diff --git a/AUTHORS.md b/AUTHORS.md index deafa641203ed9d9bd794fe92e4a91e3aaa03f63..da91933f4697f3b7c08feab20d703e22397b3757 100644 --- a/AUTHORS.md +++ b/AUTHORS.md @@ -44,6 +44,7 @@ | qingqing01 | Qing-Qing Dang | | reyoung | Yang Yu | | Sand3r- | Michal Gallus | +| sfraczek | Sylwester Fraczek | | Superjom | Chun-Wei Yan | | tensor-tang | Jian Tang | | tianbingsz | Tian-Bing Xu | @@ -54,6 +55,7 @@ | wangyang59 | Yang Wang | | wangzhen-nlp | Zhen Wang | | wen-bo-yang | Wen-Bo Yang | +| wojtuss | Wojciech Uss | | wwhu | Wei-Wei Hu | | xinghai-sun | Xing-Hai Sun | | Xreki | Yi-Qun Liu | diff --git a/CMakeLists.txt b/CMakeLists.txt index 9ec632e20690eafdc558e24f160270a89b29ee41..8e7ffe72b5fb846fb55ab8dc4809d87a40cfe06c 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -25,12 +25,18 @@ message(STATUS "CXX compiler: ${CMAKE_CXX_COMPILER}, version: " message(STATUS "C compiler: ${CMAKE_C_COMPILER}, version: " "${CMAKE_C_COMPILER_ID} ${CMAKE_C_COMPILER_VERSION}") if(WIN32) + set(CMAKE_SUPPRESS_REGENERATION ON) set(CMAKE_STATIC_LIBRARY_PREFIX lib) add_definitions("/DGOOGLE_GLOG_DLL_DECL=") set(CMAKE_C_FLAGS_DEBUG "${CMAKE_C_FLAGS_DEBUG} /bigobj /MTd") set(CMAKE_C_FLAGS_RELEASE "${CMAKE_C_FLAGS_RELEASE} /bigobj /MT") set(CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} /bigobj /MTd") set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /bigobj /MT") + add_compile_options(/wd4068 /wd4129 /wd4244 /wd4267 /wd4297 /wd4530 /wd4577 /wd4819 /wd4838) + set(PADDLE_LINK_FLAGS "/IGNORE:4006 /IGNORE:4098 /IGNORE:4217 /IGNORE:4221") + set(CMAKE_STATIC_LINKER_FLAGS "${CMAKE_STATIC_LINKER_FLAGS} ${PADDLE_LINK_FLAGS}") + set(CMAKE_SHARED_LINKER_FLAGS "${CMAKE_SHARED_LINKER_FLAGS} ${PADDLE_LINK_FLAGS}") + set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} ${PADDLE_LINK_FLAGS}") endif(WIN32) find_package(CUDA QUIET) @@ -48,23 +54,12 @@ option(WITH_NGRAPH "Compile PaddlePaddle with nGraph support." OFF) option(WITH_DSO "Compile PaddlePaddle with dynamic linked CUDA" ON) option(WITH_TESTING "Compile PaddlePaddle with unit testing" OFF) option(WITH_PYTHON "Compile PaddlePaddle with python interpreter" ON) -option(WITH_DOUBLE "Compile PaddlePaddle with double precision" OFF) -option(WITH_RDMA "Compile PaddlePaddle with RDMA support" OFF) -option(WITH_TIMER "Compile PaddlePaddle with stats timer" OFF) option(WITH_PROFILER "Compile PaddlePaddle with GPU profiler and gperftools" OFF) option(WITH_JEMALLOC "Compile PaddlePaddle with jemalloc" OFF) -option(WITH_DOC "Compile PaddlePaddle with documentation" OFF) option(WITH_COVERAGE "Compile PaddlePaddle with code coverage" OFF) option(COVERALLS_UPLOAD "Package code coverage data to coveralls" OFF) -option(ON_TRAVIS "Exclude special unit test on Travis CI" OFF) -option(WITH_FLUID_ONLY "Compile PaddlePaddle fluid only" OFF) -option(WITH_GOLANG "Compile PaddlePaddle with GOLANG" OFF) -option(GLIDE_INSTALL "Download and install go dependencies " ON) option(WITH_DISTRIBUTE "Compile with distributed support" OFF) option(WITH_PSLIB "Compile with pslib support" OFF) -option(USE_EIGEN_FOR_BLAS "Use matrix multiplication in Eigen" OFF) -option(EIGEN_USE_THREADS "Compile with multi-threaded Eigen" OFF) -option(WITH_ARM_FP16 "Use half precision support on armv8.2-a cpu" OFF) option(WITH_CONTRIB "Compile the third-party contributation" OFF) option(REPLACE_ENFORCE_GLOG "Replace PADDLE_ENFORCE with glog/CHECK for better debug." OFF) option(WITH_ANAKIN "Compile with Anakin library" OFF) @@ -99,8 +94,6 @@ endif() if (WIN32) set(WITH_DISTRIBUTE OFF CACHE STRING "Disable DISTRIBUTE when compiling for Windows" FORCE) - set(WITH_FLUID_ONLY ON CACHE STRING - "Enable FLUID_ONLY when compiling for Windows" FORCE) endif() set(THIRD_PARTY_PATH "${CMAKE_BINARY_DIR}/third_party" CACHE STRING @@ -142,7 +135,6 @@ include(external/openblas) # download, build, install openblas include(external/mkldnn) # download, build, install mkldnn include(external/ngraph) # download, build, install nGraph include(external/boost) # download boost -include(external/any) # download libn::any include(external/eigen) # download eigen3 include(external/pybind11) # download pybind11 include(external/cares) @@ -212,14 +204,13 @@ endif() if (WITH_JEMALLOC) find_package(JeMalloc REQUIRED) include_directories(${JEMALLOC_INCLUDE_DIR}) - add_definitions(-DWITH_JEMALLOC) + add_definitions(-DPADDLE_WITH_JEMALLOC) endif() include(generic) # simplify cmake module include(package) # set paddle packages include(ccache) # set ccache for compilation include(util) # set unittest and link libs -include(rdma) # set rdma libraries include(version) # set PADDLE_VERSION include(coveralls) # set code coverage include(inference_lib) # add paddle fluid inference libraries @@ -227,38 +218,11 @@ include(inference_lib) # add paddle fluid inference libraries include_directories("${PADDLE_SOURCE_DIR}") -set(EXTERNAL_LIBS - gflags - glog - ${CBLAS_LIBRARIES} - protobuf - zlib - ${PYTHON_LIBRARIES} -) - -if(WITH_PSLIB) - list(APPEND EXTERNAL_LIBS pslib) - list(APPEND EXTERNAL_LIBS pslib_brpc) - list(APPEND EXTERNAL_LIBS libmct) -endif(WITH_PSLIB) - if(WITH_AMD_GPU) find_package(HIP) include(hip) endif(WITH_AMD_GPU) -if(WITH_MKLML) - list(APPEND EXTERNAL_LIBS ${MKLML_IOMP_LIB}) -endif() - -if(WITH_LIBXSMM) - list(APPEND EXTERNAL_LIBS ${LIBXSMM_LIBS}) -endif() - -if(WITH_MKLDNN) - list(APPEND EXTERNAL_LIBS ${MKLDNN_LIB}) -endif() - set(PADDLE_PYTHON_BUILD_DIR "${CMAKE_CURRENT_BINARY_DIR}/python/build") set(CMAKE_CXX_FLAGS_RELWITHDEBINFO "-O3 -g -DNDEBUG") @@ -276,9 +240,3 @@ add_subdirectory(paddle) if(WITH_PYTHON) add_subdirectory(python) endif() - -if(WITH_DOC) - find_package(Sphinx REQUIRED) - find_python_module(recommonmark REQUIRED) - add_subdirectory(doc) -endif() diff --git a/Dockerfile b/Dockerfile index acfd091265e26d6c29c561d166fed2504c0cff1c..fe0721e9b99b5e028df2f6228ff04cb56a567a3f 100644 --- a/Dockerfile +++ b/Dockerfile @@ -11,12 +11,10 @@ RUN /bin/bash -c 'if [[ -n ${UBUNTU_MIRROR} ]]; then sed -i 's#http://archive.ub # ENV variables ARG WITH_GPU ARG WITH_AVX -ARG WITH_DOC ENV WOBOQ OFF ENV WITH_GPU=${WITH_GPU:-ON} ENV WITH_AVX=${WITH_AVX:-ON} -ENV WITH_DOC=${WITH_DOC:-OFF} ENV HOME /root # Add bash enhancements diff --git a/README.md b/README.md index 32a302cc5431a62b310d4812b545bd929f090e0a..5c428e9900762a208eebbfd053ce98663f803345 100644 --- a/README.md +++ b/README.md @@ -1,13 +1,13 @@ # PaddlePaddle +English | [简体中文](./README_cn.md) [![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://paddlepaddle.org/documentation/docs/en/1.2/getstarted/index_en.html) -[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/index.html) +[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://paddlepaddle.org/documentation/docs/en/1.3/beginners_guide/index_en.html) +[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://paddlepaddle.org/documentation/docs/zh/1.3/beginners_guide/index.html) [![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) - Welcome to the PaddlePaddle GitHub. PaddlePaddle (PArallel Distributed Deep LEarning) is an easy-to-use, @@ -18,17 +18,7 @@ learning to many products at Baidu. Our vision is to enable deep learning for everyone via PaddlePaddle. Please refer to our [release announcement](https://github.com/PaddlePaddle/Paddle/releases) to track the latest feature of PaddlePaddle. - -欢迎来到 PaddlePaddle GitHub - -PaddlePaddle (PArallel Distributed Deep LEarning) 是一个简单易用、高效灵活、可扩展的深度学习平台，最初由百度科学家和工程师共同开发，目的是将深度学习技术应用到百度的众多产品中。 - -我们的愿景是让每个人都能通过PaddlePaddle接触深度学习 - -跟进PaddlePaddle最新特性请参考我们的[版本说明](https://github.com/PaddlePaddle/Paddle/releases) - - -### Latest PaddlePaddle Release: [Fluid 1.2.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.2) +### Latest PaddlePaddle Release: [Fluid 1.3.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.3) ### Install Latest Stable Release: ``` # Linux CPU @@ -36,30 +26,13 @@ pip install paddlepaddle # Linux GPU cuda9cudnn7 pip install paddlepaddle-gpu # Linux GPU cuda8cudnn7 -pip install paddlepaddle-gpu==1.2.0.post87 +pip install paddlepaddle-gpu==1.3.0.post87 # Linux GPU cuda8cudnn5 -pip install paddlepaddle-gpu==1.2.0.post85 +pip install paddlepaddle-gpu==1.3.0.post85 # For installation on other platform, refer to http://paddlepaddle.org/ ``` - -### PaddlePaddle最新版本: [Fluid 1.2.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.2) -### 安装最新稳定版本: -``` -# Linux CPU -pip install paddlepaddle -# Linux GPU cuda9cudnn7 -pip install paddlepaddle-gpu -# Linux GPU cuda8cudnn7 -pip install paddlepaddle-gpu==1.2.0.post87 -# Linux GPU cuda8cudnn5 -pip install paddlepaddle-gpu==1.2.0.post85 - -# 其他平台上的安装指引请参考 http://paddlepaddle.org/ -``` - - ## Features - **Flexibility** @@ -100,90 +73,34 @@ pip install paddlepaddle-gpu==1.2.0.post85 Baidu and it has achieved a significant impact. We hope you can also explore the capability of PaddlePaddle to make an impact on your product. -## 特点 - -- **灵活性** - - PaddlePaddle支持丰富的神经网络架构和优化算法。易于配置复杂模型，例如带有注意力机制或复杂记忆连接的神经网络机器翻译模型。 - -- **高效性** - - 为了高效使用异步计算资源，PaddlePaddle对框架的不同层进行优化，包括计算、存储、架构和通信。下面是一些样例： - - - 通过SSE/AVX 内置函数、BLAS库(例如MKL、OpenBLAS、cuBLAS)或定制的CPU/GPU内核优化数学操作。 - - 通过MKL-DNN库优化CNN网络 - - 高度优化循环网络，无需执行 `padding` 操作即可处理 **变长** 序列 - - 针对高维稀疏数据模型，优化了局部和分布式训练。 - - -- **稳定性** - - 有了 PaddlePaddle，使得利用各种CPU/GPU和机器来加速训练变得简单。PaddlePaddle 通过优化通信可以实现巨大吞吐量和快速执行。 - -- **连接产品** - - 另外，PaddlePaddle 的设计也易于部署。在百度，PaddlePaddle 已经部署到含有巨大用户量的产品和服务上，包括广告点击率（CTR）预测、大规模图像分类、光学字符识别（OCR）、搜索排序，计算机病毒检测、推荐系统等等。PaddlePaddle广泛应用于百度产品中，产生了非常重要的影响。我们希望您也能探索 PaddlePaddle 的能力，为您的产品创造新的影响力和效果。 - ## Installation -It is recommended to read [this doc](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/install/index_cn.html) on our website. - -## 安装 - -推荐阅读官网上的[安装说明](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/install/index_cn.html) +It is recommended to read [this doc](http://paddlepaddle.org/documentation/docs/en/1.3/beginners_guide/index_en.html) on our website. ## Documentation -We provide [English](http://paddlepaddle.org/documentation/docs/en/1.2/getstarted/index_en.html) and -[Chinese](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/index.html) documentation. +We provide [English](http://paddlepaddle.org/documentation/docs/en/1.3/beginners_guide/index_en.html) and +[Chinese](http://paddlepaddle.org/documentation/docs/zh/1.3/beginners_guide/index.html) documentation. - [Deep Learning 101](https://github.com/PaddlePaddle/book) You might want to start from this online interactive book that can run in a Jupyter Notebook. -- [Distributed Training](http://paddlepaddle.org/documentation/docs/zh/1.2/user_guides/howto/training/cluster_howto.html) +- [Distributed Training](http://paddlepaddle.org/documentation/docs/en/1.3/user_guides/howto/training/multi_node_en.html) You can run distributed training jobs on MPI clusters. -- [Python API](http://paddlepaddle.org/documentation/docs/zh/1.2/api_cn/index_cn.html) +- [Python API](http://paddlepaddle.org/documentation/docs/en/1.3/api/index_en.html) Our new API enables much shorter programs. -- [How to Contribute](http://paddlepaddle.org/documentation/docs/zh/1.2/advanced_usage/development/contribute_to_paddle/index_cn.html) +- [How to Contribute](http://paddlepaddle.org/documentation/docs/en/1.3/advanced_usage/development/contribute_to_paddle/index_en.html) We appreciate your contributions! -## 文档 - -我们提供[英文](http://paddlepaddle.org/documentation/docs/en/1.2/getstarted/index_en.html)和 -[中文](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/index.html) 文档 - -- [深度学习101](https://github.com/PaddlePaddle/book) - - 或许您想从这个在线交互式书籍开始，可以在Jupyter Notebook中运行 - -- [分布式训练](http://paddlepaddle.org/documentation/docs/zh/1.2/user_guides/howto/training/cluster_howto.html) - - 可以在MPI集群上运行分布式训练任务 - -- [Python API](http://paddlepaddle.org/documentation/docs/zh/1.2/api_cn/index_cn.html) - - 新的API支持代码更少更简洁的程序 - -- [贡献方式](http://paddlepaddle.org/documentation/docs/zh/1.2/advanced_usage/development/contribute_to_paddle/index_cn.html) - - 欢迎您的贡献! - ## Ask Questions You are welcome to submit questions and bug reports as [Github Issues](https://github.com/PaddlePaddle/Paddle/issues). -## 答疑 - -欢迎您将问题和bug报告以[Github Issues](https://github.com/PaddlePaddle/Paddle/issues)的形式提交 - ## Copyright and License PaddlePaddle is provided under the [Apache-2.0 license](LICENSE). - -## 版权和许可证 -PaddlePaddle由[Apache-2.0 license](LICENSE)提供 diff --git a/README_cn.md b/README_cn.md new file mode 100644 index 0000000000000000000000000000000000000000..b7b0e75e5524cc483a8c203a382e7f339f91694f --- /dev/null +++ b/README_cn.md @@ -0,0 +1,88 @@ +# PaddlePaddle + +[English](./README.md) | 简体中文 + +[![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://paddlepaddle.org/documentation/docs/en/1.3/beginners_guide/index_en.html) +[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://paddlepaddle.org/documentation/docs/zh/1.3/beginners_guide/index.html) +[![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) + +欢迎来到 PaddlePaddle GitHub + +PaddlePaddle (PArallel Distributed Deep LEarning) 是一个简单易用、高效灵活、可扩展的深度学习平台，最初由百度科学家和工程师共同开发，目的是将深度学习技术应用到百度的众多产品中。 + +我们的愿景是让每个人都能通过PaddlePaddle接触深度学习 + +跟进PaddlePaddle最新特性请参考我们的[版本说明](https://github.com/PaddlePaddle/Paddle/releases) + +### PaddlePaddle最新版本: [Fluid 1.3.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.3) +### 安装最新稳定版本: +``` +# Linux CPU +pip install paddlepaddle +# Linux GPU cuda9cudnn7 +pip install paddlepaddle-gpu +# Linux GPU cuda8cudnn7 +pip install paddlepaddle-gpu==1.3.0.post87 +# Linux GPU cuda8cudnn5 +pip install paddlepaddle-gpu==1.3.0.post85 + +# 其他平台上的安装指引请参考 http://paddlepaddle.org/ +``` + +## 特性 + +- **灵活性** + + PaddlePaddle支持丰富的神经网络架构和优化算法。易于配置复杂模型，例如带有注意力机制或复杂记忆连接的神经网络机器翻译模型。 + +- **高效性** + + 为了高效使用异步计算资源，PaddlePaddle对框架的不同层进行优化，包括计算、存储、架构和通信。下面是一些样例： + + - 通过SSE/AVX 内置函数、BLAS库(例如MKL、OpenBLAS、cuBLAS)或定制的CPU/GPU内核优化数学操作。 + - 通过MKL-DNN库优化CNN网络 + - 高度优化循环网络，无需执行 `padding` 操作即可处理 **变长** 序列 + - 针对高维稀疏数据模型，优化了局部和分布式训练。 + + +- **稳定性** + + 有了 PaddlePaddle，使得利用各种CPU/GPU和机器来加速训练变得简单。PaddlePaddle 通过优化通信可以实现巨大吞吐量和快速执行。 + +- **与产品相连** + + 另外，PaddlePaddle 的设计也易于部署。在百度，PaddlePaddle 已经部署到含有巨大用户量的产品和服务上，包括广告点击率（CTR）预测、大规模图像分类、光学字符识别（OCR）、搜索排序，计算机病毒检测、推荐系统等等。PaddlePaddle广泛应用于百度产品中，产生了非常重要的影响。我们希望您也能探索 PaddlePaddle 的能力，为您的产品创造新的影响力和效果。 + +## 安装 + +推荐阅读官网上的[安装说明](http://paddlepaddle.org/documentation/docs/zh/1.3/beginners_guide/install/index_cn.html) + +## 文档 + +我们提供[英文](http://paddlepaddle.org/documentation/docs/en/1.3/beginners_guide/index_en.html)和 +[中文](http://paddlepaddle.org/documentation/docs/zh/1.3/beginners_guide/index.html) 文档 + +- [深度学习101](https://github.com/PaddlePaddle/book) + + 或许您想从这个在线交互式书籍开始，可以在Jupyter Notebook中运行 + +- [分布式训练](http://paddlepaddle.org/documentation/docs/zh/1.3/user_guides/howto/training/multi_node.html) + + 可以在MPI集群上运行分布式训练任务 + +- [Python API](http://paddlepaddle.org/documentation/docs/zh/1.3/api_cn/index_cn.html) + + 新的API支持代码更少更简洁的程序 + +- [贡献方式](http://paddlepaddle.org/documentation/docs/zh/1.3/advanced_usage/development/contribute_to_paddle/index_cn.html) + + 欢迎您的贡献! + +## 答疑 + +欢迎您将问题和bug报告以[Github Issues](https://github.com/PaddlePaddle/Paddle/issues)的形式提交 + +## 版权和许可证 +PaddlePaddle由[Apache-2.0 license](LICENSE)提供 diff --git a/benchmark/IntelOptimizedPaddle.md b/benchmark/IntelOptimizedPaddle.md deleted file mode 100644 index 8b7dc5b7db800896eb4de2054ab5e584aed93999..0000000000000000000000000000000000000000 --- a/benchmark/IntelOptimizedPaddle.md +++ /dev/null @@ -1,112 +0,0 @@ -# Benchmark - -Machine: - -- 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:0.11.0 (for MKLML and MKL-DNN) - - MKL-DNN tag v0.11 - - MKLML 2018.0.1.20171007 -- paddlepaddle/paddle:0.11.0-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 - -- VGG-19 - -| BatchSize | 64 | 128 | 256 | -|--------------|-------| -----| --------| -| OpenBLAS | 7.80 | 9.00 | 10.80 | -| MKLML | 12.12 | 13.70 | 16.18 | -| MKL-DNN | 28.46 | 29.83 | 30.44 | - - - - - 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 | - - - - - 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 | 45.62 | 72.79 | 107.22 | -| MKLML | 66.37 | 105.60 | 144.04 | -| MKL-DNN | 399.00 | 498.94 | 626.53 | - - - -#### 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 | 1.10 | 1.96 | 3.62 | 3.63 | 2.25 | -| 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 | 3.31 | 6.72 | 11.59 | 13.17 | 9.27 | -| 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 | 12.06 | 23.56 | 34.48 | 36.45 | 23.12 | -| 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 | 3.53 | 6.23 | 15.04 | 26.06 | 31.62 | -| MKLML | 21.32 | 36.55 | 73.06 | 131.15 | 192.77 | -| MKL-DNN | 442.91 | 656.41 | 719.10 | 847.68 | 850.51 | - - - -### Laptop -TBD diff --git a/benchmark/README.md b/benchmark/README.md deleted file mode 100644 index 367013f0457f9bbb9ae1335ea63dce181316d444..0000000000000000000000000000000000000000 --- a/benchmark/README.md +++ /dev/null @@ -1,168 +0,0 @@ -# Benchmark - -Machine: - -- CPU: 12-core Intel(R) Xeon(R) CPU E5-2620 v2 @2.10GHz -- GPU: Tesla K40m -- cuDNN: v5.1 -- system: Docker 1.12.1, all platforms are tested in docker environment. - -Platforms: - -- PaddlePaddle: paddledev/paddle:gpu-devel-v0.9.0a0 -- Tensorflow: gcr.io/tensorflow/tensorflow:0.11.0rc0-gpu -- Caffe: kaixhin/cuda-caffe - -Several convolutional neural networks and recurrent neural networks are used to test. - -## Image - -### Benchmark Model - -AlexNet, GoogleNet and a small network used in Caffe. - -- [AlexNet](https://github.com/BVLC/caffe/tree/master/models/bvlc_alexnet): but the group size is one. - -- [GoogleNet](https://github.com/BVLC/caffe/tree/master/models/bvlc_googlenet): but remove loss1 and loss2 when testing benchmark. - -- [SmallNet](https://github.com/BVLC/caffe/blob/master/examples/cifar10/cifar10\_quick\_train\_test.prototxt) - - -### Single-GPU - -- AlexNet: input - 3 * 227 * 227, Time: ms/batch - -| BatchSize | 64 | 128 | 256 | 512 | -|--------------|-----| -----| ------| -----| -| PaddlePaddle | 195 | 334 | 602 | 1629 | -| TensorFlow | 223 | 364 | 645 | 1235 | -| Caffe | 324 | 627 | 1232 | 2513 | - -**Notation** - -All platforms use cuDNN-v5.1. We see that caffe is slower in this experiment, because its workspace limit size of cuDNN-conv interface is 8 * 1024 * 1024, which is smaller in PaddlePaddle and TensorFlow. Note that Caffe will be faster if increasing the workspace limit size. - -- GoogletNet: input - 3 * 224 * 224, Time: ms/batch - - -| BatchSize | 64 | 128 | 256 | -|--------------|-------| -------| --------| -| PaddlePaddle | 613 | 1149 | 2348 | -| TensorFlow | 644 | 1176 | 2219 | -| Caffe | 694 | 1364 | out of memory | - -- SmallNet: input - 3 * 32 * 32, Time ms/batch - -| BatchSize | 64 | 128 | 256 | 512 | -|--------------|--------| -------- | --------|---------| -| PaddlePaddle | 10.463 | 18.184 | 33.113 | 63.039 | -| TensorFlow | 9 | 15 | 28 | 59 | -| Caffe | 9.373 | 16.6606 | 31.4797 | 59.719 | - -**Notation** - -All the single-GPU experiments in caffe use `caffe time` to calculate elapsed time, which does not include parameter updating time. However, both PaddlePaddle and TensorFlow experiments contain the parameter updating time. As compared with the total time, this part is relatively little on single machine, we can ignore it. - -In Tensorflow, they implement algorithm searching method instead of using the algorithm searching interface in cuDNN. - -### Multi-GPU: 4 GPUs - -- AlexNet, ms / batch - -| total-BatchSize | 128 * 4 | 256 * 4 | -|------------------|----------| -----------| -| PaddlePaddle | 347 | 622 | -| TensorFlow | 377 | 675 | -| Caffe | 1229 | 2435 | - -For example, if `total-BatchSize = 128 * 4`, the speedup ratio is calculated by - -``` - time_at_1gpu_batch_128 * 4 / time_at_4gpu_total_batch_512 -= (334 * 4)/347 -= 3.85 -``` - - - - -- GoogleNet, ms / batch - -| total-BatchSize | 128 * 4 | 256 * 4 | -|-------------------|--------------| ----------- | -| PaddlePaddle | 1178 | 2367 | -| TensorFlow | 1210 | 2292 | -| Caffe | 2007 | out of memory | - - - - -## RNN -We use lstm network for text classfication to test benchmark. - -### Dataset -- [IMDB](http://www.iro.umontreal.ca/~lisa/deep/data/imdb.pkl) -- Sequence length is 100. In fact, PaddlePaddle supports training with variable-length sequence, but TensorFlow needs to pad. Thus, we also pad sequence length to 100 in PaddlePaddle in order to compare. -- Dictionary size=30000 -- Peephole connection is used in `lstmemory` by default in PaddlePaddle. It is also configured in TensorFlow. - -### Single-GPU - -#### LSTM in Text Classification - -Testing `2 lstm layer + fc` network with different hidden size and batch size. - -- Batch size = 64, ms / batch - -| hidden_size | 256 | 512 | 1280 | -|--------------|-------| -------| --------| -| PaddlePaddle | 83 | 184 | 641 | -| TensorFlow | 175 | 280 | 818 | - -- Batch size = 128, ms / batch - -| hidden_size | 256 | 512 | 1280 | -|--------------|------- | -------| --------| -| PaddlePaddle | 110 | 261 | 1007 | -| TensorFlow | 181 | 361 | 1237 | - - -- Batch size = 256, ms / batch - -| hidden_size | 256 | 512 | 1280 | -|--------------|-------| -------| --------| -| PaddlePaddle | 170 | 414 | 1655 | -| TensorFlow | 238 | 536 | 1905 | - - - -#### Seq2Seq - -The benchmark of sequence-to-sequence network will be added later. - - -### Multi GPU: 4 GPUs - -#### LSTM in Text Classification - -- hidden_size = 256, ms / batch - -| batch_size | 256 | 512 | -|--------------| -------| --------| -| PaddlePaddle | 90 | 118 | -| TensorFlow | 226 | 118 | - - -- hidden_size = 512, ms / batch - -| batch_size | 256 | 512 | -|--------------| -------| --------| -| PaddlePaddle | 189 | 268 | -| TensorFlow | 297 | 383 | - - - - -#### Seq2Seq - -The benchmark of sequence-to-sequence network will be added later. diff --git a/benchmark/fluid/Dockerfile b/benchmark/fluid/Dockerfile index 2e1e0d376899fd664866621263db62258e7c3869..81ea870050fe5db4a60fee40221991e38de6bd2e 100644 --- a/benchmark/fluid/Dockerfile +++ b/benchmark/fluid/Dockerfile @@ -15,9 +15,6 @@ RUN ln -s /usr/lib/x86_64-linux-gnu/libcudnn.so.7 /usr/lib/libcudnn.so && ln -s RUN pip install -U pip RUN pip install -U kubernetes paddlepaddle -RUN sh -c 'echo "import paddle.v2 as paddle\npaddle.dataset.cifar.train10()\npaddle.dataset.flowers.fetch()" | python' -RUN sh -c 'echo "import paddle.v2 as paddle\npaddle.dataset.mnist.train()\npaddle.dataset.mnist.test()\npaddle.dataset.imdb.fetch()" | python' -RUN sh -c 'echo "import paddle.v2 as paddle\npaddle.dataset.imikolov.fetch()" | python' RUN pip uninstall -y paddlepaddle && mkdir /workspace ADD https://raw.githubusercontent.com/PaddlePaddle/cloud/develop/docker/paddle_k8s /usr/bin diff --git a/benchmark/paddle/image/check_env.sh b/benchmark/fluid/check_env.sh similarity index 100% rename from benchmark/paddle/image/check_env.sh rename to benchmark/fluid/check_env.sh diff --git a/benchmark/paddle/image/alexnet.py b/benchmark/paddle/image/alexnet.py deleted file mode 100644 index 9efc3f0494e4a817a7357f29e684f621bce1921e..0000000000000000000000000000000000000000 --- a/benchmark/paddle/image/alexnet.py +++ /dev/null @@ -1,93 +0,0 @@ -# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -from paddle.trainer_config_helpers import * - -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, - 'is_infer': is_infer, - 'num_samples': num_samples -} -define_py_data_sources2( - "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, - learning_rate=0.01 / batch_size, - learning_method=MomentumOptimizer(0.9), - regularization=L2Regularization(0.0005 * batch_size)) - -# conv1 -net = data_layer('data', size=height * width * 3) -net = img_conv_layer( - input=net, - filter_size=11, - num_channels=3, - num_filters=96, - stride=4, - padding=1) -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) - -# conv2 -net = img_conv_layer( - 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) - -# conv3 -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=gp) - -# conv5 -net = img_conv_layer( - 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( - input=net, - size=4096, - act=ReluActivation(), - layer_attr=ExtraAttr(drop_rate=0.5)) -net = fc_layer( - input=net, - size=4096, - act=ReluActivation(), - layer_attr=ExtraAttr(drop_rate=0.5)) -net = fc_layer(input=net, size=1000, act=SoftmaxActivation()) - -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 deleted file mode 100644 index 2a850ccb7f2c75b467554181fc5f4aa8f2b97a09..0000000000000000000000000000000000000000 --- a/benchmark/paddle/image/googlenet.py +++ /dev/null @@ -1,245 +0,0 @@ -#!/usr/bin/env python -from paddle.trainer_config_helpers import * - -height = 224 -width = 224 -num_class = 1000 -batch_size = get_config_arg('batch_size', int, 128) -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" if not is_infer else None, - "test.list" if is_infer else None, - module="provider", - obj="process", - args=args) - -settings( - batch_size=batch_size, - learning_rate=0.01 / batch_size, - 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, - filter5R, filter5, - proj): - - conv1 = name + '_1' - conv3r = name + '_3r' - conv3 = name + '_3' - conv5r = name + '_5r' - conv5 = name + '_5' - maxpool = name + '_max' - convproj = name + '_proj' - - cov1 = img_conv_layer( - name=conv1, - input=input, - filter_size=1, - num_channels=channels, - num_filters=filter1, - stride=1, - padding=0) - - cov3r = img_conv_layer( - name=conv3r, - input=input, - filter_size=1, - num_channels=channels, - num_filters=filter3R, - stride=1, - padding=0) - cov3 = img_conv_layer( - name=conv3, - input=cov3r, - filter_size=3, - num_filters=filter3, - stride=1, - padding=1) - - cov5r = img_conv_layer( - name=conv5r, - input=input, - filter_size=1, - num_channels=channels, - num_filters=filter5R, - stride=1, - padding=0) - cov5 = img_conv_layer( - name=conv5, - input=cov5r, - filter_size=5, - num_filters=filter5, - stride=1, - padding=2) - - pool1 = img_pool_layer( - name=maxpool, - input=input, - pool_size=3, - num_channels=channels, - stride=1, - padding=1) - covprj = img_conv_layer( - name=convproj, - input=pool1, - filter_size=1, - num_filters=proj, - stride=1, - padding=0) - - cat = concat_layer(name=name, input=[cov1, cov3, cov5, covprj]) - return cat - -def inception(name, input, channels, \ - filter1, - filter3R, filter3, - filter5R, filter5, - proj): - - cov1 = conv_projection( - input=input, - filter_size=1, - num_channels=channels, - num_filters=filter1, - stride=1, - padding=0) - - cov3r = img_conv_layer( - name=name + '_3r', - input=input, - filter_size=1, - num_channels=channels, - num_filters=filter3R, - stride=1, - padding=0) - cov3 = conv_projection( - input=cov3r, filter_size=3, num_filters=filter3, stride=1, padding=1) - - cov5r = img_conv_layer( - name=name + '_5r', - input=input, - filter_size=1, - num_channels=channels, - num_filters=filter5R, - stride=1, - padding=0) - cov5 = conv_projection( - input=cov5r, filter_size=5, num_filters=filter5, stride=1, padding=2) - - pool1 = img_pool_layer( - name=name + '_max', - input=input, - pool_size=3, - num_channels=channels, - stride=1, - padding=1) - covprj = conv_projection( - input=pool1, filter_size=1, num_filters=proj, stride=1, padding=0) - - cat = concat_layer( - name=name, - input=[cov1, cov3, cov5, covprj], - bias_attr=True if use_gpu else False, - act=ReluActivation()) - return cat - - -data = data_layer(name="input", size=3 * height * width) - -# stage 1 -conv1 = img_conv_layer( - name="conv1", - input=data, - filter_size=7, - num_channels=3, - num_filters=64, - stride=2, - padding=3) -pool1 = img_pool_layer( - name="pool1", input=conv1, pool_size=3, num_channels=64, stride=2) - -# stage 2 -conv2_1 = img_conv_layer( - name="conv2_1", - input=pool1, - filter_size=1, - num_filters=64, - stride=1, - padding=0) -conv2_2 = img_conv_layer( - name="conv2_2", - input=conv2_1, - filter_size=3, - num_filters=192, - stride=1, - padding=1) -pool2 = img_pool_layer( - name="pool2", input=conv2_2, pool_size=3, num_channels=192, stride=2) - -# stage 3 -ince3a = inception("ince3a", pool2, 192, 64, 96, 128, 16, 32, 32) -ince3b = inception("ince3b", ince3a, 256, 128, 128, 192, 32, 96, 64) -pool3 = img_pool_layer( - name="pool3", input=ince3b, num_channels=480, pool_size=3, stride=2) - -# stage 4 -ince4a = inception("ince4a", pool3, 480, 192, 96, 208, 16, 48, 64) -ince4b = inception("ince4b", ince4a, 512, 160, 112, 224, 24, 64, 64) -ince4c = inception("ince4c", ince4b, 512, 128, 128, 256, 24, 64, 64) -ince4d = inception("ince4d", ince4c, 512, 112, 144, 288, 32, 64, 64) -ince4e = inception("ince4e", ince4d, 528, 256, 160, 320, 32, 128, 128) -pool4 = img_pool_layer( - name="pool4", input=ince4e, num_channels=832, pool_size=3, stride=2) - -# stage 5 -ince5a = inception("ince5a", pool4, 832, 256, 160, 320, 32, 128, 128) -ince5b = inception("ince5b", ince5a, 832, 384, 192, 384, 48, 128, 128) -pool5 = img_pool_layer( - name="pool5", - input=ince5b, - num_channels=1024, - pool_size=7, - stride=7, - pool_type=AvgPooling()) - -# We remove loss1 and loss2 for all system when testing benchmark -# output 1 -# pool_o1 = img_pool_layer(name="pool_o1", input=ince4a, num_channels=512, pool_size=5, stride=3, pool_type=AvgPooling()) -# conv_o1 = img_conv_layer(name="conv_o1", input=pool_o1, filter_size=1, num_filters=128, stride=1, padding=0) -# fc_o1 = fc_layer(name="fc_o1", input=conv_o1, size=1024, layer_attr=ExtraAttr(drop_rate=0.7), act=ReluActivation()) -# out1 = fc_layer(name="output1", input=fc_o1, size=1000, act=SoftmaxActivation()) -# loss1 = cross_entropy(name='loss1', input=out1, label=lab, coeff=0.3) - -# output 2 -#pool_o2 = img_pool_layer(name="pool_o2", input=ince4d, num_channels=528, pool_size=5, stride=3, pool_type=AvgPooling()) -#conv_o2 = img_conv_layer(name="conv_o2", input=pool_o2, filter_size=1, num_filters=128, stride=1, padding=0) -#fc_o2 = fc_layer(name="fc_o2", input=conv_o2, size=1024, layer_attr=ExtraAttr(drop_rate=0.7), act=ReluActivation()) -#out2 = fc_layer(name="output2", input=fc_o2, size=1000, act=SoftmaxActivation()) -#loss2 = cross_entropy(name='loss2', input=out2, label=lab, coeff=0.3) - -# output 3 -dropout = dropout_layer(name="dropout", input=pool5, dropout_rate=0.4) -out3 = fc_layer( - name="output3", input=dropout, size=1000, act=SoftmaxActivation()) - -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/plotlog.py b/benchmark/paddle/image/plotlog.py deleted file mode 100644 index 8679d4f272d1b7aaf8d5a397f07698a6b70e4fcd..0000000000000000000000000000000000000000 --- a/benchmark/paddle/image/plotlog.py +++ /dev/null @@ -1,114 +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. - -import sys -import argparse -import matplotlib.pyplot as plt - - -def parse_args(): - parser = argparse.ArgumentParser('Parse Log') - parser.add_argument( - '--file_path', '-f', type=str, help='the path of the log file') - parser.add_argument( - '--sample_rate', - '-s', - type=float, - default=1.0, - help='the rate to take samples from log') - parser.add_argument( - '--log_period', '-p', type=int, default=1, help='the period of log') - - args = parser.parse_args() - return args - - -def parse_file(file_name): - loss = [] - error = [] - with open(file_name) as f: - for i, line in enumerate(f): - line = line.strip() - if not line.startswith('pass'): - continue - line_split = line.split(' ') - if len(line_split) != 5: - continue - - loss_str = line_split[2][:-1] - cur_loss = float(loss_str.split('=')[-1]) - loss.append(cur_loss) - - err_str = line_split[3][:-1] - cur_err = float(err_str.split('=')[-1]) - error.append(cur_err) - - accuracy = [1.0 - err for err in error] - - return loss, accuracy - - -def sample(metric, sample_rate): - interval = int(1.0 / sample_rate) - if interval > len(metric): - return metric[:1] - - num = len(metric) / interval - idx = [interval * i for i in range(num)] - metric_sample = [metric[id] for id in idx] - return metric_sample - - -def plot_metric(metric, - batch_id, - graph_title, - line_style='b-', - line_label='y', - line_num=1): - plt.figure() - plt.title(graph_title) - if line_num == 1: - plt.plot(batch_id, metric, line_style, label=line_label) - else: - for i in range(line_num): - plt.plot(batch_id, metric[i], line_style[i], label=line_label[i]) - plt.xlabel('batch') - plt.ylabel(graph_title) - plt.legend() - plt.savefig(graph_title + '.jpg') - plt.close() - - -def main(): - args = parse_args() - assert args.sample_rate > 0. and args.sample_rate <= 1.0, "The sample rate should in the range (0, 1]." - - loss, accuracy = parse_file(args.file_path) - batch = [args.log_period * i for i in range(len(loss))] - - batch_sample = sample(batch, args.sample_rate) - loss_sample = sample(loss, args.sample_rate) - accuracy_sample = sample(accuracy, args.sample_rate) - - plot_metric(loss_sample, batch_sample, 'loss', line_label='loss') - plot_metric( - accuracy_sample, - batch_sample, - 'accuracy', - line_style='g-', - line_label='accuracy') - - -if __name__ == '__main__': - main() diff --git a/benchmark/paddle/image/provider.py b/benchmark/paddle/image/provider.py deleted file mode 100644 index 6ad817ccefab3e44a8f962e907ba2110a6ed4a45..0000000000000000000000000000000000000000 --- a/benchmark/paddle/image/provider.py +++ /dev/null @@ -1,47 +0,0 @@ -# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -import io, os -import random -import numpy as np -from paddle.trainer.PyDataProvider2 import * - - -def initHook(settings, height, width, color, num_class, **kwargs): - settings.height = height - settings.width = width - settings.color = color - settings.num_class = num_class - if settings.color: - settings.data_size = settings.height * settings.width * 3 - else: - settings.data_size = settings.height * settings.width - 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(settings.num_samples): - img = np.random.rand(1, settings.data_size).reshape(-1, 1).flatten() - 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 deleted file mode 100644 index 2846e4763f1cda4602f03af5ec649d57ee6cf0d8..0000000000000000000000000000000000000000 --- a/benchmark/paddle/image/resnet.py +++ /dev/null @@ -1,230 +0,0 @@ -#!/usr/bin/env python -from paddle.trainer_config_helpers import * - -height = 224 -width = 224 -num_class = 1000 -batch_size = get_config_arg('batch_size', int, 64) -layer_num = get_config_arg("layer_num", int, 50) -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" if not is_infer else None, - "test.list" if is_infer else None, - module="provider", - obj="process", - args=args) - -settings( - batch_size=batch_size, - learning_rate=0.01 / batch_size, - learning_method=MomentumOptimizer(0.9), - regularization=L2Regularization(0.0005 * batch_size)) - - -#######################Network Configuration ############# -def conv_bn_layer(name, - input, - filter_size, - num_filters, - stride, - padding, - channels=None, - active_type=ReluActivation()): - """ - A wrapper for conv layer with batch normalization layers. - Note: - conv layer has no activation. - """ - - tmp = img_conv_layer( - name=name + "_conv", - input=input, - filter_size=filter_size, - num_channels=channels, - num_filters=num_filters, - stride=stride, - padding=padding, - act=LinearActivation(), - bias_attr=False) - return batch_norm_layer( - name=name + "_bn", - input=tmp, - act=active_type, - use_global_stats=is_infer) - - -def bottleneck_block(name, input, num_filters1, num_filters2): - """ - A wrapper for bottlenect building block in ResNet. - Last conv_bn_layer has no activation. - Addto layer has activation of relu. - """ - last_name = conv_bn_layer( - name=name + '_branch2a', - input=input, - filter_size=1, - num_filters=num_filters1, - stride=1, - padding=0) - last_name = conv_bn_layer( - name=name + '_branch2b', - input=last_name, - filter_size=3, - num_filters=num_filters1, - stride=1, - padding=1) - last_name = conv_bn_layer( - name=name + '_branch2c', - input=last_name, - filter_size=1, - num_filters=num_filters2, - stride=1, - padding=0, - active_type=LinearActivation()) - - return addto_layer( - name=name + "_addto", input=[input, last_name], act=ReluActivation()) - - -def mid_projection(name, input, num_filters1, num_filters2, stride=2): - """ - A wrapper for middile projection in ResNet. - projection shortcuts are used for increasing dimensions, - and other shortcuts are identity - branch1: projection shortcuts are used for increasing - dimensions, has no activation. - branch2x: bottleneck building block, shortcuts are identity. - """ - # stride = 2 - branch1 = conv_bn_layer( - name=name + '_branch1', - input=input, - filter_size=1, - num_filters=num_filters2, - stride=stride, - padding=0, - active_type=LinearActivation()) - - last_name = conv_bn_layer( - name=name + '_branch2a', - input=input, - filter_size=1, - num_filters=num_filters1, - stride=stride, - padding=0) - last_name = conv_bn_layer( - name=name + '_branch2b', - input=last_name, - filter_size=3, - num_filters=num_filters1, - stride=1, - padding=1) - - last_name = conv_bn_layer( - name=name + '_branch2c', - input=last_name, - filter_size=1, - num_filters=num_filters2, - stride=1, - padding=0, - active_type=LinearActivation()) - - return addto_layer( - name=name + "_addto", input=[branch1, last_name], act=ReluActivation()) - - -img = data_layer(name='image', size=height * width * 3) - - -def deep_res_net(res2_num=3, res3_num=4, res4_num=6, res5_num=3): - """ - A wrapper for 50,101,152 layers of ResNet. - res2_num: number of blocks stacked in conv2_x - res3_num: number of blocks stacked in conv3_x - res4_num: number of blocks stacked in conv4_x - res5_num: number of blocks stacked in conv5_x - """ - # For ImageNet - # conv1: 112x112 - tmp = conv_bn_layer( - "conv1", - input=img, - filter_size=7, - channels=3, - num_filters=64, - stride=2, - padding=3) - tmp = img_pool_layer(name="pool1", input=tmp, pool_size=3, stride=2) - - # conv2_x: 56x56 - tmp = mid_projection( - name="res2_1", input=tmp, num_filters1=64, num_filters2=256, stride=1) - for i in xrange(2, res2_num + 1, 1): - tmp = bottleneck_block( - name="res2_" + str(i), input=tmp, num_filters1=64, num_filters2=256) - - # conv3_x: 28x28 - tmp = mid_projection( - name="res3_1", input=tmp, num_filters1=128, num_filters2=512) - for i in xrange(2, res3_num + 1, 1): - tmp = bottleneck_block( - name="res3_" + str(i), - input=tmp, - num_filters1=128, - num_filters2=512) - - # conv4_x: 14x14 - tmp = mid_projection( - name="res4_1", input=tmp, num_filters1=256, num_filters2=1024) - for i in xrange(2, res4_num + 1, 1): - tmp = bottleneck_block( - name="res4_" + str(i), - input=tmp, - num_filters1=256, - num_filters2=1024) - - # conv5_x: 7x7 - tmp = mid_projection( - name="res5_1", input=tmp, num_filters1=512, num_filters2=2048) - for i in xrange(2, res5_num + 1, 1): - tmp = bottleneck_block( - name="res5_" + str(i), - input=tmp, - num_filters1=512, - num_filters2=2048) - - tmp = img_pool_layer( - name='avgpool', - input=tmp, - pool_size=7, - stride=1, - pool_type=AvgPooling()) - - return fc_layer(input=tmp, size=num_class, act=SoftmaxActivation()) - - -if layer_num == 50: - resnet = deep_res_net(3, 4, 6, 3) -elif layer_num == 101: - resnet = deep_res_net(3, 4, 23, 3) -elif layer_num == 152: - resnet = deep_res_net(3, 8, 36, 3) -else: - print("Wrong layer number.") - -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.sh b/benchmark/paddle/image/run.sh deleted file mode 100755 index 5b58a8d773aab795e5439b0f0e5d81bec66b5f56..0000000000000000000000000000000000000000 --- a/benchmark/paddle/image/run.sh +++ /dev/null @@ -1,53 +0,0 @@ -#!/bin/bash - -set -e - -function train() { - cfg=$1 - thread=$2 - bz=$3 - args="batch_size=$3" - prefix=$4 - paddle train --job=time \ - --config=$cfg \ - --use_gpu=True \ - --trainer_count=$thread \ - --log_period=10 \ - --test_period=100 \ - --config_args=$args \ - > logs/$prefix-${thread}gpu-$bz.log 2>&1 -} - -if [ ! -d "train.list" ]; then - echo " " > train.list -fi -if [ ! -d "logs" ]; then - mkdir logs -fi - -#========single-gpu=========# -# alexnet -train alexnet.py 1 64 alexnet -train alexnet.py 1 128 alexnet -train alexnet.py 1 256 alexnet -train alexnet.py 1 512 alexnet - -# googlenet -train googlenet.py 1 64 googlenet -train googlenet.py 1 128 googlenet -train googlenet.py 1 256 googlenet - -# smallnet -train smallnet_mnist_cifar.py 1 64 smallnet -train smallnet_mnist_cifar.py 1 128 smallnet -train smallnet_mnist_cifar.py 1 256 smallnet -train smallnet_mnist_cifar.py 1 512 smallnet - - -############################ -#========multi-gpus=========# -train alexnet.py 4 512 alexnet -train alexnet.py 4 1024 alexnet - -train googlenet.py 4 512 googlenet -train googlenet.py 4 1024 googlenet diff --git a/benchmark/paddle/image/run_mkl_infer.sh b/benchmark/paddle/image/run_mkl_infer.sh deleted file mode 100755 index 0fad5e04cc992a3ec97591d3833957bb7517a8f3..0000000000000000000000000000000000000000 --- a/benchmark/paddle/image/run_mkl_infer.sh +++ /dev/null @@ -1,89 +0,0 @@ -#!/bin/bash - -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_mkl_train.sh b/benchmark/paddle/image/run_mkl_train.sh deleted file mode 100755 index 1583bf134a276a08aa2f8e84dc63adbb205a83d6..0000000000000000000000000000000000000000 --- a/benchmark/paddle/image/run_mkl_train.sh +++ /dev/null @@ -1,54 +0,0 @@ -#!/bin/bash - -set -e - -function train() { - unset OMP_NUM_THREADS MKL_NUM_THREADS OMP_DYNAMIC KMP_AFFINITY - topology=$1 - layer_num=$2 - bs=$3 - use_mkldnn=$4 - if [ $4 == "True" ]; then - thread=1 - log="logs/train-${topology}-${layer_num}-mkldnn-${bs}.log" - elif [ $4 == "False" ]; then - thread=`nproc` - # each trainer_count use only 1 core to avoid conflict - log="logs/train-${topology}-${layer_num}-${thread}mklml-${bs}.log" - else - echo "Wrong input $4, use True or False." - exit 0 - fi - args="batch_size=${bs},layer_num=${layer_num}" - config="${topology}.py" - paddle train --job=time \ - --config=$config \ - --use_mkldnn=$use_mkldnn \ - --use_gpu=False \ - --trainer_count=$thread \ - --log_period=10 \ - --test_period=100 \ - --config_args=$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 [ ! -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 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 deleted file mode 100755 index 987381cabc2e793886099212660723c122b73bb0..0000000000000000000000000000000000000000 --- a/benchmark/paddle/image/run_openblas_infer.sh +++ /dev/null @@ -1,71 +0,0 @@ -#!/bin/bash - -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() { - export OPENBLAS_MAIN_FREE=1 - 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 deleted file mode 100755 index cc64e1d09da02087b1737190a0b75dc7758600a6..0000000000000000000000000000000000000000 --- a/benchmark/paddle/image/run_openblas_train.sh +++ /dev/null @@ -1,43 +0,0 @@ -#!/bin/bash - -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/smallnet_mnist_cifar.py b/benchmark/paddle/image/smallnet_mnist_cifar.py deleted file mode 100644 index 58879c454f37991405d83bbb593bb5d1e977ff53..0000000000000000000000000000000000000000 --- a/benchmark/paddle/image/smallnet_mnist_cifar.py +++ /dev/null @@ -1,49 +0,0 @@ -#!/usr/bin/env python - -from paddle.trainer_config_helpers import * - -height = 32 -width = 32 -num_class = 10 - -batch_size = get_config_arg('batch_size', int, 128) - -args = {'height': height, 'width': width, 'color': True, 'num_class': num_class} -define_py_data_sources2( - "train.list", None, module="provider", obj="process", args=args) - -settings( - batch_size=batch_size, - learning_rate=0.01 / batch_size, - learning_method=MomentumOptimizer(0.9), - regularization=L2Regularization(0.0005 * batch_size)) - -# conv1 -net = data_layer('data', size=height * width * 3) -net = img_conv_layer( - input=net, - filter_size=5, - num_channels=3, - num_filters=32, - stride=1, - padding=2) -net = img_pool_layer(input=net, pool_size=3, stride=2, padding=1) - -# conv2 -net = img_conv_layer( - input=net, filter_size=5, num_filters=32, stride=1, padding=2) -net = img_pool_layer( - input=net, pool_size=3, stride=2, padding=1, pool_type=AvgPooling()) - -# conv3 -net = img_conv_layer( - input=net, filter_size=3, num_filters=64, stride=1, padding=1) -net = img_pool_layer( - input=net, pool_size=3, stride=2, padding=1, pool_type=AvgPooling()) - -net = fc_layer(input=net, size=64, act=ReluActivation()) -net = fc_layer(input=net, size=10, act=SoftmaxActivation()) - -lab = data_layer('label', num_class) -loss = classification_cost(input=net, label=lab) -outputs(loss) diff --git a/benchmark/paddle/image/vgg.py b/benchmark/paddle/image/vgg.py deleted file mode 100644 index ca0a6798fb8c35b68cf84d263855955eb93ba0b0..0000000000000000000000000000000000000000 --- a/benchmark/paddle/image/vgg.py +++ /dev/null @@ -1,119 +0,0 @@ -#!/usr/bin/env python -from paddle.trainer_config_helpers import * - -height = 224 -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, - 'is_infer': is_infer, - 'num_samples': num_samples -} -define_py_data_sources2( - "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, - learning_rate=0.001 / batch_size, - learning_method=MomentumOptimizer(0.9), - regularization=L2Regularization(0.0005 * batch_size)) - -img = data_layer(name='image', size=height * width * 3) - - -def vgg_network(vgg_num=3): - tmp = img_conv_group( - input=img, - num_channels=3, - conv_padding=1, - conv_num_filter=[64, 64], - conv_filter_size=3, - conv_act=ReluActivation(), - pool_size=2, - pool_stride=2, - pool_type=MaxPooling()) - - tmp = img_conv_group( - input=tmp, - conv_num_filter=[128, 128], - conv_padding=1, - conv_filter_size=3, - conv_act=ReluActivation(), - pool_stride=2, - pool_type=MaxPooling(), - pool_size=2) - - channels = [] - for i in range(vgg_num): - channels.append(256) - tmp = img_conv_group( - input=tmp, - conv_num_filter=channels, - conv_padding=1, - conv_filter_size=3, - conv_act=ReluActivation(), - pool_stride=2, - pool_type=MaxPooling(), - pool_size=2) - channels = [] - for i in range(vgg_num): - channels.append(512) - tmp = img_conv_group( - input=tmp, - conv_num_filter=channels, - conv_padding=1, - conv_filter_size=3, - conv_act=ReluActivation(), - pool_stride=2, - pool_type=MaxPooling(), - pool_size=2) - tmp = img_conv_group( - input=tmp, - conv_num_filter=channels, - conv_padding=1, - conv_filter_size=3, - conv_act=ReluActivation(), - pool_stride=2, - pool_type=MaxPooling(), - pool_size=2) - - tmp = fc_layer( - input=tmp, - size=4096, - act=ReluActivation(), - layer_attr=ExtraAttr(drop_rate=0.5)) - - tmp = fc_layer( - input=tmp, - size=4096, - act=ReluActivation(), - layer_attr=ExtraAttr(drop_rate=0.5)) - - return fc_layer(input=tmp, size=num_class, act=SoftmaxActivation()) - - -if layer_num == 16: - vgg = vgg_network(3) -elif layer_num == 19: - vgg = vgg_network(4) -else: - print("Wrong layer number.") - -if is_infer: - outputs(vgg) -else: - lab = data_layer('label', num_class) - loss = cross_entropy(input=vgg, label=lab) - outputs(loss) diff --git a/benchmark/paddle/rnn/imdb.py b/benchmark/paddle/rnn/imdb.py deleted file mode 100755 index 2a67f9b0cf52484d9d44fe9db0b1e57cdd20fd43..0000000000000000000000000000000000000000 --- a/benchmark/paddle/rnn/imdb.py +++ /dev/null @@ -1,60 +0,0 @@ -# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -from __future__ import print_function -import six.moves.cPickle as pickle -import gzip -import os -import numpy - - -def get_dataset_file(dataset, default_dataset, origin): - data_dir, data_file = os.path.split(dataset) - if (not os.path.isfile(dataset)) and data_file == default_dataset: - from six.moves import urllib - print('Downloading data from %s' % origin) - urllib.request.urlretrieve(origin, dataset) - - return dataset - - -def create_data(path="imdb.pkl"): - - if (not os.path.isfile('imdb.train.pkl')): - path = get_dataset_file( - path, "imdb.pkl", - "http://www.iro.umontreal.ca/~lisa/deep/data/imdb.pkl") - - if path.endswith(".gz"): - f = gzip.open(path, 'rb') - else: - f = open(path, 'rb') - - train_set = pickle.load(f) - test_set = pickle.load(f) - f.close() - - pickle.dump(train_set, open('imdb.train.pkl', 'wb')) - pickle.dump(test_set, open('imdb.test.pkl', 'wb')) - - if (not os.path.isfile('train.list')): - file('train.list', 'w').write('imdb.train.pkl\n') - - -def main(): - create_data('imdb.pkl') - - -if __name__ == "__main__": - main() diff --git a/benchmark/paddle/rnn/provider.py b/benchmark/paddle/rnn/provider.py deleted file mode 100644 index 23cc0c44a98d0ae7f586d1a376a603198f2c6144..0000000000000000000000000000000000000000 --- a/benchmark/paddle/rnn/provider.py +++ /dev/null @@ -1,86 +0,0 @@ -# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -import io, os -import random -import numpy as np -import six.moves.cPickle as pickle -from paddle.trainer.PyDataProvider2 import * - - -def remove_unk(x, n_words): - return [[1 if w >= n_words else w for w in sen] for sen in x] - - -# ============================================================== -# tensorflow uses fixed length, but PaddlePaddle can process -# variable-length. Padding is used in benchmark in order to -# compare with other platform. -# ============================================================== -def pad_sequences(sequences, - maxlen=None, - dtype='int32', - padding='post', - truncating='post', - value=0.): - lengths = [len(s) for s in sequences] - - nb_samples = len(sequences) - if maxlen is None: - maxlen = np.max(lengths) - - x = (np.ones((nb_samples, maxlen)) * value).astype(dtype) - for idx, s in enumerate(sequences): - if len(s) == 0: - continue # empty list was found - if truncating == 'pre': - trunc = s[-maxlen:] - elif truncating == 'post': - trunc = s[:maxlen] - else: - raise ValueError("Truncating type '%s' not understood" % padding) - - if padding == 'post': - x[idx, :len(trunc)] = trunc - elif padding == 'pre': - x[idx, -len(trunc):] = trunc - else: - raise ValueError("Padding type '%s' not understood" % padding) - return x - - -def initHook(settings, vocab_size, pad_seq, maxlen, **kwargs): - settings.vocab_size = vocab_size - settings.pad_seq = pad_seq - settings.maxlen = maxlen - settings.input_types = [ - integer_value_sequence(vocab_size), integer_value(2) - ] - - -@provider( - init_hook=initHook, min_pool_size=-1, cache=CacheType.CACHE_PASS_IN_MEM) -def process(settings, file): - f = open(file, 'rb') - train_set = pickle.load(f) - f.close() - x, y = train_set - - # remove unk, namely remove the words out of dictionary - x = remove_unk(x, settings.vocab_size) - if settings.pad_seq: - x = pad_sequences(x, maxlen=settings.maxlen, value=0.) - - for i in range(len(y)): - yield map(int, x[i]), int(y[i]) diff --git a/benchmark/paddle/rnn/rnn.py b/benchmark/paddle/rnn/rnn.py deleted file mode 100755 index 83eb3e565473f7e7e91cddeaa3cd2aafb7e3df2c..0000000000000000000000000000000000000000 --- a/benchmark/paddle/rnn/rnn.py +++ /dev/null @@ -1,38 +0,0 @@ -#!/usr/bin/env python - -from paddle.trainer_config_helpers import * -import imdb - -num_class = 2 -vocab_size = 30000 -fixedlen = 100 -batch_size = get_config_arg('batch_size', int, 128) -lstm_num = get_config_arg('lstm_num', int, 1) -hidden_size = get_config_arg('hidden_size', int, 128) -# whether to pad sequence into fixed length -pad_seq = get_config_arg('pad_seq', bool, True) -imdb.create_data('imdb.pkl') - -args = {'vocab_size': vocab_size, 'pad_seq': pad_seq, 'maxlen': fixedlen} -define_py_data_sources2( - "train.list", None, module="provider", obj="process", args=args) - -settings( - batch_size=batch_size, - learning_rate=2e-3, - learning_method=AdamOptimizer(), - regularization=L2Regularization(8e-4), - gradient_clipping_threshold=25) - -net = data_layer('data', size=vocab_size) -net = embedding_layer(input=net, size=128) - -for i in xrange(lstm_num): - net = simple_lstm(input=net, size=hidden_size) - -net = last_seq(input=net) -net = fc_layer(input=net, size=2, act=SoftmaxActivation()) - -lab = data_layer('label', num_class) -loss = classification_cost(input=net, label=lab) -outputs(loss) diff --git a/benchmark/paddle/rnn/run.sh b/benchmark/paddle/rnn/run.sh deleted file mode 100755 index f99a562b3f88a98560f4bf7aee98ceee9daefe67..0000000000000000000000000000000000000000 --- a/benchmark/paddle/rnn/run.sh +++ /dev/null @@ -1,52 +0,0 @@ -#!/bin/bash - -set -e - -function train() { - cfg=$1 - thread=$2 - args="lstm_num=${3},seq_pad=${4},hidden_size=${5},batch_size=${6}" - paddle train --job=time \ - --config=$cfg \ - --use_gpu=1 \ - --trainer_count=$thread \ - --log_period=10 \ - --test_period=100 \ - --num_passes=1 \ - --feed_data=1 \ - --config_args=$args \ - >logs/rnn-pad${4}-${thread}gpu-lstm${3}-batch${6}-hid${5}.log 2>&1 -} - -if [ ! -d "logs" ]; then - mkdir logs -fi - -## padding, single gpu -#-----config--gpu--lstm_num--padding--hidden_size--batch_size -## lstm_num=2, batch_size=64 -train rnn.py 1 2 1 256 64 -train rnn.py 1 2 1 512 64 -train rnn.py 1 2 1 1280 64 - -## lstm_num=2, batch_size=128 -train rnn.py 1 2 1 256 128 -train rnn.py 1 2 1 512 128 -train rnn.py 1 2 1 1280 128 - -## lstm_num=4, batch_size=256 -train rnn.py 1 2 1 256 256 -train rnn.py 1 2 1 512 256 -train rnn.py 1 2 1 1280 256 - - -#==================multi gpus=====================# -# hidden_size=256, lstm_num=2, different batch size -train rnn.py 4 2 1 256 128 -train rnn.py 4 2 1 256 256 -train rnn.py 4 2 1 256 512 - -# hidden_size=512, lstm_num=4, different batch size -train rnn.py 4 2 1 512 128 -train rnn.py 4 2 1 512 256 -train rnn.py 4 2 1 512 512 diff --git a/benchmark/tensorflow/machine_translation.py b/benchmark/tensorflow/machine_translation.py index 8f77dce98353af53803246be8dc61063836b7867..7837669edc7a206c03e5b9fa2989bf45b35f0605 100644 --- a/benchmark/tensorflow/machine_translation.py +++ b/benchmark/tensorflow/machine_translation.py @@ -35,8 +35,6 @@ import os import argparse import time -import paddle.v2 as paddle - parser = argparse.ArgumentParser(description=__doc__) parser.add_argument( "--embedding_dim", diff --git a/benchmark/tensorflow/mnist.py b/benchmark/tensorflow/mnist.py index 7140eed6eaff49b5c65f9ccb2e38f113a4cdbdbf..03d533fecfededddd3956ba83ea600456782cfc9 100644 --- a/benchmark/tensorflow/mnist.py +++ b/benchmark/tensorflow/mnist.py @@ -21,7 +21,6 @@ import time import numpy as np import tensorflow as tf -import paddle.v2 as paddle DTYPE = tf.float32 diff --git a/benchmark/tensorflow/resnet.py b/benchmark/tensorflow/resnet.py index c432fa8d59571e128b9ff9e3ffa1949b792ef3a4..fdb044195766b847e16a0cc33424a999c1d9166e 100644 --- a/benchmark/tensorflow/resnet.py +++ b/benchmark/tensorflow/resnet.py @@ -27,7 +27,6 @@ import argparse import time import numpy as np -import paddle.v2 as paddle import tensorflow as tf DTYPE = tf.float32 diff --git a/benchmark/tensorflow/stacked_dynamic_lstm.py b/benchmark/tensorflow/stacked_dynamic_lstm.py index 5285033005044d907d0b7e91eb66ee7281c4f27a..1f532dc2fa082ea0f6b1da560e1a57b96d2ef1bb 100644 --- a/benchmark/tensorflow/stacked_dynamic_lstm.py +++ b/benchmark/tensorflow/stacked_dynamic_lstm.py @@ -21,8 +21,6 @@ import argparse import time import tensorflow as tf -import paddle.v2 as paddle - def parse_args(): parser = argparse.ArgumentParser("LSTM model benchmark.") diff --git a/benchmark/tensorflow/vgg.py b/benchmark/tensorflow/vgg.py index fba5ec71a46b3ac8b2e1244424c39fd5192e5458..d32c835bd7a7dafaafe0970fb6b422db3c866370 100644 --- a/benchmark/tensorflow/vgg.py +++ b/benchmark/tensorflow/vgg.py @@ -13,7 +13,6 @@ # limitations under the License. """VGG16 benchmark in TensorFlow""" import tensorflow as tf -import paddle.v2 as paddle import numpy as np import argparse import time diff --git a/cmake/FindSphinx.cmake b/cmake/FindSphinx.cmake deleted file mode 100644 index f74cd4ff8c9c2c52319b18ac37264167b3718eae..0000000000000000000000000000000000000000 --- a/cmake/FindSphinx.cmake +++ /dev/null @@ -1,147 +0,0 @@ -# - This module looks for Sphinx -# Find the Sphinx documentation generator -# -# This modules defines -# SPHINX_EXECUTABLE -# SPHINX_FOUND - -find_program(SPHINX_EXECUTABLE - NAMES sphinx-build - PATHS - /usr/bin - /usr/local/bin - /opt/local/bin - DOC "Sphinx documentation generator" -) - -if( NOT SPHINX_EXECUTABLE ) - set(_Python_VERSIONS - 2.7 2.6 2.5 2.4 2.3 2.2 2.1 2.0 1.6 1.5 - ) - - foreach( _version ${_Python_VERSIONS} ) - set( _sphinx_NAMES sphinx-build-${_version} ) - - find_program( SPHINX_EXECUTABLE - NAMES ${_sphinx_NAMES} - PATHS - /usr/bin - /usr/local/bin - /opt/loca/bin - DOC "Sphinx documentation generator" - ) - endforeach() -endif() - -include(FindPackageHandleStandardArgs) - -find_package_handle_standard_args(Sphinx DEFAULT_MSG - SPHINX_EXECUTABLE -) - - -option( SPHINX_HTML_OUTPUT "Build a single HTML with the whole content." ON ) -option( SPHINX_DIRHTML_OUTPUT "Build HTML pages, but with a single directory per document." OFF ) -option( SPHINX_HTMLHELP_OUTPUT "Build HTML pages with additional information for building a documentation collection in htmlhelp." OFF ) -option( SPHINX_QTHELP_OUTPUT "Build HTML pages with additional information for building a documentation collection in qthelp." OFF ) -option( SPHINX_DEVHELP_OUTPUT "Build HTML pages with additional information for building a documentation collection in devhelp." OFF ) -option( SPHINX_EPUB_OUTPUT "Build HTML pages with additional information for building a documentation collection in epub." OFF ) -option( SPHINX_LATEX_OUTPUT "Build LaTeX sources that can be compiled to a PDF document using pdflatex." OFF ) -option( SPHINX_MAN_OUTPUT "Build manual pages in groff format for UNIX systems." OFF ) -option( SPHINX_TEXT_OUTPUT "Build plain text files." OFF ) - - -mark_as_advanced( - SPHINX_EXECUTABLE - SPHINX_HTML_OUTPUT - SPHINX_DIRHTML_OUTPUT - SPHINX_HTMLHELP_OUTPUT - SPHINX_QTHELP_OUTPUT - SPHINX_DEVHELP_OUTPUT - SPHINX_EPUB_OUTPUT - SPHINX_LATEX_OUTPUT - SPHINX_MAN_OUTPUT - SPHINX_TEXT_OUTPUT -) - -function( Sphinx_add_target target_name builder conf cache source destination ) - add_custom_target( ${target_name} ALL - COMMAND ${SPHINX_EXECUTABLE} -b ${builder} - -d ${cache} - -c ${conf} - ${source} - ${destination} - COMMENT "Generating sphinx documentation: ${builder}" - COMMAND cd ${destination} && ln -sf ./index_*.html index.html - ) - - set_property( - DIRECTORY APPEND PROPERTY - ADDITIONAL_MAKE_CLEAN_FILES - ${destination} - ) -endfunction() - -# Target dependencies can be optionally listed at the end. -function( Sphinx_add_targets target_base_name conf source base_destination ) - - set( _dependencies ) - - foreach( arg IN LISTS ARGN ) - set( _dependencies ${_dependencies} ${arg} ) - endforeach() - - if( ${SPHINX_HTML_OUTPUT} ) - Sphinx_add_target( ${target_base_name}_html html ${conf} ${source} ${base_destination}/html ) - - add_dependencies( ${target_base_name}_html ${_dependencies} ) - endif() - - if( ${SPHINX_DIRHTML_OUTPUT} ) - Sphinx_add_target( ${target_base_name}_dirhtml dirhtml ${conf} ${source} ${base_destination}/dirhtml ) - - add_dependencies( ${target_base_name}_dirhtml ${_dependencies} ) - endif() - - if( ${SPHINX_QTHELP_OUTPUT} ) - Sphinx_add_target( ${target_base_name}_qthelp qthelp ${conf} ${source} ${base_destination}/qthelp ) - - add_dependencies( ${target_base_name}_qthelp ${_dependencies} ) - endif() - - if( ${SPHINX_DEVHELP_OUTPUT} ) - Sphinx_add_target( ${target_base_name}_devhelp devhelp ${conf} ${source} ${base_destination}/devhelp ) - - add_dependencies( ${target_base_name}_devhelp ${_dependencies} ) - endif() - - if( ${SPHINX_EPUB_OUTPUT} ) - Sphinx_add_target( ${target_base_name}_epub epub ${conf} ${source} ${base_destination}/epub ) - - add_dependencies( ${target_base_name}_epub ${_dependencies} ) - endif() - - if( ${SPHINX_LATEX_OUTPUT} ) - Sphinx_add_target( ${target_base_name}_latex latex ${conf} ${source} ${base_destination}/latex ) - - add_dependencies( ${target_base_name}_latex ${_dependencies} ) - endif() - - if( ${SPHINX_MAN_OUTPUT} ) - Sphinx_add_target( ${target_base_name}_man man ${conf} ${source} ${base_destination}/man ) - - add_dependencies( ${target_base_name}_man ${_dependencies} ) - endif() - - if( ${SPHINX_TEXT_OUTPUT} ) - Sphinx_add_target( ${target_base_name}_text text ${conf} ${source} ${base_destination}/text ) - - add_dependencies( ${target_base_name}_text ${_dependencies} ) - endif() - - if( ${BUILD_TESTING} ) - sphinx_add_target( ${target_base_name}_linkcheck linkcheck ${conf} ${source} ${base_destination}/linkcheck ) - - add_dependencies( ${target_base_name}_linkcheck ${_dependencies} ) - endif() -endfunction() diff --git a/cmake/configure.cmake b/cmake/configure.cmake index 076e839120d98d801de4374f2f8338ebd918b88f..93d74bb0a8f726ad31685cbfc7831b5441cd5108 100644 --- a/cmake/configure.cmake +++ b/cmake/configure.cmake @@ -20,31 +20,10 @@ if(WITH_DSO) add_definitions(-DPADDLE_USE_DSO) endif(WITH_DSO) -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) -if(NOT WITH_TIMER) - add_definitions(-DPADDLE_DISABLE_TIMER) -endif(NOT WITH_TIMER) - -if(USE_EIGEN_FOR_BLAS) - add_definitions(-DPADDLE_USE_EIGEN_FOR_BLAS) -endif(USE_EIGEN_FOR_BLAS) - -if(EIGEN_USE_THREADS) - add_definitions(-DEIGEN_USE_THREADS) -endif(EIGEN_USE_THREADS) - if(NOT WITH_PROFILER) add_definitions(-DPADDLE_DISABLE_PROFILER) endif(NOT WITH_PROFILER) @@ -78,10 +57,6 @@ if(WIN32) endif(NOT MSVC) endif(WIN32) -if(NOT WITH_GOLANG) - add_definitions(-DPADDLE_WITHOUT_GOLANG) -endif(NOT WITH_GOLANG) - if(WITH_PSLIB) add_definitions(-DPADDLE_WITH_PSLIB) endif() @@ -152,7 +127,12 @@ endif() if (WITH_MKLML AND MKLML_IOMP_LIB) message(STATUS "Enable Intel OpenMP with ${MKLML_IOMP_LIB}") - set(OPENMP_FLAGS "-fopenmp") + if(WIN32) + # openmp not support well for now on windows + set(OPENMP_FLAGS "") + else(WIN32) + set(OPENMP_FLAGS "-fopenmp") + endif(WIN32) set(CMAKE_C_CREATE_SHARED_LIBRARY_FORBIDDEN_FLAGS ${OPENMP_FLAGS}) set(CMAKE_CXX_CREATE_SHARED_LIBRARY_FORBIDDEN_FLAGS ${OPENMP_FLAGS}) set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${OPENMP_FLAGS}") @@ -166,55 +146,6 @@ if(WITH_DISTRIBUTE) add_definitions(-DPADDLE_WITH_DISTRIBUTE) endif() -if(WITH_GOLANG) - # we need to symlink Paddle directory into GOPATH. If we - # don't do it and we have code that depends on Paddle, go - # get ./... will download a new Paddle repo from Github, - # without the changes in our current Paddle repo that we - # want to build. - set(GOPATH "${CMAKE_CURRENT_BINARY_DIR}/go") - file(MAKE_DIRECTORY ${GOPATH}) - set(PADDLE_IN_GOPATH "${GOPATH}/src/github.com/PaddlePaddle/Paddle") - file(MAKE_DIRECTORY "${PADDLE_IN_GOPATH}") - set(PADDLE_GO_PATH "${CMAKE_SOURCE_DIR}/go") - - add_custom_target(go_path) - add_custom_command(TARGET go_path - # Symlink Paddle directory into GOPATH - COMMAND mkdir -p ${PADDLE_IN_GOPATH} - COMMAND rm -rf ${PADDLE_IN_GOPATH} - COMMAND ln -sf ${CMAKE_SOURCE_DIR} ${PADDLE_IN_GOPATH} - # Automatically get all dependencies specified in the source code - # We can't run `go get -d ./...` for every target, because - # multiple `go get` can not run concurrently, but make need to be - # able to run with multiple jobs. - WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR} - ) - - if (GLIDE_INSTALL) - if(EXISTS $ENV{GOPATH}/bin/glide) - set(GLIDE "$ENV{GOPATH}/bin/glide") - else() - message(FATAL_ERROR "no glide executeble found: $ENV{GOPATH}/bin/glide") - endif() - - # this command will only run when the file it depends is missing - # or has changed, or the output is missing. - add_custom_command(OUTPUT ${CMAKE_BINARY_DIR}/glide - COMMAND env GOPATH=${GOPATH} ${GLIDE} install - COMMAND touch ${CMAKE_BINARY_DIR}/glide - DEPENDS ${PADDLE_SOURCE_DIR}/go/glide.lock - WORKING_DIRECTORY "${PADDLE_IN_GOPATH}/go" - ) - - # depends on the custom command which outputs - # ${CMAKE_BINARY_DIR}/glide, the custom command does not need to - # run every time this target is built. - add_custom_target(go_vendor DEPENDS ${CMAKE_BINARY_DIR}/glide go_path) - endif() - -endif(WITH_GOLANG) - if(WITH_GRPC) add_definitions(-DPADDLE_WITH_GRPC) endif(WITH_GRPC) diff --git a/cmake/cuda.cmake b/cmake/cuda.cmake index ea46f6418edf1db70b2a308dd49cf2131cc89d3b..735846db1db04e3884d72ec62d911d9a0efec147 100644 --- a/cmake/cuda.cmake +++ b/cmake/cuda.cmake @@ -168,10 +168,7 @@ elseif (${CUDA_VERSION} LESS 11.0) # CUDA 10.x endif() include_directories(${CUDA_INCLUDE_DIRS}) -list(APPEND EXTERNAL_LIBS ${CUDA_LIBRARIES} ${CUDA_rt_LIBRARY}) if(NOT WITH_DSO) - # TODO(panyx0718): CUPTI only allows DSO? - list(APPEND EXTERNAL_LIBS ${CUDNN_LIBRARY} ${CUPTI_LIBRARY} ${CUDA_CUBLAS_LIBRARIES} ${CUDA_curand_LIBRARY} ${NCCL_LIBRARY}) if(WIN32) set_property(GLOBAL PROPERTY CUDA_MODULES ${CUDNN_LIBRARY} ${CUDA_CUBLAS_LIBRARIES} ${CUDA_curand_LIBRARY}) endif(WIN32) @@ -203,25 +200,26 @@ list(APPEND CUDA_NVCC_FLAGS "-w") list(APPEND CUDA_NVCC_FLAGS "--expt-relaxed-constexpr") if (NOT WIN32) -if(CMAKE_BUILD_TYPE STREQUAL "Debug") - list(APPEND CUDA_NVCC_FLAGS ${CMAKE_CXX_FLAGS_DEBUG}) -elseif(CMAKE_BUILD_TYPE STREQUAL "Release") - list(APPEND CUDA_NVCC_FLAGS ${CMAKE_CXX_FLAGS_RELEASE}) -elseif(CMAKE_BUILD_TYPE STREQUAL "RelWithDebInfo") - list(APPEND CUDA_NVCC_FLAGS ${CMAKE_CXX_FLAGS_RELWITHDEBINFO}) -elseif(CMAKE_BUILD_TYPE STREQUAL "MinSizeRel") - # nvcc 9 does not support -Os. Use Release flags instead - list(APPEND CUDA_NVCC_FLAGS ${CMAKE_CXX_FLAGS_RELEASE}) -endif() + if(CMAKE_BUILD_TYPE STREQUAL "Debug") + list(APPEND CUDA_NVCC_FLAGS ${CMAKE_CXX_FLAGS_DEBUG}) + elseif(CMAKE_BUILD_TYPE STREQUAL "Release") + list(APPEND CUDA_NVCC_FLAGS ${CMAKE_CXX_FLAGS_RELEASE}) + elseif(CMAKE_BUILD_TYPE STREQUAL "RelWithDebInfo") + list(APPEND CUDA_NVCC_FLAGS ${CMAKE_CXX_FLAGS_RELWITHDEBINFO}) + elseif(CMAKE_BUILD_TYPE STREQUAL "MinSizeRel") + # nvcc 9 does not support -Os. Use Release flags instead + list(APPEND CUDA_NVCC_FLAGS ${CMAKE_CXX_FLAGS_RELEASE}) + endif() else(NOT WIN32) -list(APPEND CUDA_NVCC_FLAGS "--compiler-options;/bigobj") -if(CMAKE_BUILD_TYPE STREQUAL "Debug") - list(APPEND CUDA_NVCC_FLAGS "-g -G") - # match the cl's _ITERATOR_DEBUG_LEVEL - list(APPEND CUDA_NVCC_FLAGS "-D_DEBUG") -elseif(CMAKE_BUILD_TYPE STREQUAL "Release") - list(APPEND CUDA_NVCC_FLAGS "-O3 -DNDEBUG") -else() + list(APPEND CUDA_NVCC_FLAGS "-Xcompiler \"/wd 4244 /wd 4267 /wd 4819\"") + list(APPEND CUDA_NVCC_FLAGS "--compiler-options;/bigobj") + if(CMAKE_BUILD_TYPE STREQUAL "Debug") + list(APPEND CUDA_NVCC_FLAGS "-g -G") + # match the cl's _ITERATOR_DEBUG_LEVEL + list(APPEND CUDA_NVCC_FLAGS "-D_DEBUG") + elseif(CMAKE_BUILD_TYPE STREQUAL "Release") + list(APPEND CUDA_NVCC_FLAGS "-O3 -DNDEBUG") + else() message(FATAL "Windows only support Release or Debug build now. Please set visual studio build type to Release/Debug, x64 build.") endif() endif(NOT WIN32) diff --git a/cmake/external/anakin.cmake b/cmake/external/anakin.cmake index 06fc6061bc98eec8c4c71860333f7d3456952aeb..77f4b34537577c7c5782675e7af19c73bc9f8e32 100644 --- a/cmake/external/anakin.cmake +++ b/cmake/external/anakin.cmake @@ -74,5 +74,3 @@ add_dependencies(anakin_shared extern_anakin) add_library(anakin_saber SHARED IMPORTED GLOBAL) set_property(TARGET anakin_saber PROPERTY IMPORTED_LOCATION ${ANAKIN_SABER_LIB}) add_dependencies(anakin_saber extern_anakin) - -list(APPEND external_project_dependencies anakin_shared anakin_saber) diff --git a/cmake/external/any.cmake b/cmake/external/any.cmake deleted file mode 100644 index 85cce80b70a1fcf57015ac7a264e4950616b2717..0000000000000000000000000000000000000000 --- a/cmake/external/any.cmake +++ /dev/null @@ -1,31 +0,0 @@ -INCLUDE(ExternalProject) - -SET(ANY_SOURCE_DIR ${THIRD_PARTY_PATH}/any) - -INCLUDE_DIRECTORIES(${ANY_SOURCE_DIR}/src/extern_lib_any) - -ExternalProject_Add( - extern_lib_any - ${EXTERNAL_PROJECT_LOG_ARGS} - GIT_REPOSITORY "https://github.com/PaddlePaddle/any.git" - GIT_TAG "15595d8324be9e8a9a80d9ae442fdd12bd66df5d" - PREFIX ${ANY_SOURCE_DIR} - UPDATE_COMMAND "" - CONFIGURE_COMMAND "" - BUILD_COMMAND "" - INSTALL_COMMAND "" - TEST_COMMAND "" -) - -if (${CMAKE_VERSION} VERSION_LESS "3.3.0") - set(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/lib_any_dummy.c) - file(WRITE ${dummyfile} "const char * dummy_any = \"${dummyfile}\";") - add_library(lib_any STATIC ${dummyfile}) -else() - add_library(lib_any INTERFACE) -endif() - -add_dependencies(lib_any extern_lib_any) - -add_definitions(-DANY_IMPL_ANY_CAST_MOVEABLE) -LIST(APPEND external_project_dependencies lib_any) diff --git a/cmake/external/boost.cmake b/cmake/external/boost.cmake index 12412a51a0fd1aaa9702bd4547fb935d94012ada..fc204dc9193bb28b654936048dd61a9b461abb2f 100644 --- a/cmake/external/boost.cmake +++ b/cmake/external/boost.cmake @@ -57,5 +57,4 @@ else() endif() add_dependencies(boost ${BOOST_PROJECT}) -list(APPEND external_project_dependencies boost) set(Boost_INCLUDE_DIR ${BOOST_INCLUDE_DIR}) diff --git a/cmake/external/brpc.cmake b/cmake/external/brpc.cmake index 6b50cff7a66a33d9413627bfbc663cca06ba86f3..989d1dbd4cf593e779b94f7bb5eda613f000859c 100644 --- a/cmake/external/brpc.cmake +++ b/cmake/external/brpc.cmake @@ -69,5 +69,3 @@ SET_PROPERTY(TARGET brpc PROPERTY IMPORTED_LOCATION ${BRPC_LIBRARIES}) ADD_DEPENDENCIES(brpc extern_brpc) add_definitions(-DBRPC_WITH_GLOG) - -LIST(APPEND external_project_dependencies brpc) diff --git a/cmake/external/cub.cmake b/cmake/external/cub.cmake index f06728de91e4509be661e56baef641d591928b66..41ad8207743201fbddd1d678fc5122afe68207ae 100644 --- a/cmake/external/cub.cmake +++ b/cmake/external/cub.cmake @@ -31,5 +31,3 @@ else() endif() add_dependencies(cub extern_cub) - -LIST(APPEND external_project_dependencies cub) diff --git a/cmake/external/dlpack.cmake b/cmake/external/dlpack.cmake index 4587475d7902a134eecd54bf8241fb96d175d0ba..63dd16b28e40a0c2d5310bec011c721285049952 100644 --- a/cmake/external/dlpack.cmake +++ b/cmake/external/dlpack.cmake @@ -27,5 +27,3 @@ else() endif() add_dependencies(dlpack extern_dlpack) - -LIST(APPEND external_project_dependencies dlpack) diff --git a/cmake/external/eigen.cmake b/cmake/external/eigen.cmake index 6aef97f21244efd09e22781f703553a19a9e1860..72441160f89d2c188d35fc6b08b5f0b6d746a1ad 100644 --- a/cmake/external/eigen.cmake +++ b/cmake/external/eigen.cmake @@ -52,5 +52,3 @@ else() endif() add_dependencies(eigen3 extern_eigen3) - -LIST(APPEND external_project_dependencies eigen3) diff --git a/cmake/external/gflags.cmake b/cmake/external/gflags.cmake index f3ca74faea3629ddce053c49ef1e629f230fdc49..911920ed6212b87aa25ba9a1faf7696fbcb22587 100644 --- a/cmake/external/gflags.cmake +++ b/cmake/external/gflags.cmake @@ -61,8 +61,6 @@ ADD_LIBRARY(gflags STATIC IMPORTED GLOBAL) SET_PROPERTY(TARGET gflags PROPERTY IMPORTED_LOCATION ${GFLAGS_LIBRARIES}) ADD_DEPENDENCIES(gflags extern_gflags) -LIST(APPEND external_project_dependencies gflags) - # On Windows (including MinGW), the Shlwapi library is used by gflags if available. if (WIN32) include(CheckIncludeFileCXX) diff --git a/cmake/external/glog.cmake b/cmake/external/glog.cmake index 7a6a4523886824a67c82f9ce978de025ddb9c2cd..7fa17ce6b7b106c47c486729d0136748c73176a7 100644 --- a/cmake/external/glog.cmake +++ b/cmake/external/glog.cmake @@ -20,8 +20,10 @@ SET(GLOG_INCLUDE_DIR "${GLOG_INSTALL_DIR}/include" CACHE PATH "glog include dire IF(WIN32) SET(GLOG_LIBRARIES "${GLOG_INSTALL_DIR}/lib/libglog.lib" CACHE FILEPATH "glog library." FORCE) + SET(GLOG_CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /wd4267 /wd4530") ELSE(WIN32) SET(GLOG_LIBRARIES "${GLOG_INSTALL_DIR}/lib/libglog.a" CACHE FILEPATH "glog library." FORCE) + SET(GLOG_CMAKE_CXX_FLAGS ${CMAKE_CXX_FLAGS}) ENDIF(WIN32) INCLUDE_DIRECTORIES(${GLOG_INCLUDE_DIR}) @@ -39,7 +41,7 @@ ExternalProject_Add( UPDATE_COMMAND "" CMAKE_ARGS -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER} -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} - -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS} + -DCMAKE_CXX_FLAGS=${GLOG_CMAKE_CXX_FLAGS} -DCMAKE_CXX_FLAGS_RELEASE=${CMAKE_CXX_FLAGS_RELEASE} -DCMAKE_CXX_FLAGS_DEBUG=${CMAKE_CXX_FLAGS_DEBUG} -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS} @@ -70,5 +72,3 @@ ADD_LIBRARY(glog STATIC IMPORTED GLOBAL) SET_PROPERTY(TARGET glog PROPERTY IMPORTED_LOCATION ${GLOG_LIBRARIES}) ADD_DEPENDENCIES(glog extern_glog gflags) LINK_LIBRARIES(glog gflags) - -LIST(APPEND external_project_dependencies glog) diff --git a/cmake/external/gtest.cmake b/cmake/external/gtest.cmake index 9be625b620287cd4c644ae6908000fd5eec5d5c7..e459526583bd5ee3c89807657f3c30376e57d971 100644 --- a/cmake/external/gtest.cmake +++ b/cmake/external/gtest.cmake @@ -79,5 +79,4 @@ IF(WITH_TESTING OR (WITH_DISTRIBUTE AND NOT WITH_GRPC)) SET_PROPERTY(TARGET gtest_main PROPERTY IMPORTED_LOCATION ${GTEST_MAIN_LIBRARIES}) ADD_DEPENDENCIES(gtest_main extern_gtest) - LIST(APPEND external_project_dependencies gtest gtest_main) ENDIF(WITH_TESTING OR (WITH_DISTRIBUTE AND NOT WITH_GRPC)) diff --git a/cmake/external/leveldb.cmake b/cmake/external/leveldb.cmake index 0df61b01ab64c8b751bdc3893dd5294ad39ab928..ac0febd076e659927a6a882ff487c61ac130437a 100644 --- a/cmake/external/leveldb.cmake +++ b/cmake/external/leveldb.cmake @@ -39,6 +39,3 @@ ADD_DEPENDENCIES(extern_leveldb snappy) ADD_LIBRARY(leveldb STATIC IMPORTED GLOBAL) SET_PROPERTY(TARGET leveldb PROPERTY IMPORTED_LOCATION ${LEVELDB_LIBRARIES}) ADD_DEPENDENCIES(leveldb extern_leveldb) - -LIST(APPEND external_project_dependencies leveldb) - diff --git a/cmake/external/libmct.cmake b/cmake/external/libmct.cmake index 27cff8cfb6315c9b4fa5677ad9062bee73a0e5d8..b944f2945b7874ca76bf1a19e0a363f564851a62 100644 --- a/cmake/external/libmct.cmake +++ b/cmake/external/libmct.cmake @@ -72,7 +72,4 @@ else() add_library(libmct INTERFACE) endif() -#ADD_LIBRARY(libmct SHARED IMPORTED GLOBAL) ADD_DEPENDENCIES(libmct ${LIBMCT_PROJECT}) -LIST(APPEND external_project_dependencies libmct) - diff --git a/cmake/external/libxsmm.cmake b/cmake/external/libxsmm.cmake index 39f49d210a20d49a06c120361ecf0a5d07d1af28..69cdba7c5921f14a87172d95791332e364045b26 100644 --- a/cmake/external/libxsmm.cmake +++ b/cmake/external/libxsmm.cmake @@ -53,5 +53,3 @@ MESSAGE(STATUS "Libxsmm library: ${LIBXSMM_LIBS}") include_directories(${LIBXSMM_INCLUDE_DIR}) ADD_DEFINITIONS(-DPADDLE_WITH_LIBXSMM) ADD_DEPENDENCIES(libxsmm extern_libxsmm) -LIST(APPEND external_project_dependencies libxsmm) - diff --git a/cmake/external/mkldnn.cmake b/cmake/external/mkldnn.cmake index 6a7be73f09a278ab0fd29c7599a7781df3d29413..94a266c50114a94d125467d55a6367a6999e3298 100644 --- a/cmake/external/mkldnn.cmake +++ b/cmake/external/mkldnn.cmake @@ -49,6 +49,8 @@ IF(NOT WIN32) SET(MKLDNN_FLAG "${MKLDNN_FLAG} -Wno-unused-result -Wno-unused-value") SET(MKLDNN_CFLAG "${CMAKE_C_FLAGS} ${MKLDNN_FLAG}") SET(MKLDNN_CXXFLAG "${CMAKE_CXX_FLAGS} ${MKLDNN_FLAG}") +ELSE() + SET(MKLDNN_CXXFLAG "${CMAKE_CXX_FLAGS} /EHsc") ENDIF(NOT WIN32) ExternalProject_Add( @@ -61,7 +63,6 @@ ExternalProject_Add( UPDATE_COMMAND "" CMAKE_ARGS -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER} CMAKE_ARGS -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} - CMAKE_ARGS -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS} CMAKE_ARGS -DCMAKE_CXX_FLAGS_RELEASE=${CMAKE_CXX_FLAGS_RELEASE} CMAKE_ARGS -DCMAKE_CXX_FLAGS_DEBUG=${CMAKE_CXX_FLAGS_DEBUG} CMAKE_ARGS -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS} @@ -88,7 +89,6 @@ SET_PROPERTY(TARGET shared_mkldnn PROPERTY IMPORTED_LOCATION ${MKLDNN_LIB}) ADD_DEPENDENCIES(shared_mkldnn ${MKLDNN_PROJECT}) MESSAGE(STATUS "MKLDNN library: ${MKLDNN_LIB}") add_definitions(-DPADDLE_WITH_MKLDNN) -LIST(APPEND external_project_dependencies shared_mkldnn) # generate a static dummy target to track mkldnn dependencies # for cc_library(xxx SRCS xxx.c DEPS mkldnn) diff --git a/cmake/external/mklml.cmake b/cmake/external/mklml.cmake index 2caff27357687018f29c1efc55b7b82c9dc3ccf6..54826cedb871690a82b535ae3ed102600277c622 100644 --- a/cmake/external/mklml.cmake +++ b/cmake/external/mklml.cmake @@ -73,4 +73,3 @@ INCLUDE_DIRECTORIES(${MKLML_INC_DIR}) ADD_LIBRARY(mklml SHARED IMPORTED GLOBAL) SET_PROPERTY(TARGET mklml PROPERTY IMPORTED_LOCATION ${MKLML_LIB}) ADD_DEPENDENCIES(mklml ${MKLML_PROJECT}) -LIST(APPEND external_project_dependencies mklml) diff --git a/cmake/external/ngraph.cmake b/cmake/external/ngraph.cmake index 14af98b2d74d4aa955aac27727e05567788a84c9..5812a61f0ddc3a3233ff212710fc1b16aa140724 100644 --- a/cmake/external/ngraph.cmake +++ b/cmake/external/ngraph.cmake @@ -77,4 +77,3 @@ add_dependencies(ngraph ${NGRAPH_PROJECT}) target_compile_definitions(ngraph INTERFACE -DPADDLE_WITH_NGRAPH) target_include_directories(ngraph INTERFACE ${NGRAPH_INC_DIR}) target_link_libraries(ngraph INTERFACE ${NGRAPH_SHARED_LIB}) -LIST(APPEND external_project_dependencies ngraph) diff --git a/cmake/external/openblas.cmake b/cmake/external/openblas.cmake index b347a592929836a473ac764c0af1153b07d54258..d8a4a0be6f5aaa3a1a4977bbc68348743f2fa742 100644 --- a/cmake/external/openblas.cmake +++ b/cmake/external/openblas.cmake @@ -11,11 +11,6 @@ # WITHOUT 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(USE_EIGEN_FOR_BLAS) - return() -ENDIF(USE_EIGEN_FOR_BLAS) - INCLUDE(cblas) IF(NOT ${CBLAS_FOUND}) @@ -91,7 +86,6 @@ ENDIF() IF(NOT ${CBLAS_FOUND}) ADD_DEPENDENCIES(cblas extern_openblas) - LIST(APPEND external_project_dependencies cblas) ELSE() IF("${CBLAS_PROVIDER}" STREQUAL "MKLML") ADD_DEPENDENCIES(cblas mklml) diff --git a/cmake/external/protobuf.cmake b/cmake/external/protobuf.cmake index e05b7694ddf1e1652b00f156cde1a2d433c9fc46..bc7fe5454f5883108e43b4ca47920995dc13a1ff 100644 --- a/cmake/external/protobuf.cmake +++ b/cmake/external/protobuf.cmake @@ -129,7 +129,6 @@ macro(PROMPT_PROTOBUF_LIB) ADD_DEPENDENCIES(protoc ${dep}) ENDFOREACH() - LIST(APPEND external_project_dependencies protobuf) RETURN() endmacro() macro(SET_PROTOBUF_VERSION) @@ -231,7 +230,7 @@ FUNCTION(build_protobuf TARGET_NAME BUILD_FOR_HOST) ) ENDFUNCTION() -SET(PROTOBUF_VERSION 3.1) +SET(PROTOBUF_VERSION 3.1.0) IF(NOT PROTOBUF_FOUND) build_protobuf(extern_protobuf FALSE) diff --git a/cmake/external/pslib.cmake b/cmake/external/pslib.cmake index b4ea268e5a48e29d00b0ec8b957b61a42553ec7e..0287e5cf2a835ed65c5fc26ff69d2653d5db217e 100644 --- a/cmake/external/pslib.cmake +++ b/cmake/external/pslib.cmake @@ -70,4 +70,3 @@ ExternalProject_Add( ADD_LIBRARY(pslib SHARED IMPORTED GLOBAL) SET_PROPERTY(TARGET pslib PROPERTY IMPORTED_LOCATION ${PSLIB_LIB}) ADD_DEPENDENCIES(pslib ${PSLIB_PROJECT}) -LIST(APPEND external_project_dependencies pslib) diff --git a/cmake/external/pslib_brpc.cmake b/cmake/external/pslib_brpc.cmake index 8b43f2ef5c999fc351543ba958c7cc4b0856625d..22c8c1b463764b6e107c5f3da25d51b36c2ce59f 100644 --- a/cmake/external/pslib_brpc.cmake +++ b/cmake/external/pslib_brpc.cmake @@ -70,4 +70,3 @@ ExternalProject_Add( ADD_LIBRARY(pslib_brpc SHARED IMPORTED GLOBAL) SET_PROPERTY(TARGET pslib_brpc PROPERTY IMPORTED_LOCATION ${PSLIB_BRPC_LIB}) ADD_DEPENDENCIES(pslib_brpc ${PSLIB_BRPC_PROJECT}) -LIST(APPEND external_project_dependencies pslib_brpc) diff --git a/cmake/external/snappy.cmake b/cmake/external/snappy.cmake index 27d075336d556528ffaf1929c34753494692f0a0..1e01057aa606af78cd722d3619a710cb35817174 100644 --- a/cmake/external/snappy.cmake +++ b/cmake/external/snappy.cmake @@ -20,6 +20,12 @@ set(SNAPPY_SOURCES_DIR ${THIRD_PARTY_PATH}/snappy) set(SNAPPY_INSTALL_DIR ${THIRD_PARTY_PATH}/install/snappy) set(SNAPPY_INCLUDE_DIR "${SNAPPY_INSTALL_DIR}/include" CACHE PATH "snappy include directory." FORCE) +if(WIN32) + SET(SNAPPY_CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /wd4244 /wd4267") +else() + SET(SNAPPY_CMAKE_CXX_FLAGS ${CMAKE_CXX_FLAGS}) +endif() + ExternalProject_Add( extern_snappy GIT_REPOSITORY "https://github.com/google/snappy" @@ -31,7 +37,7 @@ ExternalProject_Add( -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS} -DCMAKE_C_FLAGS_DEBUG=${CMAKE_C_FLAGS_DEBUG} -DCMAKE_C_FLAGS_RELEASE=${CMAKE_C_FLAGS_RELEASE} - -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS} + -DCMAKE_CXX_FLAGS=${SNAPPY_CMAKE_CXX_FLAGS} -DCMAKE_CXX_FLAGS_RELEASE=${CMAKE_CXX_FLAGS_RELEASE} -DCMAKE_CXX_FLAGS_DEBUG=${CMAKE_CXX_FLAGS_DEBUG} -DCMAKE_INSTALL_PREFIX=${SNAPPY_INSTALL_DIR} diff --git a/cmake/external/threadpool.cmake b/cmake/external/threadpool.cmake index 0159815fed81bdff6de3e561af569e9edc75f947..1f56bc7ab056ef0dd95d603ebe3461ef044b2a79 100644 --- a/cmake/external/threadpool.cmake +++ b/cmake/external/threadpool.cmake @@ -26,5 +26,3 @@ else() endif() add_dependencies(simple_threadpool extern_threadpool) - -LIST(APPEND external_project_dependencies simple_threadpool) diff --git a/cmake/external/warpctc.cmake b/cmake/external/warpctc.cmake index 7a25aaf15f2c7f46d99394d82d69bc24e4f5cb2c..6f2af8670f25c00ac0970fe4ae2b0c5b03aa0d9e 100644 --- a/cmake/external/warpctc.cmake +++ b/cmake/external/warpctc.cmake @@ -83,5 +83,3 @@ INCLUDE_DIRECTORIES(${THIRD_PARTY_PATH}/install) # For Paddle code to include wa ADD_LIBRARY(warpctc SHARED IMPORTED GLOBAL) SET_PROPERTY(TARGET warpctc PROPERTY IMPORTED_LOCATION ${WARPCTC_LIBRARIES}) ADD_DEPENDENCIES(warpctc extern_warpctc) - -LIST(APPEND external_project_dependencies warpctc) diff --git a/cmake/external/xbyak.cmake b/cmake/external/xbyak.cmake index 384c2f9328296ce6a8a6293be6cc47e5063dd3c4..1d61154c0d45dea795902d6544deb796693db263 100644 --- a/cmake/external/xbyak.cmake +++ b/cmake/external/xbyak.cmake @@ -55,4 +55,3 @@ else() endif() add_dependencies(xbyak ${XBYAK_PROJECT}) -list(APPEND external_project_dependencies xbyak) diff --git a/cmake/external/xxhash.cmake b/cmake/external/xxhash.cmake index a0f300c2e8bab9e7402f869eed1b4c2d1c579aab..23b1e02108642df561948a6faa3152effb7ca932 100644 --- a/cmake/external/xxhash.cmake +++ b/cmake/external/xxhash.cmake @@ -71,5 +71,3 @@ add_library(xxhash STATIC IMPORTED GLOBAL) set_property(TARGET xxhash PROPERTY IMPORTED_LOCATION ${XXHASH_LIBRARIES}) include_directories(${XXHASH_INCLUDE_DIR}) add_dependencies(xxhash extern_xxhash) - -LIST(APPEND external_project_dependencies xxhash) diff --git a/cmake/external/zlib.cmake b/cmake/external/zlib.cmake index 6c8d79c25e6a2655711fe4450e65600c9a584015..5569fefe992d10ad4820e51e677f40271d0214e7 100644 --- a/cmake/external/zlib.cmake +++ b/cmake/external/zlib.cmake @@ -57,5 +57,3 @@ ENDIF(WIN32) ADD_LIBRARY(zlib STATIC IMPORTED GLOBAL) SET_PROPERTY(TARGET zlib PROPERTY IMPORTED_LOCATION ${ZLIB_LIBRARIES}) ADD_DEPENDENCIES(zlib extern_zlib) - -LIST(APPEND external_project_dependencies zlib) diff --git a/cmake/flags.cmake b/cmake/flags.cmake index 9e6c47f016fe6dfd809c5b2bc88ff59d0a6b2b84..36b533aa4f7815896fb48c33fefad892b8d0d29c 100644 --- a/cmake/flags.cmake +++ b/cmake/flags.cmake @@ -21,7 +21,7 @@ function(CheckCompilerCXX11Flag) if (${CMAKE_CXX_COMPILER_VERSION} VERSION_LESS 3.3) message(FATAL_ERROR "Unsupported Clang version. Clang >= 3.3 required.") endif() - endif() + endif() endif() endfunction() @@ -147,12 +147,7 @@ set(GPU_COMMON_FLAGS -Wno-error=unused-function # Warnings in Numpy Header. -Wno-error=array-bounds # Warnings in Eigen::array ) - -else(NOT WIN32) -set(COMMON_FLAGS - "/w") #disable all warnings. -set(GPU_COMMON_FLAGS - "/w") #disable all warnings +set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -m64") endif(NOT WIN32) if (APPLE) @@ -193,8 +188,7 @@ safe_set_static_flag() CMAKE_CXX_FLAGS_MINSIZEREL CMAKE_CXX_FLAGS_RELWITHDEBINFO CMAKE_C_FLAGS CMAKE_C_FLAGS_DEBUG CMAKE_C_FLAGS_RELEASE CMAKE_C_FLAGS_MINSIZEREL CMAKE_C_FLAGS_RELWITHDEBINFO) - if(${flag_var} MATCHES "/W3") - string(REGEX REPLACE "/W3" "/w" ${flag_var} "${${flag_var}}") - endif(${flag_var} MATCHES "/W3") + string(REGEX REPLACE "(^| )/W[0-9]( |$)" " " ${flag_var} "${${flag_var}}") + set(flag_var "${flag_var} /w") endforeach(flag_var) endif(WIN32) diff --git a/cmake/hip.cmake b/cmake/hip.cmake index 4276bc5b08cd88a52bb5782bca87fc37deabd830..c3a748db502037f926dc241e4c3bc26a83ad3468 100644 --- a/cmake/hip.cmake +++ b/cmake/hip.cmake @@ -11,8 +11,6 @@ include_directories("/opt/rocm/rocrand/include") include_directories("/opt/rocm/rccl/include") include_directories("/opt/rocm/thrust") -list(APPEND EXTERNAL_LIBS "-L/opt/rocm/lib/ -lhip_hcc") - set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -fPIC -DPADDLE_WITH_HIP -std=c++11" ) if(WITH_DSO) @@ -31,22 +29,12 @@ if(WITH_GRPC) set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DPADDLE_WITH_GRPC") endif(WITH_GRPC) -if(NOT WITH_GOLANG) - set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DPADDLE_WITHOUT_GOLANG") -endif(NOT WITH_GOLANG) - if(WITH_MKLDNN) set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DPADDLE_WITH_MKLDNN") endif(WITH_MKLDNN) set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DANY_IMPL_ANY_CAST_MOVEABLE") -if(NOT WITH_RDMA) - set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DPADDLE_DISABLE_RDMA") -endif(NOT WITH_RDMA) - - - if(CMAKE_BUILD_TYPE STREQUAL "Debug") list(APPEND HIP_HCC_FLAGS ${CMAKE_CXX_FLAGS_DEBUG}) elseif(CMAKE_BUILD_TYPE STREQUAL "RelWithDebInfo") diff --git a/cmake/operators.cmake b/cmake/operators.cmake index 59c40a0e5d18b753038f2b9301d1c9494e3901be..c2d04828564e69d7ac965881057f185194aa0475 100644 --- a/cmake/operators.cmake +++ b/cmake/operators.cmake @@ -52,8 +52,8 @@ function(op_library TARGET) endif() if(WITH_MKLDNN) string(REPLACE "_op" "_mkldnn_op" MKLDNN_FILE "${TARGET}") - if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${MKLDNN_FILE}.cc) - list(APPEND mkldnn_cc_srcs ${MKLDNN_FILE}.cc) + if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/mkldnn/${MKLDNN_FILE}.cc) + list(APPEND mkldnn_cc_srcs mkldnn/${MKLDNN_FILE}.cc) endif() endif() else() diff --git a/cmake/rdma.cmake b/cmake/rdma.cmake deleted file mode 100644 index b698f3bdc3ff586a72badee3e0109e29285b457f..0000000000000000000000000000000000000000 --- a/cmake/rdma.cmake +++ /dev/null @@ -1,82 +0,0 @@ -# user should download rdma first from subversion repository - -# execute following instruction to download svn mannally -# svn co https://svn.baidu.com/sys/ip/trunk/rdma/sockrdmav1 rdma/ -# svn co https://svn.baidu.com/sys/ip/trunk/rdma/thirdparty rdma/ -# we use static output in svn repositories to avoid implict bugs from not standard runtime env. - -if(WITH_RDMA) - set(RDMA_ROOT $ENV{RDMA_ROOT} CACHE PATH "Folder contains RDMA sock library and thirdparty library") - - function(generate_rdma_links) - #redirect to current DIR to isolate the pollution from system runtime environment - #it can benifits unified control for different gcc environment. - #e.g, by default gcc48 did not refer /usr/lib64 which could contain low version - #runtime libraries that will crash process while loading it. That redirect trick - #can fix it. - execute_process( - COMMAND mkdir -p librdma - COMMAND ln -s -f /usr/lib64/libibverbs.so.1.0.0 librdma/libibverbs.so.1 - COMMAND ln -s -f /usr/lib64/libibverbs.so.1.0.0 librdma/libibverbs.so - COMMAND ln -s -f /usr/lib64/librdmacm.so.1.0.0 librdma/librdmacm.so.1 - COMMAND ln -s -f /usr/lib64/librdmacm.so.1.0.0 librdma/librdmacm.so - COMMAND ln -s -f /lib64/libnl.so.1.1.4 librdma/libnl.so.1 - COMMAND ln -s -f /lib64/libnl.so.1.1.4 librdma/libnl.so - WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR} - ) - endfunction(generate_rdma_links) - - #check and set headers - find_path(RDMA_INC_SXISOCK sxi_sock.h PATHS ${RDMA_ROOT}/sockrdmav1/output/include) - find_path(RDMA_INC_XIO libxio.h PATHS ${RDMA_ROOT}/thirdparty/output/accelio) - find_path(RDMA_INC_EVENT event2 PATHS ${RDMA_ROOT}/thirdparty/output/libevent) - find_path(RDMA_INC_NUMA numa.h PATHS ${RDMA_ROOT}/thirdparty/output/libnuma) - - #check and set libs - find_library(RDMA_LIB_SXISOCK NAMES sxisock PATHS ${RDMA_ROOT}/sockrdmav1/output) - find_library(RDMA_LIB_XIO NAMES xio PATHS ${RDMA_ROOT}/thirdparty/output/accelio) - find_library(RDMA_LIB_EVENT NAMES event PATHS ${RDMA_ROOT}/thirdparty/output/libevent) - find_library(RDMA_LIB_EVENT_CORE NAMES event_core PATHS ${RDMA_ROOT}/thirdparty/output/libevent) - find_library(RDMA_LIB_EVENT_EXTRA NAMES event_extra PATHS ${RDMA_ROOT}/thirdparty/output/libevent) - find_library(RDMA_LIB_EVENT_PTHREADS NAMES event_pthreads PATHS ${RDMA_ROOT}/thirdparty/output/libevent) - find_library(RDMA_LIB_NUMA NAMES numa PATHS ${RDMA_ROOT}/thirdparty/output/libnuma) - - if( - RDMA_INC_SXISOCK AND - RDMA_INC_XIO AND - RDMA_INC_EVENT AND - RDMA_INC_NUMA AND - RDMA_LIB_SXISOCK AND - RDMA_LIB_XIO AND - RDMA_LIB_EVENT AND - RDMA_LIB_EVENT_CORE AND - RDMA_LIB_EVENT_EXTRA AND - RDMA_LIB_EVENT_PTHREADS AND - RDMA_LIB_NUMA - ) - - set(RDMA_INC_DIR - ${RDMA_INC_SXISOCK} - ${RDMA_INC_XIO} - ${RDMA_INC_EVENT} - ${RDMA_INC_NUMA}) - set(RDMA_LIBS - ${RDMA_LIB_SXISOCK} - ${RDMA_LIB_XIO} - ${RDMA_LIB_EVENT} - ${RDMA_LIB_EVENT_CORE} - ${RDMA_LIB_EVENT_EXTRA} - ${RDMA_LIB_EVENT_PTHREADS} - ${RDMA_LIB_NUMA} - ) - set(RDMA_LD_FLAGS "-L./librdma -libverbs -lrdmacm -Xlinker -rpath ./librdma") - include_directories("${RDMA_INC_DIR}") - else() - #if this module is not called, RDMA_INC_DIR RDMA_LIBS will be null, so top module always refer this variable - message(FATAL_ERROR, "RDMA libraries are not found, try to set RDMA_ROOT or check all related libraries.") - endif() -else(WITH_RDMA) - set(RDMA_LIBS "") - set(RDMA_LD_FLAGS "") - add_definitions(-DPADDLE_DISABLE_RDMA) -endif(WITH_RDMA) diff --git a/cmake/tensorrt.cmake b/cmake/tensorrt.cmake index 3dc7171551bfb7aff8d1e75083c98b00378d247f..891ff222633741f9894c2fdb6c0096a48f8a35e1 100644 --- a/cmake/tensorrt.cmake +++ b/cmake/tensorrt.cmake @@ -33,6 +33,5 @@ if(TENSORRT_FOUND) message(STATUS "Current TensorRT header is ${TENSORRT_INCLUDE_DIR}/NvInfer.h. " "Current TensorRT version is v${TENSORRT_MAJOR_VERSION}. ") include_directories(${TENSORRT_INCLUDE_DIR}) - list(APPEND EXTERNAL_LIBS ${TENSORRT_LIBRARY}) add_definitions(-DPADDLE_WITH_TENSORRT) endif() diff --git a/cmake/version.cmake b/cmake/version.cmake index ac10bdf067be549fe90112aef73fd6e1fbe0ac48..dd57d4ab9969ce530f93ca1694350b1a26b5b543 100644 --- a/cmake/version.cmake +++ b/cmake/version.cmake @@ -31,8 +31,23 @@ while ("${PADDLE_VERSION}" STREQUAL "") set(tmp_version "${GIT_TAG_NAME}~1") endif() else() - # otherwise, we always set PADDLE_VERSION to 0.0.0 to represent latest - set(PADDLE_VERSION "0.0.0") + execute_process( + COMMAND ${GIT_EXECUTABLE} describe --exact-match --tags ${tmp_version} + WORKING_DIRECTORY ${PADDLE_SOURCE_DIR} + OUTPUT_VARIABLE GIT_EXACT_TAG_NAME + RESULT_VARIABLE GIT_EXACT_TAG_RESULT + ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE) + if (NOT ${GIT_EXACT_TAG_NAME}) + # Check if current branch is tag branch + if (${GIT_EXACT_TAG_NAME} MATCHES "v${TAG_VERSION_REGEX}") + string(REPLACE "v" "" PADDLE_VERSION ${GIT_EXACT_TAG_NAME}) + else() + set(PADDLE_VERSION "0.0.0") + endif() + else() + # otherwise, we always set PADDLE_VERSION to 0.0.0 to represent latest + set(PADDLE_VERSION "0.0.0") + endif() endif() else() set(PADDLE_VERSION "0.0.0") diff --git a/paddle/contrib/float16/run_float16_demo.sh b/paddle/contrib/float16/run_float16_demo.sh index 031225a85dabb26e5d9ea06f58909c049e7f0c08..34cb7a12db171915f2bc7df8787dd62cd381de68 100755 --- a/paddle/contrib/float16/run_float16_demo.sh +++ b/paddle/contrib/float16/run_float16_demo.sh @@ -14,9 +14,7 @@ cmake .. -DWITH_AVX=OFF \ -DWITH_MKL=OFF \ -DWITH_GPU=ON \ -DWITH_TESTING=ON \ - -DWITH_TIMER=ON \ -DWITH_PROFILER=ON \ - -DWITH_FLUID_ONLY=ON make -j `nproc` pip install -U "$WHEEL_PATH/$(ls $WHEEL_PATH)" diff --git a/paddle/fluid/API.spec b/paddle/fluid/API.spec index 7a642adac3fb4ecc0e6aa6d092026881ec141ea7..2ad4ac0cac3e1f853d6e71858aa4d7bbc65aa0e1 100644 --- a/paddle/fluid/API.spec +++ b/paddle/fluid/API.spec @@ -8,13 +8,13 @@ paddle.fluid.Program.parse_from_string ArgSpec(args=['binary_str'], varargs=None paddle.fluid.Program.to_string ArgSpec(args=['self', 'throw_on_error', 'with_details'], varargs=None, keywords=None, defaults=(False,)) paddle.fluid.default_startup_program ArgSpec(args=[], varargs=None, keywords=None, defaults=None) paddle.fluid.default_main_program ArgSpec(args=[], varargs=None, keywords=None, defaults=None) -paddle.fluid.program_guard ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None) -paddle.fluid.name_scope ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None) +paddle.fluid.program_guard ArgSpec(args=['main_program', 'startup_program'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.name_scope ArgSpec(args=['prefix'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.Executor.__init__ ArgSpec(args=['self', 'place'], varargs=None, keywords=None, defaults=None) paddle.fluid.Executor.close ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) paddle.fluid.Executor.run ArgSpec(args=['self', 'program', 'feed', 'fetch_list', 'feed_var_name', 'fetch_var_name', 'scope', 'return_numpy', 'use_program_cache'], varargs=None, keywords=None, defaults=(None, None, None, 'feed', 'fetch', None, True, False)) paddle.fluid.global_scope ArgSpec(args=[], varargs=None, keywords=None, defaults=None) -paddle.fluid.scope_guard ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None) +paddle.fluid.scope_guard ArgSpec(args=['scope'], varargs=None, keywords=None, defaults=None) paddle.fluid.DistributeTranspiler.__init__ ArgSpec(args=['self', 'config'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.DistributeTranspiler.get_pserver_program ArgSpec(args=['self', 'endpoint'], varargs=None, keywords=None, defaults=None) paddle.fluid.DistributeTranspiler.get_pserver_programs ArgSpec(args=['self', 'endpoint'], varargs=None, keywords=None, defaults=None) @@ -66,12 +66,12 @@ paddle.fluid.initializer.XavierInitializer.__init__ ArgSpec(args=['self', 'unifo paddle.fluid.initializer.BilinearInitializer.__init__ ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) paddle.fluid.initializer.MSRAInitializer.__init__ ArgSpec(args=['self', 'uniform', 'fan_in', 'seed'], varargs=None, keywords=None, defaults=(True, None, 0)) paddle.fluid.initializer.force_init_on_cpu ArgSpec(args=[], varargs=None, keywords=None, defaults=None) -paddle.fluid.initializer.init_on_cpu ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None) +paddle.fluid.initializer.init_on_cpu ArgSpec(args=[], varargs=None, keywords=None, defaults=None) paddle.fluid.initializer.NumpyArrayInitializer.__init__ ArgSpec(args=['self', 'value'], varargs=None, keywords=None, defaults=None) paddle.fluid.layers.fc ArgSpec(args=['input', 'size', 'num_flatten_dims', 'param_attr', 'bias_attr', 'act', 'is_test', 'name'], varargs=None, keywords=None, defaults=(1, None, None, None, False, None)) paddle.fluid.layers.embedding ArgSpec(args=['input', 'size', 'is_sparse', 'is_distributed', 'padding_idx', 'param_attr', 'dtype'], varargs=None, keywords=None, defaults=(False, False, None, None, 'float32')) paddle.fluid.layers.dynamic_lstm ArgSpec(args=['input', 'size', 'h_0', 'c_0', 'param_attr', 'bias_attr', 'use_peepholes', 'is_reverse', 'gate_activation', 'cell_activation', 'candidate_activation', 'dtype', 'name'], varargs=None, keywords=None, defaults=(None, None, None, None, True, False, 'sigmoid', 'tanh', 'tanh', 'float32', None)) -paddle.fluid.layers.dynamic_lstmp ArgSpec(args=['input', 'size', 'proj_size', 'param_attr', 'bias_attr', 'use_peepholes', 'is_reverse', 'gate_activation', 'cell_activation', 'candidate_activation', 'proj_activation', 'dtype', 'name'], varargs=None, keywords=None, defaults=(None, None, True, False, 'sigmoid', 'tanh', 'tanh', 'tanh', 'float32', None)) +paddle.fluid.layers.dynamic_lstmp ArgSpec(args=['input', 'size', 'proj_size', 'param_attr', 'bias_attr', 'use_peepholes', 'is_reverse', 'gate_activation', 'cell_activation', 'candidate_activation', 'proj_activation', 'dtype', 'name', 'h_0', 'c_0', 'cell_clip', 'proj_clip'], varargs=None, keywords=None, defaults=(None, None, True, False, 'sigmoid', 'tanh', 'tanh', 'tanh', 'float32', None, None, None, None, None)) paddle.fluid.layers.dynamic_gru ArgSpec(args=['input', 'size', 'param_attr', 'bias_attr', 'is_reverse', 'gate_activation', 'candidate_activation', 'h_0', 'origin_mode'], varargs=None, keywords=None, defaults=(None, None, False, 'sigmoid', 'tanh', None, False)) paddle.fluid.layers.gru_unit ArgSpec(args=['input', 'hidden', 'size', 'param_attr', 'bias_attr', 'activation', 'gate_activation', 'origin_mode'], varargs=None, keywords=None, defaults=(None, None, 'tanh', 'sigmoid', False)) paddle.fluid.layers.linear_chain_crf ArgSpec(args=['input', 'label', 'param_attr'], varargs=None, keywords=None, defaults=(None,)) @@ -121,8 +121,9 @@ paddle.fluid.layers.sequence_reshape ArgSpec(args=['input', 'new_dim'], varargs= paddle.fluid.layers.transpose ArgSpec(args=['x', 'perm', 'name'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.im2sequence ArgSpec(args=['input', 'filter_size', 'stride', 'padding', 'input_image_size', 'out_stride', 'name'], varargs=None, keywords=None, defaults=(1, 1, 0, None, 1, None)) paddle.fluid.layers.nce ArgSpec(args=['input', 'label', 'num_total_classes', 'sample_weight', 'param_attr', 'bias_attr', 'num_neg_samples', 'name', 'sampler', 'custom_dist', 'seed', 'is_sparse'], varargs=None, keywords=None, defaults=(None, None, None, None, None, 'uniform', None, 0, False)) +paddle.fluid.layers.sampled_softmax_with_cross_entropy ArgSpec(args=['logits', 'label', 'num_samples', 'num_true', 'remove_accidental_hits', 'use_customized_samples', 'customized_samples', 'customized_probabilities', 'seed'], varargs=None, keywords=None, defaults=(1, True, False, None, None, 0)) paddle.fluid.layers.hsigmoid ArgSpec(args=['input', 'label', 'num_classes', 'param_attr', 'bias_attr', 'name', 'path_table', 'path_code', 'is_custom', 'is_sparse'], varargs=None, keywords=None, defaults=(None, None, None, None, None, False, False)) -paddle.fluid.layers.beam_search ArgSpec(args=['pre_ids', 'pre_scores', 'ids', 'scores', 'beam_size', 'end_id', 'level', 'is_accumulated', 'name'], varargs=None, keywords=None, defaults=(0, True, None)) +paddle.fluid.layers.beam_search ArgSpec(args=['pre_ids', 'pre_scores', 'ids', 'scores', 'beam_size', 'end_id', 'level', 'is_accumulated', 'name', 'return_parent_idx'], varargs=None, keywords=None, defaults=(0, True, None, False)) paddle.fluid.layers.row_conv ArgSpec(args=['input', 'future_context_size', 'param_attr', 'act'], varargs=None, keywords=None, defaults=(None, None)) paddle.fluid.layers.multiplex ArgSpec(args=['inputs', 'index'], varargs=None, keywords=None, defaults=None) paddle.fluid.layers.layer_norm ArgSpec(args=['input', 'scale', 'shift', 'begin_norm_axis', 'epsilon', 'param_attr', 'bias_attr', 'act', 'name'], varargs=None, keywords=None, defaults=(True, True, 1, 1e-05, None, None, None, None)) @@ -142,10 +143,10 @@ paddle.fluid.layers.label_smooth ArgSpec(args=['label', 'prior_dist', 'epsilon', paddle.fluid.layers.roi_pool ArgSpec(args=['input', 'rois', 'pooled_height', 'pooled_width', 'spatial_scale'], varargs=None, keywords=None, defaults=(1, 1, 1.0)) paddle.fluid.layers.roi_align ArgSpec(args=['input', 'rois', 'pooled_height', 'pooled_width', 'spatial_scale', 'sampling_ratio', 'name'], varargs=None, keywords=None, defaults=(1, 1, 1.0, -1, None)) paddle.fluid.layers.dice_loss ArgSpec(args=['input', 'label', 'epsilon'], varargs=None, keywords=None, defaults=(1e-05,)) -paddle.fluid.layers.image_resize ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'resample', 'actual_shape'], varargs=None, keywords=None, defaults=(None, None, None, 'BILINEAR', None)) +paddle.fluid.layers.image_resize ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'resample', 'actual_shape', 'align_corners', 'align_mode'], varargs=None, keywords=None, defaults=(None, None, None, 'BILINEAR', None, True, 1)) paddle.fluid.layers.image_resize_short ArgSpec(args=['input', 'out_short_len', 'resample'], varargs=None, keywords=None, defaults=('BILINEAR',)) -paddle.fluid.layers.resize_bilinear ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'actual_shape'], varargs=None, keywords=None, defaults=(None, None, None, None)) -paddle.fluid.layers.resize_nearest ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'actual_shape'], varargs=None, keywords=None, defaults=(None, None, None, None)) +paddle.fluid.layers.resize_bilinear ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'actual_shape', 'align_corners', 'align_mode'], varargs=None, keywords=None, defaults=(None, None, None, None, True, 1)) +paddle.fluid.layers.resize_nearest ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'actual_shape', 'align_corners'], varargs=None, keywords=None, defaults=(None, None, None, None, True)) paddle.fluid.layers.gather ArgSpec(args=['input', 'index'], varargs=None, keywords=None, defaults=None) paddle.fluid.layers.scatter ArgSpec(args=['input', 'index', 'updates', 'name'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.sequence_scatter ArgSpec(args=['input', 'index', 'updates', 'name'], varargs=None, keywords=None, defaults=(None,)) @@ -229,7 +230,7 @@ paddle.fluid.layers.random_data_generator ArgSpec(args=['low', 'high', 'shapes', paddle.fluid.layers.py_reader ArgSpec(args=['capacity', 'shapes', 'dtypes', 'lod_levels', 'name', 'use_double_buffer'], varargs=None, keywords=None, defaults=(None, None, True)) paddle.fluid.layers.create_py_reader_by_data ArgSpec(args=['capacity', 'feed_list', 'name', 'use_double_buffer'], varargs=None, keywords=None, defaults=(None, True)) paddle.fluid.layers.Preprocessor.__init__ ArgSpec(args=['self', 'reader', 'name'], varargs=None, keywords=None, defaults=(None,)) -paddle.fluid.layers.Preprocessor.block ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None) +paddle.fluid.layers.Preprocessor.block ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) paddle.fluid.layers.Preprocessor.inputs ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) paddle.fluid.layers.Preprocessor.outputs ArgSpec(args=['self'], varargs='outs', keywords=None, defaults=None) paddle.fluid.layers.load ArgSpec(args=['out', 'file_path', 'load_as_fp16'], varargs=None, keywords=None, defaults=(None,)) @@ -261,7 +262,7 @@ paddle.fluid.layers.increment ArgSpec(args=['x', 'value', 'in_place'], varargs=N paddle.fluid.layers.array_write ArgSpec(args=['x', 'i', 'array'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.create_array ArgSpec(args=['dtype'], varargs=None, keywords=None, defaults=None) paddle.fluid.layers.less_than ArgSpec(args=['x', 'y', 'force_cpu', 'cond'], varargs=None, keywords='ignored', defaults=(None, None)) -paddle.fluid.layers.equal ArgSpec(args=['x', 'y', 'cond'], varargs=None, keywords='ignored', defaults=(None,)) +paddle.fluid.layers.equal ArgSpec(args=['x', 'y', 'cond'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.array_read ArgSpec(args=['array', 'i'], varargs=None, keywords=None, defaults=None) paddle.fluid.layers.array_length ArgSpec(args=['array'], varargs=None, keywords=None, defaults=None) paddle.fluid.layers.IfElse.__init__ ArgSpec(args=['self', 'cond', 'name'], varargs=None, keywords=None, defaults=(None,)) @@ -270,7 +271,7 @@ paddle.fluid.layers.IfElse.input ArgSpec(args=['self', 'x'], varargs=None, keywo paddle.fluid.layers.IfElse.output ArgSpec(args=['self'], varargs='outs', keywords=None, defaults=None) paddle.fluid.layers.IfElse.true_block ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) paddle.fluid.layers.DynamicRNN.__init__ ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=(None,)) -paddle.fluid.layers.DynamicRNN.block ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None) +paddle.fluid.layers.DynamicRNN.block ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) paddle.fluid.layers.DynamicRNN.memory ArgSpec(args=['self', 'init', 'shape', 'value', 'need_reorder', 'dtype'], varargs=None, keywords=None, defaults=(None, None, 0.0, False, 'float32')) paddle.fluid.layers.DynamicRNN.output ArgSpec(args=['self'], varargs='outputs', keywords=None, defaults=None) paddle.fluid.layers.DynamicRNN.static_input ArgSpec(args=['self', 'x'], varargs=None, keywords=None, defaults=None) @@ -322,9 +323,11 @@ paddle.fluid.layers.generate_proposal_labels ArgSpec(args=['rpn_rois', 'gt_class paddle.fluid.layers.generate_proposals ArgSpec(args=['scores', 'bbox_deltas', 'im_info', 'anchors', 'variances', 'pre_nms_top_n', 'post_nms_top_n', 'nms_thresh', 'min_size', 'eta', 'name'], varargs=None, keywords=None, defaults=(6000, 1000, 0.5, 0.1, 1.0, None)) paddle.fluid.layers.generate_mask_labels ArgSpec(args=['im_info', 'gt_classes', 'is_crowd', 'gt_segms', 'rois', 'labels_int32', 'num_classes', 'resolution'], varargs=None, keywords=None, defaults=None) paddle.fluid.layers.iou_similarity ArgSpec(args=['x', 'y', 'name'], varargs=None, keywords=None, defaults=(None,)) -paddle.fluid.layers.box_coder ArgSpec(args=['prior_box', 'prior_box_var', 'target_box', 'code_type', 'box_normalized', 'name'], varargs=None, keywords=None, defaults=('encode_center_size', True, None)) +paddle.fluid.layers.box_coder ArgSpec(args=['prior_box', 'prior_box_var', 'target_box', 'code_type', 'box_normalized', 'name', 'axis'], varargs=None, keywords=None, defaults=('encode_center_size', True, None, 0)) paddle.fluid.layers.polygon_box_transform ArgSpec(args=['input', 'name'], varargs=None, keywords=None, defaults=(None,)) -paddle.fluid.layers.yolov3_loss ArgSpec(args=['x', 'gtbox', 'gtlabel', 'anchors', 'class_num', 'ignore_thresh', 'loss_weight_xy', 'loss_weight_wh', 'loss_weight_conf_target', 'loss_weight_conf_notarget', 'loss_weight_class', 'name'], varargs=None, keywords=None, defaults=(None, None, None, None, None, None)) +paddle.fluid.layers.yolov3_loss ArgSpec(args=['x', 'gtbox', 'gtlabel', 'anchors', 'anchor_mask', 'class_num', 'ignore_thresh', 'downsample_ratio', 'name'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.layers.box_clip ArgSpec(args=['input', 'im_info', 'name'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.layers.multiclass_nms ArgSpec(args=['bboxes', 'scores', 'score_threshold', 'nms_top_k', 'keep_top_k', 'nms_threshold', 'normalized', 'nms_eta', 'background_label', 'name'], varargs=None, keywords=None, defaults=(0.3, True, 1.0, 0, None)) paddle.fluid.layers.accuracy ArgSpec(args=['input', 'label', 'k', 'correct', 'total'], varargs=None, keywords=None, defaults=(1, None, None)) paddle.fluid.layers.auc ArgSpec(args=['input', 'label', 'curve', 'num_thresholds', 'topk', 'slide_steps'], varargs=None, keywords=None, defaults=('ROC', 4095, 1, 1)) paddle.fluid.layers.exponential_decay ArgSpec(args=['learning_rate', 'decay_steps', 'decay_rate', 'staircase'], varargs=None, keywords=None, defaults=(False,)) @@ -344,12 +347,12 @@ paddle.fluid.contrib.StateCell.set_state ArgSpec(args=['self', 'state_name', 'st paddle.fluid.contrib.StateCell.state_updater ArgSpec(args=['self', 'updater'], varargs=None, keywords=None, defaults=None) paddle.fluid.contrib.StateCell.update_states ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) paddle.fluid.contrib.TrainingDecoder.__init__ ArgSpec(args=['self', 'state_cell', 'name'], varargs=None, keywords=None, defaults=(None,)) -paddle.fluid.contrib.TrainingDecoder.block ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None) +paddle.fluid.contrib.TrainingDecoder.block ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) paddle.fluid.contrib.TrainingDecoder.output ArgSpec(args=['self'], varargs='outputs', keywords=None, defaults=None) paddle.fluid.contrib.TrainingDecoder.static_input ArgSpec(args=['self', 'x'], varargs=None, keywords=None, defaults=None) paddle.fluid.contrib.TrainingDecoder.step_input ArgSpec(args=['self', 'x'], varargs=None, keywords=None, defaults=None) paddle.fluid.contrib.BeamSearchDecoder.__init__ ArgSpec(args=['self', 'state_cell', 'init_ids', 'init_scores', 'target_dict_dim', 'word_dim', 'input_var_dict', 'topk_size', 'sparse_emb', 'max_len', 'beam_size', 'end_id', 'name'], varargs=None, keywords=None, defaults=({}, 50, True, 100, 1, 1, None)) -paddle.fluid.contrib.BeamSearchDecoder.block ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None) +paddle.fluid.contrib.BeamSearchDecoder.block ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) paddle.fluid.contrib.BeamSearchDecoder.decode ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) paddle.fluid.contrib.BeamSearchDecoder.early_stop ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) paddle.fluid.contrib.BeamSearchDecoder.read_array ArgSpec(args=['self', 'init', 'is_ids', 'is_scores'], varargs=None, keywords=None, defaults=(False, False)) @@ -360,6 +363,9 @@ paddle.fluid.contrib.QuantizeTranspiler.__init__ ArgSpec(args=['self', 'weight_b paddle.fluid.contrib.QuantizeTranspiler.convert_to_int8 ArgSpec(args=['self', 'program', 'place', 'scope'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.contrib.QuantizeTranspiler.freeze_program ArgSpec(args=['self', 'program', 'place', 'fuse_bn', 'scope'], varargs=None, keywords=None, defaults=(False, None)) paddle.fluid.contrib.QuantizeTranspiler.training_transpile ArgSpec(args=['self', 'program', 'startup_program'], varargs=None, keywords=None, defaults=(None, None)) +paddle.fluid.contrib.Calibrator.__init__ ArgSpec(args=['self'], varargs='args', keywords='kwargs', defaults=None) +paddle.fluid.contrib.Calibrator.sample_data ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) +paddle.fluid.contrib.Calibrator.save_int8_model ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) paddle.fluid.contrib.reader.ctr_reader.ctr_reader ArgSpec(args=['feed_dict', 'file_type', 'file_format', 'dense_slot_index', 'sparse_slot_index', 'capacity', 'thread_num', 'batch_size', 'file_list', 'slots', 'name'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.contrib.build_compressor ArgSpec(args=['place', 'data_reader', 'data_feeder', 'scope', 'metrics', 'epoch', 'config'], varargs=None, keywords=None, defaults=(None, None, None, None, None, None, None)) paddle.fluid.contrib.CompressPass.__init__ ArgSpec(args=['self', 'place', 'data_reader', 'data_feeder', 'scope', 'metrics', 'epoch', 'program_exe'], varargs=None, keywords=None, defaults=(None, None, None, None, None, None, None)) @@ -422,7 +428,7 @@ paddle.fluid.optimizer.MomentumOptimizer.__init__ ArgSpec(args=['self', 'learnin paddle.fluid.optimizer.MomentumOptimizer.apply_gradients ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None) paddle.fluid.optimizer.MomentumOptimizer.backward ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)) paddle.fluid.optimizer.MomentumOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)) -paddle.fluid.optimizer.AdagradOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'epsilon', 'regularization', 'name'], varargs=None, keywords=None, defaults=(1e-06, None, None)) +paddle.fluid.optimizer.AdagradOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'epsilon', 'regularization', 'name', 'initial_accumulator_value'], varargs=None, keywords=None, defaults=(1e-06, None, None, 0.0)) paddle.fluid.optimizer.AdagradOptimizer.apply_gradients ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None) paddle.fluid.optimizer.AdagradOptimizer.backward ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)) paddle.fluid.optimizer.AdagradOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)) @@ -451,7 +457,7 @@ paddle.fluid.optimizer.AdadeltaOptimizer.apply_gradients ArgSpec(args=['self', ' paddle.fluid.optimizer.AdadeltaOptimizer.backward ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)) paddle.fluid.optimizer.AdadeltaOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)) paddle.fluid.optimizer.ModelAverage.__init__ ArgSpec(args=['self', 'average_window_rate', 'min_average_window', 'max_average_window', 'regularization', 'name'], varargs=None, keywords=None, defaults=(10000, 10000, None, None)) -paddle.fluid.optimizer.ModelAverage.apply ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None) +paddle.fluid.optimizer.ModelAverage.apply ArgSpec(args=['self', 'executor', 'need_restore'], varargs=None, keywords=None, defaults=(True,)) paddle.fluid.optimizer.ModelAverage.apply_gradients ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None) paddle.fluid.optimizer.ModelAverage.backward ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)) paddle.fluid.optimizer.ModelAverage.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)) @@ -467,12 +473,12 @@ paddle.fluid.LoDTensor.__init__ 1. __init__(self: paddle.fluid.core.LoDTensor, a paddle.fluid.LoDTensor.has_valid_recursive_sequence_lengths has_valid_recursive_sequence_lengths(self: paddle.fluid.core.LoDTensor) -> bool paddle.fluid.LoDTensor.lod lod(self: paddle.fluid.core.LoDTensor) -> List[List[int]] paddle.fluid.LoDTensor.recursive_sequence_lengths recursive_sequence_lengths(self: paddle.fluid.core.LoDTensor) -> List[List[int]] -paddle.fluid.LoDTensor.set 1. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float32], arg1: paddle::platform::CPUPlace) -> None 2. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int32], arg1: paddle::platform::CPUPlace) -> None 3. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float64], arg1: paddle::platform::CPUPlace) -> None 4. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int64], arg1: paddle::platform::CPUPlace) -> None 5. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[bool], arg1: paddle::platform::CPUPlace) -> None 6. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint16], arg1: paddle::platform::CPUPlace) -> None 7. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint8], arg1: paddle::platform::CPUPlace) -> None 8. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int8], arg1: paddle::platform::CPUPlace) -> None -paddle.fluid.LoDTensor.set_lod set_lod(self: paddle.fluid.core.LoDTensor, arg0: List[List[int]]) -> None -paddle.fluid.LoDTensor.set_recursive_sequence_lengths set_recursive_sequence_lengths(self: paddle.fluid.core.LoDTensor, arg0: List[List[int]]) -> None +paddle.fluid.LoDTensor.set 1. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float32], arg1: paddle::platform::CPUPlace) -> None 2. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int32], arg1: paddle::platform::CPUPlace) -> None 3. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float64], arg1: paddle::platform::CPUPlace) -> None 4. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int64], arg1: paddle::platform::CPUPlace) -> None 5. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[bool], arg1: paddle::platform::CPUPlace) -> None 6. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint16], arg1: paddle::platform::CPUPlace) -> None 7. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint8], arg1: paddle::platform::CPUPlace) -> None 8. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int8], arg1: paddle::platform::CPUPlace) -> None 9. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float32], arg1: paddle::platform::CUDAPlace) -> None 10. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int32], arg1: paddle::platform::CUDAPlace) -> None 11. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float64], arg1: paddle::platform::CUDAPlace) -> None 12. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int64], arg1: paddle::platform::CUDAPlace) -> None 13. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[bool], arg1: paddle::platform::CUDAPlace) -> None 14. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint16], arg1: paddle::platform::CUDAPlace) -> None 15. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint8], arg1: paddle::platform::CUDAPlace) -> None 16. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int8], arg1: paddle::platform::CUDAPlace) -> None 17. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float32], arg1: paddle::platform::CUDAPinnedPlace) -> None 18. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int32], arg1: paddle::platform::CUDAPinnedPlace) -> None 19. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float64], arg1: paddle::platform::CUDAPinnedPlace) -> None 20. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int64], arg1: paddle::platform::CUDAPinnedPlace) -> None 21. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[bool], arg1: paddle::platform::CUDAPinnedPlace) -> None 22. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint16], arg1: paddle::platform::CUDAPinnedPlace) -> None 23. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint8], arg1: paddle::platform::CUDAPinnedPlace) -> None 24. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int8], arg1: paddle::platform::CUDAPinnedPlace) -> None +paddle.fluid.LoDTensor.set_lod set_lod(self: paddle.fluid.core.LoDTensor, lod: List[List[int]]) -> None +paddle.fluid.LoDTensor.set_recursive_sequence_lengths set_recursive_sequence_lengths(self: paddle.fluid.core.LoDTensor, recursive_sequence_lengths: List[List[int]]) -> None paddle.fluid.LoDTensor.shape shape(self: paddle.fluid.core.Tensor) -> List[int] paddle.fluid.LoDTensorArray.__init__ __init__(self: paddle.fluid.core.LoDTensorArray) -> None -paddle.fluid.LoDTensorArray.append append(self: paddle.fluid.core.LoDTensorArray, arg0: paddle.fluid.core.LoDTensor) -> None +paddle.fluid.LoDTensorArray.append append(self: paddle.fluid.core.LoDTensorArray, tensor: paddle.fluid.core.LoDTensor) -> None paddle.fluid.CPUPlace.__init__ __init__(self: paddle.fluid.core.CPUPlace) -> None paddle.fluid.CUDAPlace.__init__ __init__(self: paddle.fluid.core.CUDAPlace, arg0: int) -> None paddle.fluid.CUDAPinnedPlace.__init__ __init__(self: paddle.fluid.core.CUDAPinnedPlace) -> None @@ -486,14 +492,14 @@ paddle.fluid.clip.ErrorClipByValue.__init__ ArgSpec(args=['self', 'max', 'min'], paddle.fluid.clip.GradientClipByValue.__init__ ArgSpec(args=['self', 'max', 'min'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.clip.GradientClipByNorm.__init__ ArgSpec(args=['self', 'clip_norm'], varargs=None, keywords=None, defaults=None) paddle.fluid.clip.GradientClipByGlobalNorm.__init__ ArgSpec(args=['self', 'clip_norm', 'group_name'], varargs=None, keywords=None, defaults=('default_group',)) -paddle.fluid.profiler.cuda_profiler ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None) +paddle.fluid.profiler.cuda_profiler ArgSpec(args=['output_file', 'output_mode', 'config'], varargs=None, keywords=None, defaults=(None, None)) paddle.fluid.profiler.reset_profiler ArgSpec(args=[], varargs=None, keywords=None, defaults=None) -paddle.fluid.profiler.profiler ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None) +paddle.fluid.profiler.profiler ArgSpec(args=['state', 'sorted_key', 'profile_path'], varargs=None, keywords=None, defaults=(None, '/tmp/profile')) paddle.fluid.profiler.start_profiler ArgSpec(args=['state'], varargs=None, keywords=None, defaults=None) paddle.fluid.profiler.stop_profiler ArgSpec(args=['sorted_key', 'profile_path'], varargs=None, keywords=None, defaults=(None, '/tmp/profile')) paddle.fluid.unique_name.generate ArgSpec(args=['key'], varargs=None, keywords=None, defaults=None) paddle.fluid.unique_name.switch ArgSpec(args=['new_generator'], varargs=None, keywords=None, defaults=(None,)) -paddle.fluid.unique_name.guard ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None) +paddle.fluid.unique_name.guard ArgSpec(args=['new_generator'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.recordio_writer.convert_reader_to_recordio_file ArgSpec(args=['filename', 'reader_creator', 'feeder', 'compressor', 'max_num_records', 'feed_order'], varargs=None, keywords=None, defaults=(Compressor.Snappy, 1000, None)) paddle.fluid.recordio_writer.convert_reader_to_recordio_files ArgSpec(args=['filename', 'batch_per_file', 'reader_creator', 'feeder', 'compressor', 'max_num_records', 'feed_order'], varargs=None, keywords=None, defaults=(Compressor.Snappy, 1000, None)) paddle.fluid.Scope Scope() -> paddle.fluid.core._Scope diff --git a/paddle/fluid/framework/CMakeLists.txt b/paddle/fluid/framework/CMakeLists.txt index 2ba2437de66f31549a87f20360dbb97b48ea6fbe..7ddf1ab44fe096739f4d241994e5cb686970a7c5 100644 --- a/paddle/fluid/framework/CMakeLists.txt +++ b/paddle/fluid/framework/CMakeLists.txt @@ -1,4 +1,3 @@ - #windows treat symbolic file as a real file, which is different with unix #We create a hidden file and compile it instead of origin source file. function(windows_symbolic TARGET) @@ -129,7 +128,7 @@ cc_test(version_test SRCS version_test.cc DEPS version) cc_library(proto_desc SRCS var_desc.cc op_desc.cc block_desc.cc program_desc.cc DEPS shape_inference op_info operator glog version) -cc_library(op_registry SRCS op_registry.cc DEPS op_proto_maker op_info operator glog proto_desc) +cc_library(op_registry SRCS op_registry.cc DEPS op_proto_maker op_info operator glog proto_desc memory_optimize_helper) nv_test(op_registry_test SRCS op_registry_test.cc DEPS op_registry) py_proto_compile(framework_py_proto SRCS framework.proto data_feed.proto) @@ -159,18 +158,19 @@ cc_library(variable_helper SRCS variable_helper.cc DEPS lod_tensor) cc_library(naive_executor SRCS naive_executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass variable_helper) -if(WITH_DISTRIBUTE) - cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog - lod_rank_table feed_fetch_method sendrecvop_rpc ${GLOB_DISTRIBUTE_DEPS} graph_to_program_pass variable_helper) +if(WITH_NGRAPH) + set(NGRAPH_EXE_DEPS ngraph_engine) +else() + set(NGRAPH_EXE_DEPS) +endif() - set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor") - set_source_files_properties(executor.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) +if(WITH_DISTRIBUTE) + cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog + lod_rank_table feed_fetch_method sendrecvop_rpc ${GLOB_DISTRIBUTE_DEPS} graph_to_program_pass variable_helper ${NGRAPH_EXE_DEPS}) + set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor") + set_source_files_properties(executor.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) else() - if (WITH_NGRAPH) - cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass variable_helper ngraph_engine) - else () - cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass variable_helper) - endif() + cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass variable_helper ${NGRAPH_EXE_DEPS}) cc_test(test_naive_executor SRCS naive_executor_test.cc DEPS naive_executor elementwise_add_op) endif() @@ -193,6 +193,7 @@ 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_test(var_type_inference_test SRCS var_type_inference_test.cc DEPS op_registry proto_desc) +cc_test(inplace_op_inference_test SRCS inplace_op_inference_test.cc DEPS op_registry proto_desc op_info memory_optimize_helper) cc_library(selected_rows SRCS selected_rows.cc DEPS tensor) cc_test(selected_rows_test SRCS selected_rows_test.cc DEPS selected_rows) @@ -207,3 +208,24 @@ endif (NOT WIN32) cc_library(dlpack_tensor SRCS dlpack_tensor.cc DEPS tensor dlpack) cc_test(dlpack_tensor_test SRCS dlpack_tensor_test.cc DEPS dlpack_tensor glog) + +# Get the current working branch +execute_process( + COMMAND git rev-parse --abbrev-ref HEAD + WORKING_DIRECTORY ${CMAKE_SOURCE_DIR} + OUTPUT_VARIABLE PADDLE_BRANCH + OUTPUT_STRIP_TRAILING_WHITESPACE +) + +# Get the latest abbreviated commit hash of the working branch +execute_process( + COMMAND git log -1 --format=%h + WORKING_DIRECTORY ${CMAKE_SOURCE_DIR} + OUTPUT_VARIABLE PADDLE_COMMIT + OUTPUT_STRIP_TRAILING_WHITESPACE +) + +message(STATUS "commit: ${PADDLE_COMMIT}") +message(STATUS "branch: ${PADDLE_BRANCH}") + +configure_file(commit.h.in commit.h) diff --git a/paddle/fluid/framework/async_executor.cc b/paddle/fluid/framework/async_executor.cc index 1d9678a1ba1409e5c18d3e25b3aa13dfbbf76908..60708bf609d6f8b327d46fe585cbbcf07a62eece 100644 --- a/paddle/fluid/framework/async_executor.cc +++ b/paddle/fluid/framework/async_executor.cc @@ -244,6 +244,7 @@ void AsyncExecutor::RunFromFile(const ProgramDesc& main_program, auto& block = main_program.Block(0); for (auto var_name : fetch_var_names) { auto var_desc = block.FindVar(var_name); + PADDLE_ENFORCE_NOT_NULL(var_desc, "%s is not found.", var_name); auto shapes = var_desc->GetShape(); PADDLE_ENFORCE(shapes[shapes.size() - 1] == 1, "var %s: Fetched var has wrong shape, " diff --git a/paddle/fluid/framework/commit.h.in b/paddle/fluid/framework/commit.h.in new file mode 100644 index 0000000000000000000000000000000000000000..3a33ece624443a99083ae29abb70254a5ac40a3d --- /dev/null +++ b/paddle/fluid/framework/commit.h.in @@ -0,0 +1,21 @@ +#pragma once + +#include + +namespace paddle { +namespace framework { + +static std::string paddle_commit() { + return "@PADDLE_COMMIT@"; +} + +static std::string paddle_compile_branch() { + return "@PADDLE_BRANCH@"; +} + +static std::string paddle_version() { + return "@PADDLE_VERSION@"; +} + +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/CMakeLists.txt b/paddle/fluid/framework/details/CMakeLists.txt index d5966ad5a97a97ec40c8a01d2d2c8ed5d7f90421..dc308fd2592bb158f46f6eac9dd0df25787559fe 100644 --- a/paddle/fluid/framework/details/CMakeLists.txt +++ b/paddle/fluid/framework/details/CMakeLists.txt @@ -50,10 +50,15 @@ cc_library(data_balance_op_handle SRCS data_balance_op_handle.cc DEPS op_handle_ cc_library(gather_op_handle SRCS gather_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor) cc_library(fuse_vars_op_handle SRCS fuse_vars_op_handle.cc DEPS op_handle_base scope) -cc_library(memory_optimize_pass SRCS analysis_var_pass.cc memory_reuse_types.cc DEPS graph graph_helper pass) +if(WITH_GPU) +cc_library(memory_optimize_helper SRCS memory_optimize_helper.cc DEPS graph graph_helper gpu_info) +else() +cc_library(memory_optimize_helper SRCS memory_optimize_helper.cc DEPS graph graph_helper cpu_info) +endif() + +cc_library(memory_optimize_pass SRCS memory_optimize_pass.cc DEPS memory_optimize_helper pass) +cc_library(inplace_op_pass SRCS inplace_op_pass.cc DEPS memory_optimize_pass op_info) cc_library(modify_op_lock_and_record_event_pass SRCS modify_op_lock_and_record_event_pass.cc DEPS computation_op_handle op_graph_view multi_devices_helper) -cc_library(memory_early_delete_pass SRCS memory_early_delete_pass.cc DEPS memory_optimize_pass computation_op_handle scale_loss_grad_op_handle rpc_op_handle - all_reduce_op_handle reduce_op_handle broadcast_op_handle data_balance_op_handle graph graph_helper pass) cc_library(reference_count_pass_helper SRCS reference_count_pass_helper.cc DEPS garbage_collector computation_op_handle) cc_library(eager_deletion_op_handle SRCS eager_deletion_op_handle.cc DEPS lod_tensor selected_rows reference_count_pass_helper) cc_library(eager_deletion_pass SRCS eager_deletion_pass.cc DEPS computation_op_handle eager_deletion_op_handle graph graph_helper pass) @@ -65,13 +70,11 @@ cc_library(all_reduce_deps_pass SRCS all_reduce_deps_pass.cc DEPS graph graph_he cc_library(multi_devices_graph_pass SRCS multi_devices_graph_pass.cc DEPS multi_devices_helper computation_op_handle scale_loss_grad_op_handle rpc_op_handle all_reduce_op_handle reduce_op_handle broadcast_op_handle data_balance_op_handle fused_broadcast_op_handle) -set(SSA_GRAPH_EXECUTOR_DEPS graph framework_proto sequential_execution_pass modify_op_lock_and_record_event_pass all_reduce_deps_pass reference_count_pass eager_deletion_pass memory_optimize_pass memory_early_delete_pass) +set(SSA_GRAPH_EXECUTOR_DEPS graph framework_proto sequential_execution_pass modify_op_lock_and_record_event_pass all_reduce_deps_pass reference_count_pass eager_deletion_pass memory_optimize_pass inplace_op_pass) if (WITH_GPU) list(APPEND SSA_GRAPH_EXECUTOR_DEPS reference_count_pass) endif() -cc_test(memory_reuse_types_test SRCS memory_reuse_types_test.cc memory_reuse_types.cc DEPS framework_proto graph) -cc_test(analysis_var_pass_test SRCS analysis_var_pass_test.cc analysis_var_pass.cc memory_reuse_types.cc DEPS framework_proto graph graph_helper op_registry pass) - +cc_test(memory_optimize_helper_test SRCS memory_optimize_helper_test.cc memory_optimize_helper.cc DEPS framework_proto graph graph_helper op_registry) cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS ${SSA_GRAPH_EXECUTOR_DEPS}) cc_library(threaded_ssa_graph_executor SRCS threaded_ssa_graph_executor.cc DEPS fetch_op_handle ssa_graph_executor scope diff --git a/paddle/fluid/framework/details/all_reduce_deps_pass.cc b/paddle/fluid/framework/details/all_reduce_deps_pass.cc index b7d6edd389d8e40835dadf56d7c54d53402f6f4d..2e20c436dfdb61fcda78cd044b86848c750cf22c 100644 --- a/paddle/fluid/framework/details/all_reduce_deps_pass.cc +++ b/paddle/fluid/framework/details/all_reduce_deps_pass.cc @@ -30,8 +30,6 @@ namespace paddle { namespace framework { namespace details { -static constexpr char kAllOpDescs[] = "all_op_descs"; - VarHandle* GetValidInput(const OpHandleBase* a) { for (auto p : a->Inputs()) { VarHandle* b = dynamic_cast(p); diff --git a/paddle/fluid/framework/details/all_reduce_op_handle.cc b/paddle/fluid/framework/details/all_reduce_op_handle.cc index dd77f7099f581a5b825916c4ea010023f3ad5bcd..c1f9c2b60c915370df7793f26fe83812a7ced96d 100644 --- a/paddle/fluid/framework/details/all_reduce_op_handle.cc +++ b/paddle/fluid/framework/details/all_reduce_op_handle.cc @@ -53,7 +53,7 @@ AllReduceOpHandle::AllReduceOpHandle(ir::Node *node, #endif void AllReduceOpHandle::RunImpl() { - platform::RecordEvent record_event(Name(), dev_ctxes_.cbegin()->second); + platform::RecordEvent record_event(Name()); WaitInputVarGenerated(); auto in_var_handles = DynamicCast(this->Inputs()); diff --git a/paddle/fluid/framework/details/analysis_var_pass.cc b/paddle/fluid/framework/details/analysis_var_pass.cc deleted file mode 100644 index 223b9da3cfba33fc32d1334cddccb9f503bd0bef..0000000000000000000000000000000000000000 --- a/paddle/fluid/framework/details/analysis_var_pass.cc +++ /dev/null @@ -1,656 +0,0 @@ -// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -#include "paddle/fluid/framework/details/analysis_var_pass.h" -#include -#include -#include -#include -#include -#include -#include -#include -#include -#include -#include -#include -#include "gflags/gflags.h" -#include "paddle/fluid/framework/data_type.h" -#include "paddle/fluid/framework/ir/graph.h" -#include "paddle/fluid/framework/ir/graph_helper.h" - -DEFINE_bool(enable_subgraph_optimize, false, - "SubGraph also reuse global graph variables, it will reduce the " - "memory occupation" - "but a higher risk of memory reuse error. default disabled."); -DEFINE_string(memory_optimize_debug, "", - "debug the operator output variable when do the variable reuse." - "memory reuse pass." - "only for debug, default disabled."); - -namespace paddle { -namespace framework { -namespace details { - -static inline bool IsSameDesc(OpDesc* op1, OpDesc* op2) { - return op1->Type() == op2->Type() && op1->Inputs() == op2->Inputs() && - op1->Outputs() == op2->Outputs(); -} - -template -class FilterVariableImpl { - public: - void operator()(const Container& nodes, Callback callback) { - for (auto* node : nodes) { - callback(node); - } - } -}; - -// filter var node for op->inputs/outputs -template -class FilterVariableImpl, Callback> { - public: - void operator()(const std::vector& nodes, Callback callback) { - for (auto* var : nodes) { - if (var->IsVar() && !var->IsCtrlVar()) { - callback(var); - } - } - } -}; - -template -void FilterVariables(const Container& nodes, Callback callback) { - FilterVariableImpl()(nodes, callback); -} - -std::unique_ptr AnalysisVarPass::ApplyImpl( - std::unique_ptr graph) const { - auto nodes = graph->Nodes(); - auto subblock_vars = GetSubBlockVars(nodes); - skip_set_.insert(subblock_vars.begin(), subblock_vars.end()); - - cfg_.reset(new details::ControlFlowGraph(*graph)); - cfg_->LiveVariableAnalysis(); - InitSSAGraphNodes(); - - int reuse_id = 0; - for (size_t idx = 0; idx < cfg_->Ops().size(); ++idx) { - auto& op = cfg_->Ops()[idx]; - auto* op_desc = op->Op(); - // some op in graph has no op desc - if (op_desc == nullptr) continue; - if (OpHasSubBlock(op_desc)) { - if (FLAGS_enable_subgraph_optimize) { - SubGraphOptimize(op_desc); - } else { - VLOG(3) << op->Name() - << " has subblock, but disable subgraph optimize. skipped."; - continue; - } - } - - for (auto& var : op->outputs) { - if (NodeCanReused(var) && cfg_->Use(op).count(var->Name()) == 0) { - ir::Node* cache = pool_.NodeMatch(var); - if (var->Name() == FLAGS_memory_optimize_debug) { - VLOG(3) << "start match var " << DebugString(var) << " of op " - << op->Name(); - VLOG(3) << pool_.ToString(); - VLOG(3) << "matched in pool : " - << ((cache == nullptr) ? "False" : "True"); - } - if (cache != nullptr) { - if (var->Name() == cache->Name()) { - VLOG(3) << "The same cache variable is cascade reused." - << var->Name() << " is re-filled to the pool after" - << "the reused op is finished. Current op can not " - << "replace it again. Skip this candidate."; - continue; - } - - int node_idx_in_pool = pool_.GetIndex(cache); - VLOG(3) << string::Sprintf( - "!!! %s, %s => %s, cache idx %d, pool size %d", - std::to_string(reuse_id++), DebugString(var), DebugString(cache), - node_idx_in_pool, static_cast(pool_.size())); - // update CFG Graph on the fly. - // reused var maybe re-fill into the pool - cfg_->RenameVarInCFGGraph(var->Name(), cache->Name(), idx); - // NOTE(dzhwinter): we need to both update the ProgramDesc - // and IR Graph. because op_desc/var_desc is used in CreateOp, - // CreateVar when running happens. But IR Graph - // define the dependence relationship between nodes. - RenameVarInGraphDesc(var->Name(), cache->Name(), idx); - RenameVarInGraphNode(var->Name(), cache->Name(), idx, graph.get()); - - pool_.Erase(cache); - } - } - } - // fill the pool - for (auto var : cfg_->LiveIn(op)) { - if (cfg_->LiveOut(op).count(var) == 0) { - ir::Node* var_node = cfg_->GetNodeFromVarName(var, op); - if (var_node == nullptr) continue; - if (NodeCanReused(var_node) && !pool_.Has(var_node)) { - pool_.Insert(var_node, op); - } - } - } - } - graph->ResolveHazard(var_nodes_); - - // For early delete pass. use GraphNodePool load the unlived vars. - // 1. find all deps op for each unlived var in memory pool. - for (auto& op : graph->Nodes()) { - for (auto& var : op->inputs) { - if (pool_.Has(var)) { - pool_.Insert(var, op); - } - } - } - // 2. convert ir node based memory pool to graph node - // because Node* maybe released bettwen passes. - auto& graph_pool = graph->Get(kGraphNodePool); - for (auto it = pool_.begin(); it != pool_.end(); ++it) { - std::unordered_set descs; - for (auto& op : it->second) { - PADDLE_ENFORCE(op->IsOp()); - descs.insert(op->Op()); - } - graph_pool.push_back(std::make_pair(it->first->Name(), descs)); - } - - return graph; -} - -void AnalysisVarPass::SubGraphOptimize(OpDesc* op_desc) const { - // conditional block, while op and their grad op - auto* sub_block_desc = - AttrReader(op_desc->GetAttrMap()).Get("sub_block"); - - // create a mirror block to construct an IR Graph. - ProgramDesc prog; - auto* copy_block = prog.MutableBlock(0); - for (auto* op : sub_block_desc->AllOps()) { - auto* copy_op = copy_block->AppendOp(); - copy_op->CopyFrom(*op); - copy_op->Flush(); - } - - for (auto* var : sub_block_desc->AllVars()) { - auto* copy_var = copy_block->Var(var->Name()); - copy_var->SetDataType(var->GetDataType()); - // only lod tensor can be reused. So ignore the multiple dims case. - copy_var->SetType(var->GetType()); - copy_var->SetShape(var->GetShape()); - copy_var->SetPersistable(var->Persistable()); - } - - ir::Graph sub_graph(prog); - std::unordered_set sub_graph_all_ops; - FilterVariables(sub_graph.Nodes(), [&](ir::Node* var) { - // sub_graph_all_ops.emplace(var); - if (var->IsVar() && !var->IsCtrlVar()) { - sub_graph_all_ops.emplace(var); - } - }); - int sub_reuse_id = 0; - // subgraph nodes is unordered, reuse need to follow the desc order. - // find the right op node through the descs - for (auto* sub_op_desc : sub_block_desc->AllOps()) { - ir::Node* sub_op = nullptr; - for (auto* node : sub_graph_all_ops) { - if (node->Op() == sub_op_desc) { - sub_op = node; - break; - } - } - PADDLE_ENFORCE(sub_op != nullptr); - for (auto* var : sub_op->outputs) { - if (NodeCanReused(var)) { - ir::Node* cache = pool_.NodeMatch(var); - if (cache != nullptr) { - if (var->Var()->GetDataType() != cache->Var()->GetDataType()) { - continue; - } - int node_idx_in_pool = pool_.GetIndex(cache); - VLOG(3) << string::Sprintf( - "!!! %s, %s => %s, cache idx %d, pool size %d", - std::to_string(sub_reuse_id++), DebugString(var), - DebugString(cache), node_idx_in_pool, - static_cast(pool_.size())); - // NOTE(dzh): subblock is not in IR graph. Modify the block_desc - // immediately to make the subblock variable reuse strategy take - // effect. Because it is a single op in graph. No need to - // update the ir nodes. - sub_op_desc->Rename(var->Name(), cache->Name()); - if (sub_op_desc->Block()->HasVar(var->Name())) { - sub_op_desc->Block()->RemoveVar(var->Name()); - } - } - } - } - } -} - -std::unordered_set AnalysisVarPass::GetSubBlockVars( - const std::unordered_set& nodes) const { - std::unordered_set vars; - for (auto& op : nodes) { - if (!op->IsOp() || op->Op() == nullptr) continue; - auto* op_desc = op->Op(); - if (OpHasSubBlock(op_desc)) { - auto inputs = op_desc->InputArgumentNames(); - auto outputs = op_desc->OutputArgumentNames(); - vars.insert(inputs.begin(), inputs.end()); - vars.insert(outputs.begin(), outputs.end()); - } - } - return vars; -} - -void AnalysisVarPass::RenameVarInGraphDesc(const std::string& var, - const std::string& cache_var, - size_t idx) const { - for (size_t i = idx; i < cfg_->Ops().size(); ++i) { - auto* op = cfg_->Ops()[i]; - PADDLE_ENFORCE(op->IsOp() && op->Op()); - auto* op_desc = op->Op(); - op_desc->RenameInput(var, cache_var); - op_desc->RenameOutput(var, cache_var); - if (op_desc->Block()->HasVar(var)) op_desc->Block()->RemoveVar(var); - op_desc->Flush(); - } -} - -void AnalysisVarPass::InitSSAGraphNodes() const { - std::unordered_map> all_vars; - if (var_nodes_.empty()) { - for (auto* op : cfg_->Ops()) { - for (auto* node : op->inputs) { - if (all_vars[node->Name()].count(node) == 0) { - all_vars[node->Name()].emplace(node); - var_nodes_[node->Name()].emplace_back(node); - } - } - for (auto* node : op->outputs) { - if (all_vars[node->Name()].count(node) == 0) { - all_vars[node->Name()].emplace(node); - var_nodes_[node->Name()].emplace_back(node); - } - } - } - } -} - -void AnalysisVarPass::RenameVarInGraphNode(const std::string& var, - const std::string& cache_var, - size_t idx, ir::Graph* graph) const { - // if replace happens, we need to create a newer version cache_var - // but use the same dims/data_type with var. - PADDLE_ENFORCE(var_nodes_[var].size() >= 1 && - var_nodes_[var].at(0)->Var() != nullptr); - std::unique_ptr var_desc(new VarDesc(*var_nodes_[var].at(0)->Var())); - var_desc->SetName(cache_var); - - for (size_t i = idx; i < cfg_->Ops().size(); ++i) { - auto* op = cfg_->Ops()[i]; - - // redirect the input to the latest version of cache_var - for (auto* node : op->inputs) { - if (node->Name() == var) { - ir::Node* cache_node = graph->CreateVarNode(var_desc.get()); - var_nodes_[cache_var].emplace_back(cache_node); - - // swap node to cache_node - cache_node->outputs.insert(cache_node->outputs.end(), - node->outputs.begin(), node->outputs.end()); - PADDLE_ENFORCE(node->inputs.size() == 1 && node->inputs[0]->IsOp()); - auto* prev_op = node->inputs[0]; - std::replace(prev_op->outputs.begin(), prev_op->outputs.end(), node, - cache_node); - cache_node->inputs.emplace_back(prev_op); - for (auto* next_op : node->outputs) { - std::replace(next_op->inputs.begin(), next_op->inputs.end(), node, - cache_node); - } - } - } - - // if we need to rename the output, - // always create a newer version of cache_var - for (auto* node : op->outputs) { - if (node->Name() == var) { - ir::Node* cache_node = graph->CreateVarNode(var_desc.get()); - var_nodes_[cache_var].emplace_back(cache_node); - - // swap node to cache node - cache_node->outputs.insert(cache_node->outputs.end(), - node->outputs.begin(), node->outputs.end()); - cache_node->inputs.emplace_back(op); - std::replace(op->outputs.begin(), op->outputs.end(), node, cache_node); - for (auto* next_op : node->outputs) { - std::replace(next_op->inputs.begin(), next_op->inputs.end(), node, - cache_node); - } - } - } - } - - // release node of unused var in graph - for (auto* node : var_nodes_[var]) { - graph->RemoveNode(node); - } - var_nodes_.at(var).clear(); -} - -bool AnalysisVarPass::NodeCanReused(ir::Node* node) const { - if (!node->IsVar() || node->IsCtrlVar()) return false; - auto* desc = node->Var(); - auto type = desc->GetType(); - if (desc->Persistable() || type != proto::VarType::LOD_TENSOR || - desc->GetShape().empty()) { - return false; - } - // vars can be @EMPTY@, @LR_DECAY_REUSE_ID@. For example, while_grad - std::string name = node->Name(); - if (!name.empty() && name[0] == '@' && name[name.size() - 1] == '@') - return false; - if (skip_set_.count(name)) return false; - for (auto* op : node->inputs) { - if (op->Op()->HasAttr("force_cpu")) { - // op output force generated in cpu, can not be reused. - return framework::AttrReader(op->Op()->GetAttrMap()) - .Get("force_cpu") == 0; - } - } - return true; -} - -bool AnalysisVarPass::OpHasSubBlock(OpDesc* desc) const { - const AttributeMap& attrs = desc->GetAttrMap(); - for (auto& attr : attrs) { - if (attr.second.type() == typeid(BlockDesc*) || // NOLINT - attr.second.type() == typeid(std::vector)) // NOLINT - return true; - } - return false; -} - -std::vector SortOpLikeDescOrder(const ir::Graph& graph) { - PADDLE_ENFORCE(graph.Has(kAllOpDescs), - "Graph has no attribute of kAllOpDescs."); - // 1. get op desc order - auto& op_descs = graph.Get>(kAllOpDescs); - - // 2. topology sort order - auto nodes = graph.Nodes(); - std::deque ops; - FilterVariables(nodes, [&](ir::Node* op) { - if (op->IsOp() && op->Op() != nullptr) { - ops.emplace_back(op); - } - }); - std::unordered_map op_deps; - std::list ready_ops; - std::unordered_map> pending_ops; - - for (auto* op : ops) { - std::unordered_set preceding_op; - for (auto* in : op->inputs) { - if (in->inputs.empty()) continue; - PADDLE_ENFORCE(in->inputs.size() == 1 && in->inputs[0]->IsOp()); - preceding_op.emplace(in->inputs[0]); - pending_ops[in->inputs[0]].emplace(op); - } - op_deps[op] = preceding_op.size(); - if (preceding_op.empty()) { - ready_ops.emplace_back(op); - } - } - - // 3. generated op list based desc order and the topology order - std::vector ret; - std::list op_descs_list(op_descs.begin(), op_descs.end()); - - auto update_by_found_node = [&](ir::Node* found_node) { - for (auto* pending_op : pending_ops[found_node]) { - if (--op_deps[pending_op] == 0) { - ready_ops.emplace_back(pending_op); - } - } - ready_ops.remove(found_node); - ret.emplace_back(found_node); - }; - - while (!ready_ops.empty()) { - bool all_of_ready_op_unmatched = true; - for (auto it = op_descs_list.begin(); it != op_descs_list.end();) { - auto op_desc = *it; - ir::Node* found_node = nullptr; - for (auto* op : ready_ops) { - if (IsSameDesc(op->Op(), op_desc)) { - found_node = op; - break; - } - } - - // 3.1 op desc deleted by other pass - if (found_node == nullptr) { - ++it; - continue; - } else { - all_of_ready_op_unmatched = false; - it = op_descs_list.erase(it); - } - update_by_found_node(found_node); - } - - // 3.2 op descs are added by other pass - // preceding op non empty means some new op descs are - // created, but not contained in return node list. - // these new op desc may depend on each other. - std::list prev_ready_ops(ready_ops); - if (all_of_ready_op_unmatched) { - for (auto op : prev_ready_ops) { - update_by_found_node(op); - } - } - } - - PADDLE_ENFORCE(std::all_of( - op_deps.begin(), op_deps.end(), - [&](const std::pair& p) { return p.second == 0; })); - - return ret; -} - -ControlFlowGraph::ControlFlowGraph(const ir::Graph& graph) { - ops_ = SortOpLikeDescOrder(graph); - ConnectNodes(); -} - -void ControlFlowGraph::BuildCFGGraph() { - // FIXME(dzh): same effect with ConnectNodes, but use the control - // link to build dependency graph, it goes wrong in transformer. - for (ir::Node* op : ops_) { - for (auto& input_var : op->inputs) { - if (!input_var->inputs.empty()) { - PADDLE_ENFORCE( - input_var->inputs.size() == 1 && input_var->inputs[0]->IsOp(), - "Preceding Op Node of Var Node must be unique"); - auto* pred_op = input_var->inputs[0]; - if (pred_op->Op() != nullptr) { - predecessors_[op].insert(pred_op); - successors_[pred_op].insert(op); - } - } - if (input_var->IsVar() && !input_var->IsCtrlVar()) { - uses_[op].insert(input_var->Name()); - } - } - for (auto& output_var : op->outputs) { - // output var may be used by many op - for (auto* succ_op : output_var->outputs) { - if (succ_op->Op() != nullptr) { - successors_[op].insert(succ_op); - predecessors_[succ_op].insert(op); - } - } - if (output_var->IsVar() && !output_var->IsCtrlVar()) { - defs_[op].insert(output_var->Name()); - } - } - } -} - -void ControlFlowGraph::ConnectNodes() { - for (size_t i = 0; i < ops_.size(); ++i) { - auto& op = ops_[i]; - try { - auto& next_op = ops_.at(i + 1); - successors_[op].insert(next_op); - predecessors_[next_op].insert(op); - } catch (...) { - // do nothing - } - - FilterVariables(op->inputs, - [&](ir::Node* var) { uses_[op].emplace(var->Name()); }); - - FilterVariables(op->outputs, - [&](ir::Node* var) { defs_[op].emplace(var->Name()); }); - } -} - -void ControlFlowGraph::LiveVariableAnalysis() { - // NOTE(dzh): variable liveless analysis (a.k.a reversed_ops algorithm) - // compute the liveness of for each variable though reversed_ops algorithm. - // It iterates the operators from end to begin, compute the live in/live out - // variable set for each op, then the diff between in/out will be used for - // the variable reuse. For detail refer to - // http://www.cs.cornell.edu/courses/cs4120/2013fa/lectures/lec26-fa13.pdf - std::list work_list(ops_.rbegin(), ops_.rend()); - while (!work_list.empty()) { - ir::Node* op = work_list.front(); - work_list.pop_front(); - // get the live_in calculated before. Empty if first. - auto prev_live_in = std::move(live_in_[op]); - for (auto& s : successors_[op]) { - for (auto& var : live_in_[s]) { - live_out_[op].insert(var); - } - } - for (auto& var : uses_[op]) { - live_in_[op].insert(var); - } - for (auto& var : live_out_[op]) { - live_in_[op].insert(var); - } - for (auto& var : defs_[op]) { - live_in_[op].erase(var); - } - - // If the live_in is not changed, then the liveness analysis of - // predecessors is completed. - // - // Otherwise, recalculate the predecessors liveness - if (live_in_[op] != prev_live_in) { - for (auto& pre : predecessors_[op]) { - work_list.push_back(pre); - } - } - } -} - -void ControlFlowGraph::RenameVarInCFGGraph(const std::string& old_node, - const std::string& new_node, - int begin_idx) { - // update graph from begin idx to the end - for (size_t i = begin_idx; i != ops_.size(); ++i) { - auto* op = ops_[i]; - if (uses_[op].find(old_node) != uses_[op].end()) { - uses_[op].erase(old_node); - uses_[op].insert(new_node); - } - if (defs_[op].find(old_node) != defs_[op].end()) { - defs_[op].erase(old_node); - defs_[op].insert(new_node); - } - if (live_in_[op].find(old_node) != live_in_[op].end()) { - live_in_[op].erase(old_node); - live_in_[op].insert(new_node); - } - if (live_out_[op].find(old_node) != live_out_[op].end()) { - live_out_[op].erase(old_node); - live_out_[op].insert(new_node); - } - } -} - -const std::set ControlFlowGraph::LiveIn(ir::Node* op) const { - auto it = live_in_.find(op); - PADDLE_ENFORCE( - it != live_in_.end(), - string::Sprintf("Expect %s in live_in, but Not Found.", op->Name())); - return it->second; -} - -const std::set ControlFlowGraph::LiveOut(ir::Node* op) const { - auto it = live_out_.find(op); - PADDLE_ENFORCE( - it != live_out_.end(), - string::Sprintf("Expect %s in live_out, but Not Found.", op->Name())); - return it->second; -} - -const std::set ControlFlowGraph::Use(ir::Node* op) const { - auto it = uses_.find(op); - PADDLE_ENFORCE( - it != uses_.end(), - string::Sprintf("Expect %s in live_out, but Not Found.", op->Name())); - return it->second; -} - -const std::vector ControlFlowGraph::Ops() const { return ops_; } - -std::vector& ControlFlowGraph::Ops() { return ops_; } - -ir::Node* ControlFlowGraph::GetNodeFromVarName(const std::string& name, - ir::Node* op) const { - // in ssa-graph, different version nodes have same name, - // this function get the latest version var before target op - // It may return nullptr, such as data node. - ir::Node* found_node = nullptr; - for (auto* node : ops_) { - if (node == op) break; - for (auto& output : node->outputs) { - if (output->Name() == name) { - found_node = output; - } - } - } - return found_node; -} - -} // namespace details -} // namespace framework -} // namespace paddle - -REGISTER_PASS(analysis_var_pass, paddle::framework::details::AnalysisVarPass) - .RequireGraphAttr(paddle::framework::details::kGraphNodePool) - .RequireGraphAttr(paddle::framework::details::kAllOpDescs); diff --git a/paddle/fluid/framework/details/analysis_var_pass.h b/paddle/fluid/framework/details/analysis_var_pass.h deleted file mode 100644 index 144204beafb341351172c29e3b4cd41db49be6f9..0000000000000000000000000000000000000000 --- a/paddle/fluid/framework/details/analysis_var_pass.h +++ /dev/null @@ -1,120 +0,0 @@ -// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -#pragma once - -#include -#include -#include -#include -#include -#include -#include -#include -#include - -#include "paddle/fluid/framework/data_type.h" -#include "paddle/fluid/framework/details/memory_reuse_types.h" -#include "paddle/fluid/framework/ir/graph.h" -#include "paddle/fluid/framework/ir/pass.h" - -namespace paddle { -namespace framework { -namespace details { -constexpr char kAllOpDescs[] = "all_op_descs"; - -std::vector SortOpLikeDescOrder(const ir::Graph& graph); -// sort op in bfs order -std::vector BFSSortGraphOps(const ir::Graph& graph); - -class ControlFlowGraph; - -class AnalysisVarPass : public ir::Pass { - protected: - std::unique_ptr ApplyImpl( - std::unique_ptr graph) const override; - - private: - // fill the variable map(var_nodes) by version. - void InitSSAGraphNodes() const; - // update program descs - void RenameVarInGraphDesc(const std::string& var, - const std::string& cache_var, size_t idx) const; - // update ir nodes - void RenameVarInGraphNode(const std::string& var, - const std::string& cache_var, size_t idx, - ir::Graph* graph) const; - - void SubGraphOptimize(OpDesc* op_desc) const; - // valid a tensor can be reuse or not - bool NodeCanReused(ir::Node* node) const; - // scan subblock and collect the output/input variables. - std::unordered_set GetSubBlockVars( - const std::unordered_set&) const; - // check op has subblock or not - bool OpHasSubBlock(OpDesc* desc) const; - - private: - // Reuse Node Pool, Owned. - mutable OrderedNodePairPool pool_; - // controlflow Graph - mutable std::unique_ptr cfg_; - // skip set - mutable std::unordered_set skip_set_; - // var nodes - mutable std::map> var_nodes_; -}; - -class ControlFlowGraph { - public: - ControlFlowGraph() = default; - // For IR Graph in parallelexecutor - explicit ControlFlowGraph(const ir::Graph& graph); - - void LiveVariableAnalysis(); - - void RenameVarInCFGGraph(const std::string& old_node, - const std::string& new_node, int begin_idx); - - const std::set LiveIn(ir::Node* op) const; - const std::set LiveOut(ir::Node* op) const; - const std::set Use(ir::Node* op) const; - const std::vector Ops() const; - std::vector& Ops(); - - // for ssa-graph nodes - ir::Node* GetNodeFromVarName(const std::string& name, ir::Node* op) const; - - private: - void BuildCFGGraph(); - void ConnectNodes(); - using NodeListMap = std::unordered_map>; - using VarSetMap = std::map>; - // successors ops use the output variables. - NodeListMap successors_; - // predecessors ops generated input variables. - NodeListMap predecessors_; - // variables lived before run current op. - VarSetMap live_in_; - // variables lived after run current op. - VarSetMap live_out_; - VarSetMap uses_; // op inputs - VarSetMap defs_; // op outputs - - std::vector ops_; // op sequence by topology sort -}; - -} // namespace details -} // namespace framework -} // namespace paddle diff --git a/paddle/fluid/framework/details/broadcast_op_handle.cc b/paddle/fluid/framework/details/broadcast_op_handle.cc index 89d626edddfee3d2c43a3cf2232ad4fc1611e655..fdff83b92819b39974f3b2ce0848710f1ee02a41 100644 --- a/paddle/fluid/framework/details/broadcast_op_handle.cc +++ b/paddle/fluid/framework/details/broadcast_op_handle.cc @@ -22,7 +22,7 @@ namespace framework { namespace details { void BroadcastOpHandle::RunImpl() { - platform::RecordEvent record_event(Name(), dev_ctxes_.begin()->second); + platform::RecordEvent record_event(Name()); if (places_.size() == 1) return; @@ -30,7 +30,7 @@ void BroadcastOpHandle::RunImpl() { VarHandle *in_var_handle; { auto in_var_handles = DynamicCast(inputs_); - PADDLE_ENFORCE_EQ(in_var_handles.size(), 1, + PADDLE_ENFORCE_EQ(in_var_handles.size(), 1UL, "The number of input should be one."); in_var_handle = in_var_handles[0]; } diff --git a/paddle/fluid/framework/details/build_strategy.cc b/paddle/fluid/framework/details/build_strategy.cc index ce5731a1f414e8ef6d8af22a3bb17109e82beb87..17dd1399119d190bcbc31adb34ec61deb92a9994 100644 --- a/paddle/fluid/framework/details/build_strategy.cc +++ b/paddle/fluid/framework/details/build_strategy.cc @@ -17,7 +17,7 @@ limitations under the License. */ #include #include -#include "paddle/fluid/framework/details/memory_reuse_types.h" +#include "paddle/fluid/framework/details/memory_optimize_helper.h" #include "paddle/fluid/framework/details/multi_devices_graph_pass.h" #include "paddle/fluid/framework/details/multi_devices_graph_print_pass.h" #include "paddle/fluid/framework/details/reduce_op_handle.h" @@ -34,9 +34,11 @@ namespace details { static inline bool SeqOnlyAllReduceOps(const BuildStrategy &strategy) { // Should fix the allreduce op order if scheduling // them in multiple threads or processes to avoid hang. + // NOTE: ParallelGraph would execute this pass on each graph, so + // don't need to append it here. return (!strategy.enable_sequential_execution_ && - strategy.num_trainers_ > 1) || - strategy.enable_parallel_graph_; + strategy.num_trainers_ > 1) && + !strategy.enable_parallel_graph_; } class ParallelExecutorPassBuilder : public ir::PassBuilder { @@ -47,6 +49,22 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder { AppendPass("sequential_execution_pass"); } + // Add op fusion. + if (strategy.fuse_relu_depthwise_conv_) { + AppendPass("fuse_relu_depthwise_conv_pass"); + } + + // NOTE(dzhwinter): A note for automatical inplace. + // 1. modify program desc passes should put + // before inplace pass. + // 2. manually configured inplace should put + // before inplace_pass + + // Add automatically inplace. + if (strategy_.enable_inplace_) { + AppendPass("inplace_pass"); + } + // Add a graph viz pass to record a graph. if (!strategy_.debug_graphviz_path_.empty()) { auto viz_pass = AppendPass("graph_viz_pass"); @@ -55,10 +73,6 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder { viz_pass->Set("graph_viz_path", new std::string(graph_path)); } - // Add op fusion. - if (strategy.fuse_relu_depthwise_conv_) { - AppendPass("fuse_relu_depthwise_conv_pass"); - } if (strategy.fuse_elewise_add_act_ops_) { auto fuse_elewise_add_act_pass = AppendPass("fuse_elewise_add_act_pass"); // Add a graph viz pass to record a graph. @@ -88,7 +102,7 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder { // A side-effect of that, memory optimize cannot forsee the fetched vars // , so fetchlist should be set persistable before call the Run interface. if (strategy.memory_optimize_) { - auto analysis_var_pass = AppendPass("analysis_var_pass"); + auto memory_optimize_pass = AppendPass("memory_optimize_pass"); } AppendMultiDevPass(strategy); @@ -121,12 +135,15 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder { void AppendMultiDevPass(const BuildStrategy &strategy) { ir::Pass *multi_devices_pass; if (strategy_.is_distribution_) { + VLOG(3) << "multi device parameter server mode"; multi_devices_pass = AppendPass("dist_multi_devices_pass").get(); } else { if (strategy.reduce_ == BuildStrategy::ReduceStrategy::kAllReduce) { + VLOG(3) << "multi devices collective mode with allreduce"; multi_devices_pass = AppendPass("allreduce_mode_multi_devices_pass").get(); } else if (strategy.reduce_ == BuildStrategy::ReduceStrategy::kReduce) { + VLOG(3) << "multi deivces collective mode with reduce"; multi_devices_pass = AppendPass("reduce_mode_multi_devices_pass").get(); } else { PADDLE_THROW("Unknown reduce strategy."); @@ -186,14 +203,14 @@ std::unique_ptr BuildStrategy::Apply( pass->Erase("nccl_ctxs"); pass->SetNotOwned("nccl_ctxs", nctx); #endif - - } else if (pass->Type() == "analysis_var_pass") { + } else if (pass->Type() == "memory_optimize_pass") { + if (graph->Has(kAllOpDescs)) { + graph->Erase(kAllOpDescs); + } const std::vector *all_op_descs = new std::vector(main_program.Block(0).AllOps()); graph->Set>(kAllOpDescs, all_op_descs); // take ownership - graph->Set(kGraphNodePool, - new GraphNodePool); // take ownership pass->Erase(kAllOpDescs); pass->SetNotOwned>(kAllOpDescs, all_op_descs); @@ -214,6 +231,13 @@ std::unique_ptr BuildStrategy::Apply( pass->Set>( kAllOpDescs, new std::vector(main_program.Block(0).AllOps())); + } else if (pass->Type() == "inplace_pass") { + if (graph->Has(kAllOpDescs)) { + graph->Erase(kAllOpDescs); + } + graph->Set>( + kAllOpDescs, + new std::vector(main_program.Block(0).AllOps())); } else if (pass->Type() == "fuse_relu_depthwise_conv_pass") { if (!use_cuda) { LOG(WARNING) << "fuse_relu_depthwise_conv_pass is only supported on " @@ -221,7 +245,9 @@ std::unique_ptr BuildStrategy::Apply( continue; } } + VLOG(3) << "Start Apply Pass " << pass->Type(); graph = pass->Apply(std::move(graph)); + VLOG(3) << "Finish Apply Pass " << pass->Type(); } return graph; } @@ -239,9 +265,10 @@ USE_PASS(allreduce_mode_multi_devices_pass); USE_PASS(dist_multi_devices_pass); USE_PASS(multi_devices_check_pass); USE_PASS(multi_devices_print_pass); -USE_PASS(analysis_var_pass); +USE_PASS(memory_optimize_pass); USE_PASS(sequential_execution_pass); USE_PASS(all_reduce_deps_pass); USE_PASS(modify_op_lock_and_record_event_pass); +USE_PASS(inplace_pass); USE_PASS(lock_free_optimize_pass); USE_PASS(graph_to_program_pass); diff --git a/paddle/fluid/framework/details/build_strategy.h b/paddle/fluid/framework/details/build_strategy.h index cd24a3175953bf323748bf0c7e3159761c13f0a9..e62e3edcef710df739c53b5d848f5aceb4f2db4e 100644 --- a/paddle/fluid/framework/details/build_strategy.h +++ b/paddle/fluid/framework/details/build_strategy.h @@ -77,8 +77,10 @@ struct BuildStrategy { bool fuse_relu_depthwise_conv_{false}; bool memory_optimize_{false}; - - bool memory_early_delete_{false}; + // TODO(dzhwinter): + // make enable_inplace, memory_optimize_ + // memory_early_delete_ true by default + bool enable_inplace_{false}; bool enable_sequential_execution_{false}; diff --git a/paddle/fluid/framework/details/computation_op_handle.h b/paddle/fluid/framework/details/computation_op_handle.h index 601ae4f8c6de11b0bf25d4f9a92ef8eada67be3d..1e3dbb1e44ecb16872e3bf4dee31e31cc69c9818 100644 --- a/paddle/fluid/framework/details/computation_op_handle.h +++ b/paddle/fluid/framework/details/computation_op_handle.h @@ -26,7 +26,7 @@ namespace paddle { namespace framework { namespace details { -struct ComputationOpHandle : public OpHandleBase { +class ComputationOpHandle : public OpHandleBase { public: ComputationOpHandle(ir::Node *node, Scope *scope, platform::Place place, size_t scope_idx); diff --git a/paddle/fluid/framework/details/data_balance_op_handle.cc b/paddle/fluid/framework/details/data_balance_op_handle.cc index 48dcc52623369f7b0f51cd8c8aeb198b37467d5f..c9b52b68205ade000e21a3d06b80af86cbe01f34 100644 --- a/paddle/fluid/framework/details/data_balance_op_handle.cc +++ b/paddle/fluid/framework/details/data_balance_op_handle.cc @@ -86,7 +86,7 @@ std::vector> DataBalanceOpHandle::GetBalancePlan( } void DataBalanceOpHandle::RunImpl() { - PADDLE_ENFORCE_GT(places_.size(), 1, + PADDLE_ENFORCE_GT(places_.size(), 1UL, "Data balance can only be enabled when the number of " "places to run larger than 1."); auto in_var_handles = DynamicCast(this->Inputs()); diff --git a/paddle/fluid/framework/details/fuse_vars_op_handle.cc b/paddle/fluid/framework/details/fuse_vars_op_handle.cc index d65b0920698748e8a2ded728d78fbcd69b7bae0e..14292c0a5d06aa3ff12b46b5768b136fa925752d 100644 --- a/paddle/fluid/framework/details/fuse_vars_op_handle.cc +++ b/paddle/fluid/framework/details/fuse_vars_op_handle.cc @@ -23,7 +23,7 @@ void FuseVarsOpHandle::RunImpl() { auto in_var_handles = DynamicCast(this->Inputs()); auto out_var_handles = DynamicCast(this->Outputs()); - PADDLE_ENFORCE_EQ(in_var_handles.size(), 0); + PADDLE_ENFORCE_EQ(in_var_handles.size(), 0UL); PADDLE_ENFORCE_EQ(out_var_handles.size() - 1, inputs_numel_.size(), ""); auto scope = local_scope_->FindVar(kLocalExecScopeName)->Get(); diff --git a/paddle/fluid/framework/details/fused_broadcast_op_handle.cc b/paddle/fluid/framework/details/fused_broadcast_op_handle.cc index 51dfa2d0711f49aaefab0af3549283dbf77eee4a..f48561ea32e6a3bbc7e9f2a8326b080ad21c6d61 100644 --- a/paddle/fluid/framework/details/fused_broadcast_op_handle.cc +++ b/paddle/fluid/framework/details/fused_broadcast_op_handle.cc @@ -22,7 +22,7 @@ namespace framework { namespace details { void FusedBroadcastOpHandle::RunImpl() { - platform::RecordEvent record_event(Name(), dev_ctxes_.begin()->second); + platform::RecordEvent record_event(Name()); if (places_.size() == 1UL) return; diff --git a/paddle/fluid/framework/details/fused_broadcast_op_handle_test.cc b/paddle/fluid/framework/details/fused_broadcast_op_handle_test.cc index be0d941c4f9c2fe8fbb1da8ec2c11868112fcf9b..6d53dac5c0a20b4340e71274a00a7f3c0cd08ff6 100644 --- a/paddle/fluid/framework/details/fused_broadcast_op_handle_test.cc +++ b/paddle/fluid/framework/details/fused_broadcast_op_handle_test.cc @@ -34,8 +34,8 @@ struct TestFusedBroadcastOpHandle : TestBroadcastOpHandle { ->Var(details::kLocalExecScopeName) ->GetMutable() = &local_scope; for (size_t j = 0; j < input_scope_idxes.size(); ++j) { - local_scope.Var("out_var" + j); - if (i == j) local_scope.Var("in_var" + j); + local_scope.Var("out_var" + std::to_string(j)); + if (i == j) local_scope.Var("in_var" + std::to_string(j)); } param_scopes_.emplace_back(&local_scope); } @@ -62,20 +62,21 @@ struct TestFusedBroadcastOpHandle : TestBroadcastOpHandle { for (size_t i = 0; i < input_scope_idxes.size(); ++i) { // add input var handle - nodes_.emplace_back( - ir::CreateNodeForTest("in_node" + i, ir::Node::Type::kVariable)); - VarHandle* in_var_handle = - new VarHandle(nodes_.back().get(), 1, input_scope_idxes[i], - "in_var" + i, place_list_[input_scope_idxes[i]]); + nodes_.emplace_back(ir::CreateNodeForTest("in_node" + std::to_string(i), + ir::Node::Type::kVariable)); + VarHandle* in_var_handle = new VarHandle( + nodes_.back().get(), 1, input_scope_idxes[i], + "in_var" + std::to_string(i), place_list_[input_scope_idxes[i]]); vars_.emplace_back(in_var_handle); op_handle_->AddInput(in_var_handle); // add output var handle for (size_t j = 0; j < place_list_.size(); ++j) { - nodes_.emplace_back( - ir::CreateNodeForTest("out_node" + i, ir::Node::Type::kVariable)); - VarHandle* out_var_handle = new VarHandle( - nodes_.back().get(), 2, j, "out_var" + i, place_list_[j]); + nodes_.emplace_back(ir::CreateNodeForTest( + "out_node" + std::to_string(i), ir::Node::Type::kVariable)); + VarHandle* out_var_handle = + new VarHandle(nodes_.back().get(), 2, j, + "out_var" + std::to_string(i), place_list_[j]); vars_.emplace_back(out_var_handle); op_handle_->AddOutput(out_var_handle); } @@ -86,7 +87,7 @@ struct TestFusedBroadcastOpHandle : TestBroadcastOpHandle { std::vector> send_vec; f::LoD lod{{0, 10, 20}}; for (size_t i = 0; i < input_scope_idxes.size(); ++i) { - const std::string varname("in_var" + i); + const std::string varname("in_var" + std::to_string(i)); float val_scalar = static_cast(i); send_vec.push_back( InitLoDTensor(varname, input_scope_idxes[i], lod, val_scalar)); @@ -96,7 +97,7 @@ struct TestFusedBroadcastOpHandle : TestBroadcastOpHandle { WaitAll(); for (size_t i = 0; i < input_scope_idxes.size(); ++i) { - const std::string& varname("out_var" + i); + const std::string& varname("out_var" + std::to_string(i)); for (size_t j = 0; j < place_list_.size(); ++j) { LoDTensorEqual(varname, send_vec[i], lod, param_scopes_[j]); } @@ -109,7 +110,7 @@ struct TestFusedBroadcastOpHandle : TestBroadcastOpHandle { 2, 4, 6, 3, 1, 1, 1, 1, 3, 7}; int height = static_cast(kDims[0] * 2); for (size_t i = 0; i < input_scope_idxes.size(); ++i) { - const std::string varname("in_var" + i); + const std::string varname("in_var" + std::to_string(i)); float val_scalar = static_cast(i); send_vector.push_back(InitSelectedRows(varname, input_scope_idxes[i], rows, height, val_scalar)); @@ -119,7 +120,7 @@ struct TestFusedBroadcastOpHandle : TestBroadcastOpHandle { WaitAll(); for (size_t i = 0; i < input_scope_idxes.size(); ++i) { - const std::string& varname("out_var" + i); + const std::string& varname("out_var" + std::to_string(i)); for (size_t j = 0; j < place_list_.size(); ++j) { SelectedRowsEqual(varname, input_scope_idxes[i], send_vector[i], rows, height); diff --git a/paddle/fluid/framework/details/graph_test_base.h b/paddle/fluid/framework/details/graph_test_base.h new file mode 100644 index 0000000000000000000000000000000000000000..126959bcd80a4677f76b7cff677a82a319f7cfb3 --- /dev/null +++ b/paddle/fluid/framework/details/graph_test_base.h @@ -0,0 +1,80 @@ +// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include +#include +#include +#include +#include "glog/logging.h" +#include "gtest/gtest.h" +#include "paddle/fluid/framework/ir/graph.h" +#include "paddle/fluid/framework/ir/graph_helper.