less_than(x, y, cond=None, main_program=None): - helper = LayerHelper("less_than", **locals()) - if cond is None: - cond = helper.create_tmp_variable(dtype='bool') - cond.stop_gradient = True - - helper.append_op( - type='less_than', inputs={'X': [x], - 'Y': [y]}, outputs={'Out': [cond]}) - return cond - - -def array_read(array, i, main_program=None): - """ - This function creates an operator to read the data in as a - LOD_TENSOR_ARRAY. - """ - helper = LayerHelper('array_read', **locals()) - if not isinstance( - array, - Variable) or array.type != core.VarDesc.VarType.LOD_TENSOR_ARRAY: - raise TypeError("array should be tensor array vairable") - out = helper.create_tmp_variable(dtype=array.data_type) - helper.append_op( - type='read_from_array', - inputs={'X': [array], - 'I': [i]}, - outputs={'Out': [out]}) - return out - - -def shrink_memory(x, i, table, main_program=None): - """ - This function creates an operator to shrink_rnn_memory using the RankTable - as mentioned in the input parameter. - """ - helper = LayerHelper('shrink_memory', **locals()) - out = helper.create_tmp_variable(dtype=x.data_type) - helper.append_op( - type='shrink_rnn_memory', - inputs={'X': [x], - 'I': [i], - 'RankTable': [table]}, - outputs={'Out': [out]}, - attrs={}) - return out - - -def array_length(array, main_program=None): - """ - This function creates an operator to find the length of the - LOD_TENSOR_ARRAY. - """ - helper = LayerHelper('array_length', **locals()) - tmp = helper.create_tmp_variable(dtype='int64') - tmp.stop_gradient = True - helper.append_op( - type='lod_array_length', inputs={'X': [array]}, outputs={'Out': [tmp]}) - return tmp - - -class ConditionalBlockGuard(BlockGuard): - def __init__(self, block): - if not isinstance(block, ConditionalBlock): - raise TypeError("block should be conditional block") - super(ConditionalBlockGuard, self).__init__(block.helper.main_program) - self.block = block - - def __enter__(self): - return super(ConditionalBlockGuard, self).__enter__() - - def __exit__(self, exc_type, exc_val, exc_tb): - self.block.complete() - return super(ConditionalBlockGuard, self).__exit__(exc_type, exc_val, - exc_tb) - - -class ConditionalBlock(object): - def __init__(self, inputs, name=None, main_program=None): - for each_input in inputs: - if not isinstance(each_input, Variable): - raise TypeError("Each input should be variable") - self.inputs = inputs - self.helper = LayerHelper( - 'conditional_block', name=name, main_program=main_program) - - def block(self): - return ConditionalBlockGuard(self) - - def complete(self): - inside_block = self.helper.main_program.current_block() - parent_block = self.helper.main_program.block(inside_block.parent_idx) - - intermediate = set() - params = set() - - for each_op in inside_block.ops: - assert isinstance(each_op, Operator) - for iname in each_op.input_names: - for in_var_name in each_op.input(iname): - if in_var_name not in intermediate: - params.add(in_var_name) - - for oname in each_op.output_names: - for out_var_name in each_op.output(oname): - intermediate.add(out_var_name) - input_set = set([ipt.name for ipt in self.inputs]) - - param_list = [ - parent_block.var(each_name) for each_name in params - if each_name not in input_set - ] - - out_list = [ - parent_block.var(var_name) for var_name in parent_block.vars - if var_name not in intermediate - ] - - step_scope = parent_block.create_var( - type=core.VarDesc.VarType.STEP_SCOPES) - parent_block.append_op( - type='conditional_block', - inputs={ - 'X': self.inputs, - 'Params': param_list, - }, - outputs={'Out': out_list, - 'Scope': [step_scope]}, - attrs={'block': inside_block}) diff --git a/python/paddle/v2/fluid/layers/__init__.py b/python/paddle/v2/fluid/layers/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..249f570e13b7a1b50397fb971d1c6f77e0359a5e --- /dev/null +++ b/python/paddle/v2/fluid/layers/__init__.py @@ -0,0 +1,17 @@ +import ops +from ops import * +import nn +from nn import * +import io +from io import * +import tensor +from tensor import * +import control_flow +from control_flow import * + +__all__ = [] +__all__ += nn.__all__ +__all__ += io.__all__ +__all__ += tensor.__all__ +__all__ += control_flow.__all__ +__all__ += ops.__all__ diff --git a/python/paddle/v2/fluid/layers/control_flow.py b/python/paddle/v2/fluid/layers/control_flow.py new file mode 100644 index 0000000000000000000000000000000000000000..08c52390e979b002d8d2b69a89c1e513aefeba76 --- /dev/null +++ b/python/paddle/v2/fluid/layers/control_flow.py @@ -0,0 +1,1141 @@ +from ..layer_helper import LayerHelper, unique_name +from ..framework import Program, Variable, Operator +from .. import core +from tensor import assign, fill_constant +import contextlib +from ..registry import autodoc + +__all__ = [ + 'split_lod_tensor', 'merge_lod_tensor', 'BlockGuard', 'StaticRNNGuard', + 'StaticRNNMemoryLink', 'WhileGuard', 'While', 'lod_rank_table', + 'max_sequence_len', 'topk', 'lod_tensor_to_array', 'array_to_lod_tensor', + 'increment', 'array_write', 'create_array', 'less_than', 'array_read', + 'shrink_memory', 'array_length', 'IfElse', 'DynamicRNN', 'ConditionalBlock', + 'StaticRNN', 'reorder_lod_tensor_by_rank' +] + + +def split_lod_tensor(input, mask, level=0): + helper = LayerHelper('split_lod_tensor', **locals()) + out_true = helper.create_tmp_variable(dtype=input.dtype) + out_false = helper.create_tmp_variable(dtype=input.dtype) + helper.append_op( + type='split_lod_tensor', + inputs={ + 'X': input, + 'Mask': mask, + }, + outputs={'OutTrue': out_true, + 'OutFalse': out_false}, + attrs={'level': level}) + return out_true, out_false + + +def merge_lod_tensor(in_true, in_false, x, mask, level=0): + helper = LayerHelper('merge_lod_tensor', **locals()) + out = helper.create_tmp_variable(dtype=in_true.dtype) + helper.append_op( + type='merge_lod_tensor', + inputs={'X': x, + 'Mask': mask, + 'InTrue': in_true, + 'InFalse': in_false}, + outputs={'Out': out}, + attrs={'level': level}) + return out + + +class BlockGuard(object): + """ + BlockGuard class. + + BlockGuard class is used to create a sub-block in a program by + using the Python `with` keyword. + """ + + def __init__(self, main_program): + if not isinstance(main_program, Program): + raise TypeError("BlockGuard takes a program") + self.main_program = main_program + + def __enter__(self): + self.main_program.create_block() + + def __exit__(self, exc_type, exc_val, exc_tb): + self.main_program.rollback() + if exc_type is not None: + return False # re-raise exception + return True + + +class StaticRNNGuard(BlockGuard): + """ + StaticRNNGuard class. + + StaticRNNGuard class is used to create a StaticRNN block in a program. + """ + + def __init__(self, rnn): + if not isinstance(rnn, StaticRNN): + raise TypeError("StaticRNNGuard takes a StaticRNN") + super(StaticRNNGuard, self).__init__(rnn.helper.main_program) + self.rnn = rnn + + def __enter__(self): + self.rnn.status = StaticRNN.IN_RNN_BLOCK + return super(StaticRNNGuard, self).__enter__() + + def __exit__(self, exc_type, exc_val, exc_tb): + if exc_type is not None: + return False + self.rnn.status = StaticRNN.AFTER_RNN_BLOCK + self.rnn.complete_rnn_op() + return super(StaticRNNGuard, self).__exit__(exc_type, exc_val, exc_tb) + + +class StaticRNNMemoryLink(object): + """ + StaticRNNMemoryLink class. + + Args: + init: the initial variable for Memory + init: Variable + pre_mem: the memory variable in previous time step + pre_mem: Variable + mem: the memory variable in current time step + mem: Variable + + StaticRNNMemoryLink class is used to create a link between two + memory cells of a StaticRNN. + """ + + def __init__(self, init, pre_mem, mem=None): + self.init = init + self.pre_mem = pre_mem + self.mem = mem + + +class StaticRNN(object): + """ + StaticRNN class. + + StaticRNN class is used to create a StaticRNN. The RNN will have its + own parameters like inputs, outputs, memories, status and length. + """ + BEFORE_RNN_BLOCK = 0 + IN_RNN_BLOCK = 1 + AFTER_RNN_BLOCK = 2 + + def __init__(self, name=None): + self.helper = LayerHelper("static_rnn", name=name) + self.memories = {} # memory map, from pre_mem.name --> MemoryLink + self.inputs = [] # input variable list in current block + self.outputs = [] # output variable list in parent block + self.status = StaticRNN.BEFORE_RNN_BLOCK # status flag. + # sequence length, since it is a static RNN, sequence length are fixed. + self.seq_len = None + + def step(self): + return StaticRNNGuard(self) + + def _assert_in_rnn_block_(self, method): + if self.status != StaticRNN.IN_RNN_BLOCK: + raise ValueError("You must invoke {0} in rnn block".format(method)) + + def memory(self, + init=None, + shape=None, + batch_ref=None, + init_value=0.0, + init_batch_dim_idx=0, + ref_batch_dim_idx=1): + """ + Args: + init: boot memory, if not set, a shape, batch_ref must be provided + shape: shape of the boot memory + batch_ref: batch size reference variable + init_value: the init value of boot memory + init_batch_dim_idx: the index of batch size in init's dimension + ref_batch_dim_idx: the index of batch size in batch_ref's dimension + """ + self._assert_in_rnn_block_('memory') + if init is None: + if shape is None or batch_ref is None: + raise ValueError( + "if init is None, memory at least need shape and batch_ref") + parent_block = self.parent_block() + var_name = unique_name("@".join([self.helper.name, "memory_boot"])) + boot_var = parent_block.create_var( + name=var_name, + shape=shape, + dtype=batch_ref.dtype, + persistable=False) + + parent_block.append_op( + type="fill_constant_batch_size_like", + inputs={'Input': [batch_ref]}, + outputs={'Out': [boot_var]}, + attrs={ + 'value': init_value, + 'shape': boot_var.shape, + 'dtype': boot_var.dtype, + 'input_dim_idx': ref_batch_dim_idx, + 'output_dim_idx': init_batch_dim_idx + }) + + return self.memory(init=boot_var) + else: + pre_mem = self.helper.create_variable( + name=unique_name("@".join([self.helper.name, "mem"])), + dtype=init.dtype, + shape=init.shape) + self.memories[pre_mem.name] = StaticRNNMemoryLink( + init=init, pre_mem=pre_mem) + return pre_mem + + def step_input(self, x): + self._assert_in_rnn_block_('step_input') + if not isinstance(x, Variable): + raise TypeError("step input takes a Variable") + if self.seq_len is None: + self.seq_len = x.shape[0] + elif self.seq_len != x.shape[0]: + raise ValueError("Static RNN only take fix seq_len input") + + ipt = self.helper.create_variable( + name=x.name, dtype=x.dtype, shape=list(x.shape[1:]), type=x.type) + self.inputs.append(ipt) + return ipt + + def step_output(self, o): + self._assert_in_rnn_block_('step_output') + if not isinstance(o, Variable): + raise TypeError("step output takes a Variable") + + tmp_o = self.helper.create_tmp_variable(dtype=o.dtype) + self.helper.append_op( + type='rnn_memory_helper', + inputs={'X': [o]}, + outputs={'Out': tmp_o}, + attrs={'dtype': o.dtype}) + + out_var = self.parent_block().create_var( + name=tmp_o.name, + shape=[self.seq_len] + list(tmp_o.shape), + dtype=tmp_o.dtype) + + self.outputs.append(out_var) + + def output(self, *outputs): + for each in outputs: + self.step_output(each) + + def update_memory(self, mem, var): + if not isinstance(mem, Variable) or not isinstance(var, Variable): + raise TypeError("update memory should take variables") + self.memories[mem.name].mem = var + + def parent_block(self): + prog = self.helper.main_program + parent_idx = prog.current_block().parent_idx + assert parent_idx >= 0 + parent_block = prog.block(parent_idx) + return parent_block + + def __call__(self, *args, **kwargs): + if self.status != StaticRNN.AFTER_RNN_BLOCK: + raise ValueError("RNN output can only be retrieved after rnn block") + if len(self.outputs) == 0: + raise ValueError("RNN has no output") + elif len(self.outputs) == 1: + return self.outputs[0] + else: + return self.outputs + + def complete_rnn_op(self): + main_program = self.helper.main_program + rnn_block = main_program.current_block() + parent_block = self.parent_block() + + local_inputs = set() + + for op in rnn_block.ops: + assert isinstance(op, Operator) + for oname in op.output_names: + for out_var_name in op.output(oname): + local_inputs.add(out_var_name) + + for var in self.inputs: + local_inputs.add(var.name) + for m in self.memories: + local_inputs.add(m) + + params = list() + for op in rnn_block.ops: + assert isinstance(op, Operator) + for iname in op.input_names: + for in_var_name in op.input(iname): + if in_var_name not in local_inputs: + params.append(in_var_name) + + parameters = [parent_block.var(name) for name in params] + + step_scope = parent_block.create_var( + type=core.VarDesc.VarType.STEP_SCOPES) + + inlinks = [parent_block.var(i.name) for i in self.inputs] + outlinks = self.outputs + + boot_memories = [] + pre_memories = [] + memories = [] + for _, mem in self.memories.iteritems(): + boot_memories.append(mem.init) + pre_memories.append(mem.pre_mem.name) + mem_var = rnn_block.var(mem.mem.name) + assert isinstance(mem_var, Variable) + new_mem = self.helper.create_tmp_variable(dtype=mem_var.dtype) + + rnn_block.append_op( + type='rnn_memory_helper', + inputs={'X': [mem_var]}, + outputs={'Out': [new_mem]}, + attrs={'dtype': mem_var.dtype}) + + memories.append(new_mem.name) + + parent_block.append_op( + type='recurrent', + inputs={ + 'inputs': inlinks, + 'initial_states': boot_memories, + 'parameters': parameters + }, + outputs={'outputs': outlinks, + 'step_scopes': [step_scope]}, + attrs={ + 'ex_states': pre_memories, + 'states': memories, + 'sub_block': rnn_block + }) + + +class WhileGuard(BlockGuard): + def __init__(self, while_op): + if not isinstance(while_op, While): + raise TypeError("WhileGuard takes a while op") + super(WhileGuard, self).__init__(while_op.helper.main_program) + self.while_op = while_op + + def __enter__(self): + self.while_op.status = While.IN_WHILE_BLOCK + return super(WhileGuard, self).__enter__() + + def __exit__(self, exc_type, exc_val, exc_tb): + if exc_type is not None: + return False + self.while_op.status = While.AFTER_WHILE_BLOCK + self.while_op.complete() + return super(WhileGuard, self).__exit__(exc_type, exc_val, exc_tb) + + +class While(object): + BEFORE_WHILE_BLOCK = 0 + IN_WHILE_BLOCK = 1 + AFTER_WHILE_BLOCK = 2 + + def __init__(self, cond, name=None): + self.helper = LayerHelper("while", name=name) + self.status = While.BEFORE_WHILE_BLOCK + if not isinstance(cond, Variable): + raise TypeError("condition should be a variable") + assert isinstance(cond, Variable) + if cond.dtype != core.DataType.BOOL: + raise TypeError("condition should be a bool variable") + if reduce(lambda a, b: a * b, cond.shape, 1) != 1: + raise TypeError("condition should be a bool scalar") + self.cond_var = cond + + def block(self): + return WhileGuard(self) + + def complete(self): + main_program = self.helper.main_program + while_block = main_program.current_block() + parent_block = main_program.block(main_program.current_block() + .parent_idx) + + inner_outputs = {self.cond_var.name} + x_name_list = set() + for op in while_block.ops: + for iname in op.input_names: + for in_var_name in op.input(iname): + if in_var_name not in inner_outputs: + x_name_list.add(in_var_name) + + for oname in op.output_names: + for out_var_name in op.output(oname): + inner_outputs.add(out_var_name) + + out_vars = [] + for inner_out_name in inner_outputs: + if inner_out_name in parent_block.vars: + out_vars.append(parent_block.var(inner_out_name)) + + step_scope = parent_block.create_var( + type=core.VarDesc.VarType.STEP_SCOPES) + + parent_block.append_op( + type='while', + inputs={ + 'X': [parent_block.var(x_name) for x_name in x_name_list], + 'Condition': [self.cond_var] + }, + outputs={'Out': out_vars, + 'StepScopes': [step_scope]}, + attrs={'sub_block': while_block}) + + +def lod_rank_table(x, level=0): + """LoD Rank Table Operator. Given an input variable **x** and a level number + of LoD, this layer creates a LodRankTable object. A LoDRankTable object + contains a list of bi-element tuples. Each tuple consists of an index and + a length, both of which are int type. Reffering to specified level of LoD, + the index is the sequence index number and the length representes the + sequence length. Please note that the list is ranked in descending order by + the length. The following is an example: + + .. code-block:: text + + x is a LoDTensor: + x.lod = [[0, 2, 3], + [0, 5, 6, 7]] + x.data = [a, b, c, d, e, f, g] + + 1. set level to 0: + Create lod rank table: + lod_rank_table_obj = lod_rank_table(x, level=0) + + Get: + lod_rank_table_obj.items() = [(0, 2), (1, 1)] + + 2. set level to 1: + Create lod rank table: + lod_rank_table_obj = lod_rank_table(x, level=1) + + Get: + lod_rank_table_obj.items() = [(0, 5), (1, 1), (2, 1)] + + Args: + x (Variable): Input variable, a LoDTensor based which to create the lod + rank table. + level (int): Specify the LoD level, on which to create the lod rank + table. + + Returns: + Variable: The created LoDRankTable object. + + Examples: + .. code-block:: python + + x = fluid.layers.data(name='x', shape=[10], + dtype='float32', lod_level=1) + out = layers.lod_rank_table(x=x, level=0) + """ + helper = LayerHelper("lod_rank_table", **locals()) + table = helper.create_variable( + type=core.VarDesc.VarType.LOD_RANK_TABLE, + name=unique_name("lod_rank_table")) + helper.append_op( + type='lod_rank_table', + inputs={'X': x}, + outputs={'Out': table}, + attrs={'level': level}) + return table + + +def max_sequence_len(rank_table): + """ + This function creates an operator to calculate the length of + max seqence through input rank_table(should be a lod_rank_table) + """ + helper = LayerHelper("max_seqence_len", **locals()) + res = helper.create_tmp_variable(dtype="int64") + helper.append_op( + type="max_sequence_len", + inputs={"RankTable": rank_table}, + outputs={"Out": res}) + return res + + +def topk(input, k): + helper = LayerHelper('topk', **locals()) + topk_out = helper.create_tmp_variable(dtype=input.data_type) + topk_indices = helper.create_tmp_variable(dtype='int64') + helper.append_op( + type='top_k', + inputs={'X': [input]}, + outputs={'Out': [topk_out], + 'Indices': [topk_indices]}, + attrs={'k': k}) + return topk_out, topk_indices + + +def lod_tensor_to_array(x, table): + """This function performs the operation that converts an LOD_Tensor to + an array. + + Args: + x (Variable|list): The tensor that needs to be converted to an array. + table (ParamAttr|list): The variable that stores the level of lod + which is ordered by sequence length in + descending order. + + Returns: + Variable: The variable of type array that has been converted from a + tensor. + + Examples: + .. code-block:: python + + x = fluid.layers.data(name='x', shape=[10]) + table = fluid.layers.lod_rank_table(x, level=0) + array = fluid.layers.lod_tensor_to_array(x, table) + """ + helper = LayerHelper("lod_tensor_to_array", **locals()) + array = helper.create_variable( + name=unique_name("lod_tensor_to_array"), + type=core.VarDesc.VarType.LOD_TENSOR_ARRAY, + dtype=x.dtype) + helper.append_op( + type='lod_tensor_to_array', + inputs={'X': x, + 'RankTable': table}, + outputs={'Out': array}) + return array + + +def array_to_lod_tensor(x, table): + """This function performs the operations that converts an array to + an LOD_Tensor. + + Args: + x (Variable|list): The array that needs to be converted to a tensor. + table (ParamAttr|list): The variable that stores the level of lod + which is ordered by sequence length in + descending order. + + Returns: + Variable: The variable of type tensor that has been converted + from an array. + + Examples: + .. code-block:: python + + x = fluid.layers.data(name='x', shape=[10]) + table = fluid.layers.lod_rank_table(x, level=0) + array = fluid.layers.lod_tensor_to_array(x, table) + lod_tensor = fluid.layers.array_to_lod_tensor(array, table) + """ + helper = LayerHelper("array_to_lod_tensor", **locals()) + tmp = helper.create_tmp_variable(dtype=x.dtype) + helper.append_op( + type="array_to_lod_tensor", + inputs={'X': x, + 'RankTable': table}, + outputs={'Out': tmp}) + return tmp + + +def increment(x, value=1.0, in_place=True): + """This function performs an operation that increments each value in the + input :math:`x` by an amount: :math:`value` as mentioned in the input + parameter. This operation is performed in-place by default. + + Args: + x (Variable|list): The tensor that has the input values. + value (float): The amount by which the values should be incremented. + in_place (bool): If the increment should be performed in-place. + + Returns: + Variable: The tensor variable storing the transformation of + element-wise increment of each value in the input. + + Examples: + .. code-block:: python + + data = fluid.layers.data(name='data', shape=[32, 32], dtype='float32') + data = fluid.layers.increment(x=data, value=3.0, in_place=True) + """ + helper = LayerHelper("increment", **locals()) + if not in_place: + out = helper.create_tmp_variable(dtype=x.dtype) + else: + out = x + helper.append_op( + type='increment', + inputs={'X': [x]}, + outputs={'Out': [out]}, + attrs={'step': float(value)}) + return out + + +def array_write(x, i, array=None): + """This function performs the operation to write the data out as an + LOD_TENSOR_ARRAY. + + Args: + x (Variable|list): The input tensor from which the data will be read. + i (Variable|list): The subscript index in tensor array, that points the + place from which data will be read. + array (Variable|list): The data can be read into this variable if + this is assigned. + Returns: + Variable: The tensor type variable that has the data written to it. + + Examples: + .. code-block::python + + tmp = fluid.layers.zeros(shape=[10], dtype='int32') + i = fluid.layers.fill_constant(shape=[1], dtype='int64', value=10) + arr = layers.array_write(tmp, i=i) + """ + helper = LayerHelper('array_write', **locals()) + if array is None: + array = helper.create_variable( + name="{0}.out".format(helper.name), + type=core.VarDesc.VarType.LOD_TENSOR_ARRAY, + dtype=x.dtype) + helper.append_op( + type='write_to_array', + inputs={'X': [x], + 'I': [i]}, + outputs={'Out': [array]}) + return array + + +def create_array(dtype): + """This function creates an array of type :math:`LOD_TENSOR_ARRAY` using the + LayerHelper. + + Args: + dtype (int|float): The data type of the elements in the array. + + Returns: + Variable: The tensor variable storing the elements of data type. + + Examples: + .. code-block:: python + + data = fluid.layers.create_array(dtype='float32') + + """ + helper = LayerHelper("array", **locals()) + return helper.create_variable( + name="{0}.out".format(helper.name), + type=core.VarDesc.VarType.LOD_TENSOR_ARRAY, + dtype=dtype) + + +def less_than(x, y, cond=None, **ignored): + """ + **Less than** + + This layer returns the truth value of :math:`x < y` elementwise. + + Args: + x(Variable): First operand of *less_than* + y(Variable): Second operand of *less_than* + cond(Variable|None): Optional output variable to store the result of *less_than* + + Returns: + Variable: The tensor variable storing the output of *less_than*. + + Examples: + .. code-block:: python + + less = fluid.layers.less_than(x=label, y=limit) + """ + helper = LayerHelper("less_than", **locals()) + if cond is None: + cond = helper.create_tmp_variable(dtype='bool') + cond.stop_gradient = True + + helper.append_op( + type='less_than', inputs={'X': [x], + 'Y': [y]}, outputs={'Out': [cond]}) + return cond + + +def array_read(array, i): + """This function performs the operation to read the data in as an + LOD_TENSOR_ARRAY. + Args: + array (Variable|list): The input tensor that will be written to an array. + i (Variable|list): The subscript index in tensor array, that points the + place where data will be written to. + Returns: + Variable: The tensor type variable that has the data written to it. + Examples: + .. code-block::python + tmp = fluid.layers.zeros(shape=[10], dtype='int32') + i = fluid.layers.fill_constant(shape=[1], dtype='int64', value=10) + arr = layers.array_read(tmp, i=i) + """ + helper = LayerHelper('array_read', **locals()) + if not isinstance( + array, + Variable) or array.type != core.VarDesc.VarType.LOD_TENSOR_ARRAY: + raise TypeError("array should be tensor array vairable") + out = helper.create_tmp_variable(dtype=array.dtype) + helper.append_op( + type='read_from_array', + inputs={'X': [array], + 'I': [i]}, + outputs={'Out': [out]}) + return out + + +def shrink_memory(x, i, table): + """ + This function creates an operator to shrink_rnn_memory using the RankTable + as mentioned in the input parameter. + """ + helper = LayerHelper('shrink_memory', **locals()) + out = helper.create_tmp_variable(dtype=x.dtype) + helper.append_op( + type='shrink_rnn_memory', + inputs={'X': [x], + 'I': [i], + 'RankTable': [table]}, + outputs={'Out': [out]}, + attrs={}) + return out + + +def array_length(array): + """This function performs the operation to find the length of the input + LOD_TENSOR_ARRAY. + + Args: + array (LOD_TENSOR_ARRAY): The input array that will be used + to compute the length. + + Returns: + Variable: The length of the input LoDTensorArray. + + Examples: + .. code-block::python + + tmp = fluid.layers.zeros(shape=[10], dtype='int32') + i = fluid.layers.fill_constant(shape=[1], dtype='int64', value=10) + arr = fluid.layers.array_write(tmp, i=i) + arr_len = fluid.layers.array_length(arr) + """ + helper = LayerHelper('array_length', **locals()) + tmp = helper.create_tmp_variable(dtype='int64') + tmp.stop_gradient = True + helper.append_op( + type='lod_array_length', inputs={'X': [array]}, outputs={'Out': [tmp]}) + return tmp + + +class ConditionalBlockGuard(BlockGuard): + def __init__(self, block): + if not isinstance(block, ConditionalBlock): + raise TypeError("block should be conditional block") + super(ConditionalBlockGuard, self).__init__(block.helper.main_program) + self.block = block + + def __enter__(self): + return super(ConditionalBlockGuard, self).__enter__() + + def __exit__(self, exc_type, exc_val, exc_tb): + self.block.complete() + return super(ConditionalBlockGuard, self).__exit__(exc_type, exc_val, + exc_tb) + + +class ConditionalBlock(object): + def __init__(self, inputs, name=None): + for each_input in inputs: + if not isinstance(each_input, Variable): + raise TypeError("Each input should be variable") + self.inputs = inputs + self.helper = LayerHelper('conditional_block', name=name) + + def block(self): + return ConditionalBlockGuard(self) + + def complete(self): + inside_block = self.helper.main_program.current_block() + parent_block = self.helper.main_program.block(inside_block.parent_idx) + + intermediate = set() + params = set() + + for each_op in inside_block.ops: + assert isinstance(each_op, Operator) + for iname in each_op.input_names: + for in_var_name in each_op.input(iname): + if in_var_name not in intermediate: + params.add(in_var_name) + + for oname in each_op.output_names: + for out_var_name in each_op.output(oname): + intermediate.add(out_var_name) + input_set = set([ipt.name for ipt in self.inputs]) + + param_list = [ + parent_block.var(each_name) for each_name in params + if each_name not in input_set + ] + + out_list = [ + parent_block.var(var_name) for var_name in parent_block.vars + if var_name not in intermediate + ] + + step_scope = parent_block.create_var( + type=core.VarDesc.VarType.STEP_SCOPES) + parent_block.append_op( + type='conditional_block', + inputs={ + 'X': self.inputs, + 'Params': param_list, + }, + outputs={'Out': out_list, + 'Scope': [step_scope]}, + attrs={'sub_block': inside_block}) + + +class IfElseBlockGuard(object): + def __init__(self, is_true, ifelse): + if not isinstance(ifelse, IfElse): + raise TypeError("ifelse must be an instance of IfElse class") + + if ifelse.status != IfElse.OUT_IF_ELSE_BLOCKS: + raise ValueError("You cannot invoke IfElse.block() inside a block") + + self.is_true = is_true + self.ie = ifelse + if is_true: + self.cond_block = ifelse.conditional_true_block + else: + self.cond_block = ifelse.conditional_false_block + + if not isinstance(self.cond_block, ConditionalBlock): + raise TypeError("Unexpected situation") + + self.cond_block = self.cond_block.block() + + def __enter__(self): + self.ie.status = IfElse.IN_IF_ELSE_TRUE_BLOCKS if self.is_true else IfElse.IN_IF_ELSE_FALSE_BLOCKS + self.cond_block.__enter__() + + def __exit__(self, exc_type, exc_val, exc_tb): + if not self.cond_block.__exit__(exc_type, exc_val, exc_tb): + # re-raise inside exception + return False + if len(self.ie.output_table[1 if self.is_true else 0]) == 0: + raise ValueError("Must set output inside block") + self.ie.status = IfElse.OUT_IF_ELSE_BLOCKS + + +class IfElse(object): + OUT_IF_ELSE_BLOCKS = 0 + IN_IF_ELSE_TRUE_BLOCKS = 1 + IN_IF_ELSE_FALSE_BLOCKS = 2 + + def __init__(self, cond, name=None): + if not isinstance(cond, Variable): + raise TypeError("cond must be a Variable") + self.helper = LayerHelper('ifelse', name=name) + self.cond = cond + self.input_table = {} + self.status = IfElse.OUT_IF_ELSE_BLOCKS + self.conditional_true_block = ConditionalBlock(inputs=[self.cond]) + self.conditional_false_block = ConditionalBlock(inputs=[self.cond]) + self.output_table = ([], []) # (true_outs, false_outs) + + def input(self, x): + if self.status == IfElse.OUT_IF_ELSE_BLOCKS: + raise ValueError("input must in true/false blocks") + if id(x) not in self.input_table: + parent_block = self.parent_block() + out_true = parent_block.create_var( + name=unique_name('ifelse_input' + self.helper.name), + dtype=x.dtype) + + out_false = parent_block.create_var( + name=unique_name('ifelse_input' + self.helper.name), + dtype=x.dtype) + parent_block.append_op( + type='split_lod_tensor', + inputs={ + 'X': x, + 'Mask': self.cond, + }, + outputs={'OutTrue': out_true, + 'OutFalse': out_false}, + attrs={'level': 0}) + self.input_table[id(x)] = (out_true, out_false) + else: + out_true, out_false = self.input_table[id(x)] + + if self.status == IfElse.IN_IF_ELSE_TRUE_BLOCKS: + return out_true + else: + return out_false + + def parent_block(self): + current_block = self.helper.main_program.current_block() + return self.helper.main_program.block(current_block.parent_idx) + + def true_block(self): + return IfElseBlockGuard(True, self) + + def false_block(self): + return IfElseBlockGuard(False, self) + + def output(self, *outs): + if self.status == self.OUT_IF_ELSE_BLOCKS: + raise ValueError("output can only be invoked in the sub-block") + + out_table = self.output_table[1 if self.status == + self.IN_IF_ELSE_TRUE_BLOCKS else 0] + parent_block = self.parent_block() + for each_out in outs: + if not isinstance(each_out, Variable): + raise TypeError("Each output should be a variable") + # create outside tensor + outside_out = parent_block.create_var( + name=unique_name("_".join([self.helper.name, 'output'])), + dtype=each_out.dtype) + out_table.append(outside_out) + + # assign local var to outside + assign(input=each_out, output=outside_out) + + def __call__(self): + if self.status != self.OUT_IF_ELSE_BLOCKS: + raise ValueError("IfElse::__call__ must be out of sub-block") + false_len, true_len = map(len, self.output_table) + if false_len == 0 and true_len == 0: + raise ValueError("Must invoke true_block/false_block before " + "__call__") + elif false_len != true_len and false_len != 0 and true_len != 0: + raise ValueError("The output side must be same") + elif false_len == 0 or true_len == 0: + return self.output_table[0 if false_len != 0 else 1] + + # else none of false_len/true_len is zero + # merge together + rlist = [] + for false_var, true_var in zip(*self.output_table): + rlist.append( + merge_lod_tensor( + in_true=true_var, + in_false=false_var, + mask=self.cond, + x=self.cond, + level=0)) + return rlist + + +class DynamicRNN(object): + BEFORE_RNN = 0 + IN_RNN = 1 + AFTER_RNN = 2 + + def __init__(self, name=None): + self.helper = LayerHelper('dynamic_rnn', name=name) + self.status = DynamicRNN.BEFORE_RNN + self.lod_rank_table = None + self.max_seq_len = None + self.step_idx = None + self.zero_idx = fill_constant(shape=[1], value=0, dtype='int64') + self.mem_dict = dict() + self.output_array = [] + self.outputs = [] + self.cond = self.helper.create_tmp_variable(dtype='bool') + self.cond.stop_gradient = False + self.while_op = While(self.cond) + self.input_array = [] + self.mem_link = [] + + def step_input(self, x): + self._assert_in_rnn_block_("step_input") + if not isinstance(x, Variable): + raise TypeError( + "step_input() can only take a Variable as its input") + parent_block = self._parent_block_() + if self.lod_rank_table is None: + self.lod_rank_table = parent_block.create_var( + name=unique_name('lod_rank_table'), + type=core.VarDesc.VarType.LOD_RANK_TABLE) + self.lod_rank_table.stop_gradient = True + parent_block.append_op( + type='lod_rank_table', + inputs={"X": x}, + outputs={"Out": self.lod_rank_table}) + self.max_seq_len = parent_block.create_var( + name=unique_name('dynamic_rnn_max_seq_len'), dtype='int64') + self.max_seq_len.stop_gradient = False + parent_block.append_op( + type='max_sequence_len', + inputs={'RankTable': self.lod_rank_table}, + outputs={"Out": self.max_seq_len}) + self.cond.stop_gradient = True + parent_block.append_op( + type='less_than', + inputs={'X': self.step_idx, + 'Y': self.max_seq_len}, + outputs={'Out': self.cond}) + + input_array = parent_block.create_var( + name=unique_name('dynamic_rnn_input_array'), + type=core.VarDesc.VarType.LOD_TENSOR_ARRAY, + dtype=x.dtype) + self.input_array.append((input_array, x.dtype)) + parent_block.append_op( + type='lod_tensor_to_array', + inputs={'X': x, + 'RankTable': self.lod_rank_table}, + outputs={'Out': input_array}) + return array_read(array=input_array, i=self.step_idx) + + @contextlib.contextmanager + def block(self): + if self.status != DynamicRNN.BEFORE_RNN: + raise ValueError("rnn.block() can only be invoke once") + self.step_idx = fill_constant(shape=[1], dtype='int64', value=0) + self.step_idx.stop_gradient = False + self.status = DynamicRNN.IN_RNN + with self.while_op.block(): + yield + increment(x=self.step_idx, value=1.0, in_place=True) + + for new_mem, mem_array in self.mem_link: + array_write(x=new_mem, i=self.step_idx, array=mem_array) + + less_than(x=self.step_idx, y=self.max_seq_len, cond=self.cond) + + self.status = DynamicRNN.AFTER_RNN + for each_array in self.output_array: + self.outputs.append( + array_to_lod_tensor( + x=each_array, table=self.lod_rank_table)) + + def __call__(self, *args, **kwargs): + if self.status != DynamicRNN.AFTER_RNN: + raise ValueError( + "Dynamic RNN outputs can only be retrieved after rnn block") + if len(self.outputs) == 1: + return self.outputs[0] + else: + return self.outputs + + def memory(self, init=None, shape=None, value=0.0, dtype='float32'): + self._assert_in_rnn_block_('memory') + if init is not None: + if not isinstance(init, Variable): + raise TypeError( + "The input arg `init` of memory() must be a Variable") + parent_block = self._parent_block_() + mem_array = parent_block.create_var( + name=unique_name('dynamic_rnn_mem_array'), + type=core.VarDesc.VarType.LOD_TENSOR_ARRAY, + dtype=init.dtype) + parent_block.append_op( + type='write_to_array', + inputs={'X': init, + 'I': self.zero_idx}, + outputs={'Out': mem_array}) + retv = array_read(array=mem_array, i=self.step_idx) + retv = shrink_memory( + x=retv, i=self.step_idx, table=self.lod_rank_table) + self.mem_dict[retv.name] = mem_array + return retv + else: + if len(self.input_array) == 0: + raise ValueError( + "step_input should be invoked before memory(shape=..., value=...)" + ) + parent_block = self._parent_block_() + init = parent_block.create_var( + name=unique_name('mem_init'), dtype=dtype) + arr, dtype = self.input_array[0] + in0 = parent_block.create_var(name=unique_name('in0'), dtype=dtype) + parent_block.append_op( + type='read_from_array', + inputs={'X': [arr], + 'I': [self.zero_idx]}, + outputs={'Out': [in0]}) + parent_block.append_op( + type='fill_constant_batch_size_like', + inputs={'Input': [in0]}, + outputs={'Out': [init]}, + attrs={ + 'shape': [-1] + shape, + 'value': float(value), + 'dtype': init.dtype + }) + return self.memory(init=init) + + def update_memory(self, ex_mem, new_mem): + self._assert_in_rnn_block_('update_memory') + if not isinstance(ex_mem, Variable): + raise TypeError("The input arg `ex_mem` of update_memory() must " + "be a Variable") + if not isinstance(new_mem, Variable): + raise TypeError("The input arg `new_mem` of update_memory() must " + "be a Variable") + + mem_array = self.mem_dict.get(ex_mem.name, None) + if mem_array is None: + raise ValueError("Please invoke memory before update_memory") + if self.lod_rank_table is None: + raise ValueError("Please invoke step_input before update_memory") + + self.mem_link.append((new_mem, mem_array)) + + def output(self, *outputs): + self._assert_in_rnn_block_('output') + parent_block = self._parent_block_() + for each in outputs: + outside_array = parent_block.create_var( + name=unique_name("_".join( + [self.helper.name, "output_array", each.name])), + type=core.VarDesc.VarType.LOD_TENSOR_ARRAY, + dtype=each.dtype) + array_write(x=each, i=self.step_idx, array=outside_array) + self.output_array.append(outside_array) + + def _parent_block_(self): + prog = self.helper.main_program + parent_idx = prog.current_block().parent_idx + assert parent_idx >= 0 + parent_block = prog.block(parent_idx) + + return parent_block + + def _assert_in_rnn_block_(self, method): + if self.status != DynamicRNN.IN_RNN: + raise ValueError("{0} can only be invoked inside rnn block.".format( + method)) + + +@autodoc +def reorder_lod_tensor_by_rank(x, rank_table): + helper = LayerHelper('reorder_lod_tensor_by_rank', **locals()) + helper.is_instance('x', Variable) + helper.is_instance('rank_table', Variable) + + out = helper.create_tmp_variable(dtype=x.dtype) + helper.append_op( + type='reorder_lod_tensor_by_rank', + inputs={'X': [x], + 'RankTable': [rank_table]}, + outputs={'Out': [out]}) + return out diff --git a/python/paddle/v2/fluid/layers/io.py b/python/paddle/v2/fluid/layers/io.py new file mode 100644 index 0000000000000000000000000000000000000000..56c3f7b7b7f174338bb56bc5785423ca634650a6 --- /dev/null +++ b/python/paddle/v2/fluid/layers/io.py @@ -0,0 +1,62 @@ +from .. import core +from ..layer_helper import LayerHelper + +__all__ = ['data'] + + +def data(name, + shape, + append_batch_size=True, + dtype='float32', + lod_level=0, + type=core.VarDesc.VarType.LOD_TENSOR, + stop_gradient=True): + """ + **Data Layer** + + This function takes in the input and based on whether data has + to be returned back as a minibatch, it creates the global variable using + the helper functions. The global variables can be accessed by all the + following operations and layers in the graph. + + All the input variables of this function are passed in as local variables + to the LayerHelper constructor. + + Args: + name(str): The name/alias of the function + shape(list): Tuple declaring the shape. + append_batch_size(bool): Whether or not to append the data as a batch. + 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: + Variable: The global variable that gives access to the data. + + Examples: + .. code-block:: python + + data = fluid.layers.data(name='x', shape=[784], dtype='float32') + """ + helper = LayerHelper('data', **locals()) + shape = list(shape) + for i in xrange(len(shape)): + if shape[i] is None: + shape[i] = -1 + append_batch_size = False + elif shape[i] < 0: + append_batch_size = False + + if append_batch_size: + shape = [-1] + shape # append batch size as -1 + + return helper.create_global_variable( + name=name, + shape=shape, + dtype=dtype, + type=type, + stop_gradient=stop_gradient, + lod_level=lod_level) diff --git a/python/paddle/v2/fluid/layers/nn.py b/python/paddle/v2/fluid/layers/nn.py new file mode 100644 index 0000000000000000000000000000000000000000..55b35ad543b2350915574d7f9b4ef460156ee459 --- /dev/null +++ b/python/paddle/v2/fluid/layers/nn.py @@ -0,0 +1,1414 @@ +""" +All layers just related to the neural network. +""" + +from ..layer_helper import LayerHelper +from ..initializer import Normal, Constant +from ..framework import Variable +from ..param_attr import ParamAttr +from tensor import concat + +__all__ = [ + 'fc', 'embedding', 'dynamic_lstm', 'gru_unit', 'linear_chain_crf', + 'crf_decoding', 'cos_sim', 'cross_entropy', 'square_error_cost', 'accuracy', + 'chunk_eval', 'sequence_conv', 'conv2d', 'sequence_pool', 'pool2d', + 'batch_norm', 'beam_search_decode', 'conv2d_transpose', 'sequence_expand', + 'lstm_unit', 'reduce_sum', 'reduce_mean', 'reduce_max', 'reduce_min', + 'sequence_first_step', 'sequence_last_step' +] + + +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 (one for each input tensor) 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 biases 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: + + .. math:: + + Out = Act({\sum_{i=0}^{N-1}W_iX_i + b}) + + In the above equation: + + * :math:`N`: Number of the input. + * :math:`X_i`: The input tensor. + * :math:`W`: The weights created by this layer. + * :math:`b`: The bias parameter created by this layer (if needed). + * :math:`Act`: The activation funtion. + * :math:`Out`: The output tensor. + + Args: + input(Variable|list): The input tensor(s) to the fully connected layer. + size(int): The number of output units in the fully connected layer. + num_flatten_dims(int): 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` + dimensions will be flatten to form the first + dimension of the final matrix (height of the + matrix), and the rest `rank(X) - num_col_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 + `x_num_col_dims` = 3. Then, the flattened matrix + will have a shape [2 x 3 x 4, 5 x 6] = [24, 30]. + By default, `x_num_col_dims` is set to 1. + param_attr(ParamAttr|list): The parameter attribute for learnable + parameters/weights of the fully connected + layer. + param_initializer(ParamAttr|list): The initializer used for the + weight/parameter. If set None, + XavierInitializer() will be used. + bias_attr(ParamAttr|list): The parameter attribute for the bias parameter + for this layer. If set None, no bias will be + added to the output units. + bias_initializer(ParamAttr|list): The initializer used for the bias. + If set None, then ConstantInitializer() + will be used. + act(str): Activation to be applied to the output of the fully connected + layer. + name(str): Name/alias of the fully connected layer. + + + Returns: + Variable: The output tensor variable. + + 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") + """ + + helper = LayerHelper("fc", **locals()) + + dtype = helper.input_dtype() + + mul_results = [] + for input_var, param_attr in helper.iter_inputs_and_params(): + input_shape = input_var.shape + param_shape = [ + reduce(lambda a, b: a * b, input_shape[num_flatten_dims:], 1) + ] + [size] + w = helper.create_parameter( + attr=param_attr, shape=param_shape, dtype=dtype, is_bias=False) + tmp = helper.create_tmp_variable(dtype) + helper.append_op( + type="mul", + inputs={ + "X": input_var, + "Y": w, + }, + outputs={"Out": tmp}, + attrs={"x_num_col_dims": num_flatten_dims, + "y_num_col_dims": 1}) + mul_results.append(tmp) + + # sum + if len(mul_results) == 1: + pre_bias = mul_results[0] + else: + pre_bias = helper.create_tmp_variable(dtype) + helper.append_op( + type="sum", inputs={"X": mul_results}, outputs={"Out": pre_bias}) + # add bias + pre_activation = helper.append_bias_op(pre_bias) + # add activation + return helper.append_activation(pre_activation) + + +def embedding(input, size, is_sparse=False, param_attr=None, dtype='float32'): + """ + **Embedding Layer** + + This layer is used to lookup a vector of IDs, provided by *input*, in a lookup table. + The result of this lookup is the embedding of each ID in the *input*. + + All the input variables are passed in as local variables to the LayerHelper + constructor. + + Args: + input(Variable): Input to the function + size(tuple|list|None): Shape of the look up table parameter + is_sparse(bool): Boolean flag that specifying whether the input is sparse + param_attr(ParamAttr): Parameters for this layer + dtype(np.dtype|core.DataType|str): The type of data : float32, float_16, int etc + + Returns: + Variable: The tensor variable storing the embeddings of the \ + supplied inputs. + + Examples: + .. code-block:: python + + dict_size = len(dataset.ids) + data = fluid.layers.data(name='ids', shape=[32, 32], dtype='float32') + fc = fluid.layers.embedding(input=data, size=[dict_size, 16]) + """ + + helper = LayerHelper('embedding', **locals()) + w = helper.create_parameter( + attr=helper.param_attr, shape=size, dtype=dtype, is_bias=False) + tmp = helper.create_tmp_variable(dtype) + helper.append_op( + type='lookup_table', + inputs={'Ids': input, + 'W': w}, + outputs={'Out': tmp}, + attrs={'is_sparse': is_sparse}) + return tmp + + +# TODO(qijun): expose H0 and C0 +def dynamic_lstm(input, + size, + param_attr=None, + bias_attr=None, + use_peepholes=True, + is_reverse=False, + gate_activation='sigmoid', + cell_activation='tanh', + candidate_activation='tanh', + dtype='float32'): + helper = LayerHelper('lstm', **locals()) + size = size / 4 + weight = helper.create_parameter( + attr=helper.param_attr, shape=[size, 4 * size], dtype=dtype) + bias_size = [1, 7 * size] + if not use_peepholes: + bias_size[1] = 4 * size + bias = helper.create_parameter( + attr=helper.bias_attr, shape=bias_size, dtype=dtype, is_bias=True) + + hidden = helper.create_tmp_variable(dtype) + cell = helper.create_tmp_variable(dtype) + batch_gate = helper.create_tmp_variable(dtype) + batch_cell_pre_act = helper.create_tmp_variable(dtype) + + helper.append_op( + type='lstm', + inputs={'Input': input, + 'Weight': weight, + 'Bias': bias}, + outputs={ + 'Hidden': hidden, + 'Cell': cell, + 'BatchGate': batch_gate, + 'BatchCellPreAct': batch_cell_pre_act + }, + attrs={ + 'use_peepholes': use_peepholes, + 'is_reverse': is_reverse, + 'gate_activation': gate_activation, + 'cell_activation': cell_activation, + 'candidate_activation': candidate_activation + }) + return hidden, cell + + +def