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/framework/program_desc.h" + +namespace paddle { +namespace framework { + +class DummyOp : public OperatorBase { + public: + DummyOp(const std::string& type, const VariableNameMap& inputs, + const VariableNameMap& outputs, const AttributeMap& attrs) + : OperatorBase(type, inputs, outputs, attrs) {} + + private: + void RunImpl(const Scope& scope, + const platform::Place& place) const override {} +}; + +class SumOpMaker : public OpProtoAndCheckerMaker { + public: + void Make() { + AddInput("X", "").AsDuplicable(); + AddOutput("Out", ""); + AddComment(""); + } +}; + +class AssignOpMaker : public OpProtoAndCheckerMaker { + public: + void Make() { + AddInput("X", "").AsDuplicable(); + AddOutput("Out", ""); + AddComment(""); + } +}; + +class SplitOpMaker : public OpProtoAndCheckerMaker { + public: + void Make() { + AddInput("X", ""); + AddOutput("Out", "").AsDuplicable(); + AddComment(""); + } +}; + +class DummyVarTypeInference : public VarTypeInference { + public: + void operator()(const OpDesc& op_desc, BlockDesc* block) const override { + auto& inputs = op_desc.Input("X"); + auto type = block->Var(inputs.front())->GetType(); + auto out_var_name = op_desc.Output("Out").front(); + block->Var(out_var_name)->SetType(type); + } +}; + +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/inplace_op_pass.cc b/paddle/fluid/framework/details/inplace_op_pass.cc new file mode 100644 index 0000000000000000000000000000000000000000..c91fc81b2defc9fe6b5720ce652a9aa94b27735e --- /dev/null +++ b/paddle/fluid/framework/details/inplace_op_pass.cc @@ -0,0 +1,432 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/framework/details/inplace_op_pass.h" +#include +#include +#include +#include +#include +#include +#include +#include +#include "paddle/fluid/framework/details/memory_optimize_pass.h" +#include "paddle/fluid/framework/ir/graph_helper.h" +#include "paddle/fluid/framework/op_info.h" + +// NOTE(dzhwinter): inplace means one op output variable reuse the input space. +// By our design, one operator only can read its input(const Variable), +// write its output(non-const Variable). If one operator is inplaced, means +// user have chance to write the space before reading happens. +// Especially when some optimize code writing style is applied. +// +// +// /* wrong case in operator */ +// /*In this case, a larger allocation is allocated, input content is lost*/ +// const Tensor* in = ctx.Input("In") +// Tensor* out = ctx.Output("Out"); +// auto* out_ptr = out->mutable_data(ctx.GetPlace()); +// out_ptr[0] = 0; // input contect is overwrited. + +// NOTE(dzhwinter): +// Only for backward compacity and stable. if enable_inplace_whitelist is turn +// on. +// only the ops in whitelist will be use inplace strategy. +// if not, all the op will be inplaced if it registered with InplaceClass +DEFINE_bool( + enable_inplace_whitelist, false, + "If this option turns on, only these op in whitelist can be inplaced." + "If it turns off, all of the running op can be candidate of inplaced op." + "Such as scale, elementwise_add" + "By default, it's turned off"); + +DECLARE_string(memory_optimize_debug); + +// clang-format off +const std::string kInplacedOpWhiteList[] = { // NOLINT + "sigmoid", + "exp", + "relu", + "tanh", + "sqrt", + "ceil", + "floor", + "reciprocal", + "relu6", + "soft_relu", + "hard_sigmoid", + "batch_norm", + "batch_norm_grad", + "sum", + "sum_grad", + "scale", + "reshape", + "elementwise_add", + "elementwise_add_grad", +}; +// clang-format on + +namespace paddle { +namespace framework { +namespace details { + +static inline ir::Node* GetNextCascadeInplacedVar(ir::Node* var) { + // if next op is inplaced, then return the output var + // otherwise return nullptr + PADDLE_ENFORCE(var && var->IsVar() && !var->IsCtrlVar()); + ir::Node* inplaced_var = nullptr; + for (auto* next_op : var->outputs) { + for (auto* output : next_op->outputs) { + if (output->IsVar() && !output->IsCtrlVar() && + output->Name() == var->Name()) { + inplaced_var = output; + } + } + } + return inplaced_var; +} + +static inline ir::Node* GetPrevCascadeInplacedVar(ir::Node* var) { + PADDLE_ENFORCE(var && var->IsVar() && !var->IsCtrlVar()); + if (var->inputs.empty()) return nullptr; + auto* prev_op = var->inputs.at(0); + auto input_it = std::find_if(prev_op->inputs.begin(), prev_op->inputs.end(), + [&](ir::Node* node) { + if (node->IsVar() && !node->IsCtrlVar() && + node->Name() == var->Name()) { + return true; + } else { + return false; + } + }); + return input_it == prev_op->inputs.end() ? nullptr : *input_it; +} + +InplacePass::InplacePass() : Pass() { + if (FLAGS_enable_inplace_whitelist) { + for (auto& s : kInplacedOpWhiteList) { + whitelist_.emplace(s); + } + } +} + +void InplacePass::InitSSAGraphNodes() const { + std::unordered_map> all_vars; + for (auto* op : view_.AllOps()) { + for (auto* node : op->inputs) { + if (!node->IsVar() || node->IsCtrlVar()) continue; + if (all_vars[node->Name()].count(node) == 0) { + all_vars[node->Name()].emplace(node); + var_nodes_[node->Name()].emplace_back(node); + } + } + for (auto* node : op->outputs) { + if (!node->IsVar() || node->IsCtrlVar()) continue; + if (all_vars[node->Name()].count(node) == 0) { + all_vars[node->Name()].emplace(node); + var_nodes_[node->Name()].emplace_back(node); + } + } + } +} + +std::unique_ptr InplacePass::ApplyImpl( + std::unique_ptr graph) const { + var_nodes_.clear(); + view_.Build(graph.get()); + InitSSAGraphNodes(); + + for (auto* op : view_.AllOps()) { + if (FLAGS_enable_inplace_whitelist && !whitelist_.count(op->Name())) + continue; + TryInplaceOpInputOutput(op, graph.get()); + } + graph->ResolveHazard(var_nodes_); + + return graph; +} + +void InplacePass::InplaceModifyDesc(const std::string& var, + const std::string& cache_var, + const size_t& idx) const { + for (size_t i = idx; i < view_.AllOps().size(); ++i) { + ir::Node* op = view_.AllOps()[i]; + PADDLE_ENFORCE(op->IsOp() && op->Op()); + auto* op_desc = op->Op(); + op_desc->RenameInput(var, cache_var); + op_desc->RenameOutput(var, cache_var); + if (op_desc->Block()->HasVar(var)) op_desc->Block()->RemoveVar(var); + op_desc->Flush(); + } +} + +const NodeSwapQueue InplacePass::TryInplaceModifyVar( + const std::string& var, const std::string& cache_var, const size_t& idx, + ir::Graph* graph) const { + PADDLE_ENFORCE(var_nodes_[var].size() >= 1 && + var_nodes_[var].at(0)->Var() != nullptr); + std::unique_ptr var_desc(new VarDesc(*var_nodes_[var].at(0)->Var())); + var_desc->SetName(cache_var); + + NodeSwapQueue swap_nodes; + + for (size_t i = idx; i < view_.AllOps().size(); ++i) { + auto* op = view_.AllOps()[i]; + + // redirect the input to the latest version of cache_var + for (auto* node : op->inputs) { + if (node->Name() == var) { + ir::Node* cache_node = graph->CreateVarNode(var_desc.get()); + + // swap node to cache_node + cache_node->outputs.insert(cache_node->outputs.end(), + node->outputs.begin(), node->outputs.end()); + PADDLE_ENFORCE(node->inputs.size() == 1 && node->inputs[0]->IsOp()); + auto* prev_op = node->inputs[0]; + std::replace(prev_op->outputs.begin(), prev_op->outputs.end(), node, + cache_node); + cache_node->inputs.emplace_back(prev_op); + for (auto* next_op : node->outputs) { + std::replace(next_op->inputs.begin(), next_op->inputs.end(), node, + cache_node); + } + + swap_nodes.emplace_back(std::make_pair(node, cache_node)); + } + } + + // if we need to rename the output, + // always create a newer version of cache_var + for (auto* node : op->outputs) { + if (node->Name() == var) { + ir::Node* cache_node = graph->CreateVarNode(var_desc.get()); + // swap node to cache node + cache_node->outputs.insert(cache_node->outputs.end(), + node->outputs.begin(), node->outputs.end()); + cache_node->inputs.emplace_back(op); + std::replace(op->outputs.begin(), op->outputs.end(), node, cache_node); + for (auto* next_op : node->outputs) { + std::replace(next_op->inputs.begin(), next_op->inputs.end(), node, + cache_node); + } + + swap_nodes.emplace_back(std::make_pair(node, cache_node)); + } + } + } + + return swap_nodes; +} + +void InplacePass::CommitModify(const NodeSwapQueue& swap_nodes, + ir::Graph* graph) const { + for (auto& pair : swap_nodes) { + auto *node = pair.first, *cache_node = pair.second; + const std::string var = node->Name(), cache_var = cache_node->Name(); + var_nodes_[cache_var].emplace_back(cache_node); + graph->RemoveNode(node); + auto& nodes = var_nodes_.at(var); + // release unused var in graph. Because python side memory optimize + // may reused the var in same name, so we only clear the var node + // after current inplaced index. + nodes.erase(std::remove(nodes.begin(), nodes.end(), node), nodes.end()); + } +} + +void InplacePass::WithdrawModify(const NodeSwapQueue& nodes, + ir::Graph* graph) const { + for (auto& pair : nodes) { + auto *node = pair.first, *cache_node = pair.second; + const std::string var = node->Name(), cache_var = cache_node->Name(); + auto* prev_op = node->inputs[0]; + std::replace(prev_op->outputs.begin(), prev_op->outputs.end(), cache_node, + node); + for (auto* next_op : node->outputs) { + std::replace(next_op->inputs.begin(), next_op->inputs.end(), cache_node, + node); + } + graph->RemoveNode(cache_node); + } +} + +void InplacePass::TryInplaceOpInputOutput(ir::Node* op, + ir::Graph* graph) const { + VLOG(4) << "Try to inplace op " << op->Name(); + PADDLE_ENFORCE(op->Op() != nullptr && op->Op()->Block() != nullptr, + "op_desc is nullptr"); + // some pre-requirments need to meet if the op want to inplaced. + + auto* op_desc = op->Op(); + auto& infer_inplace = + OpInfoMap::Instance().Get(op_desc->Type()).infer_inplace_; + + // 1. infer_inplace_ is registered. + if (!static_cast(infer_inplace)) return; + PADDLE_ENFORCE(static_cast(infer_inplace), + "%s's infer_inplace has not been registered", op_desc->Type()); + + auto* block = op_desc->Block(); + auto in_to_outs = infer_inplace(*op_desc, block); + + auto& all_ops = view_.AllOps(); + auto cursor = std::find(all_ops.begin(), all_ops.end(), op); + size_t idx = std::distance(all_ops.begin(), cursor); + + for (auto& pair : in_to_outs) { + auto& in_var_name = pair.first; + auto& out_var_name = pair.second; + auto* in_node = view_.GetNodeByName(in_var_name, op->inputs); + auto* out_node = view_.GetNodeByName(out_var_name, op->outputs); + + // 2. there is no external pending op on the input node + if (view_.PendingOpsOnVar(in_node).size() > 1) { + VLOG(4) << string::Sprintf( + "Skiped pair %s => %s. %s input has external dependency." + "inplace such pair will overwrite the memory.", + out_var_name, in_var_name, op->Name()); + continue; + } + + // 3. if output has been memory optimize by python(fluid.memory_optmize()). + // this candidate can not be inplaced. Will be deprecated in the future. + if (view_.InSkipSet(out_node->Name())) { + VLOG(4) << string::Sprintf( + "Skiped %s => %s reused previous memory block in python memory " + "optmize," + "it inplace may generate a circle", + out_var_name, in_var_name, op->Name()); + continue; + } + + // Debug Interface. Which would be skipped by the pass. + if (out_node->Name() == FLAGS_memory_optimize_debug) { + VLOG(3) << "Skiped var by force. FLAGS_memory_optimize_debug=" + << out_node->Name(); + continue; + } + + // NOTE(dzhwinter): + // two stage commit of inplaced process. if after inplace happens generate a + // circle, + // then withdraw the changes. Otherwise, safely add the node. + auto swap_nodes = + TryInplaceModifyVar(out_var_name, in_var_name, idx, graph); + + if (!ir::HasCircle(*graph)) { + VLOG(3) << string::Sprintf("!!! %s, %s => %s inplaced", op->Name(), + out_var_name, in_var_name); + InplaceModifyDesc(out_var_name, in_var_name, idx); + CommitModify(swap_nodes, graph); + } else { + VLOG(3) << string::Sprintf( + "Skiped pair %s => %s, inplace will generate a circle. withdraw %s", + out_var_name, in_var_name, op->Name()); + WithdrawModify(swap_nodes, graph); + } + } +} + +ir::Node* GraphView::GetNodeByName(const std::string& name, + const std::vector& nodes) const { + // nodes should be op->inputs/outputs + // node in same node do have different name. + std::unordered_set nodes_in_op; + bool has_dup_node = + std::all_of(nodes.begin(), nodes.end(), [&nodes_in_op](ir::Node* node) { + if (!node->IsVar() || node->IsCtrlVar() || node->Var() == nullptr) { + if (nodes_in_op.count(node->Name())) return true; + nodes_in_op.emplace(node->Name()); + } + return false; + }); + PADDLE_ENFORCE(has_dup_node == false, "nodes has same name!"); + ir::Node* node = nullptr; + for (auto* it : nodes) { + if (!it->IsVar() || it->IsCtrlVar() || it->Var() == nullptr) continue; + if (it->Name() == name) { + node = it; + break; + } + } + PADDLE_ENFORCE(node != nullptr, + string::Sprintf("Not found var %s in nodes!", name)); + return node; +} + +std::vector GraphView::PendingOpsOnVar(ir::Node* node) { + // get the pending ops depends on same var node. + // because node also maybe a inplaced variable, so need to backtrack all the + // previous inplaced vars. + std::vector pending_ops; + ir::Node* p = node; + while (p != nullptr) { + pending_ops.insert(pending_ops.end(), p->outputs.begin(), p->outputs.end()); + p = GetPrevCascadeInplacedVar(p); + } + return pending_ops; +} + +void GraphView::Build(ir::Graph* g) { + // track the var nodes in correct order. + // Because we insert some new created node. Which may have data race between + // nodes. + // resolve data harzards depends on the var nodes in right order. + ops_ = SortOpLikeDescOrder(*g); + + // 1. track the nodes which reused previous node in Python memory optimize. + // these node can not be inplaced, otherwise may generate a circle in graph. + std::unordered_set all_vars; + for (auto& node : g->Nodes()) { + if (node->IsVar()) continue; + for (auto& out : node->outputs) { + if (out->IsCtrlVar() || out->Var() == nullptr) continue; + if (all_vars.count(out->Name())) { + dup_nodes_.emplace(out->Name()); + } else { + all_vars.emplace(out->Name()); + } + } + } + + // 2. track the nodes which used by parameter server. + // these node can not be inplaced, otherwise trainer + // pserver can not find each other name. + auto update_skip_set = [&](ir::Node* node) { + for (auto& in : node->inputs) { + if (in->IsVar() && in->Var() != nullptr) dup_nodes_.emplace(in->Name()); + } + for (auto& out : node->outputs) { + if (out->IsVar() && out->Var() != nullptr) + dup_nodes_.emplace(out->Name()); + } + }; + for (auto& node : g->Nodes()) { + if (!node->IsOp()) continue; + if (node->Name() == "send") update_skip_set(node); + if (node->Name() == "recv") update_skip_set(node); + if (node->Name() == "prefetch") update_skip_set(node); + } +} + +const std::vector& GraphView::AllOps() { return ops_; } + +bool GraphView::InSkipSet(const std::string& var) const { + return dup_nodes_.count(var); +} + +} // namespace details +} // namespace framework +} // namespace paddle + +REGISTER_PASS(inplace_pass, paddle::framework::details::InplacePass); diff --git a/paddle/fluid/framework/details/inplace_op_pass.h b/paddle/fluid/framework/details/inplace_op_pass.h new file mode 100644 index 0000000000000000000000000000000000000000..7be7f311852d2b64ce95e1a939371760d03d296b --- /dev/null +++ b/paddle/fluid/framework/details/inplace_op_pass.h @@ -0,0 +1,94 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may abtain 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/fluid/framework/details/memory_optimize_helper.h" +#include "paddle/fluid/framework/ir/graph.h" +#include "paddle/fluid/framework/ir/pass.h" + +namespace paddle { +namespace framework { +namespace details { + +class GraphView { + public: + GraphView() = default; + + void Build(ir::Graph* g); + + const std::vector& AllOps(); + + ir::Node* GetNodeByName(const std::string& name, + const std::vector& nodes) const; + + std::vector PendingOpsOnVar(ir::Node* var); + + // Will Deperated in the future. + // NOTE(dzhwinter) : + // 1. Python memory optimize will reuse + // memory based var name, so different op output may + // have the same variable name. enable inplace on such node + // will generate a circle in ssa graph. + // 2. DistributeTranspiler will use unique name to + // map the parameter and gradient, must be skipped. + bool InSkipSet(const std::string& var) const; + + private: + std::vector ops_; + std::unordered_set dup_nodes_; // mem opt affect nodes + std::map> adj_list_; +}; + +// swap pairs in sequence +typedef std::vector> NodeSwapQueue; +class InplacePass : public ir::Pass { + public: + InplacePass(); + + protected: + std::unique_ptr ApplyImpl( + std::unique_ptr graph) const override; + + void InitSSAGraphNodes() const; + + private: + const NodeSwapQueue TryInplaceModifyVar(const std::string& var, + const std::string& cache_var, + const size_t& idx, + ir::Graph* graph) const; + + void CommitModify(const NodeSwapQueue&, ir::Graph* graph) const; + + void WithdrawModify(const NodeSwapQueue& nodes, ir::Graph* graph) const; + + void InplaceModifyDesc(const std::string& in_var, const std::string& out_var, + const size_t& idx) const; + + void TryInplaceOpInputOutput(ir::Node* op, ir::Graph* graph) const; + + mutable std::map> var_nodes_; + + mutable std::unordered_set whitelist_; + mutable GraphView view_; +}; + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/memory_early_delete_pass.cc b/paddle/fluid/framework/details/memory_early_delete_pass.cc deleted file mode 100644 index 5906b7d57ce122520a4594f1528e00982eaa1a7f..0000000000000000000000000000000000000000 --- a/paddle/fluid/framework/details/memory_early_delete_pass.cc +++ /dev/null @@ -1,117 +0,0 @@ -// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -#include "paddle/fluid/framework/details/memory_early_delete_pass.h" -#include -#include -#include -#include "paddle/fluid/framework/details/memory_reuse_types.h" -#include "paddle/fluid/framework/details/multi_devices_helper.h" -#include "paddle/fluid/framework/details/reference_count_pass_helper.h" -#include "paddle/fluid/framework/ir/graph_helper.h" - -namespace paddle { -namespace framework { -namespace details { - -static ComputationOpHandle* FindNextComputationOpHandle(VarHandle* var_in) { - std::queue queue; - queue.push(var_in); - do { - auto* var = queue.front(); - queue.pop(); - for (auto* op : var->PendingOps()) { - auto* compute_op = dynamic_cast(op); - if (compute_op != nullptr && compute_op->GetPlace() == var_in->place()) { - return compute_op; - } - for (auto* out_var : op->Outputs()) { - queue.push(out_var); - } - } - } while (!queue.empty()); - return nullptr; -} - -std::unique_ptr MemoryEarlyDeletePass::ApplyImpl( - std::unique_ptr graph) const { - auto& graph_pool = Get(kGraphNodePool); - auto& gcs = Get(kGarbageCollector); - - std::unordered_map> unlived_vars; - unlived_vars.reserve(graph_pool.size()); - for (auto& pair : graph_pool) { - unlived_vars.insert(std::make_pair(pair.first, pair.second)); - } - - auto compare_and_insert_early_delete_op = [&]( - OpHandleBase* op, const std::vector& vars) { - if (unlived_vars.empty()) return; - // unlived vars can be deleted after the last used op has finished. - auto* compute_op = dynamic_cast(op); - const auto& places = Get>(kAllPlaces); - for (auto& var : vars) { - auto* var_handle = dynamic_cast(var); - auto var_name = var->Node()->Name(); - auto& var_place = var_handle->place(); - if (unlived_vars.count(var_name) == 0) continue; - if (!unlived_vars[var_name].empty()) { - if (compute_op != nullptr && - unlived_vars[var_name].count(compute_op->Node()->Op()) != 0) { - unlived_vars[var_name].erase(compute_op->Node()->Op()); - } - continue; - } - - if (var_handle == nullptr || !var_handle->Node()->IsVar() || - var_handle->Node()->IsCtrlVar()) - continue; - - // shameless copyed from reference count pass. - if (compute_op == nullptr) { - // use next computation op scope - compute_op = FindNextComputationOpHandle(var_handle); - } - auto* early_delete_node = - graph->CreateEmptyNode("early_delete", ir::Node::Type::kOperation); - GarbageCollector* gc = gcs.at(places[compute_op->GetScopeIdx()]).get(); - auto* early_delete_handle = new EarlyDeleteOpHandle( - early_delete_node, compute_op->GetScope(), var_place, {var_name}, gc); - if (compute_op->Outputs().empty()) { - auto* dep_var = new DummyVarHandle(graph->CreateControlDepVar()); - compute_op->AddOutput(dep_var); - graph->Get(kGraphDepVars).emplace(dep_var); - } - early_delete_handle->AddInput(compute_op->Outputs().front()); - VLOG(5) << "Add early delete op " << var_name << " to Operator" - << compute_op->Name(); - } - }; - - auto all_ops = ir::FilterByNodeWrapper(*graph); - for (auto& op : all_ops) { - compare_and_insert_early_delete_op(op, op->Inputs()); - compare_and_insert_early_delete_op(op, op->Outputs()); - } - return graph; -} - -} // namespace details -} // namespace framework -} // namespace paddle - -REGISTER_PASS(memory_early_delete_pass, - paddle::framework::details::MemoryEarlyDeletePass) - .RequireGraphAttr(paddle::framework::details::kGraphNodePool) - .RequireGraphAttr(paddle::framework::details::kGarbageCollector); diff --git a/paddle/fluid/framework/details/memory_optimize_helper.cc b/paddle/fluid/framework/details/memory_optimize_helper.cc new file mode 100644 index 0000000000000000000000000000000000000000..db4e805bb692ee44ac50337fae54f8dbfe389e6f --- /dev/null +++ b/paddle/fluid/framework/details/memory_optimize_helper.cc @@ -0,0 +1,540 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/framework/details/memory_optimize_helper.h" +#include +#include +#include +#include +#include +#include +#include +#include "paddle/fluid/framework/var_desc.h" +#include "paddle/fluid/platform/cpu_info.h" + +#ifdef PADDLE_WITH_CUDA +#include "paddle/fluid/platform/gpu_info.h" +#endif // PADDLE_WITH_CUDA + +namespace paddle { +namespace framework { +namespace details { +using paddle::framework::VarDesc; + +std::vector SortOpLikeDescOrder(const ir::Graph& graph) { + PADDLE_ENFORCE(graph.Has(kAllOpDescs), + "Graph has no attribute of kAllOpDescs."); + // 1. get op desc order + auto& op_descs = graph.Get>(kAllOpDescs); + + // 2. topology sort order + auto nodes = graph.Nodes(); + std::deque ops; + FilterVariables(nodes, [&](ir::Node* op) { + if (op->IsOp() && op->Op() != nullptr) { + ops.emplace_back(op); + } + }); + std::unordered_map op_deps; + std::list ready_ops; + std::unordered_map> pending_ops; + + for (auto* op : ops) { + std::unordered_set preceding_op; + for (auto* in : op->inputs) { + if (in->inputs.empty()) continue; + PADDLE_ENFORCE(in->inputs.size() == 1 && in->inputs[0]->IsOp()); + preceding_op.emplace(in->inputs[0]); + pending_ops[in->inputs[0]].emplace(op); + } + op_deps[op] = preceding_op.size(); + if (preceding_op.empty()) { + ready_ops.emplace_back(op); + } + } + + // 3. generated op list based desc order and the topology order + std::vector ret; + std::list op_descs_list(op_descs.begin(), op_descs.end()); + + auto update_by_found_node = [&](ir::Node* found_node) { + for (auto* pending_op : pending_ops[found_node]) { + if (--op_deps[pending_op] == 0) { + ready_ops.emplace_back(pending_op); + } + } + ready_ops.remove(found_node); + ret.emplace_back(found_node); + }; + + while (!ready_ops.empty()) { + bool all_of_ready_op_unmatched = true; + for (auto it = op_descs_list.begin(); it != op_descs_list.end();) { + auto op_desc = *it; + ir::Node* found_node = nullptr; + for (auto* op : ready_ops) { + if (IsSameDesc(op->Op(), op_desc)) { + found_node = op; + break; + } + } + + // 3.1 op desc deleted by other pass + if (found_node == nullptr) { + ++it; + continue; + } else { + all_of_ready_op_unmatched = false; + it = op_descs_list.erase(it); + } + update_by_found_node(found_node); + } + + // 3.2 op descs are added by other pass + // preceding op non empty means some new op descs are + // created, but not contained in return node list. + // these new op desc may depend on each other. + std::list prev_ready_ops(ready_ops); + if (all_of_ready_op_unmatched) { + for (auto op : prev_ready_ops) { + update_by_found_node(op); + } + } + } + + PADDLE_ENFORCE(std::all_of( + op_deps.begin(), op_deps.end(), + [&](const std::pair& p) { return p.second == 0; })); + + return ret; +} + +size_t NodeSize(const VarDesc& node) { + auto shape = node.GetShape(); + int size = + std::accumulate(shape.begin(), shape.end(), 1, std::multiplies()); + size_t type_size = SizeOfType(node.GetDataType()); + return type_size * std::abs(size); +} + +size_t NodeSize(ir::Node* n) { + VarDesc* desc = nullptr; + // some op do not have block pointer + if (n->inputs[0]->Op() != nullptr) { + desc = FindVarDescInBlock(n); + } else { + desc = n->Var(); + } + return NodeSize(*desc); +} + +std::string DebugStringImpl(VarDesc* var) { + std::stringstream ss; + ss << var->Name(); + ss << "["; + try { + auto shape = var->GetShape(); + for (size_t i = 0; i < shape.size(); ++i) { + if (i != shape.size() - 1) { + ss << shape[i] << ","; + } else { + ss << shape[i]; + } + } + ss << "]"; + } catch (...) { + ss << "Var has no VarDesc !!! Name:" << var->Name(); + } + return ss.str(); +} + +std::string DebugString(ir::Node* var) { + return DebugStringImpl(FindVarDescInBlock(var)); +} + +// NOTE(dzh): based ir node, if a large node has been reused +// by a small size node, then next time it appear in pool, it will +// have the small size. Find the original node shap from blockdesc. +VarDesc* FindVarDescInBlock(ir::Node* n) { + PADDLE_ENFORCE(n->IsVar() && !n->IsCtrlVar() && n->inputs.size() == 1); + BlockDesc* block = n->inputs[0]->Op()->Block(); + PADDLE_ENFORCE(block->HasVar(n->Name()), + string::Sprintf("Block do not has var %s", n->Name())); + return block->FindVar(n->Name()); +} + +struct NodeComparator { + bool operator()(ir::Node* lhs, ir::Node* rhs) const { + auto* lhs_desc = FindVarDescInBlock(lhs); + auto* rhs_desc = FindVarDescInBlock(rhs); + // match data type + if (lhs_desc->GetDataType() != rhs_desc->GetDataType()) { + return false; + } + // match shape + auto lhs_shape = lhs_desc->GetShape(); + auto rhs_shape = rhs_desc->GetShape(); + if ((lhs_shape[0] == -1 && rhs_shape[0] == -1) || + (lhs_shape[0] != -1 && rhs_shape[0] != -1)) { + return NodeSize(lhs) <= NodeSize(rhs); + } else { + return false; + } + } +}; + +void OrderedSet::Insert(ir::Node* var) { + PADDLE_ENFORCE(var->IsVar() && !var->IsCtrlVar()); + if (mark_table_.count(var->Name()) != 0) { + mark_table_[var->Name()]->emplace_back(var); + return; + } + + auto* var_desc = FindVarDescInBlock(var); + auto var_shape = var_desc->GetShape(); + int batch_size = static_cast(var_shape[0]); + + NodeComparator functor; + Iter it = nodes_.begin(); + while (it != nodes_.end()) { + auto& prev = it->front(); + auto* cache_desc = FindVarDescInBlock(prev); + int cache_batch_size = cache_desc->GetShape()[0]; + if ((cache_batch_size == -1 && batch_size == -1) || + (cache_batch_size != -1 && batch_size != -1)) { + if (functor(prev, var)) { + ++it; + } else { + break; + } + } else if (cache_batch_size == -1 && batch_size != -1) { + ++it; + } else if (cache_batch_size != -1 && batch_size == -1) { + break; + } + } + + it = nodes_.insert(it, {var}); + mark_table_[var->Name()] = it; +} + +int OrderedSet::GetNodeIndexInPool(ir::Node* var) { + return std::distance(nodes_.begin(), mark_table_[var->Name()]); +} + +ir::Node* OrderedSet::FindBestFitNode(ir::Node* var) const { + ir::Node* found_node = nullptr; + NodeComparator functor; + + for (auto it = nodes_.begin(); it != nodes_.end(); ++it) { + auto& candidate = it->front(); + if (functor(var, candidate)) { + found_node = candidate; + break; + } + } + return found_node; +} + +ir::Node* OrderedSet::FindNextBestFitNode(ir::Node* var, ir::Node* prev) const { + ir::Node* found_node = nullptr; + NodeComparator functor; + auto it = + std::find_if(nodes_.begin(), nodes_.end(), [&](const NodeVector& v) { + if (v.front() == prev) + return true; + else + return false; + }); + PADDLE_ENFORCE(it != nodes_.end(), "Not found previous in node list!"); + for (it = std::next(it); it != nodes_.end(); ++it) { + auto& candidate = it->front(); + if (functor(var, candidate)) { + found_node = candidate; + break; + } + } + return found_node; +} + +bool OrderedSet::Has(ir::Node* var) const { + if (mark_table_.count(var->Name())) { + auto& node_in_samename = mark_table_.at(var->Name()); + auto iter = + std::find_if(node_in_samename->begin(), node_in_samename->end(), + [&](ir::Node* n) { return n->Name() == var->Name(); }); + return iter != node_in_samename->end(); + } + return false; +} + +void OrderedSet::Erase(const std::string& var) { + PADDLE_ENFORCE(mark_table_.count(var)); + nodes_.erase(mark_table_[var]); + mark_table_.erase(var); +} + +void OrderedSet::Erase(ir::Node* var) { + PADDLE_ENFORCE(var != nullptr); + Erase(var->Name()); +} + +std::string OrderedSet::ToString() const { + std::stringstream ss; + for (auto it = nodes_.begin(); it != nodes_.end(); ++it) { + for (auto& node : *it) { + ss << DebugString(node) << " "; + } + } + return ss.str(); +} + +bool NodeCanReused(ir::Node* node) { + // valid the node is a var node + if (node == nullptr || !node->IsVar() || node->IsCtrlVar()) return false; + + bool flag = true; + // op output force generated in cpu, can not be reused. + for (auto* op : node->inputs) { + if (op->Op()->HasAttr("force_cpu")) { + flag &= framework::AttrReader(op->Op()->GetAttrMap()) + .Get("force_cpu") == 0; + } + } + // var desc validation. + flag &= NodeCanReused(*node->Var()); + return flag; +} + +int MinChunkSize() { + int size{0}; +#ifdef PADDLE_WITH_CUDA + size = platform::GpuMinChunkSize(); +#else + size = platform::CpuMinChunkSize(); +#endif // PADDLE_WITH_CUDA + return size; +} + +bool NodeCanReused(const VarDesc& node) { + auto type = node.GetType(); + // only these types holds bulk of gpu memory + if (!(type == proto::VarType::LOD_TENSOR || + type == proto::VarType::SELECTED_ROWS || + type == proto::VarType::LOD_TENSOR_ARRAY)) { + return false; + } + // persistable variable is parameter + if (node.Persistable()) { + return false; + } + // shape < min_chunk_size is meaningless. + // further more, fetched loss always has size = 1 + // which should not be reused. + auto shape = node.GetShape(); + int size = std::abs( + std::accumulate(shape.begin(), shape.end(), 1, std::multiplies())); + if (shape.empty() || size < MinChunkSize()) { + return false; + } + // vars can be @EMPTY@, @LR_DECAY_REUSE_ID@. For example, while_grad + std::string name = node.Name(); + if (!name.empty() && name[0] == '@' && name[name.size() - 1] == '@') + return false; + return true; +} + +bool OpHasSubBlock(OpDesc* desc) { + const AttributeMap& attrs = desc->GetAttrMap(); + for (auto& attr : attrs) { + if (attr.second.type() == typeid(BlockDesc*) || // NOLINT + attr.second.type() == typeid(std::vector)) // NOLINT + return true; + } + return false; +} + +ControlFlowGraph::ControlFlowGraph(const ir::Graph& graph) { + ops_ = SortOpLikeDescOrder(graph); + ConnectNodes(); +} + +void ControlFlowGraph::BuildCFGGraph() { + // FIXME(dzh): same effect with ConnectNodes, but use the control + // link to build dependency graph, it goes wrong in transformer. + for (ir::Node* op : ops_) { + for (auto& input_var : op->inputs) { + if (!input_var->inputs.empty()) { + PADDLE_ENFORCE( + input_var->inputs.size() == 1 && input_var->inputs[0]->IsOp(), + "Preceding Op Node of Var Node must be unique"); + auto* pred_op = input_var->inputs[0]; + if (pred_op->Op() != nullptr) { + predecessors_[op].insert(pred_op); + successors_[pred_op].insert(op); + } + } + if (input_var->IsVar() && !input_var->IsCtrlVar()) { + uses_[op].insert(input_var->Name()); + } + } + for (auto& output_var : op->outputs) { + // output var may be used by many op + for (auto* succ_op : output_var->outputs) { + if (succ_op->Op() != nullptr) { + successors_[op].insert(succ_op); + predecessors_[succ_op].insert(op); + } + } + if (output_var->IsVar() && !output_var->IsCtrlVar()) { + defs_[op].insert(output_var->Name()); + } + } + } +} + +void ControlFlowGraph::ConnectNodes() { + for (size_t i = 0; i < ops_.size(); ++i) { + auto& op = ops_[i]; + try { + auto& next_op = ops_.at(i + 1); + successors_[op].insert(next_op); + predecessors_[next_op].insert(op); + } catch (...) { + // do nothing + } + + FilterVariables(op->inputs, + [&](ir::Node* var) { uses_[op].emplace(var->Name()); }); + + FilterVariables(op->outputs, + [&](ir::Node* var) { defs_[op].emplace(var->Name()); }); + } +} + +void ControlFlowGraph::LiveVariableAnalysis() { + // NOTE(dzh): variable liveless analysis (a.k.a reversed_ops algorithm) + // compute the liveness of for each variable though reversed_ops algorithm. + // It iterates the operators from end to begin, compute the live in/live out + // variable set for each op, then the diff between in/out will be used for + // the variable reuse. For detail refer to + // http://www.cs.cornell.edu/courses/cs4120/2013fa/lectures/lec26-fa13.pdf + std::list work_list(ops_.rbegin(), ops_.rend()); + while (!work_list.empty()) { + ir::Node* op = work_list.front(); + work_list.pop_front(); + // get the live_in calculated before. Empty if first. + auto prev_live_in = std::move(live_in_[op]); + for (auto& s : successors_[op]) { + for (auto& var : live_in_[s]) { + live_out_[op].insert(var); + } + } + for (auto& var : uses_[op]) { + live_in_[op].insert(var); + } + for (auto& var : live_out_[op]) { + live_in_[op].insert(var); + } + for (auto& var : defs_[op]) { + live_in_[op].erase(var); + } + + // If the live_in is not changed, then the liveness analysis of + // predecessors is completed. + // + // Otherwise, recalculate the predecessors liveness + if (live_in_[op] != prev_live_in) { + for (auto& pre : predecessors_[op]) { + work_list.push_back(pre); + } + } + } +} + +void ControlFlowGraph::RenameVarInCFGGraph(const std::string& old_node, + const std::string& new_node, + int begin_idx) { + // update graph from begin idx to the end + for (size_t i = begin_idx; i != ops_.size(); ++i) { + auto* op = ops_[i]; + if (uses_[op].find(old_node) != uses_[op].end()) { + uses_[op].erase(old_node); + uses_[op].insert(new_node); + } + if (defs_[op].find(old_node) != defs_[op].end()) { + defs_[op].erase(old_node); + defs_[op].insert(new_node); + } + if (live_in_[op].find(old_node) != live_in_[op].end()) { + live_in_[op].erase(old_node); + live_in_[op].insert(new_node); + } + if (live_out_[op].find(old_node) != live_out_[op].end()) { + live_out_[op].erase(old_node); + live_out_[op].insert(new_node); + } + } +} + +const std::set ControlFlowGraph::LiveIn(ir::Node* op) const { + auto it = live_in_.find(op); + PADDLE_ENFORCE( + it != live_in_.end(), + string::Sprintf("Expect %s in live_in, but Not Found.", op->Name())); + return it->second; +} + +const std::set ControlFlowGraph::LiveOut(ir::Node* op) const { + auto it = live_out_.find(op); + PADDLE_ENFORCE( + it != live_out_.end(), + string::Sprintf("Expect %s in live_out, but Not Found.", op->Name())); + return it->second; +} + +const std::set ControlFlowGraph::Use(ir::Node* op) const { + auto it = uses_.find(op); + PADDLE_ENFORCE( + it != uses_.end(), + string::Sprintf("Expect %s in live_out, but Not Found.", op->Name())); + return it->second; +} + +const std::vector ControlFlowGraph::Ops() const { return ops_; } + +std::vector& ControlFlowGraph::Ops() { return ops_; } + +ir::Node* ControlFlowGraph::GetNodeByName(const std::string& name, + ir::Node* op) const { + // in ssa-graph, different version nodes have same name, + // this function get the latest version var before target op + // It may return nullptr, such as data node. + ir::Node* found_node = nullptr; + for (auto* node : ops_) { + if (node == op) break; + for (auto& output : node->outputs) { + PADDLE_ENFORCE((output != nullptr && output->IsVar()), + "Output is empty!"); + if (output->Var() && output->Name() == name) { + found_node = output; + } + } + } + return found_node; +} + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/memory_optimize_helper.h b/paddle/fluid/framework/details/memory_optimize_helper.h new file mode 100644 index 0000000000000000000000000000000000000000..377367faf3c529496b00004f23159750cc2e4bc4 --- /dev/null +++ b/paddle/fluid/framework/details/memory_optimize_helper.h @@ -0,0 +1,182 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include "paddle/fluid/framework/data_type.h" +#include "paddle/fluid/framework/ir/graph.h" + +namespace paddle { +namespace framework { +namespace details { + +std::vector SortOpLikeDescOrder(const ir::Graph& graph); + +// NOTE(dzh): A ordered set for node reuse in memory optimize. +// the orderedset sort node in ascend order(by node bytes size). +// in fluid, -1 means the batch_size, which is determined in runtime. +// So the reuse happens between nodes who's batch_size both are -1 +// simultaneously or not. +// +// sort rule: +// rule 0 : smaller node ranking in front. +// rule 1 : batch_size equal -1 ranking in the front than the node not. +// +// For example, +// node0[-1, 1] node1[-1, 1, 1], node2[1,1], node3[1,1024], .. + +class OrderedSet { + public: + // nodes with same name exists in pool. + using NodeVector = std::vector; + using Iter = typename std::list::iterator; + using ConstIter = typename std::list::const_iterator; + + void Insert(ir::Node* var); + void Erase(ir::Node* var); + void Erase(const std::string& var); + bool Has(ir::Node* var) const; + void Clear() { + mark_table_.clear(); + nodes_.clear(); + } + // find the bestfit shape node block with var. + ir::Node* FindBestFitNode(ir::Node* var) const; + ir::Node* FindNextBestFitNode(ir::Node* var, ir::Node* prev) const; + // map store non-const iterator, can not promise const + int GetNodeIndexInPool(ir::Node* var); + // pool all node to string + std::string ToString() const; + + Iter begin() { return nodes_.begin(); } + Iter end() { return nodes_.end(); } + ConstIter begin() const { return nodes_.begin(); } + ConstIter end() const { return nodes_.end(); } + + size_t size() const { return nodes_.size(); } + + private: + // for searching. + std::unordered_map mark_table_; + // node pool + std::list nodes_; +}; + +class ControlFlowGraph { + public: + ControlFlowGraph() = default; + // IR Graph + explicit ControlFlowGraph(const ir::Graph& graph); + + void LiveVariableAnalysis(); + + void RenameVarInCFGGraph(const std::string& old_node, + const std::string& new_node, int begin_idx); + + const std::set LiveIn(ir::Node* op) const; + const std::set LiveOut(ir::Node* op) const; + const std::set Use(ir::Node* op) const; + const std::vector Ops() const; + std::vector& Ops(); + + // for ssa-graph nodes + ir::Node* GetNodeByName(const std::string& name, ir::Node* op) const; + + private: + void BuildCFGGraph(); + void ConnectNodes(); + + using NodeListMap = std::unordered_map>; + using VarSetMap = std::map>; + // successors ops use the output variables. + NodeListMap successors_; + // predecessors ops generated input variables. + NodeListMap predecessors_; + // variables lived before run current op. + VarSetMap live_in_; + // variables lived after run current op. + VarSetMap live_out_; + VarSetMap uses_; // op inputs + VarSetMap defs_; // op outputs + + std::vector ops_; // op sequence by topology sort +}; + +// valid a tensor can be reuse or not +bool NodeCanReused(ir::Node* node); + +// valid a tensor can be reuse or not. +bool NodeCanReused(const VarDesc& node); + +// check op has subblock or not +bool OpHasSubBlock(OpDesc* desc); + +// node memory size in bytes +size_t NodeSize(ir::Node* n); + +// node memory size in bytes +size_t NodeSize(const VarDesc&); + +std::string DebugString(ir::Node* var); + +// NOTE(dzhwinter) +// after node reuse, the replaced node shape is +// different with its VarDesc. So need to find the +// correct VarDesc in Block. +VarDesc* FindVarDescInBlock(ir::Node* n); + +static inline bool IsSameDesc(OpDesc* op1, OpDesc* op2) { + return op1->Type() == op2->Type() && op1->Inputs() == op2->Inputs() && + op1->Outputs() == op2->Outputs(); +} + +template +class FilterVariableImpl { + public: + void operator()(const Container& nodes, Callback callback) { + for (auto* node : nodes) { + callback(node); + } + } +}; + +// filter var node for op->inputs/outputs +template +class FilterVariableImpl, Callback> { + public: + void operator()(const std::vector& nodes, Callback callback) { + for (auto* var : nodes) { + if (var->IsVar() && !var->IsCtrlVar()) { + callback(var); + } + } + } +}; + +template +void FilterVariables(const Container& nodes, Callback callback) { + FilterVariableImpl()(nodes, callback); +} + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/analysis_var_pass_test.cc b/paddle/fluid/framework/details/memory_optimize_helper_test.cc similarity index 75% rename from paddle/fluid/framework/details/analysis_var_pass_test.cc rename to paddle/fluid/framework/details/memory_optimize_helper_test.cc index 9bc4fd33f7058949ca60983ea666a21cb4877b3e..3cfe297a73cf4128b7191cbd432cdceadc6240ec 100644 --- a/paddle/fluid/framework/details/analysis_var_pass_test.cc +++ b/paddle/fluid/framework/details/memory_optimize_helper_test.cc @@ -12,12 +12,18 @@ // See the License for the specific language governing permissions and // limitations under the License. -#include "paddle/fluid/framework/details/analysis_var_pass.h" +#include "paddle/fluid/framework/details/memory_optimize_helper.h" #include #include #include +#include +#include +#include +#include +#include #include "glog/logging.h" #include "gtest/gtest.h" +#include "paddle/fluid/framework/details/graph_test_base.h" #include "paddle/fluid/framework/ir/graph.h" #include "paddle/fluid/framework/ir/graph_helper.h" #include "paddle/fluid/framework/op_registry.h" @@ -26,46 +32,128 @@ namespace paddle { namespace framework { +namespace details { -class DummyOp : public OperatorBase { - public: - DummyOp(const std::string& type, const VariableNameMap& inputs, - const VariableNameMap& outputs, const AttributeMap& attrs) - : OperatorBase(type, inputs, outputs, attrs) {} - - private: - void RunImpl(const Scope& scope, - const platform::Place& place) const override {} -}; - -class SumOpMaker : public OpProtoAndCheckerMaker { - public: - void Make() { - AddInput("X", "").AsDuplicable(); - AddOutput("Out", ""); - AddComment(""); +TEST(OrderedSet, Normal) { + OrderedSet pool; + std::vector> nodes; + + // clang-format off + std::vector> shapes = {{-1, 10}, + {-1, 20}, + {1, 2}, + {5, 2}, + {10, 20}, + {-1, 2, 5}, + {-1, 1, 5}, + {-1, 1}}; + // clang-format on + const int COUNT = shapes.size(); + ProgramDesc prog; + BlockDesc* block_desc = prog.MutableBlock(0); + auto* op_desc = block_desc->AppendOp(); + op_desc->SetType("dummy"); + std::unique_ptr op = ir::CreateNodeForTest(op_desc); + + for (int i = 0; i < COUNT; ++i) { + auto desc = block_desc->Var(std::to_string(i)); + desc->SetShape(shapes[i]); + std::unique_ptr node = ir::CreateNodeForTest(desc); + node->inputs.emplace_back(op.get()); + nodes.emplace_back(std::move(node)); } -}; - -class AssignOpMaker : public OpProtoAndCheckerMaker { - public: - void Make() { - AddInput("X", "").AsDuplicable(); - AddOutput("Out", ""); - AddComment(""); + + // Insert + for (auto& node : nodes) { + pool.Insert(node.get()); } -}; - -class DummyVarTypeInference : public VarTypeInference { - public: - void operator()(const OpDesc& op_desc, BlockDesc* block) const override { - auto& inputs = op_desc.Input("X"); - auto type = block->Var(inputs.front())->GetType(); - auto out_var_name = op_desc.Output("Out").front(); - block->Var(out_var_name)->SetType(type); + + // Has/size + ASSERT_EQ(pool.size(), shapes.size()); + for (auto& node : nodes) { + ASSERT_TRUE(pool.Has(node.get())); } -}; + // assert its order and interface. + std::cout << pool.ToString() << std::endl; + pool.Erase(nodes.front().get()); + std::cout << pool.ToString() << std::endl; + + ASSERT_EQ(pool.size(), static_cast(COUNT - 1)); + ASSERT_EQ(pool.GetNodeIndexInPool(nodes.back().get()), 0); + + { + auto v1 = block_desc->Var("11"); + v1->SetShape({-1, 256, 56, 56}); + std::unique_ptr node1 = ir::CreateNodeForTest(v1); + node1->inputs.emplace_back(op.get()); + auto* cache = pool.FindBestFitNode(node1.get()); + ASSERT_EQ(cache, nullptr); + } + { + auto v2 = block_desc->Var("12"); + v2->SetShape({-1, 2, 5}); + std::unique_ptr node1 = ir::CreateNodeForTest(v2); + node1->inputs.emplace_back(op.get()); + auto* cache = pool.FindBestFitNode(node1.get()); + ASSERT_EQ(pool.GetNodeIndexInPool(cache), 2); // match 6:[-1,2,5] + } + { + auto v3 = block_desc->Var("13"); + v3->SetShape({2, 5}); + std::unique_ptr node1 = ir::CreateNodeForTest(v3); + node1->inputs.emplace_back(op.get()); + auto* cache = pool.FindBestFitNode(node1.get()); + ASSERT_EQ(pool.GetNodeIndexInPool(cache), 5); // match 4:[5,2] + } +} + +TEST(OrderedSet, FindBestFitNode) { + OrderedSet pool; + std::vector> nodes; + ProgramDesc prog; + BlockDesc* block_desc = prog.MutableBlock(0); + auto* op_desc = block_desc->AppendOp(); + op_desc->SetType("dummy"); + std::unique_ptr op = ir::CreateNodeForTest(op_desc); + + { + auto desc = block_desc->Var("a"); + desc->SetShape({128, 128}); + std::unique_ptr node = ir::CreateNodeForTest(desc); + node->inputs.emplace_back(op.get()); + nodes.emplace_back(std::move(node)); + } + { + auto desc = block_desc->Var("b"); + desc->SetShape({128, 129}); + std::unique_ptr node = ir::CreateNodeForTest(desc); + node->inputs.emplace_back(op.get()); + nodes.emplace_back(std::move(node)); + } + { + auto desc = block_desc->Var("c"); + desc->SetShape({128, 128}); + std::unique_ptr node = ir::CreateNodeForTest(desc); + node->inputs.emplace_back(op.get()); + nodes.emplace_back(std::move(node)); + } + + for (auto& node : nodes) { + pool.Insert(node.get()); + } + + // FindNextBestFitNode + auto* n = nodes[0].get(); + auto* cache = pool.FindBestFitNode(n); + PADDLE_ENFORCE(cache->Name() == "a"); + cache = pool.FindNextBestFitNode(n, cache); + PADDLE_ENFORCE(cache->Name() == "c"); + cache = pool.FindNextBestFitNode(n, cache); + PADDLE_ENFORCE(cache->Name() == "b"); +} + +} // namespace details } // namespace framework } // namespace paddle @@ -102,11 +190,6 @@ namespace paddle { namespace framework { namespace details { -static inline bool IsSameDesc(OpDesc* op1, OpDesc* op2) { - return op1->Type() == op2->Type() && op1->Inputs() == op2->Inputs() && - op1->Outputs() == op2->Outputs(); -} - inline static ProgramDesc FillProgramDesc() { ProgramDesc prog; prog.MutableBlock(0)->Var("a")->SetType(proto::VarType::LOD_TENSOR); @@ -141,15 +224,6 @@ inline static ProgramDesc FillProgramDesc() { return prog; } -template -inline static std::string DebugString(const Container& c) { - std::stringstream ss; - for (auto& item : c) { - ss << item << " "; - } - return ss.str(); -} - TEST(CFGGraph, IRGraph) { // prepare ir graph auto prog = FillProgramDesc(); diff --git a/paddle/fluid/framework/details/memory_optimize_pass.cc b/paddle/fluid/framework/details/memory_optimize_pass.cc new file mode 100644 index 0000000000000000000000000000000000000000..fd02bc4697e72cdd1e5af63d71931b8fe8cc29e3 --- /dev/null +++ b/paddle/fluid/framework/details/memory_optimize_pass.cc @@ -0,0 +1,340 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/framework/details/memory_optimize_pass.h" +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include "gflags/gflags.h" +#include "paddle/fluid/framework/data_type.h" +#include "paddle/fluid/framework/ir/graph.h" +#include "paddle/fluid/framework/ir/graph_helper.h" + +DEFINE_bool(enable_subgraph_optimize, false, + "SubGraph also reuse global graph variables, it will reduce the " + "memory occupation" + "but a higher risk of memory reuse error. default disabled."); +DEFINE_string(memory_optimize_debug, "", + "debug the operator output variable when do the variable reuse." + "memory reuse pass." + "only for debug, default disabled."); + +namespace paddle { +namespace framework { +namespace details { + +std::unique_ptr MemoryOptimizePass::ApplyImpl( + std::unique_ptr graph) const { + auto nodes = graph->Nodes(); + CollectSkipVarsSet(nodes); + + cfg_.reset(new details::ControlFlowGraph(*graph)); + cfg_->LiveVariableAnalysis(); + InitSSAGraphNodes(); + + int reuse_id = 0; + for (size_t idx = 0; idx < cfg_->Ops().size(); ++idx) { + auto& op = cfg_->Ops()[idx]; + auto* op_desc = op->Op(); + // some op in graph has no op desc + if (op_desc == nullptr) continue; + if (OpHasSubBlock(op_desc)) { + if (FLAGS_enable_subgraph_optimize) { + SubGraphOptimize(op_desc); + } else { + VLOG(3) << op->Name() + << " has subblock, but disable subgraph optimize. skipped."; + continue; + } + } + + for (auto& var : op->outputs) { + if (var->IsVar() && !var->IsCtrlVar() && skip_set_.count(var->Name())) { + VLOG(3) << "Skip set contains variable of " << var->Name() + << "disable reuse on it. skipped"; + continue; + } + if (NodeCanReused(var) && cfg_->Use(op).count(var->Name()) == 0) { + ir::Node* cache = pool_.FindBestFitNode(var); + while (cache != nullptr && var->Name() == cache->Name()) { + VLOG(3) << "The same cache variable is cascade reused. " + << cache->Name() << " is re-filled to the pool after " + << "the reused op is finished. Current op can not " + << "replace it again. Skip this candidate."; + cache = pool_.FindNextBestFitNode(var, cache); + } + if (var->Name() == FLAGS_memory_optimize_debug) { + VLOG(3) << "start match var " << DebugString(var) << " of op " + << op->Name(); + VLOG(3) << pool_.ToString(); + VLOG(3) << "matched in pool : " + << ((cache == nullptr) ? "False" : "True"); + } + + if (cache != nullptr) { + int node_idx_in_pool = pool_.GetNodeIndexInPool(cache); + VLOG(3) << string::Sprintf( + "!!! %s, %s => %s, cache idx %d, pool size %d", + std::to_string(reuse_id++), DebugString(var), DebugString(cache), + node_idx_in_pool, static_cast(pool_.size())); + // NOTE(dzhwinter): update the ProgramDesc/IR Graph + // and the CFG Graph on the fly. + // + // IR Graph define the dependence relationship between nodes. + // + // ProgramDesc defines the input/output vars. Its used in + // CreateOp, CreateVar when running happens. + // + // CFG Graph store the liveness information, when reuse happens + // we also need to update the variable liveness. + const std::string var_name = var->Name(); + const std::string cache_name = cache->Name(); + + cfg_->RenameVarInCFGGraph(var_name, cache_name, idx); + RenameVarInGraphDesc(var_name, cache_name, idx); + RenameVarInGraphNode(var_name, cache_name, idx, graph.get()); + pool_.Erase(cache_name); + } + } + } + // fill the pool + for (auto var : cfg_->LiveIn(op)) { + if (cfg_->LiveOut(op).count(var) == 0) { + ir::Node* var_node = cfg_->GetNodeByName(var, op); + if (var_node == nullptr || var_node->IsCtrlVar()) continue; + if (NodeCanReused(var_node) && !pool_.Has(var_node)) { + pool_.Insert(var_node); + } + } + } + } + graph->ResolveHazard(var_nodes_); + + return graph; +} + +void MemoryOptimizePass::SubGraphOptimize(OpDesc* op_desc) const { + // conditional block, while op and their grad op + auto* sub_block_desc = + AttrReader(op_desc->GetAttrMap()).Get("sub_block"); + + // create a mirror block to construct an IR Graph. + ProgramDesc prog; + auto* copy_block = prog.MutableBlock(0); + for (auto* op : sub_block_desc->AllOps()) { + auto* copy_op = copy_block->AppendOp(); + copy_op->CopyFrom(*op); + copy_op->Flush(); + } + + for (auto* var : sub_block_desc->AllVars()) { + auto* copy_var = copy_block->Var(var->Name()); + copy_var->SetDataType(var->GetDataType()); + // only lod tensor can be reused. So ignore the multiple dims case. + copy_var->SetType(var->GetType()); + copy_var->SetShape(var->GetShape()); + copy_var->SetPersistable(var->Persistable()); + } + + ir::Graph sub_graph(prog); + std::unordered_set sub_graph_all_ops; + FilterVariables(sub_graph.Nodes(), [&](ir::Node* var) { + // sub_graph_all_ops.emplace(var); + if (var->IsVar() && !var->IsCtrlVar()) { + sub_graph_all_ops.emplace(var); + } + }); + int sub_reuse_id = 0; + // subgraph nodes is unordered, reuse need to follow the desc order. + // find the right op node through the descs + for (auto* sub_op_desc : sub_block_desc->AllOps()) { + ir::Node* sub_op = nullptr; + for (auto* node : sub_graph_all_ops) { + if (node->Op() == sub_op_desc) { + sub_op = node; + break; + } + } + PADDLE_ENFORCE(sub_op != nullptr); + for (auto* var : sub_op->outputs) { + if (NodeCanReused(var)) { + ir::Node* cache = pool_.FindBestFitNode(var); + if (cache != nullptr) { + if (var->Var()->GetDataType() != cache->Var()->GetDataType()) { + continue; + } + int node_idx_in_pool = pool_.GetNodeIndexInPool(cache); + VLOG(3) << string::Sprintf( + "!!! %s, %s => %s, cache idx %d, pool size %d", + std::to_string(sub_reuse_id++), DebugString(var), + DebugString(cache), node_idx_in_pool, + static_cast(pool_.size())); + // NOTE(dzh): subblock is not in IR graph. Modify the block_desc + // immediately to make the subblock variable reuse strategy take + // effect. Because it is a single op in graph. No need to + // update the ir nodes. + sub_op_desc->Rename(var->Name(), cache->Name()); + if (sub_op_desc->Block() != nullptr && + sub_op_desc->Block()->HasVar(var->Name())) { + sub_op_desc->Block()->RemoveVar(var->Name()); + } + } + } + } + } +} + +void MemoryOptimizePass::CollectSkipVarsSet( + const std::unordered_set& nodes) const { + auto update_skip_set = [&](OpDesc* op_desc) { + auto inputs = op_desc->InputArgumentNames(); + auto outputs = op_desc->OutputArgumentNames(); + skip_set_.insert(inputs.begin(), inputs.end()); + skip_set_.insert(outputs.begin(), outputs.end()); + }; + for (auto& op : nodes) { + if (!op->IsOp() || op->Op() == nullptr) continue; + auto* op_desc = op->Op(); + // NOTE(dzhwinter): + // current block can not reuse next level block vars. + if (OpHasSubBlock(op_desc)) update_skip_set(op_desc); + // NOTE(dzhwinter): + // distributed ops input/output name need to + // keep same bettwen trainer/pserver + if (op_desc->Type() == "send") update_skip_set(op_desc); + if (op_desc->Type() == "recv") update_skip_set(op_desc); + if (op_desc->Type() == "prefetch") update_skip_set(op_desc); + } +} + +void MemoryOptimizePass::RenameVarInGraphDesc(const std::string& var, + const std::string& cache_var, + size_t idx) const { + for (size_t i = idx; i < cfg_->Ops().size(); ++i) { + auto* op = cfg_->Ops()[i]; + PADDLE_ENFORCE(op->IsOp() && op->Op()); + auto* op_desc = op->Op(); + op_desc->RenameInput(var, cache_var); + op_desc->RenameOutput(var, cache_var); + if (op_desc->Block() != nullptr) { + op_desc->Block()->RemoveVar(var); + } else { + LOG(WARNING) << "op " << op->Name() << " not know its block." + << "Is the op_desc created without block pointer? " + << "Can not find " << var << " in Block(0)"; + } + op_desc->Flush(); + } +} + +void MemoryOptimizePass::InitSSAGraphNodes() const { + std::unordered_map> all_vars; + if (var_nodes_.empty()) { + for (auto* op : cfg_->Ops()) { + for (auto* node : op->inputs) { + if (all_vars[node->Name()].count(node) == 0) { + all_vars[node->Name()].emplace(node); + var_nodes_[node->Name()].emplace_back(node); + } + } + for (auto* node : op->outputs) { + if (all_vars[node->Name()].count(node) == 0) { + all_vars[node->Name()].emplace(node); + var_nodes_[node->Name()].emplace_back(node); + } + } + } + } +} + +void MemoryOptimizePass::RenameVarInGraphNode(const std::string& var, + const std::string& cache_var, + size_t idx, + ir::Graph* graph) const { + // if replace happens, we need to create a newer version cache_var + // but use the same dims/data_type with var. + PADDLE_ENFORCE(var_nodes_[var].size() >= 1 && + var_nodes_[var].at(0)->Var() != nullptr); + std::unique_ptr var_desc(new VarDesc(*var_nodes_[var].at(0)->Var())); + var_desc->SetName(cache_var); + + for (size_t i = idx; i < cfg_->Ops().size(); ++i) { + auto* op = cfg_->Ops()[i]; + + // redirect the input to the latest version of cache_var + for (auto* node : op->inputs) { + if (node->Name() == var) { + ir::Node* cache_node = var_nodes_[cache_var].back(); + + // swap node to cache_node + cache_node->outputs.insert(cache_node->outputs.end(), + node->outputs.begin(), node->outputs.end()); + PADDLE_ENFORCE(node->inputs.size() == 1 && node->inputs[0]->IsOp()); + auto* prev_op = node->inputs[0]; + std::replace(prev_op->outputs.begin(), prev_op->outputs.end(), node, + cache_node); + for (auto* next_op : node->outputs) { + std::replace(next_op->inputs.begin(), next_op->inputs.end(), node, + cache_node); + } + + // erase unused node + auto& nodes = var_nodes_.at(var); + nodes.erase(std::remove(nodes.begin(), nodes.end(), node), nodes.end()); + graph->RemoveNode(node); + } + } + + // if we need to rename the output, + // always create a newer version of cache_var + for (auto* node : op->outputs) { + if (node->Name() == var) { + ir::Node* cache_node = graph->CreateVarNode(var_desc.get()); + var_nodes_[cache_var].emplace_back(cache_node); + + // swap node to cache node + cache_node->outputs.insert(cache_node->outputs.end(), + node->outputs.begin(), node->outputs.end()); + cache_node->inputs.emplace_back(op); + std::replace(op->outputs.begin(), op->outputs.end(), node, cache_node); + for (auto* next_op : node->outputs) { + std::replace(next_op->inputs.begin(), next_op->inputs.end(), node, + cache_node); + } + + // erase unused node + auto& nodes = var_nodes_.at(var); + nodes.erase(std::remove(nodes.begin(), nodes.end(), node), nodes.end()); + graph->RemoveNode(node); + } + } + } +} + +} // namespace details +} // namespace framework +} // namespace paddle + +REGISTER_PASS(memory_optimize_pass, + paddle::framework::details::MemoryOptimizePass) + .RequireGraphAttr(paddle::framework::details::kAllOpDescs); diff --git a/paddle/fluid/framework/details/memory_optimize_pass.h b/paddle/fluid/framework/details/memory_optimize_pass.h new file mode 100644 index 0000000000000000000000000000000000000000..593ffc10fc99d26b1ee9174ceef081581126e7e8 --- /dev/null +++ b/paddle/fluid/framework/details/memory_optimize_pass.h @@ -0,0 +1,71 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#include "paddle/fluid/framework/data_type.h" +#include "paddle/fluid/framework/details/memory_optimize_helper.h" +#include "paddle/fluid/framework/ir/graph.h" +#include "paddle/fluid/framework/ir/pass.h" + +namespace paddle { +namespace framework { +namespace details { + +class MemoryOptimizePass : public ir::Pass { + protected: + std::unique_ptr ApplyImpl( + std::unique_ptr graph) const override; + // fill the variable map(var_nodes) by version. + void InitSSAGraphNodes() const; + + private: + // update program descs + void RenameVarInGraphDesc(const std::string& var, + const std::string& cache_var, size_t idx) const; + // update ir nodes + void RenameVarInGraphNode(const std::string& var, + const std::string& cache_var, size_t idx, + ir::Graph* graph) const; + + void SubGraphOptimize(OpDesc* op_desc) const; + // 1. scan op with subblock and collect the output/input vars. + // while, while_grad, conditional_block + // 2. scan distributed ops and collect the output/input vars + void CollectSkipVarsSet(const std::unordered_set&) const; + + private: + // Reuse Node Pool, Owned. + mutable OrderedSet pool_; + // controlflow Graph + mutable std::unique_ptr cfg_; + // skip set + mutable std::unordered_set skip_set_; + // var nodes + mutable std::map> var_nodes_; +}; + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/memory_reuse_types.cc b/paddle/fluid/framework/details/memory_reuse_types.cc deleted file mode 100644 index 2b9ff518b9adcd366cc877998400a8bdc05fa033..0000000000000000000000000000000000000000 --- a/paddle/fluid/framework/details/memory_reuse_types.cc +++ /dev/null @@ -1,155 +0,0 @@ -// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -#include "paddle/fluid/framework/details/memory_reuse_types.h" -#include -#include -#include - -namespace paddle { -namespace framework { -namespace details { - -size_t NodeSizeInBytes(ir::Node* n) { - auto* desc = FindVarDescInBlock(n); - auto shape = desc->GetShape(); - size_t type_size = SizeOfType(desc->GetDataType()); - int size = 1; - for (auto& s : shape) { - size *= s; - } - return type_size * std::abs(size); -} - -std::string DebugStringImpl(VarDesc* var) { - std::stringstream ss; - ss << var->Name(); - ss << "["; - try { - auto shape = var->GetShape(); - for (size_t i = 0; i < shape.size(); ++i) { - if (i != shape.size() - 1) { - ss << shape[i] << ","; - } else { - ss << shape[i]; - } - } - ss << "]"; - } catch (...) { - ss << "Var has no VarDesc !!! Name:" << var->Name(); - } - return ss.str(); -} - -std::string DebugString(ir::Node* var) { - return DebugStringImpl(FindVarDescInBlock(var)); -} -// return DebugString(var->Var()); } - -// NOTE(dzh): based ir node, if a large node has been reused -// by a small size node, then next time it appear in pool, it will -// have the small size. Find the original node shap from blockdesc. -VarDesc* FindVarDescInBlock(ir::Node* n) { - PADDLE_ENFORCE(n->IsVar() && !n->IsCtrlVar() && n->inputs.size() == 1); - BlockDesc* block = n->inputs[0]->Op()->Block(); - PADDLE_ENFORCE(block->HasVar(n->Name()), - string::Sprintf("Block do not has var %s", n->Name())); - return block->FindVar(n->Name()); -} - -struct NodeComparator { - bool operator()(ir::Node* lhs, ir::Node* rhs) const { - auto* lhs_desc = FindVarDescInBlock(lhs); - auto* rhs_desc = FindVarDescInBlock(rhs); - auto lhs_shape = lhs_desc->GetShape(); - auto rhs_shape = rhs_desc->GetShape(); - if ((lhs_shape[0] == -1 && rhs_shape[0] == -1) || - (lhs_shape[0] != -1 && rhs_shape[0] != -1)) { - return NodeSizeInBytes(lhs) <= NodeSizeInBytes(rhs); - } else { - return false; - } - } -}; - -void OrderedNodePairPool::Insert(ir::Node* var, ir::Node* op) { - PADDLE_ENFORCE(var->IsVar() && !var->IsCtrlVar()); - PADDLE_ENFORCE(op->IsOp()); - if (mark_table_.count(var->Name()) != 0) { - mark_table_[var->Name()]->second.insert(op); - return; - } - - auto* var_desc = FindVarDescInBlock(var); - auto var_shape = var_desc->GetShape(); - int batch_size = static_cast(var_shape[0]); - - NodeComparator compare_node; - Iter it = nodes_.begin(); - while (it != nodes_.end()) { - auto* cache_desc = FindVarDescInBlock(it->first); - int cache_batch_size = cache_desc->GetShape()[0]; - if ((cache_batch_size == -1 && batch_size == -1) || - (cache_batch_size != -1 && batch_size != -1)) { - if (compare_node(it->first, var)) { - ++it; - } else { - break; - } - } else if (cache_batch_size == -1 && batch_size != -1) { - ++it; - } else if (cache_batch_size != -1 && batch_size == -1) { - break; - } - } - - it = - nodes_.insert(it, std::make_pair(var, std::unordered_set{op})); - mark_table_[var->Name()] = it; -} - -int OrderedNodePairPool::GetIndex(ir::Node* var) { - return std::distance(nodes_.begin(), mark_table_[var->Name()]); -} - -ir::Node* OrderedNodePairPool::NodeMatch(ir::Node* var) const { - ir::Node* found_node = nullptr; - NodeComparator compare_node; - - for (auto it = nodes_.begin(); it != nodes_.end(); ++it) { - if (compare_node(var, it->first)) { - found_node = it->first; - break; - } - } - return found_node; -} - -void OrderedNodePairPool::Erase(ir::Node* var) { - PADDLE_ENFORCE(mark_table_.count(var->Name())); - nodes_.erase(mark_table_[var->Name()]); - mark_table_.erase(var->Name()); -} - -std::string OrderedNodePairPool::ToString() const { - std::stringstream ss; - for (auto it = nodes_.begin(); it != nodes_.end(); ++it) { - ss << DebugString(it->first) << " "; - } - return ss.str(); -} - -} // namespace details -} // namespace framework -} // namespace paddle diff --git a/paddle/fluid/framework/details/memory_reuse_types.h b/paddle/fluid/framework/details/memory_reuse_types.h deleted file mode 100644 index 9a9c1d948e869016717fea9ff6b8236adfc29845..0000000000000000000000000000000000000000 --- a/paddle/fluid/framework/details/memory_reuse_types.h +++ /dev/null @@ -1,87 +0,0 @@ -// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -#pragma once -#include -#include -#include -#include -#include -#include -#include -#include "paddle/fluid/framework/data_type.h" -#include "paddle/fluid/framework/ir/graph.h" - -namespace paddle { -namespace framework { -namespace details { - -constexpr char kFetchedVars[] = "fetched_vars"; -constexpr char kGraphNodePool[] = "graph_node_pool"; - -// NOTE(dzh): Variable and the operators use the var. -// for early delete pass. -// Because analysis var pass build base on ir::Node, which maybe released -// or modified between passes, so we use OpDesc* to mark ops. -using GraphNodePool = std::vector< - std::pair /* ops */>>; - -// NOTE(dzh): by default, it sort node in ascend order(by node bytes size). -// in fluid, -1 means the batch_size is determined in runtime. -// the node batch_size equal -1 always ranking in the front than the node not. -// For example, -// node0[-1, 1] node1[-1, 1, 1], node2[1,1], node3[1,1024], .. -// O(1) insert, delete -class OrderedNodePairPool { - public: - using NodePair = std::pair>; - using Iter = typename std::list::iterator; - using ConstIter = typename std::list::const_iterator; - - void Insert(ir::Node* var, ir::Node* op); - - void Erase(ir::Node* var); - - bool Has(ir::Node* var) { return mark_table_.count(var->Name()); } - - ir::Node* NodeMatch(ir::Node* var) const; - // map store non-const iterator, can not promise const - int GetIndex(ir::Node* var); - // pool all node to string - std::string ToString() const; - - Iter begin() { return nodes_.begin(); } - Iter end() { return nodes_.end(); } - ConstIter begin() const { return nodes_.begin(); } - ConstIter end() const { return nodes_.end(); } - size_t size() const { return nodes_.size(); } - - private: - // for searching. - std::unordered_map mark_table_; - // node swap pairs. var -> ops dep var - std::list nodes_; -}; - -// node memory size in bytes -size_t NodeSizeInBytes(ir::Node* n); - -std::string DebugString(ir::Node* var); - -// std::string DebugString(VarDesc* var); -VarDesc* FindVarDescInBlock(ir::Node* n); - -} // namespace details -} // namespace framework -} // namespace paddle diff --git a/paddle/fluid/framework/details/memory_reuse_types_test.cc b/paddle/fluid/framework/details/memory_reuse_types_test.cc deleted file mode 100644 index d2fabf5ce068e0f752b86c0d02b971f18fc65f01..0000000000000000000000000000000000000000 --- a/paddle/fluid/framework/details/memory_reuse_types_test.cc +++ /dev/null @@ -1,99 +0,0 @@ -// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -#include "paddle/fluid/framework/details/memory_reuse_types.h" -#include -#include -#include -#include -#include -#include -#include -#include "glog/logging.h" -#include "gtest/gtest.h" - -namespace paddle { -namespace framework { -namespace details { - -TEST(OrderedNodePairPool, Normal) { - OrderedNodePairPool pool; - std::vector> nodes; - - // clang-format off - std::vector> shapes = {{-1, 10}, - {-1, 20}, - {1, 2}, - {5, 2}, - {10, 20}, - {-1, 2, 5}, - {-1, 1, 5}, - {-1, 1}}; - // clang-format on - const int COUNT = shapes.size(); - ProgramDesc prog; - BlockDesc* block_desc = prog.MutableBlock(0); - auto* op_desc = block_desc->AppendOp(); - op_desc->SetType("dummy"); - std::unique_ptr op = ir::CreateNodeForTest(op_desc); - - for (int i = 0; i < COUNT; ++i) { - auto desc = block_desc->Var(std::to_string(i)); - desc->SetShape(shapes[i]); - std::unique_ptr node = ir::CreateNodeForTest(desc); - node->inputs.emplace_back(op.get()); - nodes.emplace_back(std::move(node)); - } - - for (auto& node : nodes) { - pool.Insert(node.get(), op.get()); - } - - // assert its order and interface. - std::cout << pool.ToString() << std::endl; - pool.Erase(nodes.front().get()); - std::cout << pool.ToString() << std::endl; - - ASSERT_EQ(pool.size(), static_cast(COUNT - 1)); - ASSERT_EQ(pool.GetIndex(nodes.back().get()), 0); - - { - auto v1 = block_desc->Var("11"); - v1->SetShape({-1, 256, 56, 56}); - std::unique_ptr node1 = ir::CreateNodeForTest(v1); - node1->inputs.emplace_back(op.get()); - auto* cache = pool.NodeMatch(node1.get()); - ASSERT_EQ(cache, nullptr); - } - { - auto v2 = block_desc->Var("12"); - v2->SetShape({-1, 2, 5}); - std::unique_ptr node1 = ir::CreateNodeForTest(v2); - node1->inputs.emplace_back(op.get()); - auto* cache = pool.NodeMatch(node1.get()); - ASSERT_EQ(pool.GetIndex(cache), 2); // match 6:[-1,2,5] - } - { - auto v3 = block_desc->Var("13"); - v3->SetShape({2, 5}); - std::unique_ptr node1 = ir::CreateNodeForTest(v3); - node1->inputs.emplace_back(op.get()); - auto* cache = pool.NodeMatch(node1.get()); - ASSERT_EQ(pool.GetIndex(cache), 5); // match 4:[5,2] - } -} - -} // namespace details -} // namespace framework -} // namespace paddle diff --git a/paddle/fluid/framework/details/multi_devices_graph_pass.cc b/paddle/fluid/framework/details/multi_devices_graph_pass.cc index 75f922d2cca6855a67be7284ae407e549a1a1afb..478d2ffbcf2988487893984284d4597f018f0ca0 100644 --- a/paddle/fluid/framework/details/multi_devices_graph_pass.cc +++ b/paddle/fluid/framework/details/multi_devices_graph_pass.cc @@ -392,20 +392,32 @@ void MultiDevSSAGraphBuilderBase::CreateComputationalOp(ir::Graph *result, void MultiDevSSAGraphBuilderBase::CreateAllReduceOp( ir::Graph *result, const std::string &og) const { + OpHandleBase *op_handle = nullptr; + + auto append_allreduce_op = [&]( + const std::vector &scopes, + const std::vector &places) -> OpHandleBase * { #if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) - result->Get(kGraphOps).emplace_back(new AllReduceOpHandle( - result->CreateEmptyNode("allreduce", ir::Node::Type::kOperation), - local_scopes_, places_, nccl_ctxs_)); + result->Get(kGraphOps).emplace_back(new AllReduceOpHandle( + result->CreateEmptyNode("allreduce", ir::Node::Type::kOperation), + scopes, places, nccl_ctxs_)); #else - result->Get(kGraphOps).emplace_back(new AllReduceOpHandle( - result->CreateEmptyNode("allreduce", ir::Node::Type::kOperation), - local_scopes_, places_)); + result->Get(kGraphOps).emplace_back(new AllReduceOpHandle( + result->CreateEmptyNode("allreduce", ir::Node::Type::kOperation), + scopes, places)); #endif - auto *op_handle = result->Get(kGraphOps).back(); + return result->Get(kGraphOps).back(); + }; + + if (!strategy_.enable_parallel_graph_) + op_handle = append_allreduce_op(local_scopes_, places_); for (size_t i = 0; i < places_.size(); ++i) { - auto &p = places_[i]; - SetCommunicationContext(op_handle, p); + if (strategy_.enable_parallel_graph_) { + op_handle = append_allreduce_op({local_scopes_[i]}, {places_[i]}); + } + + SetCommunicationContext(op_handle, places_[i]); auto &vars = result->Get(kGraphVars)[i][og]; PADDLE_ENFORCE(!vars.empty()); auto &prev_grad = vars.back(); @@ -413,7 +425,7 @@ void MultiDevSSAGraphBuilderBase::CreateAllReduceOp( auto var = new VarHandle(result->CreateEmptyNode(og, ir::Node::Type::kVariable), - vars.size(), i, og, p); + vars.size(), i, og, places_[i]); vars.emplace_back(var); op_handle->AddOutput(var); } @@ -925,9 +937,21 @@ void DistSSAGraphBuilder::InsertCollectiveOp(ir::Graph *result, } void DistSSAGraphBuilder::InsertPostprocessOps(ir::Graph *result) const { - if (need_broadcast_var_ || - (UseGPU() && - strategy_.reduce_ == BuildStrategy::ReduceStrategy::kReduce)) { + // broad cast received parameters when training in parameter server mode. + if (need_broadcast_var_) { + // There are 4 conditions: + // 1. GPU && Reduce: Reduce gradient then broadcast gradient to other GPUS. + // Need to broadcast received parameters to other GPU. + // 2. GPU && AllReduce: AllReduce all graident to each GPU. Need to + // broadcast received parameters to other GPU. + // 3. CPU && AllReduce: AllReduce all gradient to each thread. Need to + // broadcast received parameters to other scope. + // 4. CPU && Reduce: because all parameters share the same memory, did not + // broadcast received parameters. + if (!UseGPU() && + strategy_.reduce_ == BuildStrategy::ReduceStrategy::kReduce) { + return; + } if (strategy_.fuse_broadcast_op_) { CreateFusedBroadcastOp(result, bcast_var_name_set_); } else { diff --git a/paddle/fluid/framework/details/multi_devices_helper.h b/paddle/fluid/framework/details/multi_devices_helper.h index 1a2b75fbc0c28984ce5cf00e0a2ce0f804349bb1..9afbb91005c9c3a9d2e185f4dfa901ebf812ee19 100644 --- a/paddle/fluid/framework/details/multi_devices_helper.h +++ b/paddle/fluid/framework/details/multi_devices_helper.h @@ -36,13 +36,14 @@ namespace details { // map from variable name to variables. The variables, who have the same name, // will have a differsent version. The offset in the // `std::vector` is the version of varaibles. -typedef std::vector>> +typedef std::vector>> GraphVars; const char kGraphVars[] = "vars"; // aux variables to represent dependency. Useful to resolve data hazard. -typedef std::unordered_set GraphDepVars; +typedef std::unordered_set GraphDepVars; const char kGraphDepVars[] = "dep_vars"; + } // namespace details } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/details/op_handle_base.h b/paddle/fluid/framework/details/op_handle_base.h index b1a82e8771b92f2d0af4a1c7732ff2da54d496a8..e0aa352e95bc3685a1f4879bffa6e86eecd7e7f9 100644 --- a/paddle/fluid/framework/details/op_handle_base.h +++ b/paddle/fluid/framework/details/op_handle_base.h @@ -70,6 +70,9 @@ class OpHandleBase { auto it = dev_ctxes_.find(place); return it != dev_ctxes_.end() ? it->second : nullptr; } + const std::map &DeviceContext() { + return dev_ctxes_; + } void SetDeviceContext(platform::Place place, platform::DeviceContext *ctx_) { dev_ctxes_[place] = ctx_; diff --git a/paddle/fluid/framework/details/op_registry.h b/paddle/fluid/framework/details/op_registry.h index eea7e712f8f6e187cdceedce77cc76d1d4ca2101..0901e59f9786b43361e7a570f8c2a07be54c1ac2 100644 --- a/paddle/fluid/framework/details/op_registry.h +++ b/paddle/fluid/framework/details/op_registry.h @@ -18,6 +18,7 @@ limitations under the License. */ #include #include #include "paddle/fluid/framework/grad_op_desc_maker.h" +#include "paddle/fluid/framework/inplace_op_inference.h" #include "paddle/fluid/framework/op_info.h" #include "paddle/fluid/framework/op_proto_maker.h" #include "paddle/fluid/framework/operator.h" @@ -32,7 +33,8 @@ enum OpInfoFillType { kOpProtoAndCheckerMaker = 1, kGradOpDescMaker = 2, kVarTypeInference = 3, - kShapeInference = 4 + kShapeInference = 4, + kInplaceOpInference = 5 }; template @@ -48,8 +50,11 @@ struct OpInfoFillTypeID { ? kVarTypeInference : (std::is_base_of::value ? kShapeInference - : static_cast( - -1))))); + : (std::is_base_of< + InplaceOpInference, T>::value + ? kInplaceOpInference + : static_cast( + -1)))))); } }; @@ -139,6 +144,16 @@ struct OpInfoFiller { } }; +template +struct OpInfoFiller { + void operator()(const char* op_type, OpInfo* info) const { + info->infer_inplace_ = [](const OpDesc& op_desc, BlockDesc* block) { + T infer; + return infer(op_desc, block); + }; + } +}; + } // namespace details } // namespace framework diff --git a/paddle/fluid/framework/details/parallel_ssa_graph_executor.cc b/paddle/fluid/framework/details/parallel_ssa_graph_executor.cc index 128aaa33a2c60e62fdca13768cdc0a815167f3ef..4c8f69c68ce17d0143c34e8adbab92cdc90058c8 100644 --- a/paddle/fluid/framework/details/parallel_ssa_graph_executor.cc +++ b/paddle/fluid/framework/details/parallel_ssa_graph_executor.cc @@ -13,22 +13,92 @@ // limitations under the License. #include "paddle/fluid/framework/details/parallel_ssa_graph_executor.h" +#include "paddle/fluid/framework/ir/graph_helper.h" namespace paddle { namespace framework { namespace details { +std::vector> +ParallelSSAGraphExecutor::SeparateMultiDevicesGraph( + std::unique_ptr &&graph) { + std::vector> graphs; + graphs.reserve(places_.size()); + for (size_t i = 0; i < places_.size(); ++i) { + ProgramDesc empty; + graphs.emplace_back(std::unique_ptr(new ir::Graph(empty))); + auto &g = graphs.back(); + g->Set(kGraphVars, new GraphVars(1UL)); + g->Set(kGraphDepVars, new GraphDepVars); + } + auto op_handles = ir::FilterByNodeWrapper(*graph); + + for (auto &op : op_handles) { + auto &dev_ctx = op->DeviceContext(); + auto &p = dev_ctx.begin()->first; + int dev_id = boost::get(p).device; + auto &dev_dummys = graphs[dev_id]->Get(kGraphDepVars); + graphs[dev_id]->AddNode(graph->RemoveNode(op->Node()).release()); + + for (auto &var : op->Inputs()) { + auto dummy_ptr = dynamic_cast(var); + if (dummy_ptr) { + dev_dummys.insert(var); + if (graph->Nodes().count(var->Node())) + graphs[dev_id]->AddNode(graph->RemoveNode(var->Node()).release()); + } + } + for (auto &var : op->Outputs()) { + auto dummy_ptr = dynamic_cast(var); + if (dummy_ptr) { + dev_dummys.insert(var); + if (graph->Nodes().count(var->Node())) + graphs[dev_id]->AddNode(graph->RemoveNode(var->Node()).release()); + } + } + } + + for (size_t dev_id = 0; dev_id < places_.size(); ++dev_id) { + auto &dev_vars = graphs[dev_id]->Get(kGraphVars)[0]; + auto &origin_vars = graph->Get(kGraphVars)[dev_id]; + for (auto &name_pair : origin_vars) { + dev_vars.emplace(name_pair.first, name_pair.second); + for (auto &version_pair : name_pair.second) { + if (graph->Nodes().count(version_pair->Node())) { + graphs[dev_id]->AddNode( + graph->RemoveNode(version_pair->Node()).release()); + } + } + } + } + + return graphs; +} + ParallelSSAGraphExecutor::ParallelSSAGraphExecutor( const ExecutionStrategy &strategy, const std::vector &local_scopes, const std::vector &places, - std::vector> &&graphs) + const framework::ProgramDesc &main_prog, std::unique_ptr &&graph) : strategy_(std::move(strategy)), local_scopes_(std::move(local_scopes)), pool_(places.size() >= 2 ? new ::ThreadPool(places.size()) : nullptr), places_(std::move(places)), - graphs_(std::move(graphs)) { + main_prog_(main_prog), + // TODO(Yancey1989): Copying graphs is not safely since it deleted the + // attrs. + graphs_(SeparateMultiDevicesGraph(std::move(graph))) { PADDLE_ENFORCE_EQ(places_.size(), local_scopes_.size()); + auto seq_allreduce_pass = + ir::PassRegistry::Instance().Get("all_reduce_deps_pass"); + seq_allreduce_pass->Erase(details::kAllOpDescs); + seq_allreduce_pass->Set>( + details::kAllOpDescs, + new std::vector(main_prog_.Block(0).AllOps())); + for (size_t i = 0; i < graphs_.size(); ++i) { + graphs_[i] = seq_allreduce_pass->Apply(std::move(graphs_[i])); + } + // set the correct size of thread pool to each device. strategy_.num_threads_ = strategy_.num_threads_ < places_.size() ? 1UL @@ -37,7 +107,7 @@ ParallelSSAGraphExecutor::ParallelSSAGraphExecutor( << " to run the operators of the graph on each device."; for (size_t i = 0; i < places.size(); ++i) { executors_.emplace_back(new details::ThreadedSSAGraphExecutor( - strategy_, {local_scopes_[i]}, {places_[i]}, std::move(graphs_[i]))); + strategy_, local_scopes_, {places_[i]}, std::move(graphs_.at(i)))); } } @@ -65,7 +135,7 @@ FeedFetchList ParallelSSAGraphExecutor::Run( if (pool_) { run_futures.emplace_back(pool_->enqueue(std::move(call))); } else { - fetch_data.emplace_back(std::move(call())); + fetch_data.emplace_back(call()); } } @@ -74,7 +144,7 @@ FeedFetchList ParallelSSAGraphExecutor::Run( if (exception_holder_.IsCaught()) { f.wait(); } else { - fetch_data.emplace_back(std::move(f.get())); + fetch_data.emplace_back(f.get()); } } } diff --git a/paddle/fluid/framework/details/parallel_ssa_graph_executor.h b/paddle/fluid/framework/details/parallel_ssa_graph_executor.h index c00c5bc2d1b4b78593f99c819b5a3d642150e773..1c35d45fdd356a867d1ad80b345379395e03172e 100644 --- a/paddle/fluid/framework/details/parallel_ssa_graph_executor.h +++ b/paddle/fluid/framework/details/parallel_ssa_graph_executor.h @@ -18,7 +18,9 @@ #include #include "ThreadPool.h" +#include "paddle/fluid/framework/details/multi_devices_helper.h" #include "paddle/fluid/framework/details/threaded_ssa_graph_executor.h" +#include "paddle/fluid/framework/ir/graph.h" namespace paddle { namespace framework { @@ -29,17 +31,23 @@ class ParallelSSAGraphExecutor : public SSAGraphExecutor { ParallelSSAGraphExecutor(const ExecutionStrategy &strategy, const std::vector &local_scopes, const std::vector &places, - std::vector> &&graphs); + const framework::ProgramDesc &main_prog, + std::unique_ptr &&graph); ~ParallelSSAGraphExecutor() final = default; + const ir::Graph &Graph() const override { return *graphs_[0]; } FeedFetchList Run(const std::vector &fetch_tensors) override; private: + std::vector> SeparateMultiDevicesGraph( + std::unique_ptr &&graph); + ExecutionStrategy strategy_; std::vector local_scopes_; std::unique_ptr<::ThreadPool> pool_{nullptr}; std::vector places_; + framework::ProgramDesc main_prog_; std::vector> graphs_; std::vector> executors_; diff --git a/paddle/fluid/framework/details/reduce_op_handle.cc b/paddle/fluid/framework/details/reduce_op_handle.cc index ee4c8a6ecf77e5d0f23f38b763917d926afdb07a..4e2477c205db5966aa0b2d0c7a608be94a69eb82 100644 --- a/paddle/fluid/framework/details/reduce_op_handle.cc +++ b/paddle/fluid/framework/details/reduce_op_handle.cc @@ -139,7 +139,7 @@ void ReduceOpHandle::GatherSelectedRows( #endif void ReduceOpHandle::RunImpl() { - platform::RecordEvent record_event(Name(), dev_ctxes_.cbegin()->second); + platform::RecordEvent record_event(Name()); if (places_.size() == 1) return; // the input and output may have dummy var. @@ -153,7 +153,7 @@ void ReduceOpHandle::RunImpl() { { auto out_var_handles = DynamicCast(outputs_); - PADDLE_ENFORCE_EQ(out_var_handles.size(), 1, + PADDLE_ENFORCE_EQ(out_var_handles.size(), 1UL, "The number of output should be one."); out_var_handle = out_var_handles.front(); } diff --git a/paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.cc b/paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.cc index 91e4f9adb418978c30f512abe6924c0ace182124..7b13112986f9ad85056a3e8a5a6ed99bd0be95d5 100644 --- a/paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.cc +++ b/paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.cc @@ -63,7 +63,7 @@ FeedFetchList ScopeBufferedSSAGraphExecutor::Run( eptr = std::current_exception(); } - platform::RecordEvent e("ScopeBufferedSSAGraphExecutorAfterRun", nullptr); + platform::RecordEvent e("ScopeBufferedSSAGraphExecutorAfterRun"); ++drop_scope_counter_; bool stream_end = false; diff --git a/paddle/fluid/framework/details/sequential_execution_pass.cc b/paddle/fluid/framework/details/sequential_execution_pass.cc index cc2c8bfef9f9f54c2e499467df0d22ce3f69d6b8..879fb29d5926941e574d0080051c195293bc60a9 100644 --- a/paddle/fluid/framework/details/sequential_execution_pass.cc +++ b/paddle/fluid/framework/details/sequential_execution_pass.cc @@ -17,6 +17,7 @@ #include #include #include +#include "paddle/fluid/framework/details/memory_optimize_helper.h" #include "paddle/fluid/framework/op_proto_maker.h" namespace paddle { diff --git a/paddle/fluid/framework/details/sequential_execution_pass.h b/paddle/fluid/framework/details/sequential_execution_pass.h index a04c08bc2eb3bae797d648b30a22a5fee7ba0eaa..ea3034877fcea80de0124df64d8d23028bdcb7b3 100644 --- a/paddle/fluid/framework/details/sequential_execution_pass.h +++ b/paddle/fluid/framework/details/sequential_execution_pass.h @@ -21,8 +21,6 @@ namespace paddle { namespace framework { namespace details { -constexpr char kAllOpDescs[] = "all_op_descs"; - class SequentialExecutionPass : public ir::Pass { protected: std::unique_ptr ApplyImpl( diff --git a/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc b/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc index 677a2937945b03fa577317cb4f26e09354d06957..72acc337b7cc4803fa010373f8817ff5fd25cb2c 100644 --- a/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc +++ b/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc @@ -37,7 +37,7 @@ ThreadedSSAGraphExecutor::ThreadedSSAGraphExecutor( FeedFetchList ThreadedSSAGraphExecutor::Run( const std::vector &fetch_tensors) { std::unique_ptr event( - new platform::RecordEvent("ThreadedSSAGraphExecutorPrepare", nullptr)); + new platform::RecordEvent("ThreadedSSAGraphExecutorPrepare")); std::unordered_map pending_ops; std::unordered_set pending_vars; auto ready_vars = std::make_shared>(); @@ -219,7 +219,7 @@ void ThreadedSSAGraphExecutor::RunOp( VLOG(10) << op << " " << op->Name() << " Done "; running_ops_--; ready_var_q->Extend(op->Outputs()); - VLOG(10) << op << " " << op->Name() << "Signal posted"; + VLOG(10) << op << " " << op->Name() << " Signal posted"; } catch (...) { exception_holder_.Catch(std::current_exception()); } diff --git a/paddle/fluid/framework/feed_fetch_method.cc b/paddle/fluid/framework/feed_fetch_method.cc index 6338be75a4b1d3c4caf7a6f7add4d05fec690340..96530b2a3f9cfd9462627a42b2bb0fea98758f92 100644 --- a/paddle/fluid/framework/feed_fetch_method.cc +++ b/paddle/fluid/framework/feed_fetch_method.cc @@ -44,6 +44,7 @@ LoDTensor& GetFetchVariable(const Scope& scope, const std::string& var_name, // Since we want to fetch LodTensor from a variable, the variable must // be created alreadly. Variable* g_fetch_value = scope.FindVar(var_name); + PADDLE_ENFORCE_NOT_NULL(g_fetch_value, "%s is not found.", var_name); PADDLE_ENFORCE(g_fetch_value->IsType(), "Only %s can be invoked by GetFetchVariable", typeid(FeedFetchList).name()); diff --git a/paddle/fluid/framework/inplace_op_inference.h b/paddle/fluid/framework/inplace_op_inference.h new file mode 100644 index 0000000000000000000000000000000000000000..a3ccf677c90e8466f6c89041979336d45c1ac942 --- /dev/null +++ b/paddle/fluid/framework/inplace_op_inference.h @@ -0,0 +1,115 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once +#include +#include +#include +#include +#include "glog/logging.h" +#include "paddle/fluid/framework/block_desc.h" +#include "paddle/fluid/framework/details/memory_optimize_helper.h" +#include "paddle/fluid/framework/op_desc.h" +#include "paddle/fluid/framework/type_defs.h" + +namespace paddle { +namespace framework { + +/* + Inplace Inference for create In->Out pairs for inplaced operator. + If we specify a pair of corresponding names. For example, X->Out. + then Out will inplaced use X's memory. The base class will do + legality validation for both variables. +*/ +class InplaceOpInference { + public: + virtual ~InplaceOpInference() {} + virtual std::unordered_map operator()( + const OpDesc& op_desc, BlockDesc* block) const = 0; +}; + +class InplaceInToOut : public InplaceOpInference { + public: + std::unordered_map operator()( + const OpDesc& op_desc, BlockDesc* block) const { + std::unordered_map ret; + auto in_out_var_names_pair = this->Apply(op_desc, block); + for (auto& pair : in_out_var_names_pair) { + PADDLE_ENFORCE(!op_desc.Input(pair.first).empty(), + string::Sprintf("op %s do not have input of %s!", + op_desc.Type(), pair.first)); + PADDLE_ENFORCE(!op_desc.Output(pair.second).empty(), + string::Sprintf("op %s do not have output of %s!", + op_desc.Type(), pair.second)); + auto& in_name = op_desc.Input(pair.first).at(0); + auto& out_name = op_desc.Output(pair.second).at(0); + + auto in = block->FindRecursiveOrCreateVar(in_name); + auto out = block->FindRecursiveOrCreateVar(out_name); + if (TryInplaceInputOutput(in, out)) ret.insert({in_name, out_name}); + } + return ret; + } + + protected: + virtual std::unordered_map Apply( + const OpDesc& op_desc, BlockDesc* block) const = 0; + + bool TryInplaceInputOutput(const VarDesc& in, const VarDesc& out) const { + return in.Name() != out.Name() && details::NodeCanReused(in) && + details::NodeCanReused(out) && + details::NodeSize(out) <= details::NodeSize(in); + } +}; + +/* + Inplace In and Out for operator only have an Input and an Output. + For example, activation op. + */ +class SingleOpInplaceInToOut : public InplaceInToOut { + protected: + std::unordered_map Apply( + const OpDesc& op_desc, BlockDesc* block) const override { + PADDLE_ENFORCE(!op_desc.InputNames().empty(), + "Op inputs must not be empty"); + PADDLE_ENFORCE(!op_desc.OutputNames().empty(), + "Op outputs must not be empty"); + auto x_name = op_desc.InputNames().at(0); + auto out_name = op_desc.OutputNames().at(0); + return std::unordered_map{{x_name, out_name}}; + } +}; + +/* + Gradient op. Inplace output use it's Input. + For example, Input@Grad->Input reuse strategy. + */ +class GradOpInplaceInToOut : public InplaceInToOut { + protected: + std::unordered_map Apply( + const OpDesc& op_desc, BlockDesc* block) const override { + std::unordered_map ret; + std::unordered_set output_names(op_desc.OutputNames().begin(), + op_desc.OutputNames().end()); + for (auto& input_name : op_desc.InputNames()) { + if (output_names.count(GradVarName(input_name))) { + ret.insert({input_name, GradVarName(input_name)}); + } + } + return ret; + } +}; + +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/inplace_op_inference_test.cc b/paddle/fluid/framework/inplace_op_inference_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..bf9d1dcd380cdff886301faf13b0015fd5a2ed5c --- /dev/null +++ b/paddle/fluid/framework/inplace_op_inference_test.cc @@ -0,0 +1,288 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include +#include +#include "gtest/gtest.h" +#include "paddle/fluid/framework/op_info.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/framework/operator.h" +#include "paddle/fluid/framework/program_desc.h" +#include "paddle/fluid/framework/var_type_inference.h" + +namespace paddle { +namespace framework { + +class NOP : public OperatorBase { + public: + NOP(const std::string& type, const VariableNameMap& inputs, + const VariableNameMap& outputs, const AttributeMap& attrs) + : OperatorBase(type, inputs, outputs, attrs) {} + + private: + void RunImpl(const Scope& scope, + const platform::Place& place) const override {} +}; + +class SingleOpMaker : public OpProtoAndCheckerMaker { + public: + void Make() { + AddInput("X", "").AsDuplicable(); + AddOutput("Out", ""); + AddComment(""); + } +}; + +class SingleGradOpMaker : public framework::SingleGradOpDescMaker { + public: + using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; + + protected: + std::unique_ptr Apply() const override { + auto* op = new framework::OpDesc(); + op->SetType("single_op_grad"); + op->SetInput("Out", OutputGrad("Out")); + op->SetOutput(framework::GradVarName("X"), InputGrad("X")); + return std::unique_ptr(op); + } +}; + +class SingleOpShapeInference : public framework::InferShapeBase { + public: + void operator()(framework::InferShapeContext* ctx) const override { + ctx->HasInput("X"); + ctx->HasOutput("Out"); + ctx->SetOutputDim("Out", ctx->GetInputDim("X")); + } +}; + +class SingleGradOpShapeInference : public framework::InferShapeBase { + public: + void operator()(framework::InferShapeContext* ctx) const override { + ctx->HasInput(framework::GradVarName("Out")); + ctx->HasOutput(framework::GradVarName("X")); + ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("Out")); + } +}; + +class MultiOutOpMaker : public OpProtoAndCheckerMaker { + public: + void Make() { + AddInput("X", "").AsDuplicable(); + AddInput("Y", "").AsDuplicable(); + AddInput("Z", "").AsDuplicable(); + AddOutput("Out", ""); + AddOutput("YOut", ""); + AddOutput("ZOut", ""); + AddOutput("NotReuseOut", ""); + AddComment(""); + } +}; + +class MultiOutShapeInference : public framework::InferShapeBase { + public: + void operator()(framework::InferShapeContext* ctx) const override { + ctx->ShareDim("X", "Out"); + ctx->ShareDim("Y", "YOut"); + ctx->ShareDim("Z", "ZOut"); + } +}; + +class MultiGradOpMaker : public framework::SingleGradOpDescMaker { + public: + using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; + + protected: + std::unique_ptr Apply() const override { + auto* op = new framework::OpDesc(); + op->SetType("multi_out_grad"); + op->SetInput("X", Input("X")); + op->SetOutput(framework::GradVarName("Y"), OutputGrad("YOut")); + op->SetOutput(framework::GradVarName("X"), OutputGrad("Out")); + op->SetOutput(framework::GradVarName("Z"), OutputGrad("ZOut")); + return std::unique_ptr(op); + } +}; + +class MultiOutGradShapeInference : public framework::InferShapeBase { + public: + void operator()(framework::InferShapeContext* ctx) const override { + ctx->SetOutputDim(framework::GradVarName("Y"), + ctx->GetInputDim(framework::GradVarName("YOut"))); + ctx->SetOutputDim(framework::GradVarName("X"), + ctx->GetInputDim(framework::GradVarName("Out"))); + ctx->SetOutputDim(framework::GradVarName("Z"), + ctx->GetInputDim(framework::GradVarName("ZOut"))); + } +}; + +class MultiOutInplaceInToOut : public framework::InplaceInToOut { + public: + using framework::InplaceInToOut::InplaceInToOut; + + protected: + std::unordered_map Apply( + const OpDesc& op_desc, BlockDesc* block) const override { + return std::unordered_map{ + {"X", "Out"}, {"Y", "YOut"}, {"Z", "ZOut"}, + }; + } +}; + +class MultiOutGradInplaceInToOut : public framework::InplaceInToOut { + public: + using framework::InplaceInToOut::InplaceInToOut; + + protected: + std::unordered_map Apply( + const OpDesc& op_desc, BlockDesc* block) const override { + return std::unordered_map{ + {framework::GradVarName("YOut"), framework::GradVarName("Y")}, + {framework::GradVarName("Out"), framework::GradVarName("X")}, + {framework::GradVarName("ZOut"), framework::GradVarName("Z")}, + }; + } +}; + +} // namespace framework +} // namespace paddle + +namespace f = paddle::framework; +REGISTER_OPERATOR(single_op, f::NOP, f::SingleOpMaker, f::SingleGradOpMaker, + f::SingleOpInplaceInToOut, f::SingleOpShapeInference); +REGISTER_OPERATOR(single_op_grad, f::NOP, f::SingleOpInplaceInToOut, + f::SingleGradOpShapeInference); +REGISTER_OPERATOR(multi_out_op, f::NOP, f::MultiOutOpMaker, f::MultiGradOpMaker, + f::MultiOutInplaceInToOut, f::MultiOutShapeInference); +REGISTER_OPERATOR(multi_out_grad, f::NOP, f::MultiOutGradInplaceInToOut, + f::MultiOutGradShapeInference); + +namespace paddle { +namespace framework { + +TEST(InferInplace, SingleOpInplaceInToOut) { + ProgramDesc prog; + auto* op = prog.MutableBlock(0)->AppendOp(); + op->SetType("single_op"); + op->SetInput("X", {"test2_a", "test2_b", "test2_c"}); + op->SetOutput("Out", {"test2_out"}); + + prog.MutableBlock(0)->Var("test2_a")->SetType(proto::VarType::LOD_TENSOR); + prog.MutableBlock(0)->Var("test2_a")->SetShape({32, 64, 128, 128}); + prog.MutableBlock(0)->Var("test2_b")->SetType(proto::VarType::LOD_TENSOR); + prog.MutableBlock(0)->Var("test2_c")->SetType(proto::VarType::LOD_TENSOR); + prog.MutableBlock(0)->Var("test2_out"); + prog.MutableBlock(0)->Var("test2_out")->SetShape({32, 16, 128, 128}); + + auto& infer_inplace = OpInfoMap::Instance().Get(op->Type()).infer_inplace_; + auto in_to_outs = infer_inplace(*op, op->Block()); + EXPECT_EQ(in_to_outs.size(), 1ul); + auto it = in_to_outs.begin(); + EXPECT_EQ(it->first, "test2_a"); + EXPECT_EQ(it->second, "test2_out"); +} + +TEST(InferInplace, SingleGradOpInplaceInToOut) { + ProgramDesc prog; + auto* op = prog.MutableBlock(0)->AppendOp(); + op->SetType("single_op_grad"); + op->SetInput(GradVarName("Out"), {"test2_out"}); + op->SetOutput(GradVarName("X"), {"test2_a", "test2_b", "test2_c"}); + + prog.MutableBlock(0)->Var("test2_a")->SetType(proto::VarType::LOD_TENSOR); + prog.MutableBlock(0)->Var("test2_a")->SetShape({32, 16, 1024, 1024}); + prog.MutableBlock(0)->Var("test2_b")->SetType(proto::VarType::LOD_TENSOR); + prog.MutableBlock(0)->Var("test2_c")->SetType(proto::VarType::LOD_TENSOR); + prog.MutableBlock(0)->Var("test2_out"); + prog.MutableBlock(0)->Var("test2_out")->SetShape({32, 16, 1024, 1024}); + + auto& infer_inplace = OpInfoMap::Instance().Get(op->Type()).infer_inplace_; + auto in_to_outs = infer_inplace(*op, op->Block()); + EXPECT_EQ(in_to_outs.size(), 1ul); + auto it = in_to_outs.begin(); + EXPECT_EQ(it->first, "test2_out"); + EXPECT_EQ(it->second, "test2_a"); +} + +TEST(InferInplace, MultiOutInplaceInToOut) { + ProgramDesc prog; + auto* op = prog.MutableBlock(0)->AppendOp(); + op->SetType("multi_out_op"); + op->SetInput("X", {"a0", "a1"}); + op->SetInput("Y", {"b0"}); + op->SetInput("Z", {"c0", "c1"}); + op->SetOutput("Out", {"o0"}); + op->SetOutput("YOut", {"y0"}); + op->SetOutput("ZOut", {"z0"}); + + prog.MutableBlock(0)->Var("a0")->SetType(proto::VarType::LOD_TENSOR); + prog.MutableBlock(0)->Var("b0")->SetType(proto::VarType::LOD_TENSOR); + prog.MutableBlock(0)->Var("c0")->SetType(proto::VarType::LOD_TENSOR); + prog.MutableBlock(0)->Var("c1")->SetType(proto::VarType::LOD_TENSOR); + prog.MutableBlock(0)->Var("o0"); + prog.MutableBlock(0)->Var("y0"); + prog.MutableBlock(0)->Var("z0"); + prog.MutableBlock(0)->Var("a0")->SetShape({32, 16, 1024, 1024}); + prog.MutableBlock(0)->Var("b0")->SetShape({32, 16, 1024, 1024}); + prog.MutableBlock(0)->Var("c0")->SetShape({32, 16, 1024, 1024}); + prog.MutableBlock(0)->Var("o0")->SetShape({32, 16, 1024, 1024}); + prog.MutableBlock(0)->Var("y0")->SetShape({32, 16, 1024, 1024}); + prog.MutableBlock(0)->Var("z0")->SetShape({32, 16, 1024, 1024}); + + auto& infer_inplace = OpInfoMap::Instance().Get(op->Type()).infer_inplace_; + auto in_to_outs = infer_inplace(*op, op->Block()); + EXPECT_EQ(in_to_outs.size(), 3ul); + std::unordered_map expects = { + {"a0", "o0"}, {"b0", "y0"}, {"c0", "z0"}, + }; + EXPECT_TRUE(expects == in_to_outs); +} + +TEST(InferInplace, MultiGradInplaceInToOut) { + ProgramDesc prog; + auto* op = prog.MutableBlock(0)->AppendOp(); + op->SetType("multi_out_grad"); + op->SetInput(GradVarName("Out"), {"o0"}); + op->SetInput(GradVarName("YOut"), {"y0"}); + op->SetInput(GradVarName("ZOut"), {"z0"}); + op->SetOutput(GradVarName("X"), {"a0", "a1"}); + op->SetOutput(GradVarName("Y"), {"b0"}); + op->SetOutput(GradVarName("Z"), {"c0", "c1"}); + + prog.MutableBlock(0)->Var("a0")->SetType(proto::VarType::LOD_TENSOR); + prog.MutableBlock(0)->Var("b0")->SetType(proto::VarType::LOD_TENSOR); + prog.MutableBlock(0)->Var("c0")->SetType(proto::VarType::LOD_TENSOR); + prog.MutableBlock(0)->Var("c1")->SetType(proto::VarType::LOD_TENSOR); + prog.MutableBlock(0)->Var("o0"); + prog.MutableBlock(0)->Var("y0"); + prog.MutableBlock(0)->Var("z0"); + prog.MutableBlock(0)->Var("a0")->SetShape({32, 16, 1024, 1024}); + prog.MutableBlock(0)->Var("b0")->SetShape({32, 16, 1024, 1024}); + prog.MutableBlock(0)->Var("c0")->SetShape({32, 16, 1024, 1024}); + prog.MutableBlock(0)->Var("o0")->SetShape({32, 16, 1024, 1024}); + prog.MutableBlock(0)->Var("y0")->SetShape({32, 16, 1024, 1024}); + prog.MutableBlock(0)->Var("z0")->SetShape({32, 16, 1024, 1024}); + + auto& infer_inplace = OpInfoMap::Instance().Get(op->Type()).infer_inplace_; + auto in_to_outs = infer_inplace(*op, op->Block()); + + EXPECT_EQ(in_to_outs.size(), 3ul); + std::unordered_map expects = { + {"o0", "a0"}, {"y0", "b0"}, {"z0", "c0"}, + }; + EXPECT_TRUE(expects == in_to_outs); +} + +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/ir/CMakeLists.txt b/paddle/fluid/framework/ir/CMakeLists.txt index b118dccd1b3de881b4791bff6cd331726c8e05da..25d9afbcc8b2bc89ec47654f0dba4cb838be55b0 100644 --- a/paddle/fluid/framework/ir/CMakeLists.txt +++ b/paddle/fluid/framework/ir/CMakeLists.txt @@ -10,8 +10,22 @@ function(pass_library TARGET DEST) set(options "") set(oneValueArgs "") set(multiValueArgs SRCS DEPS) + set(targetPrefix "") + + # Get optional argument + set(extraMacroArgs ${ARGN}) + list(LENGTH extraMacroArgs numExtraMacroArgs) + if(numExtraMacroArgs GREATER 0) + list(GET extraMacroArgs 0 targetPrefix) + endif() + cmake_parse_arguments(op_library "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) - cc_library(${TARGET} SRCS ${TARGET}.cc DEPS graph_pattern_detector pass fuse_pass_base ${op_library_DEPS}) + if(targetPrefix) + cc_library(${TARGET} SRCS ${targetPrefix}/${TARGET}.cc DEPS graph_pattern_detector pass fuse_pass_base ${op_library_DEPS}) + else() + cc_library(${TARGET} SRCS ${TARGET}.cc DEPS graph_pattern_detector pass fuse_pass_base ${op_library_DEPS}) + endif() + # add more DEST here, such as train, dist and collect USE_PASS into a file automatically. if (${DEST} STREQUAL "base" OR ${DEST} STREQUAL "inference") message(STATUS "add pass ${TARGET} ${DEST}") @@ -51,6 +65,7 @@ pass_library(conv_elementwise_add2_act_fuse_pass inference) pass_library(conv_elementwise_add_fuse_pass inference) pass_library(conv_affine_channel_fuse_pass inference) pass_library(transpose_flatten_concat_fuse_pass inference) +pass_library(identity_scale_op_clean_pass base) # There may be many transpose-flatten structures in a model, and the output of # these structures will be used as inputs to the concat Op. This pattern will @@ -62,11 +77,11 @@ foreach (index RANGE 3 6) endforeach() if(WITH_MKLDNN) - pass_library(mkldnn_placement_pass base) - pass_library(depthwise_conv_mkldnn_pass base) - pass_library(conv_bias_mkldnn_fuse_pass inference) - pass_library(conv_relu_mkldnn_fuse_pass inference) - pass_library(conv_elementwise_add_mkldnn_fuse_pass inference) + pass_library(mkldnn_placement_pass base mkldnn) + pass_library(depthwise_conv_mkldnn_pass base mkldnn) + pass_library(conv_bias_mkldnn_fuse_pass inference mkldnn) + pass_library(conv_relu_mkldnn_fuse_pass inference mkldnn) + pass_library(conv_elementwise_add_mkldnn_fuse_pass inference mkldnn) endif() cc_library(fuse_elewise_add_act_pass SRCS fuse_elewise_add_act_pass.cc DEPS pass graph_pattern_detector ) @@ -86,7 +101,8 @@ cc_test(test_fc_fuse_pass SRCS fc_fuse_pass_tester.cc DEPS fc_fuse_pass framewor cc_test(test_seqpool_concat_fuse_pass SRCS seqpool_concat_fuse_pass_tester.cc DEPS seqpool_concat_fuse_pass framework_proto) cc_test(test_is_test_pass SRCS is_test_pass_tester.cc DEPS is_test_pass) if (WITH_MKLDNN) - cc_test(test_depthwise_conv_mkldnn_pass SRCS depthwise_conv_mkldnn_pass_tester.cc DEPS depthwise_conv_mkldnn_pass) - cc_test(test_conv_relu_mkldnn_fuse_pass SRCS conv_relu_mkldnn_fuse_pass_tester.cc DEPS conv_relu_mkldnn_fuse_pass) - cc_test(test_conv_elementwise_add_mkldnn_fuse_pass SRCS conv_elementwise_add_mkldnn_fuse_pass_tester.cc DEPS conv_elementwise_add_mkldnn_fuse_pass) + cc_test(test_depthwise_conv_mkldnn_pass SRCS mkldnn/depthwise_conv_mkldnn_pass_tester.cc DEPS depthwise_conv_mkldnn_pass) + cc_test(test_conv_bias_mkldnn_fuse_pass SRCS mkldnn/conv_bias_mkldnn_fuse_pass_tester.cc DEPS conv_bias_mkldnn_fuse_pass naive_executor) + cc_test(test_conv_relu_mkldnn_fuse_pass SRCS mkldnn/conv_relu_mkldnn_fuse_pass_tester.cc DEPS conv_relu_mkldnn_fuse_pass) + cc_test(test_conv_elementwise_add_mkldnn_fuse_pass SRCS mkldnn/conv_elementwise_add_mkldnn_fuse_pass_tester.cc DEPS conv_elementwise_add_mkldnn_fuse_pass) endif () diff --git a/paddle/fluid/framework/ir/attention_lstm_fuse_pass.h b/paddle/fluid/framework/ir/attention_lstm_fuse_pass.h index a756dfc1b98e1de55c809c73e2c4df1e628950ae..39b0585d3a6f9b52c9ec4b0a24f8532a3410851a 100644 --- a/paddle/fluid/framework/ir/attention_lstm_fuse_pass.h +++ b/paddle/fluid/framework/ir/attention_lstm_fuse_pass.h @@ -22,7 +22,8 @@ namespace ir { class AttentionLSTMFusePass : public FusePassBase { protected: - std::unique_ptr ApplyImpl(std::unique_ptr graph) const; + std::unique_ptr ApplyImpl( + std::unique_ptr graph) const override; }; } // namespace ir diff --git a/paddle/fluid/framework/ir/conv_affine_channel_fuse_pass.h b/paddle/fluid/framework/ir/conv_affine_channel_fuse_pass.h index ad966e11e6222a4ed4c730089c454b0d1c7bd0b3..8c3c8b56c08cc09e66b20d17bf730edec0499f35 100644 --- a/paddle/fluid/framework/ir/conv_affine_channel_fuse_pass.h +++ b/paddle/fluid/framework/ir/conv_affine_channel_fuse_pass.h @@ -31,7 +31,8 @@ class ConvAffineChannelFusePass : public FusePassBase { virtual ~ConvAffineChannelFusePass() {} protected: - std::unique_ptr ApplyImpl(std::unique_ptr graph) const; + std::unique_ptr ApplyImpl( + std::unique_ptr graph) const override; const std::string name_scope_{"conv_affine_channel_fuse"}; }; @@ -40,7 +41,8 @@ class ConvEltwiseAddAffineChannelFusePass : public FusePassBase { virtual ~ConvEltwiseAddAffineChannelFusePass() {} protected: - std::unique_ptr ApplyImpl(std::unique_ptr graph) const; + std::unique_ptr ApplyImpl( + std::unique_ptr graph) const override; const std::string name_scope_{"conv_eltwiseadd_affine_channel_fuse"}; }; diff --git a/paddle/fluid/framework/ir/conv_bn_fuse_pass.cc b/paddle/fluid/framework/ir/conv_bn_fuse_pass.cc index 846a14e365e6bd7f056d409130a3b246371931da..04765dd1440331fb37ed2eb05a9ce762eb2b81bc 100644 --- a/paddle/fluid/framework/ir/conv_bn_fuse_pass.cc +++ b/paddle/fluid/framework/ir/conv_bn_fuse_pass.cc @@ -169,7 +169,7 @@ std::unique_ptr ConvBNFusePass::ApplyImpl( if (has_bias && conv->Op()->Input("Bias").size() > 0) { // reuse existing conv bias node auto conv_bias_names = conv->Op()->Input("Bias"); - PADDLE_ENFORCE_EQ(conv_bias_names.size(), 1); + PADDLE_ENFORCE_EQ(conv_bias_names.size(), 1UL); auto* conv_bias_var = scope->FindVar(conv_bias_names[0]); auto* conv_bias_tensor = conv_bias_var->GetMutable(); PADDLE_ENFORCE_EQ(conv_bias_tensor->dims(), diff --git a/paddle/fluid/framework/ir/conv_bn_fuse_pass.h b/paddle/fluid/framework/ir/conv_bn_fuse_pass.h index 2c9eb574fe8e054e0ae221f08f664b91f05d95c9..cf425a2730904d4ab21c33e66b72db0692cb087c 100644 --- a/paddle/fluid/framework/ir/conv_bn_fuse_pass.h +++ b/paddle/fluid/framework/ir/conv_bn_fuse_pass.h @@ -31,7 +31,8 @@ class ConvBNFusePass : public FusePassBase { virtual ~ConvBNFusePass() {} protected: - std::unique_ptr ApplyImpl(std::unique_ptr graph) const; + std::unique_ptr ApplyImpl( + std::unique_ptr graph) const override; const std::string name_scope_{"conv_bn_fuse"}; }; @@ -40,7 +41,8 @@ class ConvEltwiseAddBNFusePass : public FusePassBase { virtual ~ConvEltwiseAddBNFusePass() {} protected: - std::unique_ptr ApplyImpl(std::unique_ptr graph) const; + std::unique_ptr ApplyImpl( + std::unique_ptr graph) const override; const std::string name_scope_{"conv_eltwiseadd_bn_fuse"}; }; diff --git a/paddle/fluid/framework/ir/conv_elementwise_add2_act_fuse_pass.h b/paddle/fluid/framework/ir/conv_elementwise_add2_act_fuse_pass.h index 3b40a5a92665c07bc2b66e6a96721f573d40393f..9259a4ac5c89b1a7d1413fb2eaaa5fc6a70348f2 100644 --- a/paddle/fluid/framework/ir/conv_elementwise_add2_act_fuse_pass.h +++ b/paddle/fluid/framework/ir/conv_elementwise_add2_act_fuse_pass.h @@ -25,7 +25,8 @@ class ConvElementwiseAdd2ActFusePass : public FusePassBase { virtual ~ConvElementwiseAdd2ActFusePass() {} protected: - std::unique_ptr ApplyImpl(std::unique_ptr graph) const; + std::unique_ptr ApplyImpl( + std::unique_ptr graph) const override; }; } // namespace ir diff --git a/paddle/fluid/framework/ir/conv_elementwise_add_act_fuse_pass.h b/paddle/fluid/framework/ir/conv_elementwise_add_act_fuse_pass.h index ac69aa6458fc8c19b670dea2af1251c44dc353a8..9c0b50f155821cf2bd815a6fb087e3f6cc513641 100644 --- a/paddle/fluid/framework/ir/conv_elementwise_add_act_fuse_pass.h +++ b/paddle/fluid/framework/ir/conv_elementwise_add_act_fuse_pass.h @@ -25,7 +25,8 @@ class ConvElementwiseAddActFusePass : public FusePassBase { virtual ~ConvElementwiseAddActFusePass() {} protected: - std::unique_ptr ApplyImpl(std::unique_ptr graph) const; + std::unique_ptr ApplyImpl( + std::unique_ptr graph) const override; }; } // namespace ir diff --git a/paddle/fluid/framework/ir/conv_elementwise_add_fuse_pass.h b/paddle/fluid/framework/ir/conv_elementwise_add_fuse_pass.h index f234603f5856a9238164f7fb0e5cc81ea9b7ed60..bf43bd5ce2602a3e240c56f00f66f13b79151002 100644 --- a/paddle/fluid/framework/ir/conv_elementwise_add_fuse_pass.h +++ b/paddle/fluid/framework/ir/conv_elementwise_add_fuse_pass.h @@ -25,7 +25,8 @@ class ConvElementwiseAddFusePass : public FusePassBase { virtual ~ConvElementwiseAddFusePass() {} protected: - std::unique_ptr ApplyImpl(std::unique_ptr graph) const; + std::unique_ptr ApplyImpl( + std::unique_ptr graph) const override; }; } // namespace ir diff --git a/paddle/fluid/framework/ir/embedding_fc_lstm_fuse_pass.h b/paddle/fluid/framework/ir/embedding_fc_lstm_fuse_pass.h index e5ad3067ec4060e41f1464395f3fc76183de3e66..fde2a0a4eecdec9ad5ac58ad8e63c26cce482682 100644 --- a/paddle/fluid/framework/ir/embedding_fc_lstm_fuse_pass.h +++ b/paddle/fluid/framework/ir/embedding_fc_lstm_fuse_pass.h @@ -14,6 +14,8 @@ #pragma once +#include + #include "paddle/fluid/framework/ir/fuse_pass_base.h" #include "paddle/fluid/framework/ir/graph.h" #include "paddle/fluid/framework/ir/graph_pattern_detector.h" @@ -30,7 +32,8 @@ class EmbeddingFCLSTMFusePass : public FusePassBase { virtual ~EmbeddingFCLSTMFusePass() {} protected: - std::unique_ptr ApplyImpl(std::unique_ptr graph) const; + std::unique_ptr ApplyImpl( + std::unique_ptr graph) const override; const std::string name_scope_{"embedding_fc_lstm_fuse"}; }; diff --git a/paddle/fluid/framework/ir/fc_fuse_pass.h b/paddle/fluid/framework/ir/fc_fuse_pass.h index 6c69539d1e48268afc2435f8f73b3818d13107cd..783a052edcf84c8c437a7b2e25f0d67c0366691e 100644 --- a/paddle/fluid/framework/ir/fc_fuse_pass.h +++ b/paddle/fluid/framework/ir/fc_fuse_pass.h @@ -12,6 +12,8 @@ // See the License for the specific language governing permissions and // limitations under the License. +#pragma once + #include "paddle/fluid/framework/ir/fuse_pass_base.h" #include "paddle/fluid/framework/ir/graph.h" #include "paddle/fluid/framework/ir/graph_pattern_detector.h" @@ -29,7 +31,8 @@ class FCFusePass : public FusePassBase { virtual ~FCFusePass() {} protected: - std::unique_ptr ApplyImpl(std::unique_ptr graph) const; + std::unique_ptr ApplyImpl( + std::unique_ptr graph) const override; }; } // namespace ir diff --git a/paddle/fluid/framework/ir/fc_gru_fuse_pass.h b/paddle/fluid/framework/ir/fc_gru_fuse_pass.h index 63e1c72bfb2e2641ae5d44858b342d5e427e9045..e359a3289440fffbec622488ecf3a7f49e986574 100644 --- a/paddle/fluid/framework/ir/fc_gru_fuse_pass.h +++ b/paddle/fluid/framework/ir/fc_gru_fuse_pass.h @@ -30,7 +30,8 @@ class FCGRUFusePass : public FusePassBase { virtual ~FCGRUFusePass() {} protected: - std::unique_ptr ApplyImpl(std::unique_ptr graph) const; + std::unique_ptr ApplyImpl( + std::unique_ptr graph) const override; const std::string name_scope_{"fc_gru_fuse"}; }; @@ -41,7 +42,8 @@ class MulGRUFusePass : public FusePassBase { virtual ~MulGRUFusePass() {} protected: - std::unique_ptr ApplyImpl(std::unique_ptr graph) const; + std::unique_ptr ApplyImpl( + std::unique_ptr graph) const override; const std::string name_scope_{"fc_nobias_gru_fuse"}; }; diff --git a/paddle/fluid/framework/ir/fc_lstm_fuse_pass.h b/paddle/fluid/framework/ir/fc_lstm_fuse_pass.h index 3ee32c63a46fcc34bdccd1e14d4bbaf9668c49e9..21482615a6efef930b7328594477a51f4aaf28e7 100644 --- a/paddle/fluid/framework/ir/fc_lstm_fuse_pass.h +++ b/paddle/fluid/framework/ir/fc_lstm_fuse_pass.h @@ -14,6 +14,8 @@ #pragma once +#include + #include "paddle/fluid/framework/ir/fuse_pass_base.h" #include "paddle/fluid/framework/ir/graph.h" #include "paddle/fluid/framework/ir/graph_pattern_detector.h" @@ -30,7 +32,8 @@ class FCLstmFusePass : public FusePassBase { virtual ~FCLstmFusePass() {} protected: - std::unique_ptr ApplyImpl(std::unique_ptr graph) const; + std::unique_ptr ApplyImpl( + std::unique_ptr graph) const override; const std::string name_scope_{"fc_lstm_fuse"}; }; @@ -40,7 +43,8 @@ class MulLstmFusePass : public FusePassBase { virtual ~MulLstmFusePass() {} protected: - std::unique_ptr ApplyImpl(std::unique_ptr graph) const; + std::unique_ptr ApplyImpl( + std::unique_ptr graph) const override; const std::string name_scope_{"fc_nobias_lstm_fuse"}; }; diff --git a/paddle/fluid/framework/ir/fuse_elewise_add_act_pass.h b/paddle/fluid/framework/ir/fuse_elewise_add_act_pass.h index b2fecc076efca333539fe81e67eee222873aee2a..0fee5274478e8b8db852774077ff5979f0aaba25 100644 --- a/paddle/fluid/framework/ir/fuse_elewise_add_act_pass.h +++ b/paddle/fluid/framework/ir/fuse_elewise_add_act_pass.h @@ -32,7 +32,8 @@ class FuseElewiseAddActPass : public FusePassBase { virtual ~FuseElewiseAddActPass() {} protected: - std::unique_ptr ApplyImpl(std::unique_ptr graph) const; + std::unique_ptr ApplyImpl( + std::unique_ptr graph) const override; std::unique_ptr FuseElewiseAddAct( std::unique_ptr graph, diff --git a/paddle/fluid/framework/ir/fuse_relu_depthwise_conv_pass.cc b/paddle/fluid/framework/ir/fuse_relu_depthwise_conv_pass.cc index 0d94008ea82d0e09732d4b6448fdded94b60733c..fe844caed2e757fb080dcee398c8903b929b06e5 100644 --- a/paddle/fluid/framework/ir/fuse_relu_depthwise_conv_pass.cc +++ b/paddle/fluid/framework/ir/fuse_relu_depthwise_conv_pass.cc @@ -111,7 +111,7 @@ std::unique_ptr FuseReluDepthwiseConvPass::FuseReluDepthwiseConv( xg_var = subgraph.at(xg)->Var(); } - PADDLE_ENFORCE_EQ(layer_op->Input("Input").size(), 1); + PADDLE_ENFORCE_EQ(layer_op->Input("Input").size(), 1UL); PADDLE_ENFORCE_EQ(layer_op->Input("Input")[0], y_var->Name()); layer_op->SetInput("Input", {x_var->Name()}); subgraph.at(layer)->inputs.push_back(subgraph.at(x)); @@ -119,13 +119,13 @@ std::unique_ptr FuseReluDepthwiseConvPass::FuseReluDepthwiseConv( VLOG(4) << "replace " << y_var->Name() << " -> " << x_var->Name(); if (!only_forward) { - PADDLE_ENFORCE_EQ(layer_g_op->Input("Input").size(), 1); + PADDLE_ENFORCE_EQ(layer_g_op->Input("Input").size(), 1UL); PADDLE_ENFORCE_EQ(layer_g_op->Input("Input")[0], y_var->Name()); layer_g_op->SetInput("Input", {x_var->Name()}); subgraph.at(layer_g)->inputs.push_back(subgraph.at(x)); subgraph.at(x)->outputs.push_back(subgraph.at(layer_g)); - PADDLE_ENFORCE_EQ(layer_g_op->Output(GradVarName("Input")).size(), 1); + PADDLE_ENFORCE_EQ(layer_g_op->Output(GradVarName("Input")).size(), 1UL); PADDLE_ENFORCE_EQ(layer_g_op->Output(GradVarName("Input"))[0], yg_var->Name()); layer_g_op->SetOutput(GradVarName("Input"), {xg_var->Name()}); diff --git a/paddle/fluid/framework/ir/fuse_relu_depthwise_conv_pass.h b/paddle/fluid/framework/ir/fuse_relu_depthwise_conv_pass.h index 6bd653775e42c1ee16051e205e9fa9888ea05eaa..efb49b8300e677f17d9e205800d837b88edfd2e9 100644 --- a/paddle/fluid/framework/ir/fuse_relu_depthwise_conv_pass.h +++ b/paddle/fluid/framework/ir/fuse_relu_depthwise_conv_pass.h @@ -32,7 +32,8 @@ class FuseReluDepthwiseConvPass : public FusePassBase { virtual ~FuseReluDepthwiseConvPass() {} protected: - std::unique_ptr ApplyImpl(std::unique_ptr graph) const; + std::unique_ptr ApplyImpl( + std::unique_ptr graph) const override; std::unique_ptr FuseReluDepthwiseConv( std::unique_ptr graph, bool only_forward) const; }; diff --git a/paddle/fluid/framework/ir/graph.cc b/paddle/fluid/framework/ir/graph.cc index 3eb5bdba3b7275f45cdfc6ad47f75e7a423541d0..4b5c846f3271b2dd5e094020571069aff590cd2b 100644 --- a/paddle/fluid/framework/ir/graph.cc +++ b/paddle/fluid/framework/ir/graph.cc @@ -76,7 +76,7 @@ std::map> Graph::InitFromProgram( var->inputs.push_back(node); } } - return std::move(var_nodes); + return var_nodes; } void Graph::ResolveHazard( diff --git a/paddle/fluid/framework/ir/graph.h b/paddle/fluid/framework/ir/graph.h index 8bb3c27bdd32d07d58913db043569f6a3bf69aeb..296f3b83961c1379ee2c1237aa15784791b46878 100644 --- a/paddle/fluid/framework/ir/graph.h +++ b/paddle/fluid/framework/ir/graph.h @@ -26,6 +26,14 @@ limitations under the License. */ namespace paddle { namespace framework { + +namespace details { + +// This attr is not recommended, because the graph should not dependence +// the program once it is built. +constexpr char kAllOpDescs[] = "all_op_descs"; +} // namespace details + namespace ir { /* @@ -141,7 +149,8 @@ class Graph { ir::Node *CreateControlDepVar() { // TODO(panyx0718): control var name should be really unique. const std::string name = string::Sprintf( - "%s@%llu", ir::Node::kControlDepVarName, node_set_.size()); + "%s@%llu", static_cast(ir::Node::kControlDepVarName), + num_node_created_); auto *x = AddNode(new ir::Node(name, ir::Node::Type::kVariable)); x->SetId(num_node_created_++); return x; @@ -167,10 +176,13 @@ class Graph { return ret; } - void RemoveNode(ir::Node *node) { + std::unique_ptr RemoveNode(ir::Node *node) { PADDLE_ENFORCE(node_set_.find(node) != node_set_.end()); - node_set_.erase(node); + std::unique_ptr ret; + ret.reset(nodes_.at(node).release()); nodes_.erase(node); + node_set_.erase(node); + return ret; } // NOTE low performance, but simple and secure. @@ -183,13 +195,6 @@ class Graph { return nullptr; } - void ResolveHazard( - const std::map> &var_nodes); - - private: - std::map> InitFromProgram( - const ProgramDesc &program); - // This method takes ownership of `node`. ir::Node *AddNode(ir::Node *node) { PADDLE_ENFORCE(node_set_.find(node) == node_set_.end()); @@ -198,6 +203,13 @@ class Graph { return node; } + void ResolveHazard( + const std::map> &var_nodes); + + private: + std::map> InitFromProgram( + const ProgramDesc &program); + // NOTE: program_ shouldn't be exposed to user. const ProgramDesc program_; std::map attrs_; diff --git a/paddle/fluid/framework/ir/graph_helper.cc b/paddle/fluid/framework/ir/graph_helper.cc index 8de93cf285e4bf34c2d2bf425fa5f3459704b3d6..22d4c0a91cc1638264a8c57aa2841ff4e65a1400 100644 --- a/paddle/fluid/framework/ir/graph_helper.cc +++ b/paddle/fluid/framework/ir/graph_helper.cc @@ -52,16 +52,29 @@ bool HasCircleHelper( ir::Node *node, const std::map> &adj_list, std::unordered_set *visited, - std::unordered_set *in_trace) { + std::unordered_set *in_trace, + std::vector> *circles) { if (visited->find(node) == visited->end()) { visited->insert(node); in_trace->insert(node); for (ir::Node *in : adj_list.at(node)) { if (visited->find(in) == visited->end() && - HasCircleHelper(in, adj_list, visited, in_trace)) { + HasCircleHelper(in, adj_list, visited, in_trace, circles)) { return true; } else if (in_trace->find(in) != in_trace->end()) { + if (circles != nullptr) { + std::vector circle; + circle.emplace_back(in); + ir::Node *p = in; + for (auto &adj : adj_list.at(p)) { + if (in_trace->count(adj)) { + circle.emplace_back(adj); + p = adj; + } + } + circles->emplace_back(circle); + } return true; } } @@ -71,11 +84,12 @@ bool HasCircleHelper( } bool HasCircleInternal( - const std::map> &adj_list) { + const std::map> &adj_list, + std::vector> *circles) { std::unordered_set visited; std::unordered_set in_trace; for (auto &adj : adj_list) { - if (HasCircleHelper(adj.first, adj_list, &visited, &in_trace)) { + if (HasCircleHelper(adj.first, adj_list, &visited, &in_trace, circles)) { return true; } } @@ -84,13 +98,18 @@ bool HasCircleInternal( } // namespace bool HasCircle(const Graph &graph) { - return HasCircleInternal(BuildOperationAdjList(graph)); + return HasCircleInternal(BuildOperationAdjList(graph), nullptr); +} + +bool FindCircleSubGraph(const Graph &graph, + std::vector> *circles) { + return HasCircleInternal(BuildOperationAdjList(graph), circles); } std::vector TopologySortOperations(const Graph &graph) { std::map> adj_list = BuildOperationAdjList(graph); - PADDLE_ENFORCE(!HasCircleInternal(adj_list)); + PADDLE_ENFORCE(!HasCircleInternal(adj_list, nullptr)); std::unordered_set visited; std::vector ret; for (auto adj : adj_list) { diff --git a/paddle/fluid/framework/ir/graph_helper.h b/paddle/fluid/framework/ir/graph_helper.h index fba4936f2c5c971f6c63a452ec4480ff091db25c..214de9ec7d85aee6021b18866295777e317aa79d 100644 --- a/paddle/fluid/framework/ir/graph_helper.h +++ b/paddle/fluid/framework/ir/graph_helper.h @@ -28,6 +28,11 @@ namespace ir { // Test if the graph contains circle. bool HasCircle(const Graph &graph); +// Find All Circles for debugging, +// store all subgraph in circles. +bool FindCircleSubGraph(const Graph &graph, + std::vector> *circles); + size_t GraphNum(const Graph &graph); // Topology Sort the operations in the graph from inputs to outputs. diff --git a/paddle/fluid/framework/ir/graph_helper_test.cc b/paddle/fluid/framework/ir/graph_helper_test.cc index 260a73ae763bd2cdea9948e4d928377a7c718dda..d8973d5aeda1a2e0650a506b4c916b4346f01e2d 100644 --- a/paddle/fluid/framework/ir/graph_helper_test.cc +++ b/paddle/fluid/framework/ir/graph_helper_test.cc @@ -195,6 +195,17 @@ void BuildTwoGraphs(Graph* g) { // v4->outputs.push_back(o5); } +TEST(GraphHelperTest, Circles) { + ProgramDesc prog; + + Graph g(prog); + BuildCircleGraph(&g); + + std::vector> circles; + ASSERT_TRUE(FindCircleSubGraph(g, &circles)); + ASSERT_EQ(circles.size(), 1UL); +} + TEST(GraphHelperTest, GraphNum) { ProgramDesc prog; diff --git a/paddle/fluid/framework/ir/graph_pattern_detector.cc b/paddle/fluid/framework/ir/graph_pattern_detector.cc index 6282ced1e47329915bb3626b410e55ad8251071d..c0c34d186b00814fe6c6fd42beb78133233a1357 100644 --- a/paddle/fluid/framework/ir/graph_pattern_detector.cc +++ b/paddle/fluid/framework/ir/graph_pattern_detector.cc @@ -38,7 +38,7 @@ size_t PDPattern::id_ = 0UL; PDNode *PDPattern::NewNode(const std::string &name) { if (!name.empty()) { - PADDLE_ENFORCE_EQ(node_map_.count(name), 0, + PADDLE_ENFORCE_EQ(node_map_.count(name), 0UL, "PDNode's name should be unique, get duplicate [%s]", name); } @@ -51,7 +51,7 @@ PDNode *PDPattern::NewNode(const std::string &name) { PDNode *PDPattern::NewNode(PDNode::teller_t &&teller, const std::string &name) { if (!name.empty()) { - PADDLE_ENFORCE_EQ(node_map_.count(name), 0, + PADDLE_ENFORCE_EQ(node_map_.count(name), 0UL, "PDNode's name should be unique, get duplicate [%s]", name); } @@ -117,11 +117,6 @@ bool GraphPatternDetector::MarkPDNodesInGraph(const ir::Graph &graph) { // return false; } } - for (auto &item : pdnodes2nodes_) { - for (auto &n : item.second) { - GetMarkedNodes(const_cast(&graph)).insert(n); - } - } VLOG(3) << pdnodes2nodes_.size() << " nodes marked"; return !pdnodes2nodes_.empty(); diff --git a/paddle/fluid/framework/ir/identity_scale_op_clean_pass.cc b/paddle/fluid/framework/ir/identity_scale_op_clean_pass.cc new file mode 100644 index 0000000000000000000000000000000000000000..5bdc0c5faed7131b873edf9b43c847c010b6e3f3 --- /dev/null +++ b/paddle/fluid/framework/ir/identity_scale_op_clean_pass.cc @@ -0,0 +1,84 @@ +// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/framework/ir/identity_scale_op_clean_pass.h" +#include +#include "paddle/fluid/framework/ir/graph_pattern_detector.h" + +namespace paddle { +namespace framework { +namespace ir { + +std::unique_ptr IdentityScaleOpCleanPass::ApplyImpl( + std::unique_ptr graph) const { + FusePassBase::Init("identity_scale_op_clean", graph.get()); + + // pre_op -> scale_in -> scale_op -> scale_out + // -> + // pre_op -> scale_out + GraphPatternDetector detector; + auto pre_op = detector.mutable_pattern()->NewNode("pre_op")->assert_is_op(); + auto scale_in = detector.mutable_pattern() + ->NewNode("scale_in") + ->assert_is_op_input("scale") + ->AsIntermediate(); + auto scale_op = detector.mutable_pattern() + ->NewNode("scale_fuse") + ->assert_is_op("scale") + ->assert_op_attr("scale", 1.) + ->assert_op_attr("bias", 0.); + auto scale_out = + detector.mutable_pattern() + ->NewNode("scale_out") + ->assert_is_op_output("scale") + // scale's output var should has only one consumer, or it can't be + // removed. + ->assert_more([](Node* x) { return x->outputs.size() == 1UL; }); + + pre_op->LinksTo({scale_in}); + scale_op->LinksFrom({scale_in}).LinksTo({scale_out}); + + GraphPatternDetector::handle_t handler = [&]( + const GraphPatternDetector::subgraph_t& subgraph, Graph* graph) { + Node* scale_op_var = subgraph.at(scale_op); + Node* scale_in_var = subgraph.at(scale_in); + Node* scale_out_var = subgraph.at(scale_out); + Node* pre_op_var = subgraph.at(pre_op); + // Link pre_op directly to scale_out + const std::string scale_in_name = scale_in_var->Name(); + const std::string scale_out_name = scale_out_var->Name(); + // Remove links in graph + GraphSafeRemoveNodes(graph, {scale_in_var, scale_op_var}); + // Modify proto message + auto* pre_op_desc = pre_op_var->Op(); + for (auto& parameter : *pre_op_desc->Proto()->mutable_outputs()) { + auto* arguments = parameter.mutable_arguments(); + auto it = std::find(arguments->begin(), arguments->end(), scale_in_name); + PADDLE_ENFORCE(it != arguments->end()); + *it = scale_out_name; + } + + IR_NODE_LINK_TO(pre_op_var, scale_out_var); + }; + + detector(graph.get(), handler); + return graph; +} + +} // namespace ir +} // namespace framework +} // namespace paddle + +REGISTER_PASS(identity_scale_op_clean_pass, + paddle::framework::ir::IdentityScaleOpCleanPass); diff --git a/paddle/fluid/framework/details/memory_early_delete_pass.h b/paddle/fluid/framework/ir/identity_scale_op_clean_pass.h similarity index 71% rename from paddle/fluid/framework/details/memory_early_delete_pass.h rename to paddle/fluid/framework/ir/identity_scale_op_clean_pass.h index 8215aa1b2baa223a111f9050d5488c5fc8ac0e6e..6da592561da1e4046acbfd86c04862f69b7a97a8 100644 --- a/paddle/fluid/framework/details/memory_early_delete_pass.h +++ b/paddle/fluid/framework/ir/identity_scale_op_clean_pass.h @@ -1,4 +1,4 @@ -// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// Copyright (c) 2019 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. @@ -13,20 +13,22 @@ // limitations under the License. #pragma once -#include "paddle/fluid/framework/details/early_delete_op_handle.h" -#include "paddle/fluid/framework/ir/graph.h" -#include "paddle/fluid/framework/ir/pass.h" + +#include "paddle/fluid/framework/ir/fuse_pass_base.h" namespace paddle { namespace framework { -namespace details { +namespace ir { -class MemoryEarlyDeletePass : public ir::Pass { +class IdentityScaleOpCleanPass : public FusePassBase { protected: std::unique_ptr ApplyImpl( std::unique_ptr graph) const override; + + private: + virtual ~IdentityScaleOpCleanPass() = default; }; -} // namespace details +} // namespace ir } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/ir/infer_clean_graph_pass.cc b/paddle/fluid/framework/ir/infer_clean_graph_pass.cc index 7713ed1eab88ee4fa16d52e7425075ae66f721a3..6607c026a748576f38419b275d71217f3eee0c59 100644 --- a/paddle/fluid/framework/ir/infer_clean_graph_pass.cc +++ b/paddle/fluid/framework/ir/infer_clean_graph_pass.cc @@ -37,6 +37,7 @@ class InferCleanGraphPass : public FusePassBase { std::unordered_set invalid_nodes; int valid_op = 0; for (auto* node : graph->Nodes()) { + PADDLE_ENFORCE_NOT_NULL(node); if (is_valid_node(node)) { invalid_nodes.insert(node); } else if (node->IsOp()) { diff --git a/paddle/fluid/framework/ir/lock_free_optimize_pass.h b/paddle/fluid/framework/ir/lock_free_optimize_pass.h index 7310f596f8a3170e84840be4bab8390b780b6577..f9157b10d9554092a5da6a6f73ecf7ceac1430dd 100644 --- a/paddle/fluid/framework/ir/lock_free_optimize_pass.h +++ b/paddle/fluid/framework/ir/lock_free_optimize_pass.h @@ -60,7 +60,8 @@ class LockFreeOptimizePass : public Pass { virtual ~LockFreeOptimizePass() {} protected: - std::unique_ptr ApplyImpl(std::unique_ptr graph) const; + std::unique_ptr ApplyImpl( + std::unique_ptr graph) const override; private: // Create a new sgd node via current optimizer node diff --git a/paddle/fluid/framework/ir/conv_bias_mkldnn_fuse_pass.cc b/paddle/fluid/framework/ir/mkldnn/conv_bias_mkldnn_fuse_pass.cc similarity index 98% rename from paddle/fluid/framework/ir/conv_bias_mkldnn_fuse_pass.cc rename to paddle/fluid/framework/ir/mkldnn/conv_bias_mkldnn_fuse_pass.cc index d4a701e0b173a96d8605dff308fee7007a0ecc0c..5d0b294f6fec5f14dcddb91f8ceffb27fc833d4e 100644 --- a/paddle/fluid/framework/ir/conv_bias_mkldnn_fuse_pass.cc +++ b/paddle/fluid/framework/ir/mkldnn/conv_bias_mkldnn_fuse_pass.cc @@ -12,7 +12,7 @@ // See the License for the specific language governing permissions and // limitations under the License. -#include "paddle/fluid/framework/ir/conv_bias_mkldnn_fuse_pass.h" +#include "paddle/fluid/framework/ir/mkldnn/conv_bias_mkldnn_fuse_pass.h" #include #include #include diff --git a/paddle/fluid/framework/ir/conv_bias_mkldnn_fuse_pass.h b/paddle/fluid/framework/ir/mkldnn/conv_bias_mkldnn_fuse_pass.h similarity index 93% rename from paddle/fluid/framework/ir/conv_bias_mkldnn_fuse_pass.h rename to paddle/fluid/framework/ir/mkldnn/conv_bias_mkldnn_fuse_pass.h index f3ad9f1c2bf14db418629e0c607e2510f01908b8..0ef5c177bf98b354bb18fc1d2ec8e5bef4b58951 100644 --- a/paddle/fluid/framework/ir/conv_bias_mkldnn_fuse_pass.h +++ b/paddle/fluid/framework/ir/mkldnn/conv_bias_mkldnn_fuse_pass.h @@ -29,7 +29,8 @@ class ConvBiasFusePass : public FusePassBase { virtual bool is_conv3d() const { return false; } protected: - std::unique_ptr ApplyImpl(std::unique_ptr graph) const; + std::unique_ptr ApplyImpl( + std::unique_ptr graph) const override; const std::string name_scope_{"conv_bias_mkldnn_fuse"}; }; /* diff --git a/paddle/fluid/framework/ir/mkldnn/conv_bias_mkldnn_fuse_pass_tester.cc b/paddle/fluid/framework/ir/mkldnn/conv_bias_mkldnn_fuse_pass_tester.cc new file mode 100644 index 0000000000000000000000000000000000000000..38b7fe52037c1a264e4251b7a54ef7569ee6d765 --- /dev/null +++ b/paddle/fluid/framework/ir/mkldnn/conv_bias_mkldnn_fuse_pass_tester.cc @@ -0,0 +1,151 @@ +// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/framework/ir/mkldnn/conv_bias_mkldnn_fuse_pass.h" +#include "paddle/fluid/framework/naive_executor.h" +#include "paddle/fluid/platform/place.h" + +#include +#include "paddle/fluid/framework/op_proto_maker.h" + +namespace paddle { +namespace framework { +namespace ir { + +void SetOp(ProgramDesc* prog, const std::string& type, const std::string& name, + const std::vector& inputs, + const std::vector& outputs) { + auto* op = prog->MutableBlock(0)->AppendOp(); + op->SetType(type); + if (type == "conv2d") { + op->SetAttr("use_mkldnn", true); + op->SetAttr("name", name); + op->SetInput("Input", {inputs[0]}); + op->SetInput("Filter", {inputs[1]}); + if (inputs.size() > 2) + op->SetInput("Bias", {inputs[2]}); + else + op->SetInput("Bias", {}); + } else if (type == "elementwise_add") { + op->SetAttr("use_mkldnn", true); + op->SetInput("X", {inputs[0]}); + op->SetInput("Y", {inputs[1]}); + } + op->SetOutput("Out", outputs); + op->SetAttr(OpProtoAndCheckerMaker::OpRoleAttrName(), + static_cast(OpRole::kForward)); +} + +// (c, weights)->conv->f +// (f)->elementwise_add->g +ProgramDesc BuildProgramDesc(bool convWithExistingBias) { + ProgramDesc prog; + std::vector nodes{"c", "weights", "f", "eltwise_bias", "g"}; + if (convWithExistingBias) nodes.push_back("conv_bias"); + for (auto& v : nodes) { + auto* var = prog.MutableBlock(0)->Var(v); + var->SetType(proto::VarType::LOD_TENSOR); + if (v == "weights" || v == "conv_bias" || v == "eltwise_bias") { + var->SetPersistable(true); + } + } + + // conv+bias, both with MKL-DNN + if (convWithExistingBias) { + SetOp(&prog, "conv2d", "conv", + std::vector({"c", "weights", "conv_bias"}), + std::vector({"f"})); + } else { + SetOp(&prog, "conv2d", "conv", std::vector({"c", "weights"}), + std::vector({"f"})); + } + SetOp(&prog, "elementwise_add", "eltwise", + std::vector({"f", "eltwise_bias"}), + std::vector({"g"})); + + return prog; +} + +void InitTensorHolder(Scope* scope, const paddle::platform::Place& place, + const char* var_name) { + auto x = scope->Var(var_name); + auto tensor = x->GetMutable(); + tensor->mutable_data(place, proto::VarType::FP32, + ::paddle::memory::Allocator::kDefault, 1); +} + +void MainTest(bool convWithExistingBias) { + auto prog = BuildProgramDesc(convWithExistingBias); + std::unique_ptr graph(new ir::Graph(prog)); + auto place = paddle::platform::CPUPlace(); + NaiveExecutor exe{place}; + Scope scope; + // Init scope, as it is used in pass + exe.CreateVariables(prog, 0, true, &scope); + if (convWithExistingBias) { + InitTensorHolder(&scope, place, "conv_bias"); + InitTensorHolder(&scope, place, "eltwise_bias"); + } + graph->Set(kParamScopeAttr, new framework::Scope*(&scope)); + + auto pass = PassRegistry::Instance().Get("conv_bias_mkldnn_fuse_pass"); + + int original_nodes_num = graph->Nodes().size(); + + graph = pass->Apply(std::move(graph)); + + int current_nodes_num = graph->Nodes().size(); + + // Remove 3 Nodes: Conv, Bias, conv_out + // Add 1 Node: ConvBias + EXPECT_EQ(original_nodes_num - 2, current_nodes_num); + + // Assert conv_bias op in newly generated graph + int conv_bias_count = 0; + + for (auto* node : graph->Nodes()) { + if (node->IsOp() && node->Op()->Type() == "conv2d") { + auto* op = node->Op(); + ASSERT_TRUE(op->HasAttr("use_mkldnn")); + EXPECT_TRUE(boost::get(op->GetAttr("use_mkldnn"))); + // check if "conv" convolution is fused + auto op_name = boost::get(op->GetAttr("name")); + if (op_name == "conv") { + auto input_names = op->InputNames(); + ASSERT_TRUE(std::find(input_names.begin(), input_names.end(), "Bias") != + input_names.end()); + auto bias = boost::get>(op->Input("Bias")); + if (bias.size()) { + ++conv_bias_count; + } + } + } + } + EXPECT_EQ(conv_bias_count, 1); +} + +TEST(ConvBiasFusePass, bias_free_conv) { MainTest(false); } + +TEST(ConvBiasFusePass, conv_with_existing_bias) { MainTest(true); } + +TEST(ConvBiasFusePass, conv3d) { + Conv3DBiasFusePass pass; + ASSERT_TRUE(pass.is_conv3d()); +} + +} // namespace ir +} // namespace framework +} // namespace paddle + +USE_PASS(conv_bias_mkldnn_fuse_pass); diff --git a/paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.cc b/paddle/fluid/framework/ir/mkldnn/conv_elementwise_add_mkldnn_fuse_pass.cc similarity index 99% rename from paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.cc rename to paddle/fluid/framework/ir/mkldnn/conv_elementwise_add_mkldnn_fuse_pass.cc index a8029e67e659a269f8492cf6e2f1f09040144283..fb3db81347b102cfa264082b36a2e22ea8c22982 100644 --- a/paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.cc +++ b/paddle/fluid/framework/ir/mkldnn/conv_elementwise_add_mkldnn_fuse_pass.cc @@ -12,7 +12,7 @@ // See the License for the specific language governing permissions and // limitations under the License. -#include "paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.h" +#include "paddle/fluid/framework/ir/mkldnn/conv_elementwise_add_mkldnn_fuse_pass.h" #include #include #include diff --git a/paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.h b/paddle/fluid/framework/ir/mkldnn/conv_elementwise_add_mkldnn_fuse_pass.h similarity index 100% rename from paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.h rename to paddle/fluid/framework/ir/mkldnn/conv_elementwise_add_mkldnn_fuse_pass.h diff --git a/paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass_tester.cc b/paddle/fluid/framework/ir/mkldnn/conv_elementwise_add_mkldnn_fuse_pass_tester.cc similarity index 98% rename from paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass_tester.cc rename to paddle/fluid/framework/ir/mkldnn/conv_elementwise_add_mkldnn_fuse_pass_tester.cc index 61ba097fd8cb55e25bda1947ea97d53308c55bd3..9ef5c298b8cddfec094e9544dc6da9afdcaf0dab 100644 --- a/paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass_tester.cc +++ b/paddle/fluid/framework/ir/mkldnn/conv_elementwise_add_mkldnn_fuse_pass_tester.cc @@ -15,8 +15,8 @@ #include #include -#include "paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.h" #include "paddle/fluid/framework/ir/graph_traits.h" +#include "paddle/fluid/framework/ir/mkldnn/conv_elementwise_add_mkldnn_fuse_pass.h" namespace paddle { namespace framework { diff --git a/paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.cc b/paddle/fluid/framework/ir/mkldnn/conv_relu_mkldnn_fuse_pass.cc similarity index 97% rename from paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.cc rename to paddle/fluid/framework/ir/mkldnn/conv_relu_mkldnn_fuse_pass.cc index e359a3832ee8d549f8c58d63bc1cc6564ecadede..4f4605398a665e63662a64a3a925c32d48f10952 100644 --- a/paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.cc +++ b/paddle/fluid/framework/ir/mkldnn/conv_relu_mkldnn_fuse_pass.cc @@ -12,7 +12,7 @@ // See the License for the specific language governing permissions and // limitations under the License. -#include "paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.h" +#include "paddle/fluid/framework/ir/mkldnn/conv_relu_mkldnn_fuse_pass.h" #include #include #include "paddle/fluid/platform/enforce.h" diff --git a/paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.h b/paddle/fluid/framework/ir/mkldnn/conv_relu_mkldnn_fuse_pass.h similarity index 100% rename from paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.h rename to paddle/fluid/framework/ir/mkldnn/conv_relu_mkldnn_fuse_pass.h diff --git a/paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass_tester.cc b/paddle/fluid/framework/ir/mkldnn/conv_relu_mkldnn_fuse_pass_tester.cc similarity index 98% rename from paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass_tester.cc rename to paddle/fluid/framework/ir/mkldnn/conv_relu_mkldnn_fuse_pass_tester.cc index 19248b4dfee1da81d18cd2effac08ba68dde80fb..06d56f6222e4bb9a9969d4ab2d260c97d1ce6c72 100644 --- a/paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass_tester.cc +++ b/paddle/fluid/framework/ir/mkldnn/conv_relu_mkldnn_fuse_pass_tester.cc @@ -12,7 +12,7 @@ // See the License for the specific language governing permissions and // limitations under the License. -#include "paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.h" +#include "paddle/fluid/framework/ir/mkldnn/conv_relu_mkldnn_fuse_pass.h" #include #include "paddle/fluid/framework/op_proto_maker.h" diff --git a/paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass.cc b/paddle/fluid/framework/ir/mkldnn/depthwise_conv_mkldnn_pass.cc similarity index 96% rename from paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass.cc rename to paddle/fluid/framework/ir/mkldnn/depthwise_conv_mkldnn_pass.cc index 19056e18aa892dbc83dfbf7305b6ad8b6b6bc51c..7851e8c84bca2e3b05d3b1603eaa4c0ca5909e10 100644 --- a/paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass.cc +++ b/paddle/fluid/framework/ir/mkldnn/depthwise_conv_mkldnn_pass.cc @@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include "paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass.h" +#include "paddle/fluid/framework/ir/mkldnn/depthwise_conv_mkldnn_pass.h" #include "paddle/fluid/framework/ir/graph_pattern_detector.h" namespace paddle { diff --git a/paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass.h b/paddle/fluid/framework/ir/mkldnn/depthwise_conv_mkldnn_pass.h similarity index 100% rename from paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass.h rename to paddle/fluid/framework/ir/mkldnn/depthwise_conv_mkldnn_pass.h diff --git a/paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass_tester.cc b/paddle/fluid/framework/ir/mkldnn/depthwise_conv_mkldnn_pass_tester.cc similarity index 98% rename from paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass_tester.cc rename to paddle/fluid/framework/ir/mkldnn/depthwise_conv_mkldnn_pass_tester.cc index 09d0b15f46a7e50afb6aea46383013ce6a6c6118..1783e3322b1df8125f580f09a12aefe64d246c1a 100644 --- a/paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass_tester.cc +++ b/paddle/fluid/framework/ir/mkldnn/depthwise_conv_mkldnn_pass_tester.cc @@ -12,7 +12,7 @@ // See the License for the specific language governing permissions and // limitations under the License. -#include "paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass.h" +#include "paddle/fluid/framework/ir/mkldnn/depthwise_conv_mkldnn_pass.h" #include diff --git a/paddle/fluid/framework/ir/mkldnn_placement_pass.cc b/paddle/fluid/framework/ir/mkldnn/mkldnn_placement_pass.cc similarity index 95% rename from paddle/fluid/framework/ir/mkldnn_placement_pass.cc rename to paddle/fluid/framework/ir/mkldnn/mkldnn_placement_pass.cc index 951fcb066ce759ebfec0182e1e9dca887e343170..20e52410ffe3caa86450bc05bf3aabf5a5bce374 100644 --- a/paddle/fluid/framework/ir/mkldnn_placement_pass.cc +++ b/paddle/fluid/framework/ir/mkldnn/mkldnn_placement_pass.cc @@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include "paddle/fluid/framework/ir/mkldnn_placement_pass.h" +#include "paddle/fluid/framework/ir/mkldnn/mkldnn_placement_pass.h" #include namespace paddle { diff --git a/paddle/fluid/framework/ir/mkldnn_placement_pass.h b/paddle/fluid/framework/ir/mkldnn/mkldnn_placement_pass.h similarity index 100% rename from paddle/fluid/framework/ir/mkldnn_placement_pass.h rename to paddle/fluid/framework/ir/mkldnn/mkldnn_placement_pass.h diff --git a/paddle/fluid/framework/ir/repeated_fc_relu_fuse_pass.h b/paddle/fluid/framework/ir/repeated_fc_relu_fuse_pass.h index 3f3f0846eba1201e57a653f8e515c28d2bcdd5e3..ede0bea07ff4130a0f6b3d21d6e34222a5013170 100644 --- a/paddle/fluid/framework/ir/repeated_fc_relu_fuse_pass.h +++ b/paddle/fluid/framework/ir/repeated_fc_relu_fuse_pass.h @@ -31,7 +31,8 @@ class RepeatedFCReluFusePass : public FusePassBase { virtual ~RepeatedFCReluFusePass() {} protected: - std::unique_ptr ApplyImpl(std::unique_ptr graph) const; + std::unique_ptr ApplyImpl( + std::unique_ptr graph) const override; const std::string name_scope_{"repeated_fc_relu_fuse"}; }; diff --git a/paddle/fluid/framework/ir/seq_concat_fc_fuse_pass.h b/paddle/fluid/framework/ir/seq_concat_fc_fuse_pass.h index 9f5fd1a29adf918806d8f30097d8c7f002f48f3e..06e18f9dc327bf2ffaf8d2ab64edcbddea2eb04c 100644 --- a/paddle/fluid/framework/ir/seq_concat_fc_fuse_pass.h +++ b/paddle/fluid/framework/ir/seq_concat_fc_fuse_pass.h @@ -12,6 +12,8 @@ // See the License for the specific language governing permissions and // limitations under the License. +#pragma once + #include "paddle/fluid/framework/ir/fuse_pass_base.h" #include "paddle/fluid/framework/ir/graph.h" #include "paddle/fluid/framework/ir/pass.h" @@ -25,7 +27,8 @@ class SeqConcatFcFusePass : public FusePassBase { virtual ~SeqConcatFcFusePass() {} protected: - std::unique_ptr ApplyImpl(std::unique_ptr graph) const; + std::unique_ptr ApplyImpl( + std::unique_ptr graph) const override; }; } // namespace ir diff --git a/paddle/fluid/framework/ir/seqconv_eltadd_relu_fuse_pass.h b/paddle/fluid/framework/ir/seqconv_eltadd_relu_fuse_pass.h index dac9de71930c1768bdf416520caae6468449cd3d..c36c6b76a238dd21eb0c9308e780761aa9e4e27a 100644 --- a/paddle/fluid/framework/ir/seqconv_eltadd_relu_fuse_pass.h +++ b/paddle/fluid/framework/ir/seqconv_eltadd_relu_fuse_pass.h @@ -28,7 +28,8 @@ class SeqConvEltAddReluFusePass : public FusePassBase { virtual ~SeqConvEltAddReluFusePass() {} protected: - std::unique_ptr ApplyImpl(std::unique_ptr graph) const; + std::unique_ptr ApplyImpl( + std::unique_ptr graph) const override; const std::string name_scope_{"seqconv_eltadd_relu_fuse"}; }; diff --git a/paddle/fluid/framework/ir/seqpool_concat_fuse_pass.h b/paddle/fluid/framework/ir/seqpool_concat_fuse_pass.h index ba2154045e62c687173565c5ad30ea4d45d3c8f4..a5db3528da36ad08bb7f4d2765ee78222c569a5c 100644 --- a/paddle/fluid/framework/ir/seqpool_concat_fuse_pass.h +++ b/paddle/fluid/framework/ir/seqpool_concat_fuse_pass.h @@ -42,7 +42,8 @@ class SeqPoolConcatFusePass : public FusePassBase { virtual ~SeqPoolConcatFusePass() {} protected: - std::unique_ptr ApplyImpl(std::unique_ptr graph) const; + std::unique_ptr ApplyImpl( + std::unique_ptr graph) const override; const std::string name_scope_{"seqpool_concat_fuse"}; }; diff --git a/paddle/fluid/framework/ir/seqpool_concat_fuse_pass_tester.cc b/paddle/fluid/framework/ir/seqpool_concat_fuse_pass_tester.cc index 456a03192cc4e4a9d0dbe2dcb649b6c1b4d9cd5a..35d1d5129bba7043026e5489b806480775473257 100644 --- a/paddle/fluid/framework/ir/seqpool_concat_fuse_pass_tester.cc +++ b/paddle/fluid/framework/ir/seqpool_concat_fuse_pass_tester.cc @@ -164,7 +164,7 @@ ProgramDesc BuildProgramDesc(int num_inputs_of_concat) { }; std::vector concat_inputs; for (int i = 0; i < num_inputs_of_concat; ++i) { - std::string prefix = "seqpool_op_" + i; + std::string prefix = "seqpool_op_" + std::to_string(i); new_var(prefix + "in"); new_var(prefix + "out"); new_var(prefix + "out_unused"); diff --git a/paddle/fluid/framework/ir/squared_mat_sub_fuse_pass.h b/paddle/fluid/framework/ir/squared_mat_sub_fuse_pass.h index fb49adc3768ec99cab4321c6b90c93dfed6d32f2..c21ba65c40a8d54c315ab347e5a8a3266a143779 100644 --- a/paddle/fluid/framework/ir/squared_mat_sub_fuse_pass.h +++ b/paddle/fluid/framework/ir/squared_mat_sub_fuse_pass.h @@ -31,7 +31,8 @@ class SquaredMatSubFusePass : public FusePassBase { virtual ~SquaredMatSubFusePass() {} protected: - std::unique_ptr ApplyImpl(std::unique_ptr graph) const; + std::unique_ptr ApplyImpl( + std::unique_ptr graph) const override; const std::string name_scope_{"squared_mat_sub_fuse"}; }; diff --git a/paddle/fluid/framework/ir/transpose_flatten_concat_fuse_pass.h b/paddle/fluid/framework/ir/transpose_flatten_concat_fuse_pass.h index fb0f0ae9efdc5a25a799d6123fa658a99860cd86..a7d18ec86da1c02aef84c25c378691eb8f651015 100644 --- a/paddle/fluid/framework/ir/transpose_flatten_concat_fuse_pass.h +++ b/paddle/fluid/framework/ir/transpose_flatten_concat_fuse_pass.h @@ -30,7 +30,8 @@ class TransposeFlattenConcatFusePass : public FusePassBase { virtual ~TransposeFlattenConcatFusePass() {} protected: - std::unique_ptr ApplyImpl(std::unique_ptr graph) const; + std::unique_ptr ApplyImpl( + std::unique_ptr graph) const override; }; } // namespace ir diff --git a/paddle/fluid/framework/ngraph_operator.cc b/paddle/fluid/framework/ngraph_operator.cc deleted file mode 100644 index 7e174c7def1ffa4089a94d9cc504b18843557c53..0000000000000000000000000000000000000000 --- a/paddle/fluid/framework/ngraph_operator.cc +++ /dev/null @@ -1,545 +0,0 @@ -/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include - -#include -#include - -#include "paddle/fluid/framework/feed_fetch_type.h" -#include "paddle/fluid/framework/framework.pb.h" -#include "paddle/fluid/framework/lod_tensor.h" -#include "paddle/fluid/framework/ngraph_bridge.h" -#include "paddle/fluid/framework/ngraph_operator.h" -#include "paddle/fluid/framework/tensor.h" -#include "paddle/fluid/framework/var_desc.h" -#include "paddle/fluid/framework/var_type.h" - -#include "ngraph/ngraph.hpp" - -namespace paddle { -namespace framework { - -static ngraph::Shape Ddim2Shape(const DDim& dims) { - ngraph::Shape sp; - for (int i = 0; i < dims.size(); ++i) { - int k = dims[i]; - k = k == 0 ? 1 : k; - sp.push_back(k); - } - return sp; -} - -static std::map pd2ng_type_map = { - {proto::VarType::FP32, ngraph::element::f32}, - {proto::VarType::FP64, ngraph::element::f64}, - {proto::VarType::INT32, ngraph::element::i32}, - {proto::VarType::INT64, ngraph::element::i64}, - {proto::VarType::BOOL, ngraph::element::boolean}, -}; - -typedef enum { /* nGraph support state on ops */ - FULL_TRAIN, /* Support full ops for train */ - PARTIAL_TRAIN, /* Support partial ops for train */ - FULL_TEST, /* Support full list of ops for test */ - PARTIAL_TEST /* Support partial list of ops for test */ -} op_state; - -// perform graph build through bridge and execute computation -class NgraphEngine { - public: - explicit NgraphEngine(const Scope& scope, const platform::Place& place, - const std::vector>& ops, - const std::unordered_map< - std::string, ngraph::element::Type>& var_type_map, - const std::unordered_set& persist, - const std::unordered_set& fetches, - const std::unordered_set& post_op_inputs, - op_state ng_op_state) - : scope_(scope), - place_(place), - fused_ops_(ops), - var_type_map_(var_type_map), - persistables_(persist), - fetches_(fetches), - post_op_inputs_(post_op_inputs), - ng_op_state_(ng_op_state) { - var_in_node_map_ = std::make_shared< - std::unordered_map>>(); - - var_node_map_ = std::make_shared< - std::unordered_map>>(); - - BuildNgIO(); - - GetNgFunction(); - } - - void Run(const Scope& scope, const platform::Place& place) const; - - private: - static std::unordered_map> - func_cache_; - const Scope& scope_; - const platform::Place& place_; - std::vector> fused_ops_; - std::unordered_map var_type_map_; - std::unordered_set persistables_; - std::unordered_set fetches_; - std::unordered_set post_op_inputs_; - op_state ng_op_state_; - - // ngraph backend eg. CPU - static std::shared_ptr backend_; - // ngraph function to call and execute - std::shared_ptr ngraph_function_; - // var_name of inputs - std::vector var_in_; - // var_name of outputs from fetch in order - std::vector var_out_; - // map input vars to nodes - std::shared_ptr< - std::unordered_map>> - var_in_node_map_; - // map each var name with a ngraph node - std::shared_ptr< - std::unordered_map>> - var_node_map_; - // cache key to check if function is cached - std::shared_ptr GetCacheKey(); - // get ngraph input and define ngraph input parameters - void GetNgInputShape(std::shared_ptr op); - // Call ngraph bridge to map ops - void BuildNgNodes(); - // get the ngraph input and output var list - void BuildNgIO(); - // build ngraph function call - void BuildNgFunction(); - // Check cache for ngraph function or otherwise build the function - void GetNgFunction(); -}; - -std::vector>::iterator>> -NgraphOperator::NgraphOpIntervals( - std::vector>* ops) { - std::vector>::iterator>> - intervals; - if (ops->empty()) { - return intervals; - } - size_t size = ops->size(); - size_t left = 0; - while (left < size && ops->at(left)->Type() != kFeedOpType) { - ++left; - } - if (left == size) { - return intervals; - } - while (left < size && ops->at(left)->Type() == kFeedOpType) { - ++left; - } - - size_t right = left; - while (right < size && ops->at(right)->Type() != kFetchOpType) { - ++right; - } - if (right == size) { - return intervals; - } - if (left >= right) return intervals; - - // (left, right - 1) represents indices between feed and fetch - size_t pivot = left; - while (pivot < right) { - auto op_type = ops->at(pivot)->Type(); - if (paddle::framework::NgraphBridge::NG_NODE_MAP.find(op_type) == - paddle::framework::NgraphBridge::NG_NODE_MAP.end()) { - ++pivot; - } else { - size_t start = pivot, end = start; - while (pivot < right && - (paddle::framework::NgraphBridge::NG_NODE_MAP.find( - ops->at(pivot)->Type()) != - paddle::framework::NgraphBridge::NG_NODE_MAP.end())) { - ++pivot; - ++end; - } - std::vector>::iterator> - interval = {ops->begin() + start, ops->begin() + end}; - intervals.push_back(interval); - } - } // end while - - return intervals; -} - -NgraphOperator::NgraphOperator( - const ProgramDesc& prog, size_t block_id, - std::vector>::iterator start, - std::vector>::iterator end, - const std::string& type, const VariableNameMap& inputs, - const VariableNameMap& outputs, const AttributeMap& attrs) - : OperatorBase(type, inputs, outputs, attrs), - pdesc_(prog), - block_(block_id) { - for (std::vector>::iterator it = start; - it != end; ++it) { - fused_ops_.push_back(std::move(*it)); - } - - for (std::vector>::iterator it = end; - (*it)->Type() != kFetchOpType; ++it) { - for (auto& var_name_item : (*it)->Inputs()) { - for (auto& var_name : var_name_item.second) { - post_op_inputs_.insert(var_name); - } - } - } - - if ((*(start - 1))->Type() == kFeedOpType && (*end)->Type() == kFetchOpType) { - is_full_ = true; - } - - Process(); -} - -void NgraphOperator::Process() { - auto& bdesc = pdesc_.Block(block_); - for (auto& var : bdesc.AllVars()) { - if (!(var->GetType() == proto::VarType::SELECTED_ROWS || - var->GetType() == proto::VarType::LOD_TENSOR || - var->GetType() == proto::VarType::LOD_TENSOR_ARRAY)) { - continue; - } - - auto var_name = var->Name(); - if (var->Name() == framework::kEmptyVarName) { - continue; - } - - if (var_name != "fetch" && var_name != "feed") { - auto pd_type = var->GetDataType(); - if (pd2ng_type_map.find(pd_type) == pd2ng_type_map.end()) { - PADDLE_THROW("Data type of var %s not found in pd2ng_type_map", - var_name); - } - var_type_map_[var_name] = pd2ng_type_map[pd_type]; - } - - if (var->Persistable()) { - persistables_.insert(var->Name()); - } - } - - for (auto* op : bdesc.AllOps()) { - if (op->Type() == kFetchOpType) { - std::string fetch_target_name = op->Input("X")[0]; - fetches_.insert(fetch_target_name); - } - } -} - -void NgraphOperator::RunImpl(const Scope& scope, - const platform::Place& place) const { - op_state ng_op_state = PARTIAL_TEST; - auto& bdesc = pdesc_.Block(block_); - for (auto* op : bdesc.AllOps()) { - if (op->Type().find("_grad") != std::string::npos) { - ng_op_state = PARTIAL_TRAIN; - break; - } - } - - if (is_full_) { - ng_op_state = ng_op_state == PARTIAL_TEST ? FULL_TEST : FULL_TRAIN; - } - - NgraphEngine ngraph_engine(scope, place, fused_ops_, var_type_map_, - persistables_, fetches_, post_op_inputs_, - ng_op_state); - ngraph_engine.Run(scope, place); -} - -std::unordered_map> - NgraphEngine::func_cache_ = {}; - -std::shared_ptr NgraphEngine::backend_ = - ngraph::runtime::Backend::create("CPU"); - -void NgraphEngine::GetNgInputShape(std::shared_ptr op) { - RuntimeContext ctx(op->Inputs(), op->Outputs(), scope_); - op->RuntimeInferShape(scope_, place_, ctx); - for (auto& var_name_item : op->Inputs()) { - for (auto& var_name : var_name_item.second) { - auto* var = scope_.FindVar(var_name); - if (var && var->IsType()) { - auto* tensor_pd = GetLoDTensorOrSelectedRowsValueFromVar(*var); - auto sp = Ddim2Shape(tensor_pd->dims()); - if (std::find(var_in_.begin(), var_in_.end(), var_name) != - var_in_.end()) { - if (var_node_map_->find(var_name) == var_node_map_->end()) { - auto ng_type = var_type_map_.at(var_name); - auto prm = - std::make_shared(ng_type, sp, true); - (*var_node_map_)[var_name] = prm; - (*var_in_node_map_)[var_name] = prm; - } - } - } - } - } -} - -void NgraphEngine::BuildNgNodes() { - for (auto& var_name : var_out_) { - if (var_node_map_->find(var_name) == var_node_map_->end()) { - auto* var = scope_.FindVar(var_name); - if (var && var->IsType()) { - auto* tensor_pd = GetLoDTensorOrSelectedRowsValueFromVar(*var); - auto& ddim = tensor_pd->dims(); - auto ng_shape = Ddim2Shape(ddim); - auto ng_type = var_type_map_.at(var_name); - auto prm = - std::make_shared(ng_type, ng_shape, true); - (*var_node_map_)[var_name] = prm; - } - } - } - - paddle::framework::NgraphBridge ngb(var_node_map_); - for (auto& op : fused_ops_) { - ngb.BuildNgNode(op); - } -} - -void NgraphEngine::BuildNgIO() { - std::unordered_set inputs; - std::unordered_set outputs; - - for (auto& op : fused_ops_) { - for (auto& var_name_item : op->Inputs()) { - for (auto& var_name : var_name_item.second) { - inputs.insert(var_name); - const bool is_output = outputs.find(var_name) != outputs.end(); - if (!is_output && - std::find(var_in_.begin(), var_in_.end(), var_name) == - var_in_.end()) { - // fill var_in here to keep lhs and rhs order - var_in_.push_back(var_name); - } - } - } - - if (op->Type() != "fill_constant") { - GetNgInputShape(op); - } - - for (auto& var_name_item : op->Outputs()) { - PADDLE_ENFORCE_LE(var_name_item.second.size(), 1, - "op %s has more than 1 output - Not handling yet", - op->Type()); - for (auto& var_name : var_name_item.second) { - outputs.insert(var_name); - } - } - } - - // var_out.clear(); - for (auto& op : fused_ops_) { - for (auto& var_name_item : op->Outputs()) { - PADDLE_ENFORCE_LE(var_name_item.second.size(), 1, - "op %s has more than 1 output - Not handling yet", - op->Type()); - for (auto& var_name : var_name_item.second) { - switch (ng_op_state_) { - case PARTIAL_TEST: - if (post_op_inputs_.find(var_name) != post_op_inputs_.end() || - fetches_.find(var_name) != fetches_.end()) { - var_out_.push_back(var_name); - } - break; - case FULL_TEST: - if (fetches_.find(var_name) != fetches_.end()) { - var_out_.push_back(var_name); - } - break; - case PARTIAL_TRAIN: - if (fetches_.find(var_name) != fetches_.end() || - post_op_inputs_.find(var_name) != post_op_inputs_.end() || - persistables_.find(var_name) != persistables_.end()) { - var_out_.push_back(var_name); - } - break; - case FULL_TRAIN: - if (fetches_.find(var_name) != fetches_.end() || - persistables_.find(var_name) != persistables_.end()) { - var_out_.push_back(var_name); - } - break; - default: - var_out_.push_back(var_name); - } - } - } - } -} - -void NgraphEngine::BuildNgFunction() { - BuildNgNodes(); - ngraph_function_ = nullptr; - ngraph::NodeVector func_outputs; - ngraph::ParameterVector func_inputs; - - for (auto& vo : var_out_) { - func_outputs.push_back(var_node_map_->at(vo)); - } - - for (auto& vi : var_in_) { - std::shared_ptr prm = - std::dynamic_pointer_cast( - var_in_node_map_->at(vi)); - func_inputs.push_back(prm); - } - - ngraph_function_ = - std::make_shared(func_outputs, func_inputs); -} - -std::shared_ptr NgraphEngine::GetCacheKey() { - auto cache_key = std::make_shared(""); - *cache_key += std::to_string(fused_ops_.size()); - for (auto& op : fused_ops_) { - *cache_key += op->Type(); - } - for (auto& var_name : var_in_) { - auto shape = var_node_map_->at(var_name)->get_shape(); - *cache_key += var_name; - *cache_key += var_type_map_.at(var_name).c_type_string(); - for (size_t i = 0; i < shape.size(); ++i) { - *cache_key += std::to_string(shape.at(i)); - } - } - - for (auto& var_name : var_out_) { - auto* var = scope_.FindVar(var_name); - if (var && var->IsType()) { - auto* tensor_pd = GetLoDTensorOrSelectedRowsValueFromVar(*var); - auto& ddim = tensor_pd->dims(); - for (int i = 0; i < ddim.size(); ++i) { - *cache_key += std::to_string(ddim[i]); - } - } - } - return cache_key; -} - -void NgraphEngine::GetNgFunction() { - bool cache_on = true; - if (cache_on) { - std::string cache_key_val = *GetCacheKey(); - if (func_cache_.find(cache_key_val) != func_cache_.end()) { - ngraph_function_ = func_cache_.at(cache_key_val); - } else { - BuildNgFunction(); - func_cache_[cache_key_val] = ngraph_function_; - } - } else { - BuildNgFunction(); - } -} - -void NgraphEngine::Run(const Scope& scope, const platform::Place& place) const { - std::vector> t_in; - std::vector> t_out; - - for (size_t i = 0; i < var_in_.size(); ++i) { - auto vi = var_in_.at(i); - auto sp = var_node_map_->at(vi)->get_shape(); - std::shared_ptr ti; - auto* var = scope.FindVar(vi); - if (var && var->IsType()) { - auto* tensor_pd = GetLoDTensorOrSelectedRowsValueFromVar(*var); - PADDLE_ENFORCE(sp == Ddim2Shape(tensor_pd->dims()), - "Ensure ngraph tensor layout align with paddle tensor"); - if (tensor_pd->type() == proto::VarType::FP32) { - const float* arr = tensor_pd->data(); - ti = backend_->create_tensor(ngraph::element::f32, sp, - const_cast(arr)); - } else if (tensor_pd->type() == proto::VarType::INT32) { - const int* arr = tensor_pd->data(); - ti = backend_->create_tensor(ngraph::element::i32, sp, - const_cast(arr)); - } else if (tensor_pd->type() == proto::VarType::INT64) { - const int64_t* arr = tensor_pd->data(); - ti = backend_->create_tensor(ngraph::element::i64, sp, - const_cast(arr)); - } else if (tensor_pd->type() == proto::VarType::FP64) { - const double* arr = tensor_pd->data(); - ti = backend_->create_tensor(ngraph::element::f64, sp, - const_cast(arr)); - } else if (tensor_pd->type() == proto::VarType::BOOL) { - const bool* arr = tensor_pd->data(); - ti = backend_->create_tensor(ngraph::element::boolean, sp, - const_cast(arr)); - } else { - PADDLE_THROW("Data type not handling for var %s", vi); - } - } else { - PADDLE_THROW("Cannot find var or tensor with var name %s", vi); - } - bool is_test = (ng_op_state_ == PARTIAL_TEST || ng_op_state_ == FULL_TEST) - ? true - : false; - bool is_persistable = - (persistables_.find(vi) != persistables_.end()) ? true : false; - if (is_test && is_persistable) { - ti->set_stale(false); - } - t_in.push_back(ti); - } - - for (size_t i = 0; i < var_out_.size(); ++i) { - auto var_name = var_out_[i]; - auto* var = scope.FindVar(var_name); - std::shared_ptr to; - if (var && var->IsType()) { - auto* tensor_pd = GetMutableLoDTensorOrSelectedRowsValueFromVar(var); - auto dd = tensor_pd->dims(); - ngraph::Shape sp = Ddim2Shape(dd); - auto ng_type = var_type_map_.at(var_name); - if (ng_type == ngraph::element::f32) { - auto pd_arr = tensor_pd->mutable_data(place); - to = backend_->create_tensor(ngraph::element::f32, sp, pd_arr); - } else if (ng_type == ngraph::element::i64) { - auto pd_arr = tensor_pd->mutable_data(place); - to = backend_->create_tensor(ngraph::element::i64, sp, pd_arr); - } else if (ng_type == ngraph::element::f64) { - auto pd_arr = tensor_pd->mutable_data(place); - to = backend_->create_tensor(ngraph::element::f64, sp, pd_arr); - } else if (ng_type == ngraph::element::boolean) { - auto pd_arr = tensor_pd->mutable_data(place); - to = backend_->create_tensor(ngraph::element::boolean, sp, pd_arr); - } else { - PADDLE_THROW("Data type not handled in for var %s", var_name); - } - t_out.push_back(to); - } else { - PADDLE_THROW("Cannot find var or tensor with var name %s", var_name); - } - } - - backend_->call(backend_->compile(ngraph_function_), t_out, t_in); -} // NgraphEngine::RunImpl -} // namespace framework -} // namespace paddle diff --git a/paddle/fluid/framework/op_info.h b/paddle/fluid/framework/op_info.h index 19e5c2c73eac74dee030a4f7820531800f737e4e..4b55bd0703eee399cd841f90ea0b18d8fbdc67e8 100644 --- a/paddle/fluid/framework/op_info.h +++ b/paddle/fluid/framework/op_info.h @@ -38,6 +38,7 @@ struct OpInfo { OpAttrChecker* checker_{nullptr}; InferVarTypeFN infer_var_type_; InferShapeFN infer_shape_; + InferInplaceOpFN infer_inplace_; bool HasOpProtoAndChecker() const { return proto_ != nullptr && checker_ != nullptr; diff --git a/paddle/fluid/framework/op_proto_maker.h b/paddle/fluid/framework/op_proto_maker.h index 0a0f8f4655bc34cdb25205ff6eaec9f96c801ebd..5f3ce60e1d94ea4078cf0b709df362bad317f621 100644 --- a/paddle/fluid/framework/op_proto_maker.h +++ b/paddle/fluid/framework/op_proto_maker.h @@ -27,7 +27,7 @@ enum class OpRole { kForward = 0x0000, kBackward = 0x0001, kOptimize = 0x0002, - // RPC role is for send/recv releated op + // RPC role is for send/recv related op kRPC = 0x0004, // Dist role is for split_byref/split_selected_rows/concat // used for distributed training. diff --git a/paddle/fluid/framework/operator.cc b/paddle/fluid/framework/operator.cc index ab3cf308fc04e227d5402712f6bab226fea04711..9a0348871b050278da2ad07ac6992188a702da42 100644 --- a/paddle/fluid/framework/operator.cc +++ b/paddle/fluid/framework/operator.cc @@ -177,9 +177,7 @@ void OperatorBase::Run(const Scope& scope, const platform::Place& place) { // in concurrency scenerio. Here use an `if` to fix this issue. // Please not remove the `if`, ask @Superjomn if there are any concern. if (platform::IsProfileEnabled()) { - platform::DeviceContextPool& pool = - platform::DeviceContextPool::Instance(); - platform::RecordEvent record_event(Type(), pool.Get(place)); + platform::RecordEvent record_event(Type()); RunImpl(scope, place); } else { RunImpl(scope, place); @@ -188,14 +186,14 @@ void OperatorBase::Run(const Scope& scope, const platform::Place& place) { VLOG(3) << place << " " << DebugStringEx(&scope); } catch (platform::EnforceNotMet exception) { if (Attrs().count("sub_block") != 0) { - throw exception; + throw; } auto& callstack = Attr>( OpProtoAndCheckerMaker::OpCreationCallstackAttrName()); if (callstack.empty()) { - throw exception; + throw; } std::ostringstream sout; sout << "Invoke operator " << Type() << " error.