gru_unit(input, + hidden, + size, + weight=None, + bias=None, + activation='tanh', + gate_activation='sigmoid'): + """ + GRUUnit Operator implements partial calculations of the GRU unit as following: + + $$ + update \ gate: u_t = actGate(xu_t + W_u * h_{t-1} + b_u) \\ + reset \ gate: r_t = actGate(xr_t + W_r * h_{t-1} + b_r) \\ + output \ candidate: {h}_t = actNode(xc_t + W_c * dot(r_t, h_{t-1}) + b_c) \\ + output: h_t = dot((1 - u_t), h_{t-1}) + dot(u_t, {h}_t) + $$ + + which is same as one time step of GRU Operator. + + @note To implement the complete GRU unit, fully-connected operator must be + used before to feed xu, xr and xc as the Input of GRUUnit operator. + + TODO(ChunweiYan) add more document here + """ + activation_dict = dict( + identity=0, + sigmoid=1, + tanh=2, + relu=3, ) + activation = activation_dict[activation] + gate_activation = activation_dict[gate_activation] + + helper = LayerHelper('gru_unit', **locals()) + dtype = helper.input_dtype() + size = size / 3 + + # create weight + if weight is None: + weight = helper.create_parameter( + attr=helper.param_attr, shape=[size, 3 * size], dtype=dtype) + + # create bias + + if bias is None: + 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) + + helper.append_op( + type='gru_unit', + inputs={'Input': input, + 'HiddenPrev': hidden, + 'Weight': weight}, + outputs={ + 'Gate': gate, + 'ResetHiddenPrev': reset_hidden_pre, + 'Hidden': updated_hidden, + }, + attrs={ + 'activation': 0, + 'gate_activation': 1, + }) + + return updated_hidden, reset_hidden_pre, gate + + +def linear_chain_crf(input, label, param_attr=None): + helper = LayerHelper('linear_chain_crf', **locals()) + size = input.shape[1] + transition = helper.create_parameter( + attr=helper.param_attr, + shape=[size + 2, size], + dtype=helper.input_dtype()) + alpha = helper.create_tmp_variable(dtype=helper.input_dtype()) + emission_exps = helper.create_tmp_variable(dtype=helper.input_dtype()) + transition_exps = helper.create_tmp_variable(dtype=helper.input_dtype()) + log_likelihood = helper.create_tmp_variable(dtype=helper.input_dtype()) + helper.append_op( + type='linear_chain_crf', + inputs={"Emission": [input], + "Transition": transition, + "Label": label}, + outputs={ + "Alpha": [alpha], + "EmissionExps": [emission_exps], + "TransitionExps": transition_exps, + "LogLikelihood": log_likelihood + }) + + return log_likelihood + + +def crf_decoding(input, param_attr, label=None): + helper = LayerHelper('crf_decoding', **locals()) + transition = helper.get_parameter(param_attr.name) + viterbi_path = helper.create_tmp_variable(dtype=helper.input_dtype()) + helper.append_op( + type='crf_decoding', + inputs={"Emission": [input], + "Transition": transition, + "Label": label}, + outputs={"ViterbiPath": [viterbi_path]}) + + return viterbi_path + + +def cos_sim(X, Y, **kwargs): + """ + This function performs the cosine similarity between two tensors + X and Y and returns that as the output. + """ + helper = LayerHelper('cos_sim', **kwargs) + out = helper.create_tmp_variable(dtype=X.dtype) + xnorm = helper.create_tmp_variable(dtype=X.dtype) + ynorm = helper.create_tmp_variable(dtype=X.dtype) + helper.append_op( + type='cos_sim', + inputs={'X': [X], + 'Y': [Y]}, + outputs={'Out': [out], + 'XNorm': [xnorm], + 'YNorm': [ynorm]}) + return out + + +def cross_entropy(input, label, **kwargs): + """ + **Cross Entropy Layer** + + This layer computes the cross entropy between `input` and `label`. It supports + both standard cross-entropy and soft-label cross-entropy loss computation. + + 1) One-hot cross-entropy: + `soft_label = False`, `Label[i, 0]` indicates the class index for sample i: + + .. math:: + + Y[i] = -\log(X[i, Label[i]]) + + 2) Soft-label cross-entropy: + `soft_label = True`, `Label[i, j]` indicates the soft label of class j + for sample i: + + .. math:: + + Y[i] = \sum_j{-Label[i, j] * log(X[i, j])} + + Please make sure that in this case the summation of each row of `label` + equals one. + + 3) One-hot cross-entropy with vecterized `label`: + As a special case of 2), when each row of 'label' has only one + non-zero element which is equal to 1, soft-label cross-entropy degenerates + to a one-hot cross-entropy with one-hot label representation. + + Args: + input (Variable|list): a 2-D tensor with shape [N x D], where N is the + batch size and D is the number of classes. This input is a probability + computed by the previous operator, which is almost always the result + of a softmax operator. + label (Variable|list): the ground truth which is a 2-D tensor. When + `soft_label` is set to `False`, `label` is a tensor with shape + [N x 1]. When `soft_label` is set to `True`, `label` is a + tensor with shape [N x D]. + soft_label (bool, via `**kwargs`): a flag indicating whether to interpretate + the given labels as soft labels, default `False`. + + Returns: + A 2-D tensor with shape [N x 1], the cross entropy loss. + + Raises: + `ValueError`: 1) the 1st dimension of `input` and `label` are not equal; 2) when \ + `soft_label == True`, and the 2nd dimension of `input` and `label` are not \ + equal; 3) when `soft_label == False`, and the 2nd dimension of `label` is not 1. + + Examples: + .. code-block:: python + + predict = fluid.layers.fc(input=net, size=classdim, act='softmax') + cost = fluid.layers.cross_entropy(input=predict, label=label) + """ + helper = LayerHelper('cross_entropy', **kwargs) + out = helper.create_tmp_variable(dtype=input.dtype) + helper.append_op( + type='cross_entropy', + inputs={'X': [input], + 'Label': [label]}, + outputs={'Y': [out]}, + attrs=kwargs) + return out + + +def square_error_cost(input, label, **kwargs): + """ + This functions returns the squared error cost using the input and label. + The output is appending the op to do the above. + """ + helper = LayerHelper('square_error_cost', **kwargs) + minus_out = helper.create_tmp_variable(dtype=input.dtype) + helper.append_op( + type='elementwise_sub', + inputs={'X': [input], + 'Y': [label]}, + outputs={'Out': [minus_out]}) + + square_out = helper.create_tmp_variable(dtype=input.dtype) + helper.append_op( + type='square', inputs={'X': [minus_out]}, + outputs={'Out': [square_out]}) + return square_out + + +def accuracy(input, label, k=1, correct=None, total=None, **kwargs): + """ + This function computes the accuracy using the input and label. + The output is the top_k inputs and their indices. + """ + helper = LayerHelper("accuracy", **kwargs) + topk_out = helper.create_tmp_variable(dtype=input.dtype) + topk_indices = helper.create_tmp_variable(dtype="int64") + helper.append_op( + type="top_k", + inputs={"X": [input]}, + outputs={"Out": [topk_out], + "Indices": [topk_indices]}, + attrs={"k": k}) + acc_out = helper.create_tmp_variable(dtype="float32") + if correct is None: + correct = helper.create_tmp_variable(dtype="int64") + if total is None: + total = helper.create_tmp_variable(dtype="int64") + helper.append_op( + type="accuracy", + inputs={ + "Out": [topk_out], + "Indices": [topk_indices], + "Label": [label] + }, + outputs={ + "Accuracy": [acc_out], + "Correct": [correct], + "Total": [total], + }) + return acc_out + + +def chunk_eval(input, + label, + chunk_scheme, + num_chunk_types, + excluded_chunk_types=None, + **kwargs): + """ + This function computes and outputs the precision, recall and + F1-score of chunk detection. + """ + helper = LayerHelper("chunk_eval", **kwargs) + + # prepare output + precision = helper.create_tmp_variable(dtype="float32") + recall = helper.create_tmp_variable(dtype="float32") + f1_score = helper.create_tmp_variable(dtype="float32") + num_infer_chunks = helper.create_tmp_variable(dtype="int64") + num_label_chunks = helper.create_tmp_variable(dtype="int64") + num_correct_chunks = helper.create_tmp_variable(dtype="int64") + + helper.append_op( + type="chunk_eval", + inputs={"Inference": [input], + "Label": [label]}, + outputs={ + "Precision": [precision], + "Recall": [recall], + "F1-Score": [f1_score], + "NumInferChunks": [num_infer_chunks], + "NumLabelChunks": [num_label_chunks], + "NumCorrectChunks": [num_correct_chunks] + }, + attrs={ + "num_chunk_types": num_chunk_types, + "chunk_scheme": chunk_scheme, + "excluded_chunk_types": excluded_chunk_types or [] + }) + return precision, recall, f1_score, num_infer_chunks, num_label_chunks, num_correct_chunks + + +def sequence_conv(input, + num_filters, + filter_size=3, + filter_stride=1, + padding=None, + bias_attr=None, + param_attr=None, + act=None): + """ + This function creates the op for sequence_conv, using the inputs and + other convolutional configurations for the filters and stride as given + in the input parameters to the function. + """ + + # FIXME(dzh) : want to unify the argument of python layer + # function. So we ignore some unecessary attributes. + # such as, padding_trainable, context_start. + + helper = LayerHelper('sequence_conv', **locals()) + dtype = helper.input_dtype() + filter_shape = [filter_size * input.shape[1], num_filters] + filter_param = helper.create_parameter( + attr=helper.param_attr, shape=filter_shape, dtype=dtype) + pre_bias = helper.create_tmp_variable(dtype) + + helper.append_op( + type='sequence_conv', + inputs={ + 'X': [input], + 'Filter': [filter_param], + }, + outputs={"Out": pre_bias}, + attrs={ + 'contextStride': filter_stride, + 'contextStart': -int(filter_size / 2), + 'contextLength': filter_size + }) + pre_act = helper.append_bias_op(pre_bias) + return helper.append_activation(pre_act) + + +def conv2d(input, + num_filters, + filter_size, + stride=None, + padding=None, + groups=None, + param_attr=None, + bias_attr=None, + act=None): + """ + **Convlution2D Layer** + + The convolution2D layer calculates the output based on the input, filter + and strides, paddings, dilations, groups parameters. Input(Input) and Output(Output) + are in NCHW format. Where N is batch size, C is the number of channels, H is the height + of the feature, and W is the width of the feature. + The details of convolution layer, please refer UFLDL's `convolution, + `_ . + If bias attribution and activation type are provided, bias is added to the output of the convolution, + and the corresponding activation function is applied to the final result. + For each input :math:`X`, the equation is: + + + .. math:: + + Out = \sigma (W \\ast X + b) + + In the above equation: + + * :math:`X`: Input value, a tensor with NCHW format. + * :math:`W`: Filter value, a tensor with MCHW format. + * :math:`\\ast`: Convolution operation. + * :math:`b`: Bias value, a 2-D tensor with shape [M, 1]. + * :math:`\\sigma`: Activation function. + * :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be different. + + Example: + + Input: + Input shape: $(N, C_{in}, H_{in}, W_{in})$ + + Filter shape: $(C_{out}, C_{in}, H_f, W_f)$ + + Output: + Output shape: $(N, C_{out}, H_{out}, W_{out})$ + Where + .. math:: + + H_{out}&= \\frac{(H_{in} + 2 * paddings[0] - (dilations[0] * (H_f - 1) + 1))}{strides[0]} + 1 \\\\ + W_{out}&= \\frac{(W_{in} + 2 * paddings[1] - (dilations[1] * (W_f - 1) + 1))}{strides[1]} + 1 + + Args: + input(Variable): The input image with [N, C, H, W] format. + num_filters(int): The number of filter. It is as same as the output + image channel. + filter_size(int|tuple|None): The filter size. If filter_size is a tuple, + it must contain two integers, (filter_size_H, filter_size_W). + Otherwise, the filter will be a square. + stride(int|tuple): The stride size. If stride is a tuple, it must + contain two integers, (stride_H, stride_W). Otherwise, the + stride_H = stride_W = stride. Default: stride = 1. + padding(int|tuple): The padding size. If padding is a tuple, it must + contain two integers, (padding_H, padding_W). Otherwise, the + padding_H = padding_W = padding. Default: padding = 0. + groups(int): The groups number of the Conv2d Layer. According to grouped + convolution in Alex Krizhevsky's Deep CNN paper: 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 Layer. Default: None + bias_attr(ParamAttr): Bias parameter for the Conv2d layer. Default: None + act(str): Activation type. Default: None + + Returns: + Variable: The tensor variable storing the convolution and \ + non-linearity activation result. + + Raises: + ValueError: If the shapes of input, filter_size, stride, padding and groups mismatch. + + Examples: + .. code-block:: python + + data = fluid.layers.data(name='data', shape=[3, 32, 32], dtype='float32') + conv2d = fluid.layers.conv2d(input=data, num_filters=2, filter_size=3, act="relu") + """ + + if stride is None: + stride = [1, 1] + helper = LayerHelper('conv2d', **locals()) + dtype = helper.input_dtype() + + num_channels = input.shape[1] + if groups is None: + num_filter_channels = num_channels + else: + if num_channels % groups != 0: + raise ValueError("num_channels must be divisible by groups.") + num_filter_channels = num_channels / groups + + if isinstance(filter_size, int): + filter_size = [filter_size, filter_size] + if isinstance(stride, int): + stride = [stride, stride] + if isinstance(padding, int): + padding = [padding, padding] + + input_shape = input.shape + filter_shape = [num_filters, num_filter_channels] + filter_size + + def _get_default_param_initializer(): + std = (2.0 / (filter_size[0]**2 * num_channels))**0.5 + return Normal(0.0, std, 0) + + filter_param = helper.create_parameter( + attr=helper.param_attr, + shape=filter_shape, + dtype=dtype, + default_initializer=_get_default_param_initializer()) + + pre_bias = helper.create_tmp_variable(dtype) + + helper.append_op( + type='conv2d_cudnn', + inputs={ + 'Input': input, + 'Filter': filter_param, + }, + outputs={"Output": pre_bias}, + attrs={'strides': stride, + 'paddings': padding, + 'groups': groups}) + + pre_act = helper.append_bias_op(pre_bias, dim_start=1, dim_end=2) + + return helper.append_activation(pre_act) + + +def sequence_pool(input, pool_type, **kwargs): + """ + This function add the operator for sequence pooling. + It pools features of all time-steps of each instance, and is applied + on top of the input using pool_type mentioned in the parameters. + + It supports four pool_type: + + - average: :math:`Out[i] = \\frac{\sum_i X_i}{N}` + - sum: :math:`Out[i] = \sum_jX_{ij}` + - sqrt: :math:`Out[i] = \\frac{\sum_jX_{ij}}{\sqrt{len(X_i)}}` + - max: :math:`Out[i] = max(X_i)` + + .. code-block:: text + + x is a 1-level LoDTensor: + x.lod = [[0, 2, 5, 7]] + x.data = [1, 3, 2, 4, 6, 5, 1] + x.dims = [7, 1] + + then output is a Tensor: + out.dim = [3, 1] + with condition len(x.lod[-1]) - 1 == out.dims[0] + + for different pool_type: + average: out.data = [2, 4, 3], where 2=(1+3)/2, 4=(2+4+6)/3, 3=(5+1)/2 + sum : out.data = [4, 12, 6], where 4=1+3, 12=2+4+6, 6=5+1 + 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) + + Args: + input(variable): The input variable which is a LoDTensor. + pool_type (string): The pooling type of sequence_pool. + It supports average, sum, sqrt and max. + + Returns: + The sequence pooling variable which is a Tensor. + + Examples: + + .. code-block:: python + + x = fluid.layers.data(name='x', shape=[7, 1], + dtype='float32', lod_level=1) + avg_x = fluid.layers.sequence_pool(input=x, pool_type='average') + 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') + """ + helper = LayerHelper('sequence_pool', input=input, **kwargs) + dtype = helper.input_dtype() + pool_out = helper.create_tmp_variable(dtype) + max_index = helper.create_tmp_variable(dtype) + + helper.append_op( + type="sequence_pool", + inputs={"X": input}, + outputs={"Out": pool_out, + "MaxIndex": max_index}, + attrs={"pooltype": pool_type.upper()}) + + return pool_out + + +def sequence_first_step(input, **kwargs): + """ + This funciton get the first step of sequence. + + .. code-block:: text + + x is a 1-level LoDTensor: + x.lod = [[0, 2, 5, 7]] + x.data = [1, 3, 2, 4, 6, 5, 1] + x.dims = [7, 1] + + then output is a Tensor: + out.dim = [3, 1] + with condition len(x.lod[-1]) - 1 == out.dims[0] + 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. + + Returns: + The sequence's first step variable which is a Tensor. + + Examples: + + .. code-block:: python + + x = fluid.layers.data(name='x', shape=[7, 1], + dtype='float32', lod_level=1) + x_first_step = fluid.layers.sequence_first_step(input=x) + """ + return sequence_pool(input=input, pool_type="first") + + +def sequence_last_step(input, **kwargs): + """ + This funciton get the last step of sequence. + + .. code-block:: text + + x is a 1-level LoDTensor: + x.lod = [[0, 2, 5, 7]] + x.data = [1, 3, 2, 4, 6, 5, 1] + x.dims = [7, 1] + + then output is a Tensor: + out.dim = [3, 1] + with condition len(x.lod[-1]) - 1 == out.dims[0] + out.data = [3, 6, 1], where 3=last(1,3), 6=last(2,4,6), 1=last(5,1) + + Args: + input(variable): The input variable which is a LoDTensor. + + Returns: + The sequence's last step variable which is a Tensor. + + Examples: + + .. code-block:: python + + x = fluid.layers.data(name='x', shape=[7, 1], + dtype='float32', lod_level=1) + x_last_step = fluid.layers.sequence_last_step(input=x) + """ + return sequence_pool(input=input, pool_type="last") + + +def pool2d(input, + pool_size, + pool_type, + pool_stride=None, + pool_padding=None, + global_pooling=False): + """ + This function adds the operator for pooling in 2 dimensions, using the + pooling configurations mentioned in input parameters. + """ + if pool_padding is None: + pool_padding = [0, 0] + if pool_stride is None: + pool_stride = [1, 1] + if pool_type not in ["max", "avg"]: + raise ValueError( + "Unknown pool_type: '%s'. It can only be 'max' or 'avg'.", + str(pool_type)) + if isinstance(pool_size, int): + pool_size = [pool_size, pool_size] + if isinstance(pool_stride, int): + pool_stride = [pool_stride, pool_stride] + if isinstance(pool_padding, int): + pool_padding = [pool_padding, pool_padding] + + helper = LayerHelper('pool2d', **locals()) + dtype = helper.input_dtype() + pool_out = helper.create_tmp_variable(dtype) + + helper.append_op( + type="pool2d", + inputs={"X": input}, + outputs={"Out": pool_out}, + attrs={ + "pooling_type": pool_type, + "ksize": pool_size, + "global_pooling": global_pooling, + "strides": pool_stride, + "paddings": pool_padding + }) + + return pool_out + + +def batch_norm(input, + act=None, + is_test=False, + momentum=0.9, + epsilon=1e-05, + param_attr=None, + bias_attr=None, + data_layout='NCHW'): + """ + This function helps create an operator to implement + the BatchNorm layer using the configurations from the input parameters. + """ + helper = LayerHelper('batch_norm', **locals()) + dtype = helper.input_dtype() + + input_shape = input.shape + if data_layout == 'NCHW': + channel_num = input_shape[1] + else: + if data_layout == 'NHWC': + channel_num = input_shape[-1] + else: + raise ValueError("unsupported data layout:" + data_layout) + + param_shape = [channel_num] + + # create parameter + scale = helper.create_parameter( + attr=helper.param_attr, + shape=param_shape, + dtype=dtype, + default_initializer=Constant(1.0)) + + bias = helper.create_parameter( + attr=helper.param_attr, shape=param_shape, dtype=dtype, is_bias=True) + + mean = helper.create_global_variable( + dtype=input.dtype, shape=param_shape, persistable=True) + helper.set_variable_initializer(var=mean, initializer=Constant(0.0)) + + variance = helper.create_global_variable( + dtype=input.dtype, shape=param_shape, persistable=True) + helper.set_variable_initializer(var=variance, initializer=Constant(1.0)) + + # create output + # mean and mean_out share the same memory + mean_out = mean + # variance and variance out share the same memory + variance_out = variance + saved_mean = helper.create_tmp_variable(dtype) + saved_variance = helper.create_tmp_variable(dtype) + + batch_norm_out = helper.create_tmp_variable(dtype) + + helper.append_op( + type="batch_norm", + inputs={ + "X": input, + "Scale": scale, + "Bias": bias, + "Mean": mean, + "Variance": variance + }, + outputs={ + "Y": batch_norm_out, + "MeanOut": mean_out, + "VarianceOut": variance_out, + "SavedMean": saved_mean, + "SavedVariance": saved_variance + }, + attrs={"momentum": momentum, + "epsilon": epsilon, + "is_test": is_test}) + + return helper.append_activation(batch_norm_out) + + +def beam_search_decode(ids, scores): + helper = LayerHelper('beam_search_decode', **locals()) + sentence_ids = helper.create_tmp_variable(dtype=ids.dtype) + sentence_scores = helper.create_tmp_variable(dtype=ids.dtype) + + helper.append_op( + type="beam_search_decode", + inputs={"Ids": ids, + "Scores": scores}, + outputs={ + "SentenceIds": sentence_ids, + "SentenceScores": sentence_scores + }) + + return sentence_ids, sentence_scores + + +def conv2d_transpose(input, + num_filters, + output_size=None, + filter_size=None, + padding=None, + stride=None, + dilation=None, + param_attr=None): + """ + The transpose of conv2d layer. + + This layer is also known as deconvolution layer. + + Args: + input(Variable): The input image with [N, C, H, W] format. + num_filters(int): The number of filter. It is as same as the output + image channel. + output_size(int|tuple|None): The output image size. If output size is a + tuple, it must contain two integers, (image_H, image_W). This + parameter only works when filter_size is None. + filter_size(int|tuple|None): The filter size. If filter_size is a tuple, + it must contain two integers, (filter_size_H, filter_size_W). + Otherwise, the filter will be a square. None if use output size to + calculate filter_size + padding(int|tuple): The padding size. If padding is a tuple, it must + contain two integers, (padding_H, padding_W). Otherwise, the + padding_H = padding_W = padding. + stride(int|tuple): The stride size. If stride is a tuple, it must + contain two integers, (stride_H, stride_W). Otherwise, the + stride_H = stride_W = stride. + 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. + param_attr: Parameter Attribute. + main_program(Program): the main program + startup_program(Program): the startup program + + Returns: + Variable: Output image. + """ + helper = LayerHelper("conv2d_transpose", **locals()) + if not isinstance(input, Variable): + raise TypeError("Input of conv2d_transpose must be Variable") + input_channel = input.shape[1] + + op_attr = dict() + + if isinstance(padding, int): + op_attr['paddings'] = [padding, padding] + elif padding is not None: + op_attr['paddings'] = padding + + if isinstance(stride, int): + op_attr['strides'] = [stride, stride] + elif stride is not None: + op_attr['strides'] = stride + + if isinstance(dilation, int): + op_attr['dilations'] = [dilation, dilation] + elif dilation is not None: + op_attr['dilations'] = dilation + + if filter_size is None: + if output_size is None: + raise ValueError("output_size must be set when filter_size is None") + if isinstance(output_size, int): + output_size = [output_size, output_size] + + padding = op_attr.get('paddings', [0, 0]) + stride = op_attr.get('strides', [1, 1]) + dilation = op_attr.get('dilations', [1, 1]) + + h_in = input.shape[2] + w_in = input.shape[3] + + filter_size_h = (output_size[0] - (h_in - 1) * stride[0] + 2 * + padding[0] - 1) / dilation[0] + 1 + filter_size_w = (output_size[1] - (w_in - 1) * stride[1] + 2 * + padding[1] - 1) / dilation[1] + 1 + filter_size = [filter_size_h, filter_size_w] + + elif isinstance(filter_size, int): + filter_size = [filter_size, filter_size] + + filter_shape = [input_channel, num_filters] + filter_size + img_filter = helper.create_parameter( + dtype=input.dtype, shape=filter_shape, attr=helper.param_attr) + + out = helper.create_tmp_variable(dtype=input.dtype) + helper.append_op( + type='conv2d_transpose', + inputs={'Input': [input], + 'Filter': [img_filter]}, + outputs={'Output': out}, + attrs=op_attr) + + return out + + +def sequence_expand(x, y): + """Sequence Expand Layer. This layer will expand the input variable **x** + according to LoD information of **y**. And the following examples will + explain how sequence_expand works: + + .. code-block:: text + + * Case 1 + x is a LoDTensor: + x.lod = [[0, 2, 3], + [0, 1, 3, 4]] + x.data = [a, b, c, d] + x.dims = [4, 1] + + y is a LoDTensor: + y.lod = [[0, 2, 4], + [0, 3, 6, 7, 8]] + + with condition len(y.lod[-1]) - 1 == x.dims[0] + + then output is a 2-level LoDTensor: + out.lod = [[0, 2, 4], + [0, 3, 6, 7, 8]] + out.data = [a, a, a, b, b, b, c, d] + out.dims = [8, 1] + + * Case 2 + x is a Tensor: + x.data = [a, b, c] + x.dims = [3, 1] + + y is a LoDTensor: + y.lod = [[0, 2, 3, 6]] + + with condition len(y.lod[-1]) - 1 == x.dims[0] + + then output is a 1-level LoDTensor: + out.lod = [[0, 2, 3, 6]] + out.data = [a, a, b, c, c, c] + out.dims = [6, 1] + + Args: + x (Variable): The input variable which is a Tensor or LoDTensor. + y (Variable): The input variable which is a LoDTensor. + + Returns: + Variable: The expanded variable which is a LoDTensor. + + Examples: + .. code-block:: python + + x = fluid.layers.data(name='x', shape=[10], dtype='float32') + y = fluid.layers.data(name='y', shape=[10, 20], + dtype='float32', lod_level=1) + out = layers.sequence_expand(x=x, y=y) + """ + helper = LayerHelper('sequence_expand', input=x, **locals()) + dtype = helper.input_dtype() + tmp = helper.create_tmp_variable(dtype) + helper.append_op( + type='sequence_expand', inputs={'X': x, + 'Y': y}, outputs={'Out': tmp}) + return tmp + + +def lstm_unit(x_t, + hidden_t_prev, + cell_t_prev, + forget_bias=0.0, + param_attr=None, + bias_attr=None): + """Lstm unit layer. The equation of a lstm step is: + + .. math:: + + i_t & = \sigma(W_{x_i}x_{t} + W_{h_i}h_{t-1} + W_{c_i}c_{t-1} + b_i) + + f_t & = \sigma(W_{x_f}x_{t} + W_{h_f}h_{t-1} + W_{c_f}c_{t-1} + b_f) + + c_t & = f_tc_{t-1} + i_t tanh (W_{x_c}x_t+W_{h_c}h_{t-1} + b_c) + + o_t & = \sigma(W_{x_o}x_{t} + W_{h_o}h_{t-1} + W_{c_o}c_t + b_o) + + h_t & = o_t tanh(c_t) + + The inputs of lstm unit includes :math:`x_t`, :math:`h_{t-1}` and + :math:`c_{t-1}`. The implementation separates the linear transformation + and non-linear transformation apart. Here, we take :math:`i_t` as an + example. The linear transformation is applied by calling a `fc` layer and + the equation is: + + .. math:: + + L_{i_t} = W_{x_i}x_{t} + W_{h_i}h_{t-1} + W_{c_i}c_{t-1} + b_i + + The non-linear transformation is applied by calling `lstm_unit_op` and the + equation is: + + .. math:: + + i_t = \sigma(L_{i_t}) + + This layer has two outputs including :math:`h_t` and :math:`o_t`. + + Args: + x_t (Variable): The input value of current step. + hidden_t_prev (Variable): The hidden value of lstm unit. + cell_t_prev (Variable): The cell value of lstm unit. + forget_bias (float): The forget bias of lstm unit. + param_attr (ParamAttr): The attributes of parameter weights, used to set + initializer, name etc. + bias_attr (ParamAttr): The attributes of bias weights, if not False, + bias weights will be created and be set to default value. + + Returns: + tuple: The hidden value and cell value of lstm unit. + + Raises: + ValueError: The ranks of **x_t**, **hidden_t_prev** and **cell_t_prev**\ + not be 2 or the 1st dimensions of **x_t**, **hidden_t_prev** \ + and **cell_t_prev** not be the same. + + Examples: + + .. code-block:: python + + x_t = fluid.layers.fc(input=x_t_data, size=10) + prev_hidden = fluid.layers.fc(input=prev_hidden_data, size=20) + prev_cell = fluid.layers.fc(input=prev_cell_data, size=30) + hidden_value, cell_value = fluid.layers.lstm_unit(x_t=x_t, + hidden_t_prev=prev_hidden, + cell_t_prev=prev_cell) + """ + helper = LayerHelper('lstm_unit', **locals()) + + if len(x_t.shape) != 2: + raise ValueError("Rank of x_t must be 2.") + + if len(hidden_t_prev.shape) != 2: + raise ValueError("Rank of hidden_t_prev must be 2.") + + if len(cell_t_prev.shape) != 2: + raise ValueError("Rank of cell_t_prev must be 2.") + + if x_t.shape[0] != hidden_t_prev.shape[0] or x_t.shape[ + 0] != cell_t_prev.shape[0]: + raise ValueError("The 1s dimension of x_t, hidden_t_prev and " + "cell_t_prev must be the same.") + + if bias_attr is None: + bias_attr = ParamAttr() + + size = cell_t_prev.shape[1] + concat_out = concat(input=[x_t, hidden_t_prev], axis=1) + fc_out = fc(input=concat_out, + size=4 * size, + param_attr=param_attr, + bias_attr=bias_attr) + dtype = x_t.dtype + c = helper.create_tmp_variable(dtype) + h = helper.create_tmp_variable(dtype) + + helper.append_op( + type='lstm_unit', + inputs={"X": fc_out, + "C_prev": cell_t_prev}, + outputs={"C": c, + "H": h}, + attrs={"forget_bias": forget_bias}) + + return h, c + + +def reduce_sum(input, dim=None, keep_dim=False): + """ + Computes the sum of tensor elements over the given dimension. + + Args: + input (Variable): The input variable which is a Tensor or LoDTensor. + dim (int|None): The dimension 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`. + 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. + + Returns: + Variable: The reduced Tensor variable. + + Examples: + .. code-block:: python + + # x is a Tensor variable with following elements: + # [[0.2, 0.3, 0.5, 0.9] + # [0.1, 0.2, 0.6, 0.7]] + # Each example is followed by the correspending output tensor. + fluid.layers.reduce_sum(x) # [3.5] + 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]] + """ + helper = LayerHelper('reduce_sum', **locals()) + out = helper.create_tmp_variable(dtype=helper.input_dtype()) + helper.append_op( + type='reduce_sum', + inputs={'X': input}, + outputs={'Out': out}, + attrs={ + 'dim': dim if dim != None else 0, + 'keep_dim': keep_dim, + 'reduce_all': True if dim == None else False + }) + return out + + +def reduce_mean(input, dim=None, keep_dim=False): + """ + Computes the mean of tensor elements over the given dimension. + + Args: + input (Variable): The input variable which is a Tensor or LoDTensor. + dim (int|None): The dimension 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`. + 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. + + Returns: + Variable: The reduced Tensor variable. + + Examples: + .. code-block:: python + + # x is a Tensor variable with following elements: + # [[0.2, 0.3, 0.5, 0.9] + # [0.1, 0.2, 0.6, 0.7]] + # Each example is followed by the correspending output tensor. + 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]] + """ + helper = LayerHelper('reduce_mean', **locals()) + out = helper.create_tmp_variable(dtype=helper.input_dtype()) + helper.append_op( + type='reduce_mean', + inputs={'X': input}, + outputs={'Out': out}, + attrs={ + 'dim': dim if dim != None else 0, + 'keep_dim': keep_dim, + 'reduce_all': True if dim == None else False + }) + return out + + +def reduce_max(input, dim=None, keep_dim=False): + """ + Computes the maximum of tensor elements over the given dimension. + + Args: + input (Variable): The input variable which is a Tensor or LoDTensor. + dim (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`. + 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. + + Returns: + Variable: The reduced Tensor variable. + + Examples: + .. code-block:: python + + # x is a Tensor variable with following elements: + # [[0.2, 0.3, 0.5, 0.9] + # [0.1, 0.2, 0.6, 0.7]] + # Each example is followed by the correspending output tensor. + fluid.layers.reduce_max(x) # [0.9] + 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]] + """ + helper = LayerHelper('reduce_max', **locals()) + out = helper.create_tmp_variable(dtype=helper.input_dtype()) + helper.append_op( + type='reduce_max', + inputs={'X': input}, + outputs={'Out': out}, + attrs={ + 'dim': dim if dim != None else 0, + 'keep_dim': keep_dim, + 'reduce_all': True if dim == None else False + }) + return out + + +def reduce_min(input, dim=None, keep_dim=False): + """ + Computes the minimum of tensor elements over the given dimension. + + Args: + input (Variable): The input variable which is a Tensor or LoDTensor. + dim (int|None): The dimension 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`. + 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. + + Returns: + Variable: The reduced Tensor variable. + + Examples: + .. code-block:: python + + # x is a Tensor variable with following elements: + # [[0.2, 0.3, 0.5, 0.9] + # [0.1, 0.2, 0.6, 0.7]] + # Each example is followed by the correspending output tensor. + fluid.layers.reduce_min(x) # [0.1] + 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]] + """ + helper = LayerHelper('reduce_min', **locals()) + out = helper.create_tmp_variable(dtype=helper.input_dtype()) + helper.append_op( + type='reduce_min', + inputs={'X': input}, + outputs={'Out': out}, + attrs={ + 'dim': dim if dim != None else 0, + 'keep_dim': keep_dim, + 'reduce_all': True if dim == None else False + }) + return out diff --git a/python/paddle/v2/fluid/layers/ops.py b/python/paddle/v2/fluid/layers/ops.py new file mode 100644 index 0000000000000000000000000000000000000000..d2ff6841a317aaf6903edadc9213f69ef6c41216 --- /dev/null +++ b/python/paddle/v2/fluid/layers/ops.py @@ -0,0 +1,9 @@ +from ..registry import register_layer +__all__ = [ + 'mean', 'mul', 'dropout', 'reshape', 'sigmoid', 'scale', 'transpose', + 'sigmoid_cross_entropy_with_logits', 'elementwise_add', 'elementwise_div', + 'elementwise_sub', 'elementwise_mul', 'clip', 'abs', 'sequence_softmax' +] + +for _OP in set(__all__): + globals()[_OP] = register_layer(_OP) diff --git a/python/paddle/v2/fluid/layers/tensor.py b/python/paddle/v2/fluid/layers/tensor.py new file mode 100644 index 0000000000000000000000000000000000000000..e5820d24cd2b34ef53cbb91e2be66efc1b74d315 --- /dev/null +++ b/python/paddle/v2/fluid/layers/tensor.py @@ -0,0 +1,215 @@ +from ..layer_helper import LayerHelper + +__all__ = [ + 'create_tensor', 'cast', 'concat', 'sums', 'assign', + 'fill_constant_batch_size_like', 'fill_constant', 'ones', 'zeros' +] + + +def create_tensor(dtype, name=None): + helper = LayerHelper("create_tensor", **locals()) + return helper.create_variable(name=helper.name, dtype=dtype) + + +def cast(x, dtype): + """ + This function takes in the input with input_dtype + and casts it to the output_dtype as the output. + """ + helper = LayerHelper('cast', **locals()) + out = helper.create_tmp_variable(dtype=dtype) + helper.append_op( + type='cast', + inputs={'X': [x]}, + outputs={'Out': [out]}, + attrs={'in_dtype': x.dtype, + 'out_dtype': out.dtype}) + return out + + +def concat(input, axis=0): + """ + **Concat** + + This function concatenates the input along the axis mentioned + and returns that as the output. + + Args: + input(list): List of tensors to be concatenated + axis(int): Integer axis along which the tensors will be concatenated + + Returns: + Variable: Output variable of the concatenation + + Examples: + .. code-block:: python + out = fluid.layers.concat(input=[Efirst, Esecond, Ethird, Efourth]) + """ + helper = LayerHelper('concat', **locals()) + out = helper.create_tmp_variable(dtype=helper.input_dtype()) + helper.append_op( + type='concat', + inputs={'X': input}, + outputs={'Out': [out]}, + attrs={'axis': axis}) + return out + + +def sums(input, out=None): + """This function performs the sum operation on the input and returns the + result as the output. + + Args: + input (Variable|list): The input tensor that has the elements + that need to be summed up. + + Returns: + Variable: The tensor type variable that has the sum of input + written to it. + + Examples: + .. code-block::python + + tmp = fluid.layers.zeros(shape=[10], dtype='int32') + i = fluid.layers.fill_constant(shape=[1], dtype='int64', value=10) + a0 = layers.array_read(array=tmp, i=i) + i = layers.increment(x=i) + a1 = layers.array_read(array=tmp, i=i) + mean_a0 = layers.mean(x=a0) + mean_a1 = layers.mean(x=a1) + a_sum = layers.sums(input=[mean_a0, mean_a1]) + """ + helper = LayerHelper('sum', **locals()) + if out is None: + out = helper.create_tmp_variable(dtype=helper.input_dtype()) + helper.append_op(type='sum', inputs={'X': input}, outputs={'Out': out}) + return out + + +def assign(input, output): + """ + **Assign** + + This function copies the *input* Variable to the *output* Variable. + + Args: + input(Variable): The source variable + output(Variable): The destination variable + + Returns: + Variable: The destination variable that was supplied as the *output*. + + Examples: + .. code-block:: python + out = fluid.layers.create_tensor(dtype='float32') + hidden = fluid.layers.fc(input=data, size=10) + fluid.layers.assign(hidden, out) + """ + helper = LayerHelper('assign', **locals()) + helper.append_op( + type='scale', + inputs={'X': [input]}, + outputs={'Out': [output]}, + attrs={'scale': 1.0}) + return output + + +def fill_constant(shape, dtype, value, out=None): + """ + **fill_constant** + + This function creates a tensor of specified *shape* and + *dtype*, and initializes this with a constant supplied in *value*. + + It also sets *stop_gradient* to True. + + Args: + shape(tuple|list|None): Shape of output tensor + dtype(np.dtype|core.DataType|str): Data type of output tensor + value(float): Constant value to initialize the output tensor + out(Variable): Output Variable to initialize + + Returns: + Variable: The tensor variable storing the output + + Examples: + .. code-block:: python + + data = fluid.layers.fill_constant(shape=[1], value=0, dtype='int64') + """ + helper = LayerHelper("fill_constant", **locals()) + if out is None: + out = helper.create_tmp_variable(dtype=dtype) + helper.append_op( + type='fill_constant', + inputs={}, + outputs={'Out': [out]}, + attrs={'shape': shape, + 'dtype': out.dtype, + 'value': float(value)}) + out.stop_gradient = True + return out + + +def fill_constant_batch_size_like(input, + shape, + dtype, + value, + input_dim_idx=0, + output_dim_idx=0): + """ + **fill_constant_batch_size_like** + + This function creates a tensor of specified *shape*, *dtype* and batch size, + and initializes this with a constant supplied in *value*. The batch size is + obtained from the `input` tensor. + + It also sets *stop_gradient* to True. + + Args: + input(Variable): Tensor whose dimensions will be used to get batch size + shape(tuple|list|None): Shape of output tensor + dtype(np.dtype|core.DataType|str): Data type of output tensor + value(float): Constant value to initialize the output tensor + input_dim_idx(int): Index of input's batch size dimension + output_dim_idx(int): Index of output's batch size dimension + + Returns: + Variable: The tensor variable storing the output + + Examples: + .. code-block:: python + + data = fluid.layers.fill_constant(shape=[1], value=0, dtype='int64') + """ + helper = LayerHelper("fill_constant_batch_size_like", **locals()) + out = helper.create_tmp_variable(dtype=dtype) + helper.append_op( + type='fill_constant_batch_size_like', + inputs={'Input': input}, + outputs={'Out': [out]}, + attrs={ + 'shape': shape, + 'dtype': out.dtype, + 'value': float(value), + 'input_dim_idx': input_dim_idx, + 'output_dim_idx': output_dim_idx + }) + out.stop_gradient = True + return out + + +def ones(shape, dtype): + """ + This function performs the same function as fill_constant() declared above + with the constant value being 1.0. + """ + return fill_constant(value=1.0, **locals()) + + +def zeros(shape, dtype): + """ + This function performs the same function as fill_constant() declared above + with the constant value being 0.0. + """ + return fill_constant(value=0.0, **locals()) diff --git a/python/paddle/v2/fluid/nets.py b/python/paddle/v2/fluid/nets.py index 5e14ca594bc7965dc29039ba57bb7b26b1ce6871..54886a8f2cc63474fe82290c0a12771b4cbdba72 100644 --- a/python/paddle/v2/fluid/nets.py +++ b/python/paddle/v2/fluid/nets.py @@ -1,4 +1,4 @@ -import paddle.v2.fluid.layers as layers +import layers __all__ = ["simple_img_conv_pool", "sequence_conv_pool"] @@ -9,24 +9,20 @@ def simple_img_conv_pool(input, pool_size, pool_stride, act, - pool_type='max', - main_program=None, - startup_program=None): + param_attr=None, + pool_type='max'): conv_out = layers.conv2d( input=input, num_filters=num_filters, filter_size=filter_size, - act=act, - main_program=main_program, - startup_program=startup_program) + param_attr=param_attr, + act=act) pool_out = layers.pool2d( input=conv_out, pool_size=pool_size, pool_type=pool_type, - pool_stride=pool_stride, - main_program=main_program, - startup_program=startup_program) + pool_stride=pool_stride) return pool_out @@ -36,12 +32,11 @@ def img_conv_group(input, conv_padding=1, conv_filter_size=3, conv_act=None, + param_attr=None, conv_with_batchnorm=False, conv_batchnorm_drop_rate=None, pool_stride=1, - pool_type=None, - main_program=None, - startup_program=None): + pool_type=None): """ Image Convolution Group, Used for vgg net. """ @@ -57,6 +52,7 @@ def img_conv_group(input, conv_padding = __extend_list__(conv_padding) conv_filter_size = __extend_list__(conv_filter_size) + param_attr = __extend_list__(param_attr) conv_with_batchnorm = __extend_list__(conv_with_batchnorm) conv_batchnorm_drop_rate = __extend_list__(conv_batchnorm_drop_rate) @@ -70,52 +66,35 @@ def img_conv_group(input, num_filters=conv_num_filter[i], filter_size=conv_filter_size[i], padding=conv_padding[i], - act=local_conv_act, - main_program=main_program, - startup_program=startup_program) + param_attr=param_attr[i], + act=local_conv_act) if conv_with_batchnorm[i]: - tmp = layers.batch_norm( - input=tmp, - act=conv_act, - main_program=main_program, - startup_program=startup_program) + tmp = layers.batch_norm(input=tmp, act=conv_act) drop_rate = conv_batchnorm_drop_rate[i] if abs(drop_rate) > 1e-5: - tmp = layers.dropout( - x=tmp, - dropout_prob=drop_rate, - main_program=main_program, - startup_program=startup_program) + tmp = layers.dropout(x=tmp, dropout_prob=drop_rate) pool_out = layers.pool2d( input=tmp, pool_size=pool_size, pool_type=pool_type, - pool_stride=pool_stride, - main_program=main_program, - startup_program=startup_program) + pool_stride=pool_stride) return pool_out def sequence_conv_pool(input, num_filters, filter_size, + param_attr=None, act="sigmoid", - pool_type="max", - main_program=None, - startup_program=None): + pool_type="max"): conv_out = layers.sequence_conv( input=input, num_filters=num_filters, filter_size=filter_size, - act=act, - main_program=main_program, - startup_program=startup_program) + param_attr=param_attr, + act=act) - pool_out = layers.sequence_pool( - input=conv_out, - pool_type=pool_type, - main_program=main_program, - startup_program=startup_program) + pool_out = layers.sequence_pool(input=conv_out, pool_type=pool_type) return pool_out diff --git a/python/paddle/v2/fluid/optimizer.py b/python/paddle/v2/fluid/optimizer.py index d2841df6af7a0d860c239db952c767c995d30ba4..ff3e5315a2c2b115e4ba563f60de4139f248e93a 100644 --- a/python/paddle/v2/fluid/optimizer.py +++ b/python/paddle/v2/fluid/optimizer.py @@ -1,16 +1,14 @@ from collections import defaultdict -import paddle.v2.fluid.framework as framework -from paddle.v2.fluid.framework import unique_name, Program -from paddle.v2.fluid.backward import append_backward_ops -from paddle.v2.fluid.initializer import ConstantInitializer -from paddle.v2.fluid.regularizer import append_regularization_ops -from paddle.v2.fluid.layer_helper import LayerHelper +import framework +from backward import append_backward +from framework import unique_name, program_guard +from initializer import Constant +from layer_helper import LayerHelper +from regularizer import append_regularization_ops +from clip import append_gradient_clip_ops -__all__ = [ - 'SGDOptimizer', 'MomentumOptimizer', 'AdagradOptimizer', 'AdamOptimizer', - 'AdamaxOptimizer', 'DecayedAdagradOptimizer' -] +__all__ = ['SGD', 'Momentum', 'Adagrad', 'Adam', 'Adamax', 'DecayedAdagrad'] class Optimizer(object): @@ -21,8 +19,9 @@ class Optimizer(object): but need to use one of it's implementation. """ - def __init__(self, global_step=None): + def __init__(self, global_step=None, regularization=None): self._global_step = global_step + self.regularization = regularization # Dictionary of accumulators. Some optimizer subclasses need to # allocate and manage extra variables associated with the parameters # to train. These variables are called accumulators. @@ -48,7 +47,7 @@ class Optimizer(object): persistable=True) param_lr = param_lr * self._learning_rate self.helper.set_variable_initializer( - var=param_lr_var, initializer=ConstantInitializer(param_lr)) + var=param_lr_var, initializer=Constant(param_lr)) return param_lr_var def _create_accumulators(self, block, parameters): @@ -92,11 +91,11 @@ class Optimizer(object): var = self.helper.create_global_variable( name=unique_name(name), persistable=True, - dtype=dtype or param.data_type, + dtype=dtype or param.dtype, type=param.type, shape=param.shape) self.helper.set_variable_initializer( - var, initializer=ConstantInitializer(value=float(fill_value))) + var, initializer=Constant(value=float(fill_value))) self._accumulators[name][param.name] = var def _get_accumulator(self, name, param): @@ -161,33 +160,32 @@ class Optimizer(object): # Create any accumulators program = loss.block.program - self.helper = LayerHelper( - self.__class__.__name__, - main_program=program, - startup_program=startup_program) - self._create_accumulators(loss.block, - [p[0] for p in parameters_and_grads]) - - optimize_ops = [] - for param_and_grad in parameters_and_grads: - if param_and_grad[1] is not None: - optimize_op = self._append_optimize_op(loss.block, - param_and_grad) - optimize_ops.append(optimize_op) - - # Returned list of ops can include more ops in addition - # to optimization ops - return_ops = optimize_ops - - # Get custom finish ops for subclasses - # FIXME: Need to fix this once we figure out how to handle dependencies - finish_ops = self._finish_update(loss.block) - if finish_ops is not None: - return_ops += finish_ops - - if self._global_step is not None: - return_ops.append(self._increment_global_step(loss.block)) - return return_ops + with program_guard(program, startup_program): + self.helper = LayerHelper(self.__class__.