\n"; @@ -206,7 +204,7 @@ void OperatorBase::Run(const Scope& scope, const platform::Place& place) { sout << "C++ Callstacks: \n"; sout << exception.err_str_; exception.err_str_ = sout.str(); - throw exception; + throw; } catch (...) { std::rethrow_exception(std::current_exception()); } @@ -555,18 +553,17 @@ Tensor* ExecutionContext::LegacyOutput(const std::string& name) const { template <> std::vector ExecutionContext::MultiOutput( const std::string& name) const { - auto names = op().Outputs(name); + auto it = ctx_.outputs.find(name); + if (it == ctx_.outputs.end()) { + return {}; + } + const std::vector& vars = it->second; std::vector res; - res.reserve(names.size()); - std::transform(names.begin(), names.end(), std::back_inserter(res), - [&](const std::string& sub_name) -> Tensor* { - auto var = scope_.FindVar(sub_name); - if (var == nullptr) return nullptr; - PADDLE_ENFORCE( - var->IsType(), - "%s should be LoDTensor, but the received type is %s", - sub_name, ToTypeName(var->Type())); - return var->GetMutable(); + res.reserve(vars.size()); + std::transform(vars.begin(), vars.end(), std::back_inserter(res), + [&](Variable* var) -> Tensor* { + return var == nullptr ? nullptr + : var->GetMutable(); }); return res; } @@ -590,7 +587,7 @@ class RuntimeInferShapeContext : public InferShapeContext { public: RuntimeInferShapeContext(const OperatorBase& op, const Scope& scope, const RuntimeContext& ctx) - : op_(op), scope_(scope), ctx_(ctx) {} + : op_(op), ctx_(ctx) {} bool HasInput(const std::string& name) const override { // has only one input @@ -882,7 +879,6 @@ class RuntimeInferShapeContext : public InferShapeContext { } const OperatorBase& op_; - const Scope& scope_; const RuntimeContext& ctx_; }; @@ -991,11 +987,14 @@ void OperatorWithKernel::TransferInplaceVarsBack( const Scope& transfer_scope) const { for (auto& var_name : inplace_vars) { VLOG(3) << "share inplace var " + var_name + " back to it's original scope"; + auto* origin_var = scope.FindVar(var_name); + PADDLE_ENFORCE_NOT_NULL(origin_var, "The var[%s] should not be nullptr.", + var_name); auto* original_tensor = - GetMutableLoDTensorOrSelectedRowsValueFromVar(scope.FindVar(var_name)); + GetMutableLoDTensorOrSelectedRowsValueFromVar(origin_var); auto* var = transfer_scope.FindVar(var_name); - PADDLE_ENFORCE(var != nullptr, "The var[%s] should not be nullptr", - var_name); + PADDLE_ENFORCE_NOT_NULL(var, "The var[%s] should not be nullptr.", + var_name); auto* transformed_tensor = GetLoDTensorOrSelectedRowsValueFromVar(*var); original_tensor->ShareDataWith(*transformed_tensor); } diff --git a/paddle/fluid/framework/operator.h b/paddle/fluid/framework/operator.h index 40d935a5ff98a28b376a59b733e2929e4a128cb9..e33214b44bb5d8ea5eb32d442d597a369c198bdd 100644 --- a/paddle/fluid/framework/operator.h +++ b/paddle/fluid/framework/operator.h @@ -222,12 +222,7 @@ class ExecutionContext { if (it == ctx_.inputs.end()) { return {}; } - std::vector res; - res.reserve(it->second.size()); - std::transform(it->second.begin(), it->second.end(), - std::back_inserter(res), - [this](Variable* var) { return var; }); - return res; + return {it->second.begin(), it->second.end()}; } std::vector MultiOutputVar(const std::string& name) const { diff --git a/paddle/fluid/framework/parallel_executor.cc b/paddle/fluid/framework/parallel_executor.cc index f61c9e3a91146704faa6c5b1058137bef67d2a3e..56da5660095affa0ba49d8bc533d1da01ffd18be 100644 --- a/paddle/fluid/framework/parallel_executor.cc +++ b/paddle/fluid/framework/parallel_executor.cc @@ -21,6 +21,7 @@ limitations under the License. */ #include "paddle/fluid/framework/ir/graph.h" +#include "paddle/fluid/framework/details/all_reduce_deps_pass.h" #include "paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h" #include "paddle/fluid/framework/details/multi_devices_helper.h" #include "paddle/fluid/framework/details/parallel_ssa_graph_executor.h" @@ -171,14 +172,6 @@ std::unique_ptr ParallelExecutorPrivate::PrepareGCAndRefCnts( eager_deletion_pass->SetNotOwned(details::kAllPlaces, &places_); graph = eager_deletion_pass->Apply(std::move(graph)); VLOG(10) << "EagerDeletionPass Applied"; - - if (build_strategy_.memory_early_delete_) { - auto early_delete_pass = - ir::PassRegistry::Instance().Get("memory_early_delete_pass"); - early_delete_pass->SetNotOwned(details::kGarbageCollector, &gcs_); - graph = early_delete_pass->Apply(std::move(graph)); - } - VLOG(10) << "MemoryEarlyDeletePass Applied."; } return graph; @@ -201,7 +194,6 @@ ParallelExecutor::ParallelExecutor( member_->use_all_reduce_ = build_strategy.reduce_ == BuildStrategy::ReduceStrategy::kAllReduce; member_->nranks_ = build_strategy.num_trainers_ * places.size(); - if (!member_->use_all_reduce_) { PADDLE_ENFORCE(places.size() > 1, "If you set build_strategy.reduce with 'Reduce'," @@ -229,9 +221,10 @@ ParallelExecutor::ParallelExecutor( // choice the execution strategy. build_strategy.enable_parallel_graph_ = EnableParallelGraphExecution(main_program, exec_strategy, build_strategy); - - VLOG(1) << "Enable ParallelGraph Execution: " - << build_strategy.enable_parallel_graph_; + if (build_strategy.enable_parallel_graph_) + VLOG(0) << "The Executor would execute the graph by ParallelGraph " + "Execution which can get better performance," + << "you can force it off by env FLAGS_enable_parallel_graph=0"; if (member_->use_cuda_) { // Bcast Parameters to all GPUs @@ -265,58 +258,44 @@ ParallelExecutor::ParallelExecutor( // Step 2. Convert main_program to SSA form and dependency graph. Also, insert // ncclOp - std::vector> graphs; + std::unique_ptr graph; #if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) - if (build_strategy.enable_parallel_graph_) { - for (size_t i = 0; i < member_->places_.size(); ++i) { - std::unique_ptr graph = build_strategy.Apply( - main_program, {member_->places_[i]}, loss_var_name, - {member_->local_scopes_[i]}, member_->nranks_, member_->use_cuda_, - member_->nccl_ctxs_.get()); - graphs.push_back(std::move(graph)); - } - } else { - std::unique_ptr graph = build_strategy.Apply( - main_program, member_->places_, loss_var_name, member_->local_scopes_, - member_->nranks_, member_->use_cuda_, member_->nccl_ctxs_.get()); - graphs.push_back(std::move(graph)); - } + graph = build_strategy.Apply(main_program, member_->places_, loss_var_name, + member_->local_scopes_, member_->nranks_, + member_->use_cuda_, member_->nccl_ctxs_.get()); #else - std::unique_ptr graph = build_strategy.Apply( - main_program, member_->places_, loss_var_name, member_->local_scopes_, - member_->nranks_, member_->use_cuda_); - graphs.push_back(std::move(graph)); + graph = build_strategy.Apply(main_program, member_->places_, loss_var_name, + member_->local_scopes_, member_->nranks_, + member_->use_cuda_); #endif auto max_memory_size = GetEagerDeletionThreshold(); + VLOG(10) << "Eager Deletion Threshold " + << static_cast(max_memory_size) / (1 << 30); if (max_memory_size >= 0) { - for (size_t i = 0; i < graphs.size(); ++i) { - graphs[i] = member_->PrepareGCAndRefCnts( - std::move(graphs[i]), static_cast(max_memory_size)); - } + graph = member_->PrepareGCAndRefCnts(std::move(graph), + static_cast(max_memory_size)); } // Step 3. Create vars in each scope. Passes may also create new vars. // skip control vars and empty vars std::vector var_infos; - for (auto &graph : graphs) { - for (auto &node : graph->Nodes()) { - if (node->IsVar() && !node->IsCtrlVar() && node->Var()) { - var_infos.emplace_back(); - var_infos.back().name_ = node->Var()->Name(); - var_infos.back().type_ = node->Var()->GetType(); - var_infos.back().persistable_ = node->Var()->Persistable(); - } + for (auto &node : graph->Nodes()) { + if (node->IsVar() && !node->IsCtrlVar() && node->Var()) { + var_infos.emplace_back(); + var_infos.back().name_ = node->Var()->Name(); + var_infos.back().type_ = node->Var()->GetType(); + var_infos.back().persistable_ = node->Var()->Persistable(); } } // If the loss_var_name is given, the number of graph should be only one. if (loss_var_name.size()) { - size_t graph_num = ir::GraphNum(*graphs[0]); + size_t graph_num = ir::GraphNum(*graph); if (graph_num > 1) { LOG(WARNING) << "The number of graph should be only one, " "but the current graph has " - << ir::GraphNum(*graphs[0]) + << ir::GraphNum(*graph) << " sub_graphs. If you want to see the nodes of the " "sub_graphs, you should use 'FLAGS_print_sub_graph_dir' " "to specify the output dir. NOTES: if you not do training, " @@ -325,18 +304,25 @@ ParallelExecutor::ParallelExecutor( } if (build_strategy.enable_parallel_graph_) { +#ifdef PADDLE_WITH_CUDA + // TODO(Yancey1989): Remove passing in the main_program when + // allreduce_seq_pass doesn't need it as the attr. member_->executor_.reset(new details::ParallelSSAGraphExecutor( - exec_strategy, member_->local_scopes_, member_->places_, - std::move(graphs))); + exec_strategy, member_->local_scopes_, member_->places_, main_program, + std::move(graph))); +#else + PADDLE_THROW( + "Paddle should be compiled with CUDA for ParallelGraph Execution."); +#endif } else { if (exec_strategy.type_ == ExecutionStrategy::kDefault) { member_->executor_.reset(new details::ThreadedSSAGraphExecutor( exec_strategy, member_->local_scopes_, member_->places_, - std::move(graphs[0]))); + std::move(graph))); } else { member_->executor_.reset(new details::FastThreadedSSAGraphExecutor( exec_strategy, member_->local_scopes_, member_->places_, - std::move(graphs[0]))); + std::move(graph))); } } @@ -488,11 +474,10 @@ bool ParallelExecutor::EnableParallelGraphExecution( } if (!member_->use_all_reduce_ || !member_->use_cuda_) - enable_parallel_graph = false; - if (build_strategy.enable_sequential_execution_ || - exec_strategy.type_ == ExecutionStrategy::ExecutorType::kExperimental) - enable_parallel_graph = false; + if (build_strategy.enable_sequential_execution_ || + exec_strategy.type_ == ExecutionStrategy::ExecutorType::kExperimental) + enable_parallel_graph = false; return enable_parallel_graph; } @@ -506,6 +491,5 @@ ParallelExecutor::~ParallelExecutor() { } // namespace framework } // namespace paddle -USE_PASS(memory_early_delete_pass); USE_PASS(reference_count_pass); USE_PASS(eager_deletion_pass); diff --git a/paddle/fluid/framework/scope.cc b/paddle/fluid/framework/scope.cc index 953618560913229cd1e47659ad61e621efc10ed1..87f0f307d30bc90a43a698c3766b16c975f0635e 100644 --- a/paddle/fluid/framework/scope.cc +++ b/paddle/fluid/framework/scope.cc @@ -22,11 +22,7 @@ limitations under the License. */ #include "paddle/fluid/framework/threadpool.h" #include "paddle/fluid/string/printf.h" -DEFINE_bool(benchmark, false, - "Doing memory benchmark. It will make deleting scope synchronized, " - "and add some memory usage logs." - "Default cuda is asynchronous device, set to True will" - "force op run in synchronous mode."); +DECLARE_bool(benchmark); DEFINE_bool( eager_delete_scope, true, diff --git a/paddle/fluid/framework/type_defs.h b/paddle/fluid/framework/type_defs.h index 938e2024c3359c2acd65a1aa4af875a8350e4c58..d02c699b979d7693bd83fd43fc73f7e0aeddb0cc 100644 --- a/paddle/fluid/framework/type_defs.h +++ b/paddle/fluid/framework/type_defs.h @@ -57,5 +57,8 @@ using InferVarTypeFN = using InferShapeFN = std::function; +using InplacePair = std::unordered_map; +using InferInplaceOpFN = std::function; + } // namespace framework } // namespace paddle diff --git a/paddle/fluid/imperative/CMakeLists.txt b/paddle/fluid/imperative/CMakeLists.txt index 5db422119966948f75970874e13d416ea699158a..ec8dedd605235a2d197e6a313bd589d5b9520cdf 100644 --- a/paddle/fluid/imperative/CMakeLists.txt +++ b/paddle/fluid/imperative/CMakeLists.txt @@ -1,5 +1,5 @@ if(WITH_PYTHON) -cc_library(layer SRCS layer.cc DEPS proto_desc operator device_context blas) -cc_library(tracer SRCS tracer.cc DEPS proto_desc device_context) +cc_library(layer SRCS layer.cc DEPS proto_desc operator device_context blas pybind) +cc_library(tracer SRCS tracer.cc DEPS proto_desc device_context pybind) cc_library(engine SRCS engine.cc) endif() diff --git a/paddle/fluid/imperative/layer.cc b/paddle/fluid/imperative/layer.cc index 8029129b9a6a9fcbc0ff10daa1f25b210259e9d8..8f20f0c06e043ddc629e47c6e49280c5467b0e20 100644 --- a/paddle/fluid/imperative/layer.cc +++ b/paddle/fluid/imperative/layer.cc @@ -156,6 +156,8 @@ class Autograd { for (auto it : candidate->pre_ops_) { for (OpBase* pre_op : it.second) { if (!pre_op) continue; + VLOG(5) << "op dep " << candidate->op_desc_->Type() << " <---- " + << it.first << " <---- " << pre_op->op_desc_->Type(); if (visited.find(pre_op) == visited.end()) { visited.insert(pre_op); queue.push_back(pre_op); @@ -204,59 +206,68 @@ framework::LoDTensor& VarBase::GradValue() { } std::map> OpBase::ApplyGrad() { - if (!grad_op_desc_ && backward_id_ <= 0) { - LOG(WARNING) << "op with no grad: " << op_desc_->Type(); + if (grad_op_descs_.empty() && backward_id_ <= 0) { + VLOG(3) << "op with no grad: " << op_desc_->Type(); return {}; } - std::map> grad_outputs; + std::vector grad_outputs; if (backward_id_ > 0) { VLOG(3) << "py_layer_grad"; - grad_outputs[framework::GradVarName(PyLayer::kFwdOut)] = PyLayer::ApplyGrad( - backward_id_, - grad_input_vars_[framework::GradVarName(PyLayer::kFwdInp)]); + grad_outputs.resize(1); + grad_outputs[0][framework::GradVarName(PyLayer::kFwdOut)] = + PyLayer::ApplyGrad( + backward_id_, + grad_input_vars_[0][framework::GradVarName(PyLayer::kFwdInp)]); } else { - VLOG(3) << "op grad " << grad_op_desc_->Type(); - for (auto it : grad_output_vars_) { - auto& outputs = grad_outputs[it.first]; - for (size_t i = 0; i < it.second.size(); ++i) { - // Allocate a new variable - Variable* tmp_var = new framework::Variable(); - tmp_var->GetMutable(); - outputs.push_back(tmp_var); + grad_outputs.resize(grad_op_descs_.size()); + for (size_t k = 0; k < grad_op_descs_.size(); ++k) { + framework::OpDesc* grad_op_desc = grad_op_descs_[k]; + VLOG(3) << "op grad " << grad_op_desc->Type(); + for (auto it : grad_output_vars_[k]) { + auto& outputs = grad_outputs[k][it.first]; + for (size_t i = 0; i < it.second.size(); ++i) { + // Allocate a new variable + Variable* tmp_var = new framework::Variable(); + tmp_var->GetMutable(); + outputs.push_back(tmp_var); + } } - } - framework::RuntimeContext ctx(grad_input_vars_, grad_outputs); + framework::RuntimeContext ctx(grad_input_vars_[k], grad_outputs[k]); - // No need to do compile time infer shape here. - // grad_op_desc_->InferShape(*block_); - grad_op_desc_->InferVarType(block_); + // No need to do compile time infer shape here. + // grad_op_desc_->InferShape(*block_); + grad_op_desc->InferVarType(block_); - std::unique_ptr opbase = - framework::OpRegistry::CreateOp(*grad_op_desc_); - framework::OperatorWithKernel* op_kernel = - dynamic_cast(opbase.get()); - PADDLE_ENFORCE_NOT_NULL(op_kernel, "only support op with kernel"); + std::unique_ptr opbase = + framework::OpRegistry::CreateOp(*grad_op_desc); + framework::OperatorWithKernel* op_kernel = + dynamic_cast(opbase.get()); + PADDLE_ENFORCE_NOT_NULL(op_kernel, "only support op with kernel"); - framework::Scope scope; - PreparedOp p = PreparedOp::Prepare(ctx, *op_kernel, place_); - p.op.RuntimeInferShape(scope, place_, ctx); - p.func(framework::ExecutionContext(p.op, scope, *p.dev_ctx, p.ctx)); + framework::Scope scope; + PreparedOp p = PreparedOp::Prepare(ctx, *op_kernel, place_); + p.op.RuntimeInferShape(scope, place_, ctx); + p.func(framework::ExecutionContext(p.op, scope, *p.dev_ctx, p.ctx)); + } } - for (auto it : grad_output_vars_) { - auto& outputs = grad_outputs[it.first]; - auto& origin_outputs = it.second; - PADDLE_ENFORCE_EQ(outputs.size(), origin_outputs.size()); - - for (size_t i = 0; i < outputs.size(); ++i) { - framework::Variable* grad = outputs[i]; - framework::Variable* orig_grad = origin_outputs[i]; - AddTo(grad, orig_grad, place_); - delete grad; + for (size_t k = 0; k < grad_output_vars_.size(); ++k) { + for (auto it : grad_output_vars_[k]) { + auto& outputs = grad_outputs[k][it.first]; + auto& origin_outputs = it.second; + PADDLE_ENFORCE_EQ(outputs.size(), origin_outputs.size()); + + for (size_t i = 0; i < outputs.size(); ++i) { + framework::Variable* grad = outputs[i]; + framework::Variable* orig_grad = origin_outputs[i]; + AddTo(grad, orig_grad, place_); + delete grad; + } } } + return input_vars_; } diff --git a/paddle/fluid/imperative/layer.h b/paddle/fluid/imperative/layer.h index 633924aa417b8bd64bf4921054f82fdb7f7868fe..78205486c5534ac0c61cc6d545bdafa4dfc95695 100644 --- a/paddle/fluid/imperative/layer.h +++ b/paddle/fluid/imperative/layer.h @@ -28,6 +28,7 @@ #include "paddle/fluid/framework/var_desc.h" #include "paddle/fluid/platform/enforce.h" #include "paddle/fluid/platform/device_context.h" +#include "paddle/fluid/operators/math/math_function.h" #include "paddle/fluid/imperative/type_defs.h" @@ -140,16 +141,24 @@ class VarBase { void RunBackward(); void TrackPreOp(OpBase* pre_op, const std::string& pre_op_out_name, - int pre_op_out_idx, bool stop_gradient) { + int pre_op_out_idx, bool pre_op_stop_gradient) { pre_op_ = pre_op; pre_op_out_name_ = pre_op_out_name; pre_op_out_idx_ = pre_op_out_idx; - stop_gradient_ = stop_gradient; + if (pre_op_stop_gradient) { + stop_gradient_ = pre_op_stop_gradient; + } } void ClearGradient() { - delete grads_; - grads_ = new VarBase(true); + VLOG(1) << "clear gradient of " << var_desc_->Name(); + if (grads_ && grads_->var_ && grads_->var_->IsInitialized()) { + auto grads_t = grads_->var_->GetMutable(); + operators::math::set_constant( + *(platform::DeviceContextPool::Instance().Get( + grads_->var_->Get().place())), + grads_t, 0.0); + } } framework::LoDTensor& GradValue(); @@ -184,12 +193,13 @@ class OpBase { OpBase() : op_desc_(nullptr), forward_id_(-1), - grad_op_desc_(nullptr), backward_id_(-1), place_(platform::CPUPlace()) {} virtual ~OpBase() { - if (grad_op_desc_) delete grad_op_desc_; + for (framework::OpDesc* desc : grad_op_descs_) { + delete desc; + } } std::map> ApplyGrad(); @@ -198,9 +208,11 @@ class OpBase { // For pure python PyLayer, use `forward_id_`, otherwise, use op_desc_. framework::OpDesc* op_desc_; int forward_id_; - // When has backward, one of `grad_op_desc_` or `backward_id_` is set, + + // When has backward, one of `grad_op_descs_` or `backward_id_` is set, // not both. - framework::OpDesc* grad_op_desc_; + // Note: each fwd op corresponds to a vector of bwd ops. + std::vector grad_op_descs_; int backward_id_; platform::Place place_; @@ -210,8 +222,11 @@ class OpBase { OpBasePtrMap pre_ops_; std::map> pre_ops_out_idx_; - framework::VariableValueMap grad_input_vars_; - framework::VariableValueMap grad_output_vars_; + // Inputs to a vector of bwd ops. + std::vector grad_input_vars_; + // Outputs to a vector of bwd ops. + std::vector grad_output_vars_; + framework::BlockDesc* block_; }; diff --git a/paddle/fluid/imperative/tracer.cc b/paddle/fluid/imperative/tracer.cc index 5b87839f457c24d5d6687a27faac6c0f52f5f90b..bc39d11ba00a6a7c386162a1f9201c6f992c8692 100644 --- a/paddle/fluid/imperative/tracer.cc +++ b/paddle/fluid/imperative/tracer.cc @@ -24,15 +24,17 @@ namespace imperative { void CreateGradOp(const framework::OpDesc& op_desc, const std::unordered_set& no_grad_set, const std::vector& grad_sub_block, - framework::OpDesc** grad_op_desc, + std::vector* grad_op_descs, std::unordered_map* grad_to_var) { - std::vector> grad_op_descs = + PADDLE_ENFORCE(grad_op_descs->empty()); + std::vector> descs = framework::OpInfoMap::Instance() .Get(op_desc.Type()) .GradOpMaker()(op_desc, no_grad_set, grad_to_var, grad_sub_block); - PADDLE_ENFORCE(grad_op_descs.size() == 1, "Only support 1 grad op now."); - // TODO(panyx0718): Leak? - *grad_op_desc = grad_op_descs[0].release(); + + for (auto& desc : descs) { + grad_op_descs->emplace_back(desc.release()); + } } void InitVar(framework::Variable* var, framework::Variable* grad_var, @@ -83,11 +85,12 @@ void Tracer::Trace(OpBase* op, const VarBasePtrMap& inputs, op->input_vars_ = inputs; for (auto it : op->input_vars_) { auto& invars = invars_map[it.first]; + invars.reserve(it.second.size()); for (VarBase* inp : it.second) { PADDLE_ENFORCE_NOT_NULL(inp->var_, "op %s input %s nullptr", op->op_desc_->Type(), inp->var_desc_->Name()); - invars.push_back(inp->var_); + invars.emplace_back(inp->var_); vars[inp->var_desc_->Name()] = inp; if (inp->PreOp()) { op->pre_ops_[it.first].push_back(inp->PreOp()); @@ -104,9 +107,10 @@ void Tracer::Trace(OpBase* op, const VarBasePtrMap& inputs, for (auto it : op->output_vars_) { auto& outvars = outvars_map[it.first]; const std::vector& outputs = it.second; + outvars.reserve(outputs.size()); for (size_t i = 0; i < outputs.size(); ++i) { VarBase* out = outputs[i]; - outvars.push_back(out->var_); + outvars.emplace_back(out->var_); vars[out->var_desc_->Name()] = out; framework::VarDesc* var_desc = block->FindVar(out->var_desc_->Name()); @@ -138,49 +142,52 @@ void Tracer::Trace(OpBase* op, const VarBasePtrMap& inputs, prepared_op.op, scope, *prepared_op.dev_ctx, prepared_op.ctx)); if (!stop_gradient) { - framework::OpDesc* grad_op_desc; - // TODO(panyx): Is this leaked? std::unique_ptr> grad_to_var( new std::unordered_map()); - CreateGradOp(*op_desc, {}, {block}, &grad_op_desc, grad_to_var.get()); - op->grad_op_desc_ = grad_op_desc; - - for (auto it : grad_op_desc->Inputs()) { - auto& grad_in_vars = op->grad_input_vars_[it.first]; - for (const std::string& grad_invar : it.second) { - block->FindRecursiveOrCreateVar(grad_invar); - auto var_it = grad_to_var->find(grad_invar); - if (var_it == grad_to_var->end()) { - auto fwd_var_it = vars.find(grad_invar); - PADDLE_ENFORCE(fwd_var_it != vars.end()); - // Forward inputs or outputs. - grad_in_vars.push_back(fwd_var_it->second->var_); - } else { + CreateGradOp(*op_desc, {}, {block}, &op->grad_op_descs_, grad_to_var.get()); + + op->grad_input_vars_.resize(op->grad_op_descs_.size()); + op->grad_output_vars_.resize(op->grad_op_descs_.size()); + for (size_t i = 0; i < op->grad_op_descs_.size(); ++i) { + framework::OpDesc* grad_op_desc = op->grad_op_descs_[i]; + for (auto it : grad_op_desc->Inputs()) { + auto& grad_in_vars = op->grad_input_vars_[i][it.first]; + for (const std::string& grad_invar : it.second) { + block->FindRecursiveOrCreateVar(grad_invar); + auto var_it = grad_to_var->find(grad_invar); + if (var_it == grad_to_var->end()) { + auto fwd_var_it = vars.find(grad_invar); + PADDLE_ENFORCE(fwd_var_it != vars.end()); + // Forward inputs or outputs. + grad_in_vars.push_back(fwd_var_it->second->var_); + } else { + VarBase* var = vars[var_it->second]; + if (!var->grads_->var_->IsInitialized()) { + InitVar(var->var_, var->grads_->var_, + prepared_op.GetDeviceContext()); + } + // Douts. + grad_in_vars.push_back(var->grads_->var_); + } + } + } + + for (auto it : grad_op_desc->Outputs()) { + auto& grad_out_vars = op->grad_output_vars_[i][it.first]; + for (const std::string& grad_outvar : it.second) { + block->FindRecursiveOrCreateVar(grad_outvar); + auto var_it = grad_to_var->find(grad_outvar); + PADDLE_ENFORCE(var_it != grad_to_var->end(), + "Could not found the grad op output var, should this " + "operator %s's stop gradient be True", + op_desc->Type()); VarBase* var = vars[var_it->second]; if (!var->grads_->var_->IsInitialized()) { InitVar(var->var_, var->grads_->var_, prepared_op.GetDeviceContext()); } - // Douts. - grad_in_vars.push_back(var->grads_->var_); - } - } - } - - for (auto it : grad_op_desc->Outputs()) { - auto& grad_out_vars = op->grad_output_vars_[it.first]; - for (const std::string& grad_outvar : it.second) { - block->FindRecursiveOrCreateVar(grad_outvar); - auto var_it = grad_to_var->find(grad_outvar); - PADDLE_ENFORCE(var_it != grad_to_var->end(), - "Could not found the grad op output var, should this " - "operator %s's stop gradient be True", - op_desc->Type()); - VarBase* var = vars[var_it->second]; - if (!var->grads_->var_->IsInitialized()) { - InitVar(var->var_, var->grads_->var_, prepared_op.GetDeviceContext()); + grad_out_vars.push_back(var->grads_->var_); } - grad_out_vars.push_back(var->grads_->var_); } } } @@ -209,10 +216,12 @@ std::vector Tracer::PyTrace(OpBase* op, out->TrackPreOp(op, PyLayer::kFwdOut, i, stop_gradient); } if (!stop_gradient) { + op->grad_input_vars_.resize(1); + op->grad_output_vars_.resize(1); auto& grad_input_vars = - op->grad_input_vars_[framework::GradVarName(PyLayer::kFwdInp)]; + op->grad_input_vars_[0][framework::GradVarName(PyLayer::kFwdInp)]; auto& grad_output_vars = - op->grad_output_vars_[framework::GradVarName(PyLayer::kFwdOut)]; + op->grad_output_vars_[0][framework::GradVarName(PyLayer::kFwdOut)]; for (const VarBase* inp : inputs) { grad_input_vars.push_back(inp->var_); diff --git a/paddle/fluid/inference/CMakeLists.txt b/paddle/fluid/inference/CMakeLists.txt index 11484a647303b32a6006bef3cfe4be6b3f0d533d..157862016e3556902f6507e02417624363ed1029 100644 --- a/paddle/fluid/inference/CMakeLists.txt +++ b/paddle/fluid/inference/CMakeLists.txt @@ -58,12 +58,13 @@ if(WIN32) sep_library(paddle_fluid_shared SHARED SRCS ${SHARED_INFERENCE_SRCS} DEPS ${fluid_modules} paddle_fluid_api reset_tensor_array analysis_config paddle_pass_builder) - target_link_libraries(paddle_fluid_shared shlwapi) else(WIN32) cc_library(paddle_fluid_shared SHARED SRCS ${SHARED_INFERENCE_SRCS} DEPS ${fluid_modules} paddle_fluid_api reset_tensor_array analysis_config paddle_pass_builder) endif() +get_property(os_dependency_modules GLOBAL PROPERTY OS_DEPENDENCY_MODULES) +target_link_libraries(paddle_fluid_shared ${os_dependency_modules}) set_target_properties(paddle_fluid_shared PROPERTIES OUTPUT_NAME paddle_fluid) if(NOT APPLE AND NOT WIN32) diff --git a/paddle/fluid/inference/analysis/argument.h b/paddle/fluid/inference/analysis/argument.h index a2546ead93c3baeb8029f6451d8a60dcc75f8571..2f31b182af7293488719e41a92b2ea78709bda02 100644 --- a/paddle/fluid/inference/analysis/argument.h +++ b/paddle/fluid/inference/analysis/argument.h @@ -132,7 +132,7 @@ struct Argument { DECL_ARGUMENT_FIELD(tensorrt_workspace_size, TensorRtWorkspaceSize, int); DECL_ARGUMENT_FIELD(tensorrt_min_subgraph_size, TensorRtMinSubgraphSize, int); DECL_ARGUMENT_FIELD(tensorrt_precision_mode, TensorRtPrecisionMode, - contrib::AnalysisConfig::Precision); + AnalysisConfig::Precision); // Memory optimized related. DECL_ARGUMENT_FIELD(enable_memory_optim, EnableMemoryOptim, bool); diff --git a/paddle/fluid/inference/analysis/helper.h b/paddle/fluid/inference/analysis/helper.h index 120f6ef27d49ae59ec36304dc3742cd9ca0afa4b..59107f28080dceb0a58e17d42281db5f3773de56 100644 --- a/paddle/fluid/inference/analysis/helper.h +++ b/paddle/fluid/inference/analysis/helper.h @@ -32,7 +32,7 @@ limitations under the License. */ #ifdef _WIN32 #include #include -#define GCC_ATTRIBUTE(attr__) ; +#define GCC_ATTRIBUTE(attr__) #define MKDIR(path) _mkdir(path) #else #include diff --git a/paddle/fluid/inference/analysis/ir_pass_manager.cc b/paddle/fluid/inference/analysis/ir_pass_manager.cc index 99611ce84b23896dd173831a03d77c6e0252d998..8d5ee36ae627deccd7ddbd4bf8c5354a82c5e9db 100644 --- a/paddle/fluid/inference/analysis/ir_pass_manager.cc +++ b/paddle/fluid/inference/analysis/ir_pass_manager.cc @@ -71,7 +71,7 @@ void IRPassManager::CreatePasses(Argument *argument, new framework::ProgramDesc *(&argument->main_program())); bool enable_int8 = argument->tensorrt_precision_mode() == - contrib::AnalysisConfig::Precision::kInt8; + AnalysisConfig::Precision::kInt8; pass->Set("enable_int8", new bool(enable_int8)); std::string model_opt_cache_dir = @@ -83,7 +83,6 @@ void IRPassManager::CreatePasses(Argument *argument, new std::string(GetOrCreateModelOptCacheDir(model_opt_cache_dir))); } - // graph_ = pass->Apply(std::move(graph_)); pre_pass = pass_name; passes_.emplace_back(std::move(pass)); @@ -97,11 +96,12 @@ std::unique_ptr IRPassManager::Apply(std::unique_ptr graph) { PADDLE_ENFORCE(graph.get()); // Apply all the passes for (const auto &pass : passes_) { - if (pass->Type() == "graph_viz_pass") continue; - PrettyLogEndl(Style::H2(), "--- Running IR pass [%s]", pass->Type()); + if (pass->Type() != "graph_viz_pass") { + PrettyLogEndl(Style::H2(), "--- Running IR pass [%s]", pass->Type()); + } graph = pass->Apply(std::move(graph)); } - return std::move(graph); + return graph; } framework::proto::ProgramDesc IRPassManager::AcquireProgram( diff --git a/paddle/fluid/inference/analysis/ir_passes/CMakeLists.txt b/paddle/fluid/inference/analysis/ir_passes/CMakeLists.txt index eb6e1768a2c01f1388962eefe8e70368cae8cf8b..410a90132aa7657a23b858570763547fe53730a0 100644 --- a/paddle/fluid/inference/analysis/ir_passes/CMakeLists.txt +++ b/paddle/fluid/inference/analysis/ir_passes/CMakeLists.txt @@ -1,4 +1,7 @@ cc_library(subgraph_detector SRCS subgraph_detector.cc DEPS proto_desc) +if(WITH_TESTING) + add_dependencies(subgraph_detector gtest) +endif() if (WITH_GPU AND TENSORRT_FOUND) cc_library(tensorrt_subgraph_pass SRCS tensorrt_subgraph_pass.cc DEPS subgraph_detector tensorrt_op_teller) diff --git a/paddle/fluid/inference/analysis/ir_passes/subgraph_detector.cc b/paddle/fluid/inference/analysis/ir_passes/subgraph_detector.cc index a64f85ee9ac1a7bb8f0ed7bb8678166bbbcd5746..96befe7f8a5d16402338ac337daa96d714b4d310 100644 --- a/paddle/fluid/inference/analysis/ir_passes/subgraph_detector.cc +++ b/paddle/fluid/inference/analysis/ir_passes/subgraph_detector.cc @@ -460,77 +460,6 @@ inline bool CheckNodeIndegreeEquals(const Node &node, size_t n) { return node.inputs.size() == n; } -NodesTSIterator::NodesTSIterator(const std::vector &source) { - PADDLE_ENFORCE(!source.empty(), - "Start points of topological sorting should not be empty!"); - // CHECK all the inputs' in-degree is 0 - for (auto *node : source) { - PADDLE_ENFORCE(CheckNodeIndegreeEquals(*node, 0)); - } - - std::unordered_set visited; - std::unordered_set to_visit{source.begin(), source.end()}; - - std::vector inlink_visited; - while (!to_visit.empty()) { - std::vector queue(to_visit.begin(), to_visit.end()); - for (auto *p : queue) { - if (Agent(p).deleted()) { - visited.insert(p); - to_visit.erase(p); - } - - inlink_visited.clear(); - - std::copy_if(p->inputs.begin(), p->inputs.end(), - std::back_inserter(inlink_visited), - [&](Node *x) -> bool { return visited.count(x) != 0; }); - - if (inlink_visited.size() == p->inputs.size()) { - sorted_.push_back(p); - for (auto *_ : p->outputs) { - if (!visited.count(_)) { - to_visit.insert(_); - } - } - - to_visit.erase(p); - visited.insert(p); - } - } - } -} - -NodesTSIterator::NodesTSIterator(const NodesTSIterator &other) - : sorted_(other.sorted_), cursor_(other.cursor_) {} - -Node &NodesTSIterator::operator*() { - PADDLE_ENFORCE_LT(cursor_, sorted_.size()); - return *sorted_[cursor_]; -} - -NodesTSIterator &NodesTSIterator::operator++() { - if (++cursor_ >= sorted_.size()) { - sorted_.clear(); - cursor_ = 0; - } - return *this; -} -NodesTSIterator &NodesTSIterator::operator=(const NodesTSIterator &other) { - cursor_ = other.cursor_; - sorted_ = other.sorted_; - return *this; -} - -bool NodesTSIterator::operator==(const NodesTSIterator &other) { - return sorted_ == other.sorted_ && cursor_ == other.cursor_; -} - -Node *NodesTSIterator::operator->() { - PADDLE_ENFORCE_LT(cursor_, sorted_.size()); - return sorted_[cursor_]; -} - } // namespace analysis } // namespace inference } // namespace paddle diff --git a/paddle/fluid/inference/analysis/ir_passes/subgraph_detector.h b/paddle/fluid/inference/analysis/ir_passes/subgraph_detector.h index ea88edd042aa9d46f66af1aa92f2cb273696c118..5d11c217b69f11d45c6fb6d552dc404fa8313daf 100644 --- a/paddle/fluid/inference/analysis/ir_passes/subgraph_detector.h +++ b/paddle/fluid/inference/analysis/ir_passes/subgraph_detector.h @@ -30,6 +30,7 @@ namespace inference { namespace analysis { using framework::ir::Graph; +using framework::ir::NodesTSIterator; const char kIsFunctionNode[] = "__is_function_node__"; const char kFunctionNodeSubGraph[] = "__function_node_sub_graph__"; @@ -132,32 +133,6 @@ struct Agent { framework::ir::Node *x_; }; -// Topological sorting iterator on nodes. -struct NodesTSIterator - : public std::iterator { - NodesTSIterator() = default; - explicit NodesTSIterator(const std::vector &source); - NodesTSIterator(NodesTSIterator &&other) - : sorted_(std::move(other.sorted_)), cursor_(other.cursor_) { - other.cursor_ = 0; - } - NodesTSIterator(const NodesTSIterator &other); - - framework::ir::Node &operator*(); - NodesTSIterator &operator++(); - // TODO(Superjomn) current implementation just compare the first - // element, need to compare the graph and all the elements in the queue and - // set. - NodesTSIterator &operator=(const NodesTSIterator &other); - bool operator==(const NodesTSIterator &other); - bool operator!=(const NodesTSIterator &other) { return !(*this == other); } - framework::ir::Node *operator->(); - - private: - std::vector sorted_; - size_t cursor_{0}; -}; - // The nodes those have no input will be treated as start points. static std::vector ExtractStartPoints(const Graph &g) { std::vector result; diff --git a/paddle/fluid/inference/analysis/passes/memory_optimize_pass.cc b/paddle/fluid/inference/analysis/passes/memory_optimize_pass.cc index 3d1be9196fdeacd8ff852dbb595473a687352ccf..4b0a9d9b1c48fcb0d5e44ec1b977c817f3c70b2e 100644 --- a/paddle/fluid/inference/analysis/passes/memory_optimize_pass.cc +++ b/paddle/fluid/inference/analysis/passes/memory_optimize_pass.cc @@ -18,6 +18,7 @@ #include #include #include +#include #include #include #include "paddle/fluid/framework/ir/graph_helper.h" @@ -168,7 +169,11 @@ bool FindSuitableTensorToReuse( if (!cluster->count(candidate)) continue; size_t space = space_table.at(candidate); - size_t space_diff = std::abs(space - space_required); + PADDLE_ENFORCE( + space <= std::numeric_limits::type>::max(), + "space overload"); + size_t space_diff = + std::abs((std::make_signed::type)space - space_required); if (space_diff < best_fit.second) { best_fit.first = candidate; best_fit.second = space_diff; diff --git a/paddle/fluid/inference/analysis/passes/memory_optimize_pass.h b/paddle/fluid/inference/analysis/passes/memory_optimize_pass.h index 216f416de0d1003b944337ee98fb4e6a22c66fc5..2da565f2ae15a50a207173b10d4c350456086582 100644 --- a/paddle/fluid/inference/analysis/passes/memory_optimize_pass.h +++ b/paddle/fluid/inference/analysis/passes/memory_optimize_pass.h @@ -13,7 +13,9 @@ // limitations under the License. #pragma once - +#include +#include +#include #include "paddle/fluid/inference/analysis/analysis_pass.h" #include "paddle/fluid/platform/port.h" diff --git a/paddle/fluid/inference/api/CMakeLists.txt b/paddle/fluid/inference/api/CMakeLists.txt index ad0af4005ad154d2f5c67d00dec9d7ec397eb662..85755fc471ae3d37ec5d005882668ccf0c35b354 100644 --- a/paddle/fluid/inference/api/CMakeLists.txt +++ b/paddle/fluid/inference/api/CMakeLists.txt @@ -52,8 +52,8 @@ cc_test(test_analysis_predictor SRCS analysis_predictor_tester.cc DEPS analysis_ if (WITH_ANAKIN AND WITH_MKL) # only needed in CI # compile the libinference_anakin_api.a and anakin.so. - cc_library(inference_anakin_api SRCS api.cc api_anakin_engine.cc DEPS anakin_shared anakin_saber mklml zero_copy_tensor_dummy) - cc_library(inference_anakin_api_shared SHARED SRCS api.cc api_anakin_engine.cc DEPS anakin_shared anakin_saber zero_copy_tensor_dummy) + cc_library(inference_anakin_api SRCS api.cc api_anakin_engine.cc DEPS anakin_shared anakin_saber mklml zero_copy_tensor_dummy device_context) + cc_library(inference_anakin_api_shared SHARED SRCS api.cc api_anakin_engine.cc DEPS anakin_shared anakin_saber zero_copy_tensor_dummy device_context) function(anakin_target target_name) target_compile_options(${target_name} BEFORE PUBLIC ${ANAKIN_COMPILE_EXTRA_FLAGS}) endfunction() diff --git a/paddle/fluid/inference/api/analysis_config.cc b/paddle/fluid/inference/api/analysis_config.cc index 8efd514bd8397f099fd07321ad7e5d4ca253e229..522ab495227e9b8c52b8d38db696fa9b785ba642 100644 --- a/paddle/fluid/inference/api/analysis_config.cc +++ b/paddle/fluid/inference/api/analysis_config.cc @@ -22,7 +22,7 @@ namespace paddle { -PassStrategy *contrib::AnalysisConfig::pass_builder() const { +PassStrategy *AnalysisConfig::pass_builder() const { if (!pass_builder_.get()) { if (use_gpu_) { LOG(INFO) << "Create GPU IR passes"; @@ -42,27 +42,27 @@ PassStrategy *contrib::AnalysisConfig::pass_builder() const { return pass_builder_.get(); } -contrib::AnalysisConfig::AnalysisConfig(const std::string &model_dir) { +AnalysisConfig::AnalysisConfig(const std::string &model_dir) { model_dir_ = model_dir; Update(); } -contrib::AnalysisConfig::AnalysisConfig(const std::string &prog_file, - const std::string ¶ms_file) { +AnalysisConfig::AnalysisConfig(const std::string &prog_file, + const std::string ¶ms_file) { prog_file_ = prog_file; params_file_ = params_file; Update(); } -void contrib::AnalysisConfig::SetModel(const std::string &prog_file_path, - const std::string ¶ms_file_path) { +void AnalysisConfig::SetModel(const std::string &prog_file_path, + const std::string ¶ms_file_path) { prog_file_ = prog_file_path; params_file_ = params_file_path; Update(); } -void contrib::AnalysisConfig::EnableUseGpu(uint64_t memory_pool_init_size_mb, - int device_id) { +void AnalysisConfig::EnableUseGpu(uint64_t memory_pool_init_size_mb, + int device_id) { #ifdef PADDLE_WITH_CUDA use_gpu_ = true; memory_pool_init_size_mb_ = memory_pool_init_size_mb; @@ -74,13 +74,13 @@ void contrib::AnalysisConfig::EnableUseGpu(uint64_t memory_pool_init_size_mb, Update(); } -void contrib::AnalysisConfig::DisableGpu() { +void AnalysisConfig::DisableGpu() { use_gpu_ = false; Update(); } -contrib::AnalysisConfig::AnalysisConfig(const contrib::AnalysisConfig &other) { +AnalysisConfig::AnalysisConfig(const AnalysisConfig &other) { #define CP_MEMBER(member__) member__ = other.member__; // Model related. @@ -89,7 +89,7 @@ contrib::AnalysisConfig::AnalysisConfig(const contrib::AnalysisConfig &other) { CP_MEMBER(params_file_); CP_MEMBER(model_from_memory_); // the memory model reuses prog_file_ and // params_file_ fields. - // Gpu releated. + // Gpu related. CP_MEMBER(use_gpu_); CP_MEMBER(device_id_); CP_MEMBER(memory_pool_init_size_mb_); @@ -97,13 +97,13 @@ contrib::AnalysisConfig::AnalysisConfig(const contrib::AnalysisConfig &other) { CP_MEMBER(enable_memory_optim_); CP_MEMBER(static_memory_optim_); CP_MEMBER(static_memory_optim_force_update_); - // TensorRT releated. + // TensorRT related. CP_MEMBER(use_tensorrt_); CP_MEMBER(tensorrt_workspace_size_); CP_MEMBER(tensorrt_max_batchsize_); CP_MEMBER(tensorrt_min_subgraph_size_); CP_MEMBER(tensorrt_precision_mode_); - // MKLDNN releated. + // MKLDNN related. CP_MEMBER(use_mkldnn_); CP_MEMBER(mkldnn_enabled_op_types_); @@ -130,7 +130,7 @@ contrib::AnalysisConfig::AnalysisConfig(const contrib::AnalysisConfig &other) { Update(); } -void contrib::AnalysisConfig::EnableMKLDNN() { +void AnalysisConfig::EnableMKLDNN() { #ifdef PADDLE_WITH_MKLDNN pass_builder()->EnableMKLDNN(); use_mkldnn_ = true; @@ -142,9 +142,9 @@ void contrib::AnalysisConfig::EnableMKLDNN() { Update(); } -void contrib::AnalysisConfig::EnableTensorRtEngine( +void AnalysisConfig::EnableTensorRtEngine( int workspace_size, int max_batch_size, int min_subgraph_size, - contrib::AnalysisConfig::Precision precision_mode) { + AnalysisConfig::Precision precision_mode) { #ifdef PADDLE_WITH_CUDA if (!use_gpu()) { LOG(ERROR) << "To use TensorRT engine, please call EnableGpu() first"; @@ -165,7 +165,7 @@ void contrib::AnalysisConfig::EnableTensorRtEngine( } // TODO(Superjomn) refactor this, buggy. -void contrib::AnalysisConfig::Update() { +void AnalysisConfig::Update() { auto info = SerializeInfoCache(); if (info == serialized_info_cache_) return; @@ -225,7 +225,7 @@ void contrib::AnalysisConfig::Update() { } } -std::string contrib::AnalysisConfig::SerializeInfoCache() { +std::string AnalysisConfig::SerializeInfoCache() { std::stringstream ss; ss << model_dir_; ss << prog_file_; @@ -260,14 +260,14 @@ std::string contrib::AnalysisConfig::SerializeInfoCache() { return ss.str(); } -void contrib::AnalysisConfig::SetCpuMathLibraryNumThreads( +void AnalysisConfig::SetCpuMathLibraryNumThreads( int cpu_math_library_num_threads) { cpu_math_library_num_threads_ = cpu_math_library_num_threads; Update(); } -float contrib::AnalysisConfig::fraction_of_gpu_memory_for_pool() const { +float AnalysisConfig::fraction_of_gpu_memory_for_pool() const { #ifdef PADDLE_WITH_CUDA // Get the GPU memory details and calculate the fraction of memory for the // GPU memory pool. @@ -282,8 +282,8 @@ float contrib::AnalysisConfig::fraction_of_gpu_memory_for_pool() const { #endif } -void contrib::AnalysisConfig::EnableMemoryOptim( - bool static_optim, bool force_update_static_cache) { +void AnalysisConfig::EnableMemoryOptim(bool static_optim, + bool force_update_static_cache) { enable_memory_optim_ = true; static_memory_optim_ = static_optim; static_memory_optim_force_update_ = force_update_static_cache; @@ -291,14 +291,14 @@ void contrib::AnalysisConfig::EnableMemoryOptim( Update(); } -bool contrib::AnalysisConfig::enable_memory_optim() const { +bool AnalysisConfig::enable_memory_optim() const { return enable_memory_optim_; } -void contrib::AnalysisConfig::SetModelBuffer(const char *prog_buffer, - size_t prog_buffer_size, - const char *param_buffer, - size_t param_buffer_size) { +void AnalysisConfig::SetModelBuffer(const char *prog_buffer, + size_t prog_buffer_size, + const char *param_buffer, + size_t param_buffer_size) { prog_file_ = std::string(prog_buffer, prog_buffer + prog_buffer_size); params_file_ = std::string(param_buffer, param_buffer + param_buffer_size); model_from_memory_ = true; @@ -306,7 +306,7 @@ void contrib::AnalysisConfig::SetModelBuffer(const char *prog_buffer, Update(); } -NativeConfig contrib::AnalysisConfig::ToNativeConfig() const { +NativeConfig AnalysisConfig::ToNativeConfig() const { NativeConfig config; config.model_dir = model_dir_; config.prog_file = prog_file_; @@ -318,4 +318,9 @@ NativeConfig contrib::AnalysisConfig::ToNativeConfig() const { return config; } +void AnalysisConfig::SwitchIrDebug(int x) { + ir_debug_ = x; + Update(); +} + } // namespace paddle diff --git a/paddle/fluid/inference/api/analysis_predictor.cc b/paddle/fluid/inference/api/analysis_predictor.cc index 7d97aea714ab1bd653965c1900f900b52c3b4616..e8964c4acea0d220deca048a018eb7de42d7e4e5 100644 --- a/paddle/fluid/inference/api/analysis_predictor.cc +++ b/paddle/fluid/inference/api/analysis_predictor.cc @@ -47,7 +47,6 @@ DECLARE_bool(profile); namespace paddle { -using contrib::AnalysisConfig; using inference::Singleton; #if PADDLE_WITH_TENSORRT using inference::tensorrt::TRTInt8Calibrator; @@ -59,7 +58,8 @@ namespace { bool IsPersistable(const framework::VarDesc *var) { if (var->Persistable() && var->GetType() != framework::proto::VarType::FEED_MINIBATCH && - var->GetType() != framework::proto::VarType::FETCH_LIST) { + var->GetType() != framework::proto::VarType::FETCH_LIST && + var->GetType() != framework::proto::VarType::RAW) { return true; } return false; @@ -123,6 +123,15 @@ bool AnalysisPredictor::PrepareProgram( if (!program) { if (!LoadProgramDesc()) return false; + // If not cloned, the parameters should be loaded. + // If config_.ir_optim() is True, parameters is loaded in + // OptimizeInferenceProgram(), but other persistable variables + // (like RAW type var) are not created in scope. + // If config_.ir_optim() is False, parameters is loaded in LoadParameters(), + // still need to create other persistable variables. + // So in both case, create persistable variables at first. + executor_->CreateVariables(*inference_program_, 0, true, sub_scope_); + // Optimize the program, and load parameters and modify them in the // scope_. // This will change the scope_ address. @@ -130,15 +139,6 @@ bool AnalysisPredictor::PrepareProgram( status_ir_optim_enabled_ = true; OptimizeInferenceProgram(); } else { - // If the parent_scope is passed, we assert that the persistable variables - // are already created, so just create the no persistable variables. - - // If not cloned, the parameters should be loaded - // OptimizeInferenceProgram. - // So in both cases, just the local variables are needed to load, not the - // parematers. - executor_->CreateVariables(*inference_program_, 0, true, sub_scope_); - // Load parameters LOG(INFO) << "load parameters "; LoadParameters(); @@ -376,7 +376,7 @@ void AnalysisPredictor::OptimizeInferenceProgram() { } argument_.SetIrAnalysisPasses(passes); argument_.SetAnalysisPasses(config_.pass_builder()->AnalysisPasses()); - argument_.SetScopeNotOwned(const_cast(scope_.get())); + argument_.SetScopeNotOwned(scope_.get()); Analyzer().Run(&argument_); PADDLE_ENFORCE(argument_.scope_valid()); @@ -392,7 +392,7 @@ std::unique_ptr CreatePaddlePredictor< AnalysisConfig, PaddleEngineKind::kAnalysis>(const AnalysisConfig &config) { VLOG(3) << "create AnalysisConfig"; if (config.use_gpu()) { - // 1. GPU memeroy + // 1. GPU memory PADDLE_ENFORCE_GT(config.memory_pool_init_size_mb(), 0.f); PADDLE_ENFORCE_GE(config.gpu_device_id(), 0, "Invalid device id %d", config.gpu_device_id()); @@ -421,7 +421,7 @@ std::unique_ptr CreatePaddlePredictor< if (!dynamic_cast(predictor.get())->Init(nullptr)) { return nullptr; } - return std::move(predictor); + return predictor; } void AnalysisPredictor::PrepareFeedFetch() { @@ -726,11 +726,15 @@ bool AnalysisPredictor::need_collect_var_shapes_for_memory_optim() { return need; } +std::string AnalysisPredictor::GetSerializedProgram() const { + return inference_program_->Proto()->SerializeAsString(); +} + template <> -std::unique_ptr CreatePaddlePredictor( - const contrib::AnalysisConfig &config) { - return CreatePaddlePredictor(config); +std::unique_ptr CreatePaddlePredictor( + const AnalysisConfig &config) { + return CreatePaddlePredictor( + config); } } // namespace paddle diff --git a/paddle/fluid/inference/api/analysis_predictor.h b/paddle/fluid/inference/api/analysis_predictor.h index 921aa90952d6cf37ef84890b6ad7340cad6a8eb6..d5445c58e45ae64a8cfab03cb610e3677729338b 100644 --- a/paddle/fluid/inference/api/analysis_predictor.h +++ b/paddle/fluid/inference/api/analysis_predictor.h @@ -33,7 +33,6 @@ using inference::analysis::Argument; using inference::analysis::Analyzer; using framework::proto::ProgramDesc; using framework::NaiveExecutor; -using contrib::AnalysisConfig; /** \brief This predictor is based on the original native predictor with IR and * Analysis support. @@ -75,6 +74,8 @@ class AnalysisPredictor : public PaddlePredictor { void SetMkldnnThreadID(int tid); + std::string GetSerializedProgram() const override; + protected: // For memory optimization. bool need_collect_var_shapes_for_memory_optim(); @@ -121,7 +122,7 @@ class AnalysisPredictor : public PaddlePredictor { #endif private: - contrib::AnalysisConfig config_; + AnalysisConfig config_; Argument argument_; std::unique_ptr executor_; platform::Place place_; diff --git a/paddle/fluid/inference/api/analysis_predictor_tester.cc b/paddle/fluid/inference/api/analysis_predictor_tester.cc index 4688e93d7102109d2c7ece9ba37bc8f2d311dcf1..6696839b53fb21c274843afd86b5d8b5c2042c51 100644 --- a/paddle/fluid/inference/api/analysis_predictor_tester.cc +++ b/paddle/fluid/inference/api/analysis_predictor_tester.cc @@ -24,7 +24,6 @@ DEFINE_string(dirname, "", "dirname to tests."); namespace paddle { -using contrib::AnalysisConfig; TEST(AnalysisPredictor, analysis_off) { AnalysisConfig config; @@ -197,7 +196,7 @@ TEST(AnalysisPredictor, memory_optim) { AnalysisConfig config(FLAGS_dirname); config.DisableGpu(); config.EnableMemoryOptim(true); - config.pass_builder()->TurnOnDebug(); + config.SwitchIrDebug(); auto native_predictor = CreatePaddlePredictor(config.ToNativeConfig()); @@ -215,6 +214,8 @@ TEST(AnalysisPredictor, memory_optim) { { // The first predictor help to cache the memory optimize strategy. auto predictor = CreatePaddlePredictor(config); + LOG(INFO) << "serialized program: " << predictor->GetSerializedProgram(); + ASSERT_FALSE(predictor->GetSerializedProgram().empty()); // Run several times to check the parameters are not reused by mistake. for (int i = 0; i < 5; i++) { diff --git a/paddle/fluid/inference/api/api.cc b/paddle/fluid/inference/api/api.cc index 9be059c73e20ebeeff2c4b6e8e5502e4a56fd0d6..f83537f064187e67a08c8bbce52707d1c824abeb 100644 --- a/paddle/fluid/inference/api/api.cc +++ b/paddle/fluid/inference/api/api.cc @@ -12,6 +12,8 @@ // See the License for the specific language governing permissions and // limitations under the License. +#include +#include "paddle/fluid/framework/commit.h" #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/scope.h" #include "paddle/fluid/inference/api/paddle_inference_api.h" @@ -90,11 +92,19 @@ void PaddleBuf::Reset(void *data, size_t length) { void PaddleBuf::Free() { if (memory_owned_ && data_) { - PADDLE_ENFORCE_GT(length_, 0); + PADDLE_ENFORCE_GT(length_, 0UL); free(static_cast(data_)); data_ = nullptr; length_ = 0; } } +std::string get_version() { + std::stringstream ss; + ss << "version: " << framework::paddle_version() << "\n"; + ss << "commit: " << framework::paddle_commit() << "\n"; + ss << "branch: " << framework::paddle_compile_branch() << "\n"; + return ss.str(); +} + } // namespace paddle diff --git a/paddle/fluid/inference/api/api_impl.cc b/paddle/fluid/inference/api/api_impl.cc index e18bc02d92eb517fa20dc83811694b8ac80ae316..97c164bdef7a4b3e66be78526793f3830ada398b 100644 --- a/paddle/fluid/inference/api/api_impl.cc +++ b/paddle/fluid/inference/api/api_impl.cc @@ -290,7 +290,7 @@ std::unique_ptr CreatePaddlePredictor< NativeConfig, PaddleEngineKind::kNative>(const NativeConfig &config) { VLOG(3) << "create NativePaddlePredictor"; if (config.use_gpu) { - // 1. GPU memeroy + // 1. GPU memory PADDLE_ENFORCE_GE( config.fraction_of_gpu_memory, 0.f, "fraction_of_gpu_memory in the config should be set to range (0., 1.]"); diff --git a/paddle/fluid/inference/api/api_impl_tester.cc b/paddle/fluid/inference/api/api_impl_tester.cc index 54895679ca37362c7267677af80274b8de95e296..e82cb53bf073d3d1ab9a518218edaf430728463f 100644 --- a/paddle/fluid/inference/api/api_impl_tester.cc +++ b/paddle/fluid/inference/api/api_impl_tester.cc @@ -295,7 +295,7 @@ TEST(inference_api_native, image_classification_gpu) { #endif TEST(PassBuilder, Delete) { - contrib::AnalysisConfig config; + AnalysisConfig config; config.DisableGpu(); config.pass_builder()->DeletePass("attention_lstm_fuse_pass"); const auto& passes = config.pass_builder()->AllPasses(); diff --git a/paddle/fluid/inference/api/api_tester.cc b/paddle/fluid/inference/api/api_tester.cc index 7a579610eefda24c911edd28b5f3a178aa10ab1e..2c450ef7cead4d5c3870d5e9186eb221e5dc19a0 100644 --- a/paddle/fluid/inference/api/api_tester.cc +++ b/paddle/fluid/inference/api/api_tester.cc @@ -61,4 +61,10 @@ TEST(paddle_inference_api, demo) { predictor->Run({}, &outputs); } +TEST(paddle_inference_api, get_version) { + LOG(INFO) << "paddle version:\n" << get_version(); + auto version = get_version(); + ASSERT_FALSE(version.empty()); +} + } // namespace paddle diff --git a/paddle/fluid/inference/api/demo_ci/trt_mobilenet_demo.cc b/paddle/fluid/inference/api/demo_ci/trt_mobilenet_demo.cc index 338a0cec161f352781f132aea71dd56f68840c62..f7da55c9ae368763786c1b1fd3e86d942c5e9fe8 100644 --- a/paddle/fluid/inference/api/demo_ci/trt_mobilenet_demo.cc +++ b/paddle/fluid/inference/api/demo_ci/trt_mobilenet_demo.cc @@ -36,7 +36,7 @@ namespace demo { */ void Main() { std::unique_ptr predictor; - paddle::contrib::AnalysisConfig config; + paddle::AnalysisConfig config; config.EnableUseGpu(100, 0); config.SetModel(FLAGS_modeldir + "/__model__", FLAGS_modeldir + "/__params__"); diff --git a/paddle/fluid/inference/api/demo_ci/vis_demo.cc b/paddle/fluid/inference/api/demo_ci/vis_demo.cc index 5320992b7e78f4aa0ea8950af03038c1953dd027..0d2c418c56db620c71d99b64ee79b18be427cc34 100644 --- a/paddle/fluid/inference/api/demo_ci/vis_demo.cc +++ b/paddle/fluid/inference/api/demo_ci/vis_demo.cc @@ -34,7 +34,6 @@ DEFINE_bool(use_gpu, false, "Whether use gpu."); namespace paddle { namespace demo { -using contrib::AnalysisConfig; /* * Use the native and analysis fluid engine to inference the demo. */ diff --git a/paddle/fluid/inference/api/paddle_analysis_config.h b/paddle/fluid/inference/api/paddle_analysis_config.h index 5b899b26d60dec3634d7016c925143e1ae26992d..c1c6227cdd8b2042f6765c7932327ecae246c260 100644 --- a/paddle/fluid/inference/api/paddle_analysis_config.h +++ b/paddle/fluid/inference/api/paddle_analysis_config.h @@ -29,11 +29,6 @@ namespace paddle { class AnalysisPredictor; -// == -// -// ----------------------------------------------------------------------------------- -// NOTE: The following APIs are not mature yet, we are still working on them. -namespace contrib { // NOTE WIP, not stable yet. struct AnalysisConfig { @@ -145,9 +140,12 @@ struct AnalysisConfig { */ bool tensorrt_engine_enabled() const { return use_tensorrt_; } - /** Control whther to debug IR graph analysis phase. + /** \brief Control whether to debug IR graph analysis phase. + * + * This will generate DOT files for visualizing the computation graph after + * each analysis pass applied. */ - void SwitchIrDebug(int x = true) { ir_debug_ = x; } + void SwitchIrDebug(int x = true); /** Turn on MKLDNN. */ @@ -214,12 +212,12 @@ struct AnalysisConfig { std::string prog_file_; std::string params_file_; - // GPU releated. + // GPU related. bool use_gpu_{false}; int device_id_{0}; uint64_t memory_pool_init_size_mb_{100}; // initial size is 100MB. - // TensorRT releated. + // TensorRT related. bool use_tensorrt_{false}; // For workspace_size, refer it from here: // https://docs.nvidia.com/deeplearning/sdk/tensorrt-developer-guide/index.html#troubleshooting @@ -260,5 +258,4 @@ struct AnalysisConfig { mutable std::unique_ptr pass_builder_; }; -} // namespace contrib } // namespace paddle diff --git a/paddle/fluid/inference/api/paddle_api.h b/paddle/fluid/inference/api/paddle_api.h index 46b510fd1ec94c59032b8f41a2ac4d6aa87dc150..c9a45b4aa3b4037d3725622fc960848bc1ccfb2c 100644 --- a/paddle/fluid/inference/api/paddle_api.h +++ b/paddle/fluid/inference/api/paddle_api.h @@ -16,6 +16,12 @@ /*! \file paddle_api.h */ +/*! \mainpage Paddle Inference APIs + * \section intro_sec Introduction + * The Paddle inference library aims to offer an high performance inference SDK + * for Paddle users. + */ + #include #include #include @@ -34,26 +40,49 @@ enum PaddleDType { }; /** - *\brief Memory menager for PaddleTensor. + * \brief Memory manager for `PaddleTensor`. * - *The PaddleBuf holds a buffer for data input or output. The memory can be - *allocated by user or by PaddleBuf itself, but in any case, the PaddleBuf - *should be reused for better performance. + * The PaddleBuf holds a buffer for data input or output. The memory can be + * allocated by user or by PaddleBuf itself, but in any case, the PaddleBuf + * should be reused for better performance. * - *For user allocated memory, the following API can be used: - *- PaddleBuf(void* data, size_t length) to set an external memory by - *specifying - * the memory address and length. - *- Reset(void* data, size_t length) to reset the PaddleBuf with an external + * For user allocated memory, the following API can be used: + * - PaddleBuf(void* data, size_t length) to set an external memory by + * specifying the memory address and length. + * - Reset(void* data, size_t length) to reset the PaddleBuf with an external *memory. - *ATTENTION, for user allocated memory, deallocation should be done by users + * ATTENTION, for user allocated memory, deallocation should be done by users *externally after the program finished. The PaddleBuf won't do any allocation *or deallocation. * - *To have the PaddleBuf allocate and manage the memory: - *- PaddleBuf(size_t length) will allocate a memory of size `length`. - *- Resize(size_t length) resize the memory to no less than `length`, ATTENTION + * To have the PaddleBuf allocate and manage the memory: + * - PaddleBuf(size_t length) will allocate a memory of size `length`. + * - Resize(size_t length) resize the memory to no less than `length`, ATTENTION * if the allocated memory is larger than `length`, nothing will done. + * + * Usage: + * + * Let PaddleBuf manage the memory internally. + * \code{cpp} + * const int num_elements = 128; + * PaddleBuf buf(num_elements * sizeof(float)); + * \endcode + * + * Or + * \code{cpp} + * PaddleBuf buf; + * buf.Resize(num_elements * sizeof(float)); + * \endcode + * Works the exactly the same. + * + * One can also make the `PaddleBuf` use the external memory. + * \code{cpp} + * PaddleBuf buf; + * void* external_memory = new float[num_elements]; + * buf.Reset(external_memory, num_elements*sizeof(float)); + * ... + * delete[] external_memory; // manage the memory lifetime outside. + * \endcode */ class PaddleBuf { public: @@ -78,7 +107,7 @@ class PaddleBuf { /** Tell whether the buffer is empty. */ bool empty() const { return length_ == 0; } - /** Get the memory address. + /** Get the data's memory address. */ void* data() const { return data_; } /** Get the memory length. @@ -110,7 +139,8 @@ struct PaddleTensor { }; enum class PaddlePlace { kUNK = -1, kCPU, kGPU }; -/** Tensor without copy, currently only supports AnalysisPredictor. + +/** Tensor without copy, currently only supports `AnalysisPredictor`. */ class ZeroCopyTensor { public: @@ -215,6 +245,14 @@ class PaddlePredictor { */ virtual ~PaddlePredictor() = default; + /** \brief Get the serialized model program that executes in inference phase. + * Its data type is ProgramDesc, which is a protobuf message. + */ + virtual std::string GetSerializedProgram() const { + assert(false); // Force raise error. + return "NotImplemented"; + } + /** The common configs for all the predictors. */ struct Config { @@ -261,9 +299,11 @@ struct NativeConfig : public PaddlePredictor::Config { * * Usage: * + * \code{.cpp} * NativeConfig config; * ... // change the configs. * auto native_predictor = CreatePaddlePredictor(config); + * \endcode * * FOR EXTENSION DEVELOPER: * Different predictors are designated by config type. Similar configs can be @@ -288,4 +328,6 @@ std::unique_ptr CreatePaddlePredictor(const ConfigT& config); int PaddleDtypeSize(PaddleDType dtype); +std::string get_version(); + } // namespace paddle diff --git a/paddle/fluid/inference/api/paddle_pass_builder.cc b/paddle/fluid/inference/api/paddle_pass_builder.cc index 039389a4cf99da6c2576c148d8c294e5d79aa7a8..f9c13c2fa84b3b5d629297d3f44a6f5889a734f4 100644 --- a/paddle/fluid/inference/api/paddle_pass_builder.cc +++ b/paddle/fluid/inference/api/paddle_pass_builder.cc @@ -66,8 +66,54 @@ void GpuPassStrategy::EnableMKLDNN() { LOG(ERROR) << "GPU not support MKLDNN yet"; } +GpuPassStrategy::GpuPassStrategy() : PassStrategy({}) { + passes_.assign({ + "infer_clean_graph_pass", // + "identity_scale_op_clean_pass", // + "conv_affine_channel_fuse_pass", // + "conv_eltwiseadd_affine_channel_fuse_pass", // + "conv_bn_fuse_pass", // +#if CUDNN_VERSION >= 7100 // To run conv_fusion, the version of cudnn must be + // guaranteed at least v7 + "conv_elementwise_add_act_fuse_pass", // + "conv_elementwise_add2_act_fuse_pass", // + "conv_elementwise_add_fuse_pass", // +#endif + }); + + for (int i = 6; i >= 3; i--) { + passes_.push_back("transpose_flatten" + std::to_string(i) + + "_concat_fuse_pass"); + } + use_gpu_ = true; +} + void PaddlePassBuilder::AppendAnalysisPass(const std::string &pass) { analysis_passes_.push_back(pass); } +CpuPassStrategy::CpuPassStrategy() : PassStrategy({}) { + // NOTE the large fusions should be located in the front, so that they will + // not be damaged by smaller ones. + passes_.assign({ + "infer_clean_graph_pass", // + "attention_lstm_fuse_pass", // + "seqpool_concat_fuse_pass", // + "seqconv_eltadd_relu_fuse_pass", // + // "embedding_fc_lstm_fuse_pass", // + "fc_lstm_fuse_pass", // + "mul_lstm_fuse_pass", // + "fc_gru_fuse_pass", // + "mul_gru_fuse_pass", // + "seq_concat_fc_fuse_pass", // + "fc_fuse_pass", // + "repeated_fc_relu_fuse_pass", // + "squared_mat_sub_fuse_pass", // + "conv_bn_fuse_pass", // + "conv_eltwiseadd_bn_fuse_pass", // + "is_test_pass", // + "identity_scale_op_clean_pass", // + }); + use_gpu_ = false; +} } // namespace paddle diff --git a/paddle/fluid/inference/api/paddle_pass_builder.h b/paddle/fluid/inference/api/paddle_pass_builder.h index 391932a1ee018c45818457c55fd8f82a22ab7405..2524d89fcd1322e105ad2217347aa2380448f2bc 100644 --- a/paddle/fluid/inference/api/paddle_pass_builder.h +++ b/paddle/fluid/inference/api/paddle_pass_builder.h @@ -97,29 +97,7 @@ class PassStrategy : public PaddlePassBuilder { */ class CpuPassStrategy : public PassStrategy { public: - CpuPassStrategy() : PassStrategy({}) { - // NOTE the large fusions should be located in the front, so that they will - // not be damaged by smaller ones. - passes_.assign({ - "infer_clean_graph_pass", // - "attention_lstm_fuse_pass", // - "seqpool_concat_fuse_pass", // - "seqconv_eltadd_relu_fuse_pass", // - // "embedding_fc_lstm_fuse_pass", // - "fc_lstm_fuse_pass", // - "mul_lstm_fuse_pass", // - "fc_gru_fuse_pass", // - "mul_gru_fuse_pass", // - "seq_concat_fc_fuse_pass", // - "fc_fuse_pass", // - "repeated_fc_relu_fuse_pass", // - "squared_mat_sub_fuse_pass", // - "conv_bn_fuse_pass", // - "conv_eltwiseadd_bn_fuse_pass", // - "is_test_pass", // - }); - use_gpu_ = false; - } + CpuPassStrategy(); explicit CpuPassStrategy(const CpuPassStrategy &other) : PassStrategy(other.AllPasses()) {} @@ -152,26 +130,7 @@ class CpuPassStrategy : public PassStrategy { */ class GpuPassStrategy : public PassStrategy { public: - GpuPassStrategy() : PassStrategy({}) { - passes_.assign({ - "infer_clean_graph_pass", // - "conv_affine_channel_fuse_pass", // - "conv_eltwiseadd_affine_channel_fuse_pass", // - "conv_bn_fuse_pass", // -#if CUDNN_VERSION >= 7100 // To run conv_fusion, the version of cudnn must be - // guaranteed at least v7 - "conv_elementwise_add_act_fuse_pass", // - "conv_elementwise_add2_act_fuse_pass", // - "conv_elementwise_add_fuse_pass", // -#endif - }); - - for (int i = 6; i >= 3; i--) { - passes_.push_back("transpose_flatten" + std::to_string(i) + - "_concat_fuse_pass"); - } - use_gpu_ = true; - } + GpuPassStrategy(); explicit GpuPassStrategy(const GpuPassStrategy &other) : PassStrategy(other.AllPasses()) { diff --git a/paddle/fluid/inference/tensorrt/trt_int8_calibrator.h b/paddle/fluid/inference/tensorrt/trt_int8_calibrator.h index 919f5d55f88c3a6473f66371e2f3d91f3c4721c5..5815bc9a1464293e0a56f05e34183580eac96cea 100644 --- a/paddle/fluid/inference/tensorrt/trt_int8_calibrator.h +++ b/paddle/fluid/inference/tensorrt/trt_int8_calibrator.h @@ -13,16 +13,16 @@ // limitations under the License. #pragma once + +#include +#include #include #include -#include +#include // NOLINT #include #include #include #include - -#include -#include #include "paddle/fluid/framework/tensor.h" #include "paddle/fluid/inference/tensorrt/engine.h" #include "paddle/fluid/platform/place.h" diff --git a/paddle/fluid/inference/tests/api/CMakeLists.txt b/paddle/fluid/inference/tests/api/CMakeLists.txt index 423c39813f05af0d6aaade184914e6777c9b8a83..55ab04bfe16ec6a3d97c443f59c72e7b85fb1899 100644 --- a/paddle/fluid/inference/tests/api/CMakeLists.txt +++ b/paddle/fluid/inference/tests/api/CMakeLists.txt @@ -54,15 +54,19 @@ else() message(WARNING "These tests has been disabled in OSX or WITH_MKL=OFF before being fixed: \n test_analyzer_seq_pool1") endif() + # RNN2 set(RNN2_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/rnn2") download_model_and_data(${RNN2_INSTALL_DIR} "rnn2_model.tar.gz" "rnn2_data.txt.tar.gz") inference_analysis_api_test(test_analyzer_rnn2 ${RNN2_INSTALL_DIR} analyzer_rnn2_tester.cc) +# TODO(luotao, Superjom) Disable DAM test, temporarily fix +# https://github.com/PaddlePaddle/Paddle/issues/15032#issuecomment-455990914. +# After inference framework refactor, will reopen it. # normal DAM set(DAM_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/dam") download_model_and_data(${DAM_INSTALL_DIR} "DAM_model.tar.gz" "DAM_data.txt.tar.gz") -inference_analysis_api_test(test_analyzer_dam ${DAM_INSTALL_DIR} analyzer_dam_tester.cc EXTRA_DEPS legacy_allocator SERIAL) +#inference_analysis_api_test(test_analyzer_dam ${DAM_INSTALL_DIR} analyzer_dam_tester.cc EXTRA_DEPS legacy_allocator SERIAL) # small DAM set(DAM_SMALL_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/small_dam") @@ -115,6 +119,10 @@ if (NOT EXISTS ${MOBILENET_INSTALL_DIR}) endif() inference_analysis_api_test_with_refer_result(test_analyzer_mobilenet_transpose ${MOBILENET_INSTALL_DIR} analyzer_vis_tester.cc SERIAL) +# googlenet +inference_analysis_api_test_with_fake_data(test_analyzer_googlenet + "${INFERENCE_DEMO_INSTALL_DIR}/googlenet" analyzer_resnet50_tester.cc "googlenet.tar.gz" SERIAL) + # resnet50 inference_analysis_api_test_with_fake_data(test_analyzer_resnet50 "${INFERENCE_DEMO_INSTALL_DIR}/resnet50" analyzer_resnet50_tester.cc "resnet50_model.tar.gz" SERIAL) @@ -123,6 +131,11 @@ inference_analysis_api_test_with_fake_data(test_analyzer_resnet50 inference_analysis_api_test_with_fake_data(test_analyzer_mobilenet_depthwise_conv "${INFERENCE_DEMO_INSTALL_DIR}/mobilenet_depthwise_conv" analyzer_resnet50_tester.cc "mobilenet_model.tar.gz" SERIAL) +# bert, max_len=20, embedding_dim=128 +set(BERT_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/bert_emb128") +download_model_and_data(${BERT_INSTALL_DIR} "bert_emb128_model.tar.gz" "bert_data_len20.txt.tar.gz") +inference_analysis_api_test(test_analyzer_bert ${BERT_INSTALL_DIR} analyzer_bert_tester.cc SERIAL) + # anakin if (WITH_ANAKIN AND WITH_MKL) # only needed in CI # anakin rnn1 diff --git a/paddle/fluid/inference/tests/api/analyzer_bert_tester.cc b/paddle/fluid/inference/tests/api/analyzer_bert_tester.cc new file mode 100644 index 0000000000000000000000000000000000000000..f646fd6d91c81b6738e4fc5278739307fa5f99b5 --- /dev/null +++ b/paddle/fluid/inference/tests/api/analyzer_bert_tester.cc @@ -0,0 +1,223 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/inference/tests/api/tester_helper.h" + +namespace paddle { +namespace inference { + +using paddle::PaddleTensor; + +template +void GetValueFromStream(std::stringstream *ss, T *t) { + (*ss) >> (*t); +} + +template <> +void GetValueFromStream(std::stringstream *ss, std::string *t) { + *t = ss->str(); +} + +// Split string to vector +template +void Split(const std::string &line, char sep, std::vector *v) { + std::stringstream ss; + T t; + for (auto c : line) { + if (c != sep) { + ss << c; + } else { + GetValueFromStream(&ss, &t); + v->push_back(std::move(t)); + ss.str({}); + ss.clear(); + } + } + + if (!ss.str().empty()) { + GetValueFromStream(&ss, &t); + v->push_back(std::move(t)); + ss.str({}); + ss.clear(); + } +} + +template +constexpr paddle::PaddleDType GetPaddleDType(); + +template <> +constexpr paddle::PaddleDType GetPaddleDType() { + return paddle::PaddleDType::INT64; +} + +template <> +constexpr paddle::PaddleDType GetPaddleDType() { + return paddle::PaddleDType::FLOAT32; +} + +// Parse tensor from string +template +bool ParseTensor(const std::string &field, paddle::PaddleTensor *tensor) { + std::vector data; + Split(field, ':', &data); + if (data.size() < 2) return false; + + std::string shape_str = data[0]; + + std::vector shape; + Split(shape_str, ' ', &shape); + + std::string mat_str = data[1]; + + std::vector mat; + Split(mat_str, ' ', &mat); + + tensor->shape = shape; + auto size = + std::accumulate(shape.begin(), shape.end(), 1, std::multiplies()) * + sizeof(T); + tensor->data.Resize(size); + std::copy(mat.begin(), mat.end(), static_cast(tensor->data.data())); + tensor->dtype = GetPaddleDType(); + + return true; +} + +// Parse input tensors from string +bool ParseLine(const std::string &line, + std::vector *tensors) { + std::vector fields; + Split(line, ';', &fields); + + if (fields.size() < 5) return false; + + tensors->clear(); + tensors->reserve(5); + + int i = 0; + // src_id + paddle::PaddleTensor src_id; + ParseTensor(fields[i++], &src_id); + tensors->push_back(src_id); + + // pos_id + paddle::PaddleTensor pos_id; + ParseTensor(fields[i++], &pos_id); + tensors->push_back(pos_id); + + // segment_id + paddle::PaddleTensor segment_id; + ParseTensor(fields[i++], &segment_id); + tensors->push_back(segment_id); + + // self_attention_bias + paddle::PaddleTensor self_attention_bias; + ParseTensor(fields[i++], &self_attention_bias); + tensors->push_back(self_attention_bias); + + // next_segment_index + paddle::PaddleTensor next_segment_index; + ParseTensor(fields[i++], &next_segment_index); + tensors->push_back(next_segment_index); + + return true; +} + +bool LoadInputData(std::vector> *inputs) { + if (FLAGS_infer_data.empty()) { + LOG(ERROR) << "please set input data path"; + return false; + } + + std::ifstream fin(FLAGS_infer_data); + std::string line; + int sample = 0; + + // The unit-test dataset only have 10 samples, each sample have 5 feeds. + while (std::getline(fin, line)) { + std::vector feed_data; + ParseLine(line, &feed_data); + inputs->push_back(std::move(feed_data)); + sample++; + if (!FLAGS_test_all_data && sample == FLAGS_batch_size) break; + } + LOG(INFO) << "number of samples: " << sample; + + return true; +} + +void SetConfig(AnalysisConfig *config) { config->SetModel(FLAGS_infer_model); } + +void profile(bool use_mkldnn = false) { + AnalysisConfig config; + SetConfig(&config); + + if (use_mkldnn) { + config.EnableMKLDNN(); + } + + std::vector outputs; + std::vector> inputs; + LoadInputData(&inputs); + TestPrediction(reinterpret_cast(&config), + inputs, &outputs, FLAGS_num_threads); +} + +TEST(Analyzer_bert, profile) { profile(); } +#ifdef PADDLE_WITH_MKLDNN +TEST(Analyzer_bert, profile_mkldnn) { profile(true); } +#endif + +// Check the fuse status +TEST(Analyzer_bert, fuse_statis) { + AnalysisConfig cfg; + SetConfig(&cfg); + int num_ops; + auto predictor = CreatePaddlePredictor(cfg); + auto fuse_statis = GetFuseStatis( + static_cast(predictor.get()), &num_ops); + LOG(INFO) << "num_ops: " << num_ops; +} + +// Compare result of NativeConfig and AnalysisConfig +void compare(bool use_mkldnn = false) { + AnalysisConfig cfg; + SetConfig(&cfg); + if (use_mkldnn) { + cfg.EnableMKLDNN(); + } + + std::vector> inputs; + LoadInputData(&inputs); + CompareNativeAndAnalysis( + reinterpret_cast(&cfg), inputs); +} + +TEST(Analyzer_bert, compare) { compare(); } +#ifdef PADDLE_WITH_MKLDNN +TEST(Analyzer_bert, compare_mkldnn) { compare(true /* use_mkldnn */); } +#endif + +// Compare Deterministic result +TEST(Analyzer_bert, compare_determine) { + AnalysisConfig cfg; + SetConfig(&cfg); + + std::vector> inputs; + LoadInputData(&inputs); + CompareDeterministic(reinterpret_cast(&cfg), + inputs); +} +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/tests/api/analyzer_dam_tester.cc b/paddle/fluid/inference/tests/api/analyzer_dam_tester.cc index e78ab942d113323fecf5510dca85fb5db734efc8..735e4fb563788438ee49ff6308d11f4dbe4962be 100644 --- a/paddle/fluid/inference/tests/api/analyzer_dam_tester.cc +++ b/paddle/fluid/inference/tests/api/analyzer_dam_tester.cc @@ -19,7 +19,6 @@ DEFINE_int32(max_turn_num, 9, namespace paddle { namespace inference { -using contrib::AnalysisConfig; constexpr int32_t kMaxTurnLen = 50; @@ -165,7 +164,7 @@ void PrepareInputs(std::vector *input_slots, DataRecord *data, input_slots->push_back(std::move(response_mask_tensor)); } -void SetConfig(contrib::AnalysisConfig *cfg) { +void SetConfig(AnalysisConfig *cfg) { cfg->SetModel(FLAGS_infer_model + "/__model__", FLAGS_infer_model + "/param"); cfg->SwitchSpecifyInputNames(); cfg->SwitchIrOptim(true); @@ -187,7 +186,7 @@ void SetInput(std::vector> *inputs) { // Easy for profiling independently. void profile(bool use_mkldnn = false) { - contrib::AnalysisConfig cfg; + AnalysisConfig cfg; SetConfig(&cfg); if (use_mkldnn) { @@ -223,7 +222,7 @@ TEST(Analyzer_dam, profile_mkldnn) { profile(true /* use_mkldnn */); } // Check the fuse status TEST(Analyzer_dam, fuse_statis) { - contrib::AnalysisConfig cfg; + AnalysisConfig cfg; SetConfig(&cfg); int num_ops; @@ -256,7 +255,7 @@ void compare(bool use_mkldnn = false) { TEST(Analyzer_dam, compare_with_static_memory_optim) { // The small dam will core in CI, but works in local. if (FLAGS_max_turn_num == 9) { - contrib::AnalysisConfig cfg, cfg1; + AnalysisConfig cfg, cfg1; DataRecord data(FLAGS_infer_data, FLAGS_batch_size); std::vector> input_slots_all; @@ -282,7 +281,7 @@ TEST(Analyzer_dam, compare_with_static_memory_optim) { TEST(Analyzer_dam, compare_with_dynamic_memory_optim) { // The small dam will core in CI, but works in local. if (FLAGS_max_turn_num == 9) { - contrib::AnalysisConfig cfg, cfg1; + AnalysisConfig cfg, cfg1; DataRecord data(FLAGS_infer_data, FLAGS_batch_size); std::vector> input_slots_all; diff --git a/paddle/fluid/inference/tests/api/analyzer_lac_tester.cc b/paddle/fluid/inference/tests/api/analyzer_lac_tester.cc index b9666e01adb23e0cbd9257bc55081c3a5001e887..347672eaae314aa42096d48a3b044014f2ddbf84 100644 --- a/paddle/fluid/inference/tests/api/analyzer_lac_tester.cc +++ b/paddle/fluid/inference/tests/api/analyzer_lac_tester.cc @@ -18,8 +18,6 @@ namespace paddle { namespace inference { namespace analysis { -using contrib::AnalysisConfig; - struct DataRecord { std::vector data; std::vector lod; diff --git a/paddle/fluid/inference/tests/api/analyzer_mm_dnn_tester.cc b/paddle/fluid/inference/tests/api/analyzer_mm_dnn_tester.cc index 529a0174c8542f5226e70ef4a47bde069220ecc2..089f655c180d784af66af60277bdbf32a6019599 100644 --- a/paddle/fluid/inference/tests/api/analyzer_mm_dnn_tester.cc +++ b/paddle/fluid/inference/tests/api/analyzer_mm_dnn_tester.cc @@ -16,7 +16,6 @@ namespace paddle { namespace inference { -using contrib::AnalysisConfig; struct DataRecord { std::vector> query, title; @@ -75,7 +74,7 @@ void PrepareInputs(std::vector *input_slots, DataRecord *data, } } -void SetConfig(contrib::AnalysisConfig *cfg) { +void SetConfig(AnalysisConfig *cfg) { cfg->SetModel(FLAGS_infer_model); cfg->DisableGpu(); cfg->SwitchSpecifyInputNames(); @@ -95,7 +94,7 @@ void SetInput(std::vector> *inputs) { // Easy for profiling independently. void profile(bool use_mkldnn = false) { - contrib::AnalysisConfig cfg; + AnalysisConfig cfg; SetConfig(&cfg); std::vector outputs; @@ -130,7 +129,7 @@ TEST(Analyzer_MM_DNN, profile_mkldnn) { profile(true /* use_mkldnn */); } // Check the fuse status TEST(Analyzer_MM_DNN, fuse_statis) { - contrib::AnalysisConfig cfg; + AnalysisConfig cfg; SetConfig(&cfg); int num_ops; @@ -141,7 +140,7 @@ TEST(Analyzer_MM_DNN, fuse_statis) { // Compare result of NativeConfig and AnalysisConfig void compare(bool use_mkldnn = false) { - contrib::AnalysisConfig cfg; + AnalysisConfig cfg; SetConfig(&cfg); if (use_mkldnn) { diff --git a/paddle/fluid/inference/tests/api/analyzer_ner_tester.cc b/paddle/fluid/inference/tests/api/analyzer_ner_tester.cc index 6fef79dc4608acd6eee679ad4939e7684db98f5b..a70aa7a6ac41121a0c8ea397ebc7e24e4b206d12 100644 --- a/paddle/fluid/inference/tests/api/analyzer_ner_tester.cc +++ b/paddle/fluid/inference/tests/api/analyzer_ner_tester.cc @@ -16,7 +16,6 @@ namespace paddle { namespace inference { -using contrib::AnalysisConfig; struct DataRecord { std::vector> word, mention; @@ -76,7 +75,7 @@ void PrepareInputs(std::vector *input_slots, DataRecord *data) { } } -void SetConfig(contrib::AnalysisConfig *cfg, bool memory_load = false) { +void SetConfig(AnalysisConfig *cfg, bool memory_load = false) { if (memory_load) { std::string buffer_prog, buffer_param; ReadBinaryFile(FLAGS_infer_model + "/__model__", &buffer_prog); @@ -105,7 +104,7 @@ void SetInput(std::vector> *inputs) { // Easy for profiling independently. void profile(bool memory_load = false) { - contrib::AnalysisConfig cfg; + AnalysisConfig cfg; SetConfig(&cfg, memory_load); std::vector outputs; @@ -136,7 +135,7 @@ TEST(Analyzer_Chinese_ner, profile_memory_load) { // Check the fuse status TEST(Analyzer_Chinese_ner, fuse_statis) { - contrib::AnalysisConfig cfg; + AnalysisConfig cfg; SetConfig(&cfg); int num_ops; @@ -152,7 +151,7 @@ TEST(Analyzer_Chinese_ner, fuse_statis) { // Compare result of NativeConfig and AnalysisConfig TEST(Analyzer_Chinese_ner, compare) { - contrib::AnalysisConfig cfg; + AnalysisConfig cfg; SetConfig(&cfg); std::vector> input_slots_all; diff --git a/paddle/fluid/inference/tests/api/analyzer_pyramid_dnn_tester.cc b/paddle/fluid/inference/tests/api/analyzer_pyramid_dnn_tester.cc index ad2c46e48d5a34a457a615f313f1ac3cc916b200..3f6c933f2bcc6ed5410cb95a48f5ee6869280fe4 100644 --- a/paddle/fluid/inference/tests/api/analyzer_pyramid_dnn_tester.cc +++ b/paddle/fluid/inference/tests/api/analyzer_pyramid_dnn_tester.cc @@ -16,7 +16,6 @@ namespace paddle { namespace inference { -using contrib::AnalysisConfig; struct DataRecord { std::vector> query_basic, query_phrase, title_basic, @@ -103,7 +102,7 @@ void PrepareInputs(std::vector *input_slots, DataRecord *data, } } -void SetConfig(contrib::AnalysisConfig *cfg) { +void SetConfig(AnalysisConfig *cfg) { cfg->SetModel(FLAGS_infer_model); cfg->DisableGpu(); cfg->SwitchSpecifyInputNames(); @@ -123,7 +122,7 @@ void SetInput(std::vector> *inputs) { // Easy for profiling independently. TEST(Analyzer_Pyramid_DNN, profile) { - contrib::AnalysisConfig cfg; + AnalysisConfig cfg; SetConfig(&cfg); std::vector outputs; @@ -147,7 +146,7 @@ TEST(Analyzer_Pyramid_DNN, profile) { // Check the fuse status TEST(Analyzer_Pyramid_DNN, fuse_statis) { - contrib::AnalysisConfig cfg; + AnalysisConfig cfg; SetConfig(&cfg); int num_ops; @@ -158,7 +157,7 @@ TEST(Analyzer_Pyramid_DNN, fuse_statis) { // Compare result of NativeConfig and AnalysisConfig TEST(Analyzer_Pyramid_DNN, compare) { - contrib::AnalysisConfig cfg; + AnalysisConfig cfg; SetConfig(&cfg); std::vector> input_slots_all; diff --git a/paddle/fluid/inference/tests/api/analyzer_rnn1_tester.cc b/paddle/fluid/inference/tests/api/analyzer_rnn1_tester.cc index 22e6366fb5cba6c7a0cde9c0c5f50f56c2e23b05..c27c39f40a2067dd2bd2150e4b1e53eab7cdf06e 100644 --- a/paddle/fluid/inference/tests/api/analyzer_rnn1_tester.cc +++ b/paddle/fluid/inference/tests/api/analyzer_rnn1_tester.cc @@ -20,7 +20,6 @@ namespace paddle { namespace inference { using namespace framework; // NOLINT -using namespace contrib; // NOLINT struct DataRecord { std::vector>> link_step_data_all; @@ -223,7 +222,7 @@ void SetInput(std::vector> *inputs) { // Easy for profiling independently. TEST(Analyzer_rnn1, profile) { - contrib::AnalysisConfig cfg; + AnalysisConfig cfg; SetConfig(&cfg); cfg.DisableGpu(); cfg.SwitchIrDebug(); @@ -237,7 +236,7 @@ TEST(Analyzer_rnn1, profile) { // Check the fuse status TEST(Analyzer_rnn1, fuse_statis) { - contrib::AnalysisConfig cfg; + AnalysisConfig cfg; SetConfig(&cfg); int num_ops; @@ -254,7 +253,7 @@ TEST(Analyzer_rnn1, fuse_statis) { // Compare result of NativeConfig and AnalysisConfig TEST(Analyzer_rnn1, compare) { - contrib::AnalysisConfig cfg; + AnalysisConfig cfg; SetConfig(&cfg); std::vector> input_slots_all; @@ -276,7 +275,7 @@ TEST(Analyzer_rnn1, compare_determine) { // Test Multi-Thread. TEST(Analyzer_rnn1, multi_thread) { - contrib::AnalysisConfig cfg; + AnalysisConfig cfg; SetConfig(&cfg); std::vector outputs; diff --git a/paddle/fluid/inference/tests/api/analyzer_seq_pool1_tester.cc b/paddle/fluid/inference/tests/api/analyzer_seq_pool1_tester.cc index 8be2a6d79b2ede2c149aa523e38c3960ab30acb1..bd0059e18485c046df27d5ddbb39df9bbb249113 100644 --- a/paddle/fluid/inference/tests/api/analyzer_seq_pool1_tester.cc +++ b/paddle/fluid/inference/tests/api/analyzer_seq_pool1_tester.cc @@ -56,14 +56,14 @@ struct DataRecord { std::vector slot_data; split_to_float(data[1], ' ', &slot_data); std::string name = data[0]; - PADDLE_ENFORCE_EQ(slot_data.size() % 11, 0, + PADDLE_ENFORCE_EQ(slot_data.size() % 11, 0UL, "line %d, %s should be divisible", num_lines, name); datasets[name].emplace_back(std::move(slot_data)); } num_samples = num_lines / num_slots; PADDLE_ENFORCE_EQ(num_samples * num_slots, static_cast(num_lines), "num samples should be divisible"); - PADDLE_ENFORCE_GT(num_samples, 0); + PADDLE_ENFORCE_GT(num_samples, 0UL); } void Prepare(int bs) { @@ -142,7 +142,7 @@ void SetConfig(AnalysisConfig *cfg, bool use_mkldnn = false) { cfg->SetModel(FLAGS_infer_model + "/model", FLAGS_infer_model + "/params"); cfg->DisableGpu(); cfg->SwitchSpecifyInputNames(); - cfg->pass_builder()->TurnOnDebug(); + cfg->SwitchIrDebug(); cfg->SetCpuMathLibraryNumThreads(FLAGS_paddle_num_threads); if (use_mkldnn) { cfg->EnableMKLDNN(); diff --git a/paddle/fluid/inference/tests/api/analyzer_text_classification_tester.cc b/paddle/fluid/inference/tests/api/analyzer_text_classification_tester.cc index 2db297e2005c6b657259187d6b6b76657d9e4388..2003be82019333ca97b9fa8ef83668825fe5710d 100644 --- a/paddle/fluid/inference/tests/api/analyzer_text_classification_tester.cc +++ b/paddle/fluid/inference/tests/api/analyzer_text_classification_tester.cc @@ -69,7 +69,7 @@ void SetInput(std::vector> *inputs) { TEST(Analyzer_Text_Classification, profile) { AnalysisConfig cfg; SetConfig(&cfg); - cfg.pass_builder()->TurnOnDebug(); + cfg.SwitchIrDebug(); std::vector outputs; std::vector> input_slots_all; diff --git a/paddle/fluid/inference/tests/api/analyzer_vis_tester.cc b/paddle/fluid/inference/tests/api/analyzer_vis_tester.cc index f3e75ffbb5962885bd926af50b764bec561cc454..ca04c1365cbbffcb4a2786cde9ab240cc20aa3d8 100644 --- a/paddle/fluid/inference/tests/api/analyzer_vis_tester.cc +++ b/paddle/fluid/inference/tests/api/analyzer_vis_tester.cc @@ -20,7 +20,6 @@ limitations under the License. */ namespace paddle { namespace inference { namespace analysis { -using contrib::AnalysisConfig; struct Record { std::vector data; diff --git a/paddle/fluid/inference/tests/api/config_printer.h b/paddle/fluid/inference/tests/api/config_printer.h index ecc10bafd650e52dfb73e8dd4329c697ff4f4ccc..b0c23fbd534847c8aad244749761e9c072148796 100644 --- a/paddle/fluid/inference/tests/api/config_printer.h +++ b/paddle/fluid/inference/tests/api/config_printer.h @@ -58,9 +58,8 @@ std::ostream &operator<<(std::ostream &os, const NativeConfig &config) { return os; } -std::ostream &operator<<(std::ostream &os, - const contrib::AnalysisConfig &config) { - os << GenSpaces(num_spaces) << "contrib::AnalysisConfig {\n"; +std::ostream &operator<<(std::ostream &os, const AnalysisConfig &config) { + os << GenSpaces(num_spaces) << "AnalysisConfig {\n"; num_spaces++; os << config.ToNativeConfig(); if (!config.model_from_memory()) { diff --git a/paddle/fluid/inference/tests/api/tester_helper.h b/paddle/fluid/inference/tests/api/tester_helper.h index b1f7a3464ac6027faffe283bccaf9793eae939e1..2811eb4946ea025cf6c7ab197c4e603df86f6f2d 100644 --- a/paddle/fluid/inference/tests/api/tester_helper.h +++ b/paddle/fluid/inference/tests/api/tester_helper.h @@ -56,16 +56,9 @@ DECLARE_int32(paddle_num_threads); namespace paddle { namespace inference { -float Random(float low, float high) { - static std::random_device rd; - static std::mt19937 mt(rd()); - std::uniform_real_distribution dist(low, high); - return dist(mt); -} - void PrintConfig(const PaddlePredictor::Config *config, bool use_analysis) { const auto *analysis_config = - reinterpret_cast(config); + reinterpret_cast(config); if (use_analysis) { LOG(INFO) << *analysis_config; return; @@ -109,9 +102,9 @@ void CompareResult(const std::vector &outputs, std::unique_ptr CreateTestPredictor( const PaddlePredictor::Config *config, bool use_analysis = true) { const auto *analysis_config = - reinterpret_cast(config); + reinterpret_cast(config); if (use_analysis) { - return CreatePaddlePredictor(*analysis_config); + return CreatePaddlePredictor(*analysis_config); } auto native_config = analysis_config->ToNativeConfig(); return CreatePaddlePredictor(native_config); @@ -146,7 +139,8 @@ void SetFakeImageInput(std::vector> *inputs, const std::string &dirname, bool is_combined = true, std::string model_filename = "model", std::string params_filename = "params", - const std::vector *feed_names = nullptr) { + const std::vector *feed_names = nullptr, + const int continuous_inuput_index = 0) { // Set fake_image_data PADDLE_ENFORCE_EQ(FLAGS_test_all_data, 0, "Only have single batch of data."); std::vector> feed_target_shapes = GetFeedTargetShapes( @@ -183,7 +177,8 @@ void SetFakeImageInput(std::vector> *inputs, float *input_data = static_cast(input.data.data()); // fill input data, for profile easily, do not use random data here. for (size_t j = 0; j < len; ++j) { - *(input_data + j) = Random(0.0, 1.0) / 10.; + *(input_data + j) = + static_cast((j + continuous_inuput_index) % len) / len; } } (*inputs).emplace_back(input_slots); diff --git a/paddle/fluid/inference/tests/api/trt_models_tester.cc b/paddle/fluid/inference/tests/api/trt_models_tester.cc index db7109b7505d4fe4dcfcf88f303aa262bc5b44fb..17a433c9d98768dbda4ba93bdceb6cc1717adc07 100644 --- a/paddle/fluid/inference/tests/api/trt_models_tester.cc +++ b/paddle/fluid/inference/tests/api/trt_models_tester.cc @@ -42,9 +42,9 @@ void SetConfig(ConfigType* config, std::string model_dir, bool use_gpu, } template <> -void SetConfig(contrib::AnalysisConfig* config, - std::string model_dir, bool use_gpu, - bool use_tensorrt, int batch_size) { +void SetConfig(AnalysisConfig* config, std::string model_dir, + bool use_gpu, bool use_tensorrt, + int batch_size) { if (!FLAGS_prog_filename.empty() && !FLAGS_param_filename.empty()) { config->SetModel(model_dir + "/" + FLAGS_prog_filename, model_dir + "/" + FLAGS_param_filename); @@ -75,11 +75,11 @@ void profile(std::string model_dir, bool use_analysis, bool use_tensorrt) { std::vector outputs; if (use_analysis || use_tensorrt) { - contrib::AnalysisConfig config; + AnalysisConfig config; config.EnableUseGpu(100, 0); config.pass_builder()->TurnOnDebug(); - SetConfig(&config, model_dir, true, use_tensorrt, - FLAGS_batch_size); + SetConfig(&config, model_dir, true, use_tensorrt, + FLAGS_batch_size); TestPrediction(reinterpret_cast(&config), inputs_all, &outputs, FLAGS_num_threads, true); } else { @@ -99,18 +99,18 @@ void compare(std::string model_dir, bool use_tensorrt) { SetFakeImageInput(&inputs_all, model_dir, false, "__model__", ""); } - contrib::AnalysisConfig analysis_config; - SetConfig(&analysis_config, model_dir, true, - use_tensorrt, FLAGS_batch_size); + AnalysisConfig analysis_config; + SetConfig(&analysis_config, model_dir, true, use_tensorrt, + FLAGS_batch_size); CompareNativeAndAnalysis( reinterpret_cast(&analysis_config), inputs_all); } void compare_continuous_input(std::string model_dir, bool use_tensorrt) { - contrib::AnalysisConfig analysis_config; - SetConfig(&analysis_config, model_dir, true, - use_tensorrt, FLAGS_batch_size); + AnalysisConfig analysis_config; + SetConfig(&analysis_config, model_dir, true, use_tensorrt, + FLAGS_batch_size); auto config = reinterpret_cast(&analysis_config); auto native_pred = CreateTestPredictor(config, false); @@ -119,9 +119,10 @@ void compare_continuous_input(std::string model_dir, bool use_tensorrt) { std::vector> inputs_all; if (!FLAGS_prog_filename.empty() && !FLAGS_param_filename.empty()) { SetFakeImageInput(&inputs_all, model_dir, true, FLAGS_prog_filename, - FLAGS_param_filename); + FLAGS_param_filename, nullptr, i); } else { - SetFakeImageInput(&inputs_all, model_dir, false, "__model__", ""); + SetFakeImageInput(&inputs_all, model_dir, false, "__model__", "", nullptr, + i); } CompareNativeAndAnalysis(native_pred.get(), analysis_pred.get(), inputs_all); diff --git a/paddle/fluid/inference/tests/test.cmake b/paddle/fluid/inference/tests/test.cmake index 29f0f034a2aab50330d4d0127b870a5cb00d56a5..6c5fe043ffa3f3dcafe2dbbebd6244467f859abf 100644 --- a/paddle/fluid/inference/tests/test.cmake +++ b/paddle/fluid/inference/tests/test.cmake @@ -1,18 +1,43 @@ +include(ExternalProject) set(INFERENCE_URL "http://paddle-inference-dist.cdn.bcebos.com" CACHE STRING "inference download url") set(INFERENCE_DEMO_INSTALL_DIR "${THIRD_PARTY_PATH}/inference_demo" CACHE STRING "A path setting inference demo download directories.") -function (inference_download install_dir url filename) - message(STATUS "Download inference test stuff from ${url}/${filename}") - file(DOWNLOAD "${url}/${filename}" "${install_dir}/${filename}") - message(STATUS "finish downloading ${filename}") + +function(inference_download INSTALL_DIR URL FILENAME) + message(STATUS "Download inference test stuff from ${URL}/${FILENAME}") + string(REGEX REPLACE "[-%.]" "_" FILENAME_EX ${FILENAME}) + ExternalProject_Add( + extern_inference_download_${FILENAME_EX} + ${EXTERNAL_PROJECT_LOG_ARGS} + PREFIX ${INSTALL_DIR} + URL ${URL}/${FILENAME} + DOWNLOAD_COMMAND wget -q -O ${INSTALL_DIR}/${FILENAME} ${URL}/${FILENAME} + DOWNLOAD_DIR ${INSTALL_DIR} + DOWNLOAD_NO_PROGRESS 1 + CONFIGURE_COMMAND "" + BUILD_COMMAND "" + UPDATE_COMMAND "" + INSTALL_COMMAND "" + ) endfunction() -function (inference_download_and_uncompress install_dir url filename) - inference_download(${install_dir} ${url} ${filename}) - execute_process( - COMMAND ${CMAKE_COMMAND} -E tar xzf ${install_dir}/${filename} - WORKING_DIRECTORY ${install_dir} - ) +function(inference_download_and_uncompress INSTALL_DIR URL FILENAME) + message(STATUS "Download inference test stuff from ${URL}/${FILENAME}") + string(REGEX REPLACE "[-%.]" "_" FILENAME_EX ${FILENAME}) + set(EXTERNAL_PROJECT_NAME "extern_inference_download_${FILENAME_EX}") + set(UNPACK_DIR "${INSTALL_DIR}/src/${EXTERNAL_PROJECT_NAME}") + ExternalProject_Add( + ${EXTERNAL_PROJECT_NAME} + ${EXTERNAL_PROJECT_LOG_ARGS} + PREFIX ${INSTALL_DIR} + URL ${URL}/${FILENAME} + DOWNLOAD_DIR ${INSTALL_DIR} + DOWNLOAD_NO_PROGRESS 1 + CONFIGURE_COMMAND "" + BUILD_COMMAND "" + UPDATE_COMMAND "" + INSTALL_COMMAND ${CMAKE_COMMAND} -E copy_directory ${UNPACK_DIR} ${INSTALL_DIR} + ) endfunction() set(WORD2VEC_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/word2vec") diff --git a/paddle/fluid/inference/tests/test_helper.h b/paddle/fluid/inference/tests/test_helper.h index 75fa611c0d701dd81dfe5b33231655e0959c7dbf..861f69f4d2143b16bdec546d92ce7bd13ca53ed3 100644 --- a/paddle/fluid/inference/tests/test_helper.h +++ b/paddle/fluid/inference/tests/test_helper.h @@ -171,9 +171,7 @@ void TestInference(const std::string& dirname, // Enable the profiler paddle::platform::EnableProfiler(state); { - paddle::platform::RecordEvent record_event( - "init_program", - paddle::platform::DeviceContextPool::Instance().Get(place)); + paddle::platform::RecordEvent record_event("init_program"); inference_program = InitProgram(&executor, scope, dirname, is_combined); } @@ -230,9 +228,7 @@ void TestInference(const std::string& dirname, // Run repeat times to profile the performance for (int i = 0; i < repeat; ++i) { - paddle::platform::RecordEvent record_event( - "run_inference", - paddle::platform::DeviceContextPool::Instance().Get(place)); + paddle::platform::RecordEvent record_event("run_inference"); if (PrepareContext) { // Note: if you change the inference_program, you need to call diff --git a/paddle/fluid/inference/utils/benchmark_tester.cc b/paddle/fluid/inference/utils/benchmark_tester.cc index 80763160df3adfd8c34e66bc7a5370808b349e76..0c48c2db9b691ae8cf587f2729c2789d4ce2dbe1 100644 --- a/paddle/fluid/inference/utils/benchmark_tester.cc +++ b/paddle/fluid/inference/utils/benchmark_tester.cc @@ -34,6 +34,6 @@ TEST(Benchmark, PersistToFile) { benchmark.SetLatency(220); benchmark.PersistToFile("1.log"); - benchmark.PersistToFile("1.log"); - benchmark.PersistToFile("1.log"); + benchmark.PersistToFile("2.log"); + benchmark.PersistToFile("3.log"); } diff --git a/paddle/fluid/memory/allocation/allocator_facade.cc b/paddle/fluid/memory/allocation/allocator_facade.cc index 794d729bdc1adc7eb3fe44ffabfe0cc99719b421..ea0b729dc6f62f517877e060cb0ecbe5c1d22e61 100644 --- a/paddle/fluid/memory/allocation/allocator_facade.cc +++ b/paddle/fluid/memory/allocation/allocator_facade.cc @@ -83,7 +83,7 @@ class ChunkedAllocator : public Allocator { VLOG(1) << "Create AutoIncrementAllocator with chunk_size " << max_chunk_size_ << " and capacity " << capacity; default_allocator_ = std::make_shared( - [this] { return std::move(CreateAllocatorWithChunk()); }, capacity); + [this] { return CreateAllocatorWithChunk(); }, capacity); } } diff --git a/paddle/fluid/memory/allocation/best_fit_allocator.cc b/paddle/fluid/memory/allocation/best_fit_allocator.cc index 6f3e512fb0b68df5e86eba3e50a255c18f75214f..e3d6c2f511ef083ef9ecc1fe8df96051b2b85cc2 100644 --- a/paddle/fluid/memory/allocation/best_fit_allocator.cc +++ b/paddle/fluid/memory/allocation/best_fit_allocator.cc @@ -111,6 +111,8 @@ size_t BestFitAllocator::NumFreeChunks() const { } void BestFitAllocator::Free(Allocation* allocation) { auto* bf_allocation = dynamic_cast(allocation); + PADDLE_ENFORCE_NOT_NULL(bf_allocation, + "The input allocation is not BestFitAllocation."); auto chunk_it = bf_allocation->ChunkIterator(); PADDLE_ENFORCE(!chunk_it->is_free); chunk_it->is_free = true; diff --git a/paddle/fluid/memory/allocation/legacy_allocator.cc b/paddle/fluid/memory/allocation/legacy_allocator.cc index 5d8684f083bda8499000c9fd0a7617cf129db13b..1936f9d4cd83c53cf7b322ab29a3e0d92e042abc 100644 --- a/paddle/fluid/memory/allocation/legacy_allocator.cc +++ b/paddle/fluid/memory/allocation/legacy_allocator.cc @@ -13,9 +13,15 @@ // limitations under the License. #include "paddle/fluid/memory/allocation/legacy_allocator.h" + #include #include #include + +#ifdef PADDLE_WITH_JEMALLOC +#include +#endif + #include "glog/logging.h" #include "paddle/fluid/memory/detail/buddy_allocator.h" #include "paddle/fluid/memory/detail/system_allocator.h" @@ -30,6 +36,7 @@ DEFINE_bool(init_allocated_mem, false, "that initializing the allocated memory with a small value " "during unit testing."); DECLARE_double(fraction_of_gpu_memory_to_use); +DECLARE_bool(benchmark); namespace paddle { namespace memory { @@ -53,11 +60,6 @@ size_t memory_usage(const platform::Place &p); using BuddyAllocator = detail::BuddyAllocator; -std::unordered_map> - gpu_mem_info; - BuddyAllocator *GetCPUBuddyAllocator() { // We tried thread_local for inference::RNN1 model, but that not works much // for multi-thread test. @@ -95,7 +97,11 @@ struct NaiveAllocator { template <> void *Alloc(const platform::CPUPlace &place, size_t size) { VLOG(10) << "Allocate " << size << " bytes on " << platform::Place(place); +#ifdef PADDLE_WITH_JEMALLOC + void *p = malloc(size); +#else void *p = GetCPUBuddyAllocator()->Alloc(size); +#endif if (FLAGS_init_allocated_mem) { memset(p, 0xEF, size); } @@ -107,12 +113,21 @@ template <> void Free(const platform::CPUPlace &place, void *p, size_t size) { VLOG(10) << "Free pointer=" << p << " on " << platform::Place(place); +#ifdef PADDLE_WITH_JEMALLOC + free(p); +#else GetCPUBuddyAllocator()->Free(p); +#endif } template <> size_t Used(const platform::CPUPlace &place) { +#ifdef PADDLE_WITH_JEMALLOC + // fake the result of used memory when PADDLE_WITH_JEMALLOC is ON + return 0U; +#else return GetCPUBuddyAllocator()->Used(); +#endif } #ifdef PADDLE_WITH_CUDA @@ -125,6 +140,8 @@ BuddyAllocator *GetGPUBuddyAllocator(int gpu_id) { devices = platform::GetSelectedDevices(); int gpu_num = devices.size(); + allocation::GPUMemMonitor.Initialize(devices.size()); + a_arr = new BuddyAllocator *[gpu_num]; for (size_t i = 0; i < devices.size(); ++i) { int dev_id = devices[i]; @@ -171,25 +188,19 @@ void *Alloc(const platform::CUDAPlace &place, platform::SetDeviceId(place.device); size_t avail, total; platform::GpuMemoryUsage(&avail, &total); - LOG(WARNING) << "Cannot allocate " << string::HumanReadableSize(size) - << " in GPU " << place.device << ", available " - << string::HumanReadableSize(avail); - LOG(WARNING) << "total " << total; - LOG(WARNING) << "GpuMinChunkSize " - << string::HumanReadableSize( - buddy_allocator->GetMinChunkSize()); - LOG(WARNING) << "GpuMaxChunkSize " - << string::HumanReadableSize( - buddy_allocator->GetMaxChunkSize()); - LOG(WARNING) << "GPU memory used: " - << string::HumanReadableSize(Used(place)); + LOG(FATAL) << "Cannot allocate " << string::HumanReadableSize(size) + << " in GPU " << place.device << ", available " + << string::HumanReadableSize(avail) << "total " << total + << "GpuMinChunkSize " + << string::HumanReadableSize(buddy_allocator->GetMinChunkSize()) + << "GpuMaxChunkSize " + << string::HumanReadableSize(buddy_allocator->GetMaxChunkSize()) + << "GPU memory used: " + << string::HumanReadableSize(Used(place)); platform::SetDeviceId(cur_dev); } else { - gpu_mem_info[place.device].first += size; - if (gpu_mem_info[place.device].first > gpu_mem_info[place.device].second) { - gpu_mem_info[place.device].second = gpu_mem_info[place.device].first; - VLOG(3) << "device: " << place.device << " peak memory usage : " - << (gpu_mem_info[place.device].second >> 20) << " MiB"; + if (FLAGS_benchmark) { + allocation::GPUMemMonitor.Add(place.device, size); } if (FLAGS_init_allocated_mem) { cudaMemset(ptr, 0xEF, size); @@ -206,7 +217,9 @@ void Free(const platform::CUDAPlace &place, void *p, size_t size) { #ifdef PADDLE_WITH_CUDA GetGPUBuddyAllocator(place.device)->Free(p); - gpu_mem_info[place.device].first -= size; + if (FLAGS_benchmark) { + allocation::GPUMemMonitor.Minus(place.device, size); + } #else PADDLE_THROW("'CUDAPlace' is not supported in CPU only device."); #endif @@ -245,7 +258,7 @@ void *Alloc(const platform::CUDAPinnedPlace &place, void *ptr = buddy_allocator->Alloc(size); if (ptr == nullptr) { - LOG(WARNING) << "cudaMallocHost Cannot allocate " << size + LOG(WARNING) << "cudaHostAlloc Cannot allocate " << size << " bytes in CUDAPinnedPlace"; } if (FLAGS_init_allocated_mem) { @@ -316,6 +329,8 @@ size_t Usage::operator()(const platform::CUDAPinnedPlace &cuda_pinned) const { namespace allocation { +LegacyMemMonitor GPUMemMonitor; + Allocation *LegacyAllocator::AllocateImpl(size_t size, Allocator::Attr attr) { void *ptr = boost::apply_visitor(legacy::AllocVisitor(size), place_); return new Allocation(ptr, size, place_); @@ -327,6 +342,63 @@ void LegacyAllocator::Free(Allocation *allocation) { allocation->place()); delete allocation; } + +bool MemInfo::Add(const size_t &size) { + std::lock_guard lock(mutex_); + usage_ += size; + bool peak_point = usage_ > peak_usage_; + if (peak_point) peak_usage_ = usage_; + return peak_point; +} + +void MemInfo::Minus(const size_t &size) { + std::lock_guard lock(mutex_); + usage_ -= size; +} + +uint64_t MemInfo::GetPeakUsage() const { return peak_usage_; } + +LegacyMemMonitor::~LegacyMemMonitor() { + for (auto &item : gpu_mem_info_) delete item.second; +} + +void LegacyMemMonitor::Initialize(const int &device_num) { + for (auto i = 0; i < device_num; ++i) { + gpu_mem_info_[i] = new MemInfo(); + } +} + +void LegacyMemMonitor::Add(const int &device, const size_t &size) { + if (gpu_mem_info_[device]->Add(size)) { + VLOG(3) << "#LegacyMemMonitor# device: " << device + << " peak memory usage : " + << (gpu_mem_info_[device]->GetPeakUsage() >> 20) << " MiB"; + } +} + +void LegacyMemMonitor::Minus(const int &device, const size_t &size) { + gpu_mem_info_[device]->Minus(size); +} + +uint64_t LegacyMemMonitor::GetMemUsage(const int &device) const { + return gpu_mem_info_.find(device) == gpu_mem_info_.end() + ? 0 + : gpu_mem_info_.at(device)->GetPeakUsage(); +} + +void LegacyMemMonitor::PrintMemUsage() { + std::vector devices; + for (const auto &item : gpu_mem_info_) { + devices.emplace_back(item.first); + } + std::sort(devices.begin(), devices.end()); + for (const auto &device : devices) { + std::cout << "Device : " << device << " Peak Memory Usage : " + << (gpu_mem_info_[device]->GetPeakUsage() >> 20) << " MiB" + << std::endl; + } +} + } // namespace allocation } // namespace memory } // namespace paddle diff --git a/paddle/fluid/memory/allocation/legacy_allocator.h b/paddle/fluid/memory/allocation/legacy_allocator.h index 503a7a685cb9d8dbbbbd6c23b5b82c383893e3d8..d9bdae153da6439598f76f5cac226897e6e0c596 100644 --- a/paddle/fluid/memory/allocation/legacy_allocator.h +++ b/paddle/fluid/memory/allocation/legacy_allocator.h @@ -13,12 +13,59 @@ // limitations under the License. #pragma once +#include +#include // NOLINT +#include +#include +#include #include "paddle/fluid/memory/allocation/allocator.h" #include "paddle/fluid/platform/place.h" namespace paddle { namespace memory { namespace allocation { +class MemInfo { + public: + MemInfo() : usage_(0), peak_usage_(0) {} + + // return a flag to indicate current operation will create a peak point or not + bool Add(const size_t &); + void Minus(const size_t &); + + uint64_t GetPeakUsage() const; + + private: + /* current memory usage*/ + uint64_t usage_; + uint64_t peak_usage_; + std::mutex mutex_; + + DISABLE_COPY_AND_ASSIGN(MemInfo); +}; + +class LegacyMemMonitor { + public: + // used to store the GPU memory usage of each devices + using MemUsage = std::unordered_map; + + MemUsage GetMemUsageInfo() { return gpu_mem_info_; } + ~LegacyMemMonitor(); + + void Initialize(const int &); + void Add(const int &, const size_t &); + void Minus(const int &, const size_t &); + + uint64_t GetMemUsage(const int &) const; + + void PrintMemUsage(); + + private: + MemUsage gpu_mem_info_; +}; + +extern LegacyMemMonitor GPUMemMonitor; + class LegacyAllocatorPrivate; class LegacyAllocator : public Allocator { public: diff --git a/paddle/fluid/memory/allocation/pinned_allocator.cc b/paddle/fluid/memory/allocation/pinned_allocator.cc index 6ac3aefdd18d6d9a21dc7ce66511013dfb78bc5b..de81d12cca6ca280289371abdec225c9e2b8f4d0 100644 --- a/paddle/fluid/memory/allocation/pinned_allocator.cc +++ b/paddle/fluid/memory/allocation/pinned_allocator.cc @@ -32,7 +32,7 @@ Allocation *CPUPinnedAllocator::AllocateImpl(size_t size, // "CPUPinnedAllocator should be used for Cross-Device Communication"); void *ptr; - PADDLE_ENFORCE(cudaMallocHost(&ptr, size)); + PADDLE_ENFORCE(cudaHostAlloc(&ptr, size, cudaHostAllocPortable)); return new CPUPinnedAllocation(ptr, size); } } // namespace allocation diff --git a/paddle/fluid/memory/allocation/pinned_allocator.h b/paddle/fluid/memory/allocation/pinned_allocator.h index 26d12dd91c7fda31802226a84d883b6a6e9abbe4..42d0938f2afbb1efca8bfdd7035bc0eada30f06b 100644 --- a/paddle/fluid/memory/allocation/pinned_allocator.h +++ b/paddle/fluid/memory/allocation/pinned_allocator.h @@ -19,7 +19,7 @@ namespace paddle { namespace memory { namespace allocation { -// Allocator uses `cudaMallocHost` +// Allocator uses `cudaHostAlloc` class CPUPinnedAllocation : public Allocation { public: CPUPinnedAllocation(void *ptr, size_t size) diff --git a/paddle/fluid/memory/detail/system_allocator.cc b/paddle/fluid/memory/detail/system_allocator.cc index 3e8fb83e9d5ba2078bcf37e4a4af74708df9c11c..197d1c2f21fd818879aafe17599bc87d33caa198 100644 --- a/paddle/fluid/memory/detail/system_allocator.cc +++ b/paddle/fluid/memory/detail/system_allocator.cc @@ -173,14 +173,14 @@ void* CUDAPinnedAllocator::Alloc(size_t* index, size_t size) { void* p; // PINNED memory is visible to all CUDA contexts. - cudaError_t result = cudaMallocHost(&p, size); + cudaError_t result = cudaHostAlloc(&p, size, cudaHostAllocPortable); if (result == cudaSuccess) { *index = 1; // PINNED memory cuda_pinnd_alloc_size_ += size; return p; } else { - LOG(WARNING) << "cudaMallocHost failed."; + LOG(WARNING) << "cudaHostAlloc failed."; return nullptr; } diff --git a/paddle/fluid/operators/CMakeLists.txt b/paddle/fluid/operators/CMakeLists.txt index e099425b94221bf1229e936fc1781615d13dbc26..a3f2a69aef52b6f55aa09e6dee2c22c048626c0d 100644 --- a/paddle/fluid/operators/CMakeLists.txt +++ b/paddle/fluid/operators/CMakeLists.txt @@ -66,7 +66,7 @@ set(COMMON_OP_DEPS ${OP_HEADER_DEPS}) set(COMMON_OP_DEPS ${COMMON_OP_DEPS} selected_rows_functor selected_rows lod_tensor maxouting unpooling pooling lod_rank_table context_project sequence_pooling executor) set(COMMON_OP_DEPS ${COMMON_OP_DEPS} dynload_warpctc) -set(COMMON_OP_DEPS ${COMMON_OP_DEPS} sequence_padding sequence_scale cos_sim_functor memory jit_kernel_helper concat_and_split cross_entropy softmax vol2col im2col sampler tree2col) +set(COMMON_OP_DEPS ${COMMON_OP_DEPS} sequence_padding sequence_scale cos_sim_functor memory jit_kernel_helper concat_and_split cross_entropy softmax vol2col im2col sampler sample_prob tree2col) set(COMMON_OP_DEPS ${COMMON_OP_DEPS} sequence2batch lstm_compute matrix_bit_code gru_compute activation_functions beam_search) if (WITH_GPU) set(COMMON_OP_DEPS ${COMMON_OP_DEPS} depthwise_conv prelu) @@ -97,3 +97,4 @@ if (WITH_PYTHON) endif() set(GLOB_OP_LIB ${OP_LIBRARY} CACHE INTERNAL "Global OP library") +add_subdirectory(benchmark) diff --git a/paddle/fluid/operators/activation_op.cc b/paddle/fluid/operators/activation_op.cc index 9c5b8604f40ae56c463b54c71623feb61bd8d297..65efe2966ce12e86ba7f4944eb57ae72cdf9796f 100644 --- a/paddle/fluid/operators/activation_op.cc +++ b/paddle/fluid/operators/activation_op.cc @@ -14,7 +14,7 @@ limitations under the License. */ #include "paddle/fluid/operators/activation_op.h" #include -#include "paddle/fluid/operators/mkldnn_activation_op.h" +#include "paddle/fluid/operators/mkldnn/mkldnn_activation_op.h" #include "paddle/fluid/platform/port.h" namespace paddle { @@ -37,7 +37,7 @@ using paddle::framework::Tensor; "(bool, default false) Set to true for inference only, false " \ "for training. Some layers may run faster when this is true.") \ .SetDefault(false); \ - AddComment(#OP_COMMENT); \ + AddComment(OP_COMMENT); \ } \ } @@ -124,7 +124,7 @@ class ActivationOpGrad : public framework::OperatorWithKernel { UNUSED constexpr char SigmoidDoc[] = R"DOC( Sigmoid Activation Operator -$$out = \frac{1}{1 + e^{-x}}$$ +$$out = \\frac{1}{1 + e^{-x}}$$ )DOC"; @@ -187,14 +187,14 @@ $out = |x|$ UNUSED constexpr char CeilDoc[] = R"DOC( Ceil Activation Operator. -$out = ceil(x)$ +$out = \left \lceil x \right \rceil$ )DOC"; UNUSED constexpr char FloorDoc[] = R"DOC( Floor Activation Operator. -$out = floor(x)$ +$out = \left \lfloor x \right \rfloor$ )DOC"; @@ -252,7 +252,7 @@ $out = \ln(1 + e^{x})$ UNUSED constexpr char SoftsignDoc[] = R"DOC( Softsign Activation Operator. -$$out = \frac{x}{1 + |x|}$$ +$$out = \\frac{x}{1 + \|x\|}$$ )DOC"; @@ -547,12 +547,14 @@ namespace ops = paddle::operators; __macro(Swish, swish); \ __macro(ThresholdedRelu, thresholded_relu); -#define REGISTER_INPLACE_ACTIVATION_OP(OP_NAME, KERNEL_TYPE) \ - REGISTER_OPERATOR(KERNEL_TYPE, ::paddle::operators::ActivationOp, \ - ::paddle::operators::OP_NAME##OpMaker, \ - ::paddle::operators::ActivationOpInferVarType, \ - ::paddle::operators::OP_NAME##GradMaker); \ - REGISTER_OPERATOR(KERNEL_TYPE##_grad, ::paddle::operators::ActivationOpGrad) +#define REGISTER_INPLACE_ACTIVATION_OP(OP_NAME, KERNEL_TYPE) \ + REGISTER_OPERATOR(KERNEL_TYPE, ::paddle::operators::ActivationOp, \ + ::paddle::operators::OP_NAME##OpMaker, \ + ::paddle::operators::ActivationOpInferVarType, \ + ::paddle::operators::OP_NAME##GradMaker, \ + ::paddle::framework::SingleOpInplaceInToOut); \ + REGISTER_OPERATOR(KERNEL_TYPE##_grad, ::paddle::operators::ActivationOpGrad, \ + ::paddle::framework::SingleOpInplaceInToOut) #define REGISTER_ACTIVATION_OP(OP_NAME, KERNEL_TYPE) \ REGISTER_OPERATOR(KERNEL_TYPE, ::paddle::operators::ActivationOp, \ diff --git a/paddle/fluid/operators/attention_lstm_op.cc b/paddle/fluid/operators/attention_lstm_op.cc index b6996be4b0984bcee3b16da268d79708a68b65b3..912ec79910301b67bc520b1aa78d3fa1fd165d1f 100644 --- a/paddle/fluid/operators/attention_lstm_op.cc +++ b/paddle/fluid/operators/attention_lstm_op.cc @@ -293,7 +293,7 @@ class AttentionLSTMKernel : public framework::OpKernel { int len = x_lod[0][i + 1] - x_lod[0][i]; max_seq_len = max_seq_len < len ? len : max_seq_len; } - PADDLE_ENFORCE_EQ(x_lod.size(), 1, "Input(X)'s lod size must be 1."); + PADDLE_ENFORCE_EQ(x_lod.size(), 1UL, "Input(X)'s lod size must be 1."); PADDLE_ENFORCE_EQ(c0->dims()[0], N, "C0 dims should be %d x %d.", N, D); fc_out->Resize({max_seq_len, 1}); diff --git a/paddle/fluid/operators/batch_norm_op.cc b/paddle/fluid/operators/batch_norm_op.cc index 8b672e09b2c5c203c1a1447fbbd14a45ef7ba257..feac4125381bd897dac89943af44850012e4761d 100644 --- a/paddle/fluid/operators/batch_norm_op.cc +++ b/paddle/fluid/operators/batch_norm_op.cc @@ -589,8 +589,10 @@ class BatchNormGradMaker : public framework::SingleGradOpDescMaker { op->SetInput("SavedVariance", Output("SavedVariance")); // used when setting use_global_stats True during training - op->SetInput("Mean", Output("MeanOut")); - op->SetInput("Variance", Output("VarianceOut")); + if (boost::get(GetAttr("use_global_stats"))) { + op->SetInput("Mean", Output("MeanOut")); + op->SetInput("Variance", Output("VarianceOut")); + } op->SetAttrMap(Attrs()); @@ -602,13 +604,48 @@ class BatchNormGradMaker : public framework::SingleGradOpDescMaker { } }; +class BatchNormInplaceInToOut : public framework::InplaceInToOut { + public: + using InplaceInToOut::InplaceInToOut; + + protected: + std::unordered_map Apply( + const framework::OpDesc &op_desc, + framework::BlockDesc *block) const override { + std::unordered_map inplace_in_to_out = { + {"Mean", "MeanOut"}, {"Variance", "VarianceOut"}, {"X", "Y"}, + }; + return inplace_in_to_out; + } +}; + +class BatchNormGradInplaceInToOut : public framework::InplaceInToOut { + public: + using InplaceInToOut::InplaceInToOut; + + protected: + std::unordered_map Apply( + const framework::OpDesc &op_desc, + framework::BlockDesc *block) const override { + std::unordered_map inplace_in_to_out = { + // Scale, Bias, SavedMean, SavedVariance shape is [batch_size, C] + {framework::GradVarName("Y"), framework::GradVarName("X")}, + {"SavedMean", framework::GradVarName("Scale")}, + {"SavedVariance", framework::GradVarName("Bias")}, + }; + return inplace_in_to_out; + } +}; + } // namespace operators } // namespace paddle namespace ops = paddle::operators; REGISTER_OPERATOR(batch_norm, ops::BatchNormOp, ops::BatchNormOpMaker, - ops::BatchNormOpInferVarType, ops::BatchNormGradMaker); -REGISTER_OPERATOR(batch_norm_grad, ops::BatchNormGradOp); + ops::BatchNormOpInferVarType, ops::BatchNormGradMaker, + ops::BatchNormInplaceInToOut); +REGISTER_OPERATOR(batch_norm_grad, ops::BatchNormGradOp, + ops::BatchNormGradInplaceInToOut); REGISTER_OP_CPU_KERNEL( batch_norm, ops::BatchNormKernel, diff --git a/paddle/fluid/operators/beam_search_decode_op.h b/paddle/fluid/operators/beam_search_decode_op.h index 6aefc5446f167eebb0da673b3fbdf7ed128daa98..0b883c3158fb922caae2e731875bbb8d43a1e9ca 100644 --- a/paddle/fluid/operators/beam_search_decode_op.h +++ b/paddle/fluid/operators/beam_search_decode_op.h @@ -122,7 +122,7 @@ void BeamSearchDecoder::ConvertSentenceVectorToLodTensor( auto cpu_place = std::unique_ptr( new paddle::platform::CPUPlace()); - paddle::platform::CPUDeviceContext cpu_ctx(*cpu_place.get()); + paddle::platform::CPUDeviceContext cpu_ctx(*cpu_place); framework::LoD lod; lod.push_back(source_level_lod); diff --git a/paddle/fluid/operators/beam_search_op.cc b/paddle/fluid/operators/beam_search_op.cc index e78ecc1a12309fe084a4165e5bb0d8bfb1dcf957..e93cd8615e052e4dfc6255549bf7a9b84b7dd657 100644 --- a/paddle/fluid/operators/beam_search_op.cc +++ b/paddle/fluid/operators/beam_search_op.cc @@ -51,6 +51,9 @@ class BeamSearchOpMaker : public framework::OpProtoAndCheckerMaker { AddOutput("selected_scores", "A LoDTensor containing the accumulated scores corresponding to " "Output(selected_ids)."); + AddOutput( + "parent_idx", + "A Tensor preserving the selected_ids' parent indice in pre_ids."); // Attributes stored in AttributeMap AddAttr("level", "the level of LoDTensor"); diff --git a/paddle/fluid/operators/beam_search_op.h b/paddle/fluid/operators/beam_search_op.h index 1b939e742de06aedf187d25d002d19e0a4fafc9d..f808020cc765585d1633c6c3bf528080a7e83f07 100644 --- a/paddle/fluid/operators/beam_search_op.h +++ b/paddle/fluid/operators/beam_search_op.h @@ -41,13 +41,15 @@ class BeamSearchOpKernel : public framework::OpKernel { auto selected_ids = context.Output("selected_ids"); auto selected_scores = context.Output("selected_scores"); + auto* parent_idx = context.Output("parent_idx"); PADDLE_ENFORCE_NOT_NULL(selected_ids); PADDLE_ENFORCE_NOT_NULL(selected_scores); + PADDLE_ENFORCE_NOT_NULL(parent_idx); math::BeamSearchFunctor alg; alg(context.template device_context(), pre_ids, pre_scores, - ids, scores, selected_ids, selected_scores, level, beam_size, end_id, - is_accumulated); + ids, scores, selected_ids, selected_scores, parent_idx, level, + beam_size, end_id, is_accumulated); } }; diff --git a/paddle/fluid/operators/benchmark/CMakeLists.txt b/paddle/fluid/operators/benchmark/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..54008336a9f67f0123ba1cfa6fcea35b79b7ac4c --- /dev/null +++ b/paddle/fluid/operators/benchmark/CMakeLists.txt @@ -0,0 +1,3 @@ +cc_test(op_tester SRCS op_tester.cc op_tester_config.cc + DEPS memory timer framework_proto proto_desc lod_tensor op_registry + device_context scope ${GLOB_OP_LIB} ${GLOB_OPERATOR_DEPS}) diff --git a/paddle/fluid/operators/benchmark/op_tester.cc b/paddle/fluid/operators/benchmark/op_tester.cc new file mode 100644 index 0000000000000000000000000000000000000000..e179de56cddc5fada2e5833086d351659a7cf540 --- /dev/null +++ b/paddle/fluid/operators/benchmark/op_tester.cc @@ -0,0 +1,303 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/benchmark/op_tester.h" +#include "gflags/gflags.h" +#include "gtest/gtest.h" +#include "paddle/fluid/framework/op_info.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/framework/variable_helper.h" +#include "paddle/fluid/platform/init.h" +#include "paddle/fluid/platform/profiler.h" +#include "paddle/fluid/platform/timer.h" +#include "paddle/fluid/pybind/pybind.h" + +namespace paddle { +namespace operators { +namespace benchmark { + +DEFINE_string(op_config_list, "", "Path of op config file."); + +void OpTester::Init(const std::string &filename) { + Init(OpTesterConfig(filename)); +} + +void OpTester::Init(const OpTesterConfig &config) { + config_ = config; + + auto &op_desc_info = framework::OpInfoMap::Instance(); + // Initialize the OpDesc + if (op_desc_info.Has(config_.op_type)) { + type_ = config_.op_type; + op_desc_.SetType(config_.op_type); + + CreateInputVarDesc(); + CreateOutputVarDesc(); + } else { + LOG(FATAL) << "Op \"" << config_.op_type << "\" is not registered."; + } + + if (config_.device_id >= 0) { + place_ = paddle::platform::CUDAPlace(config_.device_id); + } else { + place_ = paddle::platform::CPUPlace(); + } + + framework::InitDevices(false); + scope_.reset(new paddle::framework::Scope()); + + op_ = framework::OpRegistry::CreateOp(op_desc_); + CreateVariables(scope_.get()); +} + +void OpTester::Run() { + if (config_.print_debug_string) { + LOG(INFO) << DebugString(); + } + + // Warm up + RunImpl(); + + platform::Timer timer; + if (config_.profile) { + if (platform::is_cpu_place(place_)) { + platform::EnableProfiler(platform::ProfilerState::kCPU); + } else { +#ifdef PADDLE_WITH_CUDA + platform::EnableProfiler(platform::ProfilerState::kAll); + platform::SetDeviceId(config_.device_id); +#else + PADDLE_THROW("'CUDAPlace' is not supported in CPU only device."); +#endif + } + + timer.Start(); + for (int i = config_.repeat; i > 0; --i) { + RunImpl(); + } + timer.Pause(); + platform::DisableProfiler(platform::EventSortingKey::kDefault, + "op_tester_profiler"); + } else { + timer.Start(); + for (int i = config_.repeat; i > 0; --i) { + RunImpl(); + } + timer.Pause(); + } + config_.runtime = timer.ElapsedMS() / config_.repeat; + LOG(INFO) << "=== Run " << config_.repeat + << " times, latency: " << config_.runtime << " ms ==="; +} + +void OpTester::RunImpl() { + op_->Run(*scope_, place_); + platform::DeviceContextPool::Instance().Get(place_)->Wait(); + scope_->DropKids(); +} + +std::vector OpTester::GetOpProtoInputNames() { + std::vector input_names; + const framework::proto::OpProto &proto = + framework::OpInfoMap::Instance().Get(type_).Proto(); + for (int i = 0; i != proto.inputs_size(); ++i) { + const auto &input = proto.inputs(i); + input_names.push_back(input.name()); + } + return input_names; +} + +std::vector OpTester::GetOpProtoOutputNames() { + std::vector output_names; + const framework::proto::OpProto &proto = + framework::OpInfoMap::Instance().Get(type_).Proto(); + for (int i = 0; i != proto.outputs_size(); ++i) { + const auto &output = proto.outputs(i); + output_names.push_back(output.name()); + } + return output_names; +} + +void OpTester::CreateInputVarDesc() { + std::vector input_names = GetOpProtoInputNames(); + for (auto &name : input_names) { + const OpInputConfig *input = config_.GetInput(name); + if (input == nullptr) { + LOG(FATAL) << "The input " << name << " of op " << config_.op_type + << " is not correctlly provided."; + } + + std::string var_name = config_.op_type + "." + name; + framework::VarDesc *var = Var(var_name); + // Need to support more type + var->SetType(framework::proto::VarType::LOD_TENSOR); + var->SetPersistable(false); + var->SetDataType(framework::proto::VarType::FP32); + var->SetShape(input->dims); + + op_desc_.SetInput(name, {var_name}); + inputs_.push_back(var_name); + } +} + +void OpTester::CreateOutputVarDesc() { + std::vector output_names = GetOpProtoOutputNames(); + for (auto &name : output_names) { + std::string var_name = config_.op_type + "." + name; + framework::VarDesc *var = Var(var_name); + // Need to support more type + var->SetType(framework::proto::VarType::LOD_TENSOR); + var->SetPersistable(false); + var->SetDataType(framework::proto::VarType::FP32); + + op_desc_.SetOutput(name, {var_name}); + outputs_.push_back(var_name); + } +} + +framework::VarDesc *OpTester::Var(const std::string &name) { + auto it = vars_.find(name); + if (it != vars_.end()) { + return it->second.get(); + } + auto *var = new framework::VarDesc(name); + vars_[name].reset(var); + return var; +} + +template +void OpTester::SetupTensor(framework::LoDTensor *tensor, + const std::vector &shape, T lower, + T upper) { + static unsigned int seed = 100; + std::mt19937 rng(seed++); + std::uniform_real_distribution uniform_dist(0, 1); + + T *ptr = tensor->mutable_data(framework::make_ddim(shape), place_); + if (platform::is_cpu_place(place_)) { + for (int i = 0; i < tensor->numel(); ++i) { + ptr[i] = static_cast(uniform_dist(rng) * (upper - lower) + lower); + } + } else { + framework::LoDTensor cpu_tensor; + T *cpu_ptr = cpu_tensor.mutable_data(framework::make_ddim(shape), + platform::CPUPlace()); + for (int i = 0; i < cpu_tensor.numel(); ++i) { + cpu_ptr[i] = static_cast(uniform_dist(rng) * (upper - lower) + lower); + } + TensorCopySync(cpu_tensor, place_, tensor); + } +} + +void OpTester::CreateVariables(framework::Scope *scope) { + for (auto &item : vars_) { + auto &var = item.second; + if (var->Name() == framework::kEmptyVarName) { + continue; + } + + auto *ptr = scope->Var(var->Name()); + framework::InitializeVariable(ptr, var->GetType()); + if (var->Persistable()) { + VLOG(3) << "Create Variable " << var->Name() + << " global, which pointer is " << ptr; + } else { + VLOG(3) << "Create Variable " << var->Name() + << " locally, which pointer is " << ptr; + } + } + + // Allocate memory for input tensor + for (auto &name : inputs_) { + VLOG(3) << "Allocate memory for tensor " << name; + auto &var_desc = vars_[name]; + std::vector shape = var_desc->GetShape(); + + auto *var = scope->Var(name); + auto *tensor = var->GetMutable(); + SetupTensor(tensor, shape, static_cast(0.0), + static_cast(1.0)); + } +} + +static std::string GenSpaces(int count) { + std::stringstream ss; + for (int i = 0; i < count; ++i) { + ss << " "; + } + return ss.str(); +} + +std::string OpTester::DebugString() { + std::stringstream ss; + int count = 0; + for (auto &item : vars_) { + auto &var = item.second; + ss << GenSpaces(count++) << "vars {\n"; + ss << GenSpaces(count) << "name: \"" << var->Name() << "\"\n"; + ss << GenSpaces(count++) << "type: {\n"; + ss << GenSpaces(count) << "type: LOD_TENSOR\n"; + ss << GenSpaces(count++) << "lod_tensor {\n"; + ss << GenSpaces(count++) << "tensor {\n"; + ss << GenSpaces(count) << "data_type: FP32\n"; + std::vector shape = var->GetShape(); + for (auto d : shape) { + ss << GenSpaces(count) << "dims: " << d << "\n"; + } + ss << GenSpaces(--count) << "}\n"; + ss << GenSpaces(--count) << "}\n"; + ss << GenSpaces(--count) << "}\n"; + ss << GenSpaces(count) << "persistable: " << var->Persistable() << "\n"; + ss << GenSpaces(--count) << "}\n"; + } + ss << GenSpaces(count++) << "ops {\n"; + for (auto &name : op_desc_.InputNames()) { + ss << GenSpaces(count++) << "inputs {\n"; + ss << GenSpaces(count) << "parameters: \"" << name << "\"\n"; + ss << GenSpaces(count) << "arguments: \"" << op_desc_.Input(name)[0] + << "\"\n"; + ss << GenSpaces(--count) << "}\n"; + } + for (auto &name : op_desc_.OutputNames()) { + ss << GenSpaces(count++) << "outputs {\n"; + ss << GenSpaces(count) << "parameters: \"" << name << "\"\n"; + ss << GenSpaces(count) << "arguments: \"" << op_desc_.Output(name)[0] + << "\"\n"; + ss << GenSpaces(--count) << "}\n"; + } + ss << GenSpaces(count) << "type: " << op_desc_.Type() << "\n"; + ss << GenSpaces(--count) << "}\n"; + return ss.str(); +} + +TEST(op_tester, base) { + OpTester tester; + if (!FLAGS_op_config_list.empty()) { + tester.Init(FLAGS_op_config_list); + } else { + OpTesterConfig config; + config.op_type = "elementwise_add"; + config.inputs.resize(2); + config.inputs[0].name = "X"; + config.inputs[0].dims = {64, 64}; + config.inputs[1].name = "Y"; + config.inputs[1].dims = {64, 1}; + tester.Init(config); + } + tester.Run(); +} + +} // namespace benchmark +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/benchmark/op_tester.h b/paddle/fluid/operators/benchmark/op_tester.h new file mode 100644 index 0000000000000000000000000000000000000000..1723d46c47ed67199713e6d726c6245f34f7c224 --- /dev/null +++ b/paddle/fluid/operators/benchmark/op_tester.h @@ -0,0 +1,69 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include +#include "paddle/fluid/framework/ddim.h" +#include "paddle/fluid/framework/op_desc.h" +#include "paddle/fluid/framework/operator.h" +#include "paddle/fluid/operators/benchmark/op_tester_config.h" + +namespace paddle { +namespace operators { +namespace benchmark { + +class OpTester { + public: + OpTester() {} + + void Init(const std::string &filename); + void Init(const OpTesterConfig &config); + + void Run(); + + std::string DebugString(); + + private: + std::vector GetOpProtoInputNames(); + std::vector GetOpProtoOutputNames(); + + void CreateInputVarDesc(); + void CreateOutputVarDesc(); + + framework::VarDesc *Var(const std::string &name); + void CreateVariables(framework::Scope *scope); + + template + void SetupTensor(framework::LoDTensor *input, + const std::vector &shape, T lower, T upper); + + void RunImpl(); + + private: + OpTesterConfig config_; + std::string type_; + framework::OpDesc op_desc_; + std::unordered_map> vars_; + std::vector inputs_; + std::vector outputs_; + std::unique_ptr op_; + platform::Place place_; + std::unique_ptr scope_; +}; + +} // namespace benchmark +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/benchmark/op_tester_config.cc b/paddle/fluid/operators/benchmark/op_tester_config.cc new file mode 100644 index 0000000000000000000000000000000000000000..3db8de7f76801eb814b57859d6b95590761c96f3 --- /dev/null +++ b/paddle/fluid/operators/benchmark/op_tester_config.cc @@ -0,0 +1,114 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/benchmark/op_tester_config.h" +#include +#include "glog/logging.h" +#include "paddle/fluid/platform/enforce.h" + +namespace paddle { +namespace operators { +namespace benchmark { + +static const char kStartSeparator[] = "{"; +static const char kEndSeparator[] = "}"; +static const char kSepBetweenItems[] = ";"; + +static bool StartWith(const std::string& str, const std::string& substr) { + return str.find(substr) == 0; +} + +static bool EndWith(const std::string& str, const std::string& substr) { + return str.rfind(substr) == (str.length() - substr.length()); +} + +static void EraseEndSep(std::string* str) { + std::string substr = kSepBetweenItems; + if (EndWith(*str, substr)) { + str->erase(str->length() - substr.length(), str->length()); + } +} + +static std::vector ParseDims(std::string dims_str) { + std::vector dims; + std::string token; + std::istringstream token_stream(dims_str); + while (std::getline(token_stream, token, 'x')) { + dims.push_back(std::stoi(token)); + } + return dims; +} + +OpInputConfig::OpInputConfig(std::istream& is) { + std::string sep; + is >> sep; + if (sep == kStartSeparator) { + while (sep != kEndSeparator) { + is >> sep; + if (sep == "name" || sep == "name:") { + is >> name; + EraseEndSep(&name); + } else if (sep == "dims" || sep == "dims:") { + std::string dims_str; + is >> dims_str; + dims = ParseDims(dims_str); + } + } + } +} + +OpTesterConfig::OpTesterConfig(const std::string& filename) { + std::ifstream fin(filename, std::ios::in | std::ios::binary); + PADDLE_ENFORCE(static_cast(fin), "Cannot open file %s", + filename.c_str()); + + Init(fin); +} + +void OpTesterConfig::Init(std::istream& is) { + std::string sep; + is >> sep; + if (sep == kStartSeparator) { + while (sep != kEndSeparator) { + is >> sep; + if (sep == "op_type" || sep == "op_type:") { + is >> op_type; + } else if (sep == "device_id" || sep == "device_id:") { + is >> device_id; + } else if (sep == "repeat" || sep == "repeat:") { + is >> repeat; + } else if (sep == "profile" || sep == "profile:") { + is >> profile; + } else if (sep == "print_debug_string" || sep == "print_debug_string:") { + is >> print_debug_string; + } else if (sep == "input" || sep == "input:") { + OpInputConfig input_config(is); + inputs.push_back(input_config); + } + } + } +} + +const OpInputConfig* OpTesterConfig::GetInput(const std::string& name) { + for (size_t i = 0; i < inputs.size(); ++i) { + if (inputs[i].name == name) { + return &inputs[i]; + } + } + return nullptr; +} + +} // namespace benchmark +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/benchmark/op_tester_config.h b/paddle/fluid/operators/benchmark/op_tester_config.h new file mode 100644 index 0000000000000000000000000000000000000000..f7b62cb8ad03b410a2ea99fe4c2a8dc8a6bea7a7 --- /dev/null +++ b/paddle/fluid/operators/benchmark/op_tester_config.h @@ -0,0 +1,51 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include +#include + +namespace paddle { +namespace operators { +namespace benchmark { + +struct OpInputConfig { + OpInputConfig() {} + explicit OpInputConfig(std::istream& is); + + std::string name; + std::vector dims; +}; + +struct OpTesterConfig { + OpTesterConfig() {} + explicit OpTesterConfig(const std::string& filename); + void Init(std::istream& is); + + const OpInputConfig* GetInput(const std::string& name); + + std::string op_type; + std::vector inputs; + int device_id{-1}; // CPU: -1 + int repeat{1}; + int profile{0}; + int print_debug_string{0}; + double runtime{0.0}; +}; + +} // namespace benchmark +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/controlflow/compare_op.cc b/paddle/fluid/operators/controlflow/compare_op.cc index 688457d4a75168577302e45817ef0463d6ff3718..5d3f9b43f8c08d356319fa0b9ccaf808811d3d39 100644 --- a/paddle/fluid/operators/controlflow/compare_op.cc +++ b/paddle/fluid/operators/controlflow/compare_op.cc @@ -51,6 +51,11 @@ class CompareOpProtoMaker : public framework::OpProtoAndCheckerMaker { comment.type)); AddInput("Y", string::Sprintf("the right hand operand of %s operator", comment.type)); + AddAttr( + "axis", + "The start dimension index for broadcasting Y onto X. [default -1]") + .SetDefault(-1) + .EqualGreaterThan(-1); AddAttr("force_cpu", "Force fill output variable to cpu " "memory. Otherwise, fill output variable to the running " @@ -64,11 +69,6 @@ N-dim tensor. X and Y could be any type. The each element of the Out tensor is calculated by $%s$ )DOC", comment.equation)); - AddAttr( - "axis", - "The start dimension index for broadcasting Y onto X. [default -1]") - .SetDefault(-1) - .EqualGreaterThan(-1); } }; diff --git a/paddle/fluid/operators/controlflow/get_places_op.cc b/paddle/fluid/operators/controlflow/get_places_op.cc index db6ff7825690176ded0ab957764ed8411d3cd804..1a157688f3d02185d18b66ff5ba3613b6cf438ad 100644 --- a/paddle/fluid/operators/controlflow/get_places_op.cc +++ b/paddle/fluid/operators/controlflow/get_places_op.cc @@ -52,7 +52,7 @@ class GetPlacesOp : public framework::OperatorBase { device_count = is_gpu ? CUDADevCount() : std::thread::hardware_concurrency(); } - PADDLE_ENFORCE_NE(device_count, 0, "Cannot indicate %s device count", + PADDLE_ENFORCE_NE(device_count, 0UL, "Cannot indicate %s device count", is_gpu ? "GPU" : "CPU"); auto out_var_name = Output("Out"); diff --git a/paddle/fluid/operators/conv_op.cc b/paddle/fluid/operators/conv_op.cc index bd788f03e7d666aad7ce6f0c63cea30f029e3491..fd9f156d070bdb1990a2fc9c63305933050e5524 100644 --- a/paddle/fluid/operators/conv_op.cc +++ b/paddle/fluid/operators/conv_op.cc @@ -222,7 +222,7 @@ void Conv2DOpMaker::Make() { .SetDefault(4096); AddAttr("exhaustive_search", "(bool, default false) cuDNN has many algorithm to calculation " - "convolution, whether enable exhaustive search ", + "convolution, whether enable exhaustive search " "for cuDNN convolution or not, defalut is False.") .SetDefault(false); AddComment(R"DOC( @@ -341,7 +341,7 @@ void Conv3DOpMaker::Make() { .SetDefault(4096); AddAttr("exhaustive_search", "(bool, default false) cuDNN has many algorithm to calculation " - "convolution, whether enable exhaustive search ", + "convolution, whether enable exhaustive search " "for cuDNN convolution or not, defalut is False.") .SetDefault(false); AddComment(R"DOC( diff --git a/paddle/fluid/operators/crf_decoding_op.cc b/paddle/fluid/operators/crf_decoding_op.cc index 81c9e9e543191d9b2d606217d726cc783be97fea..e053ae57739d3d96209e9ca180cc041f8b55396e 100644 --- a/paddle/fluid/operators/crf_decoding_op.cc +++ b/paddle/fluid/operators/crf_decoding_op.cc @@ -84,12 +84,12 @@ class CRFDecodingOp : public framework::OperatorWithKernel { "Output(ViterbiPath) should be not null."); auto emission_dims = ctx->GetInputDim("Emission"); - PADDLE_ENFORCE_EQ(emission_dims.size(), 2UL, + PADDLE_ENFORCE_EQ(emission_dims.size(), 2, "The Input(Emission) should be a 2-D tensor."); PADDLE_ENFORCE(emission_dims[0], "An empty mini-batch is not allowed."); auto transition_dims = ctx->GetInputDim("Transition"); - PADDLE_ENFORCE_EQ(transition_dims.size(), 2UL, + PADDLE_ENFORCE_EQ(transition_dims.size(), 2, "The Input(Transition) should be a 2-D tensor."); PADDLE_ENFORCE_EQ( transition_dims[0] - 2, transition_dims[1], diff --git a/paddle/fluid/operators/detection/CMakeLists.txt b/paddle/fluid/operators/detection/CMakeLists.txt index d3a61dc367c3642b8faa9085a470a302712395e5..f6fbe97565c43c306ea885c765c0a665492fa317 100644 --- a/paddle/fluid/operators/detection/CMakeLists.txt +++ b/paddle/fluid/operators/detection/CMakeLists.txt @@ -31,6 +31,8 @@ detection_library(polygon_box_transform_op SRCS polygon_box_transform_op.cc polygon_box_transform_op.cu) detection_library(rpn_target_assign_op SRCS rpn_target_assign_op.cc) detection_library(generate_proposal_labels_op SRCS generate_proposal_labels_op.cc) +detection_library(box_clip_op SRCS box_clip_op.cc box_clip_op.cu) +detection_library(yolov3_loss_op SRCS yolov3_loss_op.cc) if(WITH_GPU) detection_library(generate_proposals_op SRCS generate_proposals_op.cc generate_proposals_op.cu DEPS memory cub) diff --git a/paddle/fluid/operators/detection/anchor_generator_op.cc b/paddle/fluid/operators/detection/anchor_generator_op.cc index f2984d1af2f26d901bc30ecfd519d5268a60278a..4a333b559f82e6d39d2d4345c8ad58bc8d430c69 100644 --- a/paddle/fluid/operators/detection/anchor_generator_op.cc +++ b/paddle/fluid/operators/detection/anchor_generator_op.cc @@ -85,7 +85,7 @@ class AnchorGeneratorOpMaker : public framework::OpProtoAndCheckerMaker { " For instance, the anchor size of 64 means the area of this anchor " "equals to 64**2.") .AddCustomChecker([](const std::vector& anchor_sizes) { - PADDLE_ENFORCE_GT(anchor_sizes.size(), 0, + PADDLE_ENFORCE_GT(anchor_sizes.size(), 0UL, "Size of anchor_sizes must be at least 1."); for (size_t i = 0; i < anchor_sizes.size(); ++i) { PADDLE_ENFORCE_GT(anchor_sizes[i], 0.0, @@ -103,7 +103,7 @@ class AnchorGeneratorOpMaker : public framework::OpProtoAndCheckerMaker { "(vector) List of variances to be used " "in box regression deltas") .AddCustomChecker([](const std::vector& variances) { - PADDLE_ENFORCE_EQ(variances.size(), 4, + PADDLE_ENFORCE_EQ(variances.size(), 4UL, "Must and only provide 4 variance."); for (size_t i = 0; i < variances.size(); ++i) { PADDLE_ENFORCE_GT(variances[i], 0.0, @@ -117,7 +117,7 @@ class AnchorGeneratorOpMaker : public framework::OpProtoAndCheckerMaker { .SetDefault(std::vector(2, 16.0)) .AddCustomChecker([](const std::vector& stride) { PADDLE_ENFORCE_EQ( - stride.size(), 2, + stride.size(), 2UL, "Must and only provide 2 stride for width and height."); for (size_t i = 0; i < stride.size(); ++i) { PADDLE_ENFORCE_GT(stride[i], 0.0, diff --git a/paddle/fluid/operators/detection/bbox_util.h b/paddle/fluid/operators/detection/bbox_util.h index b99edb5bf05f94e762b377a8882e4c3fcdb5afad..a7bc3e027229884e78721d29428a8ab3f08a6ebc 100644 --- a/paddle/fluid/operators/detection/bbox_util.h +++ b/paddle/fluid/operators/detection/bbox_util.h @@ -99,5 +99,29 @@ void BboxOverlaps(const framework::Tensor& r_boxes, } } +template +void ClipTiledBoxes(const platform::DeviceContext& ctx, + const framework::Tensor& im_info, + const framework::Tensor& input_boxes, + framework::Tensor* out) { + T* out_data = out->mutable_data(ctx.GetPlace()); + const T* im_info_data = im_info.data(); + const T* input_boxes_data = input_boxes.data(); + T zero(0); + T im_w = round(im_info_data[1] / im_info_data[2]); + T im_h = round(im_info_data[0] / im_info_data[2]); + for (int64_t i = 0; i < input_boxes.numel(); ++i) { + if (i % 4 == 0) { + out_data[i] = std::max(std::min(input_boxes_data[i], im_w - 1), zero); + } else if (i % 4 == 1) { + out_data[i] = std::max(std::min(input_boxes_data[i], im_h - 1), zero); + } else if (i % 4 == 2) { + out_data[i] = std::max(std::min(input_boxes_data[i], im_w - 1), zero); + } else { + out_data[i] = std::max(std::min(input_boxes_data[i], im_h - 1), zero); + } + } +} + } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/detection/box_clip_op.cc b/paddle/fluid/operators/detection/box_clip_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..3aa766559a530bc31fbb277f2bcd474da776e63b --- /dev/null +++ b/paddle/fluid/operators/detection/box_clip_op.cc @@ -0,0 +1,86 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + http://www.apache.org/licenses/LICENSE-2.0 +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/detection/box_clip_op.h" +#include "paddle/fluid/framework/op_registry.h" + +namespace paddle { +namespace operators { + +class BoxClipOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Input"), + "Input(Input) of BoxClipOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("ImInfo"), + "Input(ImInfo) of BoxClipOp should not be null."); + + auto input_box_dims = ctx->GetInputDim("Input"); + auto im_info_dims = ctx->GetInputDim("ImInfo"); + + if (ctx->IsRuntime()) { + auto input_box_size = input_box_dims.size(); + PADDLE_ENFORCE_EQ(input_box_dims[input_box_size - 1], 4, + "The last dimension of Input must be 4"); + PADDLE_ENFORCE_EQ(im_info_dims.size(), 2, + "The rank of Input(Input) in BoxClipOp must be 2"); + PADDLE_ENFORCE_EQ(im_info_dims[1], 3, + "The last dimension of ImInfo must be 3"); + } + ctx->ShareDim("Input", /*->*/ "Output"); + ctx->ShareLoD("Input", /*->*/ "Output"); + } +}; + +class BoxClipOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("Input", + "(LoDTensor) " + "Input is a LoDTensor with shape [..., 4] holds 4 points" + "in last dimension in format [xmin, ymin, xmax, ymax]"); + AddInput("ImInfo", + "(Tensor) Information for image reshape is in shape (N, 3), " + "in format (height, width, im_scale)"); + AddOutput("Output", + "(LoDTensor) " + "Output is a LoDTensor with the same shape as Input" + "and it is the result after clip"); + AddComment(R"DOC( +This operator clips input boxes to original input images. + +For each input box, The formula is given as follows: + + $$xmin = \max(\min(xmin, im_w - 1), 0)$$ + $$ymin = \max(\min(ymin, im_h - 1), 0)$$ + $$xmax = \max(\min(xmax, im_w - 1), 0)$$ + $$ymax = \max(\min(ymax, im_h - 1), 0)$$ + +where im_w and im_h are computed from ImInfo, the formula is given as follows: + + $$im_w = \round(width / im_scale)$$ + $$im_h = \round(height / im_scale)$$ +)DOC"); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OPERATOR(box_clip, ops::BoxClipOp, ops::BoxClipOpMaker, + paddle::framework::EmptyGradOpMaker); +REGISTER_OP_CPU_KERNEL( + box_clip, ops::BoxClipKernel, + ops::BoxClipKernel); diff --git a/paddle/fluid/operators/detection/box_clip_op.cu b/paddle/fluid/operators/detection/box_clip_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..b727da5f7b736b6f22407d1dfbca708ed0cf04d9 --- /dev/null +++ b/paddle/fluid/operators/detection/box_clip_op.cu @@ -0,0 +1,74 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/detection/box_clip_op.h" +#include "paddle/fluid/operators/math/math_function.h" +#include "paddle/fluid/platform/cuda_primitives.h" +#include "paddle/fluid/platform/hostdevice.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; +using LoDTenso = framework::LoDTensor; + +static constexpr int ImInfoSize = 3; + +template +static __global__ void GPUBoxClip(const T *input, const size_t *lod, + const size_t width, const T *im_info, + T *output) { + T im_w = round(im_info[blockIdx.x * ImInfoSize + 1] / + im_info[blockIdx.x * ImInfoSize + 2]); + T im_h = round(im_info[blockIdx.x * ImInfoSize] / + im_info[blockIdx.x * ImInfoSize + 2]); + for (int i = threadIdx.x; i < (lod[blockIdx.x + 1] - lod[blockIdx.x]) * width; + i += BlockSize) { + int idx = lod[blockIdx.x] * width + i; + T im_size = (idx % 2 == 0) ? im_w : im_h; + output[idx] = max(min(input[idx], im_size - 1), T(0.)); + } +} + +template +class GPUBoxClipKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &context) const override { + PADDLE_ENFORCE(platform::is_gpu_place(context.GetPlace()), + "This kernel only runs on GPU device."); + auto *input = context.Input("Input"); + auto *im_info = context.Input("ImInfo"); + auto *output = context.Output("Output"); + const int64_t num = input->dims()[0]; + const int64_t bbox_width = input->numel() / num; + auto lod = input->lod(); + framework::LoD abs_offset_lod = framework::ToAbsOffset(lod); + auto &dev_ctx = context.template device_context(); + auto stream = dev_ctx.stream(); + const size_t batch_size = lod.back().size() - 1; + T *output_data = output->mutable_data(dev_ctx.GetPlace()); + GPUBoxClip<<>>( + input->data(), abs_offset_lod[0].CUDAMutableData(dev_ctx.GetPlace()), + bbox_width, im_info->data(), output_data); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP_CUDA_KERNEL( + box_clip, ops::GPUBoxClipKernel, + ops::GPUBoxClipKernel); diff --git a/paddle/fluid/operators/detection/box_clip_op.h b/paddle/fluid/operators/detection/box_clip_op.h new file mode 100644 index 0000000000000000000000000000000000000000..74e1f88f8d8b28e490d170934760bd9bffc807bc --- /dev/null +++ b/paddle/fluid/operators/detection/box_clip_op.h @@ -0,0 +1,50 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + http://www.apache.org/licenses/LICENSE-2.0 +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/detection/bbox_util.h" +#include "paddle/fluid/operators/math/math_function.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; +using LoDTensor = framework::LoDTensor; + +template +class BoxClipKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto* input_box = context.Input("Input"); + auto* im_info = context.Input("ImInfo"); + auto* output_box = context.Output("Output"); + auto& dev_ctx = + context.template device_context(); + output_box->mutable_data(context.GetPlace()); + if (input_box->lod().size()) { + PADDLE_ENFORCE_EQ(input_box->lod().size(), 1UL, + "Only support 1 level of LoD."); + } + auto box_lod = input_box->lod().back(); + int64_t n = static_cast(box_lod.size() - 1); + for (int i = 0; i < n; ++i) { + Tensor im_info_slice = im_info->Slice(i, i + 1); + Tensor box_slice = input_box->Slice(box_lod[i], box_lod[i + 1]); + Tensor output_slice = output_box->Slice(box_lod[i], box_lod[i + 1]); + ClipTiledBoxes(dev_ctx, im_info_slice, box_slice, &output_slice); + } + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/detection/box_coder_op.cc b/paddle/fluid/operators/detection/box_coder_op.cc index 06fbb9815c52ea69e3aa9e893512e039853b9514..0a51d50e06176e713922837861f2102c9ee8a899 100644 --- a/paddle/fluid/operators/detection/box_coder_op.cc +++ b/paddle/fluid/operators/detection/box_coder_op.cc @@ -10,6 +10,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/operators/detection/box_coder_op.h" +#include namespace paddle { namespace operators { @@ -32,32 +33,49 @@ class BoxCoderOp : public framework::OperatorWithKernel { if (ctx->IsRuntime()) { PADDLE_ENFORCE_EQ(prior_box_dims.size(), 2, - "The rank of Input of PriorBoxVar must be 2"); + "The rank of Input PriorBox must be 2"); PADDLE_ENFORCE_EQ(prior_box_dims[1], 4, "The shape of PriorBox is [N, 4]"); if (ctx->HasInput("PriorBoxVar")) { auto prior_box_var_dims = ctx->GetInputDim("PriorBoxVar"); - PADDLE_ENFORCE_EQ(prior_box_dims, prior_box_var_dims); + PADDLE_ENFORCE(prior_box_var_dims.size() == 2, + "Input(PriorBoxVar) of BoxCoderOp should be 2."); + PADDLE_ENFORCE_EQ( + prior_box_dims, prior_box_var_dims, + "The dimension of Input(PriorBoxVar) should be equal to" + "the dimension of Input(PriorBox) when the rank is 2."); } + } - auto code_type = - GetBoxCodeType(ctx->Attrs().Get("code_type")); - if (code_type == BoxCodeType::kEncodeCenterSize) { - PADDLE_ENFORCE_EQ(target_box_dims.size(), 2, - "The rank of Input of TargetBox must be 2"); - PADDLE_ENFORCE_EQ(target_box_dims[1], 4, - "The shape of TargetBox is [M, 4]"); - } else if (code_type == BoxCodeType::kDecodeCenterSize) { - PADDLE_ENFORCE_EQ(target_box_dims.size(), 3, - "The rank of Input of TargetBox must be 3"); + auto code_type = GetBoxCodeType(ctx->Attrs().Get("code_type")); + int axis = ctx->Attrs().Get("axis"); + if (code_type == BoxCodeType::kEncodeCenterSize) { + PADDLE_ENFORCE_EQ(target_box_dims.size(), 2, + "The rank of Input TargetBox must be 2"); + PADDLE_ENFORCE_EQ(target_box_dims[1], 4, + "The shape of TargetBox is [M, 4]"); + ctx->SetOutputDim( + "OutputBox", + framework::make_ddim({target_box_dims[0], prior_box_dims[0], 4})); + } else if (code_type == BoxCodeType::kDecodeCenterSize) { + PADDLE_ENFORCE_EQ(target_box_dims.size(), 3, + "The rank of Input TargetBox must be 3"); + if (axis == 0) { PADDLE_ENFORCE_EQ(target_box_dims[1], prior_box_dims[0]); - PADDLE_ENFORCE_EQ(target_box_dims[2], prior_box_dims[1]); + } else if (axis == 1) { + PADDLE_ENFORCE_EQ(target_box_dims[0], prior_box_dims[0]); + } else { + PADDLE_THROW("axis must be 0 or 1."); } + PADDLE_ENFORCE_EQ(target_box_dims[2], prior_box_dims[1]); + ctx->ShareDim("TargetBox", /*->*/ "OutputBox"); + } + + if (code_type == BoxCodeType::kDecodeCenterSize && axis == 1) { + ctx->ShareLoD("PriorBox", /*->*/ "OutputBox"); + } else { + ctx->ShareLoD("TargetBox", /*->*/ "OutputBox"); } - ctx->SetOutputDim( - "OutputBox", - framework::make_ddim({target_box_dims[0], prior_box_dims[0], 4})); - ctx->ShareLoD("TargetBox", /*->*/ "OutputBox"); } }; @@ -100,6 +118,21 @@ class BoxCoderOpMaker : public framework::OpProtoAndCheckerMaker { "(bool, default true) " "whether treat the priorbox as a noramlized box") .SetDefault(true); + AddAttr("axis", + "(int, default 0)" + "which axis in PriorBox to broadcast for box decode," + "for example, if axis is 0 and TargetBox has shape" + "[N, M, 4] and PriorBox has shape [M, 4], then PriorBox " + "will broadcast to [N, M, 4] for decoding. It is only valid" + "when code type is decode_center_size") + .SetDefault(0) + .InEnum({0, 1}); + AddAttr>( + "variance", + "(vector, default {})," + "variance of prior box with shape [4]. PriorBoxVar and variance can" + "not be provided at the same time.") + .SetDefault(std::vector{}); AddOutput("OutputBox", "(LoDTensor or Tensor) " "When code_type is 'encode_center_size', the output tensor of " @@ -138,7 +171,11 @@ where `tx`, `ty`, `tw`, `th` denote the target box's center coordinates, width and height respectively. Similarly, `px`, `py`, `pw`, `ph` denote the priorbox's (anchor) center coordinates, width and height. `pxv`, `pyv`, `pwv`, `phv` denote the variance of the priorbox and `ox`, `oy`, `ow`, `oh` denote the -encoded/decoded coordinates, width and height. +encoded/decoded coordinates, width and height. + +During Box Decoding, two modes for broadcast are supported. Say target box has +shape [N, M, 4], and the shape of prior box can be [N, 4] or [M, 4]. Then prior +box will broadcast to target box along the assigned axis. )DOC"); } }; diff --git a/paddle/fluid/operators/detection/box_coder_op.cu b/paddle/fluid/operators/detection/box_coder_op.cu index a7af111f63d654319dd1d90d2032956951dfe49e..19a5bb90fa828899ad6270c051090dd3662aeed8 100644 --- a/paddle/fluid/operators/detection/box_coder_op.cu +++ b/paddle/fluid/operators/detection/box_coder_op.cu @@ -9,6 +9,9 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ +#include +#include +#include "paddle/fluid/memory/memcpy.h" #include "paddle/fluid/operators/detection/box_coder_op.h" #include "paddle/fluid/platform/cuda_primitives.h" @@ -16,11 +19,11 @@ namespace paddle { namespace operators { template -__global__ void EncodeCenterSizeKernel(const T* prior_box_data, - const T* prior_box_var_data, - const T* target_box_data, const int row, - const int col, const int len, - const bool normalized, T* output) { +__global__ void EncodeCenterSizeKernel( + const T* prior_box_data, const T* prior_box_var_data, + const T* target_box_data, const int row, const int col, const int len, + const bool normalized, const T prior_box_var_size, const float* variance, + const int var_size, T* output) { const int idx = threadIdx.x + blockIdx.x * blockDim.x; if (idx < row * col) { const int row_idx = idx / col; @@ -30,11 +33,9 @@ __global__ void EncodeCenterSizeKernel(const T* prior_box_data, T prior_box_height = prior_box_data[col_idx * len + 3] - prior_box_data[col_idx * len + 1] + (normalized == false); - T prior_box_center_x = - (prior_box_data[col_idx * len + 2] + prior_box_data[col_idx * len]) / 2; - T prior_box_center_y = (prior_box_data[col_idx * len + 3] + - prior_box_data[col_idx * len + 1]) / - 2; + T prior_box_center_x = prior_box_data[col_idx * len] + prior_box_width / 2; + T prior_box_center_y = + prior_box_data[col_idx * len + 1] + prior_box_height / 2; T target_box_center_x = (target_box_data[row_idx * len + 2] + target_box_data[row_idx * len]) / @@ -55,58 +56,67 @@ __global__ void EncodeCenterSizeKernel(const T* prior_box_data, output[idx * len + 2] = log(fabs(target_box_width / prior_box_width)); output[idx * len + 3] = log(fabs(target_box_height / prior_box_height)); if (prior_box_var_data) { - output[idx * len] /= prior_box_var_data[col_idx * len]; - output[idx * len + 1] /= prior_box_var_data[col_idx * len + 1]; - output[idx * len + 2] /= prior_box_var_data[col_idx * len + 2]; - output[idx * len + 3] /= prior_box_var_data[col_idx * len + 3]; + int prior_var_offset = col_idx * len; + output[idx * len] /= prior_box_var_data[prior_var_offset]; + output[idx * len + 1] /= prior_box_var_data[prior_var_offset + 1]; + output[idx * len + 2] /= prior_box_var_data[prior_var_offset + 2]; + output[idx * len + 3] /= prior_box_var_data[prior_var_offset + 3]; + } else if (var_size == 4) { + for (int k = 0; k < 4; ++k) { + output[idx * len + k] /= static_cast(variance[k]); + } } } } template -__global__ void DecodeCenterSizeKernel(const T* prior_box_data, - const T* prior_box_var_data, - const T* target_box_data, const int row, - const int col, const int len, - const bool normalized, T* output) { +__global__ void DecodeCenterSizeKernel( + const T* prior_box_data, const T* prior_box_var_data, + const T* target_box_data, const int row, const int col, const int len, + const bool normalized, const T prior_box_var_size, const float* variance, + const int var_size, const int axis, T* output) { const int idx = threadIdx.x + blockIdx.x * blockDim.x; + int prior_box_offset = 0; if (idx < row * col) { const int col_idx = idx % col; - T prior_box_width = prior_box_data[col_idx * len + 2] - - prior_box_data[col_idx * len] + (normalized == false); - T prior_box_height = prior_box_data[col_idx * len + 3] - - prior_box_data[col_idx * len + 1] + + const int row_idx = idx / col; + prior_box_offset = axis == 0 ? col_idx * len : row_idx * len; + T prior_box_width = prior_box_data[prior_box_offset + 2] - + prior_box_data[prior_box_offset] + + (normalized == false); + T prior_box_height = prior_box_data[prior_box_offset + 3] - + prior_box_data[prior_box_offset + 1] + (normalized == false); T prior_box_center_x = - (prior_box_data[col_idx * len + 2] + prior_box_data[col_idx * len]) / 2; - T prior_box_center_y = (prior_box_data[col_idx * len + 3] + - prior_box_data[col_idx * len + 1]) / - 2; + prior_box_data[prior_box_offset] + prior_box_width / 2; + T prior_box_center_y = + prior_box_data[prior_box_offset + 1] + prior_box_height / 2; T target_box_width, target_box_height; T target_box_center_x, target_box_center_y; + T box_var_x = T(1), box_var_y = T(1); + T box_var_w = T(1), box_var_h = T(1); if (prior_box_var_data) { - target_box_width = exp(prior_box_var_data[col_idx * len + 2] * - target_box_data[idx * len + 2]) * - prior_box_width; - target_box_height = exp(prior_box_var_data[col_idx * len + 3] * - target_box_data[idx * len + 3]) * - prior_box_height; - target_box_center_x = prior_box_var_data[col_idx * len] * - target_box_data[idx * len] * prior_box_width + - prior_box_center_x; - target_box_center_y = prior_box_var_data[col_idx * len + 1] * - target_box_data[idx * len + 1] * - prior_box_height + - prior_box_center_y; - } else { - target_box_width = exp(target_box_data[idx * len + 2]) * prior_box_width; - target_box_height = - exp(target_box_data[idx * len + 3]) * prior_box_height; - target_box_center_x = - target_box_data[idx * len] * prior_box_width + prior_box_center_x; - target_box_center_y = target_box_data[idx * len + 1] * prior_box_height + - prior_box_center_y; + int prior_var_offset = axis == 0 ? col_idx * len : row_idx * len; + box_var_x = prior_box_var_data[prior_var_offset]; + box_var_y = prior_box_var_data[prior_var_offset + 1]; + box_var_w = prior_box_var_data[prior_var_offset + 2]; + box_var_h = prior_box_var_data[prior_var_offset + 3]; + } else if (var_size == 4) { + box_var_x = static_cast(variance[0]); + box_var_y = static_cast(variance[1]); + box_var_w = static_cast(variance[2]); + box_var_h = static_cast(variance[3]); } + target_box_width = + exp(box_var_w * target_box_data[idx * len + 2]) * prior_box_width; + target_box_height = + exp(box_var_h * target_box_data[idx * len + 3]) * prior_box_height; + target_box_center_x = + box_var_x * target_box_data[idx * len] * prior_box_width + + prior_box_center_x; + target_box_center_y = + box_var_y * target_box_data[idx * len + 1] * prior_box_height + + prior_box_center_y; output[idx * len] = target_box_center_x - target_box_width / 2; output[idx * len + 1] = target_box_center_y - target_box_height / 2; @@ -127,36 +137,64 @@ class BoxCoderCUDAKernel : public framework::OpKernel { auto* prior_box_var = context.Input("PriorBoxVar"); auto* target_box = context.Input("TargetBox"); auto* output_box = context.Output("OutputBox"); - + std::vector variance = context.Attr>("variance"); const T* prior_box_data = prior_box->data(); const T* target_box_data = target_box->data(); const T* prior_box_var_data = nullptr; - if (prior_box_var) prior_box_var_data = prior_box_var->data(); + auto prior_box_var_size = 0; + if (prior_box_var) { + PADDLE_ENFORCE(variance.empty(), + "Input 'PriorBoxVar' and attribute 'variance' should not" + "be used at the same time."); + prior_box_var_data = prior_box_var->data(); + prior_box_var_size = prior_box_var->dims().size(); + } + if (!(variance.empty())) { + PADDLE_ENFORCE(static_cast(variance.size()) == 4, + "Size of attribute 'variance' should be 4"); + } if (target_box->lod().size()) { PADDLE_ENFORCE_EQ(target_box->lod().size(), 1, "Only support 1 level of LoD."); } + const int var_size = static_cast(variance.size()); + + auto code_type = GetBoxCodeType(context.Attr("code_type")); + bool normalized = context.Attr("box_normalized"); + int axis = context.Attr("axis"); + auto row = target_box->dims()[0]; auto col = prior_box->dims()[0]; + if (code_type == BoxCodeType::kDecodeCenterSize) { + col = target_box->dims()[1]; + } auto len = prior_box->dims()[1]; int block = 512; int grid = (row * col + block - 1) / block; auto& device_ctx = context.cuda_device_context(); + auto& allocator = + platform::DeviceTemporaryAllocator::Instance().Get(device_ctx); + int bytes = var_size * sizeof(float); + auto dev_var = allocator.Allocate(bytes); + float* dev_var_data = reinterpret_cast(dev_var->ptr()); + auto cplace = platform::CPUPlace(); + const auto gplace = boost::get(context.GetPlace()); + memory::Copy(gplace, dev_var_data, cplace, &variance[0], bytes, + device_ctx.stream()); + output_box->mutable_data({row, col, len}, context.GetPlace()); T* output = output_box->data(); - auto code_type = GetBoxCodeType(context.Attr("code_type")); - bool normalized = context.Attr("box_normalized"); if (code_type == BoxCodeType::kEncodeCenterSize) { EncodeCenterSizeKernel<<>>( prior_box_data, prior_box_var_data, target_box_data, row, col, len, - normalized, output); + normalized, prior_box_var_size, dev_var_data, var_size, output); } else if (code_type == BoxCodeType::kDecodeCenterSize) { DecodeCenterSizeKernel<<>>( prior_box_data, prior_box_var_data, target_box_data, row, col, len, - normalized, output); + normalized, prior_box_var_size, dev_var_data, var_size, axis, output); } } }; diff --git a/paddle/fluid/operators/detection/box_coder_op.h b/paddle/fluid/operators/detection/box_coder_op.h index b2a2bcdce932032a761a1fc064fe622f7629f9bf..6d406f8196f9964c85bb94541fa7a7a23857539b 100644 --- a/paddle/fluid/operators/detection/box_coder_op.h +++ b/paddle/fluid/operators/detection/box_coder_op.h @@ -11,6 +11,7 @@ limitations under the License. */ #pragma once #include +#include #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/operators/math/math_function.h" @@ -34,7 +35,8 @@ class BoxCoderKernel : public framework::OpKernel { void EncodeCenterSize(const framework::Tensor* target_box, const framework::Tensor* prior_box, const framework::Tensor* prior_box_var, - const bool normalized, T* output) const { + const bool normalized, + const std::vector variance, T* output) const { int64_t row = target_box->dims()[0]; int64_t col = prior_box->dims()[0]; int64_t len = prior_box->dims()[1]; @@ -53,10 +55,9 @@ class BoxCoderKernel : public framework::OpKernel { T prior_box_height = prior_box_data[j * len + 3] - prior_box_data[j * len + 1] + (normalized == false); - T prior_box_center_x = - (prior_box_data[j * len + 2] + prior_box_data[j * len]) / 2; + T prior_box_center_x = prior_box_data[j * len] + prior_box_width / 2; T prior_box_center_y = - (prior_box_data[j * len + 3] + prior_box_data[j * len + 1]) / 2; + prior_box_data[j * len + 1] + prior_box_height / 2; T target_box_center_x = (target_box_data[i * len + 2] + target_box_data[i * len]) / 2; @@ -78,69 +79,78 @@ class BoxCoderKernel : public framework::OpKernel { output[offset + 3] = std::log(std::fabs(target_box_height / prior_box_height)); if (prior_box_var) { - output[offset] /= prior_box_var_data[j * len]; - output[offset + 1] /= prior_box_var_data[j * len + 1]; - output[offset + 2] /= prior_box_var_data[j * len + 2]; - output[offset + 3] /= prior_box_var_data[j * len + 3]; + int prior_var_offset = j * len; + output[offset] /= prior_box_var_data[prior_var_offset]; + output[offset + 1] /= prior_box_var_data[prior_var_offset + 1]; + output[offset + 2] /= prior_box_var_data[prior_var_offset + 2]; + output[offset + 3] /= prior_box_var_data[prior_var_offset + 3]; + } else if (!(variance.empty())) { + for (int k = 0; k < 4; ++k) { + output[offset + k] /= static_cast(variance[k]); + } } } } } + template void DecodeCenterSize(const framework::Tensor* target_box, const framework::Tensor* prior_box, const framework::Tensor* prior_box_var, - const bool normalized, T* output) const { + const bool normalized, std::vector variance, + T* output) const { int64_t row = target_box->dims()[0]; - int64_t col = prior_box->dims()[0]; - int64_t len = prior_box->dims()[1]; + int64_t col = target_box->dims()[1]; + int64_t len = target_box->dims()[2]; auto* target_box_data = target_box->data(); auto* prior_box_data = prior_box->data(); const T* prior_box_var_data = nullptr; - if (prior_box_var) prior_box_var_data = prior_box_var->data(); - + if (var_size == 2) prior_box_var_data = prior_box_var->data(); + int prior_box_offset = 0; + T var_data[4] = {1., 1., 1., 1.}; + T* var_ptr = var_data; #ifdef PADDLE_WITH_MKLML #pragma omp parallel for collapse(2) #endif for (int64_t i = 0; i < row; ++i) { for (int64_t j = 0; j < col; ++j) { size_t offset = i * col * len + j * len; - T prior_box_width = prior_box_data[j * len + 2] - - prior_box_data[j * len] + (normalized == false); - T prior_box_height = prior_box_data[j * len + 3] - - prior_box_data[j * len + 1] + + prior_box_offset = axis == 0 ? j * len : i * len; + T prior_box_width = prior_box_data[prior_box_offset + 2] - + prior_box_data[prior_box_offset] + + (normalized == false); + T prior_box_height = prior_box_data[prior_box_offset + 3] - + prior_box_data[prior_box_offset + 1] + (normalized == false); T prior_box_center_x = - (prior_box_data[j * len + 2] + prior_box_data[j * len]) / 2; + prior_box_data[prior_box_offset] + prior_box_width / 2; T prior_box_center_y = - (prior_box_data[j * len + 3] + prior_box_data[j * len + 1]) / 2; + prior_box_data[prior_box_offset + 1] + prior_box_height / 2; T target_box_center_x = 0, target_box_center_y = 0; T target_box_width = 0, target_box_height = 0; - if (prior_box_var) { - target_box_center_x = prior_box_var_data[j * len] * - target_box_data[offset] * prior_box_width + - prior_box_center_x; - target_box_center_y = prior_box_var_data[j * len + 1] * - target_box_data[offset + 1] * - prior_box_height + - prior_box_center_y; - target_box_width = std::exp(prior_box_var_data[j * len + 2] * - target_box_data[offset + 2]) * - prior_box_width; - target_box_height = std::exp(prior_box_var_data[j * len + 3] * - target_box_data[offset + 3]) * - prior_box_height; - } else { - target_box_center_x = - target_box_data[offset] * prior_box_width + prior_box_center_x; - target_box_center_y = target_box_data[offset + 1] * prior_box_height + - prior_box_center_y; - target_box_width = - std::exp(target_box_data[offset + 2]) * prior_box_width; - target_box_height = - std::exp(target_box_data[offset + 3]) * prior_box_height; + int prior_var_offset = axis == 0 ? j * len : i * len; + if (var_size == 2) { + std::memcpy(var_ptr, prior_box_var_data + prior_var_offset, + 4 * sizeof(T)); + } else if (var_size == 1) { + var_ptr = reinterpret_cast(variance.data()); } + T box_var_x = *var_ptr; + T box_var_y = *(var_ptr + 1); + T box_var_w = *(var_ptr + 2); + T box_var_h = *(var_ptr + 3); + + target_box_center_x = + box_var_x * target_box_data[offset] * prior_box_width + + prior_box_center_x; + target_box_center_y = + box_var_y * target_box_data[offset + 1] * prior_box_height + + prior_box_center_y; + target_box_width = + std::exp(box_var_w * target_box_data[offset + 2]) * prior_box_width; + target_box_height = std::exp(box_var_h * target_box_data[offset + 3]) * + prior_box_height; output[offset] = target_box_center_x - target_box_width / 2; output[offset + 1] = target_box_center_y - target_box_height / 2; @@ -157,26 +167,63 @@ class BoxCoderKernel : public framework::OpKernel { auto* prior_box_var = context.Input("PriorBoxVar"); auto* target_box = context.Input("TargetBox"); auto* output_box = context.Output("OutputBox"); - + std::vector variance = context.Attr>("variance"); + const int axis = context.Attr("axis"); if (target_box->lod().size()) { PADDLE_ENFORCE_EQ(target_box->lod().size(), 1UL, "Only support 1 level of LoD."); } + if (prior_box_var) { + PADDLE_ENFORCE(variance.empty(), + "Input 'PriorBoxVar' and attribute 'variance' should not" + "be used at the same time."); + } + if (!(variance.empty())) { + PADDLE_ENFORCE(static_cast(variance.size()) == 4, + "Size of attribute 'variance' should be 4"); + } + auto code_type = GetBoxCodeType(context.Attr("code_type")); + bool normalized = context.Attr("box_normalized"); + auto row = target_box->dims()[0]; auto col = prior_box->dims()[0]; + if (code_type == BoxCodeType::kDecodeCenterSize) { + col = target_box->dims()[1]; + } auto len = prior_box->dims()[1]; output_box->mutable_data({row, col, len}, context.GetPlace()); - auto code_type = GetBoxCodeType(context.Attr("code_type")); - bool normalized = context.Attr("box_normalized"); T* output = output_box->data(); if (code_type == BoxCodeType::kEncodeCenterSize) { EncodeCenterSize(target_box, prior_box, prior_box_var, normalized, - output); + variance, output); } else if (code_type == BoxCodeType::kDecodeCenterSize) { - DecodeCenterSize(target_box, prior_box, prior_box_var, normalized, - output); + if (prior_box_var) { + if (axis == 0) { + DecodeCenterSize<0, 2>(target_box, prior_box, prior_box_var, + normalized, variance, output); + } else { + DecodeCenterSize<1, 2>(target_box, prior_box, prior_box_var, + normalized, variance, output); + } + } else if (!(variance.empty())) { + if (axis == 0) { + DecodeCenterSize<0, 1>(target_box, prior_box, prior_box_var, + normalized, variance, output); + } else { + DecodeCenterSize<1, 1>(target_box, prior_box, prior_box_var, + normalized, variance, output); + } + } else { + if (axis == 0) { + DecodeCenterSize<0, 0>(target_box, prior_box, prior_box_var, + normalized, variance, output); + } else { + DecodeCenterSize<1, 0>(target_box, prior_box, prior_box_var, + normalized, variance, output); + } + } } } }; diff --git a/paddle/fluid/operators/detection/density_prior_box_op.h b/paddle/fluid/operators/detection/density_prior_box_op.h index ed2f5df80cf4d7a5a44af9b09f3b048b1b14cdb9..42137215e21af1a529563ecc995a54d610120beb 100644 --- a/paddle/fluid/operators/detection/density_prior_box_op.h +++ b/paddle/fluid/operators/detection/density_prior_box_op.h @@ -52,6 +52,10 @@ class DensityPriorBoxOpKernel : public framework::OpKernel { step_height = step_h; } int num_priors = 0; + +#ifdef PADDLE_WITH_MKLML +#pragma omp parallel for reduction(+ : num_priors) +#endif for (size_t i = 0; i < densities.size(); ++i) { num_priors += (fixed_ratios.size()) * (pow(densities[i], 2)); } @@ -64,6 +68,17 @@ class DensityPriorBoxOpKernel : public framework::OpKernel { auto e_boxes = framework::EigenTensor::From(*boxes).setConstant(0.0); int step_average = static_cast((step_width + step_height) * 0.5); + std::vector sqrt_fixed_ratios; +#ifdef PADDLE_WITH_MKLML +#pragma omp parallel for +#endif + for (size_t i = 0; i < fixed_ratios.size(); i++) { + sqrt_fixed_ratios.push_back(sqrt(fixed_ratios[i])); + } + +#ifdef PADDLE_WITH_MKLML +#pragma omp parallel for collapse(2) +#endif for (int h = 0; h < feature_height; ++h) { for (int w = 0; w < feature_width; ++w) { T center_x = (w + offset) * step_width; @@ -73,34 +88,25 @@ class DensityPriorBoxOpKernel : public framework::OpKernel { for (size_t s = 0; s < fixed_sizes.size(); ++s) { auto fixed_size = fixed_sizes[s]; int density = densities[s]; + int shift = step_average / density; // Generate density prior boxes with fixed ratios. for (size_t r = 0; r < fixed_ratios.size(); ++r) { - float ar = fixed_ratios[r]; - int shift = step_average / density; - float box_width_ratio = fixed_size * sqrt(ar); - float box_height_ratio = fixed_size / sqrt(ar); + float box_width_ratio = fixed_size * sqrt_fixed_ratios[r]; + float box_height_ratio = fixed_size / sqrt_fixed_ratios[r]; + float density_center_x = center_x - step_average / 2. + shift / 2.; + float density_center_y = center_y - step_average / 2. + shift / 2.; for (int di = 0; di < density; ++di) { for (int dj = 0; dj < density; ++dj) { - float center_x_temp = - center_x - step_average / 2. + shift / 2. + dj * shift; - float center_y_temp = - center_y - step_average / 2. + shift / 2. + di * shift; - e_boxes(h, w, idx, 0) = - (center_x_temp - box_width_ratio / 2.) / img_width >= 0 - ? (center_x_temp - box_width_ratio / 2.) / img_width - : 0; - e_boxes(h, w, idx, 1) = - (center_y_temp - box_height_ratio / 2.) / img_height >= 0 - ? (center_y_temp - box_height_ratio / 2.) / img_height - : 0; - e_boxes(h, w, idx, 2) = - (center_x_temp + box_width_ratio / 2.) / img_width <= 1 - ? (center_x_temp + box_width_ratio / 2.) / img_width - : 1; - e_boxes(h, w, idx, 3) = - (center_y_temp + box_height_ratio / 2.) / img_height <= 1 - ? (center_y_temp + box_height_ratio / 2.) / img_height - : 1; + float center_x_temp = density_center_x + dj * shift; + float center_y_temp = density_center_y + di * shift; + e_boxes(h, w, idx, 0) = std::max( + (center_x_temp - box_width_ratio / 2.) / img_width, 0.); + e_boxes(h, w, idx, 1) = std::max( + (center_y_temp - box_height_ratio / 2.) / img_height, 0.); + e_boxes(h, w, idx, 2) = std::min( + (center_x_temp + box_width_ratio / 2.) / img_width, 1.); + e_boxes(h, w, idx, 3) = std::min( + (center_y_temp + box_height_ratio / 2.) / img_height, 1.); idx++; } } @@ -109,11 +115,10 @@ class DensityPriorBoxOpKernel : public framework::OpKernel { } } if (clip) { - platform::Transform trans; - ClipFunctor clip_func; - trans(ctx.template device_context(), - boxes->data(), boxes->data() + boxes->numel(), - boxes->data(), clip_func); + T* dt = boxes->data(); + std::transform(dt, dt + boxes->numel(), dt, [](T v) -> T { + return std::min(std::max(v, 0.), 1.); + }); } framework::Tensor var_t; var_t.mutable_data( @@ -131,8 +136,14 @@ class DensityPriorBoxOpKernel : public framework::OpKernel { vars->Resize({box_num, static_cast(variances.size())}); auto e_vars = framework::EigenMatrix::From(*vars); - - e_vars = var_et.broadcast(Eigen::DSizes(box_num, 1)); +#ifdef PADDLE_WITH_MKLML +#pragma omp parallel for collapse(2) +#endif + for (int i = 0; i < box_num; ++i) { + for (size_t j = 0; j < variances.size(); ++j) { + e_vars(i, j) = variances[j]; + } + } vars->Resize(var_dim); boxes->Resize(box_dim); diff --git a/paddle/fluid/operators/detection/multiclass_nms_op.cc b/paddle/fluid/operators/detection/multiclass_nms_op.cc index 2395b181485429784e0f3dff6d056b84268ef245..f357e3ccf905309e6656f3fa87fbee45dc357c1e 100644 --- a/paddle/fluid/operators/detection/multiclass_nms_op.cc +++ b/paddle/fluid/operators/detection/multiclass_nms_op.cc @@ -9,9 +9,9 @@ http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - limitations under the License. */ +#include #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/operators/detection/poly_util.h" @@ -35,30 +35,45 @@ class MultiClassNMSOp : public framework::OperatorWithKernel { auto box_dims = ctx->GetInputDim("BBoxes"); auto score_dims = ctx->GetInputDim("Scores"); + auto score_size = score_dims.size(); if (ctx->IsRuntime()) { + PADDLE_ENFORCE(score_size == 2 || score_size == 3, + "The rank of Input(Scores) must be 2 or 3"); PADDLE_ENFORCE_EQ(box_dims.size(), 3, - "The rank of Input(BBoxes) must be 3."); - PADDLE_ENFORCE_EQ(score_dims.size(), 3, - "The rank of Input(Scores) must be 3."); - PADDLE_ENFORCE(box_dims[2] == 4 || box_dims[2] == 8 || - box_dims[2] == 16 || box_dims[2] == 24 || - box_dims[2] == 32, - "The 2nd dimension of Input(BBoxes) must be 4 or 8, " - "represents the layout of coordinate " - "[xmin, ymin, xmax, ymax] or " - "4 points: [x1, y1, x2, y2, x3, y3, x4, y4] or " - "8 points: [xi, yi] i= 1,2,...,8 or " - "12 points: [xi, yi] i= 1,2,...,12 or " - "16 points: [xi, yi] i= 1,2,...,16"); - PADDLE_ENFORCE_EQ(box_dims[1], score_dims[2], - "The 1st dimensiong of Input(BBoxes) must be equal to " - "3rd dimension of Input(Scores), which represents the " - "predicted bboxes."); + "The rank of Input(BBoxes) must be 3"); + if (score_size == 3) { + PADDLE_ENFORCE(box_dims[2] == 4 || box_dims[2] == 8 || + box_dims[2] == 16 || box_dims[2] == 24 || + box_dims[2] == 32, + "The last dimension of Input(BBoxes) must be 4 or 8, " + "represents the layout of coordinate " + "[xmin, ymin, xmax, ymax] or " + "4 points: [x1, y1, x2, y2, x3, y3, x4, y4] or " + "8 points: [xi, yi] i= 1,2,...,8 or " + "12 points: [xi, yi] i= 1,2,...,12 or " + "16 points: [xi, yi] i= 1,2,...,16"); + PADDLE_ENFORCE_EQ( + box_dims[1], score_dims[2], + "The 2nd dimension of Input(BBoxes) must be equal to " + "last dimension of Input(Scores), which represents the " + "predicted bboxes."); + } else { + PADDLE_ENFORCE(box_dims[2] == 4, + "The last dimension of Input(BBoxes) must be 4"); + PADDLE_ENFORCE_EQ(box_dims[1], score_dims[1], + "The 2nd dimension of Input(BBoxes)" + "must be equal to the 2nd dimension" + " of Input(Scores)"); + } } // Here the box_dims[0] is not the real dimension of output. // It will be rewritten in the computing kernel. - ctx->SetOutputDim("Out", {box_dims[1], box_dims[2] + 2}); + if (score_size == 3) { + ctx->SetOutputDim("Out", {box_dims[1], box_dims[2] + 2}); + } else { + ctx->SetOutputDim("Out", {-1, box_dims[2] + 2}); + } } protected: @@ -123,8 +138,9 @@ static inline T JaccardOverlap(const T* box1, const T* box2, const T inter_ymin = std::max(box1[1], box2[1]); const T inter_xmax = std::min(box1[2], box2[2]); const T inter_ymax = std::min(box1[3], box2[3]); - const T inter_w = inter_xmax - inter_xmin; - const T inter_h = inter_ymax - inter_ymin; + T norm = normalized ? static_cast(0.) : static_cast(1.); + T inter_w = inter_xmax - inter_xmin + norm; + T inter_h = inter_ymax - inter_ymin + norm; const T inter_area = inter_w * inter_h; const T bbox1_area = BBoxArea(box1, normalized); const T bbox2_area = BBoxArea(box2, normalized); @@ -139,7 +155,7 @@ T PolyIoU(const T* box1, const T* box2, const size_t box_size, T bbox2_area = PolyArea(box2, box_size, normalized); T inter_area = PolyOverlapArea(box1, box2, box_size, normalized); if (bbox1_area == 0 || bbox2_area == 0 || inter_area == 0) { - // If coordinate values are is invalid + // If coordinate values are invalid // if area size <= 0, return 0. return T(0.); } else { @@ -147,12 +163,35 @@ T PolyIoU(const T* box1, const T* box2, const size_t box_size, } } +template +void SliceOneClass(const platform::DeviceContext& ctx, + const framework::Tensor& items, const int class_id, + framework::Tensor* one_class_item) { + T* item_data = one_class_item->mutable_data(ctx.GetPlace()); + const T* items_data = items.data(); + const int64_t num_item = items.dims()[0]; + const int class_num = items.dims()[1]; + if (items.dims().size() == 3) { + int item_size = items.dims()[2]; + for (int i = 0; i < num_item; ++i) { + std::memcpy(item_data + i * item_size, + items_data + i * class_num * item_size + class_id * item_size, + sizeof(T) * item_size); + } + } else { + for (int i = 0; i < num_item; ++i) { + item_data[i] = items_data[i * class_num + class_id]; + } + } +} + template class MultiClassNMSKernel : public framework::OpKernel { public: void NMSFast(const Tensor& bbox, const Tensor& scores, const T score_threshold, const T nms_threshold, const T eta, - const int64_t top_k, std::vector* selected_indices) const { + const int64_t top_k, std::vector* selected_indices, + const bool normalized) const { // The total boxes for each instance. int64_t num_boxes = bbox.dims()[0]; // 4: [xmin ymin xmax ymax] @@ -178,15 +217,16 @@ class MultiClassNMSKernel : public framework::OpKernel { T overlap = T(0.); // 4: [xmin ymin xmax ymax] if (box_size == 4) { - overlap = JaccardOverlap(bbox_data + idx * box_size, - bbox_data + kept_idx * box_size, true); + overlap = + JaccardOverlap(bbox_data + idx * box_size, + bbox_data + kept_idx * box_size, normalized); } // 8: [x1 y1 x2 y2 x3 y3 x4 y4] or 16, 24, 32 if (box_size == 8 || box_size == 16 || box_size == 24 || box_size == 32) { - overlap = - PolyIoU(bbox_data + idx * box_size, - bbox_data + kept_idx * box_size, box_size, true); + overlap = PolyIoU(bbox_data + idx * box_size, + bbox_data + kept_idx * box_size, box_size, + normalized); } keep = overlap <= adaptive_threshold; } else { @@ -205,37 +245,58 @@ class MultiClassNMSKernel : public framework::OpKernel { void MultiClassNMS(const framework::ExecutionContext& ctx, const Tensor& scores, const Tensor& bboxes, + const int scores_size, std::map>* indices, int* num_nmsed_out) const { int64_t background_label = ctx.Attr("background_label"); int64_t nms_top_k = ctx.Attr("nms_top_k"); int64_t keep_top_k = ctx.Attr("keep_top_k"); + bool normalized = ctx.Attr("normalized"); T nms_threshold = static_cast(ctx.Attr("nms_threshold")); T nms_eta = static_cast(ctx.Attr("nms_eta")); T score_threshold = static_cast(ctx.Attr("score_threshold")); + auto& dev_ctx = ctx.template device_context(); - int64_t class_num = scores.dims()[0]; - int64_t predict_dim = scores.dims()[1]; int num_det = 0; + + int64_t class_num = scores_size == 3 ? scores.dims()[0] : scores.dims()[1]; + Tensor bbox_slice, score_slice; for (int64_t c = 0; c < class_num; ++c) { if (c == background_label) continue; - Tensor score = scores.Slice(c, c + 1); - NMSFast(bboxes, score, score_threshold, nms_threshold, nms_eta, nms_top_k, - &((*indices)[c])); + if (scores_size == 3) { + score_slice = scores.Slice(c, c + 1); + bbox_slice = bboxes; + } else { + score_slice.Resize({scores.dims()[0], 1}); + bbox_slice.Resize({scores.dims()[0], 4}); + SliceOneClass(dev_ctx, scores, c, &score_slice); + SliceOneClass(dev_ctx, bboxes, c, &bbox_slice); + } + NMSFast(bbox_slice, score_slice, score_threshold, nms_threshold, nms_eta, + nms_top_k, &((*indices)[c]), normalized); + if (scores_size == 2) { + std::stable_sort((*indices)[c].begin(), (*indices)[c].end()); + } num_det += (*indices)[c].size(); } *num_nmsed_out = num_det; const T* scores_data = scores.data(); if (keep_top_k > -1 && num_det > keep_top_k) { + const T* sdata; std::vector>> score_index_pairs; for (const auto& it : *indices) { int label = it.first; - const T* sdata = scores_data + label * predict_dim; + if (scores_size == 3) { + sdata = scores_data + label * scores.dims()[1]; + } else { + score_slice.Resize({scores.dims()[0], 1}); + SliceOneClass(dev_ctx, scores, label, &score_slice); + sdata = score_slice.data(); + } const std::vector& label_indices = it.second; for (size_t j = 0; j < label_indices.size(); ++j) { int idx = label_indices[j]; - PADDLE_ENFORCE_LT(idx, predict_dim); score_index_pairs.push_back( std::make_pair(sdata[idx], std::make_pair(label, idx))); } @@ -252,31 +313,55 @@ class MultiClassNMSKernel : public framework::OpKernel { int idx = score_index_pairs[j].second.second; new_indices[label].push_back(idx); } + if (scores_size == 2) { + for (const auto& it : new_indices) { + int label = it.first; + std::stable_sort(new_indices[label].begin(), + new_indices[label].end()); + } + } new_indices.swap(*indices); *num_nmsed_out = keep_top_k; } } - void MultiClassOutput(const Tensor& scores, const Tensor& bboxes, + void MultiClassOutput(const platform::DeviceContext& ctx, + const Tensor& scores, const Tensor& bboxes, const std::map>& selected_indices, - Tensor* outs) const { + const int scores_size, Tensor* outs) const { + int64_t class_num = scores.dims()[1]; int64_t predict_dim = scores.dims()[1]; int64_t box_size = bboxes.dims()[1]; - int64_t out_dim = bboxes.dims()[1] + 2; + if (scores_size == 2) { + box_size = bboxes.dims()[2]; + } + int64_t out_dim = box_size + 2; auto* scores_data = scores.data(); auto* bboxes_data = bboxes.data(); auto* odata = outs->data(); - + const T* sdata; + Tensor bbox; + bbox.Resize({scores.dims()[0], box_size}); int count = 0; for (const auto& it : selected_indices) { int label = it.first; - const T* sdata = scores_data + label * predict_dim; const std::vector& indices = it.second; + if (scores_size == 2) { + SliceOneClass(ctx, bboxes, label, &bbox); + } else { + sdata = scores_data + label * predict_dim; + } for (size_t j = 0; j < indices.size(); ++j) { int idx = indices[j]; - const T* bdata = bboxes_data + idx * box_size; - odata[count * out_dim] = label; // label - odata[count * out_dim + 1] = sdata[idx]; // score + odata[count * out_dim] = label; // label + const T* bdata; + if (scores_size == 3) { + bdata = bboxes_data + idx * box_size; + odata[count * out_dim + 1] = sdata[idx]; // score + } else { + bdata = bbox.data() + idx * box_size; + odata[count * out_dim + 1] = *(scores_data + idx * class_num + label); + } // xmin, ymin, xmax, ymax or multi-points coordinates std::memcpy(odata + count * out_dim + 2, bdata, box_size * sizeof(T)); count++; @@ -285,52 +370,64 @@ class MultiClassNMSKernel : public framework::OpKernel { } void Compute(const framework::ExecutionContext& ctx) const override { - auto* boxes = ctx.Input("BBoxes"); - auto* scores = ctx.Input("Scores"); + auto* boxes = ctx.Input("BBoxes"); + auto* scores = ctx.Input("Scores"); auto* outs = ctx.Output("Out"); auto score_dims = scores->dims(); - - int64_t batch_size = score_dims[0]; - int64_t class_num = score_dims[1]; - int64_t predict_dim = score_dims[2]; - int64_t box_dim = boxes->dims()[2]; - int64_t out_dim = boxes->dims()[2] + 2; + auto score_size = score_dims.size(); + auto& dev_ctx = ctx.template device_context(); std::vector>> all_indices; std::vector batch_starts = {0}; - for (int64_t i = 0; i < batch_size; ++i) { - Tensor ins_score = scores->Slice(i, i + 1); - ins_score.Resize({class_num, predict_dim}); - - Tensor ins_boxes = boxes->Slice(i, i + 1); - ins_boxes.Resize({predict_dim, box_dim}); - + int64_t batch_size = score_dims[0]; + int64_t box_dim = boxes->dims()[2]; + int64_t out_dim = box_dim + 2; + int num_nmsed_out = 0; + Tensor boxes_slice, scores_slice; + int n = score_size == 3 ? batch_size : boxes->lod().back().size() - 1; + for (int i = 0; i < n; ++i) { + if (score_size == 3) { + scores_slice = scores->Slice(i, i + 1); + scores_slice.Resize({score_dims[1], score_dims[2]}); + boxes_slice = boxes->Slice(i, i + 1); + boxes_slice.Resize({score_dims[2], box_dim}); + } else { + auto boxes_lod = boxes->lod().back(); + scores_slice = scores->Slice(boxes_lod[i], boxes_lod[i + 1]); + boxes_slice = boxes->Slice(boxes_lod[i], boxes_lod[i + 1]); + } std::map> indices; - int num_nmsed_out = 0; - MultiClassNMS(ctx, ins_score, ins_boxes, &indices, &num_nmsed_out); + MultiClassNMS(ctx, scores_slice, boxes_slice, score_size, &indices, + &num_nmsed_out); all_indices.push_back(indices); batch_starts.push_back(batch_starts.back() + num_nmsed_out); } int num_kept = batch_starts.back(); if (num_kept == 0) { - T* od = outs->mutable_data({1}, ctx.GetPlace()); + T* od = outs->mutable_data({1, 1}, ctx.GetPlace()); od[0] = -1; + batch_starts = {0, 1}; } else { outs->mutable_data({num_kept, out_dim}, ctx.GetPlace()); - for (int64_t i = 0; i < batch_size; ++i) { - Tensor ins_score = scores->Slice(i, i + 1); - ins_score.Resize({class_num, predict_dim}); - - Tensor ins_boxes = boxes->Slice(i, i + 1); - ins_boxes.Resize({predict_dim, box_dim}); - + for (int i = 0; i < n; ++i) { + if (score_size == 3) { + scores_slice = scores->Slice(i, i + 1); + boxes_slice = boxes->Slice(i, i + 1); + scores_slice.Resize({score_dims[1], score_dims[2]}); + boxes_slice.Resize({score_dims[2], box_dim}); + } else { + auto boxes_lod = boxes->lod().back(); + scores_slice = scores->Slice(boxes_lod[i], boxes_lod[i + 1]); + boxes_slice = boxes->Slice(boxes_lod[i], boxes_lod[i + 1]); + } int64_t s = batch_starts[i]; int64_t e = batch_starts[i + 1]; if (e > s) { Tensor out = outs->Slice(s, e); - MultiClassOutput(ins_score, ins_boxes, all_indices[i], &out); + MultiClassOutput(dev_ctx, scores_slice, boxes_slice, all_indices[i], + score_dims.size(), &out); } } } @@ -346,17 +443,24 @@ class MultiClassNMSOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { AddInput("BBoxes", - "(Tensor) A 3-D Tensor with shape " + "Two types of bboxes are supported:" + "1. (Tensor) A 3-D Tensor with shape " "[N, M, 4 or 8 16 24 32] represents the " "predicted locations of M bounding bboxes, N is the batch size. " "Each bounding box has four coordinate values and the layout is " - "[xmin, ymin, xmax, ymax], when box size equals to 4."); + "[xmin, ymin, xmax, ymax], when box size equals to 4." + "2. (LoDTensor) A 3-D Tensor with shape [M, C, 4]" + "M is the number of bounding boxes, C is the class number"); AddInput("Scores", - "(Tensor) A 3-D Tensor with shape [N, C, M] represents the " + "Two types of scores are supported:" + "1. (Tensor) A 3-D Tensor with shape [N, C, M] represents the " "predicted confidence predictions. N is the batch size, C is the " "class number, M is number of bounding boxes. For each category " "there are total M scores which corresponding M bounding boxes. " - " Please note, M is equal to the 1st dimension of BBoxes. "); + " Please note, M is equal to the 2nd dimension of BBoxes. " + "2. (LoDTensor) A 2-D LoDTensor with shape [M, C]. " + "M is the number of bbox, C is the class number. In this case, " + "Input BBoxes should be the second case with shape [M, C, 4]."); AddAttr( "background_label", "(int, defalut: 0) " @@ -384,6 +488,10 @@ class MultiClassNMSOpMaker : public framework::OpProtoAndCheckerMaker { "(int64_t) " "Number of total bboxes to be kept per image after NMS " "step. -1 means keeping all bboxes after NMS step."); + AddAttr("normalized", + "(bool, default true) " + "Whether detections are normalized.") + .SetDefault(true); AddOutput("Out", "(LoDTensor) A 2-D LoDTensor with shape [No, 6] represents the " "detections. Each row has 6 values: " @@ -399,24 +507,21 @@ class MultiClassNMSOpMaker : public framework::OpProtoAndCheckerMaker { AddComment(R"DOC( This operator is to do multi-class non maximum suppression (NMS) on a batched of boxes and scores. - In the NMS step, this operator greedily selects a subset of detection bounding boxes that have high scores larger than score_threshold, if providing this threshold, then selects the largest nms_top_k confidences scores if nms_top_k is larger than -1. Then this operator pruns away boxes that have high IOU (intersection over union) overlap with already selected boxes by adaptive threshold NMS based on parameters of nms_threshold and nms_eta. - Aftern NMS step, at most keep_top_k number of total bboxes are to be kept per image if keep_top_k is larger than -1. - This operator support multi-class and batched inputs. It applying NMS independently for each class. The outputs is a 2-D LoDTenosr, for each image, the offsets in first dimension of LoDTensor are called LoD, the number of offset is N + 1, where N is the batch size. If LoD[i + 1] - LoD[i] == 0, means there is no detected bbox for this image. If there is no detected boxes -for all images, all the elements in LoD are 0, and the Out only contains one -value which is -1. +for all images, all the elements in LoD are set to {1}, and the Out only +contains one value which is -1. )DOC"); } }; diff --git a/paddle/fluid/operators/detection/prior_box_op.h b/paddle/fluid/operators/detection/prior_box_op.h index 4e226abbb51c271502f0ca5419d488643b5a1a82..f84405664596ebe25983e5acbbb82bfc18c38124 100644 --- a/paddle/fluid/operators/detection/prior_box_op.h +++ b/paddle/fluid/operators/detection/prior_box_op.h @@ -46,13 +46,6 @@ inline void ExpandAspectRatios(const std::vector& input_aspect_ratior, } } -template -struct ClipFunctor { - HOSTDEVICE inline T operator()(T in) const { - return std::min(std::max(in, 0.), 1.); - } -}; - template class PriorBoxOpKernel : public framework::OpKernel { public: @@ -101,31 +94,30 @@ class PriorBoxOpKernel : public framework::OpKernel { boxes->mutable_data(ctx.GetPlace()); vars->mutable_data(ctx.GetPlace()); - auto e_boxes = framework::EigenTensor::From(*boxes); + T* b_t = boxes->data(); for (int h = 0; h < feature_height; ++h) { for (int w = 0; w < feature_width; ++w) { T center_x = (w + offset) * step_width; T center_y = (h + offset) * step_height; T box_width, box_height; - int idx = 0; for (size_t s = 0; s < min_sizes.size(); ++s) { auto min_size = min_sizes[s]; if (min_max_aspect_ratios_order) { box_width = box_height = min_size / 2.; - e_boxes(h, w, idx, 0) = (center_x - box_width) / img_width; - e_boxes(h, w, idx, 1) = (center_y - box_height) / img_height; - e_boxes(h, w, idx, 2) = (center_x + box_width) / img_width; - e_boxes(h, w, idx, 3) = (center_y + box_height) / img_height; - idx++; + b_t[0] = (center_x - box_width) / img_width; + b_t[1] = (center_y - box_height) / img_height; + b_t[2] = (center_x + box_width) / img_width; + b_t[3] = (center_y + box_height) / img_height; + b_t += 4; if (max_sizes.size() > 0) { auto max_size = max_sizes[s]; // square prior with size sqrt(minSize * maxSize) box_width = box_height = sqrt(min_size * max_size) / 2.; - e_boxes(h, w, idx, 0) = (center_x - box_width) / img_width; - e_boxes(h, w, idx, 1) = (center_y - box_height) / img_height; - e_boxes(h, w, idx, 2) = (center_x + box_width) / img_width; - e_boxes(h, w, idx, 3) = (center_y + box_height) / img_height; - idx++; + b_t[0] = (center_x - box_width) / img_width; + b_t[1] = (center_y - box_height) / img_height; + b_t[2] = (center_x + box_width) / img_width; + b_t[3] = (center_y + box_height) / img_height; + b_t += 4; } // priors with different aspect ratios for (size_t r = 0; r < aspect_ratios.size(); ++r) { @@ -135,11 +127,11 @@ class PriorBoxOpKernel : public framework::OpKernel { } box_width = min_size * sqrt(ar) / 2.; box_height = min_size / sqrt(ar) / 2.; - e_boxes(h, w, idx, 0) = (center_x - box_width) / img_width; - e_boxes(h, w, idx, 1) = (center_y - box_height) / img_height; - e_boxes(h, w, idx, 2) = (center_x + box_width) / img_width; - e_boxes(h, w, idx, 3) = (center_y + box_height) / img_height; - idx++; + b_t[0] = (center_x - box_width) / img_width; + b_t[1] = (center_y - box_height) / img_height; + b_t[2] = (center_x + box_width) / img_width; + b_t[3] = (center_y + box_height) / img_height; + b_t += 4; } } else { // priors with different aspect ratios @@ -147,21 +139,21 @@ class PriorBoxOpKernel : public framework::OpKernel { float ar = aspect_ratios[r]; box_width = min_size * sqrt(ar) / 2.; box_height = min_size / sqrt(ar) / 2.; - e_boxes(h, w, idx, 0) = (center_x - box_width) / img_width; - e_boxes(h, w, idx, 1) = (center_y - box_height) / img_height; - e_boxes(h, w, idx, 2) = (center_x + box_width) / img_width; - e_boxes(h, w, idx, 3) = (center_y + box_height) / img_height; - idx++; + b_t[0] = (center_x - box_width) / img_width; + b_t[1] = (center_y - box_height) / img_height; + b_t[2] = (center_x + box_width) / img_width; + b_t[3] = (center_y + box_height) / img_height; + b_t += 4; } if (max_sizes.size() > 0) { auto max_size = max_sizes[s]; // square prior with size sqrt(minSize * maxSize) box_width = box_height = sqrt(min_size * max_size) / 2.; - e_boxes(h, w, idx, 0) = (center_x - box_width) / img_width; - e_boxes(h, w, idx, 1) = (center_y - box_height) / img_height; - e_boxes(h, w, idx, 2) = (center_x + box_width) / img_width; - e_boxes(h, w, idx, 3) = (center_y + box_height) / img_height; - idx++; + b_t[0] = (center_x - box_width) / img_width; + b_t[1] = (center_y - box_height) / img_height; + b_t[2] = (center_x + box_width) / img_width; + b_t[3] = (center_y + box_height) / img_height; + b_t += 4; } } } @@ -169,11 +161,10 @@ class PriorBoxOpKernel : public framework::OpKernel { } if (clip) { - platform::Transform trans; - ClipFunctor clip_func; - trans(ctx.template device_context(), - boxes->data(), boxes->data() + boxes->numel(), - boxes->data(), clip_func); + T* dt = boxes->data(); + std::transform(dt, dt + boxes->numel(), dt, [](T v) -> T { + return std::min(std::max(v, 0.), 1.); + }); } framework::Tensor var_t; diff --git a/paddle/fluid/operators/yolov3_loss_op.cc b/paddle/fluid/operators/detection/yolov3_loss_op.cc similarity index 60% rename from paddle/fluid/operators/yolov3_loss_op.cc rename to paddle/fluid/operators/detection/yolov3_loss_op.cc index 60508f7ab871910c38f1e4aa04c2035075d37df5..ab01bdf7ca8c5a369bd8838b1acc734364666992 100644 --- a/paddle/fluid/operators/yolov3_loss_op.cc +++ b/paddle/fluid/operators/detection/yolov3_loss_op.cc @@ -9,7 +9,7 @@ See the License for the specific language governing permissions and limitations under the License. */ -#include "paddle/fluid/operators/yolov3_loss_op.h" +#include "paddle/fluid/operators/detection/yolov3_loss_op.h" #include "paddle/fluid/framework/op_registry.h" namespace paddle { @@ -29,23 +29,33 @@ class Yolov3LossOp : public framework::OperatorWithKernel { "Input(GTLabel) of Yolov3LossOp should not be null."); PADDLE_ENFORCE(ctx->HasOutput("Loss"), "Output(Loss) of Yolov3LossOp should not be null."); + PADDLE_ENFORCE( + ctx->HasOutput("ObjectnessMask"), + "Output(ObjectnessMask) of Yolov3LossOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("GTMatchMask"), + "Output(GTMatchMask) of Yolov3LossOp should not be null."); auto dim_x = ctx->GetInputDim("X"); auto dim_gtbox = ctx->GetInputDim("GTBox"); auto dim_gtlabel = ctx->GetInputDim("GTLabel"); auto anchors = ctx->Attrs().Get>("anchors"); + int anchor_num = anchors.size() / 2; + auto anchor_mask = ctx->Attrs().Get>("anchor_mask"); + int mask_num = anchor_mask.size(); auto class_num = ctx->Attrs().Get("class_num"); + PADDLE_ENFORCE_EQ(dim_x.size(), 4, "Input(X) should be a 4-D tensor."); PADDLE_ENFORCE_EQ(dim_x[2], dim_x[3], "Input(X) dim[3] and dim[4] should be euqal."); - PADDLE_ENFORCE_EQ(dim_x[1], anchors.size() / 2 * (5 + class_num), - "Input(X) dim[1] should be equal to (anchor_number * (5 " - "+ class_num))."); + PADDLE_ENFORCE_EQ( + dim_x[1], mask_num * (5 + class_num), + "Input(X) dim[1] should be equal to (anchor_mask_number * (5 " + "+ class_num))."); PADDLE_ENFORCE_EQ(dim_gtbox.size(), 3, "Input(GTBox) should be a 3-D tensor"); PADDLE_ENFORCE_EQ(dim_gtbox[2], 4, "Input(GTBox) dim[2] should be 5"); PADDLE_ENFORCE_EQ(dim_gtlabel.size(), 2, - "Input(GTBox) should be a 2-D tensor"); + "Input(GTLabel) should be a 2-D tensor"); PADDLE_ENFORCE_EQ(dim_gtlabel[0], dim_gtbox[0], "Input(GTBox) and Input(GTLabel) dim[0] should be same"); PADDLE_ENFORCE_EQ(dim_gtlabel[1], dim_gtbox[1], @@ -54,11 +64,22 @@ class Yolov3LossOp : public framework::OperatorWithKernel { "Attr(anchors) length should be greater then 0."); PADDLE_ENFORCE_EQ(anchors.size() % 2, 0, "Attr(anchors) length should be even integer."); + for (size_t i = 0; i < anchor_mask.size(); i++) { + PADDLE_ENFORCE_LT( + anchor_mask[i], anchor_num, + "Attr(anchor_mask) should not crossover Attr(anchors)."); + } PADDLE_ENFORCE_GT(class_num, 0, "Attr(class_num) should be an integer greater then 0."); - std::vector dim_out({1}); + std::vector dim_out({dim_x[0]}); ctx->SetOutputDim("Loss", framework::make_ddim(dim_out)); + + std::vector dim_obj_mask({dim_x[0], mask_num, dim_x[2], dim_x[3]}); + ctx->SetOutputDim("ObjectnessMask", framework::make_ddim(dim_obj_mask)); + + std::vector dim_gt_match_mask({dim_gtbox[0], dim_gtbox[1]}); + ctx->SetOutputDim("GTMatchMask", framework::make_ddim(dim_gt_match_mask)); } protected: @@ -73,11 +94,11 @@ class Yolov3LossOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { AddInput("X", - "The input tensor of YOLO v3 loss operator, " + "The input tensor of YOLOv3 loss operator, " "This is a 4-D tensor with shape of [N, C, H, W]." "H and W should be same, and the second dimention(C) stores" "box locations, confidence score and classification one-hot" - "key of each anchor box"); + "keys of each anchor box"); AddInput("GTBox", "The input tensor of ground truth boxes, " "This is a 3-D tensor with shape of [N, max_box_num, 5], " @@ -89,75 +110,99 @@ class Yolov3LossOpMaker : public framework::OpProtoAndCheckerMaker { AddInput("GTLabel", "The input tensor of ground truth label, " "This is a 2-D tensor with shape of [N, max_box_num], " - "and each element shoudl be an integer to indicate the " + "and each element should be an integer to indicate the " "box class id."); AddOutput("Loss", "The output yolov3 loss tensor, " - "This is a 1-D tensor with shape of [1]"); + "This is a 1-D tensor with shape of [N]"); + AddOutput("ObjectnessMask", + "This is an intermediate tensor with shape of [N, M, H, W], " + "M is the number of anchor masks. This parameter caches the " + "mask for calculate objectness loss in gradient kernel.") + .AsIntermediate(); + AddOutput("GTMatchMask", + "This is an intermediate tensor with shape of [N, B], " + "B is the max box number of GT boxes. This parameter caches " + "matched mask index of each GT boxes for gradient calculate.") + .AsIntermediate(); AddAttr("class_num", "The number of classes to predict."); AddAttr>("anchors", "The anchor width and height, " - "it will be parsed pair by pair."); + "it will be parsed pair by pair.") + .SetDefault(std::vector{}); + AddAttr>("anchor_mask", + "The mask index of anchors used in " + "current YOLOv3 loss calculation.") + .SetDefault(std::vector{}); + AddAttr("downsample_ratio", + "The downsample ratio from network input to YOLOv3 loss " + "input, so 32, 16, 8 should be set for the first, second, " + "and thrid YOLOv3 loss operators.") + .SetDefault(32); AddAttr("ignore_thresh", - "The ignore threshold to ignore confidence loss."); - AddAttr("loss_weight_xy", "The weight of x, y location loss.") - .SetDefault(1.0); - AddAttr("loss_weight_wh", "The weight of w, h location loss.") - .SetDefault(1.0); - AddAttr( - "loss_weight_conf_target", - "The weight of confidence score loss in locations with target object.") - .SetDefault(1.0); - AddAttr("loss_weight_conf_notarget", - "The weight of confidence score loss in locations without " - "target object.") - .SetDefault(1.0); - AddAttr("loss_weight_class", "The weight of classification loss.") - .SetDefault(1.0); + "The ignore threshold to ignore confidence loss.") + .SetDefault(0.7); AddComment(R"DOC( - This operator generate yolov3 loss by given predict result and ground + This operator generates yolov3 loss based on given predict result and ground truth boxes. The output of previous network is in shape [N, C, H, W], while H and W - should be the same, specify the grid size, each grid point predict given - number boxes, this given number is specified by anchors, it should be - half anchors length, which following will be represented as S. In the - second dimention(the channel dimention), C should be S * (class_num + 5), - class_num is the box categoriy number of source dataset(such as coco), - so in the second dimention, stores 4 box location coordinates x, y, w, h - and confidence score of the box and class one-hot key of each anchor box. + should be the same, H and W specify the grid size, each grid point predict + given number boxes, this given number, which following will be represented as S, + is specified by the number of anchors, In the second dimension(the channel + dimension), C should be equal to S * (class_num + 5), class_num is the object + category number of source dataset(such as 80 in coco dataset), so in the + second(channel) dimension, apart from 4 box location coordinates x, y, w, h, + also includes confidence score of the box and class one-hot key of each anchor box. - While the 4 location coordinates if $$tx, ty, tw, th$$, the box predictions - correspnd to: + Assume the 4 location coordinates are :math:`t_x, t_y, t_w, t_h`, the box predictions + should be as follows: $$ - b_x = \sigma(t_x) + c_x - b_y = \sigma(t_y) + c_y + b_x = \\sigma(t_x) + c_x + $$ + $$ + b_y = \\sigma(t_y) + c_y + $$ + $$ b_w = p_w e^{t_w} + $$ + $$ b_h = p_h e^{t_h} $$ - While $$c_x, c_y$$ is the left top corner of current grid and $$p_w, p_h$$ - is specified by anchors. + In the equation above, :math:`c_x, c_y` is the left top corner of current grid + and :math:`p_w, p_h` is specified by anchors. As for confidence score, it is the logistic regression value of IoU between anchor boxes and ground truth boxes, the score of the anchor box which has - the max IoU should be 1, and if the anchor box has IoU bigger then ignore + the max IoU should be 1, and if the anchor box has IoU bigger than ignore thresh, the confidence score loss of this anchor box will be ignored. Therefore, the yolov3 loss consist of three major parts, box location loss, - confidence score loss, and classification loss. The MSE loss is used for - box location, and binary cross entropy loss is used for confidence score - loss and classification loss. + confidence score loss, and classification loss. The L2 loss is used for + box coordinates (w, h), and sigmoid cross entropy loss is used for box + coordinates (x, y), confidence score loss and classification loss. + + Each groud truth box find a best matching anchor box in all anchors, + prediction of this anchor box will incur all three parts of losses, and + prediction of anchor boxes with no GT box matched will only incur objectness + loss. + + In order to trade off box coordinate losses between big boxes and small + boxes, box coordinate losses will be mutiplied by scale weight, which is + calculated as follows. + + $$ + weight_{box} = 2.0 - t_w * t_h + $$ - Final loss will be represented as follow. + Final loss will be represented as follows. $$ - loss = \loss_weight_{xy} * loss_{xy} + \loss_weight_{wh} * loss_{wh} - + \loss_weight_{conf_target} * loss_{conf_target} - + \loss_weight_{conf_notarget} * loss_{conf_notarget} - + \loss_weight_{class} * loss_{class} + loss = (loss_{xy} + loss_{wh}) * weight_{box} + + loss_{conf} + loss_{class} $$ )DOC"); } @@ -196,6 +241,8 @@ class Yolov3LossGradMaker : public framework::SingleGradOpDescMaker { op->SetInput("GTBox", Input("GTBox")); op->SetInput("GTLabel", Input("GTLabel")); op->SetInput(framework::GradVarName("Loss"), OutputGrad("Loss")); + op->SetInput("ObjectnessMask", Output("ObjectnessMask")); + op->SetInput("GTMatchMask", Output("GTMatchMask")); op->SetAttrMap(Attrs()); diff --git a/paddle/fluid/operators/detection/yolov3_loss_op.h b/paddle/fluid/operators/detection/yolov3_loss_op.h new file mode 100644 index 0000000000000000000000000000000000000000..8407d4e6e8f87a2e8d073c4fbda5691abe1bba68 --- /dev/null +++ b/paddle/fluid/operators/detection/yolov3_loss_op.h @@ -0,0 +1,447 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve. + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + http://www.apache.org/licenses/LICENSE-2.0 + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once +#include +#include +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/math/math_function.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; +template +using EigenTensor = framework::EigenTensor; +template +using EigenVector = framework::EigenVector; + +template +static inline bool LessEqualZero(T x) { + return x < 1e-6; +} + +template +static T SigmoidCrossEntropy(T x, T label) { + return (x > 0 ? x : 0.0) - x * label + std::log(1.0 + std::exp(-std::abs(x))); +} + +template +static T L2Loss(T x, T y) { + return 0.5 * (y - x) * (y - x); +} + +template +static T SigmoidCrossEntropyGrad(T x, T label) { + return 1.0 / (1.0 + std::exp(-x)) - label; +} + +template +static T L2LossGrad(T x, T y) { + return x - y; +} + +static int GetMaskIndex(std::vector mask, int val) { + for (size_t i = 0; i < mask.size(); i++) { + if (mask[i] == val) { + return i; + } + } + return -1; +} + +template +struct Box { + T x, y, w, h; +}; + +template +static inline T sigmoid(T x) { + return 1.0 / (1.0 + std::exp(-x)); +} + +template +static inline Box GetYoloBox(const T* x, std::vector anchors, int i, + int j, int an_idx, int grid_size, + int input_size, int index, int stride) { + Box b; + b.x = (i + sigmoid(x[index])) / grid_size; + b.y = (j + sigmoid(x[index + stride])) / grid_size; + b.w = std::exp(x[index + 2 * stride]) * anchors[2 * an_idx] / input_size; + b.h = std::exp(x[index + 3 * stride]) * anchors[2 * an_idx + 1] / input_size; + return b; +} + +template +static inline Box GetGtBox(const T* gt, int batch, int max_boxes, int idx) { + Box b; + b.x = gt[(batch * max_boxes + idx) * 4]; + b.y = gt[(batch * max_boxes + idx) * 4 + 1]; + b.w = gt[(batch * max_boxes + idx) * 4 + 2]; + b.h = gt[(batch * max_boxes + idx) * 4 + 3]; + return b; +} + +template +static inline T BoxOverlap(T c1, T w1, T c2, T w2) { + T l1 = c1 - w1 / 2.0; + T l2 = c2 - w2 / 2.0; + T left = l1 > l2 ? l1 : l2; + T r1 = c1 + w1 / 2.0; + T r2 = c2 + w2 / 2.0; + T right = r1 < r2 ? r1 : r2; + return right - left; +} + +template +static inline T CalcBoxIoU(Box b1, Box b2) { + T w = BoxOverlap(b1.x, b1.w, b2.x, b2.w); + T h = BoxOverlap(b1.y, b1.h, b2.y, b2.h); + T inter_area = (w < 0 || h < 0) ? 0.0 : w * h; + T union_area = b1.w * b1.h + b2.w * b2.h - inter_area; + return inter_area / union_area; +} + +static inline int GetEntryIndex(int batch, int an_idx, int hw_idx, int an_num, + int an_stride, int stride, int entry) { + return (batch * an_num + an_idx) * an_stride + entry * stride + hw_idx; +} + +template +static void CalcBoxLocationLoss(T* loss, const T* input, Box gt, + std::vector anchors, int an_idx, + int box_idx, int gi, int gj, int grid_size, + int input_size, int stride) { + T tx = gt.x * grid_size - gi; + T ty = gt.y * grid_size - gj; + T tw = std::log(gt.w * input_size / anchors[2 * an_idx]); + T th = std::log(gt.h * input_size / anchors[2 * an_idx + 1]); + + T scale = (2.0 - gt.w * gt.h); + loss[0] += SigmoidCrossEntropy(input[box_idx], tx) * scale; + loss[0] += SigmoidCrossEntropy(input[box_idx + stride], ty) * scale; + loss[0] += L2Loss(input[box_idx + 2 * stride], tw) * scale; + loss[0] += L2Loss(input[box_idx + 3 * stride], th) * scale; +} + +template +static void CalcBoxLocationLossGrad(T* input_grad, const T loss, const T* input, + Box gt, std::vector anchors, + int an_idx, int box_idx, int gi, int gj, + int grid_size, int input_size, int stride) { + T tx = gt.x * grid_size - gi; + T ty = gt.y * grid_size - gj; + T tw = std::log(gt.w * input_size / anchors[2 * an_idx]); + T th = std::log(gt.h * input_size / anchors[2 * an_idx + 1]); + + T scale = (2.0 - gt.w * gt.h); + input_grad[box_idx] = + SigmoidCrossEntropyGrad(input[box_idx], tx) * scale * loss; + input_grad[box_idx + stride] = + SigmoidCrossEntropyGrad(input[box_idx + stride], ty) * scale * loss; + input_grad[box_idx + 2 * stride] = + L2LossGrad(input[box_idx + 2 * stride], tw) * scale * loss; + input_grad[box_idx + 3 * stride] = + L2LossGrad(input[box_idx + 3 * stride], th) * scale * loss; +} + +template +static inline void CalcLabelLoss(T* loss, const T* input, const int index, + const int label, const int class_num, + const int stride) { + for (int i = 0; i < class_num; i++) { + T pred = input[index + i * stride]; + loss[0] += SigmoidCrossEntropy(pred, (i == label) ? 1.0 : 0.0); + } +} + +template +static inline void CalcLabelLossGrad(T* input_grad, const T loss, + const T* input, const int index, + const int label, const int class_num, + const int stride) { + for (int i = 0; i < class_num; i++) { + T pred = input[index + i * stride]; + input_grad[index + i * stride] = + SigmoidCrossEntropyGrad(pred, (i == label) ? 1.0 : 0.0) * loss; + } +} + +template +static inline void CalcObjnessLoss(T* loss, const T* input, const T* objness, + const int n, const int an_num, const int h, + const int w, const int stride, + const int an_stride) { + for (int i = 0; i < n; i++) { + for (int j = 0; j < an_num; j++) { + for (int k = 0; k < h; k++) { + for (int l = 0; l < w; l++) { + T obj = objness[k * w + l]; + if (obj > 1e-5) { + // positive sample: obj = 1 + loss[i] += SigmoidCrossEntropy(input[k * w + l], 1.0); + } else if (obj > -0.5) { + // negetive sample: obj = 0 + loss[i] += SigmoidCrossEntropy(input[k * w + l], 0.0); + } + } + } + objness += stride; + input += an_stride; + } + } +} + +template +static inline void CalcObjnessLossGrad(T* input_grad, const T* loss, + const T* input, const T* objness, + const int n, const int an_num, + const int h, const int w, + const int stride, const int an_stride) { + for (int i = 0; i < n; i++) { + for (int j = 0; j < an_num; j++) { + for (int k = 0; k < h; k++) { + for (int l = 0; l < w; l++) { + T obj = objness[k * w + l]; + if (obj > 1e-5) { + input_grad[k * w + l] = + SigmoidCrossEntropyGrad(input[k * w + l], 1.0) * loss[i]; + } else if (obj > -0.5) { + input_grad[k * w + l] = + SigmoidCrossEntropyGrad(input[k * w + l], 0.0) * loss[i]; + } + } + } + objness += stride; + input += an_stride; + input_grad += an_stride; + } + } +} + +template +static void inline GtValid(bool* valid, const T* gtbox, const int n, + const int b) { + for (int i = 0; i < n; i++) { + for (int j = 0; j < b; j++) { + if (LessEqualZero(gtbox[j * 4 + 2]) || LessEqualZero(gtbox[j * 4 + 3])) { + valid[j] = false; + } else { + valid[j] = true; + } + } + valid += b; + gtbox += b * 4; + } +} + +template +class Yolov3LossKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* input = ctx.Input("X"); + auto* gt_box = ctx.Input("GTBox"); + auto* gt_label = ctx.Input("GTLabel"); + auto* loss = ctx.Output("Loss"); + auto* objness_mask = ctx.Output("ObjectnessMask"); + auto* gt_match_mask = ctx.Output("GTMatchMask"); + auto anchors = ctx.Attr>("anchors"); + auto anchor_mask = ctx.Attr>("anchor_mask"); + int class_num = ctx.Attr("class_num"); + float ignore_thresh = ctx.Attr("ignore_thresh"); + int downsample_ratio = ctx.Attr("downsample_ratio"); + + const int n = input->dims()[0]; + const int h = input->dims()[2]; + const int w = input->dims()[3]; + const int an_num = anchors.size() / 2; + const int mask_num = anchor_mask.size(); + const int b = gt_box->dims()[1]; + int input_size = downsample_ratio * h; + + const int stride = h * w; + const int an_stride = (class_num + 5) * stride; + + const T* input_data = input->data(); + const T* gt_box_data = gt_box->data(); + const int* gt_label_data = gt_label->data(); + T* loss_data = loss->mutable_data({n}, ctx.GetPlace()); + memset(loss_data, 0, loss->numel() * sizeof(T)); + T* obj_mask_data = + objness_mask->mutable_data({n, mask_num, h, w}, ctx.GetPlace()); + memset(obj_mask_data, 0, objness_mask->numel() * sizeof(T)); + int* gt_match_mask_data = + gt_match_mask->mutable_data({n, b}, ctx.GetPlace()); + + // calc valid gt box mask, avoid calc duplicately in following code + Tensor gt_valid_mask; + bool* gt_valid_mask_data = + gt_valid_mask.mutable_data({n, b}, ctx.GetPlace()); + GtValid(gt_valid_mask_data, gt_box_data, n, b); + + for (int i = 0; i < n; i++) { + for (int j = 0; j < mask_num; j++) { + for (int k = 0; k < h; k++) { + for (int l = 0; l < w; l++) { + // each predict box find a best match gt box, if overlap is bigger + // then ignore_thresh, ignore the objectness loss. + int box_idx = + GetEntryIndex(i, j, k * w + l, mask_num, an_stride, stride, 0); + Box pred = GetYoloBox(input_data, anchors, l, k, anchor_mask[j], + h, input_size, box_idx, stride); + T best_iou = 0; + for (int t = 0; t < b; t++) { + if (!gt_valid_mask_data[i * b + t]) { + continue; + } + Box gt = GetGtBox(gt_box_data, i, b, t); + T iou = CalcBoxIoU(pred, gt); + if (iou > best_iou) { + best_iou = iou; + } + } + + // If best IoU is bigger then ignore_thresh, + // ignore the objectness loss. + if (best_iou > ignore_thresh) { + int obj_idx = (i * mask_num + j) * stride + k * w + l; + obj_mask_data[obj_idx] = static_cast(-1); + } + // all losses should be calculated if best IoU + // is bigger then truth thresh, but currently, + // truth thresh is an unreachable value as 1.0. + } + } + } + for (int t = 0; t < b; t++) { + if (!gt_valid_mask_data[i * b + t]) { + gt_match_mask_data[i * b + t] = -1; + continue; + } + Box gt = GetGtBox(gt_box_data, i, b, t); + int gi = static_cast(gt.x * w); + int gj = static_cast(gt.y * h); + Box gt_shift = gt; + gt_shift.x = 0.0; + gt_shift.y = 0.0; + T best_iou = 0.0; + int best_n = 0; + // each gt box find a best match anchor box as positive sample, + // for positive sample, all losses should be calculated, and for + // other samples, only objectness loss is required. + for (int an_idx = 0; an_idx < an_num; an_idx++) { + Box an_box; + an_box.x = 0.0; + an_box.y = 0.0; + an_box.w = anchors[2 * an_idx] / static_cast(input_size); + an_box.h = anchors[2 * an_idx + 1] / static_cast(input_size); + float iou = CalcBoxIoU(an_box, gt_shift); + if (iou > best_iou) { + best_iou = iou; + best_n = an_idx; + } + } + + int mask_idx = GetMaskIndex(anchor_mask, best_n); + gt_match_mask_data[i * b + t] = mask_idx; + if (mask_idx >= 0) { + int box_idx = GetEntryIndex(i, mask_idx, gj * w + gi, mask_num, + an_stride, stride, 0); + CalcBoxLocationLoss(loss_data + i, input_data, gt, anchors, best_n, + box_idx, gi, gj, h, input_size, stride); + + int obj_idx = (i * mask_num + mask_idx) * stride + gj * w + gi; + obj_mask_data[obj_idx] = 1.0; + + int label = gt_label_data[i * b + t]; + int label_idx = GetEntryIndex(i, mask_idx, gj * w + gi, mask_num, + an_stride, stride, 5); + CalcLabelLoss(loss_data + i, input_data, label_idx, label, + class_num, stride); + } + } + } + + CalcObjnessLoss(loss_data, input_data + 4 * stride, obj_mask_data, n, + mask_num, h, w, stride, an_stride); + } +}; + +template +class Yolov3LossGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* input = ctx.Input("X"); + auto* gt_box = ctx.Input("GTBox"); + auto* gt_label = ctx.Input("GTLabel"); + auto* input_grad = ctx.Output(framework::GradVarName("X")); + auto* loss_grad = ctx.Input(framework::GradVarName("Loss")); + auto* objness_mask = ctx.Input("ObjectnessMask"); + auto* gt_match_mask = ctx.Input("GTMatchMask"); + auto anchors = ctx.Attr>("anchors"); + auto anchor_mask = ctx.Attr>("anchor_mask"); + int class_num = ctx.Attr("class_num"); + int downsample_ratio = ctx.Attr("downsample_ratio"); + + const int n = input_grad->dims()[0]; + const int c = input_grad->dims()[1]; + const int h = input_grad->dims()[2]; + const int w = input_grad->dims()[3]; + const int mask_num = anchor_mask.size(); + const int b = gt_match_mask->dims()[1]; + int input_size = downsample_ratio * h; + + const int stride = h * w; + const int an_stride = (class_num + 5) * stride; + + const T* input_data = input->data(); + const T* gt_box_data = gt_box->data(); + const int* gt_label_data = gt_label->data(); + const T* loss_grad_data = loss_grad->data(); + const T* obj_mask_data = objness_mask->data(); + const int* gt_match_mask_data = gt_match_mask->data(); + T* input_grad_data = + input_grad->mutable_data({n, c, h, w}, ctx.GetPlace()); + memset(input_grad_data, 0, input_grad->numel() * sizeof(T)); + + for (int i = 0; i < n; i++) { + for (int t = 0; t < b; t++) { + int mask_idx = gt_match_mask_data[i * b + t]; + if (mask_idx >= 0) { + Box gt = GetGtBox(gt_box_data, i, b, t); + int gi = static_cast(gt.x * w); + int gj = static_cast(gt.y * h); + + int box_idx = GetEntryIndex(i, mask_idx, gj * w + gi, mask_num, + an_stride, stride, 0); + CalcBoxLocationLossGrad( + input_grad_data, loss_grad_data[i], input_data, gt, anchors, + anchor_mask[mask_idx], box_idx, gi, gj, h, input_size, stride); + + int label = gt_label_data[i * b + t]; + int label_idx = GetEntryIndex(i, mask_idx, gj * w + gi, mask_num, + an_stride, stride, 5); + CalcLabelLossGrad(input_grad_data, loss_grad_data[i], input_data, + label_idx, label, class_num, stride); + } + } + } + + CalcObjnessLossGrad(input_grad_data + 4 * stride, loss_grad_data, + input_data + 4 * stride, obj_mask_data, n, mask_num, + h, w, stride, an_stride); + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/distributed/brpc/brpc_client.cc b/paddle/fluid/operators/distributed/brpc/brpc_client.cc index b8e63f42e2040730ac79c57651d86d9e3176fa01..a1a3443348129b5cdf057592fced8fdff238ac09 100644 --- a/paddle/fluid/operators/distributed/brpc/brpc_client.cc +++ b/paddle/fluid/operators/distributed/brpc/brpc_client.cc @@ -80,7 +80,7 @@ VarHandlePtr BRPCClient::AsyncSendVar(const std::string& ep, google::protobuf::Closure* done = brpc::NewCallback( &HandleSendResponse, cntl, response, var_h, ch_ptr, ch_ctx, this); - platform::RecordRPCEvent record_event(method, p_ctx); + platform::RecordRPCEvent record_event(method); ch_ctx->stub->SendVariable(cntl, &request, response, done); @@ -184,7 +184,7 @@ VarHandlePtr BRPCClient::_AsyncGetVar(const std::string& ep, google::protobuf::Closure* done = brpc::NewCallback( &HandleGetResponse, cntl, response, var_h, ch_ptr, ch_ctx, this); - platform::RecordRPCEvent record_event(method, p_ctx); + platform::RecordRPCEvent record_event(method); if (method_name == kGetMonomerRPC) { ch_ctx->stub->GetMonomerVariable(cntl, &req, response, done); @@ -272,7 +272,7 @@ VarHandlePtr BRPCClient::AsyncPrefetchVar(const std::string& ep, &cntl->request_attachment(), out_var_name_val, false, 0, table_name_val); - platform::RecordRPCEvent record_event(method, p_ctx); + platform::RecordRPCEvent record_event(method); google::protobuf::Closure* done = brpc::NewCallback( &HandleGetResponse, cntl, response, var_h, ch_ptr, ch_ctx, this); @@ -311,7 +311,7 @@ VarHandlePtr BRPCClient::AsyncSendFetchBarrier(const std::string& ep, VarHandlePtr var_h( new VarHandle(ep, method, FETCH_BARRIER_MESSAGE, nullptr, nullptr)); - platform::RecordRPCEvent record_event(method, nullptr); + platform::RecordRPCEvent record_event(method); google::protobuf::Closure* done = brpc::NewCallback( &HandleFetchBarrierResponse, cntl, response, var_h, ch_ptr, ch_ctx, this); @@ -406,7 +406,7 @@ VarHandlePtr BRPCClient::AsyncSendVarMessage( sendrecv::VoidMessage* response = new sendrecv::VoidMessage(); cntl->set_timeout_ms(time_out); - platform::RecordRPCEvent record_event(method_name, nullptr); + platform::RecordRPCEvent record_event(method_name); VarHandlePtr var_h( new VarHandle(ep, method_name, req.varname(), nullptr, nullptr)); diff --git a/paddle/fluid/operators/distributed/grpc/grpc_client.cc b/paddle/fluid/operators/distributed/grpc/grpc_client.cc index 52310f8d04db6a5df9967c0a5ec9a5e95a24cdab..61e94dae3c7a107e10fa5e5518651014cec078bc 100644 --- a/paddle/fluid/operators/distributed/grpc/grpc_client.cc +++ b/paddle/fluid/operators/distributed/grpc/grpc_client.cc @@ -89,7 +89,7 @@ VarHandlePtr GRPCClient::AsyncSendVar(const std::string& ep, // stub context s->response_call_back_ = nullptr; - platform::RecordRPCEvent record_event(method, p_ctx); + platform::RecordRPCEvent record_event(method); auto call = s->stub_g_.PrepareUnaryCall( s->context_.get(), "/sendrecv.SendRecvService/SendVariable", req, &cq_); @@ -184,7 +184,7 @@ VarHandlePtr GRPCClient::_AsyncGetVar( // stub context s->response_call_back_ = ProcGetResponse; - platform::RecordRPCEvent record_event(method, p_ctx); + platform::RecordRPCEvent record_event(method); auto call = s->stub_g_.PrepareUnaryCall(s->context_.get(), rpc_path, buf, &cq_); @@ -235,7 +235,7 @@ VarHandlePtr GRPCClient::AsyncPrefetchVar(const std::string& ep, // stub context s->response_call_back_ = ProcGetResponse; - platform::RecordRPCEvent record_event(method, p_ctx); + platform::RecordRPCEvent record_event(method); auto call = s->stub_g_.PrepareUnaryCall( s->context_.get(), "/sendrecv.SendRecvService/PrefetchVariable", req, @@ -265,7 +265,7 @@ VarHandlePtr GRPCClient::AsyncSendBatchBarrier(const std::string& ep, sendrecv::VariableMessage req; req.set_varname(BATCH_BARRIER_MESSAGE); - platform::RecordRPCEvent record_event(method, nullptr); + platform::RecordRPCEvent record_event(method); auto rpc = s->stub_->AsyncSendVariable(s->context_.get(), req, &cq_); rpc->Finish(&s->reply_, &s->status_, reinterpret_cast(s)); @@ -290,7 +290,7 @@ VarHandlePtr GRPCClient::AsyncSendFetchBarrier(const std::string& ep, sendrecv::VariableMessage req; req.set_varname(FETCH_BARRIER_MESSAGE); - platform::RecordRPCEvent record_event(method, nullptr); + platform::RecordRPCEvent record_event(method); auto rpc = s->stub_->AsyncGetVariable(s->context_.get(), req, &cq_); rpc->Finish(&s->reply_, &s->status_, reinterpret_cast(s)); @@ -317,7 +317,7 @@ VarHandlePtr GRPCClient::AsyncGetMonomerBarrier(const std::string& ep, sendrecv::VariableMessage req; req.set_varname(var_name); - platform::RecordRPCEvent record_event(method, nullptr); + platform::RecordRPCEvent record_event(method); auto rpc = s->stub_->AsyncGetMonomerBarrier(s->context_.get(), req, &cq_); rpc->Finish(&s->reply_, &s->status_, reinterpret_cast(s)); @@ -342,7 +342,7 @@ VarHandlePtr GRPCClient::AsyncSendComplete(const std::string& ep, sendrecv::VariableMessage req; req.set_varname(COMPLETE_MESSAGE); - platform::RecordRPCEvent record_event(method, nullptr); + platform::RecordRPCEvent record_event(method); auto rpc = s->stub_->AsyncSendVariable(s->context_.get(), req, &cq_); rpc->Finish(&s->reply_, &s->status_, reinterpret_cast(s)); @@ -372,7 +372,7 @@ VarHandlePtr GRPCClient::AsyncCheckpointNotify(const std::string& ep, req.set_varname(CHECKPOINT_SAVE_MESSAGE); req.set_out_varname(dir); - platform::RecordRPCEvent record_event(method, nullptr); + platform::RecordRPCEvent record_event(method); auto rpc = s->stub_->AsyncCheckpointNotify(s->context_.get(), req, &cq_); rpc->Finish(&s->reply_, &s->status_, reinterpret_cast(s)); diff --git a/paddle/fluid/operators/distributed/grpc/grpc_serde.cc b/paddle/fluid/operators/distributed/grpc/grpc_serde.cc index 6df4fd36f95b127a0bbc0725b83c4494b160785f..6e65aa5fae83536d229be63fbaf7874bd45f967d 100644 --- a/paddle/fluid/operators/distributed/grpc/grpc_serde.cc +++ b/paddle/fluid/operators/distributed/grpc/grpc_serde.cc @@ -38,7 +38,7 @@ void SerializeToByteBuffer(const std::string& name, framework::Variable* var, ::grpc::ByteBuffer* msg, const std::string& out_name, const int trainer_id, const std::string& table_name) { - platform::RecordRPCEvent record_event("serial", &ctx); + platform::RecordRPCEvent record_event("serial"); VarMsg request; TensorPayload* payload = nullptr; @@ -147,7 +147,7 @@ void DeserializeFromByteBuffer(const ::grpc::ByteBuffer& msg, const platform::DeviceContext& ctx, const framework::Scope* scope, framework::Variable** var, int* trainer_id) { - platform::RecordRPCEvent record_event("deserial", &ctx); + platform::RecordRPCEvent record_event("deserial"); operators::distributed::GRPCVariableResponse resp(scope, &ctx); PADDLE_ENFORCE(resp.Parse(msg) == 0, "parse bytebuffer to tensor error!"); *var = resp.GetVar(); diff --git a/paddle/fluid/operators/distributed/proto_encoder_helper.h b/paddle/fluid/operators/distributed/proto_encoder_helper.h index 27ca1f4edc04f5fca54b1a6340243634a596939c..e9f06f54327875c0568c571627e9effb998e15be 100644 --- a/paddle/fluid/operators/distributed/proto_encoder_helper.h +++ b/paddle/fluid/operators/distributed/proto_encoder_helper.h @@ -85,7 +85,7 @@ class ProtoEncodeHelper { #define REPLACE_ENFORCE_GLOG 1 // Make sure callers didn't do operations that went over max_size promised if (paddle::platform::is_error(p_ <= limit_)) { - paddle::platform::throw_on_error(p_ <= limit_); + paddle::platform::throw_on_error(p_ <= limit_, ""); } #undef REPLACE_ENFORCE_GLOG } diff --git a/paddle/fluid/operators/dropout_op.cu b/paddle/fluid/operators/dropout_op.cu index d65491267de1ce3495d8b8250cf0cff570dfcc6a..7a6927d3e54b4ece8f17d7a1e7e431ba836edff9 100644 --- a/paddle/fluid/operators/dropout_op.cu +++ b/paddle/fluid/operators/dropout_op.cu @@ -114,4 +114,5 @@ REGISTER_OP_CUDA_KERNEL( ops::GPUDropoutKernel); REGISTER_OP_CUDA_KERNEL( dropout_grad, ops::DropoutGradKernel, + ops::DropoutGradKernel, ops::DropoutGradKernel); diff --git a/paddle/fluid/operators/elementwise/elementwise_add_op.cc b/paddle/fluid/operators/elementwise/elementwise_add_op.cc index 7e789cd8d9143164c2346b067855eb904e00075f..c6c658236c235f0a6767924026b0a7610071e918 100644 --- a/paddle/fluid/operators/elementwise/elementwise_add_op.cc +++ b/paddle/fluid/operators/elementwise/elementwise_add_op.cc @@ -18,6 +18,7 @@ namespace ops = paddle::operators; REGISTER_ELEMWISE_GRAD_MAKER(elementwise_add, Add); REGISTER_ELEMWISE_EXPLICIT_OP(elementwise_add, "Add", "Out = X + Y", "Out", "X"); + REGISTER_OP_CPU_KERNEL( elementwise_add, ops::ElementwiseAddKernel, diff --git a/paddle/fluid/operators/elementwise/elementwise_op.h b/paddle/fluid/operators/elementwise/elementwise_op.h index fd2a98cb45f446e80a4be1b50e94ee611cd23e62..91e44152658d87750f0b6d5826c481904085e086 100644 --- a/paddle/fluid/operators/elementwise/elementwise_op.h +++ b/paddle/fluid/operators/elementwise/elementwise_op.h @@ -250,6 +250,37 @@ class ElemwiseGradKernel : public framework::OpKernel { } }; +class ElementwiseOpInplace : public framework::InplaceInToOut { + public: + using framework::InplaceInToOut::InplaceInToOut; + + protected: + std::unordered_map Apply( + const framework::OpDesc &op_desc, + framework::BlockDesc *block) const override { + return std::unordered_map{ + {"X", "Out"}, + }; + } +}; + +class ElementwiseGradOpInplace : public framework::InplaceInToOut { + public: + using framework::InplaceInToOut::InplaceInToOut; + + protected: + std::unordered_map Apply( + const framework::OpDesc &op_desc, + framework::BlockDesc *block) const override { + std::unordered_map ret; + if (block->HasVar(framework::GradVarName("X")) && + block->HasVar(framework::GradVarName("Out"))) { + ret[framework::GradVarName("Out")] = framework::GradVarName("X"); + } + return ret; + } +}; + } // namespace operators } // namespace paddle @@ -299,6 +330,8 @@ class ElemwiseGradKernel : public framework::OpKernel { REGISTER_OPERATOR(op_type, ::paddle::operators::ElementwiseOp, \ __ElemwiseOp##op_type##Maker__, \ ::paddle::operators::ElementwiseOpInferVarType, \ - op_type##GradMaker); \ + op_type##GradMaker, \ + ::paddle::operators::ElementwiseOpInplace); \ REGISTER_OPERATOR(op_type##_grad, \ - ::paddle::operators::ElementwiseOpExplicitGrad) + ::paddle::operators::ElementwiseOpExplicitGrad, \ + ::paddle::operators::ElementwiseGradOpInplace) diff --git a/paddle/fluid/operators/elementwise/elementwise_add_mkldnn_op.cc b/paddle/fluid/operators/elementwise/mkldnn/elementwise_add_mkldnn_op.cc similarity index 100% rename from paddle/fluid/operators/elementwise/elementwise_add_mkldnn_op.cc rename to paddle/fluid/operators/elementwise/mkldnn/elementwise_add_mkldnn_op.cc diff --git a/paddle/fluid/operators/elementwise/elementwise_mul_mkldnn_op.cc b/paddle/fluid/operators/elementwise/mkldnn/elementwise_mul_mkldnn_op.cc similarity index 100% rename from paddle/fluid/operators/elementwise/elementwise_mul_mkldnn_op.cc rename to paddle/fluid/operators/elementwise/mkldnn/elementwise_mul_mkldnn_op.cc diff --git a/paddle/fluid/operators/expand_op.cc b/paddle/fluid/operators/expand_op.cc index 6aa4c76b9ce7f52f5816ea136e04b32a7d2e8d44..44a2f37b66772425a835c26e94c37b500e8a5d19 100644 --- a/paddle/fluid/operators/expand_op.cc +++ b/paddle/fluid/operators/expand_op.cc @@ -146,7 +146,11 @@ REGISTER_OPERATOR(expand, ops::ExpandOp, ops::ExpandOpMaker, paddle::framework::DefaultGradOpDescMaker); REGISTER_OPERATOR(expand_grad, ops::ExpandGradOp); REGISTER_OP_CPU_KERNEL( - expand, ops::ExpandKernel); + expand, ops::ExpandKernel, + ops::ExpandKernel, + ops::ExpandKernel, + ops::ExpandKernel); REGISTER_OP_CPU_KERNEL( expand_grad, - ops::ExpandGradKernel); + ops::ExpandGradKernel, + ops::ExpandGradKernel); diff --git a/paddle/fluid/operators/expand_op.cu b/paddle/fluid/operators/expand_op.cu index d95c9b61802b5fe7059e1c95a50776db5aa7ad93..50a506b294db14f0d170c60a0ed760dcf280ad60 100644 --- a/paddle/fluid/operators/expand_op.cu +++ b/paddle/fluid/operators/expand_op.cu @@ -15,7 +15,11 @@ limitations under the License. */ namespace ops = paddle::operators; REGISTER_OP_CUDA_KERNEL( - expand, ops::ExpandKernel); + expand, ops::ExpandKernel, + ops::ExpandKernel, + ops::ExpandKernel, + ops::ExpandKernel); REGISTER_OP_CUDA_KERNEL( expand_grad, - ops::ExpandGradKernel); + ops::ExpandGradKernel, + ops::ExpandGradKernel); diff --git a/paddle/fluid/operators/fake_quantize_op.cc b/paddle/fluid/operators/fake_quantize_op.cc index 8aff9111412030265491289bbdb03cf688d59ad8..d51eb054a96d27f6ce87ba4b4e717f49dcd8a588 100644 --- a/paddle/fluid/operators/fake_quantize_op.cc +++ b/paddle/fluid/operators/fake_quantize_op.cc @@ -21,26 +21,17 @@ limitations under the License. */ namespace paddle { namespace operators { -template -using EigenVectorArrayMap = - Eigen::TensorMap>; - -template -using ConstEigenVectorArrayMap = - Eigen::TensorMap>; +template +struct Compare { + public: + bool operator()(const T a, const T b) { return (std::abs(a) < std::abs(b)); } +}; template struct FindAbsMaxFunctor { void operator()(const platform::CPUDeviceContext& ctx, const T* in, const int num, T* out) { - Eigen::DSizes idim(num); - Eigen::DSizes odim(1); - Eigen::TensorMap> in_e(in, idim); - Eigen::TensorMap> out_e(out, odim); - - out_e = in_e.abs().maximum(); + *out = *(std::max_element(in + 0, in + num, Compare())); } }; diff --git a/paddle/fluid/operators/fc_op.cc b/paddle/fluid/operators/fc_op.cc index 38e57a41ed253eab4d0713af8bb14bac19041f6d..eb4617a9359353820fc41b9ad1c8db5327fdacde 100644 --- a/paddle/fluid/operators/fc_op.cc +++ b/paddle/fluid/operators/fc_op.cc @@ -47,7 +47,7 @@ void FCOp::InferShape(framework::InferShapeContext* ctx) const { PADDLE_ENFORCE(in_dims.size() == 2 || in_dims.size() == 4, "Fully Connected input should be 2-D or 4-D tensor."); } - PADDLE_ENFORCE_EQ(w_dims.size(), 2UL, + PADDLE_ENFORCE_EQ(w_dims.size(), 2, "Fully Connected input should be 2-D tensor."); int in_num_col_dims = ctx->Attrs().Get("in_num_col_dims"); PADDLE_ENFORCE_GT( diff --git a/paddle/fluid/operators/flatten_op.cc b/paddle/fluid/operators/flatten_op.cc index 8e80dc0e641c443923076c31e269689b5bc134a7..bb904166c4a19997a57723d9f2e50cc839aae960 100644 --- a/paddle/fluid/operators/flatten_op.cc +++ b/paddle/fluid/operators/flatten_op.cc @@ -267,6 +267,35 @@ class Flatten2GradOp : public framework::OperatorBase { } }; +class FlattenOpInplaceInToOut : public framework::InplaceInToOut { + public: + using InplaceInToOut::InplaceInToOut; + + protected: + std::unordered_map Apply( + const framework::OpDesc &op_desc, + framework::BlockDesc *block) const override { + std::unordered_map inplace_in_to_out = { + {"X", "Out"}, + }; + return inplace_in_to_out; + } +}; + +class FlattenGradInplaceinToOut : public framework::InplaceInToOut { + using InplaceInToOut::InplaceInToOut; + + protected: + std::unordered_map Apply( + const framework::OpDesc &op_desc, + framework::BlockDesc *block) const override { + std::unordered_map inplace_in_to_out = { + {framework::GradVarName("Out"), framework::GradVarName("X")}, + }; + return inplace_in_to_out; + } +}; + } // namespace operators } // namespace paddle @@ -275,10 +304,13 @@ USE_OP(reshape); namespace ops = paddle::operators; REGISTER_OPERATOR(flatten, ops::FlattenOp, ops::FlattenOpMaker, ops::FlattenOpInferShape, - paddle::framework::DefaultGradOpDescMaker); -REGISTER_OPERATOR(flatten_grad, ops::FlattenGradOp, ops::FlattenGradInferShape); + paddle::framework::DefaultGradOpDescMaker, + ops::FlattenOpInplaceInToOut); +REGISTER_OPERATOR(flatten_grad, ops::FlattenGradOp, ops::FlattenGradInferShape, + ops::FlattenGradInplaceinToOut); REGISTER_OPERATOR(flatten2, ops::Flatten2Op, ops::Flatten2OpMaker, - ops::Flatten2OpInferShape, ops::Flatten2GradOpMaker); + ops::Flatten2OpInferShape, ops::Flatten2GradOpMaker, + ops::FlattenOpInplaceInToOut); REGISTER_OPERATOR(flatten2_grad, ops::Flatten2GradOp, - ops::Flatten2GradInferShape); + ops::Flatten2GradInferShape, ops::FlattenGradInplaceinToOut); diff --git a/paddle/fluid/operators/fused/fused_embedding_seq_pool_op.h b/paddle/fluid/operators/fused/fused_embedding_seq_pool_op.h index 758432fd9e4197302e0bd8f76a1ca7c524026a70..33a1b47d150f653b84a377a61b251491aa719bee 100644 --- a/paddle/fluid/operators/fused/fused_embedding_seq_pool_op.h +++ b/paddle/fluid/operators/fused/fused_embedding_seq_pool_op.h @@ -21,6 +21,7 @@ limitations under the License. */ #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/selected_rows.h" +#include "paddle/fluid/operators/jit/kernels.h" #include "paddle/fluid/operators/math/blas.h" namespace paddle { @@ -37,32 +38,25 @@ struct EmbeddingVSumFunctor { const LoDTensor *table_t, const LoDTensor *ids_t, LoDTensor *output_t) { auto *table = table_t->data(); - int64_t row_number = table_t->dims()[0]; - int64_t row_width = table_t->dims()[1]; - int64_t last_dim = output_t->dims()[1]; + int64_t table_height = table_t->dims()[0]; + int64_t table_width = table_t->dims()[1]; + int64_t out_width = output_t->dims()[1]; const int64_t *ids = ids_t->data(); auto ids_lod = ids_t->lod()[0]; - int64_t ids_count = ids_t->numel() / ids_lod.back(); - + int64_t idx_width = ids_t->numel() / ids_lod.back(); auto *output = output_t->mutable_data(context.GetPlace()); - auto blas = math::GetBlas(context); - for (int64_t i = 0; i != ids_lod.size() - 1; ++i) { - size_t begin = ids_lod[i] * ids_count; - for (int64_t j = 0; j != ids_count; ++j) { - PADDLE_ENFORCE_LT(ids[begin], row_number); - PADDLE_ENFORCE_GE(ids[begin], 0, "ids %d", i); - blas.VCOPY(row_width, table + ids[begin + j] * row_width, - output + i * last_dim + j * row_width); - } - - for (int64_t r = (ids_lod[i] + 1) * ids_count; - r < ids_lod[i + 1] * ids_count; ++r) { - PADDLE_ENFORCE_LT(ids[r], row_number); - PADDLE_ENFORCE_GE(ids[r], 0, "ids %d", i); - blas.AXPY(row_width, 1., table + ids[r] * row_width, - output + i * last_dim + (r % ids_count) * row_width); - } + PADDLE_ENFORCE_LE(table_width * idx_width, out_width); + PADDLE_ENFORCE_GT(ids_lod.size(), 1UL); + + jit::emb_seq_pool_attr_t attr(table_height, table_width, 0, idx_width, + out_width, jit::SeqPoolType::kSum); + for (size_t i = 0; i != ids_lod.size() - 1; ++i) { + attr.index_height = ids_lod[i + 1] - ids_lod[i]; + auto emb_seqpool = jit::Get, + platform::CPUPlace>(attr); + emb_seqpool(table, ids + ids_lod[i] * idx_width, output + i * out_width, + &attr); } } }; diff --git a/paddle/fluid/operators/fused/fusion_repeated_fc_relu_op.cc b/paddle/fluid/operators/fused/fusion_repeated_fc_relu_op.cc index a35ee8a09ed5ddcc4ac465d200b84358fa65b2f3..8ecdf2ed9d40e7f5dc9226c635a8c8e6406a76ba 100644 --- a/paddle/fluid/operators/fused/fusion_repeated_fc_relu_op.cc +++ b/paddle/fluid/operators/fused/fusion_repeated_fc_relu_op.cc @@ -37,7 +37,7 @@ void FusionRepeatedFCReluOp::InferShape( "Output(Out) of FusionRepeatedFCReluOp should not be null."); auto i_dims = ctx->GetInputDim("X"); - PADDLE_ENFORCE_EQ(i_dims.size(), 2UL, "Input shape size should be 2"); + PADDLE_ENFORCE_EQ(i_dims.size(), 2, "Input shape size should be 2"); auto w_dims = ctx->GetInputsDim("W"); auto b_dims = ctx->GetInputsDim("Bias"); @@ -49,7 +49,7 @@ void FusionRepeatedFCReluOp::InferShape( "inpute width should be equal with weight height"); for (size_t i = 1; i < sz; ++i) { - PADDLE_ENFORCE_EQ(w_dims[i].size(), 2UL, + PADDLE_ENFORCE_EQ(w_dims[i].size(), 2, "Every weight shape size should be 2."); PADDLE_ENFORCE_EQ(framework::product(b_dims[i]), w_dims[i][1], "The length of Bias must be equal with w_dims[1]."); @@ -79,17 +79,17 @@ void FusionRepeatedFCReluOpMaker::Make() { } template -static void fc_relu(const T* x, const T* w, const T* b, T* y, int m, int n, - int k) { +static void fc_relu(const T* x, const T* w, const T* b, T* y, + const jit::matmul_attr_t& attr) { auto matmul = - jit::Get, platform::CPUPlace>(k); + jit::Get, platform::CPUPlace>(attr); auto addbias_relu = - jit::Get, platform::CPUPlace>(n); - matmul(x, w, y, m, n, k); + jit::Get, platform::CPUPlace>(attr.n); + matmul(x, w, y, &attr); T* dst = y; - for (int i = 0; i < m; ++i) { - addbias_relu(b, dst, dst, n); - dst += n; + for (int i = 0; i < attr.m; ++i) { + addbias_relu(b, dst, dst, attr.n); + dst += attr.n; } } @@ -107,32 +107,33 @@ class FusionRepeatedFCReluKernel : public framework::OpKernel { auto i_dims = in->dims(); auto w_dims = weights[0]->dims(); - int m = i_dims[0]; - int n = w_dims[1]; - int k = w_dims[0]; - relus[0]->Resize({m, n}); + jit::matmul_attr_t attr; + attr.m = i_dims[0]; + attr.n = w_dims[1]; + attr.k = w_dims[0]; + relus[0]->Resize({attr.m, attr.n}); fc_relu(in->data(), weights[0]->data(), biases[0]->data(), - relus[0]->mutable_data(place), m, n, k); + relus[0]->mutable_data(place), attr); for (int i = 1; i < weight_sz - 1; ++i) { auto i_dims = relus[i - 1]->dims(); auto w_dims = weights[i]->dims(); - int m = i_dims[0]; - int n = w_dims[1]; - int k = w_dims[0]; - relus[i]->Resize({m, n}); + attr.m = i_dims[0]; + attr.n = w_dims[1]; + attr.k = w_dims[0]; + relus[i]->Resize({attr.m, attr.n}); fc_relu(relus[i - 1]->data(), weights[i]->data(), - biases[i]->data(), relus[i]->mutable_data(place), m, n, k); + biases[i]->data(), relus[i]->mutable_data(place), attr); } auto i_dims_last = relus[weight_sz - 2]->dims(); auto w_dims_last = weights[weight_sz - 1]->dims(); - m = i_dims_last[0]; - n = w_dims_last[1]; - k = w_dims_last[0]; + attr.m = i_dims_last[0]; + attr.n = w_dims_last[1]; + attr.k = w_dims_last[0]; fc_relu(relus[weight_sz - 2]->data(), weights[weight_sz - 1]->data(), - biases[weight_sz - 1]->data(), out->mutable_data(place), m, n, - k); + biases[weight_sz - 1]->data(), out->mutable_data(place), + attr); } }; diff --git a/paddle/fluid/operators/fused/fusion_seqexpand_concat_fc_op.cc b/paddle/fluid/operators/fused/fusion_seqexpand_concat_fc_op.cc index aaef46de0d3b88720a762abb000e42d560fbd8cf..d091da5aa8a7e7ec30798d68021bfd2b9b87b32f 100644 --- a/paddle/fluid/operators/fused/fusion_seqexpand_concat_fc_op.cc +++ b/paddle/fluid/operators/fused/fusion_seqexpand_concat_fc_op.cc @@ -39,7 +39,7 @@ void FusionSeqExpandConcatFCOp::InferShape( auto ins_dims = ctx->GetInputsDim("X"); auto w_dims = ctx->GetInputDim("FCWeight"); // (M0+M1+M2+..) x D - PADDLE_ENFORCE_EQ(w_dims.size(), 2UL, "Input(FCWeight)'s rank must be 2."); + PADDLE_ENFORCE_EQ(w_dims.size(), 2, "Input(FCWeight)'s rank must be 2."); const int D = w_dims[1]; int sum = ins_dims[0][1]; for (size_t i = 1; i < ins_dims.size(); ++i) { diff --git a/paddle/fluid/operators/fused/fusion_seqpool_concat_op.cc b/paddle/fluid/operators/fused/fusion_seqpool_concat_op.cc index b181140db750a8d1b74c0b6cc93259a208fe5b06..d48bdafe0aa38cb860b54b2e41ebad3421b93bce 100644 --- a/paddle/fluid/operators/fused/fusion_seqpool_concat_op.cc +++ b/paddle/fluid/operators/fused/fusion_seqpool_concat_op.cc @@ -39,7 +39,7 @@ void FusionSeqPoolConcatOp::InferShape( // The output height should be confirmed in Compute, // since input lod is not accessible here. - PADDLE_ENFORCE_EQ(ins_dims[0].size(), 2UL, + PADDLE_ENFORCE_EQ(ins_dims[0].size(), 2, "The dims size of first input should be 2."); ctx->SetOutputDim("Out", {-1, ins_dims[0][axis] * static_cast(n)}); } diff --git a/paddle/fluid/operators/fused/fusion_squared_mat_sub_op.cc b/paddle/fluid/operators/fused/fusion_squared_mat_sub_op.cc index 00dafdead53bbd4614c70875441c565724fca46d..8493f4468fc994964116d99dc85dd34fb19a44cc 100644 --- a/paddle/fluid/operators/fused/fusion_squared_mat_sub_op.cc +++ b/paddle/fluid/operators/fused/fusion_squared_mat_sub_op.cc @@ -42,7 +42,7 @@ void FusionSquaredMatSubOp::InferShape( auto y_dims = ctx->GetInputDim("Y"); PADDLE_ENFORCE_EQ(x_dims.size(), y_dims.size(), "Input tensors dims size should be equal."); - PADDLE_ENFORCE_EQ(x_dims.size(), 2UL, "Input tensors should be a Matrix."); + PADDLE_ENFORCE_EQ(x_dims.size(), 2, "Input tensors should be a Matrix."); PADDLE_ENFORCE_EQ(x_dims[1], y_dims[0], "Inputs Matrix should be multiply."); ctx->SetOutputDim("SquaredX", x_dims); @@ -87,15 +87,18 @@ class FusionSquaredMatSubKernel : public framework::OpKernel { auto x_dims = x->dims(); auto y_dims = y->dims(); - int m = x_dims[0]; - int k = x_dims[1]; - int n = y_dims[1]; - int o_numel = m * n; + jit::matmul_attr_t attr; + attr.m = x_dims[0]; + attr.k = x_dims[1]; + attr.n = y_dims[1]; + int o_numel = attr.m * attr.n; auto vsquare_x = - jit::Get, platform::CPUPlace>(m * k); + jit::Get, platform::CPUPlace>(attr.m * + attr.k); auto vsquare_y = - jit::Get, platform::CPUPlace>(k * n); + jit::Get, platform::CPUPlace>(attr.k * + attr.n); auto vsquare_xy = jit::Get, platform::CPUPlace>(o_numel); auto vsub = @@ -103,7 +106,7 @@ class FusionSquaredMatSubKernel : public framework::OpKernel { auto vscal = jit::Get, platform::CPUPlace>(o_numel); auto matmul = - jit::Get, platform::CPUPlace>(k); + jit::Get, platform::CPUPlace>(attr); const T* x_data = x->data(); const T* y_data = y->data(); @@ -112,12 +115,12 @@ class FusionSquaredMatSubKernel : public framework::OpKernel { T* squared_xy_data = squared_xy->mutable_data(place); T* o_data = out->mutable_data(place); - matmul(x_data, y_data, squared_xy_data, m, n, k); + matmul(x_data, y_data, squared_xy_data, &attr); vsquare_xy(squared_xy_data, squared_xy_data, o_numel); - vsquare_x(x_data, squared_x_data, m * k); - vsquare_y(y_data, squared_y_data, k * n); - matmul(squared_x_data, squared_y_data, o_data, m, n, k); + vsquare_x(x_data, squared_x_data, attr.m * attr.k); + vsquare_y(y_data, squared_y_data, attr.k * attr.n); + matmul(squared_x_data, squared_y_data, o_data, &attr); vsub(squared_xy_data, o_data, o_data, o_numel); vscal(&scalar, o_data, o_data, o_numel); diff --git a/paddle/fluid/operators/gather_op.cu b/paddle/fluid/operators/gather_op.cu index 9f4aef08cd58e72ce344a640e6564b9e360ce169..490ba9a585ee8fac82a9e1178f506a6d39e5fd1c 100644 --- a/paddle/fluid/operators/gather_op.cu +++ b/paddle/fluid/operators/gather_op.cu @@ -31,7 +31,7 @@ class GatherOpCUDAKernel : public framework::OpKernel { auto *output = ctx.Output("Out"); output->mutable_data(ctx.GetPlace()); - + if (x->numel() == 0) return; GPUGather(ctx.device_context(), *x, *index, output); } }; @@ -45,14 +45,13 @@ class GatherGradOpCUDAKernel : public framework::OpKernel { auto *Index = ctx.Input("Index"); auto *dX = ctx.Output(framework::GradVarName("X")); auto *dO = ctx.Input(framework::GradVarName("Out")); - auto *x = ctx.Input("X"); dX->mutable_data(ctx.GetPlace()); auto dxt = framework::EigenVector::Flatten(*dX); auto &place = *ctx.template device_context() .eigen_device(); dxt.device(place) = dxt.constant(static_cast(0)); - + if (dO->numel() == 0) return; GPUScatterAssign(ctx.device_context(), *dO, *Index, dX); } }; @@ -61,11 +60,14 @@ class GatherGradOpCUDAKernel : public framework::OpKernel { } // namespace paddle namespace ops = paddle::operators; +namespace plat = paddle::platform; REGISTER_OP_CUDA_KERNEL(gather, ops::GatherOpCUDAKernel, ops::GatherOpCUDAKernel, ops::GatherOpCUDAKernel, - ops::GatherOpCUDAKernel); + ops::GatherOpCUDAKernel, + ops::GatherOpCUDAKernel); REGISTER_OP_CUDA_KERNEL(gather_grad, ops::GatherGradOpCUDAKernel, ops::GatherGradOpCUDAKernel, ops::GatherGradOpCUDAKernel, - ops::GatherGradOpCUDAKernel); + ops::GatherGradOpCUDAKernel, + ops::GatherGradOpCUDAKernel); diff --git a/paddle/fluid/operators/gather_op.h b/paddle/fluid/operators/gather_op.h index 2dd726bebb1bc2e4d83844c0b98df01c390e622f..2e18298cf8e34d5f70369c89b3b3b2a9ced0ce62 100644 --- a/paddle/fluid/operators/gather_op.h +++ b/paddle/fluid/operators/gather_op.h @@ -35,7 +35,7 @@ class GatherOpKernel : public framework::OpKernel { auto *output = ctx.Output("Out"); output->mutable_data(ctx.GetPlace()); - + if (x->numel() == 0) return; CPUGather(ctx.device_context(), *x, *index, output); } }; @@ -56,7 +56,7 @@ class GatherGradientOpKernel : public framework::OpKernel { auto &place = *ctx.template device_context() .eigen_device(); dxt.device(place) = dxt.constant(static_cast(0)); - + if (dO->numel() == 0) return; ScatterAssign(ctx.device_context(), *dO, *Index, dX); } }; diff --git a/paddle/fluid/operators/group_norm_op.cc b/paddle/fluid/operators/group_norm_op.cc index e18d9841bb87c6a684d53e1bceb6c20a37dcfcfa..cbdffa0db8277dbf7257c3b3c1d03c1b459d5b2b 100644 --- a/paddle/fluid/operators/group_norm_op.cc +++ b/paddle/fluid/operators/group_norm_op.cc @@ -170,13 +170,48 @@ class GroupNormGradMaker : public framework::SingleGradOpDescMaker { } }; +class GroupNormInplaceInToOut : public framework::InplaceInToOut { + public: + using InplaceInToOut::InplaceInToOut; + + protected: + std::unordered_map Apply( + const framework::OpDesc &op_desc, + framework::BlockDesc *block) const override { + return {{"X", "Y"}}; + } +}; + +class GroupNormGradInplaceInToOut : public framework::InplaceInToOut { + public: + using InplaceInToOut::InplaceInToOut; + + protected: + std::unordered_map Apply( + const framework::OpDesc &op_desc, + framework::BlockDesc *block) const override { + return {{framework::GradVarName("Y"), framework::GradVarName("X")}}; + } +}; + +class GroupNormOpInferVarType + : public framework::PassInDtypeAndVarTypeToOutput { + protected: + std::unordered_map GetInputOutputWithSameType() + const override { + return {{"X", /*->*/ "Y"}}; + } +}; + } // namespace operators } // namespace paddle namespace ops = paddle::operators; REGISTER_OPERATOR(group_norm, ops::GroupNormOp, ops::GroupNormOpMaker, - ops::GroupNormGradMaker); -REGISTER_OPERATOR(group_norm_grad, ops::GroupNormGradOp); + ops::GroupNormOpInferVarType, ops::GroupNormGradMaker, + ops::GroupNormInplaceInToOut); +REGISTER_OPERATOR(group_norm_grad, ops::GroupNormGradOp, + ops::GroupNormGradInplaceInToOut); REGISTER_OP_CPU_KERNEL( group_norm, ops::GroupNormKernel, ops::GroupNormKernel); diff --git a/paddle/fluid/operators/hierarchical_sigmoid_op.h b/paddle/fluid/operators/hierarchical_sigmoid_op.h index 1a7ca963010112bbcab69f1ceeb9cb8d19ca9b9e..4d5a84bcafed1ab0739349e1dbc7b5a9f9ad64ec 100644 --- a/paddle/fluid/operators/hierarchical_sigmoid_op.h +++ b/paddle/fluid/operators/hierarchical_sigmoid_op.h @@ -136,7 +136,7 @@ class HierarchicalSigmoidOpKernel : public framework::OpKernel { sum.mutable_data(framework::make_ddim(sum_dims), ctx.GetPlace()); auto sum_mat = EigenMatrix::From(sum); out->mutable_data(ctx.GetPlace()); - auto out_mat = framework::EigenVector::Flatten(*out); + auto out_mat = framework::EigenMatrix::From(*out); if (bias) { bit_code->Add(*bias, pre_out); } diff --git a/paddle/fluid/operators/interpolate_op.cc b/paddle/fluid/operators/interpolate_op.cc index 93dd3f794f6087a3158fee1f262795871f21611a..de91ba6270ac2ed22c8380878c0a0037fb1629c0 100644 --- a/paddle/fluid/operators/interpolate_op.cc +++ b/paddle/fluid/operators/interpolate_op.cc @@ -82,6 +82,18 @@ class InterpolateOpMaker : public framework::OpProtoAndCheckerMaker { "bilinear interpolation and \"nearest\" for nearest " "neighbor interpolation.") .SetDefault("bilinear"); + AddAttr( + "align_corners", + "an optinal bool. Defaults to True. " + "If True, the centers of 4 corner pixels of the input and output " + "tensors are aligned, preserving the values at the corner pixels, " + "if Flase, are not aligned") + .SetDefault(true); + AddAttr("align_mode", + "(int, default \'1\'), optional for bilinear interpolation" + "can be \'0\' for src_idx = scale*(dst_indx+0.5)-0.5 , " + "can be \'1\' for src_idx = scale*dst_index .") + .SetDefault(1); AddComment(R"DOC( This operator samples input X to given output shape by using specified interpolation method, the interpolation methods can be \"nearest\" @@ -98,6 +110,64 @@ class InterpolateOpMaker : public framework::OpProtoAndCheckerMaker { to perform linear interpolation first in one direction, and then again in the other direction. + Align_corners and align_mode are optinal parameters,the calculation method + of interpolation can be selected by them. + + Example: + + For scale: + + if align_corners = True and out_{size}>1 : + + scale_{factor} = (in_{size}-1.0)/(out_{size}-1.0) + + else: + + scale_{factor} = float(in_{size}/out_{size}) + + + Nearest neighbor interpolation: + + if: + align_corners = False + + input : (N,C,H_in,W_in) + output: (N,C,H_out,W_out) where: + + H_out = \left \lfloor {H_{in} * scale_{}factor}} \right \rfloor + W_out = \left \lfloor {W_{in} * scale_{}factor}} \right \rfloor + + else: + align_corners = True + + input : (N,C,H_in,W_in) + output: (N,C,H_out,W_out) where: + + H_out = round(H_{in} * scale_{factor}) + W_out = round(W_{in} * scale_{factor}) + + Bilinear interpolation: + + if: + align_corners = False , align_mode = 0 + + input : (N,C,H_in,W_in) + output: (N,C,H_out,W_out) where: + + H_out = (H_{in}+0.5) * scale_{factor} - 0.5 + W_out = (W_{in}+0.5) * scale_{factor} - 0.5 + + + else: + + input : (N,C,H_in,W_in) + output: (N,C,H_out,W_out) where: + + H_out = H_{in} * scale_{factor} + W_out = W_{in} * scale_{factor} + + + For details of nearest neighbor interpolation, please refer to Wikipedia: https://en.wikipedia.org/wiki/Nearest-neighbor_interpolation diff --git a/paddle/fluid/operators/interpolate_op.cu b/paddle/fluid/operators/interpolate_op.cu index 99ac725f73bf60ab0fb9a467432e8a57c646ef35..b887878ea2291d6c56fec91738784e338606b84f 100644 --- a/paddle/fluid/operators/interpolate_op.cu +++ b/paddle/fluid/operators/interpolate_op.cu @@ -23,7 +23,8 @@ __global__ void KeNearestNeighborInterpFw( const T* in, const size_t in_img_h, const size_t in_img_w, const size_t input_h, const size_t input_w, T* out, const size_t out_img_h, const size_t out_img_w, const size_t output_h, const size_t output_w, - const size_t num_channels, const float ratio_h, const float ratio_w) { + const size_t num_channels, const float ratio_h, const float ratio_w, + const bool align_corners) { int nthreads = output_h * output_w; int tid = blockIdx.x * blockDim.x + threadIdx.x; int stride = blockDim.x * gridDim.x; @@ -35,10 +36,14 @@ __global__ void KeNearestNeighborInterpFw( int channel_id = out_id_w / out_img_size; int out_img_idy = (out_id_w % out_img_size) / out_img_w; - int in_img_idy = static_cast(ratio_h * out_img_idy + 0.5); + int in_img_idy = (align_corners) + ? static_cast(ratio_h * out_img_idy + 0.5) + : static_cast(ratio_h * out_img_idy); int out_img_idx = tid % out_img_w; - int in_img_idx = static_cast(ratio_w * out_img_idx + 0.5); + int in_img_idx = (align_corners) + ? static_cast(ratio_w * out_img_idx + 0.5) + : static_cast(ratio_w * out_img_idx); out[tid] = in[out_id_h * input_w + channel_id * in_img_size + in_img_idy * in_img_w + in_img_idx]; @@ -50,7 +55,8 @@ __global__ void KeNearestNeighborInterpBw( T* in, const size_t in_img_h, const size_t in_img_w, const size_t input_h, const size_t input_w, const T* out, const size_t out_img_h, const size_t out_img_w, const size_t output_h, const size_t output_w, - const size_t num_channels, const float ratio_h, const float ratio_w) { + const size_t num_channels, const float ratio_h, const float ratio_w, + const bool align_corners) { int nthreads = output_h * output_w; int tid = blockIdx.x * blockDim.x + threadIdx.x; int stride = blockDim.x * gridDim.x; @@ -62,10 +68,14 @@ __global__ void KeNearestNeighborInterpBw( int channel_id = out_id_w / out_img_size; int out_img_idy = (out_id_w % out_img_size) / out_img_w; - int in_img_idy = static_cast(ratio_h * out_img_idy + 0.5); + int in_img_idy = (align_corners) + ? static_cast(ratio_h * out_img_idy + 0.5) + : static_cast(ratio_h * out_img_idy); int out_img_idx = tid % out_img_w; - int in_img_idx = static_cast(ratio_w * out_img_idx + 0.5); + int in_img_idx = (align_corners) + ? static_cast(ratio_w * out_img_idx + 0.5) + : static_cast(ratio_w * out_img_idx); T* in_pos = &in[out_id_h * input_w + channel_id * in_img_size + in_img_idy * in_img_w + in_img_idx]; @@ -79,10 +89,12 @@ __global__ void KeBilinearInterpFw( const T* in, const size_t in_img_h, const size_t in_img_w, const size_t input_h, const size_t input_w, T* out, const size_t out_img_h, const size_t out_img_w, const size_t output_h, const size_t output_w, - const size_t num_channels, const float ratio_h, const float ratio_w) { + const size_t num_channels, const float ratio_h, const float ratio_w, + const bool align_corners, const int align_mode) { int nthreads = output_h * output_w; int tid = blockIdx.x * blockDim.x + threadIdx.x; int stride = blockDim.x * gridDim.x; + bool align_flag = (align_mode == 0 && !align_corners); for (; tid < nthreads; tid += stride) { int out_id_h = tid / output_w; int out_id_w = tid % output_w; @@ -91,15 +103,23 @@ __global__ void KeBilinearInterpFw( int channel_id = out_id_w / out_img_size; int out_img_idy = (out_id_w % out_img_size) / out_img_w; - int in_img_idy = ratio_h * out_img_idy; + int in_img_idy = align_flag + ? static_cast(ratio_h * (out_img_idy + 0.5) - 0.5) + : static_cast(ratio_h * out_img_idy); + in_img_idy = (in_img_idy > 0) ? in_img_idy : 0; int h_id = (in_img_idy < in_img_h - 1) ? 1 : 0; - T h1lambda = ratio_h * out_img_idy - in_img_idy; + T h1lambda = align_flag ? ratio_h * (out_img_idy + 0.5) - 0.5 - in_img_idy + : ratio_h * out_img_idy - in_img_idy; T h2lambda = 1.f - h1lambda; int out_img_idx = tid % out_img_w; - int in_img_idx = ratio_w * out_img_idx; + int in_img_idx = align_flag + ? static_cast(ratio_w * (out_img_idx + 0.5) - 0.5) + : static_cast(ratio_w * out_img_idx); + in_img_idx = (in_img_idx > 0) ? in_img_idx : 0; int w_id = (in_img_idx < in_img_w - 1) ? 1 : 0; - T w1lambda = ratio_w * out_img_idx - in_img_idx; + T w1lambda = align_flag ? ratio_w * (out_img_idx + 0.5) - 0.5 - in_img_idx + : ratio_w * out_img_idx - in_img_idx; T w2lambda = 1.f - w1lambda; const T* in_pos = &in[out_id_h * input_w + channel_id * in_img_size + @@ -118,10 +138,12 @@ __global__ void KeBilinearInterpBw( T* in, const size_t in_img_h, const size_t in_img_w, const size_t input_h, const size_t input_w, const T* out, const size_t out_img_h, const size_t out_img_w, const size_t output_h, const size_t output_w, - const size_t num_channels, const T ratio_h, const T ratio_w) { + const size_t num_channels, const T ratio_h, const T ratio_w, + const bool align_corners, const int align_mode) { int nthreads = output_h * output_w; int tid = blockIdx.x * blockDim.x + threadIdx.x; int stride = blockDim.x * gridDim.x; + bool align_flag = (align_mode == 0 && !align_corners); for (; tid < nthreads; tid += stride) { int out_id_h = tid / output_w; int out_id_w = tid % output_w; @@ -130,15 +152,22 @@ __global__ void KeBilinearInterpBw( int channel_id = out_id_w / out_img_size; int out_img_idy = (out_id_w % out_img_size) / out_img_w; - int in_img_idy = ratio_h * out_img_idy; + int in_img_idy = align_flag ? ratio_h * (out_img_idy + 0.5) - 0.5 + : ratio_h * out_img_idy; + in_img_idy = (in_img_idy > 0) ? in_img_idy : 0; int h_id = (in_img_idy < in_img_h - 1) ? 1 : 0; - T h1lambda = ratio_h * out_img_idy - in_img_idy; + T h1lambda = align_flag ? ratio_h * (out_img_idy + 0.5) - 0.5 - in_img_idy + : ratio_h * out_img_idy - in_img_idy; + T h2lambda = 1.f - h1lambda; int out_img_idx = tid % out_img_w; - int in_img_idx = ratio_w * out_img_idx; + int in_img_idx = align_flag ? ratio_w * (out_img_idx + 0.5) - 0.5 + : ratio_w * out_img_idx; + in_img_idx = (in_img_idx > 0) ? in_img_idx : 0; int w_id = (in_img_idx < in_img_w - 1) ? 1 : 0; - T w1lambda = ratio_w * out_img_idx - in_img_idx; + T w1lambda = align_flag ? ratio_w * (out_img_idx + 0.5) - 0.5 - in_img_idx + : ratio_w * out_img_idx - in_img_idx; T w2lambda = 1.f - w1lambda; T* in_pos = &in[out_id_h * input_w + channel_id * in_img_size + @@ -175,6 +204,9 @@ class InterpolateOpCUDAKernel : public framework::OpKernel { out_w = size_data[1]; } + bool align_corners = ctx.Attr("align_corners"); + int align_mode = ctx.Attr("align_mode"); + int n = input->dims()[0]; int c = input->dims()[1]; int in_h = input->dims()[2]; @@ -188,10 +220,16 @@ class InterpolateOpCUDAKernel : public framework::OpKernel { int in_chw = c * in_hw; int out_chw = c * out_hw; - float ratio_h = - (out_h > 1) ? static_cast(in_h - 1) / (out_h - 1) : 0.f; - float ratio_w = - (out_w > 1) ? static_cast(in_w - 1) / (out_w - 1) : 0.f; + float ratio_h = 0.f; + float ratio_w = 0.f; + if (out_h > 1) { + ratio_h = (align_corners) ? static_cast(in_h - 1) / (out_h - 1) + : static_cast(in_h) / out_h; + } + if (out_w > 1) { + ratio_w = (align_corners) ? static_cast(in_w - 1) / (out_w - 1) + : static_cast(in_w) / out_w; + } if (in_h == out_h && in_w == out_w) { framework::TensorCopy(*input, ctx.GetPlace(), output); @@ -206,12 +244,12 @@ class InterpolateOpCUDAKernel : public framework::OpKernel { KeNearestNeighborInterpFw< T><<>>( input_data, in_h, in_w, n, in_chw, output_data, out_h, out_w, n, - out_chw, c, ratio_h, ratio_w); + out_chw, c, ratio_h, ratio_w, align_corners); } else if ("bilinear" == interp_method) { KeBilinearInterpFw< T><<>>( input_data, in_h, in_w, n, in_chw, output_data, out_h, out_w, n, - out_chw, c, ratio_h, ratio_w); + out_chw, c, ratio_h, ratio_w, align_corners, align_mode); } } }; @@ -234,6 +272,10 @@ class InterpolateGradOpCUDAKernel : public framework::OpKernel { int out_h = ctx.Attr("out_h"); int out_w = ctx.Attr("out_w"); auto out_size = ctx.Input("OutSize"); + + bool align_corners = ctx.Attr("align_corners"); + int align_mode = ctx.Attr("align_mode"); + if (out_size != nullptr) { Tensor sizes; framework::TensorCopy(*out_size, platform::CPUPlace(), &sizes); @@ -252,10 +294,16 @@ class InterpolateGradOpCUDAKernel : public framework::OpKernel { int in_chw = c * in_hw; int out_chw = c * out_hw; - float ratio_h = - (out_h > 1) ? static_cast(in_h - 1) / (out_h - 1) : 0.f; - float ratio_w = - (out_w > 1) ? static_cast(in_w - 1) / (out_w - 1) : 0.f; + float ratio_h = 0.f; + float ratio_w = 0.f; + if (out_h > 1) { + ratio_h = (align_corners) ? static_cast(in_h - 1) / (out_h - 1) + : static_cast(in_h) / out_h; + } + if (out_w > 1) { + ratio_w = (align_corners) ? static_cast(in_w - 1) / (out_w - 1) + : static_cast(in_w) / out_w; + } if (in_h == out_h && in_w == out_w) { framework::TensorCopy(*output_grad, ctx.GetPlace(), input_grad); @@ -270,12 +318,12 @@ class InterpolateGradOpCUDAKernel : public framework::OpKernel { KeNearestNeighborInterpBw< T><<>>( input_grad_data, in_h, in_w, n, in_chw, output_grad_data, out_h, - out_w, n, out_chw, c, ratio_h, ratio_w); + out_w, n, out_chw, c, ratio_h, ratio_w, align_corners); } else if ("bilinear" == interp_method) { KeBilinearInterpBw< T><<>>( input_grad_data, in_h, in_w, n, in_chw, output_grad_data, out_h, - out_w, n, out_chw, c, ratio_h, ratio_w); + out_w, n, out_chw, c, ratio_h, ratio_w, align_corners, align_mode); } } }; diff --git a/paddle/fluid/operators/interpolate_op.h b/paddle/fluid/operators/interpolate_op.h index 7fdb3e1f5a2ff82284d89dd0759e357978e1d873..c631ad1dd158ce114169602f073d69b2291b5b3b 100644 --- a/paddle/fluid/operators/interpolate_op.h +++ b/paddle/fluid/operators/interpolate_op.h @@ -26,14 +26,17 @@ template static void NearestNeighborInterpolate(const Tensor& input, Tensor* output, const float ratio_h, const float ratio_w, const int n, const int c, - const int out_h, const int out_w) { + const int out_h, const int out_w, + const bool align_corners) { auto input_t = EigenTensor::From(input); auto output_t = EigenTensor::From(*output); for (int k = 0; k < out_h; k++) { // loop for images - int in_k = static_cast(ratio_h * k + 0.5); + int in_k = (align_corners) ? static_cast(ratio_h * k + 0.5) + : static_cast(ratio_h * k); for (int l = 0; l < out_w; l++) { - int in_l = static_cast(ratio_w * l + 0.5); + int in_l = (align_corners) ? static_cast(ratio_w * l + 0.5) + : static_cast(ratio_w * l); for (int i = 0; i < n; i++) { // loop for batches for (int j = 0; j < c; j++) { // loop for channels @@ -48,20 +51,29 @@ template static void BilinearInterpolation(const Tensor& input, Tensor* output, const float ratio_h, const float ratio_w, const int in_h, const int in_w, const int n, - const int c, const int out_h, - const int out_w) { + const int c, const int out_h, const int out_w, + const bool align_corners, + const bool align_mode) { auto input_t = EigenTensor::From(input); auto output_t = EigenTensor::From(*output); + bool align_flag = (align_mode == 0 && !align_corners); for (int k = 0; k < out_h; k++) { // loop for images - int y_n = static_cast(ratio_h * k); + int y_n = align_flag ? static_cast(ratio_h * (k + 0.5) - 0.5) + : static_cast(ratio_h * k); + y_n = (y_n > 0) ? y_n : 0; int y_s = (y_n + 1) < (in_h - 1) ? (y_n + 1) : (in_h - 1); - float d_n = ratio_h * k - y_n; + float d_n = + align_flag ? ratio_h * (k + 0.5) - 0.5 - y_n : ratio_h * k - y_n; float d_s = 1.f - d_n; for (int l = 0; l < out_w; l++) { - int x_w = static_cast(ratio_w * l); + int x_w = (align_mode == 0 && !align_corners) + ? static_cast(ratio_w * (l + 0.5) - 0.5) + : static_cast(ratio_w * l); + x_w = (x_w > 0) ? x_w : 0; int x_e = (x_w + 1) < (in_w - 1) ? (x_w + 1) : (in_w - 1); - float d_w = ratio_w * l - x_w; + float d_w = + align_flag ? ratio_w * (l + 0.5) - 0.5 - x_w : ratio_w * l - x_w; float d_e = 1.f - d_w; for (int i = 0; i < n; i++) { // loop for batches @@ -78,19 +90,20 @@ static void BilinearInterpolation(const Tensor& input, Tensor* output, } template -static void NearestNeighborInterpolateGrad(const Tensor& output_grad, - Tensor* input_grad, - const float ratio_h, - const float ratio_w, const int n, - const int c, const int out_h, - const int out_w) { +static void NearestNeighborInterpolateGrad( + const Tensor& output_grad, Tensor* input_grad, const float ratio_h, + const float ratio_w, const int n, const int c, const int out_h, + const int out_w, const bool align_corners) { auto input_grad_t = EigenTensor::From(*input_grad); auto output_grad_t = EigenTensor::From(output_grad); + for (int k = 0; k < out_h; k++) { // loop for images - int in_k = static_cast(ratio_h * k + 0.5); + int in_k = (align_corners) ? static_cast(ratio_h * k + 0.5) + : static_cast(ratio_h * k); for (int l = 0; l < out_w; l++) { - int in_l = static_cast(ratio_w * l + 0.5); + int in_l = (align_corners) ? static_cast(ratio_w * l + 0.5) + : static_cast(ratio_w * l); for (int i = 0; i < n; i++) { // loop for batches for (int j = 0; j < c; j++) { // loop for channels @@ -106,19 +119,28 @@ static void BilinearInterpolationGrad(const Tensor& output_grad, Tensor* input_grad, const float ratio_h, const float ratio_w, const int in_h, const int in_w, const int n, const int c, - const int out_h, const int out_w) { + const int out_h, const int out_w, + const bool align_corners, + const int align_mode) { auto input_grad_t = EigenTensor::From(*input_grad); auto output_grad_t = EigenTensor::From(output_grad); + bool align_flag = (align_mode == 0 && !align_corners); for (int k = 0; k < out_h; k++) { // loop for images - int y_n = static_cast(ratio_h * k); + int y_n = align_flag ? static_cast(ratio_h * (k + 0.5) - 0.5) + : static_cast(ratio_h * k); + y_n = (y_n > 0) ? y_n : 0; int y_s = (y_n + 1) < (in_h - 1) ? (y_n + 1) : (in_h - 1); - float d_n = ratio_h * k - y_n; + float d_n = + align_flag ? ratio_h * (k + 0.5) - 0.5 - y_n : ratio_h * k - y_n; float d_s = 1.f - d_n; for (int l = 0; l < out_w; l++) { - int x_w = static_cast(ratio_w * l); + int x_w = align_flag ? static_cast(ratio_w * (l + 0.5) - 0.5) + : static_cast(ratio_w * l); + x_w = (x_w > 0) ? x_w : 0; int x_e = (x_w + 1) < (in_w - 1) ? (x_w + 1) : (in_w - 1); - float d_w = ratio_w * l - x_w; + float d_w = + align_flag ? ratio_w * (l + 0.5) - 0.5 - x_w : ratio_w * l - x_w; float d_e = 1.f - d_w; for (int i = 0; i < n; i++) { // loop for batches @@ -134,7 +156,6 @@ static void BilinearInterpolationGrad(const Tensor& output_grad, } } } - template class InterpolateKernel : public framework::OpKernel { public: @@ -151,6 +172,8 @@ class InterpolateKernel : public framework::OpKernel { out_h = out_size_data[0]; out_w = out_size_data[1]; } + bool align_corners = ctx.Attr("align_corners"); + int align_mode = ctx.Attr("align_mode"); const int n = input->dims()[0]; const int c = input->dims()[1]; @@ -168,17 +191,24 @@ class InterpolateKernel : public framework::OpKernel { return; } - float ratio_h = - (out_h > 1) ? static_cast(in_h - 1) / (out_h - 1) : 0.f; - float ratio_w = - (out_w > 1) ? static_cast(in_w - 1) / (out_w - 1) : 0.f; + float ratio_h = 0.f; + float ratio_w = 0.f; + + if (out_h > 1) { + ratio_h = (align_corners) ? static_cast(in_h - 1) / (out_h - 1) + : static_cast(in_h) / out_h; + } + if (out_w > 1) { + ratio_w = (align_corners) ? static_cast(in_w - 1) / (out_w - 1) + : static_cast(in_w) / out_w; + } if ("bilinear" == interp_method) { BilinearInterpolation(*input, output, ratio_h, ratio_w, in_h, in_w, n, - c, out_h, out_w); + c, out_h, out_w, align_corners, align_mode); } else if ("nearest" == interp_method) { NearestNeighborInterpolate(*input, output, ratio_h, ratio_w, n, c, - out_h, out_w); + out_h, out_w, align_corners); } } }; @@ -200,6 +230,8 @@ class InterpolateGradKernel : public framework::OpKernel { out_h = out_size_data[0]; out_w = out_size_data[1]; } + bool align_corners = ctx.Attr("align_corners"); + int align_mode = ctx.Attr("align_mode"); const int n = input->dims()[0]; const int c = input->dims()[1]; @@ -217,17 +249,26 @@ class InterpolateGradKernel : public framework::OpKernel { return; } - float ratio_h = - (out_h > 1) ? static_cast(in_h - 1) / (out_h - 1) : 0.f; - float ratio_w = - (out_w > 1) ? static_cast(in_w - 1) / (out_w - 1) : 0.f; + float ratio_h = 0.f; + float ratio_w = 0.f; + + if (out_h > 1) { + ratio_h = (align_corners) ? static_cast(in_h - 1) / (out_h - 1) + : static_cast(in_h) / out_h; + } + if (out_w > 1) { + ratio_w = (align_corners) ? static_cast(in_w - 1) / (out_w - 1) + : static_cast(in_w) / out_w; + } if ("bilinear" == interp_method) { BilinearInterpolationGrad(*output_grad, input_grad, ratio_h, ratio_w, - in_h, in_w, n, c, out_h, out_w); + in_h, in_w, n, c, out_h, out_w, + align_corners, align_mode); } else if ("nearest" == interp_method) { NearestNeighborInterpolateGrad(*output_grad, input_grad, ratio_h, - ratio_w, n, c, out_h, out_w); + ratio_w, n, c, out_h, out_w, + align_corners); } } }; diff --git a/paddle/fluid/operators/jit/benchmark.cc b/paddle/fluid/operators/jit/benchmark.cc index 186c37c56ec9410ac9a31503e33e7e334d0afc40..3348778ee782ef0cdd1df4c3c4b24060436d7d79 100644 --- a/paddle/fluid/operators/jit/benchmark.cc +++ b/paddle/fluid/operators/jit/benchmark.cc @@ -93,6 +93,7 @@ std::vector TestSizes() { template struct BenchFunc { // return this function avg time + // TODO(TJ): clear cache every time double operator()(const typename KernelTuples::func_type tgt, Args... args) { for (int i = 0; i < FLAGS_burning; ++i) { tgt(args...); @@ -158,7 +159,7 @@ void BenchAllImpls(const typename KernelTuples::attr_type& attr, Args... args) { using Tensor = paddle::framework::Tensor; -template +template void BenchXYZNKernel() { for (int d : TestSizes()) { Tensor x, y, z; @@ -172,10 +173,13 @@ void BenchXYZNKernel() { RandomVec(d, y_data); BenchAllImpls, PlaceType>(d, x.data(), y.data(), z_data, d); + // test inplace + BenchAllImpls, PlaceType>(d, x.data(), z_data, + z_data, d); } } -template +template void BenchAXYNKernel() { for (int d : TestSizes()) { const T a = static_cast(3); @@ -187,10 +191,23 @@ void BenchAXYNKernel() { RandomVec(d, x_data); BenchAllImpls, PlaceType>(d, &a, x.data(), y_data, d); + // test inplace + BenchAllImpls, PlaceType>(d, &a, x.data(), x_data, + d); + } +} + +template +void BenchXRNKernel() { + for (int d : TestSizes()) { + Tensor x; + RandomVec(d, x.mutable_data({d}, PlaceType())); + T res; + BenchAllImpls, PlaceType>(d, x.data(), &res, d); } } -template +template void BenchXYNKernel() { for (int d : TestSizes()) { Tensor x, y; @@ -203,7 +220,7 @@ void BenchXYNKernel() { } } -template +template void BenchLSTMKernel() { for (bool use_peephole : {true, false}) { for (int d : TestSizes()) { @@ -240,7 +257,7 @@ void BenchLSTMKernel() { } } -template +template void BenchGRUKernel() { for (int d : TestSizes()) { const jit::gru_attr_t attr(d, jit::kVSigmoid, jit::kVTanh); @@ -262,7 +279,7 @@ void BenchGRUKernel() { } } -template +template void BenchSeqPoolKernel() { std::vector pool_types = { jit::SeqPoolType::kSum, jit::SeqPoolType::kAvg, jit::SeqPoolType::kSqrt}; @@ -284,7 +301,38 @@ void BenchSeqPoolKernel() { } } -template +template +void BenchEmbSeqPoolKernel() { + std::vector pool_types = {jit::SeqPoolType::kSum}; + int64_t tbl_h = 1e4; + for (int tbl_w : {10, 16, 256}) { + Tensor table; + table.Resize({tbl_h, tbl_w}); + RandomVec(tbl_h * tbl_w, table.mutable_data(PlaceType()), -2.f, 2.f); + const T* table_data = table.data(); + for (auto type : pool_types) { + for (int idx_w : {1, 2, 10, 16}) { + for (int idx_h : {1, 2, 9, 13, 16}) { + int64_t out_w = tbl_w * idx_w; + jit::emb_seq_pool_attr_t attr(tbl_h, tbl_w, idx_h, idx_w, out_w, + type); + Tensor idx, out; + idx.Resize({idx_h, idx_w}); + out.Resize({out_w}); + RandomVec(idx_h * idx_w, + idx.mutable_data(PlaceType()), 0, + tbl_h - 1); + const int64_t* idx_data = idx.data(); + T* o_data = out.mutable_data(PlaceType()); + BenchAllImpls, PlaceType>( + attr, table_data, idx_data, o_data, &attr); + } + } + } + } +} + +template void BenchMatMulKernel() { for (int m : {1, 2, 3, 4}) { for (int n : TestSizes()) { @@ -298,65 +346,153 @@ void BenchMatMulKernel() { const T* a_data = a.data(); const T* b_data = b.data(); T* c_data = c.mutable_data(PlaceType()); - BenchAllImpls, PlaceType>(k, a_data, b_data, - c_data, m, n, k); + const jit::matmul_attr_t attr{m, n, k}; + BenchAllImpls, PlaceType>(attr, a_data, b_data, + c_data, &attr); } } } } -using T = float; -using PlaceType = paddle::platform::CPUPlace; +template +void BenchSoftmaxKernel() { + for (int bs : {1, 2, 10}) { + for (int n : TestSizes()) { + Tensor x, y; + x.Resize({bs, n}); + y.Resize({bs, n}); + RandomVec(bs * n, x.mutable_data(PlaceType()), -2.f, 2.f); + const T* x_data = x.data(); + T* y_data = y.mutable_data(PlaceType()); + BenchAllImpls, PlaceType>(n, x_data, y_data, n, + bs); + } + } +} -// xyzn -BENCH_FP32_CPU(kVMul) { BenchXYZNKernel(); } +template +void BenchLayerNormKernel() { + const T epsilon = 9.99999975e-06; + for (int n : {1, 2, 10}) { + for (int x_dim_0 : {1, 9, 17, 50}) { + int left = n * x_dim_0; + for (int x_dim_1 : TestSizes()) { + int right = x_dim_1; + int sz = left * right; + Tensor x, mean, var, scale, bias, out; + x.Resize({n, x_dim_0, x_dim_1}); + out.Resize({n, x_dim_0, x_dim_1}); + mean.Resize({n, x_dim_0}); + var.Resize({n, x_dim_0}); + scale.Resize({x_dim_1}); + bias.Resize({x_dim_1}); + + RandomVec(sz, x.mutable_data(PlaceType()), -2.f, 2.f); + RandomVec(left, mean.mutable_data(PlaceType()), -2.f, 2.f); + RandomVec(left, var.mutable_data(PlaceType()), -2.f, 2.f); + RandomVec(right, scale.mutable_data(PlaceType()), -2.f, 2.f); + RandomVec(right, bias.mutable_data(PlaceType()), -2.f, 2.f); + + const T* scale_data = scale.data(); + const T* bias_data = bias.data(); + T* x_data = x.data(); + T* mean_data = mean.data(); + T* var_data = var.data(); + T* out_data = out.mutable_data(PlaceType()); + + BenchAllImpls, PlaceType>( + right, x_data, out_data, mean_data, var_data, scale_data, bias_data, + left, epsilon, right); + } + } + } +} -BENCH_FP32_CPU(kVAdd) { BenchXYZNKernel(); } +template +void BenchCRFDecodingKernel() { + constexpr int state_trans_base_idx = 2; + for (int seq_len : {1, 11, 17, 50}) { + for (int tag_num : TestSizes()) { + int x_sz = seq_len * tag_num; + int w_sz = (tag_num + state_trans_base_idx) * tag_num; + Tensor x, w, alpha, track; + x.Resize({seq_len, tag_num}); + w.Resize({tag_num + state_trans_base_idx, tag_num}); + alpha.Resize({seq_len, tag_num}); + track.Resize({seq_len, tag_num}); + + RandomVec(x_sz, x.mutable_data(PlaceType()), -2.f, 2.f); + RandomVec(w_sz, w.mutable_data(PlaceType()), -2.f, 2.f); + + const T* x_data = x.data(); + const T* w_data = w.data(); + T* alpha_data = alpha.mutable_data(PlaceType()); + int* track_data = track.mutable_data(PlaceType()); + + BenchAllImpls, PlaceType>( + tag_num, seq_len, x_data, w_data, alpha_data, track_data, tag_num); + } + } +} -BENCH_FP32_CPU(kVAddRelu) { BenchXYZNKernel(); } +using T = float; +using CPUPlace = paddle::platform::CPUPlace; -BENCH_FP32_CPU(kVSub) { BenchXYZNKernel(); } +// xyzn +BENCH_FP32_CPU(kVMul) { BenchXYZNKernel(); } +BENCH_FP32_CPU(kVAdd) { BenchXYZNKernel(); } +BENCH_FP32_CPU(kVAddRelu) { BenchXYZNKernel(); } +BENCH_FP32_CPU(kVSub) { BenchXYZNKernel(); } // axyn -BENCH_FP32_CPU(kVScal) { BenchAXYNKernel(); } +BENCH_FP32_CPU(kVScal) { BenchAXYNKernel(); } +BENCH_FP32_CPU(kVAddBias) { BenchAXYNKernel(); } -BENCH_FP32_CPU(kVAddBias) { BenchAXYNKernel(); } +// xrn +BENCH_FP32_CPU(kHSum) { BenchXRNKernel(); } +BENCH_FP32_CPU(kHMax) { BenchXRNKernel(); } // xyn -BENCH_FP32_CPU(kVRelu) { BenchXYNKernel(); } - -BENCH_FP32_CPU(kVIdentity) { BenchXYNKernel(); } - -BENCH_FP32_CPU(kVSquare) { BenchXYNKernel(); } +BENCH_FP32_CPU(kVRelu) { BenchXYNKernel(); } +BENCH_FP32_CPU(kVIdentity) { BenchXYNKernel(); } +BENCH_FP32_CPU(kVSquare) { BenchXYNKernel(); } +BENCH_FP32_CPU(kVExp) { BenchXYNKernel(); } +BENCH_FP32_CPU(kVSigmoid) { BenchXYNKernel(); } +BENCH_FP32_CPU(kVTanh) { BenchXYNKernel(); } -BENCH_FP32_CPU(kVExp) { BenchXYNKernel(); } +// lstm and peephole +BENCH_FP32_CPU(kLSTMCtHt) { BenchLSTMKernel(); } +BENCH_FP32_CPU(kLSTMC1H1) { BenchLSTMKernel(); } -BENCH_FP32_CPU(kVSigmoid) { BenchXYNKernel(); } +// gru functions +BENCH_FP32_CPU(kGRUH1) { BenchGRUKernel(); } +BENCH_FP32_CPU(kGRUHtPart1) { BenchGRUKernel(); } +BENCH_FP32_CPU(kGRUHtPart2) { BenchGRUKernel(); } -BENCH_FP32_CPU(kVTanh) { BenchXYNKernel(); } +// seq pool function +BENCH_FP32_CPU(kSeqPool) { BenchSeqPoolKernel(); } -// lstm and peephole -BENCH_FP32_CPU(kLSTMCtHt) { BenchLSTMKernel(); } +// embedding seq pool function +BENCH_FP32_CPU(kEmbSeqPool) { + BenchEmbSeqPoolKernel(); +} -BENCH_FP32_CPU(kLSTMC1H1) { BenchLSTMKernel(); } +// matmul +BENCH_FP32_CPU(kMatMul) { BenchMatMulKernel(); } -// gru functions -BENCH_FP32_CPU(kGRUH1) { BenchGRUKernel(); } +// softmax +BENCH_FP32_CPU(kSoftmax) { BenchSoftmaxKernel(); } -BENCH_FP32_CPU(kGRUHtPart1) { - BenchGRUKernel(); +// layernorm +BENCH_FP32_CPU(kLayerNorm) { + BenchLayerNormKernel(); } -BENCH_FP32_CPU(kGRUHtPart2) { - BenchGRUKernel(); +// crfdecoding +BENCH_FP32_CPU(kCRFDecoding) { + BenchCRFDecodingKernel(); } -// seq pool function -BENCH_FP32_CPU(kSeqPool) { BenchSeqPoolKernel(); } - -// matmul -BENCH_FP32_CPU(kMatMul) { BenchMatMulKernel(); } - // Benchmark all jit kernels including jitcode, mkl and refer. // To use this tool, run command: ./benchmark [options...] // Options: diff --git a/paddle/fluid/operators/jit/gen/CMakeLists.txt b/paddle/fluid/operators/jit/gen/CMakeLists.txt index 40310c2d2b372a414054f75348e8e1b4471bf3d2..294f73d9646c93132e464a032e93562094663a73 100644 --- a/paddle/fluid/operators/jit/gen/CMakeLists.txt +++ b/paddle/fluid/operators/jit/gen/CMakeLists.txt @@ -9,6 +9,7 @@ function(USE_JITKERNEL_GEN TARGET) endfunction() # use gen jitcode kernel by name +USE_JITKERNEL_GEN(kMatMul) USE_JITKERNEL_GEN(kVMul) USE_JITKERNEL_GEN(kVAdd) USE_JITKERNEL_GEN(kVSub) @@ -28,3 +29,6 @@ USE_JITKERNEL_GEN(kGRUHtPart1) USE_JITKERNEL_GEN(kGRUHtPart2) USE_JITKERNEL_GEN(kNCHW16CMulNC) USE_JITKERNEL_GEN(kSeqPool) +USE_JITKERNEL_GEN(kHMax) +USE_JITKERNEL_GEN(kHSum) +USE_JITKERNEL_GEN(kEmbSeqPool) diff --git a/paddle/fluid/operators/jit/gen/act.cc b/paddle/fluid/operators/jit/gen/act.cc index a2a5661b93ad3d885983c502566860aa313d110f..e7a7375879064eb27c94315fe7b93eece7866b92 100644 --- a/paddle/fluid/operators/jit/gen/act.cc +++ b/paddle/fluid/operators/jit/gen/act.cc @@ -81,9 +81,7 @@ void VActJitCode::genCode() { #define DECLARE_ACT_CREATOR(name) \ class name##Creator : public JitCodeCreator { \ public: \ - bool UseMe(const int& attr) const override { \ - return platform::MayIUse(platform::avx); \ - } \ + bool UseMe(const int& attr) const override; \ size_t CodeSize(const int& d) const override; \ std::unique_ptr CreateJitCode(const int& attr) const override { \ return make_unique(attr, CodeSize(attr)); \ @@ -98,6 +96,30 @@ DECLARE_ACT_CREATOR(VSigmoid); DECLARE_ACT_CREATOR(VTanh); // TODO(TJ): tuning use me +bool VReluCreator::UseMe(const int& d) const { + return platform::MayIUse(platform::avx); +} + +bool VSquareCreator::UseMe(const int& d) const { + return platform::MayIUse(platform::avx); +} + +bool VIdentityCreator::UseMe(const int& d) const { + return platform::MayIUse(platform::avx); +} + +bool VExpCreator::UseMe(const int& d) const { + return platform::MayIUse(platform::avx) && d < 32; +} + +bool VSigmoidCreator::UseMe(const int& d) const { + return platform::MayIUse(platform::avx); +} + +bool VTanhCreator::UseMe(const int& d) const { + return platform::MayIUse(platform::avx); +} + size_t VReluCreator::CodeSize(const int& d) const { return 96 /* init size */ + (d / YMM_FLOAT_BLOCK + 3) * 4 /* instructions */ * diff --git a/paddle/fluid/operators/jit/gen/act.h b/paddle/fluid/operators/jit/gen/act.h index 68e66f9298c4eafabb55c20195d46fed800f4ec4..13d98577e21db9041686822f57cb4992e5ad71ec 100644 --- a/paddle/fluid/operators/jit/gen/act.h +++ b/paddle/fluid/operators/jit/gen/act.h @@ -63,7 +63,6 @@ class VActFunc : public JitCode { public: explicit VActFunc(size_t code_size, void* code_ptr) : JitCode(code_size, code_ptr) {} - virtual const char* name() const = 0; virtual void genCode() = 0; protected: @@ -269,7 +268,7 @@ class VActJitCode : public VActFunc { this->genCode(); } - const char* name() const override { + std::string name() const override { std::string base = "VActJitCode"; switch (type_) { case operand_type::RELU: @@ -293,7 +292,7 @@ class VActJitCode : public VActFunc { default: break; } - return base.c_str(); + return base; } void genCode() override; diff --git a/paddle/fluid/operators/jit/gen/blas.cc b/paddle/fluid/operators/jit/gen/blas.cc index dee6c7b9d3ee9756c1b11d10d55fdca341cbee85..5da24c359edd2df93333fe0ca8a18cdc7385aadb 100644 --- a/paddle/fluid/operators/jit/gen/blas.cc +++ b/paddle/fluid/operators/jit/gen/blas.cc @@ -155,7 +155,7 @@ class NCHW16CMulNCCreator : public JitCodeCreator { class name##Creator : public JitCodeCreator { \ public: \ bool UseMe(const int& attr) const override { \ - return platform::MayIUse(platform::avx); \ + return platform::MayIUse(platform::avx) && attr <= 1024; \ } \ size_t CodeSize(const int& d) const override { \ return 96 + d / YMM_FLOAT_BLOCK * 4 * 8; \ diff --git a/paddle/fluid/operators/jit/gen/blas.h b/paddle/fluid/operators/jit/gen/blas.h index de6b33f467279124d7acd97709516c31706ec4f9..70312bbe5e97fcf465ce13ef71e5acc9bab4874e 100644 --- a/paddle/fluid/operators/jit/gen/blas.h +++ b/paddle/fluid/operators/jit/gen/blas.h @@ -41,7 +41,7 @@ class VXXJitCode : public JitCode { this->genCode(); } - virtual const char* name() const { + std::string name() const override { std::string base = "VXXJitCode"; if (scalar_index_ == 1) { base += "_Scalar"; @@ -61,7 +61,8 @@ class VXXJitCode : public JitCode { base += "_Vec"; } base += (with_relu_ ? "_Relu" : ""); - return base.c_str(); + base += "_D" + std::to_string(num_); + return base; } void genCode() override; diff --git a/paddle/fluid/operators/jit/gen/embseqpool.cc b/paddle/fluid/operators/jit/gen/embseqpool.cc new file mode 100644 index 0000000000000000000000000000000000000000..23837a3fb9886ae8a839d4b31bd57916168ea53c --- /dev/null +++ b/paddle/fluid/operators/jit/gen/embseqpool.cc @@ -0,0 +1,149 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. */ + +#include "paddle/fluid/operators/jit/gen/embseqpool.h" +#include // offsetof +#include +#include "paddle/fluid/operators/jit/gen/act.h" // for exp_float_consts ones +#include "paddle/fluid/operators/jit/registry.h" +#include "paddle/fluid/platform/cpu_info.h" + +namespace paddle { +namespace operators { +namespace jit { +namespace gen { + +void EmbSeqPoolJitCode::genCode() { + preCode(); + constexpr int block = YMM_FLOAT_BLOCK; + constexpr int max_num_regs = 8; + const int num_block = tbl_w_ / block; + const int num_groups = num_block / max_num_regs; + const size_t block_size = sizeof(float) * block; + std::vector groups(num_groups, max_num_regs); + int rest_num_regs = num_block % max_num_regs; + if (rest_num_regs > 0) { + groups.push_back(rest_num_regs); + } + + // protect param_dst + mov(reg_ptr_param_dst, param_dst); + mov(reg_idx_width_in_byte, + qword[param_attr + offsetof(emb_seq_pool_attr_t, index_width)]); + mov(reg_idx_height, + qword[param_attr + offsetof(emb_seq_pool_attr_t, index_height)]); + mov(rax, sizeof(int64_t)); + mul(reg_idx_width_in_byte); + mov(reg_idx_width_in_byte, rax); + const size_t tbl_width_in_byte = sizeof(float) * tbl_w_; + int acc_num_regs = 0; + for (int num_regs : groups) { + Label l_next_idx_w, l_next_idx_h, l_save_now; + xor_(reg_idx_w_i_in_byte, reg_idx_w_i_in_byte); + mov(reg_ptr_dst_i, reg_ptr_param_dst); + add(reg_ptr_dst_i, acc_num_regs * block_size); + + L(l_next_idx_w); + { + // h == 0 + mov(reg_ptr_idx_i, param_idx); + add(reg_ptr_idx_i, reg_idx_w_i_in_byte); + mov(reg_idx, qword[reg_ptr_idx_i]); + mov(rax, tbl_width_in_byte); + mul(reg_idx); + mov(reg_ptr_tbl_i, rax); // reg is offset now + add(reg_ptr_tbl_i, param_tbl); // reg is ptr_i now + size_t w_offset = 0; + for (int reg_i = 0; reg_i < num_regs; ++reg_i) { + vmovups(ymm_t(reg_i + num_regs), ptr[reg_ptr_tbl_i + w_offset]); + w_offset += block_size; + } + add(reg_ptr_idx_i, reg_idx_width_in_byte); + + // end condition of idx h + mov(reg_idx_h_end, reg_idx_height); + mov(rax, reg_idx_width_in_byte); + mul(reg_idx_h_end); + mov(reg_idx_h_end, rax); + add(reg_idx_h_end, reg_idx_w_i_in_byte); + add(reg_idx_h_end, param_idx); + + cmp(reg_ptr_idx_i, reg_idx_h_end); + jge(l_save_now, T_NEAR); + L(l_next_idx_h); + { + mov(reg_idx, qword[reg_ptr_idx_i]); + mov(reg_ptr_tbl_i, reg_idx); + mov(rax, tbl_width_in_byte); + mul(reg_idx); + mov(reg_ptr_tbl_i, rax); + add(reg_ptr_tbl_i, param_tbl); + size_t w_offset = 0; + for (int reg_i = 0; reg_i < num_regs; ++reg_i) { + vmovups(ymm_t(reg_i), ptr[reg_ptr_tbl_i + w_offset]); + vaddps(ymm_t(reg_i + num_regs), ymm_t(reg_i + num_regs), + ymm_t(reg_i)); + w_offset += block_size; + } + add(reg_ptr_idx_i, reg_idx_width_in_byte); + cmp(reg_ptr_idx_i, reg_idx_h_end); + jl(l_next_idx_h, T_NEAR); + } // end of idx h + L(l_save_now); + // avg or sqrt here, if needed + w_offset = 0; + for (int reg_i = 0; reg_i < num_regs; ++reg_i) { + vmovups(ptr[reg_ptr_dst_i + w_offset], ymm_t(reg_i + num_regs)); + w_offset += block_size; + } + add(reg_ptr_dst_i, tbl_width_in_byte); + add(reg_idx_w_i_in_byte, sizeof(int64_t)); + cmp(reg_idx_w_i_in_byte, reg_idx_width_in_byte); + jl(l_next_idx_w, T_NEAR); + } // end of idx w + + acc_num_regs += num_regs; + add(param_tbl, num_regs * block_size); // do not use acc_num_regs + } // end of groups + postCode(); +} + +class EmbSeqPoolCreator : public JitCodeCreator { + public: + bool UseMe(const emb_seq_pool_attr_t& attr) const override { + return platform::MayIUse(platform::avx) && + attr.table_width % YMM_FLOAT_BLOCK == 0; + } + size_t CodeSize(const emb_seq_pool_attr_t& attr) const override { + return 96 + (attr.table_width / YMM_FLOAT_BLOCK) * 96 * 8; + } + std::unique_ptr CreateJitCode( + const emb_seq_pool_attr_t& attr) const override { + PADDLE_ENFORCE_GT(attr.table_height, 0); + PADDLE_ENFORCE_GT(attr.table_width, 0); + PADDLE_ENFORCE_GT(attr.index_height, 0); + PADDLE_ENFORCE_GT(attr.index_width, 0); + PADDLE_ENFORCE_GT(attr.out_width, 0); + return make_unique(attr, CodeSize(attr)); + } +}; + +} // namespace gen +} // namespace jit +} // namespace operators +} // namespace paddle + +namespace gen = paddle::operators::jit::gen; + +REGISTER_JITKERNEL_GEN(kEmbSeqPool, gen::EmbSeqPoolCreator); diff --git a/paddle/fluid/operators/jit/gen/embseqpool.h b/paddle/fluid/operators/jit/gen/embseqpool.h new file mode 100644 index 0000000000000000000000000000000000000000..5afcfbdc1786bef160864fcde06f8738207751be --- /dev/null +++ b/paddle/fluid/operators/jit/gen/embseqpool.h @@ -0,0 +1,81 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. */ + +#pragma once + +#include +#include "glog/logging.h" +#include "paddle/fluid/operators/jit/gen/jitcode.h" +#include "paddle/fluid/platform/enforce.h" + +namespace paddle { +namespace operators { +namespace jit { +namespace gen { + +class EmbSeqPoolJitCode : public JitCode { + public: + explicit EmbSeqPoolJitCode(const emb_seq_pool_attr_t& attr, + size_t code_size = 256 * 1024, + void* code_ptr = nullptr) + : JitCode(code_size, code_ptr), + tbl_w_(attr.table_width), + type_(attr.pool_type) { + if (type_ != SeqPoolType::kSum) { + LOG(FATAL) << "Only support sum pool yet "; + } + this->genCode(); + } + + std::string name() const override { + std::string base = "EmbSeqPoolJitCode"; + if (type_ == SeqPoolType::kSum) { + base += "_Sum"; + } else if (type_ == SeqPoolType::kAvg) { + base += "_Avg"; + } else if (type_ == SeqPoolType::kSqrt) { + base += "_Sqrt"; + } + base += ("_W" + std::to_string(tbl_w_)); + return base; + } + void genCode() override; + + private: + int tbl_w_; + SeqPoolType type_; + reg64_t param_tbl{abi_param1}; + reg64_t param_idx{abi_param2}; + reg64_t param_dst{abi_param3}; + reg64_t param_attr{abi_param4}; + + reg64_t reg_tmp{rax}; + + reg64_t reg_idx_width_in_byte{r8}; + reg64_t reg_idx_height{r9}; + + reg64_t reg_ptr_tbl_i{r10}; + reg64_t reg_idx{r10}; // could use same of reg_ptr_tbl_i + reg64_t reg_ptr_idx_i{r11}; + reg64_t reg_ptr_dst_i{r12}; + reg64_t reg_ptr_param_dst{r13}; // rdx is used in mul so protect param_dst + + reg64_t reg_idx_w_i_in_byte{r14}; + reg64_t reg_idx_h_end{r15}; +}; + +} // namespace gen +} // namespace jit +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/jit/gen/gru.h b/paddle/fluid/operators/jit/gen/gru.h index a4d7222a3459d175fc5eaf5cdf0e7a1a610f8b0c..d91f828e6aa7673265a460524dfcad119758aa77 100644 --- a/paddle/fluid/operators/jit/gen/gru.h +++ b/paddle/fluid/operators/jit/gen/gru.h @@ -49,7 +49,7 @@ class GRUJitCode : public VActFunc { this->genCode(); } - const char* name() const override { + std::string name() const override { std::string base = "GRUJitCode"; if (id_ == 0) { base += "_H1"; @@ -81,7 +81,7 @@ class GRUJitCode : public VActFunc { }; AddTypeStr(act_gate_); AddTypeStr(act_cand_); - return base.c_str(); + return base; } void genCode() override; diff --git a/paddle/fluid/operators/jit/gen/hopv.cc b/paddle/fluid/operators/jit/gen/hopv.cc new file mode 100644 index 0000000000000000000000000000000000000000..e7884017198623d996fe98a55691da6e342d656a --- /dev/null +++ b/paddle/fluid/operators/jit/gen/hopv.cc @@ -0,0 +1,103 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. */ + +#include "paddle/fluid/operators/jit/gen/hopv.h" +#include "paddle/fluid/operators/jit/registry.h" +#include "paddle/fluid/platform/cpu_info.h" + +namespace paddle { +namespace operators { +namespace jit { +namespace gen { + +void HOPVJitCode::genCode() { + const int num_blocks = num_ / YMM_FLOAT_BLOCK; + int offset = 0; + + if (num_blocks > 0) { + // load one firstly + vmovups(ymm_tmp, ptr[param_src]); + offset += sizeof(float) * YMM_FLOAT_BLOCK; + for (int i = 1; i < num_blocks; ++i) { + vmovups(ymm_src, ptr[param_src + offset]); + process(ymm_tmp, ymm_src, ymm_tmp); + offset += sizeof(float) * YMM_FLOAT_BLOCK; + } + vextractf128(xmm_dst, ymm_tmp, 1); + process(xmm_dst, xmm_dst, xmm_tmp); + } else { + if (type_ == operand_type::MAX) { + vbroadcastss(ymm_dst, ptr[param_src]); + } else if (type_ == operand_type::ADD) { + vxorps(ymm_dst, ymm_dst, ymm_dst); + } + } + + int rest = num_ % YMM_FLOAT_BLOCK; + if (rest >= 4) { + vmovups(xmm_src, ptr[param_src + offset]); + offset += sizeof(float) * 4; + rest -= 4; + process(xmm_dst, xmm_dst, xmm_src); + } + + vpermilps(xmm_tmp, xmm_dst, 16 + 8 + 3); + process(xmm_dst, xmm_dst, xmm_tmp); + + if (rest >= 2) { + vmovq(xmm_src, ptr[param_src + offset]); + offset += sizeof(float) * 2; + rest -= 2; + process(xmm_dst, xmm_dst, xmm_src); + } + + vpermilps(xmm_tmp, xmm_dst, 1); + process(xmm_dst, xmm_dst, xmm_tmp); + + if (rest >= 1) { + vmovss(xmm_src, ptr[param_src + offset]); + process(xmm_dst, xmm_dst, xmm_src); + } + vmovss(ptr[param_dst], xmm_dst); + ret(); +} + +#define DECLARE_HOP_CREATOR(name) \ + class name##Creator : public JitCodeCreator { \ + public: \ + bool UseMe(const int& attr) const override { \ + return platform::MayIUse(platform::avx); \ + } \ + size_t CodeSize(const int& d) const override { \ + return 96 + d / YMM_FLOAT_BLOCK * 4 * 8; \ + } \ + std::unique_ptr CreateJitCode(const int& attr) const override { \ + return make_unique(attr, CodeSize(attr)); \ + } \ + } + +DECLARE_HOP_CREATOR(HMax); +DECLARE_HOP_CREATOR(HSum); + +#undef DECLARE_HOP_CREATOR + +} // namespace gen +} // namespace jit +} // namespace operators +} // namespace paddle + +namespace gen = paddle::operators::jit::gen; + +REGISTER_JITKERNEL_GEN(kHMax, gen::HMaxCreator); +REGISTER_JITKERNEL_GEN(kHSum, gen::HSumCreator); diff --git a/paddle/fluid/operators/jit/gen/hopv.h b/paddle/fluid/operators/jit/gen/hopv.h new file mode 100644 index 0000000000000000000000000000000000000000..28d213e5e48749f84405454a2708d9289b9d290c --- /dev/null +++ b/paddle/fluid/operators/jit/gen/hopv.h @@ -0,0 +1,90 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. */ + +#pragma once + +#include +#include "glog/logging.h" +#include "paddle/fluid/operators/jit/gen/jitcode.h" + +namespace paddle { +namespace operators { +namespace jit { +namespace gen { + +// horizontal operand vector +class HOPVJitCode : public JitCode { + public: + explicit HOPVJitCode(int d, operand_type type, size_t code_size = 256 * 1024, + void* code_ptr = nullptr) + : JitCode(code_size, code_ptr), num_(d), type_(type) { + if (!(type_ == operand_type::MAX || type_ == operand_type::ADD)) { + LOG(FATAL) << "Do not support this operand type: " << type_; + } + this->genCode(); + } + + std::string name() const override { + std::string base = "VXXJitCode"; + if (type_ == operand_type::MAX) { + base += "_MAX"; + } else { + base += "_SUM"; + } + return base; + } + void genCode() override; + + protected: + template + void process(JMM& dst, JMM& src1, JMM& src2) { // NOLINT + if (type_ == operand_type::MAX) { + vmaxps(dst, src1, src2); + } else if (type_ == operand_type::ADD) { + vaddps(dst, src1, src2); + } + } + + private: + int num_; + operand_type type_; + reg64_t param_src{abi_param1}; + reg64_t param_dst{abi_param2}; + reg64_t param_attr{abi_param3}; + + ymm_t ymm_tmp = ymm_t(0); + ymm_t ymm_src = ymm_t(1); + ymm_t ymm_dst = ymm_t(2); + + xmm_t xmm_tmp = xmm_t(0); + xmm_t xmm_src = xmm_t(1); + xmm_t xmm_dst = xmm_t(2); +}; + +#define DECLARE_HOP_JITCODE(name, op_type) \ + class name##JitCode : public HOPVJitCode { \ + public: \ + explicit name##JitCode(int d, size_t code_size, void* code_ptr = nullptr) \ + : HOPVJitCode(d, op_type, code_size, code_ptr) {} \ + }; + +DECLARE_HOP_JITCODE(HMax, operand_type::MAX); +DECLARE_HOP_JITCODE(HSum, operand_type::ADD); + +#undef DECLARE_HOP_JITCODE + +} // namespace gen +} // namespace jit +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/jit/gen/jitcode.h b/paddle/fluid/operators/jit/gen/jitcode.h index f63d40ad5a559ab87a9b3735406671cfd936d9e4..689df8b1cbb7a928c9f9175d28a8231b56e2e82e 100644 --- a/paddle/fluid/operators/jit/gen/jitcode.h +++ b/paddle/fluid/operators/jit/gen/jitcode.h @@ -14,6 +14,7 @@ #pragma once +#include #include #include "paddle/fluid/operators/jit/gen_base.h" #include "paddle/fluid/platform/cpu_info.h" @@ -47,6 +48,7 @@ using Label = Xbyak::Label; typedef enum { MUL = 0, + MAX, ADD, SUB, RELU, @@ -58,7 +60,7 @@ typedef enum { } operand_type; #define DECLARE_JIT_CODE(codename) \ - const char* name() const override { return #codename; } + std::string name() const override { return #codename; } class JitCode : public GenBase, public Xbyak::CodeGenerator { public: @@ -67,7 +69,6 @@ class JitCode : public GenBase, public Xbyak::CodeGenerator { (code_size % 4096 != 0 ? (code_size / 4096 + 1) * 4096 : code_size), code_ptr) {} - virtual const char* name() const = 0; virtual void genCode() = 0; size_t getSize() const override { return CodeGenerator::getSize(); } diff --git a/paddle/fluid/operators/jit/gen/lstm.h b/paddle/fluid/operators/jit/gen/lstm.h index d4753bca23de91c74415d41c372cde1610712ef7..fa560b6230d7164be907f0172fb1d91860c05db2 100644 --- a/paddle/fluid/operators/jit/gen/lstm.h +++ b/paddle/fluid/operators/jit/gen/lstm.h @@ -53,7 +53,7 @@ class LSTMJitCode : public VActFunc { this->genCode(); } - const char* name() const override { + std::string name() const override { std::string base = "LSTMJitCode"; if (use_peephole_) { base += "_Peephole"; @@ -85,7 +85,7 @@ class LSTMJitCode : public VActFunc { AddTypeStr(act_gate_); AddTypeStr(act_cand_); AddTypeStr(act_cell_); - return base.c_str(); + return base; } void genCode() override; diff --git a/paddle/fluid/operators/jit/gen/matmul.cc b/paddle/fluid/operators/jit/gen/matmul.cc new file mode 100644 index 0000000000000000000000000000000000000000..ae3858eab20aeb80553d8fcec4088a6632c9c17d --- /dev/null +++ b/paddle/fluid/operators/jit/gen/matmul.cc @@ -0,0 +1,128 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. */ + +#include "paddle/fluid/operators/jit/gen/matmul.h" +#include // offsetof +#include + +#include "paddle/fluid/operators/jit/registry.h" +#include "paddle/fluid/platform/cpu_info.h" + +namespace paddle { +namespace operators { +namespace jit { +namespace gen { + +void MatMulJitCode::genCode() { + preCode(); + int block, rest; + const auto groups = packed_groups(n_, k_, &block, &rest); + PADDLE_ENFORCE_GT(groups.front(), 0); + + const int block_len = sizeof(float) * block; + const int x_reg_idx = (block == ZMM_FLOAT_BLOCK ? 32 : 16) - 1; + const int w_reg_idx = x_reg_idx - 1; + // from packed mov(reg_ptr_wgt, ptr[param_attr + offsetof(matmul_attr_t, + // packed_weight)]); + mov(reg_ptr_wgt, param_y); + size_t z_offset = 0; + size_t wgt_offset = 0; + for (size_t g = 0; g < groups.size(); ++g) { + size_t x_offset = 0; + for (int k = 0; k < k_; ++k) { + vbroadcastss(zmm_t(x_reg_idx), ptr[param_x + x_offset]); + // clean + if (k == 0) { + for (int i = 0; i < groups[g]; ++i) { + vxorps(zmm_t(i), zmm_t(i), zmm_t(i)); + } + } + for (int i = 0; i < groups[g]; ++i) { + vmovups(zmm_t(w_reg_idx), ptr[reg_ptr_wgt + wgt_offset]); + vfmadd231ps(zmm_t(i), zmm_t(w_reg_idx), zmm_t(x_reg_idx)); + wgt_offset += block_len; + } + // last one, save + if (k == k_ - 1) { + for (int i = 0; i < groups[g]; ++i) { + // only rest save should be careful + if (rest != 0 && g == groups.size() - 1 && i == groups[g] - 1) { + break; + } + vmovups(ptr[param_z + z_offset + i * block_len], zmm_t(i)); + } + } + x_offset += sizeof(float); + } + z_offset += block_len * groups[g]; + } + + if (rest != 0) { + // below should refine with mask + int reg_idx = groups.back() - 1; + z_offset = (n_ - rest) * sizeof(float); + int inner_block = 8; + while (rest > 0) { + if (rest >= 8) { + inner_block = 8; + vmovups(ptr[param_z + z_offset], ymm_t(reg_idx)); + // shift zmm of inner_block, change reg_idx if update + } else if (rest >= 4) { + inner_block = 4; + vmovups(ptr[param_z + z_offset], xmm_t(reg_idx)); + } else if (rest >= 2) { + inner_block = 2; + vmovq(ptr[param_z + z_offset], xmm_t(reg_idx)); + } else { + inner_block = 1; + vmovss(ptr[param_z + z_offset], xmm_t(reg_idx)); + } + z_offset += inner_block * sizeof(float); + rest -= inner_block; + } + } + + postCode(); +} + +class MatMulCreator : public JitCodeCreator { + public: + bool UseMe(const matmul_attr_t& attr) const override { + return attr.m == 1 && platform::MayIUse(platform::avx512f) && + attr.n % ZMM_FLOAT_BLOCK == 0 && attr.k < 512; + } + size_t CodeSize(const matmul_attr_t& attr) const override { + int block = YMM_FLOAT_BLOCK; + if (platform::MayIUse(platform::avx512f)) { + block = ZMM_FLOAT_BLOCK; + } + return 96 + 4 * attr.k * (attr.n / block + 1) * 8; + } + std::unique_ptr CreateJitCode( + const matmul_attr_t& attr) const override { + PADDLE_ENFORCE_GT(attr.m, 0); + PADDLE_ENFORCE_GT(attr.n, 0); + PADDLE_ENFORCE_GT(attr.k, 0); + return make_unique(attr, CodeSize(attr)); + } +}; + +} // namespace gen +} // namespace jit +} // namespace operators +} // namespace paddle + +namespace gen = paddle::operators::jit::gen; + +REGISTER_JITKERNEL_GEN(kMatMul, gen::MatMulCreator); diff --git a/paddle/fluid/operators/jit/gen/matmul.h b/paddle/fluid/operators/jit/gen/matmul.h new file mode 100644 index 0000000000000000000000000000000000000000..881cea581acc27a7aa7d395c041d40a4d3281947 --- /dev/null +++ b/paddle/fluid/operators/jit/gen/matmul.h @@ -0,0 +1,62 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. */ + +#pragma once + +#include // for malloc and free +#include +#include +#include "glog/logging.h" +#include "paddle/fluid/operators/jit/gen/jitcode.h" +#include "paddle/fluid/platform/enforce.h" + +namespace paddle { +namespace operators { +namespace jit { +namespace gen { + +class MatMulJitCode : public JitCode { + public: + explicit MatMulJitCode(const matmul_attr_t& attr, + size_t code_size = 256 * 1024, + void* code_ptr = nullptr) + : JitCode(code_size, code_ptr), m_(attr.m), n_(attr.n), k_(attr.k) { + PADDLE_ENFORCE_EQ(m_, 1, "Only support m==1 yet"); + this->genCode(); + } + + std::string name() const override { + std::string base = "MatMulJitCode"; + base = base + "_M" + std::to_string(m_) + "_N" + std::to_string(n_) + "_K" + + std::to_string(k_); + return base; + } + void genCode() override; + + private: + int m_, n_, k_; + + reg64_t param_x{abi_param1}; + reg64_t param_y{abi_param2}; + reg64_t param_z{abi_param3}; + reg64_t param_attr{abi_param4}; + reg64_t reg_tmp{rax}; + + reg64_t reg_ptr_wgt{r10}; +}; + +} // namespace gen +} // namespace jit +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/jit/gen/seqpool.h b/paddle/fluid/operators/jit/gen/seqpool.h index fcbbb3c84c562e2ba57110134bf07bb218b41edb..e909bc7c7939ee5cb7a2d367c7a452b96e6a91c2 100644 --- a/paddle/fluid/operators/jit/gen/seqpool.h +++ b/paddle/fluid/operators/jit/gen/seqpool.h @@ -32,13 +32,13 @@ class SeqPoolJitCode : public JitCode { : JitCode(code_size, code_ptr), w_(attr.w), type_(attr.type) { if (!(type_ == SeqPoolType::kSum || type_ == SeqPoolType::kAvg || type_ == SeqPoolType::kSqrt)) { - LOG(FATAL) << "Only support sum pool yet "; + LOG(FATAL) << "Only supported pool type: sum, avg and sqrt."; } fp_h_[0] = 1.f; this->genCode(); } - virtual const char* name() const { + std::string name() const override { std::string base = "SeqPoolJitCode"; if (type_ == SeqPoolType::kSum) { base += "_Sum"; @@ -48,7 +48,7 @@ class SeqPoolJitCode : public JitCode { base += "_Sqrt"; } base += ("_W" + std::to_string(w_)); - return base.c_str(); + return base; } void genCode() override; diff --git a/paddle/fluid/operators/jit/gen_base.cc b/paddle/fluid/operators/jit/gen_base.cc index 310da0c76f1ab251d788e54f2305f375f3fb4838..f3603875ad7bda1fc688f9c053e0d37f7bb31f02 100644 --- a/paddle/fluid/operators/jit/gen_base.cc +++ b/paddle/fluid/operators/jit/gen_base.cc @@ -16,6 +16,14 @@ #include #include #include +#include +#include "paddle/fluid/memory/allocation/cpu_allocator.h" // for posix_memalign +#include "paddle/fluid/platform/cpu_info.h" +#include "paddle/fluid/platform/enforce.h" + +#ifndef _WIN32 +#define posix_memalign_free free +#endif DEFINE_bool(dump_jitcode, false, "Whether to dump the jitcode to file"); @@ -38,6 +46,46 @@ void GenBase::dumpCode(const unsigned char* code) const { } } +void* GenBase::operator new(size_t size) { + void* ptr; + constexpr size_t alignment = 32ul; + PADDLE_ENFORCE_EQ(posix_memalign(&ptr, alignment, size), 0, + "GenBase Alloc %ld error!", size); + PADDLE_ENFORCE(ptr, "Fail to allocate GenBase CPU memory: size = %d .", size); + return ptr; +} + +void GenBase::operator delete(void* ptr) { posix_memalign_free(ptr); } + +std::vector packed_groups(int n, int k, int* block_out, int* rest_out) { + int block; + int max_num_regs; + if (platform::MayIUse(platform::avx512f)) { + block = ZMM_FLOAT_BLOCK; + max_num_regs = 32; + } else { + block = YMM_FLOAT_BLOCK; + max_num_regs = 16; + } + // one for x, one for y, others for z + const int max_used_regs_for_n = max_num_regs - 2; + const int aligned_n = n % block == 0 ? n : (n / block + 1) * block; + const int num_block = aligned_n / block; + const int num_groups = num_block / max_used_regs_for_n; + std::vector groups(num_groups, max_used_regs_for_n); + int rest_num_regs = num_block % max_used_regs_for_n; + if (rest_num_regs != 0) { + groups.push_back(rest_num_regs); + } + if (block_out) { + *block_out = block; + } + if (rest_out) { + *rest_out = n % block; + } + return groups; +} + } // namespace jit } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/jit/gen_base.h b/paddle/fluid/operators/jit/gen_base.h index 4af01a437670aa6a07d370ff23ed2abd369f69a3..a7c7a35a7ea35bd80333b04f001d4ab5b5d1e06b 100644 --- a/paddle/fluid/operators/jit/gen_base.h +++ b/paddle/fluid/operators/jit/gen_base.h @@ -16,6 +16,8 @@ #include #include // for unique_ptr +#include +#include #include "paddle/fluid/operators/jit/kernel_base.h" DECLARE_bool(dump_jitcode); @@ -27,7 +29,7 @@ namespace jit { class GenBase : public Kernel { public: virtual ~GenBase() = default; - virtual const char* name() const = 0; + virtual std::string name() const = 0; virtual size_t getSize() const = 0; virtual const unsigned char* getCodeInternal() = 0; template @@ -41,6 +43,11 @@ class GenBase : public Kernel { return reinterpret_cast(const_cast(code)); } + void* operator new(size_t size); + void operator delete(void* ptr); + void* operator new[](size_t size) { return operator new(size); } + void operator delete[](void* ptr) { operator delete(ptr); } + protected: void dumpCode(const unsigned char* code) const; }; @@ -67,6 +74,11 @@ class JitCodeCreator : public GenCreator { virtual std::unique_ptr CreateJitCode(const Attr& attr) const = 0; }; +// unify the method of packed groups +// output the packed groups which used in weights, the block size and rest size +std::vector packed_groups(int n, int k, int* block = nullptr, + int* rest = nullptr); + } // namespace jit } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/jit/helper.cc b/paddle/fluid/operators/jit/helper.cc index 5dbe22a81b4866bdf60a03710d8ffd0b7bcb597b..a76653613289892c4bb41596f998c5f4cc131fd7 100644 --- a/paddle/fluid/operators/jit/helper.cc +++ b/paddle/fluid/operators/jit/helper.cc @@ -14,6 +14,8 @@ #include "paddle/fluid/operators/jit/helper.h" #include // tolower +#include +#include #include "paddle/fluid/platform/enforce.h" namespace paddle { @@ -49,6 +51,10 @@ const char* to_string(KernelType kt) { ONE_CASE(kNCHW16CMulNC); ONE_CASE(kSeqPool); ONE_CASE(kMatMul); + ONE_CASE(kHMax); + ONE_CASE(kHSum); + ONE_CASE(kSoftmax); + ONE_CASE(kEmbSeqPool); default: PADDLE_THROW("Not support type: %d, or forget to add it.", kt); return "NOT JITKernel"; @@ -88,6 +94,41 @@ KernelType to_kerneltype(const std::string& act) { return kNone; } +template <> +void pack_weights(const float* src, float* dst, int n, int k) { + int block, rest; + const auto groups = packed_groups(n, k, &block, &rest); + std::for_each(groups.begin(), groups.end(), [&](int i) { + PADDLE_ENFORCE_GT(i, 0, "each element of groups should be larger than 0."); + }); + int sum = std::accumulate(groups.begin(), groups.end(), 0); + std::memset(dst, 0, k * sum * block * sizeof(float)); + PADDLE_ENFORCE_GE(sum * block, n, + "The packed n should be equal to or larger than n"); + + const int block_len = sizeof(float) * block; + int n_offset = 0; + + for (size_t g = 0; g < groups.size(); ++g) { + const float* from = src + n_offset; + for (int j = 0; j < k; ++j) { + size_t copy_sz = groups[g] * block_len; + if (g == groups.size() - 1 && rest != 0) { + copy_sz = (groups[g] - 1) * block_len + rest * sizeof(float); + } + std::memcpy(dst, from + j * n, copy_sz); + dst += groups[g] * block; + } + n_offset += groups[g] * block; + } +} + +template +typename std::enable_if::value>::type pack_weights( + const T* src, T* dst, int n, int k) { + PADDLE_THROW("Only support pack with float type."); +} + } // namespace jit } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/jit/helper.h b/paddle/fluid/operators/jit/helper.h index fbf34fc4b3db49596b6be0360c00e77c12fab9b8..07998588a5a560f9c2ad7cc765b66e76e87da6f6 100644 --- a/paddle/fluid/operators/jit/helper.h +++ b/paddle/fluid/operators/jit/helper.h @@ -118,6 +118,35 @@ typename KernelTuples::func_type Get( return GetRefer(); } +template +class KernelFuncs { + public: + KernelFuncs() = default; + static KernelFuncs& Cache() { + static thread_local KernelFuncs g_func_cache; + return g_func_cache; + } + + bool Has(int key) const { return funcs_.find(key) != funcs_.end(); } + + void Insert(int key, typename KernelTuples::func_type func) { + funcs_.emplace(key, func); + } + + typename KernelTuples::func_type At(int key) { + if (Has(key)) { + return funcs_.at(key); + } + auto func = Get(key); + Insert(key, func); + return func; + } + + private: + std::unordered_map funcs_; + DISABLE_COPY_AND_ASSIGN(KernelFuncs); +}; + const char* to_string(KernelType kt); const char* to_string(SeqPoolType kt); @@ -130,17 +159,37 @@ inline std::ostream& operator<<(std::ostream& os, const lstm_attr_t& attr) { << (attr.use_peephole ? "True" : "False") << "]"; return os; } + inline std::ostream& operator<<(std::ostream& os, const gru_attr_t& attr) { os << "dim_size[" << attr.d << "],act_gate[" << to_string(attr.act_gate) << "],act_cand[" << to_string(attr.act_cand) << "]"; return os; } + inline std::ostream& operator<<(std::ostream& os, const seq_pool_attr_t& attr) { os << "height_size[" << attr.h << "],width_size[" << attr.w << "],pool_type[" << to_string(attr.type) << "]"; return os; } +inline std::ostream& operator<<(std::ostream& os, + const emb_seq_pool_attr_t& attr) { + os << "table_height[" << attr.table_height << "],table_width[" + << attr.table_width << "],index_height[" << attr.index_height + << "],index_width[" << attr.index_width << "],output_width[" + << attr.out_width << "],pool_type[" << to_string(attr.pool_type) << "]"; + return os; +} + +inline std::ostream& operator<<(std::ostream& os, const matmul_attr_t& attr) { + os << "M[" << attr.m << "],N[" << attr.n << "],K[" << attr.k << "]"; + return os; +} + +// expose the method to pack matmul weight +template +void pack_weights(const T* src, T* dst, int n, int k); + } // namespace jit } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/jit/kernel_base.h b/paddle/fluid/operators/jit/kernel_base.h index adb101bd5cdf231ac330dbf44beb4c24c1fcf29e..20b6a32bef9860c52ab4423395a8e00f719b0210 100644 --- a/paddle/fluid/operators/jit/kernel_base.h +++ b/paddle/fluid/operators/jit/kernel_base.h @@ -13,6 +13,7 @@ * limitations under the License. */ #pragma once +#include #include "paddle/fluid/operators/jit/macro.h" #include "paddle/fluid/platform/macros.h" @@ -22,28 +23,33 @@ namespace jit { typedef enum { kNone = 0, - kVMul = 1, - kVAdd = 2, - kVAddRelu, - kVSub, - kVScal, - kVAddBias, - kVRelu, - kVIdentity, - kVSquare, - kVExp, - kVSigmoid, - kVTanh, - kLSTMCtHt, - kLSTMC1H1, + // sort by alphabet + kCRFDecoding = 1, + kEmbSeqPool = 2, kGRUH1, kGRUHtPart1, kGRUHtPart2, - kCRFDecoding, + kHSum, // horizontal max + kHMax, // horizontal sum + kLSTMCtHt, + kLSTMC1H1, kLayerNorm, + kMatMul, kNCHW16CMulNC, kSeqPool, - kMatMul, + kSoftmax, + kVAdd, + kVAddBias, + kVAddRelu, + kVExp, + kVIdentity, + kVMul, + kVRelu, + kVScal, + kVSigmoid, + kVSquare, + kVSub, + kVTanh, } KernelType; typedef enum { @@ -70,6 +76,10 @@ struct XYNTuples { typedef void (*func_type)(const T*, T*, int); }; +// x, return and int +template +struct XRNTuples : public XYNTuples {}; + typedef struct { void* gates; // gates: x_ch, x_ih, x_fh, x_oh const void* ct_1; @@ -137,11 +147,45 @@ struct SeqPoolTuples { typedef void (*func_type)(const T*, T*, const seq_pool_attr_t*); }; +typedef struct emb_seq_pool_attr_s { + int64_t table_height, table_width; + int64_t index_height, index_width; + int64_t out_width; + SeqPoolType pool_type; + emb_seq_pool_attr_s() = default; + explicit emb_seq_pool_attr_s(int64_t tbl_height, int64_t tbl_width, + int64_t idx_height, int64_t idx_width, + int64_t output_width, + SeqPoolType seqpool_type = SeqPoolType::kSum) + : table_height(tbl_height), + table_width(tbl_width), + index_height(idx_height), + index_width(idx_width), + out_width(output_width), + pool_type(seqpool_type) {} +} emb_seq_pool_attr_t; + +template +struct EmbSeqPoolTuples { + typedef T data_type; + typedef emb_seq_pool_attr_t attr_type; + typedef void (*func_type)(const T*, const int64_t*, T*, + const emb_seq_pool_attr_t*); +}; + +typedef struct matmul_attr_s { + int m, n, k; + void* packed_weight{nullptr}; + matmul_attr_s() = default; + explicit matmul_attr_s(int m_, int n_, int k_, void* packed_weight_ = nullptr) + : m(m_), n(n_), k(k_), packed_weight(packed_weight_) {} +} matmul_attr_t; + template struct MatMulTuples { typedef T data_type; - typedef int attr_type; - typedef void (*func_type)(const T*, const T*, T*, int, int, int); + typedef matmul_attr_t attr_type; + typedef void (*func_type)(const T*, const T*, T*, const matmul_attr_t*); }; template @@ -159,6 +203,13 @@ struct LayerNormTuples { const float, int); }; +template +struct SoftmaxTuples { + typedef T data_type; + typedef int attr_type; + typedef void (*func_type)(const T*, T*, int, int); +}; + // nChw16c = nChw16c .* NC template struct NCHW16CMulNCTuples { diff --git a/paddle/fluid/operators/jit/kernel_key.cc b/paddle/fluid/operators/jit/kernel_key.cc index 61de38688664f83775c0c4e5aa6f7e06c3602ddb..e659c6d254391f09ac8692e0b7602c65e1afd47d 100644 --- a/paddle/fluid/operators/jit/kernel_key.cc +++ b/paddle/fluid/operators/jit/kernel_key.cc @@ -49,6 +49,18 @@ size_t JitCodeKey(const seq_pool_attr_t& attr) { return (key << pool_type_shift) + static_cast(attr.type); } +template <> +size_t JitCodeKey(const matmul_attr_t& attr) { + size_t key = attr.m; + constexpr int shift = 21; + return (key << shift * 2) + ((static_cast(attr.n)) << shift) + attr.k; +} + +template <> +size_t JitCodeKey(const emb_seq_pool_attr_t& attr) { + return attr.table_width; +} + } // namespace jit } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/jit/more/mix/CMakeLists.txt b/paddle/fluid/operators/jit/more/mix/CMakeLists.txt index e05f204b1eebd03c7a00157d96d0482f4a44a7fb..dd039d29152961210958470a48f086a133ab640c 100644 --- a/paddle/fluid/operators/jit/more/mix/CMakeLists.txt +++ b/paddle/fluid/operators/jit/more/mix/CMakeLists.txt @@ -12,3 +12,4 @@ USE_JITKERNEL_MORE(kLSTMC1H1, mix) USE_JITKERNEL_MORE(kGRUH1, mix) USE_JITKERNEL_MORE(kGRUHtPart1, mix) USE_JITKERNEL_MORE(kGRUHtPart2, mix) +USE_JITKERNEL_MORE(kSoftmax, mix) diff --git a/paddle/fluid/operators/jit/more/mix/mix.cc b/paddle/fluid/operators/jit/more/mix/mix.cc index df0a85256b1f546d5f64be73925cf58b87a25bd7..0036d1c238b17768c4df61af22a85588990e1815 100644 --- a/paddle/fluid/operators/jit/more/mix/mix.cc +++ b/paddle/fluid/operators/jit/more/mix/mix.cc @@ -48,6 +48,32 @@ void VTanh(const T* x, T* y, int n) { compute_addbias(&b, y, y, n); } +void Softmax(const T* x, T* y, int n, int bs) { + auto compute_hmax = + KernelFuncs, platform::CPUPlace>::Cache().At(n); + auto compute_hsum = + KernelFuncs, platform::CPUPlace>::Cache().At(n); + auto compute_vscal = + KernelFuncs, platform::CPUPlace>::Cache().At(n); + auto compute_vaddbias = + KernelFuncs, platform::CPUPlace>::Cache().At(n); + auto compute_vexp = + KernelFuncs, platform::CPUPlace>::Cache().At(n); + + for (int i = 0; i < bs; ++i) { + T scalar; + compute_hmax(x, &scalar, n); + scalar = static_cast(0) - scalar; + compute_vaddbias(&scalar, x, y, n); // x - max + compute_vexp(y, y, n); + compute_hsum(y, &scalar, n); + scalar = static_cast(1) / scalar; + compute_vscal(&scalar, y, y, n); + x += n; + y += n; + } +} + void (*getActFunc(KernelType type, int d))(const T*, T*, int) { // NOLINT if (type == kVSigmoid) { return Get, platform::CPUPlace>(d); @@ -184,6 +210,8 @@ bool VSigmoidKernel::UseMe(const int& d) const { return true; } bool VTanhKernel::UseMe(const int& d) const { return true; } +bool SoftmaxKernel::UseMe(const int& d) const { return true; } + bool LSTMCtHtKernel::UseMe(const lstm_attr_t& attr) const { return true; } bool LSTMC1H1Kernel::UseMe(const lstm_attr_t& attr) const { return true; } @@ -207,6 +235,7 @@ namespace mix = paddle::operators::jit::more::mix; REGISTER_MORE_KERNEL(kVSigmoid, VSigmoid); REGISTER_MORE_KERNEL(kVTanh, VTanh); +REGISTER_MORE_KERNEL(kSoftmax, Softmax); REGISTER_MORE_KERNEL(kLSTMCtHt, LSTMCtHt); REGISTER_MORE_KERNEL(kLSTMC1H1, LSTMC1H1); REGISTER_MORE_KERNEL(kGRUH1, GRUH1); diff --git a/paddle/fluid/operators/jit/more/mix/mix.h b/paddle/fluid/operators/jit/more/mix/mix.h index a70ecdf9348f511311307b4c27bb4506222a7439..d64af192197a0b339a39a1862c028875da2f3900 100644 --- a/paddle/fluid/operators/jit/more/mix/mix.h +++ b/paddle/fluid/operators/jit/more/mix/mix.h @@ -26,6 +26,7 @@ using T = float; void VSigmoid(const T* x, T* y, int n); void VTanh(const T* x, T* y, int n); +void Softmax(const T* x, T* y, int n, int bs); void LSTMCtHt(lstm_t* step, const lstm_attr_t* attr); void LSTMC1H1(lstm_t* step, const lstm_attr_t* attr); @@ -45,6 +46,9 @@ void GRUHtPart2(gru_t* step, const gru_attr_t* attr); DECLARE_MORE_KERNEL(VSigmoid, XYNTuples); DECLARE_MORE_KERNEL(VTanh, XYNTuples); +// XRN +DECLARE_MORE_KERNEL(Softmax, SoftmaxTuples); + DECLARE_MORE_KERNEL(LSTMCtHt, LSTMTuples); DECLARE_MORE_KERNEL(LSTMC1H1, LSTMTuples); diff --git a/paddle/fluid/operators/jit/more/mkl/CMakeLists.txt b/paddle/fluid/operators/jit/more/mkl/CMakeLists.txt index 667c6dfad6676d00ab994564bff57c90caa0cb41..d209f31007255b3a90fdeeb4d609311b80bdc7b5 100644 --- a/paddle/fluid/operators/jit/more/mkl/CMakeLists.txt +++ b/paddle/fluid/operators/jit/more/mkl/CMakeLists.txt @@ -12,3 +12,5 @@ USE_JITKERNEL_MORE(kVSquare, mkl) USE_JITKERNEL_MORE(kVSigmoid, mkl) USE_JITKERNEL_MORE(kVTanh, mkl) USE_JITKERNEL_MORE(kSeqPool, mkl) +USE_JITKERNEL_MORE(kSoftmax, mkl) +USE_JITKERNEL_MORE(kEmbSeqPool, mkl) diff --git a/paddle/fluid/operators/jit/more/mkl/mkl.cc b/paddle/fluid/operators/jit/more/mkl/mkl.cc index fccdc68f5efa34bac6f5a34a41569d2f77416284..29a451f832fa745f8e1f5a45fd934f09e1f41e76 100644 --- a/paddle/fluid/operators/jit/more/mkl/mkl.cc +++ b/paddle/fluid/operators/jit/more/mkl/mkl.cc @@ -25,17 +25,19 @@ namespace more { namespace mkl { template <> -void MatMul(const float* a, const float* b, float* c, int m, int n, - int k) { - platform::dynload::cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans, m, - n, k, 1.f, a, k, b, n, 0.f, c, n); +void MatMul(const float* a, const float* b, float* c, + const matmul_attr_t* attr) { + platform::dynload::cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans, + attr->m, attr->n, attr->k, 1.f, a, attr->k, b, + attr->n, 0.f, c, attr->n); } template <> -void MatMul(const double* a, const double* b, double* c, int m, int n, - int k) { - platform::dynload::cblas_dgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans, m, - n, k, 1.0, a, k, b, n, 0.0, c, n); +void MatMul(const double* a, const double* b, double* c, + const matmul_attr_t* attr) { + platform::dynload::cblas_dgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans, + attr->m, attr->n, attr->k, 1.0, a, attr->k, b, + attr->n, 0.0, c, attr->n); } template <> @@ -116,12 +118,17 @@ void VAXPY(double a, const double* x, double* y, int n) { platform::dynload::cblas_daxpy(n, a, x, 1, y, 1); } -// TODO(TJ): tuning me carefully on AVX, AVX2 and AVX512 template <> -bool MatMulKernel::UseMe(const int& d) const { - return platform::MayIUse(platform::avx); +void ASum(const float* x, float* res, int n) { + res[0] = platform::dynload::cblas_sasum(n, x, 1); } +template <> +void ASum(const double* x, double* res, int n) { + res[0] = platform::dynload::cblas_dasum(n, x, 1); +} + +// TODO(TJ): tuning me carefully on AVX, AVX2 and AVX512 template <> bool VMulKernel::UseMe(const int& d) const { return platform::MayIUse(platform::avx512f) && d > 512; @@ -129,7 +136,7 @@ bool VMulKernel::UseMe(const int& d) const { template <> bool VAddKernel::UseMe(const int& d) const { - return platform::MayIUse(platform::avx512f) && d > 512; + return platform::MayIUse(platform::avx) && d > 512; } template <> @@ -167,13 +174,38 @@ bool SeqPoolKernel::UseMe(const seq_pool_attr_t& attr) const { return true; } +template <> +bool EmbSeqPoolKernel::UseMe(const emb_seq_pool_attr_t& attr) const { + return true; +} + +template <> +bool EmbSeqPoolKernel::UseMe(const emb_seq_pool_attr_t& attr) const { + return true; +} + +template <> +bool MatMulKernel::UseMe(const matmul_attr_t& attr) const { + return platform::MayIUse(platform::avx); +} + +template <> +bool MatMulKernel::UseMe(const matmul_attr_t& attr) const { + return true; +} + +template <> +bool SoftmaxKernel::UseMe(const int& d) const { + // tuned on avx2 + return platform::MayIUse(platform::avx) && d < 60; +} + #define AWALYS_USE_ME_WITH_DOUBLE(func) \ template <> \ bool func##Kernel::UseMe(const int& d) const { \ return true; \ } -AWALYS_USE_ME_WITH_DOUBLE(MatMul); AWALYS_USE_ME_WITH_DOUBLE(VMul); AWALYS_USE_ME_WITH_DOUBLE(VAdd); AWALYS_USE_ME_WITH_DOUBLE(VScal); @@ -181,6 +213,7 @@ AWALYS_USE_ME_WITH_DOUBLE(VExp); AWALYS_USE_ME_WITH_DOUBLE(VSigmoid); AWALYS_USE_ME_WITH_DOUBLE(VTanh); AWALYS_USE_ME_WITH_DOUBLE(VSquare); +AWALYS_USE_ME_WITH_DOUBLE(Softmax); #undef AWALYS_USE_ME_WITH_DOUBLE } // namespace mkl @@ -204,5 +237,7 @@ REGISTER_MKL_KERNEL(kVSquare, VSquare); REGISTER_MKL_KERNEL(kVSigmoid, VSigmoid); REGISTER_MKL_KERNEL(kVTanh, VTanh); REGISTER_MKL_KERNEL(kSeqPool, SeqPool); +REGISTER_MKL_KERNEL(kEmbSeqPool, EmbSeqPool); +REGISTER_MKL_KERNEL(kSoftmax, Softmax); #undef REGISTER_MKL_KERNEL diff --git a/paddle/fluid/operators/jit/more/mkl/mkl.h b/paddle/fluid/operators/jit/more/mkl/mkl.h index a27196fa19f1d3e9aa6c414b6b9f99a21ef49025..9a72ba83022de2beeb760772ee8489477befdd7e 100644 --- a/paddle/fluid/operators/jit/more/mkl/mkl.h +++ b/paddle/fluid/operators/jit/more/mkl/mkl.h @@ -16,7 +16,9 @@ #include #include +#include #include "paddle/fluid/operators/jit/kernel_base.h" +#include "paddle/fluid/platform/enforce.h" namespace paddle { namespace operators { @@ -25,7 +27,7 @@ namespace more { namespace mkl { template -void MatMul(const T* a, const T* b, T* c, int m, int n, int k); +void MatMul(const T* a, const T* b, T* c, const matmul_attr_t* attr); template void VMul(const T* x, const T* y, T* z, int n); @@ -90,6 +92,56 @@ void SeqPool(const T* x, T* y, const seq_pool_attr_t* attr) { } } +template +void EmbSeqPool(const T* table, const int64_t* idx, T* out, + const emb_seq_pool_attr_t* attr) { + PADDLE_ENFORCE_EQ(attr->table_width * attr->index_width, attr->out_width); + auto check_idx_value_valid = [&](int64_t i) { + PADDLE_ENFORCE_LT(idx[i], attr->table_height, "idx value: %d, i: %d", + idx[i], i); + PADDLE_ENFORCE_GE(idx[i], 0, "idx value: %d, i: %d", idx[i], i); + }; + + for (int64_t w = 0; w != attr->index_width; ++w) { + check_idx_value_valid(w); + VCopy(table + idx[w] * attr->table_width, out + w * attr->table_width, + attr->table_width); + } + + for (int64_t h = 1; h < attr->index_height; ++h) { + for (int64_t w = 0; w < attr->index_width; ++w) { + int64_t i = h * attr->index_width + w; + check_idx_value_valid(i); + VAXPY(static_cast(1), table + idx[i] * attr->table_width, + out + w * attr->table_width, attr->table_width); + } + } +} + +template +void ASum(const T* x, T* res, int n); + +template +void Softmax(const T* x, T* y, int n, int bs) { + std::vector entities(bs); + for (int i = 0; i < bs; ++i) { + entities[i] = x[i * n]; + for (int c = 1; c < n; ++c) { + entities[i] = x[i * n + c] > entities[i] ? x[i * n + c] : entities[i]; + } + for (int c = 0; c < n; ++c) { + y[i * n + c] = x[i * n + c] - entities[i]; + } + } + VExp(y, y, n * bs); + for (int i = 0; i < bs; ++i) { + T sum; + ASum(&y[i * n], &sum, n); + sum = static_cast(1) / sum; + VScal(&sum, &y[i * n], &y[i * n], n); + } +} + #define DECLARE_MKL_KERNEL(name, tuples) \ template \ class name##Kernel : public KernelMore> { \ @@ -117,6 +169,10 @@ DECLARE_MKL_KERNEL(VSquare, XYNTuples); DECLARE_MKL_KERNEL(SeqPool, SeqPoolTuples); +DECLARE_MKL_KERNEL(EmbSeqPool, EmbSeqPoolTuples); + +DECLARE_MKL_KERNEL(Softmax, SoftmaxTuples); + #undef DECLARE_MKL_KERNEL } // namespace mkl diff --git a/paddle/fluid/operators/jit/refer/CMakeLists.txt b/paddle/fluid/operators/jit/refer/CMakeLists.txt index 4b9bc5e8d49c62404d5d4ef99b7c50987fcb415a..218d801c084be455538628d1c1028d8e52142894 100644 --- a/paddle/fluid/operators/jit/refer/CMakeLists.txt +++ b/paddle/fluid/operators/jit/refer/CMakeLists.txt @@ -29,3 +29,7 @@ USE_JITKERNEL_REFER(kNCHW16CMulNC) USE_JITKERNEL_REFER(kSeqPool) USE_JITKERNEL_REFER(kMatMul) USE_JITKERNEL_REFER(kVSquare) +USE_JITKERNEL_REFER(kHSum) +USE_JITKERNEL_REFER(kHMax) +USE_JITKERNEL_REFER(kSoftmax) +USE_JITKERNEL_REFER(kEmbSeqPool) diff --git a/paddle/fluid/operators/jit/refer/refer.cc b/paddle/fluid/operators/jit/refer/refer.cc index 3512ad7fe7921381afb6152330fff6be34de5ad7..7e7dd6960b66e4e2f77eca6e96604f2a86553120 100644 --- a/paddle/fluid/operators/jit/refer/refer.cc +++ b/paddle/fluid/operators/jit/refer/refer.cc @@ -52,4 +52,11 @@ REGISTER_REFER_KERNEL(kSeqPool, SeqPool); REGISTER_REFER_KERNEL(kMatMul, MatMul); +REGISTER_REFER_KERNEL(kHMax, HMax); +REGISTER_REFER_KERNEL(kHSum, HSum); + +REGISTER_REFER_KERNEL(kSoftmax, Softmax); + +REGISTER_REFER_KERNEL(kEmbSeqPool, EmbSeqPool); + #undef REGISTER_REFER_KERNEL diff --git a/paddle/fluid/operators/jit/refer/refer.h b/paddle/fluid/operators/jit/refer/refer.h index 97d029358594d757f0e1874e9c87ecb8f97c9d50..fd1193aa41e50e3ede7f61588dc72389279bb95d 100644 --- a/paddle/fluid/operators/jit/refer/refer.h +++ b/paddle/fluid/operators/jit/refer/refer.h @@ -16,6 +16,7 @@ #include #include +#include #include "paddle/fluid/operators/jit/helper.h" #include "paddle/fluid/operators/jit/kernel_base.h" #include "paddle/fluid/platform/enforce.h" @@ -363,17 +364,84 @@ void SeqPool(const T* x, T* y, const seq_pool_attr_t* attr) { // A(M,K) * B(K,N) = C(M,N) template -void MatMul(const T* A, const T* B, T* C, int M, int N, int K) { +void MatMul(const T* A, const T* B, T* C, const matmul_attr_t* attr) { + int M = attr->m; + int N = attr->n; + int K = attr->k; for (int m = 0; m < M; ++m) { const T* pa = A + m * K; T* pc = C + m * N; for (int n = 0; n < N; ++n) { const T* pb = B + n; - T sum = static_cast(0); - for (int k = 0; k < K; ++k) { - sum += (pa[k] * pb[k * N]); + pc[n] = pa[0] * pb[0]; + for (int k = 1; k < K; ++k) { + pc[n] += pa[k] * pb[k * N]; } - *(pc + n) = sum; + } + } +} + +template +void HMax(const T* x, T* res, int n) { + res[0] = x[0]; + for (int i = 1; i < n; ++i) { + res[0] = res[0] < x[i] ? x[i] : res[0]; + } +} + +template +void HSum(const T* x, T* res, int n) { + res[0] = x[0]; + for (int i = 1; i < n; ++i) { + res[0] += x[i]; + } +} + +// y = e^(x - max(x)) +// y = y / sum(y) +template +void Softmax(const T* x, T* y, int n, int bs = 1) { + for (int i = 0; i < bs; ++i) { + T scalar; + HMax(x, &scalar, n); + scalar = static_cast(0) - scalar; + VAddBias(&scalar, x, y, n); // x - max + VExp(y, y, n); + HSum(y, &scalar, n); + scalar = static_cast(1) / scalar; + VScal(&scalar, y, y, n); + x += n; + y += n; + } +} + +// embedding seq pool +// table is a matrix with (tbl_h, tbl_w) +// idx is a matrix with (idx_h, idx_w) +// output is a vector with length tbl_w * idx_w +template +void EmbSeqPool(const T* table, const int64_t* idx, T* out, + const emb_seq_pool_attr_t* attr) { + PADDLE_ENFORCE_EQ(attr->table_width * attr->index_width, attr->out_width); + + auto check_idx_value_valid = [&](int64_t i) { + PADDLE_ENFORCE_LT(idx[i], attr->table_height, "idx value: %d, i: %d", + idx[i], i); + PADDLE_ENFORCE_GE(idx[i], 0, "idx value: %d, i: %d", idx[i], i); + }; + + for (int64_t w = 0; w != attr->index_width; ++w) { + check_idx_value_valid(w); + std::memcpy(out + w * attr->table_width, table + idx[w] * attr->table_width, + attr->table_width * sizeof(T)); + } + + for (int64_t h = 1; h < attr->index_height; ++h) { + for (int64_t w = 0; w < attr->index_width; ++w) { + int64_t i = h * attr->index_width + w; + check_idx_value_valid(i); + VAdd(table + idx[i] * attr->table_width, out + w * attr->table_width, + out + w * attr->table_width, attr->table_width); } } } @@ -421,6 +489,13 @@ DECLARE_REFER_KERNEL(SeqPool, SeqPoolTuples); DECLARE_REFER_KERNEL(MatMul, MatMulTuples); +DECLARE_REFER_KERNEL(HMax, XRNTuples); +DECLARE_REFER_KERNEL(HSum, XRNTuples); + +DECLARE_REFER_KERNEL(Softmax, SoftmaxTuples); + +DECLARE_REFER_KERNEL(EmbSeqPool, EmbSeqPoolTuples); + #undef DECLARE_REFER_KERNEL } // namespace refer diff --git a/paddle/fluid/operators/jit/test.cc b/paddle/fluid/operators/jit/test.cc index 68a79b6314e4cf86f5b715b9c6694924126b12da..356eba6f86ad180c7d23bf7fa91eb5d455ff5f08 100644 --- a/paddle/fluid/operators/jit/test.cc +++ b/paddle/fluid/operators/jit/test.cc @@ -1,16 +1,16 @@ /* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. */ + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this 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 @@ -22,7 +22,7 @@ #include "paddle/fluid/platform/cpu_info.h" #include "paddle/fluid/platform/place.h" -static double acc = 1e-5; +DEFINE_double(acc, 1e-5, "Test accuracy threshold."); template void RandomVec(const int n, T* a, const T lower = static_cast(-20.f), @@ -39,7 +39,7 @@ template void ExpectEQ(const T* target, const T* refer, int n) { if (std::is_floating_point::value) { for (int i = 0; i < n; ++i) { - EXPECT_NEAR(target[i], refer[i], acc); + EXPECT_NEAR(target[i], refer[i], FLAGS_acc); } } else { for (int i = 0; i < n; ++i) { @@ -61,6 +61,7 @@ std::vector TestSizes() { } namespace jit = paddle::operators::jit; +using CPUPlace = paddle::platform::CPUPlace; template struct TestFuncWithRefer { @@ -121,6 +122,40 @@ struct TestFuncWithRefer, T, std::vector, } }; +template +struct TestFuncWithRefer, std::vector, std::vector, + int, int> { + void operator()(const typename jit::SoftmaxTuples::func_type tgt, + const std::vector& x, const std::vector& yref, int n, + int bs) { + EXPECT_TRUE(tgt != nullptr); + EXPECT_EQ(yref.size(), x.size()); + EXPECT_EQ(x.size(), static_cast(n * bs)); + const T* x_data = x.data(); + const T* yref_data = yref.data(); + std::vector ytgt(n * bs); + T* ytgt_data = ytgt.data(); + // test normal + tgt(x_data, ytgt_data, n, bs); + ExpectEQ(ytgt_data, yref_data, n * bs); + // test inplace x + std::copy(x.begin(), x.end(), ytgt.begin()); + tgt(ytgt_data, ytgt_data, n, bs); + ExpectEQ(ytgt_data, yref_data, n * bs); + } +}; + +template +struct TestFuncWithRefer, std::vector, T> { + void operator()(const typename jit::XRNTuples::func_type tgt, + const std::vector& x, const T ref_res) { + EXPECT_TRUE(tgt != nullptr); + T tgt_res; + tgt(x.data(), &tgt_res, x.size()); + ExpectEQ(&tgt_res, &ref_res, 1); + } +}; + template struct TestFuncWithRefer, std::vector, std::vector> { void operator()(const typename jit::XYNTuples::func_type tgt, @@ -172,7 +207,7 @@ struct TestFuncWithRefer, std::vector, std::vector, T* ht_data = ht.data(); T* checked_data = checked.data(); - paddle::operators::jit::lstm_t step; + jit::lstm_t step; step.gates = x_data; step.ct_1 = ct_1_data; step.ct = ct_data; @@ -208,7 +243,7 @@ struct TestFuncWithRefer, std::vector, std::vector, const T* ht_ref_data = ht_ref.data(); T* x_data = x.data(); T* ht_data = ht.data(); - paddle::operators::jit::gru_t step; + jit::gru_t step; step.gates = x_data; step.ht_1 = ht_1_data; step.ht = ht_data; @@ -224,7 +259,7 @@ struct TestFuncWithRefer, std::vector, std::vector, const std::vector& x, const std::vector& yref, const typename jit::SeqPoolTuples::attr_type& attr) { EXPECT_TRUE(tgt != nullptr); - EXPECT_EQ(x.size() % yref.size(), 0); + EXPECT_EQ(x.size() % yref.size(), static_cast(0)); int w = yref.size(); std::vector y(w); const T* x_data = x.data(); @@ -235,28 +270,113 @@ struct TestFuncWithRefer, std::vector, std::vector, } }; +template +struct TestFuncWithRefer, std::vector, + std::vector, std::vector, + typename jit::EmbSeqPoolTuples::attr_type> { + void operator()(const typename jit::EmbSeqPoolTuples::func_type tgt, + const std::vector& table, const std::vector& idx, + const std::vector& oref, + const typename jit::EmbSeqPoolTuples::attr_type& attr) { + EXPECT_TRUE(tgt != nullptr); + EXPECT_EQ(table.size(), + static_cast(attr.table_height * attr.table_width)); + EXPECT_EQ(idx.size(), + static_cast(attr.index_height * attr.index_width)); + EXPECT_EQ(oref.size(), + static_cast(attr.table_width * attr.index_width)); + const T* table_data = table.data(); + const int64_t* idx_data = idx.data(); + const T* oref_data = oref.data(); + int o_w = oref.size(); + std::vector out(o_w); + T* o_data = out.data(); + tgt(table_data, idx_data, o_data, &attr); + ExpectEQ(o_data, oref_data, o_w); + } +}; + template struct TestFuncWithRefer, std::vector, std::vector, - std::vector, int, int, int> { + std::vector, + typename jit::MatMulTuples::attr_type> { void operator()(const typename jit::MatMulTuples::func_type tgt, const std::vector& a, const std::vector& b, - const std::vector& cref, int m, int n, int k) { + const std::vector& cref, + const typename jit::MatMulTuples::attr_type& attr) { EXPECT_TRUE(tgt != nullptr); - EXPECT_EQ(a.size(), static_cast(m * k)); - EXPECT_EQ(b.size(), static_cast(k * n)); - EXPECT_EQ(cref.size(), static_cast(m * n)); + EXPECT_EQ(a.size(), static_cast(attr.m * attr.k)); + EXPECT_EQ(b.size(), static_cast(attr.k * attr.n)); + EXPECT_EQ(cref.size(), static_cast(attr.m * attr.n)); std::vector c(cref.size()); const T* a_data = a.data(); const T* b_data = b.data(); const T* cref_data = cref.data(); T* c_data = c.data(); - tgt(a_data, b_data, c_data, m, n, k); - ExpectEQ(c_data, cref_data, m * n); + tgt(a_data, b_data, c_data, &attr); + ExpectEQ(c_data, cref_data, attr.m * attr.n); + } +}; + +template +struct TestFuncWithRefer, std::vector, + std::vector, std::vector, std::vector, + std::vector, std::vector, int, float, int> { + void operator()(const typename jit::LayerNormTuples::func_type tgt, + std::vector& x, std::vector& outref, // NOLINT + std::vector& mean, std::vector& var, // NOLINT + const std::vector& scale, const std::vector& bias, + int left, const float epsilon, int right) { + EXPECT_TRUE(tgt != nullptr); + EXPECT_EQ(x.size(), static_cast(left * right)); + EXPECT_EQ(outref.size(), static_cast(left * right)); + EXPECT_EQ(mean.size(), static_cast(left)); + EXPECT_EQ(var.size(), static_cast(left)); + EXPECT_EQ(scale.size(), static_cast(right)); + EXPECT_EQ(bias.size(), static_cast(right)); + std::vector outtgt(outref.size()); + const T* scale_data = scale.data(); + const T* bias_data = bias.data(); + T* x_data = x.data(); + T* mean_data = mean.data(); + T* var_data = var.data(); + T* outref_data = outref.data(); + T* outtgt_data = outtgt.data(); + + tgt(x_data, outtgt_data, mean_data, var_data, scale_data, bias_data, left, + epsilon, right); + ExpectEQ(outtgt_data, outref_data, left * right); } }; -template +template +struct TestFuncWithRefer, int, std::vector, + std::vector, std::vector, std::vector, + int> { + void operator()(const typename jit::CRFDecodingTuples::func_type tgt, + const int seq_len, const std::vector& x, + const std::vector& w, std::vector& alpharef, // NOLINT + std::vector& trackref, int tag_num) { // NOLINT + constexpr int state_trans_base_idx = 2; + EXPECT_TRUE(tgt != nullptr); + EXPECT_EQ(x.size(), static_cast(seq_len * tag_num)); + EXPECT_EQ(w.size(), + static_cast((tag_num + state_trans_base_idx) * tag_num)); + EXPECT_EQ(alpharef.size(), static_cast(seq_len * tag_num)); + EXPECT_EQ(trackref.size(), static_cast(seq_len * tag_num)); + std::vector alphatgt(alpharef.size()); + std::vector tracktgt(trackref.size()); + + memcpy(trackref.data(), tracktgt.data(), tag_num * sizeof(int)); + tgt(seq_len, (const T*)x.data(), (const T*)w.data(), alphatgt.data(), + tracktgt.data(), tag_num); + ExpectEQ(alpharef.data(), alphatgt.data(), seq_len * tag_num); + ExpectEQ(trackref.data(), tracktgt.data(), seq_len * tag_num); + } +}; + +template void TestAllImpls(const typename KernelTuples::attr_type& attr, Args... args) { TestFuncWithRefer test; // test jitcode @@ -286,9 +406,8 @@ void TestAllImpls(const typename KernelTuples::attr_type& attr, Args... args) { test(tgt, args...); } -template +template void TestXYZNKernel() { - namespace jit = paddle::operators::jit; VLOG(10) << "===== Test JITKernel " << jit::to_string(KT); for (int d : TestSizes()) { auto ref = jit::GetRefer>(); @@ -320,9 +439,8 @@ void TestXYZNKernel() { } } -template +template void TestAXYNKernel() { - namespace jit = paddle::operators::jit; VLOG(10) << "===== Test JITKernel " << jit::to_string(KT); for (int d : TestSizes()) { auto ref = jit::GetRefer>(); @@ -347,9 +465,26 @@ void TestAXYNKernel() { } } -template +template +void TestXRNKernel() { + VLOG(10) << "===== Test JITKernel " << jit::to_string(KT); + auto last_acc = FLAGS_acc; + FLAGS_acc = 1e-4; + for (int d : TestSizes()) { + auto ref = jit::GetRefer>(); + EXPECT_TRUE(ref != nullptr); + std::vector x(d); + RandomVec(d, x.data(), -2.f, 2.f); + T ref_res; + ref(x.data(), &ref_res, d); + TestAllImpls, PlaceType, std::vector, T>(d, x, + ref_res); + } + FLAGS_acc = last_acc; +} + +template void TestXYNKernel() { - namespace jit = paddle::operators::jit; VLOG(10) << "===== Test JITKernel " << jit::to_string(KT); for (int d : TestSizes()) { auto ref = jit::GetRefer>(); @@ -373,9 +508,8 @@ void TestXYNKernel() { } } -template +template void TestLSTMKernel() { - namespace jit = paddle::operators::jit; VLOG(10) << "===== Test JITKernel " << jit::to_string(KT); std::vector all_acts = {"sigmoid", "tanh", "relu", "identity"}; for (int d : TestSizes()) { @@ -424,9 +558,8 @@ void TestLSTMKernel() { } } -template +template void TestGRUKernel() { - namespace jit = paddle::operators::jit; VLOG(10) << "===== Test JITKernel " << jit::to_string(KT); std::vector all_acts = {"sigmoid", "tanh", "relu", "identity"}; for (int d : TestSizes()) { @@ -459,7 +592,7 @@ void TestGRUKernel() { } } -template +template void TestSeqPoolKernel() { VLOG(10) << "===== Test JITKernel " << jit::to_string(KT); std::vector pool_types = { @@ -484,12 +617,13 @@ void TestSeqPoolKernel() { } } -template +template void TestMatMulKernel() { VLOG(10) << "===== Test JITKernel " << jit::to_string(KT); - auto last_acc = acc; - // TODO(intel): this should be acc issue of MKL - acc = 1e-3; + auto last_acc = FLAGS_acc; + // TODO(intel): fix MKL acc issue + // https://github.com/PaddlePaddle/Paddle/issues/15447 + FLAGS_acc = 1e-3; for (int m : {1, 2, 3, 4}) { for (int n : {1, 2, 3, 4}) { for (int k : TestSizes()) { @@ -501,16 +635,76 @@ void TestMatMulKernel() { const T* a_data = a.data(); const T* b_data = b.data(); T* c_data = c.data(); - ref(a_data, b_data, c_data, m, n, k); + const jit::matmul_attr_t attr{m, n, k}; + ref(a_data, b_data, c_data, &attr); TestAllImpls, PlaceType, std::vector, - std::vector, std::vector>(k, a, b, c, m, n, k); + std::vector, std::vector>(attr, a, b, c, attr); } } } - acc = last_acc; + FLAGS_acc = last_acc; } -template +template +void TestSoftmaxKernel() { + VLOG(10) << "===== Test JITKernel " << jit::to_string(KT); + for (int bs : {1, 2, 10}) { + for (int n : TestSizes()) { + auto ref = jit::GetRefer>(); + EXPECT_TRUE(ref != nullptr); + std::vector x(bs * n), y(bs * n); + RandomVec(bs * n, x.data(), -2.f, 2.f); + const T* x_data = x.data(); + T* y_data = y.data(); + + std::vector xinp(x.size()); // inplace test + std::copy(x.begin(), x.end(), xinp.begin()); + ref(x_data, y_data, n, bs); + T* xinp_data = xinp.data(); + ref(xinp_data, xinp_data, n, bs); + ExpectEQ(xinp_data, y_data, n * bs); + + TestAllImpls, PlaceType, std::vector, + std::vector>(n, x, y, n, bs); + } + } +} + +template +void TestEmbSeqPoolKernel() { + VLOG(10) << "===== Test JITKernel " << jit::to_string(KT); + int64_t tbl_h = 1e4; + std::vector pool_types = { + jit::SeqPoolType::kSum}; // only support sum yet + for (int tbl_w : TestSizes()) { + std::vector table(tbl_h * tbl_w); + RandomVec(tbl_h * tbl_w, table.data(), -2.f, 2.f); + const T* table_data = table.data(); + for (auto type : pool_types) { + for (int idx_w : {1, 2, 10, 16}) { + for (int idx_h : {1, 2, 9, 13, 16}) { + auto ref = jit::GetRefer>(); + EXPECT_TRUE(ref != nullptr); + std::vector idx(idx_h * idx_w); + RandomVec(idx_h * idx_w, idx.data(), 0, tbl_h - 1); + int64_t out_w = tbl_w * idx_w; + std::vector oref(out_w); + const int64_t* idx_data = idx.data(); + T* o_data = oref.data(); + jit::emb_seq_pool_attr_t attr(tbl_h, tbl_w, idx_h, idx_w, out_w, + type); + ref(table_data, idx_data, o_data, &attr); + + TestAllImpls, PlaceType, std::vector, + std::vector, std::vector>(attr, table, idx, + oref, attr); + } + } + } + } +} + +template void TestNCHW16CMulNCKernel() { VLOG(10) << "===== Test JITKernel " << jit::to_string(KT); const int n = 3, c = 16 * 4, h = 10, w = 10; @@ -563,134 +757,207 @@ void TestNCHW16CMulNCKernel() { } } +template +void TestLayerNormKernel() { + VLOG(10) << "===== Test JITKernel " << jit::to_string(KT); + const T epsilon = 9.99999975e-06; + for (int n : {1, 2, 10}) { + for (int x_dim_0 : {1, 9, 17, 50}) { + int left = n * x_dim_0; + for (int x_dim_1 : TestSizes()) { + int right = x_dim_1; + auto ref = jit::GetRefer>(); + EXPECT_TRUE(ref != nullptr); + int sz = left * right; + std::vector x(sz), mean(left), var(left), scale(right), bias(right), + outref(sz); + RandomVec(sz, x.data(), -2.f, 2.f); + RandomVec(left, mean.data(), -2.f, 2.f); + RandomVec(left, var.data(), -2.f, 2.f); + RandomVec(right, scale.data(), -2.f, 2.f); + RandomVec(right, bias.data(), -2.f, 2.f); + + const T* scale_data = scale.data(); + const T* bias_data = bias.data(); + T* x_data = x.data(); + T* mean_data = mean.data(); + T* var_data = var.data(); + T* outref_data = outref.data(); + + ref(x_data, outref_data, mean_data, var_data, scale_data, bias_data, + left, epsilon, right); + + TestAllImpls, PlaceType, std::vector, + std::vector, std::vector, std::vector, + std::vector, std::vector, int, float>( + right, x, outref, mean, var, scale, bias, left, epsilon, right); + } + } + } +} + +template +void TestCRFDecodingKernel() { + VLOG(10) << "===== Test JITKernel " << jit::to_string(KT); + constexpr int state_trans_base_idx = 2; + for (int seq_len : {1, 11, 17, 50}) { + for (int tag_num : TestSizes()) { + auto ref = jit::GetRefer>(); + EXPECT_TRUE(ref != nullptr); + int x_sz = seq_len * tag_num; + int w_sz = (tag_num + state_trans_base_idx) * tag_num; + std::vector x(x_sz), w(w_sz), alpharef(x_sz); + std::vector trackref(x_sz); + RandomVec(x_sz, x.data(), -2.f, 2.f); + RandomVec(w_sz, w.data(), -2.f, 2.f); + + ref(seq_len, (const T*)x.data(), (const T*)w.data(), alpharef.data(), + trackref.data(), tag_num); + + TestAllImpls, PlaceType, int, + std::vector, std::vector, std::vector, + std::vector, int>(tag_num, seq_len, x, w, alpharef, + trackref, tag_num); + } + } +} + // XYZNTuple TEST(JITKernel, kVMul) { - namespace jit = paddle::operators::jit; - TestXYZNKernel(); - TestXYZNKernel(); + TestXYZNKernel(); + TestXYZNKernel(); } TEST(JITKernel, kVAdd) { - namespace jit = paddle::operators::jit; - TestXYZNKernel(); - TestXYZNKernel(); + TestXYZNKernel(); + TestXYZNKernel(); } TEST(JITKernel, kVAddRelu) { - namespace jit = paddle::operators::jit; - TestXYZNKernel(); - TestXYZNKernel(); + TestXYZNKernel(); + TestXYZNKernel(); } TEST(JITKernel, kVSub) { - namespace jit = paddle::operators::jit; - TestXYZNKernel(); - TestXYZNKernel(); + TestXYZNKernel(); + TestXYZNKernel(); } // AXYNTuples TEST(JITKernel, kVScal) { - namespace jit = paddle::operators::jit; - TestAXYNKernel(); - TestAXYNKernel(); + TestAXYNKernel(); + TestAXYNKernel(); } TEST(JITKernel, kVAddBias) { - namespace jit = paddle::operators::jit; - TestAXYNKernel(); - TestAXYNKernel(); + TestAXYNKernel(); + TestAXYNKernel(); +} + +// XRNTuples +TEST(JITKernel, kHMax) { + TestXRNKernel(); + TestXRNKernel(); +} + +TEST(JITKernel, kHSum) { + TestXRNKernel(); + TestXRNKernel(); } // XYNTuples TEST(JITKernel, kVRelu) { - namespace jit = paddle::operators::jit; - TestXYNKernel(); - TestXYNKernel(); + TestXYNKernel(); + TestXYNKernel(); } TEST(JITKernel, kVIdentity) { - namespace jit = paddle::operators::jit; - TestXYNKernel(); - TestXYNKernel(); + TestXYNKernel(); + TestXYNKernel(); } TEST(JITKernel, kVSquare) { - namespace jit = paddle::operators::jit; - TestXYNKernel(); - TestXYNKernel(); + TestXYNKernel(); + TestXYNKernel(); } TEST(JITKernel, kVExp) { - namespace