__name__) + self._create_accumulators(loss.block, + [p[0] for p in parameters_and_grads]) + + 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) + + # Returned list of ops can include more ops in addition + # to optimization ops + return_ops = optimize_ops + + # Get custom finish ops for subclasses + # FIXME: Need to fix this once we figure out how to handle dependencies + finish_ops = self._finish_update(loss.block) + if finish_ops is not None: + return_ops += finish_ops + + if self._global_step is not None: + return_ops.append(self._increment_global_step(loss.block)) + return return_ops def minimize(self, loss, @@ -196,25 +194,29 @@ class Optimizer(object): no_grad_set=None): """Add operations to minimize `loss` by updating `parameter_list`. - This method combines interface `append_backward_ops()` and + This method combines interface `append_backward()` and `create_optimization_pass()` into one. """ - params_grads = append_backward_ops(loss, parameter_list, no_grad_set or - set()) - # Add regularization if any - params_grads = append_regularization_ops(params_grads) + params_grads = append_backward(loss, parameter_list, no_grad_set) + + params_grads = append_gradient_clip_ops(params_grads) + + # Add regularization if any + params_grads = append_regularization_ops(params_grads, + self.regularization) + optimize_ops = self.create_optimization_pass(params_grads, loss, startup_program) - return optimize_ops + return optimize_ops, params_grads class SGDOptimizer(Optimizer): """ Simple SGD optimizer without any state. """ - def __init__(self, learning_rate, global_step=None): + def __init__(self, learning_rate, **kwargs): assert learning_rate is not None - super(SGDOptimizer, self).__init__(global_step) + super(SGDOptimizer, self).__init__(**kwargs) self.type = "sgd" self._learning_rate = learning_rate @@ -239,14 +241,10 @@ class MomentumOptimizer(Optimizer): """ _velocity_acc_str = "velocity" - def __init__(self, - learning_rate, - momentum, - use_nesterov=False, - global_step=None): + def __init__(self, learning_rate, momentum, use_nesterov=False, **kwargs): assert learning_rate is not None assert momentum is not None - super(MomentumOptimizer, self).__init__(global_step) + super(MomentumOptimizer, self).__init__(**kwargs) self.type = "momentum" self._learning_rate = learning_rate self._momentum = momentum @@ -287,10 +285,10 @@ class AdagradOptimizer(Optimizer): """ _moment_acc_str = "moment" - def __init__(self, learning_rate, epsilon=1.0e-6, global_step=None): + def __init__(self, learning_rate, epsilon=1.0e-6, **kwargs): assert learning_rate is not None assert epsilon is not None - super(AdagradOptimizer, self).__init__(global_step) + super(AdagradOptimizer, self).__init__(**kwargs) self.type = "adagrad" self._learning_rate = learning_rate self._epsilon = epsilon @@ -334,12 +332,12 @@ class AdamOptimizer(Optimizer): beta1=0.9, beta2=0.999, epsilon=1e-8, - global_step=None): + **kwargs): assert learning_rate is not None assert beta1 is not None assert beta2 is not None assert epsilon is not None - super(AdamOptimizer, self).__init__(global_step) + super(AdamOptimizer, self).__init__(**kwargs) self.type = "adam" self._learning_rate = learning_rate self._beta1 = beta1 @@ -359,7 +357,7 @@ class AdamOptimizer(Optimizer): lod_level=0, persistable=True) self.helper.set_variable_initializer( - self._beta1_pow_acc, initializer=ConstantInitializer(self._beta1)) + self._beta1_pow_acc, initializer=Constant(self._beta1)) self._beta2_pow_acc = self.helper.create_global_variable( name=unique_name('beta2_pow_acc'), @@ -369,7 +367,7 @@ class AdamOptimizer(Optimizer): persistable=True) self.helper.set_variable_initializer( - self._beta2_pow_acc, initializer=ConstantInitializer(self._beta2)) + self._beta2_pow_acc, initializer=Constant(self._beta2)) # Create accumulator tensors for first and second moments for p in parameters: @@ -439,12 +437,12 @@ class AdamaxOptimizer(Optimizer): beta1=0.9, beta2=0.999, epsilon=1e-8, - global_step=None): + **kwargs): assert learning_rate is not None assert beta1 is not None assert beta2 is not None assert epsilon is not None - super(AdamaxOptimizer, self).__init__() + super(AdamaxOptimizer, self).__init__(**kwargs) self.type = "adamax" self._learning_rate = learning_rate self._beta1 = beta1 @@ -461,7 +459,7 @@ class AdamaxOptimizer(Optimizer): lod_level=0, persistable=True) self.helper.set_variable_initializer( - self._beta1_pow_acc, initializer=ConstantInitializer(self._beta1)) + self._beta1_pow_acc, initializer=Constant(self._beta1)) # Create accumulator tensors for first moment and infinity norm for p in parameters: @@ -517,16 +515,12 @@ class DecayedAdagradOptimizer(Optimizer): """ _moment_acc_str = "moment" - def __init__(self, - learning_rate, - decay=0.95, - epsilon=1.0e-6, - global_step=None): + def __init__(self, learning_rate, decay=0.95, epsilon=1.0e-6, **kwargs): assert learning_rate is not None assert decay is not None assert epsilon is not None - super(DecayedAdagradOptimizer, self).__init__(global_step) + super(DecayedAdagradOptimizer, self).__init__(**kwargs) self.type = "decayed_adagrad" self._learning_rate = learning_rate self._decay = decay @@ -558,3 +552,19 @@ class DecayedAdagradOptimizer(Optimizer): attrs={"epsilon": self._epsilon}) return decayed_adagrad_op + + +# We short the class name, since users will use the optimizer with the package +# name. The sample code: +# +# import paddle.fluid as fluid +# +# sgd = fluid.optimizer.SGD(...) +# +# It is no need to add an `Optimizer` as the class suffix +SGD = SGDOptimizer +Momentum = MomentumOptimizer +Adagrad = AdagradOptimizer +Adam = AdamOptimizer +Adamax = AdamaxOptimizer +DecayedAdagrad = DecayedAdagradOptimizer diff --git a/python/paddle/v2/fluid/param_attr.py b/python/paddle/v2/fluid/param_attr.py new file mode 100644 index 0000000000000000000000000000000000000000..ab4561b0423dd73c8c0d529cbf34b52876b1b77c --- /dev/null +++ b/python/paddle/v2/fluid/param_attr.py @@ -0,0 +1,70 @@ +from initializer import Initializer, Xavier, Constant +from regularizer import WeightDecayRegularizer + +__all__ = ['ParamAttr'] + + +class ParamAttr(object): + def __init__(self, + name=None, + initializer=None, + learning_rate=1.0, + regularizer=None, + trainable=True, + clip=None): + self.name = name + self.initializer = initializer + self.learning_rate = learning_rate + self.regularizer = regularizer + self.trainable = trainable + self.clip = clip + + def set_default_initializer(self, initializer): + if initializer is None: + if self.initializer is None: + raise ValueError("ParamAttr.initializer is not set") + return + + if self.initializer is not None: + return + + self.initializer = initializer + + def set_default_param_initializer(self): + self.set_default_initializer(Xavier()) + + def set_default_bias_initializer(self): + self.set_default_initializer(Constant(0.0)) + + @staticmethod + def to_attr(arg): + if arg is None: + return ParamAttr() + elif isinstance(arg, list) or isinstance(arg, tuple): + return [ParamAttr.to_attr(a) for a in arg] + elif isinstance(arg, ParamAttr): + return arg + elif isinstance(arg, str) or isinstance(arg, unicode): + return ParamAttr(name=arg) + elif isinstance(arg, Initializer): + return ParamAttr(initializer=arg) + elif isinstance(arg, WeightDecayRegularizer): + return ParamAttr(regularizer=arg) + elif isinstance(arg, bool): + return ParamAttr.to_attr(None) if arg else False + else: + raise TypeError("{0} cast to ParamAttr".format(type(arg))) + + def to_kwargs(self, with_initializer=False): + kwargs = { + 'name': self.name, + 'optimize_attr': { + 'learning_rate': self.learning_rate + }, + 'regularizer': self.regularizer, + 'trainable': self.trainable, + 'clip_attr': self.clip + } + if with_initializer: + kwargs['initializer'] = self.initializer + return kwargs diff --git a/python/paddle/v2/fluid/profiler.py b/python/paddle/v2/fluid/profiler.py new file mode 100644 index 0000000000000000000000000000000000000000..dcecd76224e70d03ed987a5bb104a977a527d218 --- /dev/null +++ b/python/paddle/v2/fluid/profiler.py @@ -0,0 +1,51 @@ +import paddle.v2.fluid.core as core +from contextlib import contextmanager +import os + +__all__ = ['CudaProfiler'] + +NVPROF_CONFIG = [ + "gpustarttimestamp", + "gpuendtimestamp", + "gridsize3d", + "threadblocksize", + "streamid", + "enableonstart 0", + "conckerneltrace", +] + + +@contextmanager +def cuda_profiler(output_file, output_mode=None, config=None): + """The CUDA profiler. + This fuctions is used to profile CUDA program by CUDA runtime application + programming interface. The profiling result will be written into + `output_file` with Key-Value pair format or Comma separated values format. + The user can set the output mode by `output_mode` argument and set the + counters/options for profiling by `config` argument. The default config + is ['gpustarttimestamp', 'gpustarttimestamp', 'gridsize3d', + 'threadblocksize', 'streamid', 'enableonstart 0', 'conckerneltrace']. + + Args: + output_file (string) : The output file name, the result will be + written into this file. + output_mode (string) : The output mode has Key-Value pair format and + Comma separated values format. It should be 'kvp' or 'csv'. + config (list of string) : The profiler options and counters can refer + to "Compute Command Line Profiler User Guide". + """ + if output_mode is None: + output_mode = 'csv' + if output_mode not in ['kvp', 'csv']: + raise ValueError("The output mode must be 'kvp' or 'csv'.") + config = NVPROF_CONFIG if config is None else config + config_file = 'nvprof_config_file' + with open(config_file, 'wb') as fp: + fp.writelines(["%s\n" % item for item in config]) + core.nvprof_init(output_file, output_mode, config_file) + # Enables profiler collection by the active CUDA profiling tool. + core.nvprof_start() + yield + # Disables profiler collection. + core.nvprof_stop() + os.remove(config_file) diff --git a/python/paddle/v2/fluid/registry.py b/python/paddle/v2/fluid/registry.py new file mode 100644 index 0000000000000000000000000000000000000000..7aa82906114b355277185211134bb791e5dc43f9 --- /dev/null +++ b/python/paddle/v2/fluid/registry.py @@ -0,0 +1,192 @@ +import re +import cStringIO +import warnings +import functools +import inspect + +import proto.framework_pb2 as framework_pb2 +from framework import OpProtoHolder, Variable, Program, Operator +from paddle.v2.fluid.layer_helper import LayerHelper, unique_name + +__all__ = ['deprecated', 'register_layer', 'autodoc'] + + +def _convert_(name): + """ + Formatting. + + Args: + name: The name/alias + + This function takes in a name and converts it to a standard format of + group1_group2. Where as per the regular expression, group1 can have + alphabets and numbers and group2 has capital alphabets. + + """ + s1 = re.sub('(.)([A-Z][a-z]+)', r'\1_\2', name) + return re.sub('([a-z0-9])([A-Z])', r'\1_\2', s1).lower() + + +def _generate_doc_string_(op_proto): + """ + Generate docstring by OpProto + + Args: + op_proto (framework_pb2.OpProto): a protobuf message typed OpProto + + Returns: + str: the document string + """ + + def _type_to_str_(tp): + return framework_pb2.AttrType.Name(tp) + + if not isinstance(op_proto, framework_pb2.OpProto): + raise TypeError("OpProto should be `framework_pb2.OpProto`") + + buf = cStringIO.StringIO() + buf.write(op_proto.comment) + buf.write('\nArgs:\n') + for each_input in op_proto.inputs: + line_begin = ' {0}: '.format(_convert_(each_input.name)) + buf.write(line_begin) + buf.write(each_input.comment) + buf.write('\n') + buf.write(' ' * len(line_begin)) + buf.write('Duplicable: ') + buf.write(str(each_input.duplicable)) + buf.write(' Optional: ') + buf.write(str(each_input.dispensable)) + buf.write('\n') + + for each_attr in op_proto.attrs: + buf.write(' ') + buf.write(each_attr.name) + buf.write(' (') + buf.write(_type_to_str_(each_attr.type)) + buf.write('): ') + buf.write(each_attr.comment) + buf.write('\n') + + if len(op_proto.outputs) != 0: + buf.write('\nReturns:\n') + buf.write(' ') + for each_opt in op_proto.outputs: + if not each_opt.intermediate: + break + buf.write(each_opt.comment) + + return buf.getvalue() + + +def register_layer(op_type): + """ + Register an Python layer for an Operator + + Args: + op_type: The name of the operator to be created + + This function takes in the operator type (sigmoid, mean , average etc) and + creates the operator functionality. + + """ + op_proto = OpProtoHolder.instance().get_op_proto(op_type) + not_intermediate_outputs = \ + filter(lambda output: not output.intermediate, op_proto.outputs) + intermediate_outputs = \ + filter(lambda output: output.intermediate, op_proto.outputs) + + if len(not_intermediate_outputs) != 1: + raise ValueError("Only one non intermediate output operator can be", + "automatically generated") + + if not_intermediate_outputs[0].duplicable: + raise ValueError( + "Only non duplicable op can be automatically generated") + + for output in intermediate_outputs: + if output.duplicable: + raise ValueError("The op can be automatically generated only when ", + "all intermediate ops are not duplicable") + + o_name = not_intermediate_outputs[0].name + intermediate_output_names = [output.name for output in intermediate_outputs] + + def infer_and_check_dtype(op_proto, **kwargs): + """ + This function performs the sanity check for dtype and + instance type. + """ + dtype = None + for ipt in op_proto.inputs: + name = _convert_(ipt.name) + val = kwargs.pop(name, []) + if not isinstance(val, list) and not isinstance(val, tuple): + val = [val] + for each in val: + if not isinstance(each, Variable): + raise ValueError("input of {0} must be variable".format( + op_type)) + + if dtype is None: + dtype = each.dtype + elif dtype != each.dtype: + raise ValueError( + "operator {0} must input same dtype. {1} vs {2}".format( + op_type, dtype, each.dtype)) + + return dtype + + def func(**kwargs): + helper = LayerHelper(op_type, **kwargs) + + dtype = infer_and_check_dtype(op_proto, **kwargs) + + inputs = dict() + for ipt in op_proto.inputs: + name = _convert_(ipt.name) + val = kwargs.pop(name, []) + if not isinstance(val, list) and not isinstance(val, tuple): + val = [val] + inputs[ipt.name] = val + + outputs = dict() + out = helper.create_tmp_variable(dtype=dtype) + outputs[o_name] = [out] + for name in intermediate_output_names: + outputs[name] = [helper.create_tmp_variable(dtype=dtype)] + helper.append_op( + type=op_type, inputs=inputs, outputs=outputs, attrs=kwargs) + return helper.append_activation(out) + + func.__name__ = op_type + func.__doc__ = _generate_doc_string_(op_proto) + return func + + +def deprecated(func_or_class): + """ + Deprecated warning decorator. It will result a warning message. + Should be used before class or function, member function + """ + + @functools.wraps(func) + def func_wrapper(*args, **kwargs): + """ + Wrap func with deprecated warning + """ + warnings.simplefilter('always', DeprecationWarning) # turn off filter + warnings.warn( + "Call to deprecated function {}.".format(func.__name__), + category=DeprecationWarning, + stacklevel=2) + warnings.simplefilter('default', DeprecationWarning) # reset filter + return func(*args, **kwargs) + + return func_wrapper + + +def autodoc(func): + func.__doc__ = _generate_doc_string_(OpProtoHolder.instance().get_op_proto( + func.__name__)) + return func diff --git a/python/paddle/v2/fluid/regularizer.py b/python/paddle/v2/fluid/regularizer.py index 098cd0dd6439554f49e429ab75fb11bfa2c9d28c..d1955b00479676448d99603a31249aa7ac6a0d3f 100644 --- a/python/paddle/v2/fluid/regularizer.py +++ b/python/paddle/v2/fluid/regularizer.py @@ -1,11 +1,9 @@ -import paddle.v2.fluid.framework as framework +import framework -__all__ = [ - 'append_regularization_ops', 'L2DecayRegularizer', 'L1DecayRegularizer' -] +__all__ = ['append_regularization_ops', 'L1Decay', 'L2Decay'] -def append_regularization_ops(parameters_and_grads): +def append_regularization_ops(parameters_and_grads, regularization=None): """Create and add backward regularization Operators Creates and adds backward regularization operators in the BlockDesc. @@ -16,6 +14,8 @@ def append_regularization_ops(parameters_and_grads): Args: parameters_and_grads: A list of (parameters, gradients) pairs that need to be regularized. + regularization: A global regularizer. If the parameter is not + set. It will be applied with regularizer. Returns: list of (parameters, gradients) pair with the regularized gradient @@ -25,14 +25,19 @@ def append_regularization_ops(parameters_and_grads): """ params_and_grads = [] for param, grad in parameters_and_grads: + regularization_term = None + if param.regularizer is not None: + # Add variable for regularization term in grad block + regularization_term = param.regularizer(param, grad.block) + elif regularization is not None: + regularization_term = regularization(param, grad.block) + # If no gradient or no regularization specified, # then we don't need to do anything - if grad is None or param.regularizer is None: + if grad is None or regularization_term is None: params_and_grads.append((param, grad)) continue - # Add variable for regularization term in grad block - regularization_term = param.regularizer(param, grad.block) assert grad.shape == regularization_term.shape grad.block.append_op( @@ -139,3 +144,16 @@ class L1DecayRegularizer(WeightDecayRegularizer): attrs={"scale": self._regularization_coeff}) return decay + + +# We short the class name, since users will use the regulaizer with the package +# name. The sample code: +# +# import paddle.fluid as fluid +# +# hidden = fluid.layers.fc(..., +# param_attr=fluid.regularizer.Xavier()) +# +# It is no need to add a `Regularizer` as the class suffix +L1Decay = L1DecayRegularizer +L2Decay = L2DecayRegularizer diff --git a/python/paddle/v2/fluid/tests/.gitignore b/python/paddle/v2/fluid/tests/.gitignore index fcc52c04886865d96c1bfe1597a9dc99c181de1f..62f82151eb42342cd90657b1e4dfc93410950e62 100644 --- a/python/paddle/v2/fluid/tests/.gitignore +++ b/python/paddle/v2/fluid/tests/.gitignore @@ -1,2 +1,4 @@ image/ fit_a_line.model/ +tmp +cuda_profiler.txt diff --git a/python/paddle/v2/fluid/tests/__init__.py b/python/paddle/v2/fluid/tests/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/python/paddle/v2/fluid/tests/book/CMakeLists.txt b/python/paddle/v2/fluid/tests/book/CMakeLists.txt index 4d7664469e481344cf9eea84688f068b4fb99dee..a35abe3e0c436be4eaed01c9b9183344c6d3b275 100644 --- a/python/paddle/v2/fluid/tests/book/CMakeLists.txt +++ b/python/paddle/v2/fluid/tests/book/CMakeLists.txt @@ -1,5 +1,11 @@ file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py") string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}") + +list(REMOVE_ITEM TEST_OPS test_image_classification_train) +py_test(test_image_classification_train_resnet SRCS test_image_classification_train.py ARGS resnet) +py_test(test_image_classification_train_vgg SRCS test_image_classification_train.py ARGS vgg) + +# default test foreach(src ${TEST_OPS}) py_test(${src} SRCS ${src}.py) endforeach() diff --git a/python/paddle/v2/fluid/tests/book/test_fit_a_line.py b/python/paddle/v2/fluid/tests/book/test_fit_a_line.py index a7f3bfc0caf76302674a00c80c2bd9ebf834f872..fbf46ac6cba8fa4981cc8a6e8f5434a510c52d7d 100644 --- a/python/paddle/v2/fluid/tests/book/test_fit_a_line.py +++ b/python/paddle/v2/fluid/tests/book/test_fit_a_line.py @@ -1,23 +1,18 @@ import numpy as np import paddle.v2 as paddle -import paddle.v2.fluid.core as core -import paddle.v2.fluid.framework as framework -import paddle.v2.fluid.layers as layers -from paddle.v2.fluid.executor import Executor -from paddle.v2.fluid.io import save_persistables, load_persistables -from paddle.v2.fluid.optimizer import SGDOptimizer +import paddle.v2.fluid as fluid -x = layers.data(name='x', shape=[13], data_type='float32') +x = fluid.layers.data(name='x', shape=[13], dtype='float32') -y_predict = layers.fc(input=x, size=1, act=None) +y_predict = fluid.layers.fc(input=x, size=1, act=None) -y = layers.data(name='y', shape=[1], data_type='float32') +y = fluid.layers.data(name='y', shape=[1], dtype='float32') -cost = layers.square_error_cost(input=y_predict, label=y) -avg_cost = layers.mean(x=cost) +cost = fluid.layers.square_error_cost(input=y_predict, label=y) +avg_cost = fluid.layers.mean(x=cost) -sgd_optimizer = SGDOptimizer(learning_rate=0.001) -opts = sgd_optimizer.minimize(avg_cost) +sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.001) +sgd_optimizer.minimize(avg_cost) BATCH_SIZE = 20 @@ -26,32 +21,21 @@ train_reader = paddle.batch( paddle.dataset.uci_housing.train(), buf_size=500), batch_size=BATCH_SIZE) -place = core.CPUPlace() -exe = Executor(place) +place = fluid.CPUPlace() +feeder = fluid.DataFeeder(place=place, feed_list=[x, y]) +exe = fluid.Executor(place) -exe.run(framework.default_startup_program()) +exe.run(fluid.default_startup_program()) PASS_NUM = 100 for pass_id in range(PASS_NUM): - save_persistables(exe, "./fit_a_line.model/") - load_persistables(exe, "./fit_a_line.model/") + fluid.io.save_persistables(exe, "./fit_a_line.model/") + fluid.io.load_persistables(exe, "./fit_a_line.model/") for data in train_reader(): - x_data = np.array(map(lambda x: x[0], data)).astype("float32") - y_data = np.array(map(lambda x: x[1], data)).astype("float32") - - tensor_x = core.LoDTensor() - tensor_x.set(x_data, place) - # print tensor_x.get_dims() - - tensor_y = core.LoDTensor() - tensor_y.set(y_data, place) - # print tensor_y.get_dims() - outs = exe.run(framework.default_main_program(), - feed={'x': tensor_x, - 'y': tensor_y}, - fetch_list=[avg_cost]) - out = np.array(outs[0]) - - if out[0] < 10.0: + avg_loss_value, = exe.run(fluid.default_main_program(), + feed=feeder.feed(data), + fetch_list=[avg_cost]) + + if avg_loss_value[0] < 10.0: exit(0) # if avg cost less than 10.0, we think our code is good. exit(1) diff --git a/python/paddle/v2/fluid/tests/book/test_image_classification_train.py b/python/paddle/v2/fluid/tests/book/test_image_classification_train.py index b8506125501b6e533c4594b37943ec36ca8e7d30..3d336ffe9582ddd9a2031e7aa7e2c26a772820f8 100644 --- a/python/paddle/v2/fluid/tests/book/test_image_classification_train.py +++ b/python/paddle/v2/fluid/tests/book/test_image_classification_train.py @@ -1,17 +1,14 @@ -import numpy as np +from __future__ import print_function + +import sys + import paddle.v2 as paddle -import paddle.v2.fluid.core as core -import paddle.v2.fluid.framework as framework -import paddle.v2.fluid.layers as layers -import paddle.v2.fluid.nets as nets -from paddle.v2.fluid.executor import Executor -from paddle.v2.fluid.initializer import XavierInitializer -from paddle.v2.fluid.optimizer import AdamOptimizer +import paddle.v2.fluid as fluid def resnet_cifar10(input, depth=32): def conv_bn_layer(input, ch_out, filter_size, stride, padding, act='relu'): - tmp = layers.conv2d( + tmp = fluid.layers.conv2d( input=input, filter_size=filter_size, num_filters=ch_out, @@ -19,12 +16,11 @@ def resnet_cifar10(input, depth=32): padding=padding, act=None, bias_attr=False) - return layers.batch_norm(input=tmp, act=act) + return fluid.layers.batch_norm(input=tmp, act=act) - def shortcut(input, ch_in, ch_out, stride, program, init_program): + def shortcut(input, ch_in, ch_out, stride): if ch_in != ch_out: - return conv_bn_layer(input, ch_out, 1, stride, 0, None, program, - init_program) + return conv_bn_layer(input, ch_out, 1, stride, 0, None) else: return input @@ -32,7 +28,7 @@ def resnet_cifar10(input, depth=32): tmp = conv_bn_layer(input, ch_out, 3, stride, 1) tmp = conv_bn_layer(tmp, ch_out, 3, 1, 1, act=None) short = shortcut(input, ch_in, ch_out, stride) - return layers.elementwise_add(x=tmp, y=short, act='relu') + 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) @@ -47,14 +43,14 @@ def resnet_cifar10(input, depth=32): 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 = layers.pool2d( + pool = fluid.layers.pool2d( input=res3, pool_size=8, pool_type='avg', pool_stride=1) return pool def vgg16_bn_drop(input): def conv_block(input, num_filter, groups, dropouts): - return nets.img_conv_group( + return fluid.nets.img_conv_group( input=input, pool_size=2, pool_stride=2, @@ -71,44 +67,42 @@ def vgg16_bn_drop(input): conv4 = conv_block(conv3, 512, 3, [0.4, 0.4, 0]) conv5 = conv_block(conv4, 512, 3, [0.4, 0.4, 0]) - drop = layers.dropout(x=conv5, dropout_prob=0.5) - fc1 = layers.fc(input=drop, - size=512, - act=None, - param_attr={"initializer": XavierInitializer()}) - reshape1 = layers.reshape(x=fc1, shape=list(fc1.shape + (1, 1))) - bn = layers.batch_norm(input=reshape1, act='relu') - drop2 = layers.dropout(x=bn, dropout_prob=0.5) - fc2 = layers.fc(input=drop2, - size=512, - act=None, - param_attr={"initializer": XavierInitializer()}) + 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 classdim = 10 data_shape = [3, 32, 32] -images = layers.data(name='pixel', shape=data_shape, data_type='float32') -label = layers.data(name='label', shape=[1], data_type='int64') +images = fluid.layers.data(name='pixel', shape=data_shape, dtype='float32') +label = fluid.layers.data(name='label', shape=[1], dtype='int64') -# Add neural network config -# option 1. resnet -# net = resnet_cifar10(images, 32) -# option 2. vgg -net = vgg16_bn_drop(images) +net_type = "vgg" +if len(sys.argv) >= 2: + net_type = sys.argv[1] -# print(program) +if net_type == "vgg": + print("train vgg net") + net = vgg16_bn_drop(images) +elif net_type == "resnet": + print("train resnet") + net = resnet_cifar10(images, 32) +else: + raise ValueError("%s network is not supported" % net_type) -predict = layers.fc(input=net, size=classdim, act='softmax') -cost = layers.cross_entropy(input=predict, label=label) -avg_cost = layers.mean(x=cost) -accuracy = layers.accuracy(input=predict, label=label) +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) -# optimizer = SGDOptimizer(learning_rate=0.001) -optimizer = AdamOptimizer(learning_rate=0.001) +optimizer = fluid.optimizer.Adam(learning_rate=0.001) opts = optimizer.minimize(avg_cost) +accuracy = fluid.evaluator.Accuracy(input=predict, label=label) + BATCH_SIZE = 128 PASS_NUM = 1 @@ -117,39 +111,20 @@ train_reader = paddle.batch( paddle.dataset.cifar.train10(), buf_size=128 * 10), batch_size=BATCH_SIZE) -place = core.CPUPlace() -exe = Executor(place) - -exe.run(framework.default_startup_program()) +place = fluid.CPUPlace() +exe = fluid.Executor(place) +feeder = fluid.DataFeeder(place=place, feed_list=[images, label]) +exe.run(fluid.default_startup_program()) for pass_id in range(PASS_NUM): - batch_id = 0 + accuracy.reset(exe) for data in train_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") - batch_size = 1 - for i in y_data.shape: - batch_size = batch_size * i - y_data = y_data.reshape([batch_size, 1]) - - tensor_img = core.LoDTensor() - tensor_y = core.LoDTensor() - tensor_img.set(img_data, place) - tensor_y.set(y_data, place) - - outs = exe.run(framework.default_main_program(), - feed={"pixel": tensor_img, - "label": tensor_y}, - fetch_list=[avg_cost, accuracy]) - - loss = np.array(outs[0]) - acc = np.array(outs[1]) - print("pass_id:" + str(pass_id) + " batch_id:" + str(batch_id) + - " loss:" + str(loss) + " acc:" + str(acc)) - batch_id = batch_id + 1 - - if batch_id > 1: - # this model is slow, so if we can train two mini batch, we think it works properly. - exit(0) + loss, acc = exe.run(fluid.default_main_program(), + feed=feeder.feed(data), + fetch_list=[avg_cost] + accuracy.metrics) + pass_acc = accuracy.eval(exe) + print("loss:" + str(loss) + " acc:" + str(acc) + " pass_acc:" + str( + pass_acc)) + # this model is slow, so if we can train two mini batch, we think it works properly. + exit(0) exit(1) diff --git a/python/paddle/v2/fluid/tests/book/test_label_semantic_roles.py b/python/paddle/v2/fluid/tests/book/test_label_semantic_roles.py new file mode 100644 index 0000000000000000000000000000000000000000..8acd470c5ed5fa8eeda396f1e9182db4ecdd7016 --- /dev/null +++ b/python/paddle/v2/fluid/tests/book/test_label_semantic_roles.py @@ -0,0 +1,202 @@ +import math + +import numpy as np +import paddle.v2 as paddle +import paddle.v2.dataset.conll05 as conll05 +import paddle.v2.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_len = len(verb_dict) + +mark_dict_len = 2 +word_dim = 32 +mark_dim = 5 +hidden_dim = 512 +depth = 8 +mix_hidden_lr = 1e-3 + +IS_SPARSE = True +PASS_NUM = 10 +BATCH_SIZE = 20 + +embedding_name = 'emb' + + +def load_parameter(file_name, h, w): + with open(file_name, 'rb') as f: + f.read(16) # skip header. + return np.fromfile(f, dtype=np.float32).reshape(h, w) + + +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_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, 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) 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), + fluid.layers.fc(input=input_tmp[1], size=hidden_dim) + ]) + + 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), + fluid.layers.fc(input=input_tmp[1], size=label_dict_len) + ]) + + 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 main(): + # define network topology + 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=mix_hidden_lr)) + avg_cost = fluid.layers.mean(x=crf_cost) + + # TODO(qiao) + # check other optimizers and check why out will be NAN + sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.0001) + 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), + batch_size=BATCH_SIZE) + place = fluid.CPUPlace() + feeder = fluid.DataFeeder( + feed_list=[ + word, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2, predicate, mark, target + ], + place=place) + exe = fluid.Executor(place) + + exe.run(fluid.default_startup_program()) + + 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) + + 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( + fluid.default_main_program(), + feed=feeder.feed(data), + fetch_list=[avg_cost] + chunk_evaluator.metrics) + pass_precision, pass_recall, pass_f1_score = chunk_evaluator.eval( + exe) + + 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)) + + # exit early for CI + exit(0) + + batch_id = batch_id + 1 + + +if __name__ == '__main__': + main() diff --git a/python/paddle/v2/fluid/tests/book/test_machine_translation.py b/python/paddle/v2/fluid/tests/book/test_machine_translation.py new file mode 100644 index 0000000000000000000000000000000000000000..e79864b3977ed8111903f9497685ee7ebf76e1da --- /dev/null +++ b/python/paddle/v2/fluid/tests/book/test_machine_translation.py @@ -0,0 +1,119 @@ +import numpy as np +import paddle.v2 as paddle +import paddle.v2.fluid as fluid +import paddle.v2.fluid.core as core +import paddle.v2.fluid.framework as framework +import paddle.v2.fluid.layers as layers +from paddle.v2.fluid.executor import Executor + +dict_size = 30000 +source_dict_dim = target_dict_dim = dict_size +src_dict, trg_dict = paddle.dataset.wmt14.get_dict(dict_size) +hidden_dim = 32 +word_dim = 16 +IS_SPARSE = True +batch_size = 10 +max_length = 50 +topk_size = 50 +trg_dic_size = 10000 + +decoder_size = hidden_dim + + +def encoder_decoder(): + # encoder + src_word_id = layers.data( + name="src_word_id", shape=[1], dtype='int64', lod_level=1) + src_embedding = layers.embedding( + input=src_word_id, + size=[dict_size, word_dim], + dtype='float32', + is_sparse=IS_SPARSE, + param_attr=fluid.ParamAttr(name='vemb')) + + fc1 = fluid.layers.fc(input=src_embedding, size=hidden_dim * 4, act='tanh') + lstm_hidden0, lstm_0 = layers.dynamic_lstm(input=fc1, size=hidden_dim * 4) + encoder_out = layers.sequence_last_step(input=lstm_hidden0) + + # decoder + trg_language_word = layers.data( + name="target_language_word", shape=[1], dtype='int64', lod_level=1) + trg_embedding = layers.embedding( + input=trg_language_word, + size=[dict_size, word_dim], + dtype='float32', + is_sparse=IS_SPARSE, + param_attr=fluid.ParamAttr(name='vemb')) + + rnn = fluid.layers.DynamicRNN() + with rnn.block(): + current_word = rnn.step_input(trg_embedding) + mem = rnn.memory(init=encoder_out) + fc1 = fluid.layers.fc(input=[current_word, mem], + size=decoder_size, + act='tanh') + out = fluid.layers.fc(input=fc1, size=target_dict_dim, act='softmax') + rnn.update_memory(mem, fc1) + rnn.output(out) + + return rnn() + + +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 main(): + rnn_out = encoder_decoder() + label = layers.data( + name="target_language_next_word", shape=[1], dtype='int64', lod_level=1) + cost = layers.cross_entropy(input=rnn_out, label=label) + avg_cost = fluid.layers.mean(x=cost) + + optimizer = fluid.optimizer.Adagrad(learning_rate=1e-4) + optimizer.minimize(avg_cost) + + train_data = paddle.batch( + paddle.reader.shuffle( + paddle.dataset.wmt14.train(dict_size), buf_size=1000), + batch_size=batch_size) + + place = core.CPUPlace() + exe = Executor(place) + + exe.run(framework.default_startup_program()) + + 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={ + 'src_word_id': word_data, + 'target_language_word': trg_word, + 'target_language_next_word': trg_word_next + }, + fetch_list=[avg_cost]) + avg_cost_val = np.array(outs[0]) + print('pass_id=' + str(pass_id) + ' batch=' + str(batch_id) + + " avg_cost=" + str(avg_cost_val)) + if batch_id > 3: + exit(0) + batch_id += 1 + + +if __name__ == '__main__': + main() diff --git a/python/paddle/v2/fluid/tests/book/test_recognize_digits_conv.py b/python/paddle/v2/fluid/tests/book/test_recognize_digits_conv.py index 75fbaf83e8f3e62eb0d0abef9cfa267b65e72973..35bf8da924dc76475df9bd5e6a4c04f4d204426a 100644 --- a/python/paddle/v2/fluid/tests/book/test_recognize_digits_conv.py +++ b/python/paddle/v2/fluid/tests/book/test_recognize_digits_conv.py @@ -1,23 +1,18 @@ +from __future__ import print_function import numpy as np import paddle.v2 as paddle -import paddle.v2.fluid.core as core -import paddle.v2.fluid.evaluator as evaluator -import paddle.v2.fluid.framework as framework -import paddle.v2.fluid.layers as layers -import paddle.v2.fluid.nets as nets -from paddle.v2.fluid.executor import Executor -from paddle.v2.fluid.optimizer import AdamOptimizer +import paddle.v2.fluid as fluid -images = layers.data(name='pixel', shape=[1, 28, 28], data_type='float32') -label = layers.data(name='label', shape=[1], data_type='int64') -conv_pool_1 = nets.simple_img_conv_pool( +images = fluid.layers.data(name='pixel', shape=[1, 28, 28], dtype='float32') +label = fluid.layers.data(name='label', shape=[1], dtype='int64') +conv_pool_1 = fluid.nets.simple_img_conv_pool( input=images, filter_size=5, num_filters=20, pool_size=2, pool_stride=2, act="relu") -conv_pool_2 = nets.simple_img_conv_pool( +conv_pool_2 = fluid.nets.simple_img_conv_pool( input=conv_pool_1, filter_size=5, num_filters=50, @@ -25,13 +20,13 @@ conv_pool_2 = nets.simple_img_conv_pool( pool_stride=2, act="relu") -predict = layers.fc(input=conv_pool_2, size=10, act="softmax") -cost = layers.cross_entropy(input=predict, label=label) -avg_cost = layers.mean(x=cost) -optimizer = AdamOptimizer(learning_rate=0.01, beta1=0.9, beta2=0.999) -opts = optimizer.minimize(avg_cost) +predict = fluid.layers.fc(input=conv_pool_2, size=10, act="softmax") +cost = fluid.layers.cross_entropy(input=predict, label=label) +avg_cost = fluid.layers.mean(x=cost) +optimizer = fluid.optimizer.Adam(learning_rate=0.01) +optimizer.minimize(avg_cost) -accuracy, acc_out = evaluator.accuracy(input=predict, label=label) +accuracy = fluid.evaluator.Accuracy(input=predict, label=label) BATCH_SIZE = 50 PASS_NUM = 3 @@ -40,39 +35,26 @@ train_reader = paddle.batch( paddle.dataset.mnist.train(), buf_size=500), batch_size=BATCH_SIZE) -place = core.CPUPlace() -exe = Executor(place) - -exe.run(framework.default_startup_program()) +place = fluid.CPUPlace() +exe = fluid.Executor(place) +feeder = fluid.DataFeeder(feed_list=[images, label], place=place) +exe.run(fluid.default_startup_program()) for pass_id in range(PASS_NUM): - count = 0 accuracy.reset(exe) for data in train_reader(): - img_data = np.array(map(lambda x: x[0].reshape([1, 28, 28]), - data)).astype("float32") - y_data = np.array(map(lambda x: x[1], data)).astype("int64") - y_data = y_data.reshape([BATCH_SIZE, 1]) - - tensor_img = core.LoDTensor() - tensor_y = core.LoDTensor() - tensor_img.set(img_data, place) - tensor_y.set(y_data, place) - - outs = exe.run(framework.default_main_program(), - feed={"pixel": tensor_img, - "label": tensor_y}, - fetch_list=[avg_cost, acc_out]) - loss = np.array(outs[0]) - acc = np.array(outs[1]) + loss, acc = exe.run(fluid.default_main_program(), + feed=feeder.feed(data), + fetch_list=[avg_cost] + accuracy.metrics) pass_acc = accuracy.eval(exe) - print "pass id : ", pass_id, pass_acc + print("pass_id=" + str(pass_id) + " acc=" + str(acc) + " pass_acc=" + + str(pass_acc)) # print loss, acc - if loss < 10.0 and acc > 0.9: + if loss < 10.0 and pass_acc > 0.9: # if avg cost less than 10.0 and accuracy is larger than 0.9, we think our code is good. exit(0) pass_acc = accuracy.eval(exe) - print "pass id : ", pass_id, pass_acc + print("pass_id=" + str(pass_id) + " pass_acc=" + str(pass_acc)) exit(1) diff --git a/python/paddle/v2/fluid/tests/book/test_recognize_digits_mlp.py b/python/paddle/v2/fluid/tests/book/test_recognize_digits_mlp.py index cf10b1942e6a8243b18b0ae4586fdd7ec1a665fb..fc073f6be8563a363c0f98b9235ae267fa68562d 100644 --- a/python/paddle/v2/fluid/tests/book/test_recognize_digits_mlp.py +++ b/python/paddle/v2/fluid/tests/book/test_recognize_digits_mlp.py @@ -1,69 +1,77 @@ +from __future__ import print_function import numpy as np import paddle.v2 as paddle -import paddle.v2.fluid.core as core -import paddle.v2.fluid.framework as framework -import paddle.v2.fluid.layers as layers -from paddle.v2.fluid.executor import Executor -from paddle.v2.fluid.initializer import UniformInitializer -from paddle.v2.fluid.optimizer import MomentumOptimizer -from paddle.v2.fluid.regularizer import L2DecayRegularizer +import paddle.v2.fluid as fluid BATCH_SIZE = 128 -image = layers.data(name='x', shape=[784], data_type='float32') +image = fluid.layers.data(name='x', shape=[784], dtype='float32') -param_attr = { - 'name': None, - 'initializer': UniformInitializer( - low=-1.0, high=1.0), - 'regularization': L2DecayRegularizer(0.0005 * BATCH_SIZE) -} +regularizer = fluid.regularizer.L2Decay(0.0005 * BATCH_SIZE) -hidden1 = layers.fc(input=image, size=128, act='relu', param_attr=param_attr) -hidden2 = layers.fc(input=hidden1, size=64, act='relu', param_attr=param_attr) +hidden1 = fluid.layers.fc(input=image, + size=128, + act='relu', + param_attr=fluid.ParamAttr( + regularizer=regularizer, + clip=fluid.clip.ClipByValue(10))) +hidden2 = fluid.layers.fc(input=hidden1, + size=64, + act='relu', + param_attr=regularizer) -predict = layers.fc(input=hidden2, - size=10, - act='softmax', - param_attr=param_attr) +predict = fluid.layers.fc(input=hidden2, + size=10, + act='softmax', + param_attr=regularizer) -label = layers.data(name='y', shape=[1], data_type='int64') +label = fluid.layers.data(name='y', shape=[1], dtype='int64') -cost = layers.cross_entropy(input=predict, label=label) -avg_cost = layers.mean(x=cost) -accuracy = layers.accuracy(input=predict, label=label) +cost = fluid.layers.cross_entropy(input=predict, label=label) +avg_cost = fluid.layers.mean(x=cost) -optimizer = MomentumOptimizer(learning_rate=0.001, momentum=0.9) +optimizer = fluid.optimizer.Momentum(learning_rate=0.001, momentum=0.9) opts = optimizer.minimize(avg_cost) +accuracy = fluid.evaluator.Accuracy(input=predict, label=label) + +inference_program = fluid.default_main_program().clone() +with fluid.program_guard(inference_program): + test_accuracy = fluid.evaluator.Accuracy(input=predict, label=label) + test_target = [avg_cost] + test_accuracy.metrics + test_accuracy.states + inference_program = fluid.io.get_inference_program(test_target) + train_reader = paddle.batch( paddle.reader.shuffle( paddle.dataset.mnist.train(), buf_size=8192), batch_size=BATCH_SIZE) -place = core.CPUPlace() -exe = Executor(place) +test_reader = paddle.batch(paddle.dataset.mnist.test(), batch_size=128) -exe.run(framework.default_startup_program()) +place = fluid.CPUPlace() +exe = fluid.Executor(place) +feeder = fluid.DataFeeder(feed_list=[image, label], place=place) +exe.run(fluid.default_startup_program()) PASS_NUM = 100 for pass_id in range(PASS_NUM): + accuracy.reset(exe) for data in train_reader(): - x_data = np.array(map(lambda x: x[0], data)).astype("float32") - y_data = np.array(map(lambda x: x[1], data)).astype("int64") - y_data = np.expand_dims(y_data, axis=1) - - tensor_x = core.LoDTensor() - tensor_x.set(x_data, place) - - tensor_y = core.LoDTensor() - tensor_y.set(y_data, place) - - outs = exe.run(framework.default_main_program(), - feed={'x': tensor_x, - 'y': tensor_y}, - fetch_list=[avg_cost, accuracy]) - out = np.array(outs[0]) - acc = np.array(outs[1]) - if out[0] < 5.0: - exit(0) # if avg cost less than 5.0, we think our code is good. + out, acc = exe.run(fluid.default_main_program(), + feed=feeder.feed(data), + fetch_list=[avg_cost] + accuracy.metrics) + pass_acc = accuracy.eval(exe) + + test_accuracy.reset(exe) + for data in test_reader(): + out, acc = exe.run(inference_program, + feed=feeder.feed(data), + fetch_list=[avg_cost] + test_accuracy.metrics) + + test_pass_acc = test_accuracy.eval(exe) + print("pass_id=" + str(pass_id) + " train_cost=" + str( + out) + " train_acc=" + str(acc) + " train_pass_acc=" + str(pass_acc) + + " test_acc=" + str(test_pass_acc)) + + if test_pass_acc > 0.7: + exit(0) exit(1) diff --git a/python/paddle/v2/fluid/tests/book/test_recommender_system.py b/python/paddle/v2/fluid/tests/book/test_recommender_system.py index 55ded3aed3a23c8cd7795f915dc1cbd512c6d945..e3cc2a89371233014dec4ba3d730a866722d3eae 100644 --- a/python/paddle/v2/fluid/tests/book/test_recommender_system.py +++ b/python/paddle/v2/fluid/tests/book/test_recommender_system.py @@ -18,47 +18,47 @@ def get_usr_combined_features(): USR_DICT_SIZE = paddle.dataset.movielens.max_user_id() + 1 - uid = layers.data(name='user_id', shape=[1], data_type='int64') + uid = layers.data(name='user_id', shape=[1], dtype='int64') usr_emb = layers.embedding( input=uid, - data_type='float32', + dtype='float32', size=[USR_DICT_SIZE, 32], - param_attr={'name': 'user_table'}, + 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], data_type='int64') + 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={'name': 'gender_table'}, + 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], data_type="int64") + 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={'name': 'age_table'}) + 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], data_type="int64") + 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={'name': 'job_table'}, + param_attr='job_table', is_sparse=IS_SPARSE) usr_job_fc = layers.fc(input=usr_job_emb, size=16) @@ -75,20 +75,20 @@ def get_mov_combined_features(): MOV_DICT_SIZE = paddle.dataset.movielens.max_movie_id() + 1 - mov_id = layers.data(name='movie_id', shape=[1], data_type='int64') + mov_id = layers.data(name='movie_id', shape=[1], dtype='int64') mov_emb = layers.embedding( input=mov_id, - data_type='float32', + dtype='float32', size=[MOV_DICT_SIZE, 32], - param_attr={'name': 'movie_table'}, + 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], data_type='int64') + category_id = layers.data(name='category_id', shape=[1], dtype='int64') mov_categories_emb = layers.embedding( input=category_id, size=[CATEGORY_DICT_SIZE, 32], is_sparse=IS_SPARSE) @@ -98,7 +98,7 @@ def get_mov_combined_features(): MOV_TITLE_DICT_SIZE = len(paddle.dataset.movielens.get_movie_title_dict()) - mov_title_id = layers.data(name='movie_title', shape=[1], data_type='int64') + mov_title_id = layers.data(name='movie_title', shape=[1], dtype='int64') mov_title_emb = layers.embedding( input=mov_title_id, size=[MOV_TITLE_DICT_SIZE, 32], is_sparse=IS_SPARSE) @@ -125,10 +125,11 @@ def model(): # need cos sim inference = layers.cos_sim(X=usr_combined_features, Y=mov_combined_features) + scale_infer = layers.scale(x=inference, scale=5.0) - label = layers.data(name='score', shape=[1], data_type='float32') + label = layers.data(name='score', shape=[1], dtype='float32') - square_cost = layers.square_error_cost(input=inference, label=label) + square_cost = layers.square_error_cost(input=scale_infer, label=label) avg_cost = layers.mean(x=square_cost) @@ -141,7 +142,7 @@ def main(): opts = sgd_optimizer.minimize(cost) if USE_GPU: - place = core.GPUPlace(0) + place = core.CUDAPlace(0) else: place = core.CPUPlace() diff --git a/python/paddle/v2/fluid/tests/book/test_understand_sentiment_conv.py b/python/paddle/v2/fluid/tests/book/test_understand_sentiment_conv.py index e69b915a9cfaf9e06075991975563a1fc1196661..f103358edca9bbd2e28c99afd249f97b1d8069ae 100644 --- a/python/paddle/v2/fluid/tests/book/test_understand_sentiment_conv.py +++ b/python/paddle/v2/fluid/tests/book/test_understand_sentiment_conv.py @@ -1,39 +1,33 @@ +from __future__ import print_function import numpy as np import paddle.v2 as paddle -import paddle.v2.fluid.core as core -import paddle.v2.fluid.framework as framework -import paddle.v2.fluid.layers as layers -import paddle.v2.fluid.nets as nets -from paddle.v2.fluid.executor import Executor -from paddle.v2.fluid.optimizer import AdamOptimizer +import paddle.v2.fluid as fluid -def convolution_net(input_dim, class_dim=2, emb_dim=32, hid_dim=32): - data = layers.data(name="words", shape=[1], data_type="int64") - label = layers.data(name="label", shape=[1], data_type="int64") - - emb = layers.embedding(input=data, size=[input_dim, emb_dim]) - conv_3 = nets.sequence_conv_pool( +def convolution_net(data, label, input_dim, class_dim=2, emb_dim=32, + hid_dim=32): + emb = fluid.layers.embedding(input=data, size=[input_dim, emb_dim]) + conv_3 = fluid.nets.sequence_conv_pool( input=emb, num_filters=hid_dim, filter_size=3, act="tanh", pool_type="sqrt") - conv_4 = nets.sequence_conv_pool( + conv_4 = fluid.nets.sequence_conv_pool( input=emb, num_filters=hid_dim, filter_size=4, act="tanh", pool_type="sqrt") - prediction = layers.fc(input=[conv_3, conv_4], - size=class_dim, - act="softmax") - cost = layers.cross_entropy(input=prediction, label=label) - avg_cost = layers.mean(x=cost) - adam_optimizer = AdamOptimizer(learning_rate=0.002) - opts = adam_optimizer.minimize(avg_cost) - acc = layers.accuracy(input=prediction, label=label) - return avg_cost, acc + prediction = fluid.layers.fc(input=[conv_3, conv_4], + size=class_dim, + act="softmax") + cost = fluid.layers.cross_entropy(input=prediction, label=label) + avg_cost = fluid.layers.mean(x=cost) + adam_optimizer = fluid.optimizer.Adam(learning_rate=0.002) + adam_optimizer.minimize(avg_cost) + accuracy = fluid.evaluator.Accuracy(input=prediction, label=label) + return avg_cost, accuracy, accuracy.metrics[0] def to_lodtensor(data, place): @@ -45,7 +39,7 @@ def to_lodtensor(data, place): 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 = fluid.LoDTensor() res.set(flattened_data, place) res.set_lod([lod]) return res @@ -59,36 +53,32 @@ def main(): dict_dim = len(word_dict) class_dim = 2 - cost, acc = convolution_net(input_dim=dict_dim, class_dim=class_dim) + data = fluid.layers.data( + name="words", shape=[1], dtype="int64", lod_level=1) + label = fluid.layers.data(name="label", shape=[1], dtype="int64") + cost, accuracy, acc_out = convolution_net( + data, label, input_dim=dict_dim, class_dim=class_dim) train_data = paddle.batch( paddle.reader.shuffle( paddle.dataset.imdb.train(word_dict), buf_size=1000), batch_size=BATCH_SIZE) - place = core.CPUPlace() - exe = Executor(place) + place = fluid.CPUPlace() + exe = fluid.Executor(place) + feeder = fluid.DataFeeder(feed_list=[data, label], place=place) - exe.run(framework.default_startup_program()) + exe.run(fluid.default_startup_program()) for pass_id in xrange(PASS_NUM): + accuracy.reset(exe) for data in train_data(): - tensor_words = to_lodtensor(map(lambda x: x[0], data), place) - - label = np.array(map(lambda x: x[1], data)).astype("int64") - label = label.reshape([BATCH_SIZE, 1]) - - tensor_label = core.LoDTensor() - tensor_label.set(label, place) - - outs = exe.run(framework.default_main_program(), - feed={"words": tensor_words, - "label": tensor_label}, - fetch_list=[cost, acc]) - cost_val = np.array(outs[0]) - acc_val = np.array(outs[1]) - - print("cost=" + str(cost_val) + " acc=" + str(acc_val)) - if cost_val < 1.0 and acc_val > 0.7: + cost_val, acc_val = exe.run(fluid.default_main_program(), + feed=feeder.feed(data), + fetch_list=[cost, acc_out]) + pass_acc = accuracy.eval(exe) + print("cost=" + str(cost_val) + " acc=" + str(acc_val) + + " pass_acc=" + str(pass_acc)) + if cost_val < 1.0 and pass_acc > 0.8: exit(0) exit(1) diff --git a/python/paddle/v2/fluid/tests/book/test_understand_sentiment_dynamic_lstm.py b/python/paddle/v2/fluid/tests/book/test_understand_sentiment_dynamic_lstm.py index 65d44542501e6531fc1912cbc726a1d903b9c031..cd28f04b8574778316d70e7d8a03026f807c3e52 100644 --- a/python/paddle/v2/fluid/tests/book/test_understand_sentiment_dynamic_lstm.py +++ b/python/paddle/v2/fluid/tests/book/test_understand_sentiment_dynamic_lstm.py @@ -1,48 +1,44 @@ import numpy as np import paddle.v2 as paddle -import paddle.v2.fluid.core as core -import paddle.v2.fluid.framework as framework -import paddle.v2.fluid.layers as layers -from paddle.v2.fluid.executor import Executor -from paddle.v2.fluid.optimizer import AdamOptimizer +import paddle.v2.fluid as fluid -def stacked_lstm_net(input_dim, +def stacked_lstm_net(data, + label, + input_dim, class_dim=2, emb_dim=128, hid_dim=512, stacked_num=3): assert stacked_num % 2 == 1 - data = layers.data(name="words", shape=[1], data_type="int64") - label = layers.data(name="label", shape=[1], data_type="int64") - emb = layers.embedding(input=data, size=[input_dim, emb_dim]) + emb = fluid.layers.embedding(input=data, size=[input_dim, emb_dim]) # add bias attr # TODO(qijun) linear act - fc1 = layers.fc(input=emb, size=hid_dim) - lstm1, cell1 = layers.dynamic_lstm(input=fc1, size=hid_dim) + 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 = layers.fc(input=inputs, size=hid_dim) - lstm, cell = layers.dynamic_lstm( + 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 = layers.sequence_pool(input=inputs[0], pool_type='max') - lstm_last = layers.sequence_pool(input=inputs[1], pool_type='max') + 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 = layers.fc(input=[fc_last, lstm_last], - size=class_dim, - act='softmax') - cost = layers.cross_entropy(input=prediction, label=label) - avg_cost = layers.mean(x=cost) - adam_optimizer = AdamOptimizer(learning_rate=0.002) - opts = adam_optimizer.minimize(avg_cost) - acc = layers.accuracy(input=prediction, label=label) - return avg_cost, acc + prediction = fluid.layers.fc(input=[fc_last, lstm_last], + size=class_dim, + act='softmax') + cost = fluid.layers.cross_entropy(input=prediction, label=label) + avg_cost = fluid.layers.mean(x=cost) + adam_optimizer = fluid.optimizer.Adam(learning_rate=0.002) + adam_optimizer.minimize(avg_cost) + accuracy = fluid.evaluator.Accuracy(input=prediction, label=label) + return avg_cost, accuracy, accuracy.metrics[0] def to_lodtensor(data, place): @@ -54,7 +50,7 @@ def to_lodtensor(data, place): 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 = fluid.LoDTensor() res.set(flattened_data, place) res.set_lod([lod]) return res @@ -69,36 +65,32 @@ def main(): dict_dim = len(word_dict) class_dim = 2 - cost, acc = stacked_lstm_net(input_dim=dict_dim, class_dim=class_dim) + data = fluid.layers.data( + name="words", shape=[1], dtype="int64", lod_level=1) + label = fluid.layers.data(name="label", shape=[1], dtype="int64") + cost, accuracy, acc_out = stacked_lstm_net( + data, label, input_dim=dict_dim, class_dim=class_dim) train_data = paddle.batch( paddle.reader.shuffle( paddle.dataset.imdb.train(word_dict), buf_size=1000), batch_size=BATCH_SIZE) - place = core.CPUPlace() - exe = Executor(place) + place = fluid.CPUPlace() + exe = fluid.Executor(place) + feeder = fluid.DataFeeder(feed_list=[data, label], place=place) - exe.run(framework.default_startup_program()) + exe.run(fluid.default_startup_program()) for pass_id in xrange(PASS_NUM): + accuracy.reset(exe) for data in train_data(): - tensor_words = to_lodtensor(map(lambda x: x[0], data), place) - - label = np.array(map(lambda x: x[1], data)).astype("int64") - label = label.reshape([BATCH_SIZE, 1]) - - tensor_label = core.LoDTensor() - tensor_label.set(label, place) - - outs = exe.run(framework.default_main_program(), - feed={"words": tensor_words, - "label": tensor_label}, - fetch_list=[cost, acc]) - cost_val = np.array(outs[0]) - acc_val = np.array(outs[1]) - - print("cost=" + str(cost_val) + " acc=" + str(acc_val)) - if cost_val < 1.0 and acc_val > 0.7: + cost_val, acc_val = exe.run(fluid.default_main_program(), + feed=feeder.feed(data), + fetch_list=[cost, acc_out]) + pass_acc = accuracy.eval(exe) + print("cost=" + str(cost_val) + " acc=" + str(acc_val) + + " pass_acc=" + str(pass_acc)) + if cost_val < 1.0 and acc_val > 0.8: exit(0) exit(1) diff --git a/python/paddle/v2/fluid/tests/book/test_understand_sentiment_lstm.py b/python/paddle/v2/fluid/tests/book/test_understand_sentiment_lstm.py index 280f6e902c34512735a27586221c2be68963ef2b..633de66bea2af7404ab0d325b425e7b9e63d3e43 100644 --- a/python/paddle/v2/fluid/tests/book/test_understand_sentiment_lstm.py +++ b/python/paddle/v2/fluid/tests/book/test_understand_sentiment_lstm.py @@ -1,40 +1,73 @@ import numpy as np import paddle.v2 as paddle -import paddle.v2.fluid.core as core -import paddle.v2.fluid.framework as framework -import paddle.v2.fluid.layers as layers -from paddle.v2.fluid.executor import Executor -from paddle.v2.fluid.optimizer import AdamOptimizer +import paddle.v2.fluid as fluid +from paddle.v2.fluid.layer_helper import LayerHelper + + +def lstm(x, c_pre_init, hidden_dim, forget_bias=None): + """ + This function helps create an operator for the LSTM (Long Short Term + Memory) cell that can be used inside an RNN. + """ + helper = LayerHelper('lstm_unit', **locals()) + rnn = fluid.layers.StaticRNN() + with rnn.step(): + c_pre = rnn.memory(init=c_pre_init) + x_t = rnn.step_input(x) + + before_fc = fluid.layers.concat(input=[x_t, c_pre], axis=1) + after_fc = fluid.layers.fc(input=before_fc, size=hidden_dim * 4) + + dtype = x.dtype + c = helper.create_tmp_variable(dtype) + h = helper.create_tmp_variable(dtype) + + helper.append_op( + type='lstm_unit', + inputs={"X": after_fc, + "C_prev": c_pre}, + outputs={"C": c, + "H": h}, + attrs={"forget_bias": forget_bias}) + + rnn.update_memory(c_pre, c) + rnn.output(h) + + return rnn() def lstm_net(dict_dim, class_dim=2, emb_dim=32, seq_len=80, batch_size=50): - data = layers.data( + data = fluid.layers.data( name="words", shape=[seq_len * batch_size, 1], append_batch_size=False, - data_type="int64") - label = layers.data( + dtype="int64", + lod_level=1) + label = fluid.layers.data( name="label", shape=[batch_size, 1], append_batch_size=False, - data_type="int64") + dtype="int64") - emb = layers.embedding(input=data, size=[dict_dim, emb_dim]) - emb = layers.reshape(x=emb, shape=[batch_size, seq_len, emb_dim]) - emb = layers.transpose(x=emb, axis=[1, 0, 2]) + emb = fluid.layers.embedding(input=data, size=[dict_dim, emb_dim]) + emb = fluid.layers.reshape(x=emb, shape=[batch_size, seq_len, emb_dim]) + emb = fluid.layers.transpose(x=emb, axis=[1, 0, 2]) - c_pre_init = layers.fill_constant( - dtype=emb.data_type, shape=[batch_size, emb_dim], value=0.0) - layer_1_out = layers.lstm(emb, c_pre_init=c_pre_init, hidden_dim=emb_dim) - layer_1_out = layers.transpose(x=layer_1_out, axis=[1, 0, 2]) + c_pre_init = fluid.layers.fill_constant( + dtype=emb.dtype, shape=[batch_size, emb_dim], value=0.0) + c_pre_init.stop_gradient = False + layer_1_out = lstm(emb, c_pre_init=c_pre_init, hidden_dim=emb_dim) + layer_1_out = fluid.layers.transpose(x=layer_1_out, axis=[1, 0, 2]) - prediction = layers.fc(input=layer_1_out, size=class_dim, act="softmax") - cost = layers.cross_entropy(input=prediction, label=label) + prediction = fluid.layers.fc(input=layer_1_out, + size=class_dim, + act="softmax") + cost = fluid.layers.cross_entropy(input=prediction, label=label) - avg_cost = layers.mean(x=cost) - adam_optimizer = AdamOptimizer(learning_rate=0.002) - opts = adam_optimizer.minimize(avg_cost) - acc = layers.accuracy(input=prediction, label=label) + avg_cost = fluid.layers.mean(x=cost) + adam_optimizer = fluid.optimizer.Adam(learning_rate=0.002) + adam_optimizer.minimize(avg_cost) + acc = fluid.layers.accuracy(input=prediction, label=label) return avg_cost, acc @@ -48,57 +81,65 @@ def to_lodtensor(data, place): 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 = fluid.LoDTensor() res.set(flattened_data, place) res.set_lod([lod]) return res -def chop_data(data, chop_len=80, batch_len=50): +def chop_data(data, chop_len=80, batch_size=50): data = [(x[0][:chop_len], x[1]) for x in data if len(x[0]) >= chop_len] - return data[:batch_len] + return data[:batch_size] def prepare_feed_data(data, place): tensor_words = to_lodtensor(map(lambda x: x[0], data), place) label = np.array(map(lambda x: x[1], data)).astype("int64") - label = label.reshape([50, 1]) - tensor_label = core.LoDTensor() + label = label.reshape([len(label), 1]) + tensor_label = fluid.LoDTensor() tensor_label.set(label, place) return tensor_words, tensor_label def main(): + BATCH_SIZE = 100 + PASS_NUM = 5 + word_dict = paddle.dataset.imdb.word_dict() - cost, acc = lstm_net(dict_dim=len(word_dict), class_dim=2) + print "load word dict successfully" + dict_dim = len(word_dict) + class_dim = 2 + + cost, acc = lstm_net(dict_dim=dict_dim, class_dim=class_dim) - batch_size = 100 train_data = paddle.batch( - paddle.reader.buffered( - paddle.dataset.imdb.train(word_dict), size=batch_size * 10), - batch_size=batch_size) - - data = chop_data(next(train_data())) - - place = core.CPUPlace() - tensor_words, tensor_label = prepare_feed_data(data, place) - exe = Executor(place) - exe.run(framework.default_startup_program()) - - while True: - outs = exe.run(framework.default_main_program(), - feed={"words": tensor_words, - "label": tensor_label}, - fetch_list=[cost, acc]) - cost_val = np.array(outs[0]) - acc_val = np.array(outs[1]) - - print("cost=" + str(cost_val) + " acc=" + str(acc_val)) - if acc_val > 0.9: - break + paddle.reader.shuffle( + paddle.dataset.imdb.train(word_dict), buf_size=BATCH_SIZE * 10), + batch_size=BATCH_SIZE) + place = fluid.CPUPlace() + exe = fluid.Executor(place) + + exe.run(fluid.default_startup_program()) + + for pass_id in xrange(PASS_NUM): + for data in train_data(): + chopped_data = chop_data(data) + tensor_words, tensor_label = prepare_feed_data(chopped_data, place) + + outs = exe.run(fluid.default_main_program(), + feed={"words": tensor_words, + "label": tensor_label}, + fetch_list=[cost, acc]) + cost_val = np.array(outs[0]) + acc_val = np.array(outs[1]) + + print("cost=" + str(cost_val) + " acc=" + str(acc_val)) + if acc_val > 0.7: + exit(0) + exit(1) if __name__ == '__main__': diff --git a/python/paddle/v2/fluid/tests/book/test_word2vec.py b/python/paddle/v2/fluid/tests/book/test_word2vec.py index afa7b285198e0349317e123e4bd98e8336217afa..8b928ff9eed41f8945c749058b4177fd023452ba 100644 --- a/python/paddle/v2/fluid/tests/book/test_word2vec.py +++ b/python/paddle/v2/fluid/tests/book/test_word2vec.py @@ -1,10 +1,6 @@ import numpy as np import paddle.v2 as paddle -import paddle.v2.fluid.core as core -import paddle.v2.fluid.framework as framework -import paddle.v2.fluid.layers as layers -from paddle.v2.fluid.executor import Executor -from paddle.v2.fluid.optimizer import SGDOptimizer +import paddle.v2.fluid as fluid PASS_NUM = 100 EMBED_SIZE = 32 @@ -16,94 +12,62 @@ IS_SPARSE = True word_dict = paddle.dataset.imikolov.build_dict() dict_size = len(word_dict) -first_word = layers.data(name='firstw', shape=[1], data_type='int64') -second_word = layers.data(name='secondw', shape=[1], data_type='int64') -third_word = layers.data(name='thirdw', shape=[1], data_type='int64') -forth_word = layers.data(name='forthw', shape=[1], data_type='int64') -next_word = layers.data(name='nextw', shape=[1], data_type='int64') +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') +next_word = fluid.layers.data(name='nextw', shape=[1], dtype='int64') -embed_first = layers.embedding( +embed_first = fluid.layers.embedding( input=first_word, size=[dict_size, EMBED_SIZE], - data_type='float32', + dtype='float32', is_sparse=IS_SPARSE, - param_attr={'name': 'shared_w'}) -embed_second = layers.embedding( + param_attr='shared_w') +embed_second = fluid.layers.embedding( input=second_word, size=[dict_size, EMBED_SIZE], - data_type='float32', + dtype='float32', is_sparse=IS_SPARSE, - param_attr={'name': 'shared_w'}) -embed_third = layers.embedding( + param_attr='shared_w') +embed_third = fluid.layers.embedding( input=third_word, size=[dict_size, EMBED_SIZE], - data_type='float32', + dtype='float32', is_sparse=IS_SPARSE, - param_attr={'name': 'shared_w'}) -embed_forth = layers.embedding( + param_attr='shared_w') +embed_forth = fluid.layers.embedding( input=forth_word, size=[dict_size, EMBED_SIZE], - data_type='float32', + dtype='float32', is_sparse=IS_SPARSE, - param_attr={'name': 'shared_w'}) + param_attr='shared_w') -concat_embed = layers.concat( +concat_embed = fluid.layers.concat( input=[embed_first, embed_second, embed_third, embed_forth], axis=1) -hidden1 = layers.fc(input=concat_embed, size=HIDDEN_SIZE, act='sigmoid') -predict_word = layers.fc(input=hidden1, size=dict_size, act='softmax') -cost = layers.cross_entropy(input=predict_word, label=next_word) -avg_cost = layers.mean(x=cost) -sgd_optimizer = SGDOptimizer(learning_rate=0.001) -opts = sgd_optimizer.minimize(avg_cost) +hidden1 = fluid.layers.fc(input=concat_embed, size=HIDDEN_SIZE, act='sigmoid') +predict_word = fluid.layers.fc(input=hidden1, size=dict_size, act='softmax') +cost = fluid.layers.cross_entropy(input=predict_word, label=next_word) +avg_cost = fluid.layers.mean(x=cost) +sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.001) +sgd_optimizer.minimize(avg_cost) train_reader = paddle.batch( paddle.dataset.imikolov.train(word_dict, N), BATCH_SIZE) -place = core.CPUPlace() -exe = Executor(place) +place = fluid.CPUPlace() +exe = fluid.Executor(place) +feeder = fluid.DataFeeder( + feed_list=[first_word, second_word, third_word, forth_word, next_word], + place=place) -# fix https://github.com/PaddlePaddle/Paddle/issues/5434 then remove -# below exit line. -exit(0) - -exe.run(framework.default_startup_program()) +exe.run(fluid.default_startup_program()) for pass_id in range(PASS_NUM): for data in train_reader(): - input_data = [[data_idx[idx] for data_idx in data] for idx in xrange(5)] - input_data = map(lambda x: np.array(x).astype("int64"), input_data) - input_data = map(lambda x: np.expand_dims(x, axis=1), input_data) - - first_data = input_data[0] - first_tensor = core.LoDTensor() - first_tensor.set(first_data, place) - - second_data = input_data[1] - second_tensor = core.LoDTensor() - second_tensor.set(second_data, place) - - third_data = input_data[2] - third_tensor = core.LoDTensor() - third_tensor.set(third_data, place) - - forth_data = input_data[3] - forth_tensor = core.LoDTensor() - forth_tensor.set(forth_data, place) - - next_data = input_data[4] - next_tensor = core.LoDTensor() - next_tensor.set(next_data, place) - - outs = exe.run(framework.default_main_program(), - feed={ - 'firstw': first_tensor, - 'secondw': second_tensor, - 'thirdw': third_tensor, - 'forthw': forth_tensor, - 'nextw': next_tensor - }, - fetch_list=[avg_cost]) - out = np.array(outs[0]) - if out[0] < 10.0: + avg_cost_np = exe.run(fluid.default_main_program(), + feed=feeder.feed(data), + fetch_list=[avg_cost]) + if avg_cost_np[0] < 5.0: exit(0) # if avg cost less than 10.0, we think our code is good. exit(1) diff --git a/python/paddle/v2/fluid/tests/book_distribute/notest_recognize_digits_conv_dist.py b/python/paddle/v2/fluid/tests/book_distribute/notest_recognize_digits_conv_dist.py new file mode 100644 index 0000000000000000000000000000000000000000..20b4a8b34cd085ae51e6169f0d4eac58b7f3ffb2 --- /dev/null +++ b/python/paddle/v2/fluid/tests/book_distribute/notest_recognize_digits_conv_dist.py @@ -0,0 +1,80 @@ +from __future__ import print_function +import numpy as np +import paddle.v2 as paddle +import paddle.v2.fluid as fluid +import os + +images = fluid.layers.data(name='pixel', shape=[1, 28, 28], dtype='float32') +label = fluid.layers.data(name='label', shape=[1], dtype='int64') +conv_pool_1 = fluid.nets.simple_img_conv_pool( + input=images, + 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") + +predict = fluid.layers.fc(input=conv_pool_2, size=10, act="softmax") +cost = fluid.layers.cross_entropy(input=predict, label=label) +avg_cost = fluid.layers.mean(x=cost) +optimizer = fluid.optimizer.Adam(learning_rate=0.01) +optimize_ops, params_grads = optimizer.minimize(avg_cost) + +accuracy = fluid.evaluator.Accuracy(input=predict, label=label) + +BATCH_SIZE = 50 +PASS_NUM = 3 +train_reader = paddle.batch( + paddle.reader.shuffle( + paddle.dataset.mnist.train(), buf_size=500), + batch_size=BATCH_SIZE) + +place = fluid.CPUPlace() +exe = fluid.Executor(place) + +t = fluid.DistributeTranspiler() +# all parameter server endpoints list for spliting parameters +pserver_endpoints = os.getenv("PSERVERS") +# server endpoint for current node +current_endpoint = os.getenv("SERVER_ENDPOINT") +# run as trainer or parameter server +training_role = os.getenv("TRAINING_ROLE", + "TRAINER") # get the training role: trainer/pserver +t.transpile(optimize_ops, params_grads, pservers=pserver_endpoints, trainers=2) + +if training_role == "PSERVER": + if not current_endpoint: + print("need env SERVER_ENDPOINT") + exit(1) + pserver_prog = t.get_pserver_program(current_endpoint, optimize_ops) + exe.run(fluid.default_startup_program()) + exe.run(pserver_prog) +elif training_role == "TRAINER": + trainer_prog = t.get_trainer_program() + feeder = fluid.DataFeeder(feed_list=[images, label], place=place) + exe.run(fluid.default_startup_program()) + + for pass_id in range(PASS_NUM): + accuracy.reset(exe) + batch_id = 0 + for data in train_reader(): + loss, acc = exe.run(trainer_prog, + feed=feeder.feed(data), + fetch_list=[avg_cost] + accuracy.metrics) + pass_acc = accuracy.eval(exe) + if batch_id % 100 == 0: + print("batch_id %d, loss: %f, acc: %f" % + (batch_id, loss, pass_acc)) + batch_id += 1 + + pass_acc = accuracy.eval(exe) + print("pass_id=" + str(pass_id) + " pass_acc=" + str(pass_acc)) +else: + print("environment var TRAINER_ROLE should be TRAINER os PSERVER") diff --git a/python/paddle/v2/fluid/tests/decorators.py b/python/paddle/v2/fluid/tests/decorators.py new file mode 100644 index 0000000000000000000000000000000000000000..154619b0e93455922700a12d734967b4d20c4f13 --- /dev/null +++ b/python/paddle/v2/fluid/tests/decorators.py @@ -0,0 +1,29 @@ +import paddle.v2.fluid as fluid + +__all__ = ['many_times', 'prog_scope'] + + +def many_times(times): + def __impl__(fn): + def __fn__(*args, **kwargs): + for _ in range(times): + fn(*args, **kwargs) + + return __fn__ + + return __impl__ + + +def prog_scope(): + def __impl__(fn): + def __fn__(*args, **kwargs): + prog = fluid.Program() + startup_prog = fluid.Program() + scope = fluid.core.Scope() + with fluid.scope_guard(scope): + with fluid.program_guard(prog, startup_prog): + fn(*args, **kwargs) + + return __fn__ + + return __impl__ diff --git a/python/paddle/v2/fluid/tests/demo/fc_gan.py b/python/paddle/v2/fluid/tests/demo/fc_gan.py new file mode 100644 index 0000000000000000000000000000000000000000..cae959593e855f11c04585341d86478b649d17c9 --- /dev/null +++ b/python/paddle/v2/fluid/tests/demo/fc_gan.py @@ -0,0 +1,157 @@ +import errno +import math +import os + +import matplotlib +import numpy + +import paddle.v2 as paddle +import paddle.v2.fluid as fluid + +matplotlib.use('Agg') +import matplotlib.pyplot as plt +import matplotlib.gridspec as gridspec + +NOISE_SIZE = 100 +NUM_PASS = 1000 +NUM_REAL_IMGS_IN_BATCH = 121 +NUM_TRAIN_TIMES_OF_DG = 3 +LEARNING_RATE = 2e-5 + + +def D(x): + hidden = fluid.layers.fc(input=x, + size=200, + act='relu', + param_attr='D.w1', + bias_attr='D.b1') + logits = fluid.layers.fc(input=hidden, + size=1, + act=None, + param_attr='D.w2', + bias_attr='D.b2') + return logits + + +def G(x): + hidden = fluid.layers.fc(input=x, + size=200, + act='relu', + param_attr='G.w1', + bias_attr='G.b1') + img = fluid.layers.fc(input=hidden, + size=28 * 28, + act='tanh', + param_attr='G.w2', + bias_attr='G.b2') + return img + + +def plot(gen_data): + gen_data.resize(gen_data.shape[0], 28, 28) + n = int(math.ceil(math.sqrt(gen_data.shape[0]))) + fig = plt.figure(figsize=(n, n)) + gs = gridspec.GridSpec(n, n) + gs.update(wspace=0.05, hspace=0.05) + + for i, sample in enumerate(gen_data): + ax = plt.subplot(gs[i]) + plt.axis('off') + ax.set_xticklabels([]) + ax.set_yticklabels([]) + ax.set_aspect('equal') + plt.imshow(sample.reshape(28, 28), cmap='Greys_r') + + return fig + + +def main(): + try: + os.makedirs("./out") + except OSError as e: + if e.errno != errno.EEXIST: + raise + + startup_program = fluid.Program() + d_program = fluid.Program() + dg_program = fluid.Program() + + with fluid.program_guard(d_program, startup_program): + img = fluid.layers.data(name='img', shape=[784], dtype='float32') + d_loss = fluid.layers.sigmoid_cross_entropy_with_logits( + x=D(img), + label=fluid.layers.data( + name='label', shape=[1], dtype='float32')) + d_loss = fluid.layers.mean(x=d_loss) + + with fluid.program_guard(dg_program, startup_program): + noise = fluid.layers.data( + name='noise', shape=[NOISE_SIZE], dtype='float32') + g_img = G(x=noise) + g_program = dg_program.clone() + dg_loss = fluid.layers.sigmoid_cross_entropy_with_logits( + x=D(g_img), + label=fluid.layers.fill_constant_batch_size_like( + input=noise, dtype='float32', shape=[-1, 1], value=1.0)) + dg_loss = fluid.layers.mean(x=dg_loss) + + opt = fluid.optimizer.Adam(learning_rate=LEARNING_RATE) + + opt.minimize(loss=d_loss, startup_program=startup_program) + opt.minimize( + loss=dg_loss, + startup_program=startup_program, + parameter_list=[ + p.name for p in g_program.global_block().all_parameters() + ]) + exe = fluid.Executor(fluid.CPUPlace()) + exe.run(startup_program) + + num_true = NUM_REAL_IMGS_IN_BATCH + train_reader = paddle.batch( + paddle.reader.shuffle( + paddle.dataset.mnist.train(), buf_size=60000), + batch_size=num_true) + + for pass_id in range(NUM_PASS): + for batch_id, data in enumerate(train_reader()): + num_true = len(data) + n = numpy.random.uniform( + low=-1.0, high=1.0, + size=[num_true * NOISE_SIZE]).astype('float32').reshape( + [num_true, NOISE_SIZE]) + generated_img = exe.run(g_program, + feed={'noise': n}, + fetch_list={g_img})[0] + real_data = numpy.array(map(lambda x: x[0], data)).astype('float32') + real_data = real_data.reshape(num_true, 784) + total_data = numpy.concatenate([real_data, generated_img]) + total_label = numpy.concatenate([ + numpy.ones( + shape=[real_data.shape[0], 1], dtype='float32'), + numpy.zeros( + shape=[real_data.shape[0], 1], dtype='float32') + ]) + d_loss_np = exe.run(d_program, + feed={'img': total_data, + 'label': total_label}, + fetch_list={d_loss})[0] + for _ in xrange(NUM_TRAIN_TIMES_OF_DG): + n = numpy.random.uniform( + low=-1.0, high=1.0, + size=[2 * num_true * NOISE_SIZE]).astype('float32').reshape( + [2 * num_true, NOISE_SIZE, 1, 1]) + dg_loss_np = exe.run(dg_program, + feed={'noise': n}, + fetch_list={dg_loss})[0] + print("Pass ID={0}, Batch ID={1}, D-Loss={2}, DG-Loss={3}".format( + pass_id, batch_id, d_loss_np, dg_loss_np)) + # generate image each batch + fig = plot(generated_img) + plt.savefig( + 'out/{0}.png'.format(str(pass_id).zfill(3)), bbox_inches='tight') + plt.close(fig) + + +if __name__ == '__main__': + main() diff --git a/python/paddle/v2/fluid/tests/op_test.py b/python/paddle/v2/fluid/tests/op_test.py index 90269e308a31d2606b23d741ce0d0fa91a0a6aeb..b77d2b1268f27c5ec3c34839aaad9b75f0132c2e 100644 --- a/python/paddle/v2/fluid/tests/op_test.py +++ b/python/paddle/v2/fluid/tests/op_test.py @@ -4,7 +4,7 @@ import random import itertools import paddle.v2.fluid.core as core import collections -from paddle.v2.fluid.backward import append_backward_ops +from paddle.v2.fluid.backward import append_backward from paddle.v2.fluid.op import Operator from paddle.v2.fluid.executor import Executor from paddle.v2.fluid.framework import Program, OpProtoHolder @@ -90,12 +90,10 @@ def get_numeric_gradient(scope, def product(dim): return reduce(lambda a, b: a * b, dim, 1) - ctx = core.DeviceContext.create(core.CPUPlace()) - def get_output(): sum = [] for output_name in output_names: - op.run(scope, ctx) + op.run(scope, core.CPUPlace()) sum.append( np.array(scope.find_var(output_name).get_tensor()).mean()) return np.array(sum).mean() @@ -261,7 +259,10 @@ class OpTest(unittest.TestCase): feed_map = self.feed_var(inputs, place) exe = Executor(place) - outs = exe.run(program, feed=feed_map, fetch_list=fetch_list) + outs = exe.run(program, + feed=feed_map, + fetch_list=fetch_list, + return_numpy=False) for out_name, out_dup in Operator.get_op_outputs(self.op_type): if out_name not in self.outputs: @@ -315,7 +316,7 @@ class OpTest(unittest.TestCase): def check_output(self, atol=1e-5): places = [core.CPUPlace()] if core.is_compile_gpu() and core.op_support_gpu(self.op_type): - places.append(core.GPUPlace(0)) + places.append(core.CUDAPlace(0)) for place in places: self.check_output_with_place(place, atol) @@ -378,7 +379,7 @@ class OpTest(unittest.TestCase): "Gradient Check On %s" % str(cpu_place)) if core.is_compile_gpu() and self.op.support_gpu(): - gpu_place = core.GPUPlace(0) + gpu_place = core.CUDAPlace(0) gpu_analytic_grads = self._get_gradient(inputs_to_check, gpu_place, output_names, no_grad_set) @@ -458,7 +459,7 @@ class OpTest(unittest.TestCase): mean_inputs = map(block.var, output_names) if len(mean_inputs) == 1: - loss = block.create_var(dtype=mean_inputs[0].data_type, shape=[1]) + loss = block.create_var(dtype=mean_inputs[0].dtype, shape=[1]) op = block.append_op( inputs={"X": mean_inputs}, outputs={"Out": loss}, type='mean') op.desc.infer_var_type(block.desc) @@ -466,8 +467,7 @@ class OpTest(unittest.TestCase): else: avg_sum = [] for cur_loss in mean_inputs: - cur_avg_loss = block.create_var( - dtype=cur_loss.data_type, shape=[1]) + cur_avg_loss = block.create_var(dtype=cur_loss.dtype, shape=[1]) op = block.append_op( inputs={"X": [cur_loss]}, outputs={"Out": [cur_avg_loss]}, @@ -476,13 +476,13 @@ class OpTest(unittest.TestCase): op.desc.infer_shape(block.desc) avg_sum.append(cur_avg_loss) - loss_sum = block.create_var(dtype=avg_sum[0].data_type, shape=[1]) + loss_sum = block.create_var(dtype=avg_sum[0].dtype, shape=[1]) op_sum = block.append_op( inputs={"X": avg_sum}, outputs={"Out": loss_sum}, type='sum') op_sum.desc.infer_var_type(block.desc) op_sum.desc.infer_shape(block.desc) - loss = block.create_var(dtype=loss_sum.data_type, shape=[1]) + loss = block.create_var(dtype=loss_sum.dtype, shape=[1]) op_loss = block.append_op( inputs={"X": loss_sum}, outputs={"Out": loss}, @@ -491,7 +491,7 @@ class OpTest(unittest.TestCase): op_loss.desc.infer_var_type(block.desc) op_loss.desc.infer_shape(block.desc) - param_grad_list = append_backward_ops( + param_grad_list = append_backward( loss=loss, parameter_list=input_to_check, no_grad_set=no_grad_set) feed_dict = { @@ -501,5 +501,6 @@ class OpTest(unittest.TestCase): fetch_list = [g for p, g in param_grad_list] executor = Executor(place) - result = executor.run(prog, feed_dict, fetch_list) - return map(np.array, result) + return map( + np.array, + executor.run(prog, feed_dict, fetch_list, return_numpy=False)) diff --git a/python/paddle/v2/fluid/tests/test_activation_op.py b/python/paddle/v2/fluid/tests/test_activation_op.py index 7649e60a3833e34523d87cb963af3888c3cef65d..03eb7deb9a35933e5a1676a262a371c69151e6d1 100644 --- a/python/paddle/v2/fluid/tests/test_activation_op.py +++ b/python/paddle/v2/fluid/tests/test_activation_op.py @@ -1,6 +1,7 @@ import unittest import numpy as np from op_test import OpTest +from scipy.special import expit class TestExp(OpTest): @@ -9,13 +10,13 @@ class TestExp(OpTest): self.inputs = { 'X': np.random.uniform(0.1, 1, [11, 17]).astype("float32") } - self.outputs = {'Y': np.exp(self.inputs['X'])} + self.outputs = {'Out': np.exp(self.inputs['X'])} def test_check_output(self): self.check_output() def test_check_grad(self): - self.check_grad(['X'], 'Y', max_relative_error=0.007) + self.check_grad(['X'], 'Out', max_relative_error=0.007) class TestSigmoid(OpTest): @@ -24,13 +25,13 @@ class TestSigmoid(OpTest): self.inputs = { 'X': np.random.uniform(0.1, 1, [11, 17]).astype("float32") } - self.outputs = {'Y': 1 / (1 + np.exp(-self.inputs['X']))} + self.outputs = {'Out': 1 / (1 + np.exp(-self.inputs['X']))} def test_check_output(self): self.check_output() def test_check_grad(self): - self.check_grad(['X'], 'Y', max_relative_error=0.008) + self.check_grad(['X'], 'Out', max_relative_error=0.008) class TestLogSigmoid(OpTest): @@ -39,13 +40,13 @@ class TestLogSigmoid(OpTest): self.inputs = { 'X': np.random.uniform(-1, 1, [11, 17]).astype("float32") } - self.outputs = {'Y': np.log(1 / (1 + np.exp(-self.inputs['X'])))} + self.outputs = {'Out': np.log(1 / (1 + np.exp(-self.inputs['X'])))} def test_check_output(self): self.check_output() def test_check_grad(self): - self.check_grad(['X'], 'Y', max_relative_error=0.008) + self.check_grad(['X'], 'Out', max_relative_error=0.008) class TestTanh(OpTest): @@ -54,13 +55,13 @@ class TestTanh(OpTest): self.inputs = { 'X': np.random.uniform(0.1, 1, [11, 17]).astype("float32") } - self.outputs = {'Y': np.tanh(self.inputs['X'])} + self.outputs = {'Out': np.tanh(self.inputs['X'])} def test_check_output(self): self.check_output() def test_check_grad(self): - self.check_grad(['X'], 'Y', max_relative_error=0.007) + self.check_grad(['X'], 'Out', max_relative_error=0.007) class TestTanhShrink(OpTest): @@ -69,13 +70,13 @@ class TestTanhShrink(OpTest): self.inputs = { 'X': np.random.uniform(0.1, 1, [10, 17]).astype("float32") } - self.outputs = {'Y': self.inputs['X'] - np.tanh(self.inputs['X'])} + self.outputs = {'Out': self.inputs['X'] - np.tanh(self.inputs['X'])} def test_check_output(self): self.check_output() def test_check_grad(self): - self.check_grad(['X'], 'Y', max_relative_error=0.008) + self.check_grad(['X'], 'Out', max_relative_error=0.008) class TestHardShrink(OpTest): @@ -89,13 +90,13 @@ class TestHardShrink(OpTest): t = np.copy(x) t[(t >= -threshold) & (t <= threshold)] = 0 - self.outputs = {'Y': t} + self.outputs = {'Out': t} def test_check_output(self): self.check_output() def test_check_grad(self): - self.check_grad(['X'], 'Y', max_relative_error=0.005) + self.check_grad(['X'], 'Out', max_relative_error=0.005) class TestSoftShrink(OpTest): @@ -109,13 +110,13 @@ class TestSoftShrink(OpTest): y = np.copy(self.inputs['X']) y = (y < -lambda_val) * (y + lambda_val) + (y > lambda_val) * ( y - lambda_val) - self.outputs = {'Y': y} + self.outputs = {'Out': y} def test_check_output(self): self.check_output() def test_check_grad(self): - self.check_grad(['X'], 'Y', max_relative_error=0.007) + self.check_grad(['X'], 'Out', max_relative_error=0.007) class TestSqrt(OpTest): @@ -124,13 +125,13 @@ class TestSqrt(OpTest): self.inputs = { 'X': np.random.uniform(0.1, 1, [11, 17]).astype("float32") } - self.outputs = {'Y': np.sqrt(self.inputs['X'])} + self.outputs = {'Out': np.sqrt(self.inputs['X'])} def test_check_output(self): self.check_output() def test_check_grad(self): - self.check_grad(['X'], 'Y', max_relative_error=0.007) + self.check_grad(['X'], 'Out', max_relative_error=0.007) class TestAbs(OpTest): @@ -143,13 +144,56 @@ class TestAbs(OpTest): # we should avoid this x[np.abs(x) < 0.005] = 0.02 self.inputs = {'X': x} - self.outputs = {'Y': np.abs(self.inputs['X'])} + self.outputs = {'Out': np.abs(self.inputs['X'])} def test_check_output(self): self.check_output() def test_check_grad(self): - self.check_grad(['X'], 'Y', max_relative_error=0.007) + self.check_grad(['X'], 'Out', max_relative_error=0.007) + + +class TestCeil(OpTest): + def setUp(self): + self.op_type = "ceil" + x = np.random.uniform(-1, 1, [4, 4]).astype("float32") + self.inputs = {'X': x} + self.outputs = {'Out': np.ceil(self.inputs['X'])} + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad(['X'], 'Out', max_relative_error=0.007) + + +class TestFloor(OpTest): + def setUp(self): + self.op_type = "floor" + x = np.random.uniform(-1, 1, [4, 4]).astype("float32") + self.inputs = {'X': x} + # numpy floor need +1 + self.outputs = {'Out': np.floor(self.inputs['X']) + 1.0} + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad(['X'], 'Out', max_relative_error=0.007) + + +class TestRound(OpTest): + def setUp(self): + self.op_type = "round" + x = np.random.uniform(-1, 1, [4, 4]).astype("float32") + self.inputs = {'X': x} + self.outputs = {'Out': np.round(self.inputs['X'])} + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad(['X'], 'Out', max_relative_error=0.007) class TestRelu(OpTest): @@ -159,13 +203,13 @@ class TestRelu(OpTest): # The same reason with TestAbs x[np.abs(x) < 0.005] = 0.02 self.inputs = {'X': x} - self.outputs = {'Y': np.maximum(self.inputs['X'], 0)} + self.outputs = {'Out': np.maximum(self.inputs['X'], 0)} def test_check_output(self): self.check_output() def test_check_grad(self): - self.check_grad(['X'], 'Y', max_relative_error=0.007) + self.check_grad(['X'], 'Out', max_relative_error=0.007) class TestBRelu(OpTest): @@ -183,13 +227,13 @@ class TestBRelu(OpTest): t = np.copy(x) t[t < t_min] = t_min t[t > t_max] = t_max - self.outputs = {'Y': t} + self.outputs = {'Out': t} def test_check_output(self): self.check_output() def test_check_grad(self): - self.check_grad(['X'], 'Y', max_relative_error=0.02) + self.check_grad(['X'], 'Out', max_relative_error=0.02) class TestRelu6(OpTest): @@ -204,14 +248,14 @@ class TestRelu6(OpTest): self.inputs = {'X': x} self.attrs = {'threshold': threshold} self.outputs = { - 'Y': np.minimum(np.maximum(self.inputs['X'], 0), threshold) + 'Out': np.minimum(np.maximum(self.inputs['X'], 0), threshold) } def test_check_output(self): self.check_output() def test_check_grad(self): - self.check_grad(['X'], 'Y', max_relative_error=0.02) + self.check_grad(['X'], 'Out', max_relative_error=0.02) class TestSoftRelu(OpTest): @@ -227,13 +271,13 @@ class TestSoftRelu(OpTest): t = np.copy(x) t[t < -threshold] = -threshold t[t > threshold] = threshold - self.outputs = {'Y': np.log((np.exp(t) + 1))} + self.outputs = {'Out': np.log((np.exp(t) + 1))} def test_check_output(self): self.check_output() def test_check_grad(self): - self.check_grad(['X'], 'Y', max_relative_error=0.02) + self.check_grad(['X'], 'Out', max_relative_error=0.02) class TestELU(OpTest): @@ -246,27 +290,27 @@ class TestELU(OpTest): self.inputs = {'X': x} self.attrs = {'alpha': alpha} self.outputs = { - 'Y': np.maximum(0, x) + np.minimum(0, alpha * (np.exp(x) - 1)) + 'Out': np.maximum(0, x) + np.minimum(0, alpha * (np.exp(x) - 1)) } def test_check_output(self): self.check_output() def test_check_grad(self): - self.check_grad(['X'], 'Y', max_relative_error=0.02) + self.check_grad(['X'], 'Out', max_relative_error=0.02) class TestReciprocal(OpTest): def setUp(self): self.op_type = "reciprocal" self.inputs = {'X': np.random.uniform(1, 2, [11, 17]).astype("float32")} - self.outputs = {'Y': np.reciprocal(self.inputs['X'])} + self.outputs = {'Out': np.reciprocal(self.inputs['X'])} def test_check_output(self): self.check_output() def test_check_grad(self): - self.check_grad(['X'], 'Y', max_relative_error=0.01) + self.check_grad(['X'], 'Out', max_relative_error=0.01) class TestLog(OpTest): @@ -275,13 +319,13 @@ class TestLog(OpTest): self.inputs = { 'X': np.random.uniform(0.1, 1, [11, 17]).astype("float32") } - self.outputs = {'Y': np.log(self.inputs['X'])} + self.outputs = {'Out': np.log(self.inputs['X'])} def test_check_output(self): self.check_output() def test_check_grad(self): - self.check_grad(['X'], 'Y', max_relative_error=0.007) + self.check_grad(['X'], 'Out', max_relative_error=0.007) class TestSquare(OpTest): @@ -290,13 +334,13 @@ class TestSquare(OpTest): self.inputs = { 'X': np.random.uniform(0.1, 1, [11, 17]).astype("float32") } - self.outputs = {'Y': np.square(self.inputs['X'])} + self.outputs = {'Out': np.square(self.inputs['X'])} def test_check_output(self): self.check_output() def test_check_grad(self): - self.check_grad(['X'], 'Y', max_relative_error=0.007) + self.check_grad(['X'], 'Out', max_relative_error=0.007) class TestPow(OpTest): @@ -304,13 +348,13 @@ class TestPow(OpTest): self.op_type = "pow" self.inputs = {'X': np.random.uniform(1, 2, [11, 17]).astype("float32")} self.attrs = {'factor': 3.0} - self.outputs = {'Y': np.power(self.inputs['X'], 3)} + self.outputs = {'Out': np.power(self.inputs['X'], 3)} def test_check_output(self): self.check_output() def test_check_grad(self): - self.check_grad(['X'], 'Y', max_relative_error=0.02) + self.check_grad(['X'], 'Out', max_relative_error=0.02) class TestSTanh(OpTest): @@ -322,13 +366,13 @@ class TestSTanh(OpTest): scale_a = 2.0 / 3.0 scale_b = 1.7159 self.attrs = {'scale_a': scale_a, 'scale_b': scale_b} - self.outputs = {'Y': scale_b * np.tanh(self.inputs['X'] * scale_a)} + self.outputs = {'Out': scale_b * np.tanh(self.inputs['X'] * scale_a)} def test_check_output(self): self.check_output() def test_check_grad(self): - self.check_grad(['X'], 'Y', max_relative_error=0.007) + self.check_grad(['X'], 'Out', max_relative_error=0.007) class TestSoftplus(OpTest): @@ -337,13 +381,13 @@ class TestSoftplus(OpTest): self.inputs = { 'X': np.random.uniform(-1, 1, [11, 17]).astype("float64") } - self.outputs = {'Y': np.log(1 + np.exp(self.inputs['X']))} + self.outputs = {'Out': np.log(1 + np.exp(self.inputs['X']))} def test_check_output(self): self.check_output() def test_check_grad(self): - self.check_grad(['X'], 'Y', max_relative_error=0.007) + self.check_grad(['X'], 'Out', max_relative_error=0.007) class TestSoftsign(OpTest): @@ -353,14 +397,14 @@ class TestSoftsign(OpTest): 'X': np.random.uniform(-1, 1, [11, 17]).astype("float32") } self.outputs = { - 'Y': np.divide(self.inputs['X'], 1 + np.abs(self.inputs['X'])) + 'Out': np.divide(self.inputs['X'], 1 + np.abs(self.inputs['X'])) } def test_check_output(self): self.check_output() def test_check_grad(self): - self.check_grad(['X'], 'Y', max_relative_error=0.007) + self.check_grad(['X'], 'Out', max_relative_error=0.007) class TestThresholdedRelu(OpTest): @@ -375,13 +419,13 @@ class TestThresholdedRelu(OpTest): self.inputs = {'X': X} self.attrs = {'threshold': threshold} - self.outputs = {'Y': (X > threshold) * X} + self.outputs = {'Out': (X > threshold) * X} def test_check_output(self): self.check_output() def test_check_grad(self): - self.check_grad(['X'], 'Y', max_relative_error=self.relative_error) + self.check_grad(['X'], 'Out', max_relative_error=self.relative_error) class TestHardSigmoid(OpTest): @@ -403,13 +447,28 @@ class TestHardSigmoid(OpTest): upper_threshold - 0.2 temp = X * slope + offset - self.outputs = {'Y': np.maximum(0.0, np.minimum(1.0, temp))} + self.outputs = {'Out': np.maximum(0.0, np.minimum(1.0, temp))} + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad(['X'], 'Out', max_relative_error=0.002) + + +class TestSwish(OpTest): + def setUp(self): + self.op_type = "swish" + X = np.random.uniform(0.1, 1, [11, 17]).astype("float32") + self.inputs = {'X': X} + self.attrs = {'beta': 2.3} + self.outputs = {'Out': X * expit(self.attrs['beta'] * X)} def test_check_output(self): self.check_output() def test_check_grad(self): - self.check_grad(['X'], 'Y', max_relative_error=0.002) + self.check_grad(['X'], 'Out', max_relative_error=0.008) if __name__ == "__main__": diff --git a/python/paddle/v2/fluid/tests/test_adagrad_op.py b/python/paddle/v2/fluid/tests/test_adagrad_op.py index 903e84c32887100bbeef6ebf81f66f06f084fab5..7b2d02fbf4256d2c27383a3452d526271af543a3 100644 --- a/python/paddle/v2/fluid/tests/test_adagrad_op.py +++ b/python/paddle/v2/fluid/tests/test_adagrad_op.py @@ -113,8 +113,7 @@ class TestSparseAdagradOp(unittest.TestCase): LearningRate='LearningRate', epsilon=2.0) - ctx = core.DeviceContext.create(place) - adagrad_op.run(scope, ctx) + adagrad_op.run(scope, place) # get and compare moment result moment_result_array = np.array(moment) @@ -168,7 +167,7 @@ class TestSparseAdagradOp(unittest.TestCase): def test_sparse_adagrad(self): places = [core.CPUPlace()] if core.is_compile_gpu(): - places.append(core.GPUPlace(0)) + places.append(core.CUDAPlace(0)) for place in places: self.check_with_place(place) diff --git a/python/paddle/v2/fluid/tests/test_adam_op.py b/python/paddle/v2/fluid/tests/test_adam_op.py index a0d6655d4cbcff8ed3d55df0f4e68fc6591fbb11..7dbc2fa0858a68c5da9e8d48dcb187494357e940 100644 --- a/python/paddle/v2/fluid/tests/test_adam_op.py +++ b/python/paddle/v2/fluid/tests/test_adam_op.py @@ -1,6 +1,8 @@ import unittest import numpy as np from op_test import OpTest +from paddle.v2.fluid import core +from paddle.v2.fluid.op import Operator class TestAdamOp1(OpTest): @@ -176,5 +178,124 @@ def adam_step(inputs, attributes): return param_out, moment1_out, moment2_out +def adam_step_sparse(inputs, attributes, height, rows, row_numel, np_grad): + ''' + Simulate one step of the adam optimizer + :param inputs: dict of inputs + :param attributes: dict of attributes + :return tuple: tuple of output param, moment1, moment2, + beta1 power accumulator and beta2 power accumulator + ''' + param = inputs['Param'] + # grad = inputs['Grad'] + moment1 = inputs['Moment1'] + moment2 = inputs['Moment2'] + lr = inputs['LearningRate'] + beta1_pow = inputs['Beta1Pow'] + beta2_pow = inputs['Beta2Pow'] + + beta1 = attributes['beta1'] + beta2 = attributes['beta2'] + epsilon = attributes['epsilon'] + + moment1_out = np.zeros(shape=[height, row_numel]) + moment2_out = np.zeros(shape=[height, row_numel]) + param_out = np.zeros(shape=[height, row_numel]) + + for idx, row_id in enumerate(rows): + moment1_out[row_id] = beta1 * moment1[row_id] + (1 - beta1 + ) * np_grad[idx] + moment2_out[row_id] = beta2 * moment2[row_id] + ( + 1 - beta2) * np.square(np_grad[idx]) + lr_t = lr * np.sqrt(1 - beta2_pow) / (1 - beta1_pow) + param_out[row_id] = param[row_id] - lr_t * (moment1_out[row_id] / ( + np.sqrt(moment2_out[row_id]) + epsilon)) + return param_out, moment1_out, moment2_out + + +class TestSparseAdamOp(unittest.TestCase): + def setup(self, scope, place): + beta1 = 0.78 + beta2 = 0.836 + epsilon = 1e-4 + + height = 10 + rows = [0, 4, 7] + self.rows = rows + row_numel = 12 + self.row_numel = row_numel + self.dense_inputs = { + "Param": np.full((height, row_numel), 5.0).astype("float32"), + "Moment1": np.full((height, row_numel), 5.0).astype("float32"), + "Moment2": np.full((height, row_numel), 5.0).astype("float32"), + 'Beta1Pow': np.array([beta1**10]).astype("float32"), + 'Beta2Pow': np.array([beta2**10]).astype("float32"), + "LearningRate": np.full((1), 2.0).astype("float32") + } + self.attrs = {'epsilon': epsilon, 'beta1': beta1, 'beta2': beta2} + + grad_selected_rows = scope.var('Grad').get_selected_rows() + grad_selected_rows.set_height(height) + grad_selected_rows.set_rows(rows) + np_array = np.ones((len(rows), row_numel)).astype("float32") + np_array[0, 0] = 2.0 + np_array[2, 8] = 4.0 + + grad_tensor = grad_selected_rows.get_tensor() + grad_tensor.set(np_array, place) + + self.sparse_inputs = ["Grad"] + + param_out, mom1, mom2 = adam_step_sparse( + self.dense_inputs, self.attrs, height, rows, row_numel, np_array) + self.outputs = { + "ParamOut": param_out, + "Moment1Out": mom1, + "Moment2Out": mom2 + } + + def check_with_place(self, place): + scope = core.Scope() + self.setup(scope, place) + + op_args = dict() + for key, np_array in self.dense_inputs.iteritems(): + var = scope.var(key).get_tensor() + var.set(np_array, place) + op_args[key] = key + for s in self.sparse_inputs: + op_args[s] = s + for s in self.outputs: + var = scope.var(s).get_tensor() + var.set(self.outputs[s], place) + op_args[s] = s + for k in self.attrs: + op_args[k] = self.attrs[k] + + # create and run sgd operator + adam_op = Operator("adam", **op_args) + adam_op.run(scope, place) + + for key, np_array in self.outputs.iteritems(): + out_var = scope.var(key).get_tensor() + actual = np.array(out_var) + actual = actual.reshape([actual.size]) + np_array = np_array.reshape([np_array.size]) + for idx, row_id in enumerate(self.rows): + j = 0 + while j < self.row_numel: + pos = row_id * self.row_numel + j + self.assertLess((actual[pos] - np_array[pos]) / actual[pos], + 0.00001) + j += 1 + + def test_sparse_sgd(self): + places = [core.CPUPlace()] + if core.is_compile_gpu(): + places.append(core.CUDAPlace(0)) + for place in places: + self.check_with_place(place) + + if __name__ == "__main__": unittest.main() diff --git a/python/paddle/v2/fluid/tests/test_array_read_write_op.py b/python/paddle/v2/fluid/tests/test_array_read_write_op.py index e019a4e15f0e25deaedf30911b44e576c8f89013..01321de8eac34d562d99726b1f4125d1932ab40f 100644 --- a/python/paddle/v2/fluid/tests/test_array_read_write_op.py +++ b/python/paddle/v2/fluid/tests/test_array_read_write_op.py @@ -2,8 +2,8 @@ import unittest import paddle.v2.fluid.core as core import paddle.v2.fluid.layers as layers from paddle.v2.fluid.executor import Executor -from paddle.v2.fluid.backward import append_backward_ops -from paddle.v2.fluid.framework import g_main_program +from paddle.v2.fluid.backward import append_backward +from paddle.v2.fluid.framework import default_main_program import numpy @@ -52,32 +52,29 @@ class TestArrayReadWrite(unittest.TestCase): exe = Executor(cpu) - tensor = core.LoDTensor() - tensor.set(numpy.random.random(size=(100, 100)).astype('float32'), cpu) + tensor = numpy.random.random(size=(100, 100)).astype('float32') - outs = map(numpy.array, - exe.run(feed={'x0': tensor, - 'x1': tensor, - 'x2': tensor}, - fetch_list=[a_sum, x_sum], - scope=scope)) + outs = exe.run(feed={'x0': tensor, + 'x1': tensor, + 'x2': tensor}, + fetch_list=[a_sum, x_sum], + scope=scope) self.assertEqual(outs[0], outs[1]) total_sum = layers.sums(input=[a_sum, x_sum]) total_sum_scaled = layers.scale(x=total_sum, scale=1 / 6.0) - append_backward_ops(total_sum_scaled) + append_backward(total_sum_scaled) - g_vars = map(g_main_program.global_block().var, + g_vars = map(default_main_program().global_block().var, [each_x.name + "@GRAD" for each_x in x]) g_out = [ item.sum() - for item in map( - numpy.array, - exe.run(feed={'x0': tensor, - 'x1': tensor, - 'x2': tensor}, - fetch_list=g_vars)) + for item in exe.run( + feed={'x0': tensor, + 'x1': tensor, + 'x2': tensor}, + fetch_list=g_vars) ] g_out_sum = numpy.array(g_out).sum() diff --git a/python/paddle/v2/fluid/tests/test_batch_norm_op.py b/python/paddle/v2/fluid/tests/test_batch_norm_op.py index 71f9599e0de83c86808f7e62547f80d3d50ffc7d..abbd48d2b843cedb77caffc13413d2f9695defa6 100644 --- a/python/paddle/v2/fluid/tests/test_batch_norm_op.py +++ b/python/paddle/v2/fluid/tests/test_batch_norm_op.py @@ -3,10 +3,7 @@ import numpy as np from op_test import OpTest import paddle.v2.fluid.core as core from paddle.v2.fluid.op import Operator - - -def grad_var_name(var_name): - return var_name + "@GRAD" +from paddle.v2.fluid.framework import grad_var_name def get_backward_op(scope, op, no_grad_set): @@ -21,6 +18,13 @@ def get_backward_op(scope, op, no_grad_set): def _reference_training(x, scale, offset, epsilon, data_format): + x_shape = x.shape + if len(x_shape) == 2: + if data_format == "NCHW": + x = np.reshape(x, (x.shape[0], x.shape[1], 1, 1)) + else: + x = np.reshape(x, (x.shape[0], 1, 1, x.shape[1])) + if data_format == "NCHW": n, c, h, w = x.shape x_square = x * x @@ -39,6 +43,8 @@ def _reference_training(x, scale, offset, epsilon, data_format): offset_tile = np.reshape(offset, (1, c, 1, 1)) offset_tile = np.reshape(offset_tile, (1, c, 1, 1)) y = normalized * scale_tile + offset_tile + if len(x_shape) == 2: + y = np.reshape(y, (y.shape[0], y.shape[1])) return y, mean, var elif data_format == "NHWC": x_square = x * x @@ -48,7 +54,10 @@ def _reference_training(x, scale, offset, epsilon, data_format): mean = x_sum / element_count var = x_square_sum / element_count - mean * mean normalized = (x - mean) / np.sqrt(var + epsilon) - return (normalized * scale + offset), mean, var + y = normalized * scale + offset + if len(x_shape) == 2: + y = np.reshape(y, x_shape) + return y, mean, var else: raise ValueError("Unknown data order.") @@ -65,6 +74,18 @@ def _reference_grad(x, grad_y, scale, mean, var, epsilon, data_format): # (x - mean) * sum(grad_y * (x - mean)) / (var + epsilon)) # transfer from (N, C, H, W) to (N, H, W, C) to simplify computation + x_shape = x.shape + + if len(x_shape) == 2: + if data_format == "NCHW": + x = np.reshape(x, (x.shape[0], x.shape[1], 1, 1)) + grad_y = np.reshape(grad_y, + (grad_y.shape[0], grad_y.shape[1], 1, 1)) + else: + x = np.reshape(x, (x.shape[0], 1, 1, x.shape[1])) + grad_y = np.reshape(grad_y, + (grad_y.shape[0], 1, 1, grad_y.shape[1])) + if data_format == "NCHW": x = np.transpose(x, (0, 2, 3, 1)) grad_y = np.transpose(grad_y, (0, 2, 3, 1)) @@ -83,6 +104,9 @@ def _reference_grad(x, grad_y, scale, mean, var, epsilon, data_format): grad_x = np.transpose(grad_x, (0, 3, 1, 2)) x = np.transpose(x, (0, 3, 1, 2)) grad_y = np.transpose(grad_y, (0, 3, 1, 2)) + + if len(x_shape) == 2: + grad_x = np.reshape(grad_x, x_shape) return grad_x, grad_scale, grad_offset @@ -127,7 +151,7 @@ class TestBatchNormOp(OpTest): momentum = 0.9 # N, H, W, C: 2, 3, 4, 2 - n, h, w, c = 2, 3, 4, 2 + n, h, w, c = 2, 3, 4, 5 x_shape = [n, h, w, c] scale_shape = [c] @@ -184,20 +208,23 @@ class TestBatchNormOp(OpTest): print 'python: NHWC, NCHW, backward checking passed' def test_forward_backward(self): - def test_with_place(place, tensor_format): + def test_with_place(place, data_layout, shape): # attr epsilon = 0.00001 momentum = 0.9 - # N, H, W, C: 12, 3, 4, 2 - n, h, w, c = 2, 3, 4, 2 - - if data_format == "NHWC": - x_shape = [n, h, w, c] - elif data_format == "NCHW": - x_shape = [n, c, h, w] + if len(shape) == 2: + x_shape = shape + c = shape[1] else: - raise ValueError("Unknown data type.") + # n, h, w, c = 2, 3, 4, 2 + n, h, w, c = shape[0], shape[1], shape[2], shape[3] + if data_format == "NHWC": + x_shape = [n, h, w, c] + elif data_format == "NCHW": + x_shape = [n, c, h, w] + else: + raise ValueError("Unknown data type.") scale_shape = [c] x_val = np.random.random_sample(x_shape).astype(np.float32) @@ -219,7 +246,10 @@ class TestBatchNormOp(OpTest): # for gradient test # y_grad = np.ones(x_shape).astype(np.float32) y_grad = np.zeros(x_shape).astype(np.float32) - y_grad[0, 0, 0, 0] = 1. + if len(y_grad.shape) == 2: + y_grad[0, 0] = 1. + else: + y_grad[0, 0, 0, 0] = 1. # y_grad = np.random.random_sample(x_shape).astype(np.float32) x_grad_ref, scale_grad_ref, bias_grad_ref = _reference_grad( x_val, y_grad, scale_val, saved_mean, var_ref, epsilon, @@ -262,12 +292,11 @@ class TestBatchNormOp(OpTest): SavedVariance="saved_variance", # attrs is_test=False, - tensor_format=tensor_format, + data_layout=data_layout, momentum=momentum, epsilon=epsilon) - ctx = core.DeviceContext.create(place) - batch_norm_op.run(scope, ctx) + batch_norm_op.run(scope, place) # check forward result self.