jit = paddle::operators::jit; - TestXYNKernel(); - TestXYNKernel(); + TestXYNKernel(); + TestXYNKernel(); } TEST(JITKernel, kVSigmoid) { - namespace jit = paddle::operators::jit; - TestXYNKernel(); - TestXYNKernel(); + TestXYNKernel(); + TestXYNKernel(); } TEST(JITKernel, kVTanh) { - namespace jit = paddle::operators::jit; - TestXYNKernel(); - TestXYNKernel(); + TestXYNKernel(); + TestXYNKernel(); } // LSTM TEST(JITKernel, kLSTMCtHt) { - namespace jit = paddle::operators::jit; - TestLSTMKernel(); - TestLSTMKernel(); + TestLSTMKernel(); + TestLSTMKernel(); } TEST(JITKernel, kLSTMC1H1) { - namespace jit = paddle::operators::jit; - TestLSTMKernel(); - TestLSTMKernel(); + TestLSTMKernel(); + TestLSTMKernel(); } // GRU TEST(JITKernel, kGRUH1) { - namespace jit = paddle::operators::jit; - TestGRUKernel(); - TestGRUKernel(); + TestGRUKernel(); + TestGRUKernel(); } TEST(JITKernel, kGRUHtPart1) { - namespace jit = paddle::operators::jit; - TestGRUKernel(); - TestGRUKernel(); + TestGRUKernel(); + TestGRUKernel(); } TEST(JITKernel, kGRUHtPart2) { - namespace jit = paddle::operators::jit; - TestGRUKernel(); - TestGRUKernel(); + TestGRUKernel(); + TestGRUKernel(); } TEST(JITKernel, kSeqPool) { - namespace jit = paddle::operators::jit; - TestSeqPoolKernel(); - TestSeqPoolKernel(); + TestSeqPoolKernel(); + TestSeqPoolKernel(); } TEST(JITKernel, kMatMul) { - namespace jit = paddle::operators::jit; - TestMatMulKernel(); - TestMatMulKernel(); + TestMatMulKernel(); + TestMatMulKernel(); +} + +TEST(JITKernel, kSoftmax) { + TestSoftmaxKernel(); + TestSoftmaxKernel(); +} + +TEST(JITKernel, kEmbSeqPool) { + TestEmbSeqPoolKernel(); + TestEmbSeqPoolKernel(); } TEST(JITKernel, kNCHW16CMulNC) { - namespace jit = paddle::operators::jit; - TestNCHW16CMulNCKernel(); - TestNCHW16CMulNCKernel(); + TestNCHW16CMulNCKernel(); + TestNCHW16CMulNCKernel(); } -// TODO(yihua/TJ): add crf decoding and layer norm unit tests +TEST(JITKernel, kLayerNorm) { + TestLayerNormKernel(); + TestLayerNormKernel(); +} + +TEST(JITKernel, kCRFDecoding) { + TestCRFDecodingKernel(); + TestCRFDecodingKernel(); +} TEST(JITKernel, pool) { // TODO(TJ): add some test diff --git a/paddle/fluid/operators/layer_norm_op.cc b/paddle/fluid/operators/layer_norm_op.cc index f83fe355b85566d229a2673d8f27cfb5ca4831d5..b9db6daf0825b573bfc7f684266212f998c91627 100644 --- a/paddle/fluid/operators/layer_norm_op.cc +++ b/paddle/fluid/operators/layer_norm_op.cc @@ -44,11 +44,11 @@ class LayerNormOp : public framework::OperatorWithKernel { int left = static_cast(matrix_dim[0]); int right = static_cast(matrix_dim[1]); if (ctx->HasInput("Scale")) { - PADDLE_ENFORCE_EQ(ctx->GetInputDim("Scale").size(), 1UL); + PADDLE_ENFORCE_EQ(ctx->GetInputDim("Scale").size(), 1); PADDLE_ENFORCE_EQ(ctx->GetInputDim("Scale")[0], right); } if (ctx->HasInput("Bias")) { - PADDLE_ENFORCE_EQ(ctx->GetInputDim("Bias").size(), 1UL); + PADDLE_ENFORCE_EQ(ctx->GetInputDim("Bias").size(), 1); PADDLE_ENFORCE_EQ(ctx->GetInputDim("Bias")[0], right); } diff --git a/paddle/fluid/operators/linear_chain_crf_op.cc b/paddle/fluid/operators/linear_chain_crf_op.cc index 1da14631e35608d479e1b861228d52d6d57def79..e17b6cb59898524d793f3cc78a09232f5b664617 100644 --- a/paddle/fluid/operators/linear_chain_crf_op.cc +++ b/paddle/fluid/operators/linear_chain_crf_op.cc @@ -144,12 +144,12 @@ class LinearChainCRFOp : public framework::OperatorWithKernel { "Output(LogLikelihood) should be not null."); auto emission_dims = ctx->GetInputDim("Emission"); - PADDLE_ENFORCE_EQ(emission_dims.size(), 2UL, + PADDLE_ENFORCE_EQ(emission_dims.size(), 2, "The Input(Emission) should be a 2-D tensor."); PADDLE_ENFORCE(emission_dims[0], "An empty mini-batch is not allowed."); auto transition_dims = ctx->GetInputDim("Transition"); - PADDLE_ENFORCE_EQ(transition_dims.size(), 2UL, + PADDLE_ENFORCE_EQ(transition_dims.size(), 2, "The Input(Transition) should be a 2-D tensor."); PADDLE_ENFORCE_EQ( transition_dims[0] - 2, transition_dims[1], @@ -202,13 +202,13 @@ class LinearChainCRFGradOp : public framework::OperatorWithKernel { "Input(LogLikelihood@GRAD) shoudl be not null."); auto emission_exps_dims = ctx->GetInputDim("EmissionExps"); - PADDLE_ENFORCE_EQ(emission_exps_dims.size(), 2UL, + PADDLE_ENFORCE_EQ(emission_exps_dims.size(), 2, "The Input(EmissionExps) should be a 2-D tensor."); PADDLE_ENFORCE(emission_exps_dims[0], "An empty mini-batch is not allowed."); auto transition_exps_dims = ctx->GetInputDim("TransitionExps"); - PADDLE_ENFORCE_EQ(transition_exps_dims.size(), 2UL, + PADDLE_ENFORCE_EQ(transition_exps_dims.size(), 2, "The Input(TransitionExps) should be a 2-D tensor."); PADDLE_ENFORCE_EQ( transition_exps_dims[0] - 2, transition_exps_dims[1], diff --git a/paddle/fluid/operators/load_combine_op.cc b/paddle/fluid/operators/load_combine_op.cc index c4a2282e16483dbe78a32a4148c5bc4349dde3dc..f5c802986e0573e81b3ab6187b57657b52b37215 100644 --- a/paddle/fluid/operators/load_combine_op.cc +++ b/paddle/fluid/operators/load_combine_op.cc @@ -64,7 +64,7 @@ class LoadCombineOp : public framework::OperatorBase { auto *tensor = out_var->GetMutable(); // Error checking - PADDLE_ENFORCE(static_cast(buffer), "Cannot read more"); + PADDLE_ENFORCE(static_cast(*buffer), "Cannot read more"); // Get data from fin to tensor DeserializeFromStream(*buffer, tensor, dev_ctx); @@ -90,6 +90,10 @@ class LoadCombineOp : public framework::OperatorBase { tensor->ShareDataWith(fp16_tensor); } } + buffer->peek(); + PADDLE_ENFORCE(buffer->eof(), + "You are not allowed to load partial data via " + "load_combine_op, use load_op instead."); } }; diff --git a/paddle/fluid/operators/lookup_table_op.cu b/paddle/fluid/operators/lookup_table_op.cu index fd15539f7b6727496988c9b13d0d2551659a420a..0af8b9e69cfe09890f28ef2028baa19319a5c379 100644 --- a/paddle/fluid/operators/lookup_table_op.cu +++ b/paddle/fluid/operators/lookup_table_op.cu @@ -17,6 +17,7 @@ limitations under the License. */ #include "paddle/fluid/operators/lookup_table_op.h" #include "paddle/fluid/platform/assert.h" #include "paddle/fluid/platform/cuda_primitives.h" +#include "paddle/fluid/platform/float16.h" namespace paddle { namespace operators { @@ -193,8 +194,11 @@ class LookupTableGradCUDAKernel : public framework::OpKernel { } // namespace paddle namespace ops = paddle::operators; +namespace plat = paddle::platform; REGISTER_OP_CUDA_KERNEL(lookup_table, ops::LookupTableCUDAKernel, - ops::LookupTableCUDAKernel); + ops::LookupTableCUDAKernel, + ops::LookupTableCUDAKernel); REGISTER_OP_CUDA_KERNEL(lookup_table_grad, ops::LookupTableGradCUDAKernel, - ops::LookupTableGradCUDAKernel); + ops::LookupTableGradCUDAKernel, + ops::LookupTableGradCUDAKernel); diff --git a/paddle/fluid/operators/lookup_table_op.h b/paddle/fluid/operators/lookup_table_op.h index a7d0fd4856edc74237151c64f286d468ad86e7ca..56c6e37ae3c62e1f9af66ef6ed16111dc1e93d9d 100644 --- a/paddle/fluid/operators/lookup_table_op.h +++ b/paddle/fluid/operators/lookup_table_op.h @@ -129,6 +129,7 @@ class LookupTableGradKernel : public framework::OpKernel { "must be either LoDTensor or SelectedRows"); } + int64_t padding_idx = context.Attr("padding_idx"); bool is_sparse = context.Attr("is_sparse"); // Since paddings are not trainable and fixed in forward, the gradient of // paddings makes no sense and we don't deal with it in backward. @@ -187,10 +188,15 @@ class LookupTableGradKernel : public framework::OpKernel { memset(d_table_data, 0, d_table->numel() * sizeof(T)); for (int64_t i = 0; i < ids->numel(); ++i) { - PADDLE_ENFORCE_LT(ids_data[i], N); - PADDLE_ENFORCE_GE(ids_data[i], 0); - for (int j = 0; j < D; ++j) { - d_table_data[ids_data[i] * D + j] += d_output_data[i * D + j]; + if (padding_idx != kNoPadding && ids_data[i] == padding_idx) { + // the gradient of padding_idx should be 0, already done by memset, so + // do nothing. + } else { + PADDLE_ENFORCE_LT(ids_data[i], N); + PADDLE_ENFORCE_GE(ids_data[i], 0); + for (int j = 0; j < D; ++j) { + d_table_data[ids_data[i] * D + j] += d_output_data[i * D + j]; + } } } } diff --git a/paddle/fluid/operators/lstm_op.h b/paddle/fluid/operators/lstm_op.h index 7d62d2d020ec2e3a29ad8720a8f04fead3a90a63..ca998826dd0118ab4b1ecc23bed8ef882f1bcc92 100644 --- a/paddle/fluid/operators/lstm_op.h +++ b/paddle/fluid/operators/lstm_op.h @@ -151,9 +151,10 @@ class LSTMKernel : public framework::OpKernel { lstm_value.output_value = out_t.data(); lstm_value.state_value = cell_t.data(); lstm_value.state_active_value = cell_pre_act_t.data(); + T cell_clip = 0.0; math::LstmUnitFunctor::compute( - device_ctx, lstm_value, frame_size, cur_batch_size, gate_act, - cell_act, cand_act); + device_ctx, lstm_value, frame_size, cur_batch_size, cell_clip, + gate_act, cell_act, cand_act); lstm_value.prev_state_value = lstm_value.state_value; } @@ -311,10 +312,15 @@ class LSTMGradKernel : public framework::OpKernel { lstm_grad.prev_state_grad = c0_g ? ordered_c0_g.data() : nullptr; } + // lstm_value.output_value not used in bp, set to nullptr + // lstm_grad.state_active_grad not used in bp, set to nullptr + lstm_value.output_value = nullptr; + lstm_grad.state_active_grad = nullptr; int cur_batch_size = bend - bstart; + T cell_clip = 0.0; math::LstmUnitGradFunctor::compute( device_ctx, lstm_value, lstm_grad, frame_size, cur_batch_size, - gate_act, cell_act, cand_act); + cell_clip, gate_act, cell_act, cand_act); if (n > 0) { int pre_h_start = static_cast(batch_starts[n - 1]); diff --git a/paddle/fluid/operators/lstmp_op.cc b/paddle/fluid/operators/lstmp_op.cc index 7a62bc9f828e4d3485628747cdf52c60c5354144..2728aa8a4ee21a9e1fe3deddcdba4c35a6aba7bc 100644 --- a/paddle/fluid/operators/lstmp_op.cc +++ b/paddle/fluid/operators/lstmp_op.cc @@ -73,12 +73,6 @@ class LSTMPOp : public framework::OperatorWithKernel { PADDLE_ENFORCE(ctx->HasInput("C0"), "Input(C0) of LSTMP operator should not be null after " "Input(H0) provided."); - auto h_dims = ctx->GetInputDim("H0"); - auto c_dims = ctx->GetInputDim("C0"); - PADDLE_ENFORCE(h_dims == c_dims, - "The dimension of Input(H0) and Input(C0) " - "should be the same."); - ctx->SetOutputDim("OrderedP0", {h_dims[0], proj_dims[1]}); } auto b_dims = ctx->GetInputDim("Bias"); @@ -180,11 +174,6 @@ class LSTMPOpMaker : public framework::OpProtoAndCheckerMaker { "This LoDTensor is obtained in the forward and used in the " "backward.") .AsIntermediate(); - AddOutput("OrderedP0", - "(Tensor) the projection of the initial hidden state " - "H0. This is a tensor with shape (N x P), where N is the " - "batch size and P is the hidden size.") - .AsIntermediate(); AddAttr("use_peepholes", "(bool, defalut: True) " "whether to enable diagonal/peephole connections.") @@ -193,6 +182,16 @@ class LSTMPOpMaker : public framework::OpProtoAndCheckerMaker { "(bool, defalut: False) " "whether to compute reversed LSTMP.") .SetDefault(false); + AddAttr("cell_clip", + "(float, defalut: 0.0) " + "Clip for Tensor for cell state tensor when clip value is " + "greater than 0.0") + .SetDefault(0.0); + AddAttr("proj_clip", + "(float, defalut: 0.0) " + "Clip for Tensor for projection tensor when clip value is " + "greater than 0.0") + .SetDefault(0.0); AddAttr( "gate_activation", "(string, default: sigmoid)" diff --git a/paddle/fluid/operators/lstmp_op.h b/paddle/fluid/operators/lstmp_op.h index 370dd04d1449a8e211febf9a4f9e90e6f5008e20..c7d6e4205f8862526904e4fa767a2f4c4a2d8481 100644 --- a/paddle/fluid/operators/lstmp_op.h +++ b/paddle/fluid/operators/lstmp_op.h @@ -14,6 +14,7 @@ limitations under the License. */ #pragma once #include +#include #include "paddle/fluid/framework/eigen.h" #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/operators/activation_op.h" @@ -21,17 +22,50 @@ limitations under the License. */ #include "paddle/fluid/operators/math/detail/activation_functions.h" #include "paddle/fluid/operators/math/lstm_compute.h" #include "paddle/fluid/operators/math/sequence2batch.h" +#include "paddle/fluid/platform/transform.h" namespace paddle { namespace operators { using LoDTensor = framework::LoDTensor; using Tensor = framework::Tensor; +using platform::Transform; template using EigenMatrix = framework::EigenMatrix; +template +class _ClipFunctor { + public: + explicit _ClipFunctor(const T min, const T max) : min_(min), max_(max) {} + HOSTDEVICE T operator()(const T& x) const { + if (x < min_) + return min_; + else if (x > max_) + return max_; + else + return x; + } + + private: + T min_; + T max_; +}; + +template +class _ClipGradFunctor { + public: + explicit _ClipGradFunctor(const T min, const T max) : min_(min), max_(max) {} + HOSTDEVICE T operator()(const T& x, const T& y) const { + return (y > min_ && y < max_) ? x : 0; + } + + private: + T min_; + T max_; +}; + template inline void ReorderInitState(const DeviceContext& ctx, const framework::Tensor& src, @@ -67,9 +101,11 @@ class LSTMPKernel : public framework::OpKernel { auto* bias = ctx.Input("Bias"); auto* hidden_t0 = ctx.Input("H0"); - auto* ordered_proj0 = ctx.Output("OrderedP0"); auto* cell_t0 = ctx.Input("C0"); + auto proj_clip = static_cast(ctx.Attr("proj_clip")); + auto cell_clip = static_cast(ctx.Attr("cell_clip")); + auto* batch_gate = ctx.Output("BatchGate"); batch_gate->mutable_data(ctx.GetPlace()); auto* proj_out = ctx.Output("Projection"); @@ -110,6 +146,7 @@ class LSTMPKernel : public framework::OpKernel { } lstmp_value.prev_state_value = nullptr; Tensor ordered_c0; + Tensor ordered_h0; framework::Vector order(batch_gate->lod()[2]); @@ -169,18 +206,9 @@ class LSTMPKernel : public framework::OpKernel { // Since the batch computing for LSTMP reorders the input sequence // according to their length. The initialized hidden state also needs // to reorder. - - Tensor ordered_h0; - ordered_proj0->mutable_data(ctx.GetPlace()); ReorderInitState(device_ctx, *hidden_t0, order, &ordered_h0, true); - blas.MatMul(ordered_h0, false, *proj_weight, false, static_cast(1.0), - ordered_proj0, static_cast(0.0)); - if (proj_act != math::detail::ActivationType::kIdentity) { - auto proj0_dev = EigenMatrix::From(*ordered_proj0); - ActCompute(cell_act, place, proj0_dev, proj0_dev); - } - blas.MatMul(*ordered_proj0, false, *weight, false, static_cast(1.0), + blas.MatMul(ordered_h0, false, *weight, false, static_cast(1.0), &gate_t, static_cast(1.0)); } @@ -189,8 +217,8 @@ class LSTMPKernel : public framework::OpKernel { lstmp_value.state_value = cell_t.data(); lstmp_value.state_active_value = cell_pre_act_t.data(); math::LstmUnitFunctor::compute( - device_ctx, lstmp_value, frame_size, cur_batch_size, gate_act, - cell_act, cand_act); + device_ctx, lstmp_value, frame_size, cur_batch_size, cell_clip, + gate_act, cell_act, cand_act); lstmp_value.prev_state_value = lstmp_value.state_value; blas.MatMul(hidden_t, false, *proj_weight, false, static_cast(1.0), &proj_t, static_cast(0.0)); @@ -198,6 +226,14 @@ class LSTMPKernel : public framework::OpKernel { auto proj_t_dev = EigenMatrix::From(proj_t); ActCompute(cell_act, place, proj_t_dev, proj_t_dev); } + if (proj_clip && proj_clip > 0.0) { + T* x_data = proj_t.data(); + int64_t numel = proj_t.numel(); + Transform trans; + trans(ctx.template device_context(), x_data, + x_data + numel, x_data, + _ClipFunctor(-1.0 * proj_clip, proj_clip)); + } } math::Batch2LoDTensorFunctor to_seq; @@ -239,6 +275,9 @@ class LSTMPGradKernel : public framework::OpKernel { auto* proj_out = ctx.Input("Projection"); auto* cell_out = ctx.Input("Cell"); + auto proj_clip = static_cast(ctx.Attr("proj_clip")); + auto cell_clip = static_cast(ctx.Attr("cell_clip")); + auto* batch_gate = ctx.Input("BatchGate"); auto* batch_cell_pre_act = ctx.Input("BatchCellPreAct"); auto* batch_hidden = ctx.Input("BatchHidden"); @@ -253,7 +292,6 @@ class LSTMPGradKernel : public framework::OpKernel { auto* bias_g = ctx.Output(framework::GradVarName("Bias")); auto* h0 = ctx.Input("H0"); - auto* ordered_proj0 = ctx.Input("OrderedP0"); auto* c0 = ctx.Input("C0"); auto* h0_g = ctx.Output(framework::GradVarName("H0")); @@ -363,6 +401,17 @@ class LSTMPGradKernel : public framework::OpKernel { Tensor cur_proj = batch_proj.Slice(bstart, bend); Tensor proj_g = batch_proj_g.Slice(bstart, bend); + + if (proj_clip && proj_clip > 0.0) { + T* dx_data = proj_g.data(); + T* x_data = cur_proj.data(); + int64_t numel = proj_g.numel(); + Transform trans; + trans(ctx.template device_context(), dx_data, + dx_data + numel, x_data, dx_data, + _ClipGradFunctor(-1.0 * proj_clip, proj_clip)); + } + if (proj_act != math::detail::ActivationType::kIdentity) { auto cur_proj_dev = EigenMatrix::From(cur_proj); auto proj_g_dev = EigenMatrix::From(proj_g); @@ -405,9 +454,14 @@ class LSTMPGradKernel : public framework::OpKernel { } int cur_batch_size = bend - bstart; + // lstmp_value.output_value not used in bp, set to null + // lstmp_grad.state_active_grad not used in bp, set to null + lstmp_value.output_value = nullptr; + lstmp_grad.state_active_grad = nullptr; + math::LstmUnitGradFunctor::compute( device_ctx, lstmp_value, lstmp_grad, frame_size, cur_batch_size, - gate_act, cell_act, cand_act); + cell_clip, gate_act, cell_act, cand_act); if (n > 0) { int pre_h_start = static_cast(batch_starts[n - 1]); @@ -426,31 +480,14 @@ class LSTMPGradKernel : public framework::OpKernel { ReorderInitState(device_ctx, *h0, order, &ordered_h0, true); if (weight_g) { - blas.MatMul(*ordered_proj0, true, gate_g, false, - static_cast(1.0), weight_g, static_cast(1.0)); + blas.MatMul(ordered_h0, true, gate_g, false, static_cast(1.0), + weight_g, static_cast(1.0)); } } if (h0 && (h0_g || proj_weight_g)) { ordered_h0_g.mutable_data(h0_g->dims(), ctx.GetPlace()); - Tensor proj0_g; - proj0_g.Resize({in_dims[0], proj_weight->dims()[1]}); - proj0_g.mutable_data(ctx.GetPlace()); blas.MatMul(gate_g, false, *weight, true, static_cast(1.0), - &proj0_g, static_cast(0.0)); - if (proj_act != math::detail::ActivationType::kIdentity) { - auto proj0_dev = EigenMatrix::From(*ordered_proj0); - auto proj0_g_dev = EigenMatrix::From(proj0_g); - ActGradCompute(cell_act, place, proj0_dev, proj0_dev, proj0_g_dev, - proj0_g_dev); - } - if (h0_g) { - blas.MatMul(proj0_g, false, *proj_weight, true, static_cast(1.0), - &ordered_h0_g, static_cast(0.0)); - } - if (proj_weight_g) { - blas.MatMul(ordered_h0, true, proj0_g, false, static_cast(1.0), - proj_weight_g, static_cast(1.0)); - } + &ordered_h0_g, static_cast(0.0)); } } } diff --git a/paddle/fluid/operators/math/CMakeLists.txt b/paddle/fluid/operators/math/CMakeLists.txt index 6bbb7155dda9b2c844f793a63adb861c2ed956e8..d4837696241b8c4e3cca4f2afe872c6be559853c 100644 --- a/paddle/fluid/operators/math/CMakeLists.txt +++ b/paddle/fluid/operators/math/CMakeLists.txt @@ -37,8 +37,9 @@ math_library(concat_and_split) math_library(context_project DEPS im2col math_function) math_library(cross_entropy) math_library(cos_sim_functor) -math_library(depthwise_conv) +math_library(depthwise_conv DEPS cub) math_library(im2col) +math_library(sample_prob) math_library(sampler) math_library(gru_compute DEPS activation_functions math_function) @@ -53,7 +54,7 @@ math_library(sequence2batch) math_library(sequence_padding) math_library(sequence_pooling DEPS math_function jit_kernel_helper) math_library(sequence_scale) -math_library(softmax DEPS math_function) +math_library(softmax DEPS math_function jit_kernel_helper) math_library(beam_search DEPS math_function) math_library(matrix_bit_code) diff --git a/paddle/fluid/operators/math/beam_search.cc b/paddle/fluid/operators/math/beam_search.cc index fb7119273a734feba870fdabade6a4faa1d5e9a3..69971ef7423eff6bc3f8543a491edb6b0bbd00ca 100644 --- a/paddle/fluid/operators/math/beam_search.cc +++ b/paddle/fluid/operators/math/beam_search.cc @@ -29,8 +29,9 @@ class BeamSearchFunctor { const framework::LoDTensor *ids, const framework::LoDTensor *scores, framework::LoDTensor *selected_ids, - framework::LoDTensor *selected_scores, size_t level, - size_t beam_size, int end_id, bool is_accumulated) { + framework::LoDTensor *selected_scores, + framework::Tensor *parent_idx, size_t level, size_t beam_size, + int end_id, bool is_accumulated) { auto abs_lod = framework::ToAbsOffset(scores->lod()); auto &high_level = abs_lod[level]; @@ -57,11 +58,13 @@ class BeamSearchFunctor { std::vector({static_cast(num_instances), 1})); selected_ids->Resize(dims); selected_scores->Resize(dims); + parent_idx->Resize({static_cast(num_instances)}); auto *selected_ids_data = selected_ids->mutable_data(platform::CPUPlace()); auto *selected_scores_data = selected_scores->mutable_data(platform::CPUPlace()); + auto *parent_idx_data = parent_idx->mutable_data(platform::CPUPlace()); // fill in data std::vector low_level; @@ -69,6 +72,7 @@ class BeamSearchFunctor { for (auto &items : selected_items) { low_level.push_back(low_offset); for (auto &item : items) { + parent_idx_data[low_offset] = static_cast(low_level.size() - 1); selected_ids_data[low_offset] = item.id; selected_scores_data[low_offset] = item.score; low_offset++; diff --git a/paddle/fluid/operators/math/beam_search.cu b/paddle/fluid/operators/math/beam_search.cu index d94e3023ce537cb9fa456e079c4fa3cf57fb954d..61d021ef627f1ccd90b992c2078a7f3ca879422d 100644 --- a/paddle/fluid/operators/math/beam_search.cu +++ b/paddle/fluid/operators/math/beam_search.cu @@ -157,10 +157,10 @@ __device__ __forceinline__ bool PruneEndBeams(Triple* top_beam_local, } __device__ __forceinline__ void WriteBack( - int64_t* selected_ids, float* selected_scores, size_t* selected_offsets, - Triple* top_beam_local, const int seq_offset_start, - const int seq_offset_end, const int selected_seq_start, - const int selected_seq_length) { + int64_t* selected_ids, float* selected_scores, int* parent_idx, + size_t* selected_offsets, Triple* top_beam_local, + const int seq_offset_start, const int seq_offset_end, + const int selected_seq_start, const int selected_seq_length) { const int tid = threadIdx.x; // use 1 thread only for each sequence int global_index = selected_seq_start; for (int global_offset = seq_offset_start; global_offset < seq_offset_end; @@ -171,6 +171,7 @@ __device__ __forceinline__ void WriteBack( selected_ids[global_index] = static_cast(top_beam_local[local_index].id); selected_scores[global_index] = top_beam_local[local_index].score; + parent_idx[global_index] = static_cast(global_offset); global_index++; } } @@ -180,11 +181,11 @@ __device__ __forceinline__ void WriteBack( template __device__ void BeamSearchDetails( - int64_t* selected_ids, float* selected_scores, size_t* selected_offsets, - const int64_t* pre_ids, const float* pre_scores, const int64_t* ids, - const float* scores, const int seq_offset_start, const int seq_offset_end, - const int seq_width, int beam_size, int end_id, bool is_accumulated, - int num_used_threads) { + int64_t* selected_ids, float* selected_scores, int* parent_idx, + size_t* selected_offsets, const int64_t* pre_ids, const float* pre_scores, + const int64_t* ids, const float* scores, const int seq_offset_start, + const int seq_offset_end, const int seq_width, int beam_size, int end_id, + bool is_accumulated, int num_used_threads) { __shared__ Triple top_beam[MaxLength]; int num_items = 0; @@ -228,15 +229,15 @@ __device__ void BeamSearchDetails( selected_offsets[0] = 0; } - WriteBack(selected_ids, selected_scores, selected_offsets, top_beam_local, - seq_offset_start, seq_offset_end, selected_seq_start, - selected_seq_length); + WriteBack(selected_ids, selected_scores, parent_idx, selected_offsets, + top_beam_local, seq_offset_start, seq_offset_end, + selected_seq_start, selected_seq_length); } } template __global__ void BeamSearchKernel(int64_t* selected_ids, float* selected_scores, - size_t* selected_offsets, + int* parent_idx, size_t* selected_offsets, const int64_t* pre_ids, const float* pre_scores, const int64_t* ids, const float* scores, const size_t* seq_offsets, @@ -250,24 +251,25 @@ __global__ void BeamSearchKernel(int64_t* selected_ids, float* selected_scores, int seq_offset_end = static_cast(seq_offsets[seq_id + 1]); BeamSearchDetails( - selected_ids, selected_scores, selected_offsets, pre_ids, pre_scores, ids, - scores, seq_offset_start, seq_offset_end, seq_width, beam_size, end_id, - is_accumulated, num_used_threads); + selected_ids, selected_scores, parent_idx, selected_offsets, pre_ids, + pre_scores, ids, scores, seq_offset_start, seq_offset_end, seq_width, + beam_size, end_id, is_accumulated, num_used_threads); } template __global__ void BeamSearchKernelSingle( - int64_t* selected_ids, float* selected_scores, size_t* selected_offsets, - const int64_t* pre_ids, const float* pre_scores, const int64_t* ids, - const float* scores, const int seq_length, const int seq_width, - int beam_size, int end_id, bool is_accumulated, int num_used_threads) { + int64_t* selected_ids, float* selected_scores, int* parent_idx, + size_t* selected_offsets, const int64_t* pre_ids, const float* pre_scores, + const int64_t* ids, const float* scores, const int seq_length, + const int seq_width, int beam_size, int end_id, bool is_accumulated, + int num_used_threads) { const int seq_offset_start = 0; const int seq_offset_end = seq_length; BeamSearchDetails( - selected_ids, selected_scores, selected_offsets, pre_ids, pre_scores, ids, - scores, seq_offset_start, seq_offset_end, seq_width, beam_size, end_id, - is_accumulated, num_used_threads); + selected_ids, selected_scores, parent_idx, selected_offsets, pre_ids, + pre_scores, ids, scores, seq_offset_start, seq_offset_end, seq_width, + beam_size, end_id, is_accumulated, num_used_threads); } static inline int GetNumUsedThreads(const int max_threads_per_seq, @@ -300,8 +302,9 @@ class BeamSearchFunctor { const framework::LoDTensor* ids, const framework::LoDTensor* scores, framework::LoDTensor* selected_ids, - framework::LoDTensor* selected_scores, size_t level, - size_t beam_size, int end_id, bool is_accumulated) { + framework::LoDTensor* selected_scores, + framework::Tensor* parent_idx, size_t level, size_t beam_size, + int end_id, bool is_accumulated) { auto abs_lod = framework::ToAbsOffset(scores->lod()); const int64_t* pre_ids_data = pre_ids->data(); @@ -322,6 +325,8 @@ class BeamSearchFunctor { selected_ids->mutable_data(selected_dims, context.GetPlace()); float* selected_scores_data = selected_scores->mutable_data(selected_dims, context.GetPlace()); + int* parent_idx_data = parent_idx->mutable_data( + {static_cast(num_seqs * beam_size)}, context.GetPlace()); framework::LoD selected_lod(2); selected_lod[0].assign(abs_lod[level].begin(), abs_lod[level].end()); @@ -339,9 +344,9 @@ class BeamSearchFunctor { CUDA_LAUNCH_KERNEL_HELPER( BeamSearchKernelSingle<<< 1, kMaxThreadsPerSeq, 0, context.stream()>>>( - selected_ids_data, selected_scores_data, selected_offsets, - pre_ids_data, pre_scores_data, ids_data, scores_data, - seq_length, static_cast(seq_width), + selected_ids_data, selected_scores_data, parent_idx_data, + selected_offsets, pre_ids_data, pre_scores_data, ids_data, + scores_data, seq_length, static_cast(seq_width), static_cast(beam_size), static_cast(end_id), is_accumulated, num_used_threads)); } @@ -357,9 +362,9 @@ class BeamSearchFunctor { CUDA_LAUNCH_KERNEL_HELPER( BeamSearchKernel<<< 1, num_seqs * kMaxThreadsPerSeq, 0, context.stream()>>>( - selected_ids_data, selected_scores_data, selected_offsets, - pre_ids_data, pre_scores_data, ids_data, scores_data, - seq_offsets, static_cast(num_seqs), + selected_ids_data, selected_scores_data, parent_idx_data, + selected_offsets, pre_ids_data, pre_scores_data, ids_data, + scores_data, seq_offsets, static_cast(num_seqs), static_cast(seq_width), static_cast(beam_size), end_id, is_accumulated, num_used_threads)); } @@ -379,6 +384,7 @@ class BeamSearchFunctor { {static_cast(selected_lod[1].back()), 1}); selected_ids->Resize(final_selected_dims); selected_scores->Resize(final_selected_dims); + parent_idx->Resize({static_cast(selected_lod[1].back())}); } } }; diff --git a/paddle/fluid/operators/math/beam_search.h b/paddle/fluid/operators/math/beam_search.h index 3cd17f426c5596582c91f2b3f0cc5ba513e3aa4b..4474e7ea52affed792572d02202ec2577c471e50 100644 --- a/paddle/fluid/operators/math/beam_search.h +++ b/paddle/fluid/operators/math/beam_search.h @@ -104,14 +104,12 @@ class BeamSearchFunctor { * Return false if all the input tensor is empty, in machine translation task * that means no candidates is provided, and the task will stop running. */ - void operator()(const DeviceContext& context, - const framework::LoDTensor* pre_ids, - const framework::LoDTensor* pre_scores, - const framework::LoDTensor* ids, - const framework::LoDTensor* scores, - framework::LoDTensor* selected_ids, - framework::LoDTensor* selected_scores, size_t level, - size_t beam_size, int end_id, bool is_accumulated); + void operator()( + const DeviceContext& context, const framework::LoDTensor* pre_ids, + const framework::LoDTensor* pre_scores, const framework::LoDTensor* ids, + const framework::LoDTensor* scores, framework::LoDTensor* selected_ids, + framework::LoDTensor* selected_scores, framework::Tensor* parent_idx, + size_t level, size_t beam_size, int end_id, bool is_accumulated); }; } // namespace math diff --git a/paddle/fluid/operators/math/beam_search_test.cc b/paddle/fluid/operators/math/beam_search_test.cc index 1c29ee95f6b109209316e4e8c8f3cda37eac62ae..7ea8eb8b00db328ca13d3d33d751aca4eac66dae 100644 --- a/paddle/fluid/operators/math/beam_search_test.cc +++ b/paddle/fluid/operators/math/beam_search_test.cc @@ -93,13 +93,14 @@ void TestBeamSearch() { paddle::framework::LoDTensor selected_ids; paddle::framework::LoDTensor selected_scores; + paddle::framework::LoDTensor parent_idx; size_t level = 0; size_t beam_size = 2; int end_id = 0; paddle::operators::math::BeamSearchFunctor beamsearch; beamsearch(*context, &pre_ids, &pre_scores, &ids, &scores, &selected_ids, - &selected_scores, level, beam_size, end_id, true); + &selected_scores, &parent_idx, level, beam_size, end_id, true); ASSERT_EQ(selected_ids.lod(), selected_scores.lod()); diff --git a/paddle/fluid/operators/math/detail/lstm_cpu_kernel.h b/paddle/fluid/operators/math/detail/lstm_cpu_kernel.h index 2e3779ff0845294e71f27801049c010e0a585e6b..ad79c58063a8a12c703979fe32a8e671a5ade857 100644 --- a/paddle/fluid/operators/math/detail/lstm_cpu_kernel.h +++ b/paddle/fluid/operators/math/detail/lstm_cpu_kernel.h @@ -32,7 +32,8 @@ namespace detail { template void naive_lstm_forward_one_sequence(Op op, LstmMetaValue value, - int frame_size, ActivationType active_node, + int frame_size, T cell_clip, + ActivationType active_node, ActivationType active_gate, ActivationType active_state) { T r_value_in; @@ -67,7 +68,7 @@ void naive_lstm_forward_one_sequence(Op op, LstmMetaValue value, 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); + &cell_clip, active_node, active_gate, active_state); value_in[i] = r_value_in; value_ig[i] = r_value_ig; @@ -82,7 +83,7 @@ void naive_lstm_forward_one_sequence(Op op, LstmMetaValue value, template void naive_lstm_backward_one_sequence(Op op, LstmMetaValue value, LstmMetaGrad grad, int frame_size, - ActivationType active_node, + T cell_clip, ActivationType active_node, ActivationType active_gate, ActivationType active_state) { T r_value_in; @@ -135,7 +136,7 @@ void naive_lstm_backward_one_sequence(Op op, LstmMetaValue value, &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); + &cell_clip, active_node, active_gate, active_state); grad_in[i] = r_grad_in; grad_ig[i] = r_grad_ig; @@ -154,7 +155,8 @@ void naive_lstm_backward_one_sequence(Op op, LstmMetaValue value, template void avx_lstm_forward_one_sequence(Op op, LstmMetaValue value, - int frame_size, ActivationType active_node, + int frame_size, T cell_clip, + ActivationType active_node, ActivationType active_gate, ActivationType active_state) { #ifdef __AVX__ @@ -194,7 +196,7 @@ void avx_lstm_forward_one_sequence(Op op, LstmMetaValue value, 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); + &cell_clip, active_node, active_gate, active_state); value_in[i] = r_value_in; value_ig[i] = r_value_ig; @@ -210,7 +212,7 @@ void avx_lstm_forward_one_sequence(Op op, LstmMetaValue value, template void avx_lstm_backward_one_sequence(Op op, LstmMetaValue value, LstmMetaGrad grad, int frame_size, - ActivationType active_node, + T cell_clip, ActivationType active_node, ActivationType active_gate, ActivationType active_state) { #ifdef __AVX__ @@ -268,7 +270,7 @@ void avx_lstm_backward_one_sequence(Op op, LstmMetaValue value, &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); + &cell_clip, active_node, active_gate, active_state); grad_in[i] = r_grad_in; grad_ig[i] = r_grad_ig; @@ -292,27 +294,27 @@ void avx_lstm_backward_one_sequence(Op op, LstmMetaValue value, template void cpu_lstm_forward(Op op, LstmMetaValue value, int frame_size, - ActivationType active_node, ActivationType active_gate, - ActivationType active_state) { + T cell_clip, 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); + avx_lstm_forward_one_sequence(op, value, frame_size, cell_clip, + active_node, active_gate, active_state); } else { - naive_lstm_forward_one_sequence(op, value, frame_size, active_node, - active_gate, active_state); + naive_lstm_forward_one_sequence(op, value, frame_size, cell_clip, + active_node, active_gate, active_state); } } template void cpu_lstm_backward(Op op, LstmMetaValue value, LstmMetaGrad grad, - int frame_size, ActivationType active_node, + int frame_size, T cell_clip, 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); + avx_lstm_backward_one_sequence(op, value, grad, frame_size, cell_clip, + active_node, active_gate, active_state); } else { - naive_lstm_backward_one_sequence(op, value, grad, frame_size, + naive_lstm_backward_one_sequence(op, value, grad, frame_size, cell_clip, active_node, active_gate, active_state); } } diff --git a/paddle/fluid/operators/math/detail/lstm_gpu_kernel.h b/paddle/fluid/operators/math/detail/lstm_gpu_kernel.h index 2aecb69237fdf344ebc0bfe72d9c7c147f06358d..e0ca9e7f5b2f4a8bb837768d645b5103aa3e6760 100644 --- a/paddle/fluid/operators/math/detail/lstm_gpu_kernel.h +++ b/paddle/fluid/operators/math/detail/lstm_gpu_kernel.h @@ -31,7 +31,8 @@ namespace detail { */ template __global__ void KeLstmForward(Op op, LstmMetaValue value, int frame_size, - int batch_size, ActivationType active_node, + int batch_size, T cell_clip, + ActivationType active_node, ActivationType active_gate, ActivationType active_state) { const int frame_idx = blockIdx.x * blockDim.x + threadIdx.x; @@ -72,7 +73,7 @@ __global__ void KeLstmForward(Op op, LstmMetaValue value, int frame_size, 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); + &cell_clip, active_node, active_gate, active_state); value.gate_value[frame_idx] = r_value_in; value.gate_value[frame_idx + frame_size] = r_value_ig; @@ -91,7 +92,8 @@ __global__ void KeLstmForward(Op op, LstmMetaValue value, int frame_size, template __global__ void KeLstmBackward(Op op, LstmMetaValue value, LstmMetaGrad grad, int frame_size, - int batch_size, ActivationType active_node, + int batch_size, T cell_clip, + ActivationType active_node, ActivationType active_gate, ActivationType active_state) { const int frame_idx = blockIdx.x * blockDim.x + threadIdx.x; @@ -148,8 +150,8 @@ __global__ void KeLstmBackward(Op op, LstmMetaValue value, 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); + &r_checkO, &r_checkIGrad, &r_checkFGrad, &r_checkOGrad, &cell_clip, + active_node, active_gate, active_state); grad.gate_grad[frame_idx] = r_grad_in; grad.gate_grad[frame_idx + frame_size] = r_grad_ig; @@ -185,8 +187,8 @@ __global__ void KeLstmBackward(Op op, LstmMetaValue value, template void gpu_lstm_forward(const platform::DeviceContext& context, Op op, LstmMetaValue value, int frame_size, int batch_size, - ActivationType active_node, ActivationType active_gate, - ActivationType active_state) { + T cell_clip, ActivationType active_node, + ActivationType active_gate, ActivationType active_state) { dim3 threads; dim3 grid; if (batch_size == 1) { @@ -205,12 +207,12 @@ void gpu_lstm_forward(const platform::DeviceContext& context, Op op, if (batch_size == 1) { KeLstmForward<<>>( - op, value, frame_size, batch_size, active_node, active_gate, + op, value, frame_size, batch_size, cell_clip, active_node, active_gate, active_state); } else { KeLstmForward<<>>( - op, value, frame_size, batch_size, active_node, active_gate, + op, value, frame_size, batch_size, cell_clip, active_node, active_gate, active_state); } } @@ -218,7 +220,7 @@ 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 frame_size, int batch_size, + int frame_size, int batch_size, T cell_clip, ActivationType active_node, ActivationType active_gate, ActivationType active_state) { dim3 threads; @@ -239,13 +241,13 @@ void gpu_lstm_backward(const platform::DeviceContext& context, Op op, if (batch_size == 1) { KeLstmBackward<<>>( - op, value, grad, frame_size, batch_size, active_node, active_gate, - active_state); + op, value, grad, frame_size, batch_size, cell_clip, active_node, + active_gate, active_state); } else { KeLstmBackward<<>>( - op, value, grad, frame_size, batch_size, active_node, active_gate, - active_state); + op, value, grad, frame_size, batch_size, cell_clip, active_node, + active_gate, active_state); } } diff --git a/paddle/fluid/operators/math/detail/lstm_kernel.h b/paddle/fluid/operators/math/detail/lstm_kernel.h index cbe73d62938d7c4c03a2c8731665260624417fd7..8149686c97a030b91e0c4de708b9abf07f83203d 100644 --- a/paddle/fluid/operators/math/detail/lstm_kernel.h +++ b/paddle/fluid/operators/math/detail/lstm_kernel.h @@ -29,7 +29,7 @@ class lstm { public: HOSTDEVICE void operator()(T *value_in, T *value_ig, T *value_fg, T *value_og, T *prev_state, T *state, T *state_atv, T *output, - T *checkI, T *checkF, T *checkO, + T *checkI, T *checkF, T *checkO, T *cell_clip, ActivationType active_node, ActivationType active_gate, ActivationType active_state) { @@ -37,6 +37,15 @@ class lstm { *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); + + if (*cell_clip > 0.0) { + if (*state < -1.0 * (*cell_clip)) { + *state = -1.0 * (*cell_clip); + } + if (*state > *cell_clip) { + *state = *cell_clip; + } + } *value_og = activation(*value_og + (*state) * (*checkO), active_gate); *state_atv = activation(*state, active_state); *output = (*value_og) * (*state_atv); @@ -52,7 +61,7 @@ class lstm { __m256 *value_fg, __m256 *value_og, __m256 *prev_state, __m256 *state, __m256 *state_atv, __m256 *output, __m256 *checkI, - __m256 *checkF, __m256 *checkO, + __m256 *checkF, __m256 *checkO, T *cell_clip, ActivationType active_node, ActivationType active_gate, ActivationType active_state) { @@ -65,6 +74,13 @@ class lstm { active_gate); *state = _mm256_add_ps(_mm256_mul_ps(*value_in, *value_ig), _mm256_mul_ps(*prev_state, *value_fg)); + + if (*cell_clip > 0.0f) { + __m256 min = _mm256_set1_ps(0.0f - *cell_clip); + __m256 max = _mm256_set1_ps(*cell_clip); + *state = _mm256_min_ps(max, *state); + *state = _mm256_max_ps(min, *state); + } *value_og = activation( _mm256_add_ps(*value_og, _mm256_mul_ps(*state, *checkO)), active_gate); *state_atv = activation(*state, active_state); @@ -86,15 +102,26 @@ class lstm { 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, + T *checkFGrad, T *checkOGrad, T *cell_clip, 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); + if (*cell_clip > 0.0f) { + if (*state >= (*cell_clip) || *state <= (0.0f - (*cell_clip))) { + *state_grad = 0.0f; + } else { + *state_grad += + activation((*output_grad) * (*value_og), *state_atv, active_state) + + (*grad_og) * (*checkO); + } + } else { + *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 = @@ -117,15 +144,24 @@ class lstm { __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) { + __m256 *checkFGrad, __m256 *checkOGrad, T *cell_clip, + 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); + if (*cell_clip > 0.0f) { + T *state_ = reinterpret_cast(state); + if (*state_ >= (*cell_clip) || *state_ <= (0.0f - (*cell_clip))) { + *state_grad = _mm256_set1_ps(0.0f); + } else { + *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, diff --git a/paddle/fluid/operators/math/fc_compute.h b/paddle/fluid/operators/math/fc_compute.h index cddd0a18db53a7ddf9ca14d5f373180586ef6a31..0ad57c51be79cd3577b43c9af777bff710308fac 100644 --- a/paddle/fluid/operators/math/fc_compute.h +++ b/paddle/fluid/operators/math/fc_compute.h @@ -30,15 +30,17 @@ inline void FCCompute(const BlasT& blas, const int M, return; } if (relu) { - auto compute = - jit::Get, platform::CPUPlace>(N); + auto compute = jit::KernelFuncs, + platform::CPUPlace>::Cache() + .At(N); for (int i = 0; i < M; i++) { T* dst = Y + i * N; compute(B, dst, dst, N); } } else { - auto compute = - jit::Get, platform::CPUPlace>(N); + auto compute = jit::KernelFuncs, + platform::CPUPlace>::Cache() + .At(N); #ifdef PADDLE_WITH_MKLML #pragma omp parallel for #endif diff --git a/paddle/fluid/operators/math/lstm_compute.cc b/paddle/fluid/operators/math/lstm_compute.cc index b6882b4fd8e6db8592a282410888d5625bae742a..94bbcbb50670d9f0b11b77cf6a54a99c227521bf 100644 --- a/paddle/fluid/operators/math/lstm_compute.cc +++ b/paddle/fluid/operators/math/lstm_compute.cc @@ -24,12 +24,12 @@ template struct LstmUnitFunctor { static void compute(const platform::CPUDeviceContext& context, LstmMetaValue value, int frame_size, int batch_size, - const detail::ActivationType& gate_act, + T cell_clip, 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, - cand_act, gate_act, cell_act); + cell_clip, cand_act, gate_act, cell_act); value.gate_value += frame_size * 4; value.state_value += frame_size; value.state_active_value += frame_size; @@ -45,13 +45,14 @@ template struct LstmUnitGradFunctor { static void compute(const platform::CPUDeviceContext& context, LstmMetaValue value, LstmMetaGrad grad, - int frame_size, int batch_size, + int frame_size, int batch_size, T cell_clip, 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, cand_act, gate_act, cell_act); + frame_size, cell_clip, cand_act, gate_act, + cell_act); value.gate_value += frame_size * 4; value.state_value += frame_size; diff --git a/paddle/fluid/operators/math/lstm_compute.cu b/paddle/fluid/operators/math/lstm_compute.cu index 1233000083d6efc31fcbc527e8e9efb83224b4e3..e7445d3d40ae92ff66e7d33a38bfdebfc8455f0a 100644 --- a/paddle/fluid/operators/math/lstm_compute.cu +++ b/paddle/fluid/operators/math/lstm_compute.cu @@ -24,12 +24,12 @@ template struct LstmUnitFunctor { static void compute(const platform::CUDADeviceContext& context, LstmMetaValue value, int frame_size, int batch_size, - const detail::ActivationType& gate_act, + T cell_clip, 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, cand_act, gate_act, - cell_act); + frame_size, batch_size, cell_clip, cand_act, + gate_act, cell_act); } }; @@ -37,13 +37,13 @@ template struct LstmUnitGradFunctor { static void compute(const platform::CUDADeviceContext& context, LstmMetaValue value, LstmMetaGrad grad, - int frame_size, int batch_size, + int frame_size, int batch_size, T cell_clip, 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, cand_act, gate_act, - cell_act); + frame_size, batch_size, cell_clip, cand_act, + gate_act, cell_act); } }; diff --git a/paddle/fluid/operators/math/lstm_compute.h b/paddle/fluid/operators/math/lstm_compute.h index ca2f78e6f318ce39bd2272bbce20f6a6f98fe430..80af5639387aaf6a983365e13c3478353c27a617 100644 --- a/paddle/fluid/operators/math/lstm_compute.h +++ b/paddle/fluid/operators/math/lstm_compute.h @@ -50,7 +50,7 @@ template class LstmUnitFunctor { public: static void compute(const DeviceContext &context, LstmMetaValue value, - int frame_size, int batch_size, + int frame_size, int batch_size, T cell_clip, const detail::ActivationType &gate_act, const detail::ActivationType &cell_act, const detail::ActivationType &cand_act); @@ -61,7 +61,7 @@ class LstmUnitGradFunctor { public: static void compute(const DeviceContext &context, LstmMetaValue value, LstmMetaGrad grad, int frame_size, int batch_size, - const detail::ActivationType &gate_act, + T cell_clip, const detail::ActivationType &gate_act, const detail::ActivationType &cell_act, const detail::ActivationType &cand_act); }; diff --git a/paddle/fluid/operators/math/sample_prob.cc b/paddle/fluid/operators/math/sample_prob.cc new file mode 100644 index 0000000000000000000000000000000000000000..99aa318453eae161807353198a78e11085cd6237 --- /dev/null +++ b/paddle/fluid/operators/math/sample_prob.cc @@ -0,0 +1,26 @@ +/* Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/math/sample_prob.h" + +namespace paddle { +namespace operators { +namespace math { + +template class SampleWithProb; +template class SampleWithProb; + +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/math/sample_prob.cu b/paddle/fluid/operators/math/sample_prob.cu new file mode 100644 index 0000000000000000000000000000000000000000..8f9391591560cc3f76ac67f43121c4b1cff90e12 --- /dev/null +++ b/paddle/fluid/operators/math/sample_prob.cu @@ -0,0 +1,161 @@ +/* Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include +#include +#include +#include + +#include "paddle/fluid/framework/ddim.h" +#include "paddle/fluid/framework/eigen.h" +#include "paddle/fluid/framework/operator.h" +#include "paddle/fluid/framework/tensor.h" +#include "paddle/fluid/operators/math/math_function.h" +#include "paddle/fluid/operators/math/sample_prob.h" +#include "paddle/fluid/operators/math/sampler.h" + +namespace paddle { +namespace operators { +namespace math { + +using Tensor = framework::Tensor; + +template +__device__ T gpu_adjust_prob(const T prob, const int num_samples, + const int num_tries) { + if (num_samples == num_tries) { + return prob * num_samples; + } else { + return -expm1(num_tries * log1p(-prob)); + } +} + +class GPULogUniformSampler { + public: + __device__ int64_t Sample(float random, const int range, + const float log_range) const; + __device__ float Probability(int64_t value, const float log_range) const; +}; + +__device__ int64_t GPULogUniformSampler::Sample(float random, const int range, + const float log_range) const { + // Got Log Uniform distribution from uniform distribution by + // inverse_transform_sampling method + const int64_t value = static_cast(exp(random * log_range)) - 1; + // Mathematically, value should be <= range_, but might not be due to some + // floating point roundoff, so we mod by range_. + return value % range; +} + +__device__ float GPULogUniformSampler::Probability( + int64_t value, const float log_range) const { + // Given f(x) = 1/[(x+1) * log_range_] + // The value's probability is integral of f(x) from value to (value + 1) + return (log((value + 2.0) / (value + 1.0))) / log_range; +} + +template +__global__ void SamplingCondidate( + const size_t n, const int num_tries, const int range, const float log_range, + const int num_true, const std::size_t num_samples, + const int64_t* label_data, int64_t* samples_data, T* probabilities_data) { + const int num_sampled_classes = num_true + num_samples; + + int idx = blockDim.x * blockIdx.x + threadIdx.x; + int step_size = 0; + GPULogUniformSampler sampler; + + for (; idx < n; idx += blockDim.x * gridDim.x) { + int col_idx = idx % num_sampled_classes; + int row_idx = idx / num_sampled_classes; + if (col_idx < num_true) { + samples_data[idx] = label_data[row_idx * num_true + col_idx]; + } else { + samples_data[idx] = samples_data[col_idx]; + } + probabilities_data[idx] = sampler.Probability(samples_data[idx], log_range); + probabilities_data[idx] = + gpu_adjust_prob(probabilities_data[idx], num_samples, num_tries); + } +} + +template +int UniqSampler(const Sampler& sampler, const std::size_t num_samples, + int64_t* samples_data) { + // sample num_samles unique samples for an example, note that they are not + // all negative samples + std::unordered_set tmp_samples; + tmp_samples.clear(); + int num_tries = 0; + int j = 0; + while (j < num_samples) { + ++num_tries; + auto v = sampler.Sample(); + auto insert_ok = tmp_samples.insert(v).second; + if (!insert_ok) { + continue; + } + samples_data[j] = v; + ++j; + } + return num_tries; +} + +template +void GPUSampleWithProb::operator()( + const platform::CUDADeviceContext& context, const int seed, + const int dict_size, const bool uniq, const std::size_t num_samples, + const Tensor* L, Tensor* S, Tensor* P) { + // UNDERSTAND: dimension issues + const auto lbl_dim = L->dims(); + const int batch_size = lbl_dim[0]; + const int num_true = lbl_dim[1]; + const int num_sampled_classes = num_true + num_samples; + framework::DDim ret_dim{batch_size, num_sampled_classes}; + + // UNDERSTAND: raw data view + const int64_t* label_data = L->data(); + int64_t* samples_data = S->data(); + T* probabilities_data = P->data(); + + int s_size = num_samples; + framework::DDim s_dim{s_size}; + Tensor s; + int64_t* s_data = s.mutable_data(s_dim, platform::CPUPlace()); + + math::LogUniformSampler sampler(dict_size, seed); + + int range = dict_size; + float log_range = log(range + 1); + + int num_tries = UniqSampler(sampler, num_samples, s_data); + VLOG(1) << "num_tries: " << num_tries; + PADDLE_ENFORCE(cudaMemcpy(samples_data + num_true, s_data, + sizeof(int64_t) * num_samples, + cudaMemcpyHostToDevice)); + + int threads = 512; + const size_t size = batch_size * num_sampled_classes; + int grid = (batch_size * num_sampled_classes + threads - 1) / threads; + SamplingCondidate<<>>( + size, num_tries, range, log_range, num_true, num_samples, label_data, + samples_data, probabilities_data); +} + +template class GPUSampleWithProb; +template class GPUSampleWithProb; +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/math/sample_prob.h b/paddle/fluid/operators/math/sample_prob.h new file mode 100644 index 0000000000000000000000000000000000000000..e5a6d84cb2b0527c606e62a19ef02d669945ecb1 --- /dev/null +++ b/paddle/fluid/operators/math/sample_prob.h @@ -0,0 +1,118 @@ +/* Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include +#include +#include +#include "paddle/fluid/framework/ddim.h" +#include "paddle/fluid/framework/eigen.h" +#include "paddle/fluid/framework/tensor.h" +#include "paddle/fluid/operators/math/sampler.h" + +namespace paddle { +namespace operators { +namespace math { + +using Tensor = framework::Tensor; + +/* UNDERSTAND: utility function to adjust probability for unique sampling, +return whatever as it is if not using unique samping */ +template +static T adjust_prob(const T prob, const int num_samples, const int num_tries) { + if (num_samples == num_tries) { + return prob * num_samples; + } else { + return -expm1(num_tries * log1p(-prob)); + } +} + +template +class SampleWithProb { + public: + void operator()(const DeviceContext& context, const Sampler& sampler, + const std::size_t num_samples, const Tensor* L, Tensor* S, + Tensor* P) { + // UNDERSTAND: dimension issues + const auto lbl_dim = L->dims(); + const int batch_size = lbl_dim[0]; + const int num_true = lbl_dim[1]; + const int num_sampled_classes = num_true + num_samples; + framework::DDim ret_dim{batch_size, num_sampled_classes}; + + // UNDERSTAND: raw data view + const int64_t* label_data = L->data(); + int64_t* samples_data = + S->mutable_data(ret_dim, context.GetPlace()); + T* probabilities_data = P->mutable_data(ret_dim, context.GetPlace()); + + // temp sets for unique sampling + std::unordered_set tmp_samples; + int j = 0; // column index + // add true labels, not that efficient + while (j < num_true) { + for (int i = 0; i < batch_size; ++i) { + auto samples_index = i * num_sampled_classes + j; + auto v = label_data[i * num_true + j]; + samples_data[samples_index] = v; + probabilities_data[samples_index] = sampler.Probability(v); + } + ++j; + } + + // sample num_samles unique samples for an example, note that they are not + // all negative samples + tmp_samples.clear(); + int num_tries = 0; + while (j < num_sampled_classes) { + ++num_tries; + auto v = sampler.Sample(); + auto insert_ok = tmp_samples.insert(v).second; + if (!insert_ok) { + continue; + } + auto p = sampler.Probability(v); + for (int i = 0; i < batch_size; ++i) { + auto samples_index = i * num_sampled_classes + j; + samples_data[samples_index] = v; + probabilities_data[samples_index] = p; + } + ++j; + } + + // compute Q(y|x), because of unique sampling, probabilities need to be + // adjusted + for (int k = 0; k < num_sampled_classes; ++k) { + for (int i = 0; i < batch_size; ++i) { + auto samples_index = i * num_sampled_classes + k; + probabilities_data[samples_index] = adjust_prob( + probabilities_data[samples_index], num_samples, num_tries); + } + } + } +}; + +#ifdef PADDLE_WITH_CUDA +template +class GPUSampleWithProb { + public: + void operator()(const platform::CUDADeviceContext& context, const int seed, + const int dict_size, const bool uniq, + const std::size_t num_samples, const Tensor* L, Tensor* S, + Tensor* P); +}; +#endif +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/math/softmax_impl.h b/paddle/fluid/operators/math/softmax_impl.h index 1d9d98b10646af9e199f6c481740d30745888707..a1cb3f972826a67721b00ce6df0ec48cc34d6e03 100644 --- a/paddle/fluid/operators/math/softmax_impl.h +++ b/paddle/fluid/operators/math/softmax_impl.h @@ -16,8 +16,8 @@ limitations under the License. */ #include #include "paddle/fluid/framework/eigen.h" #include "paddle/fluid/framework/tensor.h" +#include "paddle/fluid/operators/jit/kernels.h" -#include "paddle/fluid/operators/math/blas.h" namespace paddle { namespace operators { namespace math { @@ -81,28 +81,11 @@ class SoftmaxFunctor> { const int kBatchDim = 0; const int kClassDim = 1; // 2D data. Batch x C - const int batch_size = in_dims[kBatchDim]; - const int num_classes = in_dims[kClassDim]; - std::vector entities(batch_size); - auto blas = math::GetBlas(context); - for (int n = 0; n < batch_size; ++n) { - entities[n] = in_data[n * num_classes]; - for (int c = 1; c < num_classes; ++c) { - entities[n] = in_data[n * num_classes + c] > entities[n] - ? in_data[n * num_classes + c] - : entities[n]; - } - for (int c = 0; c < num_classes; ++c) { - out_data[n * num_classes + c] = - in_data[n * num_classes + c] - entities[n]; - } - } - - blas.VEXP(num_classes * batch_size, out_data, out_data); - for (int n = 0; n < batch_size; ++n) { - auto sum = blas.ASUM(num_classes, &out_data[n * num_classes], 1); - blas.SCAL(num_classes, 1.0f / sum, &out_data[n * num_classes]); - } + auto compute_softmax = + jit::KernelFuncs, + platform::CPUPlace>::Cache() + .At(in_dims[kClassDim]); + compute_softmax(in_data, out_data, in_dims[kClassDim], in_dims[kBatchDim]); } }; diff --git a/paddle/fluid/operators/activation_mkldnn_op.cc b/paddle/fluid/operators/mkldnn/activation_mkldnn_op.cc similarity index 97% rename from paddle/fluid/operators/activation_mkldnn_op.cc rename to paddle/fluid/operators/mkldnn/activation_mkldnn_op.cc index e16b6f78d16ce29cc493c4c795c7fe97a4bf2550..5b7505f3c4acdef94fead04efd00b47825274117 100644 --- a/paddle/fluid/operators/activation_mkldnn_op.cc +++ b/paddle/fluid/operators/mkldnn/activation_mkldnn_op.cc @@ -52,11 +52,6 @@ class MKLDNNActivationKernel "Wrong layout/format set for Input x tensor"); Functor functor; - - auto attrs = functor.GetAttrs(); - for (auto &attr : attrs) { - *attr.second = ctx.Attr(attr.first); - } functor(ctx); } }; @@ -76,11 +71,6 @@ class MKLDNNActivationGradKernel "is_test attribute should be set to False in training phase."); Functor functor; - - auto attrs = functor.GetAttrs(); - for (auto &attr : attrs) { - *attr.second = ctx.Attr(attr.first); - } functor(ctx); } }; @@ -235,7 +225,7 @@ void eltwise_grad(const framework::ExecutionContext &ctx, std::static_pointer_cast(dev_ctx.GetBlob(key_src_mem)); PADDLE_ENFORCE(src_memory != nullptr, "Fail to find src_memory in device context"); - src_memory->set_data_handle(*p_src_data.get()); + src_memory->set_data_handle(*p_src_data); std::shared_ptr diff_src_memory; diff --git a/paddle/fluid/operators/batch_norm_mkldnn_op.cc b/paddle/fluid/operators/mkldnn/batch_norm_mkldnn_op.cc similarity index 100% rename from paddle/fluid/operators/batch_norm_mkldnn_op.cc rename to paddle/fluid/operators/mkldnn/batch_norm_mkldnn_op.cc diff --git a/paddle/fluid/operators/concat_mkldnn_op.cc b/paddle/fluid/operators/mkldnn/concat_mkldnn_op.cc similarity index 100% rename from paddle/fluid/operators/concat_mkldnn_op.cc rename to paddle/fluid/operators/mkldnn/concat_mkldnn_op.cc diff --git a/paddle/fluid/operators/conv_mkldnn_op.cc b/paddle/fluid/operators/mkldnn/conv_mkldnn_op.cc similarity index 100% rename from paddle/fluid/operators/conv_mkldnn_op.cc rename to paddle/fluid/operators/mkldnn/conv_mkldnn_op.cc diff --git a/paddle/fluid/operators/conv_transpose_mkldnn_op.cc b/paddle/fluid/operators/mkldnn/conv_transpose_mkldnn_op.cc similarity index 100% rename from paddle/fluid/operators/conv_transpose_mkldnn_op.cc rename to paddle/fluid/operators/mkldnn/conv_transpose_mkldnn_op.cc diff --git a/paddle/fluid/operators/dequantize_mkldnn_op.cc b/paddle/fluid/operators/mkldnn/dequantize_mkldnn_op.cc similarity index 100% rename from paddle/fluid/operators/dequantize_mkldnn_op.cc rename to paddle/fluid/operators/mkldnn/dequantize_mkldnn_op.cc diff --git a/paddle/fluid/operators/fc_mkldnn_op.cc b/paddle/fluid/operators/mkldnn/fc_mkldnn_op.cc similarity index 99% rename from paddle/fluid/operators/fc_mkldnn_op.cc rename to paddle/fluid/operators/mkldnn/fc_mkldnn_op.cc index e595f1a627cfefbb91b070b898046cf135dc4988..3a926a716f54a094eba11d63c3b29de27dff274b 100644 --- a/paddle/fluid/operators/fc_mkldnn_op.cc +++ b/paddle/fluid/operators/mkldnn/fc_mkldnn_op.cc @@ -282,7 +282,7 @@ class FCMKLDNNGradOpKernel : public paddle::framework::OpKernel { ? mkldnn::inner_product_backward_weights::desc( src, diff_weights, bias, diff_dst) : mkldnn::inner_product_backward_weights::desc( - src, diff_weights, bias, diff_dst); + src, diff_weights, diff_dst); return mkldnn::inner_product_backward_weights::primitive_desc( bwd_weight_desc, engine, pd); diff --git a/paddle/fluid/operators/gaussian_random_mkldnn_op.cc b/paddle/fluid/operators/mkldnn/gaussian_random_mkldnn_op.cc similarity index 100% rename from paddle/fluid/operators/gaussian_random_mkldnn_op.cc rename to paddle/fluid/operators/mkldnn/gaussian_random_mkldnn_op.cc diff --git a/paddle/fluid/operators/lrn_mkldnn_op.cc b/paddle/fluid/operators/mkldnn/lrn_mkldnn_op.cc similarity index 88% rename from paddle/fluid/operators/lrn_mkldnn_op.cc rename to paddle/fluid/operators/mkldnn/lrn_mkldnn_op.cc index 4e4f977fcc742856b877ef0b7f9a3cc9879aefce..097ba01d401dbc7969e30f576cac2567c874ed99 100644 --- a/paddle/fluid/operators/lrn_mkldnn_op.cc +++ b/paddle/fluid/operators/mkldnn/lrn_mkldnn_op.cc @@ -67,7 +67,13 @@ class LRNMKLDNNOpKernel : public paddle::framework::OpKernel { mid->mutable_data(ctx.GetPlace()); const int n = ctx.Attr("n"); - const float alpha = ctx.Attr("alpha"); + // MKL-DNN implements LRN in a caffe way: + // http://caffe.berkeleyvision.org/tutorial/layers/lrn.html + // Where sum of squares is divided by size of normalization window + // this is not the case for PaddlePaddle LRN. + // Hence we need to compensate for this diffrence by + // multipliing alpha by size of window(n) + const float alpha = ctx.Attr("alpha") * static_cast(n); const float beta = ctx.Attr("beta"); const float k = ctx.Attr("k"); const bool is_test = ctx.Attr("is_test"); @@ -78,10 +84,7 @@ class LRNMKLDNNOpKernel : public paddle::framework::OpKernel { auto dims = paddle::framework::vectorize2int(x->dims()); auto src_md = paddle::platform::MKLDNNMemDesc( - dims, mkldnn::memory::data_type::f32, mkldnn::memory::format::nchw); - - auto dst_md = paddle::platform::MKLDNNMemDesc( - dims, mkldnn::memory::data_type::f32, mkldnn::memory::format::nchw); + dims, mkldnn::memory::data_type::f32, x->format()); auto forward_desc = mkldnn::lrn_forward::desc{mkldnn::prop_kind::forward, mkldnn::lrn_across_channels, @@ -92,8 +95,6 @@ class LRNMKLDNNOpKernel : public paddle::framework::OpKernel { k}; auto src_memory_pd = mkldnn::memory::primitive_desc{src_md, mkldnn_engine}; - auto dst_memory = mkldnn::memory{{dst_md, mkldnn_engine}, - static_cast(output_data)}; if (!is_test) { const std::string key = ctx.op().Output("Out"); @@ -110,11 +111,16 @@ class LRNMKLDNNOpKernel : public paddle::framework::OpKernel { src_memory->set_data_handle( static_cast(const_cast(input_data))); + auto dst_memory = mkldnn::memory(forward_pd->dst_primitive_desc(), + static_cast(output_data)); auto workspace_memory = insert_to_context( key_workspace_memory, dev_ctx, forward_pd->workspace_primitive_desc()); run_primitive(*forward_pd, *src_memory, *workspace_memory, dst_memory); + + out->set_layout(framework::DataLayout::kMKLDNN); + out->set_format(platform::GetMKLDNNFormat(dst_memory)); } else { auto forward_pd = mkldnn::lrn_forward::primitive_desc{forward_desc, mkldnn_engine}; @@ -122,8 +128,13 @@ class LRNMKLDNNOpKernel : public paddle::framework::OpKernel { src_memory_pd, static_cast(const_cast(input_data))}; auto workspace_memory = mkldnn::memory{forward_pd.workspace_primitive_desc()}; + auto dst_memory = mkldnn::memory(forward_pd.dst_primitive_desc(), + static_cast(output_data)); run_primitive(forward_pd, src_memory, workspace_memory, dst_memory); + + out->set_layout(framework::DataLayout::kMKLDNN); + out->set_format(platform::GetMKLDNNFormat(dst_memory)); } } }; @@ -151,7 +162,7 @@ class LRNMKLDNNGradOpKernel : public paddle::framework::OpKernel { const std::string key_workspace_memory = key + "@lrn_workspace_memory"; const int n = ctx.Attr("n"); - const float alpha = ctx.Attr("alpha"); + const float alpha = ctx.Attr("alpha") * static_cast(n); const float beta = ctx.Attr("beta"); const float k = ctx.Attr("k"); diff --git a/paddle/fluid/operators/mkldnn_activation_op.h b/paddle/fluid/operators/mkldnn/mkldnn_activation_op.h similarity index 100% rename from paddle/fluid/operators/mkldnn_activation_op.h rename to paddle/fluid/operators/mkldnn/mkldnn_activation_op.h diff --git a/paddle/fluid/operators/pool_mkldnn_op.cc b/paddle/fluid/operators/mkldnn/pool_mkldnn_op.cc similarity index 98% rename from paddle/fluid/operators/pool_mkldnn_op.cc rename to paddle/fluid/operators/mkldnn/pool_mkldnn_op.cc index f4bad7b712b2b078ed68f0a3d0e751d9ae2d6191..38a65b50bd22354bea54819e8e71015202e96e9f 100644 --- a/paddle/fluid/operators/pool_mkldnn_op.cc +++ b/paddle/fluid/operators/mkldnn/pool_mkldnn_op.cc @@ -198,7 +198,7 @@ class PoolMKLDNNOpKernel : public paddle::framework::OpKernel { } // push primitive to stream and wait until it's executed - std::vector pipeline{*(pool_p.get())}; + std::vector pipeline{*pool_p}; stream(stream::kind::eager).submit(pipeline).wait(); output->set_layout(DataLayout::kMKLDNN); @@ -367,8 +367,7 @@ class PoolMKLDNNGradOpKernel : public paddle::framework::OpKernel { dev_ctx.SetBlob(key_pool_diff_dst_mem_p, diff_dst_memory); pool_bwd_p = std::make_shared( - pool_bwd_pd, *(diff_dst_memory.get()), *workspace_memory, - *(diff_src_memory)); + pool_bwd_pd, *diff_dst_memory, *workspace_memory, *diff_src_memory); dev_ctx.SetBlob(key_pool_bwd_p, pool_bwd_p); } else { @@ -404,7 +403,7 @@ class PoolMKLDNNGradOpKernel : public paddle::framework::OpKernel { if (is_diff_dst_reordered) { pipeline.push_back(reorder_diff_dst); } - pipeline.push_back(*(pool_bwd_p.get())); + pipeline.push_back(*pool_bwd_p); mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); in_x_grad->set_layout(DataLayout::kMKLDNN); diff --git a/paddle/fluid/operators/quantize_mkldnn_op.cc b/paddle/fluid/operators/mkldnn/quantize_mkldnn_op.cc similarity index 100% rename from paddle/fluid/operators/quantize_mkldnn_op.cc rename to paddle/fluid/operators/mkldnn/quantize_mkldnn_op.cc diff --git a/paddle/fluid/operators/softmax_mkldnn_op.cc b/paddle/fluid/operators/mkldnn/softmax_mkldnn_op.cc similarity index 98% rename from paddle/fluid/operators/softmax_mkldnn_op.cc rename to paddle/fluid/operators/mkldnn/softmax_mkldnn_op.cc index d2b149535426d097fea4b8fffa9efe82bd6edc64..dc1176f0848b93dd6872f676c3a71dab4f3455fd 100644 --- a/paddle/fluid/operators/softmax_mkldnn_op.cc +++ b/paddle/fluid/operators/mkldnn/softmax_mkldnn_op.cc @@ -66,8 +66,7 @@ class SoftmaxMKLDNNHandler : public platform::MKLDNNHandler { "Fail to find softmax primitive in device context"); if (softmax_p == nullptr) { softmax_p = std::make_shared( - *(softmax_pd_.get()), - *(static_cast(src_memory_p.get())), + *softmax_pd_, *(static_cast(src_memory_p.get())), *(static_cast(dst_memory_p.get()))); dev_ctx_.SetBlob(prim_key, softmax_p); } else { @@ -88,8 +87,8 @@ class SoftmaxMKLDNNHandler : public platform::MKLDNNHandler { "Fail to find softmax backward primitive in device context"); if (softmax_bwd_p == nullptr) { softmax_bwd_p = std::make_shared( - *softmax_bwd_pd_, *(dst_memory_p.get()), *(diff_dst_memory_p.get()), - *(diff_src_memory_p.get())); + *softmax_bwd_pd_, *dst_memory_p, *diff_dst_memory_p, + *diff_src_memory_p); dev_ctx_.SetBlob(prim_key, softmax_bwd_p); } else { is_reusing_ = true; diff --git a/paddle/fluid/operators/sum_mkldnn_op.cc b/paddle/fluid/operators/mkldnn/sum_mkldnn_op.cc similarity index 99% rename from paddle/fluid/operators/sum_mkldnn_op.cc rename to paddle/fluid/operators/mkldnn/sum_mkldnn_op.cc index c39f94637a1abb5bfce9a5428419282f2b870c91..fe4131df2c77ed28cd36f23002d000dac3e8a129 100644 --- a/paddle/fluid/operators/sum_mkldnn_op.cc +++ b/paddle/fluid/operators/mkldnn/sum_mkldnn_op.cc @@ -160,7 +160,7 @@ class SumMKLDNNOpKernel : public paddle::framework::OpKernel { auto get_selected_row = [&](size_t i) -> const SelectedRows& { if (i == 0 && in0) { - return *in0.get(); + return *in0; } else { return in_vars[i]->Get(); } diff --git a/paddle/fluid/operators/transpose_mkldnn_op.cc b/paddle/fluid/operators/mkldnn/transpose_mkldnn_op.cc similarity index 100% rename from paddle/fluid/operators/transpose_mkldnn_op.cc rename to paddle/fluid/operators/mkldnn/transpose_mkldnn_op.cc diff --git a/paddle/fluid/operators/ngraph/CMakeLists.txt b/paddle/fluid/operators/ngraph/CMakeLists.txt index 6b256ef02666c21ec1db3f6922b56bb23363b4a0..7559d29ce233dfcebf8b3118b4c700c35fe15d32 100644 --- a/paddle/fluid/operators/ngraph/CMakeLists.txt +++ b/paddle/fluid/operators/ngraph/CMakeLists.txt @@ -2,4 +2,5 @@ if(WITH_NGRAPH) cc_library(ngraph_bridge SRCS ngraph_bridge.cc DEPS operator framework_proto ngraph) cc_library(ngraph_engine SRCS ngraph_engine.cc DEPS ngraph_bridge framework_proto) op_library(ngraph_engine_op DEPS ngraph_engine op_registry op_info device_context) + add_subdirectory(ops) endif() diff --git a/paddle/fluid/operators/ngraph/ngraph_bridge.cc b/paddle/fluid/operators/ngraph/ngraph_bridge.cc index d6e897ed4666261cdd0bd6565f61abb218d971e5..996376c53f07b5c26eccad382e734f187f75f5a1 100644 --- a/paddle/fluid/operators/ngraph/ngraph_bridge.cc +++ b/paddle/fluid/operators/ngraph/ngraph_bridge.cc @@ -19,36 +19,21 @@ limitations under the License. */ #include "ngraph/ngraph.hpp" #include "paddle/fluid/operators/ngraph/ngraph_bridge.h" #include "paddle/fluid/operators/ngraph/ngraph_ops.h" +#include "paddle/fluid/operators/ngraph/ops/op_bridge.h" #include "paddle/fluid/platform/enforce.h" #include "paddle/fluid/platform/ngraph_helper.h" namespace paddle { namespace operators { -namespace NG_OPS = paddle::operators::ngraphs; -std::map&, - std::shared_ptr>>)>> - NgraphBridge::NG_NODE_MAP = { - {"elementwise_add", NG_OPS::BuildElementwiseAddNode}, - {"elementwise_add_grad", NG_OPS::BuildElementwiseAddGradNode}, - {"fill_constant", NG_OPS::BuildFillConstantNode}, - {"mean", NG_OPS::BuildMeanNode}, - {"mean_grad", NG_OPS::BuildMeanGradNode}, - {"mul", NG_OPS::BuildMulNode}, - {"mul_grad", NG_OPS::BuildMulGradNode}, - {"softmax", NG_OPS::BuildSoftmaxNode}, - {"softmax_grad", NG_OPS::BuildSoftmaxGradNode}, - {"scale", NG_OPS::BuildScaleNode}, - {"relu", NG_OPS::BuildUnaryNode}, - {"tanh", NG_OPS::BuildUnaryNode}, - {"top_k", NG_OPS::BuildTopKNode}}; +bool NgraphBridge::isRegister(const std::string& str) { + return ops::NgraphSingleton::Lookup(str); +} void NgraphBridge::BuildNgNode( const std::shared_ptr& op) { auto& op_type = op->Type(); - NG_NODE_MAP[op_type](op, ngb_node_map_); + ops::NgraphSingleton::BuildNode(ngb_node_map_, op, op_type); } } // namespace operators diff --git a/paddle/fluid/operators/ngraph/ngraph_bridge.h b/paddle/fluid/operators/ngraph/ngraph_bridge.h index c57988f8f6322e76678c572aa21ff5b17b9e3c22..952d5b0b4362aa1c1112782885ab5d30698f5cff 100644 --- a/paddle/fluid/operators/ngraph/ngraph_bridge.h +++ b/paddle/fluid/operators/ngraph/ngraph_bridge.h @@ -28,13 +28,6 @@ namespace operators { class NgraphBridge { public: - static std::map< - std::string, - std::function&, - std::shared_ptr>>)>> - NG_NODE_MAP; - explicit NgraphBridge( std::shared_ptr< std::unordered_map>> @@ -43,6 +36,8 @@ class NgraphBridge { void BuildNgNode(const std::shared_ptr& op); + static bool isRegister(const std::string& str); + private: std::shared_ptr< std::unordered_map>> diff --git a/paddle/fluid/operators/ngraph/ngraph_engine.cc b/paddle/fluid/operators/ngraph/ngraph_engine.cc index bec4b514a218715134d2366dd7efd7cf5b377b68..660a3298cbe4bf5d83851a916bb3ea8d260214a3 100644 --- a/paddle/fluid/operators/ngraph/ngraph_engine.cc +++ b/paddle/fluid/operators/ngraph/ngraph_engine.cc @@ -88,14 +88,12 @@ static std::vector> NgraphOpIntervals( int pivot = left; while (pivot < right) { auto op_type = ops.at(pivot)->Type(); - if (NgraphBridge::NG_NODE_MAP.find(op_type) == - NgraphBridge::NG_NODE_MAP.end()) { + if (NgraphBridge::isRegister(op_type)) { ++pivot; } else { int start = pivot, end = start; while (pivot < right && - (NgraphBridge::NG_NODE_MAP.find(ops.at(pivot)->Type()) != - NgraphBridge::NG_NODE_MAP.end())) { + (!NgraphBridge::isRegister(ops.at(pivot)->Type()))) { ++pivot; ++end; } diff --git a/paddle/fluid/operators/ngraph/ngraph_engine_op.h b/paddle/fluid/operators/ngraph/ngraph_engine_op.h index d2974298b0707575624ad2f6935e83d06b4c83bb..2f194a9b8766316fc645f7e22e21fff048fb7d63 100644 --- a/paddle/fluid/operators/ngraph/ngraph_engine_op.h +++ b/paddle/fluid/operators/ngraph/ngraph_engine_op.h @@ -35,7 +35,7 @@ class NgraphEngineOp : public framework::OperatorWithKernel { framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext& ctx) const override { framework::OpKernelType kt = framework::OpKernelType( - framework::proto::VarType::FP32, ctx.GetPlace()); + framework::proto::VarType::FP32, platform::CPUPlace()); return kt; } }; diff --git a/paddle/fluid/operators/ngraph/ngraph_ops.h b/paddle/fluid/operators/ngraph/ngraph_ops.h deleted file mode 100644 index b6c7c67c13cc54a4ccdfb4e33795cad76d8179c8..0000000000000000000000000000000000000000 --- a/paddle/fluid/operators/ngraph/ngraph_ops.h +++ /dev/null @@ -1,31 +0,0 @@ -/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - -http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -/* - * This file contains the list of the ngraph operators for Paddle. - * - * ATTENTION: It requires some C++11 features, for lower version C++ or C, we - * might release another API. - */ - -#pragma once - -#include "ops/binary_unnary_op.h" -#include "ops/elementwise_add_op.h" -#include "ops/fill_constant_op.h" -#include "ops/mean_op.h" -#include "ops/mul_op.h" -#include "ops/scale_op.h" -#include "ops/softmax_op.h" -#include "ops/top_k_op.h" diff --git a/paddle/fluid/operators/ngraph/ops/CMakeLists.txt b/paddle/fluid/operators/ngraph/ops/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..7dee3308b74a70a2daf35055d3ac80a14de99ac1 --- /dev/null +++ b/paddle/fluid/operators/ngraph/ops/CMakeLists.txt @@ -0,0 +1,8 @@ +file(GLOB LIST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "*.h") +set(pass_file ${PADDLE_BINARY_DIR}/paddle/fluid/operators/ngraph/ngraph_ops.h) +file(APPEND ${pass_file} "\#pragma once\n") +file(WRITE ${pass_file} "// Generated by the /paddle/fluid/operators/ngraph/ops/CMakeLists.txt. DO NOT EDIT!\n\n") + +foreach(OPS_NAME ${LIST_OPS}) + file(APPEND ${pass_file} "\#include \"paddle/fluid/operators/ngraph/ops/${OPS_NAME}\"\n") +endforeach(OPS_NAME) diff --git a/paddle/fluid/operators/ngraph/ops/accuracy_op.h b/paddle/fluid/operators/ngraph/ops/accuracy_op.h new file mode 100644 index 0000000000000000000000000000000000000000..d90ec97298b0f6fb8480e97ca57cb427784261e4 --- /dev/null +++ b/paddle/fluid/operators/ngraph/ops/accuracy_op.h @@ -0,0 +1,68 @@ +/*Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include +#include "ngraph/ngraph.hpp" +#include "paddle/fluid/operators/ngraph/ops/op_bridge.h" +#include "paddle/fluid/platform/ngraph_helper.h" + +namespace paddle { +namespace operators { +namespace ngraphs { + +void BuildAccuracyNode( + const std::shared_ptr& op, + std::shared_ptr< + std::unordered_map>> + ngb_node_map) { + auto indices = platform::GetInputNode(op, "Indices", ngb_node_map); + auto label = platform::GetInputNode(op, "Label", ngb_node_map); + auto inference = platform::GetInputNode(op, "Out", ngb_node_map); + auto inference_shape = inference->get_shape(); + size_t num_samples = inference_shape.at(0); + size_t k = inference_shape.at(1); + + std::shared_ptr label_k = label; + if (k > 1) { + auto label_1d = std::make_shared( + label, ngraph::AxisVector{0, 1}, ngraph::Shape{num_samples}); + label_k = std::make_shared(label_1d, inference_shape, + ngraph::AxisSet{1}); + } + + auto node_equal = std::make_shared(indices, label_k); + auto node_eq_int = + std::make_shared(node_equal, ngraph::element::i64); + auto num_correct_0d = + std::make_shared(node_eq_int, ngraph::AxisSet{0, 1}); + std::shared_ptr num_correct = + platform::NgReshaper(num_correct_0d, ngraph::Shape{1}); + std::shared_ptr n_samples = ngraph::op::Constant::create( + ngraph::element::i64, ngraph::Shape{1}, {num_samples}); + std::shared_ptr accuracy = std::make_shared( + std::make_shared(num_correct, ngraph::element::f32), + std::make_shared(n_samples, ngraph::element::f32)); + + platform::SetOutputNode(op, "Accuracy", accuracy, ngb_node_map); + platform::SetOutputNode(op, "Correct", num_correct, ngb_node_map); + platform::SetOutputNode(op, "Total", n_samples, ngb_node_map); +} +} // namespace ngraphs +} // namespace operators +} // namespace paddle + +REGISTER_NG_OP(accuracy, BuildAccuracyNode); diff --git a/paddle/fluid/operators/ngraph/ops/activation_op.h b/paddle/fluid/operators/ngraph/ops/activation_op.h new file mode 100644 index 0000000000000000000000000000000000000000..d1b0b80d227a5042219a17e35255617726aa8042 --- /dev/null +++ b/paddle/fluid/operators/ngraph/ops/activation_op.h @@ -0,0 +1,56 @@ +/*Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include + +#include "ngraph/ngraph.hpp" +#include "paddle/fluid/operators/ngraph/ops/op_bridge.h" +#include "paddle/fluid/platform/ngraph_helper.h" + +namespace paddle { +namespace operators { +namespace ngraphs { + +void BuildReluGradNode( + const std::shared_ptr& op, + std::shared_ptr< + std::unordered_map>> + ngb_node_map) { + auto out = platform::GetInputNode(op, "Out", ngb_node_map); + auto dout = platform::GetInputNode(op, "Out@GRAD", ngb_node_map); + auto relu_grad = std::make_shared(out, dout); + platform::SetOutputNode(op, "X@GRAD", relu_grad, ngb_node_map); +} + +void BuildTanhGradNode( + const std::shared_ptr& op, + std::shared_ptr< + std::unordered_map>> + ngb_node_map) { + auto out = platform::GetInputNode(op, "Out", ngb_node_map); + auto dout = platform::GetInputNode(op, "Out@GRAD", ngb_node_map); + auto shape = out->get_shape(); + auto node_const = + ngraph::op::Constant::create(ngraph::element::f32, shape, {1}); + auto result = dout * (node_const - out * out); + platform::SetOutputNode(op, "X@GRAD", result, ngb_node_map); +} +} // namespace ngraphs +} // namespace operators +} // namespace paddle + +REGISTER_NG_OP(relu_grad, BuildReluGradNode); +REGISTER_NG_OP(than_grad, BuildTanhGradNode); diff --git a/paddle/fluid/operators/ngraph/ops/batch_norm_op.h b/paddle/fluid/operators/ngraph/ops/batch_norm_op.h new file mode 100644 index 0000000000000000000000000000000000000000..2d638bb53f084ee75014d64302ec3d86b3bcf26f --- /dev/null +++ b/paddle/fluid/operators/ngraph/ops/batch_norm_op.h @@ -0,0 +1,161 @@ +/*Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include + +#include "ngraph/ngraph.hpp" +#include "paddle/fluid/operators/ngraph/ops/elementwise_node.h" +#include "paddle/fluid/operators/ngraph/ops/elementwise_scalar_op.h" +#include "paddle/fluid/operators/ngraph/ops/op_bridge.h" +#include "paddle/fluid/platform/ngraph_helper.h" + +namespace paddle { +namespace operators { +namespace ngraphs { + +void BuildBatchNormNode( + const std::shared_ptr& op, + std::shared_ptr< + std::unordered_map>> + ngb_node_map) { + auto op_attrs = paddle::framework::AttrReader(op->Attrs()); + auto& data_layout = op_attrs.Get("data_layout"); + + auto bias = paddle::platform::GetInputNode(op, "Bias", ngb_node_map); + auto mean = paddle::platform::GetInputNode(op, "Mean", ngb_node_map); + auto variance = paddle::platform::GetInputNode(op, "Variance", ngb_node_map); + auto scale = paddle::platform::GetInputNode(op, "Scale", ngb_node_map); + auto x = paddle::platform::GetInputNode(op, "X", ngb_node_map); + + const bool is_test = op_attrs.Get("is_test"); + const float epsilon = op_attrs.Get("epsilon"); + const float momentum = op_attrs.Get("momentum"); + + PADDLE_ENFORCE( + data_layout == "NHWC" || data_layout == "NCHW" || data_layout == "NC", + "The BatchNorm operator only supports NHWC/NCHW/NC data format"); + + if (data_layout == "NHWC") { + x = paddle::platform::Nhwc2Nchw(x); + } + + std::shared_ptr mean_out, saved_mean, saved_variance, + variance_out, y; + + if (!is_test) { + auto BN = std::make_shared(epsilon, scale, + bias, x); + y = std::make_shared(BN, 0); + saved_mean = std::make_shared(BN, 1); + saved_variance = std::make_shared(BN, 2); + + mean_out = std::make_shared( + paddle::operators::ngraphs::ElementwiseScalar( + momentum, mean), + paddle::operators::ngraphs::ElementwiseScalar( + 1. - momentum, saved_mean)); + variance_out = std::make_shared( + paddle::operators::ngraphs::ElementwiseScalar( + momentum, variance), + paddle::operators::ngraphs::ElementwiseScalar( + 1. - momentum, saved_variance)); + + if (data_layout == "NHWC") { + y = paddle::platform::Nchw2Nhwc(y); + } + + paddle::platform::SetOutputNode(op, "MeanOut", mean_out, ngb_node_map); + paddle::platform::SetOutputNode(op, "VarianceOut", variance_out, + ngb_node_map); + paddle::platform::SetOutputNode(op, "SavedMean", saved_mean, ngb_node_map); + paddle::platform::SetOutputNode(op, "SavedVariance", saved_variance, + ngb_node_map); + paddle::platform::SetOutputNode(op, "Y", y, ngb_node_map); + } else { + y = std::make_shared(epsilon, scale, bias, + x, mean, variance); + paddle::platform::SetOutputNode(op, "Y", y, ngb_node_map); + } +} + +void BuildBatchNormGradNode( + const std::shared_ptr& op, + std::shared_ptr< + std::unordered_map>> + ngb_node_map) { + auto op_attrs = paddle::framework::AttrReader(op->Attrs()); + auto& data_layout = op_attrs.Get("data_layout"); + + auto bias = paddle::platform::GetInputNode(op, "Bias", ngb_node_map); + auto saved_mean = + paddle::platform::GetInputNode(op, "SavedMean", ngb_node_map); + auto saved_variance = + paddle::platform::GetInputNode(op, "SavedVariance", ngb_node_map); + auto scale = paddle::platform::GetInputNode(op, "Scale", ngb_node_map); + auto x = paddle::platform::GetInputNode(op, "X", ngb_node_map); + auto dy = paddle::platform::GetInputNode(op, "Y@GRAD", ngb_node_map); + auto x_shape = x->get_shape(); + auto dy_shape = dy->get_shape(); + + PADDLE_ENFORCE(x_shape.size() == 2 || x_shape.size() == 4, + "BN grap input size needs to be 2 or 4"); + PADDLE_ENFORCE_EQ(x_shape.size(), dy_shape.size(), + "BN grap input and delta size needs to be equal"); + PADDLE_ENFORCE( + data_layout == "NHWC" || data_layout == "NCHW" || data_layout == "NC", + "The BatchNorm operator only supports NHWC/NCHW/NC data format"); + + if (x_shape.size() == 2) { + x = std::make_shared( + x, ngraph::AxisVector{0, 1}, + ngraph::Shape{x_shape.at(0), x_shape.at(1), 1, 1}); + dy = std::make_shared( + dy, ngraph::AxisVector{0, 1}, + ngraph::Shape{dy_shape.at(0), dy_shape.at(1), 1, 1}); + } + + if (data_layout == "NHWC") { + x = paddle::platform::Nhwc2Nchw(dy); + dy = paddle::platform::Nhwc2Nchw(dy); + } + const float epsilon = op_attrs.Get("epsilon"); + + auto bn_bprop = std::make_shared( + epsilon, scale, bias, x, saved_mean, saved_variance, dy); + + std::shared_ptr dx = + std::make_shared(bn_bprop, 0); + auto dscale = std::make_shared(bn_bprop, 1); + auto dbias = std::make_shared(bn_bprop, 2); + paddle::platform::SetOutputNode(op, "Bias@GRAD", dbias, ngb_node_map); + paddle::platform::SetOutputNode(op, "Scale@GRAD", dscale, ngb_node_map); + if (x_shape.size() == 2) { + paddle::platform::SetOutputNode( + op, "X@GRAD", paddle::platform::NgReshaper(dx, x_shape), ngb_node_map); + } else { + if (data_layout == "NHWC") { + dx = paddle::platform::Nchw2Nhwc(dx); + } + paddle::platform::SetOutputNode(op, "X@GRAD", dx, ngb_node_map); + } +} +} // namespace ngraphs +} // namespace operators +} // namespace paddle + +REGISTER_NG_OP(batch_norm, BuildBatchNormNode); +REGISTER_NG_OP(batch_norm_grad, BuildBatchNormGradNode); diff --git a/paddle/fluid/operators/ngraph/ops/binary_unnary_op.h b/paddle/fluid/operators/ngraph/ops/binary_unary_op.h similarity index 88% rename from paddle/fluid/operators/ngraph/ops/binary_unnary_op.h rename to paddle/fluid/operators/ngraph/ops/binary_unary_op.h index 0c0d25d0cd1ae536618057ce80388b8eeb81c68a..375f188286c123b1d652f8780989404760c8e1a4 100644 --- a/paddle/fluid/operators/ngraph/ops/binary_unnary_op.h +++ b/paddle/fluid/operators/ngraph/ops/binary_unary_op.h @@ -16,6 +16,7 @@ limitations under the License. */ #include #include "ngraph/ngraph.hpp" +#include "paddle/fluid/operators/ngraph/ops/op_bridge.h" #include "paddle/fluid/platform/ngraph_helper.h" namespace paddle { @@ -47,3 +48,7 @@ static void BuildUnaryNode( } // namespace ngraphs } // namespace operators } // namespace paddle + +REGISTER_NG_OP(relu, BuildUnaryNode); +REGISTER_NG_OP(tanh, BuildUnaryNode); +REGISTER_NG_OP(sigmoid, BuildUnaryNode); diff --git a/paddle/fluid/operators/ngraph/ops/conv2d_op.h b/paddle/fluid/operators/ngraph/ops/conv2d_op.h new file mode 100644 index 0000000000000000000000000000000000000000..d664825c53ebf17435a0ec532969978abe6d30ca --- /dev/null +++ b/paddle/fluid/operators/ngraph/ops/conv2d_op.h @@ -0,0 +1,239 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include +#include "ngraph/ngraph.hpp" +#include "paddle/fluid/operators/ngraph/ops/op_bridge.h" +#include "paddle/fluid/platform/ngraph_helper.h" + +namespace paddle { +namespace operators { +namespace ngraphs { + +std::shared_ptr GroupedConvolution( + const std::shared_ptr& data_batch, + const std::shared_ptr& filters, const ngraph::Strides strides, + const ngraph::Strides dilations, const ngraph::CoordinateDiff& paddings, + size_t groups) { + auto& data_shape = data_batch->get_shape(); + auto& filter_shape = filters->get_shape(); + ngraph::NodeVector ng_slices; + + for (size_t i = 0; i < groups; ++i) { + size_t channel_step = filter_shape.at(1); + const std::vector lower_bound{0, i * channel_step, 0, 0}; + const std::vector upper_bound{data_shape.at(0), + (i + 1) * channel_step, + data_shape.at(2), data_shape.at(3)}; + auto data_slice = std::make_shared( + data_batch, lower_bound, upper_bound); + + size_t filter_step = filter_shape.at(0) / groups; + const std::vector filter_lower_bound{i * filter_step, 0, 0, 0}; + const std::vector filter_upper_bound{ + (i + 1) * filter_step, filter_shape.at(1), filter_shape.at(2), + filter_shape.at(3)}; + auto filter_slice = std::make_shared( + filters, filter_lower_bound, filter_upper_bound); + auto ng_conv = std::make_shared( + data_slice, filter_slice, strides, dilations, paddings, paddings); + ng_slices.push_back(ng_conv); + } + + size_t concat_axis = 1; + return std::make_shared(ng_slices, concat_axis); +} + +std::shared_ptr GroupedGradConvolutionFilter( + const std::shared_ptr& data_batch, + const std::shared_ptr& filters, + const std::shared_ptr& doutput, const ngraph::Strides strides, + const ngraph::Strides dilations, const ngraph::CoordinateDiff& paddings, + size_t groups) { + auto& data_shape = data_batch->get_shape(); + auto& filter_shape = filters->get_shape(); + auto& out_shape = doutput->get_shape(); + ngraph::NodeVector ng_slices; + + for (size_t i = 0; i < groups; ++i) { + size_t channel_step = filter_shape.at(1); + const std::vector lower_bound{0, i * channel_step, 0, 0}; + const std::vector upper_bound{data_shape.at(0), + (i + 1) * channel_step, + data_shape.at(2), data_shape.at(3)}; + auto data_slice = std::make_shared( + data_batch, lower_bound, upper_bound); + + size_t filter_step = data_shape.at(0); + + const std::vector filter_lower_bound{i * filter_step, 0, 0, 0}; + const std::vector filter_upper_bound{ + (i + 1) * filter_step, filter_shape.at(1), filter_shape.at(2), + filter_shape.at(3)}; + auto filter_slice = std::make_shared( + filters, filter_lower_bound, filter_upper_bound); + + const std::vector olower_bound{0, i * filter_step, 0, 0}; + const std::vector oupper_bound{out_shape.at(0), + (i + 1) * filter_step, + out_shape.at(2), out_shape.at(3)}; + auto out_slice = std::make_shared(doutput, olower_bound, + oupper_bound); + + auto ng_conv = std::make_shared( + data_slice, filter_slice->get_shape(), out_slice, strides, dilations, + paddings, paddings, ngraph::Strides{1, 1}); + + ng_slices.push_back(ng_conv); + } + + size_t concat_axis = 0; + return std::make_shared(ng_slices, concat_axis); +} + +std::shared_ptr GroupedGradConvolutionData( + const std::shared_ptr& data_batch, + const std::shared_ptr& filters, + const std::shared_ptr& doutput, const ngraph::Strides strides, + const ngraph::Strides dilations, const ngraph::CoordinateDiff& paddings, + size_t groups) { + auto& data_shape = data_batch->get_shape(); + auto& filter_shape = filters->get_shape(); + auto& out_shape = doutput->get_shape(); + ngraph::NodeVector ng_slices; + + for (size_t i = 0; i < groups; ++i) { + size_t channel_step = filter_shape.at(1); + const std::vector lower_bound{0, i * channel_step, 0, 0}; + const std::vector upper_bound{data_shape.at(0), + (i + 1) * channel_step, + data_shape.at(2), data_shape.at(3)}; + auto data_slice = std::make_shared( + data_batch, lower_bound, upper_bound); + + size_t filter_step = data_shape.at(0); + + const std::vector filter_lower_bound{i * filter_step, 0, 0, 0}; + const std::vector filter_upper_bound{ + (i + 1) * filter_step, filter_shape.at(1), filter_shape.at(2), + filter_shape.at(3)}; + auto filter_slice = std::make_shared( + filters, filter_lower_bound, filter_upper_bound); + + const std::vector olower_bound{0, i * filter_step, 0, 0}; + const std::vector oupper_bound{out_shape.at(0), + (i + 1) * filter_step, + out_shape.at(2), out_shape.at(3)}; + auto out_slice = std::make_shared(doutput, olower_bound, + oupper_bound); + + auto ng_conv = std::make_shared( + data_slice->get_shape(), filter_slice, out_slice, strides, dilations, + paddings, paddings, ngraph::Strides{1, 1}); + ng_slices.push_back(ng_conv); + } + + size_t concat_axis = 1; + return std::make_shared(ng_slices, concat_axis); +} + +void BuildConv2dNode( + const std::shared_ptr& op, + std::shared_ptr< + std::unordered_map>> + ngb_node_map) { + auto op_attrs = paddle::framework::AttrReader(op->Attrs()); + auto filters = paddle::platform::GetInputNode(op, "Filter", ngb_node_map); + auto input = paddle::platform::GetInputNode(op, "Input", ngb_node_map); + + std::vector strides = op_attrs.Get>("strides"); + std::vector paddings = op_attrs.Get>("paddings"); + std::vector dilations = op_attrs.Get>("dilations"); + + const ngraph::Strides ng_strides{static_cast(strides.at(0)), + static_cast(strides.at(1))}; + const ngraph::Strides ng_dilations{static_cast(dilations.at(0)), + static_cast(dilations.at(1))}; + const ngraph::CoordinateDiff ng_paddings{ + static_cast(paddings.at(0)), + static_cast(paddings.at(1))}; + + int groups = static_cast(op_attrs.Get("groups")); + PADDLE_ENFORCE_GE(groups, 1, "conv groups needs be no less than 1"); + + std::shared_ptr result; + if (groups == 1) { + result = std::make_shared( + input, filters, ng_strides, ng_dilations, ng_paddings, ng_paddings); + } else { + result = GroupedConvolution(input, filters, ng_strides, ng_dilations, + ng_paddings, groups); + } + paddle::platform::SetOutputNode(op, "Output", result, ngb_node_map); +} + +void BuildConv2dGradNode( + const std::shared_ptr& op, + std::shared_ptr< + std::unordered_map>> + ngb_node_map) { + auto op_attrs = paddle::framework::AttrReader(op->Attrs()); + auto filter = paddle::platform::GetInputNode(op, "Filter", ngb_node_map); + auto input = paddle::platform::GetInputNode(op, "Input", ngb_node_map); + auto doutput = + paddle::platform::GetInputNode(op, "Output@GRAD", ngb_node_map); + + int groups = op_attrs.Get("groups"); + std::vector strides = op_attrs.Get>("strides"); + std::vector paddings = op_attrs.Get>("paddings"); + std::vector dilations = op_attrs.Get>("dilations"); + + const ngraph::Strides ng_strides{static_cast(strides.at(0)), + static_cast(strides.at(1))}; + const ngraph::Strides ng_dilations{static_cast(dilations.at(0)), + static_cast(dilations.at(1))}; + const ngraph::CoordinateDiff ng_paddings{ + static_cast(paddings.at(0)), + static_cast(paddings.at(1))}; + + std::shared_ptr dfilter; + std::shared_ptr dinput; + if (groups == 1) { + dfilter = std::make_shared( + input, filter->get_shape(), doutput, ng_strides, ng_dilations, + ng_paddings, ng_paddings, ngraph::Strides{1, 1}); + + dinput = std::make_shared( + input->get_shape(), filter, doutput, ng_strides, ng_dilations, + ng_paddings, ng_paddings, ngraph::Strides{1, 1}); + + } else { + dfilter = GroupedGradConvolutionFilter(input, filter, doutput, ng_strides, + ng_dilations, ng_paddings, groups); + dinput = GroupedGradConvolutionData(input, filter, doutput, ng_strides, + ng_dilations, ng_paddings, groups); + } + + paddle::platform::SetOutputNode(op, "Filter@GRAD", dfilter, ngb_node_map); + paddle::platform::SetOutputNode(op, "Input@GRAD", dinput, ngb_node_map); +} +} // namespace ngraphs +} // namespace operators +} // namespace paddle + +REGISTER_NG_OP(conv2d, BuildConv2dNode); +REGISTER_NG_OP(conv2d_grad, BuildConv2dGradNode); diff --git a/paddle/fluid/operators/ngraph/ops/cross_entropy_op.h b/paddle/fluid/operators/ngraph/ops/cross_entropy_op.h new file mode 100644 index 0000000000000000000000000000000000000000..3ab158f3e13a33bdb7e423919c7592831fa9831a --- /dev/null +++ b/paddle/fluid/operators/ngraph/ops/cross_entropy_op.h @@ -0,0 +1,149 @@ +/*Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include + +#include "ngraph/ngraph.hpp" +#include "paddle/fluid/operators/ngraph/ops/op_bridge.h" +#include "paddle/fluid/platform/ngraph_helper.h" + +namespace paddle { +namespace operators { +namespace ngraphs { + +void BuildCrossEntropyNode( + const std::shared_ptr& op, + std::shared_ptr< + std::unordered_map>> + ngb_node_map) { + auto x = paddle::platform::GetInputNode(op, "X", ngb_node_map); + auto label = paddle::platform::GetInputNode(op, "Label", ngb_node_map); + auto label_shape = label->get_shape(); + auto x_shape = x->get_shape(); + auto label_rank = label_shape.size(); + auto x_rank = x_shape.size(); + std::shared_ptr x_2d = x, label_2d = label; + auto label_2d_shape = label_shape, x_2d_shape = x_shape; + + if (label_rank > 2) { + label_2d_shape = paddle::platform::FlattenTo2d(label_shape, label_rank - 1); + label_2d = paddle::platform::NgReshaper(label, label_2d_shape); + } + if (x_rank > 2) { + x_2d_shape = paddle::platform::FlattenTo2d(x_shape, x_rank - 1); + x_2d = paddle::platform::NgReshaper(x, x_2d_shape); + } + + auto batch_size = x_2d_shape.at(0); + auto op_attrs = paddle::framework::AttrReader(op->Attrs()); + const bool is_soft_label = op_attrs.Get("soft_label"); + + std::shared_ptr node_1_hot = label_2d; + if (!is_soft_label) { + auto label_1d = paddle::platform::NgReshaper( + label_2d, ngraph::Shape{label_2d_shape.at(0)}); + node_1_hot = std::make_shared(label_1d, x_2d_shape, 1); + } + if (x->get_element_type() != node_1_hot->get_element_type()) { + node_1_hot = std::make_shared(node_1_hot, + x->get_element_type()); + } + + auto node_log = std::make_shared(x_2d); + auto high_clip = ngraph::op::Constant::create(node_log->get_element_type(), + node_log->get_shape(), {1e20}); + auto low_clip = ngraph::op::Constant::create(node_log->get_element_type(), + node_log->get_shape(), {-1e20}); + auto node_min = std::make_shared(node_log, high_clip); + auto node_max = std::make_shared(node_min, low_clip); + auto node_mul = node_1_hot * node_log; + auto node_sum = + std::make_shared(node_mul, ngraph::AxisSet{1}); + auto node_neg = std::make_shared(node_sum); + auto xe = + paddle::platform::NgReshaper(node_neg, ngraph::Shape{batch_size, 1}); + + if (!is_soft_label) { + auto ignore_index = op_attrs.Get("ignore_index"); + auto ignore_node = ngraph::op::Constant::create( + label->get_element_type(), label_2d_shape, {ignore_index}); + auto not_equal_node = + std::make_shared(label_2d, ignore_node); + auto mask = std::make_shared(not_equal_node, + xe->get_element_type()); + xe = xe * mask; + } + + paddle::platform::SetOutputNode(op, "Y", xe, ngb_node_map); +} + +void BuildCrossEntropyGradNode( + const std::shared_ptr& op, + std::shared_ptr< + std::unordered_map>> + ngb_node_map) { + auto op_attrs = paddle::framework::AttrReader(op->Attrs()); + const bool is_soft_label = op_attrs.Get("soft_label"); + + auto x = paddle::platform::GetInputNode(op, "X", ngb_node_map); + auto label = paddle::platform::GetInputNode(op, "Label", ngb_node_map); + auto dy = paddle::platform::GetInputNode(op, "Y@GRAD", ngb_node_map); + auto x_shape = x->get_shape(); + auto rank = x_shape.size(); + + std::shared_ptr mask; + if (!is_soft_label) { + auto label_shape = label->get_shape(); + label_shape.pop_back(); + label = paddle::platform::NgReshaper(label, label_shape); + + auto ignore_index = op_attrs.Get("ignore_index"); + auto ignore_node = ngraph::op::Constant::create( + label->get_element_type(), label_shape, {ignore_index}); + auto not_equal_node = + std::make_shared(label, ignore_node); + mask = std::make_shared(not_equal_node, + x->get_element_type()); + mask = std::make_shared(mask, x_shape, + ngraph::AxisSet{rank - 1}); + + label = std::make_shared(label, x_shape, rank - 1); + } + + auto dy_shape = dy->get_shape(); + dy_shape.pop_back(); + auto dy_reshape = paddle::platform::NgReshaper(dy, dy_shape); + auto dy_bcast = std::make_shared( + dy_reshape, x_shape, ngraph::AxisSet{rank - 1}); + if (x->get_element_type() != label->get_element_type()) { + label = std::make_shared(label, x->get_element_type()); + } + + auto xe_grad = -label * dy_bcast / x; + + if (!is_soft_label) { + xe_grad = xe_grad * mask; + } + + paddle::platform::SetOutputNode(op, "X@GRAD", xe_grad, ngb_node_map); +} +} // namespace ngraphs +} // namespace operators +} // namespace paddle + +REGISTER_NG_OP(cross_entropy, BuildCrossEntropyNode); +REGISTER_NG_OP(cross_entropy_grad, BuildCrossEntropyGradNode); diff --git a/paddle/fluid/operators/ngraph/ops/elementwise_add_op.h b/paddle/fluid/operators/ngraph/ops/elementwise_add_op.h index 868df51e16a9714a750bac64dadc3441de79165e..fb796c336a9b45966a0ff703286faa8b61752483 100644 --- a/paddle/fluid/operators/ngraph/ops/elementwise_add_op.h +++ b/paddle/fluid/operators/ngraph/ops/elementwise_add_op.h @@ -19,6 +19,7 @@ limitations under the License. */ #include "ngraph/ngraph.hpp" #include "paddle/fluid/operators/ngraph/ops/elementwise_node.h" +#include "paddle/fluid/operators/ngraph/ops/op_bridge.h" #include "paddle/fluid/platform/ngraph_helper.h" namespace paddle { @@ -85,3 +86,6 @@ void BuildElementwiseAddGradNode( } // namespace ngraphs } // namespace operators } // namespace paddle + +REGISTER_NG_OP(elementwise_add, BuildElementwiseAddNode); +REGISTER_NG_OP(elementwise_add_grad, BuildElementwiseAddGradNode); diff --git a/paddle/fluid/operators/ngraph/ops/fill_constant_op.h b/paddle/fluid/operators/ngraph/ops/fill_constant_op.h index 406a4314f89810df192280cc97de245553d5520f..bc958f2ba27cf929408d56d41bf22976caf7d6ae 100644 --- a/paddle/fluid/operators/ngraph/ops/fill_constant_op.h +++ b/paddle/fluid/operators/ngraph/ops/fill_constant_op.h @@ -17,6 +17,7 @@ limitations under the License. */ #include #include #include "ngraph/ngraph.hpp" +#include "paddle/fluid/operators/ngraph/ops/op_bridge.h" #include "paddle/fluid/platform/ngraph_helper.h" namespace paddle { @@ -46,8 +47,6 @@ void BuildFillConstantNode( ng_dtype = ngraph::element::i64; } else if (data_type == paddle::framework::proto::VarType::INT32) { ng_dtype = ngraph::element::i32; - } else if (data_type == paddle::framework::proto::VarType::BOOL) { - ng_dtype = ngraph::element::boolean; } else { PADDLE_THROW("unsupported data type: %s", data_type); } @@ -57,3 +56,5 @@ void BuildFillConstantNode( } // namespace ngraphs } // namespace operators } // namespace paddle + +REGISTER_NG_OP(fill_constant, BuildFillConstantNode); diff --git a/paddle/fluid/operators/ngraph/ops/mean_op.h b/paddle/fluid/operators/ngraph/ops/mean_op.h index 4c44bc4c112f401c2707f7babd49a33f238a768f..f839d9978d71c2967a7f2c2f22622dc615907831 100644 --- a/paddle/fluid/operators/ngraph/ops/mean_op.h +++ b/paddle/fluid/operators/ngraph/ops/mean_op.h @@ -19,6 +19,7 @@ limitations under the License. */ #include "ngraph/ngraph.hpp" #include "paddle/fluid/operators/ngraph/ops/elementwise_scalar_op.h" +#include "paddle/fluid/operators/ngraph/ops/op_bridge.h" #include "paddle/fluid/platform/ngraph_helper.h" namespace paddle { @@ -64,3 +65,6 @@ void BuildMeanGradNode( } // namespace ngraphs } // namespace operators } // namespace paddle + +REGISTER_NG_OP(mean, BuildMeanNode); +REGISTER_NG_OP(mean_grad, BuildMeanGradNode); diff --git a/paddle/fluid/operators/ngraph/ops/momentum_op.h b/paddle/fluid/operators/ngraph/ops/momentum_op.h new file mode 100644 index 0000000000000000000000000000000000000000..b8291a08a28b585a7ceb67642ba28c3314195790 --- /dev/null +++ b/paddle/fluid/operators/ngraph/ops/momentum_op.h @@ -0,0 +1,104 @@ +/*Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include +#include "ngraph/ngraph.hpp" +#include "paddle/fluid/operators/ngraph/ops/op_bridge.h" +#include "paddle/fluid/platform/ngraph_helper.h" + +namespace paddle { +namespace operators { +namespace ngraphs { + +void BuildMomentumNode( + const std::shared_ptr& op, + std::shared_ptr< + std::unordered_map>> + ngb_node_map) { + auto op_attrs = paddle::framework::AttrReader(op->Attrs()); + auto param = paddle::platform::GetInputNode(op, "Param", ngb_node_map); + auto grad = paddle::platform::GetInputNode(op, "Grad", ngb_node_map); + auto velocity = paddle::platform::GetInputNode(op, "Velocity", ngb_node_map); + auto learning_rate = + paddle::platform::GetInputNode(op, "LearningRate", ngb_node_map); + + auto mu = op_attrs.Get("mu"); + bool use_nesterov = op_attrs.Get("use_nesterov"); + + auto param_shape = param->get_shape(); + auto velocity_shape = velocity->get_shape(); + auto grad_shape = grad->get_shape(); + auto lr_shape = learning_rate->get_shape(); + + auto shape_velocity = ngraph::Shape{velocity_shape}; + auto mu_create = + ngraph::op::Constant::create(ngraph::element::f32, shape_velocity, {mu}); + + auto vel_mul = std::make_shared(velocity, mu_create); + auto vel_out = std::make_shared(vel_mul, grad); + + ngraph::NodeVector result; + if (use_nesterov) { + auto mul_res = std::make_shared(vel_out, mu_create); + auto add_res = std::make_shared(grad, mul_res); + + auto add_2d = paddle::platform::FlattenTo2d(add_res->get_shape(), 0); + auto vel_reshape = paddle::platform::NgReshaper(vel_out, add_2d); + + auto lr_bcast = std::make_shared( + learning_rate, vel_reshape->get_shape(), + ngraph::AxisSet{vel_reshape->get_shape().size() - 1}); + + auto lr_1d = paddle::platform::FlattenTo1d(lr_bcast->get_shape(), 0); + auto lr_reshape = std::make_shared( + lr_bcast, ngraph::AxisVector{0, 1}, lr_1d); + + lr_reshape = std::make_shared( + lr_reshape, ngraph::AxisVector{0}, param->get_shape()); + + auto mul_res1 = std::make_shared(add_res, lr_reshape); + auto res = std::make_shared(param, mul_res1); + paddle::platform::SetOutputNode(op, "ParamOut", res, ngb_node_map); + } else { + auto vel_2d = paddle::platform::FlattenTo2d(vel_out->get_shape(), 0); + auto vel_reshape = paddle::platform::NgReshaper(vel_out, vel_2d); + + auto lr_bcast = std::make_shared( + learning_rate, vel_reshape->get_shape(), + ngraph::AxisSet{vel_reshape->get_shape().size() - 1}); + + auto lr_1d = paddle::platform::FlattenTo1d(lr_bcast->get_shape(), 0); + auto lr_reshape = std::make_shared( + lr_bcast, ngraph::AxisVector{0, 1}, lr_1d); + + lr_reshape = std::make_shared( + lr_reshape, ngraph::AxisVector{0}, param->get_shape()); + + auto mul_result = + std::make_shared(lr_reshape, vel_out); + + auto res = std::make_shared(param, mul_result); + paddle::platform::SetOutputNode(op, "ParamOut", res, ngb_node_map); + } + paddle::platform::SetOutputNode(op, "VelocityOut", vel_out, ngb_node_map); +} + +} // namespace ngraphs +} // namespace operators +} // namespace paddle + +REGISTER_NG_OP(momentum, BuildMomentumNode); diff --git a/paddle/fluid/operators/ngraph/ops/mul_op.h b/paddle/fluid/operators/ngraph/ops/mul_op.h index 4a6cbebe245f891c6c33b2116330a41d89d50e25..98c70a1a99aa899ed8fdd3c4674668cefd14c4ae 100644 --- a/paddle/fluid/operators/ngraph/ops/mul_op.h +++ b/paddle/fluid/operators/ngraph/ops/mul_op.h @@ -16,6 +16,7 @@ limitations under the License. */ #include #include "ngraph/ngraph.hpp" +#include "paddle/fluid/operators/ngraph/ops/op_bridge.h" #include "paddle/fluid/platform/ngraph_helper.h" namespace paddle { @@ -130,3 +131,6 @@ static void BuildMulGradNode( } // namespace ngraphs } // namespace operators } // namespace paddle + +REGISTER_NG_OP(mul, BuildMulNode); +REGISTER_NG_OP(mul_grad, BuildMulGradNode); diff --git a/paddle/fluid/operators/ngraph/ops/op_bridge.h b/paddle/fluid/operators/ngraph/ops/op_bridge.h new file mode 100644 index 0000000000000000000000000000000000000000..93df0ad8062745380d9cd4ca5027bef1425083bf --- /dev/null +++ b/paddle/fluid/operators/ngraph/ops/op_bridge.h @@ -0,0 +1,84 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once +#include +#include +#include +#include + +#include "ngraph/node.hpp" +#include "paddle/fluid/framework/operator.h" +#include "paddle/fluid/operators/ngraph/ngraph_bridge.h" +#include "paddle/fluid/platform/enforce.h" + +namespace paddle { +namespace operators { +namespace ops { + +class NgraphSingleton { + NgraphSingleton() = default; + NgraphSingleton(NgraphSingleton const&) = delete; + void operator=(NgraphSingleton const) = delete; + + ~NgraphSingleton() = default; + + static std::map< + std::string, + std::function&, + std::shared_ptr>>)>> + ng_node_maps_; + + public: + template + static void Register(TF&& tf, const std::string& name) { + ng_node_maps_[name] = tf; + } + + static bool Lookup(const std::string& name) { + auto it = ng_node_maps_.find(name); + if (it == ng_node_maps_.end()) { + return true; + } + return false; + } + + static void BuildNode( + const std::shared_ptr>>& ng_maps, + const std::shared_ptr& op, + const std::string& name) { + ng_node_maps_[name](op, ng_maps); + } +}; + +std::map&, + std::shared_ptr>>)>> + NgraphSingleton::ng_node_maps_; + +} // namespace ops +} // namespace operators +} // namespace paddle + +#define REGISTER_NG_OP(op_type__, Converter__) \ + struct ng_##op_type__##_converter { \ + ng_##op_type__##_converter() { \ + paddle::operators::ops::NgraphSingleton::Register( \ + paddle::operators::ngraphs::Converter__, #op_type__); \ + } \ + }; \ + ng_##op_type__##_converter ng_##op_type__##_converter__; diff --git a/paddle/fluid/operators/ngraph/ops/pool2d_op.h b/paddle/fluid/operators/ngraph/ops/pool2d_op.h new file mode 100644 index 0000000000000000000000000000000000000000..a6371372ef10c093c41153cb0dc73f4f9e95687f --- /dev/null +++ b/paddle/fluid/operators/ngraph/ops/pool2d_op.h @@ -0,0 +1,178 @@ +/*Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include + +#include "ngraph/ngraph.hpp" +#include "paddle/fluid/operators/ngraph/ops/op_bridge.h" +#include "paddle/fluid/platform/ngraph_helper.h" + +namespace paddle { +namespace operators { +namespace ngraphs { + +void BuildPool2dNode( + const std::shared_ptr& op, + std::shared_ptr< + std::unordered_map>> + ngb_node_map) { + auto op_attrs = paddle::framework::AttrReader(op->Attrs()); + auto x = paddle::platform::GetInputNode(op, "X", ngb_node_map); + auto x_shape = x->get_shape(); + + std::string pooling_type = op_attrs.Get("pooling_type"); + std::vector ksize = op_attrs.Get>("ksize"); + std::vector strides = op_attrs.Get>("strides"); + std::vector paddings = op_attrs.Get>("paddings"); + + PADDLE_ENFORCE_EQ(x_shape.size() - 2, ksize.size(), + "Handling 2d pooling only"); + + if (op_attrs.Get("global_pooling")) { + for (size_t i = 0; i < ksize.size(); ++i) { + paddings[i] = 0; + ksize[i] = static_cast(x_shape.at(i + 2)); + } + } + + ngraph::Shape ng_padding_below{static_cast(paddings.at(0)), + static_cast(paddings.at(1))}; + ngraph::Shape ng_padding_above{static_cast(paddings.at(0)), + static_cast(paddings.at(1))}; + ngraph::Shape ng_ksize_shape{static_cast(ksize.at(0)), + static_cast(ksize.at(1))}; + ngraph::Strides ng_strides{static_cast(strides.at(0)), + static_cast(strides.at(1))}; + + auto ComputeCeiledOutput = [](size_t in, size_t k, size_t p, size_t s) { + return (in - k + 2 * p) / s + 1; + }; + + if (op_attrs.Get("ceil_mode")) { + auto dummy_out = paddle::platform::GetOutputNode(op, "Out", ngb_node_map); + auto dummpy_shape = dummy_out->get_shape(); + for (size_t i = 0; i < ng_padding_above.size(); ++i) { + auto desired_size = ComputeCeiledOutput(x_shape[i + 2], ksize[i], + paddings[i], strides[i]); + if (desired_size != dummpy_shape[i + 2]) { + ng_padding_above[i] += strides[i]; + } + } + } + + bool padding_exclusive = op_attrs.Get("exclusive"); + if (pooling_type == "max") { + auto pool2d = std::make_shared( + x, ng_ksize_shape, ng_strides, ng_padding_below, ng_padding_above); + paddle::platform::SetOutputNode(op, "Out", pool2d, ngb_node_map); + } else if (pooling_type == "avg") { + std::shared_ptr pool2d; + if (op_attrs.Get("adaptive")) { + auto ComputeAdaptive = [](size_t in, size_t k) { + return std::floor(in / k); + }; + ng_strides[0] = x_shape.size() == 4 + ? ComputeAdaptive(x_shape[3], ksize[0]) + : ng_strides[0]; + ng_strides[1] = x_shape.size() == 4 + ? ComputeAdaptive(x_shape[3], ksize[0]) + : ng_strides[1]; + pool2d = + std::make_shared(x, ng_ksize_shape, ng_strides); + } else { + pool2d = std::make_shared( + x, ng_ksize_shape, ng_strides, ng_padding_below, ng_padding_above, + !padding_exclusive); + } + paddle::platform::SetOutputNode(op, "Out", pool2d, ngb_node_map); + } else { + PADDLE_THROW("Support max and avg pooling only"); + } +} + +void BuildPool2dGradNode( + const std::shared_ptr& op, + std::shared_ptr< + std::unordered_map>> + ngb_node_map) { + auto op_attrs = paddle::framework::AttrReader(op->Attrs()); + auto out = paddle::platform::GetInputNode(op, "Out", ngb_node_map); + auto dout = paddle::platform::GetInputNode(op, "Out@GRAD", ngb_node_map); + auto x = paddle::platform::GetInputNode(op, "X", ngb_node_map); + auto x_shape = x->get_shape(); + + std::string pooling_type = op_attrs.Get("pooling_type"); + std::vector ksize = op_attrs.Get>("ksize"); + std::vector strides = op_attrs.Get>("strides"); + std::vector paddings = op_attrs.Get>("paddings"); + + PADDLE_ENFORCE_EQ(x_shape.size() - 2, ksize.size(), + "Handling 2d pooling only"); + + if (op_attrs.Get("global_pooling")) { + for (size_t i = 0; i < ksize.size(); ++i) { + paddings[i] = 0; + ksize[i] = static_cast(x_shape.at(i + 2)); + } + } + + ngraph::Shape ng_padding_below{static_cast(paddings.at(0)), + static_cast(paddings.at(1))}; + ngraph::Shape ng_padding_above{static_cast(paddings.at(0)), + static_cast(paddings.at(1))}; + ngraph::Shape ng_ksize_shape{static_cast(ksize.at(0)), + static_cast(ksize.at(1))}; + ngraph::Strides ng_strides{static_cast(strides.at(0)), + static_cast(strides.at(1))}; + + bool padding_exclusive = op_attrs.Get("exclusive"); + if (pooling_type == "max") { + auto pool2d_grad = std::make_shared( + x, dout, out, ng_ksize_shape, ng_strides, ng_padding_below, + ng_padding_above); + paddle::platform::SetOutputNode(op, "X@GRAD", pool2d_grad, ngb_node_map); + } else if (pooling_type == "avg") { + std::shared_ptr pool2d_grad; + if (op_attrs.Get("adaptive")) { + auto ComputeAdaptive = [](size_t in, size_t k) { + return std::floor(in / k); + }; + ng_strides[0] = x_shape.size() == 4 + ? ComputeAdaptive(x_shape[3], ksize[0]) + : ng_strides[0]; + ng_strides[1] = x_shape.size() == 4 + ? ComputeAdaptive(x_shape[3], ksize[0]) + : ng_strides[1]; + pool2d_grad = std::make_shared( + x->get_shape(), dout, ng_ksize_shape, ng_strides, ng_padding_below, + ng_padding_above, !padding_exclusive); + } else { + pool2d_grad = std::make_shared( + x->get_shape(), dout, ng_ksize_shape, ng_strides, ng_padding_below, + ng_padding_above, !padding_exclusive); + } + paddle::platform::SetOutputNode(op, "X@GRAD", pool2d_grad, ngb_node_map); + } else { + PADDLE_THROW("Support max and avg pooling only"); + } +} +} // namespace ngraphs +} // namespace operators +} // namespace paddle + +REGISTER_NG_OP(pool2d, BuildPool2dNode); +REGISTER_NG_OP(pool2d_grad, BuildPool2dGradNode); diff --git a/paddle/fluid/operators/ngraph/ops/scale_op.h b/paddle/fluid/operators/ngraph/ops/scale_op.h index 91a57d0be606373e985a30b7ac9c73648062d8e4..a334192419f572c429f5842cd9e418d8945eb0ef 100644 --- a/paddle/fluid/operators/ngraph/ops/scale_op.h +++ b/paddle/fluid/operators/ngraph/ops/scale_op.h @@ -17,6 +17,7 @@ limitations under the License. */ #include #include "ngraph/ngraph.hpp" #include "paddle/fluid/operators/ngraph/ops/elementwise_scalar_op.h" +#include "paddle/fluid/operators/ngraph/ops/op_bridge.h" #include "paddle/fluid/platform/ngraph_helper.h" namespace paddle { @@ -37,3 +38,5 @@ void BuildScaleNode( } // namespace ngraphs } // namespace operators } // namespace paddle + +REGISTER_NG_OP(scale, BuildScaleNode); diff --git a/paddle/fluid/operators/ngraph/ops/softmax_op.h b/paddle/fluid/operators/ngraph/ops/softmax_op.h index fc6395c08bc6b00990679c5327c3152a980be821..1df6418de06d000892d2802596df61320fcdc759 100644 --- a/paddle/fluid/operators/ngraph/ops/softmax_op.h +++ b/paddle/fluid/operators/ngraph/ops/softmax_op.h @@ -18,6 +18,7 @@ limitations under the License. */ #include #include "ngraph/ngraph.hpp" #include "paddle/fluid/operators/ngraph/ops/elementwise_scalar_op.h" +#include "paddle/fluid/operators/ngraph/ops/op_bridge.h" #include "paddle/fluid/platform/ngraph_helper.h" namespace paddle { @@ -72,3 +73,6 @@ void BuildSoftmaxGradNode( } // namespace ngraphs } // namespace operators } // namespace paddle + +REGISTER_NG_OP(softmax, BuildSoftmaxNode); +REGISTER_NG_OP(softmax_grad, BuildSoftmaxGradNode); diff --git a/paddle/fluid/operators/ngraph/ops/sum_op.h b/paddle/fluid/operators/ngraph/ops/sum_op.h new file mode 100644 index 0000000000000000000000000000000000000000..97f4ce64aa58bfa8cb70c36f9a12b7b8135da637 --- /dev/null +++ b/paddle/fluid/operators/ngraph/ops/sum_op.h @@ -0,0 +1,55 @@ +/*Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include + +#include "ngraph/ngraph.hpp" +#include "paddle/fluid/platform/ngraph_helper.h" + +namespace paddle { +namespace operators { +namespace ngraphs { + +void BuildSumNode( + const std::shared_ptr& op, + std::shared_ptr< + std::unordered_map>> + ngb_node_map) { + std::vector op_inputs; + for (auto& var_name_item : op->Inputs()) { + for (auto& var_name : var_name_item.second) { + op_inputs.push_back(var_name); + if (ngb_node_map->find(var_name) == ngb_node_map->end()) { + PADDLE_THROW("op % input varname %s is not found in var_node_map", + op->Type(), var_name); + } + } + } + std::shared_ptr& sum = ngb_node_map->at(op_inputs[0]); + for (size_t k = 1; k < op_inputs.size(); ++k) { + std::shared_ptr& nodek = ngb_node_map->at(op_inputs[k]); + if (nodek->get_element_type() != sum->get_element_type()) { + nodek = + std::make_shared(nodek, sum->get_element_type()); + } + sum = sum + nodek; + } + platform::SetOutputNode(op, "Out", sum, ngb_node_map); +} +} // namespace ngraphs +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/ngraph/ops/top_k_op.h b/paddle/fluid/operators/ngraph/ops/top_k_op.h index ea66953a125860ab1ce8309819b6c433ff32eaaa..6d10faa7c2efb9cbd87fa8ef1c6ecb4fa350d8f6 100644 --- a/paddle/fluid/operators/ngraph/ops/top_k_op.h +++ b/paddle/fluid/operators/ngraph/ops/top_k_op.h @@ -16,6 +16,7 @@ limitations under the License. */ #include #include "ngraph/ngraph.hpp" +#include "paddle/fluid/operators/ngraph/ops/op_bridge.h" #include "paddle/fluid/platform/ngraph_helper.h" namespace paddle { @@ -36,14 +37,11 @@ void BuildTopKNode( std::make_shared(top_k, 0); std::shared_ptr out = std::make_shared(top_k, 1); - auto dummy_out = paddle::platform::GetOutputNode(op, "Out", ngb_node_map); - if (dummy_out && dummy_out->get_element_type() != out->get_element_type()) { - out = std::make_shared(out, - dummy_out->get_element_type()); - } paddle::platform::SetOutputNode(op, "Indices", indices, ngb_node_map); paddle::platform::SetOutputNode(op, "Out", out, ngb_node_map); } } // namespace ngraphs } // namespace operators } // namespace paddle + +REGISTER_NG_OP(top_k, BuildTopKNode); diff --git a/paddle/fluid/operators/norm_op.h b/paddle/fluid/operators/norm_op.h index 6c95d3f3bf3a3b0448a8f39915f8b025f7d3bd46..f81cbc2c733af2a42f27e2ecb05ee2f8e2f8c17b 100644 --- a/paddle/fluid/operators/norm_op.h +++ b/paddle/fluid/operators/norm_op.h @@ -99,10 +99,10 @@ class NormGradKernel : public framework::OpKernel { auto dx_e = framework::EigenVector::Flatten(*out_dx); Eigen::DSizes shape(pre, n, post); - Eigen::DSizes norm_shape(pre, post); + Eigen::DSizes rshape(pre, 1, post); auto x = x_e.reshape(shape); auto dy = dy_e.reshape(shape); - auto norm = norm_e.reshape(norm_shape); + auto norm = norm_e.reshape(rshape); auto dx = dx_e.reshape(shape); framework::Tensor rsum; @@ -111,7 +111,6 @@ class NormGradKernel : public framework::OpKernel { Eigen::DSizes rdim(1); Eigen::DSizes bcast(1, n, 1); - Eigen::DSizes rshape(pre, 1, post); // dx = ( dy/sqrt(sum(x*x)) ) * [1 - x*sum(x) / (sum(x*x) + e)] // = [dy - dy * x * sum(x) / (sum(x*x) + e)] / sqrt(sum(x*x)) diff --git a/paddle/fluid/operators/pool_op.cc b/paddle/fluid/operators/pool_op.cc index 5399ae556e7f38a551d680704d8d825e2fdba88a..7e1df3b9efec64c3189d2cd80e761994cc061b45 100644 --- a/paddle/fluid/operators/pool_op.cc +++ b/paddle/fluid/operators/pool_op.cc @@ -259,31 +259,40 @@ Example: W_{out} = \\frac{(W_{in} - ksize[1] + 2 * paddings[1] + strides[1] - 1)}{strides[1]} + 1 $$ - For exclusive = true: + For exclusive = false: $$ hstart = i * strides[0] - paddings[0] + $$ + $$ hend = hstart + ksize[0] + $$ + $$ wstart = j * strides[1] - paddings[1] + $$ + $$ wend = wstart + ksize[1] + $$ + $$ Output(i ,j) = \\frac{sum(Input[hstart:hend, wstart:wend])}{ksize[0] * ksize[1]} $$ - For exclusive = false: + + For exclusive = true: $$ hstart = max(0, i * strides[0] - paddings[0]) + $$ + $$ hend = min(H, hstart + ksize[0]) + $$ + $$ wstart = max(0, j * strides[1] - paddings[1]) + $$ + $$ wend = min(W, wstart + ksize[1]) + $$ + $$ Output(i ,j) = \\frac{sum(Input[hstart:hend, wstart:wend])}{(hend - hstart) * (wend - wstart)} $$ - For adaptive = true: - $$ - hstart = floor(i * H_{in} / H_{out}) - hend = ceil((i + 1) * H_{in} / H_{out}) - wstart = floor(j * W_{in} / W_{out}) - wend = ceil((j + 1) * W_{in} / W_{out}) - Output(i ,j) = \\frac{sum(Input[hstart:hend, wstart:wend])}{(hend - hstart) * (wend - wstart)} - $$ )DOC"); } @@ -392,48 +401,68 @@ Example: Output: Out shape: $(N, C, D_{out}, H_{out}, W_{out})$ For ceil_mode = false: - $$ - 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 - $$ + $$ + D_{out} = \\frac{(D_{in} - ksize[0] + 2 * paddings[0])}{strides[0]} + 1 + $$ + $$ + H_{out} = \\frac{(H_{in} - ksize[1] + 2 * paddings[1])}{strides[2]} + 1 + $$ + $$ + W_{out} = \\frac{(W_{in} - ksize[2] + 2 * paddings[2])}{strides[2]} + 1 + $$ For ceil_mode = true: - $$ - D_{out} = \frac{(D_{in} - ksize[0] + 2 * paddings[0] + strides[0] -1)}{strides[0]} + 1 \\ - H_{out} = \frac{(H_{in} - ksize[1] + 2 * paddings[1] + strides[1] -1)}{strides[1]} + 1 \\ - W_{out} = \frac{(W_{in} - ksize[2] + 2 * paddings[2] + strides[2] -1)}{strides[2]} + 1 - $$ - For exclusive = true: - $$ - dstart = i * strides[0] - paddings[0] - dend = dstart + ksize[0] - hstart = j * strides[1] - paddings[1] - hend = hstart + ksize[1] - wstart = k * strides[2] - paddings[2] - wend = wstart + ksize[2] - Output(i ,j, k) = \\frac{sum(Input[dstart:dend, hstart:hend, wstart:wend])}{ksize[0] * ksize[1] * ksize[2]} - $$ + $$ + D_{out} = \\frac{(D_{in} - ksize[0] + 2 * paddings[0] + strides[0] -1)}{strides[0]} + 1 + $$ + $$ + H_{out} = \\frac{(H_{in} - ksize[1] + 2 * paddings[1] + strides[1] -1)}{strides[1]} + 1 + $$ + $$ + W_{out} = \\frac{(W_{in} - ksize[2] + 2 * paddings[2] + strides[2] -1)}{strides[2]} + 1 + $$ + For exclusive = false: - $$ - dstart = max(0, i * strides[0] - paddings[0]) - dend = min(D, dstart + ksize[0]) - hstart = max(0, j * strides[1] - paddings[1]) - hend = min(H, hstart + ksize[1]) - wstart = max(0, k * strides[2] - paddings[2]) - wend = min(W, wstart + ksize[2]) - Output(i ,j, k) = \\frac{sum(Input[dstart:dend, hstart:hend, wstart:wend])}{(dend - dstart) * (hend - hstart) * (wend - wstart)} - $$ - - For adaptive = true: - $$ - dstart = floor(i * D_{in} / D_{out}) - dend = ceil((i + 1) * D_{in} / D_{out}) - hstart = floor(j * H_{in} / H_{out}) - hend = ceil((j + 1) * H_{in} / H_{out}) - wstart = floor(k * W_{in} / W_{out}) - wend = ceil((k + 1) * W_{in} / W_{out}) - Output(i ,j, k) = \\frac{sum(Input[dstart:dend, hstart:hend, wstart:wend])}{(dend - dstart) * (hend - hstart) * (wend - wstart)} - $$ + $$ + dstart = i * strides[0] - paddings[0] + $$ + $$ + dend = dstart + ksize[0] + $$ + $$ + hstart = j * strides[1] - paddings[1] + $$ + $$ + hend = hstart + ksize[1] + $$ + $$ + wstart = k * strides[2] - paddings[2] + $$ + $$ + wend = wstart + ksize[2] + $$ + $$ + Output(i ,j, k) = \\frac{sum(Input[dstart:dend, hstart:hend, wstart:wend])}{ksize[0] * ksize[1] * ksize[2]} + $$ + + For exclusive = true: + $$ + dstart = max(0, i * strides[0] - paddings[0]) + $$ + $$ + dend = min(D, dstart + ksize[0]) + $$ + $$ + hend = min(H, hstart + ksize[1]) + $$ + $$ + wstart = max(0, k * strides[2] - paddings[2]) + $$ + $$ + wend = min(W, wstart + ksize[2]) + $$ + $$ + Output(i ,j, k) = \\frac{sum(Input[dstart:dend, hstart:hend, wstart:wend])}{(dend - dstart) * (hend - hstart) * (wend - wstart)} + $$ )DOC"); } diff --git a/paddle/fluid/operators/random_crop_op.h b/paddle/fluid/operators/random_crop_op.h index d68ba9d661698bb0d33b139f5748daec2ead6595..ee034b270527376fc268b8a868f90db52c51848a 100644 --- a/paddle/fluid/operators/random_crop_op.h +++ b/paddle/fluid/operators/random_crop_op.h @@ -121,7 +121,7 @@ struct RandomCropFunctor { HOSTDEVICE void operator()(size_t ins_idx) { typename Random::Engine engine(seed_); engine.discard(ins_idx * (rank_ - num_batchsize_dims_)); - size_t offsets[9]; + size_t offsets[9] = {}; for (int i = num_batchsize_dims_; i < rank_; ++i) { typename Random::template UniformIntDist dist( 0, x_dims_[i] - out_dims_[i]); diff --git a/paddle/fluid/operators/reader/buffered_reader.cc b/paddle/fluid/operators/reader/buffered_reader.cc index 26ff221dfa0768bd2bcc9e6485a32485f0212ac6..defc29b91f81cb851fec24c5cd9d62dc72c54147 100644 --- a/paddle/fluid/operators/reader/buffered_reader.cc +++ b/paddle/fluid/operators/reader/buffered_reader.cc @@ -14,6 +14,7 @@ #include "paddle/fluid/operators/reader/buffered_reader.h" #include +#include "paddle/fluid/framework/data_type.h" namespace paddle { namespace operators { @@ -24,6 +25,13 @@ BufferedReader::~BufferedReader() { position_.front().wait(); position_.pop(); } +#ifdef PADDLE_WITH_CUDA + if (platform::is_gpu_place(place_)) { + platform::SetDeviceId(boost::get(place_).device); + PADDLE_ENFORCE(cudaStreamDestroy(stream)); + for (auto &event : events) PADDLE_ENFORCE(cudaEventDestroy(event)); + } +#endif } BufferedReader::BufferedReader( @@ -33,6 +41,19 @@ BufferedReader::BufferedReader( thread_pool_(1), place_(place), buffer_size_(buffer_size) { +#ifdef PADDLE_WITH_CUDA + if (platform::is_gpu_place(place_)) { + platform::SetDeviceId(boost::get(place_).device); + compute_stream = + ((platform::CUDADeviceContext *)(platform::DeviceContextPool::Instance() + .Get(place_))) + ->stream(); + events.resize(buffer_size); + for (auto &event : events) + PADDLE_ENFORCE(cudaEventCreateWithFlags(&event, cudaEventDisableTiming)); + PADDLE_ENFORCE(cudaStreamCreateWithFlags(&stream, cudaStreamNonBlocking)); + } +#endif cpu_buffer_.resize(buffer_size); gpu_buffer_.resize(buffer_size); ReadTillBufferFullAsync(); @@ -46,6 +67,12 @@ void BufferedReader::ReadTillBufferFullAsync() { } void BufferedReader::ReadAsync(size_t i) { +#ifdef PADDLE_WITH_CUDA + if (platform::is_gpu_place(place_)) { + platform::SetDeviceId(boost::get(place_).device); + PADDLE_ENFORCE(cudaEventRecord(events[i], compute_stream)); + } +#endif position_.emplace(thread_pool_.enqueue([this, i]() -> size_t { TensorVec &cpu = cpu_buffer_[i]; reader_->ReadNext(&cpu); @@ -54,14 +81,41 @@ void BufferedReader::ReadAsync(size_t i) { return -1UL; } +#ifdef PADDLE_WITH_CUDA + // NOTE(liangdun): using async copy instead of TensorCopySync + // TensorCopySync would block other stream if (platform::is_gpu_place(place_)) { + platform::SetDeviceId(boost::get(place_).device); + PADDLE_ENFORCE(cudaStreamWaitEvent(stream, events[i], 0)); TensorVec &gpu = gpu_buffer_[i]; gpu.resize(cpu.size()); for (size_t i = 0; i < cpu.size(); ++i) { - framework::TensorCopySync(cpu[i], place_, &gpu[i]); + gpu[i].Resize(cpu[i].dims()); + gpu[i].set_layout(cpu[i].layout()); + auto cpu_place = cpu[i].place(); + auto cpu_ptr = cpu[i].data(); + auto gpu_ptr = gpu[i].mutable_data(place_, cpu[i].type()); + auto size = + cpu[i].numel() * paddle::framework::SizeOfType(cpu[i].type()); + if (platform::is_cuda_pinned_place(cpu_place)) + memory::Copy(boost::get(place_), gpu_ptr, + boost::get(cpu_place), + cpu_ptr, size, stream); + else if ((platform::is_gpu_place(cpu_place))) + memory::Copy(boost::get(place_), gpu_ptr, + boost::get(cpu_place), cpu_ptr, + size, stream); + else + // if cpu place is not pinned, async copy is slower than sync copy, + // so we use sync copy instead. + memory::Copy(boost::get(place_), gpu_ptr, + boost::get(cpu_place), cpu_ptr, size, + 0); gpu[i].set_lod(cpu[i].lod()); } + PADDLE_ENFORCE(cudaStreamSynchronize(stream)); } +#endif return i; })); } diff --git a/paddle/fluid/operators/reader/buffered_reader.h b/paddle/fluid/operators/reader/buffered_reader.h index cbe2bc1b5fdd69d1a843b768e3289acd621369a6..87680da01a1f51cfdfe4d100508440eda9d1877f 100644 --- a/paddle/fluid/operators/reader/buffered_reader.h +++ b/paddle/fluid/operators/reader/buffered_reader.h @@ -19,6 +19,9 @@ #include #include "ThreadPool.h" #include "paddle/fluid/framework/reader.h" +#ifdef PADDLE_WITH_CUDA +#include "paddle/fluid/platform/gpu_info.h" +#endif namespace paddle { namespace operators { @@ -59,6 +62,11 @@ class BufferedReader : public framework::DecoratedReader { std::vector cpu_buffer_; std::vector gpu_buffer_; size_t prev_pos_{-1UL}; +#ifdef PADDLE_WITH_CUDA + cudaStream_t stream; + cudaStream_t compute_stream; + std::vector events; +#endif }; } // namespace reader diff --git a/paddle/fluid/operators/reader/ctr_reader.cc b/paddle/fluid/operators/reader/ctr_reader.cc index f08798794a2f9fc042800583cbc032d6f12bf3dc..43a49de52242b96aade91013e89228fcb3247302 100644 --- a/paddle/fluid/operators/reader/ctr_reader.cc +++ b/paddle/fluid/operators/reader/ctr_reader.cc @@ -213,7 +213,7 @@ void ReadSvmData(const DataDesc& data_desc, std::shared_ptr reader, framework::LoD lod{lod_data}; lod_tensor.set_lod(lod); int64_t* tensor_data = lod_tensor.mutable_data( - framework::make_ddim({1, static_cast(batch_feasign.size())}), + framework::make_ddim({static_cast(batch_feasign.size()), 1}), platform::CPUPlace()); memcpy(tensor_data, batch_feasign.data(), batch_feasign.size() * sizeof(int64_t)); @@ -223,7 +223,7 @@ void ReadSvmData(const DataDesc& data_desc, std::shared_ptr reader, // insert label tensor framework::LoDTensor label_tensor; auto* label_tensor_data = label_tensor.mutable_data( - framework::make_ddim({1, static_cast(batch_label.size())}), + framework::make_ddim({static_cast(batch_label.size()), 1}), platform::CPUPlace()); memcpy(label_tensor_data, batch_label.data(), batch_label.size() * sizeof(int64_t)); diff --git a/paddle/fluid/operators/reader/ctr_reader_test.cc b/paddle/fluid/operators/reader/ctr_reader_test.cc index 9f3a254c84d4e04fbcd449644a7e138eff520fbc..6410439816d8ae4a9d1df507819071ce76b5308e 100644 --- a/paddle/fluid/operators/reader/ctr_reader_test.cc +++ b/paddle/fluid/operators/reader/ctr_reader_test.cc @@ -123,7 +123,7 @@ TEST(CTR_READER, read_data) { std::vector>> data_slot_6003{b1, b2, b3, b4}; - std::vector label_dims = {{1, 3}, {1, 3}, {1, 3}, {1, 1}}; + std::vector label_dims = {{3, 1}, {3, 1}, {3, 1}, {1, 1}}; LoDTensorBlockingQueueHolder queue_holder; int capacity = 64; diff --git a/paddle/fluid/operators/reader/read_op.cc b/paddle/fluid/operators/reader/read_op.cc index 8fe638ac2fdc6e0baed7d6cd3c57b72f23164129..846b2ed77e46d82fbeda8faaeed99cddf23c8824 100644 --- a/paddle/fluid/operators/reader/read_op.cc +++ b/paddle/fluid/operators/reader/read_op.cc @@ -85,9 +85,7 @@ class ReadOp : public framework::OperatorBase { std::vector ins; // For profiling - platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); - auto& ctx = *pool.Get(dev_place); - platform::RecordEvent record_event(Type(), &ctx); + platform::RecordEvent record_event(Type()); reader->ReadNext(&ins); if (ins.empty()) { diff --git a/paddle/fluid/operators/reduce_ops/CMakeLists.txt b/paddle/fluid/operators/reduce_ops/CMakeLists.txt index 5fe4d15ae2c6254a50318813c852b6c314880aba..ebd07d90ebe6b0ba008ac89c01c4f054f96a6da9 100644 --- a/paddle/fluid/operators/reduce_ops/CMakeLists.txt +++ b/paddle/fluid/operators/reduce_ops/CMakeLists.txt @@ -1,5 +1,9 @@ include(operators) -register_operators() +if(WITH_GPU) + register_operators(DEPS cub) +else() + register_operators() +endif() if(WITH_GPU) file(GLOB OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "*.part.cu") diff --git a/paddle/fluid/operators/reshape_op.cc b/paddle/fluid/operators/reshape_op.cc index 8eab3a6f891f1dfa91c5ce316f1419df2cd42248..eda54f76b898cdf893347d31cadb86dea892a4ce 100644 --- a/paddle/fluid/operators/reshape_op.cc +++ b/paddle/fluid/operators/reshape_op.cc @@ -327,13 +327,45 @@ class Reshape2GradOp : public framework::OperatorWithKernel { } }; +class ReshapeOpInplaceInToOut : public framework::InplaceInToOut { + public: + using InplaceInToOut::InplaceInToOut; + + protected: + std::unordered_map Apply( + const framework::OpDesc &op_desc, + framework::BlockDesc *block) const override { + std::unordered_map inplace_in_to_out = { + {"X", "Out"}, + }; + return inplace_in_to_out; + } +}; + +class ReshapeGradInplaceInToOut : public framework::InplaceInToOut { + using InplaceInToOut::InplaceInToOut; + + protected: + std::unordered_map Apply( + const framework::OpDesc &op_desc, + framework::BlockDesc *block) const override { + std::unordered_map inplace_in_to_out = { + {framework::GradVarName("Out"), framework::GradVarName("X")}, + }; + return inplace_in_to_out; + } +}; + } // namespace operators } // namespace paddle namespace ops = paddle::operators; +namespace plat = paddle::platform; REGISTER_OPERATOR(reshape, ops::ReshapeOp, ops::ReshapeOpMaker, - paddle::framework::DefaultGradOpDescMaker); -REGISTER_OPERATOR(reshape_grad, ops::ReshapeGradOp); + paddle::framework::DefaultGradOpDescMaker, + ops::ReshapeOpInplaceInToOut); +REGISTER_OPERATOR(reshape_grad, ops::ReshapeGradOp, + ops::ReshapeGradInplaceInToOut); REGISTER_OP_CPU_KERNEL_FUNCTOR(reshape, float, ops::ReshapeKernel, double, ops::ReshapeKernel, int, ops::ReshapeKernel, int64_t, ops::ReshapeKernel); @@ -343,8 +375,9 @@ REGISTER_OP_CPU_KERNEL_FUNCTOR(reshape_grad, float, ops::ReshapeGradKernel, ops::ReshapeGradKernel); REGISTER_OPERATOR(reshape2, ops::Reshape2Op, ops::Reshape2OpMaker, - ops::Reshape2GradMaker); -REGISTER_OPERATOR(reshape2_grad, ops::Reshape2GradOp); + ops::Reshape2GradMaker, ops::ReshapeOpInplaceInToOut); +REGISTER_OPERATOR(reshape2_grad, ops::Reshape2GradOp, + ops::ReshapeGradInplaceInToOut); REGISTER_OP_CPU_KERNEL_FUNCTOR(reshape2, float, ops::ReshapeKernel, double, ops::ReshapeKernel, int, ops::ReshapeKernel, int64_t, ops::ReshapeKernel); @@ -356,16 +389,20 @@ REGISTER_OP_CPU_KERNEL_FUNCTOR(reshape2_grad, float, ops::ReshapeGradKernel, #ifdef PADDLE_WITH_CUDA REGISTER_OP_CUDA_KERNEL_FUNCTOR(reshape, float, ops::ReshapeKernel, double, ops::ReshapeKernel, int, ops::ReshapeKernel, - int64_t, ops::ReshapeKernel); + int64_t, ops::ReshapeKernel, plat::float16, + ops::ReshapeKernel); REGISTER_OP_CUDA_KERNEL_FUNCTOR(reshape_grad, float, ops::ReshapeGradKernel, double, ops::ReshapeGradKernel, int, ops::ReshapeGradKernel, int64_t, + ops::ReshapeGradKernel, plat::float16, ops::ReshapeGradKernel); REGISTER_OP_CUDA_KERNEL_FUNCTOR(reshape2, float, ops::ReshapeKernel, double, ops::ReshapeKernel, int, ops::ReshapeKernel, - int64_t, ops::ReshapeKernel); + int64_t, ops::ReshapeKernel, plat::float16, + ops::ReshapeKernel); REGISTER_OP_CUDA_KERNEL_FUNCTOR(reshape2_grad, float, ops::ReshapeGradKernel, double, ops::ReshapeGradKernel, int, ops::ReshapeGradKernel, int64_t, + ops::ReshapeGradKernel, plat::float16, ops::ReshapeGradKernel); #endif diff --git a/paddle/fluid/operators/row_conv_op.cc b/paddle/fluid/operators/row_conv_op.cc index 10b1b0c899d833d70fa6afe51998fe210899e3c3..d283bddbe9f974ac6835ee91d5a7851453687b80 100644 --- a/paddle/fluid/operators/row_conv_op.cc +++ b/paddle/fluid/operators/row_conv_op.cc @@ -109,23 +109,23 @@ 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$, +Given an input sequence $X$ 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,:} \\cdot W_{i-j, :} +out_{i} = \\sum_{j=i}^{i + context - 1} X_{j} \\cdot W_{j-i} $$ In the above equation: * $Out_{i}$: The i-th row of output variable with shape [1, D]. -* $\\tau$: Future context size. +* $context$: Future context size. * $X_{j}$: The j-th row of input variable with shape [1, D]. -* $W_{i-j}$: The (i-j)-th row of parameters with shape [1, D]. +* $W_{j-i}$: The (j-i)-th row of parameters with shape [1, D]. More details about row_conv please refer to the design document diff --git a/paddle/fluid/operators/sample_logits_op.cc b/paddle/fluid/operators/sample_logits_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..a7f7fb26b17c77e6fe87646d3cac20c02c49b52c --- /dev/null +++ b/paddle/fluid/operators/sample_logits_op.cc @@ -0,0 +1,225 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/sample_logits_op.h" +#include "paddle/fluid/operators/math/sample_prob.h" + +namespace paddle { +namespace operators { + +class SampleLogitsOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("Logits", + "(Tensor, default: Tensor), The unscaled log probabilities " + "which is a 2-D tensor with shape [N x K]. N is the batch_size, " + "and K is the class number."); + AddInput("Labels", + "(Tensor) The ground truth which is a 2-D tensor. Labels is a " + "Tensor with shape [N x NT], where NT is the number of" + "true labels for each example."); + AddInput("CustomizedSamples", + "(Tensor, default: Tensor), A 2-D tensor with shape [N, " + "NT + S]," + " where N is the batch size, NT is the number of true labels " + "and S is the number of negtive sample for each example." + "The first NT elements of each row should be the same with true " + "labels, " + "followed by S custom negtive samples. This tensor" + "is only used when use_customized_samples is true.") + .AsDispensable(); + AddInput( + "CustomizedProbabilities", + "(Tensor, default: Tensor), A 2-D tensor with shape [N, NT + S]." + "The tensor has the same shape with CustomSamples," + "and each element represents probability of element in CustomSamples. " + "This " + "tensor is only used when use_customized_samples is true.") + .AsDispensable(); + AddOutput("Samples", + "(Tensor, default: Tensor), A 2-D tensor with shape [N, " + "NT + S]." + "The outputs value of sampler, including NT true lables and S " + "negetive samples " + "for each example. This will be used in" + "backward calculation.") + .AsIntermediate(); + AddOutput( + "Probabilities", + "(Tensor, default: Tensor), A 2-D tensor with shape [N, NT + S]." + "The probabilites of sampled positive and negtive labels.") + .AsIntermediate(); + AddOutput("SampledLogits", + "(Tensor, default: Tensor), A 2-D tensor with shape" + "[N, NT + S]. The outputs value of sampled logits, which will be" + "used in backward propagation.") + .AsIntermediate(); + AddOutput( + "SampledLabels", + "(Tensor, default: Tensor), A 2-D tensor. The sampled labels" + "with shape [N, NT]. The tonsor contains hard labels as input to " + " softmax op, that is 0, 1, ..., NT-1 because of the first NT elements" + " of Sampels are positive lables."); + AddAttr( + "use_customized_samples", + "An indicator whether to use customized samples with probabilities, if " + "True" + "the operator will use customized samples and customized probabilities" + "otherwise, the operator will generate them by itself.") + .SetDefault(false); + AddAttr( + "uniq", + "An indicator whether to sample non-repetitive negtive labels, if True" + "the operator will sample negtive labels without replacement." + "Otherwise, the operator will sample negtive labels with replacement.") + .SetDefault(true); + AddAttr( + "remove_accidental_hits", + "An indicator whether to remove accidental hits when samples hits true" + "labels, the removal is implemented by subtracting the corresponding" + "logits by float_max to subpress their softmax to be zero.") + .SetDefault(true); + AddAttr("num_samples", "The number of negative samples."); + AddAttr("seed", "Random seed for generating samples").SetDefault(0); + + AddComment(R"DOC( + """ + Computes sampled output training logits and labels suitable for implementing + sampled softmax. + """ + +)DOC"); + } +}; + +class SampleLogitsOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Logits"), + "Input(Logits) should be not null."); + PADDLE_ENFORCE(ctx->HasInput("Labels"), + "Input(Labels) should be not null."); + + PADDLE_ENFORCE(ctx->HasOutput("Samples"), + "Output(Samples) should be not null."); + PADDLE_ENFORCE(ctx->HasOutput("Probabilities"), + "Output(Probabilities) should be not null."); + PADDLE_ENFORCE(ctx->HasOutput("SampledLogits"), + "Output(SampledLogits) should be not null."); + PADDLE_ENFORCE(ctx->HasOutput("SampledLabels"), + "Output(SampledLabels) should be not null."); + + auto logits_dims = ctx->GetInputDim("Logits"); + auto labels_dims = ctx->GetInputDim("Labels"); + + PADDLE_ENFORCE_EQ( + logits_dims.size(), 2UL, + "The logits of softmax_with_cross_entropy should be a 2-D tensor."); + PADDLE_ENFORCE_EQ(labels_dims.size(), 2UL, + "The labels should be a 2-D tensor."); + + const int num_samples = ctx->Attrs().Get("num_samples"); + const int num_sampled_classes = labels_dims[1] + num_samples; + ctx->SetOutputDim("Samples", {logits_dims[0], num_sampled_classes}); + ctx->SetOutputDim("Probabilities", {logits_dims[0], num_sampled_classes}); + ctx->SetOutputDim("SampledLogits", {logits_dims[0], num_sampled_classes}); + ctx->SetOutputDim("SampledLabels", {logits_dims[0], labels_dims[1]}); + } + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + auto data_type = framework::GetDataTypeOfVar(ctx.InputVar("Logits")); + framework::OpKernelType kt = + framework::OpKernelType(data_type, ctx.device_context()); + return kt; + } +}; + +// UNDERSTAND: InferShape for Grad +class SampleLogitsOpGrad : 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 be not null."); + PADDLE_ENFORCE(ctx->HasInput("Samples"), + "Input(Samples) should be not null."); + PADDLE_ENFORCE(ctx->HasInput("SampledLogits"), + "Input(SampledLogits) should be not null."); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("SampledLogits")), + "Input(SampledLogits@Grad) should not be null."); + PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("Logits")), + "Output(Logits@Grad) should be not null."); + + auto logit_dims = ctx->GetInputDim("Logits"); + auto label_dims = ctx->GetInputDim("Labels"); + PADDLE_ENFORCE_EQ(label_dims.size(), 2UL, + "The label should be a 2-D tensor."); + PADDLE_ENFORCE_EQ(logit_dims.size(), 2UL, + "The logits should be a 2-D tensor."); + + ctx->SetOutputDim(framework::GradVarName("Logits"), + ctx->GetInputDim("Logits")); + } + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + auto data_type = framework::GetDataTypeOfVar( + ctx.InputVar(framework::GradVarName("SampledLogits"))); + framework::OpKernelType kt = + framework::OpKernelType(data_type, ctx.device_context()); + return kt; + } +}; + +// UNDERSTAND: what's the rule for making a GradMaker TODO +class SampleLogitsGradMaker : public framework::SingleGradOpDescMaker { + public: + using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; + + protected: + std::unique_ptr Apply() const override { + auto* grad_op = new framework::OpDesc(); + grad_op->SetType("sample_logits_grad"); + grad_op->SetInput("Logits", Input("Logits")); + grad_op->SetInput("Labels", Input("Labels")); + grad_op->SetInput("Samples", Output("Samples")); + grad_op->SetInput("SampledLogits", Output("SampledLogits")); + grad_op->SetInput(framework::GradVarName("SampledLogits"), + OutputGrad("SampledLogits")); + grad_op->SetOutput(framework::GradVarName("Logits"), InputGrad("Logits")); + grad_op->SetAttrMap(Attrs()); + return std::unique_ptr(grad_op); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; + +REGISTER_OPERATOR(sample_logits, ops::SampleLogitsOp, ops::SampleLogitsOpMaker, + ops::SampleLogitsGradMaker); +REGISTER_OPERATOR(sample_logits_grad, ops::SampleLogitsOpGrad); +REGISTER_OP_CPU_KERNEL(sample_logits, ops::SampleLogitsKernel, + ops::SampleLogitsKernel); +REGISTER_OP_CPU_KERNEL(sample_logits_grad, ops::SampleLogitsGradKernel, + ops::SampleLogitsGradKernel); diff --git a/paddle/fluid/operators/sample_logits_op.cu b/paddle/fluid/operators/sample_logits_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..fb49793b730f72d66dc846f233bd95ebdab37c52 --- /dev/null +++ b/paddle/fluid/operators/sample_logits_op.cu @@ -0,0 +1,257 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include +#include "paddle/fluid/framework/eigen.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/framework/tensor_util.h" +#include "paddle/fluid/operators/math/math_function.h" +#include "paddle/fluid/operators/math/sample_prob.h" +#include "paddle/fluid/operators/math/softmax.h" +#include "paddle/fluid/operators/sample_logits_op.h" + +namespace paddle { +namespace operators { + +// UNDERSTAND: something like take_along_axis in numpy. +template +__global__ void GPUTakeAlongD1(size_t size, const int batch_size, + const int array_slice_size, + const int idx_slice_size, const T* p_array, + const int64_t* p_index, T* p_value) { + const auto value_slice_size = idx_slice_size; + int idx = blockDim.x * blockIdx.x + threadIdx.x; + int step_size = blockDim.x * gridDim.x; + + for (; idx < size; idx += step_size) { + int i = idx / idx_slice_size; + auto array_index = p_index[idx]; + p_value[idx] = p_array[i * array_slice_size + array_index]; + } +} + +// UNDERSTAND: something like put_along_axis in numpy but if there is duplicate +// indices, scatter is done in += way. +template +__global__ void GPUPutAlongD1(size_t size, const int batch_size, + const int array_slice_size, + const int idx_slice_size, T* p_array, + const int64_t* p_index, const T* p_value) { + const auto value_slice_size = idx_slice_size; + int idx = blockDim.x * blockIdx.x + threadIdx.x; + int step_size = blockDim.x * gridDim.x; + + // size == batch_size + for (; idx < size; idx += step_size) { + int i = idx; + for (int j = 0; j < idx_slice_size; ++j) { + auto array_index = p_index[i * idx_slice_size + j]; + p_array[i * array_slice_size + array_index] += + p_value[i * idx_slice_size + j]; + } + } +} + +// UNDERSTAND: set label as 0,1,...,num_true-1 +template +__global__ void GPUSetLabel(size_t size, const int num_true, int64_t* p_array) { + int idx = blockDim.x * blockIdx.x + threadIdx.x; + int step_size = blockDim.x * gridDim.x; + + for (; idx < size; idx += step_size) { + p_array[idx] = idx % num_true; + } +} + +// UNDERSTAND: compute accidentdal hits from samples and minus corresponding +// logits by a float max, here 1e20 +template +__global__ void gpu_compute_remove_accidental_hits(const int size, + const int num_true, + const int idx_slice_size, + const int64_t* p_index, + T* p_value) { + const auto value_slice_size = idx_slice_size; + int idx = blockDim.x * blockIdx.x + threadIdx.x; + int step_size = blockDim.x * gridDim.x; + + for (; idx < size; idx += step_size) { + int i = idx / idx_slice_size; + if (idx % idx_slice_size < num_true) continue; + for (int j = 0; j < num_true; ++j) { + const auto true_idx = i * idx_slice_size + j; + if (p_index[true_idx] == p_index[idx]) { + p_value[idx] -= 1e20; + break; + } + } + } +} + +template +class SampleLogitsCUDAKernel : public framework::OpKernel { + public: + using Tensor = framework::Tensor; + void Compute(const framework::ExecutionContext& context) const override { + // get necessary inputs + const Tensor* logits = context.Input("Logits"); + const Tensor* labels = context.Input("Labels"); + VLOG(3) << "Enter SampleLogitsCUDAKernel"; + + // get necessary outputs + Tensor* samples = context.Output("Samples"); + Tensor* probabilities = context.Output("Probabilities"); + Tensor* sampled_logits = context.Output("SampledLogits"); + Tensor* sampled_labels = context.Output("SampledLabels"); + + // shapes + const auto batch_size = logits->dims()[0]; + const auto num_classes = logits->dims()[1]; + const auto labels_dim = labels->dims(); + const auto num_true = labels_dim[1]; + const auto samples_dim = samples->dims(); + + // attrs + const auto num_samples = context.Attr("num_samples"); + const bool use_customized_samples = + context.Attr("use_customized_samples"); + const bool uniq = context.Attr("uniq"); + const bool remove_accidental_hits = + context.Attr("remove_accidental_hits"); + + // device contexts + auto& dev_ctx = context.cuda_device_context(); + + // UNDERSTAND: allocate memories for temporaries + sampled_logits->mutable_data(samples_dim, context.GetPlace()); + math::SetConstant set_zero; + set_zero(dev_ctx, sampled_logits, static_cast(0)); + + auto sampled_labels_data = + sampled_labels->mutable_data(labels_dim, context.GetPlace()); + int threads = 512; + size_t size = batch_size * num_true; + int grid = (size + threads - 1) / threads; + GPUSetLabel< + T><<>>( + size, num_true, sampled_labels_data); + + if (use_customized_samples) { + const Tensor* customized_samples = + context.Input("CustomizedSamples"); + const Tensor* customized_probabilities = + context.Input("CustomizedProbabilities"); + samples->ShareDataWith(*customized_samples); + probabilities->ShareDataWith(*customized_probabilities); + } else { + samples->mutable_data(context.GetPlace()); + probabilities->mutable_data(samples_dim, context.GetPlace()); + // UNDERSTAND: sampling + const auto seed = context.Attr("seed"); + auto sampler_with_prob = math::GPUSampleWithProb(); + sampler_with_prob(context.cuda_device_context(), seed, num_classes, uniq, + num_samples, labels, samples, probabilities); + } + + // UNDERSTAND: gather sampled logits and remove accidental hits if needed + const auto num_take = samples->dims()[1]; + const auto array_dims = logits->dims(); + const auto idx_dims = samples->dims(); + + const T* p_array = logits->data(); + const int64_t* p_index = samples->data(); + T* p_value = sampled_logits->data(); + + // src slice size + const auto array_slice_size = array_dims[1]; + // index slice size + const auto idx_slice_size = idx_dims[1]; + + size = batch_size * num_take; + grid = (size + threads - 1) / threads; + GPUTakeAlongD1< + T><<>>( + size, batch_size, array_slice_size, idx_slice_size, p_array, p_index, + p_value); + + if (remove_accidental_hits) { + const size_t size = batch_size * (num_true + num_samples); + int grid = (size + threads - 1) / threads; + gpu_compute_remove_accidental_hits< + T><<>>( + size, num_true, idx_slice_size, p_index, p_value); + } + + // subtracted sampled logits with logQ(y|x) + auto probs = EigenMatrix::From(*probabilities); + auto smp_logits = EigenMatrix::From(*sampled_logits); + smp_logits.device(*dev_ctx.eigen_device()) = + (smp_logits - probs.log().unaryExpr(TolerableValue())) + .unaryExpr(TolerableValue()); + } +}; + +template +class SampleLogitsGradCUDAKernel : public framework::OpKernel { + public: + using Tensor = framework::Tensor; + void Compute(const framework::ExecutionContext& context) const override { + auto logits_grad = context.Output(framework::GradVarName("Logits")); + const Tensor* samples = context.Input("Samples"); + const Tensor* sampled_logits_grad = + context.Input(framework::GradVarName("SampledLogits")); + logits_grad->mutable_data(context.GetPlace()); + + auto& dev_ctx = context.cuda_device_context(); + math::SetConstant set_zero; + set_zero(dev_ctx, logits_grad, static_cast(0)); + + // UNDERSTAND: scatter it back to logit_grad + const auto batch_size = samples->dims()[0]; + const auto num_put = samples->dims()[1]; + const auto array_dims = logits_grad->dims(); + const auto idx_dims = samples->dims(); + + T* p_array = logits_grad->data(); + const int64_t* p_index = samples->data(); + const T* p_value = sampled_logits_grad->data(); + + // src slice size + const auto array_slice_size = array_dims[1]; + // index slice size + const auto idx_slice_size = idx_dims[1]; + + int threads = 128; + const size_t size = batch_size; + int grid = (size + threads - 1) / threads; + + GPUPutAlongD1< + T><<>>( + size, batch_size, array_slice_size, idx_slice_size, p_array, p_index, + p_value); + } +}; + +} // namespace operators +} // namespace paddle +namespace ops = paddle::operators; + +REGISTER_OP_CUDA_KERNEL(sample_logits, ops::SampleLogitsCUDAKernel, + ops::SampleLogitsCUDAKernel); +REGISTER_OP_CUDA_KERNEL(sample_logits_grad, + ops::SampleLogitsGradCUDAKernel, + ops::SampleLogitsGradCUDAKernel); diff --git a/paddle/fluid/operators/sample_logits_op.h b/paddle/fluid/operators/sample_logits_op.h new file mode 100644 index 0000000000000000000000000000000000000000..b55a24863cc09d5f80e07aedbbb5b3d9ac99e69e --- /dev/null +++ b/paddle/fluid/operators/sample_logits_op.h @@ -0,0 +1,245 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include "paddle/fluid/framework/eigen.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/framework/tensor_util.h" +#include "paddle/fluid/operators/math/math_function.h" +#include "paddle/fluid/operators/math/sample_prob.h" +#include "paddle/fluid/operators/math/softmax.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; +template +using EigenMatrix = framework::EigenMatrix; + +template +struct TolerableValue { + HOSTDEVICE T operator()(const T& x) const { + PADDLE_ASSERT(std::is_floating_point::value); + const T kApproInf = 1e20; + if (x == INFINITY) return kApproInf; + if (x == -INFINITY) return -kApproInf; + return x; + } +}; + +// UNDERSTAND: something like take_along_axis in numpy. +template +static void CPUTakeAlongD1(const platform::DeviceContext& ctx, + const framework::Tensor& array, + const framework::Tensor& index, + framework::Tensor* value) { + PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace())); + // UNDERSTAND: check shape src(B, C), index(B, K), out should also be (B, K) + PADDLE_ENFORCE(index.dims().size() == 2 && array.dims().size() == 2 && + index.dims()[0] == array.dims()[0] && + index.dims() == value->dims()); + + const auto batch_size = index.dims()[0]; + const auto num_take = index.dims()[1]; + const auto array_dims = array.dims(); + const auto idx_dims = index.dims(); + + // UNDERSTAND: no allocations here + const T* p_array = array.data(); + const int64_t* p_index = index.data(); + T* p_value = value->data(); + + // src slice size + const auto array_slice_size = array_dims[1]; + + // index slice size + const auto idx_slice_size = idx_dims[1]; + const auto value_slice_size = idx_slice_size; + + for (int i = 0; i < batch_size; ++i) { + for (int j = 0; j < num_take; ++j) { + auto array_index = p_index[i * idx_slice_size + j]; + p_value[i * value_slice_size + j] = + p_array[i * array_slice_size + array_index]; + } + } +} + +// UNDERSTAND: something like put_along_axis in numpy but if there is duplicate +// indices, scatter is done in += way. +template +static void CPUPutAlongD1(const platform::DeviceContext& ctx, + framework::Tensor* array, + const framework::Tensor& index, + const framework::Tensor& value) { + PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace())); + // UNDERSTAND: check shape src(B, C), index(B, K), out should also be (B, K) + PADDLE_ENFORCE(index.dims().size() == 2 && array->dims().size() == 2 && + index.dims()[0] == array->dims()[0] && + index.dims() == value.dims()); + const auto batch_size = index.dims()[0]; + const auto num_put = index.dims()[1]; + auto array_dims = array->dims(); + auto idx_dims = index.dims(); + + // UNDERSTAND: no allocations here + T* p_array = array->data(); + const int64_t* p_index = index.data(); + const T* p_value = value.data(); + + // slice sizes + const auto array_slice_size = array_dims[1]; + const auto idx_slice_size = idx_dims[1]; + const auto value_slice_size = idx_slice_size; + + for (int i = 0; i < batch_size; ++i) { + for (int j = 0; j < num_put; ++j) { + auto array_index = p_index[i * idx_slice_size + j]; + p_array[i * array_slice_size + array_index] += + p_value[i * value_slice_size + j]; + } + } +} + +// UNDERSTAND: compute accidentdal hits from samples and minus corresponding +// logits by a float max, here 1e20 +template +static void compute_remove_accidental_hits(const platform::DeviceContext& ctx, + framework::Tensor* sampled_logits, + const framework::Tensor& samples, + const int num_true) { + const auto batch_size = sampled_logits->dims()[0]; + const auto num_sampled_classes = sampled_logits->dims()[1]; + T* sampled_logits_data = sampled_logits->data(); + const auto samples_data = samples.data(); + + std::unordered_set tmp_true_labels; + for (int i = 0; i < batch_size; ++i) { + tmp_true_labels.clear(); + tmp_true_labels.insert(samples_data + i * num_sampled_classes, + samples_data + i * num_sampled_classes + num_true); + for (int j = num_true; j < num_sampled_classes; ++j) { + const auto idx = i * num_sampled_classes + j; + if (tmp_true_labels.find(samples_data[idx]) != tmp_true_labels.end()) + sampled_logits_data[idx] -= 1e20; + } + } +} + +template +class SampleLogitsKernel : public framework::OpKernel { + public: + using Tensor = framework::Tensor; + void Compute(const framework::ExecutionContext& context) const override { + PADDLE_ENFORCE(platform::is_cpu_place(context.GetPlace()), + "This kernel only runs on CPU."); + VLOG(3) << "Enter SampleLogitsKernel"; + // get necessary inputs + const Tensor* logits = context.Input("Logits"); + const Tensor* labels = context.Input("Labels"); + + // get necessary outputs + Tensor* samples = context.Output("Samples"); + Tensor* probabilities = context.Output("Probabilities"); + Tensor* sampled_logits = context.Output("SampledLogits"); + Tensor* sampled_labels = context.Output("SampledLabels"); + + // shapes + const auto batch_size = logits->dims()[0]; + const auto num_classes = logits->dims()[1]; + const auto labels_dim = labels->dims(); + const auto num_true = labels_dim[1]; + const auto samples_dim = samples->dims(); + + // attrs + const auto num_samples = context.Attr("num_samples"); + const bool use_customized_samples = + context.Attr("use_customized_samples"); + const bool remove_accidental_hits = + context.Attr("remove_accidental_hits"); + + // device contexts + auto& dev_ctx = + context.template device_context(); + + // UNDERSTAND: allocate memories for temporaries + sampled_logits->mutable_data(samples_dim, context.GetPlace()); + auto sampled_labels_data = + sampled_labels->mutable_data(labels_dim, context.GetPlace()); + for (int i = 0; i < batch_size; ++i) { + for (int j = 0; j < num_true; ++j) { + sampled_labels_data[i * num_true + j] = j; + } + } + + if (use_customized_samples) { + const Tensor* customized_samples = + context.Input("CustomizedSamples"); + const Tensor* customized_probabilities = + context.Input("CustomizedProbabilities"); + samples->ShareDataWith(*customized_samples); + probabilities->ShareDataWith(*customized_probabilities); + } else { + samples->mutable_data(context.GetPlace()); + probabilities->mutable_data(samples_dim, context.GetPlace()); + // UNDERSTAND: sampling + const auto seed = context.Attr("seed"); + auto sampler_with_prob = + math::SampleWithProb(); + sampler_with_prob(dev_ctx, math::LogUniformSampler(num_classes, seed), + num_samples, labels, samples, probabilities); + } + + // UNDERSTAND: gather sampled logits and remove accidental hits if needed + CPUTakeAlongD1(dev_ctx, *logits, *samples, sampled_logits); + if (remove_accidental_hits) { + compute_remove_accidental_hits(dev_ctx, sampled_logits, *samples, + num_true); + } + + // subtracted sampled logits with logQ(y|x) + auto probs = EigenMatrix::From(*probabilities); + auto smp_logits = EigenMatrix::From(*sampled_logits); + smp_logits.device(*dev_ctx.eigen_device()) = + (smp_logits - probs.log().unaryExpr(TolerableValue())) + .unaryExpr(TolerableValue()); + } +}; + +template +class SampleLogitsGradKernel : public framework::OpKernel { + public: + using Tensor = framework::Tensor; + void Compute(const framework::ExecutionContext& context) const override { + auto logits_grad = context.Output(framework::GradVarName("Logits")); + const Tensor* samples = context.Input("Samples"); + const Tensor* sampled_logits_grad = + context.Input(framework::GradVarName("SampledLogits")); + logits_grad->mutable_data(context.GetPlace()); + + auto& dev_ctx = + context.template device_context(); + math::SetConstant set_zero; + set_zero(dev_ctx, logits_grad, static_cast(0)); + + // UNDERSTAND: scatter it back to logit_grad + CPUPutAlongD1(dev_ctx, logits_grad, *samples, *sampled_logits_grad); + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/scale_op.cc b/paddle/fluid/operators/scale_op.cc index 981969d2aaa684731a615ec64ca7f7718b35cf09..4ea77ed30db212b694f2050952655dd1a42215bd 100644 --- a/paddle/fluid/operators/scale_op.cc +++ b/paddle/fluid/operators/scale_op.cc @@ -100,13 +100,14 @@ class ScaleGradMaker : public framework::SingleGradOpDescMaker { } }; +using ScaleOpInplace = framework::SingleOpInplaceInToOut; } // namespace operators } // namespace paddle namespace ops = paddle::operators; REGISTER_OPERATOR(scale, ops::ScaleOp, ops::ScaleOpMaker, ops::ScaleGradMaker, - ops::ScaleOpVarTypeInference); + ops::ScaleOpVarTypeInference, ops::ScaleOpInplace); REGISTER_OP_CPU_KERNEL( scale, ops::ScaleKernel, ops::ScaleKernel, diff --git a/paddle/fluid/operators/sequence_ops/sequence_enumerate_op.cc b/paddle/fluid/operators/sequence_ops/sequence_enumerate_op.cc index 1eebadc2c980ddf1cbaaefef1568dd401d0c77ed..0932211cadf30d0c464d43ca652a5c52df15747e 100644 --- a/paddle/fluid/operators/sequence_ops/sequence_enumerate_op.cc +++ b/paddle/fluid/operators/sequence_ops/sequence_enumerate_op.cc @@ -31,10 +31,10 @@ class SequenceEnumerateOp : public framework::OperatorWithKernel { const auto x_dims = ctx->GetInputDim("X"); PADDLE_ENFORCE_EQ( - x_dims.size(), 2UL, + x_dims.size(), 2, "Input(X) of SequenceEnumerate operator's rank should be 2."); PADDLE_ENFORCE_EQ( - x_dims[1], 1UL, + x_dims[1], 1, "Input(X) of SequenceEnumerate operator's 2nd dimension should be 1."); const auto win_size = ctx->Attrs().Get("win_size"); diff --git a/paddle/fluid/operators/sequence_ops/sequence_expand_op.cc b/paddle/fluid/operators/sequence_ops/sequence_expand_op.cc index 27e0201bd70df59c58eaa7567d5bb69eb1b721b4..f6c42415301bc8d6f3509bfba2ff356265643bad 100644 --- a/paddle/fluid/operators/sequence_ops/sequence_expand_op.cc +++ b/paddle/fluid/operators/sequence_ops/sequence_expand_op.cc @@ -48,10 +48,10 @@ class SequenceExpandOp : public framework::OperatorWithKernel { auto& x_lod = x_var->Get().lod(); auto& y_lod = y_var->Get().lod(); - PADDLE_ENFORCE_LE(x_lod.size(), 1, + PADDLE_ENFORCE_LE(x_lod.size(), 1UL, "Level number of Input(X)'s lod should not be " "greater than 1."); - PADDLE_ENFORCE_GT(y_lod.size(), 0, + PADDLE_ENFORCE_GT(y_lod.size(), 0UL, "Level number of Input(Y)'s lod should be " "greater than 0."); PADDLE_ENFORCE( @@ -69,7 +69,8 @@ class SequenceExpandOp : public framework::OperatorWithKernel { "size of Input(X)'s first level lod should be equal to " "size of Input(Y)'s referred level lod."); } else { - PADDLE_ENFORCE_EQ(x_dims[0], y_lod[ref_level].size() - 1, + PADDLE_ENFORCE_EQ(x_dims[0], + static_cast(y_lod[ref_level].size()) - 1, "When Input(X)'s lod is null, the dims[0] of " "Input(X) should match the " "size of Input(Y)'s referred level lod."); diff --git a/paddle/fluid/operators/shape_op.cc b/paddle/fluid/operators/shape_op.cc index 1be9fe47af71d31ce2e0eba807ea4a43601f8aca..efc497fa47d1d954bbd1e214b43f5de4c76b0714 100644 --- a/paddle/fluid/operators/shape_op.cc +++ b/paddle/fluid/operators/shape_op.cc @@ -35,14 +35,15 @@ class ShapeOp : public framework::OperatorWithKernel { class ShapeOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { - AddInput("Input", "(Tensor), The input tensor."); - AddOutput("Out", - "(Tensor), The shape of input tensor, the data type of the shape" - " is int32_t, will be on the same device with the input Tensor."); + AddInput("Input", "(LoDTensor), The input tensor."); + AddOutput( + "Out", + "(LoDTensor), The shape of input tensor, the data type of the shape" + " is int32_t, will be on the same device with the input Tensor."); AddComment(R"DOC( -Shape Operator +Shape Operator. -Get the shape of input tensor. Only support CPU input Tensor now. +Return the shape of the input. )DOC"); } }; diff --git a/paddle/fluid/operators/slice_op.cc b/paddle/fluid/operators/slice_op.cc index 789e61b2d332b9391ef45a8ebe58ad0f1a4d2bf0..94995fc99612adb1164e60f1a51747f74eacfb73 100644 --- a/paddle/fluid/operators/slice_op.cc +++ b/paddle/fluid/operators/slice_op.cc @@ -54,6 +54,9 @@ class SliceOp : public framework::OperatorWithKernel { out_dims[axes[i]] = end - start; } ctx->SetOutputDim("Out", out_dims); + if (axes[0] != 0) { + ctx->ShareLoD("Input", /*->*/ "Out"); + } } protected: diff --git a/paddle/fluid/operators/softmax_op.cc b/paddle/fluid/operators/softmax_op.cc index bc889a5a042a27838ba6ba0fccb187ec11b5f0c5..8fbf299a7c056aff3bfd4cbd3e3cc28fd3c6ccf2 100644 --- a/paddle/fluid/operators/softmax_op.cc +++ b/paddle/fluid/operators/softmax_op.cc @@ -198,6 +198,21 @@ class SoftmaxOpGradMaker : public framework::SingleGradOpDescMaker { return std::unique_ptr(op); } }; + +class SoftmaxInplaceInToOut : public framework::InplaceInToOut { + public: + using framework::InplaceInToOut::InplaceInToOut; + + protected: + std::unordered_map Apply( + const framework::OpDesc& op_desc, + framework::BlockDesc* block) const override { + return std::unordered_map{ + {"X", "Out"}, + }; + } +}; + } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/stack_op.cu b/paddle/fluid/operators/stack_op.cu index bf2a9e5b3d22996e688621727cb280dc9aed7859..24d0b2f906a8e0b360c3f477c9290ebe5d57a3ff 100644 --- a/paddle/fluid/operators/stack_op.cu +++ b/paddle/fluid/operators/stack_op.cu @@ -17,13 +17,16 @@ namespace plat = paddle::platform; namespace ops = paddle::operators; -REGISTER_OP_CUDA_KERNEL(stack, ops::StackKernel, - ops::StackKernel, - ops::StackKernel, - ops::StackKernel); +REGISTER_OP_CUDA_KERNEL( + stack, ops::StackKernel, + ops::StackKernel, + ops::StackKernel, + ops::StackKernel, + ops::StackKernel); -REGISTER_OP_CUDA_KERNEL(stack_grad, - ops::StackGradKernel, - ops::StackGradKernel, - ops::StackGradKernel, - ops::StackGradKernel); +REGISTER_OP_CUDA_KERNEL( + stack_grad, ops::StackGradKernel, + ops::StackGradKernel, + ops::StackGradKernel, + ops::StackGradKernel, + ops::StackGradKernel); diff --git a/paddle/fluid/operators/transpose_op.cu.cc b/paddle/fluid/operators/transpose_op.cu.cc index b4025350fa9f3610bde43eee91cd059f3063813f..915774e5f3624f26dbd1451a99d7bf0bf75a72c8 100644 --- a/paddle/fluid/operators/transpose_op.cu.cc +++ b/paddle/fluid/operators/transpose_op.cu.cc @@ -15,19 +15,27 @@ limitations under the License. */ #include "paddle/fluid/operators/transpose_op.h" namespace ops = paddle::operators; +namespace plat = paddle::platform; + REGISTER_OP_CUDA_KERNEL( transpose, ops::TransposeKernel, - ops::TransposeKernel); + ops::TransposeKernel, + ops::TransposeKernel); REGISTER_OP_CUDA_KERNEL( transpose_grad, ops::TransposeGradKernel, - ops::TransposeGradKernel); + ops::TransposeGradKernel, + ops::TransposeGradKernel); REGISTER_OP_CUDA_KERNEL( transpose2, ops::TransposeKernel, - ops::TransposeKernel); + ops::TransposeKernel, + ops::TransposeKernel); REGISTER_OP_CUDA_KERNEL( transpose2_grad, ops::TransposeGradKernel, - ops::TransposeGradKernel); + ops::TransposeGradKernel, + ops::TransposeGradKernel); diff --git a/paddle/fluid/operators/yolov3_loss_op.h b/paddle/fluid/operators/yolov3_loss_op.h deleted file mode 100644 index 0bb285722ddedf721d98237760ec9868e2134442..0000000000000000000000000000000000000000 --- a/paddle/fluid/operators/yolov3_loss_op.h +++ /dev/null @@ -1,483 +0,0 @@ -/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve. - Licensed under the Apache License, Version 2.0 (the "License"); - you may not use this file except in compliance with the License. - You may obtain a copy of the License at - http://www.apache.org/licenses/LICENSE-2.0 - Unless required by applicable law or agreed to in writing, software - distributed under the License is distributed on an "AS IS" BASIS, - WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - See the License for the specific language governing permissions and - limitations under the License. */ - -#pragma once -#include -#include -#include "paddle/fluid/framework/op_registry.h" - -namespace paddle { -namespace operators { - -using Tensor = framework::Tensor; -template -using EigenTensor = framework::EigenTensor; -template -using EigenVector = framework::EigenVector; - -using Array5 = Eigen::DSizes; - -template -static inline bool isZero(T x) { - return fabs(x) < 1e-6; -} - -template -static inline T sigmoid(T x) { - return 1.0 / (exp(-1.0 * x) + 1.0); -} - -template -static inline T CalcMaskPointNum(const Tensor& mask) { - auto mask_t = EigenVector::Flatten(mask); - T count = 0.0; - for (int i = 0; i < mask_t.dimensions()[0]; i++) { - if (mask_t(i)) { - count += 1.0; - } - } - return count; -} - -template -static inline T CalcMSEWithMask(const Tensor& x, const Tensor& y, - const Tensor& mask) { - auto x_t = EigenVector::Flatten(x); - auto y_t = EigenVector::Flatten(y); - auto mask_t = EigenVector::Flatten(mask); - - T error_sum = 0.0; - T points = 0.0; - for (int i = 0; i < x_t.dimensions()[0]; i++) { - if (mask_t(i)) { - error_sum += pow(x_t(i) - y_t(i), 2); - points += 1; - } - } - return (error_sum / points); -} - -template -static void CalcMSEGradWithMask(Tensor* grad, const Tensor& x, const Tensor& y, - const Tensor& mask, T mf) { - auto grad_t = EigenVector::Flatten(*grad).setConstant(0.0); - auto x_t = EigenVector::Flatten(x); - auto y_t = EigenVector::Flatten(y); - auto mask_t = EigenVector::Flatten(mask); - - for (int i = 0; i < x_t.dimensions()[0]; i++) { - if (mask_t(i)) { - grad_t(i) = 2.0 * (x_t(i) - y_t(i)) / mf; - } - } -} - -template -static inline T CalcBCEWithMask(const Tensor& x, const Tensor& y, - const Tensor& mask) { - auto x_t = EigenVector::Flatten(x); - auto y_t = EigenVector::Flatten(y); - auto mask_t = EigenVector::Flatten(mask); - - T error_sum = 0.0; - T points = 0.0; - for (int i = 0; i < x_t.dimensions()[0]; i++) { - if (mask_t(i)) { - error_sum += - -1.0 * (y_t(i) * log(x_t(i)) + (1.0 - y_t(i)) * log(1.0 - x_t(i))); - points += 1; - } - } - return (error_sum / points); -} - -template -static inline void CalcBCEGradWithMask(Tensor* grad, const Tensor& x, - const Tensor& y, const Tensor& mask, - T mf) { - auto grad_t = EigenVector::Flatten(*grad).setConstant(0.0); - auto x_t = EigenVector::Flatten(x); - auto y_t = EigenVector::Flatten(y); - auto mask_t = EigenVector::Flatten(mask); - - for (int i = 0; i < x_t.dimensions()[0]; i++) { - if (mask_t(i)) { - grad_t(i) = ((1.0 - y_t(i)) / (1.0 - x_t(i)) - y_t(i) / x_t(i)) / mf; - } - } -} - -template -static void CalcPredResult(const Tensor& input, Tensor* pred_conf, - Tensor* pred_class, Tensor* pred_x, Tensor* pred_y, - Tensor* pred_w, Tensor* pred_h, const int anchor_num, - const int class_num) { - const int n = input.dims()[0]; - const int h = input.dims()[2]; - const int w = input.dims()[3]; - const int box_attr_num = 5 + class_num; - - auto input_t = EigenTensor::From(input); - auto pred_conf_t = EigenTensor::From(*pred_conf); - auto pred_class_t = EigenTensor::From(*pred_class); - auto pred_x_t = EigenTensor::From(*pred_x); - auto pred_y_t = EigenTensor::From(*pred_y); - auto pred_w_t = EigenTensor::From(*pred_w); - auto pred_h_t = EigenTensor::From(*pred_h); - - for (int i = 0; i < n; i++) { - for (int an_idx = 0; an_idx < anchor_num; an_idx++) { - for (int j = 0; j < h; j++) { - for (int k = 0; k < w; k++) { - pred_x_t(i, an_idx, j, k) = - sigmoid(input_t(i, box_attr_num * an_idx, j, k)); - pred_y_t(i, an_idx, j, k) = - sigmoid(input_t(i, box_attr_num * an_idx + 1, j, k)); - pred_w_t(i, an_idx, j, k) = - input_t(i, box_attr_num * an_idx + 2, j, k); - pred_h_t(i, an_idx, j, k) = - input_t(i, box_attr_num * an_idx + 3, j, k); - - pred_conf_t(i, an_idx, j, k) = - sigmoid(input_t(i, box_attr_num * an_idx + 4, j, k)); - - for (int c = 0; c < class_num; c++) { - pred_class_t(i, an_idx, j, k, c) = - sigmoid(input_t(i, box_attr_num * an_idx + 5 + c, j, k)); - } - } - } - } - } -} - -template -static T CalcBoxIoU(std::vector box1, std::vector box2) { - T b1_x1 = box1[0] - box1[2] / 2; - T b1_x2 = box1[0] + box1[2] / 2; - T b1_y1 = box1[1] - box1[3] / 2; - T b1_y2 = box1[1] + box1[3] / 2; - T b2_x1 = box2[0] - box2[2] / 2; - T b2_x2 = box2[0] + box2[2] / 2; - T b2_y1 = box2[1] - box2[3] / 2; - T b2_y2 = box2[1] + box2[3] / 2; - - T b1_area = (b1_x2 - b1_x1) * (b1_y2 - b1_y1); - T b2_area = (b2_x2 - b2_x1) * (b2_y2 - b2_y1); - - T inter_rect_x1 = std::max(b1_x1, b2_x1); - T inter_rect_y1 = std::max(b1_y1, b2_y1); - T inter_rect_x2 = std::min(b1_x2, b2_x2); - T inter_rect_y2 = std::min(b1_y2, b2_y2); - T inter_area = std::max(inter_rect_x2 - inter_rect_x1, static_cast(0.0)) * - std::max(inter_rect_y2 - inter_rect_y1, static_cast(0.0)); - - return inter_area / (b1_area + b2_area - inter_area); -} - -template -static void PreProcessGTBox(const Tensor& gt_box, const Tensor& gt_label, - const float ignore_thresh, std::vector anchors, - const int grid_size, Tensor* obj_mask, - Tensor* noobj_mask, Tensor* tx, Tensor* ty, - Tensor* tw, Tensor* th, Tensor* tconf, - Tensor* tclass) { - const int n = gt_box.dims()[0]; - const int b = gt_box.dims()[1]; - const int anchor_num = anchors.size() / 2; - auto gt_box_t = EigenTensor::From(gt_box); - auto gt_label_t = EigenTensor::From(gt_label); - auto obj_mask_t = EigenTensor::From(*obj_mask).setConstant(0); - auto noobj_mask_t = EigenTensor::From(*noobj_mask).setConstant(1); - auto tx_t = EigenTensor::From(*tx).setConstant(0.0); - auto ty_t = EigenTensor::From(*ty).setConstant(0.0); - auto tw_t = EigenTensor::From(*tw).setConstant(0.0); - auto th_t = EigenTensor::From(*th).setConstant(0.0); - auto tconf_t = EigenTensor::From(*tconf).setConstant(0.0); - auto tclass_t = EigenTensor::From(*tclass).setConstant(0.0); - - for (int i = 0; i < n; i++) { - for (int j = 0; j < b; j++) { - if (isZero(gt_box_t(i, j, 0)) && isZero(gt_box_t(i, j, 1)) && - isZero(gt_box_t(i, j, 2)) && isZero(gt_box_t(i, j, 3))) { - continue; - } - - int cur_label = gt_label_t(i, j); - T gx = gt_box_t(i, j, 0) * grid_size; - T gy = gt_box_t(i, j, 1) * grid_size; - T gw = gt_box_t(i, j, 2) * grid_size; - T gh = gt_box_t(i, j, 3) * grid_size; - int gi = static_cast(gx); - int gj = static_cast(gy); - - T max_iou = static_cast(0); - T iou; - int best_an_index = -1; - std::vector gt_box_shape({0, 0, gw, gh}); - for (int an_idx = 0; an_idx < anchor_num; an_idx++) { - std::vector anchor_shape({0, 0, static_cast(anchors[2 * an_idx]), - static_cast(anchors[2 * an_idx + 1])}); - iou = CalcBoxIoU(gt_box_shape, anchor_shape); - if (iou > max_iou) { - max_iou = iou; - best_an_index = an_idx; - } - if (iou > ignore_thresh) { - noobj_mask_t(i, an_idx, gj, gi) = 0; - } - } - obj_mask_t(i, best_an_index, gj, gi) = 1; - noobj_mask_t(i, best_an_index, gj, gi) = 0; - tx_t(i, best_an_index, gj, gi) = gx - gi; - ty_t(i, best_an_index, gj, gi) = gy - gj; - tw_t(i, best_an_index, gj, gi) = log(gw / anchors[2 * best_an_index]); - th_t(i, best_an_index, gj, gi) = log(gh / anchors[2 * best_an_index + 1]); - tclass_t(i, best_an_index, gj, gi, cur_label) = 1; - tconf_t(i, best_an_index, gj, gi) = 1; - } - } -} - -static void ExpandObjMaskByClassNum(Tensor* obj_mask_expand, - const Tensor& obj_mask) { - const int n = obj_mask_expand->dims()[0]; - const int an_num = obj_mask_expand->dims()[1]; - const int h = obj_mask_expand->dims()[2]; - const int w = obj_mask_expand->dims()[3]; - const int class_num = obj_mask_expand->dims()[4]; - auto obj_mask_expand_t = EigenTensor::From(*obj_mask_expand); - auto obj_mask_t = EigenTensor::From(obj_mask); - - obj_mask_expand_t = obj_mask_t.reshape(Array5(n, an_num, h, w, 1)) - .broadcast(Array5(1, 1, 1, 1, class_num)); -} - -template -static void AddAllGradToInputGrad( - Tensor* grad, T loss, const Tensor& pred_x, const Tensor& pred_y, - const Tensor& pred_conf, const Tensor& pred_class, const Tensor& grad_x, - const Tensor& grad_y, const Tensor& grad_w, const Tensor& grad_h, - const Tensor& grad_conf_target, const Tensor& grad_conf_notarget, - const Tensor& grad_class, const int class_num, const float loss_weight_xy, - const float loss_weight_wh, const float loss_weight_conf_target, - const float loss_weight_conf_notarget, const float loss_weight_class) { - const int n = pred_x.dims()[0]; - const int an_num = pred_x.dims()[1]; - const int h = pred_x.dims()[2]; - const int w = pred_x.dims()[3]; - const int attr_num = class_num + 5; - auto grad_t = EigenTensor::From(*grad).setConstant(0.0); - auto pred_x_t = EigenTensor::From(pred_x); - auto pred_y_t = EigenTensor::From(pred_y); - auto pred_conf_t = EigenTensor::From(pred_conf); - auto pred_class_t = EigenTensor::From(pred_class); - auto grad_x_t = EigenTensor::From(grad_x); - auto grad_y_t = EigenTensor::From(grad_y); - auto grad_w_t = EigenTensor::From(grad_w); - auto grad_h_t = EigenTensor::From(grad_h); - auto grad_conf_target_t = EigenTensor::From(grad_conf_target); - auto grad_conf_notarget_t = EigenTensor::From(grad_conf_notarget); - auto grad_class_t = EigenTensor::From(grad_class); - - for (int i = 0; i < n; i++) { - for (int j = 0; j < an_num; j++) { - for (int k = 0; k < h; k++) { - for (int l = 0; l < w; l++) { - grad_t(i, j * attr_num, k, l) = - grad_x_t(i, j, k, l) * pred_x_t(i, j, k, l) * - (1.0 - pred_x_t(i, j, k, l)) * loss * loss_weight_xy; - grad_t(i, j * attr_num + 1, k, l) = - grad_y_t(i, j, k, l) * pred_y_t(i, j, k, l) * - (1.0 - pred_y_t(i, j, k, l)) * loss * loss_weight_xy; - grad_t(i, j * attr_num + 2, k, l) = - grad_w_t(i, j, k, l) * loss * loss_weight_wh; - grad_t(i, j * attr_num + 3, k, l) = - grad_h_t(i, j, k, l) * loss * loss_weight_wh; - grad_t(i, j * attr_num + 4, k, l) = - grad_conf_target_t(i, j, k, l) * pred_conf_t(i, j, k, l) * - (1.0 - pred_conf_t(i, j, k, l)) * loss * loss_weight_conf_target; - grad_t(i, j * attr_num + 4, k, l) += - grad_conf_notarget_t(i, j, k, l) * pred_conf_t(i, j, k, l) * - (1.0 - pred_conf_t(i, j, k, l)) * loss * - loss_weight_conf_notarget; - - for (int c = 0; c < class_num; c++) { - grad_t(i, j * attr_num + 5 + c, k, l) = - grad_class_t(i, j, k, l, c) * pred_class_t(i, j, k, l, c) * - (1.0 - pred_class_t(i, j, k, l, c)) * loss * loss_weight_class; - } - } - } - } - } -} - -template -class Yolov3LossKernel : public framework::OpKernel { - public: - void Compute(const framework::ExecutionContext& ctx) const override { - auto* input = ctx.Input("X"); - auto* gt_box = ctx.Input("GTBox"); - auto* gt_label = ctx.Input("GTLabel"); - auto* loss = ctx.Output("Loss"); - auto anchors = ctx.Attr>("anchors"); - int class_num = ctx.Attr("class_num"); - float ignore_thresh = ctx.Attr("ignore_thresh"); - float loss_weight_xy = ctx.Attr("loss_weight_xy"); - float loss_weight_wh = ctx.Attr("loss_weight_wh"); - float loss_weight_conf_target = ctx.Attr("loss_weight_conf_target"); - float loss_weight_conf_notarget = - ctx.Attr("loss_weight_conf_notarget"); - float loss_weight_class = ctx.Attr("loss_weight_class"); - - const int n = input->dims()[0]; - const int h = input->dims()[2]; - const int w = input->dims()[3]; - const int an_num = anchors.size() / 2; - - Tensor pred_x, pred_y, pred_w, pred_h; - Tensor pred_conf, pred_class; - pred_x.mutable_data({n, an_num, h, w}, ctx.GetPlace()); - pred_y.mutable_data({n, an_num, h, w}, ctx.GetPlace()); - pred_w.mutable_data({n, an_num, h, w}, ctx.GetPlace()); - pred_h.mutable_data({n, an_num, h, w}, ctx.GetPlace()); - pred_conf.mutable_data({n, an_num, h, w}, ctx.GetPlace()); - pred_class.mutable_data({n, an_num, h, w, class_num}, ctx.GetPlace()); - CalcPredResult(*input, &pred_conf, &pred_class, &pred_x, &pred_y, - &pred_w, &pred_h, an_num, class_num); - - Tensor obj_mask, noobj_mask; - Tensor tx, ty, tw, th, tconf, tclass; - obj_mask.mutable_data({n, an_num, h, w}, ctx.GetPlace()); - noobj_mask.mutable_data({n, an_num, h, w}, ctx.GetPlace()); - tx.mutable_data({n, an_num, h, w}, ctx.GetPlace()); - ty.mutable_data({n, an_num, h, w}, ctx.GetPlace()); - tw.mutable_data({n, an_num, h, w}, ctx.GetPlace()); - th.mutable_data({n, an_num, h, w}, ctx.GetPlace()); - tconf.mutable_data({n, an_num, h, w}, ctx.GetPlace()); - tclass.mutable_data({n, an_num, h, w, class_num}, ctx.GetPlace()); - PreProcessGTBox(*gt_box, *gt_label, ignore_thresh, anchors, h, &obj_mask, - &noobj_mask, &tx, &ty, &tw, &th, &tconf, &tclass); - - Tensor obj_mask_expand; - obj_mask_expand.mutable_data({n, an_num, h, w, class_num}, - ctx.GetPlace()); - ExpandObjMaskByClassNum(&obj_mask_expand, obj_mask); - - T loss_x = CalcMSEWithMask(pred_x, tx, obj_mask); - T loss_y = CalcMSEWithMask(pred_y, ty, obj_mask); - T loss_w = CalcMSEWithMask(pred_w, tw, obj_mask); - T loss_h = CalcMSEWithMask(pred_h, th, obj_mask); - T loss_conf_target = CalcBCEWithMask(pred_conf, tconf, obj_mask); - T loss_conf_notarget = CalcBCEWithMask(pred_conf, tconf, noobj_mask); - T loss_class = CalcBCEWithMask(pred_class, tclass, obj_mask_expand); - - auto* loss_data = loss->mutable_data({1}, ctx.GetPlace()); - loss_data[0] = loss_weight_xy * (loss_x + loss_y) + - loss_weight_wh * (loss_w + loss_h) + - loss_weight_conf_target * loss_conf_target + - loss_weight_conf_notarget * loss_conf_notarget + - loss_weight_class * loss_class; - } -}; - -template -class Yolov3LossGradKernel : public framework::OpKernel { - public: - void Compute(const framework::ExecutionContext& ctx) const override { - auto* input = ctx.Input("X"); - auto* gt_box = ctx.Input("GTBox"); - auto* gt_label = ctx.Input("GTLabel"); - auto anchors = ctx.Attr>("anchors"); - int class_num = ctx.Attr("class_num"); - float ignore_thresh = ctx.Attr("ignore_thresh"); - auto* input_grad = ctx.Output(framework::GradVarName("X")); - auto* output_grad = ctx.Input(framework::GradVarName("Loss")); - const T loss = output_grad->data()[0]; - float loss_weight_xy = ctx.Attr("loss_weight_xy"); - float loss_weight_wh = ctx.Attr("loss_weight_wh"); - float loss_weight_conf_target = ctx.Attr("loss_weight_conf_target"); - float loss_weight_conf_notarget = - ctx.Attr("loss_weight_conf_notarget"); - float loss_weight_class = ctx.Attr("loss_weight_class"); - - const int n = input->dims()[0]; - const int c = input->dims()[1]; - const int h = input->dims()[2]; - const int w = input->dims()[3]; - const int an_num = anchors.size() / 2; - - Tensor pred_x, pred_y, pred_w, pred_h; - Tensor pred_conf, pred_class; - pred_x.mutable_data({n, an_num, h, w}, ctx.GetPlace()); - pred_y.mutable_data({n, an_num, h, w}, ctx.GetPlace()); - pred_w.mutable_data({n, an_num, h, w}, ctx.GetPlace()); - pred_h.mutable_data({n, an_num, h, w}, ctx.GetPlace()); - pred_conf.mutable_data({n, an_num, h, w}, ctx.GetPlace()); - pred_class.mutable_data({n, an_num, h, w, class_num}, ctx.GetPlace()); - CalcPredResult(*input, &pred_conf, &pred_class, &pred_x, &pred_y, - &pred_w, &pred_h, an_num, class_num); - - Tensor obj_mask, noobj_mask; - Tensor tx, ty, tw, th, tconf, tclass; - obj_mask.mutable_data({n, an_num, h, w}, ctx.GetPlace()); - noobj_mask.mutable_data({n, an_num, h, w}, ctx.GetPlace()); - tx.mutable_data({n, an_num, h, w}, ctx.GetPlace()); - ty.mutable_data({n, an_num, h, w}, ctx.GetPlace()); - tw.mutable_data({n, an_num, h, w}, ctx.GetPlace()); - th.mutable_data({n, an_num, h, w}, ctx.GetPlace()); - tconf.mutable_data({n, an_num, h, w}, ctx.GetPlace()); - tclass.mutable_data({n, an_num, h, w, class_num}, ctx.GetPlace()); - PreProcessGTBox(*gt_box, *gt_label, ignore_thresh, anchors, h, &obj_mask, - &noobj_mask, &tx, &ty, &tw, &th, &tconf, &tclass); - - Tensor obj_mask_expand; - obj_mask_expand.mutable_data({n, an_num, h, w, class_num}, - ctx.GetPlace()); - ExpandObjMaskByClassNum(&obj_mask_expand, obj_mask); - - Tensor grad_x, grad_y, grad_w, grad_h; - Tensor grad_conf_target, grad_conf_notarget, grad_class; - grad_x.mutable_data({n, an_num, h, w}, ctx.GetPlace()); - grad_y.mutable_data({n, an_num, h, w}, ctx.GetPlace()); - grad_w.mutable_data({n, an_num, h, w}, ctx.GetPlace()); - grad_h.mutable_data({n, an_num, h, w}, ctx.GetPlace()); - grad_conf_target.mutable_data({n, an_num, h, w}, ctx.GetPlace()); - grad_conf_notarget.mutable_data({n, an_num, h, w}, ctx.GetPlace()); - grad_class.mutable_data({n, an_num, h, w, class_num}, ctx.GetPlace()); - T obj_mf = CalcMaskPointNum(obj_mask); - T noobj_mf = CalcMaskPointNum(noobj_mask); - T obj_expand_mf = CalcMaskPointNum(obj_mask_expand); - CalcMSEGradWithMask(&grad_x, pred_x, tx, obj_mask, obj_mf); - CalcMSEGradWithMask(&grad_y, pred_y, ty, obj_mask, obj_mf); - CalcMSEGradWithMask(&grad_w, pred_w, tw, obj_mask, obj_mf); - CalcMSEGradWithMask(&grad_h, pred_h, th, obj_mask, obj_mf); - CalcBCEGradWithMask(&grad_conf_target, pred_conf, tconf, obj_mask, - obj_mf); - CalcBCEGradWithMask(&grad_conf_notarget, pred_conf, tconf, noobj_mask, - noobj_mf); - CalcBCEGradWithMask(&grad_class, pred_class, tclass, obj_mask_expand, - obj_expand_mf); - - input_grad->mutable_data({n, c, h, w}, ctx.GetPlace()); - AddAllGradToInputGrad( - input_grad, loss, pred_x, pred_y, pred_conf, pred_class, grad_x, grad_y, - grad_w, grad_h, grad_conf_target, grad_conf_notarget, grad_class, - class_num, loss_weight_xy, loss_weight_wh, loss_weight_conf_target, - loss_weight_conf_notarget, loss_weight_class); - } -}; - -} // namespace operators -} // namespace paddle diff --git a/paddle/fluid/platform/CMakeLists.txt b/paddle/fluid/platform/CMakeLists.txt index 1f51b5bab3068cc89bffa85de28a9438359659f3..b7e84031e7b7e821d51caf405978d07995ea0e91 100644 --- a/paddle/fluid/platform/CMakeLists.txt +++ b/paddle/fluid/platform/CMakeLists.txt @@ -1,4 +1,4 @@ -proto_library(profiler_proto SRCS profiler.proto DEPS framework_proto) +proto_library(profiler_proto SRCS profiler.proto DEPS framework_proto simple_threadpool) py_proto_compile(profiler_py_proto SRCS profiler.proto) add_custom_target(profiler_py_proto_init ALL COMMAND ${CMAKE_COMMAND} -E touch __init__.py) @@ -87,8 +87,12 @@ nv_test(transform_test SRCS transform_test.cu DEPS memory place device_context) cc_library(timer SRCS timer.cc) cc_test(timer_test SRCS timer_test.cc DEPS timer) -cc_library(device_tracer SRCS device_tracer.cc DEPS boost profiler_proto framework_proto ${GPU_CTX_DEPS}) -cc_library(profiler SRCS profiler.cc DEPS device_context device_tracer) +cc_library(device_tracer SRCS device_tracer.cc DEPS boost profiler_proto framework_proto device_context ${GPU_CTX_DEPS}) +if(WITH_GPU) + nv_library(profiler SRCS profiler.cc profiler.cu DEPS device_context device_tracer) +else() + cc_library(profiler SRCS profiler.cc DEPS device_context device_tracer) +endif() cc_test(profiler_test SRCS profiler_test.cc DEPS profiler) nv_test(float16_gpu_test SRCS float16_test.cu DEPS lod_tensor) diff --git a/paddle/fluid/platform/cuda_device_function.h b/paddle/fluid/platform/cuda_device_function.h index 2ce8f141d3c51661305f4952479cf2889fc4f396..31b6c38d613cf9df8fa7e8f6a8e1cfa310280968 100644 --- a/paddle/fluid/platform/cuda_device_function.h +++ b/paddle/fluid/platform/cuda_device_function.h @@ -53,10 +53,12 @@ inline static int RoundToPowerOfTwo(int dim) { __VA_ARGS__; \ } break -#define CUDA_LAUNCH_KERNEL_HELPER(...) \ - CUDA_LAUNCH_KERNEL_BASE(256, ##__VA_ARGS__); \ - CUDA_LAUNCH_KERNEL_BASE(128, ##__VA_ARGS__); \ - CUDA_LAUNCH_KERNEL_BASE(64, ##__VA_ARGS__); \ +#define CUDA_LAUNCH_KERNEL_HELPER(...) \ + CUDA_LAUNCH_KERNEL_BASE(1024, ##__VA_ARGS__); \ + CUDA_LAUNCH_KERNEL_BASE(512, ##__VA_ARGS__); \ + CUDA_LAUNCH_KERNEL_BASE(256, ##__VA_ARGS__); \ + CUDA_LAUNCH_KERNEL_BASE(128, ##__VA_ARGS__); \ + CUDA_LAUNCH_KERNEL_BASE(64, ##__VA_ARGS__); \ CUDA_LAUNCH_KERNEL_BASE(32, ##__VA_ARGS__); template diff --git a/paddle/fluid/platform/device_context.cc b/paddle/fluid/platform/device_context.cc index 2493fb71c019f9923012afa4a46cb3e95479f860..920b43b2b1990af58b73888bf7a652d57c20563c 100644 --- a/paddle/fluid/platform/device_context.cc +++ b/paddle/fluid/platform/device_context.cc @@ -291,7 +291,7 @@ CUDADeviceContext::CUDADeviceContext(CUDAPlace place) if (dynload::HasCUDNN()) { auto local_cudnn_version = cudnn_dso_ver / 100; auto compile_cudnn_version = CUDNN_VERSION / 100; - if (local_cudnn_version < compile_cudnn_version) { + if (local_cudnn_version < static_cast(compile_cudnn_version)) { LOG_FIRST_N(WARNING, 1) << "WARNING: device: " << place_.device << ". The installed Paddle is compiled with CUDNN " @@ -394,7 +394,7 @@ void MKLDNNDeviceContext::SetBlob(const std::string& name, int tid = platform::get_cur_thread_id(); - std::lock_guard lock(*p_mutex_.get()); + std::lock_guard lock(*p_mutex_); // Find KeyBlob for current thread auto map_it = pMap->find(tid); @@ -427,7 +427,7 @@ std::shared_ptr MKLDNNDeviceContext::GetBlob( int tid = platform::get_cur_thread_id(); - std::lock_guard lock(*p_mutex_.get()); + std::lock_guard lock(*p_mutex_); // Find KeyBlob for current thread firstly auto map_it = pMap->find(tid); diff --git a/paddle/fluid/platform/device_tracer.cc b/paddle/fluid/platform/device_tracer.cc index 0a4563ead65b1e45adca1d1a1fce066a1a55d932..0179daa55715be9787bc7cc8a693319024d404b7 100644 --- a/paddle/fluid/platform/device_tracer.cc +++ b/paddle/fluid/platform/device_tracer.cc @@ -14,17 +14,23 @@ limitations under the License. */ #include "paddle/fluid/platform/device_tracer.h" #include +#include #include +#include #include #include // NOLINT #include +#include #include #include // NOLINT +#include +#include #include #include "glog/logging.h" #include "google/protobuf/text_format.h" #include "paddle/fluid/framework/block_desc.h" +#include "paddle/fluid/platform/profiler.h" #include "paddle/fluid/string/printf.h" namespace paddle { @@ -33,17 +39,31 @@ namespace { // Tracking the nested block stacks of each thread. thread_local std::deque block_id_stack; // Tracking the nested event stacks. -thread_local std::deque annotation_stack; +thread_local std::deque annotation_stack; + +std::map system_thread_id_map; std::once_flag tracer_once_flag; DeviceTracer *tracer = nullptr; + +void PrintCuptiHint() { + static bool showed = false; + if (showed) return; + showed = true; + LOG(WARNING) << "Invalid timestamp occured. Please try increasing the " + "FLAGS_multiple_of_cupti_buffer_size."; +} + } // namespace #ifdef PADDLE_WITH_CUPTI namespace { -// TODO(panyx0718): Revisit the buffer size here. -uint64_t kBufSize = 32 * 1024; +// The experimental best performance is +// the same size with CUPTI device buffer size(8M) +uint64_t kBufSize = 1024 * 1024 * 8; uint64_t kAlignSize = 8; +std::unordered_map runtime_cbid_str, + driver_cbid_str; #define ALIGN_BUFFER(buffer, align) \ (((uintptr_t)(buffer) & ((align)-1)) \ @@ -92,15 +112,33 @@ std::string MemcpyKind(CUpti_ActivityMemcpyKind kind) { return "MEMCPY"; } +std::string DriverKind(CUpti_CallbackId cbid) { + auto iter = driver_cbid_str.find(cbid); + if (iter == driver_cbid_str.end()) + return "Driver API " + std::to_string(cbid); + return iter->second; +} + +std::string RuntimeKind(CUpti_CallbackId cbid) { + auto iter = runtime_cbid_str.find(cbid); + if (iter == runtime_cbid_str.end()) + return "Runtime API " + std::to_string(cbid); + return iter->second; +} + void EnableActivity() { // Device activity record is created when CUDA initializes, so we // want to enable it before cuInit() or any CUDA runtime call. CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_MEMCPY)); - CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_KERNEL)); - CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_DEVICE)); - CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_MEMSET)); - CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_OVERHEAD)); + CUPTI_CALL( + dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_CONCURRENT_KERNEL)); + // CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_KERNEL)); + CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_DRIVER)); + CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_RUNTIME)); // We don't track these activities for now. + CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_MEMSET)); + // CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_OVERHEAD)); + // CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_DEVICE)); // CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_CONTEXT)); // CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_DRIVER)); // CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_RUNTIME)); @@ -110,16 +148,17 @@ void EnableActivity() { void DisableActivity() { CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_MEMCPY)); - CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_KERNEL)); - CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_DEVICE)); + CUPTI_CALL( + dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_CONCURRENT_KERNEL)); + // CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_DEVICE)); // Disable all other activity record kinds. - CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_CONTEXT)); + // CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_CONTEXT)); CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_DRIVER)); CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_RUNTIME)); CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_MEMSET)); - CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_NAME)); - CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_MARKER)); - CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_OVERHEAD)); + // CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_NAME)); + // CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_MARKER)); + // CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_OVERHEAD)); } void CUPTIAPI bufferRequested(uint8_t **buffer, size_t *size, @@ -132,6 +171,11 @@ void CUPTIAPI bufferRequested(uint8_t **buffer, size_t *size, void CUPTIAPI bufferCompleted(CUcontext ctx, uint32_t streamId, uint8_t *buffer, size_t size, size_t validSize) { + static std::thread::id cupti_thread_id(0); + if (cupti_thread_id == std::thread::id(0)) + cupti_thread_id = std::this_thread::get_id(); + PADDLE_ENFORCE_EQ(std::this_thread::get_id(), cupti_thread_id, + "Only one thread is allowed to call bufferCompleted()"); CUptiResult status; CUpti_Activity *record = NULL; if (validSize > 0) { @@ -168,6 +212,31 @@ void CUPTIAPI bufferCompleted(CUcontext ctx, uint32_t streamId, uint8_t *buffer, memcpy->correlationId, memcpy->bytes); break; } + case CUPTI_ACTIVITY_KIND_MEMSET: { + auto *memset = + reinterpret_cast(record); + tracer->AddKernelRecords("MEMSET", memset->start, memset->end, + memset->deviceId, memset->streamId, + memset->correlationId); + break; + } + case CUPTI_ACTIVITY_KIND_DRIVER: { + auto *api = reinterpret_cast(record); + if (api->start != 0 && api->end != 0) + // -1 device id represents CUDA api call + tracer->AddCPURecords( + DriverKind(api->cbid), api->start, api->end, -1, + GetThreadIdFromSystemThreadId(api->threadId)); + break; + } + case CUPTI_ACTIVITY_KIND_RUNTIME: { + auto *api = reinterpret_cast(record); + if (api->start != 0 && api->end != 0) + tracer->AddCPURecords( + RuntimeKind(api->cbid), api->start, api->end, -1, + GetThreadIdFromSystemThreadId(api->threadId)); + break; + } default: { break; } } } else if (status == CUPTI_ERROR_MAX_LIMIT_REACHED) { @@ -183,21 +252,35 @@ void CUPTIAPI bufferCompleted(CUcontext ctx, uint32_t streamId, uint8_t *buffer, dynload::cuptiActivityGetNumDroppedRecords(ctx, streamId, &dropped)); if (dropped != 0) { fprintf(stderr, "Dropped %u activity records\n", (unsigned int)dropped); + PrintCuptiHint(); } } free(buffer); } + +void initCuptiCbidStr(); + } // namespace #endif // PADDLE_WITH_CUPTI class DeviceTracerImpl : public DeviceTracer { public: - DeviceTracerImpl() : enabled_(false) {} + DeviceTracerImpl() : enabled_(false) { +#ifdef PADDLE_WITH_CUPTI + initCuptiCbidStr(); +#endif + } - void AddAnnotation(uint64_t id, const std::string &anno) { - std::lock_guard l(trace_mu_); - correlations_[id] = anno; + void AddAnnotation(uint32_t id, Event *event) { + thread_local std::forward_list> + *local_correlations_pairs = nullptr; + if (local_correlations_pairs == nullptr) { + std::lock_guard l(trace_mu_); + correlations_pairs.emplace_front(); + local_correlations_pairs = &correlations_pairs.front(); + } + local_correlations_pairs->push_front(std::make_pair(id, event)); } void AddCPURecords(const std::string &anno, uint64_t start_ns, @@ -206,8 +289,13 @@ class DeviceTracerImpl : public DeviceTracer { VLOG(1) << "Empty timeline annotation."; return; } - std::lock_guard l(trace_mu_); - cpu_records_.push_back( + thread_local std::forward_list *local_cpu_records_ = nullptr; + if (local_cpu_records_ == nullptr) { + std::lock_guard l(trace_mu_); + cpu_records_.emplace_front(); + local_cpu_records_ = &cpu_records_.front(); + } + local_cpu_records_->push_front( CPURecord{anno, start_ns, end_ns, device_id, thread_id}); } @@ -215,25 +303,27 @@ class DeviceTracerImpl : public DeviceTracer { uint64_t end_ns, int64_t device_id, int64_t stream_id, uint32_t correlation_id, uint64_t bytes) { // 0 means timestamp information could not be collected for the kernel. - if (start_ns == 0 || end_ns == 0) { + if (start_ns == 0 || end_ns == 0 || start_ns == end_ns) { VLOG(3) << name << " cannot be traced"; + PrintCuptiHint(); return; } - std::lock_guard l(trace_mu_); - mem_records_.push_back(MemRecord{name, start_ns, end_ns, device_id, - stream_id, correlation_id, bytes}); + // NOTE(liangdun): lock is not needed, only one thread call this function. + mem_records_.push_front(MemRecord{name, start_ns, end_ns, device_id, + stream_id, correlation_id, bytes}); } void AddKernelRecords(std::string name, uint64_t start, uint64_t end, int64_t device_id, int64_t stream_id, uint32_t correlation_id) { // 0 means timestamp information could not be collected for the kernel. - if (start == 0 || end == 0) { + if (start == 0 || end == 0 || start == end) { VLOG(3) << correlation_id << " cannot be traced"; + PrintCuptiHint(); return; } - std::lock_guard l(trace_mu_); - kernel_records_.push_back( + // NOTE(liangdun): lock is not needed, only one thread call this function. + kernel_records_.push_front( KernelRecord{name, start, end, device_id, stream_id, correlation_id}); } @@ -263,25 +353,82 @@ class DeviceTracerImpl : public DeviceTracer { } else if (ret != CUPTI_SUCCESS) { fprintf(stderr, "Failed to create CUPTI subscriber.\n"); } - CUPTI_CALL( - dynload::cuptiEnableCallback(1, subscriber_, CUPTI_CB_DOMAIN_DRIVER_API, - CUPTI_DRIVER_TRACE_CBID_cuLaunchKernel)); + const std::vector cbids { + CUPTI_RUNTIME_TRACE_CBID_cudaMemcpy_v3020, + CUPTI_RUNTIME_TRACE_CBID_cudaMemcpyAsync_v3020, + CUPTI_RUNTIME_TRACE_CBID_cudaMemset_v3020, + CUPTI_RUNTIME_TRACE_CBID_cudaMemsetAsync_v3020, + CUPTI_RUNTIME_TRACE_CBID_cudaLaunch_v3020, + CUPTI_RUNTIME_TRACE_CBID_cudaLaunchKernel_v7000 +#if CUDA_VERSION >= 9000 + , + CUPTI_RUNTIME_TRACE_CBID_cudaLaunchCooperativeKernel_v9000, + CUPTI_RUNTIME_TRACE_CBID_cudaLaunchCooperativeKernelMultiDevice_v9000 +#endif + }; + for (auto cbid : cbids) + CUPTI_CALL(dynload::cuptiEnableCallback( + 1, subscriber_, CUPTI_CB_DOMAIN_RUNTIME_API, cbid)); CUPTI_CALL(dynload::cuptiGetTimestamp(&start_ns_)); #endif // PADDLE_WITH_CUPTI enabled_ = true; } + void Reset() { +#ifdef PADDLE_WITH_CUPTI + CUPTI_CALL( + dynload::cuptiActivityFlushAll(CUPTI_ACTIVITY_FLAG_FLUSH_FORCED)); +#endif + std::lock_guard l(trace_mu_); + kernel_records_.clear(); + mem_records_.clear(); + correlations_.clear(); + for (auto &tmp : correlations_pairs) tmp.clear(); + for (auto &tmp : cpu_records_) tmp.clear(); + } + + void GenEventKernelCudaElapsedTime() { +#ifdef PADDLE_WITH_CUPTI + if (correlations_.empty()) + for (auto &tmp : correlations_pairs) + for (auto &pair : tmp) correlations_[pair.first] = pair.second; + for (const KernelRecord &r : kernel_records_) { + auto c = correlations_.find(r.correlation_id); + if (c != correlations_.end() && c->second != nullptr) { + Event *e = c->second; + e->AddCudaElapsedTime(r.start_ns, r.end_ns); + } + } + for (const auto &r : mem_records_) { + auto c = correlations_.find(r.correlation_id); + if (c != correlations_.end() && c->second != nullptr) { + Event *e = c->second; + e->AddCudaElapsedTime(r.start_ns, r.end_ns); + } + } +#endif + } + proto::Profile GenProfile(const std::string &profile_path) { + int miss = 0, find = 0; std::lock_guard l(trace_mu_); proto::Profile profile_pb; profile_pb.set_start_ns(start_ns_); profile_pb.set_end_ns(end_ns_); + if (correlations_.empty()) + for (auto &tmp : correlations_pairs) + for (auto &pair : tmp) correlations_[pair.first] = pair.second; for (const KernelRecord &r : kernel_records_) { auto *event = profile_pb.add_events(); event->set_type(proto::Event::GPUKernel); - if (correlations_.find(r.correlation_id) != correlations_.end()) { - event->set_name(correlations_.at(r.correlation_id)); + auto c = correlations_.find(r.correlation_id); + if (c != correlations_.end() && c->second != nullptr) { + event->set_name(c->second->name()); + event->set_detail_info(r.name); + find++; } else { + VLOG(10) << "Missing Kernel Event: " + r.name; + miss++; event->set_name(r.name); } event->set_start_ns(r.start_ns); @@ -289,31 +436,41 @@ class DeviceTracerImpl : public DeviceTracer { event->set_sub_device_id(r.stream_id); event->set_device_id(r.device_id); } - - for (const CPURecord &r : cpu_records_) { - auto *event = profile_pb.add_events(); - event->set_type(proto::Event::CPU); - event->set_name(r.name); - event->set_start_ns(r.start_ns); - event->set_end_ns(r.end_ns); - event->set_sub_device_id(r.thread_id); - event->set_device_id(r.device_id); - } + VLOG(1) << "KernelRecord event miss: " << miss << " find: " << find; + for (auto &tmp : cpu_records_) + for (const CPURecord &r : tmp) { + auto *event = profile_pb.add_events(); + event->set_type(proto::Event::CPU); + event->set_name(r.name); + event->set_start_ns(r.start_ns); + event->set_end_ns(r.end_ns); + event->set_sub_device_id(r.thread_id); + event->set_device_id(r.device_id); + } + miss = find = 0; for (const MemRecord &r : mem_records_) { auto *event = profile_pb.add_events(); event->set_type(proto::Event::GPUKernel); - event->set_name(r.name); + auto c = correlations_.find(r.correlation_id); + if (c != correlations_.end() && c->second != nullptr) { + event->set_name(c->second->name()); + event->set_detail_info(r.name); + find++; + } else { + miss++; + event->set_name(r.name); + } event->set_start_ns(r.start_ns); event->set_end_ns(r.end_ns); event->set_sub_device_id(r.stream_id); event->set_device_id(r.device_id); event->mutable_memcopy()->set_bytes(r.bytes); } + VLOG(1) << "MemRecord event miss: " << miss << " find: " << find; std::ofstream profile_f; - profile_f.open(profile_path, std::ios::out | std::ios::trunc); - std::string profile_str; - profile_pb.SerializeToString(&profile_str); - profile_f << profile_str; + profile_f.open(profile_path, + std::ios::out | std::ios::trunc | std::ios::binary); + profile_pb.SerializeToOstream(&profile_f); profile_f.close(); return profile_pb; } @@ -321,12 +478,13 @@ class DeviceTracerImpl : public DeviceTracer { void Disable() { #ifdef PADDLE_WITH_CUPTI // flush might cause additional calls to DeviceTracker. - dynload::cuptiActivityFlushAll(CUPTI_ACTIVITY_FLAG_FLUSH_FORCED); + CUPTI_CALL( + dynload::cuptiActivityFlushAll(CUPTI_ACTIVITY_FLAG_FLUSH_FORCED)); #endif // PADDLE_WITH_CUPTI std::lock_guard l(trace_mu_); #ifdef PADDLE_WITH_CUPTI DisableActivity(); - dynload::cuptiUnsubscribe(subscriber_); + CUPTI_CALL(dynload::cuptiUnsubscribe(subscriber_)); CUPTI_CALL(dynload::cuptiGetTimestamp(&end_ns_)); #endif // PADDLE_WITH_CUPTI enabled_ = false; @@ -337,18 +495,10 @@ class DeviceTracerImpl : public DeviceTracer { static void CUPTIAPI ApiCallback(void *userdata, CUpti_CallbackDomain domain, CUpti_CallbackId cbid, const void *cbdata) { auto *cbInfo = reinterpret_cast(cbdata); - DeviceTracer *tracer = reinterpret_cast(userdata); - - if ((domain == CUPTI_CB_DOMAIN_DRIVER_API) && - (cbid == CUPTI_DRIVER_TRACE_CBID_cuLaunchKernel)) { - if (cbInfo->callbackSite == CUPTI_API_ENTER) { - const std::string anno = !annotation_stack.empty() - ? annotation_stack.back() - : cbInfo->symbolName; - tracer->AddAnnotation(cbInfo->correlationId, anno); - } - } else { - VLOG(1) << "Unhandled API Callback for " << domain << " " << cbid; + DeviceTracerImpl *tracer = reinterpret_cast(userdata); + if (cbInfo->callbackSite == CUPTI_API_ENTER) { + Event *event = CurAnnotation(); + tracer->AddAnnotation(cbInfo->correlationId, event); } } CUpti_SubscriberHandle subscriber_; @@ -357,10 +507,12 @@ class DeviceTracerImpl : public DeviceTracer { bool enabled_; uint64_t start_ns_; uint64_t end_ns_; - std::vector kernel_records_; - std::vector mem_records_; - std::vector cpu_records_; - std::unordered_map correlations_; + std::forward_list kernel_records_; + std::forward_list mem_records_; + std::forward_list> cpu_records_; + std::forward_list>> + correlations_pairs; + std::unordered_map correlations_; }; void CreateTracer(DeviceTracer **t) { *t = new DeviceTracerImpl(); } @@ -370,21 +522,106 @@ DeviceTracer *GetDeviceTracer() { return tracer; } -void SetCurAnnotation(const std::string &anno) { - annotation_stack.push_back(anno); -} +void SetCurAnnotation(Event *event) { annotation_stack.push_back(event); } void ClearCurAnnotation() { annotation_stack.pop_back(); } -std::string CurAnnotation() { - if (annotation_stack.empty()) return ""; +Event *CurAnnotation() { + if (annotation_stack.empty()) return nullptr; return annotation_stack.back(); } +std::string CurAnnotationName() { + if (annotation_stack.empty()) return ""; + return annotation_stack.back()->name(); +} void SetCurBlock(int block_id) { block_id_stack.push_back(block_id); } void ClearCurBlock() { block_id_stack.pop_back(); } int BlockDepth() { return block_id_stack.size(); } + +uint32_t GetCurSystemThreadId() { + std::stringstream ss; + ss << std::this_thread::get_id(); + uint32_t id = static_cast(std::stoull(ss.str())); + return id; +} + +void RecoreCurThreadId(int32_t id) { + auto gid = GetCurSystemThreadId(); + VLOG(1) << "RecoreCurThreadId: " << gid << " -> " << id; + system_thread_id_map[gid] = id; +} + +int32_t GetThreadIdFromSystemThreadId(uint32_t id) { + auto it = system_thread_id_map.find(id); + if (it != system_thread_id_map.end()) return it->second; + // return origin id if no event is recorded in this thread. + return static_cast(id); +} + +#ifdef PADDLE_WITH_CUPTI +namespace { + +void initCuptiCbidStr() { + static bool called = false; + if (called) return; + called = true; +#define REGISTER_RUNTIME_CBID_STR(cbid) \ + runtime_cbid_str[CUPTI_RUNTIME_TRACE_CBID_##cbid] = #cbid + + REGISTER_RUNTIME_CBID_STR(cudaBindTexture_v3020); + REGISTER_RUNTIME_CBID_STR(cudaConfigureCall_v3020); + REGISTER_RUNTIME_CBID_STR(cudaDeviceGetAttribute_v5000); + REGISTER_RUNTIME_CBID_STR(cudaDeviceGetStreamPriorityRange_v5050); + REGISTER_RUNTIME_CBID_STR(cudaDeviceSynchronize_v3020); + REGISTER_RUNTIME_CBID_STR(cudaDriverGetVersion_v3020); + REGISTER_RUNTIME_CBID_STR(cudaEventCreateWithFlags_v3020); + REGISTER_RUNTIME_CBID_STR(cudaEventDestroy_v3020); + REGISTER_RUNTIME_CBID_STR(cudaEventDestroy_v3020); + REGISTER_RUNTIME_CBID_STR(cudaEventQuery_v3020); + REGISTER_RUNTIME_CBID_STR(cudaEventRecord_v3020); + REGISTER_RUNTIME_CBID_STR(cudaFreeHost_v3020); + REGISTER_RUNTIME_CBID_STR(cudaFree_v3020); + REGISTER_RUNTIME_CBID_STR(cudaFuncGetAttributes_v3020); + REGISTER_RUNTIME_CBID_STR(cudaGetDeviceCount_v3020); + REGISTER_RUNTIME_CBID_STR(cudaGetDeviceProperties_v3020); + REGISTER_RUNTIME_CBID_STR(cudaGetDevice_v3020); + REGISTER_RUNTIME_CBID_STR(cudaGetErrorString_v3020); + REGISTER_RUNTIME_CBID_STR(cudaGetLastError_v3020); + REGISTER_RUNTIME_CBID_STR(cudaHostAlloc_v3020); + REGISTER_RUNTIME_CBID_STR(cudaHostGetDevicePointer_v3020); + REGISTER_RUNTIME_CBID_STR(cudaLaunchKernel_v7000); + REGISTER_RUNTIME_CBID_STR(cudaMallocHost_v3020); + REGISTER_RUNTIME_CBID_STR(cudaMalloc_v3020); + REGISTER_RUNTIME_CBID_STR(cudaMemcpyAsync_v3020); + REGISTER_RUNTIME_CBID_STR(cudaMemcpy_v3020); + REGISTER_RUNTIME_CBID_STR(cudaMemsetAsync_v3020); + REGISTER_RUNTIME_CBID_STR(cudaMemset_v3020); + REGISTER_RUNTIME_CBID_STR( + cudaOccupancyMaxActiveBlocksPerMultiprocessorWithFlags_v7000); + REGISTER_RUNTIME_CBID_STR(cudaPeekAtLastError_v3020); + REGISTER_RUNTIME_CBID_STR(cudaRuntimeGetVersion_v3020); + REGISTER_RUNTIME_CBID_STR(cudaSetDevice_v3020); + REGISTER_RUNTIME_CBID_STR(cudaStreamCreate_v3020); + REGISTER_RUNTIME_CBID_STR(cudaStreamCreateWithFlags_v5000); + REGISTER_RUNTIME_CBID_STR(cudaStreamCreateWithPriority_v5050); + REGISTER_RUNTIME_CBID_STR(cudaStreamDestroy_v5050); + REGISTER_RUNTIME_CBID_STR(cudaStreamSynchronize_v3020); + REGISTER_RUNTIME_CBID_STR(cudaStreamWaitEvent_v3020); + REGISTER_RUNTIME_CBID_STR(cudaUnbindTexture_v3020); + REGISTER_RUNTIME_CBID_STR(cudaSetupArgument_v3020); + REGISTER_RUNTIME_CBID_STR(cudaLaunch_v3020); +#if CUDA_VERSION >= 9000 + REGISTER_RUNTIME_CBID_STR(cudaLaunchCooperativeKernel_v9000); + REGISTER_RUNTIME_CBID_STR(cudaLaunchCooperativeKernelMultiDevice_v9000); +#endif + +#undef REGISTER_RUNTIME_CBID_STR +} +} // namespace +#endif // PADDLE_WITH_CUPTI + } // namespace platform } // namespace paddle diff --git a/paddle/fluid/platform/device_tracer.h b/paddle/fluid/platform/device_tracer.h index bf0786be2d0fafbf4b610d16ef587ac219399203..6ee2c36146215e278c87d22dd7b8fd5cb7e33865 100644 --- a/paddle/fluid/platform/device_tracer.h +++ b/paddle/fluid/platform/device_tracer.h @@ -32,6 +32,8 @@ inline uint64_t PosixInNsec() { return 1000 * (static_cast(tv.tv_sec) * 1000000 + tv.tv_usec); } +class Event; + // DeviceTracer performs the following tasks: // 1. Register cuda callbacks for various events: kernel, memcpy, etc. // 2. Collect cuda statistics: start/end ts, memory, etc. @@ -68,11 +70,13 @@ class DeviceTracer { virtual void Enable() = 0; // Needs to be called once after use. virtual void Disable() = 0; + // Needs to be called once before reuse. + virtual void Reset() = 0; // Add a pair to correlate internal cuda id with high level - // annotation (string). So cuda statistics can be represented by + // annotation event(with string). So cuda statistics can be represented by // human-readable annotations. - virtual void AddAnnotation(uint64_t id, const std::string& anno) = 0; + virtual void AddAnnotation(uint32_t id, Event* event) = 0; virtual void AddMemRecords(const std::string& name, uint64_t start_ns, uint64_t end_ns, int64_t device_id, @@ -92,6 +96,9 @@ class DeviceTracer { // Generate a proto after done (Disabled). virtual proto::Profile GenProfile(const std::string& profile_path) = 0; + // generate kernel elapsed time into Event + virtual void GenEventKernelCudaElapsedTime() = 0; + virtual bool IsEnabled() = 0; }; @@ -99,14 +106,19 @@ class DeviceTracer { DeviceTracer* GetDeviceTracer(); // Set a name for the cuda kernel operation being launched by the thread. -void SetCurAnnotation(const std::string& anno); +void SetCurAnnotation(Event* event); // Clear the name after the operation is done. void ClearCurAnnotation(); // Current name of the operation being run in the thread. -std::string CurAnnotation(); +std::string CurAnnotationName(); +Event* CurAnnotation(); void SetCurBlock(int block_id); void ClearCurBlock(); int BlockDepth(); + +// Set current thread id, so we can map the system thread id to thread id. +void RecoreCurThreadId(int32_t id); +int32_t GetThreadIdFromSystemThreadId(uint32_t id); } // namespace platform } // namespace paddle diff --git a/paddle/fluid/platform/dynload/mklml.h b/paddle/fluid/platform/dynload/mklml.h index d0619293acf2d2df0d925e969bdeb8e45cda6e2b..a260cda49138580b209e647af459e9392d9f18f1 100644 --- a/paddle/fluid/platform/dynload/mklml.h +++ b/paddle/fluid/platform/dynload/mklml.h @@ -70,6 +70,8 @@ extern void* mklml_dso_handle; __macro(cblas_ddot); \ __macro(cblas_sasum); \ __macro(cblas_dasum); \ + __macro(cblas_isamax); \ + __macro(cblas_idamax); \ __macro(cblas_sscal); \ __macro(cblas_dscal); \ __macro(vsAdd); \ diff --git a/paddle/fluid/platform/enforce.h b/paddle/fluid/platform/enforce.h index 15413785bab3c0fd77244141e8f1840ca0cc1356..54ad18a8e4abbddc0414ed32a719d46002e99188 100644 --- a/paddle/fluid/platform/enforce.h +++ b/paddle/fluid/platform/enforce.h @@ -31,6 +31,8 @@ limitations under the License. */ #include #include #include +#include +#include #include "glog/logging.h" #include "paddle/fluid/platform/macros.h" @@ -71,9 +73,8 @@ struct EnforceNotMet : public std::exception { } } - template - EnforceNotMet(const char* f, int l, ARGS... args) { - Init(string::Sprintf(args...), f, l); + EnforceNotMet(const std::string& str, const char* f, int l) { + Init(str, f, l); } const char* what() const noexcept override { return err_str_.c_str(); } @@ -142,28 +143,23 @@ struct EOFException : public std::exception { inline bool is_error(bool stat) { return !stat; } -template -inline typename std::enable_if::type throw_on_error( - bool stat, const Args&... args) { +inline void throw_on_error(bool stat, const std::string& msg) { #ifndef REPLACE_ENFORCE_GLOG - throw std::runtime_error(string::Sprintf(args...)); + throw std::runtime_error(msg); #else - LOG(FATAL) << string::Sprintf(args...); + LOG(FATAL) << msg; #endif } #ifdef PADDLE_WITH_CUDA -inline bool is_error(cudaError_t e) { return UNLIKELY(e); } +inline bool is_error(cudaError_t e) { return e != cudaSuccess; } -template -inline typename std::enable_if::type throw_on_error( - cudaError_t e, const Args&... args) { +inline void throw_on_error(cudaError_t e, const std::string& msg) { #ifndef REPLACE_ENFORCE_GLOG - throw thrust::system_error(e, thrust::cuda_category(), - string::Sprintf(args...)); + throw thrust::system_error(e, thrust::cuda_category(), msg); #else - LOG(FATAL) << string::Sprintf(args...); + LOG(FATAL) << msg; #endif } @@ -171,14 +167,12 @@ inline bool is_error(curandStatus_t stat) { return stat != CURAND_STATUS_SUCCESS; } -template -inline typename std::enable_if::type throw_on_error( - curandStatus_t stat, const Args&... args) { +inline void throw_on_error(curandStatus_t stat, const std::string& msg) { #ifndef REPLACE_ENFORCE_GLOG throw thrust::system_error(cudaErrorLaunchFailure, thrust::cuda_category(), - string::Sprintf(args...)); + msg); #else - LOG(FATAL) << string::Sprintf(args...); + LOG(FATAL) << msg; #endif } @@ -186,14 +180,11 @@ inline bool is_error(cudnnStatus_t stat) { return stat != CUDNN_STATUS_SUCCESS; } -template -inline typename std::enable_if::type throw_on_error( - cudnnStatus_t stat, const Args&... args) { +inline void throw_on_error(cudnnStatus_t stat, const std::string& msg) { #ifndef REPLACE_ENFORCE_GLOG - throw std::runtime_error(platform::dynload::cudnnGetErrorString(stat) + - string::Sprintf(args...)); + throw std::runtime_error(platform::dynload::cudnnGetErrorString(stat) + msg); #else - LOG(FATAL) << string::Sprintf(args...); + LOG(FATAL) << platform::dynload::cudnnGetErrorString(stat) << msg; #endif } @@ -201,9 +192,7 @@ inline bool is_error(cublasStatus_t stat) { return stat != CUBLAS_STATUS_SUCCESS; } -template -inline typename std::enable_if::type throw_on_error( - cublasStatus_t stat, const Args&... args) { +inline void throw_on_error(cublasStatus_t stat, const std::string& msg) { std::string err; if (stat == CUBLAS_STATUS_NOT_INITIALIZED) { err = "CUBLAS: not initialized, "; @@ -225,87 +214,47 @@ inline typename std::enable_if::type throw_on_error( err = "CUBLAS: license error, "; } #ifndef REPLACE_ENFORCE_GLOG - throw std::runtime_error(err + string::Sprintf(args...)); + throw std::runtime_error(err + msg); #else - LOG(FATAL) << err << string::Sprintf(args...); + LOG(FATAL) << err << msg; #endif } #if !defined(__APPLE__) && !defined(_WIN32) -template -inline typename std::enable_if::type throw_on_error( - ncclResult_t stat, const Args&... args) { - if (stat == ncclSuccess) { - return; - } else { +inline bool is_error(ncclResult_t nccl_result) { + return nccl_result != ncclSuccess; +} + +inline void throw_on_error(ncclResult_t stat, const std::string& msg) { #ifndef REPLACE_ENFORCE_GLOG - throw std::runtime_error(platform::dynload::ncclGetErrorString(stat) + - string::Sprintf(args...)); + throw std::runtime_error(platform::dynload::ncclGetErrorString(stat) + msg); #else - LOG(FATAL) << platform::dynload::ncclGetErrorString(stat) - << string::Sprintf(args...); + LOG(FATAL) << platform::dynload::ncclGetErrorString(stat) << msg; #endif - } } #endif // __APPLE__ and windows #endif // PADDLE_WITH_CUDA -template -inline void throw_on_error(T e) { - throw_on_error(e, ""); -} - -#define PADDLE_THROW(...) \ - throw ::paddle::platform::EnforceNotMet(__FILE__, __LINE__, __VA_ARGS__) - -#define __PADDLE_THROW_ERROR_I(_, _9, _8, _7, _6, _5, _4, _3, _2, X_, ...) X_; - -#define __THROW_ON_ERROR_ONE_ARG(COND, ARG) \ - ::paddle::platform::throw_on_error(COND, ::paddle::string::Sprintf(ARG)); - -#ifdef _WIN32 -#define __PADDLE_THROW_ON_ERROR(COND, ...) \ - __THROW_ON_ERROR_ONE_ARG(COND, __VA_ARGS__) -#else // _WIN32 -#define __PADDLE_THROW_ON_ERROR(COND, ...) \ - __PADDLE_THROW_ERROR_I( \ - __VA_ARGS__, ::paddle::platform::throw_on_error(COND, __VA_ARGS__), \ - ::paddle::platform::throw_on_error(COND, __VA_ARGS__), \ - ::paddle::platform::throw_on_error(COND, __VA_ARGS__), \ - ::paddle::platform::throw_on_error(COND, __VA_ARGS__), \ - ::paddle::platform::throw_on_error(COND, __VA_ARGS__), \ - ::paddle::platform::throw_on_error(COND, __VA_ARGS__), \ - ::paddle::platform::throw_on_error(COND, __VA_ARGS__), \ - ::paddle::platform::throw_on_error(COND, __VA_ARGS__), \ - __THROW_ON_ERROR_ONE_ARG(COND, __VA_ARGS__)) -#endif // _WIN32 - -#define __PADDLE_UNARY_COMPARE(COND, ...) \ - do { \ - auto __cond = COND; \ - if (UNLIKELY(::paddle::platform::is_error(__cond))) { \ - __PADDLE_THROW_ON_ERROR(__cond, __VA_ARGS__); \ - } \ +#define PADDLE_THROW(...) \ + do { \ + throw ::paddle::platform::EnforceNotMet( \ + ::paddle::string::Sprintf(__VA_ARGS__), __FILE__, __LINE__); \ } while (0) -#ifndef REPLACE_ENFORCE_GLOG -#define __PADDLE_ENFORCE_I(COND, ...) \ - do { \ - try { \ - __PADDLE_UNARY_COMPARE(COND, __VA_ARGS__); \ - } catch (...) { \ - throw ::paddle::platform::EnforceNotMet(std::current_exception(), \ - __FILE__, __LINE__); \ - } \ +#define PADDLE_ENFORCE(COND, ...) \ + do { \ + auto __cond__ = (COND); \ + if (UNLIKELY(::paddle::platform::is_error(__cond__))) { \ + try { \ + ::paddle::platform::throw_on_error( \ + __cond__, ::paddle::string::Sprintf(__VA_ARGS__)); \ + } catch (...) { \ + throw ::paddle::platform::EnforceNotMet(std::current_exception(), \ + __FILE__, __LINE__); \ + } \ + } \ } while (0) -#else -#define __PADDLE_ENFORCE_I(COND, ...) __PADDLE_UNARY_COMPARE(COND, __VA_ARGS__); -#endif // REPLACE_ENFORCE_GLOG - -#define __PADDLE_ENFORCE(__args) __PADDLE_ENFORCE_I __args -#define PADDLE_ENFORCE(...) __PADDLE_ENFORCE((__VA_ARGS__)) - #define PADDLE_THROW_EOF() \ do { \ throw ::paddle::platform::EOFException("There is no next data.", __FILE__, \ @@ -325,23 +274,71 @@ inline void throw_on_error(T e) { * extra messages is also supported, for example: * PADDLE_ENFORCE(a, b, "some simple enforce failed between %d numbers", 2) */ -#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__)); \ - } \ +#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, ...) \ +namespace details { +template +inline constexpr bool IsArithmetic() { + return std::is_arithmetic::value; +} + +template +struct TypeConverterImpl { + using Type1 = typename std::common_type::type; + using Type2 = Type1; +}; + +template +struct TypeConverterImpl { + using Type1 = T1; + using Type2 = T2; +}; + +template +struct TypeConverter { + private: + static constexpr bool kIsArithmetic = + IsArithmetic() && IsArithmetic(); + + public: + using Type1 = typename TypeConverterImpl::Type1; + using Type2 = typename TypeConverterImpl::Type2; +}; + +template +using CommonType1 = typename std::add_lvalue_reference< + typename std::add_const::Type1>::type>::type; + +template +using CommonType2 = typename std::add_lvalue_reference< + typename std::add_const::Type2>::type>::type; +} // namespace details + +#define __PADDLE_BINARY_COMPARE(__VAL1, __VAL2, __CMP, __INV_CMP, ...) \ do { \ - if (UNLIKELY(!((__VAL0)__CMP(__VAL1)))) { \ + auto __val1 = (__VAL1); \ + auto __val2 = (__VAL2); \ + using __TYPE1__ = decltype(__val1); \ + using __TYPE2__ = decltype(__val2); \ + using __COMMON_TYPE1__ = \ + ::paddle::platform::details::CommonType1<__TYPE1__, __TYPE2__>; \ + using __COMMON_TYPE2__ = \ + ::paddle::platform::details::CommonType2<__TYPE1__, __TYPE2__>; \ + bool __is_not_error = (static_cast<__COMMON_TYPE1__>(__val1))__CMP( \ + static_cast<__COMMON_TYPE2__>(__val2)); \ + if (UNLIKELY(!__is_not_error)) { \ PADDLE_THROW("Enforce failed. Expected %s " #__CMP \ " %s, but received %s:%s " #__INV_CMP " %s:%s.\n%s", \ - #__VAL0, #__VAL1, #__VAL0, \ - paddle::string::to_string(__VAL0), #__VAL1, \ - paddle::string::to_string(__VAL1), \ - paddle::string::Sprintf("" __VA_ARGS__)); \ + #__VAL1, #__VAL2, #__VAL1, \ + ::paddle::string::to_string(__val1), #__VAL2, \ + ::paddle::string::to_string(__val2), \ + ::paddle::string::Sprintf(__VA_ARGS__)); \ } \ } while (0) diff --git a/paddle/fluid/platform/enforce_test.cc b/paddle/fluid/platform/enforce_test.cc index 1091badae54a809c4a9da6d0398bcbb538420af0..adcc95367f11dfa2722226e5a0386bedfa6e746e 100644 --- a/paddle/fluid/platform/enforce_test.cc +++ b/paddle/fluid/platform/enforce_test.cc @@ -118,59 +118,58 @@ TEST(ENFORCE_GT, OK) { PADDLE_ENFORCE_GT(2, 1); } TEST(ENFORCE_GT, FAIL) { bool caught_exception = false; try { - PADDLE_ENFORCE_GT(1, 2UL); + PADDLE_ENFORCE_GT(1, 2); } catch (paddle::platform::EnforceNotMet error) { caught_exception = true; - EXPECT_TRUE(HasPrefix( - StringPiece(error.what()), - "Enforce failed. Expected 1 > 2UL, but received 1:1 <= 2UL:2.")); + EXPECT_TRUE( + HasPrefix(StringPiece(error.what()), + "Enforce failed. Expected 1 > 2, but received 1:1 <= 2:2.")); } EXPECT_TRUE(caught_exception); } TEST(ENFORCE_GE, OK) { - PADDLE_ENFORCE_GE(2, 2UL); - PADDLE_ENFORCE_GE(3, 2UL); + PADDLE_ENFORCE_GE(2, 2); PADDLE_ENFORCE_GE(3, 2); - PADDLE_ENFORCE_GE(3.21, 2UL); + PADDLE_ENFORCE_GE(3.21, 2.0); } TEST(ENFORCE_GE, FAIL) { bool caught_exception = false; try { - PADDLE_ENFORCE_GE(1, 2UL); + PADDLE_ENFORCE_GE(1, 2); } catch (paddle::platform::EnforceNotMet error) { caught_exception = true; - EXPECT_TRUE(HasPrefix( - StringPiece(error.what()), - "Enforce failed. Expected 1 >= 2UL, but received 1:1 < 2UL:2.")); + EXPECT_TRUE( + HasPrefix(StringPiece(error.what()), + "Enforce failed. Expected 1 >= 2, but received 1:1 < 2:2.")); } EXPECT_TRUE(caught_exception); } TEST(ENFORCE_LE, OK) { PADDLE_ENFORCE_LE(1, 1); - PADDLE_ENFORCE_LE(1, 1UL); - PADDLE_ENFORCE_LE(2, 3UL); - PADDLE_ENFORCE_LE(2UL, 3); - PADDLE_ENFORCE_LE(2UL, 3.2); + PADDLE_ENFORCE_LE(1UL, 1UL); + PADDLE_ENFORCE_LE(2, 3); + PADDLE_ENFORCE_LE(2UL, 3UL); + PADDLE_ENFORCE_LE(2.0, 3.2); } TEST(ENFORCE_LE, FAIL) { bool caught_exception = false; try { - PADDLE_ENFORCE_GT(1, 2UL); + PADDLE_ENFORCE_GT(1, 2); } catch (paddle::platform::EnforceNotMet error) { caught_exception = true; - EXPECT_TRUE(HasPrefix( - StringPiece(error.what()), - "Enforce failed. Expected 1 > 2UL, but received 1:1 <= 2UL:2.")); + EXPECT_TRUE( + HasPrefix(StringPiece(error.what()), + "Enforce failed. Expected 1 > 2, but received 1:1 <= 2:2.")); } EXPECT_TRUE(caught_exception); } TEST(ENFORCE_LT, OK) { PADDLE_ENFORCE_LT(3, 10); - PADDLE_ENFORCE_LT(2, 3UL); - PADDLE_ENFORCE_LT(2UL, 3); + PADDLE_ENFORCE_LT(2UL, 3UL); + PADDLE_ENFORCE_LT(2, 3); } TEST(ENFORCE_LT, FAIL) { bool caught_exception = false; @@ -235,7 +234,13 @@ TEST(ENFORCE_USER_DEFINED_CLASS, EQ) { TEST(ENFORCE_USER_DEFINED_CLASS, NE) { Dims a{{1, 2, 3, 4}}, b{{5, 6, 7, 8}}; - ASSERT_THROW(PADDLE_ENFORCE_EQ(a, b), paddle::platform::EnforceNotMet); + bool caught_exception = false; + try { + PADDLE_ENFORCE_EQ(a, b); + } catch (paddle::platform::EnforceNotMet&) { + caught_exception = true; + } + EXPECT_TRUE(caught_exception); } TEST(EOF_EXCEPTION, THROW_EOF) { diff --git a/paddle/fluid/platform/init.cc b/paddle/fluid/platform/init.cc index ac86b38a61c9d8e3e946d9fb3f46d8feba7c034d..4dcf7e79043af008cb2067d90d12d629c5c2d0d9 100644 --- a/paddle/fluid/platform/init.cc +++ b/paddle/fluid/platform/init.cc @@ -22,6 +22,7 @@ limitations under the License. */ #include "paddle/fluid/string/split.h" #ifdef PADDLE_WITH_CUDA #include "paddle/fluid/platform/cuda_device_guard.h" +#include "paddle/fluid/platform/dynload/cupti.h" #endif #include "paddle/fluid/platform/device_context.h" #include "paddle/fluid/platform/init.h" @@ -30,6 +31,9 @@ limitations under the License. */ DEFINE_int32(paddle_num_threads, 1, "Number of threads for each paddle instance."); +DEFINE_int32(multiple_of_cupti_buffer_size, 1, + "Multiple of the CUPTI device buffer size. If the timestamps have " + "been dropped when you are profiling, try increasing this value."); namespace paddle { namespace framework { @@ -78,7 +82,32 @@ void InitP2P(std::vector devices) { #endif } +void InitCupti() { +#ifdef PADDLE_WITH_CUPTI + if (FLAGS_multiple_of_cupti_buffer_size == 1) return; + size_t attrValue = 0, attrValueSize = sizeof(size_t); +#define MULTIPLY_ATTR_VALUE(attr) \ + { \ + PADDLE_ENFORCE(!platform::dynload::cuptiActivityGetAttribute( \ + attr, &attrValueSize, &attrValue)); \ + attrValue *= FLAGS_multiple_of_cupti_buffer_size; \ + LOG(WARNING) << "Set " #attr " " << attrValue << " byte"; \ + PADDLE_ENFORCE(!platform::dynload::cuptiActivitySetAttribute( \ + attr, &attrValueSize, &attrValue)); \ + } + MULTIPLY_ATTR_VALUE(CUPTI_ACTIVITY_ATTR_DEVICE_BUFFER_SIZE); + MULTIPLY_ATTR_VALUE(CUPTI_ACTIVITY_ATTR_DEVICE_BUFFER_SIZE_CDP); +#if CUDA_VERSION >= 9000 + MULTIPLY_ATTR_VALUE(CUPTI_ACTIVITY_ATTR_PROFILING_SEMAPHORE_POOL_SIZE); +#endif +#undef MULTIPLY_ATTR_VALUE +#endif +} + void InitDevices(bool init_p2p) { + // CUPTI attribute should be set before any CUDA context is created (see CUPTI + // documentation about CUpti_ActivityAttribute). + InitCupti(); /*Init all available devices by default */ std::vector devices; #ifdef PADDLE_WITH_CUDA diff --git a/paddle/fluid/platform/mkldnn_reuse.h b/paddle/fluid/platform/mkldnn_reuse.h index 269280d604a13a62046fb7811d34b7c69b61b50f..908499e0d8dc679a714a332c8dfe5f16bfbdcd3d 100644 --- a/paddle/fluid/platform/mkldnn_reuse.h +++ b/paddle/fluid/platform/mkldnn_reuse.h @@ -548,9 +548,8 @@ class ConvMKLDNNTemplateHandler : public MKLDNNHandler { PADDLE_ENFORCE((conv_p != nullptr) || (is_reusing_ == false), "Fail to find convolution primitive in device context"); if (conv_p == nullptr) { - conv_p = std::make_shared(*conv_pd_, *(src_memory_p), - *(weights_memory_p.get()), - *(dst_memory_p.get())); + conv_p = std::make_shared(*conv_pd_, *src_memory_p, + *weights_memory_p, *dst_memory_p); dev_ctx_.SetBlob(prim_key, conv_p); } else { @@ -570,9 +569,9 @@ class ConvMKLDNNTemplateHandler : public MKLDNNHandler { PADDLE_ENFORCE((conv_p != nullptr) || (is_reusing_ == false), "Fail to find convolution primitive in device context"); if (conv_p == nullptr) { - conv_p = std::make_shared( - *conv_pd_, *(src_memory_p), *(weights_memory_p.get()), - *(bias_memory_p.get()), *(dst_memory_p.get())); + conv_p = std::make_shared(*conv_pd_, *src_memory_p, + *weights_memory_p, *bias_memory_p, + *dst_memory_p); dev_ctx_.SetBlob(prim_key, conv_p); } else { diff --git a/paddle/fluid/platform/nccl_helper.h b/paddle/fluid/platform/nccl_helper.h index 8df8e32098697540f02d488c873f5ae7fb29828e..6ae21ee8294bedc388f837aad3e20a2b9aca98a2 100644 --- a/paddle/fluid/platform/nccl_helper.h +++ b/paddle/fluid/platform/nccl_helper.h @@ -64,7 +64,7 @@ class NCCLGroupGuard { } inline ~NCCLGroupGuard() { - CHECK_EQ(dynload::ncclGroupEnd(), ncclSuccess); + PADDLE_ENFORCE(dynload::ncclGroupEnd()); NCCLMutex().unlock(); } }; diff --git a/paddle/fluid/platform/ngraph_helper.h b/paddle/fluid/platform/ngraph_helper.h index c5b65d6636945b85603c07aeae0290ef9cadb396..e74f57a79a66ea8fe8c9b972a9a2ec9d722731eb 100644 --- a/paddle/fluid/platform/ngraph_helper.h +++ b/paddle/fluid/platform/ngraph_helper.h @@ -23,6 +23,33 @@ limitations under the License. */ namespace paddle { namespace platform { +std::shared_ptr Nhwc2Nchw(std::shared_ptr in) { + auto in_shape = in->get_shape(); + in_shape[0] = in->get_shape()[0]; + in_shape[1] = in->get_shape()[3]; + in_shape[2] = in->get_shape()[1]; + in_shape[3] = in->get_shape()[2]; + ngraph::AxisVector axis_vec = {0, 3, 1, 2}; + return std::make_shared(in, axis_vec, in_shape); +} + +std::shared_ptr Nchw2Nhwc(std::shared_ptr in) { + auto in_shape = in->get_shape(); + in_shape[0] = in->get_shape()[0]; + in_shape[1] = in->get_shape()[2]; + in_shape[2] = in->get_shape()[3]; + in_shape[3] = in->get_shape()[1]; + ngraph::AxisVector axis_vec = {0, 2, 3, 1}; + return std::make_shared(in, axis_vec, in_shape); +} + +ngraph::Shape FlattenTo1d(ngraph::Shape sh, int num) { + auto x1 = std::accumulate(std::begin(sh), std::end(sh) + num, 1, + std::multiplies()); + size_t x1_l = (size_t)x1; + return ngraph::Shape{x1_l}; +} + ngraph::Shape FlattenTo2d(ngraph::Shape sh, int num) { auto x1 = std::accumulate(std::begin(sh), std::begin(sh) + num, 1, std::multiplies()); @@ -43,13 +70,14 @@ std::shared_ptr NgReshaper(std::shared_ptr input, std::shared_ptr GetNode( const std::shared_ptr& op, - const std::string prm, const paddle::framework::VariableNameMap& var_map, + const std::string name, const paddle::framework::VariableNameMap& var_map, std::shared_ptr< std::unordered_map>> ngb_node_map) { - auto& var_names = var_map.at(prm); + auto& var_names = var_map.at(name); PADDLE_ENFORCE_EQ(var_names.size(), 1, - "op %s prm %s expects one associated var", op->Type(), prm); + "op %s name %s expects one associated var", op->Type(), + name); if (ngb_node_map->find(var_names[0]) != ngb_node_map->end()) { return (*ngb_node_map)[var_names[0]]; } else { @@ -59,43 +87,53 @@ std::shared_ptr GetNode( std::shared_ptr GetInputNode( const std::shared_ptr& op, - const std::string prm, + const std::string name, std::shared_ptr< std::unordered_map>> ngb_node_map) { - return GetNode(op, prm, op->Inputs(), ngb_node_map); + return GetNode(op, name, op->Inputs(), ngb_node_map); } std::shared_ptr GetOutputNode( const std::shared_ptr& op, - const std::string prm, + const std::string name, std::shared_ptr< std::unordered_map>> ngb_node_map) { - return GetNode(op, prm, op->Outputs(), ngb_node_map); + return GetNode(op, name, op->Outputs(), ngb_node_map); } void SetOutputNode( const std::shared_ptr& op, - const std::string prm, std::shared_ptr node, + const std::string name, std::shared_ptr node, std::shared_ptr< std::unordered_map>> ngb_node_map) { - auto& var_names = op->Outputs().at(prm); + auto& var_names = op->Outputs().at(name); if (var_names.size() == 1) { + /* */ + auto dummy_out = GetOutputNode(op, name, ngb_node_map); + if (dummy_out && dummy_out->get_shape() != node->get_shape()) { + node = NgReshaper(node, dummy_out->get_shape()); + } + if (dummy_out && + dummy_out->get_element_type() != node->get_element_type()) { + node = std::make_shared( + node, dummy_out->get_element_type()); + } (*ngb_node_map)[var_names[0]] = node; } else if (var_names.size() == 0) { (*ngb_node_map)[""] = node; } else { - PADDLE_THROW("prm %s has more than 1 var_names.", prm); + PADDLE_THROW("name %s has more than 1 var_names.", name); } } bool HasOutput(const std::shared_ptr& op, - const std::string prm) { + const std::string name) { auto& outputs = op->Outputs(); - if (outputs.find(prm) == outputs.end()) return false; - return outputs.at(prm).size() > 0; + if (outputs.find(name) == outputs.end()) return false; + return outputs.at(name).size() > 0; } inline void GetMidDims(const ngraph::Shape& x_shape, diff --git a/paddle/fluid/platform/place.cc b/paddle/fluid/platform/place.cc index 655ce8485d4584aa0955315b045da6bf541f7fe2..60b2d83f15746eab0a4d29c7965c064690b6d46d 100644 --- a/paddle/fluid/platform/place.cc +++ b/paddle/fluid/platform/place.cc @@ -14,6 +14,12 @@ limitations under the License. */ #include "paddle/fluid/platform/place.h" +DEFINE_bool(benchmark, false, + "Doing memory benchmark. It will make deleting scope synchronized, " + "and add some memory usage logs." + "Default cuda is asynchronous device, set to True will" + "force op run in synchronous mode."); + namespace paddle { namespace platform { diff --git a/paddle/fluid/platform/profiler.cc b/paddle/fluid/platform/profiler.cc index 85977366e61c676fc5d2d3c5d22dd2f606543684..9a285a6b533dcb48013e3b3e4d34dc27186173ac 100644 --- a/paddle/fluid/platform/profiler.cc +++ b/paddle/fluid/platform/profiler.cc @@ -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 "paddle/fluid/platform/profiler.h" + #include #include #include @@ -27,7 +29,6 @@ limitations under the License. */ #include "paddle/fluid/framework/block_desc.h" #include "paddle/fluid/platform/device_tracer.h" #include "paddle/fluid/platform/port.h" -#include "paddle/fluid/platform/profiler.h" #include "paddle/fluid/string/printf.h" DEFINE_bool(enable_rpc_profiler, false, "Enable rpc profiler or not."); @@ -66,12 +67,13 @@ struct EventList { ((kEventSize + kEventAlign - 1) / kEventAlign * kEventAlign); template - void Record(Args&&... args) { + Event* Record(Args&&... args) { if (event_blocks.empty() || event_blocks.front().size() == kNumBlock) { event_blocks.emplace_front(); event_blocks.front().reserve(kNumBlock); } event_blocks.front().emplace_back(std::forward(args)...); + return &event_blocks.front().back(); } std::vector Reduce() { @@ -98,21 +100,8 @@ inline uint64_t GetTimeInNsec() { .count(); } -Event::Event(EventType type, std::string name, uint32_t thread_id, - const DeviceContext* dev_ctx) - : type_(type), name_(name), thread_id_(thread_id), has_cuda_(false) { -#ifdef PADDLE_WITH_CUDA - has_cuda_ = dev_ctx ? platform::is_gpu_place(dev_ctx->GetPlace()) : false; - if (has_cuda_) { - auto* cuda_dev_ctx = static_cast(dev_ctx); - PADDLE_ENFORCE(cudaSetDevice( - boost::get(cuda_dev_ctx->GetPlace()).device)); - PADDLE_ENFORCE(cudaGetDevice(&device_)); - PADDLE_ENFORCE(cudaEventCreate(&event_)); - auto stream = cuda_dev_ctx->stream(); - PADDLE_ENFORCE(cudaEventRecord(event_, stream)); - } -#endif +Event::Event(EventType type, std::string name, uint32_t thread_id) + : type_(type), name_(name), thread_id_(thread_id) { cpu_ns_ = GetTimeInNsec(); } @@ -123,89 +112,70 @@ double Event::CpuElapsedMs(const Event& e) const { } double Event::CudaElapsedMs(const Event& e) const { -#ifdef PADDLE_WITH_CUDA - if (!has_cuda_) return 0.0; - PADDLE_ENFORCE(e.has_cuda() && has_cuda()); - PADDLE_ENFORCE(e.device() == device()); - PADDLE_ENFORCE(cudaEventSynchronize(event_)); - PADDLE_ENFORCE(cudaEventSynchronize(e.event())); - float ms; - PADDLE_ENFORCE(cudaEventElapsedTime(&ms, event_, e.event())); - return ms; +#ifdef PADDLE_WITH_CUPTI + return gpu_ns_ / 1000000.0; #else - PADDLE_THROW("CUDA is not enabled"); + LOG_FIRST_N(WARNING, 1) << "CUDA CUPTI is not enabled"; + return 0; #endif } -#ifdef PADDLE_WITH_CUDA -static void ForEachDevice(std::function func) { - auto original_device = GetCurrentDeviceId(); - int count = GetCUDADeviceCount(); - for (int i = 0; i < count; i++) { - SetDeviceId(i); - func(i); - } - SetDeviceId(original_device); -} -#endif - inline EventList& GetEventList() { if (!g_event_list) { std::lock_guard guard(g_all_event_lists_mutex); g_event_list = std::make_shared(); g_thread_id = g_next_thread_id++; g_all_event_lists.emplace_front(g_event_list); + RecoreCurThreadId(g_thread_id); } return *g_event_list; } -void Mark(const std::string& name, const DeviceContext* dev_ctx) { - GetEventList().Record(EventType::kMark, name, g_thread_id, dev_ctx); +void Mark(const std::string& name) { + GetEventList().Record(EventType::kMark, name, g_thread_id); } -void PushEvent(const std::string& name, const DeviceContext* dev_ctx) { - GetEventList().Record(EventType::kPushRange, name, g_thread_id, dev_ctx); +Event* PushEvent(const std::string& name) { + return GetEventList().Record(EventType::kPushRange, name, g_thread_id); } -void PopEvent(const std::string& name, const DeviceContext* dev_ctx) { - GetEventList().Record(EventType::kPopRange, name, g_thread_id, dev_ctx); +void PopEvent(const std::string& name) { + GetEventList().Record(EventType::kPopRange, name, g_thread_id); } -RecordEvent::RecordEvent(const std::string& name, const DeviceContext* dev_ctx) +RecordEvent::RecordEvent(const std::string& name) : is_enabled_(false), start_ns_(PosixInNsec()) { if (g_state == ProfilerState::kDisabled) return; - std::lock_guard l(profiler_mu); + // lock is not needed, the code below is thread-safe is_enabled_ = true; - dev_ctx_ = dev_ctx; name_ = name; - PushEvent(name_, dev_ctx_); + Event* e = PushEvent(name_); // Maybe need the same push/pop behavior. - SetCurAnnotation(name_); + SetCurAnnotation(e); } RecordEvent::~RecordEvent() { if (g_state == ProfilerState::kDisabled || !is_enabled_) return; - std::lock_guard l(profiler_mu); + // lock is not needed, the code below is thread-safe DeviceTracer* tracer = GetDeviceTracer(); if (tracer) { - tracer->AddCPURecords(CurAnnotation(), start_ns_, PosixInNsec(), + tracer->AddCPURecords(CurAnnotationName(), start_ns_, PosixInNsec(), BlockDepth(), g_thread_id); } ClearCurAnnotation(); - PopEvent(name_, dev_ctx_); + PopEvent(name_); } -RecordRPCEvent::RecordRPCEvent(const std::string& name, - const DeviceContext* dev_ctx) { +RecordRPCEvent::RecordRPCEvent(const std::string& name) { if (FLAGS_enable_rpc_profiler) { - event_.reset(new platform::RecordEvent(name, dev_ctx)); + event_.reset(new platform::RecordEvent(name)); } } RecordBlock::RecordBlock(int block_id) : is_enabled_(false), start_ns_(PosixInNsec()) { - std::lock_guard l(profiler_mu); + // lock is not needed, the code below is thread-safe if (g_state == ProfilerState::kDisabled) return; is_enabled_ = true; SetCurBlock(block_id); @@ -213,7 +183,7 @@ RecordBlock::RecordBlock(int block_id) } RecordBlock::~RecordBlock() { - std::lock_guard l(profiler_mu); + // lock is not needed, the code below is thread-safe if (g_state == ProfilerState::kDisabled || !is_enabled_) return; DeviceTracer* tracer = GetDeviceTracer(); if (tracer) { @@ -225,11 +195,21 @@ RecordBlock::~RecordBlock() { ClearCurBlock(); } +void SynchronizeAllDevice() { +#ifdef PADDLE_WITH_CUDA + int count = GetCUDADeviceCount(); + for (int i = 0; i < count; i++) { + SetDeviceId(i); + PADDLE_ENFORCE(cudaDeviceSynchronize()); + } +#endif +} + void EnableProfiler(ProfilerState state) { PADDLE_ENFORCE(state != ProfilerState::kDisabled, "Can't enable profiling, since the input state is ", "ProfilerState::kDisabled"); - + SynchronizeAllDevice(); std::lock_guard l(profiler_mu); if (state == g_state) { return; @@ -238,23 +218,20 @@ void EnableProfiler(ProfilerState state) { should_send_profile_state = true; GetDeviceTracer()->Enable(); #ifdef PADDLE_WITH_CUDA - if (g_state == ProfilerState::kCUDA) { + if (g_state == ProfilerState::kCUDA || g_state == ProfilerState::kAll || + g_state == ProfilerState::kCPU) { // Generate some dummy events first to reduce the startup overhead. - for (int i = 0; i < 5; i++) { - ForEachDevice([](int d) { - DeviceContext* dev_ctx = new CUDADeviceContext(CUDAPlace(d)); - Mark("_cuda_startup_", dev_ctx); - dev_ctx->Wait(); - delete dev_ctx; - }); - } + DummyKernelAndEvent(); + GetDeviceTracer()->Reset(); } #endif // Mark the profiling start. - Mark("_start_profiler_", nullptr); + Mark("_start_profiler_"); } void ResetProfiler() { + SynchronizeAllDevice(); + GetDeviceTracer()->Reset(); std::lock_guard guard(g_all_event_lists_mutex); for (auto it = g_all_event_lists.begin(); it != g_all_event_lists.end(); ++it) { @@ -277,9 +254,11 @@ struct EventItem { std::string name; int calls; double total_time; - double min_time; double max_time; double ave_time; + double min_time; + double cpu_time; + double gpu_time; float ratio; }; @@ -313,8 +292,12 @@ void PrintProfiler(const std::vector>& events_table, // Output events table std::cout.setf(std::ios::left); std::cout << std::setw(name_width) << "Event" << std::setw(data_width) - << "Calls" << std::setw(data_width) << "Total" - << std::setw(data_width) << "Min." << std::setw(data_width) + << "Calls" << std::setw(data_width) << "Total"; + if (g_state == ProfilerState::kAll) { + std::cout << std::setw(data_width * 2) << "CPU Time (Ratio)" + << std::setw(data_width * 2) << "GPU Time (Ratio)"; + } + std::cout << std::setw(data_width) << "Min." << std::setw(data_width) << "Max." << std::setw(data_width) << "Ave." << std::setw(data_width) << "Ratio." << std::endl; for (size_t i = 0; i < events_table.size(); ++i) { @@ -322,8 +305,18 @@ void PrintProfiler(const std::vector>& events_table, const EventItem& event_item = events_table[i][j]; std::cout << std::setw(name_width) << event_item.name << std::setw(data_width) << event_item.calls - << std::setw(data_width) << event_item.total_time - << std::setw(data_width) << event_item.min_time + << std::setw(data_width) << event_item.total_time; + if (g_state == ProfilerState::kAll) { + std::cout << std::setw(data_width * 2) + << string::Sprintf( + "%f (%f)", event_item.cpu_time, + (event_item.cpu_time / event_item.total_time)) + << std::setw(data_width * 2) + << string::Sprintf( + "%f (%f)", event_item.gpu_time, + (event_item.gpu_time / event_item.total_time)); + } + std::cout << std::setw(data_width) << event_item.min_time << std::setw(data_width) << event_item.max_time << std::setw(data_width) << event_item.ave_time << std::setw(data_width) << event_item.ratio << std::endl; @@ -372,6 +365,18 @@ void ParseEvents(const std::vector>& events, return a.ave_time > b.ave_time; }; break; + case EventSortingKey::kGPUTime: + sorted_domain = "average time"; + sorted_func = [](const EventItem& a, const EventItem& b) { + return a.gpu_time > b.gpu_time; + }; + break; + case EventSortingKey::kCPUTime: + sorted_domain = "average time"; + sorted_func = [](const EventItem& a, const EventItem& b) { + return a.cpu_time > b.cpu_time; + }; + break; default: sorted_domain = "event first end time"; } @@ -410,10 +415,17 @@ void ParseEvents(const std::vector>& events, } if (rit != pushed_events.rend()) { - double event_time = (g_state == ProfilerState::kCUDA || - g_state == ProfilerState::kAll) - ? rit->CudaElapsedMs((*analyze_events)[i][j]) - : rit->CpuElapsedMs((*analyze_events)[i][j]); + double event_time = 0; + double gpu_time = rit->CudaElapsedMs((*analyze_events)[i][j]); + double cpu_time = rit->CpuElapsedMs((*analyze_events)[i][j]); + if (g_state == ProfilerState::kCUDA) { + event_time = gpu_time; + } else if (g_state == ProfilerState::kCPU) { + event_time = cpu_time; + } else { + event_time = gpu_time + cpu_time; + } + total += event_time; std::string event_name; @@ -430,7 +442,7 @@ void ParseEvents(const std::vector>& events, event_idx[event_name] = event_items.size(); EventItem event_item = {event_name, 1, event_time, event_time, event_time, event_time, - 0.}; + gpu_time, cpu_time, 0.}; event_items.push_back(event_item); } else { int index = event_idx[event_name]; @@ -443,6 +455,8 @@ void ParseEvents(const std::vector>& events, // max time event_items[index].max_time = std::max(event_time, event_items[index].max_time); + event_items[index].gpu_time += gpu_time; + event_items[index].cpu_time += cpu_time; } // remove the push marker from the list @@ -481,20 +495,23 @@ void ParseEvents(const std::vector>& events, void DisableProfiler(EventSortingKey sorted_key, const std::string& profile_path) { + SynchronizeAllDevice(); std::lock_guard l(profiler_mu); if (g_state == ProfilerState::kDisabled) return; // Mark the profiling stop. - Mark("_stop_profiler_", nullptr); + Mark("_stop_profiler_"); - std::vector> all_events = GetAllEvents(); - ParseEvents(all_events, true, sorted_key); - ParseEvents(all_events, false, sorted_key); - ResetProfiler(); DeviceTracer* tracer = GetDeviceTracer(); if (tracer->IsEnabled()) { tracer->Disable(); tracer->GenProfile(profile_path); + tracer->GenEventKernelCudaElapsedTime(); } + + std::vector> all_events = GetAllEvents(); + ParseEvents(all_events, true, sorted_key); + ParseEvents(all_events, false, sorted_key); + ResetProfiler(); g_state = ProfilerState::kDisabled; should_send_profile_state = true; } diff --git a/paddle/fluid/platform/profiler.cu b/paddle/fluid/platform/profiler.cu new file mode 100644 index 0000000000000000000000000000000000000000..e115c554caf383bad29aa5d065ec0126427f8e78 --- /dev/null +++ b/paddle/fluid/platform/profiler.cu @@ -0,0 +1,50 @@ +/* Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved. + +licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/platform/profiler.h" + +#include + +namespace paddle { +namespace platform { + +__global__ void DummyKernel(int *a) { a[0] = 0; } + +static void ForEachDevice(std::function func) { + auto original_device = GetCurrentDeviceId(); + int count = GetCUDADeviceCount(); + for (int i = 0; i < count; i++) { + SetDeviceId(i); + func(i); + } + SetDeviceId(original_device); +} + +void DummyKernelAndEvent() { + for (int i = 0; i < 5; i++) { + ForEachDevice([](int d) { + CUDADeviceContext *dev_ctx = new CUDADeviceContext(CUDAPlace(d)); + Mark("_cuda_startup_"); + int *ptr; + PADDLE_ENFORCE(cudaMalloc(&ptr, sizeof(int))); + DummyKernel<<<1, 1, 0, dev_ctx->stream()>>>(ptr); + dev_ctx->Wait(); + PADDLE_ENFORCE(cudaFree(ptr)); + delete dev_ctx; + }); + } +} + +} // namespace platform +} // namespace paddle diff --git a/paddle/fluid/platform/profiler.h b/paddle/fluid/platform/profiler.h index f5d3490634f3199a23986ec3ae13d9fe3577ac35..4057e5ea0564099667517eb2c54feb2ff13303d1 100644 --- a/paddle/fluid/platform/profiler.h +++ b/paddle/fluid/platform/profiler.h @@ -28,17 +28,17 @@ class Event { public: // The DeviceContext is used to get the cuda stream. // If CPU profiling mode, can pass nullptr. - Event(EventType type, std::string name, uint32_t thread_id, - const DeviceContext* dev_ctx); + Event(EventType type, std::string name, uint32_t thread_id); const EventType& type() const; std::string name() const { return name_; } uint32_t thread_id() const { return thread_id_; } - bool has_cuda() const { return has_cuda_; } #ifdef PADDLE_WITH_CUDA +#ifndef PADDLE_WITH_CUPTI cudaEvent_t event() const { return event_; } int device() const { return device_; } +#endif #endif double CpuElapsedMs(const Event& e) const; @@ -49,11 +49,21 @@ class Event { std::string name_; uint32_t thread_id_; int64_t cpu_ns_; - bool has_cuda_; #ifdef PADDLE_WITH_CUDA +#ifdef PADDLE_WITH_CUPTI + int64_t gpu_ns_ = 0; + + public: + void AddCudaElapsedTime(int64_t start_ns, int64_t end_ns) { + gpu_ns_ += end_ns - start_ns; + } + + private: +#else cudaEvent_t event_ = nullptr; int device_ = -1; #endif +#endif }; enum ProfilerState { @@ -63,22 +73,19 @@ enum ProfilerState { kAll, // Profile both CPU and GPU. (Currently experimental). }; -void Mark(const std::string& name, const DeviceContext* dev_ctx); +void Mark(const std::string& name); -void PushEvent(const std::string& name, const DeviceContext* dev_ctx); +Event* PushEvent(const std::string& name); -void PopEvent(const std::string& name, const DeviceContext* dev_ctx); +void PopEvent(const std::string& name); struct RecordEvent { - // dev_ctx can be set to nullptr if device is cpu. - RecordEvent(const std::string& name, const DeviceContext* dev_ctx); + explicit RecordEvent(const std::string& name); ~RecordEvent(); bool is_enabled_; uint64_t start_ns_; - // The device context is used by Event to get the current cuda stream. - const DeviceContext* dev_ctx_; // Event name std::string name_; // Need to distinguish name by op type, block_id, program_id and perhaps @@ -88,8 +95,7 @@ struct RecordEvent { class RecordRPCEvent { public: - // dev_ctx can be set to nullptr if device is cpu. - RecordRPCEvent(const std::string& name, const DeviceContext* dev_ctx); + explicit RecordRPCEvent(const std::string& name); ~RecordRPCEvent() {} private: @@ -111,7 +117,16 @@ struct RecordBlock { std::vector> GetAllEvents(); // Candidate keys to sort the profiling report -enum EventSortingKey { kDefault, kCalls, kTotal, kMin, kMax, kAve }; +enum EventSortingKey { + kDefault, + kCalls, + kTotal, + kMin, + kMax, + kAve, + kCPUTime, + kGPUTime +}; // Enable the profiling function. void EnableProfiler(ProfilerState state); @@ -132,5 +147,9 @@ bool ShouldSendProfileState(); void SetProfileListener(); int64_t ListenerId(); +#ifdef PADDLE_WITH_CUDA +void DummyKernelAndEvent(); +#endif + } // namespace platform } // namespace paddle diff --git a/paddle/fluid/platform/profiler.proto b/paddle/fluid/platform/profiler.proto index 7b42aa785ec6ad5731e3adee1e9f189127a826a1..e761d7b266e92fd5d47b5b6073ffc8bea1dc877d 100644 --- a/paddle/fluid/platform/profiler.proto +++ b/paddle/fluid/platform/profiler.proto @@ -31,6 +31,7 @@ message Event { optional int64 sub_device_id = 6; optional MemCopy memcopy = 7; + optional string detail_info = 9; } message Profile { diff --git a/paddle/fluid/platform/profiler_test.cc b/paddle/fluid/platform/profiler_test.cc index 61f467814ba4a24c8b73f1bc614cda0ab8c4debd..528fe03c67a282248c551525e899279b561ce5d2 100644 --- a/paddle/fluid/platform/profiler_test.cc +++ b/paddle/fluid/platform/profiler_test.cc @@ -23,76 +23,49 @@ TEST(Event, CpuElapsedTime) { using paddle::platform::Event; using paddle::platform::EventType; - Event start_event(EventType::kPushRange, "test", 0, nullptr); - EXPECT_TRUE(start_event.has_cuda() == false); + Event start_event(EventType::kPushRange, "test", 0); int counter = 0; while (counter != 1000) { counter++; } - Event stop_event(EventType::kPopRange, "test", 0, nullptr); + Event stop_event(EventType::kPopRange, "test", 0); EXPECT_GT(start_event.CpuElapsedMs(stop_event), 0); } -#ifdef PADDLE_WITH_CUDA -TEST(Event, CudaElapsedTime) { - using paddle::platform::DeviceContext; - using paddle::platform::CUDADeviceContext; - using paddle::platform::CUDAPlace; - using paddle::platform::Event; - using paddle::platform::EventType; - - DeviceContext* dev_ctx = new CUDADeviceContext(CUDAPlace(0)); - Event start_event(EventType::kPushRange, "test", 0, dev_ctx); - EXPECT_TRUE(start_event.has_cuda() == true); - int counter = 0; - while (counter != 1000) { - counter++; - } - Event stop_event(EventType::kPopRange, "test", 0, dev_ctx); - EXPECT_GT(start_event.CudaElapsedMs(stop_event), 0); -} -#endif - TEST(RecordEvent, RecordEvent) { using paddle::platform::DeviceContext; using paddle::platform::Event; using paddle::platform::EventType; using paddle::platform::RecordEvent; + using paddle::platform::PushEvent; + using paddle::platform::PopEvent; using paddle::platform::ProfilerState; using paddle::platform::EventSortingKey; ProfilerState state = ProfilerState::kCPU; - DeviceContext* dev_ctx = nullptr; -#ifdef PADDLE_WITH_CUDA - using paddle::platform::CUDADeviceContext; - using paddle::platform::CUDAPlace; - state = ProfilerState::kCUDA; - dev_ctx = - new paddle::platform::CUDADeviceContext(paddle::platform::CUDAPlace(0)); -#endif EnableProfiler(state); /* Usage 1: - * PushEvent(evt_name, dev_ctx); + * PushEvent(evt_name); * ... * code to be analyzed * ... - * PopEvent(evt_name, dev_ctx); + * PopEvent(evt_name); */ LOG(INFO) << "Usage 1: PushEvent & PopEvent"; for (int loop = 0; loop < 3; ++loop) { for (int i = 1; i < 5; ++i) { std::string name = "op_" + std::to_string(i); - PushEvent(name, dev_ctx); + PushEvent(name); int counter = 1; while (counter != i * 1000) counter++; - PopEvent(name, dev_ctx); + PopEvent(name); } } /* Usage 2: * { - * RecordEvent record_event(name, dev_ctx); + * RecordEvent record_event(name); * ... * code to be analyzed * ... @@ -101,7 +74,7 @@ TEST(RecordEvent, RecordEvent) { LOG(INFO) << "Usage 2: RecordEvent"; for (int i = 1; i < 5; ++i) { std::string name = "evs_op_" + std::to_string(i); - RecordEvent record_event(name, dev_ctx); + RecordEvent record_event(name); int counter = 1; while (counter != i * 1000) counter++; } @@ -123,20 +96,20 @@ TEST(RecordEvent, RecordEvent) { LOG(INFO) << "Usage 3: nested RecordEvent"; for (int i = 1; i < 5; ++i) { std::string name = "ano_evs_op_" + std::to_string(i); - RecordEvent record_event(name, dev_ctx); + RecordEvent record_event(name); int counter = 1; while (counter != i * 100) counter++; { std::string nested_name = "nested_ano_evs_op_" + std::to_string(i); - RecordEvent nested_record_event(nested_name, dev_ctx); + RecordEvent nested_record_event(nested_name); int nested_counter = 1; while (nested_counter != i * 100) nested_counter++; } } // Bad Usage: - PushEvent("event_without_pop", dev_ctx); - PopEvent("event_without_push", dev_ctx); + PushEvent("event_without_pop"); + PopEvent("event_without_push"); std::vector> events = paddle::platform::GetAllEvents(); int cuda_startup_count = 0; diff --git a/paddle/fluid/pybind/CMakeLists.txt b/paddle/fluid/pybind/CMakeLists.txt index 803ea6b26087884ad79c6bf80238953a012eaddc..4ac5b83c56b114f4e3e4c78710716adc636ebe1d 100644 --- a/paddle/fluid/pybind/CMakeLists.txt +++ b/paddle/fluid/pybind/CMakeLists.txt @@ -26,5 +26,5 @@ if(WITH_PYTHON) get_property (os_dependency_modules GLOBAL PROPERTY OS_DEPENDENCY_MODULES) target_link_libraries(paddle_pybind ${os_dependency_modules}) - cc_test(tensor_py_test SRCS tensor_py_test.cc DEPS python) + cc_test(tensor_py_test SRCS tensor_py_test.cc DEPS python pybind) endif(WITH_PYTHON) diff --git a/paddle/fluid/pybind/inference_api.cc b/paddle/fluid/pybind/inference_api.cc index e05667d2c7e9ce5c64cfacee4919cd36d7383c0c..7db2bb451b49918fd8d92a6036c132d34e965c63 100644 --- a/paddle/fluid/pybind/inference_api.cc +++ b/paddle/fluid/pybind/inference_api.cc @@ -33,7 +33,6 @@ using paddle::PaddlePredictor; using paddle::NativeConfig; using paddle::NativePaddlePredictor; using paddle::AnalysisPredictor; -using paddle::contrib::AnalysisConfig; static void BindPaddleDType(py::module *m); static void BindPaddleBuf(py::module *m); @@ -75,12 +74,12 @@ void BindPaddleBuf(py::module *m) { .def(py::init([](std::vector &data) { auto buf = PaddleBuf(data.size() * sizeof(float)); std::memcpy(buf.data(), static_cast(data.data()), buf.length()); - return std::move(buf); + return buf; })) .def(py::init([](std::vector &data) { auto buf = PaddleBuf(data.size() * sizeof(int64_t)); std::memcpy(buf.data(), static_cast(data.data()), buf.length()); - return std::move(buf); + return buf; })) .def("resize", &PaddleBuf::Resize) .def("reset", diff --git a/paddle/fluid/pybind/ir.cc b/paddle/fluid/pybind/ir.cc index 24059140ab20e24917b93a5f60936b1087797ff9..1cd1be8e8d9da8c6a82ceefc3284084bfeda0252 100644 --- a/paddle/fluid/pybind/ir.cc +++ b/paddle/fluid/pybind/ir.cc @@ -13,10 +13,12 @@ // limitations under the License. #include "paddle/fluid/pybind/ir.h" +#include #include #include #include #include "paddle/fluid/framework/ir/graph.h" +#include "paddle/fluid/framework/ir/graph_helper.h" #include "paddle/fluid/framework/ir/graph_pattern_detector.h" #include "paddle/fluid/framework/ir/node.h" #include "paddle/fluid/framework/op_desc.h" @@ -27,6 +29,10 @@ namespace py = pybind11; using paddle::framework::ir::Graph; using paddle::framework::ir::Node; using paddle::framework::ir::GraphSafeRemoveNodes; +using paddle::framework::ir::HasCircle; +using paddle::framework::ir::GraphNum; +using paddle::framework::ir::TopologySortOperations; +using paddle::framework::ir::BuildOperationAdjList; using paddle::framework::OpDesc; using paddle::framework::ProgramDesc; using paddle::framework::VarDesc; @@ -36,6 +42,12 @@ namespace paddle { namespace pybind { void BindGraph(py::module *m) { m->def("graph_safe_remove_nodes", GraphSafeRemoveNodes); + m->def("has_circle", HasCircle); + m->def("graph_num", GraphNum); + m->def("topology_sort", TopologySortOperations, + return_value_policy::reference); + m->def("build_adjacency_list", BuildOperationAdjList, + return_value_policy::reference); py::class_>( *m, "Graph", "The graph is a Directed Acyclic Single Static Assignment Graph, see " @@ -46,7 +58,6 @@ void BindGraph(py::module *m) { .def("get_float", &Graph::Get) .def("get_double", &Graph::Get) .def("get_string", &Graph::Get) - .def("get_program", &Graph::Get) .def("get_marked_nodes", &Graph::Get>) .def("set", [](Graph &self, const std::string &attr_name, int attr) { return self.Set(attr_name, new int(attr)); }) @@ -63,11 +74,6 @@ void BindGraph(py::module *m) { [](Graph &self, const std::string &attr_name, double attr) { return self.Set(attr_name, new double(attr)); }) - .def("set", - [](Graph &self, const std::string &attr_name, - const ProgramDesc &attr) { - return self.Set(attr_name, new ProgramDesc(attr)); - }) .def("set", [](Graph &self, const std::string &attr_name, const std::unordered_set &attr) { @@ -108,42 +114,42 @@ void BindNode(py::module *m) { .def("is_op", &Node::IsOp) .def("is_var", &Node::IsVar) .def("is_ctrl_var", &Node::IsCtrlVar) + .def("clear_inputs", [](Node &self) { self.inputs.clear(); }) .def("inputs_remove", [](Node &self, int node_id) { - for (auto it = self.inputs.begin(); it != self.inputs.end(); - it++) { - if ((*it)->id() == node_id) { - self.inputs.erase(it); - } + auto pos = std::find_if( + self.inputs.begin(), self.inputs.end(), + [&node_id](const Node *n) { return n->id() == node_id; }); + if (pos != self.inputs.end()) { + self.inputs.erase(pos); } }) .def("inputs_remove", [](Node &self, Node &node) { - for (auto it = self.inputs.begin(); it != self.inputs.end(); - it++) { - if (*it == &node) { - self.inputs.erase(it); - } + auto pos = + std::find(self.inputs.begin(), self.inputs.end(), &node); + if (pos != self.inputs.end()) { + self.inputs.erase(pos); } }) .def("inputs_append", [](Node &self, Node &node) { self.inputs.push_back(&node); }) + .def("clear_outputs", [](Node &self) { self.outputs.clear(); }) .def("outputs_remove", [](Node &self, int node_id) { - for (auto it = self.outputs.begin(); it != self.outputs.end(); - it++) { - if ((*it)->id() == node_id) { - self.outputs.erase(it); - } + auto pos = std::find_if( + self.outputs.begin(), self.outputs.end(), + [&node_id](const Node *n) { return n->id() == node_id; }); + if (pos != self.outputs.end()) { + self.outputs.erase(pos); } }) .def("outputs_remove", [](Node &self, Node &node) { - for (auto it = self.outputs.begin(); it != self.outputs.end(); - it++) { - if (*it == &node) { - self.outputs.erase(it); - } + auto pos = + std::find(self.outputs.begin(), self.outputs.end(), &node); + if (pos != self.outputs.end()) { + self.outputs.erase(pos); } }) .def("outputs_append", diff --git a/paddle/fluid/pybind/pybind.cc b/paddle/fluid/pybind/pybind.cc index 97e5bbaaccaf7c702a324abd708a314c72ece004..d8e57a1ac6ccfc768a7f0604fcd1f32744a126df 100644 --- a/paddle/fluid/pybind/pybind.cc +++ b/paddle/fluid/pybind/pybind.cc @@ -37,6 +37,7 @@ limitations under the License. */ #include "paddle/fluid/framework/version.h" #include "paddle/fluid/imperative/layer.h" #include "paddle/fluid/memory/allocation/allocator_strategy.h" +#include "paddle/fluid/memory/allocation/legacy_allocator.h" #include "paddle/fluid/operators/activation_op.h" #include "paddle/fluid/operators/py_func_op.h" #include "paddle/fluid/operators/reader/lod_tensor_blocking_queue.h" @@ -105,6 +106,11 @@ bool IsCompiledWithDIST() { #endif } +template +static inline bool IsSamePlace(const PlaceType1 &p1, const PlaceType2 &p2) { + return paddle::platform::Place(p1) == paddle::platform::Place(p2); +} + PYBIND11_MODULE(core, m) { // Not used, just make sure cpu_info.cc is linked. paddle::platform::CpuTotalPhysicalMemory(); @@ -127,6 +133,13 @@ PYBIND11_MODULE(core, m) { m.add_object("_cleanup", py::capsule([]() { ScopePool::Instance().Clear(); })); + m.def("get_mem_usage", [](int device) { + return memory::allocation::GPUMemMonitor.GetMemUsage(device); + }); + + m.def("print_mem_usage", + []() { return memory::allocation::GPUMemMonitor.PrintMemUsage(); }); + py::class_(m, "VarBase", R"DOC()DOC") // .def(py::init<>()) .def(py::init(), py::arg("stop_gradient") = false) @@ -287,6 +300,7 @@ PYBIND11_MODULE(core, m) { .def("_get_float_element", TensorGetElement) .def("_set_double_element", TensorSetElement) .def("_get_double_element", TensorGetElement) + .def("_place", [](Tensor &self) { return self.place(); }) .def("_dtype", [](Tensor &self) { return self.type(); }); py::class_(m, "LoDTensor", R"DOC( @@ -364,7 +378,13 @@ PYBIND11_MODULE(core, m) { PADDLE_ENFORCE(CheckLoD(new_lod, vectorize(self.dims()).front()), "the provided lod info is invalid"); self.set_lod(new_lod); - }) + }, + py::arg("lod"), R"DOC( + Set LoD of the LoDTensor. + + Args: + lod (List[List[int]]): the lod to be set. + )DOC") .def("set_recursive_sequence_lengths", [](LoDTensor &self, const std::vector> &recursive_sequence_lengths) { @@ -380,7 +400,17 @@ PYBIND11_MODULE(core, m) { CheckLoD(new_offset_lod, vectorize(self.dims()).front()), "the provided recursive_sequence_lengths info is invalid"); self.set_lod(new_offset_lod); - }) + }, + py::arg("recursive_sequence_lengths"), R"DOC( + Set LoD of the LoDTensor according to recursive sequence length. + + For example, if recursive_sequence_lengths=[[2, 3]], meaning that + there are two sequences with length 2 and 3 respectively, the + corresponding lod would be [[0, 2, 2+3]], i.e, [[0, 2, 5]]. + + Args: + recursive_sequence_lengths (List[List[int]]): sequence lengths. + )DOC") .def("lod", [](LoDTensor &self) -> std::vector> { // output the offset-based lod info @@ -389,7 +419,13 @@ PYBIND11_MODULE(core, m) { new_lod.reserve(lod.size()); std::copy(lod.begin(), lod.end(), std::back_inserter(new_lod)); return new_lod; - }) + }, + R"DOC( + Return the LoD of the LoDTensor. + + Returns: + out (List[List[int]]): the lod of the LoDTensor. + )DOC") // Set above comments of set_lod. .def("recursive_sequence_lengths", [](LoDTensor &self) -> std::vector> { @@ -399,12 +435,25 @@ PYBIND11_MODULE(core, m) { new_lod.reserve(lod.size()); std::copy(lod.begin(), lod.end(), std::back_inserter(new_lod)); return new_lod; - }) - .def("has_valid_recursive_sequence_lengths", [](LoDTensor &self) -> bool { - // Check that the lod info is valid and match the outermost - // dimension of the LoDTensor data - return CheckLoD(self.lod(), vectorize(self.dims()).front()); - }); + }, + R"DOC( + Return the sequence length of the LoDTensor corresponding to LoD. + + Returns: + out (List[List[int]): the sequence lengths. + )DOC") + .def("has_valid_recursive_sequence_lengths", + [](LoDTensor &self) -> bool { + // Check that the lod info is valid and match the outermost + // dimension of the LoDTensor data + return CheckLoD(self.lod(), vectorize(self.dims()).front()); + }, + R"DOC( + Check whether the lod of the LoDTensor is valid. + + Returns: + out (bool): whether the lod is valid. + )DOC"); py::class_(m, "SelectedRows") .def("__init__", @@ -540,11 +589,45 @@ All parameter, weight, gradient are variables in Paddle. [](Scope &self, const std::string &name) -> Variable * { return self.Var(name); }, + py::arg("name"), + R"DOC( + Find or create variable named :code:`name` in the current scope. + + If the variable named :code:`name` does not exist in the + current scope, the variable would be created. Otherwise, + return the existing variable. + + Args: + name (str): the variable name. + + Returns: + out (core.Variable): the found or created variable. + )DOC", + py::return_value_policy::reference) + .def("find_var", &Scope::FindVar, py::arg("name"), + R"DOC( + Find variable named :code:`name` in the current scope or + its parent scope. Return None if not found. + + Args: + name (str): the variable name. + + Returns: + out (core.Variable|None): the found variable or None. + )DOC", py::return_value_policy::reference) - .def("find_var", &Scope::FindVar, py::return_value_policy::reference) .def("new_scope", [](Scope &self) -> Scope * { return &self.NewScope(); }, + R"DOC( + Create a new sub-scope of the current scope. + + Returns: + out (core._Scope): the created sub-scope. + )DOC", py::return_value_policy::reference) - .def("drop_kids", &Scope::DropKids); + .def("drop_kids", &Scope::DropKids, + R"DOC( + Delete all sub-scopes of the current scope. + )DOC"); m.def("Scope", []() -> Scope * { @@ -552,6 +635,12 @@ All parameter, weight, gradient are variables in Paddle. ScopePool::Instance().Insert(std::unique_ptr(s)); return s; }, + R"DOC( + Create a new scope. + + Returns: + out (core._Scope): the created scope. + )DOC", py::return_value_policy::reference); //! @note: Be careful! PyBind will return std::string as an unicode, not @@ -648,23 +737,51 @@ All parameter, weight, gradient are variables in Paddle. PADDLE_THROW("Cannot use CUDAPlace in CPU only version"); #endif }) + .def("_equals", &IsSamePlace) + .def("_equals", &IsSamePlace) + .def("_equals", &IsSamePlace) + .def("_equals", + &IsSamePlace) .def("__str__", string::to_string); py::class_(m, "CPUPlace") .def(py::init<>()) + .def("_equals", &IsSamePlace) + .def("_equals", &IsSamePlace) + .def("_equals", &IsSamePlace) + .def("_equals", + &IsSamePlace) .def("__str__", string::to_string); py::class_(m, "CUDAPinnedPlace") .def("__init__", - [](platform::CUDAPinnedPlace &) { + [](platform::CUDAPinnedPlace &self) { #ifndef PADDLE_WITH_CUDA PADDLE_THROW("Cannot use CUDAPinnedPlace in CPU only version"); #endif + new (&self) platform::CUDAPinnedPlace(); }) + .def("_equals", &IsSamePlace) + .def("_equals", + &IsSamePlace) + .def("_equals", + &IsSamePlace) + .def("_equals", + &IsSamePlace) .def("__str__", string::to_string); py::class_(m, "Place") .def(py::init<>()) + .def("_equals", &IsSamePlace) + .def("_equals", &IsSamePlace) + .def("_equals", &IsSamePlace) + .def("_equals", &IsSamePlace) + .def("is_gpu_place", + [](platform::Place &self) { return platform::is_gpu_place(self); }) + .def("gpu_device_id", + [](platform::Place &self) { + return boost::get(self).device; + }) .def("set_place", [](platform::Place &self, const platform::CPUPlace &cpu_place) { self = cpu_place; @@ -774,11 +891,13 @@ All parameter, weight, gradient are variables in Paddle. self[i].ShareDataWith(t); self[i].set_lod(t.lod()); }) - .def("append", [](LoDTensorArray &self, const LoDTensor &t) { - self.emplace_back(); - self.back().ShareDataWith(t); - self.back().set_lod(t.lod()); - }); + .def("append", + [](LoDTensorArray &self, const LoDTensor &t) { + self.emplace_back(); + self.back().ShareDataWith(t); + self.back().set_lod(t.lod()); + }, + py::arg("tensor"), "Append a LoDensor to LoDTensorArray."); m.def("IsInplace", [](std::string op) -> bool { return operators::IsInplace(op); }); @@ -814,8 +933,7 @@ All parameter, weight, gradient are variables in Paddle. m.def("disable_profiler", platform::DisableProfiler); m.def("is_profiler_enabled", platform::IsProfileEnabled); m.def("reset_profiler", platform::ResetProfiler); - m.def("get_pass", [](const py::bytes &binary_str) { - std::string pass_type(binary_str); + m.def("get_pass", [](const std::string &pass_type) { auto pass = framework::ir::PassRegistry::Instance().Get(pass_type); return std::shared_ptr(std::move(pass)); }); @@ -823,10 +941,9 @@ All parameter, weight, gradient are variables in Paddle. py::class_> pass(m, "Pass"); pass.def(py::init()) .def("has", &ir::Pass::Has) - .def("set", - [](ir::Pass &self, const std::string &attr_name, - const ProgramDesc &attr) { - return self.Set(attr_name, new ProgramDesc(attr)); + .def("set_not_owned", + [](ir::Pass &self, const std::string &attr_name, ProgramDesc &attr) { + self.SetNotOwned(attr_name, &attr); }) .def( "set", @@ -835,7 +952,6 @@ All parameter, weight, gradient are variables in Paddle. }) .def("set", [](ir::Pass &self, const std::string &name, int val) { self.Set(name, new int(val)); }) - .def("get_program", &ir::Pass::Get) .def("type", &ir::Pass::Type) .def("apply", [](ir::Pass &self, std::shared_ptr graph) { std::unique_ptr origin_graph(graph.get()); @@ -1085,9 +1201,9 @@ All parameter, weight, gradient are variables in Paddle. [](const BuildStrategy &self) { return self.is_distribution_; }, [](BuildStrategy &self, bool b) { self.is_distribution_ = b; }) .def_property( - "memory_early_delete", - [](const BuildStrategy &self) { return self.memory_early_delete_; }, - [](BuildStrategy &self, bool b) { self.memory_early_delete_ = b; }) + "enable_inplace", + [](const BuildStrategy &self) { return self.enable_inplace_; }, + [](BuildStrategy &self, bool b) { self.enable_inplace_ = b; }) .def("_finalize_strategy_and_create_passes", [](BuildStrategy &self) -> std::shared_ptr { return self.CreatePassesFromStrategy(true); diff --git a/paddle/fluid/string/printf.h b/paddle/fluid/string/printf.h index 0b94b60018aac3a61edfda4d7ecb762e9fe70673..16bb3771f2e9bcc07028ef2039fed8691f9aab97 100644 --- a/paddle/fluid/string/printf.h +++ b/paddle/fluid/string/printf.h @@ -84,6 +84,8 @@ void Fprintf(std::ostream& out, const char* fmt, const Args&... args) { tinyformat::vformat(out, fmt, tinyformat::makeFormatList(args...)); } +inline std::string Sprintf() { return ""; } + template std::string Sprintf(const Args&... args) { std::ostringstream oss; diff --git a/paddle/fluid/train/demo/README.md b/paddle/fluid/train/demo/README.md index 191da20669e185d819ec5eed55427461cc0b10e4..bd53ab4b0c023b2591d792b504ab496a42d2835d 100644 --- a/paddle/fluid/train/demo/README.md +++ b/paddle/fluid/train/demo/README.md @@ -9,7 +9,6 @@ PADDLE_LIB=/paddle/lib/dir cmake .. -DFLUID_INSTALL_DIR=$PADDLE_LIB \ -DCMAKE_BUILD_TYPE=Release \ - -DWITH_FLUID_ONLY=ON \ -DWITH_GPU=OFF \ -DWITH_STYLE_CHECK=OFF \ -DWITH_MKL=OFF \ diff --git a/paddle/fluid/train/demo/demo_trainer.cc b/paddle/fluid/train/demo/demo_trainer.cc index a0757b53f37b29de0b3802c345b1ad9db69f16e9..1087f5672459506cc7b824127cd822c0df7ba566 100644 --- a/paddle/fluid/train/demo/demo_trainer.cc +++ b/paddle/fluid/train/demo/demo_trainer.cc @@ -73,7 +73,7 @@ int main() { PADDLE_ENFORCE_NE(loss_name, "", "loss not found"); // init all parameters - executor.Run(*startup_program.get(), &scope, 0); + executor.Run(*startup_program, &scope, 0); // prepare data auto x_var = scope.Var("x"); @@ -101,7 +101,7 @@ int main() { clock_t t1 = clock(); for (int i = 0; i < 10; ++i) { - executor.Run(*train_program.get(), &scope, 0, false, true); + executor.Run(*train_program, &scope, 0, false, true); std::cout << "step: " << i << " loss: " << loss_var->Get().data()[0] << std::endl; diff --git a/paddle/fluid/train/test_train_recognize_digits.cc b/paddle/fluid/train/test_train_recognize_digits.cc index e8731dd51ad698e53b7f10cc781c52134f2d17a8..a7846da8c191ac96e9ad7fb5b3184518e32120b2 100644 --- a/paddle/fluid/train/test_train_recognize_digits.cc +++ b/paddle/fluid/train/test_train_recognize_digits.cc @@ -74,7 +74,7 @@ void Train() { float first_loss = 0.0; float last_loss = 0.0; for (int i = 0; i < 100; ++i) { - executor.Run(*train_program.get(), &scope, 0, false, true); + executor.Run(*train_program, &scope, 0, false, true); if (i == 0) { first_loss = loss_var->Get().data()[0]; } else if (i == 99) { diff --git a/paddle/scripts/README.md b/paddle/scripts/README.md index 6c608fce3cdad38f3109e563be3ffbe2f73e5390..1db262f06d97665ee09b8e1d3485982b6b1b33d6 100644 --- a/paddle/scripts/README.md +++ b/paddle/scripts/README.md @@ -66,12 +66,10 @@ Users can specify the following Docker build arguments with either "ON" or "OFF" | `WITH_AVX` | OFF | Set to "ON" to enable AVX support. | | `WITH_TESTING` | OFF | Build unit tests binaries. | | `WITH_MKL` | ON | Build with [Intel® MKL](https://software.intel.com/en-us/mkl) and [Intel® MKL-DNN](https://github.com/01org/mkl-dnn) support. | -| `WITH_GOLANG` | OFF | Build fault-tolerant parameter server written in go. | | `WITH_PYTHON` | ON | Build with python support. Turn this off if build is only for capi. | | `WITH_STYLE_CHECK` | ON | Check the code style when building. | | `PYTHON_ABI` | "" | Build for different python ABI support, can be cp27-cp27m or cp27-cp27mu | | `RUN_TEST` | OFF | Run unit test immediently after the build. | -| `WITH_DOC` | OFF | Build docs after build binaries. | | `WOBOQ` | OFF | Generate WOBOQ code viewer under `build/woboq_out` | ## Docker Images diff --git a/paddle/scripts/fast_install.sh b/paddle/scripts/fast_install.sh new file mode 100644 index 0000000000000000000000000000000000000000..0461944ca8c6c5aeaffcac1eceac097e4d25b6d1 --- /dev/null +++ b/paddle/scripts/fast_install.sh @@ -0,0 +1,1126 @@ +#!/bin/bash + +## purple to echo +function purple(){ + echo -e "\033[35m$1\033[0m" +} + + +## green to echo +function green(){ + echo -e "\033[32m$1\033[0m" +} + +## Error to warning with blink +function bred(){ + echo -e "\033[31m\033[01m\033[05m$1\033[0m" +} + +## Error to warning with blink +function byellow(){ + echo -e "\033[33m\033[01m\033[05m$1\033[0m" +} + + +## Error +function red(){ + echo -e "\033[31m\033[01m$1\033[0m" +} + +## warning +function yellow(){ + echo -e "\033[33m\033[01m$1\033[0m" +} + +path='http://paddlepaddle.org/download?url=' +#release_version=`curl -s https://pypi.org/project/paddlepaddle/|grep -E "/project/paddlepaddle/"|grep "release"|awk -F '/' '{print $(NF-1)}'|head -1` +release_version=1.2.0 +python_list=( +"27" +"35" +"36" +"37" +) + + +function use_cpu(){ + while true + do + read -p "是否安装CPU版本的PaddlePaddle？(y/n)" cpu_option + cpu_option=`echo $cpu_option | tr 'A-Z' 'a-z'` + if [[ "$cpu_option" == "" || "$cpu_option" == "n" ]];then + echo "退出安装中..." + exit + else + GPU='cpu' + echo "将为您安装CPU版本的PaddlePaddle" + break + fi + done +} + +function checkLinuxCUDNN(){ + echo + read -n1 -p "请按回车键进行下一步..." + echo + while true + do + version_file='/usr/local/cuda/include/cudnn.h' + if [ -f "$version_file" ];then + CUDNN=`cat $version_file | grep CUDNN_MAJOR |awk 'NR==1{print $NF}'` + fi + if [ "$CUDNN" == "" ];then + version_file=`sudo find /usr -name "cudnn.h"|head -1` + if [ "$version_file" != "" ];then + CUDNN=`cat ${version_file} | grep CUDNN_MAJOR -A 2|awk 'NR==1{print $NF}'` + else + echo "检测结果：未在常规路径下找到cuda/include/cudnn.h文件" + while true + do + read -p "请核实cudnn.h位置，并在此输入路径（请注意，路径需要输入到“cudnn.h”这一级）:" cudnn_version + echo + if [ "$cudnn_version" == "" ] || [ ! -f "$cudnn_version" ];then + read -p "仍未找到cuDNN，输入y将安装CPU版本的PaddlePaddle，输入n可重新录入cuDNN路径，请输入（y/n）" cpu_option + echo + cpu_option=`echo $cpu_option | tr 'A-Z' 'a-z'` + if [ "$cpu_option" == "y" -o "$cpu_option" == "" ];then + GPU='cpu' + break + else + echo "请重新输入" + echo + fi + else + CUDNN=`cat $cudnn_version | grep CUDNN_MAJOR |awk 'NR==1{print $NF}'` + echo "检测结果：找到cudnn.h" + break + fi + done + if [ "$GPU" == "cpu" ];then + break + fi + fi + fi + if [ "$CUDA" == "9" -a "$CUDNN" != "7" ];then + echo + echo "目前CUDA9下仅支持cuDNN7，暂不支持您机器上的CUDNN${CUDNN}。您可以访问NVIDIA官网下载适合版本的CUDNN，请ctrl+c退出安装进程。按回车键将为您安装CPU版本的PaddlePaddle" + echo + use_cpu() + if [ "$GPU"=="cpu" ];then + break + fi + fi + + if [ "$CUDNN" == 5 ] || [ "$CUDNN" == 7 ];then + echo + echo "您的CUDNN版本是: CUDNN$CUDNN" + break + else + echo + read -n1 -p "目前支持的CUDNN版本为5和7,暂不支持您机器上的CUDNN${CUDNN}，将为您安装CPU版本的PaddlePaddle,请按回车键开始安装" + echo + use_cpu + if [ "$GPU"=="cpu" ];then + break + fi + fi + done +} + +function checkLinuxCUDA(){ + while true + do + CUDA=`echo ${CUDA_VERSION}|awk -F "[ .]" '{print $1}'` + if [ "$CUDA" == "" ];then + if [ -f "/usr/local/cuda/version.txt" ];then + CUDA=`cat /usr/local/cuda/version.txt | grep 'CUDA Version'|awk -F '[ .]' '{print $3}'` + tmp_cuda=$CUDA + fi + if [ -f "/usr/local/cuda8/version.txt" ];then + CUDA=`cat /usr/local/cuda8/version.txt | grep 'CUDA Version'|awk -F '[ .]' '{print $3}'` + tmp_cuda8=$CUDA + fi + if [ -f "/usr/local/cuda9/version.txt" ];then + CUDA=`cat /usr/local/cuda9/version.txt | grep 'CUDA Version'|awk -F '[ .]' '{print $3}'` + tmp_cuda9=$CUDA + fi + fi + + if [ "$tmp_cuda" != "" ];then + echo "检测结果：找到CUDA $tmp_cuda" + fi + if [ "$tmp_cudai8" != "" ];then + echo "检测结果：找到CUDA $tmp_cuda8" + fi + if [ "$tmp_cuda9" != "" ];then + echo "检测结果：找到CUDA $tmp_cuda9" + fi + + if [ "$CUDA" == "" ];then + echo "检测结果：没有在常规路径下找到cuda/version.txt文件" + while true + do + read -p "请输入cuda/version.txt的路径:" cuda_version + if [ "$cuda_version" == "" || ! -f "$cuda_version" ];then + read -p "仍未找到CUDA，输入y将安装CPU版本的PaddlePaddle，输入n可重新录入CUDA路径，请输入（y/n）" cpu_option + cpu_option=`echo $cpu_option | tr 'A-Z' 'a-z'` + if [ "$cpu_option" == "y" || "$cpu_option" == "" ];then + GPU='cpu' + break + else + echo "重新输入..." + fi + else + CUDA=`cat $cuda_version | grep 'CUDA Version'|awk -F '[ .]' '{print $3}'` + if [ "$CUDA" == "" ];then + echo "未能在version.txt中找到CUDA相关信息" + else + break + fi + fi + done + if [ "$GPU" == "cpu" ];then + break + fi + fi + + if [ "$CUDA" == "8" ] || [ "$CUDA" == "9" ];then + echo "您的CUDA版本是${CUDA}" + break + else + echo "目前支持CUDA8/9，暂不支持您的CUDA${CUDA}，将为您安装CPU版本的PaddlePaddle" + echo + use_cpu + fi + + if [ "$GPU" == "cpu" ];then + break + fi + done +} + +function checkLinuxMathLibrary(){ + while true + do + if [ "$AVX" == "" ];then + echo "正在检测您环境中是否存在AVX指令集..." + echo + echo "检测结果：您电脑上没有AVX指令集，目前针对无AVX指令集的环境，我们仅提供支持mkl数学库的PaddlePaddle，将为您安装此版本的PaddlePaddle" + math='mkl' + break + elif [ "$GPU" == "gpu" ];then + math='mkl' + echo "检测到您的机器上配备GPU，推荐您使用mkl数学库" + break + else + read -p "请输入您希望使用的数学库： + 1：openblas 一个高性能多核 BLAS 库 + 2：mkl（推荐） 英特尔数学核心函数库 + => 请输入数字1或2。如输入其他字符或直接回车，将会默认选择【 2. mkl 】 。请在这里输入并回车：" math + if [ "$math" == "" ];then + math="mkl" + echo "您选择了数字【2】" + break + fi + if [ "$math" == "1" ];then + math=openblas + echo "您选择了数字【1】" + break + elif [ "$math" == "2" ];then + math=mkl + echo "您选择了数字【2】" + break + fi + echo "输入错误，请再次输入" + fi + done +} + +function checkLinuxPaddleVersion(){ + read -n1 -p "请按回车键继续..." + while true + do + read -p " + 1. 开发版：对应Github上develop分支，如您需要开发、或希望使用PaddlePaddle最新功能，请选用此版本 + 2. 稳定版（推荐）：如您无特殊开发需求，建议使用此版本，目前最新的版本号为 ${release_version} + => 请输入数字1或2。如输入其他字符或直接回车，将会默认选择【 2. 稳定版 】 。请在这里输入并回车：" paddle_version + if [ "$paddle_version" == "" ];then + paddle_version="release-${release_version}" + echo "您选择了数字【2】，为您安装release-${release_version}" + break + fi + if [ "$paddle_version" == "1" ];then + echo "您选择了数字【1】，将为您安装开发版" + break + elif [ "$paddle_version" == "2" ];then + echo "您选择了数字【2】，为您安装release-${release_version}" + break + fi + echo "输入错误，请再次输入" + done +} + +function checkPythonVirtualenv(){ + while true + do + read -p " + 是否使用python virtualenv虚环境安装(y/n)": check_virtualenv + case $check_virtualenv in + y) + echo "为您使用python虚环境安装" + ;; + n) + break + ;; + *) + continue + ;; + esac + + virtualenv_path=`which virtualenv 2>&1` + if [ "$virtualenv_path" == "" ];then + $python_path -m pip install virtualenv + if [ "$?" != '0' ];then + echo "安装虚拟环境失败,请检查本地环境" + fi + fi + + while true + do + read -p "请输入虚拟环境名字：" virtualenv_name + if [ "$virtualenv_name" == "" ];then + echo "不能为空" + continue + fi + break + done + + virtualenv -p $python_path ${virtualenv_name} + if [ "$?" != 0 ];then + echo "创建虚环境失败,请检查环境" + exit 2 + fi + cd ${virtualenv_name} + source ./bin/activate + + if [ "$?" == 0 ];then + use_virtualenv= + python_path=`which python` + break + else + echo "创建虚环境失败,请检查环境" + exit 2 + fi + done +} + +function checkLinuxPython(){ + python_path=`which python 2>/dev/null` + while true + do + if [ "$python_path" == '' ];then + while true + do + read -p "没有找到默认的python版本,请输入要安装的python路径:" python_path + python_path=`$python_path -V` + if [ "$python_path" != "" ];then + break + else + echo "输入路径有误,未找到pyrhon" + fi + done + fi + + python_version=`$python_path -V 2>&1|awk -F '[ .]' '{print $2$3}'` + pip_version=`$python_path -m pip -V|awk -F '[ .]' '{print $2}'` + while true + do + read -p " + 找到python版本$python_version,使用请输入y,选择其他版本请输n(y/n):" check_python + case $check_python in + n) + read -p "请指定您的python路径:" new_python_path + python_V=`$new_python_path -V 2>/dev/null` + if [ "$python_V" != "" ];then + python_path=$new_python_path + python_version=`$python_path -V 2>&1|awk -F '[ .]' '{print $2$3}'` + pip_version=`python -m pip -V|awk -F '[ .]' '{print $2}'` + echo "您的python版本为${python_version}" + break + else + echo 输入有误,未找到python路径 + fi + ;; + y) + break + ;; + *) + echo "输入有误，请重新输入." + continue + ;; + esac + done + + if [ "$pip_version" -lt 9 ];then + echo "您的pip版本小于9.0.1 请升级pip (pip install --upgrade pip)" + exit 0 + fi + + if [ "$python_version" == "27" ];then + uncode=`python -c "import pip._internal;print(pip._internal.pep425tags.get_supported())"|grep "cp27mu"` + if [[ "$uncode" == "" ]];then + uncode= + else + uncode=u + fi + fi + + version_list=`echo "${python_list[@]}" | grep "$python_version" ` + if [ "$version_list" == "" ];then + echo "找不到可用的 pip, 我们只支持Python27/35/36/37及其对应的pip, 请重新输入， 或使用ctrl + c退出 " + else + break + fi + done +} + +function checkLinuxAVX(){ + while true + do + if [[ "$AVX" != "" ]];then + AVX="avx" + break + else + if [ "$CUDA" == "8" -a "$CUDNN" == "7" ] || [ "$GPU" == "cpu" ];then + AVX="noavx" + break + else + echo "Step 6. 检测是否有avx" + echo + echo "检测结果：未能找到avx，我们仅提供CPU版本或配置为CUDA8 cuDNN7的GPU版本的安装包" + break + fi + fi + done +} + +function PipLinuxInstall(){ + wheel_cpu_release="http://paddle-wheel.bj.bcebos.com/${release_version}-${GPU}-${AVX}-${math}/paddlepaddle-${release_version}-cp${python_version}-cp${python_version}m${uncode}-linux_x86_64.whl" + wheel_gpu_release="http://paddle-wheel.bj.bcebos.com/${release_version}-gpu-cuda${CUDA}-cudnn${CUDNN}-${AVX}-${math}/paddlepaddle_gpu-${release_version}.post${CUDA}${CUDNN}-cp${python_version}-cp${python_version}m${uncode}-linux_x86_64.whl" + wheel_gpu_release_noavx="http://paddle-wheel.bj.bcebos.com/${release_version}-gpu-cuda${CUDA}-cudnn${CUDNN}-${AVX}-${math}/paddlepaddle_gpu-${release_version}-cp${python_version}-cp${python_version}m${uncode}-linux_x86_64.whl" + wheel_cpu_develop="http://paddle-wheel.bj.bcebos.com/latest-cpu-${AVX}-${math}/paddlepaddle-latest-cp${python_version}-cp${python_version}m${uncode}-linux_x86_64.whl" + wheel_gpu_develop="http://paddle-wheel.bj.bcebos.com/latest-gpu-cuda${CUDA}-cudnn${CUDNN}-${AVX}-${math}/paddlepaddle_gpu-latest-cp${python_version}-cp${python_version}m${uncode}-linux_x86_64.whl" + + if [[ "$paddle_version" == "2" ]];then + if [[ "$GPU" == "gpu" ]];then + if [[ ${AVX} == "avx" ]];then + rm -rf `echo $wheel_gpu_release|awk -F '/' '{print $NF}'` + wget -q $wheel_gpu_release + if [ "$?" == "0" ];then + $python_path -m pip install ${use_virtualenv} -i https://mirrors.aliyun.com/pypi/simple --trusted-host=mirrors.aliyun.com $wheel_gpu_release + if [ "$?" == 0 ];then + echo 安装成功 + else + echo 安装失败 + exit 1 + fi + else + echo paddlepaddle whl包下载失败 + exit 1 + fi + else + rm -rf `echo $wheel_gpu_release_novax|awk -F '/' '{print $NF}'` + wget -q $wheel_gpu_release_novax + if [ "$?" == "0" ];then + $python_path -m pip install ${use_virtualenv} -i https://mirrors.aliyun.com/pypi/simple --trusted-host=mirrors.aliyun.com $wheel_gpu_release_noavx + if [ "$?" == 0 ];then + echo 安装成功 + else + echo 安装失败 + exit 1 + fi + else + echo paddlepaddle whl包下载失败 + exit 1 + fi + fi + else + rm -rf `echo $wheel_cpu_release|awk -F '/' '{print $NF}'` + wget -q $wheel_cpu_release + if [ "$?" == "0" ];then + $python_path -m pip install ${use_virtualenv} -i https://mirrors.aliyun.com/pypi/simple --trusted-host=mirrors.aliyun.com $wheel_cpu_release + if [ "$?" == 0 ];then + echo 安装成功 + else + echo 安装失败 + exit 1 + fi + else + echo paddlepaddle whl包下载失败 + exit 1 + fi + fi + else + if [[ "$GPU" == "gpu" ]];then + rm -rf `echo $wheel_gpu_develop|awk -F '/' '{print $NF}'` + wget -q $wheel_gpu_develop + if [ "$?" == "0" ];then + $python_path -m pip install ${use_virtualenv} -i https://mirrors.aliyun.com/pypi/simple --trusted-host=mirrors.aliyun.com $wheel_gpu_develop + if [ "$?" == 0 ];then + echo 安装成功 + else + echo 安装失败 + exit 1 + fi + else + echo paddlepaddle whl包下载失败 + exit 1 + fi + else + rm -rf `echo $wheel_cpu_develop|awk -F '/' '{print $NF}'` + wget -q $wheel_cpu_develop + if [ "$?" == "0" ];then + $python_path -m pip install ${use_virtualenv} -i https://mirrors.aliyun.com/pypi/simple --trusted-host=mirrors.aliyun.com $wheel_cpu_develop + if [ "$?" == 0 ];then + echo 安装成功 + else + echo 安装失败 + exit 1 + fi + else + echo paddlepaddle whl包下载失败 + exit 1 + fi + fi + fi +} + + +function checkLinuxGPU(){ + read -n1 -p "即将检测您的机器是否含GPU，请按回车键继续..." + echo + AVX=`cat /proc/cpuinfo |grep avx|tail -1|grep avx` + which nvidia-smi >/dev/null 2>&1 + if [ "$?" != "0" ];then + GPU='cpu' + echo "未在机器上找到GPU，或PaddlePaddle暂不支持此型号的GPU" + else + GPU='gpu' + echo "已在您的机器上找到GPU，即将确认CUDA和CUDNN版本..." + echo + fi + if [ "$GPU" == 'gpu' ];then + checkLinuxCUDA + checkLinuxCUDNN + fi +} + +function linux(){ +gpu_list=( +"GeForce 410M" +"GeForce 610M" +"GeForce 705M" +"GeForce 710M" +"GeForce 800M" +"GeForce 820M" +"GeForce 830M" +"GeForce 840M" +"GeForce 910M" +"GeForce 920M" +"GeForce 930M" +"GeForce 940M" +"GeForce GT 415M" +"GeForce GT 420M" +"GeForce GT 430" +"GeForce GT 435M" +"GeForce GT 440" +"GeForce GT 445M" +"GeForce GT 520" +"GeForce GT 520M" +"GeForce GT 520MX" +"GeForce GT 525M" +"GeForce GT 540M" +"GeForce GT 550M" +"GeForce GT 555M" +"GeForce GT 610" +"GeForce GT 620" +"GeForce GT 620M" +"GeForce GT 625M" +"GeForce GT 630" +"GeForce GT 630M" +"GeForce GT 635M" +"GeForce GT 640" +"GeForce GT 640 (GDDR5)" +"GeForce GT 640M" +"GeForce GT 640M LE" +"GeForce GT 645M" +"GeForce GT 650M" +"GeForce GT 705" +"GeForce GT 720" +"GeForce GT 720M" +"GeForce GT 730" +"GeForce GT 730M" +"GeForce GT 735M" +"GeForce GT 740" +"GeForce GT 740M" +"GeForce GT 745M" +"GeForce GT 750M" +"GeForce GTS 450" +"GeForce GTX 1050" +"GeForce GTX 1060" +"GeForce GTX 1070" +"GeForce GTX 1080" +"GeForce GTX 1080 Ti" +"GeForce GTX 460" +"GeForce GTX 460M" +"GeForce GTX 465" +"GeForce GTX 470" +"GeForce GTX 470M" +"GeForce GTX 480" +"GeForce GTX 480M" +"GeForce GTX 485M" +"GeForce GTX 550 Ti" +"GeForce GTX 560M" +"GeForce GTX 560 Ti" +"GeForce GTX 570" +"GeForce GTX 570M" +"GeForce GTX 580" +"GeForce GTX 580M" +"GeForce GTX 590" +"GeForce GTX 650" +"GeForce GTX 650 Ti" +"GeForce GTX 650 Ti BOOST" +"GeForce GTX 660" +"GeForce GTX 660M" +"GeForce GTX 660 Ti" +"GeForce GTX 670" +"GeForce GTX 670M" +"GeForce GTX 670MX" +"GeForce GTX 675M" +"GeForce GTX 675MX" +"GeForce GTX 680" +"GeForce GTX 680M" +"GeForce GTX 680MX" +"GeForce GTX 690" +"GeForce GTX 750" +"GeForce GTX 750 Ti" +"GeForce GTX 760" +"GeForce GTX 760M" +"GeForce GTX 765M" +"GeForce GTX 770" +"GeForce GTX 770M" +"GeForce GTX 780" +"GeForce GTX 780M" +"GeForce GTX 780 Ti" +"GeForce GTX 850M" +"GeForce GTX 860M" +"GeForce GTX 870M" +"GeForce GTX 880M" +"GeForce GTX 950" +"GeForce GTX 950M" +"GeForce GTX 960" +"GeForce GTX 960M" +"GeForce GTX 965M" +"GeForce GTX 970" +"GeForce GTX 970M" +"GeForce GTX 980" +"GeForce GTX 980M" +"GeForce GTX 980 Ti" +"GeForce GTX TITAN" +"GeForce GTX TITAN Black" +"GeForce GTX TITAN X" +"GeForce GTX TITAN Z" +"Jetson TK1" +"Jetson TX1" +"Jetson TX2" +"Mobile Products" +"NVIDIA NVS 310" +"NVIDIA NVS 315" +"NVIDIA NVS 510" +"NVIDIA NVS 810" +"NVIDIA TITAN V" +"NVIDIA TITAN X" +"NVIDIA TITAN Xp" +"NVS 4200M" +"NVS 5200M" +"NVS 5400M" +"Quadro 410" +"Quadro GP100" +"Quadro K1100M" +"Quadro K1200" +"Quadro K2000" +"Quadro K2000D" +"Quadro K2100M" +"Quadro K2200" +"Quadro K2200M" +"Quadro K3100M" +"Quadro K4000" +"Quadro K4100M" +"Quadro K420" +"Quadro K4200" +"Quadro K4200M" +"Quadro K5000" +"Quadro K500M" +"Quadro K5100M" +"Quadro K510M" +"Quadro K5200" +"Quadro K5200M" +"Quadro K600" +"Quadro K6000" +"Quadro K6000M" +"Quadro K610M" +"Quadro K620" +"Quadro K620M" +"Quadro M1000M" +"Quadro M1200" +"Quadro M2000" +"Quadro M2000M" +"Quadro M2200" +"Quadro M3000M" +"Quadro M4000" +"Quadro M4000M" +"Quadro M5000" +"Quadro M5000M" +"Quadro M500M" +"Quadro M520" +"Quadro M5500M" +"Quadro M6000" +"Quadro M6000 24GB" +"Quadro M600M" +"Quadro M620" +"Quadro Mobile Products" +"Quadro P1000" +"Quadro P2000" +"Quadro P3000" +"Quadro P400" +"Quadro P4000" +"Quadro P5000" +"Quadro P600" +"Quadro P6000" +"Quadro Plex 7000" +"Tegra K1" +"Tegra X1" +"Tesla C2050/C2070" +"Tesla C2075" +"Tesla Data Center Products" +"Tesla K10" +"Tesla K20" +"Tesla K40" +"Tesla K80" +"Tesla M40" +"Tesla M60" +"Tesla P100" +"Tesla P4" +"Tesla P40" +"Tesla V100") + + echo "Step 2. 检测GPU型号和CUDA/cuDNN版本" + echo + checkLinuxGPU + echo + echo "Step 3. 检测数学库" + echo + checkLinuxMathLibrary + echo + echo "Step 4. 选择要安装的PaddlePaddle版本" + echo + checkLinuxPaddleVersion + echo + echo "Step 5. 检测pip版本" + echo + checkLinuxPython + echo + checkLinuxAVX + echo + echo "Step 6.是否使用Python的虚拟环境" + use_virtualenv="--user" + checkPythonVirtualenv + echo "*********************2. 开始安装*****************************" + PipLinuxInstall + if [ "$check_virtualenv" == 'y' ];then + echo "虚环境创建成功，请cd 进入${virtualenv_name}, 执行 source bin/activate　进入虚环境。退出虚环境执行 deactivate命令。 + 更多虚环境使用方法请参考virtualenv官网:https://virtualenv.pypa.io/en/latest/" + fi +} + +function clearMacPythonEnv(){ + python_version="" + python_brief_version="" + python_root="" +} + +function checkMacPython2(){ + while true + do + python_version=`$python_root --version 2>&1 1>&1` + if [[ $? == "0" ]];then + if [ "$python_version" == "" ] || [ "$python_root" == "/usr/bin/python" -a "$python_version" == "Python 2.7.10" ];then + clearMacPythonEnv + else + check_python=`echo $python_version | grep "Python 2"` + if [[ -n "$check_python" ]];then + while true + do + echo -e " => 在您的环境中找到 \033[32m[ $python_version ]\033[0m, 确认使用此版本请输入y；如您希望自定义Python路径请输入n。请在这里输入（y/n）并回车: " + read -p "" use_python + echo + use_python=`echo $use_python | tr 'A-Z' 'a-z'` + if [[ "$use_python" == "y" ]]||[[ "$use_python" == "" ]];then + use_python="y" + break + elif [[ "$use_python" == "n" ]];then + clearMacPythonEnv + break + else + red " 输入错误，请重新输入(y/n)" + fi + done + if [[ "$use_python" == "y" ]];then + return 0 + fi + else + red " 您输入Python的不是Python2" + clearMacPythonEnv + fi + fi + else + clearMacPythonEnv + red " => 未能在常规路径下找到可用的Python2，请使用ctrl+c命令退出安装程序，并使用brew或pypi.org下载安装Python2（注意Python版本不能低于2.7.15）" + read -p " 如希望自定义Python路径，请输入路径 + 如果希望重新选择Python版本，请回车：" python_root + echo + if [[ "$python_root" == "" ]];then + python_V="" + clearMacPythonEnv + return 1 + fi + fi + done +} + +function checkMacPython3(){ + while true + do + python_version=`$python_root --version 2>&1 1>&1` + if [[ $? == "0" ]];then + if [ "$python_version" == "" ] || [ "$python_root" == "/usr/bin/python" -a "$python_version" == "Python 2.7.10" ] ;then + clearMacPythonEnv + else + check_python=`echo $python_version | grep "Python 3"` + if [[ -n "$check_python" ]];then + while true + do + echo -e " => 在您的环境中找到 \033[32m[ $python_version ]\033[0m, 确认使用此版本请输入y；如您希望自定义Python路径请输入n。请在这里输入（y/n）并回车: " + read -p "" use_python + echo + use_python=`echo $use_python | tr 'A-Z' 'a-z'` + if [[ "$use_python" == "y" ]]||[[ "$use_python" == "" ]];then + use_python="y" + break + elif [[ "$use_python" == "n" ]];then + clearMacPythonEnv + break + else + red " 输入错误，请重新输入(y/n)" + fi + done + if [[ "$use_python" == "y" ]];then + return 0 + fi + else + red " 您输入Python的不是Python3" + clearMacPythonEnv + fi + fi + else + clearMacPythonEnv + red " => 未能在常规路径下找到可用的Python3，请使用ctrl+c命令退出安装程序，并使用brew或pypi.org下载安装Python3（注意Python版本不能低于3.5.x)" + read -p " 如希望自定义Python路径，请输入路径 + 如果希望重新选择Python版本，请回车：" python_root + echo + if [[ "$python_root" == "" ]];then + python_V="" + clearMacPythonEnv + return 1 + fi + fi + done +} + +function checkMacPaddleVersion(){ + while true + do + read -n1 -p "Step 2. 选择PaddlePaddle的版本，请按回车键继续..." + echo + yellow " 1. 开发版：对应Github上develop分支，如您需要开发、或希望使用PaddlePaddle最新功能，请选用此版本" + yellow " 2. 稳定版（推荐）：如您无特殊开发需求，建议使用此版本，目前最新的版本号为 ${release_version}" + read -p " => 请输入数字1或2。如输入其他字符或直接回车，将会默认选择【 2. 稳定版 】 。请在这里输入并回车：" paddle_version + if [[ "$paddle_version" == "1" ]]||[[ "$paddle_version" == "2" ]];then + echo + yellow " 您选择了数字【"$paddle_version" 】" + echo + break + else + paddle_version="2" + echo + yellow " 您选择了数字【2】" + echo + break + fi + done +} +function initCheckMacPython2(){ + echo + yellow " 您选择了Python "$python_V"，正在寻找符合要求的Python 2版本" + echo + python_root=`which python2.7` + if [[ "$python_root" == "" ]];then + python_root=`which python` + fi + checkMacPython2 + if [[ "$?" == "1" ]];then + return 1 + else + return 0 + fi +} + +function initCheckMacPython3(){ + echo + yellow " 您选择了Python "$python_V"，正在寻找符合您要求的Python 2版本" + echo + python_root=`which python3` + checkMacPython3 + if [[ "$?" == "1" ]];then + return 1 + else + return 0 + fi +} + +function checkMacPip(){ + if [[ "$python_V" == "2" ]]||[[ "$python_V" == "3" ]];then + + python_brief_version=`$python_root -m pip -V |awk -F "[ |)]" '{print $6}'|sed 's#\.##g'` + if [[ ${python_brief_version} == "" ]];then + red "您输入的python：${python_root} 对应的pip不可用，请检查此pip或重新选择其他python" + echo + return 1 + fi + pip_version=`$python_root -m pip -V |awk -F '[ .]' '{print $2}'` + if [[ 9 -le ${pip_version} ]];then + : + else + red "您的pip版本过低，请安装pip 9.0.1及以上的版本" + echo + return 1 + fi + if [[ "$python_brief_version" == "" ]];then + clearMacPythonEnv + red "您的 $python_root 对应的pip存在问题，请按ctrl + c退出后重新安装pip，或切换其他python版本" + echo + return 1 + else + if [[ $python_brief_version == "27" ]];then + uncode=`python -c "import pip._internal;print(pip._internal.pep425tags.get_supported())"|grep "cp27"` + if [[ $uncode == "" ]];then + uncode="mu" + else + uncode="m" + fi + fi + version_list=`echo "${python_list[@]}" | grep "$python_brief_version" ` + if [[ "$version_list" != "" ]];then + return 0 + else + red "未找到可用的pip或pip3。PaddlePaddle目前支持：Python2.7/3.5/3.6/3.7及其对应的pip, 请重新输入，或使用ctrl + c退出" + echo + clearMacPythonEnv + return 1 + fi + + fi + fi +} + +function checkMacPythonVersion(){ + while true + do + read -n1 -p "Step 3. 选择Python版本，请按回车键继续..." + echo + yellow " 2. 使用python 2.x" + yellow " 3. 使用python 3.x" + read -p " => 请输入数字2或3。如输入其他字符或直接回车，将会默认使用【Python 2 】。请在这里输入并回车：" python_V + if [[ "$python_V" == "" ]];then + python_V="2" + fi + if [[ "$python_V" == "2" ]];then + initCheckMacPython2 + if [[ "$?" == "0" ]];then + checkMacPip + if [[ "$?" == "0" ]];then + return 0 + else + : + fi + else + : + fi + elif [[ "$python_V" == "3" ]];then + initCheckMacPython3 + if [[ "$?" == "0" ]];then + checkMacPip + if [[ "$?" == "0" ]];then + return 0 + else + : + fi + else + : + fi + else + red "输入错误，请重新输入" + fi + done +} + +function checkMacAVX(){ + read -n1 -p "Step 4. 检测您的Mac是否支持AVX指令集，请按回车键继续..." + if [[ $AVX != "" ]];then + AVX="avx" + echo "" + green " 检测结果：支持" + echo "" + return 0 + else + red " 检测结果：不支持。非常抱歉，PaddlePaddle在Mac系统暂不提供no_avx类型的安装包，您可以选择在Linux系统中安装no_avx版的PaddlePaddle, 请按回车键退出..." + echo + return 1 + fi +} + +function checkMacGPU(){ + read -n1 -p "Step 5. 选择CPU/GPU版本，请按回车键继续..." + echo + if [[ $GPU != "" ]];then + yellow " MacOS环境下，暂未提供GPU版本的PaddlePaddle安装包，将为您安装CPU版本的PaddlePaddle" + else + yellow " MacOS环境下，暂未提供GPU版本的PaddlePaddle安装包，将为您安装CPU版本的PaddlePaddle" + GPU=cpu + fi + echo +} + +function macos() { + path='http://paddlepaddle.org/download?url=' + AVX=`sysctl -a | grep cpu | grep AVX1.0 | tail -1 | grep AVX` + + while true + do + + checkMacPaddleVersion + + checkMacPythonVersion + + checkMacAVX + + checkMacGPU + + + green "*********************2. 开始安装*****************************" + echo + yellow "即将为您下载并安装PaddlePaddle，请按回车键继续..." + read -n1 -p "" + echo + if [[ $paddle_version == "2" ]];then + $python_root -m pip install paddlepaddle + if [[ $? == "0" ]];then + green "安装成功，可以使用: ${python_root} 来启动安装了PaddlePaddle的Python解释器" + break + else + rm $whl_cpu_release + red "未能正常安装PaddlePaddle，请尝试更换您输入的python路径，或者ctrl + c退出后请检查您使用的python对应的pip或pip源是否可用" + echo"" + echo "==========================================================================================" + echo"" + exit 1 + fi + else + if [[ -f $whl_cpu_develop ]];then + $python_root -m pip install $whl_cpu_develop + if [[ $? == "0" ]];then + rm -rf $whl_cpu_develop + # TODO add install success check here + green "安装成功！小提示：可以使用: ${python_root} 来启动安装了PaddlePaddle的Python解释器" + break + else + red "未能正常安装PaddlePaddle，请尝试更换您输入的python路径，或者ctrl + c退出后请检查您使用的python对应的pip或pip源是否可用" + echo"" + echo "==========================================================================================" + echo"" + exit 1 + fi + else + wget ${path}$whl_cpu_develop -O $whl_cpu_develop + if [[ $? == "0" ]];then + $python_root -m pip install $whl_cpu_develop + if [[ $? == "0" ]];then + rm $wheel_cpu_develop + green "安装成功，可以使用: ${python_root} 来启动安装了PaddlePaddle的Python解释器" + break + else + rm $whl_cpu_release + red "未能正常安装PaddlePaddle，请尝试更换您输入的python路径，或者ctrl + c退出后请检查您使用的python对应的pip或pip源是否可用" + echo"" + echo "==========================================================================================" + echo"" + exit 1 + fi + else + rm $whl_cpu_develop + red "未能正常安装PaddlePaddle，请检查您的网络 或者确认您是否安装有 wget，或者ctrl + c退出后反馈至https://github.com/PaddlePaddle/Paddle/issues" + echo"" + echo "==========================================================================================" + echo"" + exit 1 + fi + fi + fi + done +} + +function main() { + echo "*********************************" + green "欢迎使用PaddlePaddle快速安装脚本" + echo "*********************************" + echo + yellow "如果您在安装过程中遇到任何问题，请在https://github.com/PaddlePaddle/Paddle/issues反馈，我们的工作人员将会帮您答疑解惑" + echo + echo "本安装包将帮助您在Linux或Mac系统下安装PaddlePaddle,包括" + yellow "1）安装前的准备" + yellow "2）开始安装" + echo + read -n1 -p "请按回车键进行下一步..." + echo + echo + green "*********************1. 安装前的准备*****************************" + echo + echo "Step 1. 正在检测您的操作系统信息..." + echo + SYSTEM=`uname -s` + if [[ "$SYSTEM" == "Darwin" ]];then + yellow " 您的系统为：MAC OSX" + echo + macos + else + yellow " 您的系统为：Linux" + echo + OS=`cat /etc/issue|awk 'NR==1 {print $1}'` + if [[ $OS == "\S" ]] || [[ "$OS" == "CentOS" ]] || [[ $OS == "Ubuntu" ]];then + linux + else + red "您的系统不在本安装包的支持范围，如您需要在windows环境下安装PaddlePaddle，请您参考PaddlePaddle官网的windows安装文档" + fi + fi +} +main diff --git a/paddle/scripts/paddle_build.sh b/paddle/scripts/paddle_build.sh index c2156a436ec73d03082fa08b6250dc77b2cee19f..26b26c9b1faf7bd976b57f1320bff878f1a21770 100755 --- a/paddle/scripts/paddle_build.sh +++ b/paddle/scripts/paddle_build.sh @@ -87,7 +87,7 @@ function cmake_gen() { PYTHON_FLAGS="-DPYTHON_EXECUTABLE:FILEPATH=/Library/Frameworks/Python.framework/Versions/3.5/bin/python3 -DPYTHON_INCLUDE_DIR:PATH=/Library/Frameworks/Python.framework/Versions/3.5/include/python3.5m/ -DPYTHON_LIBRARY:FILEPATH=/Library/Frameworks/Python.framework/Versions/3.5/lib/libpython3.5m.dylib" - WITH_FLUID_ONLY=${WITH_FLUID_ONLY:-ON} + pip3.5 uninstall -y protobuf pip3.5 install --user -r ${PADDLE_ROOT}/python/requirements.txt else exit 1 @@ -100,7 +100,7 @@ function cmake_gen() { PYTHON_FLAGS="-DPYTHON_EXECUTABLE:FILEPATH=/Library/Frameworks/Python.framework/Versions/3.6/bin/python3 -DPYTHON_INCLUDE_DIR:PATH=/Library/Frameworks/Python.framework/Versions/3.6/include/python3.6m/ -DPYTHON_LIBRARY:FILEPATH=/Library/Frameworks/Python.framework/Versions/3.6/lib/libpython3.6m.dylib" - WITH_FLUID_ONLY=${WITH_FLUID_ONLY:-ON} + pip3.6 uninstall -y protobuf pip3.6 install --user -r ${PADDLE_ROOT}/python/requirements.txt else exit 1 @@ -113,7 +113,7 @@ function cmake_gen() { PYTHON_FLAGS="-DPYTHON_EXECUTABLE:FILEPATH=/Library/Frameworks/Python.framework/Versions/3.7/bin/python3 -DPYTHON_INCLUDE_DIR:PATH=/Library/Frameworks/Python.framework/Versions/3.7/include/python3.7m/ -DPYTHON_LIBRARY:FILEPATH=/Library/Frameworks/Python.framework/Versions/3.7/lib/libpython3.7m.dylib" - WITH_FLUID_ONLY=${WITH_FLUID_ONLY:-ON} + pip3.7 uninstall -y protobuf pip3.7 install --user -r ${PADDLE_ROOT}/python/requirements.txt else exit 1 @@ -128,31 +128,44 @@ function cmake_gen() { 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" + pip uninstall -y protobuf + pip install -r ${PADDLE_ROOT}/python/requirements.txt 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" + pip uninstall -y protobuf + pip install -r ${PADDLE_ROOT}/python/requirements.txt elif [ "$1" == "cp35-cp35m" ]; then export LD_LIBRARY_PATH=/opt/_internal/cpython-3.5.1/lib/:${LD_LIBRARY_PATH} export PATH=/opt/_internal/cpython-3.5.1/bin/:${PATH} export PYTHON_FLAGS="-DPYTHON_EXECUTABLE:FILEPATH=/opt/_internal/cpython-3.5.1/bin/python3 -DPYTHON_INCLUDE_DIR:PATH=/opt/_internal/cpython-3.5.1/include/python3.5m -DPYTHON_LIBRARIES:FILEPATH=/opt/_internal/cpython-3.5.1/lib/libpython3.so" + pip3.5 uninstall -y protobuf + pip3.5 install -r ${PADDLE_ROOT}/python/requirements.txt elif [ "$1" == "cp36-cp36m" ]; then export LD_LIBRARY_PATH=/opt/_internal/cpython-3.6.0/lib/:${LD_LIBRARY_PATH} export PATH=/opt/_internal/cpython-3.6.0/bin/:${PATH} export PYTHON_FLAGS="-DPYTHON_EXECUTABLE:FILEPATH=/opt/_internal/cpython-3.6.0/bin/python3 -DPYTHON_INCLUDE_DIR:PATH=/opt/_internal/cpython-3.6.0/include/python3.6m -DPYTHON_LIBRARIES:FILEPATH=/opt/_internal/cpython-3.6.0/lib/libpython3.so" + pip3.6 uninstall -y protobuf + pip3.6 install -r ${PADDLE_ROOT}/python/requirements.txt elif [ "$1" == "cp37-cp37m" ]; then export LD_LIBRARY_PATH=/opt/_internal/cpython-3.7.0/lib/:${LD_LIBRARY_PATH} export PATH=/opt/_internal/cpython-3.7.0/bin/:${PATH} export PYTHON_FLAGS="-DPYTHON_EXECUTABLE:FILEPATH=/opt/_internal/cpython-3.7.0/bin/python3.7 -DPYTHON_INCLUDE_DIR:PATH=/opt/_internal/cpython-3.7.0/include/python3.7m -DPYTHON_LIBRARIES:FILEPATH=/opt/_internal/cpython-3.7.0/lib/libpython3.so" + pip3.7 uninstall -y protobuf + pip3.7 install -r ${PADDLE_ROOT}/python/requirements.txt fi + else + pip uninstall -y protobuf + pip install -r ${PADDLE_ROOT}/python/requirements.txt fi fi @@ -173,7 +186,6 @@ function cmake_gen() { -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE:-Release} ${PYTHON_FLAGS} -DWITH_DSO=ON - -DWITH_DOC=${WITH_DOC:-OFF} -DWITH_GPU=${WITH_GPU:-OFF} -DWITH_AMD_GPU=${WITH_AMD_GPU:-OFF} -DWITH_DISTRIBUTE=${distibuted_flag} @@ -187,7 +199,6 @@ function cmake_gen() { -DWITH_TESTING=${WITH_TESTING:-ON} -DCMAKE_MODULE_PATH=/opt/rocm/hip/cmake -DCMAKE_EXPORT_COMPILE_COMMANDS=ON - -DWITH_FLUID_ONLY=${WITH_FLUID_ONLY:-OFF} -DCMAKE_EXPORT_COMPILE_COMMANDS=ON -DWITH_CONTRIB=${WITH_CONTRIB:-ON} -DWITH_INFERENCE_API_TEST=${WITH_INFERENCE_API_TEST:-ON} @@ -208,7 +219,6 @@ EOF -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE:-Release} \ ${PYTHON_FLAGS} \ -DWITH_DSO=ON \ - -DWITH_DOC=${WITH_DOC:-OFF} \ -DWITH_GPU=${WITH_GPU:-OFF} \ -DWITH_AMD_GPU=${WITH_AMD_GPU:-OFF} \ -DWITH_DISTRIBUTE=${distibuted_flag} \ @@ -221,7 +231,6 @@ EOF -DCUDNN_ROOT=/usr/ \ -DWITH_TESTING=${WITH_TESTING:-ON} \ -DCMAKE_MODULE_PATH=/opt/rocm/hip/cmake \ - -DWITH_FLUID_ONLY=${WITH_FLUID_ONLY:-OFF} \ -DCMAKE_EXPORT_COMPILE_COMMANDS=ON \ -DWITH_CONTRIB=${WITH_CONTRIB:-ON} \ -DWITH_INFERENCE_API_TEST=${WITH_INFERENCE_API_TEST:-ON} \ @@ -384,9 +393,7 @@ EOF pip3.7 install --user ${INSTALL_PREFIX:-/paddle/build}/opt/paddle/share/wheels/*.whl fi - if [[ ${WITH_FLUID_ONLY:-OFF} == "OFF" ]] ; then - paddle version - fi + paddle version if [ "$1" == "cp27-cp27m" ]; then pip uninstall -y paddlepaddle @@ -528,31 +535,6 @@ function bind_test() { wait } - -function gen_docs() { - mkdir -p ${PADDLE_ROOT}/build - cd ${PADDLE_ROOT}/build - cat <> ${PADDLE_ROOT}/build/Dockerfile <> ${PADDLE_ROOT}/build/Dockerfile <> ${PADDLE_ROOT}/build/Dockerfile < 1 and trainers_endpoints: assert self._build_strategy.num_trainers == len( trainers_endpoints), "num_trainers == len(end_points)" @@ -212,7 +223,7 @@ class CompiledProgram(object): if self._compiled: if scope and self._scope != scope: raise ValueError("Cannot compile with different scope") - if place and self._place != place: + if place and not self._place._equals(place): raise ValueError("Cannot compile with different place") return self self._compiled = True diff --git a/python/paddle/fluid/contrib/__init__.py b/python/paddle/fluid/contrib/__init__.py index 6127ca8a3eacd013dd258a02b9f3cc792b634137..870c57e54011361caae5265201d19f58830a87bc 100644 --- a/python/paddle/fluid/contrib/__init__.py +++ b/python/paddle/fluid/contrib/__init__.py @@ -22,6 +22,8 @@ from . import op_frequence from .op_frequence import * from . import quantize from .quantize import * +from . import int8_inference +from .int8_inference import * from . import reader from .reader import * from . import slim @@ -34,6 +36,7 @@ __all__ += decoder.__all__ __all__ += memory_usage_calc.__all__ __all__ += op_frequence.__all__ __all__ += quantize.__all__ +__all__ += int8_inference.__all__ __all__ += reader.__all__ __all__ += slim.__all__ __all__ += utils.__all__ diff --git a/python/paddle/fluid/contrib/decoder/beam_search_decoder.py b/python/paddle/fluid/contrib/decoder/beam_search_decoder.py index f2b7ac8375af25beed562b8279b6044f11c09d44..5854cadb58c76066ba4b48dc6b5dbca06fba8cba 100644 --- a/python/paddle/fluid/contrib/decoder/beam_search_decoder.py +++ b/python/paddle/fluid/contrib/decoder/beam_search_decoder.py @@ -22,7 +22,7 @@ This API is still under active development and may change drastically. from __future__ import print_function -import contextlib +from ...wrapped_decorator import signature_safe_contextmanager import numpy as np import six @@ -419,7 +419,7 @@ class TrainingDecoder(object): self._state_cell = state_cell self._state_cell._enter_decoder(self) - @contextlib.contextmanager + @signature_safe_contextmanager def block(self): """ Define the behavior of the decoder for each RNN time step. @@ -613,7 +613,7 @@ class BeamSearchDecoder(object): self._word_dim = word_dim self._input_var_dict = input_var_dict - @contextlib.contextmanager + @signature_safe_contextmanager def block(self): """ Define the behavior of the decoder for each RNN time step. diff --git a/python/paddle/fluid/contrib/inferencer.py b/python/paddle/fluid/contrib/inferencer.py index b8d5f4ffeadca0a7b103682f175d50dc46fa258a..4f37129234482189436ad71391f55394e2b8a277 100644 --- a/python/paddle/fluid/contrib/inferencer.py +++ b/python/paddle/fluid/contrib/inferencer.py @@ -14,7 +14,7 @@ from __future__ import print_function -import contextlib +from ..wrapped_decorator import signature_safe_contextmanager from .. import core @@ -105,7 +105,7 @@ class Inferencer(object): return results - @contextlib.contextmanager + @signature_safe_contextmanager def _prog_and_scope_guard(self): with framework.program_guard(main_program=self.inference_program): with executor.scope_guard(self.scope): diff --git a/python/paddle/fluid/contrib/int8_inference/README.md b/python/paddle/fluid/contrib/int8_inference/README.md new file mode 100644 index 0000000000000000000000000000000000000000..460ae393f158ae320c93601365a68b8cfe2ba50e --- /dev/null +++ b/python/paddle/fluid/contrib/int8_inference/README.md @@ -0,0 +1,72 @@ +# Offline INT8 Calibration Tool + +PaddlePaddle supports offline INT8 calibration to accelerate the inference speed. In this document, we provide the instructions on how to enable INT8 calibration and show the ResNet-50 and MobileNet-V1 results in accuracy. + +## 0. Prerequisite +You need to install at least PaddlePaddle-1.3 python package `pip install paddlepaddle==1.3`. + +## 1. How to generate INT8 model +You can refer to the unit test in [test_calibration.py](../tests/test_calibration.py). Basically, there are three steps: +* Construct calibration object. + +```python +calibrator = int8_utility.Calibrator( # Step 1 + program=infer_program, # required, FP32 program + pretrained_model=model_path, # required, FP32 pretrained model + algo=algo, # required, calibration algorithm; default is max, the alternative is KL (Kullback–Leibler divergence) + exe=exe, # required, executor + output=int8_model, # required, INT8 model + feed_var_names=feed_dict, # required, feed dict + fetch_list=fetch_targets) # required, fetch targets +``` + +* Call the calibrator.sample_data() after executor run. +```python +_, acc1, _ = exe.run( + program, + feed={feed_dict[0]: image, + feed_dict[1]: label}, + fetch_list=fetch_targets) + +calibrator.sample_data() # Step 2 +``` + +* Call the calibrator.save_int8_model() after sampling over specified iterations (e.g., iterations = 50) +```python +calibrator.save_int8_model() # Step 3 +``` + +## 2. How to run INT8 model +You can load INT8 model by load_inference_model [API](https://github.com/PaddlePaddle/Paddle/blob/8b50ad80ff6934512d3959947ac1e71ea3fb9ea3/python/paddle/fluid/io.py#L991) and run INT8 inference similar as [FP32](https://github.com/PaddlePaddle/models/blob/develop/fluid/PaddleCV/object_detection/eval.py "FP32"). + +```python +[infer_program, feed_dict, + fetch_targets] = fluid.io.load_inference_model(model_path, exe) +``` + +## 3. Result +We provide the results of accuracy measurd on [Intel® Xeon® Platinum Gold Processor](https://ark.intel.com/products/120489/Intel-Xeon-Gold-6148-Processor-27-5M-Cache-2-40-GHz- "Intel® Xeon® Gold 6148 Processor") (also known as Intel® Xeon® Skylake6148). + +| Model | Dataset | FP32 Accuracy | INT8 Accuracy | Accuracy Diff | +| ------------ | ------------ | ------------ | ------------ | ------------ | +| ResNet-50 | Small | 72.00% | 72.00% | 0.00% | +| MobileNet-V1 | Small | 62.00% | 62.00% | 0.00% | +| ResNet-50 | Full ImageNet Val | 76.63% | 76.17% | 0.46% | +| MobileNet-V1 | Full ImageNet Val | 70.78% | 70.49% | 0.29% | + +Please note that [Small](http://paddle-inference-dist.cdn.bcebos.com/int8/calibration_test_data.tar.gz "Small") is a subset of [full ImageNet validation dataset](http://www.image-net.org/challenges/LSVRC/2012/nnoupb/ILSVRC2012_img_val.tar "full ImageNet validation dataset"). + +Notes: +* The accuracy measurement requires the model with `label`. +* The INT8 theoretical speedup is ~1.33X on Intel® Xeon® Skylake Server (please refer to `This allows for 4x more input at the cost of 3x more instructions or 33.33% more compute` in [Reference](https://software.intel.com/en-us/articles/lower-numerical-precision-deep-learning-inference-and-training "Reference")). + +## 4. How to reproduce the results +* Small dataset +```bash +FLAGS_use_mkldnn=true python python/paddle/fluid/contrib/tests/test_calibration.py +``` + +* Full dataset +```bash +FLAGS_use_mkldnn=true DATASET=full python python/paddle/fluid/contrib/tests/test_calibration.py +``` diff --git a/python/paddle/fluid/contrib/int8_inference/__init__.py b/python/paddle/fluid/contrib/int8_inference/__init__.py index eca2dce114b069bf9b455d77ce670d73b5047fd2..45547201d598c809f7dcf3a1a09103ae5de3e4c6 100644 --- a/python/paddle/fluid/contrib/int8_inference/__init__.py +++ b/python/paddle/fluid/contrib/int8_inference/__init__.py @@ -11,3 +11,10 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. + +from __future__ import print_function + +from . import utility +from .utility import * + +__all__ = utility.__all__ diff --git a/python/paddle/fluid/contrib/int8_inference/utility.py b/python/paddle/fluid/contrib/int8_inference/utility.py index 40de038f28a83738e6e6cd8c77c0a9916ce68b4f..b35d9f2424ccf093f70e75b13e23f6c5ad59e859 100644 --- a/python/paddle/fluid/contrib/int8_inference/utility.py +++ b/python/paddle/fluid/contrib/int8_inference/utility.py @@ -11,11 +11,15 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. -import paddle.fluid.core as core + +from paddle.fluid import core import numpy as np import math import os -import paddle.fluid as fluid +from paddle.fluid.executor import global_scope +from paddle.fluid import io + +__all__ = ['Calibrator'] class Calibrator(object): @@ -76,8 +80,7 @@ class Calibrator(object): ''' for i in self.sampling_program.list_vars(): if i.name in self.sampling_vars: - np_data = np.array(fluid.global_scope().find_var(i.name) - .get_tensor()) + np_data = np.array(global_scope().find_var(i.name).get_tensor()) if i.name not in self._sampling_data: self._sampling_data[i.name] = [] self._sampling_data[i.name].append(np_data) @@ -86,9 +89,9 @@ class Calibrator(object): ''' Save the quantized model to the disk. ''' - fluid.io.save_inference_model(self.output, self.feed_var_names, - self.fetch_list, self.exe, - self.sampling_program) + io.save_inference_model(self.output, self.feed_var_names, + self.fetch_list, self.exe, + self.sampling_program) def __display_debug(self): if self.debug: diff --git a/python/paddle/fluid/contrib/slim/quantization/quantization_pass.py b/python/paddle/fluid/contrib/slim/quantization/quantization_pass.py index 266a106bc507104c0a8db1c882b55ac59e88195e..18b58e6f388bbe9495333b12f32d63b74fddcb3a 100644 --- a/python/paddle/fluid/contrib/slim/quantization/quantization_pass.py +++ b/python/paddle/fluid/contrib/slim/quantization/quantization_pass.py @@ -13,14 +13,19 @@ # limitations under the License. import collections +import numpy as np +import six +from ..... import compat as cpt from .... import core from ....framework import IrGraph from ....framework import Program -from ....framework import Variable from ....initializer import Constant from .... import unique_name -__all__ = ['QuantizationTransformPass'] +__all__ = [ + 'QuantizationTransformPass', 'QuantizationFreezePass', 'ConvertToInt8Pass', + 'TransformForMobilePass' +] class QuantizationTransformPass(object): @@ -35,7 +40,13 @@ class QuantizationTransformPass(object): """ Convert and rewrite the IrGraph according to weight and activation quantization type. + Args: + scope(fluid.Scope): When activation use 'range_abs_max' as the quantize + type, this pass will create some new parameters. The scope is used to + initialize these new parameters. + program_exe(fluid.Executor): program_exe is used to initialize new + parameters described above. weight_bits (int): quantization bit number for weights, the bias is not quantized. activation_bits (int): quantization bit number for activation. @@ -49,6 +60,7 @@ class QuantizationTransformPass(object): support 'abs_max'. The 'range_abs_max' usually is not used for weight, since weights are fixed once the model is well trained. window_size (int): the window size for 'range_abs_max' quantization. + Examples: .. code-block:: python # The original graph will be rewrite. @@ -88,31 +100,35 @@ class QuantizationTransformPass(object): self._quantizable_grad_ops = [ '%s_grad' % (op) for op in self._quantizable_ops ] - self._fake_quant_op_types = [ - 'fake_quantize_abs_max', 'fake_quantize_range_abs_max' - ] - self._fake_dequant_op_types = ['fake_dequantize_max_abs'] self._is_test = None self._global_step = None def apply(self, graph): + """ + Quantize the graph for training process. According to weight and + activation quantization type, the graph will be added some fake + quantize operators and fake dequantize operators. + + Args: + graph(IrGraph): the applied graph. + """ assert isinstance(graph, IrGraph), 'graph must be the instance of IrGraph.' self._need_initialized.clear() self._is_test = graph.is_test() # marked the variable which has been dequantized. dequantized_vars = collections.OrderedDict() - params = [p.name() for p in graph.all_parameters()] + persistable_vars = [p.name() for p in graph.all_persistable_vars()] def _transform_forward(graph, op): for var_node in op.inputs: if var_node.name() in dequantized_vars: dequant_var_node = dequantized_vars[var_node.name()] else: - quant_bits = self._weight_bits if var_node.name() in params \ + quant_bits = self._weight_bits if var_node.name() in persistable_vars \ else self._activation_bits quant_type = self._weight_quantize_type if var_node.name() \ - in params else self._activation_quantize_type + in persistable_vars else self._activation_quantize_type quant_var_node, scale_var_node = self._insert_quant_op( graph, var_node, quant_bits, quant_type) dequant_var_node = self._insert_dequant_op( @@ -150,9 +166,14 @@ class QuantizationTransformPass(object): assert self._program_exe is not None, \ 'The program_exe cannot be set None when activation_quantize_type equals to range_abs_max.' init_program = Program() - for var_desc, initializer in self._need_initialized.iteritems(): - var = Variable(init_program.global_block()) - var._set_desc(var_desc) + for var_desc, initializer in six.iteritems(self._need_initialized): + var = init_program.global_block().create_var( + name=var_desc.name(), + shape=var_desc.shape(), + dtype=var_desc.dtype(), + type=var_desc.type(), + lod_level=var_desc.lod_level(), + persistable=var_desc.persistable()) initializer(var, init_program.global_block()) self._program_exe.run(program=init_program, scope=self._scope) @@ -161,7 +182,7 @@ class QuantizationTransformPass(object): def _create_global_step(self, graph): if self._weight_quantize_type == 'range_abs_max' or \ self._activation_quantize_type == 'range_abs_max': - counter_name = '@STEP_COUNTER@' + counter_name = cpt.to_text('@STEP_COUNTER@') for node in graph.all_vars(): if node.name() == counter_name: self._global_step = node @@ -175,9 +196,14 @@ class QuantizationTransformPass(object): Constant(value=0, force_cpu=True) global_step_out = graph.create_var_node_from_desc( global_step_in.var()) + # The attribute of `op_role` is needed by ParallelExecutor. increment_op = graph.create_op_node( op_type='increment', - attrs={'step': 1.0}, + attrs={ + 'step': 1.0, + 'op_role': + core.op_proto_and_checker_maker.OpRole.Forward + }, inputs={'X': global_step_in}, outputs={'Out': global_step_out}) graph.link_to(global_step_in, increment_op) @@ -212,7 +238,10 @@ class QuantizationTransformPass(object): var_dtype=var_node.var().dtype()) quant_op_node = graph.create_op_node( op_type='fake_quantize_abs_max', - attrs={'bit_length': quant_bits}, + attrs={ + 'bit_length': quant_bits, + 'op_role': core.op_proto_and_checker_maker.OpRole.Forward + }, inputs={'X': var_node}, outputs={'Out': quant_var_node, 'OutScale': scale_var_node}) @@ -257,7 +286,8 @@ class QuantizationTransformPass(object): attrs = { 'window_size': self._window_size, 'bit_length': quant_bits, - 'is_test': self._is_test + 'is_test': self._is_test, + 'op_role': core.op_proto_and_checker_maker.OpRole.Forward } quant_op_node = graph.create_op_node( op_type='fake_quantize_range_abs_max', @@ -290,7 +320,10 @@ class QuantizationTransformPass(object): max_range = (1 << (quant_bits - 1)) - 1 dequant_op_node = graph.create_op_node( op_type='fake_dequantize_max_abs', - attrs={'max_range': float(max_range)}, + attrs={ + 'max_range': float(max_range), + 'op_role': core.op_proto_and_checker_maker.OpRole.Forward + }, inputs={'X': var_node, 'Scale': scale_var_node}, outputs={'Out': dequant_var_node}) @@ -316,3 +349,330 @@ class QuantizationTransformPass(object): Return the scale name of quantized variable for the input `var_name`. """ return "%s.scale" % (var_name) + + +class QuantizationFreezePass(object): + """ + The freeze pass is used to adjust the quantize operator order, for example: + 1) `activation -> quant -> dequant -> conv2d` will be freezed into + `activation -> quant -> conv2d -> dequant` + 2) `weight -> quant -> dequant -> conv2d` will be freezed into `weight -> conv2d`, + and weight will be sacled offline. + + Args: + scope(fluid.Scope): scope is used to get the weight tensor values. + place(fluid.CPUPlace|fluid.CUDAPlace): place is used to restore the weight tensors. + weight_bits (int): quantization bit number for weights. + activation_bits (int): quantization bit number for activation. + weight_quantize_type (str): quantization type for weights, support 'abs_max'. + The 'range_abs_max' usually is not used for weight, since weights are fixed once the + model is well trained. + """ + + def __init__(self, + scope, + place, + weight_bits=8, + activation_bits=8, + weight_quantize_type='abs_max'): + assert scope is not None, \ + 'The scope cannot be set None.' + assert place is not None, \ + 'The place cannot be set None.' + self._scope = scope + self._place = place + self._weight_bits = weight_bits + self._activation_bits = activation_bits + self._weight_quantize_type = weight_quantize_type + self._quantizable_ops = ['conv2d', 'depthwise_conv2d', 'mul'] + self._fake_quant_op_names = [ + 'fake_quantize_abs_max', 'fake_quantize_range_abs_max' + ] + self._fake_dequant_op_names = ['fake_dequantize_max_abs'] + self._op_input_rename_map = collections.OrderedDict() + self._op_output_rename_map = collections.OrderedDict() + self._var_scale_map = collections.OrderedDict() + + def apply(self, graph): + """ + Adjust quantize/dequantize operators order for the inference process. + + Args: + graph(IrGraph): the applied graph. + """ + persistable_vars = [p.name() for p in graph.all_persistable_vars()] + ops = graph.all_ops() + for op_node in ops: + op_name = op_node.name() + if op_name in self._fake_quant_op_names: + input_arg_name = op_node.op().input('X')[0] + if input_arg_name in persistable_vars: + if self._weight_quantize_type == 'abs_max': + param = self._load_var(input_arg_name) + scale_v = np.max(np.abs(param)) + else: + scale_v = self._load_var(op_node.op().output('OutScale') + [0])[0] + self._var_scale_map[input_arg_name] = scale_v + else: + scale_v = graph.var_node(op_node.op().output('OutScale')[0]) + self._var_scale_map[input_arg_name] = scale_v + if input_arg_name in persistable_vars: + self._remove_fake_quant_and_dequant_op(graph, op_node) + # quantize weight and restore + param_v = self._load_var(input_arg_name) + quantized_param_v = self._quant(param_v, scale_v, + self._weight_bits) + self._restore_var(input_arg_name, quantized_param_v) + + ops = graph.all_ops() + for op_node in ops: + op_name = op_node.name() + if op_name in self._fake_dequant_op_names: + self._remove_fake_quant_and_dequant_op(graph, op_node) + + ops = graph.all_ops() + for op_node in ops: + op_name = op_node.name() + if op_name in self._quantizable_ops: + self._insert_post_dequant_op(graph, op_node) + + for op_node in ops: + # insert dequant_op after fc/conv, need to rename inputs of the followed ops + for var_node in op_node.inputs: + name = var_node.name() + if name in self._op_output_rename_map: + old_in = graph.var_node(name) + new_in = self._op_output_rename_map[name] + graph.update_input_link(old_in, new_in, op_node) + + # remove the unused var node in the graph + self._remove_unused_var_nodes(graph) + return graph + + def _remove_fake_quant_and_dequant_op(self, graph, op_node): + k = op_node.op().output('Out')[0] + v = op_node.op().input('X')[0] + if v not in self._op_input_rename_map: + self._op_input_rename_map[k] = v + else: + self._op_input_rename_map[k] = self._op_input_rename_map[v] + graph.safe_remove_nodes(op_node) + + def _insert_post_dequant_op(self, graph, op_node): + max_range = None + scale_var_node = None + persistable_vars = [p.name() for p in graph.all_persistable_vars()] + for var_node in op_node.inputs: + name = var_node.name() + if name in self._op_input_rename_map: + old_in = graph.var_node(name) + new_in = graph.var_node(self._op_input_rename_map[name]) + new_in.clear_outputs() + graph.update_input_link(old_in, new_in, op_node) + original_var_name = self._original_var_name(name) + scale_v = self._var_scale_map[original_var_name] + if original_var_name in persistable_vars: + param_range = (1 << (self._weight_bits - 1)) - 1 + act_range = (1 << (self._activation_bits - 1)) - 1 + assert self._is_float( + scale_v), 'The scale of parameter %s is not a float.' % ( + original_var_name) + max_range = param_range * act_range / scale_v + else: + assert isinstance(scale_v, core.Node) + scale_var_node = self._var_scale_map[original_var_name] + + if len(op_node.outputs) != 1: + raise ValueError("Only support one output, but op %s has" + " more than one output." % (op_node.name())) + + output_var_node = op_node.outputs[0] + dequant_var_node = graph.create_var_node( + name=self._dequantized_var_name(output_var_node.name()), + var_type=output_var_node.var().type(), + shape=output_var_node.var().shape(), + var_dtype=output_var_node.var().dtype()) + dequant_op_node = graph.create_op_node( + op_type='fake_dequantize_max_abs', + attrs={ + 'max_range': float(max_range), + 'op_role': core.op_proto_and_checker_maker.OpRole.Forward + }, + inputs={'X': output_var_node, + 'Scale': scale_var_node}, + outputs={'Out': dequant_var_node}) + graph.link_to(output_var_node, dequant_op_node) + graph.link_to(scale_var_node, dequant_op_node) + graph.link_to(dequant_op_node, dequant_var_node) + self._op_output_rename_map[output_var_node.name()] = dequant_var_node + return dequant_var_node + + def _load_var(self, name): + return np.array(self._scope.find_var(name).get_tensor()) + + def _restore_var(self, name, array): + tensor = self._scope.find_var(name).get_tensor() + tensor.set(array, self._place) + + def _remove_unused_var_nodes(self, graph): + all_used_vars = set() + ops = graph.all_ops() + for op_node in ops: + for input_node in op_node.inputs: + all_used_vars.add(input_node) + for output_node in op_node.outputs: + all_used_vars.add(output_node) + + all_unused_vars = graph.all_vars() - all_used_vars + graph.safe_remove_nodes(all_unused_vars) + + def _original_var_name(self, var_name): + """ + Return the original variable name. + """ + if var_name.endswith('.quantized.dequantized'): + return var_name[:-len('.quantized.dequantized')] + if var_name.endswith('.quantized'): + return var_name[:-len('.quantized')] + if var_name.endswith('.dequantized'): + return var_name[:-len('.dequantized')] + if var_name.endswith('.scale'): + return var_name[:-len('.scale')] + else: + return var_name + + def _dequantized_var_name(self, var_name): + """ + Return dequantized variable name for the input `var_name`. + """ + return "%s.dequantized" % (var_name) + + def _is_float(self, v): + return isinstance(v, float) or isinstance(v, np.float32) \ + or isinstance(v, np.float64) + + def _quant(self, x, scale, num_bits): + return np.round(x / scale * ((1 << (num_bits - 1)) - 1)) + + +class ConvertToInt8Pass(object): + """ + Convert the weights into int8_t type. + + Args: + scope(fluid.Scope): scope is used to get the weight tensor values. + place(fluid.CPUPlace|fluid.CUDAPlace): place is used to restore the + 8bits weight tensors. + """ + + def __init__(self, scope, place): + assert scope is not None, \ + 'The scope cannot be set None.' + assert place is not None, \ + 'The place cannot be set None.' + self._scope = scope + self._place = place + self._quantizable_ops = ['conv2d', 'depthwise_conv2d', 'mul'] + + def apply(self, graph): + """ + Convert weights' tpye of the graph. After that, the data type of the + graph weigths is int8_t. + + Args: + graph(IrGraph): the applied graph. + """ + persistable_vars = [p.name() for p in graph.all_persistable_vars()] + ops = graph.all_ops() + input_map = {} + for op_node in ops: + op_name = op_node.name() + if op_name in self._quantizable_ops: + for var_node in op_node.inputs: + name = var_node.name() + if name in persistable_vars: + if name not in input_map: + int8_var_node = self._convert_to_int8(graph, + var_node) + input_map[name] = int8_var_node + graph.update_input_link(var_node, input_map[name], + op_node) + + # remove the unused var node in the graph + self._remove_unused_var_nodes(graph) + return graph + + def _convert_to_int8(self, graph, var_node): + int8_var_node_name = var_node.name() + ".int8" + int8_var_node = graph.create_param_node( + name=cpt.to_text(int8_var_node_name), + var_type=var_node.var().type(), + shape=var_node.var().shape(), + var_dtype=core.VarDesc.VarType.INT8) + array = self._load_var(var_node.name()) + self._scope.var(int8_var_node_name) + self._store_var(int8_var_node_name, array, np.int8) + return int8_var_node + + def _load_var(self, name): + return np.array(self._scope.find_var(name).get_tensor()) + + def _store_var(self, name, array, dtype): + tensor = self._scope.find_var(name).get_tensor() + tensor.set(array.astype(dtype), self._place) + + def _remove_unused_var_nodes(self, graph): + all_used_vars = set() + ops = graph.all_ops() + for op_node in ops: + for input_node in op_node.inputs: + all_used_vars.add(input_node) + for output_node in op_node.outputs: + all_used_vars.add(output_node) + + all_unused_vars = graph.all_vars() - all_used_vars + graph.safe_remove_nodes(all_unused_vars) + + +class TransformForMobilePass(object): + """ + This pass is used to convert the freezed graph for paddle-mobile execution. + """ + + def __init__(self): + self._fake_quant_op_names = [ + 'fake_quantize_abs_max', 'fake_quantize_range_abs_max' + ] + self._fake_dequant_op_names = ['fake_dequantize_max_abs'] + + def apply(self, graph): + """ + Because paddle-mobile use `quantize` an `dequantize` as the names of + quantize operator and dequantize operator, the `apply` function just + realize this logic. + + Args: + graph(IrGraph): the graph will be transformed. + """ + ops = graph.all_ops() + for op_node in ops: + name = op_node.name() + if name in self._fake_quant_op_names: + op_node.op().set_type('quantize') + quant_node = graph.create_op_node_from_desc(op_node.op()) + for input_node in op_node.inputs: + graph.link_to(input_node, quant_node) + for output_node in op_node.outputs: + graph.link_to(quant_node, output_node) + graph.safe_remove_nodes(op_node) + if name in self._fake_dequant_op_names: + op_node.op().set_type('dequantize') + dequant_node = graph.create_op_node_from_desc(op_node.op()) + for input_node in op_node.inputs: + graph.link_to(input_node, dequant_node) + for output_node in op_node.outputs: + graph.link_to(dequant_node, output_node) + graph.safe_remove_nodes(op_node) + + return graph diff --git a/python/paddle/fluid/contrib/slim/tests/CMakeLists.txt b/python/paddle/fluid/contrib/slim/tests/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..79bec8c4ad34d682895250bc29b1fddb3a569bd4 --- /dev/null +++ b/python/paddle/fluid/contrib/slim/tests/CMakeLists.txt @@ -0,0 +1,6 @@ +file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py") +string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}") + +foreach(src ${TEST_OPS}) + py_test(${src} SRCS ${src}.py) +endforeach() diff --git a/python/paddle/fluid/contrib/slim/unitest/__init__.py b/python/paddle/fluid/contrib/slim/tests/__init__.py similarity index 100% rename from python/paddle/fluid/contrib/slim/unitest/__init__.py rename to python/paddle/fluid/contrib/slim/tests/__init__.py diff --git a/python/paddle/fluid/contrib/slim/unitest/configs/config.yaml b/python/paddle/fluid/contrib/slim/tests/configs/config.yaml similarity index 88% rename from python/paddle/fluid/contrib/slim/unitest/configs/config.yaml rename to python/paddle/fluid/contrib/slim/tests/configs/config.yaml index db488b96330210df15b02b19d90abd5c9101f844..d9b49029d3e34d487ad65fe0f7e54e2cee1d5838 100644 --- a/python/paddle/fluid/contrib/slim/unitest/configs/config.yaml +++ b/python/paddle/fluid/contrib/slim/tests/configs/config.yaml @@ -1,5 +1,5 @@ version: 1.0 -include: ["./unitest/configs/pruners.yaml", "./unitest/configs/pruners_0.yaml"] +include: ["./configs/pruners.yaml", "./configs/pruners_0.yaml"] pruners: pruner_1: class: 'RatioPruner' diff --git a/python/paddle/fluid/contrib/slim/unitest/configs/pruners.yaml b/python/paddle/fluid/contrib/slim/tests/configs/pruners.yaml similarity index 100% rename from python/paddle/fluid/contrib/slim/unitest/configs/pruners.yaml rename to python/paddle/fluid/contrib/slim/tests/configs/pruners.yaml diff --git a/python/paddle/fluid/contrib/slim/unitest/configs/pruners_0.yaml b/python/paddle/fluid/contrib/slim/tests/configs/pruners_0.yaml similarity index 100% rename from python/paddle/fluid/contrib/slim/unitest/configs/pruners_0.yaml rename to python/paddle/fluid/contrib/slim/tests/configs/pruners_0.yaml diff --git a/python/paddle/fluid/contrib/slim/unitest/test_factory.py b/python/paddle/fluid/contrib/slim/tests/test_factory.py similarity index 95% rename from python/paddle/fluid/contrib/slim/unitest/test_factory.py rename to python/paddle/fluid/contrib/slim/tests/test_factory.py index 07f28aac905d1a2813dbde6143235c7916fd9278..2fc72b6475e6bdd977dafb57696046a1100d0087 100644 --- a/python/paddle/fluid/contrib/slim/unitest/test_factory.py +++ b/python/paddle/fluid/contrib/slim/tests/test_factory.py @@ -18,7 +18,7 @@ import unittest class TestFactory(unittest.TestCase): def test_parse(self): - factory = ConfigFactory('./unitest/configs/config.yaml') + factory = ConfigFactory('./configs/config.yaml') pruner = factory.instance('pruner_1') self.assertEquals(pruner.ratios['conv1_1.w'], 0.3) diff --git a/python/paddle/fluid/contrib/slim/tests/test_graph.py b/python/paddle/fluid/contrib/slim/tests/test_graph.py new file mode 100644 index 0000000000000000000000000000000000000000..75e0c95b5c3cc06d66eab9de0b85e5d7ed110837 --- /dev/null +++ b/python/paddle/fluid/contrib/slim/tests/test_graph.py @@ -0,0 +1,80 @@ +# copyright (c) 2018 paddlepaddle authors. all rights reserved. +# +# licensed under the apache license, version 2.0 (the "license"); +# you may not use this file except in compliance with the license. +# you may obtain a copy of the license at +# +# http://www.apache.org/licenses/license-2.0 +# +# unless required by applicable law or agreed to in writing, software +# distributed under the license is distributed on an "as is" basis, +# without warranties or conditions of any kind, either express or implied. +# see the license for the specific language governing permissions and +# limitations under the license. + +from __future__ import print_function +import unittest +import paddle.fluid as fluid +import six +from paddle.fluid.framework import IrGraph +from paddle.fluid import core + + +def residual_block(num): + def conv_bn_layer(input, + ch_out, + filter_size, + stride, + padding, + act='relu', + bias_attr=False): + tmp = fluid.layers.conv2d( + input=input, + filter_size=filter_size, + num_filters=ch_out, + stride=stride, + padding=padding, + act=None, + bias_attr=bias_attr) + return fluid.layers.batch_norm(input=tmp, act=act) + + data = fluid.layers.data(name='image', shape=[1, 32, 32], dtype='float32') + label = fluid.layers.data(name='label', shape=[1], dtype='int64') + hidden = data + for _ in six.moves.xrange(num): + conv = conv_bn_layer(hidden, 16, 3, 1, 1, act=None, bias_attr=True) + short = conv_bn_layer(hidden, 16, 1, 1, 0, act=None) + hidden = fluid.layers.elementwise_add(x=conv, y=short, act='relu') + fc = fluid.layers.fc(input=hidden, size=10) + loss = fluid.layers.cross_entropy(input=fc, label=label) + loss = fluid.layers.mean(loss) + return loss + + +class TestGraph(unittest.TestCase): + def test_graph_functions(self): + main = fluid.Program() + startup = fluid.Program() + with fluid.program_guard(main, startup): + loss = residual_block(2) + opt = fluid.optimizer.Adam(learning_rate=0.001) + opt.minimize(loss) + graph = IrGraph(core.Graph(main.desc), for_test=False) + marked_nodes = set() + for op in graph.all_ops(): + if op.name().find('conv2d') > -1: + marked_nodes.add(op) + graph.draw('.', 'residual', marked_nodes) + self.assertFalse(graph.has_circle()) + self.assertEqual(graph.graph_num(), 1) + nodes = graph.topology_sort() + self.assertEqual(len(nodes), len(graph.all_ops())) + nodes_map = graph.build_adjacency_list() + self.assertEqual(len(nodes_map), len(graph.all_ops())) + nodes_num = len(graph.all_nodes()) + graph.safe_remove_nodes(marked_nodes) + self.assertEqual(len(graph.all_nodes()), nodes_num - len(marked_nodes)) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/contrib/slim/tests/test_quantization_pass.py b/python/paddle/fluid/contrib/slim/tests/test_quantization_pass.py new file mode 100644 index 0000000000000000000000000000000000000000..2f291132f3049af21420f863972792c1a862b9ad --- /dev/null +++ b/python/paddle/fluid/contrib/slim/tests/test_quantization_pass.py @@ -0,0 +1,372 @@ +# copyright (c) 2018 paddlepaddle authors. all rights reserved. +# +# licensed under the apache license, version 2.0 (the "license"); +# you may not use this file except in compliance with the license. +# you may obtain a copy of the license at +# +# http://www.apache.org/licenses/license-2.0 +# +# unless required by applicable law or agreed to in writing, software +# distributed under the license is distributed on an "as is" basis, +# without warranties or conditions of any kind, either express or implied. +# see the license for the specific language governing permissions and +# limitations under the license. + +import unittest +import random +import numpy as np +import paddle.fluid as fluid +import six +import paddle +from paddle.fluid.framework import IrGraph +from paddle.fluid.contrib.slim.quantization import QuantizationTransformPass +from paddle.fluid.contrib.slim.quantization import QuantizationFreezePass +from paddle.fluid.contrib.slim.quantization import ConvertToInt8Pass +from paddle.fluid.contrib.slim.quantization import TransformForMobilePass +from paddle.fluid import core + + +def linear_fc(num): + data = fluid.layers.data(name='image', shape=[1, 32, 32], dtype='float32') + label = fluid.layers.data(name='label', shape=[1], dtype='int64') + hidden = data + for _ in six.moves.xrange(num): + hidden = fluid.layers.fc(hidden, size=128, act='relu') + loss = fluid.layers.cross_entropy(input=hidden, label=label) + loss = fluid.layers.mean(loss) + return loss + + +def residual_block(num): + def conv_bn_layer(input, + ch_out, + filter_size, + stride, + padding, + act='relu', + bias_attr=False): + tmp = fluid.layers.conv2d( + input=input, + filter_size=filter_size, + num_filters=ch_out, + stride=stride, + padding=padding, + act=None, + bias_attr=bias_attr) + return fluid.layers.batch_norm(input=tmp, act=act) + + data = fluid.layers.data(name='image', shape=[1, 32, 32], dtype='float32') + label = fluid.layers.data(name='label', shape=[1], dtype='int64') + hidden = data + for _ in six.moves.xrange(num): + conv = conv_bn_layer(hidden, 16, 3, 1, 1, act=None, bias_attr=True) + short = conv_bn_layer(hidden, 16, 1, 1, 0, act=None) + hidden = fluid.layers.elementwise_add(x=conv, y=short, act='relu') + fc = fluid.layers.fc(input=hidden, size=10) + loss = fluid.layers.cross_entropy(input=fc, label=label) + loss = fluid.layers.mean(loss) + return loss + + +def conv_net(img, label): + conv_pool_1 = fluid.nets.simple_img_conv_pool( + input=img, + filter_size=5, + num_filters=20, + pool_size=2, + pool_stride=2, + act="relu") + conv_pool_1 = fluid.layers.batch_norm(conv_pool_1) + conv_pool_2 = fluid.nets.simple_img_conv_pool( + input=conv_pool_1, + filter_size=5, + num_filters=50, + pool_size=2, + pool_stride=2, + act="relu") + prediction = fluid.layers.fc(input=conv_pool_2, size=10, act='softmax') + loss = fluid.layers.cross_entropy(input=prediction, label=label) + avg_loss = fluid.layers.mean(loss) + return avg_loss + + +class TestQuantizationTransformPass(unittest.TestCase): + def setUp(self): + self.quantizable_op_and_inputs = { + 'conv2d': ['Input', 'Filter'], + 'depthwise_conv2d': ['Input', 'Filter'], + 'mul': ['X', 'Y'] + } + self.quantizable_grad_op_inputs = { + 'conv2d_grad': ['Input', 'Filter'], + 'depthwise_conv2d_grad': ['Input', 'Filter'], + 'mul_grad': ['X', 'Y'] + } + + def check_program(self, transform_pass, program): + quantized_ops = set() + for block in program.blocks: + for op in block.ops: + # check forward + if op.type in self.quantizable_op_and_inputs: + for arg_name in op.input_arg_names: + self.assertTrue( + arg_name.endswith('.quantized.dequantized')) + quantized_ops.add(arg_name) + + for op in block.ops: + # check backward + if op.type in self.quantizable_grad_op_inputs: + for pname in self.quantizable_grad_op_inputs[op.type]: + arg_name = op.input(pname)[0] + self.assertTrue( + arg_name.endswith('.quantized.dequantized')) + self.assertTrue(arg_name in quantized_ops) + + def linear_fc_quant(self, quant_type): + main = fluid.Program() + startup = fluid.Program() + with fluid.program_guard(main, startup): + loss = linear_fc(3) + opt = fluid.optimizer.Adam(learning_rate=0.001) + opt.minimize(loss) + exe = fluid.Executor(fluid.CPUPlace()) + graph = IrGraph(core.Graph(main.desc), for_test=False) + transform_pass = QuantizationTransformPass( + scope=fluid.global_scope(), + program_exe=exe, + activation_quantize_type=quant_type) + transform_pass.apply(graph) + marked_nodes = set() + for op in graph.all_ops(): + if op.name().find('quantize') > -1: + marked_nodes.add(op) + graph.draw('.', 'quantize_fc_' + quant_type, marked_nodes) + program = graph.to_program() + self.check_program(transform_pass, program) + val_graph = IrGraph(core.Graph(program.desc), for_test=False) + val_marked_nodes = set() + for op in val_graph.all_ops(): + if op.name().find('quantize') > -1: + val_marked_nodes.add(op) + val_graph.draw('.', 'val_fc_' + quant_type, val_marked_nodes) + + def test_linear_fc_quant_abs_max(self): + self.act_quant_op_type = 'fake_quantize_abs_max' + self.linear_fc_quant('abs_max') + + def test_linear_fc_quant_range_abs_max(self): + self.act_quant_op_type = 'fake_quantize_range_abs_max' + self.linear_fc_quant('range_abs_max') + + def residual_block_quant(self, quant_type): + main = fluid.Program() + startup = fluid.Program() + with fluid.program_guard(main, startup): + loss = residual_block(2) + opt = fluid.optimizer.Adam(learning_rate=0.001) + opt.minimize(loss) + exe = fluid.Executor(fluid.CPUPlace()) + graph = IrGraph(core.Graph(main.desc), for_test=False) + transform_pass = QuantizationTransformPass( + scope=fluid.global_scope(), + program_exe=exe, + activation_quantize_type=quant_type) + transform_pass.apply(graph) + marked_nodes = set() + for op in graph.all_ops(): + if op.name().find('quantize') > -1: + marked_nodes.add(op) + graph.draw('.', 'quantize_residual_' + quant_type, marked_nodes) + program = graph.to_program() + self.check_program(transform_pass, program) + val_graph = IrGraph(core.Graph(program.desc), for_test=False) + val_marked_nodes = set() + for op in val_graph.all_ops(): + if op.name().find('quantize') > -1: + val_marked_nodes.add(op) + val_graph.draw('.', 'val_residual_' + quant_type, val_marked_nodes) + + def test_residual_block_abs_max(self): + self.act_quant_op_type = 'fake_quantize_abs_max' + self.residual_block_quant('abs_max') + + def test_residual_block_range_abs_max(self): + self.act_quant_op_type = 'fake_quantize_range_abs_max' + self.residual_block_quant('range_abs_max') + + +class TestQuantizationFreezePass(unittest.TestCase): + def freeze_graph(self, use_cuda, seed, quant_type): + def build_program(main, startup, is_test): + main.random_seed = seed + startup.random_seed = seed + with fluid.unique_name.guard(): + with fluid.program_guard(main, startup): + img = fluid.layers.data( + name='image', shape=[1, 28, 28], dtype='float32') + label = fluid.layers.data( + name='label', shape=[1], dtype='int64') + loss = conv_net(img, label) + if not is_test: + opt = fluid.optimizer.Adam(learning_rate=0.001) + opt.minimize(loss) + return [img, label], loss + + random.seed(0) + np.random.seed(0) + + main = fluid.Program() + startup = fluid.Program() + test_program = fluid.Program() + feeds, loss = build_program(main, startup, False) + build_program(test_program, startup, True) + test_program = test_program.clone(for_test=True) + main_graph = IrGraph(core.Graph(main.desc), for_test=False) + test_graph = IrGraph(core.Graph(test_program.desc), for_test=True) + + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + exe = fluid.Executor(place) + scope = fluid.Scope() + with fluid.scope_guard(scope): + exe.run(startup) + transform_pass = QuantizationTransformPass( + scope=scope, program_exe=exe, activation_quantize_type=quant_type) + transform_pass.apply(main_graph) + transform_pass.apply(test_graph) + dev_name = '_gpu_' if use_cuda else '_cpu_' + marked_nodes = set() + for op in main_graph.all_ops(): + if op.name().find('quantize') > -1: + marked_nodes.add(op) + main_graph.draw('.', 'main' + dev_name + quant_type, marked_nodes) + marked_nodes = set() + for op in test_graph.all_ops(): + if op.name().find('quantize') > -1: + marked_nodes.add(op) + test_graph.draw('.', 'test' + dev_name + quant_type, marked_nodes) + + quantized_main_program = main_graph.to_program() + quantized_test_program = test_graph.to_program() + iters = 5 + batch_size = 8 + + #train_exe = fluid.ParallelExecutor( + # main_program=quantized_main_program, + # use_cuda=bool(use_cuda), + # loss_name=loss.name, + # scope=scope) + train_reader = paddle.batch( + paddle.reader.shuffle( + paddle.dataset.mnist.train(), buf_size=500), + batch_size=batch_size) + test_reader = paddle.batch( + paddle.dataset.mnist.test(), batch_size=batch_size) + feeder = fluid.DataFeeder(feed_list=feeds, place=place) + with fluid.scope_guard(scope): + for _ in range(iters): + data = next(train_reader()) + loss_v = exe.run(program=quantized_main_program, + feed=feeder.feed(data), + fetch_list=[loss]) + #loss_v = train_exe.run(feed=feeder.feed(data), + # fetch_list=[loss.name]) + #print('{}: {}'.format('loss' + dev_name + quant_type, loss_v)) + + test_data = next(test_reader()) + with fluid.program_guard(quantized_test_program): + w_var = fluid.framework._get_var('conv2d_1.w_0.quantized', + quantized_test_program) + # Testing + with fluid.scope_guard(scope): + test_loss1, w_quant = exe.run(program=quantized_test_program, + feed=feeder.feed(test_data), + fetch_list=[loss, w_var]) + + # Freeze graph for inference, but the weight of fc/conv is still float type. + freeze_pass = QuantizationFreezePass(scope=scope, place=place) + freeze_pass.apply(test_graph) + marked_nodes = set() + for op in test_graph.all_ops(): + if op.name().find('quantize') > -1: + marked_nodes.add(op) + test_graph.draw('.', 'test_freeze' + dev_name + quant_type, + marked_nodes) + + server_program = test_graph.to_program() + with fluid.scope_guard(scope): + test_loss2, = exe.run(program=server_program, + feed=feeder.feed(test_data), + fetch_list=[loss]) + self.assertAlmostEqual(test_loss1, test_loss2, delta=5e-3) + #print('{}: {}'.format('test_loss1' + dev_name + quant_type, test_loss1)) + #print('{}: {}'.format('test_loss2' + dev_name + quant_type, test_loss2)) + w_freeze = np.array(scope.find_var('conv2d_1.w_0').get_tensor()) + # Maybe failed, this is due to the calculation precision + # self.assertAlmostEqual(np.sum(w_freeze), np.sum(w_quant)) + #print('{}: {}'.format('w_freeze' + dev_name + quant_type, + # np.sum(w_freeze))) + #print('{}: {}'.format('w_quant' + dev_name + quant_type, + # np.sum(w_quant))) + + # Convert parameter to 8-bit. + convert_int8_pass = ConvertToInt8Pass(scope=scope, place=place) + convert_int8_pass.apply(test_graph) + marked_nodes = set() + for op in test_graph.all_ops(): + if op.name().find('quantize') > -1: + marked_nodes.add(op) + test_graph.draw('.', 'test_int8' + dev_name + quant_type, marked_nodes) + server_program_int8 = test_graph.to_program() + # Save the 8-bit parameter and model file. + with fluid.scope_guard(scope): + fluid.io.save_inference_model('server_int8' + dev_name + quant_type, + ['image', 'label'], [loss], exe, + server_program_int8) + # Test whether the 8-bit parameter and model file can be loaded successfully. + [infer, feed, fetch] = fluid.io.load_inference_model( + 'server_int8' + dev_name + quant_type, exe) + # Check the loaded 8-bit weight. + w_8bit = np.array(scope.find_var('conv2d_1.w_0.int8').get_tensor()) + self.assertEqual(w_8bit.dtype, np.int8) + self.assertEqual(np.sum(w_8bit), np.sum(w_freeze)) + #print('{}: {}'.format('w_8bit' + dev_name + quant_type, np.sum(w_8bit))) + #print('{}: {}'.format('w_freeze' + dev_name + quant_type, + # np.sum(w_freeze))) + + mobile_pass = TransformForMobilePass() + mobile_pass.apply(test_graph) + marked_nodes = set() + for op in test_graph.all_ops(): + if op.name().find('quantize') > -1: + marked_nodes.add(op) + test_graph.draw('.', 'test_mobile' + dev_name + quant_type, + marked_nodes) + + mobile_program = test_graph.to_program() + with fluid.scope_guard(scope): + fluid.io.save_inference_model('mobile_int8' + dev_name + quant_type, + ['image', 'label'], [loss], exe, + mobile_program) + + def test_freeze_graph_cuda_dynamic(self): + if fluid.core.is_compiled_with_cuda(): + with fluid.unique_name.guard(): + self.freeze_graph(True, seed=1, quant_type='abs_max') + + def test_freeze_graph_cpu_dynamic(self): + with fluid.unique_name.guard(): + self.freeze_graph(False, seed=2, quant_type='abs_max') + + def test_freeze_graph_cuda_static(self): + if fluid.core.is_compiled_with_cuda(): + with fluid.unique_name.guard(): + self.freeze_graph(True, seed=1, quant_type='range_abs_max') + + def test_freeze_graph_cpu_static(self): + with fluid.unique_name.guard(): + self.freeze_graph(False, seed=2, quant_type='range_abs_max') + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/contrib/slim/unitest/test_quantization_pass.py b/python/paddle/fluid/contrib/slim/unitest/test_quantization_pass.py deleted file mode 100644 index 1bd4b95d6b90b7f16d507061190f0b463f6c4cc5..0000000000000000000000000000000000000000 --- a/python/paddle/fluid/contrib/slim/unitest/test_quantization_pass.py +++ /dev/null @@ -1,175 +0,0 @@ -# copyright (c) 2018 paddlepaddle authors. all rights reserved. -# -# licensed under the apache license, version 2.0 (the "license"); -# you may not use this file except in compliance with the license. -# you may obtain a copy of the license at -# -# http://www.apache.org/licenses/license-2.0 -# -# unless required by applicable law or agreed to in writing, software -# distributed under the license is distributed on an "as is" basis, -# without warranties or conditions of any kind, either express or implied. -# see the license for the specific language governing permissions and -# limitations under the license. - -import unittest -import random -import numpy as np -import paddle.fluid as fluid -import six -from paddle.fluid.framework import Program -from paddle.fluid.framework import IrGraph -from paddle.fluid.contrib.slim.quantization import QuantizationTransformPass -from paddle.fluid import core - - -def linear_fc(num): - data = fluid.layers.data(name='image', shape=[1, 32, 32], dtype='float32') - label = fluid.layers.data(name='label', shape=[1], dtype='int64') - hidden = data - for _ in six.moves.xrange(num): - hidden = fluid.layers.fc(hidden, size=128, act='relu') - loss = fluid.layers.cross_entropy(input=hidden, label=label) - loss = fluid.layers.mean(loss) - return loss - - -def residual_block(num): - def conv_bn_layer(input, - ch_out, - filter_size, - stride, - padding, - act='relu', - bias_attr=False): - tmp = fluid.layers.conv2d( - input=input, - filter_size=filter_size, - num_filters=ch_out, - stride=stride, - padding=padding, - act=None, - bias_attr=bias_attr) - return fluid.layers.batch_norm(input=tmp, act=act) - - data = fluid.layers.data(name='image', shape=[1, 32, 32], dtype='float32') - label = fluid.layers.data(name='label', shape=[1], dtype='int64') - hidden = data - for _ in six.moves.xrange(num): - conv = conv_bn_layer(hidden, 16, 3, 1, 1, act=None, bias_attr=True) - short = conv_bn_layer(hidden, 16, 1, 1, 0, act=None) - hidden = fluid.layers.elementwise_add(x=conv, y=short, act='relu') - fc = fluid.layers.fc(input=hidden, size=10) - loss = fluid.layers.cross_entropy(input=fc, label=label) - loss = fluid.layers.mean(loss) - return loss - - -class TestQuantizationTransformPass(unittest.TestCase): - def setUp(self): - self.quantizable_op_and_inputs = { - 'conv2d': ['Input', 'Filter'], - 'depthwise_conv2d': ['Input', 'Filter'], - 'mul': ['X', 'Y'] - } - self.quantizable_grad_op_inputs = { - 'conv2d_grad': ['Input', 'Filter'], - 'depthwise_conv2d_grad': ['Input', 'Filter'], - 'mul_grad': ['X', 'Y'] - } - - def check_program(self, transform_pass, program): - quantized_ops = set() - for block in program.blocks: - for op in block.ops: - # check forward - if op.type in self.quantizable_op_and_inputs: - for arg_name in op.input_arg_names: - self.assertTrue( - arg_name.endswith('.quantized.dequantized')) - quantized_ops.add(arg_name) - - for op in block.ops: - # check backward - if op.type in self.quantizable_grad_op_inputs: - for pname in self.quantizable_grad_op_inputs[op.type]: - arg_name = op.input(pname)[0] - self.assertTrue( - arg_name.endswith('.quantized.dequantized')) - self.assertTrue(arg_name in quantized_ops) - - def linear_fc_quant(self, quant_type): - main = fluid.Program() - startup = fluid.Program() - with fluid.program_guard(main, startup): - loss = linear_fc(3) - opt = fluid.optimizer.Adam(learning_rate=0.001) - opt.minimize(loss) - exe = fluid.Executor(fluid.CPUPlace()) - graph = IrGraph(core.Graph(main.desc), for_test=False) - transform_pass = QuantizationTransformPass( - scope=fluid.global_scope(), - program_exe=exe, - activation_quantize_type=quant_type) - transform_pass.apply(graph) - marked_nodes = set() - for op in graph.all_ops(): - if op.name().find('quantize') > -1: - marked_nodes.add(op) - graph.draw('.', 'quantize_fc_' + quant_type, marked_nodes) - program = graph.to_program() - self.check_program(transform_pass, program) - val_graph = IrGraph(core.Graph(program.desc), for_test=False) - val_marked_nodes = set() - for op in val_graph.all_ops(): - if op.name().find('quantize') > -1: - val_marked_nodes.add(op) - val_graph.draw('.', 'val_fc_' + quant_type, val_marked_nodes) - - def test_linear_fc_quant_abs_max(self): - self.act_quant_op_type = 'fake_quantize_abs_max' - self.linear_fc_quant('abs_max') - - def test_linear_fc_quant_range_abs_max(self): - self.act_quant_op_type = 'fake_quantize_range_abs_max' - self.linear_fc_quant('range_abs_max') - - def residual_block_quant(self, quant_type): - main = fluid.Program() - startup = fluid.Program() - with fluid.program_guard(main, startup): - loss = residual_block(2) - opt = fluid.optimizer.Adam(learning_rate=0.001) - opt.minimize(loss) - exe = fluid.Executor(fluid.CPUPlace()) - graph = IrGraph(core.Graph(main.desc), for_test=False) - transform_pass = QuantizationTransformPass( - scope=fluid.global_scope(), - program_exe=exe, - activation_quantize_type=quant_type) - transform_pass.apply(graph) - marked_nodes = set() - for op in graph.all_ops(): - if op.name().find('quantize') > -1: - marked_nodes.add(op) - graph.draw('.', 'quantize_residual_' + quant_type, marked_nodes) - program = graph.to_program() - self.check_program(transform_pass, program) - val_graph = IrGraph(core.Graph(program.desc), for_test=False) - val_marked_nodes = set() - for op in val_graph.all_ops(): - if op.name().find('quantize') > -1: - val_marked_nodes.add(op) - val_graph.draw('.', 'val_residual_' + quant_type, val_marked_nodes) - - def test_residual_block_abs_max(self): - self.act_quant_op_type = 'fake_quantize_abs_max' - self.residual_block_quant('abs_max') - - def test_residual_block_range_abs_max(self): - self.act_quant_op_type = 'fake_quantize_range_abs_max' - self.residual_block_quant('range_abs_max') - - -if __name__ == '__main__': - unittest.main() diff --git a/python/paddle/fluid/contrib/tests/CMakeLists.txt b/python/paddle/fluid/contrib/tests/CMakeLists.txt index 81aee1233d1db756686d1a934b94672dc5c770fe..a2c59416467e5dbe66f058666633807eb0e45047 100644 --- a/python/paddle/fluid/contrib/tests/CMakeLists.txt +++ b/python/paddle/fluid/contrib/tests/CMakeLists.txt @@ -6,5 +6,9 @@ if(APPLE OR WIN32 OR NOT WITH_MKL) endif() foreach(src ${TEST_OPS}) - py_test(${src} SRCS ${src}.py) + if(src MATCHES "test_calibration") + py_test(${src} SRCS ${src}.py ENVS FLAGS_use_mkldnn=true) + else() + py_test(${src} SRCS ${src}.py) + endif() endforeach() diff --git a/python/paddle/fluid/contrib/tests/test_calibration.py b/python/paddle/fluid/contrib/tests/test_calibration.py index ed5ea70260878ee27606dfcd0880e442b48525c1..b9f938bebed71dc9611df8d743a066858ea38bca 100644 --- a/python/paddle/fluid/contrib/tests/test_calibration.py +++ b/python/paddle/fluid/contrib/tests/test_calibration.py @@ -19,14 +19,12 @@ import sys import random import paddle import paddle.fluid as fluid -import argparse import functools import contextlib -import paddle.fluid.profiler as profiler +from paddle.dataset.common import download from PIL import Image, ImageEnhance import math -sys.path.append('..') -import int8_inference.utility as int8_utility +import paddle.fluid.contrib.int8_inference.utility as int8_utility random.seed(0) np.random.seed(0) @@ -42,7 +40,7 @@ img_mean = np.array([0.485, 0.456, 0.406]).reshape((3, 1, 1)) img_std = np.array([0.229, 0.224, 0.225]).reshape((3, 1, 1)) -# TODO(guomingz): Remove duplicated code from line 45 ~ line 114 +# TODO(guomingz): Remove duplicated code from resize_short, crop_image, process_image, _reader_creator def resize_short(img, target_size): percent = float(target_size) / min(img.size[0], img.size[1]) resized_width = int(round(img.size[0] * percent)) @@ -116,30 +114,91 @@ def val(data_dir=DATA_DIR): return _reader_creator(file_list, 'val', shuffle=False, data_dir=data_dir) -class TestCalibration(unittest.TestCase): +class TestCalibrationForResnet50(unittest.TestCase): def setUp(self): - # TODO(guomingz): Put the download process in the cmake. - # Download and unzip test data set - imagenet_dl_url = 'http://paddle-inference-dist.cdn.bcebos.com/int8/calibration_test_data.tar.gz' - zip_file_name = imagenet_dl_url.split('/')[-1] - cmd = 'rm -rf data {} && mkdir data && wget {} && tar xvf {} -C data'.format( - zip_file_name, imagenet_dl_url, zip_file_name) - os.system(cmd) - # resnet50 fp32 data - resnet50_fp32_model_url = 'http://paddle-inference-dist.cdn.bcebos.com/int8/resnet50_int8_model.tar.gz' - resnet50_zip_name = resnet50_fp32_model_url.split('/')[-1] - resnet50_unzip_folder_name = 'resnet50_fp32' - cmd = 'rm -rf {} {} && mkdir {} && wget {} && tar xvf {} -C {}'.format( - resnet50_unzip_folder_name, resnet50_zip_name, - resnet50_unzip_folder_name, resnet50_fp32_model_url, - resnet50_zip_name, resnet50_unzip_folder_name) + self.int8_download = 'int8/download' + self.cache_folder = os.path.expanduser('~/.cache/paddle/dataset/' + + self.int8_download) + + data_urls = [] + data_md5s = [] + self.data_cache_folder = '' + if os.environ.get('DATASET') == 'full': + data_urls.append( + 'https://paddle-inference-dist.bj.bcebos.com/int8/ILSVRC2012_img_val.tar.gz.partaa' + ) + data_md5s.append('60f6525b0e1d127f345641d75d41f0a8') + data_urls.append( + 'https://paddle-inference-dist.bj.bcebos.com/int8/ILSVRC2012_img_val.tar.gz.partab' + ) + data_md5s.append('1e9f15f64e015e58d6f9ec3210ed18b5') + self.data_cache_folder = self.download_data(data_urls, data_md5s, + "full_data", False) + else: + data_urls.append( + 'http://paddle-inference-dist.cdn.bcebos.com/int8/calibration_test_data.tar.gz' + ) + data_md5s.append('1b6c1c434172cca1bf9ba1e4d7a3157d') + self.data_cache_folder = self.download_data(data_urls, data_md5s, + "small_data", False) + + # reader/decorator.py requires the relative path to the data folder + cmd = 'rm -rf {0} && ln -s {1} {0}'.format("data", + self.data_cache_folder) os.system(cmd) - self.iterations = 50 + self.batch_size = 1 + self.sample_iterations = 50 + self.infer_iterations = 50000 if os.environ.get( + 'DATASET') == 'full' else 50 + + def cache_unzipping(self, target_folder, zip_path): + if not os.path.exists(target_folder): + cmd = 'mkdir {0} && tar xf {1} -C {0}'.format(target_folder, + zip_path) + os.system(cmd) + + def download_data(self, data_urls, data_md5s, folder_name, is_model=True): + data_cache_folder = os.path.join(self.cache_folder, folder_name) + zip_path = '' + if os.environ.get('DATASET') == 'full': + file_names = [] + for i in range(0, len(data_urls)): + download(data_urls[i], self.int8_download, data_md5s[i]) + file_names.append(data_urls[i].split('/')[-1]) + + zip_path = os.path.join(self.cache_folder, + 'full_imagenet_val.tar.gz') + if not os.path.exists(zip_path): + cat_command = 'cat' + for file_name in file_names: + cat_command += ' ' + os.path.join(self.cache_folder, + file_name) + cat_command += ' > ' + zip_path + os.system(cat_command) + + if os.environ.get('DATASET') != 'full' or is_model: + download(data_urls[0], self.int8_download, data_md5s[0]) + file_name = data_urls[0].split('/')[-1] + zip_path = os.path.join(self.cache_folder, file_name) + + print('Data is downloaded at {0}').format(zip_path) + self.cache_unzipping(data_cache_folder, zip_path) + return data_cache_folder + + def download_model(self): + # resnet50 fp32 data + data_urls = [ + 'http://paddle-inference-dist.cdn.bcebos.com/int8/resnet50_int8_model.tar.gz' + ] + data_md5s = ['4a5194524823d9b76da6e738e1367881'] + self.model_cache_folder = self.download_data(data_urls, data_md5s, + "resnet50_fp32") + self.model = "ResNet-50" + self.algo = "direct" def run_program(self, model_path, generate_int8=False, algo='direct'): image_shape = [3, 224, 224] - os.environ['FLAGS_use_mkldnn'] = 'True' fluid.memory_optimize(fluid.default_main_program()) @@ -151,17 +210,17 @@ class TestCalibration(unittest.TestCase): t = fluid.transpiler.InferenceTranspiler() t.transpile(infer_program, fluid.CPUPlace()) - val_reader = paddle.batch(val(), batch_size=1) + val_reader = paddle.batch(val(), self.batch_size) + iterations = self.infer_iterations if generate_int8: int8_model = os.path.join(os.getcwd(), "calibration_out") + iterations = self.sample_iterations if os.path.exists(int8_model): os.system("rm -rf " + int8_model) os.system("mkdir " + int8_model) - print("Start calibration ...") - calibrator = int8_utility.Calibrator( program=infer_program, pretrained_model=model_path, @@ -173,6 +232,7 @@ class TestCalibration(unittest.TestCase): test_info = [] cnt = 0 + periods = [] for batch_id, data in enumerate(val_reader()): image = np.array( [x[0].reshape(image_shape) for x in data]).astype("float32") @@ -180,41 +240,80 @@ class TestCalibration(unittest.TestCase): label = label.reshape([-1, 1]) running_program = calibrator.sampling_program.clone( ) if generate_int8 else infer_program.clone() - for op in running_program.current_block().ops: - if op.has_attr("use_mkldnn"): - op._set_attr("use_mkldnn", True) + t1 = time.time() _, acc1, _ = exe.run( running_program, feed={feed_dict[0]: image, feed_dict[1]: label}, fetch_list=fetch_targets) + t2 = time.time() + period = t2 - t1 + periods.append(period) + if generate_int8: calibrator.sample_data() test_info.append(np.mean(acc1) * len(data)) cnt += len(data) - if batch_id != self.iterations - 1: - continue + if (batch_id + 1) % 100 == 0: + print("{0} images,".format(batch_id + 1)) + sys.stdout.flush() - break + if (batch_id + 1) == iterations: + break if generate_int8: calibrator.save_int8_model() print( - "Calibration is done and the corresponding files were generated at {}". + "Calibration is done and the corresponding files are generated at {}". format(os.path.abspath("calibration_out"))) else: - return np.sum(test_info) / cnt - - def test_calibration_for_resnet50(self): - fp32_acc1 = self.run_program("resnet50_fp32/model") - self.run_program("resnet50_fp32/model", True) - int8_acc1 = self.run_program("calibration_out") + throughput = cnt / np.sum(periods) + latency = np.average(periods) + acc1 = np.sum(test_info) / cnt + return (throughput, latency, acc1) + + def test_calibration(self): + self.download_model() + print("Start FP32 inference for {0} on {1} images ...").format( + self.model, self.infer_iterations) + (fp32_throughput, fp32_latency, + fp32_acc1) = self.run_program(self.model_cache_folder + "/model") + print("Start INT8 calibration for {0} on {1} images ...").format( + self.model, self.sample_iterations) + self.run_program( + self.model_cache_folder + "/model", True, algo=self.algo) + print("Start INT8 inference for {0} on {1} images ...").format( + self.model, self.infer_iterations) + (int8_throughput, int8_latency, + int8_acc1) = self.run_program("calibration_out") delta_value = np.abs(fp32_acc1 - int8_acc1) self.assertLess(delta_value, 0.01) + print( + "FP32 {0}: batch_size {1}, throughput {2} images/second, latency {3} second, accuracy {4}". + format(self.model, self.batch_size, fp32_throughput, fp32_latency, + fp32_acc1)) + print( + "INT8 {0}: batch_size {1}, throughput {2} images/second, latency {3} second, accuracy {4}". + format(self.model, self.batch_size, int8_throughput, int8_latency, + int8_acc1)) + sys.stdout.flush() + + +class TestCalibrationForMobilenetv1(TestCalibrationForResnet50): + def download_model(self): + # mobilenetv1 fp32 data + data_urls = [ + 'http://paddle-inference-dist.cdn.bcebos.com/int8/mobilenetv1_int8_model.tar.gz' + ] + data_md5s = ['13892b0716d26443a8cdea15b3c6438b'] + self.model_cache_folder = self.download_data(data_urls, data_md5s, + "mobilenetv1_fp32") + self.model = "MobileNet-V1" + self.algo = "KL" if __name__ == '__main__': diff --git a/python/paddle/fluid/contrib/tests/test_quantize_transpiler.py b/python/paddle/fluid/contrib/tests/test_quantize_transpiler.py index 86fa84ad4bd7a55fb27f4e43128f0bfda6dfe6db..77fdf0087b93c3ad44a2492de68f8f57ce243ef3 100644 --- a/python/paddle/fluid/contrib/tests/test_quantize_transpiler.py +++ b/python/paddle/fluid/contrib/tests/test_quantize_transpiler.py @@ -204,9 +204,11 @@ class TestQuantizeTranspiler(unittest.TestCase): build_program(test_program, startup, True) test_program = test_program.clone(for_test=True) - quant_transpiler = QuantizeTranspiler() - quant_transpiler.training_transpile(main) - quant_transpiler.training_transpile(test_program) + quant_type = 'range_abs_max' # 'range_abs_max' or 'abs_max' + quant_transpiler = QuantizeTranspiler( + activation_quantize_type=quant_type) + quant_transpiler.training_transpile(main, startup) + quant_transpiler.training_transpile(test_program, startup) place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() exe = fluid.Executor(place) diff --git a/python/paddle/fluid/contrib/trainer.py b/python/paddle/fluid/contrib/trainer.py index 8569e486f91786b5562e84dcdccf6d91da0612cc..d27b808438d53a004db4e85345a68c35d00fff98 100644 --- a/python/paddle/fluid/contrib/trainer.py +++ b/python/paddle/fluid/contrib/trainer.py @@ -14,7 +14,7 @@ from __future__ import print_function -import contextlib +from ..wrapped_decorator import signature_safe_contextmanager import os import errno import shutil @@ -453,7 +453,7 @@ class Trainer(object): io.save_inference_model(param_path, feeded_var_names, target_vars, exe) - @contextlib.contextmanager + @signature_safe_contextmanager def _prog_and_scope_guard(self): with framework.program_guard( main_program=self.train_program, diff --git a/python/paddle/fluid/executor.py b/python/paddle/fluid/executor.py index d3ff14a17955990bff851e95bd61fbc370ea7aa5..8815911eaeb36067987c0490d7a4f3e909789499 100644 --- a/python/paddle/fluid/executor.py +++ b/python/paddle/fluid/executor.py @@ -17,7 +17,7 @@ from __future__ import print_function import os import multiprocessing import numpy as np -import contextlib +from .wrapped_decorator import signature_safe_contextmanager import six from .framework import Program, default_main_program, Variable from . import core @@ -49,7 +49,7 @@ def _switch_scope(scope): return ex -@contextlib.contextmanager +@signature_safe_contextmanager def scope_guard(scope): """ Change the global/default scope instance by Python `with` statement. All diff --git a/python/paddle/fluid/framework.py b/python/paddle/fluid/framework.py index 2bdae60db347b3d42fded138a20a505486e48dbc..f01b6ab09437a70835e84d66ab59709039b6f54f 100644 --- a/python/paddle/fluid/framework.py +++ b/python/paddle/fluid/framework.py @@ -16,7 +16,9 @@ from __future__ import print_function import collections from collections import defaultdict +from collections import Iterable import contextlib +from .wrapped_decorator import signature_safe_contextmanager import os import re import traceback @@ -85,6 +87,15 @@ def _current_expected_place(): return _imperative_current_expected_place_ +def is_pserver_mode(main_program): + main = main_program if main_program \ + else default_main_program() + for op in main.global_block().ops: + if op.type in ["send", "recv"]: + return True + return False + + class NameScope(object): def __init__(self, name="", parent=None): self._children = dict() @@ -111,7 +122,7 @@ class NameScope(object): _name_scope = NameScope() -@contextlib.contextmanager +@signature_safe_contextmanager def name_scope(prefix=None): """ Generate hierarchical name prefix for the operators. @@ -445,11 +456,16 @@ class Variable(object): @property def _stop_gradient(self): - return self._ivar.stop_gradient + if _in_imperative_mode(): + return self._ivar.stop_gradient + else: + return self.stop_gradient @_stop_gradient.setter def _stop_gradient(self, s): - self._ivar.stop_gradient = s + if _in_imperative_mode(): + self._ivar.stop_gradient = s + self.stop_gradient = s @property def persistable(self): @@ -550,7 +566,8 @@ class OpProtoHolder(object): return { core.op_proto_and_checker_maker.kOpRoleAttrName(), core.op_proto_and_checker_maker.kOpRoleVarAttrName(), - core.op_proto_and_checker_maker.kOpNameScopeAttrName() + core.op_proto_and_checker_maker.kOpNameScopeAttrName(), + core.op_proto_and_checker_maker.kOpCreationCallstackAttrName() } @@ -1310,6 +1327,9 @@ class Block(object): outputs=kwargs.get("outputs", None), attrs=kwargs.get("attrs", None)) self.ops.append(op) + + # TODO(minqiyang): add stop_gradient support in static mode too. + # currently, we only support stop_gradient in imperative mode. self._trace_op(op, kwargs.get("stop_gradient", False)) return op @@ -1521,12 +1541,16 @@ class Block(object): class IrGraph(object): """ - IrGraph uses core.Graph as the delegation to accomplish the manipulation. + Python IrGraph. Beneath it is a core.Graph, which is used for + create a c++ Ir Pass Graph. An IrGraph is just a graph view of + a Program. In an IrGraph, both Variables and Operators are graph + nodes. """ def __init__(self, graph, for_test=False): """ - Construct the IrGraph using core.Graph. + Construct an IrGraph using core.Graph. + Args: graph(core.Graph): C++ Graph. for_test(bool): True for the test graph and false for the train graph. @@ -1537,23 +1561,81 @@ class IrGraph(object): self._for_test = for_test def is_test(self): + """ + If the graph is used for testing, the function returns true. Otherwise, returns false. + """ return self._for_test - def all_parameters(self): - param_nodes = set() - for node in self.graph.nodes(): - if node.is_var() and node.var() is not None and node.var( - ).persistable(): - param_nodes.add(node) - return param_nodes + def all_nodes(self): + """ + Return all nodes included in the graph as a set. + """ + return {node for node in self.graph.nodes()} def all_vars(self): + """ + Return all variable nodes included in the graph as a set. + """ return {node for node in self.graph.nodes() if node.is_var()} + def all_persistable_vars(self): + """ + Return all persistable variable nodes included in the graph as a set. + """ + persistable_nodes = set() + for node in self.graph.nodes(): + if node.is_var() and node.var() is not None and node.var( + ).persistable(): + persistable_nodes.add(node) + return persistable_nodes + def all_ops(self): + """ + Return all operator nodes included in the graph as a set. + """ return {node for node in self.graph.nodes() if node.is_op()} + def var_node(self, name): + """ + Get a variable node by name from the graph. + + Args: + name(str): the name of the variable node. + + Raises: + ValueError: The If input's type is not str, or this graph + doesn't have a variable with the giving name. + + Returns: + core.Node: the variable node with the giving name. + """ + if not isinstance(name, six.string_types): + raise TypeError( + "var require string as parameter, but get %s instead." % + (type(name))) + target_var_node = None + var_nodes = self.all_vars() + for var_node in var_nodes: + if var_node.name() == name: + target_var_node = var_node + if target_var_node is None: + raise ValueError("var_node %s not in this graph" % name) + return target_var_node + def create_param_node(self, name, var_type, shape, var_dtype): + """ + Create a persistable variable node in the graph. In IrGraph, + it can not distinguish between persistable variables and parameters. + + Args: + name(str): the name of the persistable variable node. + vart_type(core.VarDesc.VarType): the type of the persistable variable node. + shape(list): the shape of the persistable variable node. + var_dtype(core.VarDesc.VarType): the data type of the persistable variable node. + + Returns: + core.Node: the created persistable variable node. + """ var_desc = core.VarDesc(name) var_desc.set_type(var_type) var_desc.set_shape(shape) @@ -1562,6 +1644,20 @@ class IrGraph(object): return self.graph.create_var_node(var_desc) def create_var_node(self, name, var_type, shape, var_dtype): + """ + Create a variable node in the graph. The created variable node is + not persistable. + + Args: + name(str): the name of the variable node. + vart_type(core.VarDesc.VarType): the type of the variable node. + shape(list): the shape of the variable node. + var_dtype(core.VarDesc.VarType): the data type of the variable node. + + Returns: + core.Node: the created variable node. + """ + var_desc = core.VarDesc(name) var_desc.set_type(var_type) var_desc.set_shape(shape) @@ -1569,19 +1665,41 @@ class IrGraph(object): return self.graph.create_var_node(var_desc) def create_var_node_from_desc(self, var_desc): + """ + Create a variable node by using an existing VarDesc in the graph. + Depend on the giving VarDesc, the created variable node may be persistable. + + Args: + var_desc(core.VarDesc): the giving variable description. + + Returns: + core.Node: the created variable node. + """ return self.graph.create_var_node(var_desc) def create_op_node(self, op_type, attrs, inputs, outputs): + """ + Create a operator node in the graph. + + Args: + op_type(str): the type of the operator node. + attrs(dict): the attributes of the operator node. + inputs(dict): the inputs of the operator node. + outputs(dict): the outpus of the operator node. + + Returns: + core.Node: the created operator node. + """ op_desc = core.OpDesc() op_desc.set_type(op_type) - for attr, value in attrs.iteritems(): + for attr, value in six.iteritems(attrs): self._update_desc_attr(op_desc, attr, value) - for input_name, var_nodes in inputs.iteritems(): + for input_name, var_nodes in six.iteritems(inputs): if not isinstance(var_nodes, list): var_nodes = [var_nodes] op_desc.set_input(input_name, [var_node.name() for var_node in var_nodes]) - for output_name, var_nodes in outputs.iteritems(): + for output_name, var_nodes in six.iteritems(outputs): if not isinstance(var_nodes, list): var_nodes = [var_nodes] op_desc.set_output(output_name, @@ -1589,11 +1707,29 @@ class IrGraph(object): return self.graph.create_op_node(op_desc) def create_op_node_from_desc(self, op_desc): + """ + Create a operator node by using an existing OpDesc in the graph. + + Args: + op_desc(core.VarDesc): the giving operator description. + + Returns: + core.Node: the created operator node. + """ return self.graph.create_op_node(op_desc) def update_input_link(self, old_input_node, new_input_node, op_node): - assert old_input_node in self.graph.nodes() and new_input_node in self.graph.nodes() and \ - op_node in self.graph.nodes(), 'Th three arguments must be in the graph nodes.' + """ + Update the input's link of a operator node. + + Args: + old_input_node(core.Node): the old input node of the giving op_node. + new_input_node(core.Node): the new input node of the giving op_node. + op_node(core.Node): the operator node that is needed to update input's link. + """ + assert old_input_node in self.graph.nodes() and new_input_node in \ + self.graph.nodes() and op_node in self.graph.nodes(), \ + 'The three arguments(old_input_node&new_input_node&op_node) must be in the graph nodes.' old_input_node.outputs_remove(op_node) op_node.inputs_remove(old_input_node) new_input_node.outputs_append(op_node) @@ -1601,17 +1737,85 @@ class IrGraph(object): op_node.op()._rename_input(old_input_node.name(), new_input_node.name()) def link_to(self, node_in, node_out): + """ + Connect two nodes. + + Args: + node_in(core.Node): the input node. + node_out(core.Node): the output node. + """ assert node_in in self.graph.nodes() and node_out in self.graph.nodes(), \ - 'Th two arguments must be in the graph nodes.' + 'The two arguments(node_in&node_out) must be in the graph nodes.' node_in.outputs_append(node_out) node_out.inputs_append(node_in) def safe_remove_nodes(self, remove_nodes): + """ + Remove nodes safely since links connected to these removed nodes are + also removed. + + Args: + remove_nodes(set): the nodes prepared to be removed. + """ if not isinstance(remove_nodes, set): - remove_nodes = set(remove_nodes) + if isinstance(remove_nodes, Iterable): + remove_nodes = set(remove_nodes) + else: + remove_nodes = {remove_nodes} core.graph_safe_remove_nodes(self.graph, remove_nodes) - def draw(self, save_path, name, marked_nodes=None): + def has_circle(self): + """ + Check if the graph has a circle. + + Returns: + bool: True if the graph has a circle else False. + """ + return core.has_circle(self.graph) + + def graph_num(self): + """ + Count the number of unconnected graphs in this graph. + + Returns: + int: the number of unconnected graphs. + """ + return core.graph_num(self.graph) + + def topology_sort(self): + """ + Perform the topology sort operation on the graph. + + Notes: the `graph` cannot contain a circle. + + Returns: + set(core.Node): nodes in topology order. + """ + return core.topology_sort(self.graph) + + def build_adjacency_list(self): + """ + Build an adjacency list of operations for the `graph`. + + Returns: + dict{core.Node: set(core.Node)}: the adjacency list. + """ + return core.build_adjacency_list(self.graph) + + def draw(self, save_path, name, marked_nodes=None, remove_ctr_var=True): + """ + Draw the graph. If `dot` command is installed, the drawn graph + will be saved as pdf file type, otherwise dot file type is used. + + Args: + save_path(str): the save path of drawn graph. + name(str): the name of drawn graph. + marked_nodes(set(core.Node)): nodes that are needed to be marked. + Default value is None. + remove_ctr_var(bool): If it is set True, all control variable nodes + in the graph will be removed. Default value is True. + """ + def _convert_to_pdf(dot_file_path): pdf_save_path = os.path.splitext(dot_file_path)[0] + '.pdf' exited_code = subprocess.call('dot -Tpdf ' + dot_file_path \ @@ -1621,15 +1825,17 @@ class IrGraph(object): print('The {} is saved as the dot filetype.'.format( dot_file_path)) - remove_ctr_vars = set() + if remove_ctr_var: + remove_ctr_vars = set() + for node in self.graph.nodes(): + if node.is_ctrl_var(): + remove_ctr_vars.add(node) + self.safe_remove_nodes(remove_ctr_vars) ops_num = 0 for node in self.graph.nodes(): - if node.is_ctrl_var(): - remove_ctr_vars.add(node) - elif node.is_op(): + if node.is_op(): ops_num += 1 print('Total ops num = {}.'.format(ops_num)) - self.safe_remove_nodes(remove_ctr_vars) if marked_nodes is not None: if not isinstance(marked_nodes, set): marked_nodes = set(marked_nodes) @@ -1644,10 +1850,20 @@ class IrGraph(object): _convert_to_pdf(viz_dot_path) def to_program(self): + """ + Convert the graph into a Program. + + Notes: When the graph includes backward operator nodes, the + conversion process may be failed. Usually, this function is + only used to convert a test graph. + + Returns: + Program: a program converted from the graph. + """ convert_pass = core.get_pass('graph_to_program_pass') - convert_pass.set('program', Program().desc) + desc = core.ProgramDesc() + convert_pass.set_not_owned('program', desc) convert_pass.apply(self.graph) - desc = convert_pass.get_program('program') program = Program._construct_from_desc(desc) return program @@ -1717,6 +1933,19 @@ class Program(object): self._trainers_endpoints = [] # the distributed lookup table names self._distributed_lookup_table = None + # @deprecated(the python memory optimize transpiler is deprecated) + # whether the program is optimized by memory_optimize_transpiler + self.__is_mem_optimized = False + + @property + def _is_mem_optimized(self): + # if the program is optimized, operator input/outputs + # maybe same, which conflict with save_inference_model. + return self.__is_mem_optimized + + @_is_mem_optimized.setter + def _is_mem_optimized(self, target): + self.__is_mem_optimized = target @property def op_role(self): @@ -1736,7 +1965,7 @@ class Program(object): return self._current_role @op_role.setter - def set_op_role(self, role): + def op_role(self, role): self._current_role = role @property @@ -1754,7 +1983,7 @@ class Program(object): def set_op_role_var(self, var_name): self._op_role_var = [var_name] - @contextlib.contextmanager + @signature_safe_contextmanager def _optimized_guard(self, param_and_grads): """ A with guard to set :code:`Optimization` :code:`OpRole` and @@ -1784,7 +2013,7 @@ class Program(object): self._op_role_var = tmp_var self._current_role = tmp_role - @contextlib.contextmanager + @signature_safe_contextmanager def _lr_schedule_guard(self, is_with_opt=False): """ A with guard to set :code:`LRSched` :code:`OpRole` and @@ -2438,7 +2667,7 @@ def switch_startup_program(program): return prev_program -@contextlib.contextmanager +@signature_safe_contextmanager def program_guard(main_program, startup_program=None): """ Change the global main program and startup program with `with` statement. @@ -2503,7 +2732,7 @@ def _get_var(name, program=None): return program.global_block().var(name) -@contextlib.contextmanager +@signature_safe_contextmanager def _imperative_guard(tracer): global _imperative_tracer_ tmp_trace = _imperative_tracer_ @@ -2514,7 +2743,7 @@ def _imperative_guard(tracer): _imperative_tracer_ = tmp_trace -@contextlib.contextmanager +@signature_safe_contextmanager def _imperative_place_guard(place): global _imperative_current_expected_place_ tmp_place = _imperative_current_expected_place_ diff --git a/python/paddle/fluid/imperative/base.py b/python/paddle/fluid/imperative/base.py index ff3984b11f42cf9e6ff49c8654c600c065effe1d..d4525233cc681720404770ef1d0c5d3006607a2e 100644 --- a/python/paddle/fluid/imperative/base.py +++ b/python/paddle/fluid/imperative/base.py @@ -11,7 +11,7 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. -import contextlib +from ..wrapped_decorator import signature_safe_contextmanager import numpy as np from paddle.fluid import core @@ -24,7 +24,7 @@ def enabled(): return framework._in_imperative_mode() -@contextlib.contextmanager +@signature_safe_contextmanager def guard(place=None): train = framework.Program() startup = framework.Program() diff --git a/python/paddle/fluid/imperative/layers.py b/python/paddle/fluid/imperative/layers.py index f457f56203eb2c1da62f4d8ad8915c322c822e0a..46640ce37a78f7409af7f82d3302a610ccd366b2 100644 --- a/python/paddle/fluid/imperative/layers.py +++ b/python/paddle/fluid/imperative/layers.py @@ -12,10 +12,12 @@ # See the License for the specific language governing permissions and # limitations under the License. +import collections import contextlib import sys import numpy as np - +import collections +from .. import unique_name from paddle.fluid import core from paddle.fluid import framework from paddle.fluid.imperative import base @@ -24,20 +26,69 @@ __all__ = ['Layer', 'PyLayer'] class Layer(core.Layer): - """Layers composed of operators.""" - - def __init__(self, dtype=core.VarDesc.VarType.FP32, name=None): + """Layers composed of operators. + + Args: + name_scope: prefix name used by the layer to name parameters. + If prefix is "my_model/layer_1", parameter name in MyLayer + can be "my_model/layer_1/MyLayer/w_n", where w is the parameter + base name and n is an unique suffix auto-generated. + dtype: data type for the variables in the layer. + """ + + def __init__(self, name_scope, dtype=core.VarDesc.VarType.FP32): + self._full_name = unique_name.generate(name_scope + "/" + + self.__class__.__name__) self._built = False self._dtype = dtype + self._parameters = collections.OrderedDict() + self._sub_layers = collections.OrderedDict() + + def full_name(self): + """Full name for this layers. + + Full name is composed by name_scope + "/" + MyLayer.__class__.__name__ - def parameters(self): - return [] + Returns full name of this name. + """ + return self._full_name + + def parameters(self, include_sublayers=True): + """Returns a list of Parameters from current and sub-layers. + + Args: + include_sublayers: If true, also include the parameters from + sublayers. + + Returns a list of Parameters. + """ + ret = [p for p in self._parameters.values()] + if include_sublayers: + for l in self._sub_layers.values(): + for p in l.parameters(include_sublayers): + ret.append(p) + return ret + + def sublayers(self, include_sublayers=True): + """Returns a list of sub layers. + + Args: + include_sublayers: If true, also include the layers from sublayers. + + Returns a list of sub layers. + """ + ret = [l for l in self._sub_layers.values()] + if include_sublayers: + for l in self._sub_layers.values(): + for sub_l in l.sublayers(include_sublayers): + ret.append(sub_l) + return ret def clear_gradients(self): for p in self.parameters(): p._clear_gradient() - def _build_once(self, inputs): + def _build_once(self, *args): pass def __call__(self, *inputs): @@ -54,6 +105,66 @@ class Layer(core.Layer): def backward(self, *inputs): raise ValueError("Layer shouldn't implement backward") + def add_sublayer(self, name, sublayer): + """Adds a sub Layer instance. + + Added sublayer can be access like self.name. + + Args: + name: name of this sublayer. + sublayer: an instance of Layer. + Returns: + the sublayer passed in. + """ + assert isinstance(sublayer, core.Layer) + self._sub_layers[name] = sublayer + return sublayer + + def add_parameter(self, name, parameter): + """Adds a Parameter instance. + + Added parameter can be access like self.name. + + Args: + name: name of this sublayer. + parameter: an instance of Parameter. + Returns: + the parameter passed in. + """ + assert isinstance(parameter, framework.Parameter) + self._parameters[name] = parameter + return parameter + + def __getattr__(self, name): + if name in self._parameters: + return self._parameters[name] + elif name in self._sub_layers: + return self._sub_layers[name] + + def __setattr__(self, name, value): + if isinstance(value, framework.Parameter): + params = self.__dict__.get('_parameters', None) + if params is None: + raise ValueError( + "super(YourLayer, self).__init__() should be called first") + params[name] = value + elif isinstance(value, core.Layer): + layers = self.__dict__.get('_sub_layers', None) + if layers is None: + raise ValueError( + "super(YourLayer, self).__init__() should be called first") + layers[name] = value + else: + object.__setattr__(self, name, value) + + def __delattr__(self, name): + if name in self._parameters: + del self._parameters[name] + elif name in self._sub_layers: + del self._sub_layers[name] + else: + object.__delattr__(self, name) + class PyLayer(core.PyLayer): """Layers composed of user-defined python codes.""" diff --git a/python/paddle/fluid/imperative/nn.py b/python/paddle/fluid/imperative/nn.py index 140c0ff037d453641cc119301269121025e17cbd..41655c4f54eecec55bd2c7d2b74adb51efa88b61 100644 --- a/python/paddle/fluid/imperative/nn.py +++ b/python/paddle/fluid/imperative/nn.py @@ -22,17 +22,12 @@ from . import layers from ..framework import Variable, OpProtoHolder from ..param_attr import ParamAttr from ..initializer import Normal, Constant - -__all__ = [ - 'Conv2D', - 'Pool2D', - 'FC', - 'BatchNorm', -] +__all__ = ['Conv2D', 'Pool2D', 'FC', 'BatchNorm', 'Embedding'] class Conv2D(layers.Layer): def __init__(self, + name_scope, num_channels, num_filters, filter_size, @@ -44,19 +39,17 @@ class Conv2D(layers.Layer): act=None, param_attr=None, bias_attr=None, - name=None, dtype=core.VarDesc.VarType.FP32): assert param_attr is not False, "param_attr should not be False here." - super(Conv2D, self).__init__(name=name, dtype=dtype) + super(Conv2D, self).__init__(name_scope, dtype=dtype) # TODO(minqiyang): Move this to the top. from ..layer_helper import LayerHelper self._helper = LayerHelper( - type(self).__name__, + self.full_name(), param_attr=param_attr, bias_attr=bias_attr, dtype=dtype, - name=name, act=act) self._groups = groups @@ -149,6 +142,7 @@ class Conv2D(layers.Layer): class Pool2D(layers.Layer): def __init__(self, + name_scope, pool_size=-1, pool_type="max", pool_stride=1, @@ -157,7 +151,6 @@ class Pool2D(layers.Layer): use_cudnn=True, ceil_mode=False, exclusive=True, - name=None, dtype=core.VarDesc.VarType.FP32): if pool_type not in ["max", "avg"]: raise ValueError( @@ -172,10 +165,10 @@ class Pool2D(layers.Layer): if not isinstance(use_cudnn, bool): raise ValueError("use_cudnn should be True or False") - super(Pool2D, self).__init__(name=name, dtype=dtype) + super(Pool2D, self).__init__(name_scope, dtype=dtype) from ..layer_helper import LayerHelper - self._helper = LayerHelper(type(self).__name__, dtype=dtype, name=name) + self._helper = LayerHelper(self.full_name(), dtype=dtype) self._pool_type = pool_type self._pool_size = utils.convert_to_list(pool_size, 2, 'pool_size') @@ -211,28 +204,24 @@ class Pool2D(layers.Layer): class FC(layers.Layer): def __init__(self, + name_scope, size, param_attr=None, bias_attr=None, num_flatten_dims=1, dtype=core.VarDesc.VarType.FP32, - act=None, - name=None): - super(FC, self).__init__() + act=None): + super(FC, self).__init__(name_scope) self._size = size self._num_flatten_dims = num_flatten_dims self._dtype = dtype from ..layer_helper import LayerHelper self._helper = LayerHelper( - 'FC', + self.full_name(), param_attr=param_attr, bias_attr=bias_attr, - act=act, - name=name) - - def parameters(self): - return [self._w, self._b] + act=act) def _build_once(self, input): input_shape = input.shape @@ -291,6 +280,7 @@ class FC(layers.Layer): class BatchNorm(layers.Layer): def __init__(self, + name_scope, num_channels, act=None, is_test=False, @@ -301,22 +291,20 @@ class BatchNorm(layers.Layer): dtype=core.VarDesc.VarType.FP32, data_layout='NCHW', in_place=False, - name=None, moving_mean_name=None, moving_variance_name=None, do_model_average_for_mean_and_var=False, fuse_with_relu=False, use_global_stats=False): - super(BatchNorm, self).__init__() + super(BatchNorm, self).__init__(name_scope) assert bias_attr is not False, "bias_attr should not be False in batch_norm." from ..layer_helper import LayerHelper self._helper = LayerHelper( - 'batch_norm', + self.full_name(), param_attr=param_attr, bias_attr=bias_attr, - name=name, act=act) if dtype == core.VarDesc.VarType.FP16: @@ -332,21 +320,16 @@ class BatchNorm(layers.Layer): shape=param_shape, dtype=self._dtype, default_initializer=Constant(1.0)) - - # TODO(minqiyang): change stop_gradient sign to trainable to align with static graph - # # setting stop_gradient=True to reduce computation - # if use_global_stats and self._helper.param_attr.learning_rate == 0.: - # self._scale.stop_gradient = True + if use_global_stats and self._helper.param_attr.learning_rate == 0.: + self._scale._stop_gradient = True self._bias = self._helper.create_parameter( attr=self._helper.bias_attr, shape=param_shape, dtype=self._dtype, is_bias=True) - # TODO(minqiyang): change stop_gradient sign to trainable to align with static graph - # # setting stop_gradient=True to reduce computation - # if use_global_stats and self._helper.bias_attr.learning_rate == 0.: - # self._bias.stop_gradient = True + if use_global_stats and self._helper.bias_attr.learning_rate == 0.: + self._bias._stop_gradient = True self._mean = self._helper.create_parameter( attr=ParamAttr( @@ -356,7 +339,7 @@ class BatchNorm(layers.Layer): do_model_average=do_model_average_for_mean_and_var), shape=param_shape, dtype=self._dtype) - self._mean.stop_gradient = True + self._mean._stop_gradient = True self._variance = self._helper.create_parameter( attr=ParamAttr( @@ -366,7 +349,7 @@ class BatchNorm(layers.Layer): do_model_average=do_model_average_for_mean_and_var), shape=param_shape, dtype=self._dtype) - self._variance.stop_gradient = True + self._variance._stop_gradient = True self._in_place = in_place self._momentum = momentum @@ -419,3 +402,90 @@ class BatchNorm(layers.Layer): # Currently, we don't support inplace in imperative mode return self._helper.append_activation(batch_norm_out) + + +class Embedding(layers.Layer): + """ + **Embedding Layer** + + This layer is used to lookup embeddings of IDs, provided by :attr:`input`, in + a lookup table. The result of this lookup is the embedding of each ID in the + :attr:`input`. + + All the input variables are passed in as local variables to the LayerHelper + constructor. + + Args: + name_scope: See base class. + size(tuple|list): The shape of the look up table parameter. It should + have two elements which indicate the size of the dictionary of + embeddings and the size of each embedding vector respectively. + is_sparse(bool): The flag indicating whether to use sparse update. + is_distributed(bool): Whether to run lookup table from remote parameter server. + padding_idx(int|long|None): If :attr:`None`, it makes no effect to lookup. + Otherwise the given :attr:`padding_idx` indicates padding the output + with zeros whenever lookup encounters it in :attr:`input`. If + :math:`padding_idx < 0`, the :attr:`padding_idx` to use in lookup is + :math:`size[0] + dim`. + param_attr(ParamAttr): Parameters for this layer + dtype(np.dtype|core.VarDesc.VarType|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) + input = fluid.layers.data(name='ids', shape=[32, 32], dtype='float32') + embedding = fluid.imperative.Embedding(size=[dict_size, 16]) + fc = embedding(input) + """ + + def __init__(self, + name_scope, + size, + is_sparse=False, + is_distributed=False, + padding_idx=None, + param_attr=None, + dtype='float32'): + + super(Embedding, self).__init__(name_scope) + self._size = size + self._is_sparse = is_sparse + self._is_distributed = is_distributed + + self._padding_idx = -1 if padding_idx is None else padding_idx if padding_idx >= 0 else ( + size[0] + padding_idx) + + self._param_attr = param_attr + self._dtype = dtype + self._remote_prefetch = self._is_sparse and (not self._is_distributed) + if self._remote_prefetch: + assert self._is_sparse is True and self._is_distributed is False + + from ..layer_helper import LayerHelper + self._helper = LayerHelper(self.full_name(), param_attr=param_attr) + self._w = self._helper.create_parameter( + attr=self._param_attr, + shape=self._size, + dtype=self._dtype, + is_bias=False) + + def forward(self, input): + out = self._helper.create_variable_for_type_inference(self._dtype) + self._helper.append_op( + type='lookup_table', + inputs={'Ids': input, + 'W': self._w}, + outputs={'Out': out}, + attrs={ + 'is_sparse': self._is_sparse, + 'is_distributed': self._is_distributed, + 'remote_prefetch': self._remote_prefetch, + 'padding_idx': self._padding_idx + }) + + return out diff --git a/python/paddle/fluid/initializer.py b/python/paddle/fluid/initializer.py index 4f434328e47df4363b304ff55f587018d3157c5e..e8341be28683a25971a53a37c70533a16add1593 100644 --- a/python/paddle/fluid/initializer.py +++ b/python/paddle/fluid/initializer.py @@ -16,7 +16,7 @@ from __future__ import print_function from . import framework import numpy as np -import contextlib +from .wrapped_decorator import signature_safe_contextmanager from .core import VarDesc from . import unique_name @@ -49,7 +49,7 @@ def force_init_on_cpu(): return _force_init_on_cpu_ -@contextlib.contextmanager +@signature_safe_contextmanager def init_on_cpu(): """ Force the variable to be inited on CPU. @@ -366,17 +366,40 @@ class TruncatedNormalInitializer(Initializer): # Initialization Ops should be prepended and not appended if self._seed == 0: self._seed = block.program.random_seed + + # to be compatible of fp16 initalizers + if var.dtype == VarDesc.VarType.FP16: + out_dtype = VarDesc.VarType.FP32 + out_var = block.create_var( + name=unique_name.generate(".".join( + ['truncated_gaussian_random', 'tmp'])), + shape=var.shape, + dtype=out_dtype, + type=VarDesc.VarType.LOD_TENSOR, + persistable=False) + else: + out_dtype = var.dtype + out_var = var + op = block._prepend_op( type="truncated_gaussian_random", - outputs={"Out": var}, + outputs={"Out": out_var}, attrs={ "shape": var.shape, - "dtype": int(var.dtype), + "dtype": out_dtype, "mean": self._mean, "std": self._std_dev, "seed": self._seed }, stop_gradient=True) + + if var.dtype == VarDesc.VarType.FP16: + block.append_op( + type="cast", + inputs={"X": out_var}, + outputs={"Out": var}, + attrs={"in_dtype": out_var.dtype, + "out_dtype": var.dtype}) var.op = op return op diff --git a/python/paddle/fluid/io.py b/python/paddle/fluid/io.py index 6b1d4cc34f3cd40c878740f28618f26d5e89a6bd..24e102b6c2612b58a9b8367ebbefcece535d58bb 100644 --- a/python/paddle/fluid/io.py +++ b/python/paddle/fluid/io.py @@ -16,14 +16,16 @@ from __future__ import print_function import os import errno +import warnings import time import shutil import six from functools import reduce +from paddle.fluid import layers from paddle.fluid.executor import Executor from paddle.fluid.evaluator import Evaluator -from paddle.fluid.framework import Program, Parameter, default_main_program, default_startup_program, Variable +from paddle.fluid.framework import Program, Parameter, default_main_program, default_startup_program, Variable, program_guard from . import core __all__ = [ @@ -764,7 +766,10 @@ def _load_distributed_persistables(executor, dirname, main_program=None): dtype=slice_var.dtype, persistable=True) - dim1_flatten = reduce(lambda x, y: x * y, slice.shape[1:]) + dim1_flatten = 1 + if len(slice.shape) >= 2: + dim1_flatten = reduce(lambda x, y: x * y, slice.shape[1:]) + start = int(offset / dim1_flatten) end = int(offset / dim1_flatten + slice.shape[0]) @@ -930,6 +935,24 @@ def save_inference_model(dirname, if main_program is None: main_program = default_main_program() + if main_program._is_mem_optimized: + warnings.warn( + "save_inference_model must put before you call memory_optimize. \ + the memory_optimize will modify the original program, \ + is not suitable for saving inference model \ + we save the original program as inference model.", + RuntimeWarning) + + # fix the bug that the activation op's output as target will be pruned. + # will affect the inference performance. + # TODO(Superjomn) add an IR pass to remove 1-scale op. + with program_guard(main_program): + uniq_target_vars = [] + for var in target_vars: + if isinstance(var, Variable): + var1 = layers.scale(var, 1.) + uniq_target_vars.append(var1) + target_vars = uniq_target_vars # when a pserver and a trainer running on the same machine, mkdir may conflict try: diff --git a/python/paddle/fluid/layer_helper.py b/python/paddle/fluid/layer_helper.py index 972c51938f2b2282f8de4b090f9af3bc66f89155..65864ca7e09cd4f0760637198d48154eed025c65 100644 --- a/python/paddle/fluid/layer_helper.py +++ b/python/paddle/fluid/layer_helper.py @@ -34,6 +34,9 @@ class LayerHelper(object): self.kwargs = kwargs self.layer_type = layer_type name = self.kwargs.get('name', None) + # TODO(panyx0718, minqiyang): imperative mode + # can not use both `layer_type` and `name`. Deprecate LayerHelper + # and write a Helper for imperative mode. if name is None: self.kwargs['name'] = unique_name.generate(self.layer_type) @@ -300,6 +303,18 @@ class LayerHelper(object): attr.name = unique_name.generate(".".join([self.name, suffix])) if default_initializer is None and attr.initializer is None: + if isinstance(dtype, core.VarDesc.VarType): + if dtype != core.VarDesc.VarType.FP32 and \ + dtype != core.VarDesc.VarType.FP64 and \ + dtype != core.VarDesc.VarType.FP16: + raise TypeError( + "Can not create parameter with default initializer when dtype is not float type. Set default_initializer to fit the parameter dtype!" + ) + else: + if not (dtype.startswith("float") or dtype == "double"): + raise TypeError( + "Can not create parameter with default initializer when dtype is not float type. Set default_initializer to fit the parameter dtype!" + ) if is_bias: attr._set_default_bias_initializer() else: diff --git a/python/paddle/fluid/layers/control_flow.py b/python/paddle/fluid/layers/control_flow.py index a7494aaceab42332cb4362ab1df43d9e0b139f4f..539c9675b2d69b599fc63350c0c7c3b14e32995a 100644 --- a/python/paddle/fluid/layers/control_flow.py +++ b/python/paddle/fluid/layers/control_flow.py @@ -13,7 +13,7 @@ # limitations under the License. from __future__ import print_function -import contextlib +from ..wrapped_decorator import signature_safe_contextmanager from .layer_function_generator import autodoc, templatedoc from .tensor import assign, fill_constant @@ -506,9 +506,9 @@ class While(object): while loop control flow. Args: - cond (Variable): condition used to compare. + cond(Variable): condition used to compare. is_test(bool): A flag indicating whether execution is in test phase. - name (str): The name of this layer. + name(str): The name of this layer. Examples: .. code-block:: python @@ -589,7 +589,8 @@ class While(object): def lod_rank_table(x, level=0): - """LoD Rank Table Operator. Given an input variable **x** and a level number + """ + 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. Refering to specified level of LoD, @@ -883,10 +884,8 @@ def less_than(x, y, force_cpu=None, cond=None, **ignored): return cond -def equal(x, y, cond=None, **ignored): +def equal(x, y, cond=None): """ - **equal** - This layer returns the truth value of :math:`x == y` elementwise. Args: @@ -1458,7 +1457,6 @@ class DynamicRNN(object): Returns: The current timestep in the input sequence. - """ self._assert_in_rnn_block_("step_input") if not isinstance(x, Variable): @@ -1532,11 +1530,10 @@ class DynamicRNN(object): outputs={'Out': [x_reordered]}) return shrink_memory(x_reordered, self.step_idx, self.lod_rank_table) - @contextlib.contextmanager + @signature_safe_contextmanager def block(self): """ - The block for user to define operators in RNN. See the class docstring - for more details. + The block for user to define operators in RNN. """ if self.status != DynamicRNN.BEFORE_RNN: raise ValueError("rnn.block() can only be invoke once") @@ -1640,8 +1637,7 @@ class DynamicRNN(object): dtype(str|numpy.dtype): The data type of the initialized memory. Returns: - the memory variable. - + The memory variable. """ self._assert_in_rnn_block_('memory') self._init_zero_idx_() @@ -1740,7 +1736,7 @@ class DynamicRNN(object): def output(self, *outputs): """ - mark the RNN output variables. + Mark the RNN output variables. Args: outputs: The output variables. diff --git a/python/paddle/fluid/layers/detection.py b/python/paddle/fluid/layers/detection.py index cddc302d52e0a5aea802fd7e1464f1e220c8f769..61a7d4f31d5245e635e2e1fe33e418ce20e94180 100644 --- a/python/paddle/fluid/layers/detection.py +++ b/python/paddle/fluid/layers/detection.py @@ -49,6 +49,8 @@ __all__ = [ 'box_coder', 'polygon_box_transform', 'yolov3_loss', + 'box_clip', + 'multiclass_nms', ] @@ -262,8 +264,10 @@ def detection_output(loc, number is N + 1, N is the batch size. The i-th image has `LoD[i + 1] - LoD[i]` detected results, if it is 0, the i-th image has no detected results. If all images have not detected results, - all the elements in LoD are 0, and output tensor only contains one + LoD will be set to {1}, and output tensor only contains one value, which is -1. + (After version 1.3, when no boxes detected, the lod is changed + from {0} to {1}.) Examples: .. code-block:: python @@ -343,19 +347,107 @@ def box_coder(prior_box, target_box, code_type="encode_center_size", box_normalized=True, - name=None): + name=None, + axis=0): """ - ${comment} + **Box Coder Layer** + + Encode/Decode the target bounding box with the priorbox information. + + The Encoding schema described below: + + .. math:: + + ox = (tx - px) / pw / pxv + + oy = (ty - py) / ph / pyv + + ow = \log(\abs(tw / pw)) / pwv + + oh = \log(\abs(th / ph)) / phv + + The Decoding schema described below: + + .. math:: + + ox = (pw * pxv * tx * + px) - tw / 2 + + oy = (ph * pyv * ty * + py) - th / 2 + + ow = \exp(pwv * tw) * pw + tw / 2 + + oh = \exp(phv * th) * ph + th / 2 + + where `tx`, `ty`, `tw`, `th` denote the target box's center coordinates, + width and height respectively. Similarly, `px`, `py`, `pw`, `ph` denote + the priorbox's (anchor) center coordinates, width and height. `pxv`, + `pyv`, `pwv`, `phv` denote the variance of the priorbox and `ox`, `oy`, + `ow`, `oh` denote the encoded/decoded coordinates, width and height. + + During Box Decoding, two modes for broadcast are supported. Say target + box has shape [N, M, 4], and the shape of prior box can be [N, 4] or + [M, 4]. Then prior box will broadcast to target box along the + assigned axis. Args: - prior_box(${prior_box_type}): ${prior_box_comment} - prior_box_var(${prior_box_var_type}): ${prior_box_var_comment} - target_box(${target_box_type}): ${target_box_comment} - code_type(${code_type_type}): ${code_type_comment} - box_normalized(${box_normalized_type}): ${box_normalized_comment} + prior_box(Variable): Box list prior_box is a 2-D Tensor with shape + [M, 4] holds M boxes, each box is represented as + [xmin, ymin, xmax, ymax], [xmin, ymin] is the + left top coordinate of the anchor box, if the + input is image feature map, they are close to + the origin of the coordinate system. [xmax, ymax] + is the right bottom coordinate of the anchor box. + prior_box_var(Variable|list|None): prior_box_var supports two types + of input. One is variable with shape [M, 4] + holds M group. The other one is list consist of + 4 elements shared by all boxes. + target_box(Variable): This input can be a 2-D LoDTensor with shape + [N, 4] when code_type is 'encode_center_size'. + This input also can be a 3-D Tensor with shape + [N, M, 4] when code_type is 'decode_center_size'. + Each box is represented as + [xmin, ymin, xmax, ymax]. This tensor can + contain LoD information to represent a batch + of inputs. + code_type(string): The code type used with the target box. It can be + encode_center_size or decode_center_size + box_normalized(int): Whether treat the priorbox as a noramlized box. + Set true by default. + name(string): The name of box coder. + axis(int): Which axis in PriorBox to broadcast for box decode, + for example, if axis is 0 and TargetBox has shape + [N, M, 4] and PriorBox has shape [M, 4], then PriorBox + will broadcast to [N, M, 4] for decoding. It is only valid + when code type is decode_center_size. Set 0 by default. Returns: - output_box(${output_box_type}): ${output_box_comment} + output_box(Variable): When code_type is 'encode_center_size', the + output tensor of box_coder_op with shape + [N, M, 4] representing the result of N target + boxes encoded with M Prior boxes and variances. + When code_type is 'decode_center_size', + N represents the batch size and M represents + the number of deocded boxes. + + Examples: + + .. code-block:: python + + prior_box = fluid.layers.data(name='prior_box', + shape=[512, 4], + dtype='float32', + append_batch_size=False) + target_box = fluid.layers.data(name='target_box', + shape=[512,81,4], + dtype='float32', + append_batch_size=False) + output = fluid.layers.box_coder(prior_box=prior_box, + prior_box_var=[0.1,0.1,0.2,0.2], + target_box=target_box, + code_type="decode_center_size", + box_normalized=False, + axis=1) + """ helper = LayerHelper("box_coder", **locals()) @@ -366,15 +458,22 @@ def box_coder(prior_box, output_box = helper.create_variable( name=name, dtype=prior_box.dtype, persistable=False) + inputs = {"PriorBox": prior_box, "TargetBox": target_box} + attrs = { + "code_type": code_type, + "box_normalized": box_normalized, + "axis": axis + } + if isinstance(prior_box_var, Variable): + inputs['PriorBoxVar'] = prior_box_var + elif isinstance(prior_box_var, list): + attrs['variance'] = prior_box_var + else: + raise TypeError("Input variance of box_coder must be Variable or lisz") helper.append_op( type="box_coder", - inputs={ - "PriorBox": prior_box, - "PriorBoxVar": prior_box_var, - "TargetBox": target_box - }, - attrs={"code_type": code_type, - "box_normalized": box_normalized}, + inputs=inputs, + attrs=attrs, outputs={"OutputBox": output_box}) return output_box @@ -410,13 +509,10 @@ def yolov3_loss(x, gtbox, gtlabel, anchors, + anchor_mask, class_num, ignore_thresh, - loss_weight_xy=None, - loss_weight_wh=None, - loss_weight_conf_target=None, - loss_weight_conf_notarget=None, - loss_weight_class=None, + downsample_ratio, name=None): """ ${comment} @@ -428,16 +524,13 @@ def yolov3_loss(x, and x, y, w, h should be relative value of input image. N is the batch number and B is the max box number in an image. - gtlabel (Variable): class id of ground truth boxes, shoud be ins shape + gtlabel (Variable): class id of ground truth boxes, shoud be in shape of [N, B]. anchors (list|tuple): ${anchors_comment} + anchor_mask (list|tuple): ${anchor_mask_comment} class_num (int): ${class_num_comment} ignore_thresh (float): ${ignore_thresh_comment} - loss_weight_xy (float|None): ${loss_weight_xy_comment} - loss_weight_wh (float|None): ${loss_weight_wh_comment} - loss_weight_conf_target (float|None): ${loss_weight_conf_target_comment} - loss_weight_conf_notarget (float|None): ${loss_weight_conf_notarget_comment} - loss_weight_class (float|None): ${loss_weight_class_comment} + downsample_ratio (int): ${downsample_ratio_comment} name (string): the name of yolov3 loss Returns: @@ -452,14 +545,16 @@ def yolov3_loss(x, TypeError: Attr ignore_thresh of yolov3_loss must be a float number Examples: - .. code-block:: python - - x = fluid.layers.data(name='x', shape=[255, 13, 13], dtype='float32') - gtbox = fluid.layers.data(name='gtbox', shape=[6, 5], dtype='float32') - gtlabel = fluid.layers.data(name='gtlabel', shape=[6, 1], dtype='int32') - anchors = [10, 13, 16, 30, 33, 23] - loss = fluid.layers.yolov3_loss(x=x, gtbox=gtbox, class_num=80 - anchors=anchors, ignore_thresh=0.5) + .. code-block:: python + + x = fluid.layers.data(name='x', shape=[255, 13, 13], dtype='float32') + gtbox = fluid.layers.data(name='gtbox', shape=[6, 5], dtype='float32') + gtlabel = fluid.layers.data(name='gtlabel', shape=[6, 1], dtype='int32') + anchors = [10, 13, 16, 30, 33, 23, 30, 61, 62, 45, 59, 119, 116, 90, 156, 198, 373, 326] + anchor_mask = [0, 1, 2] + loss = fluid.layers.yolov3_loss(x=x, gtbox=gtbox, gtlabel=gtlabel, anchors=anchors, + anchor_mask=anchor_mask, class_num=80, + ignore_thresh=0.7, downsample_ratio=32) """ helper = LayerHelper('yolov3_loss', **locals()) @@ -471,6 +566,8 @@ def yolov3_loss(x, raise TypeError("Input gtlabel of yolov3_loss must be Variable") if not isinstance(anchors, list) and not isinstance(anchors, tuple): raise TypeError("Attr anchors of yolov3_loss must be list or tuple") + if not isinstance(anchor_mask, list) and not isinstance(anchor_mask, tuple): + raise TypeError("Attr anchor_mask of yolov3_loss must be list or tuple") if not isinstance(class_num, int): raise TypeError("Attr class_num of yolov3_loss must be an integer") if not isinstance(ignore_thresh, float): @@ -483,31 +580,29 @@ def yolov3_loss(x, loss = helper.create_variable( name=name, dtype=x.dtype, persistable=False) + objectness_mask = helper.create_variable_for_type_inference(dtype='int32') + gt_match_mask = helper.create_variable_for_type_inference(dtype='int32') + attrs = { "anchors": anchors, + "anchor_mask": anchor_mask, "class_num": class_num, "ignore_thresh": ignore_thresh, + "downsample_ratio": downsample_ratio, } - if loss_weight_xy is not None and isinstance(loss_weight_xy, float): - self.attrs['loss_weight_xy'] = loss_weight_xy - if loss_weight_wh is not None and isinstance(loss_weight_wh, float): - self.attrs['loss_weight_wh'] = loss_weight_wh - if loss_weight_conf_target is not None and isinstance( - loss_weight_conf_target, float): - self.attrs['loss_weight_conf_target'] = loss_weight_conf_target - if loss_weight_conf_notarget is not None and isinstance( - loss_weight_conf_notarget, float): - self.attrs['loss_weight_conf_notarget'] = loss_weight_conf_notarget - if loss_weight_class is not None and isinstance(loss_weight_class, float): - self.attrs['loss_weight_class'] = loss_weight_class - helper.append_op( type='yolov3_loss', - inputs={"X": x, - "GTBox": gtbox, - "GTLabel": gtlabel}, - outputs={'Loss': loss}, + inputs={ + "X": x, + "GTBox": gtbox, + "GTLabel": gtlabel, + }, + outputs={ + 'Loss': loss, + 'ObjectnessMask': objectness_mask, + 'GTMatchMask': gt_match_mask + }, attrs=attrs) return loss @@ -1960,3 +2055,169 @@ def generate_proposals(scores, rpn_roi_probs.stop_gradient = True return rpn_rois, rpn_roi_probs + + +def box_clip(input, im_info, name=None): + """ + Clip the box into the size given by im_info + For each input box, The formula is given as follows: + + .. code-block:: text + + xmin = max(min(xmin, im_w - 1), 0) + ymin = max(min(ymin, im_h - 1), 0) + xmax = max(min(xmax, im_w - 1), 0) + ymax = max(min(ymax, im_h - 1), 0) + + where im_w and im_h are computed from im_info: + + .. code-block:: text + + im_h = round(height / scale) + im_w = round(weight / scale) + + Args: + input(variable): The input box, the last dimension is 4. + im_info(variable): The information of image with shape [N, 3] with + layout (height, width, scale). height and width + is the input size and scale is the ratio of input + size and original size. + name (str): The name of this layer. It is optional. + + Returns: + Variable: The cliped tensor variable. + + Examples: + .. code-block:: python + + boxes = fluid.layers.data( + name='data', shape=[8, 4], dtype='float32', lod_level=1) + im_info = fluid.layers.data(name='im_info', shape=[3]) + out = fluid.layers.box_clip( + input=boxes, im_info=im_info, inplace=True) + """ + + helper = LayerHelper("box_clip", **locals()) + output = helper.create_variable_for_type_inference(dtype=input.dtype) + inputs = {"Input": input, "ImInfo": im_info} + helper.append_op(type="box_clip", inputs=inputs, outputs={"Output": output}) + + return output + + +def multiclass_nms(bboxes, + scores, + score_threshold, + nms_top_k, + keep_top_k, + nms_threshold=0.3, + normalized=True, + nms_eta=1., + background_label=0, + name=None): + """ + **Multiclass NMS** + + This operator is to do multi-class non maximum suppression (NMS) on + boxes and scores. + + In the NMS step, this operator greedily selects a subset of detection bounding + boxes that have high scores larger than score_threshold, if providing this + threshold, then selects the largest nms_top_k confidences scores if nms_top_k + is larger than -1. Then this operator pruns away boxes that have high IOU + (intersection over union) overlap with already selected boxes by adaptive + threshold NMS based on parameters of nms_threshold and nms_eta. + + Aftern NMS step, at most keep_top_k number of total bboxes are to be kept + per image if keep_top_k is larger than -1. + + Args: + bboxes (Variable): Two types of bboxes are supported: + 1. (Tensor) A 3-D Tensor with shape + [N, M, 4 or 8 16 24 32] represents the + predicted locations of M bounding bboxes, + N is the batch size. Each bounding box has four + coordinate values and the layout is + [xmin, ymin, xmax, ymax], when box size equals to 4. + 2. (LoDTensor) A 3-D Tensor with shape [M, C, 4] + M is the number of bounding boxes, C is the + class number + scores (Variable): Two types of scores are supported: + 1. (Tensor) A 3-D Tensor with shape [N, C, M] + represents the predicted confidence predictions. + N is the batch size, C is the class number, M is + number of bounding boxes. For each category there + are total M scores which corresponding M bounding + boxes. Please note, M is equal to the 2nd dimension + of BBoxes. + 2. (LoDTensor) A 2-D LoDTensor with shape [M, C]. + M is the number of bbox, C is the class number. + In this case, input BBoxes should be the second + case with shape [M, C, 4]. + background_label (int): The index of background label, the background + label will be ignored. If set to -1, then all + categories will be considered. Default: 0 + score_threshold (float): Threshold to filter out bounding boxes with + low confidence score. If not provided, + consider all boxes. + nms_top_k (int): Maximum number of detections to be kept according to + the confidences aftern the filtering detections based + on score_threshold. + nms_threshold (float): The threshold to be used in NMS. Default: 0.3 + nms_eta (float): The threshold to be used in NMS. Default: 1.0 + keep_top_k (int): Number of total bboxes to be kept per image after NMS + step. -1 means keeping all bboxes after NMS step. + normalized (bool): Whether detections are normalized. Default: True + name(str): Name of the multiclass nms op. Default: None. + + Returns: + Out: A 2-D LoDTensor with shape [No, 6] represents the detections. + Each row has 6 values: [label, confidence, xmin, ymin, xmax, ymax] + or A 2-D LoDTensor with shape [No, 10] represents the detections. + Each row has 10 values: + [label, confidence, x1, y1, x2, y2, x3, y3, x4, y4]. No is the + total number of detections. If there is no detected boxes for all + images, lod will be set to {1} and Out only contains one value + which is -1. + (After version 1.3, when no boxes detected, the lod is changed + from {0} to {1}) + + + Examples: + .. code-block:: python + + + boxes = fluid.layers.data(name='bboxes', shape=[81, 4], + dtype='float32', lod_level=1) + scores = fluid.layers.data(name='scores', shape=[81], + dtype='float32', lod_level=1) + out = fluid.layers.multiclass_nms(bboxes=boxes, + scores=scores, + background_label=0, + score_threshold=0.5, + nms_top_k=400, + nms_threshold=0.3, + keep_top_k=200, + normalized=False) + """ + helper = LayerHelper('multiclass_nms', **locals()) + + output = helper.create_variable_for_type_inference(dtype=bboxes.dtype) + helper.append_op( + type="multiclass_nms", + inputs={'BBoxes': bboxes, + 'Scores': scores}, + attrs={ + 'background_label': background_label, + 'score_threshold': score_threshold, + 'nms_top_k': nms_top_k, + 'nms_threshold': nms_threshold, + 'nms_eta': nms_eta, + 'keep_top_k': keep_top_k, + 'nms_eta': nms_eta, + 'normalized': normalized + }, + outputs={'Out': output}) + output.stop_gradient = True + + return output diff --git a/python/paddle/fluid/layers/io.py b/python/paddle/fluid/layers/io.py index 1762bd3e343e8af6768dd23f8fbc58cd0182d3c9..a9b391fd53a98dc05ee2d909a38dcf82cd5880ea 100644 --- a/python/paddle/fluid/layers/io.py +++ b/python/paddle/fluid/layers/io.py @@ -13,7 +13,7 @@ # limitations under the License. from __future__ import print_function -import contextlib +from ..wrapped_decorator import signature_safe_contextmanager import multiprocessing import os import six @@ -56,7 +56,10 @@ def data(name, Args: name(str): The name/alias of the function - shape(list): Tuple declaring the shape. + shape(list): Tuple declaring the shape. If :code:`append_batch_size` is + True and there is no -1 inside :code:`shape`, it should be + considered as the shape of the each sample. Otherwise, it + should be considered as the shape of the batched data. append_batch_size(bool): 1. If true, it prepends -1 to the shape. For example if shape=[1], the resulting shape is [-1, 1]. @@ -1116,7 +1119,7 @@ class Preprocessor(object): def _is_completed(self): return self.sub_block and self.source_var_names and self.sink_var_names - @contextlib.contextmanager + @signature_safe_contextmanager def block(self): self.status = Preprocessor.IN_SUB_BLOCK self.sub_block = self.main_prog._create_block() diff --git a/python/paddle/fluid/layers/layer_function_generator.py b/python/paddle/fluid/layers/layer_function_generator.py index 09b1b30216b03e71253ca8da1d462db897e1a607..da6c24100452ba26896c8e7c06a76d874b3f51a2 100644 --- a/python/paddle/fluid/layers/layer_function_generator.py +++ b/python/paddle/fluid/layers/layer_function_generator.py @@ -24,7 +24,7 @@ from ..framework import OpProtoHolder, Variable, core, convert_np_dtype_to_dtype from ..layer_helper import LayerHelper __all__ = [ - 'deprecated', 'generate_layer_fn', 'generate_layer_fn_noattr', 'autodoc', + 'deprecated', 'generate_layer_fn', 'generate_activation_fn', 'autodoc', 'templatedoc' ] @@ -89,6 +89,9 @@ def _generate_doc_string_(op_proto, additional_args_lines=None): buf.write('\n') skip_attrs = OpProtoHolder.generated_op_attr_names() + # attr use_mkldnn and is_test also should not be visible to users. + skip_attrs.add("use_mkldnn") + skip_attrs.add("is_test") for each_attr in op_proto.attrs: if each_attr.name in skip_attrs: @@ -226,7 +229,7 @@ def generate_layer_fn(op_type): return func -def generate_layer_fn_noattr(op_type): +def generate_activation_fn(op_type): """Register the Python layer for an Operator without Attribute. Args: @@ -246,6 +249,7 @@ def generate_layer_fn_noattr(op_type): func.__name__ = op_type func.__doc__ = _generate_doc_string_(op_proto) + return func diff --git a/python/paddle/fluid/layers/nn.py b/python/paddle/fluid/layers/nn.py index c63f34aaaaf296e7df5f171140120798a9629c50..fad67b8f376654676e439de173d6b461916f5a11 100644 --- a/python/paddle/fluid/layers/nn.py +++ b/python/paddle/fluid/layers/nn.py @@ -87,6 +87,7 @@ __all__ = [ 'transpose', 'im2sequence', 'nce', + 'sampled_softmax_with_cross_entropy', 'hsigmoid', 'beam_search', 'row_conv', @@ -668,7 +669,11 @@ def dynamic_lstmp(input, candidate_activation='tanh', proj_activation='tanh', dtype='float32', - name=None): + name=None, + h_0=None, + c_0=None, + cell_clip=None, + proj_clip=None): """ **Dynamic LSTMP Layer** @@ -785,6 +790,17 @@ def dynamic_lstmp(input, dtype(str): Data type. Choices = ["float32", "float64"], default "float32". name(str|None): A name for this layer(optional). If set None, the layer will be named automatically. + h_0(Variable): The initial hidden state is an optional input, default is zero. + This is a tensor with shape (N x D), where N is the + batch size and D is the projection size. + c_0(Variable): The initial cell state is an optional input, default is zero. + This is a tensor with shape (N x D), where N is the + batch size. `h_0` and `c_0` can be NULL but only at the same time. + cell_clip(float): If provided the cell state is clipped + by this value prior to the cell output activation. + proj_clip(float): If `num_proj > 0` and `proj_clip` is + provided, then the projected values are clipped elementwise to within + `[-proj_clip, proj_clip]`. Returns: tuple: A tuple of two output variable: the projection of hidden state, \ @@ -831,25 +847,41 @@ def dynamic_lstmp(input, batch_hidden = helper.create_variable_for_type_inference(dtype) batch_gate = helper.create_variable_for_type_inference(dtype) batch_cell_pre_act = helper.create_variable_for_type_inference(dtype) + inputs = { + 'Input': input, + 'Weight': weight, + 'ProjWeight': proj_weight, + 'Bias': bias + } + batch_size = input.shape[0] + if h_0: + assert h_0.shape == (batch_size, proj_size), \ + 'The shape of h0 should be (batch_size, %d)' % proj_size + inputs['H0'] = h_0 + if c_0: + assert c_0.shape == (batch_size, size), \ + 'The shape of c0 should be (batch_size, %d)' % size + inputs['C0'] = c_0 + + if cell_clip: + assert cell_clip >= 0, "cell_clip should not be negtive." + if proj_clip: + assert proj_clip >= 0, "proj_clip should not be negtive." helper.append_op( type='lstmp', - inputs={ - 'Input': input, - 'Weight': weight, - 'ProjWeight': proj_weight, - 'Bias': bias - }, + inputs=inputs, outputs={ 'Projection': projection, 'Cell': cell, - 'OrderedP0': ordered_proj0, 'BatchHidden': batch_hidden, 'BatchGate': batch_gate, 'BatchCellPreAct': batch_cell_pre_act }, attrs={ 'use_peepholes': use_peepholes, + 'cell_clip': cell_clip, + 'proj_clip': proj_clip, 'is_reverse': is_reverse, 'gate_activation': gate_activation, 'cell_activation': cell_activation, @@ -932,7 +964,7 @@ def dynamic_gru(input, create ParamAttr as param_attr. If the Initializer of the param_attr is not set, the parameter is initialized with Xavier. Default: None. bias_attr (ParamAttr|bool|None): The parameter attribute for the bias - of GRU. Note that the bias with :math:`(1 \\times 3D)` concatenates + of GRU.Note that the bias with :math:`(1 \\times 3D)` concatenates the bias in the update gate, reset gate and candidate calculations. If it is set to False, no bias will be applied to the update gate, reset gate and candidate calculations. If it is set to None or one @@ -1073,7 +1105,7 @@ def gru_unit(input, create ParamAttr as param_attr. If the Initializer of the param_attr is not set, the parameter is initialized with Xavier. Default: None. bias_attr (ParamAttr|bool|None): The parameter attribute for the bias - of GRU. Note that the bias with :math:`(1 \\times 3D)` concatenates + of GRU.Note that the bias with :math:`(1 \\times 3D)` concatenates the bias in the update gate, reset gate and candidate calculations. If it is set to False, no bias will be applied to the update gate, reset gate and candidate calculations. If it is set to None or one @@ -2569,7 +2601,27 @@ def adaptive_pool2d(input, require_index=False, name=None): """ - ${comment} + **Adaptive Pool2d Operator** + The adaptive_pool2d operation calculates the output based on the input, pool_size, + pool_type parameters. Input(X) and output(Out) 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(pool_size) should contain two elements which + represent height and width, respectively. Also the H and W dimensions of output(Out) + is same as Parameter(pool_size). + + For average adaptive pool2d: + + .. math:: + + hstart &= floor(i * H_{in} / H_{out}) + + hend &= ceil((i + 1) * H_{in} / H_{out}) + + wstart &= floor(j * W_{in} / W_{out}) + + wend &= ceil((j + 1) * W_{in} / W_{out}) + + Output(i ,j) &= \\frac{sum(Input[hstart:hend, wstart:wend])}{(hend - hstart) * (wend - wstart)} Args: input (Variable): The input tensor of pooling operator. The format of @@ -2579,8 +2631,8 @@ def adaptive_pool2d(input, pool_size (int|list|tuple): The pool kernel size. If pool kernel size is a tuple or list, it must contain two integers, (pool_size_Height, pool_size_Width). pool_type: ${pooling_type_comment} - require_index (bool): If true, the index of max pooling point along with outputs. - it cannot be set in average pooling type. + require_index (bool): If true, the index of max pooling point will be returned along + with outputs. It cannot be set in average pooling type. name (str|None): A name for this layer(optional). If set None, the layer will be named automatically. @@ -2661,18 +2713,42 @@ def adaptive_pool3d(input, require_index=False, name=None): """ - ${comment} + **Adaptive Pool3d Operator** + The adaptive_pool3d operation calculates the output based on the input, pool_size, + pool_type parameters. Input(X) and output(Out) 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(pool_size) should contain + three elements which represent height and width, respectively. Also the D, H and W + dimensions of output(Out) is same as Parameter(pool_size). + + For average adaptive pool3d: + + .. math:: + + dstart &= floor(i * D_{in} / D_{out}) + + dend &= ceil((i + 1) * D_{in} / D_{out}) + + hstart &= floor(j * H_{in} / H_{out}) + + hend &= ceil((j + 1) * H_{in} / H_{out}) + + wstart &= floor(k * W_{in} / W_{out}) + + wend &= ceil((k + 1) * W_{in} / W_{out}) + + Output(i ,j, k) &= \\frac{sum(Input[dstart:dend, hstart:hend, wstart:wend])}{(dend - dstart) * (hend - hstart) * (wend - wstart)} Args: input (Variable): The input tensor of pooling operator. The format of - input tensor is NCHW, 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. + input tensor is NCDHW, 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. pool_size (int|list|tuple): The pool kernel size. If pool kernel size is a tuple or list, - it must contain two integers, (Depth, Height, Width). + it must contain three integers, (Depth, Height, Width). pool_type: ${pooling_type_comment} - require_index (bool): If true, the index of max pooling point along with outputs. - it cannot be set in average pooling type. + require_index (bool): If true, the index of max pooling point will be returned along + with outputs. It cannot be set in average pooling type. name (str|None): A name for this layer(optional). If set None, the layer will be named automatically. @@ -2709,7 +2785,7 @@ def adaptive_pool3d(input, name='data', shape=[3, 32, 32], dtype='float32') pool_out, mask = fluid.layers.adaptive_pool3d( input=data, - pool_size=[3, 3], + pool_size=[3, 3, 3], pool_type='avg') """ if pool_type not in ["max", "avg"]: @@ -2930,6 +3006,7 @@ def batch_norm(input, "momentum": momentum, "epsilon": epsilon, "is_test": is_test, + "data_layout": data_layout, "use_mkldnn": False, "fuse_with_relu": fuse_with_relu, "use_global_stats": use_global_stats @@ -3232,7 +3309,7 @@ def group_norm(input, # create output mean_out = helper.create_variable(dtype=dtype, stop_gradient=True) variance_out = helper.create_variable(dtype=dtype, stop_gradient=True) - group_norm_out = helper.create_variable(dtype) + group_norm_out = helper.create_variable(dtype=dtype) helper.append_op( type="group_norm", @@ -3874,7 +3951,8 @@ def beam_search(pre_ids, end_id, level=0, is_accumulated=True, - name=None): + name=None, + return_parent_idx=False): """ Beam search is a classical algorithm for selecting candidate words in a machine translation task. @@ -3930,10 +4008,16 @@ def beam_search(pre_ids, accumulated scores. name(str|None): A name for this layer(optional). If set None, the layer will be named automatically. + return_parent_idx(bool): Whether to return an extra Tensor variable + preserving the selected_ids' parent indice in pre_ids + in output, which can be used to gather cell states at + the next time step. Returns: - Variable: The LodTensor pair containing the selected ids and the \ - corresponding scores. + Variable: The LodTensor tuple containing the selected ids and the \ + corresponding scores. If :attr:`return_parent_idx` is :attr:`True`, \ + an extra Tensor variable preserving the selected_ids' parent indice \ + is included. Examples: .. code-block:: python @@ -3966,6 +4050,11 @@ def beam_search(pre_ids, selected_scores = helper.create_variable_for_type_inference( dtype=score_type) selected_ids = helper.create_variable_for_type_inference(dtype=id_type) + # parent_idx is a tensor used to gather cell states at the next time + # step. Though lod in selected_ids can also be used to gather by + # sequence_expand, it is not efficient. + # gather_op's index input only supports int32 dtype currently + parent_idx = helper.create_variable_for_type_inference(dtype="int32") helper.append_op( type='beam_search', @@ -3973,6 +4062,7 @@ def beam_search(pre_ids, outputs={ 'selected_ids': selected_ids, 'selected_scores': selected_scores, + 'parent_idx': parent_idx }, attrs={ # TODO(ChunweiYan) to assure other value support @@ -3981,8 +4071,10 @@ def beam_search(pre_ids, 'end_id': end_id, 'is_accumulated': is_accumulated, }) - - return selected_ids, selected_scores + if return_parent_idx: + return selected_ids, selected_scores, parent_idx + else: + return selected_ids, selected_scores def beam_search_decode(ids, scores, beam_size, end_id, name=None): @@ -5400,7 +5492,7 @@ def transpose(x, perm, name=None): Examples: .. code-block:: python - # use append_batch_size=False to avoid prepending extra + # use append_batch_size=False to avoid prepending extra # batch size in shape x = fluid.layers.data(name='x', shape=[5, 10, 15], dtype='float32', append_batch_size=False) @@ -5746,6 +5838,132 @@ def softmax_with_cross_entropy(logits, return loss +def sampled_softmax_with_cross_entropy(logits, + label, + num_samples, + num_true=1, + remove_accidental_hits=True, + use_customized_samples=False, + customized_samples=None, + customized_probabilities=None, + seed=0): + """ + **Sampled Softmax With Cross Entropy Operator.** + + Cross entropy loss with sampled softmax is used as the output layer for + larger output classes extensively. This operator samples a number of samples + for all examples, and computes the softmax normalized values for each + row of the sampled tensor, after which cross-entropy loss is computed. + + Because this operator performs a softmax on logits internally, it expects + unscaled logits. This operator should not be used with the output of + softmax operator since that would produce incorrect results. + + For examples with T true labels (T >= 1), we assume that each true label has + a probability of 1/T. For each sample, S samples are generated using a + log uniform distribution. True labels are concatenated with these samples to + form T + S samples for each example. So, assume the shape of logits is + [N x K], the shape for samples is [N x (T+S)]. For each sampled label, a + probability is calculated, which corresponds to the Q(y|x) in + [Jean et al., 2014](http://arxiv.org/abs/1412.2007). + + Logits are sampled according to the sampled labels. Then if + remove_accidental_hits is True, if a sample[i, j] accidentally hits true + labels, then the corresponding sampled_logits[i, j] is minus by 1e20 to + make its softmax result close to zero. Then sampled logits are subtracted by + logQ(y|x), these sampled logits and re-indexed labels are used to compute + a softmax with cross entropy. + + Args: + logits (Variable): The unscaled log probabilities, which is a 2-D tensor + with shape [N x K]. N is the batch_size, and K is the class number. + label (Variable): The ground truth which is a 2-D tensor. Label is a + Tensor with shape [N x T], where T is the number of true + labels per example. + num_samples (int): The number for each example, num_samples should be + less than the number of class. + num_true(int): The number of target classes per training example. + remove_accidental_hits (bool): A flag indicating whether to remove + accidental hits when sampling. If True and if a sample[i, j] + accidentally hits true labels, then the corresponding + sampled_logits[i, j] is minus by 1e20 to make its softmax result + close to zero. Default is True. + use_customized_samples (bool): Whether to use custom samples and probabities to sample + logits. + customized_samples (Variable): User defined samples, which is a 2-D tensor + with shape [N, T + S]. S is the num_samples, and T is the number of true + labels per example. + customized_probabilities (Variable): User defined probabilities of samples, + a 2-D tensor which has the same shape with customized_samples. + seed (int): The random seed for generating random number, which is used + in the process of sampling. Default is 0. + + Returns: + Variable: Return the cross entropy loss which is a 2-D tensor with shape + [N x 1]. + + Examples: + .. code-block:: python + + logits = fluid.layers.data(name='data', shape=[256], dtype='float32') + label = fluid.layers.data(name='label', shape=[5], dtype='int64') + fc = fluid.layers.fc(input=data, size=100) + out = fluid.layers.sampled_softmax_with_cross_entropy( + logits=fc, label=label, num_samples=25) + """ + helper = LayerHelper('sample_logits', **locals()) + samples = helper.create_variable_for_type_inference(dtype='int64') + probabilities = helper.create_variable_for_type_inference( + dtype=logits.dtype) + sampled_logits \ + = helper.create_variable_for_type_inference(dtype=logits.dtype) + sampled_label = helper.create_variable_for_type_inference(dtype='int64') + sampled_softlabel = helper.create_variable_for_type_inference( + dtype=logits.dtype) + + helper.append_op( + type='sample_logits', + inputs={ + 'Logits': logits, + 'Labels': label, + 'CustomizedSamples': customized_samples, + 'CustomizedProbabilities': customized_probabilities + }, + outputs={ + 'Samples': samples, + 'Probabilities': probabilities, + 'SampledLabels': sampled_label, + 'SampledLogits': sampled_logits + }, + attrs={ + 'use_customized_samples': use_customized_samples, + 'uniq': True, + 'remove_accidental_hits': remove_accidental_hits, + 'num_samples': num_samples, + 'seed': seed + }) + loss = helper.create_variable_for_type_inference(dtype=logits.dtype) + softmax = helper.create_variable_for_type_inference(dtype=logits.dtype) + helper.append_op( + type='one_hot', + inputs={'X': sampled_label}, + attrs={'depth': num_samples + 1}, + outputs={'Out': sampled_softlabel}) + + helper.append_op( + type='softmax_with_cross_entropy', + inputs={'Logits': sampled_logits, + 'Label': sampled_softlabel}, + outputs={'Softmax': softmax, + 'Loss': loss}, + attrs={ + 'soft_label': True, + 'ignore_index': False, + 'numeric_stable_mode': False + }) + return loss / num_true + + def smooth_l1(x, y, inside_weight=None, outside_weight=None, sigma=None): """ This layer computes the smooth L1 loss for Variable :attr:`x` and :attr:`y`. @@ -5917,13 +6135,10 @@ def reshape(x, shape, actual_shape=None, act=None, inplace=False, name=None): than :attr:`shape`. act (str): The non-linear activation to be applied to the reshaped tensor variable. - inplace(bool): Must use :attr:`False` if :attr:`x` is used in multiple - operators. If this flag is set :attr:`True`, reuse input - :attr:`x` to reshape, which will change the shape of - tensor variable :attr:`x` and might cause errors when - :attr:`x` is used in multiple operators. If :attr:`False`, - preserve the shape :attr:`x` and create a new output tensor - variable whose data is copied from input x but reshaped. + inplace(bool): If ``inplace`` is `True`, the input and output of ``layers.reshape`` + are the same variable, otherwise, the input and output of + ``layers.reshape`` are different variables. Note that if :attr:`x` + is more than one layer's input, ``inplace`` must be :attr:`False`. name (str): The name of this layer. It is optional. Returns: @@ -6578,7 +6793,9 @@ def image_resize(input, scale=None, name=None, resample='BILINEAR', - actual_shape=None): + actual_shape=None, + align_corners=True, + align_mode=1): """ **Resize a Batch of Images** @@ -6591,6 +6808,80 @@ def image_resize(input, 'NEAREST' : Nearest neighbor interpolation + Nearest neighbor interpolation is to perform nearest neighbor interpolation + in both the 3rd dimention(in height direction) and the 4th dimention(in width + direction) on input tensor. + + Bilinear interpolation is an extension of linear interpolation for + interpolating functions of two variables (e.g. H-direction and + W-direction in this op) on a rectilinear 2D grid. The key idea is + to perform linear interpolation first in one direction, and then + again in the other direction. + + Align_corners and align_mode are optinal parameters,the calculation method + of interpolation can be selected by them. + + Example: + + For scale: + + if align_corners = True && out_size > 1 : + + scale_factor = (in_size-1.0)/(out_size-1.0) + + else: + + scale_factor = float(in_size/out_size) + + + Nearest neighbor interpolation: + + if: + align_corners = False + + input : (N,C,H_in,W_in) + output: (N,C,H_out,W_out) where: + + H_out = \left \lfloor {H_{in} * scale_{}factor}} \right \rfloor + W_out = \left \lfloor {W_{in} * scale_{}factor}} \right \rfloor + + else: + align_corners = True + + input : (N,C,H_in,W_in) + output: (N,C,H_out,W_out) where: + + H_out = round(H_{in} * scale_{factor}) + W_out = round(W_{in} * scale_{factor}) + + Bilinear interpolation: + + if: + align_corners = False , align_mode = 0 + + input : (N,C,H_in,W_in) + output: (N,C,H_out,W_out) where: + + H_out = (H_{in}+0.5) * scale_{factor} - 0.5 + W_out = (W_{in}+0.5) * scale_{factor} - 0.5 + + + else: + + input : (N,C,H_in,W_in) + output: (N,C,H_out,W_out) where: + + H_out = H_{in} * scale_{factor} + W_out = W_{in} * scale_{factor} + + For details of nearest neighbor interpolation, please refer to Wikipedia: + https://en.wikipedia.org/wiki/Nearest-neighbor_interpolation. + + For details of bilinear interpolation, please refer to Wikipedia: + https://en.wikipedia.org/wiki/Bilinear_interpolation. + + + Args: input (Variable): The input tensor of image resize layer, This is a 4-D tensor of the shape @@ -6620,6 +6911,13 @@ def image_resize(input, set, otherwise errors would be occured in graph constructing stage. Default: None + align_corners(bool) : An optional bool, If True, the centers of the 4 corner pixels of the + input and output tensors are aligned, preserving the values at the + corner pixels. + Default: True + align_mode(int) : An optional for bilinear interpolation. can be \'0\' + for src_idx = scale*(dst_indx+0.5)-0.5 , can be \'1\' for + src_idx = scale*dst_index . Returns: Variable: The output is a 4-D tensor of the shape @@ -6632,6 +6930,8 @@ def image_resize(input, or 'NEAREST' currently. ValueError: One of out_shape and scale must not be None. ValueError: out_shape length should be 2. + TypeError: align_corners shoule be a bool value + ValueError: align_mode can only be '0' or '1' Examples: .. code-block:: python @@ -6647,6 +6947,12 @@ def image_resize(input, "The 'resample' of image_resize can only be 'BILINEAR' or 'NEAREST' currently." ) resample_type = resample_methods[resample] + + if not isinstance(align_corners, bool): + raise TypeError("Attr align_corners should be a bool value") + if align_mode != 0 and align_mode != 1: + raise ValueError("align_mode can only be 0 or 1") + if out_shape is None and scale is None: raise ValueError("One of out_shape and scale must not be None.") helper = LayerHelper('{}_interp'.format(resample_type), **locals()) @@ -6686,9 +6992,13 @@ def image_resize(input, type='{}_interp'.format(resample_type), inputs=inputs, outputs={"Out": out}, - attrs={"out_h": out_h, - "out_w": out_w, - "interp_method": resample_type}) + attrs={ + "out_h": out_h, + "out_w": out_w, + "interp_method": resample_type, + "align_corners": align_corners, + "align_mode": align_mode + }) return out @@ -6697,7 +7007,9 @@ def resize_bilinear(input, out_shape=None, scale=None, name=None, - actual_shape=None): + actual_shape=None, + align_corners=True, + align_mode=1): """ Resize input by performing bilinear interpolation based on given output shape which specified by actual_shape, out_shape and scale @@ -6712,6 +7024,47 @@ def resize_bilinear(input, For details of bilinear interpolation, please refer to Wikipedia: https://en.wikipedia.org/wiki/Bilinear_interpolation + Align_corners and align_mode are optinal parameters,the calculation + method of interpolation can be selected by them. + + + Align_corners and align_mode are optinal parameters,the calculation method + of interpolation can be selected by them. + + Example: + + For scale: + + if align_corners = True && out_size > 1 : + + scale_factor = (in_size-1.0)/(out_size-1.0) + + else: + + scale_factor = float(in_size/out_size) + + Bilinear interpolation: + + if: + align_corners = False , align_mode = 0 + + input : (N,C,H_in,W_in) + output: (N,C,H_out,W_out) where: + + H_out = (H_{in}+0.5) * scale_{factor} - 0.5 + W_out = (W_{in}+0.5) * scale_{factor} - 0.5 + + + else: + + input : (N,C,H_in,W_in) + output: (N,C,H_out,W_out) where: + + H_out = H_{in} * scale_{factor} + W_out = W_{in} * scale_{factor} + + + Args: input(${x_type}): ${x_comment}. @@ -6735,6 +7088,8 @@ def resize_bilinear(input, set, otherwise errors would be occured in graph constructing stage. Default: None + align_corners(bool): ${align_corners_comment} + align_mode(bool): ${align_mode_comment} Returns: ${out_comment}. @@ -6745,7 +7100,8 @@ def resize_bilinear(input, out = fluid.layers.resize_bilinear(input, out_shape=[12, 12]) """ - return image_resize(input, out_shape, scale, name, 'BILINEAR', actual_shape) + return image_resize(input, out_shape, scale, name, 'BILINEAR', actual_shape, + align_corners, align_mode) @templatedoc(op_type="nearest_interp") @@ -6753,13 +7109,48 @@ def resize_nearest(input, out_shape=None, scale=None, name=None, - actual_shape=None): + actual_shape=None, + align_corners=True): """ Resize input by performing nearest neighbor interpolation in both the 3rd dimention(in height direction) and the 4th dimention(in width direction) based on given output shape which specified by actual_shape, out_shape and scale in priority order. + Example: + + For scale: + + if align_corners = True && out_size > 1 : + + scale_factor = (in_size-1.0)/(out_size-1.0) + + else: + + scale_factor = float(in_size/out_size) + + + Nearest neighbor interpolation: + + if: + align_corners = False + + input : (N,C,H_in,W_in) + output: (N,C,H_out,W_out) where: + + H_out = \left \lfloor {H_{in} * scale_{}factor}} \right \rfloor + W_out = \left \lfloor {W_{in} * scale_{}factor}} \right \rfloor + + else: + align_corners = True + + input : (N,C,H_in,W_in) + output: (N,C,H_out,W_out) where: + + H_out = round(H_{in} * scale_{factor}) + W_out = round(W_{in} * scale_{factor}) + + For details of nearest neighbor interpolation, please refer to Wikipedia: https://en.wikipedia.org/wiki/Nearest-neighbor_interpolation @@ -6786,6 +7177,7 @@ def resize_nearest(input, set, otherwise errors would be occured in graph constructing stage. Default: None + align_corners(bool): ${align_corners_comment} Returns: ${out_comment}. @@ -6796,7 +7188,8 @@ def resize_nearest(input, out = fluid.layers.resize_nearest(input, out_shape=[12, 12]) """ - return image_resize(input, out_shape, scale, name, 'NEAREST', actual_shape) + return image_resize(input, out_shape, scale, name, 'NEAREST', actual_shape, + align_corners) def image_resize_short(input, out_short_len, resample='BILINEAR'): @@ -8138,6 +8531,46 @@ def stack(x, axis=0): If :code:`axis` < 0, it would be replaced with :code:`axis+rank(x[0])+1`. If :code:`axis` is None, it would be replaced with 0. + For Example: + + .. code-block:: text + + Case 1: + Input: + x[0].data = [ [1.0 , 2.0 ] ] + x[0].dims = [1, 2] + x[1].data = [ [3.0 , 4.0 ] ] + x[1].dims = [1, 2] + x[2].data = [ [5.0 , 6.0 ] ] + x[2].dims = [1, 2] + + Attrs: + axis = 0 + + Output: + Out.data =[ [ [1.0, 2.0] ], + [ [3.0, 4.0] ], + [ [5.0, 6.0] ] ] + Out.dims = [3, 1, 2] + + Case 2: + Given + x[0].data = [ [1.0 , 2.0 ] ] + x[0].dims = [1, 2] + x[1].data = [ [3.0 , 4.0 ] ] + x[1].dims = [1, 2] + x[2].data = [ [5.0 , 6.0 ] ] + x[2].dims = [1, 2] + + Attrs: + axis = 1 or axis = -2 + + Output: + Out.data =[ [ [1.0, 2.0] + [3.0, 4.0] + [5.0, 6.0] ] ] + Out.dims = [1, 3, 2] + Args: x (Variable|list(Variable)|tuple(Variable)): Input variables. axis (int|None): The axis along which all inputs are stacked. @@ -8510,16 +8943,17 @@ def slice(input, axes, starts, ends): return out -@templatedoc() def shape(input): """ - ${comment} + **Shape Layer** + + Get the shape of the input. Args: - input (Variable): ${input_comment} + input (Variable): The input variable. Returns: - out (Variable): ${out_comment} + Variable: The shape of the input variable. Examples: .. code-block:: python diff --git a/python/paddle/fluid/layers/ops.py b/python/paddle/fluid/layers/ops.py index 6c18af7283e19bd431c8d543255d900dc89cba09..6b4dc4ac89af43271ff4cbb51b02fe78af754a7b 100644 --- a/python/paddle/fluid/layers/ops.py +++ b/python/paddle/fluid/layers/ops.py @@ -14,7 +14,7 @@ from __future__ import print_function import os -from .layer_function_generator import generate_layer_fn, generate_layer_fn_noattr +from .layer_function_generator import generate_layer_fn, generate_activation_fn from .. import core from ..framework import convert_np_dtype_to_dtype_ @@ -53,7 +53,7 @@ globals()['_elementwise_div'] = generate_layer_fn('elementwise_div') __all__ += __activations_noattr__ for _OP in set(__activations_noattr__): - globals()[_OP] = generate_layer_fn_noattr(_OP) + globals()[_OP] = generate_activation_fn(_OP) __all__ += ["uniform_random"] @@ -135,7 +135,7 @@ def thresholded_relu(x, threshold=None): if val is not None: kwargs[name] = val - _thresholded_relu_(**kwargs) + return _thresholded_relu_(**kwargs) thresholded_relu.__doc__ = _thresholded_relu_.__doc__ + """ diff --git a/python/paddle/fluid/layers/tensor.py b/python/paddle/fluid/layers/tensor.py index 2153ca254f0e286a77160a2d53473e1bc76109d5..af747c3cecac66492bb2e2642a88f66a5cfae3db 100644 --- a/python/paddle/fluid/layers/tensor.py +++ b/python/paddle/fluid/layers/tensor.py @@ -567,7 +567,7 @@ def ones(shape, dtype, force_cpu=False): It also sets *stop_gradient* to True. Args: - shape(tuple|list|None): Shape of output tensor + shape(tuple|list): Shape of output tensor dtype(np.dtype|core.VarDesc.VarType|str): Data type of output tensor Returns: @@ -578,6 +578,10 @@ def ones(shape, dtype, force_cpu=False): data = fluid.layers.ones(shape=[1], dtype='int64') """ + assert isinstance(shape, list) or isinstance( + shape, tuple), "The shape's type should be list or tuple." + assert reduce(lambda x, y: x * y, + shape) > 0, "The shape is invalid: %s." % (str(shape)) return fill_constant(value=1.0, **locals()) diff --git a/python/paddle/fluid/optimizer.py b/python/paddle/fluid/optimizer.py index 14f4276e2f4fc4a24d701ef05c94b88c4f0336da..cb799b639648fc0af64a890ffe788d23e7f4f9eb 100644 --- a/python/paddle/fluid/optimizer.py +++ b/python/paddle/fluid/optimizer.py @@ -15,7 +15,7 @@ from __future__ import print_function from collections import defaultdict -from contextlib import contextmanager +from .wrapped_decorator import signature_safe_contextmanager from paddle.fluid.framework import Program, Variable, name_scope, default_main_program from paddle.fluid.distribute_lookup_table import find_distributed_lookup_table @@ -387,7 +387,7 @@ class Optimizer(object): params_grads = [] for param in parameters: - if param.stop_gradient: + if param.stop_gradient or not param.trainable: continue # create gradient variable grad_var = Variable( @@ -649,6 +649,7 @@ class AdagradOptimizer(Optimizer): regularization: A Regularizer, such as fluid.regularizer.L2DecayRegularizer. name: A optional name prefix. + initial_accumulator_value (float): Initial value for moment accumulator. Examples: .. code-block:: python @@ -662,7 +663,8 @@ class AdagradOptimizer(Optimizer): learning_rate, epsilon=1.0e-6, regularization=None, - name=None): + name=None, + initial_accumulator_value=0.0): assert learning_rate is not None assert epsilon is not None super(AdagradOptimizer, self).__init__( @@ -671,6 +673,7 @@ class AdagradOptimizer(Optimizer): name=name) self.type = "adagrad" self._epsilon = epsilon + self.initial_accumulator_value = initial_accumulator_value def _create_accumulators(self, block, parameters): assert isinstance(block, framework.Block) @@ -683,6 +686,16 @@ class AdagradOptimizer(Optimizer): moment_acc = self._get_accumulator(self._moment_acc_str, param_and_grad[0]) + startup_block = framework.default_startup_program().global_block() + startup_block.append_op( + type='fill_constant', + inputs={}, + outputs={'Out': [moment_acc]}, + attrs={ + 'dtype': moment_acc.dtype, + 'value': self.initial_accumulator_value, + 'shape': moment_acc.shape, + }) # Create the adagrad optimizer op adagrad_op = block.append_op( @@ -1368,9 +1381,9 @@ class FtrlOptimizer(Optimizer): Args: learning_rate (float|Variable): global learning rate. - l1 (float): - l2 (float): - lr_power (float): + l1 (float): L1 regularization strength. + l2 (float): L2 regularization strength. + lr_power (float): Learning Rate Power. regularization: A Regularizer, such as fluid.regularizer.L2DecayRegularizer. name: A optional name prefix. @@ -1610,7 +1623,7 @@ class ModelAverage(Optimizer): }, stop_gradient=True) - @contextmanager + @signature_safe_contextmanager def apply(self, executor, need_restore=True): """Apply average values to parameters of current model. """ diff --git a/python/paddle/fluid/parallel_executor.py b/python/paddle/fluid/parallel_executor.py index a07ff6ac69ca20c8c68659a67606076ce8cdf027..648bf69273f7ce31431bf8006c4540580cb94b61 100644 --- a/python/paddle/fluid/parallel_executor.py +++ b/python/paddle/fluid/parallel_executor.py @@ -29,15 +29,6 @@ ExecutionStrategy = core.ParallelExecutor.ExecutionStrategy BuildStrategy = core.ParallelExecutor.BuildStrategy -def _is_pserver_mode(main_program): - main = main_program if main_program \ - else framework.default_main_program() - for op in main.global_block().ops: - if op.type in ["send", "recv"]: - return True - return False - - class ParallelExecutor(object): """ ParallelExecutor is designed for data parallelism, which focuses on distributing @@ -140,12 +131,18 @@ class ParallelExecutor(object): # FIXME(zcd): is_distribution_ is a temporary field, because in pserver mode, # num_trainers is 1, so the current fields of build_strategy doesn't tell if # it's distributed model. - build_strategy.is_distribution = _is_pserver_mode( + build_strategy.is_distribution = framework.is_pserver_mode( main_program) or num_trainers > 1 # step4: get main_program, scope, local_scopes main = main_program if main_program \ else framework.default_main_program() + # FIXME(dzhwinter): enable_inplace should be after memory_optimize + # if turn on python memory optimize, turn off the inplace_pass. + if build_strategy.memory_optimize is None: + build_strategy.memory_optimize = False if main._is_mem_optimized else True + if build_strategy.enable_inplace is None: + build_strategy.enable_inplace = False if main._is_mem_optimized else True scope = scope if scope is not None else executor.global_scope() if share_vars_from and not isinstance(share_vars_from, diff --git a/python/paddle/fluid/profiler.py b/python/paddle/fluid/profiler.py index e05885f5f5bfc169828c1c6e723dffff098c3c2e..d5670dbc823c5d317f27f768c596ed2e009e71b6 100644 --- a/python/paddle/fluid/profiler.py +++ b/python/paddle/fluid/profiler.py @@ -15,7 +15,7 @@ from __future__ import print_function from . import core -from contextlib import contextmanager +from .wrapped_decorator import signature_safe_contextmanager import os import six @@ -35,7 +35,7 @@ NVPROF_CONFIG = [ ] -@contextmanager +@signature_safe_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 @@ -217,7 +217,7 @@ def stop_profiler(sorted_key=None, profile_path='/tmp/profile'): core.disable_profiler(key_map[sorted_key], profile_path) -@contextmanager +@signature_safe_contextmanager def profiler(state, sorted_key=None, profile_path='/tmp/profile'): """The profiler interface. Different from cuda_profiler, this profiler can be used to profile both CPU diff --git a/python/paddle/fluid/recordio_writer.py b/python/paddle/fluid/recordio_writer.py index 076a942cdde5623faa570bf98f889e8145b60f8b..aa581f23a191639fdc026e7781897d5d996823a9 100644 --- a/python/paddle/fluid/recordio_writer.py +++ b/python/paddle/fluid/recordio_writer.py @@ -15,14 +15,14 @@ from __future__ import print_function import os -import contextlib +from .wrapped_decorator import signature_safe_contextmanager from . import core __all__ = [ 'convert_reader_to_recordio_file', 'convert_reader_to_recordio_files' ] -@contextlib.contextmanager +@signature_safe_contextmanager def create_recordio_writer(filename, compressor=core.RecordIOWriter.Compressor.Snappy, max_num_records=1000): diff --git a/python/paddle/fluid/tests/demo/file_reader/convert_data_to_recordio.py b/python/paddle/fluid/tests/demo/file_reader/convert_data_to_recordio.py index 45a104ec9625eacfcb87ea6eae619e3d71410da9..b00af91a9dce637e312c9dc5d7d3824106b5a051 100644 --- a/python/paddle/fluid/tests/demo/file_reader/convert_data_to_recordio.py +++ b/python/paddle/fluid/tests/demo/file_reader/convert_data_to_recordio.py @@ -16,7 +16,6 @@ from __future__ import print_function import sys import paddle.fluid as fluid -import paddle.v2 as paddle def load_vocab(filename): diff --git a/python/paddle/fluid/tests/demo/pyreader.py b/python/paddle/fluid/tests/demo/pyreader.py index ec61e0ebae4feb1a2177da916b77b2ba2d3981b9..bbcef4c3ff23d955662be10b5f4b96a66da4c7d8 100644 --- a/python/paddle/fluid/tests/demo/pyreader.py +++ b/python/paddle/fluid/tests/demo/pyreader.py @@ -20,7 +20,6 @@ import six import paddle import paddle.dataset.mnist as mnist import paddle.fluid as fluid -import paddle.v2 def network(is_train): @@ -72,7 +71,7 @@ def main(): use_cuda=use_cuda, share_vars_from=trainer, main_program=test_prog) train_reader.decorate_paddle_reader( - paddle.v2.reader.shuffle( + paddle.reader.shuffle( paddle.batch(mnist.train(), 512), buf_size=8192)) test_reader.decorate_paddle_reader(paddle.batch(mnist.test(), 512)) diff --git a/python/paddle/fluid/tests/test_detection.py b/python/paddle/fluid/tests/test_detection.py index 2d9ed9f9c69a15af454bfec5918fd8bab27d6e4c..0d39a139eed87f900b1f59fd0569b6acaec0962b 100644 --- a/python/paddle/fluid/tests/test_detection.py +++ b/python/paddle/fluid/tests/test_detection.py @@ -50,6 +50,19 @@ class TestDetection(unittest.TestCase): self.assertEqual(out.shape[-1], 6) print(str(program)) + def test_box_coder_api(self): + program = Program() + with program_guard(program): + x = layers.data(name='x', shape=[4], dtype='float32') + y = layers.data(name='z', shape=[4], dtype='float32', lod_level=1) + bcoder = layers.box_coder( + prior_box=x, + prior_box_var=[0.1, 0.2, 0.1, 0.2], + target_box=y, + code_type='encode_center_size') + self.assertIsNotNone(bcoder) + print(str(program)) + def test_detection_api(self): program = Program() with program_guard(program): @@ -463,11 +476,33 @@ class TestYoloDetection(unittest.TestCase): x = layers.data(name='x', shape=[30, 7, 7], dtype='float32') gtbox = layers.data(name='gtbox', shape=[10, 4], dtype='float32') gtlabel = layers.data(name='gtlabel', shape=[10], dtype='int32') - loss = layers.yolov3_loss(x, gtbox, gtlabel, [10, 13, 30, 13], 10, - 0.5) + loss = layers.yolov3_loss(x, gtbox, gtlabel, [10, 13, 30, 13], + [0, 1], 10, 0.7, 32) self.assertIsNotNone(loss) +class TestBoxClip(unittest.TestCase): + def test_box_clip(self): + program = Program() + with program_guard(program): + input_box = layers.data( + name='input_box', shape=[7, 4], dtype='float32', lod_level=1) + im_info = layers.data(name='im_info', shape=[3], dtype='float32') + out = layers.box_clip(input_box, im_info) + self.assertIsNotNone(out) + + +class TestMulticlassNMS(unittest.TestCase): + def test_multiclass_nms(self): + program = Program() + with program_guard(program): + bboxes = layers.data( + name='bboxes', shape=[-1, 10, 4], dtype='float32') + scores = layers.data(name='scores', shape=[-1, 10], dtype='float32') + output = layers.multiclass_nms(bboxes, scores, 0.3, 400, 200, 0.7) + self.assertIsNotNone(output) + + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/unittests/CMakeLists.txt b/python/paddle/fluid/tests/unittests/CMakeLists.txt index c23dfa01e76c21d0d162f2fed986e2eaf3a70a6d..a1cf5fad138f068c9eac5fe8d681c9f08b192270 100644 --- a/python/paddle/fluid/tests/unittests/CMakeLists.txt +++ b/python/paddle/fluid/tests/unittests/CMakeLists.txt @@ -1,15 +1,6 @@ file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py") string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}") -# The MKLDNN tests are skiped when the MKLDNN flag is OFF -if(NOT WITH_MKLDNN) - foreach(src ${TEST_OPS}) - if(${src} MATCHES ".*_mkldnn_op$") - list(REMOVE_ITEM TEST_OPS ${src}) - endif() - endforeach() -endif(NOT WITH_MKLDNN) - if(NOT WITH_DISTRIBUTE) list(REMOVE_ITEM TEST_OPS test_recv_op) list(REMOVE_ITEM TEST_OPS test_dist_transpiler) @@ -85,6 +76,8 @@ list(REMOVE_ITEM TEST_OPS test_image_classification_resnet) list(REMOVE_ITEM TEST_OPS test_bilinear_interp_op) list(REMOVE_ITEM TEST_OPS test_nearest_interp_op) list(REMOVE_ITEM TEST_OPS test_imperative_resnet) +list(REMOVE_ITEM TEST_OPS test_imperative_optimizer) +list(REMOVE_ITEM TEST_OPS test_ir_memory_optimize_transformer) foreach(TEST_OP ${TEST_OPS}) py_test_modules(${TEST_OP} MODULES ${TEST_OP}) endforeach(TEST_OP) @@ -94,6 +87,8 @@ py_test_modules(test_bilinear_interp_op MODULES test_bilinear_interp_op SERIAL) py_test_modules(test_nearest_interp_op MODULES test_nearest_interp_op SERIAL) py_test_modules(test_imperative_resnet MODULES test_imperative_resnet ENVS FLAGS_cudnn_deterministic=1) +py_test_modules(test_imperative_optimizer MODULES test_imperative_optimizer ENVS + FLAGS_cudnn_deterministic=1) if(WITH_DISTRIBUTE) py_test_modules(test_dist_train MODULES test_dist_train SERIAL) set_tests_properties(test_listen_and_serv_op PROPERTIES TIMEOUT 20) @@ -113,10 +108,21 @@ py_test_modules(test_parallel_executor_crf MODULES test_parallel_executor_crf SE py_test_modules(test_parallel_executor_fetch_feed MODULES test_parallel_executor_fetch_feed SERIAL) set_tests_properties(test_parallel_executor_fetch_feed PROPERTIES TIMEOUT 450) py_test_modules(test_parallel_executor_transformer MODULES test_parallel_executor_transformer SERIAL) +if(NOT WIN32) +py_test_modules(test_ir_memory_optimize_transformer MODULES test_ir_memory_optimize_transformer SERIAL) +endif() if(NOT APPLE) py_test_modules(test_image_classification_resnet MODULES test_image_classification_resnet SERIAL) endif() +if(CMAKE_BUILD_TYPE STREQUAL "Debug") + # change the timeout from 600 to 1200, because in debug mode, this test need more time. + set_tests_properties(test_parallel_executor_seresnext PROPERTIES TIMEOUT 1200) +endif() if (WITH_NGRAPH) add_subdirectory(ngraph) endif() + +if (WITH_MKLDNN) + add_subdirectory(mkldnn) +endif() diff --git a/python/paddle/fluid/tests/unittests/mkldnn/CMakeLists.txt b/python/paddle/fluid/tests/unittests/mkldnn/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..f71e04c09aa38b8cf7b3a167b84d4dc0e6cc3ec7 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/mkldnn/CMakeLists.txt @@ -0,0 +1,6 @@ +file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py") +string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}") + +foreach(TEST_OP ${TEST_OPS}) + py_test_modules(${TEST_OP} MODULES ${TEST_OP}) +endforeach(TEST_OP) diff --git a/python/paddle/fluid/tests/unittests/mkldnn/__init__.py b/python/paddle/fluid/tests/unittests/mkldnn/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..b94a21a7e406b833797f8f521c62a2351c2bc30a --- /dev/null +++ b/python/paddle/fluid/tests/unittests/mkldnn/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/python/paddle/fluid/tests/unittests/test_activation_mkldnn_op.py b/python/paddle/fluid/tests/unittests/mkldnn/test_activation_mkldnn_op.py similarity index 52% rename from python/paddle/fluid/tests/unittests/test_activation_mkldnn_op.py rename to python/paddle/fluid/tests/unittests/mkldnn/test_activation_mkldnn_op.py index 611d0dd076b827b0f528f2e3a31182cc4939d1f1..0f301de47f53f3fcacd38d1415ebdbd7b4efc8f1 100644 --- a/python/paddle/fluid/tests/unittests/test_activation_mkldnn_op.py +++ b/python/paddle/fluid/tests/unittests/mkldnn/test_activation_mkldnn_op.py @@ -17,9 +17,9 @@ from __future__ import print_function import unittest import numpy as np import paddle.fluid.core as core -from op_test import OpTest -from scipy.special import expit -from test_activation_op import TestRelu, TestTanh, TestSqrt, TestAbs +from paddle.fluid.tests.unittests.op_test import OpTest +from paddle.fluid.tests.unittests.test_activation_op import TestRelu, TestTanh, TestSqrt, TestAbs +import paddle.fluid as fluid class TestMKLDNNReluDim2(TestRelu): @@ -97,5 +97,64 @@ class TestMKLDNNAbsDim4(TestAbs): self.attrs = {"use_mkldnn": True} +# Check if primitives already exist in backward +class TestMKLDNNReluPrimitivesAlreadyExist(unittest.TestCase): + def __assert_close(self, tensor, np_array, msg, atol=1e-4): + self.assertTrue(np.allclose(np.array(tensor), np_array, atol=atol), msg) + + def test_check_forward_backward(self): + place = core.CPUPlace() + + np.random.seed(123) + x = np.random.uniform(-1, 1, [2, 2]).astype(np.float32) + out = np.abs(x) + + out_grad = np.random.random_sample(x.shape).astype(np.float32) + x_grad = out_grad * np.sign(x) # Abs grad calculation + + var_dict = {'x': x, 'out': out, 'out@GRAD': out_grad, 'x@GRAD': x_grad} + var_names = list(var_dict.keys()) + ground_truth = {name: var_dict[name] for name in var_names} + + program = fluid.Program() + with fluid.program_guard(program): + block = program.global_block() + for name in ground_truth: + block.create_var( + name=name, dtype='float32', shape=ground_truth[name].shape) + + relu_op = block.append_op( + type="abs", + inputs={"X": block.var('x'), }, + outputs={"Out": block.var('out')}, + attrs={"use_mkldnn": True}) + + # Generate backward op_desc + grad_op_desc_list, op_grad_to_var = core.get_grad_op_desc( + relu_op.desc, set(), []) + grad_op_desc = grad_op_desc_list[0] + new_op_desc = block.desc.append_op() + new_op_desc.copy_from(grad_op_desc) + for var_name in grad_op_desc.output_arg_names(): + block.desc.var(var_name.encode("ascii")) + grad_op_desc.infer_var_type(block.desc) + grad_op_desc.infer_shape(block.desc) + for arg in grad_op_desc.output_arg_names(): + grad_var = block.desc.find_var(arg.encode("ascii")) + grad_var.set_dtype(core.VarDesc.VarType.FP32) + + exe = fluid.Executor(place) + + # Do at least 2 iterations + for i in range(2): + out = exe.run( + program, + feed={name: var_dict[name] + for name in ['x', 'out@GRAD']}, + fetch_list=['x@GRAD']) + + self.__assert_close(x_grad, out[0], "x@GRAD") + + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_batch_norm_mkldnn_op.py b/python/paddle/fluid/tests/unittests/mkldnn/test_batch_norm_mkldnn_op.py similarity index 92% rename from python/paddle/fluid/tests/unittests/test_batch_norm_mkldnn_op.py rename to python/paddle/fluid/tests/unittests/mkldnn/test_batch_norm_mkldnn_op.py index 1286cee8dc1855c1b1695da46ae0b5222c065114..5fce90372d9beda9b04ab68d0a8ac5ef5c124421 100644 --- a/python/paddle/fluid/tests/unittests/test_batch_norm_mkldnn_op.py +++ b/python/paddle/fluid/tests/unittests/mkldnn/test_batch_norm_mkldnn_op.py @@ -19,9 +19,9 @@ import numpy as np import paddle.fluid.core as core from paddle.fluid.op import Operator import paddle.fluid as fluid -from op_test import OpTest +from paddle.fluid.tests.unittests.op_test import OpTest from paddle.fluid.framework import grad_var_name -from test_batch_norm_op import TestBatchNormOpInference, TestBatchNormOpTraining, _reference_training, _reference_grad +from paddle.fluid.tests.unittests.test_batch_norm_op import TestBatchNormOpInference, TestBatchNormOpTraining, _reference_training, _reference_grad class TestMKLDNNBatchNormOpTraining(TestBatchNormOpTraining): diff --git a/python/paddle/fluid/tests/unittests/test_concat_mkldnn_op.py b/python/paddle/fluid/tests/unittests/mkldnn/test_concat_mkldnn_op.py similarity index 94% rename from python/paddle/fluid/tests/unittests/test_concat_mkldnn_op.py rename to python/paddle/fluid/tests/unittests/mkldnn/test_concat_mkldnn_op.py index 0f2130f9049c7ee294444282e59c654551f76603..1a399740692eab8ccea0c984a1a4f2ac984eb045 100644 --- a/python/paddle/fluid/tests/unittests/test_concat_mkldnn_op.py +++ b/python/paddle/fluid/tests/unittests/mkldnn/test_concat_mkldnn_op.py @@ -15,7 +15,7 @@ from __future__ import print_function import unittest -from test_concat_op import TestConcatOp, TestConcatOp2, TestConcatOp3 +from paddle.fluid.tests.unittests.test_concat_op import TestConcatOp, TestConcatOp2, TestConcatOp3 class TestMKLDNNConcatOp(TestConcatOp): diff --git a/python/paddle/fluid/tests/unittests/test_conv2d_int8_mkldnn_op.py b/python/paddle/fluid/tests/unittests/mkldnn/test_conv2d_int8_mkldnn_op.py similarity index 98% rename from python/paddle/fluid/tests/unittests/test_conv2d_int8_mkldnn_op.py rename to python/paddle/fluid/tests/unittests/mkldnn/test_conv2d_int8_mkldnn_op.py index 5ad376cb08e488e85be6369a91d4e81031e9e9db..100a03cea0f740a615c4a08810d4ad9e8c974d7a 100644 --- a/python/paddle/fluid/tests/unittests/test_conv2d_int8_mkldnn_op.py +++ b/python/paddle/fluid/tests/unittests/mkldnn/test_conv2d_int8_mkldnn_op.py @@ -18,8 +18,8 @@ import unittest import numpy as np import paddle.fluid.core as core -from op_test import OpTest -from test_conv2d_op import conv2d_forward_naive, TestConv2dOp +from paddle.fluid.tests.unittests.op_test import OpTest +from paddle.fluid.tests.unittests.test_conv2d_op import conv2d_forward_naive, TestConv2dOp def conv2d_forward_refer(input, filter, group, conv_param): diff --git a/python/paddle/fluid/tests/unittests/test_conv2d_mkldnn_op.py b/python/paddle/fluid/tests/unittests/mkldnn/test_conv2d_mkldnn_op.py similarity index 91% rename from python/paddle/fluid/tests/unittests/test_conv2d_mkldnn_op.py rename to python/paddle/fluid/tests/unittests/mkldnn/test_conv2d_mkldnn_op.py index 438d45b84033b697c3210acc44392b93bf436df0..0542eef80070cbf281ee013c28b7092a2dd17eaa 100644 --- a/python/paddle/fluid/tests/unittests/test_conv2d_mkldnn_op.py +++ b/python/paddle/fluid/tests/unittests/mkldnn/test_conv2d_mkldnn_op.py @@ -16,7 +16,7 @@ from __future__ import print_function import unittest -from test_conv2d_op import TestConv2dOp, TestWithPad, TestWithStride, TestWithGroup, TestWith1x1, TestWithInput1x1Filter1x1 +from paddle.fluid.tests.unittests.test_conv2d_op import TestConv2dOp, TestWithPad, TestWithStride, TestWithGroup, TestWith1x1, TestWithInput1x1Filter1x1 class TestMKLDNN(TestConv2dOp): diff --git a/python/paddle/fluid/tests/unittests/test_conv2d_transpose_mkldnn_op.py b/python/paddle/fluid/tests/unittests/mkldnn/test_conv2d_transpose_mkldnn_op.py similarity index 94% rename from python/paddle/fluid/tests/unittests/test_conv2d_transpose_mkldnn_op.py rename to python/paddle/fluid/tests/unittests/mkldnn/test_conv2d_transpose_mkldnn_op.py index deefdd09abe6b9f9ca362654f21850f598337245..9bcdb7b2a975b648471714ab628caf91b6b6f3a9 100644 --- a/python/paddle/fluid/tests/unittests/test_conv2d_transpose_mkldnn_op.py +++ b/python/paddle/fluid/tests/unittests/mkldnn/test_conv2d_transpose_mkldnn_op.py @@ -16,7 +16,7 @@ from __future__ import print_function import unittest -from test_conv2d_transpose_op import TestConv2dTransposeOp, TestWithPad, TestWithStride +from paddle.fluid.tests.unittests.test_conv2d_transpose_op import TestConv2dTransposeOp, TestWithPad, TestWithStride class TestMKLDNN(TestConv2dTransposeOp): diff --git a/python/paddle/fluid/tests/unittests/test_conv3d_mkldnn_op.py b/python/paddle/fluid/tests/unittests/mkldnn/test_conv3d_mkldnn_op.py similarity index 91% rename from python/paddle/fluid/tests/unittests/test_conv3d_mkldnn_op.py rename to python/paddle/fluid/tests/unittests/mkldnn/test_conv3d_mkldnn_op.py index f0e1265e142b800587599783367eca2203033bf1..080b74502fbe83e97e88a65866e0d9b66b37033e 100644 --- a/python/paddle/fluid/tests/unittests/test_conv3d_mkldnn_op.py +++ b/python/paddle/fluid/tests/unittests/mkldnn/test_conv3d_mkldnn_op.py @@ -16,7 +16,7 @@ from __future__ import print_function import unittest -from test_conv3d_op import TestConv3dOp, TestCase1, TestWithGroup1, TestWithGroup2, TestWith1x1, TestWithInput1x1Filter1x1 +from paddle.fluid.tests.unittests.test_conv3d_op import TestConv3dOp, TestCase1, TestWithGroup1, TestWithGroup2, TestWith1x1, TestWithInput1x1Filter1x1 class TestMKLDNN(TestConv3dOp): diff --git a/python/paddle/fluid/tests/unittests/test_dequantize_mkldnn_op.py b/python/paddle/fluid/tests/unittests/mkldnn/test_dequantize_mkldnn_op.py similarity index 97% rename from python/paddle/fluid/tests/unittests/test_dequantize_mkldnn_op.py rename to python/paddle/fluid/tests/unittests/mkldnn/test_dequantize_mkldnn_op.py index 0c5e1abd7c8fb010357998c0ceaebaf21619fda9..9a54f927cbde648bbbb06d043bbc1391ee43c314 100644 --- a/python/paddle/fluid/tests/unittests/test_dequantize_mkldnn_op.py +++ b/python/paddle/fluid/tests/unittests/mkldnn/test_dequantize_mkldnn_op.py @@ -16,7 +16,7 @@ from __future__ import print_function import unittest import numpy as np -from op_test import OpTest +from paddle.fluid.tests.unittests.op_test import OpTest class TestDeQuantizeOp(OpTest): diff --git a/python/paddle/fluid/tests/unittests/test_elementwise_add_mkldnn_op.py b/python/paddle/fluid/tests/unittests/mkldnn/test_elementwise_add_mkldnn_op.py similarity index 97% rename from python/paddle/fluid/tests/unittests/test_elementwise_add_mkldnn_op.py rename to python/paddle/fluid/tests/unittests/mkldnn/test_elementwise_add_mkldnn_op.py index d85cc1f856df8eaa73cef318b48a292042488edf..c3a42656b71d09dbc22abf8ce2ddc243b43b422f 100644 --- a/python/paddle/fluid/tests/unittests/test_elementwise_add_mkldnn_op.py +++ b/python/paddle/fluid/tests/unittests/mkldnn/test_elementwise_add_mkldnn_op.py @@ -16,8 +16,8 @@ from __future__ import print_function import unittest import numpy as np import paddle.fluid.core as core -from op_test import OpTest -from test_elementwise_add_op import * +from paddle.fluid.tests.unittests.op_test import OpTest +from paddle.fluid.tests.unittests.test_elementwise_add_op import * ''' Some tests differ from the tests defined in test_elementwise_add_op.py because MKLDNN does not support tensors of number of dimensions 3. diff --git a/python/paddle/fluid/tests/unittests/test_elementwise_mul_mkldnn_op.py b/python/paddle/fluid/tests/unittests/mkldnn/test_elementwise_mul_mkldnn_op.py similarity index 98% rename from python/paddle/fluid/tests/unittests/test_elementwise_mul_mkldnn_op.py rename to python/paddle/fluid/tests/unittests/mkldnn/test_elementwise_mul_mkldnn_op.py index 536e9a1c58ec4a8b1b5a7c1d3a5fe737b38d24ab..738715dd70181988028adff1c50be3a52199c312 100644 --- a/python/paddle/fluid/tests/unittests/test_elementwise_mul_mkldnn_op.py +++ b/python/paddle/fluid/tests/unittests/mkldnn/test_elementwise_mul_mkldnn_op.py @@ -15,10 +15,10 @@ from __future__ import print_function import unittest import numpy as np -from op_test import OpTest +from paddle.fluid.tests.unittests.op_test import OpTest import paddle.fluid.core as core from paddle.fluid.op import Operator -from test_elementwise_mul_op import * +from paddle.fluid.tests.unittests.test_elementwise_mul_op import * class TestElementwiseMulMKLDNNOp_BroadcastNCHW16c(ElementwiseMulOp): diff --git a/python/paddle/fluid/tests/unittests/test_fc_mkldnn_op.py b/python/paddle/fluid/tests/unittests/mkldnn/test_fc_mkldnn_op.py similarity index 98% rename from python/paddle/fluid/tests/unittests/test_fc_mkldnn_op.py rename to python/paddle/fluid/tests/unittests/mkldnn/test_fc_mkldnn_op.py index 45951a34d6f61a242cb2dc004d6801a6c1c9dd92..84229a5cffbb466ef3c69cd997adacfb21f6aae2 100644 --- a/python/paddle/fluid/tests/unittests/test_fc_mkldnn_op.py +++ b/python/paddle/fluid/tests/unittests/mkldnn/test_fc_mkldnn_op.py @@ -16,7 +16,7 @@ from __future__ import print_function import unittest import numpy as np -from op_test import OpTest +from paddle.fluid.tests.unittests.op_test import OpTest def fully_connected_naive(input, weights, bias_data=None): diff --git a/python/paddle/fluid/tests/unittests/test_gaussian_random_mkldnn_op.py b/python/paddle/fluid/tests/unittests/mkldnn/test_gaussian_random_mkldnn_op.py similarity index 90% rename from python/paddle/fluid/tests/unittests/test_gaussian_random_mkldnn_op.py rename to python/paddle/fluid/tests/unittests/mkldnn/test_gaussian_random_mkldnn_op.py index 9777ec390656d3f6166bf9f5de7bbad8b6bd786d..c18bd77bd3e6de08283f3ac3a31c73453f3c9129 100644 --- a/python/paddle/fluid/tests/unittests/test_gaussian_random_mkldnn_op.py +++ b/python/paddle/fluid/tests/unittests/mkldnn/test_gaussian_random_mkldnn_op.py @@ -16,7 +16,7 @@ from __future__ import print_function import unittest -from test_gaussian_random_op import TestGaussianRandomOp +from paddle.fluid.tests.unittests.test_gaussian_random_op import TestGaussianRandomOp class TestMKLDNN(TestGaussianRandomOp): diff --git a/python/paddle/fluid/tests/unittests/test_lrn_mkldnn_op.py b/python/paddle/fluid/tests/unittests/mkldnn/test_lrn_mkldnn_op.py similarity index 96% rename from python/paddle/fluid/tests/unittests/test_lrn_mkldnn_op.py rename to python/paddle/fluid/tests/unittests/mkldnn/test_lrn_mkldnn_op.py index f6bb2ab7a696c40cb61dd5b38ca702b577fe7ea2..a5e6e116a5f1bc1e051ce3cfdac8cd1e5f3ed90e 100644 --- a/python/paddle/fluid/tests/unittests/test_lrn_mkldnn_op.py +++ b/python/paddle/fluid/tests/unittests/mkldnn/test_lrn_mkldnn_op.py @@ -15,7 +15,7 @@ from __future__ import print_function import unittest -from test_lrn_op import TestLRNOp +from paddle.fluid.tests.unittests.test_lrn_op import TestLRNOp class TestLRNMKLDNNOp(TestLRNOp): diff --git a/python/paddle/fluid/tests/unittests/test_pool2d_int8_mkldnn_op.py b/python/paddle/fluid/tests/unittests/mkldnn/test_pool2d_int8_mkldnn_op.py similarity index 94% rename from python/paddle/fluid/tests/unittests/test_pool2d_int8_mkldnn_op.py rename to python/paddle/fluid/tests/unittests/mkldnn/test_pool2d_int8_mkldnn_op.py index f4495d0bc8198189962d033ec18b8b67f1f47c84..fca906fecc5fe8d25b9251c886398f8df778043f 100644 --- a/python/paddle/fluid/tests/unittests/test_pool2d_int8_mkldnn_op.py +++ b/python/paddle/fluid/tests/unittests/mkldnn/test_pool2d_int8_mkldnn_op.py @@ -19,8 +19,8 @@ import unittest import numpy as np import paddle.fluid.core as core -from op_test import OpTest -from test_pool2d_op import TestPool2D_Op, avg_pool2D_forward_naive, max_pool2D_forward_naive +from paddle.fluid.tests.unittests.op_test import OpTest +from paddle.fluid.tests.unittests.test_pool2d_op import TestPool2D_Op, avg_pool2D_forward_naive, max_pool2D_forward_naive class TestPool2dMKLDNNInt8_Op(TestPool2D_Op): diff --git a/python/paddle/fluid/tests/unittests/test_pool2d_mkldnn_op.py b/python/paddle/fluid/tests/unittests/mkldnn/test_pool2d_mkldnn_op.py similarity index 90% rename from python/paddle/fluid/tests/unittests/test_pool2d_mkldnn_op.py rename to python/paddle/fluid/tests/unittests/mkldnn/test_pool2d_mkldnn_op.py index 7de5fefc148021d4109da2ac9f4b36c93a05a23f..6de43dd46e5d184ec934f2d85e0c87137e9702e0 100644 --- a/python/paddle/fluid/tests/unittests/test_pool2d_mkldnn_op.py +++ b/python/paddle/fluid/tests/unittests/mkldnn/test_pool2d_mkldnn_op.py @@ -15,7 +15,7 @@ from __future__ import print_function import unittest -from test_pool2d_op import TestPool2D_Op, TestCase1, TestCase2, TestCase3, TestCase4, TestCase5 +from paddle.fluid.tests.unittests.test_pool2d_op import TestPool2D_Op, TestCase1, TestCase2, TestCase3, TestCase4, TestCase5 def create_test_mkldnn_class(parent): diff --git a/python/paddle/fluid/tests/unittests/test_quantize_mkldnn_op.py b/python/paddle/fluid/tests/unittests/mkldnn/test_quantize_mkldnn_op.py similarity index 97% rename from python/paddle/fluid/tests/unittests/test_quantize_mkldnn_op.py rename to python/paddle/fluid/tests/unittests/mkldnn/test_quantize_mkldnn_op.py index 99607928648be437b7f944f86a0c28b99d1775c4..132f7bd039f7797fb0fc332d6f7b8c242af46535 100644 --- a/python/paddle/fluid/tests/unittests/test_quantize_mkldnn_op.py +++ b/python/paddle/fluid/tests/unittests/mkldnn/test_quantize_mkldnn_op.py @@ -16,7 +16,7 @@ from __future__ import print_function import unittest import numpy as np -from op_test import OpTest +from paddle.fluid.tests.unittests.op_test import OpTest class TestQuantizeOp(OpTest): diff --git a/python/paddle/fluid/tests/unittests/test_sum_mkldnn_op.py b/python/paddle/fluid/tests/unittests/mkldnn/test_sum_mkldnn_op.py similarity index 92% rename from python/paddle/fluid/tests/unittests/test_sum_mkldnn_op.py rename to python/paddle/fluid/tests/unittests/mkldnn/test_sum_mkldnn_op.py index 55820f31b81df9f3618d1004f6d21565564efa29..5928047b5171bcf33b024040ce79577b8aa0b53a 100644 --- a/python/paddle/fluid/tests/unittests/test_sum_mkldnn_op.py +++ b/python/paddle/fluid/tests/unittests/mkldnn/test_sum_mkldnn_op.py @@ -16,7 +16,7 @@ from __future__ import print_function import unittest -from test_sum_op import TestSumOp +from paddle.fluid.tests.unittests.test_sum_op import TestSumOp class TestMKLDNN(TestSumOp): diff --git a/python/paddle/fluid/tests/unittests/test_transpose_mkldnn_op.py b/python/paddle/fluid/tests/unittests/mkldnn/test_transpose_mkldnn_op.py similarity index 95% rename from python/paddle/fluid/tests/unittests/test_transpose_mkldnn_op.py rename to python/paddle/fluid/tests/unittests/mkldnn/test_transpose_mkldnn_op.py index 0c201b9e4f48df94924a248d820ae2cf73367560..4845eefe367f1ad6a2eb6ffd1f9b0598b1b4fbbd 100644 --- a/python/paddle/fluid/tests/unittests/test_transpose_mkldnn_op.py +++ b/python/paddle/fluid/tests/unittests/mkldnn/test_transpose_mkldnn_op.py @@ -16,7 +16,7 @@ from __future__ import print_function import unittest -from test_transpose_op import TestTransposeOp +from paddle.fluid.tests.unittests.test_transpose_op import TestTransposeOp class TestTransposeMKLDNN(TestTransposeOp): diff --git a/python/paddle/fluid/tests/unittests/ngraph/test_accuracy_ngraph_op.py b/python/paddle/fluid/tests/unittests/ngraph/test_accuracy_ngraph_op.py new file mode 100644 index 0000000000000000000000000000000000000000..5298c3c2f6f0113977342ab3e09830027585ada1 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/ngraph/test_accuracy_ngraph_op.py @@ -0,0 +1,21 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import print_function + +import unittest +from paddle.fluid.tests.unittests.test_accuracy_op import TestAccuracyOp + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/ngraph/test_activation_ngraph_op.py b/python/paddle/fluid/tests/unittests/ngraph/test_activation_ngraph_op.py index 2bd9bf843039573862a22c85557d416bf82b41f6..034d7792c13efb432e6bef6c95ee554584f29519 100644 --- a/python/paddle/fluid/tests/unittests/ngraph/test_activation_ngraph_op.py +++ b/python/paddle/fluid/tests/unittests/ngraph/test_activation_ngraph_op.py @@ -18,17 +18,7 @@ import unittest import numpy as np import paddle.fluid.core as core from paddle.fluid.tests.unittests.op_test import OpTest -from paddle.fluid.tests.unittests.test_activation_op import TestRelu, TestTanh - - -class TestNGRAPHReluDim2(TestRelu): - def setUp(self): - super(TestNGRAPHReluDim2, self).setUp() - - -class TestNGRAPHTanhDim2(TestTanh): - def setUp(self): - super(TestNGRAPHTanhDim2, self).setUp() +from paddle.fluid.tests.unittests.test_activation_op import TestSigmoid, TestRelu, TestTanh class TestNGRAPHReluDim4(TestRelu): diff --git a/python/paddle/fluid/tests/unittests/ngraph/test_batch_norm_ngraph_op.py b/python/paddle/fluid/tests/unittests/ngraph/test_batch_norm_ngraph_op.py new file mode 100644 index 0000000000000000000000000000000000000000..34fb73f3cf7e8b3d906ed4e04d151923aa219ab1 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/ngraph/test_batch_norm_ngraph_op.py @@ -0,0 +1,21 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import print_function + +import unittest +from paddle.fluid.tests.unittests.test_batch_norm_op import TestBatchNormOpTraining, TestBatchNormOpInference + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/ngraph/test_conv2d_ngraph_op.py b/python/paddle/fluid/tests/unittests/ngraph/test_conv2d_ngraph_op.py new file mode 100644 index 0000000000000000000000000000000000000000..ff2e865b66a5f1166281c267392b0964ca5b3082 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/ngraph/test_conv2d_ngraph_op.py @@ -0,0 +1,21 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import print_function + +import unittest +from paddle.fluid.tests.unittests.test_conv2d_op import TestConv2dOp, TestWithPad, TestWithStride, TestWithGroup, TestWith1x1, TestWithInput1x1Filter1x1 + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/ngraph/test_cross_entropy_ngraph_op.py b/python/paddle/fluid/tests/unittests/ngraph/test_cross_entropy_ngraph_op.py new file mode 100644 index 0000000000000000000000000000000000000000..3057218a1d80deffe7eb3164c2350143fc38007d --- /dev/null +++ b/python/paddle/fluid/tests/unittests/ngraph/test_cross_entropy_ngraph_op.py @@ -0,0 +1,21 @@ +# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import print_function + +import unittest +from paddle.fluid.tests.unittests.test_cross_entropy_op import TestCrossEntropyOp, TestCrossEntropyOp2, TestCrossEntropyOp3, TestCrossEntropyOp4, TestCrossEntropyOp5, TestCrossEntropyOp6, TestCrossEntropyOp7 + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/ngraph/test_elementwise_add_ngraph_op.py b/python/paddle/fluid/tests/unittests/ngraph/test_elementwise_add_ngraph_op.py index 67722db89bc9007c6247b8fc108f6df177157b7d..3fb9af3a542d5e6b0de7d8d839408759abdaedcb 100644 --- a/python/paddle/fluid/tests/unittests/ngraph/test_elementwise_add_ngraph_op.py +++ b/python/paddle/fluid/tests/unittests/ngraph/test_elementwise_add_ngraph_op.py @@ -13,75 +13,9 @@ # limitations under the License. from __future__ import print_function -import unittest -from paddle.fluid.tests.unittests.test_elementwise_add_op import * - - -class TestNGRAPHElementwiseAddOp(TestElementwiseAddOp): - def init_input_output(self): - super(TestNGRAPHElementwiseAddOp, self).init_input_output() - - -class TestNGRAPHElementwiseAddOp_scalar(TestElementwiseAddOp_scalar): - def init_input_output(self): - super(TestNGRAPHElementwiseAddOp_scalar, self).init_input_output() - - -class TestNGRAPHElementwiseAddOp_scalar2(TestElementwiseAddOp_scalar2): - def init_input_output(self): - super(TestNGRAPHElementwiseAddOp_scalar2, self).init_input_output() - - -class TestNGRAPHElementwiseAddOp_Vector(TestElementwiseAddOp_Vector): - def init_input_output(self): - super(TestNGRAPHElementwiseAddOp_Vector, self).init_input_output() - - -class TesNGRAPHtElementwiseAddOp_broadcast_0(TestElementwiseAddOp_broadcast_0): - def init_input_output(self): - super(TesNGRAPHtElementwiseAddOp_broadcast_0, self).init_input_output() - - -class TestNGRAPHElementwiseAddOp_broadcast_1(TestElementwiseAddOp_broadcast_1): - def init_input_output(self): - super(TestNGRAPHElementwiseAddOp_broadcast_1, self).init_input_output() - - -class TestNGRAPHElementwiseAddOp_broadcast_2(TestElementwiseAddOp_broadcast_2): - def init_input_output(self): - super(TestNGRAPHElementwiseAddOp_broadcast_2, self).init_input_output() - - -class TestNGRAPHElementwiseAddOp_broadcast_3(TestElementwiseAddOp_broadcast_3): - def init_input_output(self): - super(TestNGRAPHElementwiseAddOp_broadcast_3, self).init_input_output() - - -class TestNGRAPHElementwiseAddOp_broadcast_4(TestElementwiseAddOp_broadcast_4): - def init_input_output(self): - super(TestNGRAPHElementwiseAddOp_broadcast_4, self).init_input_output() - - -class TestNGRAPHElementwiseAddOp_rowwise_add_0( - TestElementwiseAddOp_rowwise_add_0): - def init_input_output(self): - super(TestNGRAPHElementwiseAddOp_rowwise_add_0, - self).init_input_output() - - -class TestNGRAPHElementwiseAddOp_rowwise_add_1( - TestElementwiseAddOp_rowwise_add_1): - def init_input_output(self): - super(TestNGRAPHElementwiseAddOp_rowwise_add_1, - self).init_input_output() - - -class TestNGRAPHElementwiseAddOp_channelwise_add( - TestElementwiseAddOp_channelwise_add): - def init_input_output(self): - super(TestNGRAPHElementwiseAddOp_channelwise_add, - self).init_input_output() +import unittest +from paddle.fluid.tests.unittests.test_elementwise_add_op import TestElementwiseAddOp, TestElementwiseAddOp_broadcast_0 if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/unittests/ngraph/test_fill_constant_ngraph_op.py b/python/paddle/fluid/tests/unittests/ngraph/test_fill_constant_ngraph_op.py index 835376ffe78f9119a9be6c379998e3a3b50aab43..2b10b8f7a3ac0f978c13bd86824b939e69c5336a 100644 --- a/python/paddle/fluid/tests/unittests/ngraph/test_fill_constant_ngraph_op.py +++ b/python/paddle/fluid/tests/unittests/ngraph/test_fill_constant_ngraph_op.py @@ -13,24 +13,34 @@ # limitations under the License. from __future__ import print_function + import unittest +import numpy as np from paddle.fluid.tests.unittests.test_fill_constant_op import TestFillConstantOp1, TestFillConstantOp2, TestFillConstantOpWithSelectedRows -class TestNGRAPHFillConstantOp1(TestFillConstantOp1): +class TestNGRAPHFillConstantFP64(TestFillConstantOp1): def setUp(self): - super(TestNGRAPHFillConstantOp1, self).setUp() + super(TestNGRAPHFillConstantFP64, self).setUp() + + self.attrs = {'shape': [123, 92], 'value': 3.8, 'dtype': 6} + self.outputs = {'Out': np.full((123, 92), 3.8)} -class TestNGRAPHFillConstantOp2(TestFillConstantOp2): +class TestNGRAPHFillConstantINT32(TestFillConstantOp2): def setUp(self): - super(TestNGRAPHFillConstantOp2, self).setUp() + super(TestNGRAPHFillConstantINT32, self).setUp() + self.attrs = {'shape': [123, 92], 'dtype': 2} + self.outputs = {'Out': np.full((123, 92), 0)} -class TestNGRAPHFillConstantOpWithSelectedRows( - TestFillConstantOpWithSelectedRows): + +class TestNGRAPHFillConstantINT64(TestFillConstantOp2): def setUp(self): - super(TestFillConstantOpWithSelectedRows, self).setUp() + super(TestNGRAPHFillConstantINT64, self).setUp() + + self.attrs = {'shape': [123, 92], 'dtype': 3} + self.outputs = {'Out': np.full((123, 92), 0)} if __name__ == "__main__": diff --git a/python/paddle/fluid/tests/unittests/ngraph/test_mean_ngraph_op.py b/python/paddle/fluid/tests/unittests/ngraph/test_mean_ngraph_op.py index 5535427ea8a93fdc5818cdc058aedb6fe72165ee..b4894734cbcc11cf5eec7401297dc35545aa7268 100644 --- a/python/paddle/fluid/tests/unittests/ngraph/test_mean_ngraph_op.py +++ b/python/paddle/fluid/tests/unittests/ngraph/test_mean_ngraph_op.py @@ -14,18 +14,7 @@ from __future__ import print_function import unittest -from paddle.fluid.tests.unittests.test_mean_op import TestMeanOp, TestFP16MeanOp - - -class TestNGRAPHMeanOp(TestMeanOp): - def setUp(self): - super(TestNGRAPHMeanOp, self).setUp() - - -class TestNGRAPHFP16MeanOp(TestFP16MeanOp): - def setUp(self): - super(TestNGRAPHFP16MeanOp, self).setUp() - +from paddle.fluid.tests.unittests.test_mean_op import TestMeanOp if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/ngraph/test_momentum_ngraph_op.py b/python/paddle/fluid/tests/unittests/ngraph/test_momentum_ngraph_op.py new file mode 100644 index 0000000000000000000000000000000000000000..2c3549d907f5f67abc0cbd448a492d95b8ae6c32 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/ngraph/test_momentum_ngraph_op.py @@ -0,0 +1,21 @@ +# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import print_function + +import unittest +from paddle.fluid.tests.unittests.test_momentum_op import TestMomentumOp1, TestMomentumOp2, TestLarsMomentumOp, TestSparseMomentumOp, TestSparseMomentumOp2 + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/ngraph/test_mul_ngraph_op.py b/python/paddle/fluid/tests/unittests/ngraph/test_mul_ngraph_op.py index 6aba62f7c08e3fe646372c851622f2e321b3aee2..549d03f6e92dc7e88ec8618e5f97287bb68ed0d9 100644 --- a/python/paddle/fluid/tests/unittests/ngraph/test_mul_ngraph_op.py +++ b/python/paddle/fluid/tests/unittests/ngraph/test_mul_ngraph_op.py @@ -15,28 +15,7 @@ from __future__ import print_function import unittest -from paddle.fluid.tests.unittests.test_mul_op import TestMulOp, TestMulOp2, TestFP16MulOp1, TestFP16MulOp2 - - -class TestNGRAPHMulOp(TestMulOp): - def init_dtype_type(self): - pass - - -class TestNGRAPHMulOp2(TestMulOp2): - def init_dtype_type(self): - pass - - -class TestNGRAPHFP16MulOp1(TestFP16MulOp1): - def init_dtype_type(self): - pass - - -class TestNGRAPHFP16MulOp2(TestFP16MulOp2): - def init_dtype_type(self): - pass - +from paddle.fluid.tests.unittests.test_mul_op import TestMulOp, TestMulOp2 if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/ngraph/test_pool2d_ngraph_op.py b/python/paddle/fluid/tests/unittests/ngraph/test_pool2d_ngraph_op.py new file mode 100644 index 0000000000000000000000000000000000000000..ff82e9fa1d3d343aa7faf56a0bd27d2c9edc1ea4 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/ngraph/test_pool2d_ngraph_op.py @@ -0,0 +1,39 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import print_function + +import unittest + +from paddle.fluid.tests.unittests.test_pool2d_op import TestPool2D_Op, TestCase1, TestCase2, TestCase3, TestCase4, TestCase5 + + +class TestNGRAPHCeilMode(TestCase1): + def setUp(self): + super(TestNGRAPHCeilMode, self).setUp() + + def init_ceil_mode(self): + self.ceil_mode = True + + +class TestNGRAPHAdaptive(TestCase1): + def setUp(self): + super(TestNGRAPHAdaptive, self).setUp() + + def init_adaptive(self): + self.adaptive = True + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/ngraph/test_scale_ngraph_op.py b/python/paddle/fluid/tests/unittests/ngraph/test_scale_ngraph_op.py index b42a1f73fa72b0dab936a3bb61a8893978b229ec..8beb44f55e487eef5f1957e9284d4a711c9770aa 100644 --- a/python/paddle/fluid/tests/unittests/ngraph/test_scale_ngraph_op.py +++ b/python/paddle/fluid/tests/unittests/ngraph/test_scale_ngraph_op.py @@ -13,28 +13,7 @@ # limitations under the License. from __future__ import print_function import unittest -from paddle.fluid.tests.unittests.test_scale_op import TestScaleOp, TestScaleOpSelectedRows, TestScaleFp16Op, TestScaleFp16OpSelectedRows - - -class TestNGRAPHScaleOp(TestScaleOp): - def init_dtype_type(self): - pass - - -class TestNGRAPHScaleOpSelectedRows(TestScaleOpSelectedRows): - def init_dtype_type(self): - pass - - -class TestNGRAPHScaleFp16Op(TestScaleFp16Op): - def init_dtype_type(self): - pass - - -class TestNGRAPHScaleFp16OpSelectedRows(TestScaleFp16OpSelectedRows): - def init_dtype_type(self): - pass - +from paddle.fluid.tests.unittests.test_scale_op import TestScaleOp, TestScaleOpSelectedRows if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/ngraph/test_softmax_ngraph_op.py b/python/paddle/fluid/tests/unittests/ngraph/test_softmax_ngraph_op.py index 81894c6e3872e4617085c6bb4b0219a49c9986fd..0cb08842df0797952c47a63ba2bbb8614c0e8a22 100644 --- a/python/paddle/fluid/tests/unittests/ngraph/test_softmax_ngraph_op.py +++ b/python/paddle/fluid/tests/unittests/ngraph/test_softmax_ngraph_op.py @@ -16,11 +16,5 @@ from __future__ import print_function import unittest from paddle.fluid.tests.unittests.test_softmax_op import TestSoftmaxOp - -class TestSoftmaxNGRAPHOp(TestSoftmaxOp): - def setUp(self): - super(TestSoftmaxNGRAPHOp, self).setUp() - - if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/ngraph/test_sum_ngraph_op.py b/python/paddle/fluid/tests/unittests/ngraph/test_sum_ngraph_op.py new file mode 100644 index 0000000000000000000000000000000000000000..ed9fb618024301818a12fd0d02b09c6f3a5f2c53 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/ngraph/test_sum_ngraph_op.py @@ -0,0 +1,19 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from __future__ import print_function +import unittest +from paddle.fluid.tests.unittests.test_sum_op import TestSumOp, TestSelectedRowsSumOp, TestLoDTensorAndSelectedRowsOp + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/ngraph/test_top_k_ngraph_op.py b/python/paddle/fluid/tests/unittests/ngraph/test_top_k_ngraph_op.py index 3a0171087dce5d4c7b72eca7f7e4fb955af94812..d2319c4d921fccb950b1a3059fdecd3b3b044182 100644 --- a/python/paddle/fluid/tests/unittests/ngraph/test_top_k_ngraph_op.py +++ b/python/paddle/fluid/tests/unittests/ngraph/test_top_k_ngraph_op.py @@ -16,26 +16,5 @@ from __future__ import print_function import unittest from paddle.fluid.tests.unittests.test_top_k_op import TestTopkOp, TestTopkOp3d, TestTopkOp2, TestTopkOp3, TestTopkOp4 - -class TestNGRAPHTopkOp(TestTopkOp): - def setUp(self): - super(TestNGRAPHTopkOp, self).setUp() - - -class TestNGRAPHTopkOp2(TestTopkOp2): - def setUp(self): - super(TestNGRAPHTopkOp2, self).setUp() - - -class TestNGRAPHTopkOp3(TestTopkOp3): - def setUp(self): - super(TestNGRAPHTopkOp3, self).setUp() - - -class TestNGRAPHTopkOp4(TestTopkOp4): - def setUp(self): - super(TestNGRAPHTopkOp4, self).setUp() - - if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/op_test.py b/python/paddle/fluid/tests/unittests/op_test.py index 0fe836683b029698b670bbb9f9bb258c2f3b68a0..823445724302dbde47bc36122c62ef44a7e2394f 100644 --- a/python/paddle/fluid/tests/unittests/op_test.py +++ b/python/paddle/fluid/tests/unittests/op_test.py @@ -14,6 +14,7 @@ from __future__ import print_function +import os import unittest import numpy as np import random @@ -374,6 +375,9 @@ class OpTest(unittest.TestCase): return [] places = [fluid.CPUPlace()] cpu_only = self._cpu_only if hasattr(self, '_cpu_only') else False + use_ngraph = bool(os.getenv("FLAGS_use_ngraph", False)) + if use_ngraph: + cpu_only = True if core.is_compiled_with_cuda() and core.op_support_gpu(self.op_type)\ and not cpu_only: places.append(core.CUDAPlace(0)) diff --git a/python/paddle/fluid/tests/unittests/parallel_executor_test_base.py b/python/paddle/fluid/tests/unittests/parallel_executor_test_base.py index fdacd241f9e1f8d442f55098e2d192a3d57fdaf1..a94487e67dc90d4df935867f841bc567c37c8aa2 100644 --- a/python/paddle/fluid/tests/unittests/parallel_executor_test_base.py +++ b/python/paddle/fluid/tests/unittests/parallel_executor_test_base.py @@ -40,7 +40,8 @@ class TestParallelExecutorBase(unittest.TestCase): seed=None, use_parallel_executor=True, use_reduce=False, - use_ir_memory_optimize=False, + use_ir_memory_optimize=True, + enable_inplace=True, fuse_elewise_add_act_ops=False, fuse_relu_depthwise_conv=False, optimizer=fluid.optimizer.Adam, @@ -60,63 +61,65 @@ class TestParallelExecutorBase(unittest.TestCase): main.random_seed = seed loss = method(use_feed=feed_dict is not None) - if optimizer: optimizer().minimize(loss) if memory_opt: fluid.memory_optimize(main) - place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() - exe = fluid.Executor(place) - exe.run(startup) - exec_strategy = fluid.ExecutionStrategy() - exec_strategy.allow_op_delay = allow_op_delay - if use_fast_executor: - exec_strategy.use_experimental_executor = True - build_strategy = fluid.BuildStrategy() - build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce \ - if use_reduce else fluid.BuildStrategy.ReduceStrategy.AllReduce - build_strategy.fuse_elewise_add_act_ops = fuse_elewise_add_act_ops - build_strategy.fuse_relu_depthwise_conv = fuse_relu_depthwise_conv - build_strategy.memory_optimize = use_ir_memory_optimize - build_strategy.enable_sequential_execution = enable_sequential_execution - if use_cuda and core.is_compiled_with_cuda(): - build_strategy.remove_unnecessary_lock = True - if use_parallel_executor: - binary = compiler.CompiledProgram(main).with_data_parallel( - loss_name=loss.name, - build_strategy=build_strategy, - exec_strategy=exec_strategy) - else: - binary = compiler.CompiledProgram(main) - - if batch_size is not None: - batch_size *= fluid.core.get_cuda_device_count( - ) if use_cuda else int( - os.environ.get('CPU_NUM', multiprocessing.cpu_count())) - begin = time.time() - first_loss, = run_executor( - exe=exe, binary=binary, feed=feed_dict, fetch_list=[loss.name]) - - for i in range(iter): - run_executor( - exe=exe, binary=binary, feed=feed_dict, fetch_list=[]) - - last_loss, = run_executor( - exe=exe, binary=binary, feed=feed_dict, fetch_list=[loss.name]) - end = time.time() - - if batch_size is not None: - print("%.4f Instance per second" % ( - (batch_size * iter + 2) / (end - begin))) - - avg_last_loss_val = np.array(last_loss).mean() - avg_first_loss_val = np.array(first_loss).mean() - if math.isnan(float(avg_last_loss_val)) or math.isnan( - float(avg_first_loss_val)): - sys.exit("got NaN loss, training failed.") - - print(first_loss, last_loss) - # self.assertGreater(first_loss[0], last_loss[0]) - return first_loss, last_loss + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + exe = fluid.Executor(place) + exe.run(startup) + exec_strategy = fluid.ExecutionStrategy() + exec_strategy.allow_op_delay = allow_op_delay + if use_fast_executor: + exec_strategy.use_experimental_executor = True + build_strategy = fluid.BuildStrategy() + build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce \ + if use_reduce else fluid.BuildStrategy.ReduceStrategy.AllReduce + build_strategy.fuse_elewise_add_act_ops = fuse_elewise_add_act_ops + build_strategy.fuse_relu_depthwise_conv = fuse_relu_depthwise_conv + build_strategy.memory_optimize = False if memory_opt else use_ir_memory_optimize + # python memory optimization is conflict with inplace pass. + # Use ir graph memory optimization after inplace pass is the correct way. + build_strategy.enable_inplace = False if memory_opt else enable_inplace + build_strategy.enable_sequential_execution = enable_sequential_execution + + if use_cuda and core.is_compiled_with_cuda(): + build_strategy.remove_unnecessary_lock = True + if use_parallel_executor: + binary = compiler.CompiledProgram(main).with_data_parallel( + loss_name=loss.name, + build_strategy=build_strategy, + exec_strategy=exec_strategy) + else: + binary = compiler.CompiledProgram(main) + + if batch_size is not None: + batch_size *= fluid.core.get_cuda_device_count( + ) if use_cuda else int( + os.environ.get('CPU_NUM', multiprocessing.cpu_count())) + begin = time.time() + first_loss, = run_executor( + exe=exe, binary=binary, feed=feed_dict, fetch_list=[loss.name]) + + for i in range(iter): + run_executor(exe=exe, binary=binary, feed=feed_dict, fetch_list=[]) + + last_loss, = run_executor( + exe=exe, binary=binary, feed=feed_dict, fetch_list=[loss.name]) + end = time.time() + + if batch_size is not None: + print("%.4f Instance per second" % ( + (batch_size * iter + 2) / (end - begin))) + + avg_last_loss_val = np.array(last_loss).mean() + avg_first_loss_val = np.array(first_loss).mean() + if math.isnan(float(avg_last_loss_val)) or math.isnan( + float(avg_first_loss_val)): + sys.exit("got NaN loss, training failed.") + + print(first_loss, last_loss) + # self.assertGreater(first_loss[0], last_loss[0]) + return first_loss, last_loss diff --git a/python/paddle/fluid/tests/unittests/test_base_layer.py b/python/paddle/fluid/tests/unittests/test_base_layer.py new file mode 100644 index 0000000000000000000000000000000000000000..caf9750e58889ac40c7cdde022f0b6aa5e77fc42 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_base_layer.py @@ -0,0 +1,89 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest +import numpy as np + +import paddle.fluid as fluid +from paddle.fluid.layer_helper import LayerHelper + + +class L1(fluid.imperative.Layer): + def __init__(self, prefix): + super(L1, self).__init__(prefix) + self._helper = LayerHelper( + self.full_name(), + param_attr=fluid.ParamAttr( + initializer=fluid.initializer.Constant(value=0.1))) + + self.w1 = self._helper.create_parameter( + attr=self._helper.param_attr, + shape=[2, 2], + dtype='float32', + is_bias=False) + self.w2 = self._helper.create_parameter( + attr=self._helper.param_attr, + shape=[2, 2], + dtype='float32', + is_bias=False) + + def forward(self): + return self.w1 + self.w2 + + +class L2(fluid.imperative.Layer): + def __init__(self, prefix): + super(L2, self).__init__(prefix) + self.layer1 = L1(self.full_name()) + self.layer2 = L1(self.full_name()) + + def forward(self): + return self.layer1() + self.layer2() + + +class L3(fluid.imperative.Layer): + def __init__(self, prefix): + super(L3, self).