__assert_close(y_tensor, y_out, "y_out") @@ -275,13 +304,13 @@ class TestBatchNormOp(OpTest): self.__assert_close(saved_variance_tensor, saved_variance, "saved_variance") self.__assert_close(mean_out_tensor, mean_out, "mean_out") - if isinstance(place, core.GPUPlace): + if isinstance(place, core.CUDAPlace): atol = 5e-2 else: atol = 1e-4 self.__assert_close(variance_out_tensor, variance_out, "variance_out", atol) - print "op test forward passed: ", str(place), tensor_format + print "op test forward passed: ", str(place), data_layout # run backward batch_norm_op_grad = get_backward_op(scope, batch_norm_op, set()) @@ -290,7 +319,7 @@ class TestBatchNormOp(OpTest): ["y_out", "mean", "variance", "saved_mean", "saved_variance"], place, feed_dict={"y_out": y_grad}) - batch_norm_op_grad.run(scope, ctx) + batch_norm_op_grad.run(scope, place) x_grad_tensor = create_or_get_tensor(scope, grad_var_name("x_val"), None, @@ -306,14 +335,19 @@ class TestBatchNormOp(OpTest): self.__assert_close(x_grad_tensor, x_grad_ref, "x_grad") self.__assert_close(scale_grad_tensor, scale_grad_ref, "scale_grad") self.__assert_close(bias_grad_tensor, bias_grad_ref, "bias_grad") - print "op test backward passed: ", str(place), tensor_format + print "op test backward passed: ", str(place), data_layout places = [core.CPUPlace()] if core.is_compile_gpu() and core.op_support_gpu("batch_norm"): - places.append(core.GPUPlace(0)) + places.append(core.CUDAPlace(0)) + + core.init_devices(["CPU", "GPU:0"]) + else: + core.init_devices(["CPU"]) for place in places: for data_format in ["NCHW", "NHWC"]: - test_with_place(place, data_format) + test_with_place(place, data_format, [2, 3, 4, 5]) + test_with_place(place, data_format, [2, 3]) if __name__ == '__main__': diff --git a/python/paddle/v2/fluid/tests/test_beam_search_decode_op.py b/python/paddle/v2/fluid/tests/test_beam_search_decode_op.py index 8a11820d2aba2dd4d17d925f0e0fe9f324100418..f329214dce407fe0382c51b29f0f4c33b562541a 100644 --- a/python/paddle/v2/fluid/tests/test_beam_search_decode_op.py +++ b/python/paddle/v2/fluid/tests/test_beam_search_decode_op.py @@ -35,15 +35,15 @@ class TestBeamSearchDecodeOp(unittest.TestCase): self.append_lod_tensor( scores, [[0, 3, 6], [0, 1, 2, 3, 4, 5, 6]], np.array( - [1, 2, 3, 4, 5, 6], dtype="float32")) + [1, 2, 3, 4, 5, 6], dtype="float64")) self.append_lod_tensor( scores, [[0, 3, 6], [0, 1, 1, 3, 5, 5, 6]], np.array( - [0, 1, 2, 3, 4, 5], dtype="float32")) + [0, 1, 2, 3, 4, 5], dtype="float64")) self.append_lod_tensor( scores, [[0, 3, 6], [0, 0, 1, 2, 3, 4, 5]], np.array( - [0, 1, 2, 3, 4], dtype="float32")) + [0, 1, 2, 3, 4], dtype="float64")) sentence_ids = self.scope.var("sentence_ids").get_tensor() sentence_scores = self.scope.var("sentence_scores").get_tensor() @@ -57,8 +57,7 @@ class TestBeamSearchDecodeOp(unittest.TestCase): SentenceIds="sentence_ids", SentenceScores="sentence_scores") - ctx = core.DeviceContext.create(self.cpu_place) - beam_search_decode_op.run(self.scope, ctx) + beam_search_decode_op.run(self.scope, self.cpu_place) expected_lod = [[0, 4, 8], [0, 1, 3, 6, 9, 10, 13, 16, 19]] self.assertEqual(sentence_ids.lod(), expected_lod) diff --git a/python/paddle/v2/fluid/tests/test_beam_search_op.py b/python/paddle/v2/fluid/tests/test_beam_search_op.py index cc7c09bb59de3f83e47b4d95c1203f7f050c5132..595f132fa85f0a65f15d9ac31ad320e567c96358 100644 --- a/python/paddle/v2/fluid/tests/test_beam_search_op.py +++ b/python/paddle/v2/fluid/tests/test_beam_search_op.py @@ -14,7 +14,6 @@ def create_tensor(scope, name, np_data): class BeamSearchOpTester(unittest.TestCase): def setUp(self): self.scope = core.Scope() - self.ctx = core.DeviceContext.create(core.CPUPlace()) self._create_ids() self._create_scores() self._create_pre_ids() @@ -32,7 +31,7 @@ class BeamSearchOpTester(unittest.TestCase): level=0, beam_size=2, end_id=0, ) - op.run(self.scope, self.ctx) + op.run(self.scope, core.CPUPlace()) selected_ids = self.scope.find_var("selected_ids").get_tensor() print 'selected_ids', np.array(selected_ids) print 'lod', selected_ids.lod() diff --git a/python/paddle/v2/fluid/tests/test_cast_op.py b/python/paddle/v2/fluid/tests/test_cast_op.py index 0c4b6310652e84d3dd7f281a8b98ae0435072afb..4e431bb88da6070718d64a68467be20ca87f8fb9 100644 --- a/python/paddle/v2/fluid/tests/test_cast_op.py +++ b/python/paddle/v2/fluid/tests/test_cast_op.py @@ -10,8 +10,8 @@ class TestCastOp(op_test.OpTest): self.inputs = {'X': ipt.astype('float32')} self.outputs = {'Out': ipt.astype('float64')} self.attrs = { - 'in_data_type': int(core.DataType.FP32), - 'out_data_type': int(core.DataType.FP64) + 'in_dtype': int(core.DataType.FP32), + 'out_dtype': int(core.DataType.FP64) } self.op_type = 'cast' diff --git a/python/paddle/v2/fluid/tests/test_chunk_eval_op.py b/python/paddle/v2/fluid/tests/test_chunk_eval_op.py index 48673296a67716c4de804da533f0fd2567f10e2e..53bf6f815b8c7baf4c92d9fd488b69722ab0bef5 100644 --- a/python/paddle/v2/fluid/tests/test_chunk_eval_op.py +++ b/python/paddle/v2/fluid/tests/test_chunk_eval_op.py @@ -120,7 +120,7 @@ class TestChunkEvalOp(OpTest): self.num_correct_chunks, self.num_infer_chunks, self.num_label_chunks = 4, 5, 9 def set_data(self): - infer = np.zeros((self.batch_size, )).astype('int32') + infer = np.zeros((self.batch_size, )).astype('int64') infer.fill(self.num_chunk_types * self.num_tag_types) label = np.copy(infer) starts = np.random.choice( @@ -147,7 +147,13 @@ class TestChunkEvalOp(OpTest): 'Recall': np.asarray( [recall], dtype='float32'), 'F1-Score': np.asarray( - [f1], dtype='float32') + [f1], dtype='float32'), + 'NumInferChunks': np.asarray( + [self.num_infer_chunks], dtype='int64'), + 'NumLabelChunks': np.asarray( + [self.num_label_chunks], dtype='int64'), + 'NumCorrectChunks': np.asarray( + [self.num_correct_chunks], dtype='int64') } def setUp(self): diff --git a/python/paddle/v2/fluid/tests/test_cond_op.py b/python/paddle/v2/fluid/tests/test_cond_op.py index 9d1df44b9065f8101e90b87815660f8c0818645f..32e54084e48cf77c569db4dee54a0c89d5108373 100644 --- a/python/paddle/v2/fluid/tests/test_cond_op.py +++ b/python/paddle/v2/fluid/tests/test_cond_op.py @@ -65,8 +65,7 @@ class TestCondOp(unittest.TestCase): self.create_global_variables() self.create_cond_op() self.create_sub_net() - ctx = core.DeviceContext.create(core.CPUPlace()) - self.condop.run(self.scope, ctx) + self.condop.run(self.scope, core.CPUPlace()) return np.array(self.scope.find_var("Out").get_tensor()) def create_global_variables(self): diff --git a/python/paddle/v2/fluid/tests/test_conditional_block.py b/python/paddle/v2/fluid/tests/test_conditional_block.py index 293803f004a1513611fba30634d5552e1da84fef..7d815123f3454d1457f59202219f9a93bf3d8c31 100644 --- a/python/paddle/v2/fluid/tests/test_conditional_block.py +++ b/python/paddle/v2/fluid/tests/test_conditional_block.py @@ -1,15 +1,15 @@ import unittest import paddle.v2.fluid.layers as layers import paddle.v2.fluid.core as core -from paddle.v2.fluid.framework import g_startup_program, g_main_program +from paddle.v2.fluid.framework import default_startup_program, default_main_program from paddle.v2.fluid.executor import Executor -from paddle.v2.fluid.backward import append_backward_ops +from paddle.v2.fluid.backward import append_backward import numpy class ConditionalBlock(unittest.TestCase): def test_forward(self): - data = layers.data(name='X', shape=[1], data_type='float32') + data = layers.data(name='X', shape=[1], dtype='float32') data.stop_gradient = False cond = layers.ConditionalBlock(inputs=[data]) out = layers.create_tensor(dtype='float32') @@ -19,20 +19,19 @@ class ConditionalBlock(unittest.TestCase): cpu = core.CPUPlace() exe = Executor(cpu) - exe.run(g_startup_program) + exe.run(default_startup_program()) - x = core.LoDTensor() - x.set(numpy.random.random(size=(10, 1)).astype('float32'), cpu) + x = numpy.random.random(size=(10, 1)).astype('float32') - outs = map(numpy.array, exe.run(feed={'X': x}, fetch_list=[out]))[0] + outs = exe.run(feed={'X': x}, fetch_list=[out])[0] print outs loss = layers.mean(x=out) - append_backward_ops(loss=loss) - outs = map(numpy.array, - exe.run(feed={'X': x}, - fetch_list=[ - g_main_program.block(0).var(data.name + "@GRAD") - ]))[0] + append_backward(loss=loss) + outs = exe.run( + feed={'X': x}, + fetch_list=[ + default_main_program().block(0).var(data.name + "@GRAD") + ])[0] print outs diff --git a/python/paddle/v2/fluid/tests/test_const_value.py b/python/paddle/v2/fluid/tests/test_const_value.py new file mode 100644 index 0000000000000000000000000000000000000000..f8c17c2c98674fa67458efa090e166e37f5a6a8a --- /dev/null +++ b/python/paddle/v2/fluid/tests/test_const_value.py @@ -0,0 +1,14 @@ +import unittest +import paddle.v2.fluid.framework as framework + + +class ConditionalBlock(unittest.TestCase): + def test_const_value(self): + self.assertEqual(framework.GRAD_VAR_SUFFIX, "@GRAD") + self.assertEqual(framework.TEMP_VAR_NAME, "@TEMP@") + self.assertEqual(framework.GRAD_VAR_SUFFIX, "@GRAD") + self.assertEqual(framework.ZERO_VAR_SUFFIX, "@ZERO") + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/v2/fluid/tests/test_conv2d_op.py b/python/paddle/v2/fluid/tests/test_conv2d_op.py index 2240dc73cdd31f320fed174dd811e93c6640137f..e82e3ab0c9c0bc75a13a8948fda925bc4f0b6512 100644 --- a/python/paddle/v2/fluid/tests/test_conv2d_op.py +++ b/python/paddle/v2/fluid/tests/test_conv2d_op.py @@ -16,8 +16,8 @@ def conv2d_forward_naive(input, filter, group, conv_param): out_w = 1 + (in_w + 2 * pad[1] - (dilation[1] * (f_w - 1) + 1)) / stride[1] out = np.zeros((in_n, out_c, out_h, out_w)) - d_bolck_w = (dilation[0] * (f_h - 1) + 1) - d_bolck_h = (dilation[1] * (f_w - 1) + 1) + d_bolck_h = (dilation[0] * (f_h - 1) + 1) + d_bolck_w = (dilation[1] * (f_w - 1) + 1) input_pad = np.pad(input, ((0, ), (0, ), (pad[0], ), (pad[1], )), mode='constant', @@ -167,27 +167,27 @@ class TestWithDilation(TestConv2dOp): #----------------Conv2dCudnn---------------- class TestCudnn(TestConv2dOp): def init_op_type(self): - self.op_type = "conv_cudnn" + self.op_type = "conv2d_cudnn" class TestCudnnWithPad(TestWithPad): def init_op_type(self): - self.op_type = "conv_cudnn" + self.op_type = "conv2d_cudnn" class TestCudnnWithStride(TestWithStride): def init_op_type(self): - self.op_type = "conv_cudnn" + self.op_type = "conv2d_cudnn" class TestCudnnWithGroup(TestWithGroup): def init_op_type(self): - self.op_type = "conv_cudnn" + self.op_type = "conv2d_cudnn" class TestCudnnWith1x1(TestWith1x1): def init_op_type(self): - self.op_type = "conv_cudnn" + self.op_type = "conv2d_cudnn" # cudnn v5 does not support dilation conv. diff --git a/python/paddle/v2/fluid/tests/test_conv2d_transpose_op.py b/python/paddle/v2/fluid/tests/test_conv2d_transpose_op.py index d7b1f2f2a3abf6335998742dbbef8e17794170fa..d59537b924d57d40f7d740d99eb814c95f528e5f 100644 --- a/python/paddle/v2/fluid/tests/test_conv2d_transpose_op.py +++ b/python/paddle/v2/fluid/tests/test_conv2d_transpose_op.py @@ -3,14 +3,17 @@ import numpy as np from op_test import OpTest -def conv2dtranspose_forward_naive(input_, filter_, conv2dtranspose_param): +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 assert in_c == f_c - stride, pad = conv2dtranspose_param['stride'], conv2dtranspose_param['pad'] - out_h = (in_h - 1) * stride[0] + f_h - out_w = (in_w - 1) * stride[1] + f_w + stride, pad, dilations = attrs['strides'], attrs['paddings'], attrs[ + 'dilations'] + d_bolck_h = dilations[0] * (f_h - 1) + 1 + d_bolck_w = dilations[1] * (f_w - 1) + 1 + out_h = (in_h - 1) * stride[0] + d_bolck_h + out_w = (in_w - 1) * stride[1] + d_bolck_w out = np.zeros((in_n, out_c, out_h, out_w)) @@ -23,9 +26,9 @@ def conv2dtranspose_forward_naive(input_, filter_, conv2dtranspose_param): for k in range(out_c): tmp_out = np.sum(input_masked * filter_[:, k, :, :], axis=0) - i1, i2 = i * stride[0], i * stride[0] + f_h - j1, j2 = j * stride[0], j * stride[0] + f_w - out[n, k, i1:i2, j1:j2] += tmp_out + 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 out = out[:, :, pad[0]:out_h - pad[0], pad[1]:out_w - pad[1]] return out @@ -37,11 +40,8 @@ class TestConv2dTransposeOp(OpTest): self.init_op_type() self.init_test_case() - conv2dtranspose_param = {'stride': self.stride, 'pad': self.pad} input_ = np.random.random(self.input_size).astype("float32") filter_ = np.random.random(self.filter_size).astype("float32") - output = conv2dtranspose_forward_naive( - input_, filter_, conv2dtranspose_param).astype('float32') self.inputs = {'Input': input_, 'Filter': filter_} self.attrs = { @@ -49,6 +49,10 @@ class TestConv2dTransposeOp(OpTest): 'paddings': self.pad, 'dilations': self.dilations } + + output = conv2dtranspose_forward_naive(input_, filter_, + self.attrs).astype('float32') + self.outputs = {'Output': output} def test_check_output(self): @@ -104,11 +108,60 @@ class TestWithStride(TestConv2dTransposeOp): self.filter_size = [f_c, 6, 3, 3] +class TestWithDilation(TestConv2dTransposeOp): + def init_test_case(self): + self.pad = [1, 1] + self.stride = [1, 1] + self.dilations = [2, 2] + self.input_size = [2, 3, 5, 5] # NCHW + f_c = self.input_size[1] + self.filter_size = [f_c, 6, 3, 3] + + # ------------ test_cudnn ------------ class TestCudnn(TestConv2dTransposeOp): def init_op_type(self): self.op_type = "conv2d_transpose_cudnn" +class TestCudnnWithPad(TestWithPad): + def init_test_case(self): + self.pad = [1, 1] + self.stride = [1, 1] + self.dilations = [1, 1] + self.input_size = [2, 3, 5, 5] # NCHW + f_c = self.input_size[1] + self.filter_size = [f_c, 6, 3, 3] + + def init_op_type(self): + self.op_type = "conv2d_transpose_cudnn" + + +class TestCudnnWithStride(TestWithStride): + def init_test_case(self): + self.pad = [1, 1] + self.stride = [2, 2] + self.dilations = [1, 1] + self.input_size = [2, 3, 5, 5] # NCHW + f_c = self.input_size[1] + self.filter_size = [f_c, 6, 3, 3] + + def init_op_type(self): + self.op_type = "conv2d_transpose_cudnn" + + +# #cudnn v5 does not support dilation conv. +# class TestCudnnWithDilation(TestWithDilation): +# def init_test_case(self): +# self.pad = [1, 1] +# self.stride = [2, 2] +# self.dilations = [2, 2] +# self.input_size = [2, 3, 5, 5] # NCHW +# f_c = self.input_size[1] +# self.filter_size = [f_c, 6, 3, 3] +# +# def init_op_type(self): +# self.op_type = "conv2d_transpose_cudnn" + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/v2/fluid/tests/test_conv3d_op.py b/python/paddle/v2/fluid/tests/test_conv3d_op.py index 934ea46437d67b78309a86a2779e0c6577399136..8593dff20b5c283d5862206dfb0c0d2501039d07 100644 --- a/python/paddle/v2/fluid/tests/test_conv3d_op.py +++ b/python/paddle/v2/fluid/tests/test_conv3d_op.py @@ -169,5 +169,31 @@ class TestWithDilation(TestConv3dOp): self.groups = 3 +class TestCudnn(TestConv3dOp): + def init_op_type(self): + self.op_type = "conv3d_cudnn" + + +class TestWithGroup1Cudnn(TestWithGroup1): + def init_op_type(self): + self.op_type = "conv3d_cudnn" + + +class TestWithGroup2Cudnn(TestWithGroup2): + def init_op_type(self): + self.op_type = "conv3d_cudnn" + + +class TestWith1x1Cudnn(TestWith1x1): + def init_op_type(self): + self.op_type = "conv3d_cudnn" + + +# FIXME(typhoonzero): find a way to determine if +# using cudnn > 6 in python +# class TestWithDilationCudnn(TestWithDilation): +# def init_op_type(self): +# self.op_type = "conv3d_cudnn" + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/v2/fluid/tests/test_conv3d_transpose_op.py b/python/paddle/v2/fluid/tests/test_conv3d_transpose_op.py index 8fd34b87bfea91307f52fdcbb9f71f2e1a9c6c56..a353f9b4d40233de46237005138f21430f4d865a 100644 --- a/python/paddle/v2/fluid/tests/test_conv3d_transpose_op.py +++ b/python/paddle/v2/fluid/tests/test_conv3d_transpose_op.py @@ -3,15 +3,20 @@ import numpy as np from op_test import OpTest -def conv3dtranspose_forward_naive(input_, filter_, conv3dtranspose_param): +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 assert in_c == f_c - stride, pad = conv3dtranspose_param['stride'], conv3dtranspose_param['pad'] - out_d = (in_d - 1) * stride[0] + f_d - out_h = (in_h - 1) * stride[1] + f_h - out_w = (in_w - 1) * stride[2] + f_w + stride, pad, dilations = attrs['strides'], attrs['paddings'], attrs[ + 'dilations'] + + d_bolck_d = dilations[0] * (f_d - 1) + 1 + d_bolck_h = dilations[1] * (f_h - 1) + 1 + d_bolck_w = dilations[2] * (f_w - 1) + 1 + out_d = (in_d - 1) * stride[0] + d_bolck_d + out_h = (in_h - 1) * stride[1] + d_bolck_h + out_w = (in_w - 1) * stride[2] + d_bolck_w out = np.zeros((in_n, out_c, out_d, out_h, out_w)) for n in range(in_n): @@ -25,10 +30,11 @@ def conv3dtranspose_forward_naive(input_, filter_, conv3dtranspose_param): for k in range(out_c): tmp_out = np.sum(input_masked * filter_[:, k, :, :, :], axis=0) - d1, d2 = d * stride[0], d * stride[0] + f_d - i1, i2 = i * stride[1], i * stride[1] + f_h - j1, j2 = j * stride[2], j * stride[2] + f_w - out[n, k, d1:d2, i1:i2, j1:j2] += tmp_out + 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 out = out[:, :, pad[0]:out_d - pad[0], pad[1]:out_h - pad[1], pad[2]:out_w - pad[2]] @@ -41,18 +47,19 @@ class TestConv3dTransposeOp(OpTest): self.init_op_type() self.init_test_case() - conv3dtranspose_param = {'stride': self.stride, 'pad': self.pad} input_ = np.random.random(self.input_size).astype("float32") filter_ = np.random.random(self.filter_size).astype("float32") - output = conv3dtranspose_forward_naive( - input_, filter_, conv3dtranspose_param).astype("float32") self.inputs = {'Input': input_, 'Filter': filter_} self.attrs = { 'strides': self.stride, 'paddings': self.pad, - # 'dilations': self.dilations + 'dilations': self.dilations } + + output = conv3dtranspose_forward_naive(input_, filter_, + self.attrs).astype("float32") + self.outputs = {'Output': output} def test_check_output(self): @@ -108,11 +115,60 @@ class TestWithStride(TestConv3dTransposeOp): self.filter_size = [f_c, 6, 3, 3, 3] +class TestWithDilation(TestConv3dTransposeOp): + def init_test_case(self): + self.pad = [1, 1, 1] + self.stride = [1, 1, 1] + self.dilations = [2, 2, 2] + self.input_size = [2, 3, 5, 5, 5] # NCDHW + f_c = self.input_size[1] + self.filter_size = [f_c, 6, 3, 3, 3] + + # ------------ test_cudnn ------------ class TestCudnn(TestConv3dTransposeOp): def init_op_type(self): self.op_type = "conv3d_transpose_cudnn" +class TestCudnnWithPad(TestWithPad): + def init_test_case(self): + self.pad = [1, 1, 1] + self.stride = [1, 1, 1] + self.dilations = [1, 1, 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] + + def init_op_type(self): + self.op_type = "conv3d_transpose_cudnn" + + +class TestCudnnWithStride(TestWithStride): + def init_test_case(self): + self.pad = [1, 1, 1] + self.stride = [2, 2, 2] + self.dilations = [1, 1, 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] + + def init_op_type(self): + self.op_type = "conv3d_transpose_cudnn" + + +# #cudnn v5 does not support dilation conv. +# class TestCudnnWithDilation(TestWithDilation): +# def init_test_case(self): +# self.pad = [1, 1, 1] +# self.stride = [2, 2, 2] +# self.dilations = [2, 2, 2] +# self.input_size = [2, 3, 5, 5, 5] # NCDHW +# f_c = self.input_size[1] +# self.filter_size = [f_c, 6, 3, 3, 3] +# +# def init_op_type(self): +# self.op_type = "conv3d_transpose_cudnn" + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/v2/fluid/tests/test_crf_decoding_op.py b/python/paddle/v2/fluid/tests/test_crf_decoding_op.py index ee2b996bf430d5a0edaa0de459a937adffd9f8f6..ab573da31dfb9d7b40e44a79465a61cdc6b62a46 100644 --- a/python/paddle/v2/fluid/tests/test_crf_decoding_op.py +++ b/python/paddle/v2/fluid/tests/test_crf_decoding_op.py @@ -20,14 +20,14 @@ class CRFDecoding(object): self.w = transition_weights[2:, :] self.track = np.zeros( - (seq_start_positions[-1], self.tag_num), dtype="int32") + (seq_start_positions[-1], self.tag_num), dtype="int64") self.decoded_path = np.zeros( - (seq_start_positions[-1], 1), dtype="int32") + (seq_start_positions[-1], 1), dtype="int64") def _decode_one_sequence(self, decoded_path, x): seq_len, tag_num = x.shape alpha = np.zeros((seq_len, tag_num), dtype="float64") - track = np.zeros((seq_len, tag_num), dtype="int32") + track = np.zeros((seq_len, tag_num), dtype="int64") for i in range(tag_num): alpha[0, i] = self.a[i] + x[0, i] @@ -125,10 +125,10 @@ class TestCRFDecodingOp2(OpTest): axis=0) labels = np.random.randint( - low=0, high=TAG_NUM, size=(lod[-1][-1], 1), dtype="int32") + low=0, high=TAG_NUM, size=(lod[-1][-1], 1), dtype="int64") predicted_labels = np.ones( - (lod[-1][-1], 1), dtype="int32") * (TAG_NUM - 1) - expected_output = (labels == predicted_labels).astype("int32") + (lod[-1][-1], 1), dtype="int64") * (TAG_NUM - 1) + expected_output = (labels == predicted_labels).astype("int64") self.inputs = { "Emission": (emission, lod), diff --git a/python/paddle/v2/fluid/tests/test_data_feeder.py b/python/paddle/v2/fluid/tests/test_data_feeder.py new file mode 100644 index 0000000000000000000000000000000000000000..454969320321b72342803f507f0054f79f276669 --- /dev/null +++ b/python/paddle/v2/fluid/tests/test_data_feeder.py @@ -0,0 +1,13 @@ +import paddle.v2.fluid as fluid + + +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) + + +if __name__ == '__main__': + test_converter() diff --git a/python/paddle/v2/fluid/tests/test_detection_output_op.py b/python/paddle/v2/fluid/tests/test_detection_output_op.py new file mode 100644 index 0000000000000000000000000000000000000000..080a9743b0182cb7e6dd0030fc306a7f82510a05 --- /dev/null +++ b/python/paddle/v2/fluid/tests/test_detection_output_op.py @@ -0,0 +1,57 @@ +import unittest +import numpy as np +from op_test import OpTest + + +class TestUnpoolOp(OpTest): + def setUp(self): + self.op_type = "detection_output" + self.init_test_case() + + #loc.shape ((1, 4, 4, 1, 1)) + #conf.shape ((1, 4, 2, 1, 1)) + + loc = np.array([[[[[0.1]], [[0.1]], [[0.1]], [[0.1]]], + [[[0.1]], [[0.1]], [[0.1]], [[0.1]]], + [[[0.1]], [[0.1]], [[0.1]], [[0.1]]], + [[[0.1]], [[0.1]], [[0.1]], [[0.1]]]]]) + conf = np.array([[[[[0.1]], [[0.9]]], [[[0.2]], [[0.8]]], + [[[0.3]], [[0.7]]], [[[0.4]], [[0.6]]]]]) + priorbox = np.array([ + 0.1, 0.1, 0.5, 0.5, 0.1, 0.1, 0.2, 0.2, 0.2, 0.2, 0.6, 0.6, 0.1, + 0.1, 0.2, 0.2, 0.3, 0.3, 0.7, 0.7, 0.1, 0.1, 0.2, 0.2, 0.4, 0.4, + 0.8, 0.8, 0.1, 0.1, 0.2, 0.2 + ]) + + output = np.array([ + 0, 1, 0.68997443, 0.099959746, 0.099959746, 0.50804031, 0.50804031 + ]) + self.inputs = { + 'Loc': loc.astype('float32'), + 'Conf': conf.astype('float32'), + 'PriorBox': priorbox.astype('float32') + } + self.attrs = { + 'num_classes': self.num_classes, + 'top_k': self.top_k, + 'nms_top_k': self.nms_top_k, + 'background_label_id': self.background_label_id, + 'nms_threshold': self.nms_threshold, + 'confidence_threshold': self.confidence_threshold, + } + self.outputs = {'Out': output.astype('float32')} + + def test_check_output(self): + self.check_output() + + def init_test_case(self): + self.num_classes = 2 + self.top_k = 10 + self.nms_top_k = 20 + self.background_label_id = 0 + self.nms_threshold = 0.01 + self.confidence_threshold = 0.01 + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/v2/fluid/tests/test_dropout_op.py b/python/paddle/v2/fluid/tests/test_dropout_op.py index b14a366fcad7f4bf6968b6013c6cfbb57090071d..2483200212686caf9c46f9c1129b5d8ffdcc9145 100644 --- a/python/paddle/v2/fluid/tests/test_dropout_op.py +++ b/python/paddle/v2/fluid/tests/test_dropout_op.py @@ -7,7 +7,7 @@ class TestDropoutOp(OpTest): def setUp(self): self.op_type = "dropout" self.inputs = {'X': np.random.random((32, 64)).astype("float32")} - self.attrs = {'dropout_prob': 0.0, 'is_training': True} + self.attrs = {'dropout_prob': 0.0, 'is_test': False} self.outputs = { 'Out': self.inputs['X'], 'Mask': np.ones((32, 64)).astype('float32') @@ -24,7 +24,7 @@ class TestDropoutOp2(TestDropoutOp): def setUp(self): self.op_type = "dropout" self.inputs = {'X': np.random.random((32, 64)).astype("float32")} - self.attrs = {'dropout_prob': 1.0, 'is_training': True} + self.attrs = {'dropout_prob': 1.0, 'is_test': False} self.outputs = { 'Out': np.zeros((32, 64)).astype('float32'), 'Mask': np.zeros((32, 64)).astype('float32') @@ -35,7 +35,7 @@ class TestDropoutOp3(TestDropoutOp): def setUp(self): self.op_type = "dropout" self.inputs = {'X': np.random.random((32, 64, 2)).astype("float32")} - self.attrs = {'dropout_prob': 0.0, 'is_training': True} + self.attrs = {'dropout_prob': 0.0, 'is_test': False} self.outputs = { 'Out': self.inputs['X'], 'Mask': np.ones((32, 64, 2)).astype('float32') @@ -46,8 +46,10 @@ class TestDropoutOp4(OpTest): def setUp(self): self.op_type = "dropout" self.inputs = {'X': np.random.random((32, 64)).astype("float32")} - self.attrs = {'dropout_prob': 0.35, 'is_training': False} - self.outputs = {'Out': self.inputs['X'] * self.attrs['dropout_prob']} + self.attrs = {'dropout_prob': 0.35, 'is_test': True} + self.outputs = { + 'Out': self.inputs['X'] * (1.0 - self.attrs['dropout_prob']) + } def test_check_output(self): self.check_output() @@ -57,8 +59,10 @@ class TestDropoutOp5(OpTest): def setUp(self): self.op_type = "dropout" self.inputs = {'X': np.random.random((32, 64, 3)).astype("float32")} - self.attrs = {'dropout_prob': 0.75, 'is_training': False} - self.outputs = {'Out': self.inputs['X'] * self.attrs['dropout_prob']} + self.attrs = {'dropout_prob': 0.75, 'is_test': True} + self.outputs = { + 'Out': self.inputs['X'] * (1.0 - self.attrs['dropout_prob']) + } def test_check_output(self): self.check_output() diff --git a/python/paddle/v2/fluid/tests/test_dyn_rnn.py b/python/paddle/v2/fluid/tests/test_dyn_rnn.py new file mode 100644 index 0000000000000000000000000000000000000000..8090c5f47814c60034f2f46f00e56c530e0f2c19 --- /dev/null +++ b/python/paddle/v2/fluid/tests/test_dyn_rnn.py @@ -0,0 +1,129 @@ +import paddle.v2.fluid as fluid +import paddle.v2 as paddle +import unittest +import numpy + + +class TestDynRNN(unittest.TestCase): + def setUp(self): + self.word_dict = paddle.dataset.imdb.word_dict() + self.BATCH_SIZE = 2 + self.train_data = paddle.batch( + paddle.dataset.imdb.train(self.word_dict), + batch_size=self.BATCH_SIZE) + + def test_plain_while_op(self): + main_program = fluid.Program() + startup_program = fluid.Program() + + with fluid.program_guard(main_program, startup_program): + sentence = fluid.layers.data( + name='word', shape=[1], dtype='int64', lod_level=1) + sent_emb = fluid.layers.embedding( + input=sentence, size=[len(self.word_dict), 32], dtype='float32') + + label = fluid.layers.data(name='label', shape=[1], dtype='float32') + + rank_table = fluid.layers.lod_rank_table(x=sent_emb) + + sent_emb_array = fluid.layers.lod_tensor_to_array( + x=sent_emb, table=rank_table) + + seq_len = fluid.layers.max_sequence_len(rank_table=rank_table) + i = fluid.layers.fill_constant(shape=[1], dtype='int64', value=0) + i.stop_gradient = False + + boot_mem = fluid.layers.fill_constant_batch_size_like( + input=fluid.layers.array_read( + array=sent_emb_array, i=i), + value=0, + shape=[-1, 100], + dtype='float32') + boot_mem.stop_gradient = False + + mem_array = fluid.layers.array_write(x=boot_mem, i=i) + + cond = fluid.layers.less_than(x=i, y=seq_len) + cond.stop_gradient = False + while_op = fluid.layers.While(cond=cond) + out = fluid.layers.create_array(dtype='float32') + + with while_op.block(): + mem = fluid.layers.array_read(array=mem_array, i=i) + ipt = fluid.layers.array_read(array=sent_emb_array, i=i) + + mem = fluid.layers.shrink_memory(x=mem, i=i, table=rank_table) + + hidden = fluid.layers.fc(input=[mem, ipt], size=100, act='tanh') + + fluid.layers.array_write(x=hidden, i=i, array=out) + fluid.layers.increment(x=i, in_place=True) + fluid.layers.array_write(x=hidden, i=i, array=mem_array) + fluid.layers.less_than(x=i, y=seq_len, cond=cond) + + all_timesteps = fluid.layers.array_to_lod_tensor( + x=out, table=rank_table) + last = fluid.layers.sequence_last_step(input=all_timesteps) + logits = fluid.layers.fc(input=last, size=1, act=None) + loss = fluid.layers.sigmoid_cross_entropy_with_logits( + x=logits, label=label) + loss = fluid.layers.mean(x=loss) + sgd = fluid.optimizer.SGD(1e-4) + sgd.minimize(loss=loss) + cpu = fluid.CPUPlace() + exe = fluid.Executor(cpu) + exe.run(startup_program) + feeder = fluid.DataFeeder(feed_list=[sentence, label], place=cpu) + + data = next(self.train_data()) + val = exe.run(main_program, feed=feeder.feed(data), + fetch_list=[loss])[0] + self.assertEqual((1, ), val.shape) + print(val) + self.assertFalse(numpy.isnan(val)) + + def test_train_dyn_rnn(self): + main_program = fluid.Program() + startup_program = fluid.Program() + with fluid.program_guard(main_program, startup_program): + sentence = fluid.layers.data( + name='word', shape=[1], dtype='int64', lod_level=1) + sent_emb = fluid.layers.embedding( + input=sentence, size=[len(self.word_dict), 32], dtype='float32') + + rnn = fluid.layers.DynamicRNN() + + with rnn.block(): + in_ = rnn.step_input(sent_emb) + mem = rnn.memory(shape=[100], dtype='float32') + out_ = fluid.layers.fc(input=[in_, mem], size=100, act='tanh') + rnn.update_memory(mem, out_) + rnn.output(out_) + + last = fluid.layers.sequence_last_step(input=rnn()) + logits = fluid.layers.fc(input=last, size=1, act=None) + label = fluid.layers.data(name='label', shape=[1], dtype='float32') + loss = fluid.layers.sigmoid_cross_entropy_with_logits( + x=logits, label=label) + loss = fluid.layers.mean(x=loss) + sgd = fluid.optimizer.Adam(1e-3) + sgd.minimize(loss=loss) + + cpu = fluid.CPUPlace() + exe = fluid.Executor(cpu) + exe.run(startup_program) + feeder = fluid.DataFeeder(feed_list=[sentence, label], place=cpu) + data = next(self.train_data()) + loss_0 = exe.run(main_program, + feed=feeder.feed(data), + fetch_list=[loss])[0] + for _ in xrange(100): + val = exe.run(main_program, + feed=feeder.feed(data), + fetch_list=[loss])[0] + # loss should be small after 100 mini-batch + self.assertLess(val[0], loss_0[0]) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/v2/fluid/tests/test_dynamic_recurrent_op.py b/python/paddle/v2/fluid/tests/test_dynamic_recurrent_op.py deleted file mode 100644 index c2d8b48ea944ae40a451492b8e9fad38dda0835c..0000000000000000000000000000000000000000 --- a/python/paddle/v2/fluid/tests/test_dynamic_recurrent_op.py +++ /dev/null @@ -1,171 +0,0 @@ -import logging -import paddle.v2.fluid.core as core -import unittest -from paddle.v2.fluid.op import Operator, DynamicRecurrentOp -import numpy as np - -# for siplicity, just one level LoD -lod_py = [[0, 4, 7, 9, 10]] -input_dim = 30 -num_sents = len(lod_py[0]) - 1 -weight_dim = 15 - - -def create_tensor(scope, name, shape, np_data): - tensor = scope.var(name).get_tensor() - tensor.set_dims(shape) - tensor.set(np_data, core.CPUPlace()) - return tensor - - -class PyRNNStep(object): - def __init__(self): - - self.x = np.random.normal(size=(lod_py[0][-1], - input_dim)).astype("float32") - self.W = np.random.normal(size=(input_dim, input_dim)).astype("float32") - self.U = np.random.normal(size=(input_dim, input_dim)).astype("float32") - self.h_boot = np.random.normal(size=(num_sents, - input_dim)).astype("float32") - - -class DynamicRecurrentOpTest(unittest.TestCase): - ''' - Test RNNOp - - equation: - h_t = \sigma (W x_t + U h_{t-1}) - weights: - - W - - U - vars: - - x - states: - - h - outputs: - - h - ''' - - py = PyRNNStep() - - def forward(self): - self.scope = core.Scope() - self.create_global_variables() - self.create_rnn_op() - self.create_step_net() - ctx = core.DeviceContext.create(core.CPUPlace()) - self.rnnop.run(self.scope, ctx) - state = self.rnnop.get_state("h@state") - print 'state size: ', state.size() - - step_inputs = self.rnnop.get_step_input("x") - print "x size ", step_inputs.size() - for i in range(step_inputs.size()): - print "x %d" % i, np.array(step_inputs.read(i).get_dims()) - step_outputs = self.rnnop.get_step_output('h@state') - print 'step_outputs.size ', step_outputs.size() - output = self.scope.find_var("h@state").get_tensor() - print 'output', np.array(output).shape - - def create_global_variables(self): - # create inlink - x_tensor = create_tensor(self.scope, "x", [num_sents, input_dim], - self.py.x) - x_tensor.set_lod(lod_py) - create_tensor(self.scope, "W", [input_dim, input_dim], self.py.W) - create_tensor(self.scope, "U", [input_dim, input_dim], self.py.U) - create_tensor(self.scope, "h_boot", [num_sents, input_dim], - self.py.h_boot) - self.scope.var("step_scopes") - self.scope.var("h@state") - - def create_rnn_op(self): - # create RNNOp - self.rnnop = DynamicRecurrentOp( - # inputs - inputs=["x"], - initial_states=["h_boot"], - step_net="step_unit", - # outputs - outputs=["h@state"], - step_scopes="step_scopes", - # attributes - ex_states=["h@pre"], - states=["h@state"]) - - def create_step_net(self): - step_unit = core.Net.create() - x_fc_op = Operator("mul", X="x", Y="W", Out="Wx") - h_fc_op = Operator("mul", X="h@pre", Y="U", Out="Uh") - sum_op = Operator("sum", X=["Wx", "Uh"], Out="sum") - sig_op = Operator("sigmoid", X="sum", Y="h@state") - - for op in [x_fc_op, h_fc_op, sum_op, sig_op]: - step_unit.append_op(op) - step_unit.complete_add_op(True) - self.rnnop.set_step_unit(step_unit) - - def test_forward(self): - print 'test recurrent op forward' - pd_output = self.forward() - print 'pd_output', pd_output - - -class RecurrentGradientOpTest(unittest.TestCase): - py = PyRNNStep() - - def create_forward_op(self): - # create RNNOp - self.forward_op = DynamicRecurrentOp( - # inputs - inputs=["x"], - initial_states=["h_boot"], - step_net="step_unit", - # outputs - outputs=["h@state"], - step_scopes="step_scopes", - # attributes - ex_states=["h@pre"], - states=["h@state"]) - - def create_gradient_op(self): - a = set() - backward_op = core.DynamicRecurrentOp.backward(self.forward_op, a) - - def create_step_net(self): - step_unit = core.Net.create() - x_fc_op = Operator("mul", X="x", Y="W", Out="Wx") - h_fc_op = Operator("mul", X="h@pre", Y="U", Out="Uh") - sum_op = Operator("sum", X=["Wx", "Uh"], Out="sum") - sig_op = Operator("sigmoid", X="sum", Y="h@state") - - for op in [x_fc_op, h_fc_op, sum_op, sig_op]: - step_unit.append_op(op) - step_unit.complete_add_op(True) - self.forward_op.set_step_unit(step_unit) - - def create_global_variables(self): - # create inlink - x_tensor = create_tensor(self.scope, "x", [num_sents, input_dim], - self.py.x) - x_tensor.set_lod(lod_py) - create_tensor(self.scope, "W", [input_dim, input_dim], self.py.W) - create_tensor(self.scope, "U", [input_dim, input_dim], self.py.U) - create_tensor(self.scope, "h_boot", [num_sents, input_dim], - self.py.h_boot) - self.scope.var("step_scopes") - self.scope.var("h@state") - - def test_grad(self): - self.scope = core.Scope() - self.create_forward_op() - self.create_global_variables() - self.create_step_net() - self.create_gradient_op() - - -if __name__ == '__main__': - exit( - 0 - ) # FIXME(qijun): https://github.com/PaddlePaddle/Paddle/issues/5101#issuecomment-339814957 - unittest.main() diff --git a/python/paddle/v2/fluid/tests/test_dynrnn_gradient_check.py b/python/paddle/v2/fluid/tests/test_dynrnn_gradient_check.py new file mode 100644 index 0000000000000000000000000000000000000000..c02c59284e1ca2e28ba2f6c5ec13b241c15fc288 --- /dev/null +++ b/python/paddle/v2/fluid/tests/test_dynrnn_gradient_check.py @@ -0,0 +1,347 @@ +import numpy +import random +import collections +import paddle.v2.fluid as fluid +import unittest +from decorators import * + + +class Memory(object): + def __init__(self, shape, dtype='float32'): + self.ex = numpy.zeros(shape=shape, dtype=dtype) + self.cur = None + + def update(self, val): + assert val.shape == self.ex.shape + assert val.dtype == self.ex.dtype + self.cur = val + + def ex(self): + return self.ex + + def next(self): + self.ex = self.cur + self.cur = None + + def __next__(self): + self.next() + + def reset(self): + self.ex = numpy.zeros(shape=self.ex.shape, dtype=self.ex.dtype) + self.cur = None + + +class Output(object): + def __init__(self): + self.outs = [] + + def next_sequence(self): + self.outs.append([]) + + def out(self, val): + self.outs[-1].append(val) + + def last(self): + return self.outs[-1][-1] + + +class BaseRNN(object): + def __init__(self, ins, mems, params, outs, num_seq=5, max_seq_len=15): + self.num_seq = num_seq + self.inputs = collections.defaultdict(list) + + for _ in xrange(num_seq): + seq_len = random.randint(1, max_seq_len - 1) + for iname in ins: + ishape = ins[iname].get('shape', None) + idtype = ins[iname].get('dtype', 'float32') + lst = [] + for _ in xrange(seq_len): + lst.append(numpy.random.random(size=ishape).astype(idtype)) + self.inputs[iname].append(lst) + + self.mems = dict() + for mname in mems: + mshape = mems[mname].get('shape', None) + mdtype = mems[mname].get('dtype', 'float32') + self.mems[mname] = Memory(shape=mshape, dtype=mdtype) + + self.params = dict() + for pname in params: + pshape = params[pname].get('shape', None) + pdtype = params[pname].get('dtype', 'float32') + self.params[pname] = numpy.random.random(size=pshape).astype(pdtype) + + self.outputs = dict() + + for oname in outs: + self.outputs[oname] = Output() + + def step(self, **kwargs): + raise NotImplementedError() + + def exe(self): + retv = dict() + for out in self.outputs: + retv[out] = [] + + for seq_id in xrange(self.num_seq): + for mname in self.mems: + self.mems[mname].reset() + for out in self.outputs: + self.outputs[out].next_sequence() + + iname0 = self.inputs.keys()[0] + seq_len = len(self.inputs[iname0][seq_id]) + + for step_id in xrange(seq_len): + xargs = dict() + + for iname in self.inputs: + xargs[iname] = self.inputs[iname][seq_id][step_id] + + for mname in self.mems: + xargs[mname] = self.mems[mname] + + for pname in self.params: + xargs[pname] = self.params[pname] + + for out in self.outputs: + xargs[out] = self.outputs[out] + + self.step(**xargs) + + for mname in self.mems: + next(self.mems[mname]) + + for out in self.outputs: + retv[out].append(self.outputs[out].last()) + + for out in retv: + retv[out] = numpy.array(retv[out]) + return retv + + def to_feed(self, place): + feed_dict = dict() + + for iname in self.inputs: + lod = [0] + np_flatten = [] + for seq_id in xrange(len(self.inputs[iname])): + seq_len = len(self.inputs[iname][seq_id]) + lod.append(lod[-1] + seq_len) + np_flatten.extend(self.inputs[iname][seq_id]) + + t = fluid.Tensor() + t.set(numpy.array(np_flatten), place) + t.set_lod([lod]) + feed_dict[iname] = t + + for pname in self.params: + feed_dict[pname] = self.params[pname] + return feed_dict + + def get_numeric_gradient_of_param(self, param_name, delta=0.001): + p = self.params[param_name] + if len(p.shape) != 2: + raise ValueError("Not support get numeric gradient of an parameter," + " which is not matrix") + g = numpy.zeros(shape=p.shape, dtype=p.dtype) + + for i in xrange(p.shape[0]): + for j in xrange(p.shape[1]): + o = p[i][j] + p[i][j] += delta + pos = self._exe_mean_out_() + p[i][j] -= 2 * delta + neg = self._exe_mean_out_() + p[i][j] = o + g[i][j] = (pos - neg) / (delta * 2) + return g + + def get_numeric_gradient_of_input(self, + input_name, + delta=0.001, + return_one_tensor=True): + ipt = self.inputs[input_name] + grad = [] + + for seq in ipt: + seq_grad = [] + for item in seq: + item_grad = numpy.zeros(shape=item.shape, dtype=item.dtype) + if len(item.shape) != 1: + raise ValueError("Not support") + + for i in xrange(len(item)): + o = item[i] + item[i] += delta + pos = self._exe_mean_out_() + item[i] -= 2 * delta + neg = self._exe_mean_out_() + item[i] = o + item_grad[i] = (pos - neg) / (delta * 2) + seq_grad.append(item_grad) + grad.append(seq_grad) + + if not return_one_tensor: + return grad + + for i in xrange(len(grad)): + grad[i] = numpy.concatenate(grad[i]) + grad = numpy.concatenate(grad) + return grad + + def _exe_mean_out_(self): + outs = self.exe() + return numpy.array([o.mean() for o in outs.itervalues()]).mean() + + +class TestSimpleMul(unittest.TestCase): + DATA_NAME = 'X' + DATA_WIDTH = 32 + PARAM_NAME = 'W' + HIDDEN_WIDTH = 10 + OUT_NAME = 'Out' + + class SimpleMul(BaseRNN): + def __init__(self): + base = TestSimpleMul + super(base.SimpleMul, self).