__init__(prefix) + self.layer1 = L2(self.full_name()) + self.layer2 = L2(self.full_name()) + + def forward(self): + return self.layer1() + self.layer2() + + +class TestBaseLayer(unittest.TestCase): + def test_one_level(self): + with fluid.imperative.guard(): + l = L1('test_one_level') + ret = l() + self.assertEqual(l.w1.name, "test_one_level/L1_0_0.w_0") + self.assertEqual(l.w2.name, "test_one_level/L1_0_0.w_1") + self.assertTrue(np.allclose(ret._numpy(), 0.2 * np.ones([2, 2]))) + + def test_three_level(self): + with fluid.imperative.guard(): + l = L3('test_three_level') + names = [p.name for p in l.parameters()] + ret = l() + self.assertEqual(names[0], "test_three_level/L3_0/L2_0/L1_0_0.w_0") + self.assertEqual(names[1], "test_three_level/L3_0/L2_0/L1_0_0.w_1") + self.assertEqual(names[2], "test_three_level/L3_0/L2_0/L1_1_0.w_0") + self.assertEqual(names[3], "test_three_level/L3_0/L2_0/L1_1_0.w_1") + self.assertEqual(names[4], "test_three_level/L3_0/L2_1/L1_0_0.w_0") + self.assertEqual(names[5], "test_three_level/L3_0/L2_1/L1_0_0.w_1") + self.assertTrue(np.allclose(ret._numpy(), 0.8 * np.ones([2, 2]))) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_beam_search_op.py b/python/paddle/fluid/tests/unittests/test_beam_search_op.py index c28dda4b53ce5d394ff11222e5df8d257b4e80da..1d9f4b78f30fefa21c189036c3731e0afe39ea9e 100644 --- a/python/paddle/fluid/tests/unittests/test_beam_search_op.py +++ b/python/paddle/fluid/tests/unittests/test_beam_search_op.py @@ -38,6 +38,7 @@ class BeamSearchOpTester(unittest.TestCase): self._create_pre_ids() self.scope.var('selected_ids') self.scope.var('selected_scores') + self.scope.var('parent_idx') def test_run(self): op = Operator( @@ -48,12 +49,14 @@ class BeamSearchOpTester(unittest.TestCase): scores='scores', selected_ids='selected_ids', selected_scores='selected_scores', + parent_idx='parent_idx', level=0, beam_size=2, end_id=0, ) op.run(self.scope, core.CPUPlace()) selected_ids = self.scope.find_var("selected_ids").get_tensor() selected_scores = self.scope.find_var("selected_scores").get_tensor() + parent_idx = self.scope.find_var("parent_idx").get_tensor() self.assertTrue( np.allclose( np.array(selected_ids), np.array([4, 2, 3, 8])[:, np.newaxis])) @@ -62,6 +65,8 @@ class BeamSearchOpTester(unittest.TestCase): np.array(selected_scores), np.array([0.5, 0.6, 0.9, 0.7])[:, np.newaxis])) self.assertEqual(selected_ids.lod(), [[0, 2, 4], [0, 1, 2, 3, 4]]) + self.assertTrue( + np.allclose(np.array(parent_idx), np.array([0, 1, 2, 3]))) def _create_pre_ids(self): np_data = np.array([[1, 2, 3, 4]], dtype='int64') diff --git a/python/paddle/fluid/tests/unittests/test_bilinear_interp_op.py b/python/paddle/fluid/tests/unittests/test_bilinear_interp_op.py index c8a7063dc1cd3e5cc7cd3458b51f5e74981aa75c..f60ed1d79ae5778f751d6101fde386ae3a90c0f7 100644 --- a/python/paddle/fluid/tests/unittests/test_bilinear_interp_op.py +++ b/python/paddle/fluid/tests/unittests/test_bilinear_interp_op.py @@ -20,7 +20,13 @@ from op_test import OpTest import paddle.fluid.core as core -def bilinear_interp_np(input, out_h, out_w, out_size=None, actual_shape=None): +def bilinear_interp_np(input, + out_h, + out_w, + out_size=None, + actual_shape=None, + align_corners=True, + align_mode=0): """bilinear interpolation implement in shape [N, C, H, W]""" if out_size is not None: out_h = out_size[0] @@ -29,25 +35,45 @@ def bilinear_interp_np(input, out_h, out_w, out_size=None, actual_shape=None): out_h = actual_shape[0] out_w = actual_shape[1] batch_size, channel, in_h, in_w = input.shape + + ratio_h = ratio_w = 0.0 if out_h > 1: - ratio_h = (in_h - 1.0) / (out_h - 1.0) - else: - ratio_h = 0.0 + if (align_corners): + ratio_h = (in_h - 1.0) / (out_h - 1.0) + else: + ratio_h = 1.0 * in_h / out_h if out_w > 1: - ratio_w = (in_w - 1.0) / (out_w - 1.0) - else: - ratio_w = 0.0 + if (align_corners): + ratio_w = (in_w - 1.0) / (out_w - 1.0) + else: + ratio_w = 1.0 * in_w / out_w out = np.zeros((batch_size, channel, out_h, out_w)) + for i in range(out_h): - h = int(ratio_h * i) + if (align_mode == 0 and not align_corners): + h = int(ratio_h * (i + 0.5) - 0.5) + else: + h = int(ratio_h * i) + + h = max(0, h) hid = 1 if h < in_h - 1 else 0 - h1lambda = ratio_h * i - h + if (align_mode == 0 and not align_corners): + h1lambda = ratio_h * (i + 0.5) - 0.5 - h + else: + h1lambda = ratio_h * i - h h2lambda = 1.0 - h1lambda for j in range(out_w): - w = int(ratio_w * j) + if (align_mode == 0 and not align_corners): + w = int(ratio_w * (j + 0.5) - 0.5) + else: + w = int(ratio_w * j) + w = max(0, w) wid = 1 if w < in_w - 1 else 0 - w1lambda = ratio_w * j - w + if (align_mode == 0 and not align_corners): + w1lambda = ratio_w * (j + 0.5) - 0.5 - w + else: + w1lambda = ratio_w * j - w w2lambda = 1.0 - w1lambda out[:, :, i, j] = h2lambda*(w2lambda*input[:, :, h, w] + @@ -66,7 +92,8 @@ class TestBilinearInterpOp(OpTest): input_np = np.random.random(self.input_shape).astype("float32") output_np = bilinear_interp_np(input_np, self.out_h, self.out_w, - self.out_size, self.actual_shape) + self.out_size, self.actual_shape, + self.align_corners, self.align_mode) self.inputs = {'X': input_np} if self.out_size is not None: self.inputs['OutSize'] = self.out_size @@ -75,7 +102,9 @@ class TestBilinearInterpOp(OpTest): self.attrs = { 'out_h': self.out_h, 'out_w': self.out_w, - 'interp_method': self.interp_method + 'interp_method': self.interp_method, + 'align_corners': self.align_corners, + 'align_mode': self.align_mode } self.outputs = {'Out': output_np} @@ -91,6 +120,8 @@ class TestBilinearInterpOp(OpTest): self.out_h = 2 self.out_w = 2 self.out_size = np.array([3, 3]).astype("int32") + self.align_corners = True + self.align_mode = 1 class TestBilinearInterpCase1(TestBilinearInterpOp): @@ -99,6 +130,8 @@ class TestBilinearInterpCase1(TestBilinearInterpOp): self.input_shape = [4, 1, 7, 8] self.out_h = 1 self.out_w = 1 + self.align_corners = True + self.align_mode = 1 class TestBilinearInterpCase2(TestBilinearInterpOp): @@ -107,6 +140,8 @@ class TestBilinearInterpCase2(TestBilinearInterpOp): self.input_shape = [3, 3, 9, 6] self.out_h = 12 self.out_w = 12 + self.align_corners = True + self.align_mode = 1 class TestBilinearInterpCase3(TestBilinearInterpOp): @@ -115,6 +150,8 @@ class TestBilinearInterpCase3(TestBilinearInterpOp): self.input_shape = [1, 1, 128, 64] self.out_h = 64 self.out_w = 128 + self.align_corners = True + self.align_mode = 1 class TestBilinearInterpCase4(TestBilinearInterpOp): @@ -124,6 +161,8 @@ class TestBilinearInterpCase4(TestBilinearInterpOp): self.out_h = 1 self.out_w = 1 self.out_size = np.array([2, 2]).astype("int32") + self.align_corners = True + self.align_mode = 1 class TestBilinearInterpCase5(TestBilinearInterpOp): @@ -133,6 +172,8 @@ class TestBilinearInterpCase5(TestBilinearInterpOp): self.out_h = 12 self.out_w = 12 self.out_size = np.array([11, 11]).astype("int32") + self.align_corners = True + self.align_mode = 1 class TestBilinearInterpCase6(TestBilinearInterpOp): @@ -142,6 +183,8 @@ class TestBilinearInterpCase6(TestBilinearInterpOp): self.out_h = 64 self.out_w = 128 self.out_size = np.array([65, 129]).astype("int32") + self.align_corners = True + self.align_mode = 1 class TestBilinearInterpActualShape(TestBilinearInterpOp): @@ -151,6 +194,8 @@ class TestBilinearInterpActualShape(TestBilinearInterpOp): self.out_h = 64 self.out_w = 32 self.out_size = np.array([66, 40]).astype("int32") + self.align_corners = True + self.align_mode = 1 class TestBilinearInterpOpUint8(OpTest): @@ -162,14 +207,17 @@ class TestBilinearInterpOpUint8(OpTest): input_np = np.random.randint( low=0, high=256, size=self.input_shape).astype("uint8") output_np = bilinear_interp_np(input_np, self.out_h, self.out_w, - self.out_size, self.actual_shape) + self.out_size, self.actual_shape, + self.align_corners, self.align_mode) self.inputs = {'X': input_np} if self.out_size is not None: self.inputs['OutSize'] = self.out_size self.attrs = { 'out_h': self.out_h, 'out_w': self.out_w, - 'interp_method': self.interp_method + 'interp_method': self.interp_method, + 'align_corners': self.align_corners, + 'align_mode': self.align_mode } self.outputs = {'Out': output_np} @@ -181,6 +229,8 @@ class TestBilinearInterpOpUint8(OpTest): self.input_shape = [1, 3, 9, 6] self.out_h = 10 self.out_w = 9 + self.align_corners = True + self.align_mode = 1 class TestBilinearInterpCase1Uint8(TestBilinearInterpOpUint8): @@ -189,6 +239,8 @@ class TestBilinearInterpCase1Uint8(TestBilinearInterpOpUint8): self.input_shape = [2, 3, 128, 64] self.out_h = 120 self.out_w = 50 + self.align_corners = True + self.align_mode = 1 class TestBilinearInterpCase2Uint8(TestBilinearInterpOpUint8): @@ -198,6 +250,26 @@ class TestBilinearInterpCase2Uint8(TestBilinearInterpOpUint8): self.out_h = 5 self.out_w = 13 self.out_size = np.array([6, 15]).astype("int32") + self.align_corners = True + self.align_mode = 1 + + +class TestBilinearInterpOtherMethod1(TestBilinearInterpOp): + def set_align_mode(self): + self.align_corners = False + self.align_mode = 1 + + +class TestBilinearInterpWithMethod2(TestBilinearInterpOp): + def set_align_mode(self): + self.align_corners = False + self.align_mode = 0 + + +class TestBilinearInterpWithMethod3(TestBilinearInterpOp): + def set_align_mode(self): + self.align_corners = True + self.align_mode = 0 if __name__ == "__main__": diff --git a/python/paddle/fluid/tests/unittests/test_box_clip_op.py b/python/paddle/fluid/tests/unittests/test_box_clip_op.py new file mode 100644 index 0000000000000000000000000000000000000000..b2b0598f31dd27e12e5ce329129129b5e0f1caf0 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_box_clip_op.py @@ -0,0 +1,70 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import print_function + +import unittest +import numpy as np +import sys +import math +from op_test import OpTest +import copy + + +def box_clip(input_box, im_info, output_box): + im_w = round(im_info[1] / im_info[2]) + im_h = round(im_info[0] / im_info[2]) + output_box[:, :, 0] = np.maximum( + np.minimum(input_box[:, :, 0], im_w - 1), 0) + output_box[:, :, 1] = np.maximum( + np.minimum(input_box[:, :, 1], im_h - 1), 0) + output_box[:, :, 2] = np.maximum( + np.minimum(input_box[:, :, 2], im_w - 1), 0) + output_box[:, :, 3] = np.maximum( + np.minimum(input_box[:, :, 3], im_h - 1), 0) + + +def batch_box_clip(input_boxes, im_info, lod): + n = input_boxes.shape[0] + m = input_boxes.shape[1] + output_boxes = np.zeros((n, m, 4), dtype=np.float32) + cur_offset = 0 + for i in range(len(lod)): + box_clip(input_boxes[cur_offset:(cur_offset + lod[i]), :, :], + im_info[i, :], + output_boxes[cur_offset:(cur_offset + lod[i]), :, :]) + cur_offset += lod[i] + return output_boxes + + +class TestBoxClipOp(OpTest): + def test_check_output(self): + self.check_output() + + def setUp(self): + self.op_type = "box_clip" + lod = [[1, 2, 3]] + input_boxes = np.random.random((6, 10, 4)) * 5 + im_info = np.array([[5, 8, 1.], [6, 6, 1.], [7, 5, 1.]]) + output_boxes = batch_box_clip(input_boxes, im_info, lod[0]) + + self.inputs = { + 'Input': (input_boxes.astype('float32'), lod), + 'ImInfo': im_info.astype('float32'), + } + self.outputs = {'Output': output_boxes} + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_box_coder_op.py b/python/paddle/fluid/tests/unittests/test_box_coder_op.py index 2511c5c22e012babdeb71a71d3546456ea2ceaf3..220bffebe83925c60af65aa9594ddd8a29c38145 100644 --- a/python/paddle/fluid/tests/unittests/test_box_coder_op.py +++ b/python/paddle/fluid/tests/unittests/test_box_coder_op.py @@ -21,80 +21,82 @@ import math from op_test import OpTest -def box_coder(target_box, prior_box, prior_box_var, output_box, code_type, - box_normalized): - prior_box_x = ( - (prior_box[:, 2] + prior_box[:, 0]) / 2).reshape(1, prior_box.shape[0]) - prior_box_y = ( - (prior_box[:, 3] + prior_box[:, 1]) / 2).reshape(1, prior_box.shape[0]) - prior_box_width = ( - (prior_box[:, 2] - prior_box[:, 0])).reshape(1, prior_box.shape[0]) - prior_box_height = ( - (prior_box[:, 3] - prior_box[:, 1])).reshape(1, prior_box.shape[0]) - prior_box_var = prior_box_var.reshape(1, prior_box_var.shape[0], - prior_box_var.shape[1]) - if not box_normalized: - prior_box_height = prior_box_height + 1 - prior_box_width = prior_box_width + 1 - - if (code_type == "EncodeCenterSize"): - target_box_x = ((target_box[:, 2] + target_box[:, 0]) / 2).reshape( - target_box.shape[0], 1) - target_box_y = ((target_box[:, 3] + target_box[:, 1]) / 2).reshape( - target_box.shape[0], 1) - target_box_width = ((target_box[:, 2] - target_box[:, 0])).reshape( - target_box.shape[0], 1) - target_box_height = ((target_box[:, 3] - target_box[:, 1])).reshape( - target_box.shape[0], 1) - if not box_normalized: - target_box_height = target_box_height + 1 - target_box_width = target_box_width + 1 - - output_box[:,:,0] = (target_box_x - prior_box_x) / prior_box_width / \ - prior_box_var[:,:,0] - output_box[:,:,1] = (target_box_y - prior_box_y) / prior_box_height / \ - prior_box_var[:,:,1] - output_box[:,:,2] = np.log(np.fabs(target_box_width / prior_box_width)) / \ - prior_box_var[:,:,2] - output_box[:,:,3] = np.log(np.fabs(target_box_height / prior_box_height)) / \ - prior_box_var[:,:,3] - - elif (code_type == "DecodeCenterSize"): - target_box_x = prior_box_var[:,:,0] * target_box[:,:,0] * \ - prior_box_width + prior_box_x - target_box_y = prior_box_var[:,:,1] * target_box[:,:,1] * \ - prior_box_height + prior_box_y - target_box_width = np.exp(prior_box_var[:,:,2] * target_box[:,:,2]) * \ - prior_box_width - target_box_height = np.exp(prior_box_var[:,:,3] * target_box[:,:,3]) * \ - prior_box_height - - output_box[:, :, 0] = target_box_x - target_box_width / 2 - output_box[:, :, 1] = target_box_y - target_box_height / 2 - output_box[:, :, 2] = target_box_x + target_box_width / 2 - output_box[:, :, 3] = target_box_y + target_box_height / 2 - if not box_normalized: - output_box[:, :, 2] = output_box[:, :, 2] - 1 - output_box[:, :, 3] = output_box[:, :, 3] - 1 - - -def batch_box_coder(prior_box, prior_box_var, target_box, lod, code_type, - box_normalized): - n = target_box.shape[0] - m = prior_box.shape[0] +def box_decoder(t_box, p_box, pb_v, output_box, norm, axis=0): + pb_w = p_box[:, 2] - p_box[:, 0] + (norm == False) + pb_h = p_box[:, 3] - p_box[:, 1] + (norm == False) + pb_x = pb_w * 0.5 + p_box[:, 0] + pb_y = pb_h * 0.5 + p_box[:, 1] + shape = (1, p_box.shape[0]) if axis == 0 else (p_box.shape[0], 1) + + pb_w = pb_w.reshape(shape) + pb_h = pb_h.reshape(shape) + pb_x = pb_x.reshape(shape) + pb_y = pb_y.reshape(shape) + + if pb_v.ndim == 2: + var_shape = (1, pb_v.shape[0], pb_v.shape[1]) if axis == 0 else ( + pb_v.shape[0], 1, pb_v.shape[1]) + pb_v = pb_v.reshape(var_shape) + if pb_v.ndim == 1: + tb_x = pb_v[0] * t_box[:, :, 0] * pb_w + pb_x + tb_y = pb_v[1] * t_box[:, :, 1] * pb_h + pb_y + tb_w = np.exp(pb_v[2] * t_box[:, :, 2]) * pb_w + tb_h = np.exp(pb_v[3] * t_box[:, :, 3]) * pb_h + else: + tb_x = pb_v[:, :, 0] * t_box[:, :, 0] * pb_w + pb_x + tb_y = pb_v[:, :, 1] * t_box[:, :, 1] * pb_h + pb_y + tb_w = np.exp(pb_v[:, :, 2] * t_box[:, :, 2]) * pb_w + tb_h = np.exp(pb_v[:, :, 3] * t_box[:, :, 3]) * pb_h + output_box[:, :, 0] = tb_x - tb_w / 2 + output_box[:, :, 1] = tb_y - tb_h / 2 + output_box[:, :, 2] = tb_x + tb_w / 2 - (not norm) + output_box[:, :, 3] = tb_y + tb_h / 2 - (not norm) + + +def box_encoder(t_box, p_box, pb_v, output_box, norm): + pb_w = p_box[:, 2] - p_box[:, 0] + (norm == False) + pb_h = p_box[:, 3] - p_box[:, 1] + (norm == False) + pb_x = pb_w * 0.5 + p_box[:, 0] + pb_y = pb_h * 0.5 + p_box[:, 1] + shape = (1, p_box.shape[0]) + + pb_w = pb_w.reshape(shape) + pb_h = pb_h.reshape(shape) + pb_x = pb_x.reshape(shape) + pb_y = pb_y.reshape(shape) + + if pb_v.ndim == 2: + pb_v = pb_v.reshape(1, pb_v.shape[0], pb_v.shape[1]) + tb_x = ((t_box[:, 2] + t_box[:, 0]) / 2).reshape(t_box.shape[0], 1) + tb_y = ((t_box[:, 3] + t_box[:, 1]) / 2).reshape(t_box.shape[0], 1) + tb_w = (t_box[:, 2] - t_box[:, 0]).reshape(t_box.shape[0], 1) + (not norm) + tb_h = (t_box[:, 3] - t_box[:, 1]).reshape(t_box.shape[0], 1) + (not norm) + if pb_v.ndim == 1: + output_box[:, :, 0] = (tb_x - pb_x) / pb_w / pb_v[0] + output_box[:, :, 1] = (tb_y - pb_y) / pb_h / pb_v[1] + output_box[:, :, 2] = np.log(np.fabs(tb_w / pb_w)) / pb_v[2] + output_box[:, :, 3] = np.log(np.fabs(tb_h / pb_h)) / pb_v[3] + else: + output_box[:, :, 0] = (tb_x - pb_x) / pb_w / pb_v[:, :, 0] + output_box[:, :, 1] = (tb_y - pb_y) / pb_h / pb_v[:, :, 1] + output_box[:, :, 2] = np.log(np.fabs(tb_w / pb_w)) / pb_v[:, :, 2] + output_box[:, :, 3] = np.log(np.fabs(tb_h / pb_h)) / pb_v[:, :, 3] + + +def batch_box_coder(p_box, pb_v, t_box, lod, code_type, norm, axis=0): + n = t_box.shape[0] + m = p_box.shape[0] + if code_type == "DecodeCenterSize": + m = t_box.shape[1] output_box = np.zeros((n, m, 4), dtype=np.float32) cur_offset = 0 for i in range(len(lod)): if (code_type == "EncodeCenterSize"): - box_coder(target_box[cur_offset:(cur_offset + lod[i]), :], - prior_box, prior_box_var, - output_box[cur_offset:(cur_offset + lod[i]), :, :], - code_type, box_normalized) + box_encoder(t_box[cur_offset:(cur_offset + lod[i]), :], p_box, pb_v, + output_box[cur_offset:(cur_offset + lod[i]), :, :], + norm) elif (code_type == "DecodeCenterSize"): - box_coder(target_box[cur_offset:(cur_offset + lod[i]), :, :], - prior_box, prior_box_var, - output_box[cur_offset:(cur_offset + lod[i]), :, :], - code_type, box_normalized) + box_decoder(t_box, p_box, pb_v, output_box, norm, axis) cur_offset += lod[i] return output_box @@ -106,14 +108,13 @@ class TestBoxCoderOp(OpTest): def setUp(self): self.op_type = "box_coder" lod = [[1, 1, 1, 1, 1]] - prior_box = np.random.random((10, 4)).astype('float32') - prior_box_var = np.random.random((10, 4)).astype('float32') - target_box = np.random.random((5, 10, 4)).astype('float32') + prior_box = np.random.random((81, 4)).astype('float32') + prior_box_var = np.random.random((81, 4)).astype('float32') + target_box = np.random.random((20, 81, 4)).astype('float32') code_type = "DecodeCenterSize" box_normalized = False output_box = batch_box_coder(prior_box, prior_box_var, target_box, lod[0], code_type, box_normalized) - self.inputs = { 'PriorBox': prior_box, 'PriorBoxVar': prior_box_var, @@ -133,9 +134,9 @@ class TestBoxCoderOpWithoutBoxVar(OpTest): def setUp(self): self.op_type = "box_coder" lod = [[0, 1, 2, 3, 4, 5]] - prior_box = np.random.random((10, 4)).astype('float32') - prior_box_var = np.ones((10, 4)).astype('float32') - target_box = np.random.random((5, 10, 4)).astype('float32') + prior_box = np.random.random((81, 4)).astype('float32') + prior_box_var = np.ones((81, 4)).astype('float32') + target_box = np.random.random((20, 81, 4)).astype('float32') code_type = "DecodeCenterSize" box_normalized = False output_box = batch_box_coder(prior_box, prior_box_var, target_box, @@ -158,10 +159,10 @@ class TestBoxCoderOpWithLoD(OpTest): def setUp(self): self.op_type = "box_coder" - lod = [[4, 8, 8]] - prior_box = np.random.random((10, 4)).astype('float32') - prior_box_var = np.random.random((10, 4)).astype('float32') - target_box = np.random.random((20, 4)).astype('float32') + lod = [[10, 20, 20]] + prior_box = np.random.random((20, 4)).astype('float32') + prior_box_var = np.random.random((20, 4)).astype('float32') + target_box = np.random.random((50, 4)).astype('float32') code_type = "EncodeCenterSize" box_normalized = True output_box = batch_box_coder(prior_box, prior_box_var, target_box, @@ -176,5 +177,63 @@ class TestBoxCoderOpWithLoD(OpTest): self.outputs = {'OutputBox': output_box} +class TestBoxCoderOpWithAxis(OpTest): + def test_check_output(self): + self.check_output() + + def setUp(self): + self.op_type = "box_coder" + lod = [[1, 1, 1, 1, 1]] + prior_box = np.random.random((30, 4)).astype('float32') + prior_box_var = np.random.random((30, 4)).astype('float32') + target_box = np.random.random((30, 81, 4)).astype('float32') + code_type = "DecodeCenterSize" + box_normalized = False + axis = 1 + output_box = batch_box_coder(prior_box, prior_box_var, target_box, + lod[0], code_type, box_normalized, axis) + + self.inputs = { + 'PriorBox': prior_box, + 'PriorBoxVar': prior_box_var, + 'TargetBox': target_box, + } + self.attrs = { + 'code_type': 'decode_center_size', + 'box_normalized': False, + 'axis': axis + } + self.outputs = {'OutputBox': output_box} + + +class TestBoxCoderOpWithVariance(OpTest): + def test_check_output(self): + self.check_output() + + def setUp(self): + self.op_type = "box_coder" + lod = [[1, 1, 1, 1, 1]] + prior_box = np.random.random((30, 4)).astype('float32') + prior_box_var = np.random.random((4)).astype('float32') + target_box = np.random.random((30, 81, 4)).astype('float32') + code_type = "DecodeCenterSize" + box_normalized = False + axis = 1 + output_box = batch_box_coder(prior_box, prior_box_var, target_box, + lod[0], code_type, box_normalized, axis) + + self.inputs = { + 'PriorBox': prior_box, + 'TargetBox': target_box, + } + self.attrs = { + 'code_type': 'decode_center_size', + 'box_normalized': False, + 'variance': prior_box_var.astype(np.float).flatten(), + 'axis': axis + } + self.outputs = {'OutputBox': output_box} + + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_dist_transpiler.py b/python/paddle/fluid/tests/unittests/test_dist_transpiler.py index 3566fed215229223f4d2ecd1bbb66cb297dd7716..12132477d28c74c7da718321140a3ddef784fc30 100644 --- a/python/paddle/fluid/tests/unittests/test_dist_transpiler.py +++ b/python/paddle/fluid/tests/unittests/test_dist_transpiler.py @@ -22,6 +22,9 @@ import six import unittest import numpy as np +import gc +gc.set_debug(gc.DEBUG_COLLECTABLE) + import paddle.fluid as fluid @@ -99,6 +102,12 @@ class TranspilerTest(unittest.TestCase): with fluid.unique_name.guard(): with fluid.program_guard(main, startup): self.transpiler_test_impl() + # NOTE: run gc.collect to eliminate pybind side objects to + # prevent random double-deallocate when inherited in python. + del self.transpiler + del main + del startup + gc.collect() class TestBasicModel(TranspilerTest): @@ -797,6 +806,7 @@ class TestNCCL2Transpile(TranspilerTest): print([op.type for op in startup.global_block().ops]) self.assertEqual(startup.global_block().ops[-1].type, "gen_nccl_id") self.assertIsNotNone(startup.global_block().vars.get("NCCLID")) + gc.collect() else: pass diff --git a/python/paddle/fluid/tests/unittests/test_eager_deletion_transformer.py b/python/paddle/fluid/tests/unittests/test_eager_deletion_transformer.py index 754d5fd40953311a5deb466fa42216f72671a65a..603c8e74885d2a050e6e1e3101dce880b6eabe9c 100644 --- a/python/paddle/fluid/tests/unittests/test_eager_deletion_transformer.py +++ b/python/paddle/fluid/tests/unittests/test_eager_deletion_transformer.py @@ -16,12 +16,10 @@ import os import unittest os.environ['FLAGS_eager_delete_tensor_gb'] = "0.0" -from test_parallel_executor_transformer import TestTransformer - - -class EagerDeletionTestTransformer(TestTransformer): - pass +os.environ[ + 'RECORDIO_FILENAME'] = '/tmp/eager_deletion_transformer.wmt16.recordio' +from test_parallel_executor_transformer import TestTransformer if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_expand_op.py b/python/paddle/fluid/tests/unittests/test_expand_op.py index 67a8d8f0721c2c75b432d68d64be8fc1035ffc74..690875662e666aab63ac5eb62df0fb52823b8dff 100644 --- a/python/paddle/fluid/tests/unittests/test_expand_op.py +++ b/python/paddle/fluid/tests/unittests/test_expand_op.py @@ -109,5 +109,32 @@ class TestExpandOpRank4(OpTest): self.check_grad(['X'], 'Out') +class TestExpandOpInteger(OpTest): + def setUp(self): + self.op_type = "expand" + self.inputs = { + 'X': np.random.randint( + 10, size=(2, 4, 5)).astype("int32") + } + self.attrs = {'expand_times': [2, 1, 4]} + output = np.tile(self.inputs['X'], (2, 1, 4)) + self.outputs = {'Out': output} + + def test_check_output(self): + self.check_output() + + +class TestExpandOpBoolean(OpTest): + def setUp(self): + self.op_type = "expand" + self.inputs = {'X': np.random.randint(2, size=(2, 4, 5)).astype("bool")} + self.attrs = {'expand_times': [2, 1, 4]} + output = np.tile(self.inputs['X'], (2, 1, 4)) + self.outputs = {'Out': output} + + def test_check_output(self): + self.check_output() + + if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_fuse_elewise_add_act_pass.py b/python/paddle/fluid/tests/unittests/test_fuse_elewise_add_act_pass.py index 03471a4432f2b6bf6220e79e99aa506628b1535b..c1fb53ecf52d953fa470998c120930b2bec6325b 100644 --- a/python/paddle/fluid/tests/unittests/test_fuse_elewise_add_act_pass.py +++ b/python/paddle/fluid/tests/unittests/test_fuse_elewise_add_act_pass.py @@ -121,6 +121,8 @@ class TestMNIST(TestParallelExecutorBase): regularization=fluid.regularizer.L2Decay(1e-6)) return optimizer + # NOTE(dzh): + # need to make it compatible with elewise fuse act not_fuse_op_first_loss, not_fuse_op_last_loss = self.check_network_convergence( model, feed_dict={"image": img, @@ -128,6 +130,7 @@ class TestMNIST(TestParallelExecutorBase): use_cuda=use_cuda, fuse_elewise_add_act_ops=False, memory_opt=False, + use_ir_memory_optimize=False, optimizer=_optimizer) fuse_op_first_loss, fuse_op_last_loss = self.check_network_convergence( model, @@ -136,6 +139,7 @@ class TestMNIST(TestParallelExecutorBase): use_cuda=use_cuda, fuse_elewise_add_act_ops=True, memory_opt=False, + use_ir_memory_optimize=False, optimizer=_optimizer) for loss in zip(not_fuse_op_first_loss, fuse_op_first_loss): diff --git a/python/paddle/fluid/tests/unittests/test_imperative.py b/python/paddle/fluid/tests/unittests/test_imperative.py index 7533ab9fdbff3d3f44b4eef88b7c805ff51cae95..dae0c466ee5ea919688b29100f77f17f5f3b8c6d 100644 --- a/python/paddle/fluid/tests/unittests/test_imperative.py +++ b/python/paddle/fluid/tests/unittests/test_imperative.py @@ -15,7 +15,6 @@ import contextlib import unittest import numpy as np -import sys import paddle.fluid as fluid from paddle.fluid import core @@ -24,8 +23,8 @@ from test_imperative_base import new_program_scope class MyLayer(fluid.imperative.Layer): - def __init__(self): - super(MyLayer, self).__init__() + def __init__(self, name_scope): + super(MyLayer, self).__init__(name_scope) def forward(self, inputs): x = fluid.layers.relu(inputs) @@ -50,12 +49,14 @@ class MyPyLayer(fluid.imperative.PyLayer): class MLP(fluid.imperative.Layer): - def __init__(self): - super(MLP, self).__init__() - self._fc1 = FC(3, + def __init__(self, name_scope): + super(MLP, self).__init__(name_scope) + self._fc1 = FC(self.full_name(), + 3, fluid.ParamAttr( initializer=fluid.initializer.Constant(value=0.1))) - self._fc2 = FC(4, + self._fc2 = FC(self.full_name(), + 4, fluid.ParamAttr( initializer=fluid.initializer.Constant(value=0.1))) @@ -66,12 +67,148 @@ class MLP(fluid.imperative.Layer): return x +class SimpleRNNCell(fluid.imperative.Layer): + def __init__(self, name_scope, step_input_size, hidden_size, output_size, + param_attr): + super(SimpleRNNCell, self).__init__(name_scope) + self.step_input_size = step_input_size + self.hidden_size = hidden_size + self.output_size = output_size + self._dype = core.VarDesc.VarType.FP32 + from paddle.fluid.layer_helper import LayerHelper + self._helper = LayerHelper( + 'SimpleRNNCell', act="tanh", param_attr=param_attr) + + def _build_once(self, inputs, pre_hidden): + i2h_param_shape = [self.step_input_size, self.hidden_size] + h2h_param_shape = [self.hidden_size, self.hidden_size] + h2o_param_shape = [self.output_size, self.hidden_size] + self._i2h_w = self._helper.create_parameter( + attr=self._helper.param_attr, + shape=i2h_param_shape, + dtype=self._dtype, + is_bias=False) + self._h2h_w = self._helper.create_parameter( + attr=self._helper.param_attr, + shape=h2h_param_shape, + dtype=self._dtype, + is_bias=False) + self._h2o_w = self._helper.create_parameter( + attr=self._helper.param_attr, + shape=h2o_param_shape, + dtype=self._dtype, + is_bias=False) + + def forward(self, input, pre_hidden): + + tmp_i2h = self._helper.create_variable_for_type_inference(self._dtype) + tmp_h2h = self._helper.create_variable_for_type_inference(self._dtype) + hidden = self._helper.create_variable_for_type_inference(self._dype) + out = self._helper.create_variable_for_type_inference(self._dype) + softmax_out = self._helper.create_variable_for_type_inference( + self._dtype) + reduce_out = self._helper.create_variable_for_type_inference( + self._dtype) + self._helper.append_op( + type="mul", + inputs={"X": input, + "Y": self._i2h_w}, + outputs={"Out": tmp_i2h}, + attrs={"x_num_col_dims": 1, + "y_num_col_dims": 1}) + + self._helper.append_op( + type="mul", + inputs={"X": pre_hidden, + "Y": self._h2h_w}, + outputs={"Out": tmp_h2h}, + attrs={"x_num_col_dims": 1, + "y_num_col_dims": 1}) + + self._helper.append_op( + type="elementwise_add", + inputs={'X': tmp_h2h, + 'Y': tmp_i2h}, + outputs={'Out': hidden}, + attrs={'axis': -1, + 'use_mkldnn': False}) + hidden = self._helper.append_activation(hidden) + + self._helper.append_op( + type="mul", + inputs={"X": hidden, + "Y": self._h2o_w}, + outputs={"Out": out}, + attrs={"x_num_col_dims": 1, + "y_num_col_dims": 1}) + + self._helper.append_op( + type="softmax", + inputs={"X": out}, + outputs={"Out": softmax_out}, + attrs={"use_cudnn": False}) + + self._helper.append_op( + type='reduce_sum', + inputs={'X': softmax_out}, + outputs={'Out': reduce_out}, + attrs={'dim': None, + 'keep_dim': False, + 'reduce_all': True}) + + return reduce_out, hidden + + +class SimpleRNN(fluid.imperative.Layer): + def __init__(self, name_scope): + super(SimpleRNN, self).__init__(name_scope) + self.seq_len = 4 + self._cell = SimpleRNNCell( + self.full_name(), + 3, + 3, + 3, + fluid.ParamAttr(initializer=fluid.initializer.Constant(value=0.1))) + + def forward(self, inputs): + outs = list() + pre_hiddens = list() + + init_hidden = fluid.layers.tensor.create_parameter( + attr=fluid.ParamAttr( + initializer=fluid.initializer.Constant(value=0.1)), + shape=[1, 3], + dtype='float32', + is_bias=False) + pre_hidden = init_hidden + for i in range(self.seq_len): + input = fluid.layers.slice( + inputs, axes=[1], starts=[i], ends=[i + 1]) + input = fluid.layers.reshape(input, shape=[1, 3]) + out_softmax, pre_hidden = self._cell(input, pre_hidden) + outs.append(out_softmax) + + return outs, pre_hiddens + + class TestImperative(unittest.TestCase): + def test_sum_op(self): + x = np.ones([2, 2], np.float32) + with fluid.imperative.guard(): + inputs = [] + for _ in range(10): + inputs.append(fluid.imperative.base.to_variable(x)) + ret = fluid.layers.sums(inputs) + loss = fluid.layers.reduce_sum(ret) + loss._backward() + self.assertTrue(np.allclose(ret._numpy(), x * 10)) + self.assertTrue(np.allclose(inputs[0]._gradient(), x)) + def test_layer(self): with fluid.imperative.guard(): cl = core.Layer() cl.forward([]) - l = fluid.imperative.Layer() + l = fluid.imperative.Layer("l") self.assertRaises(NotImplementedError, l.forward, []) def test_pylayer_func_id(self): @@ -147,7 +284,7 @@ class TestImperative(unittest.TestCase): np_inp = np.array([1.0, 2.0, -1.0], dtype=np.float32) with fluid.imperative.guard(): var_inp = fluid.imperative.base.to_variable(np_inp) - l = MyLayer() + l = MyLayer("my_layer") x = l(var_inp)[0] self.assertIsNotNone(x) dy_out = x._numpy() @@ -157,7 +294,7 @@ class TestImperative(unittest.TestCase): with new_program_scope(): inp = fluid.layers.data( name="inp", shape=[3], append_batch_size=False) - l = MyLayer() + l = MyLayer("my_layer") x = l(inp)[0] param_grads = fluid.backward.append_backward( x, parameter_list=[l._x_for_debug.name])[0] @@ -175,7 +312,7 @@ class TestImperative(unittest.TestCase): np_inp = np.array([[1.0, 2.0], [3.0, 4.0]], dtype=np.float32) with fluid.imperative.guard(): var_inp = fluid.imperative.base.to_variable(np_inp) - mlp = MLP() + mlp = MLP("mlp") out = mlp(var_inp) dy_out = out._numpy() out._backward() @@ -184,7 +321,7 @@ class TestImperative(unittest.TestCase): with new_program_scope(): inp = fluid.layers.data( name="inp", shape=[2, 2], append_batch_size=False) - mlp = MLP() + mlp = MLP("mlp") out = mlp(inp) param_grads = fluid.backward.append_backward( out, parameter_list=[mlp._fc1._w.name])[0] @@ -199,6 +336,53 @@ class TestImperative(unittest.TestCase): self.assertTrue(np.allclose(dy_out, static_out)) self.assertTrue(np.allclose(dy_grad, static_grad)) + params = mlp.parameters(True) + self.assertEqual("mlp/MLP_0/FC_0_0.w_0", params[0].name) + self.assertEqual("mlp/MLP_0/FC_0_0.b_0", params[1].name) + self.assertEqual("mlp/MLP_0/FC_1_0.w_0", params[2].name) + self.assertEqual("mlp/MLP_0/FC_1_0.b_0", params[3].name) + self.assertEqual(len(params), 4) + + sublayers = mlp.sublayers(True) + self.assertEqual(mlp._fc1, sublayers[0]) + self.assertEqual(mlp._fc2, sublayers[1]) + self.assertEqual(len(sublayers), 2) + + def test_rnn(self): + np_inp = np.array([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0], [7.0, 8.0, 9.0], + [10.0, 11.0, 12.0]]) + np_inp = np_inp.reshape((1, 4, 3)) + np_inp = np_inp.astype(np.float32) + with fluid.imperative.guard(): + var_inp = fluid.imperative.base.to_variable(np_inp) + var_inp = fluid.layers.reshape(var_inp, shape=[1, 4, 3]) + simple_rnn = SimpleRNN("simple_rnn") + outs, pre_hiddens = simple_rnn.forward(var_inp) + dy_out = outs[3]._numpy() + outs[3]._backward() + dy_grad_h2o = simple_rnn._cell._h2o_w._gradient() + dy_grad_h2h = simple_rnn._cell._h2h_w._gradient() + dy_grad_i2h = simple_rnn._cell._i2h_w._gradient() + + with new_program_scope(): + inp = fluid.layers.data( + name="inp", shape=[1, 4, 3], append_batch_size=False) + simple_rnn = SimpleRNN("simple_rnn") + outs, pre_hiddens = simple_rnn(inp) + param_grads = fluid.backward.append_backward(outs[3]) + exe = fluid.Executor(fluid.CPUPlace()) + exe.run(fluid.default_startup_program()) + static_out, static_grad_h2o, static_grad_h2h, static_grad_i2h = exe.run( + feed={inp.name: np_inp}, + fetch_list=[ + outs[3].name, param_grads[0][1].name, + param_grads[1][1].name, param_grads[2][1].name + ]) + self.assertTrue(np.allclose(dy_out, static_out)) + self.assertTrue(np.allclose(dy_grad_h2o, static_grad_h2o)) + self.assertTrue(np.allclose(dy_grad_h2h, static_grad_h2h)) + self.assertTrue(np.allclose(dy_grad_i2h, static_grad_i2h)) + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_imperative_gan.py b/python/paddle/fluid/tests/unittests/test_imperative_gan.py index 681661bfc63db95653be371688a047efe96f3866..a80202d6dddacaa4cb6fa3efd3c3dfd5b0ab4400 100644 --- a/python/paddle/fluid/tests/unittests/test_imperative_gan.py +++ b/python/paddle/fluid/tests/unittests/test_imperative_gan.py @@ -28,13 +28,10 @@ from paddle.fluid.imperative.base import to_variable class Discriminator(fluid.imperative.Layer): - def __init__(self): - super(Discriminator, self).__init__() - self._fc1 = FC(size=32, act='elu', name="d_fc1") - self._fc2 = FC(size=1, name="d_fc2") - - def parameters(self): - return self._fc1.parameters() + self._fc2.parameters() + def __init__(self, name_scope): + super(Discriminator, self).__init__(name_scope) + self._fc1 = FC(self.full_name(), size=32, act='elu') + self._fc2 = FC(self.full_name(), size=1) def forward(self, inputs): x = self._fc1(inputs) @@ -42,15 +39,11 @@ class Discriminator(fluid.imperative.Layer): class Generator(fluid.imperative.Layer): - def __init__(self): - super(Generator, self).__init__() - self._fc1 = FC(size=64, act='elu', name="g_fc1") - self._fc2 = FC(size=64, act='elu', name="g_fc2") - self._fc3 = FC(size=1, name="g_fc3") - - def parameters(self): - return self._fc1.parameters() + self._fc2.parameters( - ) + self._fc3.parameters() + def __init__(self, name_scope): + super(Generator, self).__init__(name_scope) + self._fc1 = FC(self.full_name(), size=64, act='elu') + self._fc2 = FC(self.full_name(), size=64, act='elu') + self._fc3 = FC(self.full_name(), size=1) def forward(self, inputs): x = self._fc1(inputs) @@ -72,8 +65,8 @@ class TestImperativeMnist(unittest.TestCase): scope = fluid.core.Scope() with new_program_scope( main=discriminate_p, startup=startup, scope=scope): - discriminator = Discriminator() - generator = Generator() + discriminator = Discriminator("d") + generator = Generator("g") img = fluid.layers.data( name="img", shape=[2, 1], append_batch_size=False) @@ -100,8 +93,8 @@ class TestImperativeMnist(unittest.TestCase): sgd.minimize(d_loss) with new_program_scope(main=generate_p, startup=startup, scope=scope): - discriminator = Discriminator() - generator = Generator() + discriminator = Discriminator("d") + generator = Generator("g") noise = fluid.layers.data( name="noise", shape=[2, 2], append_batch_size=False) @@ -141,8 +134,8 @@ class TestImperativeMnist(unittest.TestCase): fluid.default_startup_program().random_seed = seed fluid.default_main_program().random_seed = seed - discriminator = Discriminator() - generator = Generator() + discriminator = Discriminator("d") + generator = Generator("g") sgd = SGDOptimizer(learning_rate=1e-3) d_real = discriminator(to_variable(np.ones([2, 1], np.float32))) diff --git a/python/paddle/fluid/tests/unittests/test_imperative_optimizer.py b/python/paddle/fluid/tests/unittests/test_imperative_optimizer.py index d0a5a883174cb33a035b344f9489b2ba02ba99f1..780c6a6be567c9f60f472c27cebd5300d56eb378 100644 --- a/python/paddle/fluid/tests/unittests/test_imperative_optimizer.py +++ b/python/paddle/fluid/tests/unittests/test_imperative_optimizer.py @@ -28,6 +28,7 @@ from test_imperative_base import new_program_scope class SimpleImgConvPool(fluid.imperative.Layer): def __init__(self, + name_scope, num_channels, num_filters, filter_size, @@ -44,9 +45,10 @@ class SimpleImgConvPool(fluid.imperative.Layer): use_cudnn=False, param_attr=None, bias_attr=None): - super(SimpleImgConvPool, self).__init__() + super(SimpleImgConvPool, self).__init__(name_scope) self._conv2d = Conv2D( + self.full_name(), num_channels=num_channels, num_filters=num_filters, filter_size=filter_size, @@ -59,6 +61,7 @@ class SimpleImgConvPool(fluid.imperative.Layer): use_cudnn=use_cudnn) self._pool2d = Pool2D( + self.full_name(), pool_size=pool_size, pool_type=pool_type, pool_stride=pool_stride, @@ -73,22 +76,24 @@ class SimpleImgConvPool(fluid.imperative.Layer): class MNIST(fluid.imperative.Layer): - def __init__(self, param_attr=None, bias_attr=None): - super(MNIST, self).__init__() + def __init__(self, name_scope, param_attr=None, bias_attr=None): + super(MNIST, self).__init__(name_scope) self._simple_img_conv_pool_1 = SimpleImgConvPool( - 1, 20, 5, 2, 2, act="relu") + self.full_name(), 1, 20, 5, 2, 2, act="relu") self._simple_img_conv_pool_2 = SimpleImgConvPool( - 20, 50, 5, 2, 2, act="relu") + self.full_name(), 20, 50, 5, 2, 2, act="relu") - pool_2_shape = 50 * 8 * 8 + pool_2_shape = 50 * 4 * 4 SIZE = 10 scale = (2.0 / (pool_2_shape**2 * SIZE))**0.5 - self._fc = FC(10, + self._fc = FC(self.full_name(), + 10, param_attr=fluid.param_attr.ParamAttr( initializer=fluid.initializer.NormalInitializer( - loc=0.0, scale=scale))) + loc=0.0, scale=scale)), + act="softmax") def forward(self, inputs): x = self._simple_img_conv_pool_1(inputs) @@ -98,29 +103,29 @@ class MNIST(fluid.imperative.Layer): class TestImperativeMnist(unittest.TestCase): - def test_mnist_cpu_float32(self): + def test_mnist_float32(self): seed = 90 - + batch_num = 2 with fluid.imperative.guard(): fluid.default_startup_program().random_seed = seed fluid.default_main_program().random_seed = seed - mnist = MNIST() + mnist = MNIST("mnist") sgd = SGDOptimizer(learning_rate=1e-3) train_reader = paddle.batch( paddle.dataset.mnist.train(), batch_size=128) dy_param_init_value = {} for batch_id, data in enumerate(train_reader()): - if batch_id >= 2: + if batch_id >= batch_num: break - x_data = np.array( + dy_x_data = np.array( [x[0].reshape(1, 28, 28) for x in data]).astype('float32') y_data = np.array([x[1] for x in data]).astype('int64').reshape( 128, 1) - img = to_variable(x_data) + img = to_variable(dy_x_data) label = to_variable(y_data) label._stop_gradient = True @@ -136,6 +141,7 @@ class TestImperativeMnist(unittest.TestCase): avg_loss._backward() sgd.minimize(avg_loss) + mnist.clear_gradients() dy_param_value = {} for param in fluid.default_main_program().global_block( ).all_parameters(): @@ -148,7 +154,7 @@ class TestImperativeMnist(unittest.TestCase): exe = fluid.Executor(fluid.CPUPlace( ) if not core.is_compiled_with_cuda() else fluid.CUDAPlace(0)) - mnist = MNIST() + mnist = MNIST("mnist") sgd = SGDOptimizer(learning_rate=1e-3) train_reader = paddle.batch( paddle.dataset.mnist.train(), batch_size=128) @@ -175,10 +181,10 @@ class TestImperativeMnist(unittest.TestCase): static_param_init_value[static_param_name_list[i]] = out[i] for batch_id, data in enumerate(train_reader()): - if batch_id >= 2: + if batch_id >= batch_num: break - x_data = np.array( + static_x_data = np.array( [x[0].reshape(1, 28, 28) for x in data]).astype('float32') y_data = np.array([x[1] for x in data]).astype('int64').reshape( [128, 1]) @@ -186,7 +192,7 @@ class TestImperativeMnist(unittest.TestCase): fetch_list = [avg_loss.name] fetch_list.extend(static_param_name_list) out = exe.run(fluid.default_main_program(), - feed={"pixel": x_data, + feed={"pixel": static_x_data, "label": y_data}, fetch_list=fetch_list) @@ -196,11 +202,12 @@ class TestImperativeMnist(unittest.TestCase): static_param_value[static_param_name_list[i - 1]] = out[i] for key, value in six.iteritems(static_param_init_value): - self.assertTrue( - np.allclose(value.all(), dy_param_init_value[key].all())) - self.assertTrue(np.allclose(static_out.all(), dy_out.all())) + self.assertTrue(np.allclose(value, dy_param_init_value[key])) + + self.assertTrue(np.allclose(static_out, dy_out)) + for key, value in six.iteritems(static_param_value): - self.assertTrue(np.allclose(value.all(), dy_param_value[key].all())) + self.assertTrue(np.allclose(value, dy_param_value[key])) if __name__ == '__main__': diff --git a/python/paddle/fluid/tests/unittests/test_imperative_ptb_rnn.py b/python/paddle/fluid/tests/unittests/test_imperative_ptb_rnn.py new file mode 100644 index 0000000000000000000000000000000000000000..c8e42d5ede57896b0d5c09a2334709ced2d16a3f --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_imperative_ptb_rnn.py @@ -0,0 +1,347 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import print_function + +import unittest +import paddle.fluid as fluid +from paddle.fluid.imperative.nn import Embedding +import paddle.fluid.framework as framework +from paddle.fluid.optimizer import SGDOptimizer +from paddle.fluid.imperative.base import to_variable +from test_imperative_base import new_program_scope +import numpy as np +import six +from paddle.fluid.backward import append_backward + + +class SimpleLSTMRNN(fluid.imperative.Layer): + def __init__(self, + name_scope, + hidden_size, + num_steps, + num_layers=2, + init_scale=0.1, + dropout=None): + super(SimpleLSTMRNN, self).__init__(name_scope) + self._hidden_size = hidden_size + self._num_layers = num_layers + self._init_scale = init_scale + self._dropout = dropout + self._input = None + self._num_steps = num_steps + from paddle.fluid.layer_helper import LayerHelper + self._helper = LayerHelper('SimpleLSTMRNN', act="tanh") + + def _build_once(self, input_embedding, init_hidden=None, init_cell=None): + self.weight_1_arr = [] + self.weight_2_arr = [] + self.bias_arr = [] + self.hidden_array = [] + self.cell_array = [] + self.mask_array = [] + + for i in range(self._num_layers): + weight_1 = self._helper.create_parameter( + attr=fluid.ParamAttr( + initializer=fluid.initializer.UniformInitializer( + low=-self._init_scale, high=self._init_scale)), + shape=[self._hidden_size * 2, self._hidden_size * 4], + dtype="float32", + default_initializer=fluid.initializer.UniformInitializer( + low=-self._init_scale, high=self._init_scale)) + self.weight_1_arr.append(weight_1) + bias_1 = self._helper.create_parameter( + attr=fluid.ParamAttr( + initializer=fluid.initializer.UniformInitializer( + low=-self._init_scale, high=self._init_scale)), + shape=[self._hidden_size * 4], + dtype="float32", + default_initializer=fluid.initializer.Constant(0.0)) + self.bias_arr.append(bias_1) + + pre_hidden = fluid.layers.slice( + init_hidden, axes=[0], starts=[i], ends=[i + 1]) + pre_cell = fluid.layers.slice( + init_cell, axes=[0], starts=[i], ends=[i + 1]) + pre_hidden = fluid.layers.reshape( + pre_hidden, shape=[-1, self._hidden_size]) + pre_cell = fluid.layers.reshape( + pre_cell, shape=[-1, self._hidden_size]) + self.hidden_array.append(pre_hidden) + self.cell_array.append(pre_cell) + + def forward(self, input_embedding, init_hidden=None, init_cell=None): + res = [] + for index in range(self._num_steps): + self._input = fluid.layers.slice( + input_embedding, axes=[1], starts=[index], ends=[index + 1]) + self._input = fluid.layers.reshape( + self._input, shape=[-1, self._hidden_size]) + for k in range(self._num_layers): + pre_hidden = self.hidden_array[k] + pre_cell = self.cell_array[k] + weight_1 = self.weight_1_arr[k] + bias = self.bias_arr[k] + + nn = fluid.layers.concat([self._input, pre_hidden], 1) + gate_input = fluid.layers.matmul(x=nn, y=weight_1) + + gate_input = fluid.layers.elementwise_add(gate_input, bias) + i, j, f, o = fluid.layers.split( + gate_input, num_or_sections=4, dim=-1) + c = pre_cell * fluid.layers.sigmoid(f) + fluid.layers.sigmoid( + i) * fluid.layers.tanh(j) + m = fluid.layers.tanh(c) * fluid.layers.sigmoid(o) + self.hidden_array[k] = m + self.cell_array[k] = c + self._input = m + + if self._dropout is not None and self._dropout > 0.0: + self._input = fluid.layers.dropout( + self._input, + dropout_prob=self._dropout, + dropout_implementation='upscale_in_train') + res.append( + fluid.layers.reshape( + self._input, shape=[1, -1, self._hidden_size])) + real_res = fluid.layers.concat(res, 0) + real_res = fluid.layers.transpose(x=real_res, perm=[1, 0, 2]) + last_hidden = fluid.layers.concat(self.hidden_array, 1) + last_hidden = fluid.layers.reshape( + last_hidden, shape=[-1, self._num_layers, self._hidden_size]) + last_hidden = fluid.layers.transpose(x=last_hidden, perm=[1, 0, 2]) + last_cell = fluid.layers.concat(self.cell_array, 1) + last_cell = fluid.layers.reshape( + last_cell, shape=[-1, self._num_layers, self._hidden_size]) + last_cell = fluid.layers.transpose(x=last_cell, perm=[1, 0, 2]) + return real_res, last_hidden, last_cell + + +class PtbModel(fluid.imperative.Layer): + def __init__(self, + name_scope, + hidden_size, + vocab_size, + num_layers=2, + num_steps=20, + init_scale=0.1, + dropout=None): + super(PtbModel, self).__init__(name_scope) + self.hidden_size = hidden_size + self.vocab_size = vocab_size + self.init_scale = init_scale + self.num_layers = num_layers + self.num_steps = num_steps + self.dropout = dropout + from paddle.fluid.layer_helper import LayerHelper + self._helper = LayerHelper('PtbModel', act="tanh") + self.simple_lstm_rnn = SimpleLSTMRNN( + self.full_name(), + hidden_size, + num_steps, + num_layers=num_layers, + init_scale=init_scale, + dropout=dropout) + self.embedding = Embedding( + self.full_name(), + size=[vocab_size, hidden_size], + dtype='float32', + is_sparse=False, + param_attr=fluid.ParamAttr( + name='embedding_para', + initializer=fluid.initializer.UniformInitializer( + low=-init_scale, high=init_scale))) + self.softmax_weight = self._helper.create_parameter( + attr=fluid.ParamAttr(), + shape=[self.hidden_size, self.vocab_size], + dtype="float32", + default_initializer=fluid.initializer.UniformInitializer( + low=-self.init_scale, high=self.init_scale)) + self.softmax_bias = self._helper.create_parameter( + attr=fluid.ParamAttr(), + shape=[self.vocab_size], + dtype="float32", + default_initializer=fluid.initializer.UniformInitializer( + low=-self.init_scale, high=self.init_scale)) + + def _build_once(self, input, label, init_hidden, init_cell): + pass + + def forward(self, input, label, init_hidden, init_cell): + + init_h = fluid.layers.reshape( + init_hidden, shape=[self.num_layers, -1, self.hidden_size]) + + init_c = fluid.layers.reshape( + init_cell, shape=[self.num_layers, -1, self.hidden_size]) + + x_emb = self.embedding(input) + x_emb = fluid.layers.reshape( + x_emb, shape=[-1, self.num_steps, self.hidden_size]) + if self.dropout is not None and self.dropout > 0.0: + x_emb = fluid.layers.dropout( + x_emb, + dropout_prob=self.drop_out, + dropout_implementation='upscale_in_train') + rnn_out, last_hidden, last_cell = self.simple_lstm_rnn(x_emb, init_h, + init_c) + rnn_out = fluid.layers.reshape( + rnn_out, shape=[-1, self.num_steps, self.hidden_size]) + projection = fluid.layers.matmul(rnn_out, self.softmax_weight) + projection = fluid.layers.elementwise_add(projection, self.softmax_bias) + projection = fluid.layers.reshape( + projection, shape=[-1, self.vocab_size]) + projection = fluid.layers.reshape( + projection, shape=[-1, self.vocab_size]) + loss = fluid.layers.softmax_with_cross_entropy( + logits=projection, label=label, soft_label=False) + loss = fluid.layers.reshape(loss, shape=[-1, self.num_steps]) + loss = fluid.layers.reduce_mean(loss, dim=[0]) + loss = fluid.layers.reduce_sum(loss) + loss.permissions = True + + return loss, last_hidden, last_cell + + +class TestImperativePtbRnn(unittest.TestCase): + def test_ptb_rnn_cpu_float32(self): + seed = 90 + hidden_size = 10 + vocab_size = 1000 + num_layers = 1 + num_steps = 3 + init_scale = 0.1 + batch_size = 4 + + with fluid.imperative.guard(): + fluid.default_startup_program().random_seed = seed + fluid.default_main_program().random_seed = seed + # TODO: marsyang1993 Change seed to + ptb_model = PtbModel( + "ptb_model", + hidden_size=hidden_size, + vocab_size=vocab_size, + num_layers=num_layers, + num_steps=num_steps, + init_scale=init_scale) + + sgd = SGDOptimizer(learning_rate=1e-3) + dy_param_updated = dict() + dy_param_init = dict() + dy_loss = None + last_hidden = None + last_cell = None + for i in range(2): + x_data = np.arange(12).reshape(4, 3).astype('int64') + y_data = np.arange(1, 13).reshape(4, 3).astype('int64') + x_data = x_data.reshape((-1, num_steps, 1)) + y_data = y_data.reshape((-1, 1)) + init_hidden_data = np.zeros( + (num_layers, batch_size, hidden_size), dtype='float32') + init_cell_data = np.zeros( + (num_layers, batch_size, hidden_size), dtype='float32') + x = to_variable(x_data) + y = to_variable(y_data) + init_hidden = to_variable(init_hidden_data) + init_cell = to_variable(init_cell_data) + dy_loss, last_hidden, last_cell = ptb_model(x, y, init_hidden, + init_cell) + if i == 0: + for param in ptb_model.parameters(): + dy_param_init[param.name] = param._numpy() + dy_loss._backward() + sgd.minimize(dy_loss) + for param in ptb_model.parameters(): + dy_param_updated[param.name] = param._numpy() + + with new_program_scope(): + fluid.default_startup_program().random_seed = seed + fluid.default_main_program().random_seed = seed + ptb_model = PtbModel( + "ptb_model", + hidden_size=hidden_size, + vocab_size=vocab_size, + num_layers=num_layers, + num_steps=num_steps, + init_scale=init_scale) + + exe = fluid.Executor(fluid.CPUPlace()) + sgd = SGDOptimizer(learning_rate=1e-3) + x = fluid.layers.data(name="x", shape=[-1, 3, 1], dtype='int64') + y = fluid.layers.data(name="y", shape=[-1, 1], dtype='float32') + init_hidden = fluid.layers.data( + name="init_hidden", shape=[1], dtype='float32') + init_cell = fluid.layers.data( + name="init_cell", shape=[1], dtype='float32') + + static_loss, static_last_hidden, static_last_cell = ptb_model( + x, y, init_hidden, init_cell) + sgd.minimize(static_loss) + static_param_updated = dict() + static_param_init = dict() + static_param_name_list = list() + for param in ptb_model.parameters(): + static_param_name_list.append(param.name) + + out = exe.run(framework.default_startup_program(), + fetch_list=static_param_name_list) + for i in range(len(static_param_name_list)): + static_param_init[static_param_name_list[i]] = out[i] + static_loss_value = None + static_last_cell_value = None + static_last_hidden_value = None + for i in range(2): + x_data = np.arange(12).reshape(4, 3).astype('int64') + y_data = np.arange(1, 13).reshape(4, 3).astype('int64') + x_data = x_data.reshape((-1, num_steps, 1)) + y_data = y_data.reshape((-1, 1)) + init_hidden_data = np.zeros( + (num_layers, batch_size, hidden_size), dtype='float32') + init_cell_data = np.zeros( + (num_layers, batch_size, hidden_size), dtype='float32') + fetch_list = [static_loss, static_last_hidden, static_last_cell] + fetch_list.extend(static_param_name_list) + out = exe.run(fluid.default_main_program(), + feed={ + "x": x_data, + "y": y_data, + "init_hidden": init_hidden_data, + "init_cell": init_cell_data + }, + fetch_list=fetch_list) + static_loss_value = out[0] + static_last_cell_value = out[1] + static_last_hidden_value = out[2] + for k in range(3, len(out)): + static_param_updated[static_param_name_list[k - 3]] = out[k] + + self.assertTrue( + np.allclose(static_loss_value.all(), dy_loss._numpy().all())) + self.assertTrue( + np.allclose(static_last_cell_value.all(), + last_cell._numpy().all())) + self.assertTrue( + np.allclose(static_last_hidden_value.all(), + last_hidden._numpy().all())) + for key, value in six.iteritems(static_param_init): + self.assertTrue( + np.allclose(value.all(), dy_param_init[key].all())) + for key, value in six.iteritems(static_param_updated): + self.assertTrue( + np.allclose(value.all(), dy_param_updated[key].all())) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_imperative_resnet.py b/python/paddle/fluid/tests/unittests/test_imperative_resnet.py index 87a72dd04e376cf9225e275d862b0cbbb9774e2c..0e134742a7e80c462206072644bb4bf196397b38 100644 --- a/python/paddle/fluid/tests/unittests/test_imperative_resnet.py +++ b/python/paddle/fluid/tests/unittests/test_imperative_resnet.py @@ -21,7 +21,6 @@ import paddle import paddle.fluid as fluid from paddle.fluid import core from paddle.fluid.layer_helper import LayerHelper -from paddle.fluid.optimizer import SGDOptimizer from paddle.fluid.imperative.nn import Conv2D, Pool2D, BatchNorm, FC from paddle.fluid.imperative.base import to_variable from test_imperative_base import new_program_scope @@ -71,15 +70,17 @@ def optimizer_setting(params): class ConvBNLayer(fluid.imperative.Layer): def __init__(self, + name_scope, num_channels, num_filters, filter_size, stride=1, groups=1, act=None): - super(ConvBNLayer, self).__init__() + super(ConvBNLayer, self).__init__(name_scope) self._conv = Conv2D( + self.full_name(), num_channels=num_channels, num_filters=num_filters, filter_size=filter_size, @@ -89,7 +90,7 @@ class ConvBNLayer(fluid.imperative.Layer): act=None, bias_attr=None) - self._batch_norm = BatchNorm(num_filters, act=act) + self._batch_norm = BatchNorm(self.full_name(), num_filters, act=act) def forward(self, inputs): y = self._conv(inputs) @@ -99,21 +100,29 @@ class ConvBNLayer(fluid.imperative.Layer): class BottleneckBlock(fluid.imperative.Layer): - def __init__(self, num_channels, num_filters, stride, shortcut=True): - super(BottleneckBlock, self).__init__() + def __init__(self, + name_scope, + num_channels, + num_filters, + stride, + shortcut=True): + super(BottleneckBlock, self).__init__(name_scope) self.conv0 = ConvBNLayer( + self.full_name(), num_channels=num_channels, num_filters=num_filters, filter_size=1, act='relu') self.conv1 = ConvBNLayer( + self.full_name(), num_channels=num_filters, num_filters=num_filters, filter_size=3, stride=stride, act='relu') self.conv2 = ConvBNLayer( + self.full_name(), num_channels=num_filters, num_filters=num_filters * 4, filter_size=1, @@ -121,6 +130,7 @@ class BottleneckBlock(fluid.imperative.Layer): if not shortcut: self.short = ConvBNLayer( + self.full_name(), num_channels=num_channels, num_filters=num_filters * 4, filter_size=1, @@ -142,13 +152,13 @@ class BottleneckBlock(fluid.imperative.Layer): y = fluid.layers.elementwise_add(x=short, y=conv2) - layer_helper = LayerHelper('elementwise_add_activation', act='relu') + layer_helper = LayerHelper(self.full_name(), act='relu') return layer_helper.append_activation(y) class ResNet(fluid.imperative.Layer): - def __init__(self, layers=50, class_dim=102): - super(ResNet, self).__init__() + def __init__(self, name_scope, layers=50, class_dim=102): + super(ResNet, self).__init__(name_scope) self.layers = layers supported_layers = [50, 101, 152] @@ -164,31 +174,44 @@ class ResNet(fluid.imperative.Layer): num_filters = [64, 128, 256, 512] self.conv = ConvBNLayer( - num_channels=3, num_filters=64, filter_size=7, stride=2, act='relu') + self.full_name(), + num_channels=3, + num_filters=64, + filter_size=7, + stride=2, + act='relu') self.pool2d_max = Pool2D( - pool_size=3, pool_stride=2, pool_padding=1, pool_type='max') + self.full_name(), + pool_size=3, + pool_stride=2, + pool_padding=1, + pool_type='max') self.bottleneck_block_list = [] num_channels = 64 for block in range(len(depth)): shortcut = False for i in range(depth[block]): - bottleneck_block = BottleneckBlock( - num_channels=num_channels, - num_filters=num_filters[block], - stride=2 if i == 0 and block != 0 else 1, - shortcut=shortcut) + bottleneck_block = self.add_sublayer( + 'bb_%d_%d' % (block, i), + BottleneckBlock( + self.full_name(), + num_channels=num_channels, + num_filters=num_filters[block], + stride=2 if i == 0 and block != 0 else 1, + shortcut=shortcut)) num_channels = bottleneck_block._num_channels_out self.bottleneck_block_list.append(bottleneck_block) shortcut = True self.pool2d_avg = Pool2D( - pool_size=7, pool_type='avg', global_pooling=True) + self.full_name(), pool_size=7, pool_type='avg', global_pooling=True) import math stdv = 1.0 / math.sqrt(2048 * 1.0) - self.out = FC(size=class_dim, + self.out = FC(self.full_name(), + size=class_dim, act='softmax', param_attr=fluid.param_attr.ParamAttr( initializer=fluid.initializer.Uniform(-stdv, stdv))) @@ -213,7 +236,7 @@ class TestImperativeResnet(unittest.TestCase): fluid.default_startup_program().random_seed = seed fluid.default_main_program().random_seed = seed - resnet = ResNet() + resnet = ResNet("resnet") optimizer = optimizer_setting(train_parameters) np.random.seed(seed) import random @@ -223,8 +246,7 @@ class TestImperativeResnet(unittest.TestCase): batch_size=batch_size) dy_param_init_value = {} - for param in fluid.default_main_program().global_block( - ).all_parameters(): + for param in resnet.parameters(): dy_param_init_value[param.name] = param._numpy() for batch_id, data in enumerate(train_reader()): @@ -247,16 +269,14 @@ class TestImperativeResnet(unittest.TestCase): dy_out = avg_loss._numpy() if batch_id == 0: - for param in fluid.default_main_program().global_block( - ).all_parameters(): + for param in resnet.parameters(): if param.name not in dy_param_init_value: dy_param_init_value[param.name] = param._numpy() avg_loss._backward() dy_grad_value = {} - for param in fluid.default_main_program().global_block( - ).all_parameters(): + for param in resnet.parameters(): if not param.stop_gradient: np_array = np.array(param._ivar._grad_ivar().value() .get_tensor()) @@ -264,10 +284,10 @@ class TestImperativeResnet(unittest.TestCase): )] = np_array optimizer.minimize(avg_loss) + resnet.clear_gradients() dy_param_value = {} - for param in fluid.default_main_program().global_block( - ).all_parameters(): + for param in resnet.parameters(): dy_param_value[param.name] = param._numpy() with new_program_scope(): @@ -277,7 +297,7 @@ class TestImperativeResnet(unittest.TestCase): exe = fluid.Executor(fluid.CPUPlace( ) if not core.is_compiled_with_cuda() else fluid.CUDAPlace(0)) - resnet = ResNet() + resnet = ResNet("resnet") optimizer = optimizer_setting(train_parameters) np.random.seed(seed) @@ -348,6 +368,7 @@ class TestImperativeResnet(unittest.TestCase): self.assertTrue(np.allclose(static_out, dy_out)) self.assertEqual(len(dy_param_init_value), len(static_param_init_value)) + for key, value in six.iteritems(static_param_init_value): self.assertTrue(np.allclose(value, dy_param_init_value[key])) self.assertTrue(np.isfinite(value.all())) diff --git a/python/paddle/fluid/tests/unittests/test_inference_model_io.py b/python/paddle/fluid/tests/unittests/test_inference_model_io.py index 9962702f69644b7aef7d868f086abb390441f617..9c9f86330704466c7a8801af6ab0fb2bba23f931 100644 --- a/python/paddle/fluid/tests/unittests/test_inference_model_io.py +++ b/python/paddle/fluid/tests/unittests/test_inference_model_io.py @@ -25,6 +25,7 @@ import paddle.fluid.layers as layers import paddle.fluid.optimizer as optimizer from paddle.fluid.framework import Program, program_guard from paddle.fluid.io import save_inference_model, load_inference_model +from paddle.fluid.transpiler import memory_optimize class TestBook(unittest.TestCase): @@ -82,9 +83,36 @@ class TestBook(unittest.TestCase): self.assertEqual(feed_var_names, ["x", "y"]) self.assertEqual(len(fetch_vars), 1) - self.assertEqual(str(fetch_vars[0]), str(avg_cost)) + print("fetch %s" % str(fetch_vars[0])) + self.assertTrue("scale" in str(fetch_vars[0])) self.assertEqual(expected, actual) +class TestSaveInferenceModel(unittest.TestCase): + def test_save_inference_model(self): + MODEL_DIR = "./tmp/inference_model2" + init_program = Program() + program = Program() + + # fake program without feed/fetch + 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(cost) + + place = core.CPUPlace() + exe = executor.Executor(place) + exe.run(init_program, feed={}, fetch_list=[]) + + memory_optimize(program, print_log=True) + self.assertEqual(program._is_mem_optimized, True) + # will print warning message + save_inference_model(MODEL_DIR, ["x", "y"], [avg_cost], exe, program) + + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_ir_inplace_pass.py b/python/paddle/fluid/tests/unittests/test_ir_inplace_pass.py new file mode 100644 index 0000000000000000000000000000000000000000..4e196758efc990506957089fb5b88ebb099cca29 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_ir_inplace_pass.py @@ -0,0 +1,76 @@ +# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import print_function + +import os +import unittest +import numpy as np +import paddle.fluid.core as core +import paddle.fluid as fluid +from parallel_executor_test_base import TestParallelExecutorBase + + +def fc_with_batchnorm(use_feed): + img = fluid.layers.data(name='image', shape=[784], dtype='float32') + label = fluid.layers.data(name='label', shape=[1], dtype='int64') + + hidden = img + for _ in range(3): + hidden = fluid.layers.fc( + hidden, + size=200, + act='tanh', + bias_attr=fluid.ParamAttr( + initializer=fluid.initializer.Constant(value=1.0))) + + hidden = fluid.layers.batch_norm(input=hidden) + prediction = fluid.layers.fc(hidden, size=10, act='softmax') + loss = fluid.layers.cross_entropy(input=prediction, label=label) + loss = fluid.layers.mean(loss) + return loss + + +class TestIrInplace(TestParallelExecutorBase): + @classmethod + def setUpClass(cls): + os.environ['CPU_NUM'] = str(4) + + def _fc_with_batchnorm(self, + ir_memory_optimize, + enable_inplace, + memory_opt=False): + + if not core.is_compiled_with_cuda(): + return + np.random.seed(5) + img = np.random.random(size=[32, 784]).astype(np.float32) + label = np.ones(shape=[32, 1], dtype='int64') + self.check_network_convergence( + fc_with_batchnorm, + feed_dict={"image": img, + "label": label}, + use_cuda=True, + memory_opt=memory_opt, + use_ir_memory_optimize=ir_memory_optimize, + enable_inplace=enable_inplace) + + def test_fc_with_batchnorm(self, delta=1e-3): + loss00 = self._fc_with_batchnorm(False, False) + loss10 = self._fc_with_batchnorm(True, False) + loss01 = self._fc_with_batchnorm(False, True) + loss11 = self._fc_with_batchnorm(True, True) + self.assertAlmostEqual(loss00, loss10, delta=delta) + self.assertAlmostEqual(loss00, loss01, delta=delta) + self.assertAlmostEqual(loss00, loss11, delta=delta) diff --git a/python/paddle/fluid/tests/unittests/test_ir_memory_optimize_transformer.py b/python/paddle/fluid/tests/unittests/test_ir_memory_optimize_transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..c0f480e34dcac3351ba3008ad632a29943afdb81 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_ir_memory_optimize_transformer.py @@ -0,0 +1,48 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import os +import unittest +import paddle.fluid as fluid +import paddle.fluid.core as core + +os.environ['FLAGS_eager_delete_tensor_gb'] = "0.0" +os.environ[ + 'RECORDIO_FILENAME'] = '/tmp/ir_memory_optimize_transformer.wmt16.recordio' + +from test_parallel_executor_transformer import TestTransformer +from test_parallel_executor_transformer import transformer + + +# NOTE(dzhwinter): test diferent strategy colisions. +# open the eager delete tensor strategy by default. +class TestTransformerWithIR(TestTransformer): + def test_main(self): + if core.is_compiled_with_cuda(): + # check python transpiler + self.check_network_convergence( + transformer, + use_cuda=True, + memory_opt=True, + use_ir_memory_optimize=False) + # check IR memory optimize + self.check_network_convergence( + transformer, + use_cuda=True, + memory_opt=False, + use_ir_memory_optimize=True) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_layers.py b/python/paddle/fluid/tests/unittests/test_layers.py index c13f03e86f3e375026b04a31d51ac1a5223360ef..30194f8cacfea2361ffe4afe537287a261cf470b 100644 --- a/python/paddle/fluid/tests/unittests/test_layers.py +++ b/python/paddle/fluid/tests/unittests/test_layers.py @@ -58,7 +58,8 @@ class TestBook(unittest.TestCase): def test_simple_conv2d(self): program = Program() with program_guard(program, startup_program=Program()): - images = layers.data(name='pixel', shape=[3, 48, 48], dtype='int32') + images = layers.data( + name='pixel', shape=[3, 48, 48], dtype='float32') layers.conv2d(input=images, num_filters=3, filter_size=[4, 4]) print(str(program)) @@ -373,6 +374,17 @@ class TestBook(unittest.TestCase): self.assertIsNotNone(output) print(str(program)) + def test_sampled_softmax_with_cross_entropy(self): + program = Program() + with program_guard(program): + logits = layers.data(name='Logits', shape=[256], dtype='float64') + label = layers.data(name='Label', shape=[1], dtype='int64') + num_samples = 25 + output = layers.sampled_softmax_with_cross_entropy(logits, label, + num_samples) + self.assertIsNotNone(output) + print(str(program)) + @decorators.prog_scope() def test_nce(self): window_size = 5 diff --git a/python/paddle/fluid/tests/unittests/test_lstmp_op.py b/python/paddle/fluid/tests/unittests/test_lstmp_op.py index 9c3ec45515ffe0a07541fd9cfb7e92b079264071..0645cfedb8089f5618c54672cac91343e5dee285 100644 --- a/python/paddle/fluid/tests/unittests/test_lstmp_op.py +++ b/python/paddle/fluid/tests/unittests/test_lstmp_op.py @@ -36,12 +36,14 @@ def lstmp( w_b=None, # 1 x 4D w_c=None, # 1 x 3D is_reverse=False, + proj_clip=0.0, + cell_clip=0.0, act_gate=None, act_cell=None, act_cand=None, act_proj=None): - def _step(x, w_r, w_rh, w_c, r_pre, c_pre, act_gate, act_cell, act_cand, - act_proj): + def _step(x, w_r, w_rh, w_c, r_pre, c_pre, proj_clip, cell_clip, act_gate, + act_cell, act_cand, act_proj): g = np.dot(r_pre, w_r) # 1 x 4D g = g + x g = np.reshape(g, (1, g.size)) @@ -55,6 +57,17 @@ def lstmp( g_f = act_gate(g_f + w_fc * c_pre) # 1 x D c = g_f * c_pre + g_i * act_cand(c) # 1 x D + def array_clip(a, clip): + size = np.prod(a.shape) + new_a = np.reshape(a, (size)) + for i in range(size): + new_a[i] = max(new_a[i], -1.0 * clip) + new_a[i] = min(new_a[i], clip) + new_a = np.reshape(new_a, a.shape) + return new_a + + if cell_clip > 0.0: + c = array_clip(c, cell_clip) if w_c is None: g_o = act_gate(g_o) # 1 x D else: @@ -64,6 +77,8 @@ def lstmp( # projection r = np.dot(h, w_rh) r = act_proj(r) + if proj_clip > 0.0: + r = array_clip(r, proj_clip) return r, c def _reverse(x, offset): @@ -87,13 +102,13 @@ def lstmp( # compute one sequence seq_len = lod[0][i] x = input[offset[i]:offset[i + 1], :] - r_pre = np.dot(h0[i], w_rh) # 1 x P - r_pre = act_proj(r_pre) + r_pre = h0[i] c_pre = c0[i] # 1 x D for j in range(seq_len): # compute one step - r_pre, c_pre = _step(x[j], w_r, w_rh, w_c, r_pre, c_pre, act_gate, - act_cell, act_cand, act_proj) + r_pre, c_pre = _step(x[j], w_r, w_rh, w_c, r_pre, c_pre, proj_clip, + cell_clip, act_gate, act_cell, act_cand, + act_proj) projection.append(r_pre.flatten()) cell.append(c_pre.flatten()) @@ -123,13 +138,12 @@ class TestLstmpOp(LstmTest.TestLstmOp): T = sum(self.lod[0]) N = len(self.lod[0]) - x = np.random.normal(size=(T, 4 * self.D)).astype('float64') if self.has_initial_state: - h0 = np.random.normal(size=(N, self.D)).astype('float64') + h0 = np.random.normal(size=(N, self.P)).astype('float64') c0 = np.random.normal(size=(N, self.D)).astype('float64') else: - h0 = np.zeros((N, self.D)).astype('float64') + h0 = np.zeros((N, self.P)).astype('float64') c0 = np.zeros((N, self.D)).astype('float64') w = np.random.normal(size=(self.P, 4 * self.D)).astype('float64') if self.use_peepholes: @@ -140,9 +154,12 @@ class TestLstmpOp(LstmTest.TestLstmOp): w_b = b[:, 0:4 * self.D] w_c = b[:, 4 * self.D:] if self.use_peepholes else None w_rh = np.random.normal(size=(self.D, self.P)).astype('float64') + proj_clip = 0.1 + cell_clip = 0.1 r, c = lstmp(x, self.lod, h0, c0, w, w_rh, w_b, w_c, self.is_reverse, - ACTIVATION[self.act_gate], ACTIVATION[self.act_cell], - ACTIVATION[self.act_cand], ACTIVATION[self.act_proj]) + proj_clip, cell_clip, ACTIVATION[self.act_gate], + ACTIVATION[self.act_cell], ACTIVATION[self.act_cand], + ACTIVATION[self.act_proj]) self.inputs = {'Input': (x, self.lod), 'Weight': w, 'ProjWeight': w_rh} @@ -159,6 +176,8 @@ class TestLstmpOp(LstmTest.TestLstmOp): self.attrs = { 'use_peepholes': self.use_peepholes, 'is_reverse': self.is_reverse, + 'proj_clip': proj_clip, + 'cell_clip': cell_clip, 'gate_activation': self.act_gate, 'cell_activation': self.act_cell, 'candidate_activation': self.act_cand, @@ -171,14 +190,14 @@ class TestLstmpOp(LstmTest.TestLstmOp): def test_check_grad(self): # TODO(qingqing) remove folowing lines after the check_grad is refined. N = len(self.lod[0]) - self.outputs['OrderedP0'] = np.zeros((N, self.P)).astype('float64') self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64') self.outputs['BatchHidden'] = np.zeros((N, self.D)).astype('float64') self.outputs['BatchCellPreAct'] = np.zeros( (N, self.D)).astype('float64') self.check_grad( ['Input', 'Weight', 'ProjWeight', 'Bias'], ['Projection'], - max_relative_error=1e-2) + max_relative_error=1e-2, + numeric_grad_delta=0.0000005) class TestLstmpOpHasInitial(TestLstmpOp): @@ -188,7 +207,6 @@ class TestLstmpOpHasInitial(TestLstmpOp): def test_check_grad(self): # TODO(qingqing) remove folowing lines after the check_grad is refined. N = len(self.lod[0]) - self.outputs['OrderedP0'] = np.zeros((N, self.P)).astype('float64') self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64') self.outputs['BatchHidden'] = np.zeros((N, self.D)).astype('float64') self.outputs['BatchCellPreAct'] = np.zeros( @@ -196,11 +214,11 @@ class TestLstmpOpHasInitial(TestLstmpOp): self.check_grad( ['Input', 'Weight', 'ProjWeight', 'Bias', 'H0', 'C0'], ['Projection'], + numeric_grad_delta=0.0000005, max_relative_error=1e-2) def test_check_grad_ingore_bias(self): N = len(self.lod[0]) - self.outputs['OrderedP0'] = np.zeros((N, self.P)).astype('float64') self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64') self.outputs['BatchHidden'] = np.zeros((N, self.D)).astype('float64') self.outputs['BatchCellPreAct'] = np.zeros( @@ -208,11 +226,11 @@ class TestLstmpOpHasInitial(TestLstmpOp): self.check_grad( ['Input', 'ProjWeight', 'Weight'], ['Projection'], max_relative_error=1e-2, + numeric_grad_delta=0.0000005, no_grad_set=set('Bias')) def test_check_grad_ingore_weight(self): N = len(self.lod[0]) - self.outputs['OrderedP0'] = np.zeros((N, self.P)).astype('float64') self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64') self.outputs['BatchHidden'] = np.zeros((N, self.D)).astype('float64') self.outputs['BatchCellPreAct'] = np.zeros( @@ -220,11 +238,11 @@ class TestLstmpOpHasInitial(TestLstmpOp): self.check_grad( ['Input', 'ProjWeight', 'Bias'], ['Projection'], max_relative_error=1e-2, + numeric_grad_delta=0.0000005, no_grad_set=set('Weight')) def test_check_grad_ingore_proj_weight(self): N = len(self.lod[0]) - self.outputs['OrderedP0'] = np.zeros((N, self.P)).astype('float64') self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64') self.outputs['BatchHidden'] = np.zeros((N, self.D)).astype('float64') self.outputs['BatchCellPreAct'] = np.zeros( @@ -232,11 +250,11 @@ class TestLstmpOpHasInitial(TestLstmpOp): self.check_grad( ['Input', 'Weight', 'Bias'], ['Projection'], max_relative_error=1e-2, + numeric_grad_delta=0.0000005, no_grad_set=set('ProjWeight')) def test_check_grad_ingore_input(self): N = len(self.lod[0]) - self.outputs['OrderedP0'] = np.zeros((N, self.P)).astype('float64') self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64') self.outputs['BatchHidden'] = np.zeros((N, self.D)).astype('float64') self.outputs['BatchCellPreAct'] = np.zeros( @@ -244,11 +262,11 @@ class TestLstmpOpHasInitial(TestLstmpOp): self.check_grad( ['Weight', 'ProjWeight', 'Bias'], ['Projection'], max_relative_error=1e-2, + numeric_grad_delta=0.0000005, no_grad_set=set('Input')) def test_check_grad_ingore_h0(self): N = len(self.lod[0]) - self.outputs['OrderedP0'] = np.zeros((N, self.P)).astype('float64') self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64') self.outputs['BatchHidden'] = np.zeros((N, self.D)).astype('float64') self.outputs['BatchCellPreAct'] = np.zeros( @@ -256,11 +274,11 @@ class TestLstmpOpHasInitial(TestLstmpOp): self.check_grad( ['Input', 'Weight', 'ProjWeight', 'Bias', 'C0'], ['Projection'], max_relative_error=1e-2, + numeric_grad_delta=0.0000005, no_grad_set=set('H0')) def test_check_grad_ingore_c0(self): N = len(self.lod[0]) - self.outputs['OrderedP0'] = np.zeros((N, self.P)).astype('float64') self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64') self.outputs['BatchHidden'] = np.zeros((N, self.D)).astype('float64') self.outputs['BatchCellPreAct'] = np.zeros( @@ -268,6 +286,7 @@ class TestLstmpOpHasInitial(TestLstmpOp): self.check_grad( ['Input', 'Weight', 'ProjWeight', 'Bias', 'H0'], ['Projection'], max_relative_error=1e-2, + numeric_grad_delta=0.0000005, no_grad_set=set('C0')) diff --git a/python/paddle/fluid/tests/unittests/test_multiclass_nms_op.py b/python/paddle/fluid/tests/unittests/test_multiclass_nms_op.py index 9778bd694de4b21f3ff723846c77a8ad0dceb57b..69e060341ed9dbb711f13f860e047e19f741b336 100644 --- a/python/paddle/fluid/tests/unittests/test_multiclass_nms_op.py +++ b/python/paddle/fluid/tests/unittests/test_multiclass_nms_op.py @@ -19,7 +19,7 @@ import copy from op_test import OpTest -def iou(box_a, box_b): +def iou(box_a, box_b, norm): """Apply intersection-over-union overlap between box_a and box_b """ xmin_a = min(box_a[0], box_a[2]) @@ -32,8 +32,10 @@ def iou(box_a, box_b): xmax_b = max(box_b[0], box_b[2]) ymax_b = max(box_b[1], box_b[3]) - area_a = (ymax_a - ymin_a) * (xmax_a - xmin_a) - area_b = (ymax_b - ymin_b) * (xmax_b - xmin_b) + area_a = (ymax_a - ymin_a + (norm == False)) * (xmax_a - xmin_a + + (norm == False)) + area_b = (ymax_b - ymin_b + (norm == False)) * (xmax_b - xmin_b + + (norm == False)) if area_a <= 0 and area_b <= 0: return 0.0 @@ -42,17 +44,21 @@ def iou(box_a, box_b): xb = min(xmax_a, xmax_b) yb = min(ymax_a, ymax_b) - inter_area = max(xb - xa, 0.0) * max(yb - ya, 0.0) - - box_a_area = (box_a[2] - box_a[0]) * (box_a[3] - box_a[1]) - box_b_area = (box_b[2] - box_b[0]) * (box_b[3] - box_b[1]) + inter_area = max(xb - xa + (norm == False), + 0.0) * max(yb - ya + (norm == False), 0.0) iou_ratio = inter_area / (area_a + area_b - inter_area) return iou_ratio -def nms(boxes, scores, score_threshold, nms_threshold, top_k=200, eta=1.0): +def nms(boxes, + scores, + score_threshold, + nms_threshold, + top_k=200, + normalized=True, + eta=1.0): """Apply non-maximum suppression at test time to avoid detecting too many overlapping bounding boxes for a given object. Args: @@ -87,7 +93,7 @@ def nms(boxes, scores, score_threshold, nms_threshold, top_k=200, eta=1.0): for k in range(len(selected_indices)): if keep: kept_idx = selected_indices[k] - overlap = iou(boxes[idx], boxes[kept_idx]) + overlap = iou(boxes[idx], boxes[kept_idx], normalized) keep = True if overlap <= adaptive_threshold else False else: break @@ -99,16 +105,24 @@ def nms(boxes, scores, score_threshold, nms_threshold, top_k=200, eta=1.0): def multiclass_nms(boxes, scores, background, score_threshold, nms_threshold, - nms_top_k, keep_top_k): - class_num = scores.shape[0] - priorbox_num = scores.shape[1] + nms_top_k, keep_top_k, normalized, shared): + if shared: + class_num = scores.shape[0] + priorbox_num = scores.shape[1] + else: + box_num = scores.shape[0] + class_num = scores.shape[1] selected_indices = {} num_det = 0 for c in range(class_num): if c == background: continue - indices = nms(boxes, scores[c], score_threshold, nms_threshold, - nms_top_k) + if shared: + indices = nms(boxes, scores[c], score_threshold, nms_threshold, + nms_top_k, normalized) + else: + indices = nms(boxes[:, c, :], scores[:, c], score_threshold, + nms_threshold, nms_top_k, normalized) selected_indices[c] = indices num_det += len(indices) @@ -116,7 +130,10 @@ def multiclass_nms(boxes, scores, background, score_threshold, nms_threshold, score_index = [] for c, indices in selected_indices.items(): for idx in indices: - score_index.append((scores[c][idx], c, idx)) + if shared: + score_index.append((scores[c][idx], c, idx)) + else: + score_index.append((scores[idx][c], c, idx)) sorted_score_index = sorted( score_index, key=lambda tup: tup[0], reverse=True) @@ -127,24 +144,78 @@ def multiclass_nms(boxes, scores, background, score_threshold, nms_threshold, selected_indices[c] = [] for s, c, idx in sorted_score_index: selected_indices[c].append(idx) + if not shared: + for labels in selected_indices: + selected_indices[labels].sort() num_det = keep_top_k return selected_indices, num_det -def batched_multiclass_nms(boxes, scores, background, score_threshold, - nms_threshold, nms_top_k, keep_top_k): +def lod_multiclass_nms(boxes, scores, background, score_threshold, + nms_threshold, nms_top_k, keep_top_k, box_lod, + normalized): + det_outs = [] + lod = [] + head = 0 + for n in range(len(box_lod[0])): + box = boxes[head:head + box_lod[0][n]] + score = scores[head:head + box_lod[0][n]] + head = head + box_lod[0][n] + nmsed_outs, nmsed_num = multiclass_nms( + box, + score, + background, + score_threshold, + nms_threshold, + nms_top_k, + keep_top_k, + normalized, + shared=False) + if nmsed_num == 0: + continue + lod.append(nmsed_num) + tmp_det_out = [] + for c, indices in nmsed_outs.items(): + for idx in indices: + xmin, ymin, xmax, ymax = box[idx, c, :] + tmp_det_out.append([c, score[idx][c], xmin, ymin, xmax, ymax]) + sorted_det_out = sorted( + tmp_det_out, key=lambda tup: tup[0], reverse=False) + det_outs.extend(sorted_det_out) + if len(lod) == 0: + lod.append(1) + + return det_outs, lod + + +def batched_multiclass_nms(boxes, + scores, + background, + score_threshold, + nms_threshold, + nms_top_k, + keep_top_k, + normalized=True): batch_size = scores.shape[0] det_outs = [] lod = [] for n in range(batch_size): - nmsed_outs, nmsed_num = multiclass_nms(boxes[n], scores[n], background, - score_threshold, nms_threshold, - nms_top_k, keep_top_k) - lod.append(nmsed_num) - if nmsed_num == 0: continue + nmsed_outs, nmsed_num = multiclass_nms( + boxes[n], + scores[n], + background, + score_threshold, + nms_threshold, + nms_top_k, + keep_top_k, + normalized, + shared=True) + if nmsed_num == 0: + continue + lod.append(nmsed_num) tmp_det_out = [] for c, indices in nmsed_outs.items(): for idx in indices: @@ -154,7 +225,8 @@ def batched_multiclass_nms(boxes, scores, background, score_threshold, sorted_det_out = sorted( tmp_det_out, key=lambda tup: tup[0], reverse=False) det_outs.extend(sorted_det_out) - + if len(lod) == 0: + lod += [1] return det_outs, lod @@ -168,7 +240,6 @@ class TestMulticlassNMSOp(OpTest): M = 1200 C = 21 BOX_SIZE = 4 - background = 0 nms_threshold = 0.3 nms_top_k = 400 @@ -206,6 +277,7 @@ class TestMulticlassNMSOp(OpTest): 'keep_top_k': keep_top_k, 'score_threshold': score_threshold, 'nms_eta': 1.0, + 'normalized': True, } def test_check_output(self): @@ -219,13 +291,70 @@ class TestMulticlassNMSOpNoOutput(TestMulticlassNMSOp): self.score_threshold = 2.0 +class TestMulticlassNMSLoDInput(OpTest): + def set_argument(self): + self.score_threshold = 0.01 + + def setUp(self): + self.set_argument() + M = 1200 + C = 21 + BOX_SIZE = 4 + box_lod = [[1200]] + background = 0 + nms_threshold = 0.3 + nms_top_k = 400 + keep_top_k = 200 + score_threshold = self.score_threshold + normalized = False + + scores = np.random.random((M, C)).astype('float32') + + def softmax(x): + shiftx = x - np.max(x).clip(-64.) + exps = np.exp(shiftx) + return exps / np.sum(exps) + + scores = np.apply_along_axis(softmax, 1, scores) + + boxes = np.random.random((M, C, BOX_SIZE)).astype('float32') + boxes[:, :, 0] = boxes[:, :, 0] * 10 + boxes[:, :, 1] = boxes[:, :, 1] * 10 + boxes[:, :, 2] = boxes[:, :, 2] * 10 + 10 + boxes[:, :, 3] = boxes[:, :, 3] * 10 + 10 + + nmsed_outs, lod = lod_multiclass_nms( + boxes, scores, background, score_threshold, nms_threshold, + nms_top_k, keep_top_k, box_lod, normalized) + nmsed_outs = [-1] if not nmsed_outs else nmsed_outs + nmsed_outs = np.array(nmsed_outs).astype('float32') + self.op_type = 'multiclass_nms' + self.inputs = { + 'BBoxes': (boxes, box_lod), + 'Scores': (scores, box_lod), + } + self.outputs = {'Out': (nmsed_outs, [lod])} + self.attrs = { + 'background_label': 0, + 'nms_threshold': nms_threshold, + 'nms_top_k': nms_top_k, + 'keep_top_k': keep_top_k, + 'score_threshold': score_threshold, + 'nms_eta': 1.0, + 'normalized': normalized, + } + + def test_check_output(self): + self.check_output() + + class TestIOU(unittest.TestCase): def test_iou(self): box1 = np.array([4.0, 3.0, 7.0, 5.0]).astype('float32') box2 = np.array([3.0, 4.0, 6.0, 8.0]).astype('float32') expt_output = np.array([2.0 / 16.0]).astype('float32') - calc_output = np.array([iou(box1, box2)]).astype('float32') + calc_output = np.array([iou(box1, box2, True)]).astype('float32') self.assertTrue(np.allclose(calc_output, expt_output)) diff --git a/python/paddle/fluid/tests/unittests/test_nearest_interp_op.py b/python/paddle/fluid/tests/unittests/test_nearest_interp_op.py index 242709425f2d3f190d3c1ed795d30938fb8e23fe..5bb2260ef7a143670dd75fc88769603d1437173d 100644 --- a/python/paddle/fluid/tests/unittests/test_nearest_interp_op.py +++ b/python/paddle/fluid/tests/unittests/test_nearest_interp_op.py @@ -24,7 +24,8 @@ def nearest_neighbor_interp_np(X, out_h, out_w, out_size=None, - actual_shape=None): + actual_shape=None, + align_corners=True): """nearest neighbor interpolation implement in shape [N, C, H, W]""" if out_size is not None: out_h = out_size[0] @@ -35,17 +36,31 @@ def nearest_neighbor_interp_np(X, n, c, in_h, in_w = X.shape ratio_h = ratio_w = 0.0 - if out_h > 1: - ratio_h = (in_h - 1.0) / (out_h - 1.0) - if out_w > 1: - ratio_w = (in_w - 1.0) / (out_w - 1.0) + if (out_h > 1): + if (align_corners): + ratio_h = (in_h - 1.0) / (out_h - 1.0) + else: + ratio_h = 1.0 * in_h / out_h + if (out_w > 1): + if (align_corners): + ratio_w = (in_w - 1.0) / (out_w - 1.0) + else: + ratio_w = 1.0 * in_w / out_w out = np.zeros((n, c, out_h, out_w)) - for i in range(out_h): - in_i = int(ratio_h * i + 0.5) - for j in range(out_w): - in_j = int(ratio_w * j + 0.5) - out[:, :, i, j] = X[:, :, in_i, in_j] + + if align_corners: + for i in range(out_h): + in_i = int(ratio_h * i + 0.5) + for j in range(out_w): + in_j = int(ratio_w * j + 0.5) + out[:, :, i, j] = X[:, :, in_i, in_j] + else: + for i in range(out_h): + in_i = int(ratio_h * i) + for j in range(out_w): + in_j = int(ratio_w * j) + out[:, :, i, j] = X[:, :, in_i, in_j] return out.astype(X.dtype) @@ -59,7 +74,8 @@ class TestNearestInterpOp(OpTest): input_np = np.random.random(self.input_shape).astype("float32") output_np = nearest_neighbor_interp_np(input_np, self.out_h, self.out_w, - self.out_size, self.actual_shape) + self.out_size, self.actual_shape, + self.align_corners) self.inputs = {'X': input_np} if self.out_size is not None: self.inputs['OutSize'] = self.out_size @@ -68,7 +84,8 @@ class TestNearestInterpOp(OpTest): self.attrs = { 'out_h': self.out_h, 'out_w': self.out_w, - 'interp_method': self.interp_method + 'interp_method': self.interp_method, + 'align_corners': self.align_corners, } self.outputs = {'Out': output_np} @@ -84,6 +101,7 @@ class TestNearestInterpOp(OpTest): self.out_h = 2 self.out_w = 2 self.out_size = np.array([3, 3]).astype("int32") + self.align_corners = True class TestNearestNeighborInterpCase1(TestNearestInterpOp): @@ -92,6 +110,7 @@ class TestNearestNeighborInterpCase1(TestNearestInterpOp): self.input_shape = [4, 1, 7, 8] self.out_h = 1 self.out_w = 1 + self.align_corners = True class TestNearestNeighborInterpCase2(TestNearestInterpOp): @@ -100,6 +119,7 @@ class TestNearestNeighborInterpCase2(TestNearestInterpOp): self.input_shape = [3, 3, 9, 6] self.out_h = 12 self.out_w = 12 + self.align_corners = True class TestNearestNeighborInterpCase3(TestNearestInterpOp): @@ -108,6 +128,7 @@ class TestNearestNeighborInterpCase3(TestNearestInterpOp): self.input_shape = [1, 1, 128, 64] self.out_h = 64 self.out_w = 128 + self.align_corners = True class TestNearestNeighborInterpCase4(TestNearestInterpOp): @@ -117,6 +138,7 @@ class TestNearestNeighborInterpCase4(TestNearestInterpOp): self.out_h = 1 self.out_w = 1 self.out_size = np.array([2, 2]).astype("int32") + self.align_corners = True class TestNearestNeighborInterpCase5(TestNearestInterpOp): @@ -126,6 +148,7 @@ class TestNearestNeighborInterpCase5(TestNearestInterpOp): self.out_h = 12 self.out_w = 12 self.out_size = np.array([11, 11]).astype("int32") + self.align_corners = True class TestNearestNeighborInterpCase6(TestNearestInterpOp): @@ -135,6 +158,7 @@ class TestNearestNeighborInterpCase6(TestNearestInterpOp): self.out_h = 64 self.out_w = 128 self.out_size = np.array([65, 129]).astype("int32") + self.align_corners = True class TestNearestNeighborInterpActualShape(TestNearestInterpOp): @@ -144,6 +168,7 @@ class TestNearestNeighborInterpActualShape(TestNearestInterpOp): self.out_h = 64 self.out_w = 32 self.out_size = np.array([66, 40]).astype("int32") + self.align_corners = True class TestNearestInterpOpUint8(OpTest): @@ -155,14 +180,16 @@ class TestNearestInterpOpUint8(OpTest): input_np = np.random.randint( low=0, high=256, size=self.input_shape).astype("uint8") output_np = nearest_neighbor_interp_np(input_np, self.out_h, self.out_w, - self.out_size, self.actual_shape) + self.out_size, self.actual_shape, + self.align_corners) self.inputs = {'X': input_np} if self.out_size is not None: self.inputs['OutSize'] = self.out_size self.attrs = { 'out_h': self.out_h, 'out_w': self.out_w, - 'interp_method': self.interp_method + 'interp_method': self.interp_method, + 'align_corners': self.align_corners } self.outputs = {'Out': output_np} @@ -174,6 +201,7 @@ class TestNearestInterpOpUint8(OpTest): self.input_shape = [1, 3, 9, 6] self.out_h = 10 self.out_w = 9 + self.align_corners = True class TestNearestNeighborInterpCase1Uint8(TestNearestInterpOpUint8): @@ -182,6 +210,7 @@ class TestNearestNeighborInterpCase1Uint8(TestNearestInterpOpUint8): self.input_shape = [2, 3, 128, 64] self.out_h = 120 self.out_w = 50 + self.align_corners = True class TestNearestNeighborInterpCase2Uint8(TestNearestInterpOpUint8): @@ -191,6 +220,12 @@ class TestNearestNeighborInterpCase2Uint8(TestNearestInterpOpUint8): self.out_h = 5 self.out_w = 13 self.out_size = np.array([6, 15]).astype("int32") + self.align_corners = True + + +class TestNearestInterpWithoutCorners(TestNearestInterpOp): + def set_align_corners(self): + self.align_corners = False if __name__ == "__main__": diff --git a/python/paddle/fluid/tests/unittests/test_optimizer.py b/python/paddle/fluid/tests/unittests/test_optimizer.py index 34c9b7e006950f1c10fb265ce903b1e836281de7..95ddc135b3da5bc144f64f20dab5dfd2b5bd3215 100644 --- a/python/paddle/fluid/tests/unittests/test_optimizer.py +++ b/python/paddle/fluid/tests/unittests/test_optimizer.py @@ -274,7 +274,7 @@ class TestAdagradOptimizer(unittest.TestCase): # Check init_program init_ops = init_program.global_block().ops - self.assertEqual(len(init_ops), 2) + self.assertEqual(len(init_ops), 3) self.assertEqual(init_ops[0].type, "fill_constant") self.assertAlmostEqual(init_ops[0].attr('value'), learning_rate) self.assertEqual(init_ops[1].type, "fill_constant") diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor_pg.py b/python/paddle/fluid/tests/unittests/test_parallel_executor_pg.py new file mode 100644 index 0000000000000000000000000000000000000000..041c56fce11e6f6abb0a941a9e9c9ad1cb60ab42 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_parallel_executor_pg.py @@ -0,0 +1,107 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import print_function + +import unittest + +import numpy as np +import os +os.environ['FLAGS_enable_parallel_graph'] = str(1) +import paddle.fluid.core as core +import os +import paddle.fluid as fluid +from parallel_executor_test_base import TestParallelExecutorBase + + +def simple_fc_net(use_feed): + img = fluid.layers.data(name='image', shape=[784], dtype='float32') + label = fluid.layers.data(name='label', shape=[1], dtype='int64') + hidden = img + for _ in range(4): + hidden = fluid.layers.fc( + hidden, + size=200, + act='tanh', + bias_attr=fluid.ParamAttr( + initializer=fluid.initializer.Constant(value=1.0))) + prediction = fluid.layers.fc(hidden, size=10, act='softmax') + loss = fluid.layers.cross_entropy(input=prediction, label=label) + loss = fluid.layers.mean(loss) + return loss + + +class TestMNIST(TestParallelExecutorBase): + @classmethod + def setUpClass(cls): + os.environ['CPU_NUM'] = str(4) + + def _init_data(self): + np.random.seed(5) + img = np.random.random(size=[32, 784]).astype(np.float32) + label = np.ones(shape=[32, 1], dtype='int64') + return img, label + + # simple_fc + def check_simple_fc_convergence(self, use_cuda, use_reduce=False): + if use_cuda and not core.is_compiled_with_cuda(): + return + + img, label = self._init_data() + + self.check_network_convergence( + simple_fc_net, + feed_dict={"image": img, + "label": label}, + use_cuda=use_cuda, + use_reduce=use_reduce) + + def test_simple_fc(self): + # use_cuda + self.check_simple_fc_convergence(True) + + def check_simple_fc_parallel_accuracy(self, use_cuda): + if use_cuda and not core.is_compiled_with_cuda(): + return + + img, label = self._init_data() + + single_first_loss, single_last_loss = self.check_network_convergence( + method=simple_fc_net, + seed=1, + feed_dict={"image": img, + "label": label}, + use_cuda=use_cuda, + use_parallel_executor=False) + parallel_first_loss, parallel_last_loss = self.check_network_convergence( + method=simple_fc_net, + seed=1, + feed_dict={"image": img, + "label": label}, + use_cuda=use_cuda, + use_parallel_executor=True) + + self.assertAlmostEquals( + np.mean(parallel_first_loss), + single_first_loss, + delta=1e-6, ) + self.assertAlmostEquals( + np.mean(parallel_last_loss), single_last_loss, delta=1e-6) + + def test_simple_fc_parallel_accuracy(self): + self.check_simple_fc_parallel_accuracy(True) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor_seresnext.py b/python/paddle/fluid/tests/unittests/test_parallel_executor_seresnext.py index e7a56bb6386a812e43e5c1b5c08cd0682aa9223a..9548598d75367ed1f1863d1f6ae50b83d58f8c7f 100644 --- a/python/paddle/fluid/tests/unittests/test_parallel_executor_seresnext.py +++ b/python/paddle/fluid/tests/unittests/test_parallel_executor_seresnext.py @@ -200,7 +200,7 @@ class TestResnet(TestParallelExecutorBase): model, use_cuda, iter=20, - delta2=1e-6): + delta2=1e-5): if use_cuda and not core.is_compiled_with_cuda(): return @@ -228,7 +228,7 @@ class TestResnet(TestParallelExecutorBase): optimizer=optimizer) for loss in zip(all_reduce_first_loss, reduce_first_loss): - self.assertAlmostEquals(loss[0], loss[1], delta=1e-6) + self.assertAlmostEquals(loss[0], loss[1], delta=1e-5) for loss in zip(all_reduce_last_loss, reduce_last_loss): self.assertAlmostEquals(loss[0], loss[1], delta=delta2) @@ -258,17 +258,17 @@ class TestResnet(TestParallelExecutorBase): enable_sequential_execution=True) for loss in zip(all_reduce_first_loss, all_reduce_first_loss_seq): - self.assertAlmostEquals(loss[0], loss[1], delta=1e-6) + self.assertAlmostEquals(loss[0], loss[1], delta=1e-5) for loss in zip(all_reduce_last_loss, all_reduce_last_loss_seq): self.assertAlmostEquals(loss[0], loss[1], delta=delta2) for loss in zip(reduce_first_loss, reduce_first_loss_seq): - self.assertAlmostEquals(loss[0], loss[1], delta=1e-6) + self.assertAlmostEquals(loss[0], loss[1], delta=1e-5) for loss in zip(reduce_last_loss, reduce_last_loss_seq): self.assertAlmostEquals(loss[0], loss[1], delta=delta2) for loss in zip(all_reduce_first_loss_seq, reduce_first_loss_seq): - self.assertAlmostEquals(loss[0], loss[1], delta=1e-6) + self.assertAlmostEquals(loss[0], loss[1], delta=1e-5) for loss in zip(all_reduce_last_loss_seq, reduce_last_loss_seq): self.assertAlmostEquals(loss[0], loss[1], delta=delta2) @@ -277,7 +277,7 @@ class TestResnet(TestParallelExecutorBase): use_cuda=True, use_reduce=False, iter=20, - delta2=1e-6): + delta2=1e-5): if use_cuda and not core.is_compiled_with_cuda(): return @@ -308,7 +308,7 @@ class TestResnet(TestParallelExecutorBase): optimizer=optimizer) self.assertAlmostEquals( - np.mean(parallel_first_loss), single_first_loss[0], delta=1e-6) + np.mean(parallel_first_loss), single_first_loss[0], delta=1e-5) self.assertAlmostEquals( np.mean(parallel_last_loss), single_last_loss[0], delta=delta2) diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor_transformer.py b/python/paddle/fluid/tests/unittests/test_parallel_executor_transformer.py index 3827743908c1d76931572277323d1dd5ddd05523..aacc1c3ecda8c25dec9f08827a856d38c37b1b2f 100644 --- a/python/paddle/fluid/tests/unittests/test_parallel_executor_transformer.py +++ b/python/paddle/fluid/tests/unittests/test_parallel_executor_transformer.py @@ -24,7 +24,7 @@ import paddle.fluid.core as core import paddle.dataset.wmt16 as wmt16 import os -WMT16_RECORDIO_FILE = "/tmp/wmt16.recordio" +WMT16_RECORDIO_FILE = os.environ.get('RECORDIO_FILENAME', '/tmp/wmt16.recordio') class ModelHyperParams(object): diff --git a/python/paddle/fluid/tests/unittests/test_peak_gpumem_monitor.py b/python/paddle/fluid/tests/unittests/test_peak_gpumem_monitor.py new file mode 100644 index 0000000000000000000000000000000000000000..3673fd10c4d14ba8d7a9644dcc024f58cf92a099 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_peak_gpumem_monitor.py @@ -0,0 +1,59 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import print_function + +import unittest +import os +os.environ['FLAGS_benchmark'] = 'True' + +import numpy +import paddle.fluid.core as core +from paddle.fluid.executor import Executor +from paddle.fluid.layers import mul, data + + +class TestPeakMemoryMonitoring(unittest.TestCase): + def test_mul(self): + + a = data(name='a', shape=[784], dtype='float32') + b = data( + name='b', + shape=[784, 100], + dtype='float32', + append_batch_size=False) + out = mul(x=a, y=b) + + if core.is_compiled_with_cuda(): + place = core.CUDAPlace(0) + + a_np = numpy.random.random((100, 784)).astype('float32') + b_np = numpy.random.random((784, 100)).astype('float32') + self.assertEqual(0, core.get_mem_usage(0)) + exe = Executor(place) + outs = exe.run(feed={'a': a_np, 'b': b_np}, fetch_list=[out]) + out = outs[0] + #disable this assert since ctest will ignore the os.environ setting + #self.assertGreater(core.get_mem_usage(0), 0) + + raised = False + try: + core.print_mem_usage() + except: + raised = True + self.assertFalse(raised, 'Exception raised') + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_profiler.py b/python/paddle/fluid/tests/unittests/test_profiler.py index 7934164b84931f886967982ce0cb65c406bbf800..39d778b82a04f403bea030381ff220a68b1ff0ef 100644 --- a/python/paddle/fluid/tests/unittests/test_profiler.py +++ b/python/paddle/fluid/tests/unittests/test_profiler.py @@ -16,15 +16,19 @@ from __future__ import print_function import unittest import os +import tempfile import numpy as np import paddle.fluid as fluid import paddle.fluid.profiler as profiler import paddle.fluid.layers as layers import paddle.fluid.core as core +import paddle.fluid.proto.profiler.profiler_pb2 as profiler_pb2 class TestProfiler(unittest.TestCase): - def net_profiler(self, state, profile_path='/tmp/profile'): + def net_profiler(self, state, use_parallel_executor=False): + profile_path = os.path.join(tempfile.gettempdir(), "profile") + open(profile_path, "w").write("") startup_program = fluid.Program() main_program = fluid.Program() @@ -60,6 +64,11 @@ class TestProfiler(unittest.TestCase): place = fluid.CPUPlace() if state == 'CPU' else fluid.CUDAPlace(0) exe = fluid.Executor(place) exe.run(startup_program) + if use_parallel_executor: + pe = fluid.ParallelExecutor( + state != 'CPU', + loss_name=avg_cost.name, + main_program=main_program) pass_acc_calculator = fluid.average.WeightedAverage() with profiler.profiler(state, 'total', profile_path) as prof: @@ -69,6 +78,9 @@ class TestProfiler(unittest.TestCase): x = np.random.random((32, 784)).astype("float32") y = np.random.randint(0, 10, (32, 1)).astype("int64") + if use_parallel_executor: + pe.run(feed={'x': x, 'y': y}, fetch_list=[avg_cost.name]) + continue outs = exe.run(main_program, feed={'x': x, 'y': y}, @@ -77,21 +89,37 @@ class TestProfiler(unittest.TestCase): b_size = np.array(outs[2]) pass_acc_calculator.add(value=acc, weight=b_size) pass_acc = pass_acc_calculator.eval() + data = open(profile_path, 'rb').read() + self.assertGreater(len(data), 0) + profile_pb = profiler_pb2.Profile() + profile_pb.ParseFromString(data) + self.assertGreater(len(profile_pb.events), 0) + for event in profile_pb.events: + if event.type == profiler_pb2.Event.GPUKernel: + if not event.detail_info and not event.name.startswith("MEM"): + raise Exception( + "Kernel %s missing event. Has this kernel been recorded by RecordEvent?" + % event.name) + elif event.type == profiler_pb2.Event.CPU and ( + event.name.startswith("Driver API") or + event.name.startswith("Runtime API")): + print("Warning: unregister", event.name) def test_cpu_profiler(self): self.net_profiler('CPU') + self.net_profiler('CPU', use_parallel_executor=True) @unittest.skipIf(not core.is_compiled_with_cuda(), "profiler is enabled only with GPU") def test_cuda_profiler(self): self.net_profiler('GPU') + self.net_profiler('GPU', use_parallel_executor=True) @unittest.skipIf(not core.is_compiled_with_cuda(), "profiler is enabled only with GPU") def test_all_profiler(self): - self.net_profiler('All', '/tmp/profile_out') - with open('/tmp/profile_out', 'rb') as f: - self.assertGreater(len(f.read()), 0) + self.net_profiler('All') + self.net_profiler('All', use_parallel_executor=True) if __name__ == '__main__': diff --git a/python/paddle/fluid/tests/unittests/test_yolov3_loss_op.py b/python/paddle/fluid/tests/unittests/test_yolov3_loss_op.py index 544fe4b4f81909b69a05d9751316e3d3137fdc45..020c1139230a9177c4d7765367359d91839d7d46 100644 --- a/python/paddle/fluid/tests/unittests/test_yolov3_loss_op.py +++ b/python/paddle/fluid/tests/unittests/test_yolov3_loss_op.py @@ -16,174 +16,179 @@ from __future__ import division import unittest import numpy as np +from scipy.special import logit +from scipy.special import expit from op_test import OpTest from paddle.fluid import core -def sigmoid(x): - return 1.0 / (1.0 + np.exp(-1.0 * x)) +def l2loss(x, y): + return 0.5 * (y - x) * (y - x) -def mse(x, y, num): - return ((y - x)**2).sum() / num +def sce(x, label): + sigmoid_x = expit(x) + term1 = label * np.log(sigmoid_x) + term2 = (1.0 - label) * np.log(1.0 - sigmoid_x) + return -term1 - term2 -def bce(x, y, mask): - x = x.reshape((-1)) - y = y.reshape((-1)) - mask = mask.reshape((-1)) +def sigmoid(x): + return 1.0 / (1.0 + np.exp(-1.0 * x)) - error_sum = 0.0 - count = 0 - for i in range(x.shape[0]): - if mask[i] > 0: - error_sum += y[i] * np.log(x[i]) + (1 - y[i]) * np.log(1 - x[i]) - count += 1 - return error_sum / (-1.0 * count) +def batch_xywh_box_iou(box1, box2): + b1_left = box1[:, :, 0] - box1[:, :, 2] / 2 + b1_right = box1[:, :, 0] + box1[:, :, 2] / 2 + b1_top = box1[:, :, 1] - box1[:, :, 3] / 2 + b1_bottom = box1[:, :, 1] + box1[:, :, 3] / 2 -def box_iou(box1, box2): - b1_x1 = box1[0] - box1[2] / 2 - b1_x2 = box1[0] + box1[2] / 2 - b1_y1 = box1[1] - box1[3] / 2 - b1_y2 = box1[1] + box1[3] / 2 - b2_x1 = box2[0] - box2[2] / 2 - b2_x2 = box2[0] + box2[2] / 2 - b2_y1 = box2[1] - box2[3] / 2 - b2_y2 = box2[1] + box2[3] / 2 + b2_left = box2[:, :, 0] - box2[:, :, 2] / 2 + b2_right = box2[:, :, 0] + box2[:, :, 2] / 2 + b2_top = box2[:, :, 1] - box2[:, :, 3] / 2 + b2_bottom = box2[:, :, 1] + box2[:, :, 3] / 2 - b1_area = (b1_x2 - b1_x1) * (b1_y2 - b1_y1) - b2_area = (b2_x2 - b2_x1) * (b2_y2 - b2_y1) + left = np.maximum(b1_left[:, :, np.newaxis], b2_left[:, np.newaxis, :]) + right = np.minimum(b1_right[:, :, np.newaxis], b2_right[:, np.newaxis, :]) + top = np.maximum(b1_top[:, :, np.newaxis], b2_top[:, np.newaxis, :]) + bottom = np.minimum(b1_bottom[:, :, np.newaxis], + b2_bottom[:, np.newaxis, :]) - inter_rect_x1 = max(b1_x1, b2_x1) - inter_rect_y1 = max(b1_y1, b2_y1) - inter_rect_x2 = min(b1_x2, b2_x2) - inter_rect_y2 = min(b1_y2, b2_y2) - inter_area = max(inter_rect_x2 - inter_rect_x1, 0) * max( - inter_rect_y2 - inter_rect_y1, 0) + inter_w = np.clip(right - left, 0., 1.) + inter_h = np.clip(bottom - top, 0., 1.) + inter_area = inter_w * inter_h - return inter_area / (b1_area + b2_area + inter_area) + b1_area = (b1_right - b1_left) * (b1_bottom - b1_top) + b2_area = (b2_right - b2_left) * (b2_bottom - b2_top) + union = b1_area[:, :, np.newaxis] + b2_area[:, np.newaxis, :] - inter_area + return inter_area / union -def build_target(gtboxs, gtlabel, attrs, grid_size): - n, b, _ = gtboxs.shape - ignore_thresh = attrs["ignore_thresh"] - anchors = attrs["anchors"] - class_num = attrs["class_num"] - an_num = len(anchors) // 2 - obj_mask = np.zeros((n, an_num, grid_size, grid_size)).astype('float32') - noobj_mask = np.ones((n, an_num, grid_size, grid_size)).astype('float32') - tx = np.zeros((n, an_num, grid_size, grid_size)).astype('float32') - ty = np.zeros((n, an_num, grid_size, grid_size)).astype('float32') - tw = np.zeros((n, an_num, grid_size, grid_size)).astype('float32') - th = np.zeros((n, an_num, grid_size, grid_size)).astype('float32') - tconf = np.zeros((n, an_num, grid_size, grid_size)).astype('float32') - tcls = np.zeros( - (n, an_num, grid_size, grid_size, class_num)).astype('float32') +def YOLOv3Loss(x, gtbox, gtlabel, attrs): + n, c, h, w = x.shape + b = gtbox.shape[1] + anchors = attrs['anchors'] + an_num = len(anchors) // 2 + anchor_mask = attrs['anchor_mask'] + mask_num = len(anchor_mask) + class_num = attrs["class_num"] + ignore_thresh = attrs['ignore_thresh'] + downsample = attrs['downsample'] + input_size = downsample * h + x = x.reshape((n, mask_num, 5 + class_num, h, w)).transpose((0, 1, 3, 4, 2)) + loss = np.zeros((n)).astype('float32') + + pred_box = x[:, :, :, :, :4].copy() + grid_x = np.tile(np.arange(w).reshape((1, w)), (h, 1)) + grid_y = np.tile(np.arange(h).reshape((h, 1)), (1, w)) + pred_box[:, :, :, :, 0] = (grid_x + sigmoid(pred_box[:, :, :, :, 0])) / w + pred_box[:, :, :, :, 1] = (grid_y + sigmoid(pred_box[:, :, :, :, 1])) / h + + x[:, :, :, :, 5:] = np.where(x[:, :, :, :, 5:] < -0.5, x[:, :, :, :, 5:], + np.ones_like(x[:, :, :, :, 5:]) * 1.0 / + class_num) + + mask_anchors = [] + for m in anchor_mask: + mask_anchors.append((anchors[2 * m], anchors[2 * m + 1])) + anchors_s = np.array( + [(an_w / input_size, an_h / input_size) for an_w, an_h in mask_anchors]) + anchor_w = anchors_s[:, 0:1].reshape((1, mask_num, 1, 1)) + anchor_h = anchors_s[:, 1:2].reshape((1, mask_num, 1, 1)) + pred_box[:, :, :, :, 2] = np.exp(pred_box[:, :, :, :, 2]) * anchor_w + pred_box[:, :, :, :, 3] = np.exp(pred_box[:, :, :, :, 3]) * anchor_h + + pred_box = pred_box.reshape((n, -1, 4)) + pred_obj = x[:, :, :, :, 4].reshape((n, -1)) + objness = np.zeros(pred_box.shape[:2]).astype('float32') + ious = batch_xywh_box_iou(pred_box, gtbox) + ious_max = np.max(ious, axis=-1) + objness = np.where(ious_max > ignore_thresh, -np.ones_like(objness), + objness) + + gtbox_shift = gtbox.copy() + gtbox_shift[:, :, 0] = 0 + gtbox_shift[:, :, 1] = 0 + + anchors = [(anchors[2 * i], anchors[2 * i + 1]) for i in range(0, an_num)] + anchors_s = np.array( + [(an_w / input_size, an_h / input_size) for an_w, an_h in anchors]) + anchor_boxes = np.concatenate( + [np.zeros_like(anchors_s), anchors_s], axis=-1) + anchor_boxes = np.tile(anchor_boxes[np.newaxis, :, :], (n, 1, 1)) + ious = batch_xywh_box_iou(gtbox_shift, anchor_boxes) + iou_matches = np.argmax(ious, axis=-1) + gt_matches = iou_matches.copy() for i in range(n): for j in range(b): - if gtboxs[i, j, :].sum() == 0: + if gtbox[i, j, 2:].sum() == 0: + gt_matches[i, j] = -1 continue + if iou_matches[i, j] not in anchor_mask: + gt_matches[i, j] = -1 + continue + an_idx = anchor_mask.index(iou_matches[i, j]) + gt_matches[i, j] = an_idx + gi = int(gtbox[i, j, 0] * w) + gj = int(gtbox[i, j, 1] * h) - gt_label = gtlabel[i, j] - gx = gtboxs[i, j, 0] * grid_size - gy = gtboxs[i, j, 1] * grid_size - gw = gtboxs[i, j, 2] * grid_size - gh = gtboxs[i, j, 3] * grid_size - - gi = int(gx) - gj = int(gy) - - gtbox = [0, 0, gw, gh] - max_iou = 0 - for k in range(an_num): - anchor_box = [0, 0, anchors[2 * k], anchors[2 * k + 1]] - iou = box_iou(gtbox, anchor_box) - if iou > max_iou: - max_iou = iou - best_an_index = k - if iou > ignore_thresh: - noobj_mask[i, best_an_index, gj, gi] = 0 - - obj_mask[i, best_an_index, gj, gi] = 1 - noobj_mask[i, best_an_index, gj, gi] = 0 - tx[i, best_an_index, gj, gi] = gx - gi - ty[i, best_an_index, gj, gi] = gy - gj - tw[i, best_an_index, gj, gi] = np.log(gw / anchors[2 * - best_an_index]) - th[i, best_an_index, gj, gi] = np.log( - gh / anchors[2 * best_an_index + 1]) - tconf[i, best_an_index, gj, gi] = 1 - tcls[i, best_an_index, gj, gi, gt_label] = 1 - - return (tx, ty, tw, th, tconf, tcls, obj_mask, noobj_mask) - - -def YoloV3Loss(x, gtbox, gtlabel, attrs): - n, c, h, w = x.shape - an_num = len(attrs['anchors']) // 2 - class_num = attrs["class_num"] - x = x.reshape((n, an_num, 5 + class_num, h, w)).transpose((0, 1, 3, 4, 2)) - pred_x = sigmoid(x[:, :, :, :, 0]) - pred_y = sigmoid(x[:, :, :, :, 1]) - pred_w = x[:, :, :, :, 2] - pred_h = x[:, :, :, :, 3] - pred_conf = sigmoid(x[:, :, :, :, 4]) - pred_cls = sigmoid(x[:, :, :, :, 5:]) - - tx, ty, tw, th, tconf, tcls, obj_mask, noobj_mask = build_target( - gtbox, gtlabel, attrs, x.shape[2]) - - obj_mask_expand = np.tile( - np.expand_dims(obj_mask, 4), (1, 1, 1, 1, int(attrs['class_num']))) - loss_x = mse(pred_x * obj_mask, tx * obj_mask, obj_mask.sum()) - loss_y = mse(pred_y * obj_mask, ty * obj_mask, obj_mask.sum()) - loss_w = mse(pred_w * obj_mask, tw * obj_mask, obj_mask.sum()) - loss_h = mse(pred_h * obj_mask, th * obj_mask, obj_mask.sum()) - loss_conf_target = bce(pred_conf * obj_mask, tconf * obj_mask, obj_mask) - loss_conf_notarget = bce(pred_conf * noobj_mask, tconf * noobj_mask, - noobj_mask) - loss_class = bce(pred_cls * obj_mask_expand, tcls * obj_mask_expand, - obj_mask_expand) - - return attrs['loss_weight_xy'] * (loss_x + loss_y) \ - + attrs['loss_weight_wh'] * (loss_w + loss_h) \ - + attrs['loss_weight_conf_target'] * loss_conf_target \ - + attrs['loss_weight_conf_notarget'] * loss_conf_notarget \ - + attrs['loss_weight_class'] * loss_class + tx = gtbox[i, j, 0] * w - gi + ty = gtbox[i, j, 1] * w - gj + tw = np.log(gtbox[i, j, 2] * input_size / mask_anchors[an_idx][0]) + th = np.log(gtbox[i, j, 3] * input_size / mask_anchors[an_idx][1]) + scale = (2.0 - gtbox[i, j, 2] * gtbox[i, j, 3]) + loss[i] += sce(x[i, an_idx, gj, gi, 0], tx) * scale + loss[i] += sce(x[i, an_idx, gj, gi, 1], ty) * scale + loss[i] += l2loss(x[i, an_idx, gj, gi, 2], tw) * scale + loss[i] += l2loss(x[i, an_idx, gj, gi, 3], th) * scale + + objness[i, an_idx * h * w + gj * w + gi] = 1.0 + + for label_idx in range(class_num): + loss[i] += sce(x[i, an_idx, gj, gi, 5 + label_idx], + float(label_idx == gtlabel[i, j])) + + for j in range(mask_num * h * w): + if objness[i, j] > 0: + loss[i] += sce(pred_obj[i, j], 1.0) + elif objness[i, j] == 0: + loss[i] += sce(pred_obj[i, j], 0.0) + + return (loss, objness.reshape((n, mask_num, h, w)).astype('float32'), \ + gt_matches.astype('int32')) class TestYolov3LossOp(OpTest): def setUp(self): - self.loss_weight_xy = 1.0 - self.loss_weight_wh = 1.0 - self.loss_weight_conf_target = 1.0 - self.loss_weight_conf_notarget = 1.0 - self.loss_weight_class = 1.0 self.initTestCase() self.op_type = 'yolov3_loss' - x = np.random.random(size=self.x_shape).astype('float32') + x = logit(np.random.uniform(0, 1, self.x_shape).astype('float32')) gtbox = np.random.random(size=self.gtbox_shape).astype('float32') - gtlabel = np.random.randint(0, self.class_num, - self.gtbox_shape[:2]).astype('int32') + gtlabel = np.random.randint(0, self.class_num, self.gtbox_shape[:2]) + gtmask = np.random.randint(0, 2, self.gtbox_shape[:2]) + gtbox = gtbox * gtmask[:, :, np.newaxis] + gtlabel = gtlabel * gtmask self.attrs = { "anchors": self.anchors, + "anchor_mask": self.anchor_mask, "class_num": self.class_num, "ignore_thresh": self.ignore_thresh, - "loss_weight_xy": self.loss_weight_xy, - "loss_weight_wh": self.loss_weight_wh, - "loss_weight_conf_target": self.loss_weight_conf_target, - "loss_weight_conf_notarget": self.loss_weight_conf_notarget, - "loss_weight_class": self.loss_weight_class, + "downsample": self.downsample, } - self.inputs = {'X': x, 'GTBox': gtbox, 'GTLabel': gtlabel} + self.inputs = { + 'X': x, + 'GTBox': gtbox.astype('float32'), + 'GTLabel': gtlabel.astype('int32'), + } + loss, objness, gt_matches = YOLOv3Loss(x, gtbox, gtlabel, self.attrs) self.outputs = { - 'Loss': np.array( - [YoloV3Loss(x, gtbox, gtlabel, self.attrs)]).astype('float32') + 'Loss': loss, + 'ObjectnessMask': objness, + "GTMatchMask": gt_matches } def test_check_output(self): @@ -196,19 +201,16 @@ class TestYolov3LossOp(OpTest): place, ['X'], 'Loss', no_grad_set=set(["GTBox", "GTLabel"]), - max_relative_error=0.06) + max_relative_error=0.3) def initTestCase(self): - self.anchors = [10, 13, 12, 12] - self.class_num = 10 + self.anchors = [10, 13, 16, 30, 33, 23] + self.anchor_mask = [1, 2] + self.class_num = 5 self.ignore_thresh = 0.5 - self.x_shape = (5, len(self.anchors) // 2 * (5 + self.class_num), 7, 7) - self.gtbox_shape = (5, 10, 4) - self.loss_weight_xy = 2.5 - self.loss_weight_wh = 0.8 - self.loss_weight_conf_target = 1.5 - self.loss_weight_conf_notarget = 0.5 - self.loss_weight_class = 1.2 + self.downsample = 32 + self.x_shape = (3, len(self.anchor_mask) * (5 + self.class_num), 5, 5) + self.gtbox_shape = (3, 5, 4) if __name__ == "__main__": diff --git a/python/paddle/fluid/tests/unittests/transformer_model.py b/python/paddle/fluid/tests/unittests/transformer_model.py index 143d187edc3a154418f9e639b7d492c8ce994d42..905b7d6fe75ab0080e3e97fbd4710ad913a05a38 100644 --- a/python/paddle/fluid/tests/unittests/transformer_model.py +++ b/python/paddle/fluid/tests/unittests/transformer_model.py @@ -17,6 +17,7 @@ from __future__ import print_function from functools import partial import numpy as np +import os import paddle.fluid as fluid import paddle.fluid.layers as layers from paddle.fluid.layers.io import open_recordio_file @@ -408,7 +409,7 @@ def transformer( trg_pad_idx, pos_pad_idx, ): file_obj = open_recordio_file( - filename='/tmp/wmt16.recordio', + filename=os.environ.get('RECORDIO_FILENAME', '/tmp/wmt16.recordio'), shapes=[ [batch_size * max_length, 1], [batch_size * max_length, 1], diff --git a/python/paddle/fluid/transpiler/details/__init__.py b/python/paddle/fluid/transpiler/details/__init__.py index f33c05ed2f48c2498b98fc486d6ff7471088d77e..82d0d336e523ec48c5ceca3b92ff0963c4499123 100644 --- a/python/paddle/fluid/transpiler/details/__init__.py +++ b/python/paddle/fluid/transpiler/details/__init__.py @@ -17,3 +17,4 @@ from __future__ import print_function from .program_utils import * from .ufind import * from .checkport import * +from .vars_distributed import * diff --git a/python/paddle/fluid/transpiler/details/vars_distributed.py b/python/paddle/fluid/transpiler/details/vars_distributed.py new file mode 100644 index 0000000000000000000000000000000000000000..05e7f6e3e706376efc8af870a780d96c45642514 --- /dev/null +++ b/python/paddle/fluid/transpiler/details/vars_distributed.py @@ -0,0 +1,269 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from __future__ import print_function +from paddle.fluid.framework import Variable + + +class VarStruct(object): + """ + record part properties of a Variable in python. + """ + + def __init__(self, name, shape, dtype, type, lod_level, persistable): + self.name = name + self.shape = shape + self.dtype = dtype + self.type = type + self.lod_level = lod_level + self.persistable = persistable + + +class VarDistributed(object): + """ + a class to record the var distributed on parameter servers. + the class will record the relationship between origin var and slice var. + the slice var's properties, such as type/shape/offset/endpoint. + """ + + def __init__(self, + origin_var, + slice_var, + is_slice=None, + block_id=None, + offset=None, + vtype=None, + endpoint=None): + """ + Args: + origin_var(Variable|VarStruct): origin var properties + slice_var(Variable|VarStruct): slice var properties + is_slice(bool|None): slice or not, slice_var=True/False and its block size > 8192 are the judgement standard. + block_id(int|None): the number about the slice var. + offset(int|None): if the slice var is sliced, offset is the numel before the var. + vtype(str|None): a tag, such as Optimizer/Param/RemoteProfetch. + endpoint(str|None): which parameter the slice var on, such as "127.0.0.1:1001" + """ + + if isinstance(origin_var, Variable): + self.origin = self.__create_var_struct(origin_var) + else: + self.origin = origin_var + + if isinstance(slice_var, Variable): + self.slice = self.__create_var_struct(slice_var) + else: + self.slice = slice_var + + if self.equal(self.origin, self.slice): + self.is_slice = False + self.block_id = 0 + self.offset = 0 + else: + self.is_slice = True + self.block_id = 0 + self.offset = 0 + + if is_slice is not None: + self.is_slice = is_slice + if block_id is not None: + self.block_id = block_id + if offset is not None: + self.offset = offset + + self.vtype = vtype + self.endpoint = endpoint + + @staticmethod + def __create_var_struct(var): + return VarStruct(var.name, var.shape, var.dtype, var.type, + var.lod_level, var.persistable) + + @staticmethod + def equal(var1, var2): + """ + the two var is equal or not. + Returns: + bool: equal will return True else False + """ + assert isinstance(var1, VarStruct) and isinstance(var2, VarStruct) + + return var1.name == var2.name and \ + var1.type == var2.type and \ + var1.shape == var2.shape and \ + var1.dtype == var2.dtype and \ + var1.lod_level == var2.lod_level and \ + var1.persistable == var2.persistable + + def __str__(self): + origin_var_str = "{name} : fluid.{type}.shape{shape}.astype({dtype})". \ + format(i="{", e="}", name=self.origin.name, type=self.origin.type, + shape=self.origin.shape, dtype=self.origin.dtype) + + slice_var_str = "{name} : fluid.{type}.shape{shape}.astype({dtype})" \ + ".slice({is_slice}).block({block_id}).offset({offset})". \ + format(i="{", e="}", name=self.slice.name, type=self.slice.type, + shape=self.slice.shape, dtype=self.slice.dtype, + is_slice=self.is_slice, block_id=self.block_id, offset=self.offset) + + return "var owned: {}, origin var: ( {} ), slice var: ( {} ), endpoint: {} ".format( + self.vtype, origin_var_str, slice_var_str, self.endpoint) + + +class VarsDistributed(object): + """ + a gather about VarDistributed with many methods to find distributed vars. + through the class, we can get overview about the distributed parameters on parameter servers. + this class may centralized and convenient for developer to manage and get variable's distribute. + other module can also use this to find variables such io.py. + """ + + def __init__(self): + self.distributed_vars = [] + + def add_distributed_var(self, + origin_var, + slice_var, + is_slice=None, + block_id=None, + offset=None, + vtype=None, + endpoint=None): + """ + add distributed var in this. + + Args: + origin_var(Variable|VarStruct): origin var properties + slice_var(Variable|VarStruct): slice var properties + is_slice(bool|None): slice or not, slice_var=True/False and its block size > 8192 are the judgement standard. + block_id(int|None): the number about the slice var. + offset(int|None): if the slice var is sliced, offset is the numel before the var. + vtype(str|None): a tag, such as Optimizer/Param/RemoteProfetch. + endpoint(str|None): which parameter the slice var on, such as "127.0.0.1:1001" + Returns: + None + """ + self.distributed_vars.append( + VarDistributed(origin_var, slice_var, is_slice, block_id, offset, + vtype, endpoint)) + + def get_distributed_var_by_slice(self, var_name): + """ + get distributed var by conditions. + + Args: + var_name(str): slice var name, such as "w.traier0.block1" + Returns: + VarDistributed: distributed var. + """ + for dist_var in self.distributed_vars: + if dist_var.slice.name == var_name: + return dist_var + return None + + @staticmethod + def equal(var1, var2): + """ + the two var is equal or not. + Returns: + bool: equal will return True else False + """ + return var1.name == var2.name and \ + var1.type == var2.type and \ + var1.shape == var2.shape and \ + var1.dtype == var2.dtype and \ + var1.lod_level == var2.lod_level and \ + var1.persistable == var2.persistable + + def get_distributed_var_by_origin_and_ep(self, origin_var_name, endpoint): + """ + get distributed var by conditions. + + Args: + origin_var_name(str): + endpoint(str): the parameter endpoint, such as "127.0.0.1:1001" + Returns: + VarDistributed: distributed var. + """ + for dist_var in self.distributed_vars: + if dist_var.origin.name == origin_var_name and dist_var.endpoint == endpoint: + return dist_var + return None + + def get_distributed_vars_by_vtypes(self, vtypes, groupby=False): + """ + get distributed vars by conditions. + + Args: + vtype(str|None): distributed var's vtype, such as "Optimizer", "RemotePrefetch" + groupby(bool|False): group by origin var or not. + + Returns: + list: distributed var list. + dict: distributed var map when groupby=True + """ + vtype_vars = [] + for var in self.distributed_vars: + if var.vtype in vtypes: + vtype_vars.append(var) + if not groupby: + return vtype_vars + + params_map = {} + for var in vtype_vars: + origin_var_name = var.origin.name + + if origin_var_name in params_map.keys(): + optimizers = params_map.get(origin_var_name) + else: + optimizers = [] + optimizers.append(var) + params_map[origin_var_name] = optimizers + return params_map + + def get_distributed_vars_by_ep(self, endpoint, vtype=None): + """ + get distributed vars by conditions. + + Args: + endpoint(str): the parameter server endpoint, such as "127.0.0.1:2001" + vtype(str|None): distributed var's vtype, such as "Optimizer", "RemotePrefetch" + + Returns: + list: distributed var list. + """ + endpoint_vars = [] + for var in self.distributed_vars: + if var.endpoint == endpoint: + endpoint_vars.append(var) + if not vtype: + return endpoint_vars + + vtype_vars = [] + for var in endpoint_vars: + if var.vtype == vtype: + vtype_vars.append(var) + return vtype_vars + + def overview(self): + """ + get the overview string about all params on all parameter servers. + + Returns: + Str: overview string. + + """ + vars_str = [] + for var in self.distributed_vars: + vars_str.append(str(var)) + return "\n".join(vars_str) diff --git a/python/paddle/fluid/transpiler/distribute_transpiler.py b/python/paddle/fluid/transpiler/distribute_transpiler.py index e58f34e3750803669149685003ea5858fa775ed7..eb54068650e8b3f4e64317778e2ad7c7aa7fe1b2 100644 --- a/python/paddle/fluid/transpiler/distribute_transpiler.py +++ b/python/paddle/fluid/transpiler/distribute_transpiler.py @@ -30,19 +30,23 @@ Steps to transpile pserver: 5. add listen_and_serv op """ +import sys import math -import numpy as np +from functools import reduce + import collections +import six import logging +import numpy as np + from .ps_dispatcher import RoundRobin, PSDispatcher from .. import core, framework, unique_name from ..framework import Program, default_main_program, \ - default_startup_program, Block, \ - Parameter, Variable, grad_var_name -from .details import * + default_startup_program, Block, Parameter, grad_var_name +from .details import wait_server_ready, UnionFind, VarStruct, VarsDistributed +from .details import delete_ops, find_op_by_output_arg from ..distribute_lookup_table import find_distributed_lookup_table -from functools import reduce LOOKUP_TABLE_TYPE = "lookup_table" LOOKUP_TABLE_GRAD_TYPE = "lookup_table_grad" @@ -62,260 +66,6 @@ def log(*args): print(args) -class VarStruct(object): - """ - record part properties of a Variable in python. - """ - - def __init__(self, name, shape, dtype, type, lod_level, persistable): - self.name = name - self.shape = shape - self.dtype = dtype - self.type = type - self.lod_level = lod_level - self.persistable = persistable - - -class VarDistributed(object): - """ - a class to record the var distributed on parameter servers. - the class will record the relationship between origin var and slice var. - the slice var's properties, such as type/shape/offset/endpoint. - """ - - def __init__(self, - origin_var, - slice_var, - is_slice=None, - block_id=None, - offset=None, - vtype=None, - endpoint=None): - """ - Args: - origin_var(Variable|VarStruct): origin var properties - slice_var(Variable|VarStruct): slice var properties - is_slice(bool|None): slice or not, slice_var=True/False and its block size > 8192 are the judgement standard. - block_id(int|None): the number about the slice var. - offset(int|None): if the slice var is sliced, offset is the numel before the var. - vtype(str|None): a tag, such as Optimizer/Param/RemoteProfetch. - endpoint(str|None): which parameter the slice var on, such as "127.0.0.1:1001" - """ - - if isinstance(origin_var, Variable): - self.origin = self.__create_var_struct(origin_var) - else: - self.origin = origin_var - - if isinstance(slice_var, Variable): - self.slice = self.__create_var_struct(slice_var) - else: - self.slice = slice_var - - if self.equal(self.origin, self.slice): - self.is_slice = False - self.block_id = 0 - self.offset = 0 - else: - self.is_slice = True - self.block_id = 0 - self.offset = 0 - - if is_slice is not None: - self.is_slice = is_slice - if block_id is not None: - self.block_id = block_id - if offset is not None: - self.offset = offset - - self.vtype = vtype - self.endpoint = endpoint - - @staticmethod - def __create_var_struct(var): - return VarStruct(var.name, var.shape, var.dtype, var.type, - var.lod_level, var.persistable) - - @staticmethod - def equal(var1, var2): - """ - the two var is equal or not. - Returns: - bool: equal will return True else False - """ - assert isinstance(var1, VarStruct) and isinstance(var2, VarStruct) - - return var1.name == var2.name and \ - var1.type == var2.type and \ - var1.shape == var2.shape and \ - var1.dtype == var2.dtype and \ - var1.lod_level == var2.lod_level and \ - var1.persistable == var2.persistable - - def __str__(self): - origin_var_str = "{name} : fluid.{type}.shape{shape}.astype({dtype})". \ - format(i="{", e="}", name=self.origin.name, type=self.origin.type, - shape=self.origin.shape, dtype=self.origin.dtype) - - slice_var_str = "{name} : fluid.{type}.shape{shape}.astype({dtype})" \ - ".slice({is_slice}).block({block_id}).offset({offset})". \ - format(i="{", e="}", name=self.slice.name, type=self.slice.type, - shape=self.slice.shape, dtype=self.slice.dtype, - is_slice=self.is_slice, block_id=self.block_id, offset=self.offset) - - return "var owned: {}, origin var: ( {} ), slice var: ( {} ), endpoint: {} ".format( - self.vtype, origin_var_str, slice_var_str, self.endpoint) - - -class VarsDistributed(object): - """ - a gather about VarDistributed with many methods to find distributed vars. - through the class, we can get overview about the distributed parameters on parameter servers. - this class may centralized and convenient for developer to manage and get variable's distribute. - other module can also use this to find variables such io.py. - """ - - def __init__(self): - self.distributed_vars = [] - - def add_distributed_var(self, - origin_var, - slice_var, - is_slice=None, - block_id=None, - offset=None, - vtype=None, - endpoint=None): - """ - add distributed var in this. - - Args: - origin_var(Variable|VarStruct): origin var properties - slice_var(Variable|VarStruct): slice var properties - is_slice(bool|None): slice or not, slice_var=True/False and its block size > 8192 are the judgement standard. - block_id(int|None): the number about the slice var. - offset(int|None): if the slice var is sliced, offset is the numel before the var. - vtype(str|None): a tag, such as Optimizer/Param/RemoteProfetch. - endpoint(str|None): which parameter the slice var on, such as "127.0.0.1:1001" - Returns: - None - """ - self.distributed_vars.append( - VarDistributed(origin_var, slice_var, is_slice, block_id, offset, - vtype, endpoint)) - - def get_distributed_var_by_slice(self, var_name): - """ - get distributed var by conditions. - - Args: - var_name(str): slice var name, such as "w.traier0.block1" - Returns: - VarDistributed: distributed var. - """ - for dist_var in self.distributed_vars: - if dist_var.slice.name == var_name: - return dist_var - return None - - @staticmethod - def equal(var1, var2): - """ - the two var is equal or not. - Returns: - bool: equal will return True else False - """ - return var1.name == var2.name and \ - var1.type == var2.type and \ - var1.shape == var2.shape and \ - var1.dtype == var2.dtype and \ - var1.lod_level == var2.lod_level and \ - var1.persistable == var2.persistable - - def get_distributed_var_by_origin_and_ep(self, origin_var_name, endpoint): - """ - get distributed var by conditions. - - Args: - origin_var_name(str): - endpoint(str): the parameter endpoint, such as "127.0.0.1:1001" - Returns: - VarDistributed: distributed var. - """ - for dist_var in self.distributed_vars: - if dist_var.origin.name == origin_var_name and dist_var.endpoint == endpoint: - return dist_var - return None - - def get_distributed_vars_by_vtypes(self, vtypes, groupby=False): - """ - get distributed vars by conditions. - - Args: - vtype(str|None): distributed var's vtype, such as "Optimizer", "RemotePrefetch" - groupby(bool|False): group by origin var or not. - - Returns: - list: distributed var list. - dict: distributed var map when groupby=True - """ - vtype_vars = [] - for var in self.distributed_vars: - if var.vtype in vtypes: - vtype_vars.append(var) - if not groupby: - return vtype_vars - - params_map = {} - for var in vtype_vars: - origin_var_name = var.origin.name - - if origin_var_name in params_map.keys(): - optimizers = params_map.get(origin_var_name) - else: - optimizers = [] - optimizers.append(var) - params_map[origin_var_name] = optimizers - return params_map - - def get_distributed_vars_by_ep(self, endpoint, vtype=None): - """ - get distributed vars by conditions. - - Args: - endpoint(str): the parameter server endpoint, such as "127.0.0.1:2001" - vtype(str|None): distributed var's vtype, such as "Optimizer", "RemotePrefetch" - - Returns: - list: distributed var list. - """ - endpoint_vars = [] - for var in self.distributed_vars: - if var.endpoint == endpoint: - endpoint_vars.append(var) - if not vtype: - return endpoint_vars - - vtype_vars = [] - for var in endpoint_vars: - if var.vtype == vtype: - vtype_vars.append(var) - return vtype_vars - - def overview(self): - """ - get the overview string about all params on all parameter servers. - - Returns: - Str: overview string. - - """ - vars_str = [] - for var in self.distributed_vars: - vars_str.append(str(var)) - return "\n".join(vars_str) - - class VarBlock: def __init__(self, varname, offset, size): self.varname = varname @@ -1270,7 +1020,11 @@ class DistributeTranspiler(object): skip_dim0 = 0 slice_vars = self.param_var_mapping[orig_var_name] - orig_dim1_flatten = reduce(lambda x, y: x * y, slice_vars[0].shape[1:]) + orig_dim1_flatten = 1 + + if len(slice_vars[0].shape) >= 2: + orig_dim1_flatten = reduce(lambda x, y: x * y, + slice_vars[0].shape[1:]) for slice_var in slice_vars[:block_idx]: skip_dim0 += slice_var.shape[0] diff --git a/python/paddle/fluid/transpiler/memory_optimization_transpiler.py b/python/paddle/fluid/transpiler/memory_optimization_transpiler.py index e5d48d3d19ed71624d528144f13e23770a09362a..ee8cde441ffc63ebd923bd579a7f44d1e2218cf0 100755 --- a/python/paddle/fluid/transpiler/memory_optimization_transpiler.py +++ b/python/paddle/fluid/transpiler/memory_optimization_transpiler.py @@ -355,6 +355,10 @@ class ControlFlowGraph(object): is_forward).dtype() cache_dtype = self._find_var(block_desc, cache_var, is_forward).dtype() + if x_dtype != cache_dtype: + if PRINT_LOG: + print("x_dtype and cache_dtype are different") + continue if not compare_shape(x_shape, cache_shape, level): continue @@ -540,6 +544,7 @@ def memory_optimize(input_program, if skip_opt_set is not None: skip_opt_set = set(map(to_name_str, skip_opt_set)) cfgs = _get_cfgs(input_program) + input_program._is_mem_optimized = True for cfg in cfgs: cfg.memory_optimize(skip_opt_set=skip_opt_set, level=level) @@ -559,5 +564,6 @@ def release_memory(input_program, skip_opt_set=None): None """ cfgs = _get_cfgs(input_program) + input_program._is_mem_optimized = True for cfg in cfgs: cfg.release_memory(skip_opt_set=skip_opt_set) diff --git a/python/paddle/fluid/unique_name.py b/python/paddle/fluid/unique_name.py index b9957a699e597898bee75ce0e7283f7224293f0c..324257c13ff9828b341ca9affe8186387688c0bf 100644 --- a/python/paddle/fluid/unique_name.py +++ b/python/paddle/fluid/unique_name.py @@ -15,7 +15,7 @@ from __future__ import print_function import collections -import contextlib +from .wrapped_decorator import signature_safe_contextmanager import six import sys @@ -68,7 +68,7 @@ def switch(new_generator=None): return old -@contextlib.contextmanager +@signature_safe_contextmanager def guard(new_generator=None): if isinstance(new_generator, six.string_types): new_generator = UniqueNameGenerator(new_generator) diff --git a/python/paddle/fluid/wrapped_decorator.py b/python/paddle/fluid/wrapped_decorator.py new file mode 100644 index 0000000000000000000000000000000000000000..7e7dbff65611e947d1a11a0c33c6ecc27e6df636 --- /dev/null +++ b/python/paddle/fluid/wrapped_decorator.py @@ -0,0 +1,30 @@ +# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import decorator +import contextlib + +__all__ = ['wrap_decorator', 'signature_safe_contextmanager'] + + +def wrap_decorator(decorator_func): + @decorator.decorator + def __impl__(func, *args, **kwargs): + wrapped_func = decorator_func(func) + return wrapped_func(*args, **kwargs) + + return __impl__ + + +signature_safe_contextmanager = wrap_decorator(contextlib.contextmanager) diff --git a/python/paddle/utils/dump_config.py b/python/paddle/utils/dump_config.py deleted file mode 100644 index 6a96a0a78fc77c50904ee7822c725c41e646c5e6..0000000000000000000000000000000000000000 --- a/python/paddle/utils/dump_config.py +++ /dev/null @@ -1,45 +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 paddle.trainer.config_parser import parse_config -from paddle.proto import TrainerConfig_pb2 -import sys - -__all__ = [] - -if __name__ == '__main__': - whole_conf = False - binary = False - if len(sys.argv) == 2: - conf = parse_config(sys.argv[1], '') - elif len(sys.argv) == 3: - conf = parse_config(sys.argv[1], sys.argv[2]) - elif len(sys.argv) == 4: - conf = parse_config(sys.argv[1], sys.argv[2]) - if sys.argv[3] == '--whole': - whole_conf = True - elif sys.argv[3] == '--binary': - binary = True - else: - raise RuntimeError() - - assert isinstance(conf, TrainerConfig_pb2.TrainerConfig) - - if whole_conf: - print(conf) - else: - if binary: - sys.stdout.write(conf.model_config.SerializeToString()) - else: - print(conf.model_config) diff --git a/python/paddle/utils/dump_v2_config.py b/python/paddle/utils/dump_v2_config.py deleted file mode 100644 index 5dc2111e379fd39b40e1e9bcf2e577b57b101a68..0000000000000000000000000000000000000000 --- a/python/paddle/utils/dump_v2_config.py +++ /dev/null @@ -1,62 +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. -import collections - -from paddle.trainer_config_helpers.layers import LayerOutput -from paddle.v2.layer import parse_network -from paddle.proto import TrainerConfig_pb2 - -__all__ = ["dump_v2_config"] - - -def dump_v2_config(topology, save_path, binary=False): - """ Dump the network topology to a specified file. - - This function is only used to dump network defined by using PaddlePaddle V2 - APIs. This function will NOT dump configurations related to PaddlePaddle - optimizer. - - :param topology: The output layers (can be more than one layers given in a - Python List or Tuple) of the entire network. Using the - specified layers (if more than one layer is given) as root, - traversing back to the data layer(s), all the layers - connected to the specified output layers will be dumped. - Layers not connceted to the specified will not be dumped. - :type topology: LayerOutput|List|Tuple - :param save_path: The path to save the dumped network topology. - :type save_path: str - :param binary: Whether to dump the serialized network topology or not. - The default value is false. NOTE that, if you call this - function to generate network topology for PaddlePaddle C-API, - a serialized version of network topology is required. When - using PaddlePaddle C-API, this flag MUST be set to True. - :type binary: bool - """ - - if isinstance(topology, LayerOutput): - topology = [topology] - elif isinstance(topology, collections.Sequence): - for out_layer in topology: - assert isinstance(out_layer, LayerOutput), ( - "The type of each element in the parameter topology " - "should be LayerOutput.") - else: - raise RuntimeError("Error input type for parameter topology.") - - model_str = parse_network(topology) - with open(save_path, "w") as fout: - if binary: - fout.write(model_str.SerializeToString()) - else: - fout.write(str(model_str)) diff --git a/python/paddle/utils/image_multiproc.py b/python/paddle/utils/image_multiproc.py deleted file mode 100644 index d1bbda3fd3562efe486377d41a9fb7359bafa4e7..0000000000000000000000000000000000000000 --- a/python/paddle/utils/image_multiproc.py +++ /dev/null @@ -1,278 +0,0 @@ -# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -import os, sys -import numpy as np -from PIL import Image -import six -from six.moves import cStringIO as StringIO -import multiprocessing -import functools -import itertools - -from paddle.utils.image_util import * -from paddle.trainer.config_parser import logger - -try: - import cv2 -except ImportError: - logger.warning("OpenCV2 is not installed, using PIL to process") - cv2 = None - -__all__ = ["CvTransformer", "PILTransformer", "MultiProcessImageTransformer"] - - -class CvTransformer(ImageTransformer): - """ - CvTransformer used python-opencv to process image. - """ - - def __init__( - self, - min_size=None, - crop_size=None, - transpose=(2, 0, 1), # transpose to C * H * W - channel_swap=None, - mean=None, - is_train=True, - is_color=True): - ImageTransformer.__init__(self, transpose, channel_swap, mean, is_color) - self.min_size = min_size - self.crop_size = crop_size - self.is_train = is_train - - def resize(self, im, min_size): - row, col = im.shape[:2] - new_row, new_col = min_size, min_size - if row > col: - new_row = min_size * row / col - else: - new_col = min_size * col / row - im = cv2.resize(im, (new_row, new_col), interpolation=cv2.INTER_CUBIC) - return im - - def crop_and_flip(self, im): - """ - Return cropped image. - The size of the cropped image is inner_size * inner_size. - im: (H x W x K) ndarrays - """ - row, col = im.shape[:2] - start_h, start_w = 0, 0 - if self.is_train: - start_h = np.random.randint(0, row - self.crop_size + 1) - start_w = np.random.randint(0, col - self.crop_size + 1) - else: - start_h = (row - self.crop_size) / 2 - start_w = (col - self.crop_size) / 2 - end_h, end_w = start_h + self.crop_size, start_w + self.crop_size - if self.is_color: - im = im[start_h:end_h, start_w:end_w, :] - else: - im = im[start_h:end_h, start_w:end_w] - if (self.is_train) and (np.random.randint(2) == 0): - if self.is_color: - im = im[:, ::-1, :] - else: - im = im[:, ::-1] - return im - - def transform(self, im): - im = self.resize(im, self.min_size) - im = self.crop_and_flip(im) - # transpose, swap channel, sub mean - im = im.astype('float32') - ImageTransformer.transformer(self, im) - return im - - def load_image_from_string(self, data): - flag = cv2.CV_LOAD_IMAGE_COLOR if self.is_color else cv2.CV_LOAD_IMAGE_GRAYSCALE - im = cv2.imdecode(np.fromstring(data, np.uint8), flag) - return im - - def transform_from_string(self, data): - im = self.load_image_from_string(data) - return self.transform(im) - - def load_image_from_file(self, file): - flag = cv2.CV_LOAD_IMAGE_COLOR if self.is_color else cv2.CV_LOAD_IMAGE_GRAYSCALE - im = cv2.imread(file, flag) - return im - - def transform_from_file(self, file): - im = self.load_image_from_file(file) - return self.transform(im) - - -class PILTransformer(ImageTransformer): - """ - PILTransformer used PIL to process image. - """ - - def __init__( - self, - min_size=None, - crop_size=None, - transpose=(2, 0, 1), # transpose to C * H * W - channel_swap=None, - mean=None, - is_train=True, - is_color=True): - ImageTransformer.__init__(self, transpose, channel_swap, mean, is_color) - self.min_size = min_size - self.crop_size = crop_size - self.is_train = is_train - - def resize(self, im, min_size): - row, col = im.size[:2] - new_row, new_col = min_size, min_size - if row > col: - new_row = min_size * row / col - else: - new_col = min_size * col / row - im = im.resize((new_row, new_col), Image.ANTIALIAS) - return im - - def crop_and_flip(self, im): - """ - Return cropped image. - The size of the cropped image is inner_size * inner_size. - """ - row, col = im.size[:2] - start_h, start_w = 0, 0 - if self.is_train: - start_h = np.random.randint(0, row - self.crop_size + 1) - start_w = np.random.randint(0, col - self.crop_size + 1) - else: - start_h = (row - self.crop_size) / 2 - start_w = (col - self.crop_size) / 2 - end_h, end_w = start_h + self.crop_size, start_w + self.crop_size - im = im.crop((start_h, start_w, end_h, end_w)) - if (self.is_train) and (np.random.randint(2) == 0): - im = im.transpose(Image.FLIP_LEFT_RIGHT) - return im - - def transform(self, im): - im = self.resize(im, self.min_size) - im = self.crop_and_flip(im) - im = np.array(im, dtype=np.float32) # convert to numpy.array - # transpose, swap channel, sub mean - ImageTransformer.transformer(self, im) - return im - - def load_image_from_string(self, data): - im = Image.open(StringIO(data)) - return im - - def transform_from_string(self, data): - im = self.load_image_from_string(data) - return self.transform(im) - - def load_image_from_file(self, file): - im = Image.open(file) - return im - - def transform_from_file(self, file): - im = self.load_image_from_file(file) - return self.transform(im) - - -def job(is_img_string, transformer, data_label_pack): - (data, label) = data_label_pack - if is_img_string: - return transformer.transform_from_string(data), label - else: - return transformer.transform_from_file(data), label - - -class MultiProcessImageTransformer(object): - def __init__(self, - procnum=10, - resize_size=None, - crop_size=None, - transpose=(2, 0, 1), - channel_swap=None, - mean=None, - is_train=True, - is_color=True, - is_img_string=True): - """ - Processing image with multi-process. If it is used in PyDataProvider, - the simple usage for CNN is as follows: - - .. code-block:: python - - def hool(settings, is_train, **kwargs): - settings.is_train = is_train - settings.mean_value = np.array([103.939,116.779,123.68], dtype=np.float32) - settings.input_types = [ - dense_vector(3 * 224 * 224), - integer_value(1)] - settings.transformer = MultiProcessImageTransformer( - procnum=10, - resize_size=256, - crop_size=224, - transpose=(2, 0, 1), - mean=settings.mean_values, - is_train=settings.is_train) - - - @provider(init_hook=hook, pool_size=20480) - def process(settings, file_list): - with open(file_list, 'r') as fdata: - for line in fdata: - data_dic = np.load(line.strip()) # load the data batch pickled by Pickle. - data = data_dic['data'] - labels = data_dic['label'] - labels = np.array(labels, dtype=np.float32) - for im, lab in settings.dp.run(data, labels): - yield [im.astype('float32'), int(lab)] - - :param procnum: processor number. - :type procnum: int - :param resize_size: the shorter edge size of image after resizing. - :type resize_size: int - :param crop_size: the croping size. - :type crop_size: int - :param transpose: the transpose order, Paddle only allow C * H * W order. - :type transpose: tuple or list - :param channel_swap: the channel swap order, RGB or BRG. - :type channel_swap: tuple or list - :param mean: the mean values of image, per-channel mean or element-wise mean. - :type mean: array, The dimension is 1 for per-channel mean. - The dimension is 3 for element-wise mean. - :param is_train: training peroid or testing peroid. - :type is_train: bool. - :param is_color: the image is color or gray. - :type is_color: bool. - :param is_img_string: The input can be the file name of image or image string. - :type is_img_string: bool. - """ - - self.procnum = procnum - self.pool = multiprocessing.Pool(procnum) - self.is_img_string = is_img_string - if cv2 is not None: - self.transformer = CvTransformer(resize_size, crop_size, transpose, - channel_swap, mean, is_train, - is_color) - else: - self.transformer = PILTransformer(resize_size, crop_size, transpose, - channel_swap, mean, is_train, - is_color) - - def run(self, data, label): - fun = functools.partial(job, self.is_img_string, self.transformer) - return self.pool.imap_unordered( - fun, six.moves.zip(data, label), chunksize=100 * self.procnum) diff --git a/python/paddle/utils/make_model_diagram.py b/python/paddle/utils/make_model_diagram.py deleted file mode 100644 index 52759d3ad230c3a5a5488a8bc46a2e8f8fae1025..0000000000000000000000000000000000000000 --- a/python/paddle/utils/make_model_diagram.py +++ /dev/null @@ -1,140 +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. - -# Generate dot diagram file for the given paddle model config -# The generated file can be viewed using Graphviz (http://graphviz.org) - -from __future__ import print_function - -import six -import sys -import traceback - -from paddle.trainer.config_parser import parse_config - - -def make_layer_label(layer_config): - label = '%s type=%s' % (layer_config.name, layer_config.type) - if layer_config.reversed: - label += ' <==' - - label2 = '' - if layer_config.active_type: - label2 += 'act=%s ' % layer_config.active_type - if layer_config.bias_parameter_name: - label2 += 'bias=%s ' % layer_config.bias_parameter_name - - if label2: - label += '\l' + label2 - return label - - -def make_diagram(config_file, dot_file, config_arg_str): - config = parse_config(config_file, config_arg_str) - make_diagram_from_proto(config.model_config, dot_file) - - -def make_diagram_from_proto(model_config, dot_file): - # print >> sys.stderr, config - name2id = {} - f = open(dot_file, 'w') - submodel_layers = set() - - def make_link(link): - return 'l%s -> l%s;' % (name2id[link.layer_name], - name2id[link.link_name]) - - def make_mem(mem): - s = '' - if mem.boot_layer_name: - s += 'l%s -> l%s;\n' % (name2id[mem.boot_layer_name], - name2id[mem.layer_name]) - s += 'l%s -> l%s [style=dashed];' % (name2id[mem.layer_name], - name2id[mem.link_name]) - return s - - print('digraph graphname {', file=f) - print('node [width=0.375,height=0.25];', file=f) - for i in six.moves.xrange(len(model_config.layers)): - l = model_config.layers[i] - name2id[l.name] = i - - i = 0 - for sub_model in model_config.sub_models: - if sub_model.name == 'root': - continue - print('subgraph cluster_%s {' % i, file=f) - print('style=dashed;', file=f) - label = '%s ' % sub_model.name - if sub_model.reversed: - label += '<==' - print('label = "%s";' % label, file=f) - i += 1 - submodel_layers.add(sub_model.name) - for layer_name in sub_model.layer_names: - submodel_layers.add(layer_name) - lid = name2id[layer_name] - layer_config = model_config.layers[lid] - label = make_layer_label(layer_config) - print('l%s [label="%s", shape=box];' % (lid, label), file=f) - print('}', file=f) - - for i in six.moves.xrange(len(model_config.layers)): - l = model_config.layers[i] - if l.name not in submodel_layers: - label = make_layer_label(l) - print('l%s [label="%s", shape=box];' % (i, label), file=f) - - for sub_model in model_config.sub_models: - if sub_model.name == 'root': - continue - for link in sub_model.in_links: - print(make_link(link), file=f) - for link in sub_model.out_links: - print(make_link(link), file=f) - for mem in sub_model.memories: - print(make_mem(mem), file=f) - - for i in six.moves.xrange(len(model_config.layers)): - for l in model_config.layers[i].inputs: - print( - 'l%s -> l%s [label="%s"];' % (name2id[l.input_layer_name], i, - l.input_parameter_name), - file=f) - - print('}', file=f) - f.close() - - -def usage(): - print( - ("Usage: python show_model_diagram.py" + - " CONFIG_FILE DOT_FILE [config_str]"), - file=sys.stderr) - exit(1) - - -if __name__ == '__main__': - if len(sys.argv) < 3 or len(sys.argv) > 4: - usage() - - config_file = sys.argv[1] - dot_file = sys.argv[2] - config_arg_str = sys.argv[3] if len(sys.argv) == 4 else '' - - try: - make_diagram(config_file, dot_file, config_arg_str) - except: - traceback.print_exc() - raise diff --git a/python/paddle/utils/merge_model.py b/python/paddle/utils/merge_model.py deleted file mode 100644 index b74649e93640c3600636034d58792b8d12dffeda..0000000000000000000000000000000000000000 --- a/python/paddle/utils/merge_model.py +++ /dev/null @@ -1,73 +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. - -import gzip -import struct -import os - -from paddle.trainer_config_helpers.layers import LayerOutput -from paddle.v2.parameters import Parameters -from paddle.proto import ModelConfig_pb2 -from paddle.v2.topology import Topology - - -def merge_v2_model(net, param_file, output_file): - '''Merge the model config and parameters into one file. - - The model configuration file describes the model structure which - ends with .py. The parameters file stores the parameters of the model - which ends with .tar.gz. - - @param net The output layer of the network for inference. - @param param_file Path of the parameters (.tar.gz) which is stored by - v2 api. - @param output_file Path of the merged file which will be generated. - - Usage: - - from paddle.utils.merge_model import merge_v2_model - # import your network configuration - from example_net import net_conf - - net = net_conf(is_predict=True) - param_file = './param_pass_00000.tar.gz' - output_file = './output.paddle' - - merge_v2_model(net, param_file, output_file) - - ''' - - assert isinstance(net, LayerOutput), \ - "The net should be the output of the network for inference" - assert os.path.exists(param_file), \ - "The model parameters file %s does not exists " % (param_file) - - model_proto = Topology(net).proto() - assert isinstance(model_proto, ModelConfig_pb2.ModelConfig) - - with gzip.open(param_file) as f: - params = Parameters.from_tar(f) - - if os.path.exists(output_file): - os.remove(output_file) - - with open(output_file, 'w') as f: - param_names = [param.name for param in model_proto.parameters] - conf_str = model_proto.SerializeToString() - f.write(struct.pack('q', len(conf_str))) - f.write(conf_str) - for pname in param_names: - params.serialize(pname, f) - - print('Generate %s success!' % (output_file)) diff --git a/python/paddle/utils/predefined_net.py b/python/paddle/utils/predefined_net.py deleted file mode 100644 index 2801f4877c079615239b92be146b3e33df16b37f..0000000000000000000000000000000000000000 --- a/python/paddle/utils/predefined_net.py +++ /dev/null @@ -1,381 +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. - -import numpy as np -import six -import os -from paddle.trainer.config_parser import * -from paddle.utils.preprocess_img import \ - ImageClassificationDatasetCreater -from paddle.trainer_config_helpers import * - - -def image_data(data_dir, - processed_image_size, - overwrite=False, - color=True, - train_list="batches/train.list", - test_list="batches/test.list", - meta_file="batches/batches.meta", - use_jpeg=1): - """ - Predefined image data provider for image classification. - train_list: a text file containing a list of training batches. - test_list: a text file containing a list of test batches. - processed_image_size: all the input images will be resized into this size. - If the image is not square. Then the shorter edge will be resized into - this size, and the aspect ratio is kept the same. - color: whether the images are color or gray. - meta_path: the path of the meta file that stores the mean image file and - other dataset information, such as the size of images, - the size of the mean image, the number of classes. - async_load_data: whether to load image data asynchronuously. - """ - data_creator = ImageClassificationDatasetCreater( - data_dir, processed_image_size, color) - batch_data_dir = data_dir - train_list = os.path.join(batch_data_dir, train_list) - test_list = os.path.join(batch_data_dir, test_list) - meta_path = os.path.join(batch_data_dir, meta_file) - image_size = processed_image_size - conf = np.load(meta_path) - mean_image_size = conf["mean_image_size"] - is_color = conf["color"] - num_classes = conf["num_classes"] - color_string = "color" if is_color else "gray" - - args = { - 'meta': meta_path, - 'mean_img_size': mean_image_size, - 'img_size': image_size, - 'num_classes': num_classes, - 'use_jpeg': use_jpeg != 0, - 'color': color_string - } - - define_py_data_sources2( - train_list, - test_list, - module='image_provider', - obj='processData', - args=args) - return { - "image_size": image_size, - "num_classes": num_classes, - "is_color": is_color - } - - -def get_extra_layer_attr(drop_rate): - if drop_rate == 0: - return None - else: - return ExtraLayerAttribute(drop_rate=drop_rate) - - -def image_data_layers(image_size, num_classes, is_color=False, - is_predict=False): - """ - Data layers for image classification. - image_size: image size. - num_classes: num of classes. - is_color: whether the input images are color. - is_predict: whether the network is used for prediction. - """ - num_image_channels = 3 if is_color else 1 - data_input = data_layer("input", - image_size * image_size * num_image_channels) - if is_predict: - return data_input, None, num_image_channels - else: - label_input = data_layer("label", 1) - return data_input, label_input, num_image_channels - - -def simple_conv_net(data_conf, is_color=False): - """ - A Wrapper for a simple network for MNIST digit recognition. - It contains two convolutional layers, one fully conencted layer, and - one softmax layer. - data_conf is a dictionary with the following keys: - image_size: image size. - num_classes: num of classes. - is_color: whether the input images are color. - """ - for k, v in six.iteritems(data_conf): - globals()[k] = v - data_input, label_input, num_image_channels = \ - image_data_layers(image_size, num_classes, is_color, is_predict) - filter_sizes = [5, 5] - num_channels = [32, 64] - strides = [1, 1] - fc_dims = [500] - conv_bn_pool1 = img_conv_bn_pool( - name="g1", - input=data_input, - filter_size=filter_sizes[0], - num_channel=num_image_channels, - num_filters=num_channels[0], - conv_stride=1, - conv_padding=0, - pool_size=3, - pool_stride=2, - act=ReluActivation()) - conv_bn_pool2 = img_conv_bn_pool( - name="g2", - input=conv_bn_pool1, - filter_size=filter_sizes[1], - num_channel=num_channels[0], - num_filters=num_channels[1], - conv_stride=1, - conv_padding=0, - pool_size=3, - pool_stride=2, - act=ReluActivation()) - fc3 = fc_layer( - name="fc3", input=conv_bn_pool2, dim=fc_dims[0], act=ReluActivation()) - fc3_dropped = dropout_layer(name="fc3_dropped", input=fc3, dropout_rate=0.5) - output = fc_layer( - name="output", - input=fc3_dropped, - dim=fc_dims[0], - act=SoftmaxActivation()) - if is_predict: - end_of_network(output) - else: - cost = classify(name="cost", input=output, label=label_input) - end_of_network(cost) - - -def conv_layer_group(prefix_num, - num_layers, - input, - input_channels, - output_channels, - drop_rates=[], - strides=[], - with_bn=[]): - """ - A set of convolution layers, and batch normalization layers, - followed by one pooling layer. - It is utilized in VGG network for image classifcation. - prefix_num: the prefix number of the layer names. - For example, if prefix_num = 1, the first convolutioal layer's - name will be conv_1_1. - num_layers: number of the convolutional layers. - input: the name of the input layer. - input_channels: the number of channels of the input feature map. - output_channels: the number of channels of the output feature map. - drop_rates: the drop rates of the BN layers. It will be all zero by default. - strides: the stride of the convolution for the layers. - It will be all 1 by default. - with_bn: whether to use Batch Normalization for Conv layers. - By default, it is all false. - """ - if len(drop_rates) == 0: drop_rates = [0] * num_layers - if len(strides) == 0: strides = [1] * num_layers - if len(with_bn) == 0: with_bn = [False] * num_layers - assert (len(drop_rates) == num_layers) - assert (len(strides) == num_layers) - - for i in range(1, num_layers + 1): - if i == 1: - i_conv_in = input - else: - i_conv_in = group_output - i_channels_conv = input_channels if i == 1 else output_channels - conv_act = LinearActivation() if with_bn[i - 1] else ReluActivation() - conv_output = img_conv_layer( - name="conv%d_%d" % (prefix_num, i), - input=i_conv_in, - filter_size=3, - num_channels=i_channels_conv, - num_filters=output_channels, - stride=strides[i - 1], - padding=1, - act=conv_act) - if with_bn[i - 1]: - bn = batch_norm_layer( - name="conv%d_%d_bn" % (prefix_num, i), - input=conv_output, - num_channels=output_channels, - act=ReluActivation(), - layer_attr=get_extra_layer_attr(drop_rate=drop_rates[i - 1])) - group_output = bn - else: - group_output = conv_output - pool = img_pool_layer( - name="pool%d" % prefix_num, - input=group_output, - pool_size=2, - num_channels=output_channels, - stride=2) - return pool - - -def vgg_conv_net(image_size, - num_classes, - num_layers, - channels, - strides, - with_bn, - fc_dims, - drop_rates, - drop_rates_fc=[], - is_color=True, - is_predict=False): - """ - A Wrapper for a VGG network for image classification. - It is a set of convolutional groups followed by several fully - connected layers, and a cross-entropy classifiation loss. - The detailed architecture of the paper can be found here: - Very Deep Convolutional Networks for Large-Scale Visual Recognition - http://www.robots.ox.ac.uk/~vgg/research/very_deep/ - image_size: image size. - num_classes: num of classes. - num_layers: the number of layers for all the convolution groups. - channels: the number of output filters for all the convolution groups. - with_bn: whether each layer of a convolution group is followed by a - batch normalization. - drop_rates: the dropout rates for all the convolutional layers. - fc_dims: the dimension for all the fully connected layers. - is_color: whether the input images are color. - """ - data_input, label_input, num_image_channels = \ - image_data_layers(image_size, num_classes, is_color, is_predict) - assert (len(num_layers) == len(channels)) - assert (len(num_layers) == len(strides)) - assert (len(num_layers) == len(with_bn)) - num_fc_layers = len(fc_dims) - assert (num_fc_layers + 1 == len(drop_rates_fc)) - - for i in range(len(num_layers)): - input_layer = data_input if i == 0 else group_output - input_channels = 3 if i == 0 else channels[i - 1] - group_output = conv_layer_group( - prefix_num=i + 1, - num_layers=num_layers[i], - input=input_layer, - input_channels=input_channels, - output_channels=channels[i], - drop_rates=drop_rates[i], - strides=strides[i], - with_bn=with_bn[i]) - conv_output_name = group_output - if drop_rates_fc[0] != 0.0: - dropped_pool_name = "pool_dropped" - conv_output_name = dropout_layer( - name=dropped_pool_name, - input=conv_output_name, - dropout_rate=drop_rates_fc[0]) - for i in range(len(fc_dims)): - input_layer_name = conv_output_name if i == 0 else fc_output - active_type = LinearActivation() if i == len( - fc_dims) - 1 else ReluActivation() - drop_rate = 0.0 if i == len(fc_dims) - 1 else drop_rates_fc[i + 1] - fc_output = fc_layer( - name="fc%d" % (i + 1), - input=input_layer_name, - size=fc_dims[i], - act=active_type, - layer_attr=get_extra_layer_attr(drop_rate)) - bn = batch_norm_layer( - name="fc_bn", - input=fc_output, - num_channels=fc_dims[len(fc_dims) - 1], - act=ReluActivation(), - layer_attr=get_extra_layer_attr(drop_rate=drop_rates_fc[-1])) - output = fc_layer( - name="output", input=bn, size=num_classes, act=SoftmaxActivation()) - if is_predict: - outputs(output) - else: - cost = classification_cost(name="cost", input=output, label=label_input) - outputs(cost) - - -def vgg16_conv_net(image_size, num_classes, is_color=True, is_predict=False): - """ - A Wrapper for a 16 layers VGG network for image classification. - The detailed architecture of the paper can be found here: - Very Deep Convolutional Networks for Large-Scale Visual Recognition - http://www.robots.ox.ac.uk/~vgg/research/very_deep/ - image_size: image size. - num_classes: num of classes. - is_color: whether the input images are color. - """ - vgg_conv_net(image_size, num_classes, - num_layers=[2, 2, 3, 3, 3], - channels=[64, 128, 256, 512, 512], - strides=[[], [], [], [], []], - with_bn=[[False, True], [False, True], [False, False, True], \ - [False, False, True], [False, False, True]], - drop_rates=[[]] * 5, - drop_rates_fc=[0.0, 0.5, 0.5], - fc_dims=[4096, 4096], - is_predict=is_predict) - - -def small_vgg(data_conf, is_predict=False): - """ - A Wrapper for a small VGG network for CIFAR-10 image classification. - The detailed architecture of the paper can be found here: - 92.45% on CIFAR-10 in Torch - http://torch.ch/blog/2015/07/30/cifar.html - Due to the constraints of CuDNN, it only has four convolutional groups - rather than five. - Thus, it only achieves 91.2% test accuracy and 98.1% training accuracy. - data_conf is a dictionary with the following keys: - image_size: image size. - num_classes: num of classes. - is_color: whether the input images are color. - """ - for k, v in six.iteritems(data_conf): - globals()[k] = v - vgg_conv_net(image_size, num_classes, - num_layers=[2, 2, 3, 3], - channels=[64, 128, 256, 512], - strides=[[], [], [], []], - with_bn=[[True, True], [True, True], [True, True, True], \ - [True, True, True]], - drop_rates=[[0.3, 0.0], [0.4, 0.0], - [0.4, 0.4, 0.0], [0.4, 0.4, 0.0]], - drop_rates_fc=[0.5, 0.5], - fc_dims=[512], - is_predict=is_predict) - - -def training_settings(learning_rate=0.1, - batch_size=128, - algorithm="sgd", - momentum=0.9, - decay_rate=0.001): - """ - Training settings. - learning_rate: learning rate of the training. - batch_size: the size of each training batch. - algorithm: training algorithm, can be - - sgd - - adagrad - - adadelta - - rmsprop - momentum: momentum of the training algorithm. - decay_rate: weight decay rate. - """ - Settings( - algorithm=algorithm, - batch_size=batch_size, - learning_rate=learning_rate / float(batch_size)) - default_momentum(momentum) - default_decay_rate(decay_rate * batch_size) diff --git a/python/requirements.txt b/python/requirements.txt index 03d5e33e88cd5f1138ca8f6a6e885d6acfbc260e..36bd5d4261cc7aa78d26b8c8ddfd87abd4f4e2e2 100644 --- a/python/requirements.txt +++ b/python/requirements.txt @@ -1,6 +1,6 @@ requests==2.9.2 numpy>=1.12 -protobuf==3.1 +protobuf>=3.1.0 recordio>=0.1.0 matplotlib==2.2.3 # TODO: let python3 paddlepaddle package use latest matplotlib rarfile @@ -11,3 +11,4 @@ graphviz six funcsigs pyyaml +decorator diff --git a/python/setup.py.in b/python/setup.py.in index c947785cbf7517be56c3e43120db65284ab22d10..a7c1e91f9c3a9597d799659a0abe3c9f56e54a57 100644 --- a/python/setup.py.in +++ b/python/setup.py.in @@ -100,6 +100,7 @@ packages=['paddle', 'paddle.utils', 'paddle.dataset', 'paddle.reader', + 'paddle.distributed', 'paddle.fluid', 'paddle.fluid.imperative', 'paddle.fluid.proto', @@ -109,6 +110,7 @@ packages=['paddle', 'paddle.fluid.contrib', 'paddle.fluid.contrib.decoder', 'paddle.fluid.contrib.quantize', + 'paddle.fluid.contrib.int8_inference', 'paddle.fluid.contrib.reader', 'paddle.fluid.contrib.slim', 'paddle.fluid.contrib.slim.core', diff --git a/tools/manylinux1/Dockerfile.x64 b/tools/manylinux1/Dockerfile.x64 index 48fd145e5fe6735fca3096752f801b1ec1cb39f0..c2fd743f62f536ab7443ca215d100478021d8f7c 100644 --- a/tools/manylinux1/Dockerfile.x64 +++ b/tools/manylinux1/Dockerfile.x64 @@ -31,10 +31,10 @@ RUN wget --no-check-certificate -qO- https://storage.googleapis.com/golang/go1.8 ENV GOROOT=/usr/local/go GOPATH=/root/gopath ENV PATH=${GOROOT}/bin:${GOPATH}/bin:${PATH} -# protobuf 3.1.0 -RUN cd /opt && wget -q --no-check-certificate https://github.com/google/protobuf/releases/download/v3.1.0/protobuf-cpp-3.1.0.tar.gz && \ - tar xzf protobuf-cpp-3.1.0.tar.gz && \ - cd protobuf-3.1.0 && ./configure && make -j4 && make install && cd .. && rm -f protobuf-cpp-3.1.0.tar.gz +# protobuf 3.6.1 +RUN cd /opt && wget -q --no-check-certificate https://github.com/google/protobuf/releases/download/v3.6.1/protobuf-cpp-3.6.1.tar.gz && \ + tar xzf protobuf-cpp-3.6.1.tar.gz && \ + cd protobuf-3.6.1 && ./configure && make -j4 && make install && cd .. && rm -f protobuf-cpp-3.6.1.tar.gz RUN wget https://raw.githubusercontent.com/PaddlePaddle/Paddle/develop/python/requirements.txt -O /root/requirements.txt diff --git a/tools/manylinux1/build_all.sh b/tools/manylinux1/build_all.sh index 097bedb5265d00f8aa362bb0272af633c97192ba..caf21722158b749ffe8d026a98a8b7d015e555d8 100755 --- a/tools/manylinux1/build_all.sh +++ b/tools/manylinux1/build_all.sh @@ -24,3 +24,8 @@ sed 's//9.0-cudnn7-devel-centos6/g' Dockerfile.x64 | \ sed 's//NVCC_GENCODE="-gencode=arch=compute_35,code=sm_35 -gencode=arch=compute_50,code=sm_50 -gencode=arch=compute_52,code=sm_52 -gencode=arch=compute_60,code=sm_60 -gencode=arch=compute_60,code=compute_60 -gencode=arch=compute_61,code=sm_61 -gencode=arch=compute_62,code=sm_62 -gencode=arch=compute_70,code=sm_70"/g'> Dockerfile.tmp docker build -t ${REPO}/paddle_manylinux_devel:cuda9.0_cudnn7 -f Dockerfile.tmp . docker push ${REPO}/paddle_manylinux_devel:cuda9.0_cudnn7 + +sed 's//10.0-devel-centos6/g' Dockerfile.x64 | \ +sed 's//NVCC_GENCODE="-gencode=arch=compute_35,code=sm_35 -gencode=arch=compute_50,code=sm_50 -gencode=arch=compute_52,code=sm_52 -gencode=arch=compute_60,code=sm_60 -gencode=arch=compute_60,code=compute_60 -gencode=arch=compute_61,code=sm_61 -gencode=arch=compute_62,code=sm_62 -gencode=arch=compute_70,code=sm_70 -gencode=arch=compute_75,code=sm_75"/g'> Dockerfile.tmp +docker build -t ${REPO}/paddle_manylinux_devel:cuda10.0_cudnn7 -f Dockerfile.tmp . +docker push ${REPO}/paddle_manylinux_devel:cuda10.0_cudnn7 diff --git a/tools/manylinux1/build_scripts/build.sh b/tools/manylinux1/build_scripts/build.sh index 6c551eceb4543bf33229b9e5b5124522f3ee134c..1b0059a8c69fca93ecbf1db570a6092ca5c908b1 100644 --- a/tools/manylinux1/build_scripts/build.sh +++ b/tools/manylinux1/build_scripts/build.sh @@ -17,7 +17,7 @@ OPENSSL_ROOT=openssl-1.1.0i OPENSSL_HASH=ebbfc844a8c8cc0ea5dc10b86c9ce97f401837f3fa08c17b2cdadc118253cf99 EPEL_RPM_HASH=e5ed9ecf22d0c4279e92075a64c757ad2b38049bcf5c16c4f2b75d5f6860dc0d DEVTOOLS_HASH=a8ebeb4bed624700f727179e6ef771dafe47651131a00a78b342251415646acc -PATCHELF_HASH=d9afdff4baeacfbc64861454f368b7f2c15c44d245293f7587bbf726bfe722fb +PATCHELF_HASH=f2aa40a6148cb3b0ca807a1bf836b081793e55ec9e5540a5356d800132be7e0a CURL_ROOT=curl-7.49.1 CURL_HASH=eb63cec4bef692eab9db459033f409533e6d10e20942f4b060b32819e81885f1 AUTOCONF_ROOT=autoconf-2.69 @@ -107,11 +107,13 @@ curl-config --features rm -rf /usr/local/ssl # Install patchelf (latest with unreleased bug fixes) -curl -sLO http://nipy.bic.berkeley.edu/manylinux/patchelf-0.9njs2.tar.gz -check_sha256sum patchelf-0.9njs2.tar.gz $PATCHELF_HASH -tar -xzf patchelf-0.9njs2.tar.gz -(cd patchelf-0.9njs2 && ./configure && make && make install) -rm -rf patchelf-0.9njs2.tar.gz patchelf-0.9njs2 +# FIXME(typhoonzero): restore this when the link is fixed. +# curl -sLO http://nipy.bic.berkeley.edu/manylinux/patchelf-0.9njs2.tar.gz +# check_sha256sum patchelf-0.9njs2.tar.gz $PATCHELF_HASH +# tar -xzf patchelf-0.9njs2.tar.gz +# (cd patchelf-0.9njs2 && ./configure && make && make install) +# rm -rf patchelf-0.9njs2.tar.gz patchelf-0.9njs2 +yum install -y patchelf # Install latest pypi release of auditwheel LD_LIBRARY_PATH="${ORIGINAL_LD_LIBRARY_PATH}:$(dirname ${PY35_BIN})/lib" $PY35_BIN/pip install auditwheel diff --git a/tools/manylinux1/build_scripts/build_utils.sh b/tools/manylinux1/build_scripts/build_utils.sh index 48cce15a145138376177731009c61157d1d4d0c8..083101249cd8560f63c95b3fe2aef610b01dd6ac 100755 --- a/tools/manylinux1/build_scripts/build_utils.sh +++ b/tools/manylinux1/build_scripts/build_utils.sh @@ -87,6 +87,8 @@ function do_cpython_build { # NOTE Make libpython shared library visible to python calls below LD_LIBRARY_PATH="${prefix}/lib" ${prefix}/bin/python get-pip.py LD_LIBRARY_PATH="${prefix}/lib" ${prefix}/bin/pip install wheel + cd / + ls ${MY_DIR} local abi_tag=$(LD_LIBRARY_PATH="${prefix}/lib" ${prefix}/bin/python ${MY_DIR}/python-tag-abi-tag.py) ln -s ${prefix} /opt/python/${abi_tag} } diff --git a/tools/timeline.py b/tools/timeline.py index f850476831d84787bf5cc7c7f7c91ff9dd6a2d5b..ebadb29bdbe00caeb3fb16a95b7dde6f418db155 100644 --- a/tools/timeline.py +++ b/tools/timeline.py @@ -131,8 +131,12 @@ class Timeline(object): if (k, event.device_id, "CPU") not in self._devices: pid = self._allocate_pid() self._devices[(k, event.device_id, "CPU")] = pid - self._chrome_trace.emit_pid("%s:cpu:block:%d" % - (k, event.device_id), pid) + # -1 device id represents CUDA api call + if event.device_id == -1: + self._chrome_trace.emit_pid("%s:cuda_api" % k, pid) + else: + self._chrome_trace.emit_pid( + "%s:cpu:block:%d" % (k, event.device_id), pid) elif event.type == profiler_pb2.Event.GPUKernel: if (k, event.device_id, "GPUKernel") not in self._devices: pid = self._allocate_pid() @@ -150,7 +154,9 @@ class Timeline(object): pid = self._devices[(k, event.device_id, type)] args = {'name': event.name} if event.memcopy.bytes > 0: - args = {'mem_bytes': event.memcopy.bytes} + args['mem_bytes'] = event.memcopy.bytes + if event.detail_info: + args['detail_info'] = event.detail_info # TODO(panyx0718): Chrome tracing only handles ms. However, some # ops takes micro-seconds. Hence, we keep the ns here. self._chrome_trace.emit_region( @@ -173,7 +179,7 @@ if args.timeline_path: profile_paths = profile_path.split(',') profile_dict = dict() if len(profile_paths) == 1: - with open(profile_path, 'r') as f: + with open(profile_path, 'rb') as f: profile_s = f.read() profile_pb = profiler_pb2.Profile() profile_pb.ParseFromString(profile_s) @@ -181,7 +187,7 @@ if len(profile_paths) == 1: else: for profile_path in profile_paths: k, v = profile_path.split('=') - with open(v, 'r') as f: + with open(v, 'rb') as f: profile_s = f.read() profile_pb = profiler_pb2.Profile() profile_pb.ParseFromString(profile_s)