__init__({ + base.DATA_NAME: { + 'shape': [base.DATA_WIDTH] + } + }, {}, { + base.PARAM_NAME: { + 'shape': [base.DATA_WIDTH, base.HIDDEN_WIDTH] + } + }, [base.OUT_NAME]) + + def step(self, X, W, Out): + Out.out(numpy.matmul(X, W)) + + # Test many times in local to ensure the random seed cannot breaks CI + # @many_times(10) + @prog_scope() + def test_forward_backward(self): + py_rnn = TestSimpleMul.SimpleMul() + dat = fluid.layers.data( + name=self.DATA_NAME, shape=[self.DATA_WIDTH], lod_level=1) + dat.stop_gradient = False + + rnn = fluid.layers.DynamicRNN() + with rnn.block(): + d = rnn.step_input(dat) + o = fluid.layers.fc(input=d, + param_attr=self.PARAM_NAME, + bias_attr=False, + size=self.HIDDEN_WIDTH, + act=None) + rnn.output(o) + + out = rnn() + out = fluid.layers.sequence_pool(out, pool_type='last') + loss = fluid.layers.mean(x=out) + fluid.backward.append_backward(loss) + + cpu = fluid.CPUPlace() + exe = fluid.Executor(cpu) + out, w_g, i_g = map(numpy.array, + exe.run(feed=py_rnn.to_feed(cpu), + fetch_list=[ + out, self.PARAM_NAME + "@GRAD", + self.DATA_NAME + "@GRAD" + ], + return_numpy=False)) + out_by_python = py_rnn.exe()[self.OUT_NAME] + self.assertTrue(numpy.allclose(out, out_by_python)) + w_g_num = py_rnn.get_numeric_gradient_of_param(self.PARAM_NAME) + self.assertTrue(numpy.allclose(w_g_num, w_g, rtol=0.05)) + i_g_num = py_rnn.get_numeric_gradient_of_input( + input_name=self.DATA_NAME) + i_g_num = i_g_num.reshape(i_g.shape) + self.assertTrue(numpy.allclose(i_g_num, i_g, rtol=0.05)) + + +class TestSimpleMulWithMemory(unittest.TestCase): + DATA_WIDTH = 32 + HIDDEN_WIDTH = 20 + DATA_NAME = 'X' + PARAM_NAME = 'W' + + class SimpleMulWithMemory(BaseRNN): + def __init__(self): + super(TestSimpleMulWithMemory.SimpleMulWithMemory, self).__init__({ + TestSimpleMulWithMemory.DATA_NAME: { + 'shape': [TestSimpleMulWithMemory.DATA_WIDTH] + } + }, {'Mem': { + 'shape': [TestSimpleMulWithMemory.HIDDEN_WIDTH] + }}, { + TestSimpleMulWithMemory.PARAM_NAME: { + 'shape': [ + TestSimpleMulWithMemory.DATA_WIDTH, + TestSimpleMulWithMemory.HIDDEN_WIDTH + ] + } + }, ['Out']) + + def step(self, X, Mem, W, Out): + o = numpy.matmul(X, W) + assert isinstance(Mem, Memory) + o += Mem.ex + Mem.update(o) + assert isinstance(Out, Output) + Out.out(o) + + # many_times used locally for debug. Make sure the calculation is stable. + # @many_times(10) + @prog_scope() + def test_forward_backward(self): + py_rnn = TestSimpleMulWithMemory.SimpleMulWithMemory() + data = fluid.layers.data( + name=self.DATA_NAME, shape=[self.DATA_WIDTH], lod_level=1) + data.stop_gradient = False + rnn = fluid.layers.DynamicRNN() + with rnn.block(): + d = rnn.step_input(data) + mem = rnn.memory(value=0.0, shape=[self.HIDDEN_WIDTH]) + hidden = fluid.layers.fc(input=d, + size=self.HIDDEN_WIDTH, + param_attr=self.PARAM_NAME, + bias_attr=False, + act=None) + o = fluid.layers.elementwise_add(x=hidden, y=mem) + rnn.update_memory(mem, o) + rnn.output(o) + + out = rnn() + last = fluid.layers.sequence_pool(input=out, pool_type='last') + loss = fluid.layers.mean(x=last) + fluid.backward.append_backward(loss) + + cpu = fluid.CPUPlace() + exe = fluid.Executor(cpu) + feed = py_rnn.to_feed(cpu) + last_np, w_g, i_g = map(numpy.array, + exe.run(feed=feed, + fetch_list=[ + last, self.PARAM_NAME + "@GRAD", + self.DATA_NAME + "@GRAD" + ], + return_numpy=False)) + last_by_py, = py_rnn.exe().values() + w_g_num = py_rnn.get_numeric_gradient_of_param(self.PARAM_NAME) + self.assertTrue(numpy.allclose(last_np, last_by_py)) + + self.assertTrue(numpy.allclose(w_g_num, w_g, rtol=0.1)) + i_g_num = py_rnn.get_numeric_gradient_of_input(self.DATA_NAME) + i_g_num = i_g_num.reshape(i_g.shape) + + # Since this RNN has many float add. The number could be not stable. + # rtol = 0.1 + self.assertTrue(numpy.allclose(i_g_num, i_g, rtol=0.1)) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/v2/fluid/tests/test_executor_and_mul.py b/python/paddle/v2/fluid/tests/test_executor_and_mul.py index 709250d0c86dde84ac22c37d8e2385ca4a80a40a..b1ef87c5cb1711c419b401c5950839816f7f4160 100644 --- a/python/paddle/v2/fluid/tests/test_executor_and_mul.py +++ b/python/paddle/v2/fluid/tests/test_executor_and_mul.py @@ -1,33 +1,27 @@ import unittest -from paddle.v2.fluid.layers import mul, data + +import numpy import paddle.v2.fluid.core as core + from paddle.v2.fluid.executor import Executor -from paddle.v2.fluid.framework import g_main_program -import numpy +from paddle.v2.fluid.layers import mul, data class TestExecutor(unittest.TestCase): def test_mul(self): - a = data(name='a', shape=[784], data_type='float32') + a = data(name='a', shape=[784], dtype='float32') b = data( name='b', shape=[784, 100], - data_type='float32', + dtype='float32', append_batch_size=False) out = mul(x=a, y=b) place = core.CPUPlace() a_np = numpy.random.random((100, 784)).astype('float32') - tensor_a = core.LoDTensor() - tensor_a.set(a_np, place) b_np = numpy.random.random((784, 100)).astype('float32') - tensor_b = core.LoDTensor() - tensor_b.set(b_np, place) exe = Executor(place) - outs = exe.run(g_main_program, - feed={'a': tensor_a, - 'b': tensor_b}, - fetch_list=[out]) - out = numpy.array(outs[0]) + outs = exe.run(feed={'a': a_np, 'b': b_np}, fetch_list=[out]) + out = outs[0] self.assertEqual((100, 100), out.shape) self.assertTrue(numpy.allclose(out, numpy.dot(a_np, b_np))) diff --git a/python/paddle/v2/fluid/tests/test_fill_op.py b/python/paddle/v2/fluid/tests/test_fill_op.py new file mode 100644 index 0000000000000000000000000000000000000000..88337598c895a5a663ef45fd0800fa950fee1253 --- /dev/null +++ b/python/paddle/v2/fluid/tests/test_fill_op.py @@ -0,0 +1,24 @@ +import unittest +import numpy as np +from op_test import OpTest +import paddle.v2.fluid.core as core + + +class TestFillOp(OpTest): + def setUp(self): + self.op_type = "fill" + val = np.random.random(size=[100, 200]) + self.inputs = {} + self.attrs = { + 'value': val.flatten().tolist(), + 'shape': [100, 200], + 'dtype': int(core.DataType.FP64) + } + self.outputs = {'Out': val.astype('float64')} + + def test_check_output(self): + self.check_output() + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/v2/fluid/tests/test_fill_zeros_like_op.py b/python/paddle/v2/fluid/tests/test_fill_zeros_like_op.py index eff8fa87d9c0dafc6935604101e94ee6c8b081ce..cd91769a22f8d6af193efabd8d997913676fbba6 100644 --- a/python/paddle/v2/fluid/tests/test_fill_zeros_like_op.py +++ b/python/paddle/v2/fluid/tests/test_fill_zeros_like_op.py @@ -7,7 +7,7 @@ class TestFillZerosLikeOp(OpTest): def setUp(self): self.op_type = "fill_zeros_like" self.inputs = {'X': np.random.random((219, 232)).astype("float32")} - self.outputs = {'Y': np.zeros_like(self.inputs["X"])} + self.outputs = {'Out': np.zeros_like(self.inputs["X"])} def test_check_output(self): self.check_output() diff --git a/python/paddle/v2/fluid/tests/test_ftrl_op.py b/python/paddle/v2/fluid/tests/test_ftrl_op.py new file mode 100644 index 0000000000000000000000000000000000000000..f77ac4659a9b877829f7ae52dd005d9dd11dac07 --- /dev/null +++ b/python/paddle/v2/fluid/tests/test_ftrl_op.py @@ -0,0 +1,62 @@ +import unittest +import numpy as np +from op_test import OpTest + + +class TestFTRLOp(OpTest): + def setUp(self): + self.op_type = "ftrl" + w = np.random.random((102, 105)).astype("float32") + g = np.random.random((102, 105)).astype("float32") + sq_accum = np.full((102, 105), 0.1).astype("float32") + linear_accum = np.full((102, 105), 0.1).astype("float32") + lr = np.array([0.01]).astype("float32") + l1 = 0.1 + l2 = 0.2 + lr_power = -0.5 + + self.inputs = { + 'Param': w, + 'SquaredAccumulator': sq_accum, + 'LinearAccumulator': linear_accum, + 'Grad': g, + 'LearningRate': lr + } + self.attrs = { + 'l1': l1, + 'l2': l2, + 'lr_power': lr_power, + 'learning_rate': lr + } + new_accum = sq_accum + g * g + if lr_power == -0.5: + linear_out = linear_accum + g - ( + (np.sqrt(new_accum) - np.sqrt(sq_accum)) / lr) * w + else: + linear_out = linear_accum + g - ((np.power( + new_accum, -lr_power) - np.power(sq_accum, -lr_power)) / lr) * w + + x = (l1 * np.sign(linear_out) - linear_out) + if lr_power == -0.5: + y = (np.sqrt(new_accum) / lr) + (2 * l2) + pre_shrink = x / y + param_out = np.where(np.abs(linear_out) > l1, pre_shrink, 0.0) + else: + y = (np.power(new_accum, -lr_power) / lr) + (2 * l2) + pre_shrink = x / y + param_out = np.where(np.abs(linear_out) > l1, pre_shrink, 0.0) + + sq_accum_out = sq_accum + g * g + + self.outputs = { + 'ParamOut': param_out, + 'SquaredAccumOut': sq_accum_out, + 'LinearAccumOut': linear_out + } + + def test_check_output(self): + self.check_output() + + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/v2/fluid/tests/test_gaussian_random_op.py b/python/paddle/v2/fluid/tests/test_gaussian_random_op.py index 627ab4e23562f14538d85f2e21edeb7d72d940bb..6f6a60ccb3ff17f6a12eec6974b8b2d73885c29f 100644 --- a/python/paddle/v2/fluid/tests/test_gaussian_random_op.py +++ b/python/paddle/v2/fluid/tests/test_gaussian_random_op.py @@ -1,32 +1,46 @@ import unittest +import numpy + +import paddle.v2.fluid as fluid import paddle.v2.fluid.core as core from paddle.v2.fluid.op import Operator -import numpy +from paddle.v2.fluid.executor import Executor class TestGaussianRandomOp(unittest.TestCase): + def setUp(self): + self.op_type = "gaussian_random" + self.inputs = {} + self.attrs = {"shape": [1000, 784], "mean": .0, "std": 1., "seed": 10} + + self.outputs = ["Out"] + def test_cpu(self): - self.gaussian_random_test(place=core.CPUPlace()) + self.gaussian_random_test(place=fluid.CPUPlace()) def test_gpu(self): if core.is_compile_gpu(): - self.gaussian_random_test(place=core.GPUPlace(0)) + self.gaussian_random_test(place=fluid.CUDAPlace(0)) def gaussian_random_test(self, place): - scope = core.Scope() - scope.var('Out').get_tensor() - - op = Operator( - "gaussian_random", - Out='Out', - shape=[1000, 784], - mean=.0, - std=1., - seed=10) - - context = core.DeviceContext.create(place) - op.run(scope, context) - tensor = numpy.array(scope.find_var('Out').get_tensor()) + + program = fluid.Program() + block = program.global_block() + vout = block.create_var(name="Out") + op = block.append_op( + type=self.op_type, outputs={"Out": vout}, attrs=self.attrs) + + op.desc.infer_var_type(block.desc) + op.desc.infer_shape(block.desc) + + fetch_list = [] + for var_name in self.outputs: + fetch_list.append(block.var(var_name)) + + exe = Executor(place) + outs = exe.run(program, fetch_list=fetch_list) + tensor = outs[0] + self.assertAlmostEqual(numpy.mean(tensor), .0, delta=0.1) self.assertAlmostEqual(numpy.std(tensor), 1., delta=0.1) diff --git a/python/paddle/v2/fluid/tests/test_gru_unit_op.py b/python/paddle/v2/fluid/tests/test_gru_unit_op.py index f356f6e9ec0da2d3e1fb67638d81e8d54c544f53..501d5aa5797d6def708338692f0861657f951ef7 100644 --- a/python/paddle/v2/fluid/tests/test_gru_unit_op.py +++ b/python/paddle/v2/fluid/tests/test_gru_unit_op.py @@ -28,8 +28,8 @@ def relu(x): class TestGRUUnitOp(OpTest): - batch_size = 3 - frame_size = 5 + batch_size = 5 + frame_size = 10 activate = { GRUActivationType.identity: identity, GRUActivationType.sigmoid: sigmoid, @@ -77,7 +77,7 @@ class TestGRUUnitOp(OpTest): c = self.activate[self.attrs['activation']](np.dot(r_h_p, w_c) + g[:, frame_size * 2:]) g = np.hstack((u_r, c)) - h = u * h_p + (1 - u) * c + h = u * c + (1 - u) * h_p self.outputs = { 'Gate': g.astype('float64'), 'ResetHiddenPrev': r_h_p.astype('float64'), @@ -92,10 +92,7 @@ class TestGRUUnitOp(OpTest): self.check_output() def test_check_grad(self): - self.check_grad( - ['Input', 'HiddenPrev', 'Weight'], - ['Hidden', 'ResetHiddenPrev', 'Gate'], - max_relative_error=0.007) + self.check_grad(['Input', 'HiddenPrev', 'Weight'], ['Hidden']) class TestGRUUnitOpWithBias(TestGRUUnitOp): @@ -104,18 +101,20 @@ class TestGRUUnitOpWithBias(TestGRUUnitOp): frame_size = self.frame_size super(TestGRUUnitOpWithBias, self).set_inputs() self.inputs['Bias'] = np.random.uniform( - -0.1, 0.1, (1, frame_size * 3)).astype('float32') + -0.1, 0.1, (1, frame_size * 3)).astype('float64') self.attrs = { 'activation': GRUActivationType.identity, 'gate_activation': GRUActivationType.sigmoid } def test_check_grad(self): + self.check_grad(['Input', 'HiddenPrev', 'Weight', 'Bias'], ['Hidden']) + + def test_check_grad_ingore_input(self): self.check_grad( - ['Input', 'HiddenPrev', 'Weight', 'Bias'], ['Hidden'], - max_relative_error=0.007) + ['HiddenPrev', 'Weight', 'Bias'], ['Hidden'], + no_grad_set=set('Input')) if __name__ == '__main__': - exit(0) # FIXME(yuyang18): This unittest is not pass. Fix it later unittest.main() diff --git a/python/paddle/v2/fluid/tests/test_hinge_loss_op.py b/python/paddle/v2/fluid/tests/test_hinge_loss_op.py new file mode 100644 index 0000000000000000000000000000000000000000..a8757a891faa01413dc6858451f1a988a3e030b5 --- /dev/null +++ b/python/paddle/v2/fluid/tests/test_hinge_loss_op.py @@ -0,0 +1,28 @@ +import unittest +import numpy as np +from op_test import OpTest + + +class TestHingeLossOp(OpTest): + def setUp(self): + self.op_type = 'hinge_loss' + samples_num = 64 + logits = np.random.uniform(-10, 10, (samples_num, 1)).astype('float32') + labels = np.random.randint(0, 2, (samples_num, 1)).astype('float32') + + self.inputs = { + 'Logits': logits, + 'Labels': labels, + } + loss = np.maximum(1.0 - (2 * labels - 1) * logits, 0) + self.outputs = {'Loss': loss} + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad(['Logits'], 'Loss', max_relative_error=0.008) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/v2/fluid/tests/test_image_classification_layer.py b/python/paddle/v2/fluid/tests/test_image_classification_layer.py index bf5444107fa1609e67b09823b82e5fb92234b0a4..b621d1525e33693869e24e2bb233bc8e257b077f 100644 --- a/python/paddle/v2/fluid/tests/test_image_classification_layer.py +++ b/python/paddle/v2/fluid/tests/test_image_classification_layer.py @@ -1,16 +1,11 @@ import unittest -import paddle.v2.fluid.layers as layers +import paddle.v2.fluid as fluid import paddle.v2.fluid.nets as nets from paddle.v2.fluid.framework import Program -def conv_block(input, - num_filter, - groups, - dropouts, - main_program=None, - startup_program=None): +def conv_block(input, num_filter, groups, dropouts): return nets.img_conv_group( input=input, pool_size=2, @@ -20,82 +15,54 @@ def conv_block(input, conv_act='relu', conv_with_batchnorm=True, conv_batchnorm_drop_rate=dropouts, - pool_type='max', - main_program=main_program, - startup_program=startup_program) + pool_type='max') class TestLayer(unittest.TestCase): def test_batch_norm_layer(self): main_program = Program() startup_program = Program() - images = layers.data( - name='pixel', - shape=[3, 48, 48], - data_type='float32', - main_program=main_program) - layers.batch_norm( - input=images, - main_program=main_program, - startup_program=startup_program) + with fluid.program_guard(main_program, startup_program): + images = fluid.layers.data( + name='pixel', shape=[3, 48, 48], dtype='float32') + hidden1 = fluid.layers.batch_norm(input=images) + hidden2 = fluid.layers.fc(input=hidden1, size=128, act='relu') + fluid.layers.batch_norm(input=hidden2) - # print str(main_program) + print str(main_program) def test_dropout_layer(self): main_program = Program() startup_program = Program() - images = layers.data( - name='pixel', - shape=[3, 48, 48], - data_type='float32', - main_program=main_program) - layers.dropout( - x=images, - dropout_prob=0.5, - main_program=main_program, - startup_program=startup_program) + with fluid.program_guard(main_program, startup_program): + images = fluid.layers.data( + name='pixel', shape=[3, 48, 48], dtype='float32') + fluid.layers.dropout(x=images, dropout_prob=0.5) - # print str(main_program) + print str(main_program) def test_img_conv_group(self): main_program = Program() startup_program = Program() - images = layers.data( - name='pixel', - shape=[3, 48, 48], - data_type='float32', - main_program=main_program, - startup_program=startup_program) - conv1 = conv_block(images, 64, 2, [0.3, 0], main_program, - startup_program) - conv2 = conv_block(conv1, 256, 3, [0.4, 0.4, 0], main_program, - startup_program) + with fluid.program_guard(main_program, startup_program): + images = fluid.layers.data( + name='pixel', shape=[3, 48, 48], dtype='float32') + conv1 = conv_block(images, 64, 2, [0.3, 0]) + conv_block(conv1, 256, 3, [0.4, 0.4, 0]) - # print str(main_program) + print str(main_program) def test_elementwise_add_with_act(self): main_program = Program() startup_program = Program() - image1 = layers.data( - name='pixel1', - shape=[3, 48, 48], - data_type='float32', - main_program=main_program, - startup_program=startup_program) - image2 = layers.data( - name='pixel2', - shape=[3, 48, 48], - data_type='float32', - main_program=main_program, - startup_program=startup_program) - out = layers.elementwise_add( - x=image1, - y=image2, - act='relu', - main_program=main_program, - startup_program=startup_program) - # print(main_program) + with fluid.program_guard(main_program, startup_program): + image1 = fluid.layers.data( + name='pixel1', shape=[3, 48, 48], dtype='float32') + image2 = fluid.layers.data( + name='pixel2', shape=[3, 48, 48], dtype='float32') + fluid.layers.elementwise_add(x=image1, y=image2, act='relu') + print(main_program) if __name__ == '__main__': diff --git a/python/paddle/v2/fluid/tests/test_inference_model_io.py b/python/paddle/v2/fluid/tests/test_inference_model_io.py index 98b95713b73e8eba93bd6a58eaaed603cfae7952..71ca3e6c105c4437470f8e9f596e723d879b65e4 100644 --- a/python/paddle/v2/fluid/tests/test_inference_model_io.py +++ b/python/paddle/v2/fluid/tests/test_inference_model_io.py @@ -1,13 +1,13 @@ -import paddle.v2 as paddle -import paddle.v2.fluid.layers as layers +import unittest + +import numpy as np import paddle.v2.fluid.core as core -import paddle.v2.fluid.optimizer as optimizer -from paddle.v2.fluid.framework import Program -from paddle.v2.fluid.io import save_inference_model, load_inference_model import paddle.v2.fluid.executor as executor -import unittest -import numpy as np +import paddle.v2.fluid.layers as layers +import paddle.v2.fluid.optimizer as optimizer +from paddle.v2.fluid.framework import Program, program_guard +from paddle.v2.fluid.io import save_inference_model, load_inference_model class TestBook(unittest.TestCase): @@ -16,35 +16,18 @@ class TestBook(unittest.TestCase): init_program = Program() program = Program() - x = layers.data( - name='x', - shape=[2], - data_type='float32', - main_program=program, - startup_program=init_program) - y = layers.data( - name='y', - shape=[1], - data_type='float32', - main_program=program, - startup_program=init_program) - - y_predict = layers.fc(input=x, - size=1, - act=None, - main_program=program, - startup_program=init_program) - - cost = layers.square_error_cost( - input=y_predict, - label=y, - main_program=program, - startup_program=init_program) - avg_cost = layers.mean( - x=cost, main_program=program, startup_program=init_program) - - sgd_optimizer = optimizer.SGDOptimizer(learning_rate=0.001) - opts = sgd_optimizer.minimize(avg_cost, init_program) + + with program_guard(program, init_program): + x = layers.data(name='x', shape=[2], dtype='float32') + y = layers.data(name='y', shape=[1], dtype='float32') + + y_predict = layers.fc(input=x, size=1, act=None) + + cost = layers.square_error_cost(input=y_predict, label=y) + avg_cost = layers.mean(x=cost) + + sgd_optimizer = optimizer.SGDOptimizer(learning_rate=0.001) + sgd_optimizer.minimize(avg_cost, init_program) place = core.CPUPlace() exe = executor.Executor(place) @@ -52,25 +35,20 @@ class TestBook(unittest.TestCase): exe.run(init_program, feed={}, fetch_list=[]) for i in xrange(100): - x_data = np.array( + tensor_x = np.array( [[1, 1], [1, 2], [3, 4], [5, 2]]).astype("float32") - y_data = np.array([[-2], [-3], [-7], [-7]]).astype("float32") + tensor_y = np.array([[-2], [-3], [-7], [-7]]).astype("float32") - tensor_x = core.LoDTensor() - tensor_x.set(x_data, place) - tensor_y = core.LoDTensor() - tensor_y.set(y_data, place) exe.run(program, feed={'x': tensor_x, 'y': tensor_y}, fetch_list=[avg_cost]) save_inference_model(MODEL_DIR, ["x", "y"], [avg_cost], exe, program) - outs = exe.run(program, - feed={'x': tensor_x, - 'y': tensor_y}, - fetch_list=[avg_cost]) - expected = np.array(outs[0]) + expected = exe.run(program, + feed={'x': tensor_x, + 'y': tensor_y}, + fetch_list=[avg_cost])[0] reload(executor) # reload to build a new scope exe = executor.Executor(place) @@ -83,7 +61,7 @@ class TestBook(unittest.TestCase): feed={feed_var_names[0]: tensor_x, feed_var_names[1]: tensor_y}, fetch_list=fetch_vars) - actual = np.array(outs[0]) + actual = outs[0] self.assertEqual(feed_var_names, ["x", "y"]) self.assertEqual(len(fetch_vars), 1) diff --git a/python/paddle/v2/fluid/tests/test_initializer.py b/python/paddle/v2/fluid/tests/test_initializer.py index f2eb79b209627f5814847db6d96c0a17300d9b5a..3175010f48229d04421fc0068af0f0ed90e63af4 100644 --- a/python/paddle/v2/fluid/tests/test_initializer.py +++ b/python/paddle/v2/fluid/tests/test_initializer.py @@ -60,6 +60,29 @@ class TestUniformInitializer(unittest.TestCase): self.assertAlmostEqual(init_op.attr('max'), 1.0, delta=DELTA) self.assertEqual(init_op.attr('seed'), 0) + def test_uniform_initializer_random_seed(self): + """Test the uniform initializer with manually setting seed + """ + program = framework.Program() + program.random_seed = 123 + block = program.global_block() + block.create_parameter( + dtype="float32", + shape=[5, 10], + lod_level=0, + name="param", + initializer=initializer.UniformInitializer()) + block.create_parameter( + dtype="float32", + shape=[5, 10], + lod_level=0, + name="param", + initializer=initializer.UniformInitializer(seed=456)) + init_op = block.ops[1] + self.assertEqual(init_op.attr("seed"), 123) + init_op1 = block.ops[0] + self.assertEqual(init_op1.attr("seed"), 456) + def test_uniform_initializer(self): """Test uniform initializer with supplied attributes """ @@ -223,5 +246,109 @@ class TestXavierInitializer(unittest.TestCase): self.assertEqual(init_op.attr('seed'), 134) +class TestMSRAInitializer(unittest.TestCase): + def test_uniform_msra_initializer(self): + """Test MSRA initializer with uniform distribution on + for matrix multiply. + """ + program = framework.Program() + block = program.global_block() + param = block.create_parameter( + dtype="float32", + shape=[5, 10], + lod_level=0, + name="param", + initializer=initializer.MSRAInitializer()) + self.assertEqual(len(block.ops), 1) + init_op = block.ops[0] + self.assertEqual(init_op.type, 'uniform_random') + limit = np.sqrt(6.0 / param.shape[0]) + self.assertAlmostEqual(init_op.attr('min'), -limit, delta=DELTA) + self.assertAlmostEqual(init_op.attr('max'), limit, delta=DELTA) + self.assertEqual(init_op.attr('seed'), 0) + + def test_uniform_msra_initializer_conv(self): + """Test MSRA initializer with uniform distribution on + for convolutions. + """ + program = framework.Program() + block = program.global_block() + param = block.create_parameter( + dtype="float32", + shape=[5, 10, 15, 20], + lod_level=0, + name="param", + initializer=initializer.MSRAInitializer()) + self.assertEqual(len(block.ops), 1) + init_op = block.ops[0] + self.assertEqual(init_op.type, 'uniform_random') + receptive_field_size = float(15 * 20) + limit = np.sqrt(6.0 / (param.shape[1] * receptive_field_size)) + self.assertAlmostEqual(init_op.attr('min'), -limit, delta=DELTA) + self.assertAlmostEqual(init_op.attr('max'), limit, delta=DELTA) + self.assertEqual(init_op.attr('seed'), 0) + + def test_normal_msra_initializer(self): + """Test MSRA initializer with normal distribution on + for matrix multiply. + """ + program = framework.Program() + block = program.global_block() + param = block.create_parameter( + dtype="float32", + shape=[5, 10], + lod_level=0, + name="param", + initializer=initializer.MSRAInitializer(uniform=False)) + self.assertEqual(len(block.ops), 1) + init_op = block.ops[0] + self.assertEqual(init_op.type, 'gaussian_random') + std = np.sqrt(2.0 / param.shape[0]) + self.assertAlmostEqual(init_op.attr('mean'), 0.0, delta=DELTA) + self.assertAlmostEqual(init_op.attr('std'), std, delta=DELTA) + self.assertEqual(init_op.attr('seed'), 0) + + def test_normal_msra_initializer_conv(self): + """Test MSRA initializer with normal distribution on + for convolutions. + """ + program = framework.Program() + block = program.global_block() + param = block.create_parameter( + dtype="float32", + shape=[5, 10, 15, 20], + lod_level=0, + name="param", + initializer=initializer.MSRAInitializer(uniform=False)) + self.assertEqual(len(block.ops), 1) + init_op = block.ops[0] + self.assertEqual(init_op.type, 'gaussian_random') + receptive_field_size = float(15 * 20) + std = np.sqrt(2.0 / (param.shape[1] * receptive_field_size)) + self.assertAlmostEqual(init_op.attr('mean'), 0.0, delta=DELTA) + self.assertAlmostEqual(init_op.attr('std'), std, delta=DELTA) + self.assertEqual(init_op.attr('seed'), 0) + + def test_msra_initializer_supplied_arguments(self): + """Test the MSRA initializer with supplied arguments + """ + program = framework.Program() + block = program.global_block() + block.create_parameter( + dtype="float32", + shape=[5, 10], + lod_level=0, + name="param", + initializer=initializer.MSRAInitializer( + fan_in=12, seed=134)) + self.assertEqual(len(block.ops), 1) + init_op = block.ops[0] + self.assertEqual(init_op.type, 'uniform_random') + limit = np.sqrt(6.0 / 12) + self.assertAlmostEqual(init_op.attr('min'), -limit, delta=DELTA) + self.assertAlmostEqual(init_op.attr('max'), limit, delta=DELTA) + self.assertEqual(init_op.attr('seed'), 134) + + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/v2/fluid/tests/test_is_empty_op.py b/python/paddle/v2/fluid/tests/test_is_empty_op.py index ed6e3fe24f6333c9c90d760787eb13241a7e1868..0a4dd0f4faf370161e5695d97f0ed4bf73b6ec26 100644 --- a/python/paddle/v2/fluid/tests/test_is_empty_op.py +++ b/python/paddle/v2/fluid/tests/test_is_empty_op.py @@ -33,8 +33,7 @@ class TestIsEmptyOp(unittest.TestCase): def one_case(self, input, target): op = Operator(type="is_empty", X=input, Out="out") - ctx = core.DeviceContext.create(core.CPUPlace()) - op.run(self.scope, ctx) + op.run(self.scope, core.CPUPlace()) out = self.scope.var("out").get_tensor() self.assertEqual(np.array(out)[0], target) diff --git a/python/paddle/v2/fluid/tests/test_layers.py b/python/paddle/v2/fluid/tests/test_layers.py index 3d18e7ce3a4dc6c6b917a1000de39fca71f6ac18..9d2dcca56dd1361b9e2448be9f1d5403f8ee17e3 100644 --- a/python/paddle/v2/fluid/tests/test_layers.py +++ b/python/paddle/v2/fluid/tests/test_layers.py @@ -1,170 +1,200 @@ +from __future__ import print_function +import unittest + import paddle.v2.fluid.layers as layers import paddle.v2.fluid.nets as nets -from paddle.v2.fluid.framework import Program -import paddle.v2.fluid.core as core -import unittest +from paddle.v2.fluid.framework import Program, program_guard +from paddle.v2.fluid.param_attr import ParamAttr class TestBook(unittest.TestCase): def test_fit_a_line(self): program = Program() - x = layers.data( - name='x', shape=[13], data_type='float32', main_program=program) - y_predict = layers.fc(input=x, size=1, act=None, main_program=program) - - y = layers.data( - name='y', shape=[1], data_type='float32', main_program=program) - cost = layers.square_error_cost( - input=y_predict, label=y, main_program=program) + with program_guard(program, startup_program=Program()): + x = layers.data(name='x', shape=[13], dtype='float32') + y_predict = layers.fc(input=x, size=1, act=None) + y = layers.data(name='y', shape=[1], dtype='float32') + cost = layers.square_error_cost(input=y_predict, label=y) + avg_cost = layers.mean(x=cost) + self.assertIsNotNone(avg_cost) + program.append_backward(avg_cost) - avg_cost = layers.mean(x=cost, main_program=program) - self.assertIsNotNone(avg_cost) - program.append_backward(avg_cost) - print str(program) + print(str(program)) def test_recognize_digits_mlp(self): program = Program() - - # Change g_program, so the rest layers use `g_program` - images = layers.data( - name='pixel', - shape=[784], - data_type='float32', - main_program=program) - label = layers.data( - name='label', shape=[1], data_type='int32', main_program=program) - hidden1 = layers.fc(input=images, - size=128, - act='relu', - main_program=program) - hidden2 = layers.fc(input=hidden1, - size=64, - act='relu', - main_program=program) - predict = layers.fc(input=hidden2, - size=10, - act='softmax', - main_program=program) - cost = layers.cross_entropy( - input=predict, label=label, main_program=program) - avg_cost = layers.mean(x=cost, main_program=program) - self.assertIsNotNone(avg_cost) - print str(program) + with program_guard(program, startup_program=Program()): + # Change g_program, so the rest layers use `g_program` + images = layers.data(name='pixel', shape=[784], dtype='float32') + label = layers.data(name='label', shape=[1], dtype='int32') + hidden1 = layers.fc(input=images, size=128, act='relu') + hidden2 = layers.fc(input=hidden1, size=64, act='relu') + predict = layers.fc(input=[hidden2, hidden1], + size=10, + act='softmax', + param_attr=["sftmax.w1", "sftmax.w2"]) + cost = layers.cross_entropy(input=predict, label=label) + avg_cost = layers.mean(x=cost) + self.assertIsNotNone(avg_cost) + + print(str(program)) def test_simple_conv2d(self): program = Program() - images = layers.data( - name='pixel', - shape=[3, 48, 48], - data_type='int32', - main_program=program) - layers.conv2d( - input=images, - num_filters=3, - filter_size=[4, 4], - main_program=program) - - print str(program) + with program_guard(program, startup_program=Program()): + images = layers.data(name='pixel', shape=[3, 48, 48], dtype='int32') + layers.conv2d(input=images, num_filters=3, filter_size=[4, 4]) - def test_recognize_digits_conv(self): + print(str(program)) + + def test_conv2d_transpose(self): program = Program() + with program_guard(program): + img = layers.data(name='pixel', shape=[3, 2, 2], dtype='float32') + layers.conv2d_transpose(input=img, num_filters=10, output_size=28) + print(str(program)) - images = layers.data( - name='pixel', - shape=[1, 28, 28], - data_type='float32', - main_program=program) - label = layers.data( - name='label', shape=[1], data_type='int32', main_program=program) - conv_pool_1 = nets.simple_img_conv_pool( - input=images, - filter_size=5, - num_filters=2, - pool_size=2, - pool_stride=2, - act="relu", - main_program=program) - conv_pool_2 = nets.simple_img_conv_pool( - input=conv_pool_1, - filter_size=5, - num_filters=4, - pool_size=2, - pool_stride=2, - act="relu", - main_program=program) - - predict = layers.fc(input=conv_pool_2, - size=10, - act="softmax", - main_program=program) - cost = layers.cross_entropy( - input=predict, label=label, main_program=program) - avg_cost = layers.mean(x=cost, main_program=program) - - program.append_backward(avg_cost) - - print str(program) + def test_recognize_digits_conv(self): + program = Program() + with program_guard(program, startup_program=Program()): + images = layers.data( + name='pixel', shape=[1, 28, 28], dtype='float32') + label = layers.data(name='label', shape=[1], dtype='int32') + conv_pool_1 = nets.simple_img_conv_pool( + input=images, + filter_size=5, + num_filters=2, + pool_size=2, + pool_stride=2, + act="relu") + conv_pool_2 = nets.simple_img_conv_pool( + input=conv_pool_1, + filter_size=5, + num_filters=4, + pool_size=2, + pool_stride=2, + act="relu") + + predict = layers.fc(input=conv_pool_2, size=10, act="softmax") + cost = layers.cross_entropy(input=predict, label=label) + avg_cost = layers.mean(x=cost) + + program.append_backward(avg_cost) + + print(str(program)) def test_word_embedding(self): program = Program() - dict_size = 10000 - embed_size = 32 - first_word = layers.data( - name='firstw', shape=[1], data_type='int64', main_program=program) - second_word = layers.data( - name='secondw', shape=[1], data_type='int64', main_program=program) - third_word = layers.data( - name='thirdw', shape=[1], data_type='int64', main_program=program) - forth_word = layers.data( - name='forthw', shape=[1], data_type='int64', main_program=program) - next_word = layers.data( - name='nextw', shape=[1], data_type='int64', main_program=program) - - embed_first = layers.embedding( - input=first_word, - size=[dict_size, embed_size], - data_type='float32', - param_attr={'name': 'shared_w'}, - main_program=program) - embed_second = layers.embedding( - input=second_word, - size=[dict_size, embed_size], - data_type='float32', - param_attr={'name': 'shared_w'}, - main_program=program) - - embed_third = layers.embedding( - input=third_word, - size=[dict_size, embed_size], - data_type='float32', - param_attr={'name': 'shared_w'}, - main_program=program) - embed_forth = layers.embedding( - input=forth_word, - size=[dict_size, embed_size], - data_type='float32', - param_attr={'name': 'shared_w'}, - main_program=program) - - concat_embed = layers.concat( - input=[embed_first, embed_second, embed_third, embed_forth], - axis=1, - main_program=program) - - hidden1 = layers.fc(input=concat_embed, - size=256, - act='sigmoid', - main_program=program) - predict_word = layers.fc(input=hidden1, - size=dict_size, - act='softmax', - main_program=program) - cost = layers.cross_entropy( - input=predict_word, label=next_word, main_program=program) - avg_cost = layers.mean(x=cost, main_program=program) - self.assertIsNotNone(avg_cost) - - print str(program) + with program_guard(program, startup_program=Program()): + dict_size = 10000 + embed_size = 32 + first_word = layers.data(name='firstw', shape=[1], dtype='int64') + second_word = layers.data(name='secondw', shape=[1], dtype='int64') + third_word = layers.data(name='thirdw', shape=[1], dtype='int64') + forth_word = layers.data(name='forthw', shape=[1], dtype='int64') + next_word = layers.data(name='nextw', shape=[1], dtype='int64') + + embed_first = layers.embedding( + input=first_word, + size=[dict_size, embed_size], + dtype='float32', + param_attr='shared_w') + embed_second = layers.embedding( + input=second_word, + size=[dict_size, embed_size], + dtype='float32', + param_attr='shared_w') + + embed_third = layers.embedding( + input=third_word, + size=[dict_size, embed_size], + dtype='float32', + param_attr='shared_w') + embed_forth = layers.embedding( + input=forth_word, + size=[dict_size, embed_size], + dtype='float32', + param_attr='shared_w') + + concat_embed = layers.concat( + input=[embed_first, embed_second, embed_third, embed_forth], + axis=1) + + hidden1 = layers.fc(input=concat_embed, size=256, act='sigmoid') + predict_word = layers.fc(input=hidden1, + size=dict_size, + act='softmax') + cost = layers.cross_entropy(input=predict_word, label=next_word) + avg_cost = layers.mean(x=cost) + self.assertIsNotNone(avg_cost) + + print(str(program)) + + def test_linear_chain_crf(self): + program = Program() + with program_guard(program, startup_program=Program()): + label_dict_len = 10 + images = layers.data(name='pixel', shape=[784], dtype='float32') + label = layers.data(name='label', shape=[1], dtype='int32') + hidden = layers.fc(input=images, size=128) + crf = layers.linear_chain_crf( + input=hidden, label=label, param_attr=ParamAttr(name="crfw")) + crf_decode = layers.crf_decoding( + input=hidden, param_attr=ParamAttr(name="crfw")) + layers.chunk_eval( + input=crf_decode, + label=label, + chunk_scheme="IOB", + num_chunk_types=(label_dict_len - 1) / 2) + self.assertNotEqual(crf, None) + self.assertNotEqual(crf_decode, None) + + print(str(program)) + + def test_sigmoid_cross_entropy(self): + program = Program() + with program_guard(program): + dat = layers.data(name='data', shape=[10], dtype='float32') + lbl = layers.data(name='label', shape=[10], dtype='float32') + self.assertIsNotNone( + layers.sigmoid_cross_entropy_with_logits( + x=dat, label=lbl)) + print(str(program)) + + def test_sequence_expand(self): + program = Program() + with program_guard(program): + x = layers.data(name='x', shape=[10], dtype='float32') + y = layers.data( + name='y', shape=[10, 20], dtype='float32', lod_level=1) + self.assertIsNotNone(layers.sequence_expand(x=x, y=y)) + print(str(program)) + + def test_lstm_unit(self): + program = Program() + with program_guard(program): + x_t_data = layers.data( + name='x_t_data', shape=[10, 10], dtype='float32') + x_t = layers.fc(input=x_t_data, size=10) + prev_hidden_data = layers.data( + name='prev_hidden_data', shape=[10, 20], dtype='float32') + prev_hidden = layers.fc(input=prev_hidden_data, size=20) + prev_cell_data = layers.data( + name='prev_cell', shape=[10, 30], dtype='float32') + prev_cell = layers.fc(input=prev_cell_data, size=30) + self.assertIsNotNone( + layers.lstm_unit( + x_t=x_t, hidden_t_prev=prev_hidden, cell_t_prev=prev_cell)) + print(str(program)) + + def test_sequence_softmax(self): + program = Program() + with program_guard(program): + seq_data = layers.data( + name='seq_data', shape=[10, 10], dtype='float32', lod_level=1) + seq = layers.fc(input=seq_data, size=20) + self.assertIsNotNone(layers.sequence_softmax(x=seq)) + print(str(program)) if __name__ == '__main__': diff --git a/python/paddle/v2/fluid/tests/test_linear_chain_crf_op.py b/python/paddle/v2/fluid/tests/test_linear_chain_crf_op.py index 6f06a66c825b37ee91214efc0a29a58f0b9057f9..c26634ff20c46e484d600c758be386ec8327d1c1 100644 --- a/python/paddle/v2/fluid/tests/test_linear_chain_crf_op.py +++ b/python/paddle/v2/fluid/tests/test_linear_chain_crf_op.py @@ -104,7 +104,7 @@ class TestLinearChainCrfOp(OpTest): transition_exps = np.exp(transition) labels = np.random.randint( - low=0, high=TAG_NUM, size=(lod[-1][-1], 1), dtype="int32") + low=0, high=TAG_NUM, size=(lod[-1][-1], 1), dtype="int64") self.inputs = { "Emission": (emission, lod), diff --git a/python/paddle/v2/fluid/tests/test_lod_array_length_op.py b/python/paddle/v2/fluid/tests/test_lod_array_length_op.py index a01ae83772185df218b8c453557dc0cac719673b..8a4be545eda841dbda33b7c8cae9f91a4199f2f8 100644 --- a/python/paddle/v2/fluid/tests/test_lod_array_length_op.py +++ b/python/paddle/v2/fluid/tests/test_lod_array_length_op.py @@ -13,7 +13,7 @@ class TestLoDArrayLength(unittest.TestCase): arr_len = layers.array_length(arr) cpu = core.CPUPlace() exe = Executor(cpu) - result = numpy.array(exe.run(fetch_list=[arr_len])[0]) + result = exe.run(fetch_list=[arr_len])[0] self.assertEqual(11, result[0]) diff --git a/python/paddle/v2/fluid/tests/test_lod_rank_table.py b/python/paddle/v2/fluid/tests/test_lod_rank_table.py index bbc11930b9e804c2769cc590c298c6e90dc36ca6..30d619fe318517345195281b17f88e9916b6afb3 100644 --- a/python/paddle/v2/fluid/tests/test_lod_rank_table.py +++ b/python/paddle/v2/fluid/tests/test_lod_rank_table.py @@ -1,6 +1,5 @@ from paddle.v2.fluid.layers import lod_rank_table, data from paddle.v2.fluid.executor import Executor -from paddle.v2.fluid.framework import g_main_program import paddle.v2.fluid.core as core import numpy import unittest @@ -18,7 +17,7 @@ class TestLoDRankTable(unittest.TestCase): tensor = core.LoDTensor() tensor.set(numpy.random.random(size=(17, 100)), cpu) tensor.set_lod([[0, 1, 3], [0, 5, 6, 7], [0, 3, 4, 9, 10, 13, 16, 17]]) - exe.run(g_main_program, scope=scope, feed={'x': tensor}) + exe.run(scope=scope, feed={'x': tensor}) var = scope.find_var(rank_table.name) table = var.get_lod_rank_table() self.assertEqual([(0, 5), (1, 1), (2, 1)], table.items()) diff --git a/python/paddle/v2/fluid/tests/test_lod_tensor_array_ops.py b/python/paddle/v2/fluid/tests/test_lod_tensor_array_ops.py index b18cb6b49fa41f26e1b6de1128690507c5a2f099..c552cb033f1ec8f5843490083edee7b2762b5703 100644 --- a/python/paddle/v2/fluid/tests/test_lod_tensor_array_ops.py +++ b/python/paddle/v2/fluid/tests/test_lod_tensor_array_ops.py @@ -2,9 +2,9 @@ import unittest import paddle.v2.fluid.core as core import numpy import paddle.v2.fluid.layers as layers -from paddle.v2.fluid.framework import Program +from paddle.v2.fluid.framework import Program, program_guard from paddle.v2.fluid.executor import Executor -from paddle.v2.fluid.backward import append_backward_ops +from paddle.v2.fluid.backward import append_backward class TestCPULoDTensorArrayOps(unittest.TestCase): @@ -18,7 +18,11 @@ class TestCPULoDTensorArrayOps(unittest.TestCase): tensor.set_lod([[0, 3, 9, 10]]) expect = map(lambda x: numpy.array(x).astype('int32'), [[3, 0, 9], [4, 1], [5, 2], [6], [7], [8]]) - self.main(tensor=tensor, expect_array=expect, expect_lod=[] * 6) + self.main( + tensor=tensor, + expect_array=expect, + expect_lod=[] * 6, + expect_max_len=6) def test_lod_tensor_to_array_level_0_empty_seq(self): tensor = core.LoDTensor() @@ -27,7 +31,11 @@ class TestCPULoDTensorArrayOps(unittest.TestCase): tensor.set_lod([[0, 3, 9, 9, 10]]) expect = map(lambda x: numpy.array(x).astype('int32'), [[3, 0, 9], [4, 1], [5, 2], [6], [7], [8]]) - self.main(tensor=tensor, expect_array=expect, expect_lod=[] * 6) + self.main( + tensor=tensor, + expect_array=expect, + expect_lod=[] * 6, + expect_max_len=6) def test_lod_tensor_to_array_level_1(self): tensor = core.LoDTensor() @@ -44,7 +52,11 @@ class TestCPULoDTensorArrayOps(unittest.TestCase): ] lod = [[[0, 2, 5]], [[0, 6, 12]], [[0, 3]]] - self.main(tensor=tensor, expect_array=expect, expect_lod=lod) + self.main( + tensor=tensor, + expect_array=expect, + expect_lod=lod, + expect_max_len=3) def test_lod_tensor_to_array_level_1_empty_seq(self): tensor = core.LoDTensor() @@ -63,7 +75,11 @@ class TestCPULoDTensorArrayOps(unittest.TestCase): ] lod = [[[0, 5, 8, 8, 15]], [[0, 2, 6, 7, 8]], [[0, 2, 6]], [[0, 2]]] - self.main(tensor=tensor, expect_array=expect, expect_lod=lod) + self.main( + tensor=tensor, + expect_array=expect, + expect_lod=lod, + expect_max_len=4) def test_lod_tensor_to_array_level_2(self): tensor = core.LoDTensor() @@ -80,7 +96,11 @@ class TestCPULoDTensorArrayOps(unittest.TestCase): ] lod = [[[0, 1, 3, 4], [0, 1, 4, 8, 12]], [[0, 4, 7], [0, 1, 5, 9, 17, 21, 27, 31]], [[0, 2], [0, 6, 7]]] - self.main(tensor=tensor, expect_array=expect, expect_lod=lod) + self.main( + tensor=tensor, + expect_array=expect, + expect_lod=lod, + expect_max_len=3) def test_lod_tensor_to_array_level_2_skip_level(self): tensor = core.LoDTensor() @@ -88,19 +108,27 @@ class TestCPULoDTensorArrayOps(unittest.TestCase): numpy.arange(50).reshape(50, 1).astype('int32'), self.place()) tensor.set_lod([[0, 2, 5, 6], [0, 2, 5, 6, 10, 12, 13], [0, 3, 7, 11, 17, 21, 22, 23, 27, 31, 39, 45, 46, 50]]) - self.main(tensor=tensor, expect_array=None, expect_lod=None, level=1) - - def main(self, tensor, expect_array, expect_lod, level=0): + self.main( + tensor=tensor, + expect_array=None, + expect_lod=None, + expect_max_len=4, + level=1) + + def main(self, tensor, expect_array, expect_lod, expect_max_len, level=0): place = self.place() program = Program() - x = layers.data(name='x', shape=[10], main_program=program) - x.persistable = True - table = layers.lod_rank_table(x, level=level, main_program=program) - array = layers.lod_tensor_to_array(x, table, main_program=program) - array.persistable = True - - result = layers.array_to_lod_tensor(array, table, main_program=program) - result.persistable = True + with program_guard(program): + x = layers.data(name='x', shape=[10]) + x.persistable = True + table = layers.lod_rank_table(x, level=level) + max_len = layers.max_sequence_len(table) + max_len.persistable = True + array = layers.lod_tensor_to_array(x, table) + array.persistable = True + + result = layers.array_to_lod_tensor(array, table) + result.persistable = True exe = Executor(place) scope = core.Scope() exe.run(program, feed={'x': tensor}, scope=scope) @@ -110,6 +138,10 @@ class TestCPULoDTensorArrayOps(unittest.TestCase): self.check_array_same(array, expect_array, expect_lod) self.check_tensor_same(scope.find_var(result.name).get_tensor(), tensor) + self.assertEqual( + numpy.array(scope.find_var(max_len.name).get_tensor())[0], + expect_max_len) + def check_array_same(self, array, expect_tensor, expect_lod): self.assertEqual(len(expect_tensor), len(array)) for i, exp in enumerate(zip(expect_tensor, expect_lod)): @@ -129,19 +161,16 @@ class TestCPULoDTensorArrayOpGrad(unittest.TestCase): place = core.CPUPlace() program = Program() - x = layers.data( - name='x', - shape=[1], - data_type='float32', - main_program=program, - stop_gradient=False) - table = layers.lod_rank_table(x, level=0, main_program=program) - array = layers.lod_tensor_to_array(x, table, main_program=program) - result = layers.array_to_lod_tensor(array, table, main_program=program) + with program_guard(program): + x = layers.data( + name='x', shape=[1], dtype='float32', stop_gradient=False) + table = layers.lod_rank_table(x, level=0) + array = layers.lod_tensor_to_array(x, table) + result = layers.array_to_lod_tensor(array, table) - mean = layers.mean(x=result, main_program=program) + mean = layers.mean(x=result) - append_backward_ops(mean) + append_backward(mean) tensor = core.LoDTensor() tensor.set(numpy.arange(10).reshape(10, 1).astype('float32'), place) @@ -151,10 +180,11 @@ class TestCPULoDTensorArrayOpGrad(unittest.TestCase): exe = Executor(place) g_out = [ - item.sum() - for item in map( - numpy.array, - exe.run(program, feed={'x': tensor}, fetch_list=[g_vars])) + numpy.array(item).sum() + for item in exe.run(program, + feed={'x': tensor}, + fetch_list=[g_vars], + return_numpy=False) ] g_out_sum = numpy.array(g_out).sum() diff --git a/python/paddle/v2/fluid/tests/test_log_loss_op.py b/python/paddle/v2/fluid/tests/test_log_loss_op.py new file mode 100644 index 0000000000000000000000000000000000000000..2eeaa90758c57ef0d92a8ad7b0a4c1b1f2c38be3 --- /dev/null +++ b/python/paddle/v2/fluid/tests/test_log_loss_op.py @@ -0,0 +1,33 @@ +import unittest +import numpy as np +from op_test import OpTest + + +class TestLogLossOp(OpTest): + def setUp(self): + self.op_type = 'log_loss' + samples_num = 32 + + predicted = np.random.uniform(0.1, 1.0, + (samples_num, 1)).astype("float32") + labels = np.random.randint(0, 2, (samples_num, 1)).astype("float32") + epsilon = 1e-4 + self.inputs = { + 'Predicted': predicted, + 'Labels': labels, + } + + self.attrs = {'epsilon': epsilon} + loss = -labels * np.log(predicted + epsilon) - ( + 1 - labels) * np.log(1 - predicted + epsilon) + self.outputs = {'Loss': loss} + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad(['Predicted'], 'Loss', max_relative_error=0.03) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/v2/fluid/tests/test_lrn_op.py b/python/paddle/v2/fluid/tests/test_lrn_op.py index 7e34b3c91c16c440f12c51415c509400e1f315dc..9abb09e53a7af8eec69f9bd501c6883dd9df9930 100644 --- a/python/paddle/v2/fluid/tests/test_lrn_op.py +++ b/python/paddle/v2/fluid/tests/test_lrn_op.py @@ -23,7 +23,7 @@ class TestLRNOp(OpTest): start = -(self.n - 1) / 2 end = start + self.n - mid = np.empty((self.N, self.C, self.H, self.W), dtype=float) + mid = np.empty((self.N, self.C, self.H, self.W)).astype("float32") mid.fill(self.k) for m in range(0, self.N): for i in range(0, self.C): @@ -74,5 +74,4 @@ class TestLRNOp(OpTest): if __name__ == "__main__": - exit(0) # LRN grad implement wrong unittest.main() diff --git a/python/paddle/v2/fluid/tests/test_maxout_op.py b/python/paddle/v2/fluid/tests/test_maxout_op.py new file mode 100644 index 0000000000000000000000000000000000000000..5fbed43e254b811d38e441e946a73c24f87373de --- /dev/null +++ b/python/paddle/v2/fluid/tests/test_maxout_op.py @@ -0,0 +1,37 @@ +import unittest +import numpy as np +from op_test import OpTest + + +def maxout_forward_naive(input, groups): + s0, s1, s2, s3 = input.shape + return np.ndarray([s0, s1 / groups, groups, s2, s3], \ + buffer = input, dtype=input.dtype).max(axis=(2)) + + +class TestMaxOutOp(OpTest): + def setUp(self): + self.op_type = "maxout" + self.init_test_case() + input = np.random.random(self.shape).astype("float32") + output = self.MaxOut_forward_naive(input, self.groups).astype("float32") + + self.inputs = {'X': input} + self.attrs = {'groups': self.groups} + + self.outputs = {'Out': output.astype('float32')} + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad(['X'], 'Out') + + def init_test_case(self): + self.MaxOut_forward_naive = maxout_forward_naive + self.shape = [100, 6, 2, 2] + self.groups = 2 + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/v2/fluid/tests/test_mnist_if_else_op.py b/python/paddle/v2/fluid/tests/test_mnist_if_else_op.py new file mode 100644 index 0000000000000000000000000000000000000000..33558c6105442b169b2e26abc7f39e15b7fe7322 --- /dev/null +++ b/python/paddle/v2/fluid/tests/test_mnist_if_else_op.py @@ -0,0 +1,134 @@ +import paddle.v2.fluid.layers as layers +from paddle.v2.fluid.framework import Program, program_guard, default_main_program, default_startup_program +from paddle.v2.fluid.executor import Executor +from paddle.v2.fluid.optimizer import MomentumOptimizer +import paddle.v2.fluid.core as core +import paddle.v2 as paddle +import unittest +import numpy as np + + +class TestMNISTIfElseOp(unittest.TestCase): + def test_raw_api(self): + prog = Program() + startup_prog = Program() + with program_guard(prog, startup_prog): + image = layers.data(name='x', shape=[784], dtype='float32') + + label = layers.data(name='y', shape=[1], dtype='int64') + + limit = layers.fill_constant_batch_size_like( + input=label, dtype='int64', shape=[1], value=5.0) + cond = layers.less_than(x=label, y=limit) + true_image, false_image = layers.split_lod_tensor( + input=image, mask=cond) + + true_out = layers.create_tensor(dtype='float32') + true_cond = layers.ConditionalBlock([true_image]) + + with true_cond.block(): + hidden = layers.fc(input=true_image, size=100, act='tanh') + prob = layers.fc(input=hidden, size=10, act='softmax') + layers.assign(input=prob, output=true_out) + + false_out = layers.create_tensor(dtype='float32') + false_cond = layers.ConditionalBlock([false_image]) + + with false_cond.block(): + hidden = layers.fc(input=false_image, size=200, act='tanh') + prob = layers.fc(input=hidden, size=10, act='softmax') + layers.assign(input=prob, output=false_out) + + prob = layers.merge_lod_tensor( + in_true=true_out, in_false=false_out, mask=cond, x=image) + loss = layers.cross_entropy(input=prob, label=label) + avg_loss = layers.mean(x=loss) + + optimizer = MomentumOptimizer(learning_rate=0.001, momentum=0.9) + optimizer.minimize(avg_loss, startup_prog) + + train_reader = paddle.batch( + paddle.reader.shuffle( + paddle.dataset.mnist.train(), buf_size=8192), + batch_size=200) + + place = core.CPUPlace() + exe = Executor(place) + + exe.run(startup_prog) + PASS_NUM = 100 + for pass_id in range(PASS_NUM): + for data in train_reader(): + x_data = np.array(map(lambda x: x[0], data)).astype("float32") + y_data = np.array(map(lambda x: x[1], data)).astype("int64") + y_data = np.expand_dims(y_data, axis=1) + + outs = exe.run(prog, + feed={'x': x_data, + 'y': y_data}, + fetch_list=[avg_loss]) + print outs[0] + if outs[0] < 1.0: + return + self.assertFalse(True) + + def test_ifelse(self): + prog = Program() + startup_prog = Program() + with program_guard(prog, startup_prog): + image = layers.data(name='x', shape=[784], dtype='float32') + + label = layers.data(name='y', shape=[1], dtype='int64') + + limit = layers.fill_constant_batch_size_like( + input=label, dtype='int64', shape=[1], value=5.0) + cond = layers.less_than(x=label, y=limit) + ie = layers.IfElse(cond) + + with ie.true_block(): + true_image = ie.input(image) + hidden = layers.fc(input=true_image, size=100, act='tanh') + prob = layers.fc(input=hidden, size=10, act='softmax') + ie.output(prob) + + with ie.false_block(): + false_image = ie.input(image) + hidden = layers.fc(input=false_image, size=200, act='tanh') + prob = layers.fc(input=hidden, size=10, act='softmax') + ie.output(prob) + + prob = ie() + loss = layers.cross_entropy(input=prob[0], label=label) + avg_loss = layers.mean(x=loss) + + optimizer = MomentumOptimizer(learning_rate=0.001, momentum=0.9) + optimizer.minimize(avg_loss, startup_prog) + train_reader = paddle.batch( + paddle.reader.shuffle( + paddle.dataset.mnist.train(), buf_size=8192), + batch_size=200) + + place = core.CPUPlace() + exe = Executor(place) + + exe.run(kwargs['startup_program']) + PASS_NUM = 100 + for pass_id in range(PASS_NUM): + for data in train_reader(): + x_data = np.array(map(lambda x: x[0], data)).astype("float32") + y_data = np.array(map(lambda x: x[1], data)).astype("int64") + y_data = y_data.reshape((y_data.shape[0], 1)) + + outs = exe.run(kwargs['main_program'], + feed={'x': x_data, + 'y': y_data}, + fetch_list=[avg_loss]) + print outs[0] + if outs[0] < 1.0: + return + self.assertFalse(True) + + +if __name__ == '__main__': + # temp disable if else unittest since it could be buggy. + exit(0) diff --git a/python/paddle/v2/fluid/tests/test_nccl_init_op.py b/python/paddle/v2/fluid/tests/test_nccl_init_op.py deleted file mode 100644 index a536800ccd81fdc2f3b7c8320cede4f8ecf3a8cb..0000000000000000000000000000000000000000 --- a/python/paddle/v2/fluid/tests/test_nccl_init_op.py +++ /dev/null @@ -1,39 +0,0 @@ -import unittest, os -import numpy as np -import paddle.v2 as paddle -from paddle.v2.fluid.op import Operator -import paddle.v2.fluid.core as core -from op_test import OpTest, create_op, set_input - -if not core.is_compile_gpu(): - exit(0) - -gpu_count = core.get_cuda_device_count() - -if gpu_count <= 1: - exit(0) - -g_scope = core.Scope() -g_ctx = core.DeviceContext.create(core.CPUPlace()) - - -class TestNCCLInit(unittest.TestCase): - def test_init(self): - self.op_type = "ncclInit" - self.gpus = range(gpu_count) - - self.inputs = {} - self.attrs = {"gpus": self.gpus} - g_scope.var("Communicator").get_communicator() - self.outputs = {"Communicator": g_scope.find_var("Communicator")} - nccl_init = create_op( - g_scope, - op_type=self.op_type, - inputs=self.inputs, - outputs=self.outputs, - attrs=self.attrs) - nccl_init.run(g_scope, g_ctx) - - -if __name__ == "__main__": - unittest.main() diff --git a/python/paddle/v2/fluid/tests/test_nce.py b/python/paddle/v2/fluid/tests/test_nce.py new file mode 100644 index 0000000000000000000000000000000000000000..8aeba69769525935c26576ec50035ed50d2ce44f --- /dev/null +++ b/python/paddle/v2/fluid/tests/test_nce.py @@ -0,0 +1,98 @@ +import unittest +import numpy as np +from op_test import OpTest + + +def nce(input, weight, bias, sample_weight, labels, num_classes, + num_sample_class): + samples = [] + sample_labels = [] + batch_size = input.shape[0] + num_true_class = labels.shape[1] + for i in range(batch_size): + w = 1 if sample_weight is None else sample_weight[i] + for label in labels[i]: + samples.append((i, label, True, w)) + sample_labels.append(label) + for num in range(num_sample_class): + samples.append((i, num, False, w)) + sample_labels.append(num) + # forward bias + sample_out = np.zeros(len(samples)).astype(np.float32) + if bias is not None: + for i in range(len(samples)): + sample_out[i] = bias[samples[i][1]] + # forward weight + for i in range(len(samples)): + sample_out[i] += np.dot(input[samples[i][0]], weight[samples[i][1]]) + + # forward activation + sample_out = 1.0 / (1.0 + np.exp(-sample_out)) + # forward cost + out = np.zeros(batch_size).astype(np.float32) + b = 1.0 / num_classes * num_sample_class + for i in range(len(samples)): + o = sample_out[i] + cost = -np.log(o / (o + b)) if samples[i][2] else -np.log(b / (o + b)) + out[samples[i][0]] += cost * samples[i][3] + return (out[:, np.newaxis], np.array(sample_out).reshape( + batch_size, num_sample_class + num_true_class), + np.array(sample_labels).reshape(batch_size, + num_sample_class + num_true_class)) + + +class TestNCE(OpTest): + def generate_data(self, dim, batch_size, num_classes, num_true_class, + num_neg_samples): + input = np.random.randn(batch_size, dim).astype(np.float32) + weight = np.random.randn(num_classes, dim).astype(np.float32) + bias = np.random.randn(num_classes).astype(np.float32) + sample_weight = np.random.randn(batch_size).astype(np.float32) + labels = np.random.randint(0, num_classes, (batch_size, num_true_class)) + self.attrs = { + 'num_total_classes': num_classes, + 'num_neg_samples': num_neg_samples, + 'custom_neg_classes': range(num_neg_samples) + } + self.inputs = { + 'Input': input, + 'Label': labels, + 'Weight': weight, + 'Bias': bias, + 'SampleWeight': sample_weight + } + + def set_data(self): + self.generate_data(5, 5, 4, 1, 2) + + def compute(self): + out = nce(self.inputs['Input'], self.inputs['Weight'], + self.inputs['Bias'], self.inputs['SampleWeight'], + self.inputs['Label'], self.attrs['num_total_classes'], + self.attrs['num_neg_samples']) + self.outputs = { + 'Cost': out[0], + 'SampleLogits': out[1], + 'SampleLabels': out[2] + } + + def setUp(self): + self.op_type = 'nce' + self.set_data() + self.compute() + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad( + ["Input", "Weight", "Bias"], "Cost", max_relative_error=0.02) + + +class TestNCECase1(TestNCE): + def set_data(self): + self.generate_data(10, 20, 10, 2, 5) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/v2/fluid/tests/test_net.py b/python/paddle/v2/fluid/tests/test_net.py index 318df08a9e73ac95cab73c34182bc6220ef6c681..d9fe55a8af5c750c5c926e875ddbb645f8abb1a0 100644 --- a/python/paddle/v2/fluid/tests/test_net.py +++ b/python/paddle/v2/fluid/tests/test_net.py @@ -7,7 +7,7 @@ def fc(X, W, Y): ret_v = core.Net.create() ret_v.append_op(Operator("mul", X="X", Y="W", Out="pre_activation")) - ret_v.append_op(Operator("sigmoid", X="pre_activation", Y=Y)) + ret_v.append_op(Operator("sigmoid", X="pre_activation", Out=Y)) ret_v.complete_add_op(True) return ret_v @@ -30,7 +30,7 @@ Op(plain_net), inputs:{all[W, X, Y]}, outputs:{all[Out, fc.out, pre_activation]} Op(plain_net), inputs:{all[W, X]}, outputs:{all[fc.out, pre_activation]}. Op(plain_net), inputs:{all[W, X]}, outputs:{all[fc.out, pre_activation]}. Op(mul), inputs:{X[X], Y[W]}, outputs:{Out[pre_activation]}. - Op(sigmoid), inputs:{X[pre_activation]}, outputs:{Y[fc.out]}. + Op(sigmoid), inputs:{X[pre_activation]}, outputs:{Out[fc.out]}. ''' self.assertEqual(expected, "\n" + str(net)) diff --git a/python/paddle/v2/fluid/tests/test_operator.py b/python/paddle/v2/fluid/tests/test_operator.py index 4aa022ef90159cd96eed4e4dbe30cf5d1e8a41a7..c059a2b88b1324935f871b6e9c11efd5652ddd65 100644 --- a/python/paddle/v2/fluid/tests/test_operator.py +++ b/python/paddle/v2/fluid/tests/test_operator.py @@ -1,6 +1,6 @@ import unittest + import paddle.v2.fluid.op as op -import paddle.v2.fluid.core as core import paddle.v2.fluid.proto.framework_pb2 as framework_pb2 diff --git a/python/paddle/v2/fluid/tests/test_operator_desc.py b/python/paddle/v2/fluid/tests/test_operator_desc.py index e8362d2e9c6038c04c24dce35de8c53bfde78142..ce34d95ac8cb2644dee9c551cd8e85b33609919a 100644 --- a/python/paddle/v2/fluid/tests/test_operator_desc.py +++ b/python/paddle/v2/fluid/tests/test_operator_desc.py @@ -1,11 +1,15 @@ import unittest -from paddle.v2.fluid.framework import Variable, Program, g_main_program + import paddle.v2.fluid.core as core +from paddle.v2.fluid.framework import Program, default_startup_program + +main_program = default_startup_program() + class TestOperator(unittest.TestCase): def test_error_type(self): - block = g_main_program.create_block() + block = main_program.create_block() try: block.append_op() self.assertFail() diff --git a/python/paddle/v2/fluid/tests/test_optimizer.py b/python/paddle/v2/fluid/tests/test_optimizer.py index 2459dfd664300d405edb36c4ca906c1769b5e7d2..1eadb7d912629024ee21e30b0a5fa4910bb96e06 100644 --- a/python/paddle/v2/fluid/tests/test_optimizer.py +++ b/python/paddle/v2/fluid/tests/test_optimizer.py @@ -2,7 +2,7 @@ import unittest import paddle.v2.fluid.framework as framework import paddle.v2.fluid.optimizer as optimizer -from paddle.v2.fluid.backward import append_backward_ops +from paddle.v2.fluid.backward import append_backward class TestOptimizer(unittest.TestCase): @@ -27,7 +27,7 @@ class TestOptimizer(unittest.TestCase): block.append_op( type="mean", inputs={"X": mul_out}, outputs={"Out": mean_out}) sgd_optimizer = optimizer.SGDOptimizer(learning_rate=0.01) - opts = sgd_optimizer.minimize(mean_out, init_program) + opts, _ = sgd_optimizer.minimize(mean_out, init_program) self.assertEqual(len(opts), 1) sgd_op = opts[0] self.assertEqual(sgd_op.type, "sgd") @@ -57,7 +57,7 @@ class TestOptimizer(unittest.TestCase): learning_rate = 0.01 sgd_optimizer = optimizer.SGDOptimizer( learning_rate=learning_rate, global_step=global_step) - opts = sgd_optimizer.minimize(mean_out, init_program) + opts, _ = sgd_optimizer.minimize(mean_out, init_program) self.assertEqual(len(opts), 2) sgd_op = opts[0] self.assertEqual(sgd_op.type, "sgd") @@ -102,7 +102,7 @@ class TestMomentumOptimizer(unittest.TestCase): dtype="float32", shape=[1], lod_level=0, name="mean.out") block.append_op( type="mean", inputs={"X": mul_out}, outputs={"Out": mean_out}) - params_grads = append_backward_ops(mean_out) + params_grads = append_backward(mean_out) self.assertEqual(len(params_grads), 1) self.assertEqual(len(momentum_optimizer.get_accumulators()), 0) opts = momentum_optimizer.create_optimization_pass( @@ -151,7 +151,7 @@ class TestMomentumOptimizer(unittest.TestCase): learning_rate = 0.01 momentum_optimizer = self.MockMomentum( learning_rate=learning_rate, momentum=0.2, use_nesterov=True) - params_grads = append_backward_ops(mean_out) + params_grads = append_backward(mean_out) self.assertEqual(len(params_grads), 1) self.assertEqual(len(momentum_optimizer.get_accumulators()), 0) opts = momentum_optimizer.create_optimization_pass( @@ -209,7 +209,7 @@ class TestAdagradOptimizer(unittest.TestCase): learning_rate = 0.01 adagrad_optimizer = self.MockAdagrad( learning_rate=learning_rate, epsilon=1.0e-6) - params_grads = append_backward_ops(mean_out) + params_grads = append_backward(mean_out) self.assertEqual(len(params_grads), 1) self.assertEqual(len(adagrad_optimizer.get_accumulators()), 0) opts = adagrad_optimizer.create_optimization_pass(params_grads, mul_out, @@ -269,7 +269,7 @@ class TestAdamOptimizer(unittest.TestCase): learning_rate = 0.01 adam_optimizer = self.MockAdam( learning_rate=learning_rate, beta1=0.9, beta2=0.999) - params_grads = append_backward_ops(mean_out) + params_grads = append_backward(mean_out) self.assertEqual(len(params_grads), 1) self.assertEqual(len(adam_optimizer.get_accumulators()), 0) opts = adam_optimizer.create_optimization_pass(params_grads, mul_out, @@ -331,7 +331,7 @@ class TestAdamaxOptimizer(unittest.TestCase): learning_rate = 0.01 adamax_optimizer = self.MockAdamax( learning_rate=learning_rate, beta1=0.9, beta2=0.999) - params_grads = append_backward_ops(mean_out) + params_grads = append_backward(mean_out) self.assertEqual(len(params_grads), 1) self.assertEqual(len(adamax_optimizer.get_accumulators()), 0) opts = adamax_optimizer.create_optimization_pass(params_grads, mul_out, @@ -390,7 +390,7 @@ class TestDecayedAdagradOptimizer(unittest.TestCase): learning_rate = 0.01 decayed_adagrad_optimizer = self.MockDecayedAdagrad( learning_rate=learning_rate, decay=0.95, epsilon=1.0e-6) - params_grads = append_backward_ops(mean_out) + params_grads = append_backward(mean_out) self.assertEqual(len(params_grads), 1) self.assertEqual(len(decayed_adagrad_optimizer.get_accumulators()), 0) opts = decayed_adagrad_optimizer.create_optimization_pass( diff --git a/python/paddle/v2/fluid/tests/test_parameter.py b/python/paddle/v2/fluid/tests/test_parameter.py index a633d22c2b1db2728b6eb767078ce4aec6cce163..694344acbbd3b7c80cb0ff48ada843f794061282 100644 --- a/python/paddle/v2/fluid/tests/test_parameter.py +++ b/python/paddle/v2/fluid/tests/test_parameter.py @@ -1,17 +1,19 @@ import unittest -from paddle.v2.fluid.framework import g_main_program +from paddle.v2.fluid.framework import default_main_program import paddle.v2.fluid.core as core from paddle.v2.fluid.executor import Executor import paddle.v2.fluid.io as io from paddle.v2.fluid.initializer import ConstantInitializer import numpy as np +main_program = default_main_program() + class TestParameter(unittest.TestCase): def test_param(self): shape = [784, 100] val = 1.0625 - b = g_main_program.global_block() + b = main_program.global_block() param = b.create_parameter( name='fc.w', shape=shape, @@ -20,12 +22,12 @@ class TestParameter(unittest.TestCase): self.assertIsNotNone(param) self.assertEqual('fc.w', param.name) self.assertEqual((784, 100), param.shape) - self.assertEqual(core.DataType.FP32, param.data_type) + self.assertEqual(core.DataType.FP32, param.dtype) self.assertEqual(0, param.block.idx) exe = Executor(core.CPUPlace()) - p = exe.run(g_main_program, fetch_list=[param])[0] - self.assertTrue(np.allclose(np.array(p), np.ones(shape) * val)) - p = io.get_parameter_value_by_name('fc.w', exe, g_main_program) + p = exe.run(main_program, fetch_list=[param])[0] + self.assertTrue(np.allclose(p, np.ones(shape) * val)) + p = io.get_parameter_value_by_name('fc.w', exe, main_program) self.assertTrue(np.allclose(np.array(p), np.ones(shape) * val)) diff --git a/python/paddle/v2/fluid/tests/test_pool2d_op.py b/python/paddle/v2/fluid/tests/test_pool2d_op.py index ac3fa6aa87835b3cd6fb9bbf6fe66b1d0c577ca2..5dff6270f455395ce6ca8ae2428236f630467095 100644 --- a/python/paddle/v2/fluid/tests/test_pool2d_op.py +++ b/python/paddle/v2/fluid/tests/test_pool2d_op.py @@ -3,8 +3,7 @@ import numpy as np from op_test import OpTest -def max_pool2D_forward_naive(x, ksize, strides, paddings=[0, 0], global_pool=0): - +def max_pool2D_forward_naive(x, ksize, strides, paddings, global_pool=0): N, C, H, W = x.shape if global_pool == 1: ksize = [H, W] @@ -23,8 +22,7 @@ def max_pool2D_forward_naive(x, ksize, strides, paddings=[0, 0], global_pool=0): return out -def avg_pool2D_forward_naive(x, ksize, strides, paddings=[0, 0], global_pool=0): - +def avg_pool2D_forward_naive(x, ksize, strides, paddings, global_pool=0): N, C, H, W = x.shape if global_pool == 1: ksize = [H, W] @@ -47,6 +45,7 @@ def avg_pool2D_forward_naive(x, ksize, strides, paddings=[0, 0], global_pool=0): class TestPool2d_Op(OpTest): def setUp(self): self.init_test_case() + self.init_global_pool() self.init_op_type() self.init_pool_type() if self.global_pool: @@ -75,8 +74,6 @@ class TestPool2d_Op(OpTest): self.check_grad(set(['X']), 'Out', max_relative_error=0.07) def init_test_case(self): - self.global_pool = True - self.pool2D_forward_naive = avg_pool2D_forward_naive self.shape = [2, 3, 5, 5] self.ksize = [3, 3] self.strides = [1, 1] @@ -87,12 +84,14 @@ class TestPool2d_Op(OpTest): def init_pool_type(self): self.pool_type = "avg" + self.pool2D_forward_naive = avg_pool2D_forward_naive + + def init_global_pool(self): + self.global_pool = True class TestCase1(TestPool2d_Op): def init_test_case(self): - self.global_pool = False - self.pool2D_forward_naive = avg_pool2D_forward_naive self.shape = [2, 3, 7, 7] self.ksize = [3, 3] self.strides = [1, 1] @@ -103,12 +102,14 @@ class TestCase1(TestPool2d_Op): def init_pool_type(self): self.pool_type = "avg" + self.pool2D_forward_naive = avg_pool2D_forward_naive + + def init_global_pool(self): + self.global_pool = False class TestCase2(TestPool2d_Op): def init_test_case(self): - self.global_pool = False - self.pool2D_forward_naive = avg_pool2D_forward_naive self.shape = [2, 3, 7, 7] self.ksize = [3, 3] self.strides = [1, 1] @@ -119,152 +120,69 @@ class TestCase2(TestPool2d_Op): def init_pool_type(self): self.pool_type = "avg" + self.pool2D_forward_naive = avg_pool2D_forward_naive + def init_global_pool(self): + self.global_pool = False -class TestCase3(TestPool2d_Op): - def init_test_case(self): - self.global_pool = True - self.pool2D_forward_naive = max_pool2D_forward_naive - self.shape = [2, 3, 5, 5] - self.ksize = [3, 3] - self.strides = [1, 1] - self.paddings = [0, 0] +class TestCase3(TestPool2d_Op): def init_op_type(self): self.op_type = "pool2d" def init_pool_type(self): self.pool_type = "max" - - -class TestCase4(TestPool2d_Op): - def init_test_case(self): - self.global_pool = False self.pool2D_forward_naive = max_pool2D_forward_naive - self.shape = [2, 3, 7, 7] - self.ksize = [3, 3] - self.strides = [1, 1] - self.paddings = [0, 0] + +class TestCase4(TestCase1): def init_op_type(self): self.op_type = "pool2d" def init_pool_type(self): self.pool_type = "max" - - -class TestCase5(TestPool2d_Op): - def init_test_case(self): - self.global_pool = False self.pool2D_forward_naive = max_pool2D_forward_naive - self.shape = [2, 3, 7, 7] - self.ksize = [3, 3] - self.strides = [1, 1] - self.paddings = [1, 1] + +class TestCase5(TestCase2): def init_op_type(self): self.op_type = "pool2d" def init_pool_type(self): self.pool_type = "max" + self.pool2D_forward_naive = max_pool2D_forward_naive #--------------------test pool2d_cudnn-------------------- -class TestCaseCudnn1(TestPool2d_Op): - def init_test_case(self): - self.global_pool = True - self.pool2D_forward_naive = avg_pool2D_forward_naive - self.shape = [2, 3, 5, 5] - self.ksize = [3, 3] - self.strides = [1, 1] - self.paddings = [0, 0] - +class TestCudnnCase1(TestPool2d_Op): def init_op_type(self): self.op_type = "pool2d_cudnn" - def init_pool_type(self): - self.pool_type = "avg" - - -class TestCaseCudnn2(TestPool2d_Op): - def init_test_case(self): - self.global_pool = False - self.pool2D_forward_naive = avg_pool2D_forward_naive - self.shape = [2, 3, 7, 7] - self.ksize = [3, 3] - self.strides = [1, 1] - self.paddings = [0, 0] +class TestCudnnCase2(TestCase1): def init_op_type(self): self.op_type = "pool2d_cudnn" - def init_pool_type(self): - self.pool_type = "avg" - - -class TestCaseCudnn3(TestPool2d_Op): - def init_test_case(self): - self.global_pool = False - self.pool2D_forward_naive = avg_pool2D_forward_naive - self.shape = [2, 3, 7, 7] - self.ksize = [3, 3] - self.strides = [1, 1] - self.paddings = [1, 1] +class TestCudnnCase3(TestCase2): def init_op_type(self): self.op_type = "pool2d_cudnn" - def init_pool_type(self): - self.pool_type = "avg" - - -class TestCaseCudnn4(TestPool2d_Op): - def init_test_case(self): - self.global_pool = True - self.pool2D_forward_naive = max_pool2D_forward_naive - self.shape = [2, 3, 5, 5] - self.ksize = [3, 3] - self.strides = [1, 1] - self.paddings = [0, 0] +class TestCudnnCase4(TestCase3): def init_op_type(self): self.op_type = "pool2d_cudnn" - def init_pool_type(self): - self.pool_type = "max" - - -class TestCaseCudnn5(TestPool2d_Op): - def init_test_case(self): - self.global_pool = False - self.pool2D_forward_naive = max_pool2D_forward_naive - self.shape = [2, 3, 7, 7] - self.ksize = [3, 3] - self.strides = [1, 1] - self.paddings = [0, 0] +class TestCudnnCase5(TestCase4): def init_op_type(self): self.op_type = "pool2d_cudnn" - def init_pool_type(self): - self.pool_type = "max" - - -class TestCaseCudnn6(TestPool2d_Op): - def init_test_case(self): - self.global_pool = False - self.pool2D_forward_naive = max_pool2D_forward_naive - self.shape = [2, 3, 7, 7] - self.ksize = [3, 3] - self.strides = [1, 1] - self.paddings = [1, 1] +class TestCudnnCase6(TestCase5): def init_op_type(self): self.op_type = "pool2d_cudnn" - def init_pool_type(self): - self.pool_type = "max" - if __name__ == '__main__': unittest.main() diff --git a/python/paddle/v2/fluid/tests/test_pool3d_op.py b/python/paddle/v2/fluid/tests/test_pool3d_op.py index 87483ae5e568c01141ff789f37e84069cb8e827d..2ba86665a7d207e61159c02643fa40daca3be080 100644 --- a/python/paddle/v2/fluid/tests/test_pool3d_op.py +++ b/python/paddle/v2/fluid/tests/test_pool3d_op.py @@ -3,8 +3,7 @@ import numpy as np from op_test import OpTest -def max_pool3D_forward_naive(x, ksize, strides, paddings=[0, 0], global_pool=0): - +def max_pool3D_forward_naive(x, ksize, strides, paddings, global_pool=0): N, C, D, H, W = x.shape if global_pool == 1: ksize = [D, H, W] @@ -27,8 +26,7 @@ def max_pool3D_forward_naive(x, ksize, strides, paddings=[0, 0], global_pool=0): return out -def avg_pool3D_forward_naive(x, ksize, strides, paddings=[0, 0], global_pool=0): - +def avg_pool3D_forward_naive(x, ksize, strides, paddings, global_pool=0): N, C, D, H, W = x.shape if global_pool == 1: ksize = [D, H, W] @@ -55,6 +53,10 @@ def avg_pool3D_forward_naive(x, ksize, strides, paddings=[0, 0], global_pool=0): class TestPool3d_Op(OpTest): def setUp(self): self.init_test_case() + self.init_global_pool() + self.init_op_type() + self.init_pool_type() + if self.global_pool: self.paddings = [0 for _ in range(len(self.paddings))] input = np.random.random(self.shape).astype("float32") @@ -81,74 +83,115 @@ class TestPool3d_Op(OpTest): self.check_grad(set(['X']), 'Out', max_relative_error=0.07) def init_test_case(self): - self.global_pool = True - self.op_type = "pool3d" - self.pool_type = "avg" - self.pool3D_forward_naive = avg_pool3D_forward_naive self.shape = [2, 3, 5, 5, 5] 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 + + def init_global_pool(self): + self.global_pool = True + class TestCase1(TestPool3d_Op): def init_test_case(self): - self.global_pool = False self.op_type = "pool3d" - self.pool_type = "avg" - self.pool3D_forward_naive = avg_pool3D_forward_naive self.shape = [2, 3, 7, 7, 7] self.ksize = [3, 3, 3] self.strides = [1, 1, 1] self.paddings = [0, 0, 0] - -class TestCase2(TestPool3d_Op): - def init_test_case(self): - self.global_pool = False + 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 + + def init_global_pool(self): + self.global_pool = False + + +class TestCase2(TestPool3d_Op): + def init_test_case(self): self.shape = [2, 3, 7, 7, 7] self.ksize = [3, 3, 3] 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 + + def init_global_pool(self): + self.global_pool = False + class TestCase3(TestPool3d_Op): - def init_test_case(self): - self.global_pool = True + 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 - self.shape = [2, 3, 5, 5, 5] - self.ksize = [3, 3, 3] - self.strides = [1, 1, 1] - self.paddings = [0, 0, 0] -class TestCase4(TestPool3d_Op): - def init_test_case(self): - self.global_pool = False +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 - self.shape = [2, 3, 7, 7, 7] - self.ksize = [3, 3, 3] - self.strides = [1, 1, 1] - self.paddings = [0, 0, 0] -class TestCase5(TestPool3d_Op): - def init_test_case(self): - self.global_pool = False +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 - self.shape = [2, 3, 7, 7, 7] - self.ksize = [3, 3, 3] - self.strides = [1, 1, 1] - self.paddings = [1, 1, 1] + + +#--------------------test pool3d_cudnn-------------------- +class TestCudnnCase1(TestPool3d_Op): + def init_op_type(self): + self.op_type = "pool3d_cudnn" + + +class TestCudnnCase2(TestCase1): + def init_op_type(self): + self.op_type = "pool3d_cudnn" + + +class TestCudnnCase3(TestCase2): + def init_op_type(self): + self.op_type = "pool3d_cudnn" + + +class TestCudnnCase4(TestCase3): + def init_op_type(self): + self.op_type = "pool3d_cudnn" + + +class TestCudnnCase5(TestCase4): + def init_op_type(self): + self.op_type = "pool3d_cudnn" + + +class TestCudnnCase6(TestCase5): + def init_op_type(self): + self.op_type = "pool3d_cudnn" if __name__ == '__main__': diff --git a/python/paddle/v2/fluid/tests/test_pool_max_op.py b/python/paddle/v2/fluid/tests/test_pool_max_op.py index 04843a28ac19e076e097d1aa1034bcf9378aa495..9d2d61c43868701392e90542f3b7fb2c4ea07548 100644 --- a/python/paddle/v2/fluid/tests/test_pool_max_op.py +++ b/python/paddle/v2/fluid/tests/test_pool_max_op.py @@ -3,11 +3,13 @@ import numpy as np from op_test import OpTest -def max_pool3D_forward_naive(x, ksize, strides, paddings, global_pool=0): +def max_pool3D_forward_naive(x, ksize, strides, paddings, global_pool=False): N, C, D, H, W = x.shape - if global_pool == 1: + if global_pool: ksize = [D, H, W] + paddings = [0, 0, 0] + D_out = (D - ksize[0] + 2 * paddings[0]) / strides[0] + 1 H_out = (H - ksize[1] + 2 * paddings[1]) / strides[1] + 1 W_out = (W - ksize[2] + 2 * paddings[2]) / strides[2] + 1 @@ -40,11 +42,13 @@ def max_pool3D_forward_naive(x, ksize, strides, paddings, global_pool=0): return out, mask -def max_pool2D_forward_naive(x, ksize, strides, paddings, global_pool=0): +def max_pool2D_forward_naive(x, ksize, strides, paddings, global_pool=False): N, C, H, W = x.shape - if global_pool == 1: + if global_pool: ksize = [H, W] + paddings = [0, 0] + H_out = (H - ksize[0] + 2 * paddings[0]) / strides[0] + 1 W_out = (W - ksize[1] + 2 * paddings[1]) / strides[1] + 1 out = np.zeros((N, C, H_out, W_out)) @@ -74,13 +78,13 @@ def max_pool2D_forward_naive(x, ksize, strides, paddings, global_pool=0): class TestMaxPoolWithIndex_Op(OpTest): def setUp(self): self.init_test_case() - if self.global_pool: - self.paddings = [0 for _ in range(len(self.paddings))] + self.init_global() + input = np.random.random(self.shape).astype("float32") output, mask = self.pool_forward_naive(input, self.ksize, self.strides, self.paddings, self.global_pool) output = output.astype("float32") - mask = mask.astype("float32") + mask = mask.astype("int32") self.attrs = { 'strides': self.strides, @@ -99,41 +103,24 @@ class TestMaxPoolWithIndex_Op(OpTest): # self.check_grad(set(['X']), ['Out'], max_relative_error=0.07) def init_test_case(self): - self.global_pool = True - self.index = "max_pool3d_with_index" - self.op_type = "%s" % self.index + self.op_type = "max_pool3d_with_index" self.pool_forward_naive = max_pool3D_forward_naive self.shape = [2, 3, 5, 5, 5] self.ksize = [3, 3, 3] self.strides = [1, 1, 1] self.paddings = [1, 1, 1] + def init_global(self): + self.global_pool = False + class TestCase1(TestMaxPoolWithIndex_Op): - def init_test_case(self): + def init_global(self): self.global_pool = True - self.op_type = "max_pool3d_with_index" - self.pool_forward_naive = max_pool3D_forward_naive - self.shape = [2, 3, 5, 5, 5] - self.ksize = [3, 3, 3] - self.strides = [1, 1, 1] - self.paddings = [1, 1, 1] class TestCase2(TestMaxPoolWithIndex_Op): def init_test_case(self): - self.global_pool = False - self.op_type = "max_pool3d_with_index" - self.pool_forward_naive = max_pool3D_forward_naive - self.shape = [2, 3, 7, 7, 7] - self.ksize = [3, 3, 3] - self.strides = [1, 1, 1] - self.paddings = [1, 1, 1] - - -class TestCase3(TestMaxPoolWithIndex_Op): - def init_test_case(self): - self.global_pool = False self.op_type = "max_pool3d_with_index" self.pool_forward_naive = max_pool3D_forward_naive self.shape = [2, 3, 7, 7, 7] @@ -141,32 +128,18 @@ class TestCase3(TestMaxPoolWithIndex_Op): self.strides = [2, 2, 2] self.paddings = [0, 0, 0] - -class TestCase4(TestMaxPoolWithIndex_Op): - def init_test_case(self): + def init_global(self): self.global_pool = True - self.op_type = "max_pool3d_with_index" - self.pool_forward_naive = max_pool3D_forward_naive - self.shape = [2, 3, 5, 5, 5] - self.ksize = [3, 3, 3] - self.strides = [1, 1, 1] - self.paddings = [1, 1, 1] -class TestCase5(TestMaxPoolWithIndex_Op): - def init_test_case(self): - self.global_pool = True - self.op_type = "max_pool3d_with_index" - self.pool_forward_naive = max_pool3D_forward_naive - self.shape = [2, 3, 5, 5, 5] - self.ksize = [3, 3, 3] - self.strides = [2, 2, 2] - self.paddings = [0, 0, 0] +class TestCase3(TestCase2): + def init_global(self): + self.global_pool = False -class TestCase6(TestMaxPoolWithIndex_Op): +#----------------max_pool2d_with_index---------------- +class TestCase4(TestMaxPoolWithIndex_Op): def init_test_case(self): - self.global_pool = False self.op_type = "max_pool2d_with_index" self.pool_forward_naive = max_pool2D_forward_naive self.shape = [2, 3, 7, 7] @@ -174,10 +147,17 @@ class TestCase6(TestMaxPoolWithIndex_Op): self.strides = [1, 1] self.paddings = [1, 1] + def init_global(self): + self.global_pool = True + -class TestCase7(TestMaxPoolWithIndex_Op): - def init_test_case(self): +class TestCase5(TestCase4): + def init_global(self): self.global_pool = False + + +class TestCase6(TestMaxPoolWithIndex_Op): + def init_test_case(self): self.op_type = "max_pool2d_with_index" self.pool_forward_naive = max_pool2D_forward_naive self.shape = [2, 3, 7, 7] @@ -185,27 +165,13 @@ class TestCase7(TestMaxPoolWithIndex_Op): self.strides = [2, 2] self.paddings = [0, 0] - -class TestCase8(TestMaxPoolWithIndex_Op): - def init_test_case(self): + def init_global(self): self.global_pool = True - self.op_type = "max_pool2d_with_index" - self.pool_forward_naive = max_pool2D_forward_naive - self.shape = [2, 3, 5, 5] - self.ksize = [3, 3] - self.strides = [1, 1] - self.paddings = [1, 1] -class TestCase9(TestMaxPoolWithIndex_Op): - def init_test_case(self): - self.global_pool = True - self.op_type = "max_pool2d_with_index" - self.pool_forward_naive = max_pool2D_forward_naive - self.shape = [2, 3, 5, 5] - self.ksize = [3, 3] - self.strides = [2, 2] - self.paddings = [0, 0] +class TestCase7(TestCase6): + def init_global(self): + self.global_pool = False if __name__ == '__main__': diff --git a/python/paddle/v2/fluid/tests/test_profiler.py b/python/paddle/v2/fluid/tests/test_profiler.py new file mode 100644 index 0000000000000000000000000000000000000000..e3f3ac58ef9b30864849770510f7339749dab84f --- /dev/null +++ b/python/paddle/v2/fluid/tests/test_profiler.py @@ -0,0 +1,31 @@ +import unittest +import numpy as np +import paddle.v2.fluid as fluid +import paddle.v2.fluid.profiler as profiler +import paddle.v2.fluid.layers as layers +import os + + +class TestProfiler(unittest.TestCase): + def test_nvprof(self): + if not fluid.core.is_compile_gpu(): + return + epoc = 8 + dshape = [4, 3, 28, 28] + data = layers.data(name='data', shape=[3, 28, 28], dtype='float32') + conv = layers.conv2d(data, 20, 3, stride=[1, 1], padding=[1, 1]) + + place = fluid.CUDAPlace(0) + exe = fluid.Executor(place) + exe.run(fluid.default_startup_program()) + + output_file = 'cuda_profiler.txt' + with profiler.cuda_profiler(output_file, 'csv') as nvprof: + for i in range(epoc): + input = np.random.random(dshape).astype('float32') + exe.run(fluid.default_main_program(), feed={'data': input}) + os.remove(output_file) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/v2/fluid/tests/test_program.py b/python/paddle/v2/fluid/tests/test_program.py index e9bcefd21569aaa9225c676ea03b5c8e37d00333..447c746aacc1c9455d7a023bca625d548ab2638b 100644 --- a/python/paddle/v2/fluid/tests/test_program.py +++ b/python/paddle/v2/fluid/tests/test_program.py @@ -1,35 +1,38 @@ +from __future__ import print_function import unittest -from paddle.v2.fluid.framework import Program -from paddle.v2.fluid.framework import g_main_program +from paddle.v2.fluid.framework import Program, default_main_program, program_guard, grad_var_name +import paddle.v2.fluid.layers as layers + +main_program = default_main_program() class TestProgram(unittest.TestCase): def test_program(self): - b = g_main_program.current_block() + b = main_program.current_block() self.assertEqual(-1, b.parent_idx) self.assertEqual(0, b.idx) - b = g_main_program.create_block() + b = main_program.create_block() self.assertEqual(1, b.idx) self.assertEqual(0, b.parent_idx) - b = g_main_program.create_block() + b = main_program.create_block() self.assertEqual(2, b.idx) self.assertEqual(1, b.parent_idx) - g_main_program.rollback() + main_program.rollback() - b = g_main_program.current_block() + b = main_program.current_block() self.assertEqual(1, b.idx) self.assertEqual(0, b.parent_idx) - b = g_main_program.create_block() + b = main_program.create_block() self.assertEqual(3, b.idx) self.assertEqual(1, b.parent_idx) - g_main_program.rollback() - b = g_main_program.current_block() + main_program.rollback() + b = main_program.current_block() self.assertEqual(1, b.idx) self.assertEqual(0, b.parent_idx) @@ -48,8 +51,8 @@ class TestProgram(unittest.TestCase): # FIXME(yuyang18): We manual compare the output string, since the order # of variable could be changed. - print prog - print prog.clone() + print(prog) + print(prog.clone()) def test_parse_program_from_string(self): prog = Program() @@ -67,8 +70,8 @@ class TestProgram(unittest.TestCase): binary_str = prog.desc.serialize_to_string() prog_restored = Program.parse_from_string(binary_str) - print prog - print prog_restored + print(prog) + print(prog_restored) def test_append_backward(self): prog = Program() @@ -106,12 +109,10 @@ class TestProgram(unittest.TestCase): self.assertEqual(add_op.idx, 1) param_to_grad = prog.append_backward(mean_out, set()) - def grad_name(name): - return name + "@GRAD" - for var_name in ("mul.x", "mul.y", "mul.out", "add.y", "add.out", "mean.out"): - self.assertEqual(param_to_grad[var_name][0], grad_name(var_name)) + self.assertEqual(param_to_grad[var_name][0], + grad_var_name(var_name)) self.assertEqual(param_to_grad[var_name][1], 0) expect_ops = [ @@ -123,6 +124,17 @@ class TestProgram(unittest.TestCase): actual_ops.append(op.type) self.assertEqual(actual_ops, expect_ops) + def test_program_clone_with_parameter(self): + main_program = Program() + startup_program = Program() + with program_guard(main_program, startup_program): + d = layers.data(name='x', shape=[784], dtype='float32') + hidden = layers.fc(input=d, size=100) + layers.fc(input=hidden, size=100) + + new_program = main_program.clone() + self.assertNotEqual(0, len(new_program.blocks[0].all_parameters())) + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/v2/fluid/tests/test_protobuf_descs.py b/python/paddle/v2/fluid/tests/test_protobuf_descs.py index 098a9802dfc6763ce2a2356b7267a439145b7939..d8abe17606c4ddb2ff51d5f918b1e5d7e110f7fa 100644 --- a/python/paddle/v2/fluid/tests/test_protobuf_descs.py +++ b/python/paddle/v2/fluid/tests/test_protobuf_descs.py @@ -101,13 +101,13 @@ class TestVarDesc(unittest.TestCase): self.assertEqual(src_shape, res_shape) self.assertEqual(core.VarDesc.VarType.SELECTED_ROWS, var.type()) - def test_data_type(self): + def test_dtype(self): program_desc = core.ProgramDesc() block = program_desc.block(0) var = block.var('my_var') var.set_type(core.VarDesc.VarType.LOD_TENSOR) - var.set_data_type(core.DataType.INT32) - self.assertEqual(core.DataType.INT32, var.data_type()) + var.set_dtype(core.DataType.INT32) + self.assertEqual(core.DataType.INT32, var.dtype()) self.assertEqual(core.VarDesc.VarType.LOD_TENSOR, var.type()) diff --git a/python/paddle/v2/fluid/tests/test_recurrent_op.py b/python/paddle/v2/fluid/tests/test_recurrent_op.py index b623d1231838faff9e91c9234befb1f647fe8ec2..84f4e36fa7312fbcb96cc66ff26e234c3016df30 100644 --- a/python/paddle/v2/fluid/tests/test_recurrent_op.py +++ b/python/paddle/v2/fluid/tests/test_recurrent_op.py @@ -1,9 +1,9 @@ import unittest import paddle.v2.fluid.layers as layers -from paddle.v2.fluid.framework import Program +from paddle.v2.fluid.framework import Program, grad_var_name from paddle.v2.fluid.executor import Executor -from paddle.v2.fluid.backward import append_backward_ops +from paddle.v2.fluid.backward import append_backward import numpy as np import paddle.v2.fluid.core as core @@ -118,14 +118,14 @@ class RecurrentOpTest1(unittest.TestCase): def create_rnn_op(self): x = layers.data( shape=[self.sent_len, self.batch_size, self.input_dim], - data_type='float32', + dtype='float32', name='x', append_batch_size=False, **self.p_info) x.stop_gradient = False h_boot = layers.data( shape=[self.input_dim], - data_type='float32', + dtype='float32', name='h_boot', **self.p_info) h_boot.stop_gradient = False @@ -156,7 +156,7 @@ class RecurrentOpTest1(unittest.TestCase): feed=self.feed_map, fetch_list=[self.output]) - return np.array(out[0]) + return out[0] def backward(self): self.feed_map = { @@ -164,19 +164,20 @@ class RecurrentOpTest1(unittest.TestCase): for x in self.data_field } fetch_list = [ - self.main_program.global_block().var(x + "@GRAD") + self.main_program.global_block().var(grad_var_name(x)) for x in self.data_field ] exe = Executor(self.place) return exe.run(self.main_program, feed=self.feed_map, - fetch_list=fetch_list) + fetch_list=fetch_list, + return_numpy=False) def test_backward(self): self.check_forward() - append_backward_ops(self.output) + append_backward(self.output) ana_grad = [np.array(x) for x in self.backward()] @@ -251,14 +252,14 @@ class RecurrentOpTest2(RecurrentOpTest1): def create_rnn_op(self): x = layers.data( shape=[self.sent_len, self.batch_size, self.input_dim], - data_type='float32', + dtype='float32', name='x', append_batch_size=False, **self.p_info) x.stop_gradient = False h_boot = layers.data( shape=[self.input_dim], - data_type='float32', + dtype='float32', name='h_boot', **self.p_info) h_boot.stop_gradient = False @@ -270,12 +271,12 @@ class RecurrentOpTest2(RecurrentOpTest1): temp_l = layers.fc(input=x_t, size=self.input_dim, - param_attr={'name': 'W'}, + param_attr='W', bias_attr=False, **self.p_info) temp_r = layers.fc(input=h_pre, size=self.input_dim, - param_attr={'name': 'U'}, + param_attr='U', bias_attr=False, **self.p_info) @@ -350,21 +351,21 @@ class RecurrentOpMultipleMemoryTest(RecurrentOpTest1): def create_rnn_op(self): x = layers.data( shape=[self.sent_len, self.batch_size, self.input_dim], - data_type='float32', + dtype='float32', name='x', append_batch_size=False, **self.p_info) x.stop_gradient = False h_boot1 = layers.data( shape=[self.batch_size, self.input_dim], - data_type='float32', + dtype='float32', name='h_boot1', append_batch_size=False, **self.p_info) h_boot1.stop_gradient = False h_boot2 = layers.data( shape=[self.batch_size, self.input_dim], - data_type='float32', + dtype='float32', name='h_boot2', append_batch_size=False, **self.p_info) @@ -435,7 +436,7 @@ class RecurrentOpNoMemBootTest(RecurrentOpTest1): def create_rnn_op(self): x = layers.data( shape=[self.sent_len, self.batch_size, self.input_dim], - data_type='float32', + dtype='float32', name='x', append_batch_size=False, **self.p_info) @@ -453,4 +454,6 @@ class RecurrentOpNoMemBootTest(RecurrentOpTest1): if __name__ == '__main__': + # FIXME(qijun) https://github.com/PaddlePaddle/Paddle/issues/6152 + exit(0) unittest.main() diff --git a/python/paddle/v2/fluid/tests/test_reduce_op.py b/python/paddle/v2/fluid/tests/test_reduce_op.py index 70359d60cbe656150877673c63e81eae92d8ab9a..a021d4dd91bb9cc1e5d85411b3813b966ef5b296 100644 --- a/python/paddle/v2/fluid/tests/test_reduce_op.py +++ b/python/paddle/v2/fluid/tests/test_reduce_op.py @@ -85,5 +85,19 @@ class Test1DReduce(OpTest): self.check_grad(['X'], 'Out') +class TestReduceAll(OpTest): + def setUp(self): + self.op_type = "reduce_sum" + self.inputs = {'X': np.random.random((5, 6, 2, 10)).astype("float32")} + self.attrs = {'reduce_all': True} + self.outputs = {'Out': self.inputs['X'].sum()} + + 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/v2/fluid/tests/test_registry.py b/python/paddle/v2/fluid/tests/test_registry.py new file mode 100644 index 0000000000000000000000000000000000000000..f8328f31cf8203f5ea8af2c14417879616ccab71 --- /dev/null +++ b/python/paddle/v2/fluid/tests/test_registry.py @@ -0,0 +1,22 @@ +import unittest +import warnings + +import paddle.v2.fluid as fluid +import paddle.v2.fluid.framework as framework +import paddle.v2.fluid.layers as layers +import paddle.v2.fluid.registry as registry + + +class TestRegistry(unittest.TestCase): + def test_registry_layer(self): + self.layer_type = "mean" + program = framework.Program() + + x = fluid.layers.data(name='X', shape=[10, 10], dtype='float32') + output = layers.mean(x) + place = fluid.CPUPlace() + exe = fluid.Executor(place) + + X = np.random.random((10, 10)).astype("float32") + mean_out = exe.run(program, feed={"X": X}, fetch_list=[output]) + self.assertAlmostEqual(np.mean(X), mean_out) diff --git a/python/paddle/v2/fluid/tests/test_regularizer.py b/python/paddle/v2/fluid/tests/test_regularizer.py index 24baf55e90c98f39bab926e8c85a791eee5ed4a4..890c881a126a32344128652691c6cad45e02e82d 100644 --- a/python/paddle/v2/fluid/tests/test_regularizer.py +++ b/python/paddle/v2/fluid/tests/test_regularizer.py @@ -3,7 +3,7 @@ import unittest import paddle.v2.fluid.framework as framework import paddle.v2.fluid.optimizer as optimizer import paddle.v2.fluid.regularizer as regularizer -from paddle.v2.fluid.backward import append_backward_ops +from paddle.v2.fluid.backward import append_backward class TestL2DecayRegularizer(unittest.TestCase): @@ -33,7 +33,7 @@ class TestL2DecayRegularizer(unittest.TestCase): dtype="float32", shape=[1], lod_level=0, name="mean.out") block.append_op( type="mean", inputs={"X": mul_out}, outputs={"Out": mean_out}) - params_grads = append_backward_ops(mean_out) + params_grads = append_backward(mean_out) self.assertEqual(len(params_grads), 1) count_ops = len(block.ops) params_grads = optimizer.append_regularization_ops(params_grads) @@ -70,7 +70,7 @@ class TestL1DecayRegularizer(unittest.TestCase): dtype="float32", shape=[1], lod_level=0, name="mean.out") block.append_op( type="mean", inputs={"X": mul_out}, outputs={"Out": mean_out}) - params_grads = append_backward_ops(mean_out) + params_grads = append_backward(mean_out) self.assertEqual(len(params_grads), 1) count_ops = len(block.ops) params_grads = optimizer.append_regularization_ops(params_grads) diff --git a/python/paddle/v2/fluid/tests/test_reorder_lod_tensor.py b/python/paddle/v2/fluid/tests/test_reorder_lod_tensor.py new file mode 100644 index 0000000000000000000000000000000000000000..7c136f6360ce73a7c532b5486e544796e6853bcb --- /dev/null +++ b/python/paddle/v2/fluid/tests/test_reorder_lod_tensor.py @@ -0,0 +1,47 @@ +import unittest +import paddle.v2.fluid as fluid +import numpy + + +class TestReorderLoDTensor(unittest.TestCase): + def test_reorder(self): + dat = fluid.layers.data(name='input', shape=[1], lod_level=2) + dat.stop_gradient = False + rank_dat = fluid.layers.data(name='ref', shape=[1], lod_level=1) + table = fluid.layers.lod_rank_table(rank_dat) + new_dat = fluid.layers.reorder_lod_tensor_by_rank( + x=dat, rank_table=table) + loss = fluid.layers.mean(x=new_dat) + fluid.backward.append_backward(loss=loss) + + cpu = fluid.CPUPlace() + exe = fluid.Executor(cpu) + exe.run(fluid.default_startup_program()) + + ref = fluid.Tensor() + ref_lod = [0, 3, 4, 7, 8, 14] + ref.set_lod([ref_lod]) + + ref.set(numpy.random.random(size=[14, 1]).astype('float32'), cpu) + input = fluid.Tensor() + lod_level_0 = numpy.random.randint(low=1, high=5, size=5) + lod_level_0 = [0] + numpy.cumsum(lod_level_0).tolist() + lod_level_1 = numpy.random.randint(low=1, high=5, size=lod_level_0[-1]) + lod_level_1 = [0] + numpy.cumsum(lod_level_1).tolist() + + input.set_lod([lod_level_0, lod_level_1]) + input.set( + numpy.random.random(size=[lod_level_1[-1], 1]).astype('float32'), + cpu) + + ig = exe.run(fluid.default_main_program(), + feed={'input': input, + 'ref': ref}, + fetch_list=['input@GRAD'], + return_numpy=False)[0] + self.assertAlmostEqual(numpy.array(ig).sum(), 1.0, delta=0.001) + self.assertEqual(input.lod(), ig.lod()) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/v2/fluid/tests/test_reshape_op.py b/python/paddle/v2/fluid/tests/test_reshape_op.py index 16bb6bb2af67f7d32a2fafc1cb37412084ec0829..18ee3aece656276fec9671df9baf298b7fd3c9b1 100644 --- a/python/paddle/v2/fluid/tests/test_reshape_op.py +++ b/python/paddle/v2/fluid/tests/test_reshape_op.py @@ -17,5 +17,19 @@ class TestReshapeOp(OpTest): self.check_grad(["X"], "Out") +class TestReshapeOpDimInfer(OpTest): + def setUp(self): + self.op_type = "reshape" + self.inputs = {'X': np.random.random((10, 20)).astype("float32")} + self.attrs = {'shape': [4, -1, 5]} + self.outputs = {'Out': self.inputs['X'].reshape(self.attrs['shape'])} + + 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/v2/fluid/tests/test_rnn_memory_helper_op.py b/python/paddle/v2/fluid/tests/test_rnn_memory_helper_op.py index a3cba92504a28590083df57e69f7662a887d94a6..d1bb20f37a3785f70bee072b9df282bba4012c16 100644 --- a/python/paddle/v2/fluid/tests/test_rnn_memory_helper_op.py +++ b/python/paddle/v2/fluid/tests/test_rnn_memory_helper_op.py @@ -2,17 +2,11 @@ import unittest from paddle.v2.fluid.framework import Program from paddle.v2.fluid.executor import Executor -from paddle.v2.fluid.backward import append_backward_ops +from paddle.v2.fluid.backward import append_backward import numpy as np import paddle.v2.fluid.core as core -def create_tensor(np_data, place): - tensor = core.LoDTensor() - tensor.set(np_data, place) - return tensor - - class RNNMemoryHelperOpTest(unittest.TestCase): def setUp(self): self.program = Program() @@ -30,13 +24,13 @@ class RNNMemoryHelperOpTest(unittest.TestCase): def test_forward(self): x_np = np.random.normal(size=(2, 3)).astype("float32") - self.feed_map = {'X': create_tensor(x_np, self.place)} + self.feed_map = {'X': x_np} self.fetch_list = [self.Out] exe = Executor(self.place) out = exe.run(self.program, feed=self.feed_map, fetch_list=self.fetch_list) - np.isclose(np.array(out[0]), x_np, rtol=1e-5) + self.assertTrue(np.allclose(out[0], x_np, rtol=1e-5)) class RNNMemoryHelperGradOpTest(unittest.TestCase): @@ -66,8 +60,7 @@ class RNNMemoryHelperGradOpTest(unittest.TestCase): def test_backward(self): self.feed_map = { - name: create_tensor( - np.random.normal(size=(2, 3)).astype("float32"), self.place) + name: np.random.normal(size=(2, 3)).astype("float32") for name in self.input_names } self.fetch_list = [self.output_vars['X@GRAD']] @@ -76,7 +69,7 @@ class RNNMemoryHelperGradOpTest(unittest.TestCase): out = exe.run(self.program, feed=self.feed_map, fetch_list=self.fetch_list) - np.isclose(np.array(out[0]), self.feed_map['Out@GRAD'], rtol=1e-5) + np.isclose(out[0], self.feed_map['Out@GRAD'], rtol=1e-5) class RNNMemoryHelperGradOpWithoutInputTest(unittest.TestCase): @@ -110,8 +103,7 @@ class RNNMemoryHelperGradOpWithoutInputTest(unittest.TestCase): def test_backward(self): self.feed_map = { - name: create_tensor( - np.random.normal(size=(2, 3)).astype("float32"), self.place) + name: np.random.normal(size=(2, 3)).astype("float32") for name in ['X', 'Out'] } self.fetch_list = [self.output_vars['X@GRAD']] @@ -120,10 +112,9 @@ class RNNMemoryHelperGradOpWithoutInputTest(unittest.TestCase): out = exe.run(self.program, feed=self.feed_map, fetch_list=self.fetch_list) - np.isclose( - np.array(out[0]), - np.zeros(shape=(2, 3)).astype("float32"), - rtol=1e-5) + self.assertTrue( + np.allclose( + out[0], np.zeros(shape=(2, 3)).astype("float32"), rtol=1e-5)) if __name__ == '__main__': diff --git a/python/paddle/v2/fluid/tests/test_roi_pool_op.py b/python/paddle/v2/fluid/tests/test_roi_pool_op.py new file mode 100644 index 0000000000000000000000000000000000000000..a28d9c7f82d3735c410369eb61e350168c267cea --- /dev/null +++ b/python/paddle/v2/fluid/tests/test_roi_pool_op.py @@ -0,0 +1,123 @@ +import unittest +import numpy as np +import math +import sys +from op_test import OpTest + + +class TestROIPoolOp(OpTest): + def set_data(self): + self.init_test_case() + self.make_rois() + self.calc_roi_pool() + + self.inputs = {'X': self.x, 'ROIs': self.rois} + + self.attrs = { + 'spatial_scale': self.spatial_scale, + 'pooled_height': self.pooled_height, + 'pooled_width': self.pooled_width + } + + self.outputs = {'Out': self.outs, 'Argmax': self.argmaxes} + + def init_test_case(self): + self.batch_size = 5 + self.channels = 3 + self.height = 6 + self.width = 4 + + # n, c, h, w + self.x_dim = (self.batch_size, self.channels, self.height, self.width) + + 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') + + def calc_roi_pool(self): + out_data = np.zeros((self.rois_num, self.channels, self.pooled_height, + self.pooled_width)) + argmax_data = np.zeros((self.rois_num, self.channels, + self.pooled_height, self.pooled_width)) + + for i in range(self.rois_num): + roi = self.rois[i] + roi_batch_id = roi[0] + roi_start_w = int(round(roi[1] * self.spatial_scale)) + roi_start_h = int(round(roi[2] * self.spatial_scale)) + roi_end_w = int(round(roi[3] * self.spatial_scale)) + roi_end_h = int(round(roi[4] * self.spatial_scale)) + + roi_height = int(max(roi_end_h - roi_start_h + 1, 1)) + roi_width = int(max(roi_end_w - roi_start_w + 1, 1)) + + x_i = self.x[roi_batch_id] + + bin_size_h = float(roi_height) / float(self.pooled_height) + bin_size_w = float(roi_width) / float(self.pooled_width) + + for c in range(self.channels): + for ph in range(self.pooled_height): + for pw in range(self.pooled_width): + hstart = int(math.floor(ph * bin_size_h)) + wstart = int(math.floor(pw * bin_size_w)) + hend = int(math.ceil((ph + 1) * bin_size_h)) + wend = int(math.ceil((pw + 1) * bin_size_w)) + + hstart = min(max(hstart + roi_start_h, 0), self.height) + hend = min(max(hend + roi_start_h, 0), self.height) + wstart = min(max(wstart + roi_start_w, 0), self.width) + wend = min(max(wend + roi_start_w, 0), self.width) + + is_empty = (hend <= hstart) or (wend <= wstart) + if is_empty: + out_data[i, c, ph, pw] = 0 + else: + out_data[i, c, ph, pw] = -sys.float_info.max + + argmax_data[i, c, ph, pw] = -1 + + for h in range(hstart, hend): + for w in range(wstart, wend): + if x_i[c, h, w] > out_data[i, c, ph, pw]: + out_data[i, c, ph, pw] = x_i[c, h, w] + argmax_data[i, c, ph, pw] = h * \ + self.width + w + + self.outs = out_data.astype('float32') + self.argmaxes = argmax_data.astype('int64') + + 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 = np.array(rois).astype("int64") + + def setUp(self): + self.op_type = "roi_pool" + self.set_data() + + 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/v2/fluid/tests/test_row_conv_op.py b/python/paddle/v2/fluid/tests/test_row_conv_op.py new file mode 100644 index 0000000000000000000000000000000000000000..1ed86e23ac28a575cdc3388e9da547918eb8a1be --- /dev/null +++ b/python/paddle/v2/fluid/tests/test_row_conv_op.py @@ -0,0 +1,95 @@ +import unittest +import numpy as np +from op_test import OpTest + + +def row_conv_forward(x, lod, wt): + out = np.zeros_like(x) + seq_info = lod[0] + num_sequences = len(seq_info) - 1 + context_length = wt.shape[0] + + for i in range(num_sequences): # loop over number of sequences + start = seq_info[i] + end = seq_info[i + 1] + curinput = x[start:end, :] + curoutput = out[start:end, :] + + cur_timesteps = end - start + for j in range(cur_timesteps): # loop over different timesteps + for k in range(context_length): + + if j + k >= cur_timesteps: + continue + curoutput[j, :] += curinput[j + k, :] * wt[k, :] + + return out + + +class TestRowConvOp1(OpTest): + def setUp(self): + + self.op_type = "row_conv" + lod = [[0, 2, 5, 7]] + T = lod[0][-1] + D = 16 + context_length = 2 + + x = np.random.random((T, D)).astype("float32") + wt = np.random.random((context_length, D)).astype("float32") + self.inputs = {'X': (x, lod), 'Filter': wt} + + out = row_conv_forward(x, lod, wt) + self.outputs = {'Out': (out, lod)} + + def test_check_output(self): + self.check_output() + + def test_check_grad_normal(self): + self.check_grad(['X', 'Filter'], 'Out', max_relative_error=0.05) + + def test_check_grad_ignore_x(self): + self.check_grad( + ['Filter'], 'Out', max_relative_error=0.05, no_grad_set=set('X')) + + def test_check_grad_ignore_wt(self): + self.check_grad( + ['X'], 'Out', max_relative_error=0.05, no_grad_set=set('Filter')) + + +class TestRowConvOp2(OpTest): + def setUp(self): + + self.op_type = "row_conv" + lod = [[0, 20, 50, 100]] + T = lod[0][-1] + D = 35 + context_length = 35 + + x = np.random.random((T, D)).astype("float32") + wt = np.random.random((context_length, D)).astype("float32") + self.inputs = {'X': (x, lod), 'Filter': wt} + + out = row_conv_forward(x, lod, wt) + self.outputs = {'Out': (out, lod)} + + def test_check_output(self): + self.check_output() + + #max_relative_error is increased from 0.05 to 0.06 as for higher + #dimensional input, the dX on CPU for some values has max_rel_error + #slightly more than 0.05 + def test_check_grad_normal(self): + self.check_grad(['X', 'Filter'], 'Out', max_relative_error=0.06) + + def test_check_grad_ignore_x(self): + self.check_grad( + ['Filter'], 'Out', max_relative_error=0.06, no_grad_set=set('X')) + + def test_check_grad_ignore_wt(self): + self.check_grad( + ['X'], 'Out', max_relative_error=0.06, no_grad_set=set('Filter')) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/v2/fluid/tests/test_seq_expand.py b/python/paddle/v2/fluid/tests/test_sequence_expand.py similarity index 89% rename from python/paddle/v2/fluid/tests/test_seq_expand.py rename to python/paddle/v2/fluid/tests/test_sequence_expand.py index ff17edd04bfd34ab8449a0ae05aacf66632dabc8..0f22612d3dbe483e4d5a8638636e44e172160156 100644 --- a/python/paddle/v2/fluid/tests/test_seq_expand.py +++ b/python/paddle/v2/fluid/tests/test_sequence_expand.py @@ -3,7 +3,7 @@ import numpy as np from op_test import OpTest -class TestSeqExpand(OpTest): +class TestSequenceExpand(OpTest): def set_data(self): x_data = np.random.uniform(0.1, 1, [3, 1]).astype('float32') y_data = np.random.uniform(0.1, 1, [8, 1]).astype('float32') @@ -21,7 +21,7 @@ class TestSeqExpand(OpTest): self.outputs = {'Out': out} def setUp(self): - self.op_type = 'seq_expand' + self.op_type = 'sequence_expand' self.set_data() self.compute() @@ -32,7 +32,7 @@ class TestSeqExpand(OpTest): self.check_grad(["X"], "Out") -class TestSeqExpandCase1(TestSeqExpand): +class TestSequenceExpandCase1(TestSequenceExpand): def set_data(self): x_data = np.random.uniform(0.1, 1, [5, 1]).astype('float32') x_lod = [[0, 2, 5]] @@ -41,7 +41,7 @@ class TestSeqExpandCase1(TestSeqExpand): self.inputs = {'X': (x_data, x_lod), 'Y': (y_data, y_lod)} -class TestSeqExpandCase2(TestSeqExpand): +class TestSequenceExpandCase2(TestSequenceExpand): def set_data(self): x_data = np.random.uniform(0.1, 1, [1, 2, 2]).astype('float32') x_lod = [[0, 1]] @@ -50,7 +50,7 @@ class TestSeqExpandCase2(TestSeqExpand): self.inputs = {'X': (x_data, x_lod), 'Y': (y_data, y_lod)} -class TestSeqExpandCase3(TestSeqExpand): +class TestSequenceExpandCase3(TestSequenceExpand): def set_data(self): x_data = np.random.uniform(0.1, 1, [4, 1]).astype('float32') x_lod = [[0, 1, 2, 3, 4]] diff --git a/python/paddle/v2/fluid/tests/test_sequence_slice_op.py b/python/paddle/v2/fluid/tests/test_sequence_slice_op.py old mode 100755 new mode 100644 index 4351d8e6d77c16e0012f9ae163b118fdbb793a8f..ccd9a05343b0c4aa05b258959665c0662f271512 --- a/python/paddle/v2/fluid/tests/test_sequence_slice_op.py +++ b/python/paddle/v2/fluid/tests/test_sequence_slice_op.py @@ -3,6 +3,7 @@ import numpy as np import sys from op_test import OpTest + class TestSequenceSliceOp(OpTest): def set_data(self): self.init_test_case() @@ -13,12 +14,12 @@ class TestSequenceSliceOp(OpTest): length = np.array(self.length).astype("int64") self.inputs = {'X': (x, lod), 'Offset': offset, 'Length': length} - outs = [] #np.zeros((100, 3, 2)).astype('float32') + outs = [] #np.zeros((100, 3, 2)).astype('float32') out_lod = [[0]] out_lod_offset = 0 for i in range(len(offset)): - sub_x = x[lod[0][i] + offset[i, 0]: lod[0] - [i] + offset[i, 0] + length[i, 0], :] + sub_x = x[lod[0][i] + offset[i, 0]:lod[0][i] + offset[i, 0] + + length[i, 0], :] out_lod_offset = out_lod_offset + len(sub_x) outs.append(sub_x) out_lod[0].append(out_lod_offset) @@ -41,5 +42,6 @@ class TestSequenceSliceOp(OpTest): def test_check_grad(self): self.check_grad(['X'], 'Out') + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/v2/fluid/tests/test_sgd_op.py b/python/paddle/v2/fluid/tests/test_sgd_op.py index ca05a381f06cfd40b7939dbda8d4f1f4aacd0271..14d41e172a22c677235ab3fa997ef6f0b6e39778 100644 --- a/python/paddle/v2/fluid/tests/test_sgd_op.py +++ b/python/paddle/v2/fluid/tests/test_sgd_op.py @@ -55,8 +55,7 @@ class TestSparseSGDOp(unittest.TestCase): Grad='Grad', ParamOut='Param', LearningRate='LearningRate') - ctx = core.DeviceContext.create(place) - sgd_op.run(scope, ctx) + sgd_op.run(scope, place) # get and compare result result_array = np.array(param) @@ -79,7 +78,7 @@ class TestSparseSGDOp(unittest.TestCase): def test_sparse_sgd(self): places = [core.CPUPlace()] if core.is_compile_gpu(): - places.append(core.GPUPlace(0)) + places.append(core.CUDAPlace(0)) for place in places: self.check_with_place(place) diff --git a/python/paddle/v2/fluid/tests/test_shrink_rnn_memory.py b/python/paddle/v2/fluid/tests/test_shrink_rnn_memory.py index 1a3b88e18e38b88d75ad17a0bb6a2965d1e60406..be1588fc2d09fa58882425eb3d080ef1560ebc79 100644 --- a/python/paddle/v2/fluid/tests/test_shrink_rnn_memory.py +++ b/python/paddle/v2/fluid/tests/test_shrink_rnn_memory.py @@ -2,14 +2,16 @@ import unittest import paddle.v2.fluid.core as core from paddle.v2.fluid.executor import Executor import paddle.v2.fluid.layers as layers -from paddle.v2.fluid.backward import append_backward_ops -from paddle.v2.fluid.framework import g_main_program +from paddle.v2.fluid.backward import append_backward +from paddle.v2.fluid.framework import default_main_program import numpy +main_program = default_main_program() + class TestShrinkRNNMemory(unittest.TestCase): def test_shrink_rnn_memory(self): - x = layers.data('x', shape=[100], data_type='float32') + x = layers.data('x', shape=[100], dtype='float32') x.stop_gradient = False table = layers.lod_rank_table(x=x) i = layers.zeros(dtype='int64', shape=[1]) @@ -27,19 +29,16 @@ class TestShrinkRNNMemory(unittest.TestCase): tensor_np = numpy.random.random(size=(3, 100)).astype('float32') tensor.set(tensor_np, cpu) exe = Executor(cpu) - outs = map(numpy.array, - exe.run(feed={'x': tensor}, fetch_list=[mem1, mem2, mem3])) + outs = exe.run(feed={'x': tensor}, fetch_list=[mem1, mem2, mem3]) self.assertTrue(numpy.allclose(tensor_np[0:3], outs[0])) self.assertTrue(numpy.allclose(tensor_np[0:2], outs[1])) self.assertTrue(numpy.allclose(tensor_np[0:1], outs[2])) mem3_mean = layers.mean(x=mem3) - append_backward_ops(loss=mem3_mean) - x_grad = map(numpy.array, - exe.run(feed={'x': tensor}, - fetch_list=[ - g_main_program.global_block().var('x@GRAD') - ]))[0] + append_backward(loss=mem3_mean) + x_grad = exe.run( + feed={'x': tensor}, + fetch_list=[main_program.global_block().var('x@GRAD')])[0] self.assertAlmostEqual(1.0, x_grad.sum(), delta=0.1) diff --git a/python/paddle/v2/fluid/tests/test_sigmoid_cross_entropy_with_logits_op.py b/python/paddle/v2/fluid/tests/test_sigmoid_cross_entropy_with_logits_op.py index e53856b38aa5ddd6061b350a66e9fe86bc23923c..c42f578f72cb121a24d6b852334cbd8a977f2730 100644 --- a/python/paddle/v2/fluid/tests/test_sigmoid_cross_entropy_with_logits_op.py +++ b/python/paddle/v2/fluid/tests/test_sigmoid_cross_entropy_with_logits_op.py @@ -2,11 +2,12 @@ import numpy as np from op_test import OpTest from scipy.special import logit from scipy.special import expit +import unittest class TestSigmoidCrossEntropyWithLogitsOp1(OpTest): - '''Test sigmoid_cross_entropy_with_logit_op with binary labels - ''' + """Test sigmoid_cross_entropy_with_logit_op with binary label + """ def setUp(self): self.op_type = "sigmoid_cross_entropy_with_logits" @@ -16,16 +17,16 @@ class TestSigmoidCrossEntropyWithLogitsOp1(OpTest): 'X': logit( np.random.uniform(0, 1, (batch_size, num_classes)) .astype("float32")), - 'Labels': np.random.randint(0, 2, (batch_size, num_classes)) + 'Label': np.random.randint(0, 2, (batch_size, num_classes)) .astype("float32") } # Fw Pass is implemented as elementwise sigmoid followed by # elementwise logistic loss - # Labels * -log(sigmoid(X)) + (1 - labels) * -log(1 - sigmoid(X)) + # Label * -log(sigmoid(X)) + (1 - label) * -log(1 - sigmoid(X)) sigmoid_X = expit(self.inputs['X']) - term1 = self.inputs['Labels'] * np.log(sigmoid_X) - term2 = (1 - self.inputs['Labels']) * np.log(1 - sigmoid_X) + term1 = self.inputs['Label'] * np.log(sigmoid_X) + term2 = (1 - self.inputs['Label']) * np.log(1 - sigmoid_X) self.outputs = {'Out': -term1 - term2} def test_check_output(self): @@ -36,8 +37,8 @@ class TestSigmoidCrossEntropyWithLogitsOp1(OpTest): class TestSigmoidCrossEntropyWithLogitsOp2(OpTest): - '''Test sigmoid_cross_entropy_with_logit_op with probabalistic labels - ''' + """Test sigmoid_cross_entropy_with_logit_op with probabalistic label + """ def setUp(self): self.op_type = "sigmoid_cross_entropy_with_logits" @@ -47,16 +48,16 @@ class TestSigmoidCrossEntropyWithLogitsOp2(OpTest): 'X': logit( np.random.uniform(0, 1, (batch_size, num_classes)) .astype("float32")), - 'Labels': np.random.uniform(0, 1, (batch_size, num_classes)) + 'Label': np.random.uniform(0, 1, (batch_size, num_classes)) .astype("float32") } # Fw Pass is implemented as elementwise sigmoid followed by # elementwise logistic loss - # Labels * -log(sigmoid(X)) + (1 - labels) * -log(1 - sigmoid(X)) + # Label * -log(sigmoid(X)) + (1 - label) * -log(1 - sigmoid(X)) sigmoid_X = expit(self.inputs['X']) - term1 = self.inputs['Labels'] * np.log(sigmoid_X) - term2 = (1 - self.inputs['Labels']) * np.log(1 - sigmoid_X) + term1 = self.inputs['Label'] * np.log(sigmoid_X) + term2 = (1 - self.inputs['Label']) * np.log(1 - sigmoid_X) self.outputs = {'Out': -term1 - term2} def test_check_output(self): @@ -64,3 +65,7 @@ class TestSigmoidCrossEntropyWithLogitsOp2(OpTest): def test_check_grad(self): self.check_grad(['X'], 'Out') + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/v2/fluid/tests/test_softmax_op.py b/python/paddle/v2/fluid/tests/test_softmax_op.py index b41c810d9a6269c934a434b085748a86deccb475..136fc0283afd6acf1de4baae5e681789662295ce 100644 --- a/python/paddle/v2/fluid/tests/test_softmax_op.py +++ b/python/paddle/v2/fluid/tests/test_softmax_op.py @@ -17,14 +17,14 @@ class TestSoftmaxOp(OpTest): 'X': np.random.uniform(0.1, 1, [10, 10]).astype("float32") } self.outputs = { - 'Y': np.apply_along_axis(stable_softmax, 1, self.inputs['X']) + 'Out': np.apply_along_axis(stable_softmax, 1, self.inputs['X']) } def test_check_output(self): self.check_output() def test_check_grad(self): - self.check_grad(['X'], 'Y') + self.check_grad(['X'], 'Out') if __name__ == "__main__": diff --git a/python/paddle/v2/fluid/tests/test_split_and_merge_lod_tensor_op.py b/python/paddle/v2/fluid/tests/test_split_and_merge_lod_tensor_op.py index 3aed83b2ea3418c54f9540279ae6e2e0045421fa..2e4defd55d75c2012f39bea30a6c4de12528e77c 100644 --- a/python/paddle/v2/fluid/tests/test_split_and_merge_lod_tensor_op.py +++ b/python/paddle/v2/fluid/tests/test_split_and_merge_lod_tensor_op.py @@ -2,9 +2,9 @@ import unittest import paddle.v2.fluid.core as core import numpy as np import paddle.v2.fluid.layers as layers -from paddle.v2.fluid.framework import Program +from paddle.v2.fluid.framework import Program, program_guard from paddle.v2.fluid.executor import Executor -from paddle.v2.fluid.backward import append_backward_ops +from paddle.v2.fluid.backward import append_backward class TestCPULoDTensorArrayOps(unittest.TestCase): @@ -75,30 +75,30 @@ class TestCPULoDTensorArrayOps(unittest.TestCase): level=0): place = self.place() program = Program() - x = layers.data(name='x', shape=[1], main_program=program) - x.persistable = True + with program_guard(program): + x = layers.data(name='x', shape=[1]) + x.persistable = True - y = layers.data(name='y', shape=[1], main_program=program) - y.persistable = True + y = layers.data(name='y', shape=[1]) + y.persistable = True - out_true, out_false = layers.split_lod_tensor( - input=x, mask=y, level=level, main_program=program) - out_true.persistable = True - out_false.persistable = True + out_true, out_false = layers.split_lod_tensor( + input=x, mask=y, level=level) + out_true.persistable = True + out_false.persistable = True - out = layers.merge_lod_tensor( - in_true=out_true, - in_false=out_false, - mask=y, - x=x, - level=level, - main_program=program) + out = layers.merge_lod_tensor( + in_true=out_true, in_false=out_false, mask=y, x=x, level=level) - out.persistable = True + out.persistable = True exe = Executor(place) scope = core.Scope() - exe.run(program, feed={'x': tensor, 'y': mask}, scope=scope) + exe.run(program, + feed={'x': tensor, + 'y': mask}, + scope=scope, + return_numpy=False) var_true = scope.find_var(out_true.name).get_tensor() @@ -119,34 +119,21 @@ class TestCPUSplitMergeLoDTensorGrad(unittest.TestCase): def test_grad(self): place = core.CPUPlace() program = Program() + with program_guard(program): + x = layers.data( + name='x', shape=[1], dtype='float32', stop_gradient=False) + y = layers.data( + name='y', shape=[1], dtype='bool', stop_gradient=False) - x = layers.data( - name='x', - shape=[1], - data_type='float32', - main_program=program, - stop_gradient=False) - y = layers.data( - name='y', - shape=[1], - data_type='bool', - main_program=program, - stop_gradient=False) - - level = 0 - - out_true, out_false = layers.split_lod_tensor( - input=x, mask=y, level=level, main_program=program) - out = layers.merge_lod_tensor( - in_true=out_true, - in_false=out_false, - mask=y, - x=x, - level=level, - main_program=program) - mean = layers.mean(x=out, main_program=program) - - append_backward_ops(mean) + level = 0 + + out_true, out_false = layers.split_lod_tensor( + input=x, mask=y, level=level) + out = layers.merge_lod_tensor( + in_true=out_true, in_false=out_false, mask=y, x=x, level=level) + mean = layers.mean(x=out) + + append_backward(mean) tensor = core.LoDTensor() tensor.set(np.arange(10).reshape(10, 1).astype('float32'), place) @@ -169,7 +156,8 @@ class TestCPUSplitMergeLoDTensorGrad(unittest.TestCase): feed={'x': tensor, 'y': mask}, fetch_list=[g_vars], - scope=scope)) + scope=scope, + return_numpy=False)) ] g_out_sum = np.array(g_out).sum() diff --git a/python/paddle/v2/fluid/tests/test_spp_op.py b/python/paddle/v2/fluid/tests/test_spp_op.py new file mode 100644 index 0000000000000000000000000000000000000000..007723f0e35ad194c427401337bc9b13756576de --- /dev/null +++ b/python/paddle/v2/fluid/tests/test_spp_op.py @@ -0,0 +1,68 @@ +import unittest +import numpy as np +from op_test import OpTest +from test_pool2d_op import max_pool2D_forward_naive +from test_pool2d_op import avg_pool2D_forward_naive + + +class TestSppOp(OpTest): + def setUp(self): + self.op_type = "spp" + self.init_test_case() + input = np.random.random(self.shape).astype("float32") + nsize, csize, hsize, wsize = input.shape + out_level_flatten = [] + for i in xrange(self.pyramid_height): + bins = np.power(2, i) + kernel_size = [0, 0] + padding = [0, 0] + kernel_size[0] = np.ceil(hsize / + bins.astype("double")).astype("int32") + padding[0] = ( + (kernel_size[0] * bins - hsize + 1) / 2).astype("int32") + + kernel_size[1] = np.ceil(wsize / + bins.astype("double")).astype("int32") + padding[1] = ( + (kernel_size[1] * bins - wsize + 1) / 2).astype("int32") + out_level = self.pool2D_forward_naive(input, kernel_size, + kernel_size, padding) + out_level_flatten.append( + out_level.reshape(nsize, bins * bins * csize)) + if i == 0: + output = out_level_flatten[i] + else: + output = np.concatenate((output, out_level_flatten[i]), 1) + # output = np.concatenate(out_level_flatten.tolist(), 0); + self.inputs = {'X': input.astype('float32'), } + self.attrs = { + 'pyramid_height': self.pyramid_height, + 'pooling_type': self.pool_type + } + + self.outputs = {'Out': output.astype('float32')} + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + if self.pool_type != "avg": + self.check_grad(['X'], 'Out', max_relative_error=0.05) + + def init_test_case(self): + self.shape = [3, 2, 4, 4] + self.pyramid_height = 3 + self.pool2D_forward_naive = max_pool2D_forward_naive + self.pool_type = "max" + + +class TestCase2(TestSppOp): + def init_test_case(self): + self.shape = [3, 2, 4, 4] + self.pyramid_height = 3 + self.pool2D_forward_naive = avg_pool2D_forward_naive + self.pool_type = "avg" + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/v2/fluid/tests/test_tensor_array.py b/python/paddle/v2/fluid/tests/test_tensor_array.py deleted file mode 100644 index d6929ba16e4dae0c57adcceb4f0e78c094eee55c..0000000000000000000000000000000000000000 --- a/python/paddle/v2/fluid/tests/test_tensor_array.py +++ /dev/null @@ -1,106 +0,0 @@ -import logging -import paddle.v2.fluid.core as core -import unittest -import numpy as np - - -class TestTensorArray(unittest.TestCase): - def setUp(self): - self.ta = core.TensorArray() - - self.batch_size = 10 - self.dim = 2 - - # create a LoDTensor - self.scope = core.Scope() - var = self.scope.var("test_tensor") - self.place = core.CPUPlace() - tensor = var.get_tensor() - tensor.set_dims([self.batch_size, self.dim]) - tensor.alloc_float(self.place) - tensor_array = np.array(tensor) - tensor_array[0, 0] = 0 - tensor_array[1, 0] = 1 - tensor_array[2, 0] = 2 - tensor_array[3, 0] = 3 - tensor_array[4, 0] = 4 - tensor_array[5, 0] = 5 - tensor_array[6, 0] = 6 - tensor_array[7, 0] = 7 - tensor_array[8, 0] = 8 - tensor_array[9, 0] = 9 - - lod_py = [[0, 2, 5, 10]] - lod_tensor = core.LoDTensor(lod_py) - lod_tensor.set(tensor_array, self.place) - - self.py_seq_meta = [[5, 10, 2], [2, 5, 1], [0, 2, 0]] - - self.tensor = lod_tensor - - def test_unstack(self): - self.ta.unstack(self.tensor) - self.assertEqual(self.tensor.get_dims()[0], self.ta.size()) - - def test_read(self): - self.ta.unstack(self.tensor) - for i in range(self.batch_size): - tensor = self.ta.read(i) - - def test_write(self): - self.ta.unstack(self.tensor) - - # create a tensor with shape of [1, self.dim] - var = self.scope.var("hell") - tensor = var.get_tensor() - tensor.set_dims([1, self.dim]) - tensor.alloc_float(self.place) - tensor_array = np.array(tensor) - for i in range(self.dim): - tensor_array[0, i] = i - tensor.set(tensor_array, self.place) - - self.ta.write(2, tensor) - - ta_tensor = self.ta.read(2) - ta_tensor_array = np.array(ta_tensor) - self.assertEqual(ta_tensor.get_dims(), [1, self.dim]) - self.assertTrue((tensor_array == ta_tensor_array).all()) - - def test_write_shared(self): - self.ta.unstack(self.tensor) - - # create a tensor with shape of [1, self.dim] - var = self.scope.var("hell") - tensor = var.get_tensor() - tensor.set_dims([1, self.dim]) - tensor.alloc_float(self.place) - tensor_array = np.array(tensor) - for i in range(self.dim): - tensor_array[0, i] = i - tensor.set(tensor_array, self.place) - - self.ta.write_shared(2, tensor) - - ta_tensor = self.ta.read(2) - ta_tensor_array = np.array(ta_tensor) - self.assertEqual(ta_tensor.get_dims(), [1, self.dim]) - self.assertTrue((tensor_array == ta_tensor_array).all()) - - def test_unpack(self): - meta = self.ta.unpack(self.tensor, 0, True) - self.assertEqual(self.ta.size(), 5) - self.assertEqual(meta, self.py_seq_meta) - - def test_pack(self): - meta = self.ta.unpack(self.tensor, 0, True) - print "meta", meta - tensor = self.ta.pack(0, meta, self.tensor.lod()) - print np.array(self.tensor) - print np.array(tensor) - self.assertTrue((np.array(self.tensor) == np.array(tensor)).all()) - self.assertTrue(tensor.lod(), self.tensor.lod()) - - -if __name__ == '__main__': - unittest.main() diff --git a/python/paddle/v2/fluid/tests/test_uniform_random_op.py b/python/paddle/v2/fluid/tests/test_uniform_random_op.py index f736dfb2e85552b321403c961da517f3b3efb100..dbe4d6bcd069d2088b3cc1b4efd575d14afd4198 100644 --- a/python/paddle/v2/fluid/tests/test_uniform_random_op.py +++ b/python/paddle/v2/fluid/tests/test_uniform_random_op.py @@ -1,32 +1,49 @@ import unittest +import numpy + from paddle.v2.fluid.op import Operator import paddle.v2.fluid.core as core -import numpy +import paddle.v2.fluid as fluid class TestUniformRandomOp(unittest.TestCase): - def test_uniform_random_cpu(self): + def setUp(self): + self.op_type = "uniform_random" + self.inputs = {} + self.attrs = { + "shape": [1000, 784], + "min": -5.0, + "max": 10.0, + "seed": 10 + } + self.outputs = ["Out"] + + def test_cpu(self): self.uniform_random_test(place=core.CPUPlace()) - def test_uniform_random_gpu(self): + def test_gpu(self): if core.is_compile_gpu(): - self.uniform_random_test(place=core.GPUPlace(0)) + self.uniform_random_test(place=core.CUDAPlace(0)) def uniform_random_test(self, place): - scope = core.Scope() - scope.var('X').get_tensor() - - op = Operator( - "uniform_random", - Out='X', - shape=[1000, 784], - min=-5.0, - max=10.0, - seed=10) - - ctx = core.DeviceContext.create(place) - op.run(scope, ctx) - tensor = numpy.array(scope.find_var('X').get_tensor()) + program = fluid.Program() + block = program.global_block() + vout = block.create_var(name="Out") + op = block.append_op( + type=self.op_type, outputs={"Out": vout}, attrs=self.attrs) + + op.desc.infer_var_type(block.desc) + op.desc.infer_shape(block.desc) + + fetch_list = [] + for var_name in self.outputs: + fetch_list.append(block.var(var_name)) + + exe = fluid.Executor(place) + outs = exe.run(program, fetch_list=fetch_list) + + tensor = outs[0] + self.assertAlmostEqual(tensor.mean(), 2.5, delta=0.1) diff --git a/python/paddle/v2/fluid/tests/test_unpool_op.py b/python/paddle/v2/fluid/tests/test_unpool_op.py new file mode 100644 index 0000000000000000000000000000000000000000..e87f283042c081ed9f232d140ff8c303cd3d1858 --- /dev/null +++ b/python/paddle/v2/fluid/tests/test_unpool_op.py @@ -0,0 +1,83 @@ +import unittest +import numpy as np +from op_test import OpTest + + +def unpool2dmax_forward_naive(input, indices, ksize, strides, paddings): + s0, s1, s2, s3 = input.shape + out_hsize = (s2 - 1) * strides[0] - 2 * paddings[0] + ksize[0] + out_wsize = (s2 - 1) * strides[1] - 2 * paddings[1] + ksize[1] + out = np.zeros((s0, s1, out_hsize, out_wsize)) + for nidx in xrange(s0): + for cidx in xrange(s1): + for h in xrange(s2): + for w in xrange(s3): + index = indices[nidx, cidx, h, w] + hidx = (index - index % out_wsize) / out_wsize + widx = index % out_wsize + out[nidx, cidx, int(hidx), int(widx)] = \ + input[nidx, cidx, h, w] + + return out + + +class TestUnpoolOp(OpTest): + def setUp(self): + self.op_type = "unpool" + self.init_test_case() + pre_input = np.random.random(self.shape).astype("float32") + nsize, csize, hsize, wsize = pre_input.shape + hsize_out = (hsize - self.ksize[0] + 2 * self.paddings[0]) / \ + self.strides[0] + 1 + wsize_out = (wsize - self.ksize[1] + 2 * self.paddings[1]) / \ + self.strides[1] + 1 + input = np.zeros((nsize, csize, hsize_out, wsize_out)) + indices = np.zeros((nsize, csize, hsize_out, wsize_out)) + for i in xrange(hsize_out): + for j in xrange(wsize_out): + r_start = np.max((i * self.strides[0] - self.paddings[0], 0)) + r_end = np.min((i * self.strides[0] + self.ksize[0] - \ + self.paddings[0], hsize)) + c_start = np.max((j * self.strides[1] - self.paddings[1], 0)) + c_end = np.min((j * self.strides[1] + self.ksize[1] - \ + self.paddings[1], wsize)) + for nidx in xrange(nsize): + for cidx in xrange(csize): + x_masked = pre_input[nidx, cidx, r_start:r_end, \ + c_start:c_end] + input[nidx, cidx, i, j] = x_masked.max() + arg = x_masked.argmax() + indices[nidx, cidx, i, j] = \ + (r_start + arg / self.ksize[1]) * wsize + \ + c_start + arg % self.ksize[1] + output = self.unpool2d_forward_naive(input, indices, self.ksize, \ + self.strides, self.paddings).astype("float32") + self.inputs = { + 'X': input.astype('float32'), + 'Indices': indices.astype('int32') + } + self.attrs = { + 'strides': self.strides, + 'paddings': self.paddings, + 'ksize': self.ksize, + 'unpooling_type': self.unpooling_type, + } + self.outputs = {'Out': output.astype('float32')} + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad(['X'], 'Out') + + def init_test_case(self): + self.unpool2d_forward_naive = unpool2dmax_forward_naive + self.unpooling_type = "max" + self.shape = [6, 4, 5, 5] + self.ksize = [3, 3] + self.strides = [2, 2] + self.paddings = [0, 0] + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/v2/fluid/tests/test_variable.py b/python/paddle/v2/fluid/tests/test_variable.py index a3e60a751719666bdca56a3096b688125d09f4b2..f1e4c0ba21d5c4f10d2b5011bdb5abaebaec5431 100644 --- a/python/paddle/v2/fluid/tests/test_variable.py +++ b/python/paddle/v2/fluid/tests/test_variable.py @@ -1,5 +1,5 @@ import unittest -from paddle.v2.fluid.framework import Variable, g_main_program, Program +from paddle.v2.fluid.framework import default_main_program, Program, convert_np_dtype_to_dtype_ import paddle.v2.fluid.core as core import numpy as np @@ -7,7 +7,7 @@ import numpy as np class TestVariable(unittest.TestCase): def test_np_dtype_convert(self): DT = core.DataType - convert = Variable._convert_np_dtype_to_dtype_ + convert = convert_np_dtype_to_dtype_ self.assertEqual(DT.FP32, convert(np.float32)) self.assertEqual(DT.FP16, convert("float16")) self.assertEqual(DT.FP64, convert("float64")) @@ -18,17 +18,17 @@ class TestVariable(unittest.TestCase): self.assertRaises(ValueError, lambda: convert("int8")) def test_var(self): - b = g_main_program.current_block() + b = default_main_program().current_block() w = b.create_var( dtype="float64", shape=[784, 100], lod_level=0, name="fc.w") self.assertNotEqual(str(w), "") - self.assertEqual(core.DataType.FP64, w.data_type) + self.assertEqual(core.DataType.FP64, w.dtype) self.assertEqual((784, 100), w.shape) self.assertEqual("fc.w", w.name) self.assertEqual(0, w.lod_level) w = b.create_var(name='fc.w') - self.assertEqual(core.DataType.FP64, w.data_type) + self.assertEqual(core.DataType.FP64, w.dtype) self.assertEqual((784, 100), w.shape) self.assertEqual("fc.w", w.name) self.assertEqual(0, w.lod_level) diff --git a/python/paddle/v2/fluid/tests/test_while_op.py b/python/paddle/v2/fluid/tests/test_while_op.py index 84b432333f950f754a97bc1a051b59c16fb22aed..7c5593cc5e5a66d4ccb237e3706ff3e544adf033 100644 --- a/python/paddle/v2/fluid/tests/test_while_op.py +++ b/python/paddle/v2/fluid/tests/test_while_op.py @@ -2,18 +2,18 @@ import unittest import paddle.v2.fluid.layers as layers from paddle.v2.fluid.executor import Executor import paddle.v2.fluid.core as core -from paddle.v2.fluid.backward import append_backward_ops +from paddle.v2.fluid.backward import append_backward import numpy class TestWhileOp(unittest.TestCase): def test_simple_forward(self): d0 = layers.data( - "d0", shape=[10], append_batch_size=False, data_type='float32') + "d0", shape=[10], append_batch_size=False, dtype='float32') d1 = layers.data( - "d1", shape=[10], append_batch_size=False, data_type='float32') + "d1", shape=[10], append_batch_size=False, dtype='float32') d2 = layers.data( - "d2", shape=[10], append_batch_size=False, data_type='float32') + "d2", shape=[10], append_batch_size=False, dtype='float32') i = layers.zeros(shape=[1], dtype='int64') i.stop_gradient = True init = layers.zeros(shape=[10], dtype='float32') @@ -46,7 +46,7 @@ class TestWhileOp(unittest.TestCase): sum_result = layers.array_read(array=mem_array, i=i) loss = layers.mean(x=sum_result) - append_backward_ops(loss) + append_backward(loss) cpu = core.CPUPlace() exe = Executor(cpu) @@ -55,19 +55,10 @@ class TestWhileOp(unittest.TestCase): for i in xrange(3): d.append(numpy.random.random(size=[10]).astype('float32')) - d_tensor = [] - for item in d: - t = core.LoDTensor() - t.set(item, cpu) - d_tensor.append(t) - - outs = map(numpy.array, - exe.run(feed={ - 'd0': d_tensor[0], - 'd1': d_tensor[1], - 'd2': d_tensor[2] - }, - fetch_list=[sum_result])) + outs = exe.run(feed={'d0': d[0], + 'd1': d[1], + 'd2': d[2]}, + fetch_list=[sum_result]) self.assertAlmostEqual(numpy.sum(d), numpy.sum(outs[0]), delta=0.01) diff --git a/python/paddle/v2/parameters.py b/python/paddle/v2/parameters.py index bd97dc1199fedc8ac91c1c6086957e8cce88bdc4..7b7d1a1d1672802e0e91a857100604758683224e 100644 --- a/python/paddle/v2/parameters.py +++ b/python/paddle/v2/parameters.py @@ -383,19 +383,22 @@ class Parameters(object): params.deserialize(param_name, f) return params - def init_from_tar(self, f): + def init_from_tar(self, f, exclude_params=[]): """ Different from `from_tar`, this interface can be used to init partial network parameters from another saved model. :param f: the initialized model file. :type f: tar file + :param exclude_params: the names of parameters that should + not be initialized from the model file. + :type exclude_params: list of strings :return: Nothing. """ tar_param = Parameters.from_tar(f) for pname in tar_param.names(): - if pname in self.names(): + if pname in self.names() and pname not in exclude_params: self.set(pname, tar_param.get(pname)) diff --git a/python/paddle/v2/reader/decorator.py b/python/paddle/v2/reader/decorator.py index 45a4288751e37b99dd1005ec78f30a98044926ff..44a6e344630bb35d28ee29078bf8727053a24bef 100644 --- a/python/paddle/v2/reader/decorator.py +++ b/python/paddle/v2/reader/decorator.py @@ -14,13 +14,16 @@ __all__ = [ 'map_readers', 'buffered', 'compose', 'chain', 'shuffle', - 'ComposeNotAligned', 'firstn', 'xmap_readers' + 'ComposeNotAligned', 'firstn', 'xmap_readers', 'PipeReader' ] +from threading import Thread +import subprocess + +from Queue import Queue import itertools import random -from Queue import Queue -from threading import Thread +import zlib def map_readers(func, *readers): @@ -323,3 +326,80 @@ def xmap_readers(mapper, reader, process_num, buffer_size, order=False): yield sample return xreader + + +def _buf2lines(buf, line_break="\n"): + # FIXME: line_break should be automatically configured. + lines = buf.split(line_break) + return lines[:-1], lines[-1] + + +class PipeReader: + """ + PipeReader read data by stream from a command, take it's + stdout into a pipe buffer and redirect it to the parser to + parse, then yield data as your desired format. + + You can using standard linux command or call another program + to read data, from HDFS, Ceph, URL, AWS S3 etc: + + .. code-block:: python + cmd = "hadoop fs -cat /path/to/some/file" + cmd = "cat sample_file.tar.gz" + cmd = "curl http://someurl" + cmd = "python print_s3_bucket.py" + + An example: + + .. code-block:: python + + def example_reader(): + for f in myfiles: + pr = PipeReader("cat %s"%f) + for l in pr.get_line(): + sample = l.split(" ") + yield sample + """ + + def __init__(self, command, bufsize=8192, file_type="plain"): + if not isinstance(command, str): + raise TypeError("left_cmd must be a string") + if file_type == "gzip": + self.dec = zlib.decompressobj( + 32 + zlib.MAX_WBITS) # offset 32 to skip the header + self.file_type = file_type + self.bufsize = bufsize + self.process = subprocess.Popen( + command.split(" "), bufsize=bufsize, stdout=subprocess.PIPE) + + def get_line(self, cut_lines=True, line_break="\n"): + """ + :param cut_lines: cut buffer to lines + :type cut_lines: bool + :param line_break: line break of the file, like \n or \r + :type line_break: string + + :return: one line or a buffer of bytes + :rtype: string + """ + remained = "" + while True: + buff = self.process.stdout.read(self.bufsize) + if buff: + if self.file_type == "gzip": + decomp_buff = self.dec.decompress(buff) + elif self.file_type == "plain": + decomp_buff = buff + else: + raise TypeError("file_type %s is not allowed" % + self.file_type) + + if cut_lines: + lines, remained = _buf2lines(''.join( + [remained, decomp_buff]), line_break) + for line in lines: + yield line + else: + yield decomp_buff + else: + break diff --git a/python/paddle/v2/reader/tests/decorator_test.py b/python/paddle/v2/reader/tests/decorator_test.py index 5a92951b100fa51ab6df7039d9c6b54d1f9d963e..4ba71969dffe7447b6c5b70aeb752a4e5469fb36 100644 --- a/python/paddle/v2/reader/tests/decorator_test.py +++ b/python/paddle/v2/reader/tests/decorator_test.py @@ -145,5 +145,33 @@ class TestXmap(unittest.TestCase): self.assertEqual(e, mapper(idx)) +class TestPipeReader(unittest.TestCase): + def test_pipe_reader(self): + def example_reader(myfiles): + for f in myfiles: + pr = paddle.v2.reader.PipeReader("cat %s" % f, bufsize=128) + for l in pr.get_line(): + yield l + + import tempfile + + records = [str(i) for i in xrange(5)] + temp = tempfile.NamedTemporaryFile() + try: + with open(temp.name, 'w') as f: + for r in records: + f.write('%s\n' % r) + + result = [] + for r in example_reader([temp.name]): + result.append(r) + + for idx, e in enumerate(records): + self.assertEqual(e, result[idx]) + finally: + # delete the temporary file + temp.close() + + if __name__ == '__main__': unittest.main() diff --git a/python/setup.py.in b/python/setup.py.in index fe91df10daf303bb14d1e5f28817984d261e0880..66ccfe808763d0e157f866ce08868e3fdebdea79 100644 --- a/python/setup.py.in +++ b/python/setup.py.in @@ -1,8 +1,61 @@ from setuptools import setup, Distribution, Extension +import subprocess class BinaryDistribution(Distribution): def has_ext_modules(foo): return True +MAJOR = 0 +MINOR = 11 +PATCH = 0 +RC = 0 +ISTAGED = False + + + +def git_commit(): + try: + cmd = ['git', 'rev-parse', 'HEAD'] + git_commit = subprocess.Popen(cmd, stdout = subprocess.PIPE).communicate()[0].strip() + except: + git_commit = 'Unknown' + return git_commit + +def write_version_py(filename='paddle/version.py'): + cnt = ''' +# THIS FILE IS GENERATED FROM PADDLEPADDLE SETUP.PY +# +full_version = '%(major)d.%(minor)d.%(patch)d' +major = '%(major)d' +minor = '%(minor)d' +patch = '%(patch)d' +rc = '%(rc)d' +istaged = %(istaged)s +commit = '%(commit)s' + +def show(): + if istaged: + print 'full_version:', full_version + print 'major:', major + print 'minor:', minor + print 'patch:', patch + print 'rc:', rc + else: + print 'commit:', commit +''' + commit = git_commit() + with open(filename, 'w') as f: + f.write(cnt % { + 'major': MAJOR, + 'minor': MINOR, + 'patch': PATCH, + 'rc': RC, + 'version': '${PADDLE_VERSION}', + 'commit': commit, + 'istaged': ISTAGED}) + +write_version_py(filename='@PADDLE_SOURCE_DIR@/python/paddle/version.py') + + packages=['paddle', 'paddle.proto', 'paddle.trainer', @@ -15,18 +68,18 @@ packages=['paddle', 'paddle.v2.plot', 'paddle.v2.fluid', 'paddle.v2.fluid.proto', + 'paddle.v2.fluid.layers', 'py_paddle'] with open('@PADDLE_SOURCE_DIR@/python/requirements.txt') as f: setup_requires = f.read().splitlines() if '${CMAKE_SYSTEM_PROCESSOR}' not in ['arm', 'armv7-a', 'aarch64']: - setup_requires+=["opencv-python"] + setup_requires+=['opencv-python'] # the prefix is sys.prefix which should always be usr paddle_bin_dir = 'opt/paddle/bin' -paddle_bins = ['${PADDLE_BINARY_DIR}/paddle/scripts/paddle_usage', - '${PADDLE_BINARY_DIR}/paddle/trainer/paddle_trainer', +paddle_bins = ['${PADDLE_BINARY_DIR}/paddle/trainer/paddle_trainer', '${PADDLE_BINARY_DIR}/paddle/trainer/paddle_merge_model', '${PADDLE_BINARY_DIR}/paddle/pserver/paddle_pserver_main', '${PADDLE_BINARY_DIR}/paddle/scripts/paddle'] @@ -36,7 +89,7 @@ paddle_rt_libs = ['${WARPCTC_LIBRARIES}'] if '${MKL_SHARED_LIBS}'!= '': paddle_rt_libs += '${MKL_SHARED_LIBS}'.split(';') -setup(name='paddlepaddle', +setup(name='${PACKAGE_NAME}', version='${PADDLE_VERSION}', description='Parallel Distributed Deep Learning', install_requires=setup_requires,